Debian Codegraph @

llvm-project/clang/lib/CodeGen/CGBuiltin.cpp

raw
1 //===---- CGBuiltin.cpp - Emit LLVM Code for builtins ---------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This contains code to emit Builtin calls as LLVM code.
10 //
11 //===----------------------------------------------------------------------===//
12
13 #include "ABIInfo.h"
14 #include "CGCUDARuntime.h"
15 #include "CGCXXABI.h"
16 #include "CGObjCRuntime.h"
17 #include "CGOpenCLRuntime.h"
18 #include "CGRecordLayout.h"
19 #include "CodeGenFunction.h"
20 #include "CodeGenModule.h"
21 #include "ConstantEmitter.h"
22 #include "PatternInit.h"
23 #include "TargetInfo.h"
24 #include "clang/AST/ASTContext.h"
25 #include "clang/AST/Attr.h"
26 #include "clang/AST/Decl.h"
27 #include "clang/AST/OSLog.h"
28 #include "clang/Basic/TargetBuiltins.h"
29 #include "clang/Basic/TargetInfo.h"
30 #include "clang/CodeGen/CGFunctionInfo.h"
31 #include "llvm/ADT/APFloat.h"
32 #include "llvm/ADT/APInt.h"
33 #include "llvm/ADT/SmallPtrSet.h"
34 #include "llvm/ADT/StringExtras.h"
35 #include "llvm/Analysis/ValueTracking.h"
36 #include "llvm/IR/DataLayout.h"
37 #include "llvm/IR/InlineAsm.h"
38 #include "llvm/IR/Intrinsics.h"
39 #include "llvm/IR/IntrinsicsAArch64.h"
40 #include "llvm/IR/IntrinsicsAMDGPU.h"
41 #include "llvm/IR/IntrinsicsARM.h"
42 #include "llvm/IR/IntrinsicsBPF.h"
43 #include "llvm/IR/IntrinsicsHexagon.h"
44 #include "llvm/IR/IntrinsicsLoongArch.h"
45 #include "llvm/IR/IntrinsicsNVPTX.h"
46 #include "llvm/IR/IntrinsicsPowerPC.h"
47 #include "llvm/IR/IntrinsicsR600.h"
48 #include "llvm/IR/IntrinsicsRISCV.h"
49 #include "llvm/IR/IntrinsicsS390.h"
50 #include "llvm/IR/IntrinsicsVE.h"
51 #include "llvm/IR/IntrinsicsWebAssembly.h"
52 #include "llvm/IR/IntrinsicsX86.h"
53 #include "llvm/IR/MDBuilder.h"
54 #include "llvm/IR/MatrixBuilder.h"
55 #include "llvm/Support/AArch64TargetParser.h"
56 #include "llvm/Support/ConvertUTF.h"
57 #include "llvm/Support/ScopedPrinter.h"
58 #include "llvm/Support/X86TargetParser.h"
59 #include <optional>
60 #include <sstream>
61
62 using namespace clang;
63 using namespace CodeGen;
64 using namespace llvm;
65
66 static void initializeAlloca(CodeGenFunction &CGF, AllocaInst *AI, Value *Size,
67 Align AlignmentInBytes) {
68 ConstantInt *Byte;
69 switch (CGF.getLangOpts().getTrivialAutoVarInit()) {
70 case LangOptions::TrivialAutoVarInitKind::Uninitialized:
71 // Nothing to initialize.
72 return;
73 case LangOptions::TrivialAutoVarInitKind::Zero:
74 Byte = CGF.Builder.getInt8(0x00);
75 break;
76 case LangOptions::TrivialAutoVarInitKind::Pattern: {
77 llvm::Type *Int8 = llvm::IntegerType::getInt8Ty(CGF.CGM.getLLVMContext());
78 Byte = llvm::dyn_cast<llvm::ConstantInt>(
79 initializationPatternFor(CGF.CGM, Int8));
80 break;
81 }
82 }
83 if (CGF.CGM.stopAutoInit())
84 return;
85 auto *I = CGF.Builder.CreateMemSet(AI, Byte, Size, AlignmentInBytes);
86 I->addAnnotationMetadata("auto-init");
87 }
88
89 /// getBuiltinLibFunction - Given a builtin id for a function like
90 /// "__builtin_fabsf", return a Function* for "fabsf".
91 llvm::Constant *CodeGenModule::getBuiltinLibFunction(const FunctionDecl *FD,
92 unsigned BuiltinID) {
93 assert(Context.BuiltinInfo.isLibFunction(BuiltinID));
94
95 // Get the name, skip over the __builtin_ prefix (if necessary).
96 StringRef Name;
97 GlobalDecl D(FD);
98
99 // TODO: This list should be expanded or refactored after all GCC-compatible
100 // std libcall builtins are implemented.
101 static SmallDenseMap<unsigned, StringRef, 8> F128Builtins{
102 {Builtin::BI__builtin_printf, "__printfieee128"},
103 {Builtin::BI__builtin_vsnprintf, "__vsnprintfieee128"},
104 {Builtin::BI__builtin_vsprintf, "__vsprintfieee128"},
105 {Builtin::BI__builtin_sprintf, "__sprintfieee128"},
106 {Builtin::BI__builtin_snprintf, "__snprintfieee128"},
107 {Builtin::BI__builtin_fprintf, "__fprintfieee128"},
108 {Builtin::BI__builtin_nexttowardf128, "__nexttowardieee128"},
109 };
110
111 // The AIX library functions frexpl, ldexpl, and modfl are for 128-bit
112 // IBM 'long double' (i.e. __ibm128). Map to the 'double' versions
113 // if it is 64-bit 'long double' mode.
114 static SmallDenseMap<unsigned, StringRef, 4> AIXLongDouble64Builtins{
115 {Builtin::BI__builtin_frexpl, "frexp"},
116 {Builtin::BI__builtin_ldexpl, "ldexp"},
117 {Builtin::BI__builtin_modfl, "modf"},
118 };
119
120 // If the builtin has been declared explicitly with an assembler label,
121 // use the mangled name. This differs from the plain label on platforms
122 // that prefix labels.
123 if (FD->hasAttr<AsmLabelAttr>())
124 Name = getMangledName(D);
125 else {
126 // TODO: This mutation should also be applied to other targets other than
127 // PPC, after backend supports IEEE 128-bit style libcalls.
128 if (getTriple().isPPC64() &&
129 &getTarget().getLongDoubleFormat() == &llvm::APFloat::IEEEquad() &&
130 F128Builtins.find(BuiltinID) != F128Builtins.end())
131 Name = F128Builtins[BuiltinID];
132 else if (getTriple().isOSAIX() &&
133 &getTarget().getLongDoubleFormat() ==
134 &llvm::APFloat::IEEEdouble() &&
135 AIXLongDouble64Builtins.find(BuiltinID) !=
136 AIXLongDouble64Builtins.end())
137 Name = AIXLongDouble64Builtins[BuiltinID];
138 else
139 Name = Context.BuiltinInfo.getName(BuiltinID).substr(10);
140 }
141
142 llvm::FunctionType *Ty =
143 cast<llvm::FunctionType>(getTypes().ConvertType(FD->getType()));
144
145 return GetOrCreateLLVMFunction(Name, Ty, D, /*ForVTable=*/false);
146 }
147
148 /// Emit the conversions required to turn the given value into an
149 /// integer of the given size.
150 static Value *EmitToInt(CodeGenFunction &CGF, llvm::Value *V,
151 QualType T, llvm::IntegerType *IntType) {
152 V = CGF.EmitToMemory(V, T);
153
154 if (V->getType()->isPointerTy())
155 return CGF.Builder.CreatePtrToInt(V, IntType);
156
157 assert(V->getType() == IntType);
158 return V;
159 }
160
161 static Value *EmitFromInt(CodeGenFunction &CGF, llvm::Value *V,
162 QualType T, llvm::Type *ResultType) {
163 V = CGF.EmitFromMemory(V, T);
164
165 if (ResultType->isPointerTy())
166 return CGF.Builder.CreateIntToPtr(V, ResultType);
167
168 assert(V->getType() == ResultType);
169 return V;
170 }
171
172 /// Utility to insert an atomic instruction based on Intrinsic::ID
173 /// and the expression node.
174 static Value *MakeBinaryAtomicValue(
175 CodeGenFunction &CGF, llvm::AtomicRMWInst::BinOp Kind, const CallExpr *E,
176 AtomicOrdering Ordering = AtomicOrdering::SequentiallyConsistent) {
177
178 QualType T = E->getType();
179 assert(E->getArg(0)->getType()->isPointerType());
180 assert(CGF.getContext().hasSameUnqualifiedType(T,
181 E->getArg(0)->getType()->getPointeeType()));
182 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType()));
183
184 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0));
185 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace();
186
187 llvm::IntegerType *IntType =
188 llvm::IntegerType::get(CGF.getLLVMContext(),
189 CGF.getContext().getTypeSize(T));
190 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace);
191
192 llvm::Value *Args[2];
193 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType);
194 Args[1] = CGF.EmitScalarExpr(E->getArg(1));
195 llvm::Type *ValueType = Args[1]->getType();
196 Args[1] = EmitToInt(CGF, Args[1], T, IntType);
197
198 llvm::Value *Result = CGF.Builder.CreateAtomicRMW(
199 Kind, Args[0], Args[1], Ordering);
200 return EmitFromInt(CGF, Result, T, ValueType);
201 }
202
203 static Value *EmitNontemporalStore(CodeGenFunction &CGF, const CallExpr *E) {
204 Value *Val = CGF.EmitScalarExpr(E->getArg(0));
205 Value *Address = CGF.EmitScalarExpr(E->getArg(1));
206
207 // Convert the type of the pointer to a pointer to the stored type.
208 Val = CGF.EmitToMemory(Val, E->getArg(0)->getType());
209 unsigned SrcAddrSpace = Address->getType()->getPointerAddressSpace();
210 Value *BC = CGF.Builder.CreateBitCast(
211 Address, llvm::PointerType::get(Val->getType(), SrcAddrSpace), "cast");
212 LValue LV = CGF.MakeNaturalAlignAddrLValue(BC, E->getArg(0)->getType());
213 LV.setNontemporal(true);
214 CGF.EmitStoreOfScalar(Val, LV, false);
215 return nullptr;
216 }
217
218 static Value *EmitNontemporalLoad(CodeGenFunction &CGF, const CallExpr *E) {
219 Value *Address = CGF.EmitScalarExpr(E->getArg(0));
220
221 LValue LV = CGF.MakeNaturalAlignAddrLValue(Address, E->getType());
222 LV.setNontemporal(true);
223 return CGF.EmitLoadOfScalar(LV, E->getExprLoc());
224 }
225
226 static RValue EmitBinaryAtomic(CodeGenFunction &CGF,
227 llvm::AtomicRMWInst::BinOp Kind,
228 const CallExpr *E) {
229 return RValue::get(MakeBinaryAtomicValue(CGF, Kind, E));
230 }
231
232 /// Utility to insert an atomic instruction based Intrinsic::ID and
233 /// the expression node, where the return value is the result of the
234 /// operation.
235 static RValue EmitBinaryAtomicPost(CodeGenFunction &CGF,
236 llvm::AtomicRMWInst::BinOp Kind,
237 const CallExpr *E,
238 Instruction::BinaryOps Op,
239 bool Invert = false) {
240 QualType T = E->getType();
241 assert(E->getArg(0)->getType()->isPointerType());
242 assert(CGF.getContext().hasSameUnqualifiedType(T,
243 E->getArg(0)->getType()->getPointeeType()));
244 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType()));
245
246 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0));
247 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace();
248
249 llvm::IntegerType *IntType =
250 llvm::IntegerType::get(CGF.getLLVMContext(),
251 CGF.getContext().getTypeSize(T));
252 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace);
253
254 llvm::Value *Args[2];
255 Args[1] = CGF.EmitScalarExpr(E->getArg(1));
256 llvm::Type *ValueType = Args[1]->getType();
257 Args[1] = EmitToInt(CGF, Args[1], T, IntType);
258 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType);
259
260 llvm::Value *Result = CGF.Builder.CreateAtomicRMW(
261 Kind, Args[0], Args[1], llvm::AtomicOrdering::SequentiallyConsistent);
262 Result = CGF.Builder.CreateBinOp(Op, Result, Args[1]);
263 if (Invert)
264 Result =
265 CGF.Builder.CreateBinOp(llvm::Instruction::Xor, Result,
266 llvm::ConstantInt::getAllOnesValue(IntType));
267 Result = EmitFromInt(CGF, Result, T, ValueType);
268 return RValue::get(Result);
269 }
270
271 /// Utility to insert an atomic cmpxchg instruction.
272 ///
273 /// @param CGF The current codegen function.
274 /// @param E Builtin call expression to convert to cmpxchg.
275 /// arg0 - address to operate on
276 /// arg1 - value to compare with
277 /// arg2 - new value
278 /// @param ReturnBool Specifies whether to return success flag of
279 /// cmpxchg result or the old value.
280 ///
281 /// @returns result of cmpxchg, according to ReturnBool
282 ///
283 /// Note: In order to lower Microsoft's _InterlockedCompareExchange* intrinsics
284 /// invoke the function EmitAtomicCmpXchgForMSIntrin.
285 static Value *MakeAtomicCmpXchgValue(CodeGenFunction &CGF, const CallExpr *E,
286 bool ReturnBool) {
287 QualType T = ReturnBool ? E->getArg(1)->getType() : E->getType();
288 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0));
289 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace();
290
291 llvm::IntegerType *IntType = llvm::IntegerType::get(
292 CGF.getLLVMContext(), CGF.getContext().getTypeSize(T));
293 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace);
294
295 Value *Args[3];
296 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType);
297 Args[1] = CGF.EmitScalarExpr(E->getArg(1));
298 llvm::Type *ValueType = Args[1]->getType();
299 Args[1] = EmitToInt(CGF, Args[1], T, IntType);
300 Args[2] = EmitToInt(CGF, CGF.EmitScalarExpr(E->getArg(2)), T, IntType);
301
302 Value *Pair = CGF.Builder.CreateAtomicCmpXchg(
303 Args[0], Args[1], Args[2], llvm::AtomicOrdering::SequentiallyConsistent,
304 llvm::AtomicOrdering::SequentiallyConsistent);
305 if (ReturnBool)
306 // Extract boolean success flag and zext it to int.
307 return CGF.Builder.CreateZExt(CGF.Builder.CreateExtractValue(Pair, 1),
308 CGF.ConvertType(E->getType()));
309 else
310 // Extract old value and emit it using the same type as compare value.
311 return EmitFromInt(CGF, CGF.Builder.CreateExtractValue(Pair, 0), T,
312 ValueType);
313 }
314
315 /// This function should be invoked to emit atomic cmpxchg for Microsoft's
316 /// _InterlockedCompareExchange* intrinsics which have the following signature:
317 /// T _InterlockedCompareExchange(T volatile *Destination,
318 /// T Exchange,
319 /// T Comparand);
320 ///
321 /// Whereas the llvm 'cmpxchg' instruction has the following syntax:
322 /// cmpxchg *Destination, Comparand, Exchange.
323 /// So we need to swap Comparand and Exchange when invoking
324 /// CreateAtomicCmpXchg. That is the reason we could not use the above utility
325 /// function MakeAtomicCmpXchgValue since it expects the arguments to be
326 /// already swapped.
327
328 static
329 Value *EmitAtomicCmpXchgForMSIntrin(CodeGenFunction &CGF, const CallExpr *E,
330 AtomicOrdering SuccessOrdering = AtomicOrdering::SequentiallyConsistent) {
331 assert(E->getArg(0)->getType()->isPointerType());
332 assert(CGF.getContext().hasSameUnqualifiedType(
333 E->getType(), E->getArg(0)->getType()->getPointeeType()));
334 assert(CGF.getContext().hasSameUnqualifiedType(E->getType(),
335 E->getArg(1)->getType()));
336 assert(CGF.getContext().hasSameUnqualifiedType(E->getType(),
337 E->getArg(2)->getType()));
338
339 auto *Destination = CGF.EmitScalarExpr(E->getArg(0));
340 auto *Comparand = CGF.EmitScalarExpr(E->getArg(2));
341 auto *Exchange = CGF.EmitScalarExpr(E->getArg(1));
342
343 // For Release ordering, the failure ordering should be Monotonic.
344 auto FailureOrdering = SuccessOrdering == AtomicOrdering::Release ?
345 AtomicOrdering::Monotonic :
346 SuccessOrdering;
347
348 // The atomic instruction is marked volatile for consistency with MSVC. This
349 // blocks the few atomics optimizations that LLVM has. If we want to optimize
350 // _Interlocked* operations in the future, we will have to remove the volatile
351 // marker.
352 auto *Result = CGF.Builder.CreateAtomicCmpXchg(
353 Destination, Comparand, Exchange,
354 SuccessOrdering, FailureOrdering);
355 Result->setVolatile(true);
356 return CGF.Builder.CreateExtractValue(Result, 0);
357 }
358
359 // 64-bit Microsoft platforms support 128 bit cmpxchg operations. They are
360 // prototyped like this:
361 //
362 // unsigned char _InterlockedCompareExchange128...(
363 // __int64 volatile * _Destination,
364 // __int64 _ExchangeHigh,
365 // __int64 _ExchangeLow,
366 // __int64 * _ComparandResult);
367 static Value *EmitAtomicCmpXchg128ForMSIntrin(CodeGenFunction &CGF,
368 const CallExpr *E,
369 AtomicOrdering SuccessOrdering) {
370 assert(E->getNumArgs() == 4);
371 llvm::Value *Destination = CGF.EmitScalarExpr(E->getArg(0));
372 llvm::Value *ExchangeHigh = CGF.EmitScalarExpr(E->getArg(1));
373 llvm::Value *ExchangeLow = CGF.EmitScalarExpr(E->getArg(2));
374 llvm::Value *ComparandPtr = CGF.EmitScalarExpr(E->getArg(3));
375
376 assert(Destination->getType()->isPointerTy());
377 assert(!ExchangeHigh->getType()->isPointerTy());
378 assert(!ExchangeLow->getType()->isPointerTy());
379 assert(ComparandPtr->getType()->isPointerTy());
380
381 // For Release ordering, the failure ordering should be Monotonic.
382 auto FailureOrdering = SuccessOrdering == AtomicOrdering::Release
383 ? AtomicOrdering::Monotonic
384 : SuccessOrdering;
385
386 // Convert to i128 pointers and values.
387 llvm::Type *Int128Ty = llvm::IntegerType::get(CGF.getLLVMContext(), 128);
388 llvm::Type *Int128PtrTy = Int128Ty->getPointerTo();
389 Destination = CGF.Builder.CreateBitCast(Destination, Int128PtrTy);
390 Address ComparandResult(CGF.Builder.CreateBitCast(ComparandPtr, Int128PtrTy),
391 Int128Ty, CGF.getContext().toCharUnitsFromBits(128));
392
393 // (((i128)hi) << 64) | ((i128)lo)
394 ExchangeHigh = CGF.Builder.CreateZExt(ExchangeHigh, Int128Ty);
395 ExchangeLow = CGF.Builder.CreateZExt(ExchangeLow, Int128Ty);
396 ExchangeHigh =
397 CGF.Builder.CreateShl(ExchangeHigh, llvm::ConstantInt::get(Int128Ty, 64));
398 llvm::Value *Exchange = CGF.Builder.CreateOr(ExchangeHigh, ExchangeLow);
399
400 // Load the comparand for the instruction.
401 llvm::Value *Comparand = CGF.Builder.CreateLoad(ComparandResult);
402
403 auto *CXI = CGF.Builder.CreateAtomicCmpXchg(Destination, Comparand, Exchange,
404 SuccessOrdering, FailureOrdering);
405
406 // The atomic instruction is marked volatile for consistency with MSVC. This
407 // blocks the few atomics optimizations that LLVM has. If we want to optimize
408 // _Interlocked* operations in the future, we will have to remove the volatile
409 // marker.
410 CXI->setVolatile(true);
411
412 // Store the result as an outparameter.
413 CGF.Builder.CreateStore(CGF.Builder.CreateExtractValue(CXI, 0),
414 ComparandResult);
415
416 // Get the success boolean and zero extend it to i8.
417 Value *Success = CGF.Builder.CreateExtractValue(CXI, 1);
418 return CGF.Builder.CreateZExt(Success, CGF.Int8Ty);
419 }
420
421 static Value *EmitAtomicIncrementValue(CodeGenFunction &CGF, const CallExpr *E,
422 AtomicOrdering Ordering = AtomicOrdering::SequentiallyConsistent) {
423 assert(E->getArg(0)->getType()->isPointerType());
424
425 auto *IntTy = CGF.ConvertType(E->getType());
426 auto *Result = CGF.Builder.CreateAtomicRMW(
427 AtomicRMWInst::Add,
428 CGF.EmitScalarExpr(E->getArg(0)),
429 ConstantInt::get(IntTy, 1),
430 Ordering);
431 return CGF.Builder.CreateAdd(Result, ConstantInt::get(IntTy, 1));
432 }
433
434 static Value *EmitAtomicDecrementValue(CodeGenFunction &CGF, const CallExpr *E,
435 AtomicOrdering Ordering = AtomicOrdering::SequentiallyConsistent) {
436 assert(E->getArg(0)->getType()->isPointerType());
437
438 auto *IntTy = CGF.ConvertType(E->getType());
439 auto *Result = CGF.Builder.CreateAtomicRMW(
440 AtomicRMWInst::Sub,
441 CGF.EmitScalarExpr(E->getArg(0)),
442 ConstantInt::get(IntTy, 1),
443 Ordering);
444 return CGF.Builder.CreateSub(Result, ConstantInt::get(IntTy, 1));
445 }
446
447 // Build a plain volatile load.
448 static Value *EmitISOVolatileLoad(CodeGenFunction &CGF, const CallExpr *E) {
449 Value *Ptr = CGF.EmitScalarExpr(E->getArg(0));
450 QualType ElTy = E->getArg(0)->getType()->getPointeeType();
451 CharUnits LoadSize = CGF.getContext().getTypeSizeInChars(ElTy);
452 llvm::Type *ITy =
453 llvm::IntegerType::get(CGF.getLLVMContext(), LoadSize.getQuantity() * 8);
454 Ptr = CGF.Builder.CreateBitCast(Ptr, ITy->getPointerTo());
455 llvm::LoadInst *Load = CGF.Builder.CreateAlignedLoad(ITy, Ptr, LoadSize);
456 Load->setVolatile(true);
457 return Load;
458 }
459
460 // Build a plain volatile store.
461 static Value *EmitISOVolatileStore(CodeGenFunction &CGF, const CallExpr *E) {
462 Value *Ptr = CGF.EmitScalarExpr(E->getArg(0));
463 Value *Value = CGF.EmitScalarExpr(E->getArg(1));
464 QualType ElTy = E->getArg(0)->getType()->getPointeeType();
465 CharUnits StoreSize = CGF.getContext().getTypeSizeInChars(ElTy);
466 llvm::Type *ITy =
467 llvm::IntegerType::get(CGF.getLLVMContext(), StoreSize.getQuantity() * 8);
468 Ptr = CGF.Builder.CreateBitCast(Ptr, ITy->getPointerTo());
469 llvm::StoreInst *Store =
470 CGF.Builder.CreateAlignedStore(Value, Ptr, StoreSize);
471 Store->setVolatile(true);
472 return Store;
473 }
474
475 // Emit a simple mangled intrinsic that has 1 argument and a return type
476 // matching the argument type. Depending on mode, this may be a constrained
477 // floating-point intrinsic.
478 static Value *emitUnaryMaybeConstrainedFPBuiltin(CodeGenFunction &CGF,
479 const CallExpr *E, unsigned IntrinsicID,
480 unsigned ConstrainedIntrinsicID) {
481 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0));
482
483 if (CGF.Builder.getIsFPConstrained()) {
484 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(CGF, E);
485 Function *F = CGF.CGM.getIntrinsic(ConstrainedIntrinsicID, Src0->getType());
486 return CGF.Builder.CreateConstrainedFPCall(F, { Src0 });
487 } else {
488 Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType());
489 return CGF.Builder.CreateCall(F, Src0);
490 }
491 }
492
493 // Emit an intrinsic that has 2 operands of the same type as its result.
494 // Depending on mode, this may be a constrained floating-point intrinsic.
495 static Value *emitBinaryMaybeConstrainedFPBuiltin(CodeGenFunction &CGF,
496 const CallExpr *E, unsigned IntrinsicID,
497 unsigned ConstrainedIntrinsicID) {
498 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0));
499 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1));
500
501 if (CGF.Builder.getIsFPConstrained()) {
502 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(CGF, E);
503 Function *F = CGF.CGM.getIntrinsic(ConstrainedIntrinsicID, Src0->getType());
504 return CGF.Builder.CreateConstrainedFPCall(F, { Src0, Src1 });
505 } else {
506 Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType());
507 return CGF.Builder.CreateCall(F, { Src0, Src1 });
508 }
509 }
510
511 // Emit an intrinsic that has 3 operands of the same type as its result.
512 // Depending on mode, this may be a constrained floating-point intrinsic.
513 static Value *emitTernaryMaybeConstrainedFPBuiltin(CodeGenFunction &CGF,
514 const CallExpr *E, unsigned IntrinsicID,
515 unsigned ConstrainedIntrinsicID) {
516 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0));
517 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1));
518 llvm::Value *Src2 = CGF.EmitScalarExpr(E->getArg(2));
519
520 if (CGF.Builder.getIsFPConstrained()) {
521 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(CGF, E);
522 Function *F = CGF.CGM.getIntrinsic(ConstrainedIntrinsicID, Src0->getType());
523 return CGF.Builder.CreateConstrainedFPCall(F, { Src0, Src1, Src2 });
524 } else {
525 Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType());
526 return CGF.Builder.CreateCall(F, { Src0, Src1, Src2 });
527 }
528 }
529
530 // Emit an intrinsic where all operands are of the same type as the result.
531 // Depending on mode, this may be a constrained floating-point intrinsic.
532 static Value *emitCallMaybeConstrainedFPBuiltin(CodeGenFunction &CGF,
533 unsigned IntrinsicID,
534 unsigned ConstrainedIntrinsicID,
535 llvm::Type *Ty,
536 ArrayRef<Value *> Args) {
537 Function *F;
538 if (CGF.Builder.getIsFPConstrained())
539 F = CGF.CGM.getIntrinsic(ConstrainedIntrinsicID, Ty);
540 else
541 F = CGF.CGM.getIntrinsic(IntrinsicID, Ty);
542
543 if (CGF.Builder.getIsFPConstrained())
544 return CGF.Builder.CreateConstrainedFPCall(F, Args);
545 else
546 return CGF.Builder.CreateCall(F, Args);
547 }
548
549 // Emit a simple mangled intrinsic that has 1 argument and a return type
550 // matching the argument type.
551 static Value *emitUnaryBuiltin(CodeGenFunction &CGF, const CallExpr *E,
552 unsigned IntrinsicID,
553 llvm::StringRef Name = "") {
554 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0));
555
556 Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType());
557 return CGF.Builder.CreateCall(F, Src0, Name);
558 }
559
560 // Emit an intrinsic that has 2 operands of the same type as its result.
561 static Value *emitBinaryBuiltin(CodeGenFunction &CGF,
562 const CallExpr *E,
563 unsigned IntrinsicID) {
564 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0));
565 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1));
566
567 Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType());
568 return CGF.Builder.CreateCall(F, { Src0, Src1 });
569 }
570
571 // Emit an intrinsic that has 3 operands of the same type as its result.
572 static Value *emitTernaryBuiltin(CodeGenFunction &CGF,
573 const CallExpr *E,
574 unsigned IntrinsicID) {
575 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0));
576 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1));
577 llvm::Value *Src2 = CGF.EmitScalarExpr(E->getArg(2));
578
579 Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType());
580 return CGF.Builder.CreateCall(F, { Src0, Src1, Src2 });
581 }
582
583 // Emit an intrinsic that has 1 float or double operand, and 1 integer.
584 static Value *emitFPIntBuiltin(CodeGenFunction &CGF,
585 const CallExpr *E,
586 unsigned IntrinsicID) {
587 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0));
588 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1));
589
590 Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType());
591 return CGF.Builder.CreateCall(F, {Src0, Src1});
592 }
593
594 // Emit an intrinsic that has overloaded integer result and fp operand.
595 static Value *
596 emitMaybeConstrainedFPToIntRoundBuiltin(CodeGenFunction &CGF, const CallExpr *E,
597 unsigned IntrinsicID,
598 unsigned ConstrainedIntrinsicID) {
599 llvm::Type *ResultType = CGF.ConvertType(E->getType());
600 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0));
601
602 if (CGF.Builder.getIsFPConstrained()) {
603 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(CGF, E);
604 Function *F = CGF.CGM.getIntrinsic(ConstrainedIntrinsicID,
605 {ResultType, Src0->getType()});
606 return CGF.Builder.CreateConstrainedFPCall(F, {Src0});
607 } else {
608 Function *F =
609 CGF.CGM.getIntrinsic(IntrinsicID, {ResultType, Src0->getType()});
610 return CGF.Builder.CreateCall(F, Src0);
611 }
612 }
613
614 /// EmitFAbs - Emit a call to @llvm.fabs().
615 static Value *EmitFAbs(CodeGenFunction &CGF, Value *V) {
616 Function *F = CGF.CGM.getIntrinsic(Intrinsic::fabs, V->getType());
617 llvm::CallInst *Call = CGF.Builder.CreateCall(F, V);
618 Call->setDoesNotAccessMemory();
619 return Call;
620 }
621
622 /// Emit the computation of the sign bit for a floating point value. Returns
623 /// the i1 sign bit value.
624 static Value *EmitSignBit(CodeGenFunction &CGF, Value *V) {
625 LLVMContext &C = CGF.CGM.getLLVMContext();
626
627 llvm::Type *Ty = V->getType();
628 int Width = Ty->getPrimitiveSizeInBits();
629 llvm::Type *IntTy = llvm::IntegerType::get(C, Width);
630 V = CGF.Builder.CreateBitCast(V, IntTy);
631 if (Ty->isPPC_FP128Ty()) {
632 // We want the sign bit of the higher-order double. The bitcast we just
633 // did works as if the double-double was stored to memory and then
634 // read as an i128. The "store" will put the higher-order double in the
635 // lower address in both little- and big-Endian modes, but the "load"
636 // will treat those bits as a different part of the i128: the low bits in
637 // little-Endian, the high bits in big-Endian. Therefore, on big-Endian
638 // we need to shift the high bits down to the low before truncating.
639 Width >>= 1;
640 if (CGF.getTarget().isBigEndian()) {
641 Value *ShiftCst = llvm::ConstantInt::get(IntTy, Width);
642 V = CGF.Builder.CreateLShr(V, ShiftCst);
643 }
644 // We are truncating value in order to extract the higher-order
645 // double, which we will be using to extract the sign from.
646 IntTy = llvm::IntegerType::get(C, Width);
647 V = CGF.Builder.CreateTrunc(V, IntTy);
648 }
649 Value *Zero = llvm::Constant::getNullValue(IntTy);
650 return CGF.Builder.CreateICmpSLT(V, Zero);
651 }
652
653 static RValue emitLibraryCall(CodeGenFunction &CGF, const FunctionDecl *FD,
654 const CallExpr *E, llvm::Constant *calleeValue) {
655 CGCallee callee = CGCallee::forDirect(calleeValue, GlobalDecl(FD));
656 return CGF.EmitCall(E->getCallee()->getType(), callee, E, ReturnValueSlot());
657 }
658
659 /// Emit a call to llvm.{sadd,uadd,ssub,usub,smul,umul}.with.overflow.*
660 /// depending on IntrinsicID.
661 ///
662 /// \arg CGF The current codegen function.
663 /// \arg IntrinsicID The ID for the Intrinsic we wish to generate.
664 /// \arg X The first argument to the llvm.*.with.overflow.*.
665 /// \arg Y The second argument to the llvm.*.with.overflow.*.
666 /// \arg Carry The carry returned by the llvm.*.with.overflow.*.
667 /// \returns The result (i.e. sum/product) returned by the intrinsic.
668 static llvm::Value *EmitOverflowIntrinsic(CodeGenFunction &CGF,
669 const llvm::Intrinsic::ID IntrinsicID,
670 llvm::Value *X, llvm::Value *Y,
671 llvm::Value *&Carry) {
672 // Make sure we have integers of the same width.
673 assert(X->getType() == Y->getType() &&
674 "Arguments must be the same type. (Did you forget to make sure both "
675 "arguments have the same integer width?)");
676
677 Function *Callee = CGF.CGM.getIntrinsic(IntrinsicID, X->getType());
678 llvm::Value *Tmp = CGF.Builder.CreateCall(Callee, {X, Y});
679 Carry = CGF.Builder.CreateExtractValue(Tmp, 1);
680 return CGF.Builder.CreateExtractValue(Tmp, 0);
681 }
682
683 static Value *emitRangedBuiltin(CodeGenFunction &CGF,
684 unsigned IntrinsicID,
685 int low, int high) {
686 llvm::MDBuilder MDHelper(CGF.getLLVMContext());
687 llvm::MDNode *RNode = MDHelper.createRange(APInt(32, low), APInt(32, high));
688 Function *F = CGF.CGM.getIntrinsic(IntrinsicID, {});
689 llvm::Instruction *Call = CGF.Builder.CreateCall(F);
690 Call->setMetadata(llvm::LLVMContext::MD_range, RNode);
691 Call->setMetadata(llvm::LLVMContext::MD_noundef,
692 llvm::MDNode::get(CGF.getLLVMContext(), std::nullopt));
693 return Call;
694 }
695
696 namespace {
697 struct WidthAndSignedness {
698 unsigned Width;
699 bool Signed;
700 };
701 }
702
703 static WidthAndSignedness
704 getIntegerWidthAndSignedness(const clang::ASTContext &context,
705 const clang::QualType Type) {
706 assert(Type->isIntegerType() && "Given type is not an integer.");
707 unsigned Width = Type->isBooleanType() ? 1
708 : Type->isBitIntType() ? context.getIntWidth(Type)
709 : context.getTypeInfo(Type).Width;
710 bool Signed = Type->isSignedIntegerType();
711 return {Width, Signed};
712 }
713
714 // Given one or more integer types, this function produces an integer type that
715 // encompasses them: any value in one of the given types could be expressed in
716 // the encompassing type.
717 static struct WidthAndSignedness
718 EncompassingIntegerType(ArrayRef<struct WidthAndSignedness> Types) {
719 assert(Types.size() > 0 && "Empty list of types.");
720
721 // If any of the given types is signed, we must return a signed type.
722 bool Signed = false;
723 for (const auto &Type : Types) {
724 Signed |= Type.Signed;
725 }
726
727 // The encompassing type must have a width greater than or equal to the width
728 // of the specified types. Additionally, if the encompassing type is signed,
729 // its width must be strictly greater than the width of any unsigned types
730 // given.
731 unsigned Width = 0;
732 for (const auto &Type : Types) {
733 unsigned MinWidth = Type.Width + (Signed && !Type.Signed);
734 if (Width < MinWidth) {
735 Width = MinWidth;
736 }
737 }
738
739 return {Width, Signed};
740 }
741
742 Value *CodeGenFunction::EmitVAStartEnd(Value *ArgValue, bool IsStart) {
743 llvm::Type *DestType = Int8PtrTy;
744 if (ArgValue->getType() != DestType)
745 ArgValue =
746 Builder.CreateBitCast(ArgValue, DestType, ArgValue->getName().data());
747
748 Intrinsic::ID inst = IsStart ? Intrinsic::vastart : Intrinsic::vaend;
749 return Builder.CreateCall(CGM.getIntrinsic(inst), ArgValue);
750 }
751
752 /// Checks if using the result of __builtin_object_size(p, @p From) in place of
753 /// __builtin_object_size(p, @p To) is correct
754 static bool areBOSTypesCompatible(int From, int To) {
755 // Note: Our __builtin_object_size implementation currently treats Type=0 and
756 // Type=2 identically. Encoding this implementation detail here may make
757 // improving __builtin_object_size difficult in the future, so it's omitted.
758 return From == To || (From == 0 && To == 1) || (From == 3 && To == 2);
759 }
760
761 static llvm::Value *
762 getDefaultBuiltinObjectSizeResult(unsigned Type, llvm::IntegerType *ResType) {
763 return ConstantInt::get(ResType, (Type & 2) ? 0 : -1, /*isSigned=*/true);
764 }
765
766 llvm::Value *
767 CodeGenFunction::evaluateOrEmitBuiltinObjectSize(const Expr *E, unsigned Type,
768 llvm::IntegerType *ResType,
769 llvm::Value *EmittedE,
770 bool IsDynamic) {
771 uint64_t ObjectSize;
772 if (!E->tryEvaluateObjectSize(ObjectSize, getContext(), Type))
773 return emitBuiltinObjectSize(E, Type, ResType, EmittedE, IsDynamic);
774 return ConstantInt::get(ResType, ObjectSize, /*isSigned=*/true);
775 }
776
777 /// Returns a Value corresponding to the size of the given expression.
778 /// This Value may be either of the following:
779 /// - A llvm::Argument (if E is a param with the pass_object_size attribute on
780 /// it)
781 /// - A call to the @llvm.objectsize intrinsic
782 ///
783 /// EmittedE is the result of emitting `E` as a scalar expr. If it's non-null
784 /// and we wouldn't otherwise try to reference a pass_object_size parameter,
785 /// we'll call @llvm.objectsize on EmittedE, rather than emitting E.
786 llvm::Value *
787 CodeGenFunction::emitBuiltinObjectSize(const Expr *E, unsigned Type,
788 llvm::IntegerType *ResType,
789 llvm::Value *EmittedE, bool IsDynamic) {
790 // We need to reference an argument if the pointer is a parameter with the
791 // pass_object_size attribute.
792 if (auto *D = dyn_cast<DeclRefExpr>(E->IgnoreParenImpCasts())) {
793 auto *Param = dyn_cast<ParmVarDecl>(D->getDecl());
794 auto *PS = D->getDecl()->getAttr<PassObjectSizeAttr>();
795 if (Param != nullptr && PS != nullptr &&
796 areBOSTypesCompatible(PS->getType(), Type)) {
797 auto Iter = SizeArguments.find(Param);
798 assert(Iter != SizeArguments.end());
799
800 const ImplicitParamDecl *D = Iter->second;
801 auto DIter = LocalDeclMap.find(D);
802 assert(DIter != LocalDeclMap.end());
803
804 return EmitLoadOfScalar(DIter->second, /*Volatile=*/false,
805 getContext().getSizeType(), E->getBeginLoc());
806 }
807 }
808
809 // LLVM can't handle Type=3 appropriately, and __builtin_object_size shouldn't
810 // evaluate E for side-effects. In either case, we shouldn't lower to
811 // @llvm.objectsize.
812 if (Type == 3 || (!EmittedE && E->HasSideEffects(getContext())))
813 return getDefaultBuiltinObjectSizeResult(Type, ResType);
814
815 Value *Ptr = EmittedE ? EmittedE : EmitScalarExpr(E);
816 assert(Ptr->getType()->isPointerTy() &&
817 "Non-pointer passed to __builtin_object_size?");
818
819 Function *F =
820 CGM.getIntrinsic(Intrinsic::objectsize, {ResType, Ptr->getType()});
821
822 // LLVM only supports 0 and 2, make sure that we pass along that as a boolean.
823 Value *Min = Builder.getInt1((Type & 2) != 0);
824 // For GCC compatibility, __builtin_object_size treat NULL as unknown size.
825 Value *NullIsUnknown = Builder.getTrue();
826 Value *Dynamic = Builder.getInt1(IsDynamic);
827 return Builder.CreateCall(F, {Ptr, Min, NullIsUnknown, Dynamic});
828 }
829
830 namespace {
831 /// A struct to generically describe a bit test intrinsic.
832 struct BitTest {
833 enum ActionKind : uint8_t { TestOnly, Complement, Reset, Set };
834 enum InterlockingKind : uint8_t {
835 Unlocked,
836 Sequential,
837 Acquire,
838 Release,
839 NoFence
840 };
841
842 ActionKind Action;
843 InterlockingKind Interlocking;
844 bool Is64Bit;
845
846 static BitTest decodeBitTestBuiltin(unsigned BuiltinID);
847 };
848 } // namespace
849
850 BitTest BitTest::decodeBitTestBuiltin(unsigned BuiltinID) {
851 switch (BuiltinID) {
852 // Main portable variants.
853 case Builtin::BI_bittest:
854 return {TestOnly, Unlocked, false};
855 case Builtin::BI_bittestandcomplement:
856 return {Complement, Unlocked, false};
857 case Builtin::BI_bittestandreset:
858 return {Reset, Unlocked, false};
859 case Builtin::BI_bittestandset:
860 return {Set, Unlocked, false};
861 case Builtin::BI_interlockedbittestandreset:
862 return {Reset, Sequential, false};
863 case Builtin::BI_interlockedbittestandset:
864 return {Set, Sequential, false};
865
866 // X86-specific 64-bit variants.
867 case Builtin::BI_bittest64:
868 return {TestOnly, Unlocked, true};
869 case Builtin::BI_bittestandcomplement64:
870 return {Complement, Unlocked, true};
871 case Builtin::BI_bittestandreset64:
872 return {Reset, Unlocked, true};
873 case Builtin::BI_bittestandset64:
874 return {Set, Unlocked, true};
875 case Builtin::BI_interlockedbittestandreset64:
876 return {Reset, Sequential, true};
877 case Builtin::BI_interlockedbittestandset64:
878 return {Set, Sequential, true};
879
880 // ARM/AArch64-specific ordering variants.
881 case Builtin::BI_interlockedbittestandset_acq:
882 return {Set, Acquire, false};
883 case Builtin::BI_interlockedbittestandset_rel:
884 return {Set, Release, false};
885 case Builtin::BI_interlockedbittestandset_nf:
886 return {Set, NoFence, false};
887 case Builtin::BI_interlockedbittestandreset_acq:
888 return {Reset, Acquire, false};
889 case Builtin::BI_interlockedbittestandreset_rel:
890 return {Reset, Release, false};
891 case Builtin::BI_interlockedbittestandreset_nf:
892 return {Reset, NoFence, false};
893 }
894 llvm_unreachable("expected only bittest intrinsics");
895 }
896
897 static char bitActionToX86BTCode(BitTest::ActionKind A) {
898 switch (A) {
899 case BitTest::TestOnly: return '\0';
900 case BitTest::Complement: return 'c';
901 case BitTest::Reset: return 'r';
902 case BitTest::Set: return 's';
903 }
904 llvm_unreachable("invalid action");
905 }
906
907 static llvm::Value *EmitX86BitTestIntrinsic(CodeGenFunction &CGF,
908 BitTest BT,
909 const CallExpr *E, Value *BitBase,
910 Value *BitPos) {
911 char Action = bitActionToX86BTCode(BT.Action);
912 char SizeSuffix = BT.Is64Bit ? 'q' : 'l';
913
914 // Build the assembly.
915 SmallString<64> Asm;
916 raw_svector_ostream AsmOS(Asm);
917 if (BT.Interlocking != BitTest::Unlocked)
918 AsmOS << "lock ";
919 AsmOS << "bt";
920 if (Action)
921 AsmOS << Action;
922 AsmOS << SizeSuffix << " $2, ($1)";
923
924 // Build the constraints. FIXME: We should support immediates when possible.
925 std::string Constraints = "={@ccc},r,r,~{cc},~{memory}";
926 std::string MachineClobbers = CGF.getTarget().getClobbers();
927 if (!MachineClobbers.empty()) {
928 Constraints += ',';
929 Constraints += MachineClobbers;
930 }
931 llvm::IntegerType *IntType = llvm::IntegerType::get(
932 CGF.getLLVMContext(),
933 CGF.getContext().getTypeSize(E->getArg(1)->getType()));
934 llvm::Type *IntPtrType = IntType->getPointerTo();
935 llvm::FunctionType *FTy =
936 llvm::FunctionType::get(CGF.Int8Ty, {IntPtrType, IntType}, false);
937
938 llvm::InlineAsm *IA =
939 llvm::InlineAsm::get(FTy, Asm, Constraints, /*hasSideEffects=*/true);
940 return CGF.Builder.CreateCall(IA, {BitBase, BitPos});
941 }
942
943 static llvm::AtomicOrdering
944 getBitTestAtomicOrdering(BitTest::InterlockingKind I) {
945 switch (I) {
946 case BitTest::Unlocked: return llvm::AtomicOrdering::NotAtomic;
947 case BitTest::Sequential: return llvm::AtomicOrdering::SequentiallyConsistent;
948 case BitTest::Acquire: return llvm::AtomicOrdering::Acquire;
949 case BitTest::Release: return llvm::AtomicOrdering::Release;
950 case BitTest::NoFence: return llvm::AtomicOrdering::Monotonic;
951 }
952 llvm_unreachable("invalid interlocking");
953 }
954
955 /// Emit a _bittest* intrinsic. These intrinsics take a pointer to an array of
956 /// bits and a bit position and read and optionally modify the bit at that
957 /// position. The position index can be arbitrarily large, i.e. it can be larger
958 /// than 31 or 63, so we need an indexed load in the general case.
959 static llvm::Value *EmitBitTestIntrinsic(CodeGenFunction &CGF,
960 unsigned BuiltinID,
961 const CallExpr *E) {
962 Value *BitBase = CGF.EmitScalarExpr(E->getArg(0));
963 Value *BitPos = CGF.EmitScalarExpr(E->getArg(1));
964
965 BitTest BT = BitTest::decodeBitTestBuiltin(BuiltinID);
966
967 // X86 has special BT, BTC, BTR, and BTS instructions that handle the array
968 // indexing operation internally. Use them if possible.
969 if (CGF.getTarget().getTriple().isX86())
970 return EmitX86BitTestIntrinsic(CGF, BT, E, BitBase, BitPos);
971
972 // Otherwise, use generic code to load one byte and test the bit. Use all but
973 // the bottom three bits as the array index, and the bottom three bits to form
974 // a mask.
975 // Bit = BitBaseI8[BitPos >> 3] & (1 << (BitPos & 0x7)) != 0;
976 Value *ByteIndex = CGF.Builder.CreateAShr(
977 BitPos, llvm::ConstantInt::get(BitPos->getType(), 3), "bittest.byteidx");
978 Value *BitBaseI8 = CGF.Builder.CreatePointerCast(BitBase, CGF.Int8PtrTy);
979 Address ByteAddr(CGF.Builder.CreateInBoundsGEP(CGF.Int8Ty, BitBaseI8,
980 ByteIndex, "bittest.byteaddr"),
981 CGF.Int8Ty, CharUnits::One());
982 Value *PosLow =
983 CGF.Builder.CreateAnd(CGF.Builder.CreateTrunc(BitPos, CGF.Int8Ty),
984 llvm::ConstantInt::get(CGF.Int8Ty, 0x7));
985
986 // The updating instructions will need a mask.
987 Value *Mask = nullptr;
988 if (BT.Action != BitTest::TestOnly) {
989 Mask = CGF.Builder.CreateShl(llvm::ConstantInt::get(CGF.Int8Ty, 1), PosLow,
990 "bittest.mask");
991 }
992
993 // Check the action and ordering of the interlocked intrinsics.
994 llvm::AtomicOrdering Ordering = getBitTestAtomicOrdering(BT.Interlocking);
995
996 Value *OldByte = nullptr;
997 if (Ordering != llvm::AtomicOrdering::NotAtomic) {
998 // Emit a combined atomicrmw load/store operation for the interlocked
999 // intrinsics.
1000 llvm::AtomicRMWInst::BinOp RMWOp = llvm::AtomicRMWInst::Or;
1001 if (BT.Action == BitTest::Reset) {
1002 Mask = CGF.Builder.CreateNot(Mask);
1003 RMWOp = llvm::AtomicRMWInst::And;
1004 }
1005 OldByte = CGF.Builder.CreateAtomicRMW(RMWOp, ByteAddr.getPointer(), Mask,
1006 Ordering);
1007 } else {
1008 // Emit a plain load for the non-interlocked intrinsics.
1009 OldByte = CGF.Builder.CreateLoad(ByteAddr, "bittest.byte");
1010 Value *NewByte = nullptr;
1011 switch (BT.Action) {
1012 case BitTest::TestOnly:
1013 // Don't store anything.
1014 break;
1015 case BitTest::Complement:
1016 NewByte = CGF.Builder.CreateXor(OldByte, Mask);
1017 break;
1018 case BitTest::Reset:
1019 NewByte = CGF.Builder.CreateAnd(OldByte, CGF.Builder.CreateNot(Mask));
1020 break;
1021 case BitTest::Set:
1022 NewByte = CGF.Builder.CreateOr(OldByte, Mask);
1023 break;
1024 }
1025 if (NewByte)
1026 CGF.Builder.CreateStore(NewByte, ByteAddr);
1027 }
1028
1029 // However we loaded the old byte, either by plain load or atomicrmw, shift
1030 // the bit into the low position and mask it to 0 or 1.
1031 Value *ShiftedByte = CGF.Builder.CreateLShr(OldByte, PosLow, "bittest.shr");
1032 return CGF.Builder.CreateAnd(
1033 ShiftedByte, llvm::ConstantInt::get(CGF.Int8Ty, 1), "bittest.res");
1034 }
1035
1036 static llvm::Value *emitPPCLoadReserveIntrinsic(CodeGenFunction &CGF,
1037 unsigned BuiltinID,
1038 const CallExpr *E) {
1039 Value *Addr = CGF.EmitScalarExpr(E->getArg(0));
1040
1041 SmallString<64> Asm;
1042 raw_svector_ostream AsmOS(Asm);
1043 llvm::IntegerType *RetType = CGF.Int32Ty;
1044
1045 switch (BuiltinID) {
1046 case clang::PPC::BI__builtin_ppc_ldarx:
1047 AsmOS << "ldarx ";
1048 RetType = CGF.Int64Ty;
1049 break;
1050 case clang::PPC::BI__builtin_ppc_lwarx:
1051 AsmOS << "lwarx ";
1052 RetType = CGF.Int32Ty;
1053 break;
1054 case clang::PPC::BI__builtin_ppc_lharx:
1055 AsmOS << "lharx ";
1056 RetType = CGF.Int16Ty;
1057 break;
1058 case clang::PPC::BI__builtin_ppc_lbarx:
1059 AsmOS << "lbarx ";
1060 RetType = CGF.Int8Ty;
1061 break;
1062 default:
1063 llvm_unreachable("Expected only PowerPC load reserve intrinsics");
1064 }
1065
1066 AsmOS << "$0, ${1:y}";
1067
1068 std::string Constraints = "=r,*Z,~{memory}";
1069 std::string MachineClobbers = CGF.getTarget().getClobbers();
1070 if (!MachineClobbers.empty()) {
1071 Constraints += ',';
1072 Constraints += MachineClobbers;
1073 }
1074
1075 llvm::Type *IntPtrType = RetType->getPointerTo();
1076 llvm::FunctionType *FTy =
1077 llvm::FunctionType::get(RetType, {IntPtrType}, false);
1078
1079 llvm::InlineAsm *IA =
1080 llvm::InlineAsm::get(FTy, Asm, Constraints, /*hasSideEffects=*/true);
1081 llvm::CallInst *CI = CGF.Builder.CreateCall(IA, {Addr});
1082 CI->addParamAttr(
1083 0, Attribute::get(CGF.getLLVMContext(), Attribute::ElementType, RetType));
1084 return CI;
1085 }
1086
1087 namespace {
1088 enum class MSVCSetJmpKind {
1089 _setjmpex,
1090 _setjmp3,
1091 _setjmp
1092 };
1093 }
1094
1095 /// MSVC handles setjmp a bit differently on different platforms. On every
1096 /// architecture except 32-bit x86, the frame address is passed. On x86, extra
1097 /// parameters can be passed as variadic arguments, but we always pass none.
1098 static RValue EmitMSVCRTSetJmp(CodeGenFunction &CGF, MSVCSetJmpKind SJKind,
1099 const CallExpr *E) {
1100 llvm::Value *Arg1 = nullptr;
1101 llvm::Type *Arg1Ty = nullptr;
1102 StringRef Name;
1103 bool IsVarArg = false;
1104 if (SJKind == MSVCSetJmpKind::_setjmp3) {
1105 Name = "_setjmp3";
1106 Arg1Ty = CGF.Int32Ty;
1107 Arg1 = llvm::ConstantInt::get(CGF.IntTy, 0);
1108 IsVarArg = true;
1109 } else {
1110 Name = SJKind == MSVCSetJmpKind::_setjmp ? "_setjmp" : "_setjmpex";
1111 Arg1Ty = CGF.Int8PtrTy;
1112 if (CGF.getTarget().getTriple().getArch() == llvm::Triple::aarch64) {
1113 Arg1 = CGF.Builder.CreateCall(
1114 CGF.CGM.getIntrinsic(Intrinsic::sponentry, CGF.AllocaInt8PtrTy));
1115 } else
1116 Arg1 = CGF.Builder.CreateCall(
1117 CGF.CGM.getIntrinsic(Intrinsic::frameaddress, CGF.AllocaInt8PtrTy),
1118 llvm::ConstantInt::get(CGF.Int32Ty, 0));
1119 }
1120
1121 // Mark the call site and declaration with ReturnsTwice.
1122 llvm::Type *ArgTypes[2] = {CGF.Int8PtrTy, Arg1Ty};
1123 llvm::AttributeList ReturnsTwiceAttr = llvm::AttributeList::get(
1124 CGF.getLLVMContext(), llvm::AttributeList::FunctionIndex,
1125 llvm::Attribute::ReturnsTwice);
1126 llvm::FunctionCallee SetJmpFn = CGF.CGM.CreateRuntimeFunction(
1127 llvm::FunctionType::get(CGF.IntTy, ArgTypes, IsVarArg), Name,
1128 ReturnsTwiceAttr, /*Local=*/true);
1129
1130 llvm::Value *Buf = CGF.Builder.CreateBitOrPointerCast(
1131 CGF.EmitScalarExpr(E->getArg(0)), CGF.Int8PtrTy);
1132 llvm::Value *Args[] = {Buf, Arg1};
1133 llvm::CallBase *CB = CGF.EmitRuntimeCallOrInvoke(SetJmpFn, Args);
1134 CB->setAttributes(ReturnsTwiceAttr);
1135 return RValue::get(CB);
1136 }
1137
1138 // Many of MSVC builtins are on x64, ARM and AArch64; to avoid repeating code,
1139 // we handle them here.
1140 enum class CodeGenFunction::MSVCIntrin {
1141 _BitScanForward,
1142 _BitScanReverse,
1143 _InterlockedAnd,
1144 _InterlockedDecrement,
1145 _InterlockedExchange,
1146 _InterlockedExchangeAdd,
1147 _InterlockedExchangeSub,
1148 _InterlockedIncrement,
1149 _InterlockedOr,
1150 _InterlockedXor,
1151 _InterlockedExchangeAdd_acq,
1152 _InterlockedExchangeAdd_rel,
1153 _InterlockedExchangeAdd_nf,
1154 _InterlockedExchange_acq,
1155 _InterlockedExchange_rel,
1156 _InterlockedExchange_nf,
1157 _InterlockedCompareExchange_acq,
1158 _InterlockedCompareExchange_rel,
1159 _InterlockedCompareExchange_nf,
1160 _InterlockedCompareExchange128,
1161 _InterlockedCompareExchange128_acq,
1162 _InterlockedCompareExchange128_rel,
1163 _InterlockedCompareExchange128_nf,
1164 _InterlockedOr_acq,
1165 _InterlockedOr_rel,
1166 _InterlockedOr_nf,
1167 _InterlockedXor_acq,
1168 _InterlockedXor_rel,
1169 _InterlockedXor_nf,
1170 _InterlockedAnd_acq,
1171 _InterlockedAnd_rel,
1172 _InterlockedAnd_nf,
1173 _InterlockedIncrement_acq,
1174 _InterlockedIncrement_rel,
1175 _InterlockedIncrement_nf,
1176 _InterlockedDecrement_acq,
1177 _InterlockedDecrement_rel,
1178 _InterlockedDecrement_nf,
1179 __fastfail,
1180 };
1181
1182 static std::optional<CodeGenFunction::MSVCIntrin>
1183 translateArmToMsvcIntrin(unsigned BuiltinID) {
1184 using MSVCIntrin = CodeGenFunction::MSVCIntrin;
1185 switch (BuiltinID) {
1186 default:
1187 return std::nullopt;
1188 case clang::ARM::BI_BitScanForward:
1189 case clang::ARM::BI_BitScanForward64:
1190 return MSVCIntrin::_BitScanForward;
1191 case clang::ARM::BI_BitScanReverse:
1192 case clang::ARM::BI_BitScanReverse64:
1193 return MSVCIntrin::_BitScanReverse;
1194 case clang::ARM::BI_InterlockedAnd64:
1195 return MSVCIntrin::_InterlockedAnd;
1196 case clang::ARM::BI_InterlockedExchange64:
1197 return MSVCIntrin::_InterlockedExchange;
1198 case clang::ARM::BI_InterlockedExchangeAdd64:
1199 return MSVCIntrin::_InterlockedExchangeAdd;
1200 case clang::ARM::BI_InterlockedExchangeSub64:
1201 return MSVCIntrin::_InterlockedExchangeSub;
1202 case clang::ARM::BI_InterlockedOr64:
1203 return MSVCIntrin::_InterlockedOr;
1204 case clang::ARM::BI_InterlockedXor64:
1205 return MSVCIntrin::_InterlockedXor;
1206 case clang::ARM::BI_InterlockedDecrement64:
1207 return MSVCIntrin::_InterlockedDecrement;
1208 case clang::ARM::BI_InterlockedIncrement64:
1209 return MSVCIntrin::_InterlockedIncrement;
1210 case clang::ARM::BI_InterlockedExchangeAdd8_acq:
1211 case clang::ARM::BI_InterlockedExchangeAdd16_acq:
1212 case clang::ARM::BI_InterlockedExchangeAdd_acq:
1213 case clang::ARM::BI_InterlockedExchangeAdd64_acq:
1214 return MSVCIntrin::_InterlockedExchangeAdd_acq;
1215 case clang::ARM::BI_InterlockedExchangeAdd8_rel:
1216 case clang::ARM::BI_InterlockedExchangeAdd16_rel:
1217 case clang::ARM::BI_InterlockedExchangeAdd_rel:
1218 case clang::ARM::BI_InterlockedExchangeAdd64_rel:
1219 return MSVCIntrin::_InterlockedExchangeAdd_rel;
1220 case clang::ARM::BI_InterlockedExchangeAdd8_nf:
1221 case clang::ARM::BI_InterlockedExchangeAdd16_nf:
1222 case clang::ARM::BI_InterlockedExchangeAdd_nf:
1223 case clang::ARM::BI_InterlockedExchangeAdd64_nf:
1224 return MSVCIntrin::_InterlockedExchangeAdd_nf;
1225 case clang::ARM::BI_InterlockedExchange8_acq:
1226 case clang::ARM::BI_InterlockedExchange16_acq:
1227 case clang::ARM::BI_InterlockedExchange_acq:
1228 case clang::ARM::BI_InterlockedExchange64_acq:
1229 return MSVCIntrin::_InterlockedExchange_acq;
1230 case clang::ARM::BI_InterlockedExchange8_rel:
1231 case clang::ARM::BI_InterlockedExchange16_rel:
1232 case clang::ARM::BI_InterlockedExchange_rel:
1233 case clang::ARM::BI_InterlockedExchange64_rel:
1234 return MSVCIntrin::_InterlockedExchange_rel;
1235 case clang::ARM::BI_InterlockedExchange8_nf:
1236 case clang::ARM::BI_InterlockedExchange16_nf:
1237 case clang::ARM::BI_InterlockedExchange_nf:
1238 case clang::ARM::BI_InterlockedExchange64_nf:
1239 return MSVCIntrin::_InterlockedExchange_nf;
1240 case clang::ARM::BI_InterlockedCompareExchange8_acq:
1241 case clang::ARM::BI_InterlockedCompareExchange16_acq:
1242 case clang::ARM::BI_InterlockedCompareExchange_acq:
1243 case clang::ARM::BI_InterlockedCompareExchange64_acq:
1244 return MSVCIntrin::_InterlockedCompareExchange_acq;
1245 case clang::ARM::BI_InterlockedCompareExchange8_rel:
1246 case clang::ARM::BI_InterlockedCompareExchange16_rel:
1247 case clang::ARM::BI_InterlockedCompareExchange_rel:
1248 case clang::ARM::BI_InterlockedCompareExchange64_rel:
1249 return MSVCIntrin::_InterlockedCompareExchange_rel;
1250 case clang::ARM::BI_InterlockedCompareExchange8_nf:
1251 case clang::ARM::BI_InterlockedCompareExchange16_nf:
1252 case clang::ARM::BI_InterlockedCompareExchange_nf:
1253 case clang::ARM::BI_InterlockedCompareExchange64_nf:
1254 return MSVCIntrin::_InterlockedCompareExchange_nf;
1255 case clang::ARM::BI_InterlockedOr8_acq:
1256 case clang::ARM::BI_InterlockedOr16_acq:
1257 case clang::ARM::BI_InterlockedOr_acq:
1258 case clang::ARM::BI_InterlockedOr64_acq:
1259 return MSVCIntrin::_InterlockedOr_acq;
1260 case clang::ARM::BI_InterlockedOr8_rel:
1261 case clang::ARM::BI_InterlockedOr16_rel:
1262 case clang::ARM::BI_InterlockedOr_rel:
1263 case clang::ARM::BI_InterlockedOr64_rel:
1264 return MSVCIntrin::_InterlockedOr_rel;
1265 case clang::ARM::BI_InterlockedOr8_nf:
1266 case clang::ARM::BI_InterlockedOr16_nf:
1267 case clang::ARM::BI_InterlockedOr_nf:
1268 case clang::ARM::BI_InterlockedOr64_nf:
1269 return MSVCIntrin::_InterlockedOr_nf;
1270 case clang::ARM::BI_InterlockedXor8_acq:
1271 case clang::ARM::BI_InterlockedXor16_acq:
1272 case clang::ARM::BI_InterlockedXor_acq:
1273 case clang::ARM::BI_InterlockedXor64_acq:
1274 return MSVCIntrin::_InterlockedXor_acq;
1275 case clang::ARM::BI_InterlockedXor8_rel:
1276 case clang::ARM::BI_InterlockedXor16_rel:
1277 case clang::ARM::BI_InterlockedXor_rel:
1278 case clang::ARM::BI_InterlockedXor64_rel:
1279 return MSVCIntrin::_InterlockedXor_rel;
1280 case clang::ARM::BI_InterlockedXor8_nf:
1281 case clang::ARM::BI_InterlockedXor16_nf:
1282 case clang::ARM::BI_InterlockedXor_nf:
1283 case clang::ARM::BI_InterlockedXor64_nf:
1284 return MSVCIntrin::_InterlockedXor_nf;
1285 case clang::ARM::BI_InterlockedAnd8_acq:
1286 case clang::ARM::BI_InterlockedAnd16_acq:
1287 case clang::ARM::BI_InterlockedAnd_acq:
1288 case clang::ARM::BI_InterlockedAnd64_acq:
1289 return MSVCIntrin::_InterlockedAnd_acq;
1290 case clang::ARM::BI_InterlockedAnd8_rel:
1291 case clang::ARM::BI_InterlockedAnd16_rel:
1292 case clang::ARM::BI_InterlockedAnd_rel:
1293 case clang::ARM::BI_InterlockedAnd64_rel:
1294 return MSVCIntrin::_InterlockedAnd_rel;
1295 case clang::ARM::BI_InterlockedAnd8_nf:
1296 case clang::ARM::BI_InterlockedAnd16_nf:
1297 case clang::ARM::BI_InterlockedAnd_nf:
1298 case clang::ARM::BI_InterlockedAnd64_nf:
1299 return MSVCIntrin::_InterlockedAnd_nf;
1300 case clang::ARM::BI_InterlockedIncrement16_acq:
1301 case clang::ARM::BI_InterlockedIncrement_acq:
1302 case clang::ARM::BI_InterlockedIncrement64_acq:
1303 return MSVCIntrin::_InterlockedIncrement_acq;
1304 case clang::ARM::BI_InterlockedIncrement16_rel:
1305 case clang::ARM::BI_InterlockedIncrement_rel:
1306 case clang::ARM::BI_InterlockedIncrement64_rel:
1307 return MSVCIntrin::_InterlockedIncrement_rel;
1308 case clang::ARM::BI_InterlockedIncrement16_nf:
1309 case clang::ARM::BI_InterlockedIncrement_nf:
1310 case clang::ARM::BI_InterlockedIncrement64_nf:
1311 return MSVCIntrin::_InterlockedIncrement_nf;
1312 case clang::ARM::BI_InterlockedDecrement16_acq:
1313 case clang::ARM::BI_InterlockedDecrement_acq:
1314 case clang::ARM::BI_InterlockedDecrement64_acq:
1315 return MSVCIntrin::_InterlockedDecrement_acq;
1316 case clang::ARM::BI_InterlockedDecrement16_rel:
1317 case clang::ARM::BI_InterlockedDecrement_rel:
1318 case clang::ARM::BI_InterlockedDecrement64_rel:
1319 return MSVCIntrin::_InterlockedDecrement_rel;
1320 case clang::ARM::BI_InterlockedDecrement16_nf:
1321 case clang::ARM::BI_InterlockedDecrement_nf:
1322 case clang::ARM::BI_InterlockedDecrement64_nf:
1323 return MSVCIntrin::_InterlockedDecrement_nf;
1324 }
1325 llvm_unreachable("must return from switch");
1326 }
1327
1328 static std::optional<CodeGenFunction::MSVCIntrin>
1329 translateAarch64ToMsvcIntrin(unsigned BuiltinID) {
1330 using MSVCIntrin = CodeGenFunction::MSVCIntrin;
1331 switch (BuiltinID) {
1332 default:
1333 return std::nullopt;
1334 case clang::AArch64::BI_BitScanForward:
1335 case clang::AArch64::BI_BitScanForward64:
1336 return MSVCIntrin::_BitScanForward;
1337 case clang::AArch64::BI_BitScanReverse:
1338 case clang::AArch64::BI_BitScanReverse64:
1339 return MSVCIntrin::_BitScanReverse;
1340 case clang::AArch64::BI_InterlockedAnd64:
1341 return MSVCIntrin::_InterlockedAnd;
1342 case clang::AArch64::BI_InterlockedExchange64:
1343 return MSVCIntrin::_InterlockedExchange;
1344 case clang::AArch64::BI_InterlockedExchangeAdd64:
1345 return MSVCIntrin::_InterlockedExchangeAdd;
1346 case clang::AArch64::BI_InterlockedExchangeSub64:
1347 return MSVCIntrin::_InterlockedExchangeSub;
1348 case clang::AArch64::BI_InterlockedOr64:
1349 return MSVCIntrin::_InterlockedOr;
1350 case clang::AArch64::BI_InterlockedXor64:
1351 return MSVCIntrin::_InterlockedXor;
1352 case clang::AArch64::BI_InterlockedDecrement64:
1353 return MSVCIntrin::_InterlockedDecrement;
1354 case clang::AArch64::BI_InterlockedIncrement64:
1355 return MSVCIntrin::_InterlockedIncrement;
1356 case clang::AArch64::BI_InterlockedExchangeAdd8_acq:
1357 case clang::AArch64::BI_InterlockedExchangeAdd16_acq:
1358 case clang::AArch64::BI_InterlockedExchangeAdd_acq:
1359 case clang::AArch64::BI_InterlockedExchangeAdd64_acq:
1360 return MSVCIntrin::_InterlockedExchangeAdd_acq;
1361 case clang::AArch64::BI_InterlockedExchangeAdd8_rel:
1362 case clang::AArch64::BI_InterlockedExchangeAdd16_rel:
1363 case clang::AArch64::BI_InterlockedExchangeAdd_rel:
1364 case clang::AArch64::BI_InterlockedExchangeAdd64_rel:
1365 return MSVCIntrin::_InterlockedExchangeAdd_rel;
1366 case clang::AArch64::BI_InterlockedExchangeAdd8_nf:
1367 case clang::AArch64::BI_InterlockedExchangeAdd16_nf:
1368 case clang::AArch64::BI_InterlockedExchangeAdd_nf:
1369 case clang::AArch64::BI_InterlockedExchangeAdd64_nf:
1370 return MSVCIntrin::_InterlockedExchangeAdd_nf;
1371 case clang::AArch64::BI_InterlockedExchange8_acq:
1372 case clang::AArch64::BI_InterlockedExchange16_acq:
1373 case clang::AArch64::BI_InterlockedExchange_acq:
1374 case clang::AArch64::BI_InterlockedExchange64_acq:
1375 return MSVCIntrin::_InterlockedExchange_acq;
1376 case clang::AArch64::BI_InterlockedExchange8_rel:
1377 case clang::AArch64::BI_InterlockedExchange16_rel:
1378 case clang::AArch64::BI_InterlockedExchange_rel:
1379 case clang::AArch64::BI_InterlockedExchange64_rel:
1380 return MSVCIntrin::_InterlockedExchange_rel;
1381 case clang::AArch64::BI_InterlockedExchange8_nf:
1382 case clang::AArch64::BI_InterlockedExchange16_nf:
1383 case clang::AArch64::BI_InterlockedExchange_nf:
1384 case clang::AArch64::BI_InterlockedExchange64_nf:
1385 return MSVCIntrin::_InterlockedExchange_nf;
1386 case clang::AArch64::BI_InterlockedCompareExchange8_acq:
1387 case clang::AArch64::BI_InterlockedCompareExchange16_acq:
1388 case clang::AArch64::BI_InterlockedCompareExchange_acq:
1389 case clang::AArch64::BI_InterlockedCompareExchange64_acq:
1390 return MSVCIntrin::_InterlockedCompareExchange_acq;
1391 case clang::AArch64::BI_InterlockedCompareExchange8_rel:
1392 case clang::AArch64::BI_InterlockedCompareExchange16_rel:
1393 case clang::AArch64::BI_InterlockedCompareExchange_rel:
1394 case clang::AArch64::BI_InterlockedCompareExchange64_rel:
1395 return MSVCIntrin::_InterlockedCompareExchange_rel;
1396 case clang::AArch64::BI_InterlockedCompareExchange8_nf:
1397 case clang::AArch64::BI_InterlockedCompareExchange16_nf:
1398 case clang::AArch64::BI_InterlockedCompareExchange_nf:
1399 case clang::AArch64::BI_InterlockedCompareExchange64_nf:
1400 return MSVCIntrin::_InterlockedCompareExchange_nf;
1401 case clang::AArch64::BI_InterlockedCompareExchange128:
1402 return MSVCIntrin::_InterlockedCompareExchange128;
1403 case clang::AArch64::BI_InterlockedCompareExchange128_acq:
1404 return MSVCIntrin::_InterlockedCompareExchange128_acq;
1405 case clang::AArch64::BI_InterlockedCompareExchange128_nf:
1406 return MSVCIntrin::_InterlockedCompareExchange128_nf;
1407 case clang::AArch64::BI_InterlockedCompareExchange128_rel:
1408 return MSVCIntrin::_InterlockedCompareExchange128_rel;
1409 case clang::AArch64::BI_InterlockedOr8_acq:
1410 case clang::AArch64::BI_InterlockedOr16_acq:
1411 case clang::AArch64::BI_InterlockedOr_acq:
1412 case clang::AArch64::BI_InterlockedOr64_acq:
1413 return MSVCIntrin::_InterlockedOr_acq;
1414 case clang::AArch64::BI_InterlockedOr8_rel:
1415 case clang::AArch64::BI_InterlockedOr16_rel:
1416 case clang::AArch64::BI_InterlockedOr_rel:
1417 case clang::AArch64::BI_InterlockedOr64_rel:
1418 return MSVCIntrin::_InterlockedOr_rel;
1419 case clang::AArch64::BI_InterlockedOr8_nf:
1420 case clang::AArch64::BI_InterlockedOr16_nf:
1421 case clang::AArch64::BI_InterlockedOr_nf:
1422 case clang::AArch64::BI_InterlockedOr64_nf:
1423 return MSVCIntrin::_InterlockedOr_nf;
1424 case clang::AArch64::BI_InterlockedXor8_acq:
1425 case clang::AArch64::BI_InterlockedXor16_acq:
1426 case clang::AArch64::BI_InterlockedXor_acq:
1427 case clang::AArch64::BI_InterlockedXor64_acq:
1428 return MSVCIntrin::_InterlockedXor_acq;
1429 case clang::AArch64::BI_InterlockedXor8_rel:
1430 case clang::AArch64::BI_InterlockedXor16_rel:
1431 case clang::AArch64::BI_InterlockedXor_rel:
1432 case clang::AArch64::BI_InterlockedXor64_rel:
1433 return MSVCIntrin::_InterlockedXor_rel;
1434 case clang::AArch64::BI_InterlockedXor8_nf:
1435 case clang::AArch64::BI_InterlockedXor16_nf:
1436 case clang::AArch64::BI_InterlockedXor_nf:
1437 case clang::AArch64::BI_InterlockedXor64_nf:
1438 return MSVCIntrin::_InterlockedXor_nf;
1439 case clang::AArch64::BI_InterlockedAnd8_acq:
1440 case clang::AArch64::BI_InterlockedAnd16_acq:
1441 case clang::AArch64::BI_InterlockedAnd_acq:
1442 case clang::AArch64::BI_InterlockedAnd64_acq:
1443 return MSVCIntrin::_InterlockedAnd_acq;
1444 case clang::AArch64::BI_InterlockedAnd8_rel:
1445 case clang::AArch64::BI_InterlockedAnd16_rel:
1446 case clang::AArch64::BI_InterlockedAnd_rel:
1447 case clang::AArch64::BI_InterlockedAnd64_rel:
1448 return MSVCIntrin::_InterlockedAnd_rel;
1449 case clang::AArch64::BI_InterlockedAnd8_nf:
1450 case clang::AArch64::BI_InterlockedAnd16_nf:
1451 case clang::AArch64::BI_InterlockedAnd_nf:
1452 case clang::AArch64::BI_InterlockedAnd64_nf:
1453 return MSVCIntrin::_InterlockedAnd_nf;
1454 case clang::AArch64::BI_InterlockedIncrement16_acq:
1455 case clang::AArch64::BI_InterlockedIncrement_acq:
1456 case clang::AArch64::BI_InterlockedIncrement64_acq:
1457 return MSVCIntrin::_InterlockedIncrement_acq;
1458 case clang::AArch64::BI_InterlockedIncrement16_rel:
1459 case clang::AArch64::BI_InterlockedIncrement_rel:
1460 case clang::AArch64::BI_InterlockedIncrement64_rel:
1461 return MSVCIntrin::_InterlockedIncrement_rel;
1462 case clang::AArch64::BI_InterlockedIncrement16_nf:
1463 case clang::AArch64::BI_InterlockedIncrement_nf:
1464 case clang::AArch64::BI_InterlockedIncrement64_nf:
1465 return MSVCIntrin::_InterlockedIncrement_nf;
1466 case clang::AArch64::BI_InterlockedDecrement16_acq:
1467 case clang::AArch64::BI_InterlockedDecrement_acq:
1468 case clang::AArch64::BI_InterlockedDecrement64_acq:
1469 return MSVCIntrin::_InterlockedDecrement_acq;
1470 case clang::AArch64::BI_InterlockedDecrement16_rel:
1471 case clang::AArch64::BI_InterlockedDecrement_rel:
1472 case clang::AArch64::BI_InterlockedDecrement64_rel:
1473 return MSVCIntrin::_InterlockedDecrement_rel;
1474 case clang::AArch64::BI_InterlockedDecrement16_nf:
1475 case clang::AArch64::BI_InterlockedDecrement_nf:
1476 case clang::AArch64::BI_InterlockedDecrement64_nf:
1477 return MSVCIntrin::_InterlockedDecrement_nf;
1478 }
1479 llvm_unreachable("must return from switch");
1480 }
1481
1482 static std::optional<CodeGenFunction::MSVCIntrin>
1483 translateX86ToMsvcIntrin(unsigned BuiltinID) {
1484 using MSVCIntrin = CodeGenFunction::MSVCIntrin;
1485 switch (BuiltinID) {
1486 default:
1487 return std::nullopt;
1488 case clang::X86::BI_BitScanForward:
1489 case clang::X86::BI_BitScanForward64:
1490 return MSVCIntrin::_BitScanForward;
1491 case clang::X86::BI_BitScanReverse:
1492 case clang::X86::BI_BitScanReverse64:
1493 return MSVCIntrin::_BitScanReverse;
1494 case clang::X86::BI_InterlockedAnd64:
1495 return MSVCIntrin::_InterlockedAnd;
1496 case clang::X86::BI_InterlockedCompareExchange128:
1497 return MSVCIntrin::_InterlockedCompareExchange128;
1498 case clang::X86::BI_InterlockedExchange64:
1499 return MSVCIntrin::_InterlockedExchange;
1500 case clang::X86::BI_InterlockedExchangeAdd64:
1501 return MSVCIntrin::_InterlockedExchangeAdd;
1502 case clang::X86::BI_InterlockedExchangeSub64:
1503 return MSVCIntrin::_InterlockedExchangeSub;
1504 case clang::X86::BI_InterlockedOr64:
1505 return MSVCIntrin::_InterlockedOr;
1506 case clang::X86::BI_InterlockedXor64:
1507 return MSVCIntrin::_InterlockedXor;
1508 case clang::X86::BI_InterlockedDecrement64:
1509 return MSVCIntrin::_InterlockedDecrement;
1510 case clang::X86::BI_InterlockedIncrement64:
1511 return MSVCIntrin::_InterlockedIncrement;
1512 }
1513 llvm_unreachable("must return from switch");
1514 }
1515
1516 // Emit an MSVC intrinsic. Assumes that arguments have *not* been evaluated.
1517 Value *CodeGenFunction::EmitMSVCBuiltinExpr(MSVCIntrin BuiltinID,
1518 const CallExpr *E) {
1519 switch (BuiltinID) {
1520 case MSVCIntrin::_BitScanForward:
1521 case MSVCIntrin::_BitScanReverse: {
1522 Address IndexAddress(EmitPointerWithAlignment(E->getArg(0)));
1523 Value *ArgValue = EmitScalarExpr(E->getArg(1));
1524
1525 llvm::Type *ArgType = ArgValue->getType();
1526 llvm::Type *IndexType = IndexAddress.getElementType();
1527 llvm::Type *ResultType = ConvertType(E->getType());
1528
1529 Value *ArgZero = llvm::Constant::getNullValue(ArgType);
1530 Value *ResZero = llvm::Constant::getNullValue(ResultType);
1531 Value *ResOne = llvm::ConstantInt::get(ResultType, 1);
1532
1533 BasicBlock *Begin = Builder.GetInsertBlock();
1534 BasicBlock *End = createBasicBlock("bitscan_end", this->CurFn);
1535 Builder.SetInsertPoint(End);
1536 PHINode *Result = Builder.CreatePHI(ResultType, 2, "bitscan_result");
1537
1538 Builder.SetInsertPoint(Begin);
1539 Value *IsZero = Builder.CreateICmpEQ(ArgValue, ArgZero);
1540 BasicBlock *NotZero = createBasicBlock("bitscan_not_zero", this->CurFn);
1541 Builder.CreateCondBr(IsZero, End, NotZero);
1542 Result->addIncoming(ResZero, Begin);
1543
1544 Builder.SetInsertPoint(NotZero);
1545
1546 if (BuiltinID == MSVCIntrin::_BitScanForward) {
1547 Function *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType);
1548 Value *ZeroCount = Builder.CreateCall(F, {ArgValue, Builder.getTrue()});
1549 ZeroCount = Builder.CreateIntCast(ZeroCount, IndexType, false);
1550 Builder.CreateStore(ZeroCount, IndexAddress, false);
1551 } else {
1552 unsigned ArgWidth = cast<llvm::IntegerType>(ArgType)->getBitWidth();
1553 Value *ArgTypeLastIndex = llvm::ConstantInt::get(IndexType, ArgWidth - 1);
1554
1555 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType);
1556 Value *ZeroCount = Builder.CreateCall(F, {ArgValue, Builder.getTrue()});
1557 ZeroCount = Builder.CreateIntCast(ZeroCount, IndexType, false);
1558 Value *Index = Builder.CreateNSWSub(ArgTypeLastIndex, ZeroCount);
1559 Builder.CreateStore(Index, IndexAddress, false);
1560 }
1561 Builder.CreateBr(End);
1562 Result->addIncoming(ResOne, NotZero);
1563
1564 Builder.SetInsertPoint(End);
1565 return Result;
1566 }
1567 case MSVCIntrin::_InterlockedAnd:
1568 return MakeBinaryAtomicValue(*this, AtomicRMWInst::And, E);
1569 case MSVCIntrin::_InterlockedExchange:
1570 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xchg, E);
1571 case MSVCIntrin::_InterlockedExchangeAdd:
1572 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Add, E);
1573 case MSVCIntrin::_InterlockedExchangeSub:
1574 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Sub, E);
1575 case MSVCIntrin::_InterlockedOr:
1576 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Or, E);
1577 case MSVCIntrin::_InterlockedXor:
1578 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xor, E);
1579 case MSVCIntrin::_InterlockedExchangeAdd_acq:
1580 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Add, E,
1581 AtomicOrdering::Acquire);
1582 case MSVCIntrin::_InterlockedExchangeAdd_rel:
1583 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Add, E,
1584 AtomicOrdering::Release);
1585 case MSVCIntrin::_InterlockedExchangeAdd_nf:
1586 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Add, E,
1587 AtomicOrdering::Monotonic);
1588 case MSVCIntrin::_InterlockedExchange_acq:
1589 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xchg, E,
1590 AtomicOrdering::Acquire);
1591 case MSVCIntrin::_InterlockedExchange_rel:
1592 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xchg, E,
1593 AtomicOrdering::Release);
1594 case MSVCIntrin::_InterlockedExchange_nf:
1595 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xchg, E,
1596 AtomicOrdering::Monotonic);
1597 case MSVCIntrin::_InterlockedCompareExchange_acq:
1598 return EmitAtomicCmpXchgForMSIntrin(*this, E, AtomicOrdering::Acquire);
1599 case MSVCIntrin::_InterlockedCompareExchange_rel:
1600 return EmitAtomicCmpXchgForMSIntrin(*this, E, AtomicOrdering::Release);
1601 case MSVCIntrin::_InterlockedCompareExchange_nf:
1602 return EmitAtomicCmpXchgForMSIntrin(*this, E, AtomicOrdering::Monotonic);
1603 case MSVCIntrin::_InterlockedCompareExchange128:
1604 return EmitAtomicCmpXchg128ForMSIntrin(
1605 *this, E, AtomicOrdering::SequentiallyConsistent);
1606 case MSVCIntrin::_InterlockedCompareExchange128_acq:
1607 return EmitAtomicCmpXchg128ForMSIntrin(*this, E, AtomicOrdering::Acquire);
1608 case MSVCIntrin::_InterlockedCompareExchange128_rel:
1609 return EmitAtomicCmpXchg128ForMSIntrin(*this, E, AtomicOrdering::Release);
1610 case MSVCIntrin::_InterlockedCompareExchange128_nf:
1611 return EmitAtomicCmpXchg128ForMSIntrin(*this, E, AtomicOrdering::Monotonic);
1612 case MSVCIntrin::_InterlockedOr_acq:
1613 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Or, E,
1614 AtomicOrdering::Acquire);
1615 case MSVCIntrin::_InterlockedOr_rel:
1616 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Or, E,
1617 AtomicOrdering::Release);
1618 case MSVCIntrin::_InterlockedOr_nf:
1619 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Or, E,
1620 AtomicOrdering::Monotonic);
1621 case MSVCIntrin::_InterlockedXor_acq:
1622 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xor, E,
1623 AtomicOrdering::Acquire);
1624 case MSVCIntrin::_InterlockedXor_rel:
1625 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xor, E,
1626 AtomicOrdering::Release);
1627 case MSVCIntrin::_InterlockedXor_nf:
1628 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xor, E,
1629 AtomicOrdering::Monotonic);
1630 case MSVCIntrin::_InterlockedAnd_acq:
1631 return MakeBinaryAtomicValue(*this, AtomicRMWInst::And, E,
1632 AtomicOrdering::Acquire);
1633 case MSVCIntrin::_InterlockedAnd_rel:
1634 return MakeBinaryAtomicValue(*this, AtomicRMWInst::And, E,
1635 AtomicOrdering::Release);
1636 case MSVCIntrin::_InterlockedAnd_nf:
1637 return MakeBinaryAtomicValue(*this, AtomicRMWInst::And, E,
1638 AtomicOrdering::Monotonic);
1639 case MSVCIntrin::_InterlockedIncrement_acq:
1640 return EmitAtomicIncrementValue(*this, E, AtomicOrdering::Acquire);
1641 case MSVCIntrin::_InterlockedIncrement_rel:
1642 return EmitAtomicIncrementValue(*this, E, AtomicOrdering::Release);
1643 case MSVCIntrin::_InterlockedIncrement_nf:
1644 return EmitAtomicIncrementValue(*this, E, AtomicOrdering::Monotonic);
1645 case MSVCIntrin::_InterlockedDecrement_acq:
1646 return EmitAtomicDecrementValue(*this, E, AtomicOrdering::Acquire);
1647 case MSVCIntrin::_InterlockedDecrement_rel:
1648 return EmitAtomicDecrementValue(*this, E, AtomicOrdering::Release);
1649 case MSVCIntrin::_InterlockedDecrement_nf:
1650 return EmitAtomicDecrementValue(*this, E, AtomicOrdering::Monotonic);
1651
1652 case MSVCIntrin::_InterlockedDecrement:
1653 return EmitAtomicDecrementValue(*this, E);
1654 case MSVCIntrin::_InterlockedIncrement:
1655 return EmitAtomicIncrementValue(*this, E);
1656
1657 case MSVCIntrin::__fastfail: {
1658 // Request immediate process termination from the kernel. The instruction
1659 // sequences to do this are documented on MSDN:
1660 // https://msdn.microsoft.com/en-us/library/dn774154.aspx
1661 llvm::Triple::ArchType ISA = getTarget().getTriple().getArch();
1662 StringRef Asm, Constraints;
1663 switch (ISA) {
1664 default:
1665 ErrorUnsupported(E, "__fastfail call for this architecture");
1666 break;
1667 case llvm::Triple::x86:
1668 case llvm::Triple::x86_64:
1669 Asm = "int $$0x29";
1670 Constraints = "{cx}";
1671 break;
1672 case llvm::Triple::thumb:
1673 Asm = "udf #251";
1674 Constraints = "{r0}";
1675 break;
1676 case llvm::Triple::aarch64:
1677 Asm = "brk #0xF003";
1678 Constraints = "{w0}";
1679 }
1680 llvm::FunctionType *FTy = llvm::FunctionType::get(VoidTy, {Int32Ty}, false);
1681 llvm::InlineAsm *IA =
1682 llvm::InlineAsm::get(FTy, Asm, Constraints, /*hasSideEffects=*/true);
1683 llvm::AttributeList NoReturnAttr = llvm::AttributeList::get(
1684 getLLVMContext(), llvm::AttributeList::FunctionIndex,
1685 llvm::Attribute::NoReturn);
1686 llvm::CallInst *CI = Builder.CreateCall(IA, EmitScalarExpr(E->getArg(0)));
1687 CI->setAttributes(NoReturnAttr);
1688 return CI;
1689 }
1690 }
1691 llvm_unreachable("Incorrect MSVC intrinsic!");
1692 }
1693
1694 namespace {
1695 // ARC cleanup for __builtin_os_log_format
1696 struct CallObjCArcUse final : EHScopeStack::Cleanup {
1697 CallObjCArcUse(llvm::Value *object) : object(object) {}
1698 llvm::Value *object;
1699
1700 void Emit(CodeGenFunction &CGF, Flags flags) override {
1701 CGF.EmitARCIntrinsicUse(object);
1702 }
1703 };
1704 }
1705
1706 Value *CodeGenFunction::EmitCheckedArgForBuiltin(const Expr *E,
1707 BuiltinCheckKind Kind) {
1708 assert((Kind == BCK_CLZPassedZero || Kind == BCK_CTZPassedZero)
1709 && "Unsupported builtin check kind");
1710
1711 Value *ArgValue = EmitScalarExpr(E);
1712 if (!SanOpts.has(SanitizerKind::Builtin) || !getTarget().isCLZForZeroUndef())
1713 return ArgValue;
1714
1715 SanitizerScope SanScope(this);
1716 Value *Cond = Builder.CreateICmpNE(
1717 ArgValue, llvm::Constant::getNullValue(ArgValue->getType()));
1718 EmitCheck(std::make_pair(Cond, SanitizerKind::Builtin),
1719 SanitizerHandler::InvalidBuiltin,
1720 {EmitCheckSourceLocation(E->getExprLoc()),
1721 llvm::ConstantInt::get(Builder.getInt8Ty(), Kind)},
1722 std::nullopt);
1723 return ArgValue;
1724 }
1725
1726 /// Get the argument type for arguments to os_log_helper.
1727 static CanQualType getOSLogArgType(ASTContext &C, int Size) {
1728 QualType UnsignedTy = C.getIntTypeForBitwidth(Size * 8, /*Signed=*/false);
1729 return C.getCanonicalType(UnsignedTy);
1730 }
1731
1732 llvm::Function *CodeGenFunction::generateBuiltinOSLogHelperFunction(
1733 const analyze_os_log::OSLogBufferLayout &Layout,
1734 CharUnits BufferAlignment) {
1735 ASTContext &Ctx = getContext();
1736
1737 llvm::SmallString<64> Name;
1738 {
1739 raw_svector_ostream OS(Name);
1740 OS << "__os_log_helper";
1741 OS << "_" << BufferAlignment.getQuantity();
1742 OS << "_" << int(Layout.getSummaryByte());
1743 OS << "_" << int(Layout.getNumArgsByte());
1744 for (const auto &Item : Layout.Items)
1745 OS << "_" << int(Item.getSizeByte()) << "_"
1746 << int(Item.getDescriptorByte());
1747 }
1748
1749 if (llvm::Function *F = CGM.getModule().getFunction(Name))
1750 return F;
1751
1752 llvm::SmallVector<QualType, 4> ArgTys;
1753 FunctionArgList Args;
1754 Args.push_back(ImplicitParamDecl::Create(
1755 Ctx, nullptr, SourceLocation(), &Ctx.Idents.get("buffer"), Ctx.VoidPtrTy,
1756 ImplicitParamDecl::Other));
1757 ArgTys.emplace_back(Ctx.VoidPtrTy);
1758
1759 for (unsigned int I = 0, E = Layout.Items.size(); I < E; ++I) {
1760 char Size = Layout.Items[I].getSizeByte();
1761 if (!Size)
1762 continue;
1763
1764 QualType ArgTy = getOSLogArgType(Ctx, Size);
1765 Args.push_back(ImplicitParamDecl::Create(
1766 Ctx, nullptr, SourceLocation(),
1767 &Ctx.Idents.get(std::string("arg") + llvm::to_string(I)), ArgTy,
1768 ImplicitParamDecl::Other));
1769 ArgTys.emplace_back(ArgTy);
1770 }
1771
1772 QualType ReturnTy = Ctx.VoidTy;
1773
1774 // The helper function has linkonce_odr linkage to enable the linker to merge
1775 // identical functions. To ensure the merging always happens, 'noinline' is
1776 // attached to the function when compiling with -Oz.
1777 const CGFunctionInfo &FI =
1778 CGM.getTypes().arrangeBuiltinFunctionDeclaration(ReturnTy, Args);
1779 llvm::FunctionType *FuncTy = CGM.getTypes().GetFunctionType(FI);
1780 llvm::Function *Fn = llvm::Function::Create(
1781 FuncTy, llvm::GlobalValue::LinkOnceODRLinkage, Name, &CGM.getModule());
1782 Fn->setVisibility(llvm::GlobalValue::HiddenVisibility);
1783 CGM.SetLLVMFunctionAttributes(GlobalDecl(), FI, Fn, /*IsThunk=*/false);
1784 CGM.SetLLVMFunctionAttributesForDefinition(nullptr, Fn);
1785 Fn->setDoesNotThrow();
1786
1787 // Attach 'noinline' at -Oz.
1788 if (CGM.getCodeGenOpts().OptimizeSize == 2)
1789 Fn->addFnAttr(llvm::Attribute::NoInline);
1790
1791 auto NL = ApplyDebugLocation::CreateEmpty(*this);
1792 StartFunction(GlobalDecl(), ReturnTy, Fn, FI, Args);
1793
1794 // Create a scope with an artificial location for the body of this function.
1795 auto AL = ApplyDebugLocation::CreateArtificial(*this);
1796
1797 CharUnits Offset;
1798 Address BufAddr =
1799 Address(Builder.CreateLoad(GetAddrOfLocalVar(Args[0]), "buf"), Int8Ty,
1800 BufferAlignment);
1801 Builder.CreateStore(Builder.getInt8(Layout.getSummaryByte()),
1802 Builder.CreateConstByteGEP(BufAddr, Offset++, "summary"));
1803 Builder.CreateStore(Builder.getInt8(Layout.getNumArgsByte()),
1804 Builder.CreateConstByteGEP(BufAddr, Offset++, "numArgs"));
1805
1806 unsigned I = 1;
1807 for (const auto &Item : Layout.Items) {
1808 Builder.CreateStore(
1809 Builder.getInt8(Item.getDescriptorByte()),
1810 Builder.CreateConstByteGEP(BufAddr, Offset++, "argDescriptor"));
1811 Builder.CreateStore(
1812 Builder.getInt8(Item.getSizeByte()),
1813 Builder.CreateConstByteGEP(BufAddr, Offset++, "argSize"));
1814
1815 CharUnits Size = Item.size();
1816 if (!Size.getQuantity())
1817 continue;
1818
1819 Address Arg = GetAddrOfLocalVar(Args[I]);
1820 Address Addr = Builder.CreateConstByteGEP(BufAddr, Offset, "argData");
1821 Addr =
1822 Builder.CreateElementBitCast(Addr, Arg.getElementType(), "argDataCast");
1823 Builder.CreateStore(Builder.CreateLoad(Arg), Addr);
1824 Offset += Size;
1825 ++I;
1826 }
1827
1828 FinishFunction();
1829
1830 return Fn;
1831 }
1832
1833 RValue CodeGenFunction::emitBuiltinOSLogFormat(const CallExpr &E) {
1834 assert(E.getNumArgs() >= 2 &&
1835 "__builtin_os_log_format takes at least 2 arguments");
1836 ASTContext &Ctx = getContext();
1837 analyze_os_log::OSLogBufferLayout Layout;
1838 analyze_os_log::computeOSLogBufferLayout(Ctx, &E, Layout);
1839 Address BufAddr = EmitPointerWithAlignment(E.getArg(0));
1840 llvm::SmallVector<llvm::Value *, 4> RetainableOperands;
1841
1842 // Ignore argument 1, the format string. It is not currently used.
1843 CallArgList Args;
1844 Args.add(RValue::get(BufAddr.getPointer()), Ctx.VoidPtrTy);
1845
1846 for (const auto &Item : Layout.Items) {
1847 int Size = Item.getSizeByte();
1848 if (!Size)
1849 continue;
1850
1851 llvm::Value *ArgVal;
1852
1853 if (Item.getKind() == analyze_os_log::OSLogBufferItem::MaskKind) {
1854 uint64_t Val = 0;
1855 for (unsigned I = 0, E = Item.getMaskType().size(); I < E; ++I)
1856 Val |= ((uint64_t)Item.getMaskType()[I]) << I * 8;
1857 ArgVal = llvm::Constant::getIntegerValue(Int64Ty, llvm::APInt(64, Val));
1858 } else if (const Expr *TheExpr = Item.getExpr()) {
1859 ArgVal = EmitScalarExpr(TheExpr, /*Ignore*/ false);
1860
1861 // If a temporary object that requires destruction after the full
1862 // expression is passed, push a lifetime-extended cleanup to extend its
1863 // lifetime to the end of the enclosing block scope.
1864 auto LifetimeExtendObject = [&](const Expr *E) {
1865 E = E->IgnoreParenCasts();
1866 // Extend lifetimes of objects returned by function calls and message
1867 // sends.
1868
1869 // FIXME: We should do this in other cases in which temporaries are
1870 // created including arguments of non-ARC types (e.g., C++
1871 // temporaries).
1872 if (isa<CallExpr>(E) || isa<ObjCMessageExpr>(E))
1873 return true;
1874 return false;
1875 };
1876
1877 if (TheExpr->getType()->isObjCRetainableType() &&
1878 getLangOpts().ObjCAutoRefCount && LifetimeExtendObject(TheExpr)) {
1879 assert(getEvaluationKind(TheExpr->getType()) == TEK_Scalar &&
1880 "Only scalar can be a ObjC retainable type");
1881 if (!isa<Constant>(ArgVal)) {
1882 CleanupKind Cleanup = getARCCleanupKind();
1883 QualType Ty = TheExpr->getType();
1884 Address Alloca = Address::invalid();
1885 Address Addr = CreateMemTemp(Ty, "os.log.arg", &Alloca);
1886 ArgVal = EmitARCRetain(Ty, ArgVal);
1887 Builder.CreateStore(ArgVal, Addr);
1888 pushLifetimeExtendedDestroy(Cleanup, Alloca, Ty,
1889 CodeGenFunction::destroyARCStrongPrecise,
1890 Cleanup & EHCleanup);
1891
1892 // Push a clang.arc.use call to ensure ARC optimizer knows that the
1893 // argument has to be alive.
1894 if (CGM.getCodeGenOpts().OptimizationLevel != 0)
1895 pushCleanupAfterFullExpr<CallObjCArcUse>(Cleanup, ArgVal);
1896 }
1897 }
1898 } else {
1899 ArgVal = Builder.getInt32(Item.getConstValue().getQuantity());
1900 }
1901
1902 unsigned ArgValSize =
1903 CGM.getDataLayout().getTypeSizeInBits(ArgVal->getType());
1904 llvm::IntegerType *IntTy = llvm::Type::getIntNTy(getLLVMContext(),
1905 ArgValSize);
1906 ArgVal = Builder.CreateBitOrPointerCast(ArgVal, IntTy);
1907 CanQualType ArgTy = getOSLogArgType(Ctx, Size);
1908 // If ArgVal has type x86_fp80, zero-extend ArgVal.
1909 ArgVal = Builder.CreateZExtOrBitCast(ArgVal, ConvertType(ArgTy));
1910 Args.add(RValue::get(ArgVal), ArgTy);
1911 }
1912
1913 const CGFunctionInfo &FI =
1914 CGM.getTypes().arrangeBuiltinFunctionCall(Ctx.VoidTy, Args);
1915 llvm::Function *F = CodeGenFunction(CGM).generateBuiltinOSLogHelperFunction(
1916 Layout, BufAddr.getAlignment());
1917 EmitCall(FI, CGCallee::forDirect(F), ReturnValueSlot(), Args);
1918 return RValue::get(BufAddr.getPointer());
1919 }
1920
1921 static bool isSpecialUnsignedMultiplySignedResult(
1922 unsigned BuiltinID, WidthAndSignedness Op1Info, WidthAndSignedness Op2Info,
1923 WidthAndSignedness ResultInfo) {
1924 return BuiltinID == Builtin::BI__builtin_mul_overflow &&
1925 Op1Info.Width == Op2Info.Width && Op2Info.Width == ResultInfo.Width &&
1926 !Op1Info.Signed && !Op2Info.Signed && ResultInfo.Signed;
1927 }
1928
1929 static RValue EmitCheckedUnsignedMultiplySignedResult(
1930 CodeGenFunction &CGF, const clang::Expr *Op1, WidthAndSignedness Op1Info,
1931 const clang::Expr *Op2, WidthAndSignedness Op2Info,
1932 const clang::Expr *ResultArg, QualType ResultQTy,
1933 WidthAndSignedness ResultInfo) {
1934 assert(isSpecialUnsignedMultiplySignedResult(
1935 Builtin::BI__builtin_mul_overflow, Op1Info, Op2Info, ResultInfo) &&
1936 "Cannot specialize this multiply");
1937
1938 llvm::Value *V1 = CGF.EmitScalarExpr(Op1);
1939 llvm::Value *V2 = CGF.EmitScalarExpr(Op2);
1940
1941 llvm::Value *HasOverflow;
1942 llvm::Value *Result = EmitOverflowIntrinsic(
1943 CGF, llvm::Intrinsic::umul_with_overflow, V1, V2, HasOverflow);
1944
1945 // The intrinsic call will detect overflow when the value is > UINT_MAX,
1946 // however, since the original builtin had a signed result, we need to report
1947 // an overflow when the result is greater than INT_MAX.
1948 auto IntMax = llvm::APInt::getSignedMaxValue(ResultInfo.Width);
1949 llvm::Value *IntMaxValue = llvm::ConstantInt::get(Result->getType(), IntMax);
1950
1951 llvm::Value *IntMaxOverflow = CGF.Builder.CreateICmpUGT(Result, IntMaxValue);
1952 HasOverflow = CGF.Builder.CreateOr(HasOverflow, IntMaxOverflow);
1953
1954 bool isVolatile =
1955 ResultArg->getType()->getPointeeType().isVolatileQualified();
1956 Address ResultPtr = CGF.EmitPointerWithAlignment(ResultArg);
1957 CGF.Builder.CreateStore(CGF.EmitToMemory(Result, ResultQTy), ResultPtr,
1958 isVolatile);
1959 return RValue::get(HasOverflow);
1960 }
1961
1962 /// Determine if a binop is a checked mixed-sign multiply we can specialize.
1963 static bool isSpecialMixedSignMultiply(unsigned BuiltinID,
1964 WidthAndSignedness Op1Info,
1965 WidthAndSignedness Op2Info,
1966 WidthAndSignedness ResultInfo) {
1967 return BuiltinID == Builtin::BI__builtin_mul_overflow &&
1968 std::max(Op1Info.Width, Op2Info.Width) >= ResultInfo.Width &&
1969 Op1Info.Signed != Op2Info.Signed;
1970 }
1971
1972 /// Emit a checked mixed-sign multiply. This is a cheaper specialization of
1973 /// the generic checked-binop irgen.
1974 static RValue
1975 EmitCheckedMixedSignMultiply(CodeGenFunction &CGF, const clang::Expr *Op1,
1976 WidthAndSignedness Op1Info, const clang::Expr *Op2,
1977 WidthAndSignedness Op2Info,
1978 const clang::Expr *ResultArg, QualType ResultQTy,
1979 WidthAndSignedness ResultInfo) {
1980 assert(isSpecialMixedSignMultiply(Builtin::BI__builtin_mul_overflow, Op1Info,
1981 Op2Info, ResultInfo) &&
1982 "Not a mixed-sign multipliction we can specialize");
1983
1984 // Emit the signed and unsigned operands.
1985 const clang::Expr *SignedOp = Op1Info.Signed ? Op1 : Op2;
1986 const clang::Expr *UnsignedOp = Op1Info.Signed ? Op2 : Op1;
1987 llvm::Value *Signed = CGF.EmitScalarExpr(SignedOp);
1988 llvm::Value *Unsigned = CGF.EmitScalarExpr(UnsignedOp);
1989 unsigned SignedOpWidth = Op1Info.Signed ? Op1Info.Width : Op2Info.Width;
1990 unsigned UnsignedOpWidth = Op1Info.Signed ? Op2Info.Width : Op1Info.Width;
1991
1992 // One of the operands may be smaller than the other. If so, [s|z]ext it.
1993 if (SignedOpWidth < UnsignedOpWidth)
1994 Signed = CGF.Builder.CreateSExt(Signed, Unsigned->getType(), "op.sext");
1995 if (UnsignedOpWidth < SignedOpWidth)
1996 Unsigned = CGF.Builder.CreateZExt(Unsigned, Signed->getType(), "op.zext");
1997
1998 llvm::Type *OpTy = Signed->getType();
1999 llvm::Value *Zero = llvm::Constant::getNullValue(OpTy);
2000 Address ResultPtr = CGF.EmitPointerWithAlignment(ResultArg);
2001 llvm::Type *ResTy = ResultPtr.getElementType();
2002 unsigned OpWidth = std::max(Op1Info.Width, Op2Info.Width);
2003
2004 // Take the absolute value of the signed operand.
2005 llvm::Value *IsNegative = CGF.Builder.CreateICmpSLT(Signed, Zero);
2006 llvm::Value *AbsOfNegative = CGF.Builder.CreateSub(Zero, Signed);
2007 llvm::Value *AbsSigned =
2008 CGF.Builder.CreateSelect(IsNegative, AbsOfNegative, Signed);
2009
2010 // Perform a checked unsigned multiplication.
2011 llvm::Value *UnsignedOverflow;
2012 llvm::Value *UnsignedResult =
2013 EmitOverflowIntrinsic(CGF, llvm::Intrinsic::umul_with_overflow, AbsSigned,
2014 Unsigned, UnsignedOverflow);
2015
2016 llvm::Value *Overflow, *Result;
2017 if (ResultInfo.Signed) {
2018 // Signed overflow occurs if the result is greater than INT_MAX or lesser
2019 // than INT_MIN, i.e when |Result| > (INT_MAX + IsNegative).
2020 auto IntMax =
2021 llvm::APInt::getSignedMaxValue(ResultInfo.Width).zext(OpWidth);
2022 llvm::Value *MaxResult =
2023 CGF.Builder.CreateAdd(llvm::ConstantInt::get(OpTy, IntMax),
2024 CGF.Builder.CreateZExt(IsNegative, OpTy));
2025 llvm::Value *SignedOverflow =
2026 CGF.Builder.CreateICmpUGT(UnsignedResult, MaxResult);
2027 Overflow = CGF.Builder.CreateOr(UnsignedOverflow, SignedOverflow);
2028
2029 // Prepare the signed result (possibly by negating it).
2030 llvm::Value *NegativeResult = CGF.Builder.CreateNeg(UnsignedResult);
2031 llvm::Value *SignedResult =
2032 CGF.Builder.CreateSelect(IsNegative, NegativeResult, UnsignedResult);
2033 Result = CGF.Builder.CreateTrunc(SignedResult, ResTy);
2034 } else {
2035 // Unsigned overflow occurs if the result is < 0 or greater than UINT_MAX.
2036 llvm::Value *Underflow = CGF.Builder.CreateAnd(
2037 IsNegative, CGF.Builder.CreateIsNotNull(UnsignedResult));
2038 Overflow = CGF.Builder.CreateOr(UnsignedOverflow, Underflow);
2039 if (ResultInfo.Width < OpWidth) {
2040 auto IntMax =
2041 llvm::APInt::getMaxValue(ResultInfo.Width).zext(OpWidth);
2042 llvm::Value *TruncOverflow = CGF.Builder.CreateICmpUGT(
2043 UnsignedResult, llvm::ConstantInt::get(OpTy, IntMax));
2044 Overflow = CGF.Builder.CreateOr(Overflow, TruncOverflow);
2045 }
2046
2047 // Negate the product if it would be negative in infinite precision.
2048 Result = CGF.Builder.CreateSelect(
2049 IsNegative, CGF.Builder.CreateNeg(UnsignedResult), UnsignedResult);
2050
2051 Result = CGF.Builder.CreateTrunc(Result, ResTy);
2052 }
2053 assert(Overflow && Result && "Missing overflow or result");
2054
2055 bool isVolatile =
2056 ResultArg->getType()->getPointeeType().isVolatileQualified();
2057 CGF.Builder.CreateStore(CGF.EmitToMemory(Result, ResultQTy), ResultPtr,
2058 isVolatile);
2059 return RValue::get(Overflow);
2060 }
2061
2062 static bool
2063 TypeRequiresBuiltinLaunderImp(const ASTContext &Ctx, QualType Ty,
2064 llvm::SmallPtrSetImpl<const Decl *> &Seen) {
2065 if (const auto *Arr = Ctx.getAsArrayType(Ty))
2066 Ty = Ctx.getBaseElementType(Arr);
2067
2068 const auto *Record = Ty->getAsCXXRecordDecl();
2069 if (!Record)
2070 return false;
2071
2072 // We've already checked this type, or are in the process of checking it.
2073 if (!Seen.insert(Record).second)
2074 return false;
2075
2076 assert(Record->hasDefinition() &&
2077 "Incomplete types should already be diagnosed");
2078
2079 if (Record->isDynamicClass())
2080 return true;
2081
2082 for (FieldDecl *F : Record->fields()) {
2083 if (TypeRequiresBuiltinLaunderImp(Ctx, F->getType(), Seen))
2084 return true;
2085 }
2086 return false;
2087 }
2088
2089 /// Determine if the specified type requires laundering by checking if it is a
2090 /// dynamic class type or contains a subobject which is a dynamic class type.
2091 static bool TypeRequiresBuiltinLaunder(CodeGenModule &CGM, QualType Ty) {
2092 if (!CGM.getCodeGenOpts().StrictVTablePointers)
2093 return false;
2094 llvm::SmallPtrSet<const Decl *, 16> Seen;
2095 return TypeRequiresBuiltinLaunderImp(CGM.getContext(), Ty, Seen);
2096 }
2097
2098 RValue CodeGenFunction::emitRotate(const CallExpr *E, bool IsRotateRight) {
2099 llvm::Value *Src = EmitScalarExpr(E->getArg(0));
2100 llvm::Value *ShiftAmt = EmitScalarExpr(E->getArg(1));
2101
2102 // The builtin's shift arg may have a different type than the source arg and
2103 // result, but the LLVM intrinsic uses the same type for all values.
2104 llvm::Type *Ty = Src->getType();
2105 ShiftAmt = Builder.CreateIntCast(ShiftAmt, Ty, false);
2106
2107 // Rotate is a special case of LLVM funnel shift - 1st 2 args are the same.
2108 unsigned IID = IsRotateRight ? Intrinsic::fshr : Intrinsic::fshl;
2109 Function *F = CGM.getIntrinsic(IID, Ty);
2110 return RValue::get(Builder.CreateCall(F, { Src, Src, ShiftAmt }));
2111 }
2112
2113 // Map math builtins for long-double to f128 version.
2114 static unsigned mutateLongDoubleBuiltin(unsigned BuiltinID) {
2115 switch (BuiltinID) {
2116 #define MUTATE_LDBL(func) \
2117 case Builtin::BI__builtin_##func##l: \
2118 return Builtin::BI__builtin_##func##f128;
2119 MUTATE_LDBL(sqrt)
2120 MUTATE_LDBL(cbrt)
2121 MUTATE_LDBL(fabs)
2122 MUTATE_LDBL(log)
2123 MUTATE_LDBL(log2)
2124 MUTATE_LDBL(log10)
2125 MUTATE_LDBL(log1p)
2126 MUTATE_LDBL(logb)
2127 MUTATE_LDBL(exp)
2128 MUTATE_LDBL(exp2)
2129 MUTATE_LDBL(expm1)
2130 MUTATE_LDBL(fdim)
2131 MUTATE_LDBL(hypot)
2132 MUTATE_LDBL(ilogb)
2133 MUTATE_LDBL(pow)
2134 MUTATE_LDBL(fmin)
2135 MUTATE_LDBL(fmax)
2136 MUTATE_LDBL(ceil)
2137 MUTATE_LDBL(trunc)
2138 MUTATE_LDBL(rint)
2139 MUTATE_LDBL(nearbyint)
2140 MUTATE_LDBL(round)
2141 MUTATE_LDBL(floor)
2142 MUTATE_LDBL(lround)
2143 MUTATE_LDBL(llround)
2144 MUTATE_LDBL(lrint)
2145 MUTATE_LDBL(llrint)
2146 MUTATE_LDBL(fmod)
2147 MUTATE_LDBL(modf)
2148 MUTATE_LDBL(nan)
2149 MUTATE_LDBL(nans)
2150 MUTATE_LDBL(inf)
2151 MUTATE_LDBL(fma)
2152 MUTATE_LDBL(sin)
2153 MUTATE_LDBL(cos)
2154 MUTATE_LDBL(tan)
2155 MUTATE_LDBL(sinh)
2156 MUTATE_LDBL(cosh)
2157 MUTATE_LDBL(tanh)
2158 MUTATE_LDBL(asin)
2159 MUTATE_LDBL(acos)
2160 MUTATE_LDBL(atan)
2161 MUTATE_LDBL(asinh)
2162 MUTATE_LDBL(acosh)
2163 MUTATE_LDBL(atanh)
2164 MUTATE_LDBL(atan2)
2165 MUTATE_LDBL(erf)
2166 MUTATE_LDBL(erfc)
2167 MUTATE_LDBL(ldexp)
2168 MUTATE_LDBL(frexp)
2169 MUTATE_LDBL(huge_val)
2170 MUTATE_LDBL(copysign)
2171 MUTATE_LDBL(nextafter)
2172 MUTATE_LDBL(nexttoward)
2173 MUTATE_LDBL(remainder)
2174 MUTATE_LDBL(remquo)
2175 MUTATE_LDBL(scalbln)
2176 MUTATE_LDBL(scalbn)
2177 MUTATE_LDBL(tgamma)
2178 MUTATE_LDBL(lgamma)
2179 #undef MUTATE_LDBL
2180 default:
2181 return BuiltinID;
2182 }
2183 }
2184
2185 RValue CodeGenFunction::EmitBuiltinExpr(const GlobalDecl GD, unsigned BuiltinID,
2186 const CallExpr *E,
2187 ReturnValueSlot ReturnValue) {
2188 const FunctionDecl *FD = GD.getDecl()->getAsFunction();
2189 // See if we can constant fold this builtin. If so, don't emit it at all.
2190 // TODO: Extend this handling to all builtin calls that we can constant-fold.
2191 Expr::EvalResult Result;
2192 if (E->isPRValue() && E->EvaluateAsRValue(Result, CGM.getContext()) &&
2193 !Result.hasSideEffects()) {
2194 if (Result.Val.isInt())
2195 return RValue::get(llvm::ConstantInt::get(getLLVMContext(),
2196 Result.Val.getInt()));
2197 if (Result.Val.isFloat())
2198 return RValue::get(llvm::ConstantFP::get(getLLVMContext(),
2199 Result.Val.getFloat()));
2200 }
2201
2202 // If current long-double semantics is IEEE 128-bit, replace math builtins
2203 // of long-double with f128 equivalent.
2204 // TODO: This mutation should also be applied to other targets other than PPC,
2205 // after backend supports IEEE 128-bit style libcalls.
2206 if (getTarget().getTriple().isPPC64() &&
2207 &getTarget().getLongDoubleFormat() == &llvm::APFloat::IEEEquad())
2208 BuiltinID = mutateLongDoubleBuiltin(BuiltinID);
2209
2210 // If the builtin has been declared explicitly with an assembler label,
2211 // disable the specialized emitting below. Ideally we should communicate the
2212 // rename in IR, or at least avoid generating the intrinsic calls that are
2213 // likely to get lowered to the renamed library functions.
2214 const unsigned BuiltinIDIfNoAsmLabel =
2215 FD->hasAttr<AsmLabelAttr>() ? 0 : BuiltinID;
2216
2217 // There are LLVM math intrinsics/instructions corresponding to math library
2218 // functions except the LLVM op will never set errno while the math library
2219 // might. Also, math builtins have the same semantics as their math library
2220 // twins. Thus, we can transform math library and builtin calls to their
2221 // LLVM counterparts if the call is marked 'const' (known to never set errno).
2222 // In case FP exceptions are enabled, the experimental versions of the
2223 // intrinsics model those.
2224 bool ConstWithoutErrnoAndExceptions =
2225 getContext().BuiltinInfo.isConstWithoutErrnoAndExceptions(BuiltinID);
2226 bool ConstWithoutExceptions =
2227 getContext().BuiltinInfo.isConstWithoutExceptions(BuiltinID);
2228 if (FD->hasAttr<ConstAttr>() ||
2229 ((ConstWithoutErrnoAndExceptions || ConstWithoutExceptions) &&
2230 (!ConstWithoutErrnoAndExceptions || (!getLangOpts().MathErrno)))) {
2231 switch (BuiltinIDIfNoAsmLabel) {
2232 case Builtin::BIceil:
2233 case Builtin::BIceilf:
2234 case Builtin::BIceill:
2235 case Builtin::BI__builtin_ceil:
2236 case Builtin::BI__builtin_ceilf:
2237 case Builtin::BI__builtin_ceilf16:
2238 case Builtin::BI__builtin_ceill:
2239 case Builtin::BI__builtin_ceilf128:
2240 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E,
2241 Intrinsic::ceil,
2242 Intrinsic::experimental_constrained_ceil));
2243
2244 case Builtin::BIcopysign:
2245 case Builtin::BIcopysignf:
2246 case Builtin::BIcopysignl:
2247 case Builtin::BI__builtin_copysign:
2248 case Builtin::BI__builtin_copysignf:
2249 case Builtin::BI__builtin_copysignf16:
2250 case Builtin::BI__builtin_copysignl:
2251 case Builtin::BI__builtin_copysignf128:
2252 return RValue::get(emitBinaryBuiltin(*this, E, Intrinsic::copysign));
2253
2254 case Builtin::BIcos:
2255 case Builtin::BIcosf:
2256 case Builtin::BIcosl:
2257 case Builtin::BI__builtin_cos:
2258 case Builtin::BI__builtin_cosf:
2259 case Builtin::BI__builtin_cosf16:
2260 case Builtin::BI__builtin_cosl:
2261 case Builtin::BI__builtin_cosf128:
2262 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E,
2263 Intrinsic::cos,
2264 Intrinsic::experimental_constrained_cos));
2265
2266 case Builtin::BIexp:
2267 case Builtin::BIexpf:
2268 case Builtin::BIexpl:
2269 case Builtin::BI__builtin_exp:
2270 case Builtin::BI__builtin_expf:
2271 case Builtin::BI__builtin_expf16:
2272 case Builtin::BI__builtin_expl:
2273 case Builtin::BI__builtin_expf128:
2274 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E,
2275 Intrinsic::exp,
2276 Intrinsic::experimental_constrained_exp));
2277
2278 case Builtin::BIexp2:
2279 case Builtin::BIexp2f:
2280 case Builtin::BIexp2l:
2281 case Builtin::BI__builtin_exp2:
2282 case Builtin::BI__builtin_exp2f:
2283 case Builtin::BI__builtin_exp2f16:
2284 case Builtin::BI__builtin_exp2l:
2285 case Builtin::BI__builtin_exp2f128:
2286 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E,
2287 Intrinsic::exp2,
2288 Intrinsic::experimental_constrained_exp2));
2289
2290 case Builtin::BIfabs:
2291 case Builtin::BIfabsf:
2292 case Builtin::BIfabsl:
2293 case Builtin::BI__builtin_fabs:
2294 case Builtin::BI__builtin_fabsf:
2295 case Builtin::BI__builtin_fabsf16:
2296 case Builtin::BI__builtin_fabsl:
2297 case Builtin::BI__builtin_fabsf128:
2298 return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::fabs));
2299
2300 case Builtin::BIfloor:
2301 case Builtin::BIfloorf:
2302 case Builtin::BIfloorl:
2303 case Builtin::BI__builtin_floor:
2304 case Builtin::BI__builtin_floorf:
2305 case Builtin::BI__builtin_floorf16:
2306 case Builtin::BI__builtin_floorl:
2307 case Builtin::BI__builtin_floorf128:
2308 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E,
2309 Intrinsic::floor,
2310 Intrinsic::experimental_constrained_floor));
2311
2312 case Builtin::BIfma:
2313 case Builtin::BIfmaf:
2314 case Builtin::BIfmal:
2315 case Builtin::BI__builtin_fma:
2316 case Builtin::BI__builtin_fmaf:
2317 case Builtin::BI__builtin_fmaf16:
2318 case Builtin::BI__builtin_fmal:
2319 case Builtin::BI__builtin_fmaf128:
2320 return RValue::get(emitTernaryMaybeConstrainedFPBuiltin(*this, E,
2321 Intrinsic::fma,
2322 Intrinsic::experimental_constrained_fma));
2323
2324 case Builtin::BIfmax:
2325 case Builtin::BIfmaxf:
2326 case Builtin::BIfmaxl:
2327 case Builtin::BI__builtin_fmax:
2328 case Builtin::BI__builtin_fmaxf:
2329 case Builtin::BI__builtin_fmaxf16:
2330 case Builtin::BI__builtin_fmaxl:
2331 case Builtin::BI__builtin_fmaxf128:
2332 return RValue::get(emitBinaryMaybeConstrainedFPBuiltin(*this, E,
2333 Intrinsic::maxnum,
2334 Intrinsic::experimental_constrained_maxnum));
2335
2336 case Builtin::BIfmin:
2337 case Builtin::BIfminf:
2338 case Builtin::BIfminl:
2339 case Builtin::BI__builtin_fmin:
2340 case Builtin::BI__builtin_fminf:
2341 case Builtin::BI__builtin_fminf16:
2342 case Builtin::BI__builtin_fminl:
2343 case Builtin::BI__builtin_fminf128:
2344 return RValue::get(emitBinaryMaybeConstrainedFPBuiltin(*this, E,
2345 Intrinsic::minnum,
2346 Intrinsic::experimental_constrained_minnum));
2347
2348 // fmod() is a special-case. It maps to the frem instruction rather than an
2349 // LLVM intrinsic.
2350 case Builtin::BIfmod:
2351 case Builtin::BIfmodf:
2352 case Builtin::BIfmodl:
2353 case Builtin::BI__builtin_fmod:
2354 case Builtin::BI__builtin_fmodf:
2355 case Builtin::BI__builtin_fmodf16:
2356 case Builtin::BI__builtin_fmodl:
2357 case Builtin::BI__builtin_fmodf128: {
2358 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E);
2359 Value *Arg1 = EmitScalarExpr(E->getArg(0));
2360 Value *Arg2 = EmitScalarExpr(E->getArg(1));
2361 return RValue::get(Builder.CreateFRem(Arg1, Arg2, "fmod"));
2362 }
2363
2364 case Builtin::BIlog:
2365 case Builtin::BIlogf:
2366 case Builtin::BIlogl:
2367 case Builtin::BI__builtin_log:
2368 case Builtin::BI__builtin_logf:
2369 case Builtin::BI__builtin_logf16:
2370 case Builtin::BI__builtin_logl:
2371 case Builtin::BI__builtin_logf128:
2372 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E,
2373 Intrinsic::log,
2374 Intrinsic::experimental_constrained_log));
2375
2376 case Builtin::BIlog10:
2377 case Builtin::BIlog10f:
2378 case Builtin::BIlog10l:
2379 case Builtin::BI__builtin_log10:
2380 case Builtin::BI__builtin_log10f:
2381 case Builtin::BI__builtin_log10f16:
2382 case Builtin::BI__builtin_log10l:
2383 case Builtin::BI__builtin_log10f128:
2384 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E,
2385 Intrinsic::log10,
2386 Intrinsic::experimental_constrained_log10));
2387
2388 case Builtin::BIlog2:
2389 case Builtin::BIlog2f:
2390 case Builtin::BIlog2l:
2391 case Builtin::BI__builtin_log2:
2392 case Builtin::BI__builtin_log2f:
2393 case Builtin::BI__builtin_log2f16:
2394 case Builtin::BI__builtin_log2l:
2395 case Builtin::BI__builtin_log2f128:
2396 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E,
2397 Intrinsic::log2,
2398 Intrinsic::experimental_constrained_log2));
2399
2400 case Builtin::BInearbyint:
2401 case Builtin::BInearbyintf:
2402 case Builtin::BInearbyintl:
2403 case Builtin::BI__builtin_nearbyint:
2404 case Builtin::BI__builtin_nearbyintf:
2405 case Builtin::BI__builtin_nearbyintl:
2406 case Builtin::BI__builtin_nearbyintf128:
2407 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E,
2408 Intrinsic::nearbyint,
2409 Intrinsic::experimental_constrained_nearbyint));
2410
2411 case Builtin::BIpow:
2412 case Builtin::BIpowf:
2413 case Builtin::BIpowl:
2414 case Builtin::BI__builtin_pow:
2415 case Builtin::BI__builtin_powf:
2416 case Builtin::BI__builtin_powf16:
2417 case Builtin::BI__builtin_powl:
2418 case Builtin::BI__builtin_powf128:
2419 return RValue::get(emitBinaryMaybeConstrainedFPBuiltin(*this, E,
2420 Intrinsic::pow,
2421 Intrinsic::experimental_constrained_pow));
2422
2423 case Builtin::BIrint:
2424 case Builtin::BIrintf:
2425 case Builtin::BIrintl:
2426 case Builtin::BI__builtin_rint:
2427 case Builtin::BI__builtin_rintf:
2428 case Builtin::BI__builtin_rintf16:
2429 case Builtin::BI__builtin_rintl:
2430 case Builtin::BI__builtin_rintf128:
2431 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E,
2432 Intrinsic::rint,
2433 Intrinsic::experimental_constrained_rint));
2434
2435 case Builtin::BIround:
2436 case Builtin::BIroundf:
2437 case Builtin::BIroundl:
2438 case Builtin::BI__builtin_round:
2439 case Builtin::BI__builtin_roundf:
2440 case Builtin::BI__builtin_roundf16:
2441 case Builtin::BI__builtin_roundl:
2442 case Builtin::BI__builtin_roundf128:
2443 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E,
2444 Intrinsic::round,
2445 Intrinsic::experimental_constrained_round));
2446
2447 case Builtin::BIsin:
2448 case Builtin::BIsinf:
2449 case Builtin::BIsinl:
2450 case Builtin::BI__builtin_sin:
2451 case Builtin::BI__builtin_sinf:
2452 case Builtin::BI__builtin_sinf16:
2453 case Builtin::BI__builtin_sinl:
2454 case Builtin::BI__builtin_sinf128:
2455 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E,
2456 Intrinsic::sin,
2457 Intrinsic::experimental_constrained_sin));
2458
2459 case Builtin::BIsqrt:
2460 case Builtin::BIsqrtf:
2461 case Builtin::BIsqrtl:
2462 case Builtin::BI__builtin_sqrt:
2463 case Builtin::BI__builtin_sqrtf:
2464 case Builtin::BI__builtin_sqrtf16:
2465 case Builtin::BI__builtin_sqrtl:
2466 case Builtin::BI__builtin_sqrtf128:
2467 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E,
2468 Intrinsic::sqrt,
2469 Intrinsic::experimental_constrained_sqrt));
2470
2471 case Builtin::BItrunc:
2472 case Builtin::BItruncf:
2473 case Builtin::BItruncl:
2474 case Builtin::BI__builtin_trunc:
2475 case Builtin::BI__builtin_truncf:
2476 case Builtin::BI__builtin_truncf16:
2477 case Builtin::BI__builtin_truncl:
2478 case Builtin::BI__builtin_truncf128:
2479 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E,
2480 Intrinsic::trunc,
2481 Intrinsic::experimental_constrained_trunc));
2482
2483 case Builtin::BIlround:
2484 case Builtin::BIlroundf:
2485 case Builtin::BIlroundl:
2486 case Builtin::BI__builtin_lround:
2487 case Builtin::BI__builtin_lroundf:
2488 case Builtin::BI__builtin_lroundl:
2489 case Builtin::BI__builtin_lroundf128:
2490 return RValue::get(emitMaybeConstrainedFPToIntRoundBuiltin(
2491 *this, E, Intrinsic::lround,
2492 Intrinsic::experimental_constrained_lround));
2493
2494 case Builtin::BIllround:
2495 case Builtin::BIllroundf:
2496 case Builtin::BIllroundl:
2497 case Builtin::BI__builtin_llround:
2498 case Builtin::BI__builtin_llroundf:
2499 case Builtin::BI__builtin_llroundl:
2500 case Builtin::BI__builtin_llroundf128:
2501 return RValue::get(emitMaybeConstrainedFPToIntRoundBuiltin(
2502 *this, E, Intrinsic::llround,
2503 Intrinsic::experimental_constrained_llround));
2504
2505 case Builtin::BIlrint:
2506 case Builtin::BIlrintf:
2507 case Builtin::BIlrintl:
2508 case Builtin::BI__builtin_lrint:
2509 case Builtin::BI__builtin_lrintf:
2510 case Builtin::BI__builtin_lrintl:
2511 case Builtin::BI__builtin_lrintf128:
2512 return RValue::get(emitMaybeConstrainedFPToIntRoundBuiltin(
2513 *this, E, Intrinsic::lrint,
2514 Intrinsic::experimental_constrained_lrint));
2515
2516 case Builtin::BIllrint:
2517 case Builtin::BIllrintf:
2518 case Builtin::BIllrintl:
2519 case Builtin::BI__builtin_llrint:
2520 case Builtin::BI__builtin_llrintf:
2521 case Builtin::BI__builtin_llrintl:
2522 case Builtin::BI__builtin_llrintf128:
2523 return RValue::get(emitMaybeConstrainedFPToIntRoundBuiltin(
2524 *this, E, Intrinsic::llrint,
2525 Intrinsic::experimental_constrained_llrint));
2526
2527 default:
2528 break;
2529 }
2530 }
2531
2532 switch (BuiltinIDIfNoAsmLabel) {
2533 default: break;
2534 case Builtin::BI__builtin___CFStringMakeConstantString:
2535 case Builtin::BI__builtin___NSStringMakeConstantString:
2536 return RValue::get(ConstantEmitter(*this).emitAbstract(E, E->getType()));
2537 case Builtin::BI__builtin_stdarg_start:
2538 case Builtin::BI__builtin_va_start:
2539 case Builtin::BI__va_start:
2540 case Builtin::BI__builtin_va_end:
2541 EmitVAStartEnd(BuiltinID == Builtin::BI__va_start
2542 ? EmitScalarExpr(E->getArg(0))
2543 : EmitVAListRef(E->getArg(0)).getPointer(),
2544 BuiltinID != Builtin::BI__builtin_va_end);
2545 return RValue::get(nullptr);
2546 case Builtin::BI__builtin_va_copy: {
2547 Value *DstPtr = EmitVAListRef(E->getArg(0)).getPointer();
2548 Value *SrcPtr = EmitVAListRef(E->getArg(1)).getPointer();
2549
2550 llvm::Type *Type = Int8PtrTy;
2551
2552 DstPtr = Builder.CreateBitCast(DstPtr, Type);
2553 SrcPtr = Builder.CreateBitCast(SrcPtr, Type);
2554 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::vacopy), {DstPtr, SrcPtr});
2555 return RValue::get(nullptr);
2556 }
2557 case Builtin::BI__builtin_abs:
2558 case Builtin::BI__builtin_labs:
2559 case Builtin::BI__builtin_llabs: {
2560 // X < 0 ? -X : X
2561 // The negation has 'nsw' because abs of INT_MIN is undefined.
2562 Value *ArgValue = EmitScalarExpr(E->getArg(0));
2563 Value *NegOp = Builder.CreateNSWNeg(ArgValue, "neg");
2564 Constant *Zero = llvm::Constant::getNullValue(ArgValue->getType());
2565 Value *CmpResult = Builder.CreateICmpSLT(ArgValue, Zero, "abscond");
2566 Value *Result = Builder.CreateSelect(CmpResult, NegOp, ArgValue, "abs");
2567 return RValue::get(Result);
2568 }
2569 case Builtin::BI__builtin_complex: {
2570 Value *Real = EmitScalarExpr(E->getArg(0));
2571 Value *Imag = EmitScalarExpr(E->getArg(1));
2572 return RValue::getComplex({Real, Imag});
2573 }
2574 case Builtin::BI__builtin_conj:
2575 case Builtin::BI__builtin_conjf:
2576 case Builtin::BI__builtin_conjl:
2577 case Builtin::BIconj:
2578 case Builtin::BIconjf:
2579 case Builtin::BIconjl: {
2580 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0));
2581 Value *Real = ComplexVal.first;
2582 Value *Imag = ComplexVal.second;
2583 Imag = Builder.CreateFNeg(Imag, "neg");
2584 return RValue::getComplex(std::make_pair(Real, Imag));
2585 }
2586 case Builtin::BI__builtin_creal:
2587 case Builtin::BI__builtin_crealf:
2588 case Builtin::BI__builtin_creall:
2589 case Builtin::BIcreal:
2590 case Builtin::BIcrealf:
2591 case Builtin::BIcreall: {
2592 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0));
2593 return RValue::get(ComplexVal.first);
2594 }
2595
2596 case Builtin::BI__builtin_preserve_access_index: {
2597 // Only enabled preserved access index region when debuginfo
2598 // is available as debuginfo is needed to preserve user-level
2599 // access pattern.
2600 if (!getDebugInfo()) {
2601 CGM.Error(E->getExprLoc(), "using builtin_preserve_access_index() without -g");
2602 return RValue::get(EmitScalarExpr(E->getArg(0)));
2603 }
2604
2605 // Nested builtin_preserve_access_index() not supported
2606 if (IsInPreservedAIRegion) {
2607 CGM.Error(E->getExprLoc(), "nested builtin_preserve_access_index() not supported");
2608 return RValue::get(EmitScalarExpr(E->getArg(0)));
2609 }
2610
2611 IsInPreservedAIRegion = true;
2612 Value *Res = EmitScalarExpr(E->getArg(0));
2613 IsInPreservedAIRegion = false;
2614 return RValue::get(Res);
2615 }
2616
2617 case Builtin::BI__builtin_cimag:
2618 case Builtin::BI__builtin_cimagf:
2619 case Builtin::BI__builtin_cimagl:
2620 case Builtin::BIcimag:
2621 case Builtin::BIcimagf:
2622 case Builtin::BIcimagl: {
2623 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0));
2624 return RValue::get(ComplexVal.second);
2625 }
2626
2627 case Builtin::BI__builtin_clrsb:
2628 case Builtin::BI__builtin_clrsbl:
2629 case Builtin::BI__builtin_clrsbll: {
2630 // clrsb(x) -> clz(x < 0 ? ~x : x) - 1 or
2631 Value *ArgValue = EmitScalarExpr(E->getArg(0));
2632
2633 llvm::Type *ArgType = ArgValue->getType();
2634 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType);
2635
2636 llvm::Type *ResultType = ConvertType(E->getType());
2637 Value *Zero = llvm::Constant::getNullValue(ArgType);
2638 Value *IsNeg = Builder.CreateICmpSLT(ArgValue, Zero, "isneg");
2639 Value *Inverse = Builder.CreateNot(ArgValue, "not");
2640 Value *Tmp = Builder.CreateSelect(IsNeg, Inverse, ArgValue);
2641 Value *Ctlz = Builder.CreateCall(F, {Tmp, Builder.getFalse()});
2642 Value *Result = Builder.CreateSub(Ctlz, llvm::ConstantInt::get(ArgType, 1));
2643 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
2644 "cast");
2645 return RValue::get(Result);
2646 }
2647 case Builtin::BI__builtin_ctzs:
2648 case Builtin::BI__builtin_ctz:
2649 case Builtin::BI__builtin_ctzl:
2650 case Builtin::BI__builtin_ctzll: {
2651 Value *ArgValue = EmitCheckedArgForBuiltin(E->getArg(0), BCK_CTZPassedZero);
2652
2653 llvm::Type *ArgType = ArgValue->getType();
2654 Function *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType);
2655
2656 llvm::Type *ResultType = ConvertType(E->getType());
2657 Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef());
2658 Value *Result = Builder.CreateCall(F, {ArgValue, ZeroUndef});
2659 if (Result->getType() != ResultType)
2660 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
2661 "cast");
2662 return RValue::get(Result);
2663 }
2664 case Builtin::BI__builtin_clzs:
2665 case Builtin::BI__builtin_clz:
2666 case Builtin::BI__builtin_clzl:
2667 case Builtin::BI__builtin_clzll: {
2668 Value *ArgValue = EmitCheckedArgForBuiltin(E->getArg(0), BCK_CLZPassedZero);
2669
2670 llvm::Type *ArgType = ArgValue->getType();
2671 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType);
2672
2673 llvm::Type *ResultType = ConvertType(E->getType());
2674 Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef());
2675 Value *Result = Builder.CreateCall(F, {ArgValue, ZeroUndef});
2676 if (Result->getType() != ResultType)
2677 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
2678 "cast");
2679 return RValue::get(Result);
2680 }
2681 case Builtin::BI__builtin_ffs:
2682 case Builtin::BI__builtin_ffsl:
2683 case Builtin::BI__builtin_ffsll: {
2684 // ffs(x) -> x ? cttz(x) + 1 : 0
2685 Value *ArgValue = EmitScalarExpr(E->getArg(0));
2686
2687 llvm::Type *ArgType = ArgValue->getType();
2688 Function *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType);
2689
2690 llvm::Type *ResultType = ConvertType(E->getType());
2691 Value *Tmp =
2692 Builder.CreateAdd(Builder.CreateCall(F, {ArgValue, Builder.getTrue()}),
2693 llvm::ConstantInt::get(ArgType, 1));
2694 Value *Zero = llvm::Constant::getNullValue(ArgType);
2695 Value *IsZero = Builder.CreateICmpEQ(ArgValue, Zero, "iszero");
2696 Value *Result = Builder.CreateSelect(IsZero, Zero, Tmp, "ffs");
2697 if (Result->getType() != ResultType)
2698 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
2699 "cast");
2700 return RValue::get(Result);
2701 }
2702 case Builtin::BI__builtin_parity:
2703 case Builtin::BI__builtin_parityl:
2704 case Builtin::BI__builtin_parityll: {
2705 // parity(x) -> ctpop(x) & 1
2706 Value *ArgValue = EmitScalarExpr(E->getArg(0));
2707
2708 llvm::Type *ArgType = ArgValue->getType();
2709 Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType);
2710
2711 llvm::Type *ResultType = ConvertType(E->getType());
2712 Value *Tmp = Builder.CreateCall(F, ArgValue);
2713 Value *Result = Builder.CreateAnd(Tmp, llvm::ConstantInt::get(ArgType, 1));
2714 if (Result->getType() != ResultType)
2715 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
2716 "cast");
2717 return RValue::get(Result);
2718 }
2719 case Builtin::BI__lzcnt16:
2720 case Builtin::BI__lzcnt:
2721 case Builtin::BI__lzcnt64: {
2722 Value *ArgValue = EmitScalarExpr(E->getArg(0));
2723
2724 llvm::Type *ArgType = ArgValue->getType();
2725 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType);
2726
2727 llvm::Type *ResultType = ConvertType(E->getType());
2728 Value *Result = Builder.CreateCall(F, {ArgValue, Builder.getFalse()});
2729 if (Result->getType() != ResultType)
2730 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
2731 "cast");
2732 return RValue::get(Result);
2733 }
2734 case Builtin::BI__popcnt16:
2735 case Builtin::BI__popcnt:
2736 case Builtin::BI__popcnt64:
2737 case Builtin::BI__builtin_popcount:
2738 case Builtin::BI__builtin_popcountl:
2739 case Builtin::BI__builtin_popcountll: {
2740 Value *ArgValue = EmitScalarExpr(E->getArg(0));
2741
2742 llvm::Type *ArgType = ArgValue->getType();
2743 Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType);
2744
2745 llvm::Type *ResultType = ConvertType(E->getType());
2746 Value *Result = Builder.CreateCall(F, ArgValue);
2747 if (Result->getType() != ResultType)
2748 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
2749 "cast");
2750 return RValue::get(Result);
2751 }
2752 case Builtin::BI__builtin_unpredictable: {
2753 // Always return the argument of __builtin_unpredictable. LLVM does not
2754 // handle this builtin. Metadata for this builtin should be added directly
2755 // to instructions such as branches or switches that use it.
2756 return RValue::get(EmitScalarExpr(E->getArg(0)));
2757 }
2758 case Builtin::BI__builtin_expect: {
2759 Value *ArgValue = EmitScalarExpr(E->getArg(0));
2760 llvm::Type *ArgType = ArgValue->getType();
2761
2762 Value *ExpectedValue = EmitScalarExpr(E->getArg(1));
2763 // Don't generate llvm.expect on -O0 as the backend won't use it for
2764 // anything.
2765 // Note, we still IRGen ExpectedValue because it could have side-effects.
2766 if (CGM.getCodeGenOpts().OptimizationLevel == 0)
2767 return RValue::get(ArgValue);
2768
2769 Function *FnExpect = CGM.getIntrinsic(Intrinsic::expect, ArgType);
2770 Value *Result =
2771 Builder.CreateCall(FnExpect, {ArgValue, ExpectedValue}, "expval");
2772 return RValue::get(Result);
2773 }
2774 case Builtin::BI__builtin_expect_with_probability: {
2775 Value *ArgValue = EmitScalarExpr(E->getArg(0));
2776 llvm::Type *ArgType = ArgValue->getType();
2777
2778 Value *ExpectedValue = EmitScalarExpr(E->getArg(1));
2779 llvm::APFloat Probability(0.0);
2780 const Expr *ProbArg = E->getArg(2);
2781 bool EvalSucceed = ProbArg->EvaluateAsFloat(Probability, CGM.getContext());
2782 assert(EvalSucceed && "probability should be able to evaluate as float");
2783 (void)EvalSucceed;
2784 bool LoseInfo = false;
2785 Probability.convert(llvm::APFloat::IEEEdouble(),
2786 llvm::RoundingMode::Dynamic, &LoseInfo);
2787 llvm::Type *Ty = ConvertType(ProbArg->getType());
2788 Constant *Confidence = ConstantFP::get(Ty, Probability);
2789 // Don't generate llvm.expect.with.probability on -O0 as the backend
2790 // won't use it for anything.
2791 // Note, we still IRGen ExpectedValue because it could have side-effects.
2792 if (CGM.getCodeGenOpts().OptimizationLevel == 0)
2793 return RValue::get(ArgValue);
2794
2795 Function *FnExpect =
2796 CGM.getIntrinsic(Intrinsic::expect_with_probability, ArgType);
2797 Value *Result = Builder.CreateCall(
2798 FnExpect, {ArgValue, ExpectedValue, Confidence}, "expval");
2799 return RValue::get(Result);
2800 }
2801 case Builtin::BI__builtin_assume_aligned: {
2802 const Expr *Ptr = E->getArg(0);
2803 Value *PtrValue = EmitScalarExpr(Ptr);
2804 if (PtrValue->getType() != VoidPtrTy)
2805 PtrValue = EmitCastToVoidPtr(PtrValue);
2806 Value *OffsetValue =
2807 (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : nullptr;
2808
2809 Value *AlignmentValue = EmitScalarExpr(E->getArg(1));
2810 ConstantInt *AlignmentCI = cast<ConstantInt>(AlignmentValue);
2811 if (AlignmentCI->getValue().ugt(llvm::Value::MaximumAlignment))
2812 AlignmentCI = ConstantInt::get(AlignmentCI->getType(),
2813 llvm::Value::MaximumAlignment);
2814
2815 emitAlignmentAssumption(PtrValue, Ptr,
2816 /*The expr loc is sufficient.*/ SourceLocation(),
2817 AlignmentCI, OffsetValue);
2818 return RValue::get(PtrValue);
2819 }
2820 case Builtin::BI__assume:
2821 case Builtin::BI__builtin_assume: {
2822 if (E->getArg(0)->HasSideEffects(getContext()))
2823 return RValue::get(nullptr);
2824
2825 Value *ArgValue = EmitScalarExpr(E->getArg(0));
2826 Function *FnAssume = CGM.getIntrinsic(Intrinsic::assume);
2827 Builder.CreateCall(FnAssume, ArgValue);
2828 return RValue::get(nullptr);
2829 }
2830 case Builtin::BI__arithmetic_fence: {
2831 // Create the builtin call if FastMath is selected, and the target
2832 // supports the builtin, otherwise just return the argument.
2833 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E);
2834 llvm::FastMathFlags FMF = Builder.getFastMathFlags();
2835 bool isArithmeticFenceEnabled =
2836 FMF.allowReassoc() &&
2837 getContext().getTargetInfo().checkArithmeticFenceSupported();
2838 QualType ArgType = E->getArg(0)->getType();
2839 if (ArgType->isComplexType()) {
2840 if (isArithmeticFenceEnabled) {
2841 QualType ElementType = ArgType->castAs<ComplexType>()->getElementType();
2842 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0));
2843 Value *Real = Builder.CreateArithmeticFence(ComplexVal.first,
2844 ConvertType(ElementType));
2845 Value *Imag = Builder.CreateArithmeticFence(ComplexVal.second,
2846 ConvertType(ElementType));
2847 return RValue::getComplex(std::make_pair(Real, Imag));
2848 }
2849 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0));
2850 Value *Real = ComplexVal.first;
2851 Value *Imag = ComplexVal.second;
2852 return RValue::getComplex(std::make_pair(Real, Imag));
2853 }
2854 Value *ArgValue = EmitScalarExpr(E->getArg(0));
2855 if (isArithmeticFenceEnabled)
2856 return RValue::get(
2857 Builder.CreateArithmeticFence(ArgValue, ConvertType(ArgType)));
2858 return RValue::get(ArgValue);
2859 }
2860 case Builtin::BI__builtin_bswap16:
2861 case Builtin::BI__builtin_bswap32:
2862 case Builtin::BI__builtin_bswap64:
2863 case Builtin::BI_byteswap_ushort:
2864 case Builtin::BI_byteswap_ulong:
2865 case Builtin::BI_byteswap_uint64: {
2866 return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::bswap));
2867 }
2868 case Builtin::BI__builtin_bitreverse8:
2869 case Builtin::BI__builtin_bitreverse16:
2870 case Builtin::BI__builtin_bitreverse32:
2871 case Builtin::BI__builtin_bitreverse64: {
2872 return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::bitreverse));
2873 }
2874 case Builtin::BI__builtin_rotateleft8:
2875 case Builtin::BI__builtin_rotateleft16:
2876 case Builtin::BI__builtin_rotateleft32:
2877 case Builtin::BI__builtin_rotateleft64:
2878 case Builtin::BI_rotl8: // Microsoft variants of rotate left
2879 case Builtin::BI_rotl16:
2880 case Builtin::BI_rotl:
2881 case Builtin::BI_lrotl:
2882 case Builtin::BI_rotl64:
2883 return emitRotate(E, false);
2884
2885 case Builtin::BI__builtin_rotateright8:
2886 case Builtin::BI__builtin_rotateright16:
2887 case Builtin::BI__builtin_rotateright32:
2888 case Builtin::BI__builtin_rotateright64:
2889 case Builtin::BI_rotr8: // Microsoft variants of rotate right
2890 case Builtin::BI_rotr16:
2891 case Builtin::BI_rotr:
2892 case Builtin::BI_lrotr:
2893 case Builtin::BI_rotr64:
2894 return emitRotate(E, true);
2895
2896 case Builtin::BI__builtin_constant_p: {
2897 llvm::Type *ResultType = ConvertType(E->getType());
2898
2899 const Expr *Arg = E->getArg(0);
2900 QualType ArgType = Arg->getType();
2901 // FIXME: The allowance for Obj-C pointers and block pointers is historical
2902 // and likely a mistake.
2903 if (!ArgType->isIntegralOrEnumerationType() && !ArgType->isFloatingType() &&
2904 !ArgType->isObjCObjectPointerType() && !ArgType->isBlockPointerType())
2905 // Per the GCC documentation, only numeric constants are recognized after
2906 // inlining.
2907 return RValue::get(ConstantInt::get(ResultType, 0));
2908
2909 if (Arg->HasSideEffects(getContext()))
2910 // The argument is unevaluated, so be conservative if it might have
2911 // side-effects.
2912 return RValue::get(ConstantInt::get(ResultType, 0));
2913
2914 Value *ArgValue = EmitScalarExpr(Arg);
2915 if (ArgType->isObjCObjectPointerType()) {
2916 // Convert Objective-C objects to id because we cannot distinguish between
2917 // LLVM types for Obj-C classes as they are opaque.
2918 ArgType = CGM.getContext().getObjCIdType();
2919 ArgValue = Builder.CreateBitCast(ArgValue, ConvertType(ArgType));
2920 }
2921 Function *F =
2922 CGM.getIntrinsic(Intrinsic::is_constant, ConvertType(ArgType));
2923 Value *Result = Builder.CreateCall(F, ArgValue);
2924 if (Result->getType() != ResultType)
2925 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/false);
2926 return RValue::get(Result);
2927 }
2928 case Builtin::BI__builtin_dynamic_object_size:
2929 case Builtin::BI__builtin_object_size: {
2930 unsigned Type =
2931 E->getArg(1)->EvaluateKnownConstInt(getContext()).getZExtValue();
2932 auto *ResType = cast<llvm::IntegerType>(ConvertType(E->getType()));
2933
2934 // We pass this builtin onto the optimizer so that it can figure out the
2935 // object size in more complex cases.
2936 bool IsDynamic = BuiltinID == Builtin::BI__builtin_dynamic_object_size;
2937 return RValue::get(emitBuiltinObjectSize(E->getArg(0), Type, ResType,
2938 /*EmittedE=*/nullptr, IsDynamic));
2939 }
2940 case Builtin::BI__builtin_prefetch: {
2941 Value *Locality, *RW, *Address = EmitScalarExpr(E->getArg(0));
2942 // FIXME: Technically these constants should of type 'int', yes?
2943 RW = (E->getNumArgs() > 1) ? EmitScalarExpr(E->getArg(1)) :
2944 llvm::ConstantInt::get(Int32Ty, 0);
2945 Locality = (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) :
2946 llvm::ConstantInt::get(Int32Ty, 3);
2947 Value *Data = llvm::ConstantInt::get(Int32Ty, 1);
2948 Function *F = CGM.getIntrinsic(Intrinsic::prefetch, Address->getType());
2949 Builder.CreateCall(F, {Address, RW, Locality, Data});
2950 return RValue::get(nullptr);
2951 }
2952 case Builtin::BI__builtin_readcyclecounter: {
2953 Function *F = CGM.getIntrinsic(Intrinsic::readcyclecounter);
2954 return RValue::get(Builder.CreateCall(F));
2955 }
2956 case Builtin::BI__builtin___clear_cache: {
2957 Value *Begin = EmitScalarExpr(E->getArg(0));
2958 Value *End = EmitScalarExpr(E->getArg(1));
2959 Function *F = CGM.getIntrinsic(Intrinsic::clear_cache);
2960 return RValue::get(Builder.CreateCall(F, {Begin, End}));
2961 }
2962 case Builtin::BI__builtin_trap:
2963 EmitTrapCall(Intrinsic::trap);
2964 return RValue::get(nullptr);
2965 case Builtin::BI__debugbreak:
2966 EmitTrapCall(Intrinsic::debugtrap);
2967 return RValue::get(nullptr);
2968 case Builtin::BI__builtin_unreachable: {
2969 EmitUnreachable(E->getExprLoc());
2970
2971 // We do need to preserve an insertion point.
2972 EmitBlock(createBasicBlock("unreachable.cont"));
2973
2974 return RValue::get(nullptr);
2975 }
2976
2977 case Builtin::BI__builtin_powi:
2978 case Builtin::BI__builtin_powif:
2979 case Builtin::BI__builtin_powil: {
2980 llvm::Value *Src0 = EmitScalarExpr(E->getArg(0));
2981 llvm::Value *Src1 = EmitScalarExpr(E->getArg(1));
2982
2983 if (Builder.getIsFPConstrained()) {
2984 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E);
2985 Function *F = CGM.getIntrinsic(Intrinsic::experimental_constrained_powi,
2986 Src0->getType());
2987 return RValue::get(Builder.CreateConstrainedFPCall(F, { Src0, Src1 }));
2988 }
2989
2990 Function *F = CGM.getIntrinsic(Intrinsic::powi,
2991 { Src0->getType(), Src1->getType() });
2992 return RValue::get(Builder.CreateCall(F, { Src0, Src1 }));
2993 }
2994 case Builtin::BI__builtin_isgreater:
2995 case Builtin::BI__builtin_isgreaterequal:
2996 case Builtin::BI__builtin_isless:
2997 case Builtin::BI__builtin_islessequal:
2998 case Builtin::BI__builtin_islessgreater:
2999 case Builtin::BI__builtin_isunordered: {
3000 // Ordered comparisons: we know the arguments to these are matching scalar
3001 // floating point values.
3002 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E);
3003 Value *LHS = EmitScalarExpr(E->getArg(0));
3004 Value *RHS = EmitScalarExpr(E->getArg(1));
3005
3006 switch (BuiltinID) {
3007 default: llvm_unreachable("Unknown ordered comparison");
3008 case Builtin::BI__builtin_isgreater:
3009 LHS = Builder.CreateFCmpOGT(LHS, RHS, "cmp");
3010 break;
3011 case Builtin::BI__builtin_isgreaterequal:
3012 LHS = Builder.CreateFCmpOGE(LHS, RHS, "cmp");
3013 break;
3014 case Builtin::BI__builtin_isless:
3015 LHS = Builder.CreateFCmpOLT(LHS, RHS, "cmp");
3016 break;
3017 case Builtin::BI__builtin_islessequal:
3018 LHS = Builder.CreateFCmpOLE(LHS, RHS, "cmp");
3019 break;
3020 case Builtin::BI__builtin_islessgreater:
3021 LHS = Builder.CreateFCmpONE(LHS, RHS, "cmp");
3022 break;
3023 case Builtin::BI__builtin_isunordered:
3024 LHS = Builder.CreateFCmpUNO(LHS, RHS, "cmp");
3025 break;
3026 }
3027 // ZExt bool to int type.
3028 return RValue::get(Builder.CreateZExt(LHS, ConvertType(E->getType())));
3029 }
3030 case Builtin::BI__builtin_isnan: {
3031 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E);
3032 Value *V = EmitScalarExpr(E->getArg(0));
3033 llvm::Type *Ty = V->getType();
3034 const llvm::fltSemantics &Semantics = Ty->getFltSemantics();
3035 if (!Builder.getIsFPConstrained() ||
3036 Builder.getDefaultConstrainedExcept() == fp::ebIgnore ||
3037 !Ty->isIEEE()) {
3038 V = Builder.CreateFCmpUNO(V, V, "cmp");
3039 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType())));
3040 }
3041
3042 if (Value *Result = getTargetHooks().testFPKind(V, BuiltinID, Builder, CGM))
3043 return RValue::get(Result);
3044
3045 // NaN has all exp bits set and a non zero significand. Therefore:
3046 // isnan(V) == ((exp mask - (abs(V) & exp mask)) < 0)
3047 unsigned bitsize = Ty->getScalarSizeInBits();
3048 llvm::IntegerType *IntTy = Builder.getIntNTy(bitsize);
3049 Value *IntV = Builder.CreateBitCast(V, IntTy);
3050 APInt AndMask = APInt::getSignedMaxValue(bitsize);
3051 Value *AbsV =
3052 Builder.CreateAnd(IntV, llvm::ConstantInt::get(IntTy, AndMask));
3053 APInt ExpMask = APFloat::getInf(Semantics).bitcastToAPInt();
3054 Value *Sub =
3055 Builder.CreateSub(llvm::ConstantInt::get(IntTy, ExpMask), AbsV);
3056 // V = sign bit (Sub) <=> V = (Sub < 0)
3057 V = Builder.CreateLShr(Sub, llvm::ConstantInt::get(IntTy, bitsize - 1));
3058 if (bitsize > 32)
3059 V = Builder.CreateTrunc(V, ConvertType(E->getType()));
3060 return RValue::get(V);
3061 }
3062
3063 case Builtin::BI__builtin_elementwise_abs: {
3064 Value *Result;
3065 QualType QT = E->getArg(0)->getType();
3066
3067 if (auto *VecTy = QT->getAs<VectorType>())
3068 QT = VecTy->getElementType();
3069 if (QT->isIntegerType())
3070 Result = Builder.CreateBinaryIntrinsic(
3071 llvm::Intrinsic::abs, EmitScalarExpr(E->getArg(0)),
3072 Builder.getFalse(), nullptr, "elt.abs");
3073 else
3074 Result = emitUnaryBuiltin(*this, E, llvm::Intrinsic::fabs, "elt.abs");
3075
3076 return RValue::get(Result);
3077 }
3078
3079 case Builtin::BI__builtin_elementwise_ceil:
3080 return RValue::get(
3081 emitUnaryBuiltin(*this, E, llvm::Intrinsic::ceil, "elt.ceil"));
3082 case Builtin::BI__builtin_elementwise_cos:
3083 return RValue::get(
3084 emitUnaryBuiltin(*this, E, llvm::Intrinsic::cos, "elt.cos"));
3085 case Builtin::BI__builtin_elementwise_floor:
3086 return RValue::get(
3087 emitUnaryBuiltin(*this, E, llvm::Intrinsic::floor, "elt.floor"));
3088 case Builtin::BI__builtin_elementwise_roundeven:
3089 return RValue::get(emitUnaryBuiltin(*this, E, llvm::Intrinsic::roundeven,
3090 "elt.roundeven"));
3091 case Builtin::BI__builtin_elementwise_sin:
3092 return RValue::get(
3093 emitUnaryBuiltin(*this, E, llvm::Intrinsic::sin, "elt.sin"));
3094
3095 case Builtin::BI__builtin_elementwise_trunc:
3096 return RValue::get(
3097 emitUnaryBuiltin(*this, E, llvm::Intrinsic::trunc, "elt.trunc"));
3098 case Builtin::BI__builtin_elementwise_canonicalize:
3099 return RValue::get(
3100 emitUnaryBuiltin(*this, E, llvm::Intrinsic::canonicalize, "elt.trunc"));
3101 case Builtin::BI__builtin_elementwise_copysign:
3102 return RValue::get(emitBinaryBuiltin(*this, E, llvm::Intrinsic::copysign));
3103 case Builtin::BI__builtin_elementwise_add_sat:
3104 case Builtin::BI__builtin_elementwise_sub_sat: {
3105 Value *Op0 = EmitScalarExpr(E->getArg(0));
3106 Value *Op1 = EmitScalarExpr(E->getArg(1));
3107 Value *Result;
3108 assert(Op0->getType()->isIntOrIntVectorTy() && "integer type expected");
3109 QualType Ty = E->getArg(0)->getType();
3110 if (auto *VecTy = Ty->getAs<VectorType>())
3111 Ty = VecTy->getElementType();
3112 bool IsSigned = Ty->isSignedIntegerType();
3113 unsigned Opc;
3114 if (BuiltinIDIfNoAsmLabel == Builtin::BI__builtin_elementwise_add_sat)
3115 Opc = IsSigned ? llvm::Intrinsic::sadd_sat : llvm::Intrinsic::uadd_sat;
3116 else
3117 Opc = IsSigned ? llvm::Intrinsic::ssub_sat : llvm::Intrinsic::usub_sat;
3118 Result = Builder.CreateBinaryIntrinsic(Opc, Op0, Op1, nullptr, "elt.sat");
3119 return RValue::get(Result);
3120 }
3121
3122 case Builtin::BI__builtin_elementwise_max: {
3123 Value *Op0 = EmitScalarExpr(E->getArg(0));
3124 Value *Op1 = EmitScalarExpr(E->getArg(1));
3125 Value *Result;
3126 if (Op0->getType()->isIntOrIntVectorTy()) {
3127 QualType Ty = E->getArg(0)->getType();
3128 if (auto *VecTy = Ty->getAs<VectorType>())
3129 Ty = VecTy->getElementType();
3130 Result = Builder.CreateBinaryIntrinsic(Ty->isSignedIntegerType()
3131 ? llvm::Intrinsic::smax
3132 : llvm::Intrinsic::umax,
3133 Op0, Op1, nullptr, "elt.max");
3134 } else
3135 Result = Builder.CreateMaxNum(Op0, Op1, "elt.max");
3136 return RValue::get(Result);
3137 }
3138 case Builtin::BI__builtin_elementwise_min: {
3139 Value *Op0 = EmitScalarExpr(E->getArg(0));
3140 Value *Op1 = EmitScalarExpr(E->getArg(1));
3141 Value *Result;
3142 if (Op0->getType()->isIntOrIntVectorTy()) {
3143 QualType Ty = E->getArg(0)->getType();
3144 if (auto *VecTy = Ty->getAs<VectorType>())
3145 Ty = VecTy->getElementType();
3146 Result = Builder.CreateBinaryIntrinsic(Ty->isSignedIntegerType()
3147 ? llvm::Intrinsic::smin
3148 : llvm::Intrinsic::umin,
3149 Op0, Op1, nullptr, "elt.min");
3150 } else
3151 Result = Builder.CreateMinNum(Op0, Op1, "elt.min");
3152 return RValue::get(Result);
3153 }
3154
3155 case Builtin::BI__builtin_reduce_max: {
3156 auto GetIntrinsicID = [](QualType QT) {
3157 if (auto *VecTy = QT->getAs<VectorType>())
3158 QT = VecTy->getElementType();
3159 if (QT->isSignedIntegerType())
3160 return llvm::Intrinsic::vector_reduce_smax;
3161 if (QT->isUnsignedIntegerType())
3162 return llvm::Intrinsic::vector_reduce_umax;
3163 assert(QT->isFloatingType() && "must have a float here");
3164 return llvm::Intrinsic::vector_reduce_fmax;
3165 };
3166 return RValue::get(emitUnaryBuiltin(
3167 *this, E, GetIntrinsicID(E->getArg(0)->getType()), "rdx.min"));
3168 }
3169
3170 case Builtin::BI__builtin_reduce_min: {
3171 auto GetIntrinsicID = [](QualType QT) {
3172 if (auto *VecTy = QT->getAs<VectorType>())
3173 QT = VecTy->getElementType();
3174 if (QT->isSignedIntegerType())
3175 return llvm::Intrinsic::vector_reduce_smin;
3176 if (QT->isUnsignedIntegerType())
3177 return llvm::Intrinsic::vector_reduce_umin;
3178 assert(QT->isFloatingType() && "must have a float here");
3179 return llvm::Intrinsic::vector_reduce_fmin;
3180 };
3181
3182 return RValue::get(emitUnaryBuiltin(
3183 *this, E, GetIntrinsicID(E->getArg(0)->getType()), "rdx.min"));
3184 }
3185
3186 case Builtin::BI__builtin_reduce_add:
3187 return RValue::get(emitUnaryBuiltin(
3188 *this, E, llvm::Intrinsic::vector_reduce_add, "rdx.add"));
3189 case Builtin::BI__builtin_reduce_mul:
3190 return RValue::get(emitUnaryBuiltin(
3191 *this, E, llvm::Intrinsic::vector_reduce_mul, "rdx.mul"));
3192 case Builtin::BI__builtin_reduce_xor:
3193 return RValue::get(emitUnaryBuiltin(
3194 *this, E, llvm::Intrinsic::vector_reduce_xor, "rdx.xor"));
3195 case Builtin::BI__builtin_reduce_or:
3196 return RValue::get(emitUnaryBuiltin(
3197 *this, E, llvm::Intrinsic::vector_reduce_or, "rdx.or"));
3198 case Builtin::BI__builtin_reduce_and:
3199 return RValue::get(emitUnaryBuiltin(
3200 *this, E, llvm::Intrinsic::vector_reduce_and, "rdx.and"));
3201
3202 case Builtin::BI__builtin_matrix_transpose: {
3203 auto *MatrixTy = E->getArg(0)->getType()->castAs<ConstantMatrixType>();
3204 Value *MatValue = EmitScalarExpr(E->getArg(0));
3205 MatrixBuilder MB(Builder);
3206 Value *Result = MB.CreateMatrixTranspose(MatValue, MatrixTy->getNumRows(),
3207 MatrixTy->getNumColumns());
3208 return RValue::get(Result);
3209 }
3210
3211 case Builtin::BI__builtin_matrix_column_major_load: {
3212 MatrixBuilder MB(Builder);
3213 // Emit everything that isn't dependent on the first parameter type
3214 Value *Stride = EmitScalarExpr(E->getArg(3));
3215 const auto *ResultTy = E->getType()->getAs<ConstantMatrixType>();
3216 auto *PtrTy = E->getArg(0)->getType()->getAs<PointerType>();
3217 assert(PtrTy && "arg0 must be of pointer type");
3218 bool IsVolatile = PtrTy->getPointeeType().isVolatileQualified();
3219
3220 Address Src = EmitPointerWithAlignment(E->getArg(0));
3221 EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(0)->getType(),
3222 E->getArg(0)->getExprLoc(), FD, 0);
3223 Value *Result = MB.CreateColumnMajorLoad(
3224 Src.getElementType(), Src.getPointer(),
3225 Align(Src.getAlignment().getQuantity()), Stride, IsVolatile,
3226 ResultTy->getNumRows(), ResultTy->getNumColumns(),
3227 "matrix");
3228 return RValue::get(Result);
3229 }
3230
3231 case Builtin::BI__builtin_matrix_column_major_store: {
3232 MatrixBuilder MB(Builder);
3233 Value *Matrix = EmitScalarExpr(E->getArg(0));
3234 Address Dst = EmitPointerWithAlignment(E->getArg(1));
3235 Value *Stride = EmitScalarExpr(E->getArg(2));
3236
3237 const auto *MatrixTy = E->getArg(0)->getType()->getAs<ConstantMatrixType>();
3238 auto *PtrTy = E->getArg(1)->getType()->getAs<PointerType>();
3239 assert(PtrTy && "arg1 must be of pointer type");
3240 bool IsVolatile = PtrTy->getPointeeType().isVolatileQualified();
3241
3242 EmitNonNullArgCheck(RValue::get(Dst.getPointer()), E->getArg(1)->getType(),
3243 E->getArg(1)->getExprLoc(), FD, 0);
3244 Value *Result = MB.CreateColumnMajorStore(
3245 Matrix, Dst.getPointer(), Align(Dst.getAlignment().getQuantity()),
3246 Stride, IsVolatile, MatrixTy->getNumRows(), MatrixTy->getNumColumns());
3247 return RValue::get(Result);
3248 }
3249
3250 case Builtin::BIfinite:
3251 case Builtin::BI__finite:
3252 case Builtin::BIfinitef:
3253 case Builtin::BI__finitef:
3254 case Builtin::BIfinitel:
3255 case Builtin::BI__finitel:
3256 case Builtin::BI__builtin_isinf:
3257 case Builtin::BI__builtin_isfinite: {
3258 // isinf(x) --> fabs(x) == infinity
3259 // isfinite(x) --> fabs(x) != infinity
3260 // x != NaN via the ordered compare in either case.
3261 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E);
3262 Value *V = EmitScalarExpr(E->getArg(0));
3263 llvm::Type *Ty = V->getType();
3264 if (!Builder.getIsFPConstrained() ||
3265 Builder.getDefaultConstrainedExcept() == fp::ebIgnore ||
3266 !Ty->isIEEE()) {
3267 Value *Fabs = EmitFAbs(*this, V);
3268 Constant *Infinity = ConstantFP::getInfinity(V->getType());
3269 CmpInst::Predicate Pred = (BuiltinID == Builtin::BI__builtin_isinf)
3270 ? CmpInst::FCMP_OEQ
3271 : CmpInst::FCMP_ONE;
3272 Value *FCmp = Builder.CreateFCmp(Pred, Fabs, Infinity, "cmpinf");
3273 return RValue::get(Builder.CreateZExt(FCmp, ConvertType(E->getType())));
3274 }
3275
3276 if (Value *Result = getTargetHooks().testFPKind(V, BuiltinID, Builder, CGM))
3277 return RValue::get(Result);
3278
3279 // Inf values have all exp bits set and a zero significand. Therefore:
3280 // isinf(V) == ((V << 1) == ((exp mask) << 1))
3281 // isfinite(V) == ((V << 1) < ((exp mask) << 1)) using unsigned comparison
3282 unsigned bitsize = Ty->getScalarSizeInBits();
3283 llvm::IntegerType *IntTy = Builder.getIntNTy(bitsize);
3284 Value *IntV = Builder.CreateBitCast(V, IntTy);
3285 Value *Shl1 = Builder.CreateShl(IntV, 1);
3286 const llvm::fltSemantics &Semantics = Ty->getFltSemantics();
3287 APInt ExpMask = APFloat::getInf(Semantics).bitcastToAPInt();
3288 Value *ExpMaskShl1 = llvm::ConstantInt::get(IntTy, ExpMask.shl(1));
3289 if (BuiltinID == Builtin::BI__builtin_isinf)
3290 V = Builder.CreateICmpEQ(Shl1, ExpMaskShl1);
3291 else
3292 V = Builder.CreateICmpULT(Shl1, ExpMaskShl1);
3293 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType())));
3294 }
3295
3296 case Builtin::BI__builtin_isinf_sign: {
3297 // isinf_sign(x) -> fabs(x) == infinity ? (signbit(x) ? -1 : 1) : 0
3298 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E);
3299 // FIXME: for strictfp/IEEE-754 we need to not trap on SNaN here.
3300 Value *Arg = EmitScalarExpr(E->getArg(0));
3301 Value *AbsArg = EmitFAbs(*this, Arg);
3302 Value *IsInf = Builder.CreateFCmpOEQ(
3303 AbsArg, ConstantFP::getInfinity(Arg->getType()), "isinf");
3304 Value *IsNeg = EmitSignBit(*this, Arg);
3305
3306 llvm::Type *IntTy = ConvertType(E->getType());
3307 Value *Zero = Constant::getNullValue(IntTy);
3308 Value *One = ConstantInt::get(IntTy, 1);
3309 Value *NegativeOne = ConstantInt::get(IntTy, -1);
3310 Value *SignResult = Builder.CreateSelect(IsNeg, NegativeOne, One);
3311 Value *Result = Builder.CreateSelect(IsInf, SignResult, Zero);
3312 return RValue::get(Result);
3313 }
3314
3315 case Builtin::BI__builtin_isnormal: {
3316 // isnormal(x) --> x == x && fabsf(x) < infinity && fabsf(x) >= float_min
3317 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E);
3318 // FIXME: for strictfp/IEEE-754 we need to not trap on SNaN here.
3319 Value *V = EmitScalarExpr(E->getArg(0));
3320 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq");
3321
3322 Value *Abs = EmitFAbs(*this, V);
3323 Value *IsLessThanInf =
3324 Builder.CreateFCmpULT(Abs, ConstantFP::getInfinity(V->getType()),"isinf");
3325 APFloat Smallest = APFloat::getSmallestNormalized(
3326 getContext().getFloatTypeSemantics(E->getArg(0)->getType()));
3327 Value *IsNormal =
3328 Builder.CreateFCmpUGE(Abs, ConstantFP::get(V->getContext(), Smallest),
3329 "isnormal");
3330 V = Builder.CreateAnd(Eq, IsLessThanInf, "and");
3331 V = Builder.CreateAnd(V, IsNormal, "and");
3332 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType())));
3333 }
3334
3335 case Builtin::BI__builtin_flt_rounds: {
3336 Function *F = CGM.getIntrinsic(Intrinsic::get_rounding);
3337
3338 llvm::Type *ResultType = ConvertType(E->getType());
3339 Value *Result = Builder.CreateCall(F);
3340 if (Result->getType() != ResultType)
3341 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
3342 "cast");
3343 return RValue::get(Result);
3344 }
3345
3346 case Builtin::BI__builtin_fpclassify: {
3347 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E);
3348 // FIXME: for strictfp/IEEE-754 we need to not trap on SNaN here.
3349 Value *V = EmitScalarExpr(E->getArg(5));
3350 llvm::Type *Ty = ConvertType(E->getArg(5)->getType());
3351
3352 // Create Result
3353 BasicBlock *Begin = Builder.GetInsertBlock();
3354 BasicBlock *End = createBasicBlock("fpclassify_end", this->CurFn);
3355 Builder.SetInsertPoint(End);
3356 PHINode *Result =
3357 Builder.CreatePHI(ConvertType(E->getArg(0)->getType()), 4,
3358 "fpclassify_result");
3359
3360 // if (V==0) return FP_ZERO
3361 Builder.SetInsertPoint(Begin);
3362 Value *IsZero = Builder.CreateFCmpOEQ(V, Constant::getNullValue(Ty),
3363 "iszero");
3364 Value *ZeroLiteral = EmitScalarExpr(E->getArg(4));
3365 BasicBlock *NotZero = createBasicBlock("fpclassify_not_zero", this->CurFn);
3366 Builder.CreateCondBr(IsZero, End, NotZero);
3367 Result->addIncoming(ZeroLiteral, Begin);
3368
3369 // if (V != V) return FP_NAN
3370 Builder.SetInsertPoint(NotZero);
3371 Value *IsNan = Builder.CreateFCmpUNO(V, V, "cmp");
3372 Value *NanLiteral = EmitScalarExpr(E->getArg(0));
3373 BasicBlock *NotNan = createBasicBlock("fpclassify_not_nan", this->CurFn);
3374 Builder.CreateCondBr(IsNan, End, NotNan);
3375 Result->addIncoming(NanLiteral, NotZero);
3376
3377 // if (fabs(V) == infinity) return FP_INFINITY
3378 Builder.SetInsertPoint(NotNan);
3379 Value *VAbs = EmitFAbs(*this, V);
3380 Value *IsInf =
3381 Builder.CreateFCmpOEQ(VAbs, ConstantFP::getInfinity(V->getType()),
3382 "isinf");
3383 Value *InfLiteral = EmitScalarExpr(E->getArg(1));
3384 BasicBlock *NotInf = createBasicBlock("fpclassify_not_inf", this->CurFn);
3385 Builder.CreateCondBr(IsInf, End, NotInf);
3386 Result->addIncoming(InfLiteral, NotNan);
3387
3388 // if (fabs(V) >= MIN_NORMAL) return FP_NORMAL else FP_SUBNORMAL
3389 Builder.SetInsertPoint(NotInf);
3390 APFloat Smallest = APFloat::getSmallestNormalized(
3391 getContext().getFloatTypeSemantics(E->getArg(5)->getType()));
3392 Value *IsNormal =
3393 Builder.CreateFCmpUGE(VAbs, ConstantFP::get(V->getContext(), Smallest),
3394 "isnormal");
3395 Value *NormalResult =
3396 Builder.CreateSelect(IsNormal, EmitScalarExpr(E->getArg(2)),
3397 EmitScalarExpr(E->getArg(3)));
3398 Builder.CreateBr(End);
3399 Result->addIncoming(NormalResult, NotInf);
3400
3401 // return Result
3402 Builder.SetInsertPoint(End);
3403 return RValue::get(Result);
3404 }
3405
3406 case Builtin::BIalloca:
3407 case Builtin::BI_alloca:
3408 case Builtin::BI__builtin_alloca_uninitialized:
3409 case Builtin::BI__builtin_alloca: {
3410 Value *Size = EmitScalarExpr(E->getArg(0));
3411 const TargetInfo &TI = getContext().getTargetInfo();
3412 // The alignment of the alloca should correspond to __BIGGEST_ALIGNMENT__.
3413 const Align SuitableAlignmentInBytes =
3414 CGM.getContext()
3415 .toCharUnitsFromBits(TI.getSuitableAlign())
3416 .getAsAlign();
3417 AllocaInst *AI = Builder.CreateAlloca(Builder.getInt8Ty(), Size);
3418 AI->setAlignment(SuitableAlignmentInBytes);
3419 if (BuiltinID != Builtin::BI__builtin_alloca_uninitialized)
3420 initializeAlloca(*this, AI, Size, SuitableAlignmentInBytes);
3421 return RValue::get(AI);
3422 }
3423
3424 case Builtin::BI__builtin_alloca_with_align_uninitialized:
3425 case Builtin::BI__builtin_alloca_with_align: {
3426 Value *Size = EmitScalarExpr(E->getArg(0));
3427 Value *AlignmentInBitsValue = EmitScalarExpr(E->getArg(1));
3428 auto *AlignmentInBitsCI = cast<ConstantInt>(AlignmentInBitsValue);
3429 unsigned AlignmentInBits = AlignmentInBitsCI->getZExtValue();
3430 const Align AlignmentInBytes =
3431 CGM.getContext().toCharUnitsFromBits(AlignmentInBits).getAsAlign();
3432 AllocaInst *AI = Builder.CreateAlloca(Builder.getInt8Ty(), Size);
3433 AI->setAlignment(AlignmentInBytes);
3434 if (BuiltinID != Builtin::BI__builtin_alloca_with_align_uninitialized)
3435 initializeAlloca(*this, AI, Size, AlignmentInBytes);
3436 return RValue::get(AI);
3437 }
3438
3439 case Builtin::BIbzero:
3440 case Builtin::BI__builtin_bzero: {
3441 Address Dest = EmitPointerWithAlignment(E->getArg(0));
3442 Value *SizeVal = EmitScalarExpr(E->getArg(1));
3443 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(),
3444 E->getArg(0)->getExprLoc(), FD, 0);
3445 Builder.CreateMemSet(Dest, Builder.getInt8(0), SizeVal, false);
3446 return RValue::get(nullptr);
3447 }
3448 case Builtin::BImemcpy:
3449 case Builtin::BI__builtin_memcpy:
3450 case Builtin::BImempcpy:
3451 case Builtin::BI__builtin_mempcpy: {
3452 Address Dest = EmitPointerWithAlignment(E->getArg(0));
3453 Address Src = EmitPointerWithAlignment(E->getArg(1));
3454 Value *SizeVal = EmitScalarExpr(E->getArg(2));
3455 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(),
3456 E->getArg(0)->getExprLoc(), FD, 0);
3457 EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(1)->getType(),
3458 E->getArg(1)->getExprLoc(), FD, 1);
3459 Builder.CreateMemCpy(Dest, Src, SizeVal, false);
3460 if (BuiltinID == Builtin::BImempcpy ||
3461 BuiltinID == Builtin::BI__builtin_mempcpy)
3462 return RValue::get(Builder.CreateInBoundsGEP(Dest.getElementType(),
3463 Dest.getPointer(), SizeVal));
3464 else
3465 return RValue::get(Dest.getPointer());
3466 }
3467
3468 case Builtin::BI__builtin_memcpy_inline: {
3469 Address Dest = EmitPointerWithAlignment(E->getArg(0));
3470 Address Src = EmitPointerWithAlignment(E->getArg(1));
3471 uint64_t Size =
3472 E->getArg(2)->EvaluateKnownConstInt(getContext()).getZExtValue();
3473 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(),
3474 E->getArg(0)->getExprLoc(), FD, 0);
3475 EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(1)->getType(),
3476 E->getArg(1)->getExprLoc(), FD, 1);
3477 Builder.CreateMemCpyInline(Dest, Src, Size);
3478 return RValue::get(nullptr);
3479 }
3480
3481 case Builtin::BI__builtin_char_memchr:
3482 BuiltinID = Builtin::BI__builtin_memchr;
3483 break;
3484
3485 case Builtin::BI__builtin___memcpy_chk: {
3486 // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memcpy iff cst1<=cst2.
3487 Expr::EvalResult SizeResult, DstSizeResult;
3488 if (!E->getArg(2)->EvaluateAsInt(SizeResult, CGM.getContext()) ||
3489 !E->getArg(3)->EvaluateAsInt(DstSizeResult, CGM.getContext()))
3490 break;
3491 llvm::APSInt Size = SizeResult.Val.getInt();
3492 llvm::APSInt DstSize = DstSizeResult.Val.getInt();
3493 if (Size.ugt(DstSize))
3494 break;
3495 Address Dest = EmitPointerWithAlignment(E->getArg(0));
3496 Address Src = EmitPointerWithAlignment(E->getArg(1));
3497 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size);
3498 Builder.CreateMemCpy(Dest, Src, SizeVal, false);
3499 return RValue::get(Dest.getPointer());
3500 }
3501
3502 case Builtin::BI__builtin_objc_memmove_collectable: {
3503 Address DestAddr = EmitPointerWithAlignment(E->getArg(0));
3504 Address SrcAddr = EmitPointerWithAlignment(E->getArg(1));
3505 Value *SizeVal = EmitScalarExpr(E->getArg(2));
3506 CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this,
3507 DestAddr, SrcAddr, SizeVal);
3508 return RValue::get(DestAddr.getPointer());
3509 }
3510
3511 case Builtin::BI__builtin___memmove_chk: {
3512 // fold __builtin_memmove_chk(x, y, cst1, cst2) to memmove iff cst1<=cst2.
3513 Expr::EvalResult SizeResult, DstSizeResult;
3514 if (!E->getArg(2)->EvaluateAsInt(SizeResult, CGM.getContext()) ||
3515 !E->getArg(3)->EvaluateAsInt(DstSizeResult, CGM.getContext()))
3516 break;
3517 llvm::APSInt Size = SizeResult.Val.getInt();
3518 llvm::APSInt DstSize = DstSizeResult.Val.getInt();
3519 if (Size.ugt(DstSize))
3520 break;
3521 Address Dest = EmitPointerWithAlignment(E->getArg(0));
3522 Address Src = EmitPointerWithAlignment(E->getArg(1));
3523 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size);
3524 Builder.CreateMemMove(Dest, Src, SizeVal, false);
3525 return RValue::get(Dest.getPointer());
3526 }
3527
3528 case Builtin::BImemmove:
3529 case Builtin::BI__builtin_memmove: {
3530 Address Dest = EmitPointerWithAlignment(E->getArg(0));
3531 Address Src = EmitPointerWithAlignment(E->getArg(1));
3532 Value *SizeVal = EmitScalarExpr(E->getArg(2));
3533 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(),
3534 E->getArg(0)->getExprLoc(), FD, 0);
3535 EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(1)->getType(),
3536 E->getArg(1)->getExprLoc(), FD, 1);
3537 Builder.CreateMemMove(Dest, Src, SizeVal, false);
3538 return RValue::get(Dest.getPointer());
3539 }
3540 case Builtin::BImemset:
3541 case Builtin::BI__builtin_memset: {
3542 Address Dest = EmitPointerWithAlignment(E->getArg(0));
3543 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)),
3544 Builder.getInt8Ty());
3545 Value *SizeVal = EmitScalarExpr(E->getArg(2));
3546 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(),
3547 E->getArg(0)->getExprLoc(), FD, 0);
3548 Builder.CreateMemSet(Dest, ByteVal, SizeVal, false);
3549 return RValue::get(Dest.getPointer());
3550 }
3551 case Builtin::BI__builtin_memset_inline: {
3552 Address Dest = EmitPointerWithAlignment(E->getArg(0));
3553 Value *ByteVal =
3554 Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), Builder.getInt8Ty());
3555 uint64_t Size =
3556 E->getArg(2)->EvaluateKnownConstInt(getContext()).getZExtValue();
3557 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(),
3558 E->getArg(0)->getExprLoc(), FD, 0);
3559 Builder.CreateMemSetInline(Dest, ByteVal, Size);
3560 return RValue::get(nullptr);
3561 }
3562 case Builtin::BI__builtin___memset_chk: {
3563 // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2.
3564 Expr::EvalResult SizeResult, DstSizeResult;
3565 if (!E->getArg(2)->EvaluateAsInt(SizeResult, CGM.getContext()) ||
3566 !E->getArg(3)->EvaluateAsInt(DstSizeResult, CGM.getContext()))
3567 break;
3568 llvm::APSInt Size = SizeResult.Val.getInt();
3569 llvm::APSInt DstSize = DstSizeResult.Val.getInt();
3570 if (Size.ugt(DstSize))
3571 break;
3572 Address Dest = EmitPointerWithAlignment(E->getArg(0));
3573 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)),
3574 Builder.getInt8Ty());
3575 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size);
3576 Builder.CreateMemSet(Dest, ByteVal, SizeVal, false);
3577 return RValue::get(Dest.getPointer());
3578 }
3579 case Builtin::BI__builtin_wmemchr: {
3580 // The MSVC runtime library does not provide a definition of wmemchr, so we
3581 // need an inline implementation.
3582 if (!getTarget().getTriple().isOSMSVCRT())
3583 break;
3584
3585 llvm::Type *WCharTy = ConvertType(getContext().WCharTy);
3586 Value *Str = EmitScalarExpr(E->getArg(0));
3587 Value *Chr = EmitScalarExpr(E->getArg(1));
3588 Value *Size = EmitScalarExpr(E->getArg(2));
3589
3590 BasicBlock *Entry = Builder.GetInsertBlock();
3591 BasicBlock *CmpEq = createBasicBlock("wmemchr.eq");
3592 BasicBlock *Next = createBasicBlock("wmemchr.next");
3593 BasicBlock *Exit = createBasicBlock("wmemchr.exit");
3594 Value *SizeEq0 = Builder.CreateICmpEQ(Size, ConstantInt::get(SizeTy, 0));
3595 Builder.CreateCondBr(SizeEq0, Exit, CmpEq);
3596
3597 EmitBlock(CmpEq);
3598 PHINode *StrPhi = Builder.CreatePHI(Str->getType(), 2);
3599 StrPhi->addIncoming(Str, Entry);
3600 PHINode *SizePhi = Builder.CreatePHI(SizeTy, 2);
3601 SizePhi->addIncoming(Size, Entry);
3602 CharUnits WCharAlign =
3603 getContext().getTypeAlignInChars(getContext().WCharTy);
3604 Value *StrCh = Builder.CreateAlignedLoad(WCharTy, StrPhi, WCharAlign);
3605 Value *FoundChr = Builder.CreateConstInBoundsGEP1_32(WCharTy, StrPhi, 0);
3606 Value *StrEqChr = Builder.CreateICmpEQ(StrCh, Chr);
3607 Builder.CreateCondBr(StrEqChr, Exit, Next);
3608
3609 EmitBlock(Next);
3610 Value *NextStr = Builder.CreateConstInBoundsGEP1_32(WCharTy, StrPhi, 1);
3611 Value *NextSize = Builder.CreateSub(SizePhi, ConstantInt::get(SizeTy, 1));
3612 Value *NextSizeEq0 =
3613 Builder.CreateICmpEQ(NextSize, ConstantInt::get(SizeTy, 0));
3614 Builder.CreateCondBr(NextSizeEq0, Exit, CmpEq);
3615 StrPhi->addIncoming(NextStr, Next);
3616 SizePhi->addIncoming(NextSize, Next);
3617
3618 EmitBlock(Exit);
3619 PHINode *Ret = Builder.CreatePHI(Str->getType(), 3);
3620 Ret->addIncoming(llvm::Constant::getNullValue(Str->getType()), Entry);
3621 Ret->addIncoming(llvm::Constant::getNullValue(Str->getType()), Next);
3622 Ret->addIncoming(FoundChr, CmpEq);
3623 return RValue::get(Ret);
3624 }
3625 case Builtin::BI__builtin_wmemcmp: {
3626 // The MSVC runtime library does not provide a definition of wmemcmp, so we
3627 // need an inline implementation.
3628 if (!getTarget().getTriple().isOSMSVCRT())
3629 break;
3630
3631 llvm::Type *WCharTy = ConvertType(getContext().WCharTy);
3632
3633 Value *Dst = EmitScalarExpr(E->getArg(0));
3634 Value *Src = EmitScalarExpr(E->getArg(1));
3635 Value *Size = EmitScalarExpr(E->getArg(2));
3636
3637 BasicBlock *Entry = Builder.GetInsertBlock();
3638 BasicBlock *CmpGT = createBasicBlock("wmemcmp.gt");
3639 BasicBlock *CmpLT = createBasicBlock("wmemcmp.lt");
3640 BasicBlock *Next = createBasicBlock("wmemcmp.next");
3641 BasicBlock *Exit = createBasicBlock("wmemcmp.exit");
3642 Value *SizeEq0 = Builder.CreateICmpEQ(Size, ConstantInt::get(SizeTy, 0));
3643 Builder.CreateCondBr(SizeEq0, Exit, CmpGT);
3644
3645 EmitBlock(CmpGT);
3646 PHINode *DstPhi = Builder.CreatePHI(Dst->getType(), 2);
3647 DstPhi->addIncoming(Dst, Entry);
3648 PHINode *SrcPhi = Builder.CreatePHI(Src->getType(), 2);
3649 SrcPhi->addIncoming(Src, Entry);
3650 PHINode *SizePhi = Builder.CreatePHI(SizeTy, 2);
3651 SizePhi->addIncoming(Size, Entry);
3652 CharUnits WCharAlign =
3653 getContext().getTypeAlignInChars(getContext().WCharTy);
3654 Value *DstCh = Builder.CreateAlignedLoad(WCharTy, DstPhi, WCharAlign);
3655 Value *SrcCh = Builder.CreateAlignedLoad(WCharTy, SrcPhi, WCharAlign);
3656 Value *DstGtSrc = Builder.CreateICmpUGT(DstCh, SrcCh);
3657 Builder.CreateCondBr(DstGtSrc, Exit, CmpLT);
3658
3659 EmitBlock(CmpLT);
3660 Value *DstLtSrc = Builder.CreateICmpULT(DstCh, SrcCh);
3661 Builder.CreateCondBr(DstLtSrc, Exit, Next);
3662
3663 EmitBlock(Next);
3664 Value *NextDst = Builder.CreateConstInBoundsGEP1_32(WCharTy, DstPhi, 1);
3665 Value *NextSrc = Builder.CreateConstInBoundsGEP1_32(WCharTy, SrcPhi, 1);
3666 Value *NextSize = Builder.CreateSub(SizePhi, ConstantInt::get(SizeTy, 1));
3667 Value *NextSizeEq0 =
3668 Builder.CreateICmpEQ(NextSize, ConstantInt::get(SizeTy, 0));
3669 Builder.CreateCondBr(NextSizeEq0, Exit, CmpGT);
3670 DstPhi->addIncoming(NextDst, Next);
3671 SrcPhi->addIncoming(NextSrc, Next);
3672 SizePhi->addIncoming(NextSize, Next);
3673
3674 EmitBlock(Exit);
3675 PHINode *Ret = Builder.CreatePHI(IntTy, 4);
3676 Ret->addIncoming(ConstantInt::get(IntTy, 0), Entry);
3677 Ret->addIncoming(ConstantInt::get(IntTy, 1), CmpGT);
3678 Ret->addIncoming(ConstantInt::get(IntTy, -1), CmpLT);
3679 Ret->addIncoming(ConstantInt::get(IntTy, 0), Next);
3680 return RValue::get(Ret);
3681 }
3682 case Builtin::BI__builtin_dwarf_cfa: {
3683 // The offset in bytes from the first argument to the CFA.
3684 //
3685 // Why on earth is this in the frontend? Is there any reason at
3686 // all that the backend can't reasonably determine this while
3687 // lowering llvm.eh.dwarf.cfa()?
3688 //
3689 // TODO: If there's a satisfactory reason, add a target hook for
3690 // this instead of hard-coding 0, which is correct for most targets.
3691 int32_t Offset = 0;
3692
3693 Function *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa);
3694 return RValue::get(Builder.CreateCall(F,
3695 llvm::ConstantInt::get(Int32Ty, Offset)));
3696 }
3697 case Builtin::BI__builtin_return_address: {
3698 Value *Depth = ConstantEmitter(*this).emitAbstract(E->getArg(0),
3699 getContext().UnsignedIntTy);
3700 Function *F = CGM.getIntrinsic(Intrinsic::returnaddress);
3701 return RValue::get(Builder.CreateCall(F, Depth));
3702 }
3703 case Builtin::BI_ReturnAddress: {
3704 Function *F = CGM.getIntrinsic(Intrinsic::returnaddress);
3705 return RValue::get(Builder.CreateCall(F, Builder.getInt32(0)));
3706 }
3707 case Builtin::BI__builtin_frame_address: {
3708 Value *Depth = ConstantEmitter(*this).emitAbstract(E->getArg(0),
3709 getContext().UnsignedIntTy);
3710 Function *F = CGM.getIntrinsic(Intrinsic::frameaddress, AllocaInt8PtrTy);
3711 return RValue::get(Builder.CreateCall(F, Depth));
3712 }
3713 case Builtin::BI__builtin_extract_return_addr: {
3714 Value *Address = EmitScalarExpr(E->getArg(0));
3715 Value *Result = getTargetHooks().decodeReturnAddress(*this, Address);
3716 return RValue::get(Result);
3717 }
3718 case Builtin::BI__builtin_frob_return_addr: {
3719 Value *Address = EmitScalarExpr(E->getArg(0));
3720 Value *Result = getTargetHooks().encodeReturnAddress(*this, Address);
3721 return RValue::get(Result);
3722 }
3723 case Builtin::BI__builtin_dwarf_sp_column: {
3724 llvm::IntegerType *Ty
3725 = cast<llvm::IntegerType>(ConvertType(E->getType()));
3726 int Column = getTargetHooks().getDwarfEHStackPointer(CGM);
3727 if (Column == -1) {
3728 CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column");
3729 return RValue::get(llvm::UndefValue::get(Ty));
3730 }
3731 return RValue::get(llvm::ConstantInt::get(Ty, Column, true));
3732 }
3733 case Builtin::BI__builtin_init_dwarf_reg_size_table: {
3734 Value *Address = EmitScalarExpr(E->getArg(0));
3735 if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address))
3736 CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table");
3737 return RValue::get(llvm::UndefValue::get(ConvertType(E->getType())));
3738 }
3739 case Builtin::BI__builtin_eh_return: {
3740 Value *Int = EmitScalarExpr(E->getArg(0));
3741 Value *Ptr = EmitScalarExpr(E->getArg(1));
3742
3743 llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType());
3744 assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) &&
3745 "LLVM's __builtin_eh_return only supports 32- and 64-bit variants");
3746 Function *F =
3747 CGM.getIntrinsic(IntTy->getBitWidth() == 32 ? Intrinsic::eh_return_i32
3748 : Intrinsic::eh_return_i64);
3749 Builder.CreateCall(F, {Int, Ptr});
3750 Builder.CreateUnreachable();
3751
3752 // We do need to preserve an insertion point.
3753 EmitBlock(createBasicBlock("builtin_eh_return.cont"));
3754
3755 return RValue::get(nullptr);
3756 }
3757 case Builtin::BI__builtin_unwind_init: {
3758 Function *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init);
3759 Builder.CreateCall(F);
3760 return RValue::get(nullptr);
3761 }
3762 case Builtin::BI__builtin_extend_pointer: {
3763 // Extends a pointer to the size of an _Unwind_Word, which is
3764 // uint64_t on all platforms. Generally this gets poked into a
3765 // register and eventually used as an address, so if the
3766 // addressing registers are wider than pointers and the platform
3767 // doesn't implicitly ignore high-order bits when doing
3768 // addressing, we need to make sure we zext / sext based on
3769 // the platform's expectations.
3770 //
3771 // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html
3772
3773 // Cast the pointer to intptr_t.
3774 Value *Ptr = EmitScalarExpr(E->getArg(0));
3775 Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast");
3776
3777 // If that's 64 bits, we're done.
3778 if (IntPtrTy->getBitWidth() == 64)
3779 return RValue::get(Result);
3780
3781 // Otherwise, ask the codegen data what to do.
3782 if (getTargetHooks().extendPointerWithSExt())
3783 return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext"));
3784 else
3785 return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext"));
3786 }
3787 case Builtin::BI__builtin_setjmp: {
3788 // Buffer is a void**.
3789 Address Buf = EmitPointerWithAlignment(E->getArg(0));
3790
3791 // Store the frame pointer to the setjmp buffer.
3792 Value *FrameAddr = Builder.CreateCall(
3793 CGM.getIntrinsic(Intrinsic::frameaddress, AllocaInt8PtrTy),
3794 ConstantInt::get(Int32Ty, 0));
3795 Builder.CreateStore(FrameAddr, Buf);
3796
3797 // Store the stack pointer to the setjmp buffer.
3798 Value *StackAddr =
3799 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave));
3800 Address StackSaveSlot = Builder.CreateConstInBoundsGEP(Buf, 2);
3801 Builder.CreateStore(StackAddr, StackSaveSlot);
3802
3803 // Call LLVM's EH setjmp, which is lightweight.
3804 Function *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp);
3805 Buf = Builder.CreateElementBitCast(Buf, Int8Ty);
3806 return RValue::get(Builder.CreateCall(F, Buf.getPointer()));
3807 }
3808 case Builtin::BI__builtin_longjmp: {
3809 Value *Buf = EmitScalarExpr(E->getArg(0));
3810 Buf = Builder.CreateBitCast(Buf, Int8PtrTy);
3811
3812 // Call LLVM's EH longjmp, which is lightweight.
3813 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf);
3814
3815 // longjmp doesn't return; mark this as unreachable.
3816 Builder.CreateUnreachable();
3817
3818 // We do need to preserve an insertion point.
3819 EmitBlock(createBasicBlock("longjmp.cont"));
3820
3821 return RValue::get(nullptr);
3822 }
3823 case Builtin::BI__builtin_launder: {
3824 const Expr *Arg = E->getArg(0);
3825 QualType ArgTy = Arg->getType()->getPointeeType();
3826 Value *Ptr = EmitScalarExpr(Arg);
3827 if (TypeRequiresBuiltinLaunder(CGM, ArgTy))
3828 Ptr = Builder.CreateLaunderInvariantGroup(Ptr);
3829
3830 return RValue::get(Ptr);
3831 }
3832 case Builtin::BI__sync_fetch_and_add:
3833 case Builtin::BI__sync_fetch_and_sub:
3834 case Builtin::BI__sync_fetch_and_or:
3835 case Builtin::BI__sync_fetch_and_and:
3836 case Builtin::BI__sync_fetch_and_xor:
3837 case Builtin::BI__sync_fetch_and_nand:
3838 case Builtin::BI__sync_add_and_fetch:
3839 case Builtin::BI__sync_sub_and_fetch:
3840 case Builtin::BI__sync_and_and_fetch:
3841 case Builtin::BI__sync_or_and_fetch:
3842 case Builtin::BI__sync_xor_and_fetch:
3843 case Builtin::BI__sync_nand_and_fetch:
3844 case Builtin::BI__sync_val_compare_and_swap:
3845 case Builtin::BI__sync_bool_compare_and_swap:
3846 case Builtin::BI__sync_lock_test_and_set:
3847 case Builtin::BI__sync_lock_release:
3848 case Builtin::BI__sync_swap:
3849 llvm_unreachable("Shouldn't make it through sema");
3850 case Builtin::BI__sync_fetch_and_add_1:
3851 case Builtin::BI__sync_fetch_and_add_2:
3852 case Builtin::BI__sync_fetch_and_add_4:
3853 case Builtin::BI__sync_fetch_and_add_8:
3854 case Builtin::BI__sync_fetch_and_add_16:
3855 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E);
3856 case Builtin::BI__sync_fetch_and_sub_1:
3857 case Builtin::BI__sync_fetch_and_sub_2:
3858 case Builtin::BI__sync_fetch_and_sub_4:
3859 case Builtin::BI__sync_fetch_and_sub_8:
3860 case Builtin::BI__sync_fetch_and_sub_16:
3861 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E);
3862 case Builtin::BI__sync_fetch_and_or_1:
3863 case Builtin::BI__sync_fetch_and_or_2:
3864 case Builtin::BI__sync_fetch_and_or_4:
3865 case Builtin::BI__sync_fetch_and_or_8:
3866 case Builtin::BI__sync_fetch_and_or_16:
3867 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E);
3868 case Builtin::BI__sync_fetch_and_and_1:
3869 case Builtin::BI__sync_fetch_and_and_2:
3870 case Builtin::BI__sync_fetch_and_and_4:
3871 case Builtin::BI__sync_fetch_and_and_8:
3872 case Builtin::BI__sync_fetch_and_and_16:
3873 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E);
3874 case Builtin::BI__sync_fetch_and_xor_1:
3875 case Builtin::BI__sync_fetch_and_xor_2:
3876 case Builtin::BI__sync_fetch_and_xor_4:
3877 case Builtin::BI__sync_fetch_and_xor_8:
3878 case Builtin::BI__sync_fetch_and_xor_16:
3879 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E);
3880 case Builtin::BI__sync_fetch_and_nand_1:
3881 case Builtin::BI__sync_fetch_and_nand_2:
3882 case Builtin::BI__sync_fetch_and_nand_4:
3883 case Builtin::BI__sync_fetch_and_nand_8:
3884 case Builtin::BI__sync_fetch_and_nand_16:
3885 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Nand, E);
3886
3887 // Clang extensions: not overloaded yet.
3888 case Builtin::BI__sync_fetch_and_min:
3889 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E);
3890 case Builtin::BI__sync_fetch_and_max:
3891 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E);
3892 case Builtin::BI__sync_fetch_and_umin:
3893 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E);
3894 case Builtin::BI__sync_fetch_and_umax:
3895 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E);
3896
3897 case Builtin::BI__sync_add_and_fetch_1:
3898 case Builtin::BI__sync_add_and_fetch_2:
3899 case Builtin::BI__sync_add_and_fetch_4:
3900 case Builtin::BI__sync_add_and_fetch_8:
3901 case Builtin::BI__sync_add_and_fetch_16:
3902 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E,
3903 llvm::Instruction::Add);
3904 case Builtin::BI__sync_sub_and_fetch_1:
3905 case Builtin::BI__sync_sub_and_fetch_2:
3906 case Builtin::BI__sync_sub_and_fetch_4:
3907 case Builtin::BI__sync_sub_and_fetch_8:
3908 case Builtin::BI__sync_sub_and_fetch_16:
3909 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E,
3910 llvm::Instruction::Sub);
3911 case Builtin::BI__sync_and_and_fetch_1:
3912 case Builtin::BI__sync_and_and_fetch_2:
3913 case Builtin::BI__sync_and_and_fetch_4:
3914 case Builtin::BI__sync_and_and_fetch_8:
3915 case Builtin::BI__sync_and_and_fetch_16:
3916 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E,
3917 llvm::Instruction::And);
3918 case Builtin::BI__sync_or_and_fetch_1:
3919 case Builtin::BI__sync_or_and_fetch_2:
3920 case Builtin::BI__sync_or_and_fetch_4:
3921 case Builtin::BI__sync_or_and_fetch_8:
3922 case Builtin::BI__sync_or_and_fetch_16:
3923 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E,
3924 llvm::Instruction::Or);
3925 case Builtin::BI__sync_xor_and_fetch_1:
3926 case Builtin::BI__sync_xor_and_fetch_2:
3927 case Builtin::BI__sync_xor_and_fetch_4:
3928 case Builtin::BI__sync_xor_and_fetch_8:
3929 case Builtin::BI__sync_xor_and_fetch_16:
3930 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E,
3931 llvm::Instruction::Xor);
3932 case Builtin::BI__sync_nand_and_fetch_1:
3933 case Builtin::BI__sync_nand_and_fetch_2:
3934 case Builtin::BI__sync_nand_and_fetch_4:
3935 case Builtin::BI__sync_nand_and_fetch_8:
3936 case Builtin::BI__sync_nand_and_fetch_16:
3937 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Nand, E,
3938 llvm::Instruction::And, true);
3939
3940 case Builtin::BI__sync_val_compare_and_swap_1:
3941 case Builtin::BI__sync_val_compare_and_swap_2:
3942 case Builtin::BI__sync_val_compare_and_swap_4:
3943 case Builtin::BI__sync_val_compare_and_swap_8:
3944 case Builtin::BI__sync_val_compare_and_swap_16:
3945 return RValue::get(MakeAtomicCmpXchgValue(*this, E, false));
3946
3947 case Builtin::BI__sync_bool_compare_and_swap_1:
3948 case Builtin::BI__sync_bool_compare_and_swap_2:
3949 case Builtin::BI__sync_bool_compare_and_swap_4:
3950 case Builtin::BI__sync_bool_compare_and_swap_8:
3951 case Builtin::BI__sync_bool_compare_and_swap_16:
3952 return RValue::get(MakeAtomicCmpXchgValue(*this, E, true));
3953
3954 case Builtin::BI__sync_swap_1:
3955 case Builtin::BI__sync_swap_2:
3956 case Builtin::BI__sync_swap_4:
3957 case Builtin::BI__sync_swap_8:
3958 case Builtin::BI__sync_swap_16:
3959 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E);
3960
3961 case Builtin::BI__sync_lock_test_and_set_1:
3962 case Builtin::BI__sync_lock_test_and_set_2:
3963 case Builtin::BI__sync_lock_test_and_set_4:
3964 case Builtin::BI__sync_lock_test_and_set_8:
3965 case Builtin::BI__sync_lock_test_and_set_16:
3966 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E);
3967
3968 case Builtin::BI__sync_lock_release_1:
3969 case Builtin::BI__sync_lock_release_2:
3970 case Builtin::BI__sync_lock_release_4:
3971 case Builtin::BI__sync_lock_release_8:
3972 case Builtin::BI__sync_lock_release_16: {
3973 Value *Ptr = EmitScalarExpr(E->getArg(0));
3974 QualType ElTy = E->getArg(0)->getType()->getPointeeType();
3975 CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy);
3976 llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(),
3977 StoreSize.getQuantity() * 8);
3978 Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo());
3979 llvm::StoreInst *Store =
3980 Builder.CreateAlignedStore(llvm::Constant::getNullValue(ITy), Ptr,
3981 StoreSize);
3982 Store->setAtomic(llvm::AtomicOrdering::Release);
3983 return RValue::get(nullptr);
3984 }
3985
3986 case Builtin::BI__sync_synchronize: {
3987 // We assume this is supposed to correspond to a C++0x-style
3988 // sequentially-consistent fence (i.e. this is only usable for
3989 // synchronization, not device I/O or anything like that). This intrinsic
3990 // is really badly designed in the sense that in theory, there isn't
3991 // any way to safely use it... but in practice, it mostly works
3992 // to use it with non-atomic loads and stores to get acquire/release
3993 // semantics.
3994 Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent);
3995 return RValue::get(nullptr);
3996 }
3997
3998 case Builtin::BI__builtin_nontemporal_load:
3999 return RValue::get(EmitNontemporalLoad(*this, E));
4000 case Builtin::BI__builtin_nontemporal_store:
4001 return RValue::get(EmitNontemporalStore(*this, E));
4002 case Builtin::BI__c11_atomic_is_lock_free:
4003 case Builtin::BI__atomic_is_lock_free: {
4004 // Call "bool __atomic_is_lock_free(size_t size, void *ptr)". For the
4005 // __c11 builtin, ptr is 0 (indicating a properly-aligned object), since
4006 // _Atomic(T) is always properly-aligned.
4007 const char *LibCallName = "__atomic_is_lock_free";
4008 CallArgList Args;
4009 Args.add(RValue::get(EmitScalarExpr(E->getArg(0))),
4010 getContext().getSizeType());
4011 if (BuiltinID == Builtin::BI__atomic_is_lock_free)
4012 Args.add(RValue::get(EmitScalarExpr(E->getArg(1))),
4013 getContext().VoidPtrTy);
4014 else
4015 Args.add(RValue::get(llvm::Constant::getNullValue(VoidPtrTy)),
4016 getContext().VoidPtrTy);
4017 const CGFunctionInfo &FuncInfo =
4018 CGM.getTypes().arrangeBuiltinFunctionCall(E->getType(), Args);
4019 llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FuncInfo);
4020 llvm::FunctionCallee Func = CGM.CreateRuntimeFunction(FTy, LibCallName);
4021 return EmitCall(FuncInfo, CGCallee::forDirect(Func),
4022 ReturnValueSlot(), Args);
4023 }
4024
4025 case Builtin::BI__atomic_test_and_set: {
4026 // Look at the argument type to determine whether this is a volatile
4027 // operation. The parameter type is always volatile.
4028 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType();
4029 bool Volatile =
4030 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified();
4031
4032 Value *Ptr = EmitScalarExpr(E->getArg(0));
4033 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace();
4034 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace));
4035 Value *NewVal = Builder.getInt8(1);
4036 Value *Order = EmitScalarExpr(E->getArg(1));
4037 if (isa<llvm::ConstantInt>(Order)) {
4038 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue();
4039 AtomicRMWInst *Result = nullptr;
4040 switch (ord) {
4041 case 0: // memory_order_relaxed
4042 default: // invalid order
4043 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal,
4044 llvm::AtomicOrdering::Monotonic);
4045 break;
4046 case 1: // memory_order_consume
4047 case 2: // memory_order_acquire
4048 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal,
4049 llvm::AtomicOrdering::Acquire);
4050 break;
4051 case 3: // memory_order_release
4052 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal,
4053 llvm::AtomicOrdering::Release);
4054 break;
4055 case 4: // memory_order_acq_rel
4056
4057 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal,
4058 llvm::AtomicOrdering::AcquireRelease);
4059 break;
4060 case 5: // memory_order_seq_cst
4061 Result = Builder.CreateAtomicRMW(
4062 llvm::AtomicRMWInst::Xchg, Ptr, NewVal,
4063 llvm::AtomicOrdering::SequentiallyConsistent);
4064 break;
4065 }
4066 Result->setVolatile(Volatile);
4067 return RValue::get(Builder.CreateIsNotNull(Result, "tobool"));
4068 }
4069
4070 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn);
4071
4072 llvm::BasicBlock *BBs[5] = {
4073 createBasicBlock("monotonic", CurFn),
4074 createBasicBlock("acquire", CurFn),
4075 createBasicBlock("release", CurFn),
4076 createBasicBlock("acqrel", CurFn),
4077 createBasicBlock("seqcst", CurFn)
4078 };
4079 llvm::AtomicOrdering Orders[5] = {
4080 llvm::AtomicOrdering::Monotonic, llvm::AtomicOrdering::Acquire,
4081 llvm::AtomicOrdering::Release, llvm::AtomicOrdering::AcquireRelease,
4082 llvm::AtomicOrdering::SequentiallyConsistent};
4083
4084 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false);
4085 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]);
4086
4087 Builder.SetInsertPoint(ContBB);
4088 PHINode *Result = Builder.CreatePHI(Int8Ty, 5, "was_set");
4089
4090 for (unsigned i = 0; i < 5; ++i) {
4091 Builder.SetInsertPoint(BBs[i]);
4092 AtomicRMWInst *RMW = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg,
4093 Ptr, NewVal, Orders[i]);
4094 RMW->setVolatile(Volatile);
4095 Result->addIncoming(RMW, BBs[i]);
4096 Builder.CreateBr(ContBB);
4097 }
4098
4099 SI->addCase(Builder.getInt32(0), BBs[0]);
4100 SI->addCase(Builder.getInt32(1), BBs[1]);
4101 SI->addCase(Builder.getInt32(2), BBs[1]);
4102 SI->addCase(Builder.getInt32(3), BBs[2]);
4103 SI->addCase(Builder.getInt32(4), BBs[3]);
4104 SI->addCase(Builder.getInt32(5), BBs[4]);
4105
4106 Builder.SetInsertPoint(ContBB);
4107 return RValue::get(Builder.CreateIsNotNull(Result, "tobool"));
4108 }
4109
4110 case Builtin::BI__atomic_clear: {
4111 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType();
4112 bool Volatile =
4113 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified();
4114
4115 Address Ptr = EmitPointerWithAlignment(E->getArg(0));
4116 Ptr = Builder.CreateElementBitCast(Ptr, Int8Ty);
4117 Value *NewVal = Builder.getInt8(0);
4118 Value *Order = EmitScalarExpr(E->getArg(1));
4119 if (isa<llvm::ConstantInt>(Order)) {
4120 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue();
4121 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile);
4122 switch (ord) {
4123 case 0: // memory_order_relaxed
4124 default: // invalid order
4125 Store->setOrdering(llvm::AtomicOrdering::Monotonic);
4126 break;
4127 case 3: // memory_order_release
4128 Store->setOrdering(llvm::AtomicOrdering::Release);
4129 break;
4130 case 5: // memory_order_seq_cst
4131 Store->setOrdering(llvm::AtomicOrdering::SequentiallyConsistent);
4132 break;
4133 }
4134 return RValue::get(nullptr);
4135 }
4136
4137 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn);
4138
4139 llvm::BasicBlock *BBs[3] = {
4140 createBasicBlock("monotonic", CurFn),
4141 createBasicBlock("release", CurFn),
4142 createBasicBlock("seqcst", CurFn)
4143 };
4144 llvm::AtomicOrdering Orders[3] = {
4145 llvm::AtomicOrdering::Monotonic, llvm::AtomicOrdering::Release,
4146 llvm::AtomicOrdering::SequentiallyConsistent};
4147
4148 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false);
4149 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]);
4150
4151 for (unsigned i = 0; i < 3; ++i) {
4152 Builder.SetInsertPoint(BBs[i]);
4153 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile);
4154 Store->setOrdering(Orders[i]);
4155 Builder.CreateBr(ContBB);
4156 }
4157
4158 SI->addCase(Builder.getInt32(0), BBs[0]);
4159 SI->addCase(Builder.getInt32(3), BBs[1]);
4160 SI->addCase(Builder.getInt32(5), BBs[2]);
4161
4162 Builder.SetInsertPoint(ContBB);
4163 return RValue::get(nullptr);
4164 }
4165
4166 case Builtin::BI__atomic_thread_fence:
4167 case Builtin::BI__atomic_signal_fence:
4168 case Builtin::BI__c11_atomic_thread_fence:
4169 case Builtin::BI__c11_atomic_signal_fence: {
4170 llvm::SyncScope::ID SSID;
4171 if (BuiltinID == Builtin::BI__atomic_signal_fence ||
4172 BuiltinID == Builtin::BI__c11_atomic_signal_fence)
4173 SSID = llvm::SyncScope::SingleThread;
4174 else
4175 SSID = llvm::SyncScope::System;
4176 Value *Order = EmitScalarExpr(E->getArg(0));
4177 if (isa<llvm::ConstantInt>(Order)) {
4178 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue();
4179 switch (ord) {
4180 case 0: // memory_order_relaxed
4181 default: // invalid order
4182 break;
4183 case 1: // memory_order_consume
4184 case 2: // memory_order_acquire
4185 Builder.CreateFence(llvm::AtomicOrdering::Acquire, SSID);
4186 break;
4187 case 3: // memory_order_release
4188 Builder.CreateFence(llvm::AtomicOrdering::Release, SSID);
4189 break;
4190 case 4: // memory_order_acq_rel
4191 Builder.CreateFence(llvm::AtomicOrdering::AcquireRelease, SSID);
4192 break;
4193 case 5: // memory_order_seq_cst
4194 Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent, SSID);
4195 break;
4196 }
4197 return RValue::get(nullptr);
4198 }
4199
4200 llvm::BasicBlock *AcquireBB, *ReleaseBB, *AcqRelBB, *SeqCstBB;
4201 AcquireBB = createBasicBlock("acquire", CurFn);
4202 ReleaseBB = createBasicBlock("release", CurFn);
4203 AcqRelBB = createBasicBlock("acqrel", CurFn);
4204 SeqCstBB = createBasicBlock("seqcst", CurFn);
4205 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn);
4206
4207 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false);
4208 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, ContBB);
4209
4210 Builder.SetInsertPoint(AcquireBB);
4211 Builder.CreateFence(llvm::AtomicOrdering::Acquire, SSID);
4212 Builder.CreateBr(ContBB);
4213 SI->addCase(Builder.getInt32(1), AcquireBB);
4214 SI->addCase(Builder.getInt32(2), AcquireBB);
4215
4216 Builder.SetInsertPoint(ReleaseBB);
4217 Builder.CreateFence(llvm::AtomicOrdering::Release, SSID);
4218 Builder.CreateBr(ContBB);
4219 SI->addCase(Builder.getInt32(3), ReleaseBB);
4220
4221 Builder.SetInsertPoint(AcqRelBB);
4222 Builder.CreateFence(llvm::AtomicOrdering::AcquireRelease, SSID);
4223 Builder.CreateBr(ContBB);
4224 SI->addCase(Builder.getInt32(4), AcqRelBB);
4225
4226 Builder.SetInsertPoint(SeqCstBB);
4227 Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent, SSID);
4228 Builder.CreateBr(ContBB);
4229 SI->addCase(Builder.getInt32(5), SeqCstBB);
4230
4231 Builder.SetInsertPoint(ContBB);
4232 return RValue::get(nullptr);
4233 }
4234
4235 case Builtin::BI__builtin_signbit:
4236 case Builtin::BI__builtin_signbitf:
4237 case Builtin::BI__builtin_signbitl: {
4238 return RValue::get(
4239 Builder.CreateZExt(EmitSignBit(*this, EmitScalarExpr(E->getArg(0))),
4240 ConvertType(E->getType())));
4241 }
4242 case Builtin::BI__warn_memset_zero_len:
4243 return RValue::getIgnored();
4244 case Builtin::BI__annotation: {
4245 // Re-encode each wide string to UTF8 and make an MDString.
4246 SmallVector<Metadata *, 1> Strings;
4247 for (const Expr *Arg : E->arguments()) {
4248 const auto *Str = cast<StringLiteral>(Arg->IgnoreParenCasts());
4249 assert(Str->getCharByteWidth() == 2);
4250 StringRef WideBytes = Str->getBytes();
4251 std::string StrUtf8;
4252 if (!convertUTF16ToUTF8String(
4253 ArrayRef(WideBytes.data(), WideBytes.size()), StrUtf8)) {
4254 CGM.ErrorUnsupported(E, "non-UTF16 __annotation argument");
4255 continue;
4256 }
4257 Strings.push_back(llvm::MDString::get(getLLVMContext(), StrUtf8));
4258 }
4259
4260 // Build and MDTuple of MDStrings and emit the intrinsic call.
4261 llvm::Function *F =
4262 CGM.getIntrinsic(llvm::Intrinsic::codeview_annotation, {});
4263 MDTuple *StrTuple = MDTuple::get(getLLVMContext(), Strings);
4264 Builder.CreateCall(F, MetadataAsValue::get(getLLVMContext(), StrTuple));
4265 return RValue::getIgnored();
4266 }
4267 case Builtin::BI__builtin_annotation: {
4268 llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0));
4269 llvm::Function *F =
4270 CGM.getIntrinsic(llvm::Intrinsic::annotation,
4271 {AnnVal->getType(), CGM.ConstGlobalsPtrTy});
4272
4273 // Get the annotation string, go through casts. Sema requires this to be a
4274 // non-wide string literal, potentially casted, so the cast<> is safe.
4275 const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts();
4276 StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString();
4277 return RValue::get(
4278 EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc(), nullptr));
4279 }
4280 case Builtin::BI__builtin_addcb:
4281 case Builtin::BI__builtin_addcs:
4282 case Builtin::BI__builtin_addc:
4283 case Builtin::BI__builtin_addcl:
4284 case Builtin::BI__builtin_addcll:
4285 case Builtin::BI__builtin_subcb:
4286 case Builtin::BI__builtin_subcs:
4287 case Builtin::BI__builtin_subc:
4288 case Builtin::BI__builtin_subcl:
4289 case Builtin::BI__builtin_subcll: {
4290
4291 // We translate all of these builtins from expressions of the form:
4292 // int x = ..., y = ..., carryin = ..., carryout, result;
4293 // result = __builtin_addc(x, y, carryin, &carryout);
4294 //
4295 // to LLVM IR of the form:
4296 //
4297 // %tmp1 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %x, i32 %y)
4298 // %tmpsum1 = extractvalue {i32, i1} %tmp1, 0
4299 // %carry1 = extractvalue {i32, i1} %tmp1, 1
4300 // %tmp2 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %tmpsum1,
4301 // i32 %carryin)
4302 // %result = extractvalue {i32, i1} %tmp2, 0
4303 // %carry2 = extractvalue {i32, i1} %tmp2, 1
4304 // %tmp3 = or i1 %carry1, %carry2
4305 // %tmp4 = zext i1 %tmp3 to i32
4306 // store i32 %tmp4, i32* %carryout
4307
4308 // Scalarize our inputs.
4309 llvm::Value *X = EmitScalarExpr(E->getArg(0));
4310 llvm::Value *Y = EmitScalarExpr(E->getArg(1));
4311 llvm::Value *Carryin = EmitScalarExpr(E->getArg(2));
4312 Address CarryOutPtr = EmitPointerWithAlignment(E->getArg(3));
4313
4314 // Decide if we are lowering to a uadd.with.overflow or usub.with.overflow.
4315 llvm::Intrinsic::ID IntrinsicId;
4316 switch (BuiltinID) {
4317 default: llvm_unreachable("Unknown multiprecision builtin id.");
4318 case Builtin::BI__builtin_addcb:
4319 case Builtin::BI__builtin_addcs:
4320 case Builtin::BI__builtin_addc:
4321 case Builtin::BI__builtin_addcl:
4322 case Builtin::BI__builtin_addcll:
4323 IntrinsicId = llvm::Intrinsic::uadd_with_overflow;
4324 break;
4325 case Builtin::BI__builtin_subcb:
4326 case Builtin::BI__builtin_subcs:
4327 case Builtin::BI__builtin_subc:
4328 case Builtin::BI__builtin_subcl:
4329 case Builtin::BI__builtin_subcll:
4330 IntrinsicId = llvm::Intrinsic::usub_with_overflow;
4331 break;
4332 }
4333
4334 // Construct our resulting LLVM IR expression.
4335 llvm::Value *Carry1;
4336 llvm::Value *Sum1 = EmitOverflowIntrinsic(*this, IntrinsicId,
4337 X, Y, Carry1);
4338 llvm::Value *Carry2;
4339 llvm::Value *Sum2 = EmitOverflowIntrinsic(*this, IntrinsicId,
4340 Sum1, Carryin, Carry2);
4341 llvm::Value *CarryOut = Builder.CreateZExt(Builder.CreateOr(Carry1, Carry2),
4342 X->getType());
4343 Builder.CreateStore(CarryOut, CarryOutPtr);
4344 return RValue::get(Sum2);
4345 }
4346
4347 case Builtin::BI__builtin_add_overflow:
4348 case Builtin::BI__builtin_sub_overflow:
4349 case Builtin::BI__builtin_mul_overflow: {
4350 const clang::Expr *LeftArg = E->getArg(0);
4351 const clang::Expr *RightArg = E->getArg(1);
4352 const clang::Expr *ResultArg = E->getArg(2);
4353
4354 clang::QualType ResultQTy =
4355 ResultArg->getType()->castAs<PointerType>()->getPointeeType();
4356
4357 WidthAndSignedness LeftInfo =
4358 getIntegerWidthAndSignedness(CGM.getContext(), LeftArg->getType());
4359 WidthAndSignedness RightInfo =
4360 getIntegerWidthAndSignedness(CGM.getContext(), RightArg->getType());
4361 WidthAndSignedness ResultInfo =
4362 getIntegerWidthAndSignedness(CGM.getContext(), ResultQTy);
4363
4364 // Handle mixed-sign multiplication as a special case, because adding
4365 // runtime or backend support for our generic irgen would be too expensive.
4366 if (isSpecialMixedSignMultiply(BuiltinID, LeftInfo, RightInfo, ResultInfo))
4367 return EmitCheckedMixedSignMultiply(*this, LeftArg, LeftInfo, RightArg,
4368 RightInfo, ResultArg, ResultQTy,
4369 ResultInfo);
4370
4371 if (isSpecialUnsignedMultiplySignedResult(BuiltinID, LeftInfo, RightInfo,
4372 ResultInfo))
4373 return EmitCheckedUnsignedMultiplySignedResult(
4374 *this, LeftArg, LeftInfo, RightArg, RightInfo, ResultArg, ResultQTy,
4375 ResultInfo);
4376
4377 WidthAndSignedness EncompassingInfo =
4378 EncompassingIntegerType({LeftInfo, RightInfo, ResultInfo});
4379
4380 llvm::Type *EncompassingLLVMTy =
4381 llvm::IntegerType::get(CGM.getLLVMContext(), EncompassingInfo.Width);
4382
4383 llvm::Type *ResultLLVMTy = CGM.getTypes().ConvertType(ResultQTy);
4384
4385 llvm::Intrinsic::ID IntrinsicId;
4386 switch (BuiltinID) {
4387 default:
4388 llvm_unreachable("Unknown overflow builtin id.");
4389 case Builtin::BI__builtin_add_overflow:
4390 IntrinsicId = EncompassingInfo.Signed
4391 ? llvm::Intrinsic::sadd_with_overflow
4392 : llvm::Intrinsic::uadd_with_overflow;
4393 break;
4394 case Builtin::BI__builtin_sub_overflow:
4395 IntrinsicId = EncompassingInfo.Signed
4396 ? llvm::Intrinsic::ssub_with_overflow
4397 : llvm::Intrinsic::usub_with_overflow;
4398 break;
4399 case Builtin::BI__builtin_mul_overflow:
4400 IntrinsicId = EncompassingInfo.Signed
4401 ? llvm::Intrinsic::smul_with_overflow
4402 : llvm::Intrinsic::umul_with_overflow;
4403 break;
4404 }
4405
4406 llvm::Value *Left = EmitScalarExpr(LeftArg);
4407 llvm::Value *Right = EmitScalarExpr(RightArg);
4408 Address ResultPtr = EmitPointerWithAlignment(ResultArg);
4409
4410 // Extend each operand to the encompassing type.
4411 Left = Builder.CreateIntCast(Left, EncompassingLLVMTy, LeftInfo.Signed);
4412 Right = Builder.CreateIntCast(Right, EncompassingLLVMTy, RightInfo.Signed);
4413
4414 // Perform the operation on the extended values.
4415 llvm::Value *Overflow, *Result;
4416 Result = EmitOverflowIntrinsic(*this, IntrinsicId, Left, Right, Overflow);
4417
4418 if (EncompassingInfo.Width > ResultInfo.Width) {
4419 // The encompassing type is wider than the result type, so we need to
4420 // truncate it.
4421 llvm::Value *ResultTrunc = Builder.CreateTrunc(Result, ResultLLVMTy);
4422
4423 // To see if the truncation caused an overflow, we will extend
4424 // the result and then compare it to the original result.
4425 llvm::Value *ResultTruncExt = Builder.CreateIntCast(
4426 ResultTrunc, EncompassingLLVMTy, ResultInfo.Signed);
4427 llvm::Value *TruncationOverflow =
4428 Builder.CreateICmpNE(Result, ResultTruncExt);
4429
4430 Overflow = Builder.CreateOr(Overflow, TruncationOverflow);
4431 Result = ResultTrunc;
4432 }
4433
4434 // Finally, store the result using the pointer.
4435 bool isVolatile =
4436 ResultArg->getType()->getPointeeType().isVolatileQualified();
4437 Builder.CreateStore(EmitToMemory(Result, ResultQTy), ResultPtr, isVolatile);
4438
4439 return RValue::get(Overflow);
4440 }
4441
4442 case Builtin::BI__builtin_uadd_overflow:
4443 case Builtin::BI__builtin_uaddl_overflow:
4444 case Builtin::BI__builtin_uaddll_overflow:
4445 case Builtin::BI__builtin_usub_overflow:
4446 case Builtin::BI__builtin_usubl_overflow:
4447 case Builtin::BI__builtin_usubll_overflow:
4448 case Builtin::BI__builtin_umul_overflow:
4449 case Builtin::BI__builtin_umull_overflow:
4450 case Builtin::BI__builtin_umulll_overflow:
4451 case Builtin::BI__builtin_sadd_overflow:
4452 case Builtin::BI__builtin_saddl_overflow:
4453 case Builtin::BI__builtin_saddll_overflow:
4454 case Builtin::BI__builtin_ssub_overflow:
4455 case Builtin::BI__builtin_ssubl_overflow:
4456 case Builtin::BI__builtin_ssubll_overflow:
4457 case Builtin::BI__builtin_smul_overflow:
4458 case Builtin::BI__builtin_smull_overflow:
4459 case Builtin::BI__builtin_smulll_overflow: {
4460
4461 // We translate all of these builtins directly to the relevant llvm IR node.
4462
4463 // Scalarize our inputs.
4464 llvm::Value *X = EmitScalarExpr(E->getArg(0));
4465 llvm::Value *Y = EmitScalarExpr(E->getArg(1));
4466 Address SumOutPtr = EmitPointerWithAlignment(E->getArg(2));
4467
4468 // Decide which of the overflow intrinsics we are lowering to:
4469 llvm::Intrinsic::ID IntrinsicId;
4470 switch (BuiltinID) {
4471 default: llvm_unreachable("Unknown overflow builtin id.");
4472 case Builtin::BI__builtin_uadd_overflow:
4473 case Builtin::BI__builtin_uaddl_overflow:
4474 case Builtin::BI__builtin_uaddll_overflow:
4475 IntrinsicId = llvm::Intrinsic::uadd_with_overflow;
4476 break;
4477 case Builtin::BI__builtin_usub_overflow:
4478 case Builtin::BI__builtin_usubl_overflow:
4479 case Builtin::BI__builtin_usubll_overflow:
4480 IntrinsicId = llvm::Intrinsic::usub_with_overflow;
4481 break;
4482 case Builtin::BI__builtin_umul_overflow:
4483 case Builtin::BI__builtin_umull_overflow:
4484 case Builtin::BI__builtin_umulll_overflow:
4485 IntrinsicId = llvm::Intrinsic::umul_with_overflow;
4486 break;
4487 case Builtin::BI__builtin_sadd_overflow:
4488 case Builtin::BI__builtin_saddl_overflow:
4489 case Builtin::BI__builtin_saddll_overflow:
4490 IntrinsicId = llvm::Intrinsic::sadd_with_overflow;
4491 break;
4492 case Builtin::BI__builtin_ssub_overflow:
4493 case Builtin::BI__builtin_ssubl_overflow:
4494 case Builtin::BI__builtin_ssubll_overflow:
4495 IntrinsicId = llvm::Intrinsic::ssub_with_overflow;
4496 break;
4497 case Builtin::BI__builtin_smul_overflow:
4498 case Builtin::BI__builtin_smull_overflow:
4499 case Builtin::BI__builtin_smulll_overflow:
4500 IntrinsicId = llvm::Intrinsic::smul_with_overflow;
4501 break;
4502 }
4503
4504
4505 llvm::Value *Carry;
4506 llvm::Value *Sum = EmitOverflowIntrinsic(*this, IntrinsicId, X, Y, Carry);
4507 Builder.CreateStore(Sum, SumOutPtr);
4508
4509 return RValue::get(Carry);
4510 }
4511 case Builtin::BIaddressof:
4512 case Builtin::BI__addressof:
4513 case Builtin::BI__builtin_addressof:
4514 return RValue::get(EmitLValue(E->getArg(0)).getPointer(*this));
4515 case Builtin::BI__builtin_function_start:
4516 return RValue::get(CGM.GetFunctionStart(
4517 E->getArg(0)->getAsBuiltinConstantDeclRef(CGM.getContext())));
4518 case Builtin::BI__builtin_operator_new:
4519 return EmitBuiltinNewDeleteCall(
4520 E->getCallee()->getType()->castAs<FunctionProtoType>(), E, false);
4521 case Builtin::BI__builtin_operator_delete:
4522 EmitBuiltinNewDeleteCall(
4523 E->getCallee()->getType()->castAs<FunctionProtoType>(), E, true);
4524 return RValue::get(nullptr);
4525
4526 case Builtin::BI__builtin_is_aligned:
4527 return EmitBuiltinIsAligned(E);
4528 case Builtin::BI__builtin_align_up:
4529 return EmitBuiltinAlignTo(E, true);
4530 case Builtin::BI__builtin_align_down:
4531 return EmitBuiltinAlignTo(E, false);
4532
4533 case Builtin::BI__noop:
4534 // __noop always evaluates to an integer literal zero.
4535 return RValue::get(ConstantInt::get(IntTy, 0));
4536 case Builtin::BI__builtin_call_with_static_chain: {
4537 const CallExpr *Call = cast<CallExpr>(E->getArg(0));
4538 const Expr *Chain = E->getArg(1);
4539 return EmitCall(Call->getCallee()->getType(),
4540 EmitCallee(Call->getCallee()), Call, ReturnValue,
4541 EmitScalarExpr(Chain));
4542 }
4543 case Builtin::BI_InterlockedExchange8:
4544 case Builtin::BI_InterlockedExchange16:
4545 case Builtin::BI_InterlockedExchange:
4546 case Builtin::BI_InterlockedExchangePointer:
4547 return RValue::get(
4548 EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchange, E));
4549 case Builtin::BI_InterlockedCompareExchangePointer:
4550 case Builtin::BI_InterlockedCompareExchangePointer_nf: {
4551 llvm::Type *RTy;
4552 llvm::IntegerType *IntType =
4553 IntegerType::get(getLLVMContext(),
4554 getContext().getTypeSize(E->getType()));
4555 llvm::Type *IntPtrType = IntType->getPointerTo();
4556
4557 llvm::Value *Destination =
4558 Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), IntPtrType);
4559
4560 llvm::Value *Exchange = EmitScalarExpr(E->getArg(1));
4561 RTy = Exchange->getType();
4562 Exchange = Builder.CreatePtrToInt(Exchange, IntType);
4563
4564 llvm::Value *Comparand =
4565 Builder.CreatePtrToInt(EmitScalarExpr(E->getArg(2)), IntType);
4566
4567 auto Ordering =
4568 BuiltinID == Builtin::BI_InterlockedCompareExchangePointer_nf ?
4569 AtomicOrdering::Monotonic : AtomicOrdering::SequentiallyConsistent;
4570
4571 auto Result = Builder.CreateAtomicCmpXchg(Destination, Comparand, Exchange,
4572 Ordering, Ordering);
4573 Result->setVolatile(true);
4574
4575 return RValue::get(Builder.CreateIntToPtr(Builder.CreateExtractValue(Result,
4576 0),
4577 RTy));
4578 }
4579 case Builtin::BI_InterlockedCompareExchange8:
4580 case Builtin::BI_InterlockedCompareExchange16:
4581 case Builtin::BI_InterlockedCompareExchange:
4582 case Builtin::BI_InterlockedCompareExchange64:
4583 return RValue::get(EmitAtomicCmpXchgForMSIntrin(*this, E));
4584 case Builtin::BI_InterlockedIncrement16:
4585 case Builtin::BI_InterlockedIncrement:
4586 return RValue::get(
4587 EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedIncrement, E));
4588 case Builtin::BI_InterlockedDecrement16:
4589 case Builtin::BI_InterlockedDecrement:
4590 return RValue::get(
4591 EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedDecrement, E));
4592 case Builtin::BI_InterlockedAnd8:
4593 case Builtin::BI_InterlockedAnd16:
4594 case Builtin::BI_InterlockedAnd:
4595 return RValue::get(EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedAnd, E));
4596 case Builtin::BI_InterlockedExchangeAdd8:
4597 case Builtin::BI_InterlockedExchangeAdd16:
4598 case Builtin::BI_InterlockedExchangeAdd:
4599 return RValue::get(
4600 EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeAdd, E));
4601 case Builtin::BI_InterlockedExchangeSub8:
4602 case Builtin::BI_InterlockedExchangeSub16:
4603 case Builtin::BI_InterlockedExchangeSub:
4604 return RValue::get(
4605 EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeSub, E));
4606 case Builtin::BI_InterlockedOr8:
4607 case Builtin::BI_InterlockedOr16:
4608 case Builtin::BI_InterlockedOr:
4609 return RValue::get(EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedOr, E));
4610 case Builtin::BI_InterlockedXor8:
4611 case Builtin::BI_InterlockedXor16:
4612 case Builtin::BI_InterlockedXor:
4613 return RValue::get(EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedXor, E));
4614
4615 case Builtin::BI_bittest64:
4616 case Builtin::BI_bittest:
4617 case Builtin::BI_bittestandcomplement64:
4618 case Builtin::BI_bittestandcomplement:
4619 case Builtin::BI_bittestandreset64:
4620 case Builtin::BI_bittestandreset:
4621 case Builtin::BI_bittestandset64:
4622 case Builtin::BI_bittestandset:
4623 case Builtin::BI_interlockedbittestandreset:
4624 case Builtin::BI_interlockedbittestandreset64:
4625 case Builtin::BI_interlockedbittestandset64:
4626 case Builtin::BI_interlockedbittestandset:
4627 case Builtin::BI_interlockedbittestandset_acq:
4628 case Builtin::BI_interlockedbittestandset_rel:
4629 case Builtin::BI_interlockedbittestandset_nf:
4630 case Builtin::BI_interlockedbittestandreset_acq:
4631 case Builtin::BI_interlockedbittestandreset_rel:
4632 case Builtin::BI_interlockedbittestandreset_nf:
4633 return RValue::get(EmitBitTestIntrinsic(*this, BuiltinID, E));
4634
4635 // These builtins exist to emit regular volatile loads and stores not
4636 // affected by the -fms-volatile setting.
4637 case Builtin::BI__iso_volatile_load8:
4638 case Builtin::BI__iso_volatile_load16:
4639 case Builtin::BI__iso_volatile_load32:
4640 case Builtin::BI__iso_volatile_load64:
4641 return RValue::get(EmitISOVolatileLoad(*this, E));
4642 case Builtin::BI__iso_volatile_store8:
4643 case Builtin::BI__iso_volatile_store16:
4644 case Builtin::BI__iso_volatile_store32:
4645 case Builtin::BI__iso_volatile_store64:
4646 return RValue::get(EmitISOVolatileStore(*this, E));
4647
4648 case Builtin::BI__exception_code:
4649 case Builtin::BI_exception_code:
4650 return RValue::get(EmitSEHExceptionCode());
4651 case Builtin::BI__exception_info:
4652 case Builtin::BI_exception_info:
4653 return RValue::get(EmitSEHExceptionInfo());
4654 case Builtin::BI__abnormal_termination:
4655 case Builtin::BI_abnormal_termination:
4656 return RValue::get(EmitSEHAbnormalTermination());
4657 case Builtin::BI_setjmpex:
4658 if (getTarget().getTriple().isOSMSVCRT() && E->getNumArgs() == 1 &&
4659 E->getArg(0)->getType()->isPointerType())
4660 return EmitMSVCRTSetJmp(*this, MSVCSetJmpKind::_setjmpex, E);
4661 break;
4662 case Builtin::BI_setjmp:
4663 if (getTarget().getTriple().isOSMSVCRT() && E->getNumArgs() == 1 &&
4664 E->getArg(0)->getType()->isPointerType()) {
4665 if (getTarget().getTriple().getArch() == llvm::Triple::x86)
4666 return EmitMSVCRTSetJmp(*this, MSVCSetJmpKind::_setjmp3, E);
4667 else if (getTarget().getTriple().getArch() == llvm::Triple::aarch64)
4668 return EmitMSVCRTSetJmp(*this, MSVCSetJmpKind::_setjmpex, E);
4669 return EmitMSVCRTSetJmp(*this, MSVCSetJmpKind::_setjmp, E);
4670 }
4671 break;
4672
4673 // C++ std:: builtins.
4674 case Builtin::BImove:
4675 case Builtin::BImove_if_noexcept:
4676 case Builtin::BIforward:
4677 case Builtin::BIas_const:
4678 return RValue::get(EmitLValue(E->getArg(0)).getPointer(*this));
4679 case Builtin::BI__GetExceptionInfo: {
4680 if (llvm::GlobalVariable *GV =
4681 CGM.getCXXABI().getThrowInfo(FD->getParamDecl(0)->getType()))
4682 return RValue::get(llvm::ConstantExpr::getBitCast(GV, CGM.Int8PtrTy));
4683 break;
4684 }
4685
4686 case Builtin::BI__fastfail:
4687 return RValue::get(EmitMSVCBuiltinExpr(MSVCIntrin::__fastfail, E));
4688
4689 case Builtin::BI__builtin_coro_id:
4690 return EmitCoroutineIntrinsic(E, Intrinsic::coro_id);
4691 case Builtin::BI__builtin_coro_promise:
4692 return EmitCoroutineIntrinsic(E, Intrinsic::coro_promise);
4693 case Builtin::BI__builtin_coro_resume:
4694 EmitCoroutineIntrinsic(E, Intrinsic::coro_resume);
4695 return RValue::get(nullptr);
4696 case Builtin::BI__builtin_coro_frame:
4697 return EmitCoroutineIntrinsic(E, Intrinsic::coro_frame);
4698 case Builtin::BI__builtin_coro_noop:
4699 return EmitCoroutineIntrinsic(E, Intrinsic::coro_noop);
4700 case Builtin::BI__builtin_coro_free:
4701 return EmitCoroutineIntrinsic(E, Intrinsic::coro_free);
4702 case Builtin::BI__builtin_coro_destroy:
4703 EmitCoroutineIntrinsic(E, Intrinsic::coro_destroy);
4704 return RValue::get(nullptr);
4705 case Builtin::BI__builtin_coro_done:
4706 return EmitCoroutineIntrinsic(E, Intrinsic::coro_done);
4707 case Builtin::BI__builtin_coro_alloc:
4708 return EmitCoroutineIntrinsic(E, Intrinsic::coro_alloc);
4709 case Builtin::BI__builtin_coro_begin:
4710 return EmitCoroutineIntrinsic(E, Intrinsic::coro_begin);
4711 case Builtin::BI__builtin_coro_end:
4712 return EmitCoroutineIntrinsic(E, Intrinsic::coro_end);
4713 case Builtin::BI__builtin_coro_suspend:
4714 return EmitCoroutineIntrinsic(E, Intrinsic::coro_suspend);
4715 case Builtin::BI__builtin_coro_size:
4716 return EmitCoroutineIntrinsic(E, Intrinsic::coro_size);
4717 case Builtin::BI__builtin_coro_align:
4718 return EmitCoroutineIntrinsic(E, Intrinsic::coro_align);
4719
4720 // OpenCL v2.0 s6.13.16.2, Built-in pipe read and write functions
4721 case Builtin::BIread_pipe:
4722 case Builtin::BIwrite_pipe: {
4723 Value *Arg0 = EmitScalarExpr(E->getArg(0)),
4724 *Arg1 = EmitScalarExpr(E->getArg(1));
4725 CGOpenCLRuntime OpenCLRT(CGM);
4726 Value *PacketSize = OpenCLRT.getPipeElemSize(E->getArg(0));
4727 Value *PacketAlign = OpenCLRT.getPipeElemAlign(E->getArg(0));
4728
4729 // Type of the generic packet parameter.
4730 unsigned GenericAS =
4731 getContext().getTargetAddressSpace(LangAS::opencl_generic);
4732 llvm::Type *I8PTy = llvm::PointerType::get(
4733 llvm::Type::getInt8Ty(getLLVMContext()), GenericAS);
4734
4735 // Testing which overloaded version we should generate the call for.
4736 if (2U == E->getNumArgs()) {
4737 const char *Name = (BuiltinID == Builtin::BIread_pipe) ? "__read_pipe_2"
4738 : "__write_pipe_2";
4739 // Creating a generic function type to be able to call with any builtin or
4740 // user defined type.
4741 llvm::Type *ArgTys[] = {Arg0->getType(), I8PTy, Int32Ty, Int32Ty};
4742 llvm::FunctionType *FTy = llvm::FunctionType::get(
4743 Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false);
4744 Value *BCast = Builder.CreatePointerCast(Arg1, I8PTy);
4745 return RValue::get(
4746 EmitRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name),
4747 {Arg0, BCast, PacketSize, PacketAlign}));
4748 } else {
4749 assert(4 == E->getNumArgs() &&
4750 "Illegal number of parameters to pipe function");
4751 const char *Name = (BuiltinID == Builtin::BIread_pipe) ? "__read_pipe_4"
4752 : "__write_pipe_4";
4753
4754 llvm::Type *ArgTys[] = {Arg0->getType(), Arg1->getType(), Int32Ty, I8PTy,
4755 Int32Ty, Int32Ty};
4756 Value *Arg2 = EmitScalarExpr(E->getArg(2)),
4757 *Arg3 = EmitScalarExpr(E->getArg(3));
4758 llvm::FunctionType *FTy = llvm::FunctionType::get(
4759 Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false);
4760 Value *BCast = Builder.CreatePointerCast(Arg3, I8PTy);
4761 // We know the third argument is an integer type, but we may need to cast
4762 // it to i32.
4763 if (Arg2->getType() != Int32Ty)
4764 Arg2 = Builder.CreateZExtOrTrunc(Arg2, Int32Ty);
4765 return RValue::get(
4766 EmitRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name),
4767 {Arg0, Arg1, Arg2, BCast, PacketSize, PacketAlign}));
4768 }
4769 }
4770 // OpenCL v2.0 s6.13.16 ,s9.17.3.5 - Built-in pipe reserve read and write
4771 // functions
4772 case Builtin::BIreserve_read_pipe:
4773 case Builtin::BIreserve_write_pipe:
4774 case Builtin::BIwork_group_reserve_read_pipe:
4775 case Builtin::BIwork_group_reserve_write_pipe:
4776 case Builtin::BIsub_group_reserve_read_pipe:
4777 case Builtin::BIsub_group_reserve_write_pipe: {
4778 // Composing the mangled name for the function.
4779 const char *Name;
4780 if (BuiltinID == Builtin::BIreserve_read_pipe)
4781 Name = "__reserve_read_pipe";
4782 else if (BuiltinID == Builtin::BIreserve_write_pipe)
4783 Name = "__reserve_write_pipe";
4784 else if (BuiltinID == Builtin::BIwork_group_reserve_read_pipe)
4785 Name = "__work_group_reserve_read_pipe";
4786 else if (BuiltinID == Builtin::BIwork_group_reserve_write_pipe)
4787 Name = "__work_group_reserve_write_pipe";
4788 else if (BuiltinID == Builtin::BIsub_group_reserve_read_pipe)
4789 Name = "__sub_group_reserve_read_pipe";
4790 else
4791 Name = "__sub_group_reserve_write_pipe";
4792
4793 Value *Arg0 = EmitScalarExpr(E->getArg(0)),
4794 *Arg1 = EmitScalarExpr(E->getArg(1));
4795 llvm::Type *ReservedIDTy = ConvertType(getContext().OCLReserveIDTy);
4796 CGOpenCLRuntime OpenCLRT(CGM);
4797 Value *PacketSize = OpenCLRT.getPipeElemSize(E->getArg(0));
4798 Value *PacketAlign = OpenCLRT.getPipeElemAlign(E->getArg(0));
4799
4800 // Building the generic function prototype.
4801 llvm::Type *ArgTys[] = {Arg0->getType(), Int32Ty, Int32Ty, Int32Ty};
4802 llvm::FunctionType *FTy = llvm::FunctionType::get(
4803 ReservedIDTy, llvm::ArrayRef<llvm::Type *>(ArgTys), false);
4804 // We know the second argument is an integer type, but we may need to cast
4805 // it to i32.
4806 if (Arg1->getType() != Int32Ty)
4807 Arg1 = Builder.CreateZExtOrTrunc(Arg1, Int32Ty);
4808 return RValue::get(EmitRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name),
4809 {Arg0, Arg1, PacketSize, PacketAlign}));
4810 }
4811 // OpenCL v2.0 s6.13.16, s9.17.3.5 - Built-in pipe commit read and write
4812 // functions
4813 case Builtin::BIcommit_read_pipe:
4814 case Builtin::BIcommit_write_pipe:
4815 case Builtin::BIwork_group_commit_read_pipe:
4816 case Builtin::BIwork_group_commit_write_pipe:
4817 case Builtin::BIsub_group_commit_read_pipe:
4818 case Builtin::BIsub_group_commit_write_pipe: {
4819 const char *Name;
4820 if (BuiltinID == Builtin::BIcommit_read_pipe)
4821 Name = "__commit_read_pipe";
4822 else if (BuiltinID == Builtin::BIcommit_write_pipe)
4823 Name = "__commit_write_pipe";
4824 else if (BuiltinID == Builtin::BIwork_group_commit_read_pipe)
4825 Name = "__work_group_commit_read_pipe";
4826 else if (BuiltinID == Builtin::BIwork_group_commit_write_pipe)
4827 Name = "__work_group_commit_write_pipe";
4828 else if (BuiltinID == Builtin::BIsub_group_commit_read_pipe)
4829 Name = "__sub_group_commit_read_pipe";
4830 else
4831 Name = "__sub_group_commit_write_pipe";
4832
4833 Value *Arg0 = EmitScalarExpr(E->getArg(0)),
4834 *Arg1 = EmitScalarExpr(E->getArg(1));
4835 CGOpenCLRuntime OpenCLRT(CGM);
4836 Value *PacketSize = OpenCLRT.getPipeElemSize(E->getArg(0));
4837 Value *PacketAlign = OpenCLRT.getPipeElemAlign(E->getArg(0));
4838
4839 // Building the generic function prototype.
4840 llvm::Type *ArgTys[] = {Arg0->getType(), Arg1->getType(), Int32Ty, Int32Ty};
4841 llvm::FunctionType *FTy =
4842 llvm::FunctionType::get(llvm::Type::getVoidTy(getLLVMContext()),
4843 llvm::ArrayRef<llvm::Type *>(ArgTys), false);
4844
4845 return RValue::get(EmitRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name),
4846 {Arg0, Arg1, PacketSize, PacketAlign}));
4847 }
4848 // OpenCL v2.0 s6.13.16.4 Built-in pipe query functions
4849 case Builtin::BIget_pipe_num_packets:
4850 case Builtin::BIget_pipe_max_packets: {
4851 const char *BaseName;
4852 const auto *PipeTy = E->getArg(0)->getType()->castAs<PipeType>();
4853 if (BuiltinID == Builtin::BIget_pipe_num_packets)
4854 BaseName = "__get_pipe_num_packets";
4855 else
4856 BaseName = "__get_pipe_max_packets";
4857 std::string Name = std::string(BaseName) +
4858 std::string(PipeTy->isReadOnly() ? "_ro" : "_wo");
4859
4860 // Building the generic function prototype.
4861 Value *Arg0 = EmitScalarExpr(E->getArg(0));
4862 CGOpenCLRuntime OpenCLRT(CGM);
4863 Value *PacketSize = OpenCLRT.getPipeElemSize(E->getArg(0));
4864 Value *PacketAlign = OpenCLRT.getPipeElemAlign(E->getArg(0));
4865 llvm::Type *ArgTys[] = {Arg0->getType(), Int32Ty, Int32Ty};
4866 llvm::FunctionType *FTy = llvm::FunctionType::get(
4867 Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false);
4868
4869 return RValue::get(EmitRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name),
4870 {Arg0, PacketSize, PacketAlign}));
4871 }
4872
4873 // OpenCL v2.0 s6.13.9 - Address space qualifier functions.
4874 case Builtin::BIto_global:
4875 case Builtin::BIto_local:
4876 case Builtin::BIto_private: {
4877 auto Arg0 = EmitScalarExpr(E->getArg(0));
4878 auto NewArgT = llvm::PointerType::get(Int8Ty,
4879 CGM.getContext().getTargetAddressSpace(LangAS::opencl_generic));
4880 auto NewRetT = llvm::PointerType::get(Int8Ty,
4881 CGM.getContext().getTargetAddressSpace(
4882 E->getType()->getPointeeType().getAddressSpace()));
4883 auto FTy = llvm::FunctionType::get(NewRetT, {NewArgT}, false);
4884 llvm::Value *NewArg;
4885 if (Arg0->getType()->getPointerAddressSpace() !=
4886 NewArgT->getPointerAddressSpace())
4887 NewArg = Builder.CreateAddrSpaceCast(Arg0, NewArgT);
4888 else
4889 NewArg = Builder.CreateBitOrPointerCast(Arg0, NewArgT);
4890 auto NewName = std::string("__") + E->getDirectCallee()->getName().str();
4891 auto NewCall =
4892 EmitRuntimeCall(CGM.CreateRuntimeFunction(FTy, NewName), {NewArg});
4893 return RValue::get(Builder.CreateBitOrPointerCast(NewCall,
4894 ConvertType(E->getType())));
4895 }
4896
4897 // OpenCL v2.0, s6.13.17 - Enqueue kernel function.
4898 // It contains four different overload formats specified in Table 6.13.17.1.
4899 case Builtin::BIenqueue_kernel: {
4900 StringRef Name; // Generated function call name
4901 unsigned NumArgs = E->getNumArgs();
4902
4903 llvm::Type *QueueTy = ConvertType(getContext().OCLQueueTy);
4904 llvm::Type *GenericVoidPtrTy = Builder.getInt8PtrTy(
4905 getContext().getTargetAddressSpace(LangAS::opencl_generic));
4906
4907 llvm::Value *Queue = EmitScalarExpr(E->getArg(0));
4908 llvm::Value *Flags = EmitScalarExpr(E->getArg(1));
4909 LValue NDRangeL = EmitAggExprToLValue(E->getArg(2));
4910 llvm::Value *Range = NDRangeL.getAddress(*this).getPointer();
4911 llvm::Type *RangeTy = NDRangeL.getAddress(*this).getType();
4912
4913 if (NumArgs == 4) {
4914 // The most basic form of the call with parameters:
4915 // queue_t, kernel_enqueue_flags_t, ndrange_t, block(void)
4916 Name = "__enqueue_kernel_basic";
4917 llvm::Type *ArgTys[] = {QueueTy, Int32Ty, RangeTy, GenericVoidPtrTy,
4918 GenericVoidPtrTy};
4919 llvm::FunctionType *FTy = llvm::FunctionType::get(
4920 Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false);
4921
4922 auto Info =
4923 CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(3));
4924 llvm::Value *Kernel =
4925 Builder.CreatePointerCast(Info.Kernel, GenericVoidPtrTy);
4926 llvm::Value *Block =
4927 Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy);
4928
4929 AttrBuilder B(Builder.getContext());
4930 B.addByValAttr(NDRangeL.getAddress(*this).getElementType());
4931 llvm::AttributeList ByValAttrSet =
4932 llvm::AttributeList::get(CGM.getModule().getContext(), 3U, B);
4933
4934 auto RTCall =
4935 EmitRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name, ByValAttrSet),
4936 {Queue, Flags, Range, Kernel, Block});
4937 RTCall->setAttributes(ByValAttrSet);
4938 return RValue::get(RTCall);
4939 }
4940 assert(NumArgs >= 5 && "Invalid enqueue_kernel signature");
4941
4942 // Create a temporary array to hold the sizes of local pointer arguments
4943 // for the block. \p First is the position of the first size argument.
4944 auto CreateArrayForSizeVar = [=](unsigned First)
4945 -> std::tuple<llvm::Value *, llvm::Value *, llvm::Value *> {
4946 llvm::APInt ArraySize(32, NumArgs - First);
4947 QualType SizeArrayTy = getContext().getConstantArrayType(
4948 getContext().getSizeType(), ArraySize, nullptr, ArrayType::Normal,
4949 /*IndexTypeQuals=*/0);
4950 auto Tmp = CreateMemTemp(SizeArrayTy, "block_sizes");
4951 llvm::Value *TmpPtr = Tmp.getPointer();
4952 llvm::Value *TmpSize = EmitLifetimeStart(
4953 CGM.getDataLayout().getTypeAllocSize(Tmp.getElementType()), TmpPtr);
4954 llvm::Value *ElemPtr;
4955 // Each of the following arguments specifies the size of the corresponding
4956 // argument passed to the enqueued block.
4957 auto *Zero = llvm::ConstantInt::get(IntTy, 0);
4958 for (unsigned I = First; I < NumArgs; ++I) {
4959 auto *Index = llvm::ConstantInt::get(IntTy, I - First);
4960 auto *GEP = Builder.CreateGEP(Tmp.getElementType(), TmpPtr,
4961 {Zero, Index});
4962 if (I == First)
4963 ElemPtr = GEP;
4964 auto *V =
4965 Builder.CreateZExtOrTrunc(EmitScalarExpr(E->getArg(I)), SizeTy);
4966 Builder.CreateAlignedStore(
4967 V, GEP, CGM.getDataLayout().getPrefTypeAlign(SizeTy));
4968 }
4969 return std::tie(ElemPtr, TmpSize, TmpPtr);
4970 };
4971
4972 // Could have events and/or varargs.
4973 if (E->getArg(3)->getType()->isBlockPointerType()) {
4974 // No events passed, but has variadic arguments.
4975 Name = "__enqueue_kernel_varargs";
4976 auto Info =
4977 CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(3));
4978 llvm::Value *Kernel =
4979 Builder.CreatePointerCast(Info.Kernel, GenericVoidPtrTy);
4980 auto *Block = Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy);
4981 llvm::Value *ElemPtr, *TmpSize, *TmpPtr;
4982 std::tie(ElemPtr, TmpSize, TmpPtr) = CreateArrayForSizeVar(4);
4983
4984 // Create a vector of the arguments, as well as a constant value to
4985 // express to the runtime the number of variadic arguments.
4986 llvm::Value *const Args[] = {Queue, Flags,
4987 Range, Kernel,
4988 Block, ConstantInt::get(IntTy, NumArgs - 4),
4989 ElemPtr};
4990 llvm::Type *const ArgTys[] = {
4991 QueueTy, IntTy, RangeTy, GenericVoidPtrTy,
4992 GenericVoidPtrTy, IntTy, ElemPtr->getType()};
4993
4994 llvm::FunctionType *FTy = llvm::FunctionType::get(Int32Ty, ArgTys, false);
4995 auto Call = RValue::get(
4996 EmitRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Args));
4997 if (TmpSize)
4998 EmitLifetimeEnd(TmpSize, TmpPtr);
4999 return Call;
5000 }
5001 // Any calls now have event arguments passed.
5002 if (NumArgs >= 7) {
5003 llvm::Type *EventTy = ConvertType(getContext().OCLClkEventTy);
5004 llvm::PointerType *EventPtrTy = EventTy->getPointerTo(
5005 CGM.getContext().getTargetAddressSpace(LangAS::opencl_generic));
5006
5007 llvm::Value *NumEvents =
5008 Builder.CreateZExtOrTrunc(EmitScalarExpr(E->getArg(3)), Int32Ty);
5009
5010 // Since SemaOpenCLBuiltinEnqueueKernel allows fifth and sixth arguments
5011 // to be a null pointer constant (including `0` literal), we can take it
5012 // into account and emit null pointer directly.
5013 llvm::Value *EventWaitList = nullptr;
5014 if (E->getArg(4)->isNullPointerConstant(
5015 getContext(), Expr::NPC_ValueDependentIsNotNull)) {
5016 EventWaitList = llvm::ConstantPointerNull::get(EventPtrTy);
5017 } else {
5018 EventWaitList = E->getArg(4)->getType()->isArrayType()
5019 ? EmitArrayToPointerDecay(E->getArg(4)).getPointer()
5020 : EmitScalarExpr(E->getArg(4));
5021 // Convert to generic address space.
5022 EventWaitList = Builder.CreatePointerCast(EventWaitList, EventPtrTy);
5023 }
5024 llvm::Value *EventRet = nullptr;
5025 if (E->getArg(5)->isNullPointerConstant(
5026 getContext(), Expr::NPC_ValueDependentIsNotNull)) {
5027 EventRet = llvm::ConstantPointerNull::get(EventPtrTy);
5028 } else {
5029 EventRet =
5030 Builder.CreatePointerCast(EmitScalarExpr(E->getArg(5)), EventPtrTy);
5031 }
5032
5033 auto Info =
5034 CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(6));
5035 llvm::Value *Kernel =
5036 Builder.CreatePointerCast(Info.Kernel, GenericVoidPtrTy);
5037 llvm::Value *Block =
5038 Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy);
5039
5040 std::vector<llvm::Type *> ArgTys = {
5041 QueueTy, Int32Ty, RangeTy, Int32Ty,
5042 EventPtrTy, EventPtrTy, GenericVoidPtrTy, GenericVoidPtrTy};
5043
5044 std::vector<llvm::Value *> Args = {Queue, Flags, Range,
5045 NumEvents, EventWaitList, EventRet,
5046 Kernel, Block};
5047
5048 if (NumArgs == 7) {
5049 // Has events but no variadics.
5050 Name = "__enqueue_kernel_basic_events";
5051 llvm::FunctionType *FTy = llvm::FunctionType::get(
5052 Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false);
5053 return RValue::get(
5054 EmitRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name),
5055 llvm::ArrayRef<llvm::Value *>(Args)));
5056 }
5057 // Has event info and variadics
5058 // Pass the number of variadics to the runtime function too.
5059 Args.push_back(ConstantInt::get(Int32Ty, NumArgs - 7));
5060 ArgTys.push_back(Int32Ty);
5061 Name = "__enqueue_kernel_events_varargs";
5062
5063 llvm::Value *ElemPtr, *TmpSize, *TmpPtr;
5064 std::tie(ElemPtr, TmpSize, TmpPtr) = CreateArrayForSizeVar(7);
5065 Args.push_back(ElemPtr);
5066 ArgTys.push_back(ElemPtr->getType());
5067
5068 llvm::FunctionType *FTy = llvm::FunctionType::get(
5069 Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false);
5070 auto Call =
5071 RValue::get(EmitRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name),
5072 llvm::ArrayRef<llvm::Value *>(Args)));
5073 if (TmpSize)
5074 EmitLifetimeEnd(TmpSize, TmpPtr);
5075 return Call;
5076 }
5077 [[fallthrough]];
5078 }
5079 // OpenCL v2.0 s6.13.17.6 - Kernel query functions need bitcast of block
5080 // parameter.
5081 case Builtin::BIget_kernel_work_group_size: {
5082 llvm::Type *GenericVoidPtrTy = Builder.getInt8PtrTy(
5083 getContext().getTargetAddressSpace(LangAS::opencl_generic));
5084 auto Info =
5085 CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(0));
5086 Value *Kernel = Builder.CreatePointerCast(Info.Kernel, GenericVoidPtrTy);
5087 Value *Arg = Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy);
5088 return RValue::get(EmitRuntimeCall(
5089 CGM.CreateRuntimeFunction(
5090 llvm::FunctionType::get(IntTy, {GenericVoidPtrTy, GenericVoidPtrTy},
5091 false),
5092 "__get_kernel_work_group_size_impl"),
5093 {Kernel, Arg}));
5094 }
5095 case Builtin::BIget_kernel_preferred_work_group_size_multiple: {
5096 llvm::Type *GenericVoidPtrTy = Builder.getInt8PtrTy(
5097 getContext().getTargetAddressSpace(LangAS::opencl_generic));
5098 auto Info =
5099 CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(0));
5100 Value *Kernel = Builder.CreatePointerCast(Info.Kernel, GenericVoidPtrTy);
5101 Value *Arg = Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy);
5102 return RValue::get(EmitRuntimeCall(
5103 CGM.CreateRuntimeFunction(
5104 llvm::FunctionType::get(IntTy, {GenericVoidPtrTy, GenericVoidPtrTy},
5105 false),
5106 "__get_kernel_preferred_work_group_size_multiple_impl"),
5107 {Kernel, Arg}));
5108 }
5109 case Builtin::BIget_kernel_max_sub_group_size_for_ndrange:
5110 case Builtin::BIget_kernel_sub_group_count_for_ndrange: {
5111 llvm::Type *GenericVoidPtrTy = Builder.getInt8PtrTy(
5112 getContext().getTargetAddressSpace(LangAS::opencl_generic));
5113 LValue NDRangeL = EmitAggExprToLValue(E->getArg(0));
5114 llvm::Value *NDRange = NDRangeL.getAddress(*this).getPointer();
5115 auto Info =
5116 CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(1));
5117 Value *Kernel = Builder.CreatePointerCast(Info.Kernel, GenericVoidPtrTy);
5118 Value *Block = Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy);
5119 const char *Name =
5120 BuiltinID == Builtin::BIget_kernel_max_sub_group_size_for_ndrange
5121 ? "__get_kernel_max_sub_group_size_for_ndrange_impl"
5122 : "__get_kernel_sub_group_count_for_ndrange_impl";
5123 return RValue::get(EmitRuntimeCall(
5124 CGM.CreateRuntimeFunction(
5125 llvm::FunctionType::get(
5126 IntTy, {NDRange->getType(), GenericVoidPtrTy, GenericVoidPtrTy},
5127 false),
5128 Name),
5129 {NDRange, Kernel, Block}));
5130 }
5131
5132 case Builtin::BI__builtin_store_half:
5133 case Builtin::BI__builtin_store_halff: {
5134 Value *Val = EmitScalarExpr(E->getArg(0));
5135 Address Address = EmitPointerWithAlignment(E->getArg(1));
5136 Value *HalfVal = Builder.CreateFPTrunc(Val, Builder.getHalfTy());
5137 Builder.CreateStore(HalfVal, Address);
5138 return RValue::get(nullptr);
5139 }
5140 case Builtin::BI__builtin_load_half: {
5141 Address Address = EmitPointerWithAlignment(E->getArg(0));
5142 Value *HalfVal = Builder.CreateLoad(Address);
5143 return RValue::get(Builder.CreateFPExt(HalfVal, Builder.getDoubleTy()));
5144 }
5145 case Builtin::BI__builtin_load_halff: {
5146 Address Address = EmitPointerWithAlignment(E->getArg(0));
5147 Value *HalfVal = Builder.CreateLoad(Address);
5148 return RValue::get(Builder.CreateFPExt(HalfVal, Builder.getFloatTy()));
5149 }
5150 case Builtin::BIprintf:
5151 if (getTarget().getTriple().isNVPTX() ||
5152 getTarget().getTriple().isAMDGCN()) {
5153 if (getLangOpts().OpenMPIsDevice)
5154 return EmitOpenMPDevicePrintfCallExpr(E);
5155 if (getTarget().getTriple().isNVPTX())
5156 return EmitNVPTXDevicePrintfCallExpr(E);
5157 if (getTarget().getTriple().isAMDGCN() && getLangOpts().HIP)
5158 return EmitAMDGPUDevicePrintfCallExpr(E);
5159 }
5160
5161 break;
5162 case Builtin::BI__builtin_canonicalize:
5163 case Builtin::BI__builtin_canonicalizef:
5164 case Builtin::BI__builtin_canonicalizef16:
5165 case Builtin::BI__builtin_canonicalizel:
5166 return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::canonicalize));
5167
5168 case Builtin::BI__builtin_thread_pointer: {
5169 if (!getContext().getTargetInfo().isTLSSupported())
5170 CGM.ErrorUnsupported(E, "__builtin_thread_pointer");
5171 // Fall through - it's already mapped to the intrinsic by ClangBuiltin.
5172 break;
5173 }
5174 case Builtin::BI__builtin_os_log_format:
5175 return emitBuiltinOSLogFormat(*E);
5176
5177 case Builtin::BI__xray_customevent: {
5178 if (!ShouldXRayInstrumentFunction())
5179 return RValue::getIgnored();
5180
5181 if (!CGM.getCodeGenOpts().XRayInstrumentationBundle.has(
5182 XRayInstrKind::Custom))
5183 return RValue::getIgnored();
5184
5185 if (const auto *XRayAttr = CurFuncDecl->getAttr<XRayInstrumentAttr>())
5186 if (XRayAttr->neverXRayInstrument() && !AlwaysEmitXRayCustomEvents())
5187 return RValue::getIgnored();
5188
5189 Function *F = CGM.getIntrinsic(Intrinsic::xray_customevent);
5190 auto FTy = F->getFunctionType();
5191 auto Arg0 = E->getArg(0);
5192 auto Arg0Val = EmitScalarExpr(Arg0);
5193 auto Arg0Ty = Arg0->getType();
5194 auto PTy0 = FTy->getParamType(0);
5195 if (PTy0 != Arg0Val->getType()) {
5196 if (Arg0Ty->isArrayType())
5197 Arg0Val = EmitArrayToPointerDecay(Arg0).getPointer();
5198 else
5199 Arg0Val = Builder.CreatePointerCast(Arg0Val, PTy0);
5200 }
5201 auto Arg1 = EmitScalarExpr(E->getArg(1));
5202 auto PTy1 = FTy->getParamType(1);
5203 if (PTy1 != Arg1->getType())
5204 Arg1 = Builder.CreateTruncOrBitCast(Arg1, PTy1);
5205 return RValue::get(Builder.CreateCall(F, {Arg0Val, Arg1}));
5206 }
5207
5208 case Builtin::BI__xray_typedevent: {
5209 // TODO: There should be a way to always emit events even if the current
5210 // function is not instrumented. Losing events in a stream can cripple
5211 // a trace.
5212 if (!ShouldXRayInstrumentFunction())
5213 return RValue::getIgnored();
5214
5215 if (!CGM.getCodeGenOpts().XRayInstrumentationBundle.has(
5216 XRayInstrKind::Typed))
5217 return RValue::getIgnored();
5218
5219 if (const auto *XRayAttr = CurFuncDecl->getAttr<XRayInstrumentAttr>())
5220 if (XRayAttr->neverXRayInstrument() && !AlwaysEmitXRayTypedEvents())
5221 return RValue::getIgnored();
5222
5223 Function *F = CGM.getIntrinsic(Intrinsic::xray_typedevent);
5224 auto FTy = F->getFunctionType();
5225 auto Arg0 = EmitScalarExpr(E->getArg(0));
5226 auto PTy0 = FTy->getParamType(0);
5227 if (PTy0 != Arg0->getType())
5228 Arg0 = Builder.CreateTruncOrBitCast(Arg0, PTy0);
5229 auto Arg1 = E->getArg(1);
5230 auto Arg1Val = EmitScalarExpr(Arg1);
5231 auto Arg1Ty = Arg1->getType();
5232 auto PTy1 = FTy->getParamType(1);
5233 if (PTy1 != Arg1Val->getType()) {
5234 if (Arg1Ty->isArrayType())
5235 Arg1Val = EmitArrayToPointerDecay(Arg1).getPointer();
5236 else
5237 Arg1Val = Builder.CreatePointerCast(Arg1Val, PTy1);
5238 }
5239 auto Arg2 = EmitScalarExpr(E->getArg(2));
5240 auto PTy2 = FTy->getParamType(2);
5241 if (PTy2 != Arg2->getType())
5242 Arg2 = Builder.CreateTruncOrBitCast(Arg2, PTy2);
5243 return RValue::get(Builder.CreateCall(F, {Arg0, Arg1Val, Arg2}));
5244 }
5245
5246 case Builtin::BI__builtin_ms_va_start:
5247 case Builtin::BI__builtin_ms_va_end:
5248 return RValue::get(
5249 EmitVAStartEnd(EmitMSVAListRef(E->getArg(0)).getPointer(),
5250 BuiltinID == Builtin::BI__builtin_ms_va_start));
5251
5252 case Builtin::BI__builtin_ms_va_copy: {
5253 // Lower this manually. We can't reliably determine whether or not any
5254 // given va_copy() is for a Win64 va_list from the calling convention
5255 // alone, because it's legal to do this from a System V ABI function.
5256 // With opaque pointer types, we won't have enough information in LLVM
5257 // IR to determine this from the argument types, either. Best to do it
5258 // now, while we have enough information.
5259 Address DestAddr = EmitMSVAListRef(E->getArg(0));
5260 Address SrcAddr = EmitMSVAListRef(E->getArg(1));
5261
5262 llvm::Type *BPP = Int8PtrPtrTy;
5263
5264 DestAddr = Address(Builder.CreateBitCast(DestAddr.getPointer(), BPP, "cp"),
5265 Int8PtrTy, DestAddr.getAlignment());
5266 SrcAddr = Address(Builder.CreateBitCast(SrcAddr.getPointer(), BPP, "ap"),
5267 Int8PtrTy, SrcAddr.getAlignment());
5268
5269 Value *ArgPtr = Builder.CreateLoad(SrcAddr, "ap.val");
5270 return RValue::get(Builder.CreateStore(ArgPtr, DestAddr));
5271 }
5272
5273 case Builtin::BI__builtin_get_device_side_mangled_name: {
5274 auto Name = CGM.getCUDARuntime().getDeviceSideName(
5275 cast<DeclRefExpr>(E->getArg(0)->IgnoreImpCasts())->getDecl());
5276 auto Str = CGM.GetAddrOfConstantCString(Name, "");
5277 llvm::Constant *Zeros[] = {llvm::ConstantInt::get(SizeTy, 0),
5278 llvm::ConstantInt::get(SizeTy, 0)};
5279 auto *Ptr = llvm::ConstantExpr::getGetElementPtr(Str.getElementType(),
5280 Str.getPointer(), Zeros);
5281 return RValue::get(Ptr);
5282 }
5283 }
5284
5285 // If this is an alias for a lib function (e.g. __builtin_sin), emit
5286 // the call using the normal call path, but using the unmangled
5287 // version of the function name.
5288 if (getContext().BuiltinInfo.isLibFunction(BuiltinID))
5289 return emitLibraryCall(*this, FD, E,
5290 CGM.getBuiltinLibFunction(FD, BuiltinID));
5291
5292 // If this is a predefined lib function (e.g. malloc), emit the call
5293 // using exactly the normal call path.
5294 if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID))
5295 return emitLibraryCall(*this, FD, E,
5296 cast<llvm::Constant>(EmitScalarExpr(E->getCallee())));
5297
5298 // Check that a call to a target specific builtin has the correct target
5299 // features.
5300 // This is down here to avoid non-target specific builtins, however, if
5301 // generic builtins start to require generic target features then we
5302 // can move this up to the beginning of the function.
5303 checkTargetFeatures(E, FD);
5304
5305 if (unsigned VectorWidth = getContext().BuiltinInfo.getRequiredVectorWidth(BuiltinID))
5306 LargestVectorWidth = std::max(LargestVectorWidth, VectorWidth);
5307
5308 // See if we have a target specific intrinsic.
5309 StringRef Name = getContext().BuiltinInfo.getName(BuiltinID);
5310 Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic;
5311 StringRef Prefix =
5312 llvm::Triple::getArchTypePrefix(getTarget().getTriple().getArch());
5313 if (!Prefix.empty()) {
5314 IntrinsicID = Intrinsic::getIntrinsicForClangBuiltin(Prefix.data(), Name);
5315 // NOTE we don't need to perform a compatibility flag check here since the
5316 // intrinsics are declared in Builtins*.def via LANGBUILTIN which filter the
5317 // MS builtins via ALL_MS_LANGUAGES and are filtered earlier.
5318 if (IntrinsicID == Intrinsic::not_intrinsic)
5319 IntrinsicID = Intrinsic::getIntrinsicForMSBuiltin(Prefix.data(), Name);
5320 }
5321
5322 if (IntrinsicID != Intrinsic::not_intrinsic) {
5323 SmallVector<Value*, 16> Args;
5324
5325 // Find out if any arguments are required to be integer constant
5326 // expressions.
5327 unsigned ICEArguments = 0;
5328 ASTContext::GetBuiltinTypeError Error;
5329 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments);
5330 assert(Error == ASTContext::GE_None && "Should not codegen an error");
5331
5332 Function *F = CGM.getIntrinsic(IntrinsicID);
5333 llvm::FunctionType *FTy = F->getFunctionType();
5334
5335 for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) {
5336 Value *ArgValue;
5337 // If this is a normal argument, just emit it as a scalar.
5338 if ((ICEArguments & (1 << i)) == 0) {
5339 ArgValue = EmitScalarExpr(E->getArg(i));
5340 } else {
5341 // If this is required to be a constant, constant fold it so that we
5342 // know that the generated intrinsic gets a ConstantInt.
5343 ArgValue = llvm::ConstantInt::get(
5344 getLLVMContext(),
5345 *E->getArg(i)->getIntegerConstantExpr(getContext()));
5346 }
5347
5348 // If the intrinsic arg type is different from the builtin arg type
5349 // we need to do a bit cast.
5350 llvm::Type *PTy = FTy->getParamType(i);
5351 if (PTy != ArgValue->getType()) {
5352 // XXX - vector of pointers?
5353 if (auto *PtrTy = dyn_cast<llvm::PointerType>(PTy)) {
5354 if (PtrTy->getAddressSpace() !=
5355 ArgValue->getType()->getPointerAddressSpace()) {
5356 ArgValue = Builder.CreateAddrSpaceCast(
5357 ArgValue,
5358 ArgValue->getType()->getPointerTo(PtrTy->getAddressSpace()));
5359 }
5360 }
5361
5362 assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) &&
5363 "Must be able to losslessly bit cast to param");
5364 // Cast vector type (e.g., v256i32) to x86_amx, this only happen
5365 // in amx intrinsics.
5366 if (PTy->isX86_AMXTy())
5367 ArgValue = Builder.CreateIntrinsic(Intrinsic::x86_cast_vector_to_tile,
5368 {ArgValue->getType()}, {ArgValue});
5369 else
5370 ArgValue = Builder.CreateBitCast(ArgValue, PTy);
5371 }
5372
5373 Args.push_back(ArgValue);
5374 }
5375
5376 Value *V = Builder.CreateCall(F, Args);
5377 QualType BuiltinRetType = E->getType();
5378
5379 llvm::Type *RetTy = VoidTy;
5380 if (!BuiltinRetType->isVoidType())
5381 RetTy = ConvertType(BuiltinRetType);
5382
5383 if (RetTy != V->getType()) {
5384 // XXX - vector of pointers?
5385 if (auto *PtrTy = dyn_cast<llvm::PointerType>(RetTy)) {
5386 if (PtrTy->getAddressSpace() != V->getType()->getPointerAddressSpace()) {
5387 V = Builder.CreateAddrSpaceCast(
5388 V, V->getType()->getPointerTo(PtrTy->getAddressSpace()));
5389 }
5390 }
5391
5392 assert(V->getType()->canLosslesslyBitCastTo(RetTy) &&
5393 "Must be able to losslessly bit cast result type");
5394 // Cast x86_amx to vector type (e.g., v256i32), this only happen
5395 // in amx intrinsics.
5396 if (V->getType()->isX86_AMXTy())
5397 V = Builder.CreateIntrinsic(Intrinsic::x86_cast_tile_to_vector, {RetTy},
5398 {V});
5399 else
5400 V = Builder.CreateBitCast(V, RetTy);
5401 }
5402
5403 if (RetTy->isVoidTy())
5404 return RValue::get(nullptr);
5405
5406 return RValue::get(V);
5407 }
5408
5409 // Some target-specific builtins can have aggregate return values, e.g.
5410 // __builtin_arm_mve_vld2q_u32. So if the result is an aggregate, force
5411 // ReturnValue to be non-null, so that the target-specific emission code can
5412 // always just emit into it.
5413 TypeEvaluationKind EvalKind = getEvaluationKind(E->getType());
5414 if (EvalKind == TEK_Aggregate && ReturnValue.isNull()) {
5415 Address DestPtr = CreateMemTemp(E->getType(), "agg.tmp");
5416 ReturnValue = ReturnValueSlot(DestPtr, false);
5417 }
5418
5419 // Now see if we can emit a target-specific builtin.
5420 if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E, ReturnValue)) {
5421 switch (EvalKind) {
5422 case TEK_Scalar:
5423 if (V->getType()->isVoidTy())
5424 return RValue::get(nullptr);
5425 return RValue::get(V);
5426 case TEK_Aggregate:
5427 return RValue::getAggregate(ReturnValue.getValue(),
5428 ReturnValue.isVolatile());
5429 case TEK_Complex:
5430 llvm_unreachable("No current target builtin returns complex");
5431 }
5432 llvm_unreachable("Bad evaluation kind in EmitBuiltinExpr");
5433 }
5434
5435 ErrorUnsupported(E, "builtin function");
5436
5437 // Unknown builtin, for now just dump it out and return undef.
5438 return GetUndefRValue(E->getType());
5439 }
5440
5441 static Value *EmitTargetArchBuiltinExpr(CodeGenFunction *CGF,
5442 unsigned BuiltinID, const CallExpr *E,
5443 ReturnValueSlot ReturnValue,
5444 llvm::Triple::ArchType Arch) {
5445 switch (Arch) {
5446 case llvm::Triple::arm:
5447 case llvm::Triple::armeb:
5448 case llvm::Triple::thumb:
5449 case llvm::Triple::thumbeb:
5450 return CGF->EmitARMBuiltinExpr(BuiltinID, E, ReturnValue, Arch);
5451 case llvm::Triple::aarch64:
5452 case llvm::Triple::aarch64_32:
5453 case llvm::Triple::aarch64_be:
5454 return CGF->EmitAArch64BuiltinExpr(BuiltinID, E, Arch);
5455 case llvm::Triple::bpfeb:
5456 case llvm::Triple::bpfel:
5457 return CGF->EmitBPFBuiltinExpr(BuiltinID, E);
5458 case llvm::Triple::x86:
5459 case llvm::Triple::x86_64:
5460 return CGF->EmitX86BuiltinExpr(BuiltinID, E);
5461 case llvm::Triple::ppc:
5462 case llvm::Triple::ppcle:
5463 case llvm::Triple::ppc64:
5464 case llvm::Triple::ppc64le:
5465 return CGF->EmitPPCBuiltinExpr(BuiltinID, E);
5466 case llvm::Triple::r600:
5467 case llvm::Triple::amdgcn:
5468 return CGF->EmitAMDGPUBuiltinExpr(BuiltinID, E);
5469 case llvm::Triple::systemz:
5470 return CGF->EmitSystemZBuiltinExpr(BuiltinID, E);
5471 case llvm::Triple::nvptx:
5472 case llvm::Triple::nvptx64:
5473 return CGF->EmitNVPTXBuiltinExpr(BuiltinID, E);
5474 case llvm::Triple::wasm32:
5475 case llvm::Triple::wasm64:
5476 return CGF->EmitWebAssemblyBuiltinExpr(BuiltinID, E);
5477 case llvm::Triple::hexagon:
5478 return CGF->EmitHexagonBuiltinExpr(BuiltinID, E);
5479 case llvm::Triple::riscv32:
5480 case llvm::Triple::riscv64:
5481 return CGF->EmitRISCVBuiltinExpr(BuiltinID, E, ReturnValue);
5482 case llvm::Triple::loongarch32:
5483 case llvm::Triple::loongarch64:
5484 return CGF->EmitLoongArchBuiltinExpr(BuiltinID, E);
5485 default:
5486 return nullptr;
5487 }
5488 }
5489
5490 Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID,
5491 const CallExpr *E,
5492 ReturnValueSlot ReturnValue) {
5493 if (getContext().BuiltinInfo.isAuxBuiltinID(BuiltinID)) {
5494 assert(getContext().getAuxTargetInfo() && "Missing aux target info");
5495 return EmitTargetArchBuiltinExpr(
5496 this, getContext().BuiltinInfo.getAuxBuiltinID(BuiltinID), E,
5497 ReturnValue, getContext().getAuxTargetInfo()->getTriple().getArch());
5498 }
5499
5500 return EmitTargetArchBuiltinExpr(this, BuiltinID, E, ReturnValue,
5501 getTarget().getTriple().getArch());
5502 }
5503
5504 static llvm::FixedVectorType *GetNeonType(CodeGenFunction *CGF,
5505 NeonTypeFlags TypeFlags,
5506 bool HasLegalHalfType = true,
5507 bool V1Ty = false,
5508 bool AllowBFloatArgsAndRet = true) {
5509 int IsQuad = TypeFlags.isQuad();
5510 switch (TypeFlags.getEltType()) {
5511 case NeonTypeFlags::Int8:
5512 case NeonTypeFlags::Poly8:
5513 return llvm::FixedVectorType::get(CGF->Int8Ty, V1Ty ? 1 : (8 << IsQuad));
5514 case NeonTypeFlags::Int16:
5515 case NeonTypeFlags::Poly16:
5516 return llvm::FixedVectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad));
5517 case NeonTypeFlags::BFloat16:
5518 if (AllowBFloatArgsAndRet)
5519 return llvm::FixedVectorType::get(CGF->BFloatTy, V1Ty ? 1 : (4 << IsQuad));
5520 else
5521 return llvm::FixedVectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad));
5522 case NeonTypeFlags::Float16:
5523 if (HasLegalHalfType)
5524 return llvm::FixedVectorType::get(CGF->HalfTy, V1Ty ? 1 : (4 << IsQuad));
5525 else
5526 return llvm::FixedVectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad));
5527 case NeonTypeFlags::Int32:
5528 return llvm::FixedVectorType::get(CGF->Int32Ty, V1Ty ? 1 : (2 << IsQuad));
5529 case NeonTypeFlags::Int64:
5530 case NeonTypeFlags::Poly64:
5531 return llvm::FixedVectorType::get(CGF->Int64Ty, V1Ty ? 1 : (1 << IsQuad));
5532 case NeonTypeFlags::Poly128:
5533 // FIXME: i128 and f128 doesn't get fully support in Clang and llvm.
5534 // There is a lot of i128 and f128 API missing.
5535 // so we use v16i8 to represent poly128 and get pattern matched.
5536 return llvm::FixedVectorType::get(CGF->Int8Ty, 16);
5537 case NeonTypeFlags::Float32:
5538 return llvm::FixedVectorType::get(CGF->FloatTy, V1Ty ? 1 : (2 << IsQuad));
5539 case NeonTypeFlags::Float64:
5540 return llvm::FixedVectorType::get(CGF->DoubleTy, V1Ty ? 1 : (1 << IsQuad));
5541 }
5542 llvm_unreachable("Unknown vector element type!");
5543 }
5544
5545 static llvm::VectorType *GetFloatNeonType(CodeGenFunction *CGF,
5546 NeonTypeFlags IntTypeFlags) {
5547 int IsQuad = IntTypeFlags.isQuad();
5548 switch (IntTypeFlags.getEltType()) {
5549 case NeonTypeFlags::Int16:
5550 return llvm::FixedVectorType::get(CGF->HalfTy, (4 << IsQuad));
5551 case NeonTypeFlags::Int32:
5552 return llvm::FixedVectorType::get(CGF->FloatTy, (2 << IsQuad));
5553 case NeonTypeFlags::Int64:
5554 return llvm::FixedVectorType::get(CGF->DoubleTy, (1 << IsQuad));
5555 default:
5556 llvm_unreachable("Type can't be converted to floating-point!");
5557 }
5558 }
5559
5560 Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C,
5561 const ElementCount &Count) {
5562 Value *SV = llvm::ConstantVector::getSplat(Count, C);
5563 return Builder.CreateShuffleVector(V, V, SV, "lane");
5564 }
5565
5566 Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) {
5567 ElementCount EC = cast<llvm::VectorType>(V->getType())->getElementCount();
5568 return EmitNeonSplat(V, C, EC);
5569 }
5570
5571 Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops,
5572 const char *name,
5573 unsigned shift, bool rightshift) {
5574 unsigned j = 0;
5575 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end();
5576 ai != ae; ++ai, ++j) {
5577 if (F->isConstrainedFPIntrinsic())
5578 if (ai->getType()->isMetadataTy())
5579 continue;
5580 if (shift > 0 && shift == j)
5581 Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift);
5582 else
5583 Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name);
5584 }
5585
5586 if (F->isConstrainedFPIntrinsic())
5587 return Builder.CreateConstrainedFPCall(F, Ops, name);
5588 else
5589 return Builder.CreateCall(F, Ops, name);
5590 }
5591
5592 Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty,
5593 bool neg) {
5594 int SV = cast<ConstantInt>(V)->getSExtValue();
5595 return ConstantInt::get(Ty, neg ? -SV : SV);
5596 }
5597
5598 // Right-shift a vector by a constant.
5599 Value *CodeGenFunction::EmitNeonRShiftImm(Value *Vec, Value *Shift,
5600 llvm::Type *Ty, bool usgn,
5601 const char *name) {
5602 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty);
5603
5604 int ShiftAmt = cast<ConstantInt>(Shift)->getSExtValue();
5605 int EltSize = VTy->getScalarSizeInBits();
5606
5607 Vec = Builder.CreateBitCast(Vec, Ty);
5608
5609 // lshr/ashr are undefined when the shift amount is equal to the vector
5610 // element size.
5611 if (ShiftAmt == EltSize) {
5612 if (usgn) {
5613 // Right-shifting an unsigned value by its size yields 0.
5614 return llvm::ConstantAggregateZero::get(VTy);
5615 } else {
5616 // Right-shifting a signed value by its size is equivalent
5617 // to a shift of size-1.
5618 --ShiftAmt;
5619 Shift = ConstantInt::get(VTy->getElementType(), ShiftAmt);
5620 }
5621 }
5622
5623 Shift = EmitNeonShiftVector(Shift, Ty, false);
5624 if (usgn)
5625 return Builder.CreateLShr(Vec, Shift, name);
5626 else
5627 return Builder.CreateAShr(Vec, Shift, name);
5628 }
5629
5630 enum {
5631 AddRetType = (1 << 0),
5632 Add1ArgType = (1 << 1),
5633 Add2ArgTypes = (1 << 2),
5634
5635 VectorizeRetType = (1 << 3),
5636 VectorizeArgTypes = (1 << 4),
5637
5638 InventFloatType = (1 << 5),
5639 UnsignedAlts = (1 << 6),
5640
5641 Use64BitVectors = (1 << 7),
5642 Use128BitVectors = (1 << 8),
5643
5644 Vectorize1ArgType = Add1ArgType | VectorizeArgTypes,
5645 VectorRet = AddRetType | VectorizeRetType,
5646 VectorRetGetArgs01 =
5647 AddRetType | Add2ArgTypes | VectorizeRetType | VectorizeArgTypes,
5648 FpCmpzModifiers =
5649 AddRetType | VectorizeRetType | Add1ArgType | InventFloatType
5650 };
5651
5652 namespace {
5653 struct ARMVectorIntrinsicInfo {
5654 const char *NameHint;
5655 unsigned BuiltinID;
5656 unsigned LLVMIntrinsic;
5657 unsigned AltLLVMIntrinsic;
5658 uint64_t TypeModifier;
5659
5660 bool operator<(unsigned RHSBuiltinID) const {
5661 return BuiltinID < RHSBuiltinID;
5662 }
5663 bool operator<(const ARMVectorIntrinsicInfo &TE) const {
5664 return BuiltinID < TE.BuiltinID;
5665 }
5666 };
5667 } // end anonymous namespace
5668
5669 #define NEONMAP0(NameBase) \
5670 { #NameBase, NEON::BI__builtin_neon_ ## NameBase, 0, 0, 0 }
5671
5672 #define NEONMAP1(NameBase, LLVMIntrinsic, TypeModifier) \
5673 { #NameBase, NEON:: BI__builtin_neon_ ## NameBase, \
5674 Intrinsic::LLVMIntrinsic, 0, TypeModifier }
5675
5676 #define NEONMAP2(NameBase, LLVMIntrinsic, AltLLVMIntrinsic, TypeModifier) \
5677 { #NameBase, NEON:: BI__builtin_neon_ ## NameBase, \
5678 Intrinsic::LLVMIntrinsic, Intrinsic::AltLLVMIntrinsic, \
5679 TypeModifier }
5680
5681 static const ARMVectorIntrinsicInfo ARMSIMDIntrinsicMap [] = {
5682 NEONMAP1(__a32_vcvt_bf16_f32, arm_neon_vcvtfp2bf, 0),
5683 NEONMAP0(splat_lane_v),
5684 NEONMAP0(splat_laneq_v),
5685 NEONMAP0(splatq_lane_v),
5686 NEONMAP0(splatq_laneq_v),
5687 NEONMAP2(vabd_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts),
5688 NEONMAP2(vabdq_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts),
5689 NEONMAP1(vabs_v, arm_neon_vabs, 0),
5690 NEONMAP1(vabsq_v, arm_neon_vabs, 0),
5691 NEONMAP0(vadd_v),
5692 NEONMAP0(vaddhn_v),
5693 NEONMAP0(vaddq_v),
5694 NEONMAP1(vaesdq_u8, arm_neon_aesd, 0),
5695 NEONMAP1(vaeseq_u8, arm_neon_aese, 0),
5696 NEONMAP1(vaesimcq_u8, arm_neon_aesimc, 0),
5697 NEONMAP1(vaesmcq_u8, arm_neon_aesmc, 0),
5698 NEONMAP1(vbfdot_f32, arm_neon_bfdot, 0),
5699 NEONMAP1(vbfdotq_f32, arm_neon_bfdot, 0),
5700 NEONMAP1(vbfmlalbq_f32, arm_neon_bfmlalb, 0),
5701 NEONMAP1(vbfmlaltq_f32, arm_neon_bfmlalt, 0),
5702 NEONMAP1(vbfmmlaq_f32, arm_neon_bfmmla, 0),
5703 NEONMAP1(vbsl_v, arm_neon_vbsl, AddRetType),
5704 NEONMAP1(vbslq_v, arm_neon_vbsl, AddRetType),
5705 NEONMAP1(vcadd_rot270_f16, arm_neon_vcadd_rot270, Add1ArgType),
5706 NEONMAP1(vcadd_rot270_f32, arm_neon_vcadd_rot270, Add1ArgType),
5707 NEONMAP1(vcadd_rot90_f16, arm_neon_vcadd_rot90, Add1ArgType),
5708 NEONMAP1(vcadd_rot90_f32, arm_neon_vcadd_rot90, Add1ArgType),
5709 NEONMAP1(vcaddq_rot270_f16, arm_neon_vcadd_rot270, Add1ArgType),
5710 NEONMAP1(vcaddq_rot270_f32, arm_neon_vcadd_rot270, Add1ArgType),
5711 NEONMAP1(vcaddq_rot270_f64, arm_neon_vcadd_rot270, Add1ArgType),
5712 NEONMAP1(vcaddq_rot90_f16, arm_neon_vcadd_rot90, Add1ArgType),
5713 NEONMAP1(vcaddq_rot90_f32, arm_neon_vcadd_rot90, Add1ArgType),
5714 NEONMAP1(vcaddq_rot90_f64, arm_neon_vcadd_rot90, Add1ArgType),
5715 NEONMAP1(vcage_v, arm_neon_vacge, 0),
5716 NEONMAP1(vcageq_v, arm_neon_vacge, 0),
5717 NEONMAP1(vcagt_v, arm_neon_vacgt, 0),
5718 NEONMAP1(vcagtq_v, arm_neon_vacgt, 0),
5719 NEONMAP1(vcale_v, arm_neon_vacge, 0),
5720 NEONMAP1(vcaleq_v, arm_neon_vacge, 0),
5721 NEONMAP1(vcalt_v, arm_neon_vacgt, 0),
5722 NEONMAP1(vcaltq_v, arm_neon_vacgt, 0),
5723 NEONMAP0(vceqz_v),
5724 NEONMAP0(vceqzq_v),
5725 NEONMAP0(vcgez_v),
5726 NEONMAP0(vcgezq_v),
5727 NEONMAP0(vcgtz_v),
5728 NEONMAP0(vcgtzq_v),
5729 NEONMAP0(vclez_v),
5730 NEONMAP0(vclezq_v),
5731 NEONMAP1(vcls_v, arm_neon_vcls, Add1ArgType),
5732 NEONMAP1(vclsq_v, arm_neon_vcls, Add1ArgType),
5733 NEONMAP0(vcltz_v),
5734 NEONMAP0(vcltzq_v),
5735 NEONMAP1(vclz_v, ctlz, Add1ArgType),
5736 NEONMAP1(vclzq_v, ctlz, Add1ArgType),
5737 NEONMAP1(vcnt_v, ctpop, Add1ArgType),
5738 NEONMAP1(vcntq_v, ctpop, Add1ArgType),
5739 NEONMAP1(vcvt_f16_f32, arm_neon_vcvtfp2hf, 0),
5740 NEONMAP0(vcvt_f16_s16),
5741 NEONMAP0(vcvt_f16_u16),
5742 NEONMAP1(vcvt_f32_f16, arm_neon_vcvthf2fp, 0),
5743 NEONMAP0(vcvt_f32_v),
5744 NEONMAP1(vcvt_n_f16_s16, arm_neon_vcvtfxs2fp, 0),
5745 NEONMAP1(vcvt_n_f16_u16, arm_neon_vcvtfxu2fp, 0),
5746 NEONMAP2(vcvt_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0),
5747 NEONMAP1(vcvt_n_s16_f16, arm_neon_vcvtfp2fxs, 0),
5748 NEONMAP1(vcvt_n_s32_v, arm_neon_vcvtfp2fxs, 0),
5749 NEONMAP1(vcvt_n_s64_v, arm_neon_vcvtfp2fxs, 0),
5750 NEONMAP1(vcvt_n_u16_f16, arm_neon_vcvtfp2fxu, 0),
5751 NEONMAP1(vcvt_n_u32_v, arm_neon_vcvtfp2fxu, 0),
5752 NEONMAP1(vcvt_n_u64_v, arm_neon_vcvtfp2fxu, 0),
5753 NEONMAP0(vcvt_s16_f16),
5754 NEONMAP0(vcvt_s32_v),
5755 NEONMAP0(vcvt_s64_v),
5756 NEONMAP0(vcvt_u16_f16),
5757 NEONMAP0(vcvt_u32_v),
5758 NEONMAP0(vcvt_u64_v),
5759 NEONMAP1(vcvta_s16_f16, arm_neon_vcvtas, 0),
5760 NEONMAP1(vcvta_s32_v, arm_neon_vcvtas, 0),
5761 NEONMAP1(vcvta_s64_v, arm_neon_vcvtas, 0),
5762 NEONMAP1(vcvta_u16_f16, arm_neon_vcvtau, 0),
5763 NEONMAP1(vcvta_u32_v, arm_neon_vcvtau, 0),
5764 NEONMAP1(vcvta_u64_v, arm_neon_vcvtau, 0),
5765 NEONMAP1(vcvtaq_s16_f16, arm_neon_vcvtas, 0),
5766 NEONMAP1(vcvtaq_s32_v, arm_neon_vcvtas, 0),
5767 NEONMAP1(vcvtaq_s64_v, arm_neon_vcvtas, 0),
5768 NEONMAP1(vcvtaq_u16_f16, arm_neon_vcvtau, 0),
5769 NEONMAP1(vcvtaq_u32_v, arm_neon_vcvtau, 0),
5770 NEONMAP1(vcvtaq_u64_v, arm_neon_vcvtau, 0),
5771 NEONMAP1(vcvth_bf16_f32, arm_neon_vcvtbfp2bf, 0),
5772 NEONMAP1(vcvtm_s16_f16, arm_neon_vcvtms, 0),
5773 NEONMAP1(vcvtm_s32_v, arm_neon_vcvtms, 0),
5774 NEONMAP1(vcvtm_s64_v, arm_neon_vcvtms, 0),
5775 NEONMAP1(vcvtm_u16_f16, arm_neon_vcvtmu, 0),
5776 NEONMAP1(vcvtm_u32_v, arm_neon_vcvtmu, 0),
5777 NEONMAP1(vcvtm_u64_v, arm_neon_vcvtmu, 0),
5778 NEONMAP1(vcvtmq_s16_f16, arm_neon_vcvtms, 0),
5779 NEONMAP1(vcvtmq_s32_v, arm_neon_vcvtms, 0),
5780 NEONMAP1(vcvtmq_s64_v, arm_neon_vcvtms, 0),
5781 NEONMAP1(vcvtmq_u16_f16, arm_neon_vcvtmu, 0),
5782 NEONMAP1(vcvtmq_u32_v, arm_neon_vcvtmu, 0),
5783 NEONMAP1(vcvtmq_u64_v, arm_neon_vcvtmu, 0),
5784 NEONMAP1(vcvtn_s16_f16, arm_neon_vcvtns, 0),
5785 NEONMAP1(vcvtn_s32_v, arm_neon_vcvtns, 0),
5786 NEONMAP1(vcvtn_s64_v, arm_neon_vcvtns, 0),
5787 NEONMAP1(vcvtn_u16_f16, arm_neon_vcvtnu, 0),
5788 NEONMAP1(vcvtn_u32_v, arm_neon_vcvtnu, 0),
5789 NEONMAP1(vcvtn_u64_v, arm_neon_vcvtnu, 0),
5790 NEONMAP1(vcvtnq_s16_f16, arm_neon_vcvtns, 0),
5791 NEONMAP1(vcvtnq_s32_v, arm_neon_vcvtns, 0),
5792 NEONMAP1(vcvtnq_s64_v, arm_neon_vcvtns, 0),
5793 NEONMAP1(vcvtnq_u16_f16, arm_neon_vcvtnu, 0),
5794 NEONMAP1(vcvtnq_u32_v, arm_neon_vcvtnu, 0),
5795 NEONMAP1(vcvtnq_u64_v, arm_neon_vcvtnu, 0),
5796 NEONMAP1(vcvtp_s16_f16, arm_neon_vcvtps, 0),
5797 NEONMAP1(vcvtp_s32_v, arm_neon_vcvtps, 0),
5798 NEONMAP1(vcvtp_s64_v, arm_neon_vcvtps, 0),
5799 NEONMAP1(vcvtp_u16_f16, arm_neon_vcvtpu, 0),
5800 NEONMAP1(vcvtp_u32_v, arm_neon_vcvtpu, 0),
5801 NEONMAP1(vcvtp_u64_v, arm_neon_vcvtpu, 0),
5802 NEONMAP1(vcvtpq_s16_f16, arm_neon_vcvtps, 0),
5803 NEONMAP1(vcvtpq_s32_v, arm_neon_vcvtps, 0),
5804 NEONMAP1(vcvtpq_s64_v, arm_neon_vcvtps, 0),
5805 NEONMAP1(vcvtpq_u16_f16, arm_neon_vcvtpu, 0),
5806 NEONMAP1(vcvtpq_u32_v, arm_neon_vcvtpu, 0),
5807 NEONMAP1(vcvtpq_u64_v, arm_neon_vcvtpu, 0),
5808 NEONMAP0(vcvtq_f16_s16),
5809 NEONMAP0(vcvtq_f16_u16),
5810 NEONMAP0(vcvtq_f32_v),
5811 NEONMAP1(vcvtq_n_f16_s16, arm_neon_vcvtfxs2fp, 0),
5812 NEONMAP1(vcvtq_n_f16_u16, arm_neon_vcvtfxu2fp, 0),
5813 NEONMAP2(vcvtq_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0),
5814 NEONMAP1(vcvtq_n_s16_f16, arm_neon_vcvtfp2fxs, 0),
5815 NEONMAP1(vcvtq_n_s32_v, arm_neon_vcvtfp2fxs, 0),
5816 NEONMAP1(vcvtq_n_s64_v, arm_neon_vcvtfp2fxs, 0),
5817 NEONMAP1(vcvtq_n_u16_f16, arm_neon_vcvtfp2fxu, 0),
5818 NEONMAP1(vcvtq_n_u32_v, arm_neon_vcvtfp2fxu, 0),
5819 NEONMAP1(vcvtq_n_u64_v, arm_neon_vcvtfp2fxu, 0),
5820 NEONMAP0(vcvtq_s16_f16),
5821 NEONMAP0(vcvtq_s32_v),
5822 NEONMAP0(vcvtq_s64_v),
5823 NEONMAP0(vcvtq_u16_f16),
5824 NEONMAP0(vcvtq_u32_v),
5825 NEONMAP0(vcvtq_u64_v),
5826 NEONMAP1(vdot_s32, arm_neon_sdot, 0),
5827 NEONMAP1(vdot_u32, arm_neon_udot, 0),
5828 NEONMAP1(vdotq_s32, arm_neon_sdot, 0),
5829 NEONMAP1(vdotq_u32, arm_neon_udot, 0),
5830 NEONMAP0(vext_v),
5831 NEONMAP0(vextq_v),
5832 NEONMAP0(vfma_v),
5833 NEONMAP0(vfmaq_v),
5834 NEONMAP2(vhadd_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts),
5835 NEONMAP2(vhaddq_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts),
5836 NEONMAP2(vhsub_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts),
5837 NEONMAP2(vhsubq_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts),
5838 NEONMAP0(vld1_dup_v),
5839 NEONMAP1(vld1_v, arm_neon_vld1, 0),
5840 NEONMAP1(vld1_x2_v, arm_neon_vld1x2, 0),
5841 NEONMAP1(vld1_x3_v, arm_neon_vld1x3, 0),
5842 NEONMAP1(vld1_x4_v, arm_neon_vld1x4, 0),
5843 NEONMAP0(vld1q_dup_v),
5844 NEONMAP1(vld1q_v, arm_neon_vld1, 0),
5845 NEONMAP1(vld1q_x2_v, arm_neon_vld1x2, 0),
5846 NEONMAP1(vld1q_x3_v, arm_neon_vld1x3, 0),
5847 NEONMAP1(vld1q_x4_v, arm_neon_vld1x4, 0),
5848 NEONMAP1(vld2_dup_v, arm_neon_vld2dup, 0),
5849 NEONMAP1(vld2_lane_v, arm_neon_vld2lane, 0),
5850 NEONMAP1(vld2_v, arm_neon_vld2, 0),
5851 NEONMAP1(vld2q_dup_v, arm_neon_vld2dup, 0),
5852 NEONMAP1(vld2q_lane_v, arm_neon_vld2lane, 0),
5853 NEONMAP1(vld2q_v, arm_neon_vld2, 0),
5854 NEONMAP1(vld3_dup_v, arm_neon_vld3dup, 0),
5855 NEONMAP1(vld3_lane_v, arm_neon_vld3lane, 0),
5856 NEONMAP1(vld3_v, arm_neon_vld3, 0),
5857 NEONMAP1(vld3q_dup_v, arm_neon_vld3dup, 0),
5858 NEONMAP1(vld3q_lane_v, arm_neon_vld3lane, 0),
5859 NEONMAP1(vld3q_v, arm_neon_vld3, 0),
5860 NEONMAP1(vld4_dup_v, arm_neon_vld4dup, 0),
5861 NEONMAP1(vld4_lane_v, arm_neon_vld4lane, 0),
5862 NEONMAP1(vld4_v, arm_neon_vld4, 0),
5863 NEONMAP1(vld4q_dup_v, arm_neon_vld4dup, 0),
5864 NEONMAP1(vld4q_lane_v, arm_neon_vld4lane, 0),
5865 NEONMAP1(vld4q_v, arm_neon_vld4, 0),
5866 NEONMAP2(vmax_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts),
5867 NEONMAP1(vmaxnm_v, arm_neon_vmaxnm, Add1ArgType),
5868 NEONMAP1(vmaxnmq_v, arm_neon_vmaxnm, Add1ArgType),
5869 NEONMAP2(vmaxq_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts),
5870 NEONMAP2(vmin_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts),
5871 NEONMAP1(vminnm_v, arm_neon_vminnm, Add1ArgType),
5872 NEONMAP1(vminnmq_v, arm_neon_vminnm, Add1ArgType),
5873 NEONMAP2(vminq_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts),
5874 NEONMAP1(vmmlaq_s32, arm_neon_smmla, 0),
5875 NEONMAP1(vmmlaq_u32, arm_neon_ummla, 0),
5876 NEONMAP0(vmovl_v),
5877 NEONMAP0(vmovn_v),
5878 NEONMAP1(vmul_v, arm_neon_vmulp, Add1ArgType),
5879 NEONMAP0(vmull_v),
5880 NEONMAP1(vmulq_v, arm_neon_vmulp, Add1ArgType),
5881 NEONMAP2(vpadal_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts),
5882 NEONMAP2(vpadalq_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts),
5883 NEONMAP1(vpadd_v, arm_neon_vpadd, Add1ArgType),
5884 NEONMAP2(vpaddl_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts),
5885 NEONMAP2(vpaddlq_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts),
5886 NEONMAP1(vpaddq_v, arm_neon_vpadd, Add1ArgType),
5887 NEONMAP2(vpmax_v, arm_neon_vpmaxu, arm_neon_vpmaxs, Add1ArgType | UnsignedAlts),
5888 NEONMAP2(vpmin_v, arm_neon_vpminu, arm_neon_vpmins, Add1ArgType | UnsignedAlts),
5889 NEONMAP1(vqabs_v, arm_neon_vqabs, Add1ArgType),
5890 NEONMAP1(vqabsq_v, arm_neon_vqabs, Add1ArgType),
5891 NEONMAP2(vqadd_v, uadd_sat, sadd_sat, Add1ArgType | UnsignedAlts),
5892 NEONMAP2(vqaddq_v, uadd_sat, sadd_sat, Add1ArgType | UnsignedAlts),
5893 NEONMAP2(vqdmlal_v, arm_neon_vqdmull, sadd_sat, 0),
5894 NEONMAP2(vqdmlsl_v, arm_neon_vqdmull, ssub_sat, 0),
5895 NEONMAP1(vqdmulh_v, arm_neon_vqdmulh, Add1ArgType),
5896 NEONMAP1(vqdmulhq_v, arm_neon_vqdmulh, Add1ArgType),
5897 NEONMAP1(vqdmull_v, arm_neon_vqdmull, Add1ArgType),
5898 NEONMAP2(vqmovn_v, arm_neon_vqmovnu, arm_neon_vqmovns, Add1ArgType | UnsignedAlts),
5899 NEONMAP1(vqmovun_v, arm_neon_vqmovnsu, Add1ArgType),
5900 NEONMAP1(vqneg_v, arm_neon_vqneg, Add1ArgType),
5901 NEONMAP1(vqnegq_v, arm_neon_vqneg, Add1ArgType),
5902 NEONMAP1(vqrdmlah_s16, arm_neon_vqrdmlah, Add1ArgType),
5903 NEONMAP1(vqrdmlah_s32, arm_neon_vqrdmlah, Add1ArgType),
5904 NEONMAP1(vqrdmlahq_s16, arm_neon_vqrdmlah, Add1ArgType),
5905 NEONMAP1(vqrdmlahq_s32, arm_neon_vqrdmlah, Add1ArgType),
5906 NEONMAP1(vqrdmlsh_s16, arm_neon_vqrdmlsh, Add1ArgType),
5907 NEONMAP1(vqrdmlsh_s32, arm_neon_vqrdmlsh, Add1ArgType),
5908 NEONMAP1(vqrdmlshq_s16, arm_neon_vqrdmlsh, Add1ArgType),
5909 NEONMAP1(vqrdmlshq_s32, arm_neon_vqrdmlsh, Add1ArgType),
5910 NEONMAP1(vqrdmulh_v, arm_neon_vqrdmulh, Add1ArgType),
5911 NEONMAP1(vqrdmulhq_v, arm_neon_vqrdmulh, Add1ArgType),
5912 NEONMAP2(vqrshl_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts),
5913 NEONMAP2(vqrshlq_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts),
5914 NEONMAP2(vqshl_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts),
5915 NEONMAP2(vqshl_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts),
5916 NEONMAP2(vqshlq_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts),
5917 NEONMAP2(vqshlq_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts),
5918 NEONMAP1(vqshlu_n_v, arm_neon_vqshiftsu, 0),
5919 NEONMAP1(vqshluq_n_v, arm_neon_vqshiftsu, 0),
5920 NEONMAP2(vqsub_v, usub_sat, ssub_sat, Add1ArgType | UnsignedAlts),
5921 NEONMAP2(vqsubq_v, usub_sat, ssub_sat, Add1ArgType | UnsignedAlts),
5922 NEONMAP1(vraddhn_v, arm_neon_vraddhn, Add1ArgType),
5923 NEONMAP2(vrecpe_v, arm_neon_vrecpe, arm_neon_vrecpe, 0),
5924 NEONMAP2(vrecpeq_v, arm_neon_vrecpe, arm_neon_vrecpe, 0),
5925 NEONMAP1(vrecps_v, arm_neon_vrecps, Add1ArgType),
5926 NEONMAP1(vrecpsq_v, arm_neon_vrecps, Add1ArgType),
5927 NEONMAP2(vrhadd_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts),
5928 NEONMAP2(vrhaddq_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts),
5929 NEONMAP1(vrnd_v, arm_neon_vrintz, Add1ArgType),
5930 NEONMAP1(vrnda_v, arm_neon_vrinta, Add1ArgType),
5931 NEONMAP1(vrndaq_v, arm_neon_vrinta, Add1ArgType),
5932 NEONMAP0(vrndi_v),
5933 NEONMAP0(vrndiq_v),
5934 NEONMAP1(vrndm_v, arm_neon_vrintm, Add1ArgType),
5935 NEONMAP1(vrndmq_v, arm_neon_vrintm, Add1ArgType),
5936 NEONMAP1(vrndn_v, arm_neon_vrintn, Add1ArgType),
5937 NEONMAP1(vrndnq_v, arm_neon_vrintn, Add1ArgType),
5938 NEONMAP1(vrndp_v, arm_neon_vrintp, Add1ArgType),
5939 NEONMAP1(vrndpq_v, arm_neon_vrintp, Add1ArgType),
5940 NEONMAP1(vrndq_v, arm_neon_vrintz, Add1ArgType),
5941 NEONMAP1(vrndx_v, arm_neon_vrintx, Add1ArgType),
5942 NEONMAP1(vrndxq_v, arm_neon_vrintx, Add1ArgType),
5943 NEONMAP2(vrshl_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts),
5944 NEONMAP2(vrshlq_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts),
5945 NEONMAP2(vrshr_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts),
5946 NEONMAP2(vrshrq_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts),
5947 NEONMAP2(vrsqrte_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0),
5948 NEONMAP2(vrsqrteq_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0),
5949 NEONMAP1(vrsqrts_v, arm_neon_vrsqrts, Add1ArgType),
5950 NEONMAP1(vrsqrtsq_v, arm_neon_vrsqrts, Add1ArgType),
5951 NEONMAP1(vrsubhn_v, arm_neon_vrsubhn, Add1ArgType),
5952 NEONMAP1(vsha1su0q_u32, arm_neon_sha1su0, 0),
5953 NEONMAP1(vsha1su1q_u32, arm_neon_sha1su1, 0),
5954 NEONMAP1(vsha256h2q_u32, arm_neon_sha256h2, 0),
5955 NEONMAP1(vsha256hq_u32, arm_neon_sha256h, 0),
5956 NEONMAP1(vsha256su0q_u32, arm_neon_sha256su0, 0),
5957 NEONMAP1(vsha256su1q_u32, arm_neon_sha256su1, 0),
5958 NEONMAP0(vshl_n_v),
5959 NEONMAP2(vshl_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts),
5960 NEONMAP0(vshll_n_v),
5961 NEONMAP0(vshlq_n_v),
5962 NEONMAP2(vshlq_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts),
5963 NEONMAP0(vshr_n_v),
5964 NEONMAP0(vshrn_n_v),
5965 NEONMAP0(vshrq_n_v),
5966 NEONMAP1(vst1_v, arm_neon_vst1, 0),
5967 NEONMAP1(vst1_x2_v, arm_neon_vst1x2, 0),
5968 NEONMAP1(vst1_x3_v, arm_neon_vst1x3, 0),
5969 NEONMAP1(vst1_x4_v, arm_neon_vst1x4, 0),
5970 NEONMAP1(vst1q_v, arm_neon_vst1, 0),
5971 NEONMAP1(vst1q_x2_v, arm_neon_vst1x2, 0),
5972 NEONMAP1(vst1q_x3_v, arm_neon_vst1x3, 0),
5973 NEONMAP1(vst1q_x4_v, arm_neon_vst1x4, 0),
5974 NEONMAP1(vst2_lane_v, arm_neon_vst2lane, 0),
5975 NEONMAP1(vst2_v, arm_neon_vst2, 0),
5976 NEONMAP1(vst2q_lane_v, arm_neon_vst2lane, 0),
5977 NEONMAP1(vst2q_v, arm_neon_vst2, 0),
5978 NEONMAP1(vst3_lane_v, arm_neon_vst3lane, 0),
5979 NEONMAP1(vst3_v, arm_neon_vst3, 0),
5980 NEONMAP1(vst3q_lane_v, arm_neon_vst3lane, 0),
5981 NEONMAP1(vst3q_v, arm_neon_vst3, 0),
5982 NEONMAP1(vst4_lane_v, arm_neon_vst4lane, 0),
5983 NEONMAP1(vst4_v, arm_neon_vst4, 0),
5984 NEONMAP1(vst4q_lane_v, arm_neon_vst4lane, 0),
5985 NEONMAP1(vst4q_v, arm_neon_vst4, 0),
5986 NEONMAP0(vsubhn_v),
5987 NEONMAP0(vtrn_v),
5988 NEONMAP0(vtrnq_v),
5989 NEONMAP0(vtst_v),
5990 NEONMAP0(vtstq_v),
5991 NEONMAP1(vusdot_s32, arm_neon_usdot, 0),
5992 NEONMAP1(vusdotq_s32, arm_neon_usdot, 0),
5993 NEONMAP1(vusmmlaq_s32, arm_neon_usmmla, 0),
5994 NEONMAP0(vuzp_v),
5995 NEONMAP0(vuzpq_v),
5996 NEONMAP0(vzip_v),
5997 NEONMAP0(vzipq_v)
5998 };
5999
6000 static const ARMVectorIntrinsicInfo AArch64SIMDIntrinsicMap[] = {
6001 NEONMAP1(__a64_vcvtq_low_bf16_f32, aarch64_neon_bfcvtn, 0),
6002 NEONMAP0(splat_lane_v),
6003 NEONMAP0(splat_laneq_v),
6004 NEONMAP0(splatq_lane_v),
6005 NEONMAP0(splatq_laneq_v),
6006 NEONMAP1(vabs_v, aarch64_neon_abs, 0),
6007 NEONMAP1(vabsq_v, aarch64_neon_abs, 0),
6008 NEONMAP0(vadd_v),
6009 NEONMAP0(vaddhn_v),
6010 NEONMAP0(vaddq_p128),
6011 NEONMAP0(vaddq_v),
6012 NEONMAP1(vaesdq_u8, aarch64_crypto_aesd, 0),
6013 NEONMAP1(vaeseq_u8, aarch64_crypto_aese, 0),
6014 NEONMAP1(vaesimcq_u8, aarch64_crypto_aesimc, 0),
6015 NEONMAP1(vaesmcq_u8, aarch64_crypto_aesmc, 0),
6016 NEONMAP2(vbcaxq_s16, aarch64_crypto_bcaxu, aarch64_crypto_bcaxs, Add1ArgType | UnsignedAlts),
6017 NEONMAP2(vbcaxq_s32, aarch64_crypto_bcaxu, aarch64_crypto_bcaxs, Add1ArgType | UnsignedAlts),
6018 NEONMAP2(vbcaxq_s64, aarch64_crypto_bcaxu, aarch64_crypto_bcaxs, Add1ArgType | UnsignedAlts),
6019 NEONMAP2(vbcaxq_s8, aarch64_crypto_bcaxu, aarch64_crypto_bcaxs, Add1ArgType | UnsignedAlts),
6020 NEONMAP2(vbcaxq_u16, aarch64_crypto_bcaxu, aarch64_crypto_bcaxs, Add1ArgType | UnsignedAlts),
6021 NEONMAP2(vbcaxq_u32, aarch64_crypto_bcaxu, aarch64_crypto_bcaxs, Add1ArgType | UnsignedAlts),
6022 NEONMAP2(vbcaxq_u64, aarch64_crypto_bcaxu, aarch64_crypto_bcaxs, Add1ArgType | UnsignedAlts),
6023 NEONMAP2(vbcaxq_u8, aarch64_crypto_bcaxu, aarch64_crypto_bcaxs, Add1ArgType | UnsignedAlts),
6024 NEONMAP1(vbfdot_f32, aarch64_neon_bfdot, 0),
6025 NEONMAP1(vbfdotq_f32, aarch64_neon_bfdot, 0),
6026 NEONMAP1(vbfmlalbq_f32, aarch64_neon_bfmlalb, 0),
6027 NEONMAP1(vbfmlaltq_f32, aarch64_neon_bfmlalt, 0),
6028 NEONMAP1(vbfmmlaq_f32, aarch64_neon_bfmmla, 0),
6029 NEONMAP1(vcadd_rot270_f16, aarch64_neon_vcadd_rot270, Add1ArgType),
6030 NEONMAP1(vcadd_rot270_f32, aarch64_neon_vcadd_rot270, Add1ArgType),
6031 NEONMAP1(vcadd_rot90_f16, aarch64_neon_vcadd_rot90, Add1ArgType),
6032 NEONMAP1(vcadd_rot90_f32, aarch64_neon_vcadd_rot90, Add1ArgType),
6033 NEONMAP1(vcaddq_rot270_f16, aarch64_neon_vcadd_rot270, Add1ArgType),
6034 NEONMAP1(vcaddq_rot270_f32, aarch64_neon_vcadd_rot270, Add1ArgType),
6035 NEONMAP1(vcaddq_rot270_f64, aarch64_neon_vcadd_rot270, Add1ArgType),
6036 NEONMAP1(vcaddq_rot90_f16, aarch64_neon_vcadd_rot90, Add1ArgType),
6037 NEONMAP1(vcaddq_rot90_f32, aarch64_neon_vcadd_rot90, Add1ArgType),
6038 NEONMAP1(vcaddq_rot90_f64, aarch64_neon_vcadd_rot90, Add1ArgType),
6039 NEONMAP1(vcage_v, aarch64_neon_facge, 0),
6040 NEONMAP1(vcageq_v, aarch64_neon_facge, 0),
6041 NEONMAP1(vcagt_v, aarch64_neon_facgt, 0),
6042 NEONMAP1(vcagtq_v, aarch64_neon_facgt, 0),
6043 NEONMAP1(vcale_v, aarch64_neon_facge, 0),
6044 NEONMAP1(vcaleq_v, aarch64_neon_facge, 0),
6045 NEONMAP1(vcalt_v, aarch64_neon_facgt, 0),
6046 NEONMAP1(vcaltq_v, aarch64_neon_facgt, 0),
6047 NEONMAP0(vceqz_v),
6048 NEONMAP0(vceqzq_v),
6049 NEONMAP0(vcgez_v),
6050 NEONMAP0(vcgezq_v),
6051 NEONMAP0(vcgtz_v),
6052 NEONMAP0(vcgtzq_v),
6053 NEONMAP0(vclez_v),
6054 NEONMAP0(vclezq_v),
6055 NEONMAP1(vcls_v, aarch64_neon_cls, Add1ArgType),
6056 NEONMAP1(vclsq_v, aarch64_neon_cls, Add1ArgType),
6057 NEONMAP0(vcltz_v),
6058 NEONMAP0(vcltzq_v),
6059 NEONMAP1(vclz_v, ctlz, Add1ArgType),
6060 NEONMAP1(vclzq_v, ctlz, Add1ArgType),
6061 NEONMAP1(vcmla_f16, aarch64_neon_vcmla_rot0, Add1ArgType),
6062 NEONMAP1(vcmla_f32, aarch64_neon_vcmla_rot0, Add1ArgType),
6063 NEONMAP1(vcmla_rot180_f16, aarch64_neon_vcmla_rot180, Add1ArgType),
6064 NEONMAP1(vcmla_rot180_f32, aarch64_neon_vcmla_rot180, Add1ArgType),
6065 NEONMAP1(vcmla_rot270_f16, aarch64_neon_vcmla_rot270, Add1ArgType),
6066 NEONMAP1(vcmla_rot270_f32, aarch64_neon_vcmla_rot270, Add1ArgType),
6067 NEONMAP1(vcmla_rot90_f16, aarch64_neon_vcmla_rot90, Add1ArgType),
6068 NEONMAP1(vcmla_rot90_f32, aarch64_neon_vcmla_rot90, Add1ArgType),
6069 NEONMAP1(vcmlaq_f16, aarch64_neon_vcmla_rot0, Add1ArgType),
6070 NEONMAP1(vcmlaq_f32, aarch64_neon_vcmla_rot0, Add1ArgType),
6071 NEONMAP1(vcmlaq_f64, aarch64_neon_vcmla_rot0, Add1ArgType),
6072 NEONMAP1(vcmlaq_rot180_f16, aarch64_neon_vcmla_rot180, Add1ArgType),
6073 NEONMAP1(vcmlaq_rot180_f32, aarch64_neon_vcmla_rot180, Add1ArgType),
6074 NEONMAP1(vcmlaq_rot180_f64, aarch64_neon_vcmla_rot180, Add1ArgType),
6075 NEONMAP1(vcmlaq_rot270_f16, aarch64_neon_vcmla_rot270, Add1ArgType),
6076 NEONMAP1(vcmlaq_rot270_f32, aarch64_neon_vcmla_rot270, Add1ArgType),
6077 NEONMAP1(vcmlaq_rot270_f64, aarch64_neon_vcmla_rot270, Add1ArgType),
6078 NEONMAP1(vcmlaq_rot90_f16, aarch64_neon_vcmla_rot90, Add1ArgType),
6079 NEONMAP1(vcmlaq_rot90_f32, aarch64_neon_vcmla_rot90, Add1ArgType),
6080 NEONMAP1(vcmlaq_rot90_f64, aarch64_neon_vcmla_rot90, Add1ArgType),
6081 NEONMAP1(vcnt_v, ctpop, Add1ArgType),
6082 NEONMAP1(vcntq_v, ctpop, Add1ArgType),
6083 NEONMAP1(vcvt_f16_f32, aarch64_neon_vcvtfp2hf, 0),
6084 NEONMAP0(vcvt_f16_s16),
6085 NEONMAP0(vcvt_f16_u16),
6086 NEONMAP1(vcvt_f32_f16, aarch64_neon_vcvthf2fp, 0),
6087 NEONMAP0(vcvt_f32_v),
6088 NEONMAP1(vcvt_n_f16_s16, aarch64_neon_vcvtfxs2fp, 0),
6089 NEONMAP1(vcvt_n_f16_u16, aarch64_neon_vcvtfxu2fp, 0),
6090 NEONMAP2(vcvt_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
6091 NEONMAP2(vcvt_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
6092 NEONMAP1(vcvt_n_s16_f16, aarch64_neon_vcvtfp2fxs, 0),
6093 NEONMAP1(vcvt_n_s32_v, aarch64_neon_vcvtfp2fxs, 0),
6094 NEONMAP1(vcvt_n_s64_v, aarch64_neon_vcvtfp2fxs, 0),
6095 NEONMAP1(vcvt_n_u16_f16, aarch64_neon_vcvtfp2fxu, 0),
6096 NEONMAP1(vcvt_n_u32_v, aarch64_neon_vcvtfp2fxu, 0),
6097 NEONMAP1(vcvt_n_u64_v, aarch64_neon_vcvtfp2fxu, 0),
6098 NEONMAP0(vcvtq_f16_s16),
6099 NEONMAP0(vcvtq_f16_u16),
6100 NEONMAP0(vcvtq_f32_v),
6101 NEONMAP1(vcvtq_high_bf16_f32, aarch64_neon_bfcvtn2, 0),
6102 NEONMAP1(vcvtq_n_f16_s16, aarch64_neon_vcvtfxs2fp, 0),
6103 NEONMAP1(vcvtq_n_f16_u16, aarch64_neon_vcvtfxu2fp, 0),
6104 NEONMAP2(vcvtq_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
6105 NEONMAP2(vcvtq_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0),
6106 NEONMAP1(vcvtq_n_s16_f16, aarch64_neon_vcvtfp2fxs, 0),
6107 NEONMAP1(vcvtq_n_s32_v, aarch64_neon_vcvtfp2fxs, 0),
6108 NEONMAP1(vcvtq_n_s64_v, aarch64_neon_vcvtfp2fxs, 0),
6109 NEONMAP1(vcvtq_n_u16_f16, aarch64_neon_vcvtfp2fxu, 0),
6110 NEONMAP1(vcvtq_n_u32_v, aarch64_neon_vcvtfp2fxu, 0),
6111 NEONMAP1(vcvtq_n_u64_v, aarch64_neon_vcvtfp2fxu, 0),
6112 NEONMAP1(vcvtx_f32_v, aarch64_neon_fcvtxn, AddRetType | Add1ArgType),
6113 NEONMAP1(vdot_s32, aarch64_neon_sdot, 0),
6114 NEONMAP1(vdot_u32, aarch64_neon_udot, 0),
6115 NEONMAP1(vdotq_s32, aarch64_neon_sdot, 0),
6116 NEONMAP1(vdotq_u32, aarch64_neon_udot, 0),
6117 NEONMAP2(veor3q_s16, aarch64_crypto_eor3u, aarch64_crypto_eor3s, Add1ArgType | UnsignedAlts),
6118 NEONMAP2(veor3q_s32, aarch64_crypto_eor3u, aarch64_crypto_eor3s, Add1ArgType | UnsignedAlts),
6119 NEONMAP2(veor3q_s64, aarch64_crypto_eor3u, aarch64_crypto_eor3s, Add1ArgType | UnsignedAlts),
6120 NEONMAP2(veor3q_s8, aarch64_crypto_eor3u, aarch64_crypto_eor3s, Add1ArgType | UnsignedAlts),
6121 NEONMAP2(veor3q_u16, aarch64_crypto_eor3u, aarch64_crypto_eor3s, Add1ArgType | UnsignedAlts),
6122 NEONMAP2(veor3q_u32, aarch64_crypto_eor3u, aarch64_crypto_eor3s, Add1ArgType | UnsignedAlts),
6123 NEONMAP2(veor3q_u64, aarch64_crypto_eor3u, aarch64_crypto_eor3s, Add1ArgType | UnsignedAlts),
6124 NEONMAP2(veor3q_u8, aarch64_crypto_eor3u, aarch64_crypto_eor3s, Add1ArgType | UnsignedAlts),
6125 NEONMAP0(vext_v),
6126 NEONMAP0(vextq_v),
6127 NEONMAP0(vfma_v),
6128 NEONMAP0(vfmaq_v),
6129 NEONMAP1(vfmlal_high_f16, aarch64_neon_fmlal2, 0),
6130 NEONMAP1(vfmlal_low_f16, aarch64_neon_fmlal, 0),
6131 NEONMAP1(vfmlalq_high_f16, aarch64_neon_fmlal2, 0),
6132 NEONMAP1(vfmlalq_low_f16, aarch64_neon_fmlal, 0),
6133 NEONMAP1(vfmlsl_high_f16, aarch64_neon_fmlsl2, 0),
6134 NEONMAP1(vfmlsl_low_f16, aarch64_neon_fmlsl, 0),
6135 NEONMAP1(vfmlslq_high_f16, aarch64_neon_fmlsl2, 0),
6136 NEONMAP1(vfmlslq_low_f16, aarch64_neon_fmlsl, 0),
6137 NEONMAP2(vhadd_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts),
6138 NEONMAP2(vhaddq_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts),
6139 NEONMAP2(vhsub_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts),
6140 NEONMAP2(vhsubq_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts),
6141 NEONMAP1(vld1_x2_v, aarch64_neon_ld1x2, 0),
6142 NEONMAP1(vld1_x3_v, aarch64_neon_ld1x3, 0),
6143 NEONMAP1(vld1_x4_v, aarch64_neon_ld1x4, 0),
6144 NEONMAP1(vld1q_x2_v, aarch64_neon_ld1x2, 0),
6145 NEONMAP1(vld1q_x3_v, aarch64_neon_ld1x3, 0),
6146 NEONMAP1(vld1q_x4_v, aarch64_neon_ld1x4, 0),
6147 NEONMAP1(vmmlaq_s32, aarch64_neon_smmla, 0),
6148 NEONMAP1(vmmlaq_u32, aarch64_neon_ummla, 0),
6149 NEONMAP0(vmovl_v),
6150 NEONMAP0(vmovn_v),
6151 NEONMAP1(vmul_v, aarch64_neon_pmul, Add1ArgType),
6152 NEONMAP1(vmulq_v, aarch64_neon_pmul, Add1ArgType),
6153 NEONMAP1(vpadd_v, aarch64_neon_addp, Add1ArgType),
6154 NEONMAP2(vpaddl_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts),
6155 NEONMAP2(vpaddlq_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts),
6156 NEONMAP1(vpaddq_v, aarch64_neon_addp, Add1ArgType),
6157 NEONMAP1(vqabs_v, aarch64_neon_sqabs, Add1ArgType),
6158 NEONMAP1(vqabsq_v, aarch64_neon_sqabs, Add1ArgType),
6159 NEONMAP2(vqadd_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts),
6160 NEONMAP2(vqaddq_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts),
6161 NEONMAP2(vqdmlal_v, aarch64_neon_sqdmull, aarch64_neon_sqadd, 0),
6162 NEONMAP2(vqdmlsl_v, aarch64_neon_sqdmull, aarch64_neon_sqsub, 0),
6163 NEONMAP1(vqdmulh_lane_v, aarch64_neon_sqdmulh_lane, 0),
6164 NEONMAP1(vqdmulh_laneq_v, aarch64_neon_sqdmulh_laneq, 0),
6165 NEONMAP1(vqdmulh_v, aarch64_neon_sqdmulh, Add1ArgType),
6166 NEONMAP1(vqdmulhq_lane_v, aarch64_neon_sqdmulh_lane, 0),
6167 NEONMAP1(vqdmulhq_laneq_v, aarch64_neon_sqdmulh_laneq, 0),
6168 NEONMAP1(vqdmulhq_v, aarch64_neon_sqdmulh, Add1ArgType),
6169 NEONMAP1(vqdmull_v, aarch64_neon_sqdmull, Add1ArgType),
6170 NEONMAP2(vqmovn_v, aarch64_neon_uqxtn, aarch64_neon_sqxtn, Add1ArgType | UnsignedAlts),
6171 NEONMAP1(vqmovun_v, aarch64_neon_sqxtun, Add1ArgType),
6172 NEONMAP1(vqneg_v, aarch64_neon_sqneg, Add1ArgType),
6173 NEONMAP1(vqnegq_v, aarch64_neon_sqneg, Add1ArgType),
6174 NEONMAP1(vqrdmlah_s16, aarch64_neon_sqrdmlah, Add1ArgType),
6175 NEONMAP1(vqrdmlah_s32, aarch64_neon_sqrdmlah, Add1ArgType),
6176 NEONMAP1(vqrdmlahq_s16, aarch64_neon_sqrdmlah, Add1ArgType),
6177 NEONMAP1(vqrdmlahq_s32, aarch64_neon_sqrdmlah, Add1ArgType),
6178 NEONMAP1(vqrdmlsh_s16, aarch64_neon_sqrdmlsh, Add1ArgType),
6179 NEONMAP1(vqrdmlsh_s32, aarch64_neon_sqrdmlsh, Add1ArgType),
6180 NEONMAP1(vqrdmlshq_s16, aarch64_neon_sqrdmlsh, Add1ArgType),
6181 NEONMAP1(vqrdmlshq_s32, aarch64_neon_sqrdmlsh, Add1ArgType),
6182 NEONMAP1(vqrdmulh_lane_v, aarch64_neon_sqrdmulh_lane, 0),
6183 NEONMAP1(vqrdmulh_laneq_v, aarch64_neon_sqrdmulh_laneq, 0),
6184 NEONMAP1(vqrdmulh_v, aarch64_neon_sqrdmulh, Add1ArgType),
6185 NEONMAP1(vqrdmulhq_lane_v, aarch64_neon_sqrdmulh_lane, 0),
6186 NEONMAP1(vqrdmulhq_laneq_v, aarch64_neon_sqrdmulh_laneq, 0),
6187 NEONMAP1(vqrdmulhq_v, aarch64_neon_sqrdmulh, Add1ArgType),
6188 NEONMAP2(vqrshl_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts),
6189 NEONMAP2(vqrshlq_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts),
6190 NEONMAP2(vqshl_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl, UnsignedAlts),
6191 NEONMAP2(vqshl_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts),
6192 NEONMAP2(vqshlq_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl,UnsignedAlts),
6193 NEONMAP2(vqshlq_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts),
6194 NEONMAP1(vqshlu_n_v, aarch64_neon_sqshlu, 0),
6195 NEONMAP1(vqshluq_n_v, aarch64_neon_sqshlu, 0),
6196 NEONMAP2(vqsub_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts),
6197 NEONMAP2(vqsubq_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts),
6198 NEONMAP1(vraddhn_v, aarch64_neon_raddhn, Add1ArgType),
6199 NEONMAP1(vrax1q_u64, aarch64_crypto_rax1, 0),
6200 NEONMAP2(vrecpe_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0),
6201 NEONMAP2(vrecpeq_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0),
6202 NEONMAP1(vrecps_v, aarch64_neon_frecps, Add1ArgType),
6203 NEONMAP1(vrecpsq_v, aarch64_neon_frecps, Add1ArgType),
6204 NEONMAP2(vrhadd_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts),
6205 NEONMAP2(vrhaddq_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts),
6206 NEONMAP1(vrnd32x_f32, aarch64_neon_frint32x, Add1ArgType),
6207 NEONMAP1(vrnd32xq_f32, aarch64_neon_frint32x, Add1ArgType),
6208 NEONMAP1(vrnd32z_f32, aarch64_neon_frint32z, Add1ArgType),
6209 NEONMAP1(vrnd32zq_f32, aarch64_neon_frint32z, Add1ArgType),
6210 NEONMAP1(vrnd64x_f32, aarch64_neon_frint64x, Add1ArgType),
6211 NEONMAP1(vrnd64xq_f32, aarch64_neon_frint64x, Add1ArgType),
6212 NEONMAP1(vrnd64z_f32, aarch64_neon_frint64z, Add1ArgType),
6213 NEONMAP1(vrnd64zq_f32, aarch64_neon_frint64z, Add1ArgType),
6214 NEONMAP0(vrndi_v),
6215 NEONMAP0(vrndiq_v),
6216 NEONMAP2(vrshl_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts),
6217 NEONMAP2(vrshlq_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts),
6218 NEONMAP2(vrshr_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts),
6219 NEONMAP2(vrshrq_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts),
6220 NEONMAP2(vrsqrte_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0),
6221 NEONMAP2(vrsqrteq_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0),
6222 NEONMAP1(vrsqrts_v, aarch64_neon_frsqrts, Add1ArgType),
6223 NEONMAP1(vrsqrtsq_v, aarch64_neon_frsqrts, Add1ArgType),
6224 NEONMAP1(vrsubhn_v, aarch64_neon_rsubhn, Add1ArgType),
6225 NEONMAP1(vsha1su0q_u32, aarch64_crypto_sha1su0, 0),
6226 NEONMAP1(vsha1su1q_u32, aarch64_crypto_sha1su1, 0),
6227 NEONMAP1(vsha256h2q_u32, aarch64_crypto_sha256h2, 0),
6228 NEONMAP1(vsha256hq_u32, aarch64_crypto_sha256h, 0),
6229 NEONMAP1(vsha256su0q_u32, aarch64_crypto_sha256su0, 0),
6230 NEONMAP1(vsha256su1q_u32, aarch64_crypto_sha256su1, 0),
6231 NEONMAP1(vsha512h2q_u64, aarch64_crypto_sha512h2, 0),
6232 NEONMAP1(vsha512hq_u64, aarch64_crypto_sha512h, 0),
6233 NEONMAP1(vsha512su0q_u64, aarch64_crypto_sha512su0, 0),
6234 NEONMAP1(vsha512su1q_u64, aarch64_crypto_sha512su1, 0),
6235 NEONMAP0(vshl_n_v),
6236 NEONMAP2(vshl_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts),
6237 NEONMAP0(vshll_n_v),
6238 NEONMAP0(vshlq_n_v),
6239 NEONMAP2(vshlq_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts),
6240 NEONMAP0(vshr_n_v),
6241 NEONMAP0(vshrn_n_v),
6242 NEONMAP0(vshrq_n_v),
6243 NEONMAP1(vsm3partw1q_u32, aarch64_crypto_sm3partw1, 0),
6244 NEONMAP1(vsm3partw2q_u32, aarch64_crypto_sm3partw2, 0),
6245 NEONMAP1(vsm3ss1q_u32, aarch64_crypto_sm3ss1, 0),
6246 NEONMAP1(vsm3tt1aq_u32, aarch64_crypto_sm3tt1a, 0),
6247 NEONMAP1(vsm3tt1bq_u32, aarch64_crypto_sm3tt1b, 0),
6248 NEONMAP1(vsm3tt2aq_u32, aarch64_crypto_sm3tt2a, 0),
6249 NEONMAP1(vsm3tt2bq_u32, aarch64_crypto_sm3tt2b, 0),
6250 NEONMAP1(vsm4ekeyq_u32, aarch64_crypto_sm4ekey, 0),
6251 NEONMAP1(vsm4eq_u32, aarch64_crypto_sm4e, 0),
6252 NEONMAP1(vst1_x2_v, aarch64_neon_st1x2, 0),
6253 NEONMAP1(vst1_x3_v, aarch64_neon_st1x3, 0),
6254 NEONMAP1(vst1_x4_v, aarch64_neon_st1x4, 0),
6255 NEONMAP1(vst1q_x2_v, aarch64_neon_st1x2, 0),
6256 NEONMAP1(vst1q_x3_v, aarch64_neon_st1x3, 0),
6257 NEONMAP1(vst1q_x4_v, aarch64_neon_st1x4, 0),
6258 NEONMAP0(vsubhn_v),
6259 NEONMAP0(vtst_v),
6260 NEONMAP0(vtstq_v),
6261 NEONMAP1(vusdot_s32, aarch64_neon_usdot, 0),
6262 NEONMAP1(vusdotq_s32, aarch64_neon_usdot, 0),
6263 NEONMAP1(vusmmlaq_s32, aarch64_neon_usmmla, 0),
6264 NEONMAP1(vxarq_u64, aarch64_crypto_xar, 0),
6265 };
6266
6267 static const ARMVectorIntrinsicInfo AArch64SISDIntrinsicMap[] = {
6268 NEONMAP1(vabdd_f64, aarch64_sisd_fabd, Add1ArgType),
6269 NEONMAP1(vabds_f32, aarch64_sisd_fabd, Add1ArgType),
6270 NEONMAP1(vabsd_s64, aarch64_neon_abs, Add1ArgType),
6271 NEONMAP1(vaddlv_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType),
6272 NEONMAP1(vaddlv_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType),
6273 NEONMAP1(vaddlvq_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType),
6274 NEONMAP1(vaddlvq_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType),
6275 NEONMAP1(vaddv_f32, aarch64_neon_faddv, AddRetType | Add1ArgType),
6276 NEONMAP1(vaddv_s32, aarch64_neon_saddv, AddRetType | Add1ArgType),
6277 NEONMAP1(vaddv_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType),
6278 NEONMAP1(vaddvq_f32, aarch64_neon_faddv, AddRetType | Add1ArgType),
6279 NEONMAP1(vaddvq_f64, aarch64_neon_faddv, AddRetType | Add1ArgType),
6280 NEONMAP1(vaddvq_s32, aarch64_neon_saddv, AddRetType | Add1ArgType),
6281 NEONMAP1(vaddvq_s64, aarch64_neon_saddv, AddRetType | Add1ArgType),
6282 NEONMAP1(vaddvq_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType),
6283 NEONMAP1(vaddvq_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType),
6284 NEONMAP1(vcaged_f64, aarch64_neon_facge, AddRetType | Add1ArgType),
6285 NEONMAP1(vcages_f32, aarch64_neon_facge, AddRetType | Add1ArgType),
6286 NEONMAP1(vcagtd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType),
6287 NEONMAP1(vcagts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType),
6288 NEONMAP1(vcaled_f64, aarch64_neon_facge, AddRetType | Add1ArgType),
6289 NEONMAP1(vcales_f32, aarch64_neon_facge, AddRetType | Add1ArgType),
6290 NEONMAP1(vcaltd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType),
6291 NEONMAP1(vcalts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType),
6292 NEONMAP1(vcvtad_s64_f64, aarch64_neon_fcvtas, AddRetType | Add1ArgType),
6293 NEONMAP1(vcvtad_u64_f64, aarch64_neon_fcvtau, AddRetType | Add1ArgType),
6294 NEONMAP1(vcvtas_s32_f32, aarch64_neon_fcvtas, AddRetType | Add1ArgType),
6295 NEONMAP1(vcvtas_u32_f32, aarch64_neon_fcvtau, AddRetType | Add1ArgType),
6296 NEONMAP1(vcvtd_n_f64_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType),
6297 NEONMAP1(vcvtd_n_f64_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType),
6298 NEONMAP1(vcvtd_n_s64_f64, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType),
6299 NEONMAP1(vcvtd_n_u64_f64, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType),
6300 NEONMAP1(vcvtd_s64_f64, aarch64_neon_fcvtzs, AddRetType | Add1ArgType),
6301 NEONMAP1(vcvtd_u64_f64, aarch64_neon_fcvtzu, AddRetType | Add1ArgType),
6302 NEONMAP1(vcvth_bf16_f32, aarch64_neon_bfcvt, 0),
6303 NEONMAP1(vcvtmd_s64_f64, aarch64_neon_fcvtms, AddRetType | Add1ArgType),
6304 NEONMAP1(vcvtmd_u64_f64, aarch64_neon_fcvtmu, AddRetType | Add1ArgType),
6305 NEONMAP1(vcvtms_s32_f32, aarch64_neon_fcvtms, AddRetType | Add1ArgType),
6306 NEONMAP1(vcvtms_u32_f32, aarch64_neon_fcvtmu, AddRetType | Add1ArgType),
6307 NEONMAP1(vcvtnd_s64_f64, aarch64_neon_fcvtns, AddRetType | Add1ArgType),
6308 NEONMAP1(vcvtnd_u64_f64, aarch64_neon_fcvtnu, AddRetType | Add1ArgType),
6309 NEONMAP1(vcvtns_s32_f32, aarch64_neon_fcvtns, AddRetType | Add1ArgType),
6310 NEONMAP1(vcvtns_u32_f32, aarch64_neon_fcvtnu, AddRetType | Add1ArgType),
6311 NEONMAP1(vcvtpd_s64_f64, aarch64_neon_fcvtps, AddRetType | Add1ArgType),
6312 NEONMAP1(vcvtpd_u64_f64, aarch64_neon_fcvtpu, AddRetType | Add1ArgType),
6313 NEONMAP1(vcvtps_s32_f32, aarch64_neon_fcvtps, AddRetType | Add1ArgType),
6314 NEONMAP1(vcvtps_u32_f32, aarch64_neon_fcvtpu, AddRetType | Add1ArgType),
6315 NEONMAP1(vcvts_n_f32_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType),
6316 NEONMAP1(vcvts_n_f32_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType),
6317 NEONMAP1(vcvts_n_s32_f32, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType),
6318 NEONMAP1(vcvts_n_u32_f32, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType),
6319 NEONMAP1(vcvts_s32_f32, aarch64_neon_fcvtzs, AddRetType | Add1ArgType),
6320 NEONMAP1(vcvts_u32_f32, aarch64_neon_fcvtzu, AddRetType | Add1ArgType),
6321 NEONMAP1(vcvtxd_f32_f64, aarch64_sisd_fcvtxn, 0),
6322 NEONMAP1(vmaxnmv_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
6323 NEONMAP1(vmaxnmvq_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
6324 NEONMAP1(vmaxnmvq_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
6325 NEONMAP1(vmaxv_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
6326 NEONMAP1(vmaxv_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType),
6327 NEONMAP1(vmaxv_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType),
6328 NEONMAP1(vmaxvq_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
6329 NEONMAP1(vmaxvq_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
6330 NEONMAP1(vmaxvq_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType),
6331 NEONMAP1(vmaxvq_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType),
6332 NEONMAP1(vminnmv_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
6333 NEONMAP1(vminnmvq_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
6334 NEONMAP1(vminnmvq_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
6335 NEONMAP1(vminv_f32, aarch64_neon_fminv, AddRetType | Add1ArgType),
6336 NEONMAP1(vminv_s32, aarch64_neon_sminv, AddRetType | Add1ArgType),
6337 NEONMAP1(vminv_u32, aarch64_neon_uminv, AddRetType | Add1ArgType),
6338 NEONMAP1(vminvq_f32, aarch64_neon_fminv, AddRetType | Add1ArgType),
6339 NEONMAP1(vminvq_f64, aarch64_neon_fminv, AddRetType | Add1ArgType),
6340 NEONMAP1(vminvq_s32, aarch64_neon_sminv, AddRetType | Add1ArgType),
6341 NEONMAP1(vminvq_u32, aarch64_neon_uminv, AddRetType | Add1ArgType),
6342 NEONMAP1(vmull_p64, aarch64_neon_pmull64, 0),
6343 NEONMAP1(vmulxd_f64, aarch64_neon_fmulx, Add1ArgType),
6344 NEONMAP1(vmulxs_f32, aarch64_neon_fmulx, Add1ArgType),
6345 NEONMAP1(vpaddd_s64, aarch64_neon_uaddv, AddRetType | Add1ArgType),
6346 NEONMAP1(vpaddd_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType),
6347 NEONMAP1(vpmaxnmqd_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
6348 NEONMAP1(vpmaxnms_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType),
6349 NEONMAP1(vpmaxqd_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
6350 NEONMAP1(vpmaxs_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType),
6351 NEONMAP1(vpminnmqd_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
6352 NEONMAP1(vpminnms_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType),
6353 NEONMAP1(vpminqd_f64, aarch64_neon_fminv, AddRetType | Add1ArgType),
6354 NEONMAP1(vpmins_f32, aarch64_neon_fminv, AddRetType | Add1ArgType),
6355 NEONMAP1(vqabsb_s8, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors),
6356 NEONMAP1(vqabsd_s64, aarch64_neon_sqabs, Add1ArgType),
6357 NEONMAP1(vqabsh_s16, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors),
6358 NEONMAP1(vqabss_s32, aarch64_neon_sqabs, Add1ArgType),
6359 NEONMAP1(vqaddb_s8, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors),
6360 NEONMAP1(vqaddb_u8, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors),
6361 NEONMAP1(vqaddd_s64, aarch64_neon_sqadd, Add1ArgType),
6362 NEONMAP1(vqaddd_u64, aarch64_neon_uqadd, Add1ArgType),
6363 NEONMAP1(vqaddh_s16, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors),
6364 NEONMAP1(vqaddh_u16, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors),
6365 NEONMAP1(vqadds_s32, aarch64_neon_sqadd, Add1ArgType),
6366 NEONMAP1(vqadds_u32, aarch64_neon_uqadd, Add1ArgType),
6367 NEONMAP1(vqdmulhh_s16, aarch64_neon_sqdmulh, Vectorize1ArgType | Use64BitVectors),
6368 NEONMAP1(vqdmulhs_s32, aarch64_neon_sqdmulh, Add1ArgType),
6369 NEONMAP1(vqdmullh_s16, aarch64_neon_sqdmull, VectorRet | Use128BitVectors),
6370 NEONMAP1(vqdmulls_s32, aarch64_neon_sqdmulls_scalar, 0),
6371 NEONMAP1(vqmovnd_s64, aarch64_neon_scalar_sqxtn, AddRetType | Add1ArgType),
6372 NEONMAP1(vqmovnd_u64, aarch64_neon_scalar_uqxtn, AddRetType | Add1ArgType),
6373 NEONMAP1(vqmovnh_s16, aarch64_neon_sqxtn, VectorRet | Use64BitVectors),
6374 NEONMAP1(vqmovnh_u16, aarch64_neon_uqxtn, VectorRet | Use64BitVectors),
6375 NEONMAP1(vqmovns_s32, aarch64_neon_sqxtn, VectorRet | Use64BitVectors),
6376 NEONMAP1(vqmovns_u32, aarch64_neon_uqxtn, VectorRet | Use64BitVectors),
6377 NEONMAP1(vqmovund_s64, aarch64_neon_scalar_sqxtun, AddRetType | Add1ArgType),
6378 NEONMAP1(vqmovunh_s16, aarch64_neon_sqxtun, VectorRet | Use64BitVectors),
6379 NEONMAP1(vqmovuns_s32, aarch64_neon_sqxtun, VectorRet | Use64BitVectors),
6380 NEONMAP1(vqnegb_s8, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors),
6381 NEONMAP1(vqnegd_s64, aarch64_neon_sqneg, Add1ArgType),
6382 NEONMAP1(vqnegh_s16, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors),
6383 NEONMAP1(vqnegs_s32, aarch64_neon_sqneg, Add1ArgType),
6384 NEONMAP1(vqrdmlahh_s16, aarch64_neon_sqrdmlah, Vectorize1ArgType | Use64BitVectors),
6385 NEONMAP1(vqrdmlahs_s32, aarch64_neon_sqrdmlah, Add1ArgType),
6386 NEONMAP1(vqrdmlshh_s16, aarch64_neon_sqrdmlsh, Vectorize1ArgType | Use64BitVectors),
6387 NEONMAP1(vqrdmlshs_s32, aarch64_neon_sqrdmlsh, Add1ArgType),
6388 NEONMAP1(vqrdmulhh_s16, aarch64_neon_sqrdmulh, Vectorize1ArgType | Use64BitVectors),
6389 NEONMAP1(vqrdmulhs_s32, aarch64_neon_sqrdmulh, Add1ArgType),
6390 NEONMAP1(vqrshlb_s8, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors),
6391 NEONMAP1(vqrshlb_u8, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors),
6392 NEONMAP1(vqrshld_s64, aarch64_neon_sqrshl, Add1ArgType),
6393 NEONMAP1(vqrshld_u64, aarch64_neon_uqrshl, Add1ArgType),
6394 NEONMAP1(vqrshlh_s16, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors),
6395 NEONMAP1(vqrshlh_u16, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors),
6396 NEONMAP1(vqrshls_s32, aarch64_neon_sqrshl, Add1ArgType),
6397 NEONMAP1(vqrshls_u32, aarch64_neon_uqrshl, Add1ArgType),
6398 NEONMAP1(vqrshrnd_n_s64, aarch64_neon_sqrshrn, AddRetType),
6399 NEONMAP1(vqrshrnd_n_u64, aarch64_neon_uqrshrn, AddRetType),
6400 NEONMAP1(vqrshrnh_n_s16, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors),
6401 NEONMAP1(vqrshrnh_n_u16, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors),
6402 NEONMAP1(vqrshrns_n_s32, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors),
6403 NEONMAP1(vqrshrns_n_u32, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors),
6404 NEONMAP1(vqrshrund_n_s64, aarch64_neon_sqrshrun, AddRetType),
6405 NEONMAP1(vqrshrunh_n_s16, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors),
6406 NEONMAP1(vqrshruns_n_s32, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors),
6407 NEONMAP1(vqshlb_n_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors),
6408 NEONMAP1(vqshlb_n_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors),
6409 NEONMAP1(vqshlb_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors),
6410 NEONMAP1(vqshlb_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors),
6411 NEONMAP1(vqshld_s64, aarch64_neon_sqshl, Add1ArgType),
6412 NEONMAP1(vqshld_u64, aarch64_neon_uqshl, Add1ArgType),
6413 NEONMAP1(vqshlh_n_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors),
6414 NEONMAP1(vqshlh_n_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors),
6415 NEONMAP1(vqshlh_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors),
6416 NEONMAP1(vqshlh_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors),
6417 NEONMAP1(vqshls_n_s32, aarch64_neon_sqshl, Add1ArgType),
6418 NEONMAP1(vqshls_n_u32, aarch64_neon_uqshl, Add1ArgType),
6419 NEONMAP1(vqshls_s32, aarch64_neon_sqshl, Add1ArgType),
6420 NEONMAP1(vqshls_u32, aarch64_neon_uqshl, Add1ArgType),
6421 NEONMAP1(vqshlub_n_s8, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors),
6422 NEONMAP1(vqshluh_n_s16, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors),
6423 NEONMAP1(vqshlus_n_s32, aarch64_neon_sqshlu, Add1ArgType),
6424 NEONMAP1(vqshrnd_n_s64, aarch64_neon_sqshrn, AddRetType),
6425 NEONMAP1(vqshrnd_n_u64, aarch64_neon_uqshrn, AddRetType),
6426 NEONMAP1(vqshrnh_n_s16, aarch64_neon_sqshrn, VectorRet | Use64BitVectors),
6427 NEONMAP1(vqshrnh_n_u16, aarch64_neon_uqshrn, VectorRet | Use64BitVectors),
6428 NEONMAP1(vqshrns_n_s32, aarch64_neon_sqshrn, VectorRet | Use64BitVectors),
6429 NEONMAP1(vqshrns_n_u32, aarch64_neon_uqshrn, VectorRet | Use64BitVectors),
6430 NEONMAP1(vqshrund_n_s64, aarch64_neon_sqshrun, AddRetType),
6431 NEONMAP1(vqshrunh_n_s16, aarch64_neon_sqshrun, VectorRet | Use64BitVectors),
6432 NEONMAP1(vqshruns_n_s32, aarch64_neon_sqshrun, VectorRet | Use64BitVectors),
6433 NEONMAP1(vqsubb_s8, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors),
6434 NEONMAP1(vqsubb_u8, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors),
6435 NEONMAP1(vqsubd_s64, aarch64_neon_sqsub, Add1ArgType),
6436 NEONMAP1(vqsubd_u64, aarch64_neon_uqsub, Add1ArgType),
6437 NEONMAP1(vqsubh_s16, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors),
6438 NEONMAP1(vqsubh_u16, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors),
6439 NEONMAP1(vqsubs_s32, aarch64_neon_sqsub, Add1ArgType),
6440 NEONMAP1(vqsubs_u32, aarch64_neon_uqsub, Add1ArgType),
6441 NEONMAP1(vrecped_f64, aarch64_neon_frecpe, Add1ArgType),
6442 NEONMAP1(vrecpes_f32, aarch64_neon_frecpe, Add1ArgType),
6443 NEONMAP1(vrecpxd_f64, aarch64_neon_frecpx, Add1ArgType),
6444 NEONMAP1(vrecpxs_f32, aarch64_neon_frecpx, Add1ArgType),
6445 NEONMAP1(vrshld_s64, aarch64_neon_srshl, Add1ArgType),
6446 NEONMAP1(vrshld_u64, aarch64_neon_urshl, Add1ArgType),
6447 NEONMAP1(vrsqrted_f64, aarch64_neon_frsqrte, Add1ArgType),
6448 NEONMAP1(vrsqrtes_f32, aarch64_neon_frsqrte, Add1ArgType),
6449 NEONMAP1(vrsqrtsd_f64, aarch64_neon_frsqrts, Add1ArgType),
6450 NEONMAP1(vrsqrtss_f32, aarch64_neon_frsqrts, Add1ArgType),
6451 NEONMAP1(vsha1cq_u32, aarch64_crypto_sha1c, 0),
6452 NEONMAP1(vsha1h_u32, aarch64_crypto_sha1h, 0),
6453 NEONMAP1(vsha1mq_u32, aarch64_crypto_sha1m, 0),
6454 NEONMAP1(vsha1pq_u32, aarch64_crypto_sha1p, 0),
6455 NEONMAP1(vshld_s64, aarch64_neon_sshl, Add1ArgType),
6456 NEONMAP1(vshld_u64, aarch64_neon_ushl, Add1ArgType),
6457 NEONMAP1(vslid_n_s64, aarch64_neon_vsli, Vectorize1ArgType),
6458 NEONMAP1(vslid_n_u64, aarch64_neon_vsli, Vectorize1ArgType),
6459 NEONMAP1(vsqaddb_u8, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors),
6460 NEONMAP1(vsqaddd_u64, aarch64_neon_usqadd, Add1ArgType),
6461 NEONMAP1(vsqaddh_u16, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors),
6462 NEONMAP1(vsqadds_u32, aarch64_neon_usqadd, Add1ArgType),
6463 NEONMAP1(vsrid_n_s64, aarch64_neon_vsri, Vectorize1ArgType),
6464 NEONMAP1(vsrid_n_u64, aarch64_neon_vsri, Vectorize1ArgType),
6465 NEONMAP1(vuqaddb_s8, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors),
6466 NEONMAP1(vuqaddd_s64, aarch64_neon_suqadd, Add1ArgType),
6467 NEONMAP1(vuqaddh_s16, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors),
6468 NEONMAP1(vuqadds_s32, aarch64_neon_suqadd, Add1ArgType),
6469 // FP16 scalar intrinisics go here.
6470 NEONMAP1(vabdh_f16, aarch64_sisd_fabd, Add1ArgType),
6471 NEONMAP1(vcvtah_s32_f16, aarch64_neon_fcvtas, AddRetType | Add1ArgType),
6472 NEONMAP1(vcvtah_s64_f16, aarch64_neon_fcvtas, AddRetType | Add1ArgType),
6473 NEONMAP1(vcvtah_u32_f16, aarch64_neon_fcvtau, AddRetType | Add1ArgType),
6474 NEONMAP1(vcvtah_u64_f16, aarch64_neon_fcvtau, AddRetType | Add1ArgType),
6475 NEONMAP1(vcvth_n_f16_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType),
6476 NEONMAP1(vcvth_n_f16_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType),
6477 NEONMAP1(vcvth_n_f16_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType),
6478 NEONMAP1(vcvth_n_f16_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType),
6479 NEONMAP1(vcvth_n_s32_f16, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType),
6480 NEONMAP1(vcvth_n_s64_f16, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType),
6481 NEONMAP1(vcvth_n_u32_f16, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType),
6482 NEONMAP1(vcvth_n_u64_f16, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType),
6483 NEONMAP1(vcvth_s32_f16, aarch64_neon_fcvtzs, AddRetType | Add1ArgType),
6484 NEONMAP1(vcvth_s64_f16, aarch64_neon_fcvtzs, AddRetType | Add1ArgType),
6485 NEONMAP1(vcvth_u32_f16, aarch64_neon_fcvtzu, AddRetType | Add1ArgType),
6486 NEONMAP1(vcvth_u64_f16, aarch64_neon_fcvtzu, AddRetType | Add1ArgType),
6487 NEONMAP1(vcvtmh_s32_f16, aarch64_neon_fcvtms, AddRetType | Add1ArgType),
6488 NEONMAP1(vcvtmh_s64_f16, aarch64_neon_fcvtms, AddRetType | Add1ArgType),
6489 NEONMAP1(vcvtmh_u32_f16, aarch64_neon_fcvtmu, AddRetType | Add1ArgType),
6490 NEONMAP1(vcvtmh_u64_f16, aarch64_neon_fcvtmu, AddRetType | Add1ArgType),
6491 NEONMAP1(vcvtnh_s32_f16, aarch64_neon_fcvtns, AddRetType | Add1ArgType),
6492 NEONMAP1(vcvtnh_s64_f16, aarch64_neon_fcvtns, AddRetType | Add1ArgType),
6493 NEONMAP1(vcvtnh_u32_f16, aarch64_neon_fcvtnu, AddRetType | Add1ArgType),
6494 NEONMAP1(vcvtnh_u64_f16, aarch64_neon_fcvtnu, AddRetType | Add1ArgType),
6495 NEONMAP1(vcvtph_s32_f16, aarch64_neon_fcvtps, AddRetType | Add1ArgType),
6496 NEONMAP1(vcvtph_s64_f16, aarch64_neon_fcvtps, AddRetType | Add1ArgType),
6497 NEONMAP1(vcvtph_u32_f16, aarch64_neon_fcvtpu, AddRetType | Add1ArgType),
6498 NEONMAP1(vcvtph_u64_f16, aarch64_neon_fcvtpu, AddRetType | Add1ArgType),
6499 NEONMAP1(vmulxh_f16, aarch64_neon_fmulx, Add1ArgType),
6500 NEONMAP1(vrecpeh_f16, aarch64_neon_frecpe, Add1ArgType),
6501 NEONMAP1(vrecpxh_f16, aarch64_neon_frecpx, Add1ArgType),
6502 NEONMAP1(vrsqrteh_f16, aarch64_neon_frsqrte, Add1ArgType),
6503 NEONMAP1(vrsqrtsh_f16, aarch64_neon_frsqrts, Add1ArgType),
6504 };
6505
6506 // Some intrinsics are equivalent for codegen.
6507 static const std::pair<unsigned, unsigned> NEONEquivalentIntrinsicMap[] = {
6508 { NEON::BI__builtin_neon_splat_lane_bf16, NEON::BI__builtin_neon_splat_lane_v, },
6509 { NEON::BI__builtin_neon_splat_laneq_bf16, NEON::BI__builtin_neon_splat_laneq_v, },
6510 { NEON::BI__builtin_neon_splatq_lane_bf16, NEON::BI__builtin_neon_splatq_lane_v, },
6511 { NEON::BI__builtin_neon_splatq_laneq_bf16, NEON::BI__builtin_neon_splatq_laneq_v, },
6512 { NEON::BI__builtin_neon_vabd_f16, NEON::BI__builtin_neon_vabd_v, },
6513 { NEON::BI__builtin_neon_vabdq_f16, NEON::BI__builtin_neon_vabdq_v, },
6514 { NEON::BI__builtin_neon_vabs_f16, NEON::BI__builtin_neon_vabs_v, },
6515 { NEON::BI__builtin_neon_vabsq_f16, NEON::BI__builtin_neon_vabsq_v, },
6516 { NEON::BI__builtin_neon_vbsl_f16, NEON::BI__builtin_neon_vbsl_v, },
6517 { NEON::BI__builtin_neon_vbslq_f16, NEON::BI__builtin_neon_vbslq_v, },
6518 { NEON::BI__builtin_neon_vcage_f16, NEON::BI__builtin_neon_vcage_v, },
6519 { NEON::BI__builtin_neon_vcageq_f16, NEON::BI__builtin_neon_vcageq_v, },
6520 { NEON::BI__builtin_neon_vcagt_f16, NEON::BI__builtin_neon_vcagt_v, },
6521 { NEON::BI__builtin_neon_vcagtq_f16, NEON::BI__builtin_neon_vcagtq_v, },
6522 { NEON::BI__builtin_neon_vcale_f16, NEON::BI__builtin_neon_vcale_v, },
6523 { NEON::BI__builtin_neon_vcaleq_f16, NEON::BI__builtin_neon_vcaleq_v, },
6524 { NEON::BI__builtin_neon_vcalt_f16, NEON::BI__builtin_neon_vcalt_v, },
6525 { NEON::BI__builtin_neon_vcaltq_f16, NEON::BI__builtin_neon_vcaltq_v, },
6526 { NEON::BI__builtin_neon_vceqz_f16, NEON::BI__builtin_neon_vceqz_v, },
6527 { NEON::BI__builtin_neon_vceqzq_f16, NEON::BI__builtin_neon_vceqzq_v, },
6528 { NEON::BI__builtin_neon_vcgez_f16, NEON::BI__builtin_neon_vcgez_v, },
6529 { NEON::BI__builtin_neon_vcgezq_f16, NEON::BI__builtin_neon_vcgezq_v, },
6530 { NEON::BI__builtin_neon_vcgtz_f16, NEON::BI__builtin_neon_vcgtz_v, },
6531 { NEON::BI__builtin_neon_vcgtzq_f16, NEON::BI__builtin_neon_vcgtzq_v, },
6532 { NEON::BI__builtin_neon_vclez_f16, NEON::BI__builtin_neon_vclez_v, },
6533 { NEON::BI__builtin_neon_vclezq_f16, NEON::BI__builtin_neon_vclezq_v, },
6534 { NEON::BI__builtin_neon_vcltz_f16, NEON::BI__builtin_neon_vcltz_v, },
6535 { NEON::BI__builtin_neon_vcltzq_f16, NEON::BI__builtin_neon_vcltzq_v, },
6536 { NEON::BI__builtin_neon_vext_f16, NEON::BI__builtin_neon_vext_v, },
6537 { NEON::BI__builtin_neon_vextq_f16, NEON::BI__builtin_neon_vextq_v, },
6538 { NEON::BI__builtin_neon_vfma_f16, NEON::BI__builtin_neon_vfma_v, },
6539 { NEON::BI__builtin_neon_vfma_lane_f16, NEON::BI__builtin_neon_vfma_lane_v, },
6540 { NEON::BI__builtin_neon_vfma_laneq_f16, NEON::BI__builtin_neon_vfma_laneq_v, },
6541 { NEON::BI__builtin_neon_vfmaq_f16, NEON::BI__builtin_neon_vfmaq_v, },
6542 { NEON::BI__builtin_neon_vfmaq_lane_f16, NEON::BI__builtin_neon_vfmaq_lane_v, },
6543 { NEON::BI__builtin_neon_vfmaq_laneq_f16, NEON::BI__builtin_neon_vfmaq_laneq_v, },
6544 { NEON::BI__builtin_neon_vld1_bf16_x2, NEON::BI__builtin_neon_vld1_x2_v },
6545 { NEON::BI__builtin_neon_vld1_bf16_x3, NEON::BI__builtin_neon_vld1_x3_v },
6546 { NEON::BI__builtin_neon_vld1_bf16_x4, NEON::BI__builtin_neon_vld1_x4_v },
6547 { NEON::BI__builtin_neon_vld1_bf16, NEON::BI__builtin_neon_vld1_v },
6548 { NEON::BI__builtin_neon_vld1_dup_bf16, NEON::BI__builtin_neon_vld1_dup_v },
6549 { NEON::BI__builtin_neon_vld1_lane_bf16, NEON::BI__builtin_neon_vld1_lane_v },
6550 { NEON::BI__builtin_neon_vld1q_bf16_x2, NEON::BI__builtin_neon_vld1q_x2_v },
6551 { NEON::BI__builtin_neon_vld1q_bf16_x3, NEON::BI__builtin_neon_vld1q_x3_v },
6552 { NEON::BI__builtin_neon_vld1q_bf16_x4, NEON::BI__builtin_neon_vld1q_x4_v },
6553 { NEON::BI__builtin_neon_vld1q_bf16, NEON::BI__builtin_neon_vld1q_v },
6554 { NEON::BI__builtin_neon_vld1q_dup_bf16, NEON::BI__builtin_neon_vld1q_dup_v },
6555 { NEON::BI__builtin_neon_vld1q_lane_bf16, NEON::BI__builtin_neon_vld1q_lane_v },
6556 { NEON::BI__builtin_neon_vld2_bf16, NEON::BI__builtin_neon_vld2_v },
6557 { NEON::BI__builtin_neon_vld2_dup_bf16, NEON::BI__builtin_neon_vld2_dup_v },
6558 { NEON::BI__builtin_neon_vld2_lane_bf16, NEON::BI__builtin_neon_vld2_lane_v },
6559 { NEON::BI__builtin_neon_vld2q_bf16, NEON::BI__builtin_neon_vld2q_v },
6560 { NEON::BI__builtin_neon_vld2q_dup_bf16, NEON::BI__builtin_neon_vld2q_dup_v },
6561 { NEON::BI__builtin_neon_vld2q_lane_bf16, NEON::BI__builtin_neon_vld2q_lane_v },
6562 { NEON::BI__builtin_neon_vld3_bf16, NEON::BI__builtin_neon_vld3_v },
6563 { NEON::BI__builtin_neon_vld3_dup_bf16, NEON::BI__builtin_neon_vld3_dup_v },
6564 { NEON::BI__builtin_neon_vld3_lane_bf16, NEON::BI__builtin_neon_vld3_lane_v },
6565 { NEON::BI__builtin_neon_vld3q_bf16, NEON::BI__builtin_neon_vld3q_v },
6566 { NEON::BI__builtin_neon_vld3q_dup_bf16, NEON::BI__builtin_neon_vld3q_dup_v },
6567 { NEON::BI__builtin_neon_vld3q_lane_bf16, NEON::BI__builtin_neon_vld3q_lane_v },
6568 { NEON::BI__builtin_neon_vld4_bf16, NEON::BI__builtin_neon_vld4_v },
6569 { NEON::BI__builtin_neon_vld4_dup_bf16, NEON::BI__builtin_neon_vld4_dup_v },
6570 { NEON::BI__builtin_neon_vld4_lane_bf16, NEON::BI__builtin_neon_vld4_lane_v },
6571 { NEON::BI__builtin_neon_vld4q_bf16, NEON::BI__builtin_neon_vld4q_v },
6572 { NEON::BI__builtin_neon_vld4q_dup_bf16, NEON::BI__builtin_neon_vld4q_dup_v },
6573 { NEON::BI__builtin_neon_vld4q_lane_bf16, NEON::BI__builtin_neon_vld4q_lane_v },
6574 { NEON::BI__builtin_neon_vmax_f16, NEON::BI__builtin_neon_vmax_v, },
6575 { NEON::BI__builtin_neon_vmaxnm_f16, NEON::BI__builtin_neon_vmaxnm_v, },
6576 { NEON::BI__builtin_neon_vmaxnmq_f16, NEON::BI__builtin_neon_vmaxnmq_v, },
6577 { NEON::BI__builtin_neon_vmaxq_f16, NEON::BI__builtin_neon_vmaxq_v, },
6578 { NEON::BI__builtin_neon_vmin_f16, NEON::BI__builtin_neon_vmin_v, },
6579 { NEON::BI__builtin_neon_vminnm_f16, NEON::BI__builtin_neon_vminnm_v, },
6580 { NEON::BI__builtin_neon_vminnmq_f16, NEON::BI__builtin_neon_vminnmq_v, },
6581 { NEON::BI__builtin_neon_vminq_f16, NEON::BI__builtin_neon_vminq_v, },
6582 { NEON::BI__builtin_neon_vmulx_f16, NEON::BI__builtin_neon_vmulx_v, },
6583 { NEON::BI__builtin_neon_vmulxq_f16, NEON::BI__builtin_neon_vmulxq_v, },
6584 { NEON::BI__builtin_neon_vpadd_f16, NEON::BI__builtin_neon_vpadd_v, },
6585 { NEON::BI__builtin_neon_vpaddq_f16, NEON::BI__builtin_neon_vpaddq_v, },
6586 { NEON::BI__builtin_neon_vpmax_f16, NEON::BI__builtin_neon_vpmax_v, },
6587 { NEON::BI__builtin_neon_vpmaxnm_f16, NEON::BI__builtin_neon_vpmaxnm_v, },
6588 { NEON::BI__builtin_neon_vpmaxnmq_f16, NEON::BI__builtin_neon_vpmaxnmq_v, },
6589 { NEON::BI__builtin_neon_vpmaxq_f16, NEON::BI__builtin_neon_vpmaxq_v, },
6590 { NEON::BI__builtin_neon_vpmin_f16, NEON::BI__builtin_neon_vpmin_v, },
6591 { NEON::BI__builtin_neon_vpminnm_f16, NEON::BI__builtin_neon_vpminnm_v, },
6592 { NEON::BI__builtin_neon_vpminnmq_f16, NEON::BI__builtin_neon_vpminnmq_v, },
6593 { NEON::BI__builtin_neon_vpminq_f16, NEON::BI__builtin_neon_vpminq_v, },
6594 { NEON::BI__builtin_neon_vrecpe_f16, NEON::BI__builtin_neon_vrecpe_v, },
6595 { NEON::BI__builtin_neon_vrecpeq_f16, NEON::BI__builtin_neon_vrecpeq_v, },
6596 { NEON::BI__builtin_neon_vrecps_f16, NEON::BI__builtin_neon_vrecps_v, },
6597 { NEON::BI__builtin_neon_vrecpsq_f16, NEON::BI__builtin_neon_vrecpsq_v, },
6598 { NEON::BI__builtin_neon_vrnd_f16, NEON::BI__builtin_neon_vrnd_v, },
6599 { NEON::BI__builtin_neon_vrnda_f16, NEON::BI__builtin_neon_vrnda_v, },
6600 { NEON::BI__builtin_neon_vrndaq_f16, NEON::BI__builtin_neon_vrndaq_v, },
6601 { NEON::BI__builtin_neon_vrndi_f16, NEON::BI__builtin_neon_vrndi_v, },
6602 { NEON::BI__builtin_neon_vrndiq_f16, NEON::BI__builtin_neon_vrndiq_v, },
6603 { NEON::BI__builtin_neon_vrndm_f16, NEON::BI__builtin_neon_vrndm_v, },
6604 { NEON::BI__builtin_neon_vrndmq_f16, NEON::BI__builtin_neon_vrndmq_v, },
6605 { NEON::BI__builtin_neon_vrndn_f16, NEON::BI__builtin_neon_vrndn_v, },
6606 { NEON::BI__builtin_neon_vrndnq_f16, NEON::BI__builtin_neon_vrndnq_v, },
6607 { NEON::BI__builtin_neon_vrndp_f16, NEON::BI__builtin_neon_vrndp_v, },
6608 { NEON::BI__builtin_neon_vrndpq_f16, NEON::BI__builtin_neon_vrndpq_v, },
6609 { NEON::BI__builtin_neon_vrndq_f16, NEON::BI__builtin_neon_vrndq_v, },
6610 { NEON::BI__builtin_neon_vrndx_f16, NEON::BI__builtin_neon_vrndx_v, },
6611 { NEON::BI__builtin_neon_vrndxq_f16, NEON::BI__builtin_neon_vrndxq_v, },
6612 { NEON::BI__builtin_neon_vrsqrte_f16, NEON::BI__builtin_neon_vrsqrte_v, },
6613 { NEON::BI__builtin_neon_vrsqrteq_f16, NEON::BI__builtin_neon_vrsqrteq_v, },
6614 { NEON::BI__builtin_neon_vrsqrts_f16, NEON::BI__builtin_neon_vrsqrts_v, },
6615 { NEON::BI__builtin_neon_vrsqrtsq_f16, NEON::BI__builtin_neon_vrsqrtsq_v, },
6616 { NEON::BI__builtin_neon_vsqrt_f16, NEON::BI__builtin_neon_vsqrt_v, },
6617 { NEON::BI__builtin_neon_vsqrtq_f16, NEON::BI__builtin_neon_vsqrtq_v, },
6618 { NEON::BI__builtin_neon_vst1_bf16_x2, NEON::BI__builtin_neon_vst1_x2_v },
6619 { NEON::BI__builtin_neon_vst1_bf16_x3, NEON::BI__builtin_neon_vst1_x3_v },
6620 { NEON::BI__builtin_neon_vst1_bf16_x4, NEON::BI__builtin_neon_vst1_x4_v },
6621 { NEON::BI__builtin_neon_vst1_bf16, NEON::BI__builtin_neon_vst1_v },
6622 { NEON::BI__builtin_neon_vst1_lane_bf16, NEON::BI__builtin_neon_vst1_lane_v },
6623 { NEON::BI__builtin_neon_vst1q_bf16_x2, NEON::BI__builtin_neon_vst1q_x2_v },
6624 { NEON::BI__builtin_neon_vst1q_bf16_x3, NEON::BI__builtin_neon_vst1q_x3_v },
6625 { NEON::BI__builtin_neon_vst1q_bf16_x4, NEON::BI__builtin_neon_vst1q_x4_v },
6626 { NEON::BI__builtin_neon_vst1q_bf16, NEON::BI__builtin_neon_vst1q_v },
6627 { NEON::BI__builtin_neon_vst1q_lane_bf16, NEON::BI__builtin_neon_vst1q_lane_v },
6628 { NEON::BI__builtin_neon_vst2_bf16, NEON::BI__builtin_neon_vst2_v },
6629 { NEON::BI__builtin_neon_vst2_lane_bf16, NEON::BI__builtin_neon_vst2_lane_v },
6630 { NEON::BI__builtin_neon_vst2q_bf16, NEON::BI__builtin_neon_vst2q_v },
6631 { NEON::BI__builtin_neon_vst2q_lane_bf16, NEON::BI__builtin_neon_vst2q_lane_v },
6632 { NEON::BI__builtin_neon_vst3_bf16, NEON::BI__builtin_neon_vst3_v },
6633 { NEON::BI__builtin_neon_vst3_lane_bf16, NEON::BI__builtin_neon_vst3_lane_v },
6634 { NEON::BI__builtin_neon_vst3q_bf16, NEON::BI__builtin_neon_vst3q_v },
6635 { NEON::BI__builtin_neon_vst3q_lane_bf16, NEON::BI__builtin_neon_vst3q_lane_v },
6636 { NEON::BI__builtin_neon_vst4_bf16, NEON::BI__builtin_neon_vst4_v },
6637 { NEON::BI__builtin_neon_vst4_lane_bf16, NEON::BI__builtin_neon_vst4_lane_v },
6638 { NEON::BI__builtin_neon_vst4q_bf16, NEON::BI__builtin_neon_vst4q_v },
6639 { NEON::BI__builtin_neon_vst4q_lane_bf16, NEON::BI__builtin_neon_vst4q_lane_v },
6640 { NEON::BI__builtin_neon_vtrn_f16, NEON::BI__builtin_neon_vtrn_v, },
6641 { NEON::BI__builtin_neon_vtrnq_f16, NEON::BI__builtin_neon_vtrnq_v, },
6642 { NEON::BI__builtin_neon_vuzp_f16, NEON::BI__builtin_neon_vuzp_v, },
6643 { NEON::BI__builtin_neon_vuzpq_f16, NEON::BI__builtin_neon_vuzpq_v, },
6644 { NEON::BI__builtin_neon_vzip_f16, NEON::BI__builtin_neon_vzip_v, },
6645 { NEON::BI__builtin_neon_vzipq_f16, NEON::BI__builtin_neon_vzipq_v, },
6646 };
6647
6648 #undef NEONMAP0
6649 #undef NEONMAP1
6650 #undef NEONMAP2
6651
6652 #define SVEMAP1(NameBase, LLVMIntrinsic, TypeModifier) \
6653 { \
6654 #NameBase, SVE::BI__builtin_sve_##NameBase, Intrinsic::LLVMIntrinsic, 0, \
6655 TypeModifier \
6656 }
6657
6658 #define SVEMAP2(NameBase, TypeModifier) \
6659 { #NameBase, SVE::BI__builtin_sve_##NameBase, 0, 0, TypeModifier }
6660 static const ARMVectorIntrinsicInfo AArch64SVEIntrinsicMap[] = {
6661 #define GET_SVE_LLVM_INTRINSIC_MAP
6662 #include "clang/Basic/arm_sve_builtin_cg.inc"
6663 #include "clang/Basic/BuiltinsAArch64NeonSVEBridge_cg.def"
6664 #undef GET_SVE_LLVM_INTRINSIC_MAP
6665 };
6666
6667 #undef SVEMAP1
6668 #undef SVEMAP2
6669
6670 static bool NEONSIMDIntrinsicsProvenSorted = false;
6671
6672 static bool AArch64SIMDIntrinsicsProvenSorted = false;
6673 static bool AArch64SISDIntrinsicsProvenSorted = false;
6674 static bool AArch64SVEIntrinsicsProvenSorted = false;
6675
6676 static const ARMVectorIntrinsicInfo *
6677 findARMVectorIntrinsicInMap(ArrayRef<ARMVectorIntrinsicInfo> IntrinsicMap,
6678 unsigned BuiltinID, bool &MapProvenSorted) {
6679
6680 #ifndef NDEBUG
6681 if (!MapProvenSorted) {
6682 assert(llvm::is_sorted(IntrinsicMap));
6683 MapProvenSorted = true;
6684 }
6685 #endif
6686
6687 const ARMVectorIntrinsicInfo *Builtin =
6688 llvm::lower_bound(IntrinsicMap, BuiltinID);
6689
6690 if (Builtin != IntrinsicMap.end() && Builtin->BuiltinID == BuiltinID)
6691 return Builtin;
6692
6693 return nullptr;
6694 }
6695
6696 Function *CodeGenFunction::LookupNeonLLVMIntrinsic(unsigned IntrinsicID,
6697 unsigned Modifier,
6698 llvm::Type *ArgType,
6699 const CallExpr *E) {
6700 int VectorSize = 0;
6701 if (Modifier & Use64BitVectors)
6702 VectorSize = 64;
6703 else if (Modifier & Use128BitVectors)
6704 VectorSize = 128;
6705
6706 // Return type.
6707 SmallVector<llvm::Type *, 3> Tys;
6708 if (Modifier & AddRetType) {
6709 llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext()));
6710 if (Modifier & VectorizeRetType)
6711 Ty = llvm::FixedVectorType::get(
6712 Ty, VectorSize ? VectorSize / Ty->getPrimitiveSizeInBits() : 1);
6713
6714 Tys.push_back(Ty);
6715 }
6716
6717 // Arguments.
6718 if (Modifier & VectorizeArgTypes) {
6719 int Elts = VectorSize ? VectorSize / ArgType->getPrimitiveSizeInBits() : 1;
6720 ArgType = llvm::FixedVectorType::get(ArgType, Elts);
6721 }
6722
6723 if (Modifier & (Add1ArgType | Add2ArgTypes))
6724 Tys.push_back(ArgType);
6725
6726 if (Modifier & Add2ArgTypes)
6727 Tys.push_back(ArgType);
6728
6729 if (Modifier & InventFloatType)
6730 Tys.push_back(FloatTy);
6731
6732 return CGM.getIntrinsic(IntrinsicID, Tys);
6733 }
6734
6735 static Value *EmitCommonNeonSISDBuiltinExpr(
6736 CodeGenFunction &CGF, const ARMVectorIntrinsicInfo &SISDInfo,
6737 SmallVectorImpl<Value *> &Ops, const CallExpr *E) {
6738 unsigned BuiltinID = SISDInfo.BuiltinID;
6739 unsigned int Int = SISDInfo.LLVMIntrinsic;
6740 unsigned Modifier = SISDInfo.TypeModifier;
6741 const char *s = SISDInfo.NameHint;
6742
6743 switch (BuiltinID) {
6744 case NEON::BI__builtin_neon_vcled_s64:
6745 case NEON::BI__builtin_neon_vcled_u64:
6746 case NEON::BI__builtin_neon_vcles_f32:
6747 case NEON::BI__builtin_neon_vcled_f64:
6748 case NEON::BI__builtin_neon_vcltd_s64:
6749 case NEON::BI__builtin_neon_vcltd_u64:
6750 case NEON::BI__builtin_neon_vclts_f32:
6751 case NEON::BI__builtin_neon_vcltd_f64:
6752 case NEON::BI__builtin_neon_vcales_f32:
6753 case NEON::BI__builtin_neon_vcaled_f64:
6754 case NEON::BI__builtin_neon_vcalts_f32:
6755 case NEON::BI__builtin_neon_vcaltd_f64:
6756 // Only one direction of comparisons actually exist, cmle is actually a cmge
6757 // with swapped operands. The table gives us the right intrinsic but we
6758 // still need to do the swap.
6759 std::swap(Ops[0], Ops[1]);
6760 break;
6761 }
6762
6763 assert(Int && "Generic code assumes a valid intrinsic");
6764
6765 // Determine the type(s) of this overloaded AArch64 intrinsic.
6766 const Expr *Arg = E->getArg(0);
6767 llvm::Type *ArgTy = CGF.ConvertType(Arg->getType());
6768 Function *F = CGF.LookupNeonLLVMIntrinsic(Int, Modifier, ArgTy, E);
6769
6770 int j = 0;
6771 ConstantInt *C0 = ConstantInt::get(CGF.SizeTy, 0);
6772 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end();
6773 ai != ae; ++ai, ++j) {
6774 llvm::Type *ArgTy = ai->getType();
6775 if (Ops[j]->getType()->getPrimitiveSizeInBits() ==
6776 ArgTy->getPrimitiveSizeInBits())
6777 continue;
6778
6779 assert(ArgTy->isVectorTy() && !Ops[j]->getType()->isVectorTy());
6780 // The constant argument to an _n_ intrinsic always has Int32Ty, so truncate
6781 // it before inserting.
6782 Ops[j] = CGF.Builder.CreateTruncOrBitCast(
6783 Ops[j], cast<llvm::VectorType>(ArgTy)->getElementType());
6784 Ops[j] =
6785 CGF.Builder.CreateInsertElement(PoisonValue::get(ArgTy), Ops[j], C0);
6786 }
6787
6788 Value *Result = CGF.EmitNeonCall(F, Ops, s);
6789 llvm::Type *ResultType = CGF.ConvertType(E->getType());
6790 if (ResultType->getPrimitiveSizeInBits().getFixedValue() <
6791 Result->getType()->getPrimitiveSizeInBits().getFixedValue())
6792 return CGF.Builder.CreateExtractElement(Result, C0);
6793
6794 return CGF.Builder.CreateBitCast(Result, ResultType, s);
6795 }
6796
6797 Value *CodeGenFunction::EmitCommonNeonBuiltinExpr(
6798 unsigned BuiltinID, unsigned LLVMIntrinsic, unsigned AltLLVMIntrinsic,
6799 const char *NameHint, unsigned Modifier, const CallExpr *E,
6800 SmallVectorImpl<llvm::Value *> &Ops, Address PtrOp0, Address PtrOp1,
6801 llvm::Triple::ArchType Arch) {
6802 // Get the last argument, which specifies the vector type.
6803 const Expr *Arg = E->getArg(E->getNumArgs() - 1);
6804 std::optional<llvm::APSInt> NeonTypeConst =
6805 Arg->getIntegerConstantExpr(getContext());
6806 if (!NeonTypeConst)
6807 return nullptr;
6808
6809 // Determine the type of this overloaded NEON intrinsic.
6810 NeonTypeFlags Type(NeonTypeConst->getZExtValue());
6811 bool Usgn = Type.isUnsigned();
6812 bool Quad = Type.isQuad();
6813 const bool HasLegalHalfType = getTarget().hasLegalHalfType();
6814 const bool AllowBFloatArgsAndRet =
6815 getTargetHooks().getABIInfo().allowBFloatArgsAndRet();
6816
6817 llvm::FixedVectorType *VTy =
6818 GetNeonType(this, Type, HasLegalHalfType, false, AllowBFloatArgsAndRet);
6819 llvm::Type *Ty = VTy;
6820 if (!Ty)
6821 return nullptr;
6822
6823 auto getAlignmentValue32 = [&](Address addr) -> Value* {
6824 return Builder.getInt32(addr.getAlignment().getQuantity());
6825 };
6826
6827 unsigned Int = LLVMIntrinsic;
6828 if ((Modifier & UnsignedAlts) && !Usgn)
6829 Int = AltLLVMIntrinsic;
6830
6831 switch (BuiltinID) {
6832 default: break;
6833 case NEON::BI__builtin_neon_splat_lane_v:
6834 case NEON::BI__builtin_neon_splat_laneq_v:
6835 case NEON::BI__builtin_neon_splatq_lane_v:
6836 case NEON::BI__builtin_neon_splatq_laneq_v: {
6837 auto NumElements = VTy->getElementCount();
6838 if (BuiltinID == NEON::BI__builtin_neon_splatq_lane_v)
6839 NumElements = NumElements * 2;
6840 if (BuiltinID == NEON::BI__builtin_neon_splat_laneq_v)
6841 NumElements = NumElements.divideCoefficientBy(2);
6842
6843 Ops[0] = Builder.CreateBitCast(Ops[0], VTy);
6844 return EmitNeonSplat(Ops[0], cast<ConstantInt>(Ops[1]), NumElements);
6845 }
6846 case NEON::BI__builtin_neon_vpadd_v:
6847 case NEON::BI__builtin_neon_vpaddq_v:
6848 // We don't allow fp/int overloading of intrinsics.
6849 if (VTy->getElementType()->isFloatingPointTy() &&
6850 Int == Intrinsic::aarch64_neon_addp)
6851 Int = Intrinsic::aarch64_neon_faddp;
6852 break;
6853 case NEON::BI__builtin_neon_vabs_v:
6854 case NEON::BI__builtin_neon_vabsq_v:
6855 if (VTy->getElementType()->isFloatingPointTy())
6856 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::fabs, Ty), Ops, "vabs");
6857 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vabs");
6858 case NEON::BI__builtin_neon_vadd_v:
6859 case NEON::BI__builtin_neon_vaddq_v: {
6860 llvm::Type *VTy = llvm::FixedVectorType::get(Int8Ty, Quad ? 16 : 8);
6861 Ops[0] = Builder.CreateBitCast(Ops[0], VTy);
6862 Ops[1] = Builder.CreateBitCast(Ops[1], VTy);
6863 Ops[0] = Builder.CreateXor(Ops[0], Ops[1]);
6864 return Builder.CreateBitCast(Ops[0], Ty);
6865 }
6866 case NEON::BI__builtin_neon_vaddhn_v: {
6867 llvm::FixedVectorType *SrcTy =
6868 llvm::FixedVectorType::getExtendedElementVectorType(VTy);
6869
6870 // %sum = add <4 x i32> %lhs, %rhs
6871 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy);
6872 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy);
6873 Ops[0] = Builder.CreateAdd(Ops[0], Ops[1], "vaddhn");
6874
6875 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16>
6876 Constant *ShiftAmt =
6877 ConstantInt::get(SrcTy, SrcTy->getScalarSizeInBits() / 2);
6878 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vaddhn");
6879
6880 // %res = trunc <4 x i32> %high to <4 x i16>
6881 return Builder.CreateTrunc(Ops[0], VTy, "vaddhn");
6882 }
6883 case NEON::BI__builtin_neon_vcale_v:
6884 case NEON::BI__builtin_neon_vcaleq_v:
6885 case NEON::BI__builtin_neon_vcalt_v:
6886 case NEON::BI__builtin_neon_vcaltq_v:
6887 std::swap(Ops[0], Ops[1]);
6888 [[fallthrough]];
6889 case NEON::BI__builtin_neon_vcage_v:
6890 case NEON::BI__builtin_neon_vcageq_v:
6891 case NEON::BI__builtin_neon_vcagt_v:
6892 case NEON::BI__builtin_neon_vcagtq_v: {
6893 llvm::Type *Ty;
6894 switch (VTy->getScalarSizeInBits()) {
6895 default: llvm_unreachable("unexpected type");
6896 case 32:
6897 Ty = FloatTy;
6898 break;
6899 case 64:
6900 Ty = DoubleTy;
6901 break;
6902 case 16:
6903 Ty = HalfTy;
6904 break;
6905 }
6906 auto *VecFlt = llvm::FixedVectorType::get(Ty, VTy->getNumElements());
6907 llvm::Type *Tys[] = { VTy, VecFlt };
6908 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys);
6909 return EmitNeonCall(F, Ops, NameHint);
6910 }
6911 case NEON::BI__builtin_neon_vceqz_v:
6912 case NEON::BI__builtin_neon_vceqzq_v:
6913 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OEQ,
6914 ICmpInst::ICMP_EQ, "vceqz");
6915 case NEON::BI__builtin_neon_vcgez_v:
6916 case NEON::BI__builtin_neon_vcgezq_v:
6917 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGE,
6918 ICmpInst::ICMP_SGE, "vcgez");
6919 case NEON::BI__builtin_neon_vclez_v:
6920 case NEON::BI__builtin_neon_vclezq_v:
6921 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLE,
6922 ICmpInst::ICMP_SLE, "vclez");
6923 case NEON::BI__builtin_neon_vcgtz_v:
6924 case NEON::BI__builtin_neon_vcgtzq_v:
6925 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGT,
6926 ICmpInst::ICMP_SGT, "vcgtz");
6927 case NEON::BI__builtin_neon_vcltz_v:
6928 case NEON::BI__builtin_neon_vcltzq_v:
6929 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLT,
6930 ICmpInst::ICMP_SLT, "vcltz");
6931 case NEON::BI__builtin_neon_vclz_v:
6932 case NEON::BI__builtin_neon_vclzq_v:
6933 // We generate target-independent intrinsic, which needs a second argument
6934 // for whether or not clz of zero is undefined; on ARM it isn't.
6935 Ops.push_back(Builder.getInt1(getTarget().isCLZForZeroUndef()));
6936 break;
6937 case NEON::BI__builtin_neon_vcvt_f32_v:
6938 case NEON::BI__builtin_neon_vcvtq_f32_v:
6939 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
6940 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, Quad),
6941 HasLegalHalfType);
6942 return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt")
6943 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt");
6944 case NEON::BI__builtin_neon_vcvt_f16_s16:
6945 case NEON::BI__builtin_neon_vcvt_f16_u16:
6946 case NEON::BI__builtin_neon_vcvtq_f16_s16:
6947 case NEON::BI__builtin_neon_vcvtq_f16_u16:
6948 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
6949 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float16, false, Quad),
6950 HasLegalHalfType);
6951 return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt")
6952 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt");
6953 case NEON::BI__builtin_neon_vcvt_n_f16_s16:
6954 case NEON::BI__builtin_neon_vcvt_n_f16_u16:
6955 case NEON::BI__builtin_neon_vcvtq_n_f16_s16:
6956 case NEON::BI__builtin_neon_vcvtq_n_f16_u16: {
6957 llvm::Type *Tys[2] = { GetFloatNeonType(this, Type), Ty };
6958 Function *F = CGM.getIntrinsic(Int, Tys);
6959 return EmitNeonCall(F, Ops, "vcvt_n");
6960 }
6961 case NEON::BI__builtin_neon_vcvt_n_f32_v:
6962 case NEON::BI__builtin_neon_vcvt_n_f64_v:
6963 case NEON::BI__builtin_neon_vcvtq_n_f32_v:
6964 case NEON::BI__builtin_neon_vcvtq_n_f64_v: {
6965 llvm::Type *Tys[2] = { GetFloatNeonType(this, Type), Ty };
6966 Int = Usgn ? LLVMIntrinsic : AltLLVMIntrinsic;
6967 Function *F = CGM.getIntrinsic(Int, Tys);
6968 return EmitNeonCall(F, Ops, "vcvt_n");
6969 }
6970 case NEON::BI__builtin_neon_vcvt_n_s16_f16:
6971 case NEON::BI__builtin_neon_vcvt_n_s32_v:
6972 case NEON::BI__builtin_neon_vcvt_n_u16_f16:
6973 case NEON::BI__builtin_neon_vcvt_n_u32_v:
6974 case NEON::BI__builtin_neon_vcvt_n_s64_v:
6975 case NEON::BI__builtin_neon_vcvt_n_u64_v:
6976 case NEON::BI__builtin_neon_vcvtq_n_s16_f16:
6977 case NEON::BI__builtin_neon_vcvtq_n_s32_v:
6978 case NEON::BI__builtin_neon_vcvtq_n_u16_f16:
6979 case NEON::BI__builtin_neon_vcvtq_n_u32_v:
6980 case NEON::BI__builtin_neon_vcvtq_n_s64_v:
6981 case NEON::BI__builtin_neon_vcvtq_n_u64_v: {
6982 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) };
6983 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys);
6984 return EmitNeonCall(F, Ops, "vcvt_n");
6985 }
6986 case NEON::BI__builtin_neon_vcvt_s32_v:
6987 case NEON::BI__builtin_neon_vcvt_u32_v:
6988 case NEON::BI__builtin_neon_vcvt_s64_v:
6989 case NEON::BI__builtin_neon_vcvt_u64_v:
6990 case NEON::BI__builtin_neon_vcvt_s16_f16:
6991 case NEON::BI__builtin_neon_vcvt_u16_f16:
6992 case NEON::BI__builtin_neon_vcvtq_s32_v:
6993 case NEON::BI__builtin_neon_vcvtq_u32_v:
6994 case NEON::BI__builtin_neon_vcvtq_s64_v:
6995 case NEON::BI__builtin_neon_vcvtq_u64_v:
6996 case NEON::BI__builtin_neon_vcvtq_s16_f16:
6997 case NEON::BI__builtin_neon_vcvtq_u16_f16: {
6998 Ops[0] = Builder.CreateBitCast(Ops[0], GetFloatNeonType(this, Type));
6999 return Usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt")
7000 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt");
7001 }
7002 case NEON::BI__builtin_neon_vcvta_s16_f16:
7003 case NEON::BI__builtin_neon_vcvta_s32_v:
7004 case NEON::BI__builtin_neon_vcvta_s64_v:
7005 case NEON::BI__builtin_neon_vcvta_u16_f16:
7006 case NEON::BI__builtin_neon_vcvta_u32_v:
7007 case NEON::BI__builtin_neon_vcvta_u64_v:
7008 case NEON::BI__builtin_neon_vcvtaq_s16_f16:
7009 case NEON::BI__builtin_neon_vcvtaq_s32_v:
7010 case NEON::BI__builtin_neon_vcvtaq_s64_v:
7011 case NEON::BI__builtin_neon_vcvtaq_u16_f16:
7012 case NEON::BI__builtin_neon_vcvtaq_u32_v:
7013 case NEON::BI__builtin_neon_vcvtaq_u64_v:
7014 case NEON::BI__builtin_neon_vcvtn_s16_f16:
7015 case NEON::BI__builtin_neon_vcvtn_s32_v:
7016 case NEON::BI__builtin_neon_vcvtn_s64_v:
7017 case NEON::BI__builtin_neon_vcvtn_u16_f16:
7018 case NEON::BI__builtin_neon_vcvtn_u32_v:
7019 case NEON::BI__builtin_neon_vcvtn_u64_v:
7020 case NEON::BI__builtin_neon_vcvtnq_s16_f16:
7021 case NEON::BI__builtin_neon_vcvtnq_s32_v:
7022 case NEON::BI__builtin_neon_vcvtnq_s64_v:
7023 case NEON::BI__builtin_neon_vcvtnq_u16_f16:
7024 case NEON::BI__builtin_neon_vcvtnq_u32_v:
7025 case NEON::BI__builtin_neon_vcvtnq_u64_v:
7026 case NEON::BI__builtin_neon_vcvtp_s16_f16:
7027 case NEON::BI__builtin_neon_vcvtp_s32_v:
7028 case NEON::BI__builtin_neon_vcvtp_s64_v:
7029 case NEON::BI__builtin_neon_vcvtp_u16_f16:
7030 case NEON::BI__builtin_neon_vcvtp_u32_v:
7031 case NEON::BI__builtin_neon_vcvtp_u64_v:
7032 case NEON::BI__builtin_neon_vcvtpq_s16_f16:
7033 case NEON::BI__builtin_neon_vcvtpq_s32_v:
7034 case NEON::BI__builtin_neon_vcvtpq_s64_v:
7035 case NEON::BI__builtin_neon_vcvtpq_u16_f16:
7036 case NEON::BI__builtin_neon_vcvtpq_u32_v:
7037 case NEON::BI__builtin_neon_vcvtpq_u64_v:
7038 case NEON::BI__builtin_neon_vcvtm_s16_f16:
7039 case NEON::BI__builtin_neon_vcvtm_s32_v:
7040 case NEON::BI__builtin_neon_vcvtm_s64_v:
7041 case NEON::BI__builtin_neon_vcvtm_u16_f16:
7042 case NEON::BI__builtin_neon_vcvtm_u32_v:
7043 case NEON::BI__builtin_neon_vcvtm_u64_v:
7044 case NEON::BI__builtin_neon_vcvtmq_s16_f16:
7045 case NEON::BI__builtin_neon_vcvtmq_s32_v:
7046 case NEON::BI__builtin_neon_vcvtmq_s64_v:
7047 case NEON::BI__builtin_neon_vcvtmq_u16_f16:
7048 case NEON::BI__builtin_neon_vcvtmq_u32_v:
7049 case NEON::BI__builtin_neon_vcvtmq_u64_v: {
7050 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) };
7051 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, NameHint);
7052 }
7053 case NEON::BI__builtin_neon_vcvtx_f32_v: {
7054 llvm::Type *Tys[2] = { VTy->getTruncatedElementVectorType(VTy), Ty};
7055 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, NameHint);
7056
7057 }
7058 case NEON::BI__builtin_neon_vext_v:
7059 case NEON::BI__builtin_neon_vextq_v: {
7060 int CV = cast<ConstantInt>(Ops[2])->getSExtValue();
7061 SmallVector<int, 16> Indices;
7062 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i)
7063 Indices.push_back(i+CV);
7064
7065 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
7066 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
7067 return Builder.CreateShuffleVector(Ops[0], Ops[1], Indices, "vext");
7068 }
7069 case NEON::BI__builtin_neon_vfma_v:
7070 case NEON::BI__builtin_neon_vfmaq_v: {
7071 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
7072 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
7073 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
7074
7075 // NEON intrinsic puts accumulator first, unlike the LLVM fma.
7076 return emitCallMaybeConstrainedFPBuiltin(
7077 *this, Intrinsic::fma, Intrinsic::experimental_constrained_fma, Ty,
7078 {Ops[1], Ops[2], Ops[0]});
7079 }
7080 case NEON::BI__builtin_neon_vld1_v:
7081 case NEON::BI__builtin_neon_vld1q_v: {
7082 llvm::Type *Tys[] = {Ty, Int8PtrTy};
7083 Ops.push_back(getAlignmentValue32(PtrOp0));
7084 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, "vld1");
7085 }
7086 case NEON::BI__builtin_neon_vld1_x2_v:
7087 case NEON::BI__builtin_neon_vld1q_x2_v:
7088 case NEON::BI__builtin_neon_vld1_x3_v:
7089 case NEON::BI__builtin_neon_vld1q_x3_v:
7090 case NEON::BI__builtin_neon_vld1_x4_v:
7091 case NEON::BI__builtin_neon_vld1q_x4_v: {
7092 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getElementType());
7093 Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
7094 llvm::Type *Tys[2] = { VTy, PTy };
7095 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys);
7096 Ops[1] = Builder.CreateCall(F, Ops[1], "vld1xN");
7097 Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
7098 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
7099 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
7100 }
7101 case NEON::BI__builtin_neon_vld2_v:
7102 case NEON::BI__builtin_neon_vld2q_v:
7103 case NEON::BI__builtin_neon_vld3_v:
7104 case NEON::BI__builtin_neon_vld3q_v:
7105 case NEON::BI__builtin_neon_vld4_v:
7106 case NEON::BI__builtin_neon_vld4q_v:
7107 case NEON::BI__builtin_neon_vld2_dup_v:
7108 case NEON::BI__builtin_neon_vld2q_dup_v:
7109 case NEON::BI__builtin_neon_vld3_dup_v:
7110 case NEON::BI__builtin_neon_vld3q_dup_v:
7111 case NEON::BI__builtin_neon_vld4_dup_v:
7112 case NEON::BI__builtin_neon_vld4q_dup_v: {
7113 llvm::Type *Tys[] = {Ty, Int8PtrTy};
7114 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys);
7115 Value *Align = getAlignmentValue32(PtrOp1);
7116 Ops[1] = Builder.CreateCall(F, {Ops[1], Align}, NameHint);
7117 Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
7118 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
7119 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
7120 }
7121 case NEON::BI__builtin_neon_vld1_dup_v:
7122 case NEON::BI__builtin_neon_vld1q_dup_v: {
7123 Value *V = PoisonValue::get(Ty);
7124 PtrOp0 = Builder.CreateElementBitCast(PtrOp0, VTy->getElementType());
7125 LoadInst *Ld = Builder.CreateLoad(PtrOp0);
7126 llvm::Constant *CI = ConstantInt::get(SizeTy, 0);
7127 Ops[0] = Builder.CreateInsertElement(V, Ld, CI);
7128 return EmitNeonSplat(Ops[0], CI);
7129 }
7130 case NEON::BI__builtin_neon_vld2_lane_v:
7131 case NEON::BI__builtin_neon_vld2q_lane_v:
7132 case NEON::BI__builtin_neon_vld3_lane_v:
7133 case NEON::BI__builtin_neon_vld3q_lane_v:
7134 case NEON::BI__builtin_neon_vld4_lane_v:
7135 case NEON::BI__builtin_neon_vld4q_lane_v: {
7136 llvm::Type *Tys[] = {Ty, Int8PtrTy};
7137 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys);
7138 for (unsigned I = 2; I < Ops.size() - 1; ++I)
7139 Ops[I] = Builder.CreateBitCast(Ops[I], Ty);
7140 Ops.push_back(getAlignmentValue32(PtrOp1));
7141 Ops[1] = Builder.CreateCall(F, ArrayRef(Ops).slice(1), NameHint);
7142 Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
7143 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
7144 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
7145 }
7146 case NEON::BI__builtin_neon_vmovl_v: {
7147 llvm::FixedVectorType *DTy =
7148 llvm::FixedVectorType::getTruncatedElementVectorType(VTy);
7149 Ops[0] = Builder.CreateBitCast(Ops[0], DTy);
7150 if (Usgn)
7151 return Builder.CreateZExt(Ops[0], Ty, "vmovl");
7152 return Builder.CreateSExt(Ops[0], Ty, "vmovl");
7153 }
7154 case NEON::BI__builtin_neon_vmovn_v: {
7155 llvm::FixedVectorType *QTy =
7156 llvm::FixedVectorType::getExtendedElementVectorType(VTy);
7157 Ops[0] = Builder.CreateBitCast(Ops[0], QTy);
7158 return Builder.CreateTrunc(Ops[0], Ty, "vmovn");
7159 }
7160 case NEON::BI__builtin_neon_vmull_v:
7161 // FIXME: the integer vmull operations could be emitted in terms of pure
7162 // LLVM IR (2 exts followed by a mul). Unfortunately LLVM has a habit of
7163 // hoisting the exts outside loops. Until global ISel comes along that can
7164 // see through such movement this leads to bad CodeGen. So we need an
7165 // intrinsic for now.
7166 Int = Usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls;
7167 Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int;
7168 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull");
7169 case NEON::BI__builtin_neon_vpadal_v:
7170 case NEON::BI__builtin_neon_vpadalq_v: {
7171 // The source operand type has twice as many elements of half the size.
7172 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits();
7173 llvm::Type *EltTy =
7174 llvm::IntegerType::get(getLLVMContext(), EltBits / 2);
7175 auto *NarrowTy =
7176 llvm::FixedVectorType::get(EltTy, VTy->getNumElements() * 2);
7177 llvm::Type *Tys[2] = { Ty, NarrowTy };
7178 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, NameHint);
7179 }
7180 case NEON::BI__builtin_neon_vpaddl_v:
7181 case NEON::BI__builtin_neon_vpaddlq_v: {
7182 // The source operand type has twice as many elements of half the size.
7183 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits();
7184 llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2);
7185 auto *NarrowTy =
7186 llvm::FixedVectorType::get(EltTy, VTy->getNumElements() * 2);
7187 llvm::Type *Tys[2] = { Ty, NarrowTy };
7188 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl");
7189 }
7190 case NEON::BI__builtin_neon_vqdmlal_v:
7191 case NEON::BI__builtin_neon_vqdmlsl_v: {
7192 SmallVector<Value *, 2> MulOps(Ops.begin() + 1, Ops.end());
7193 Ops[1] =
7194 EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), MulOps, "vqdmlal");
7195 Ops.resize(2);
7196 return EmitNeonCall(CGM.getIntrinsic(AltLLVMIntrinsic, Ty), Ops, NameHint);
7197 }
7198 case NEON::BI__builtin_neon_vqdmulhq_lane_v:
7199 case NEON::BI__builtin_neon_vqdmulh_lane_v:
7200 case NEON::BI__builtin_neon_vqrdmulhq_lane_v:
7201 case NEON::BI__builtin_neon_vqrdmulh_lane_v: {
7202 auto *RTy = cast<llvm::FixedVectorType>(Ty);
7203 if (BuiltinID == NEON::BI__builtin_neon_vqdmulhq_lane_v ||
7204 BuiltinID == NEON::BI__builtin_neon_vqrdmulhq_lane_v)
7205 RTy = llvm::FixedVectorType::get(RTy->getElementType(),
7206 RTy->getNumElements() * 2);
7207 llvm::Type *Tys[2] = {
7208 RTy, GetNeonType(this, NeonTypeFlags(Type.getEltType(), false,
7209 /*isQuad*/ false))};
7210 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, NameHint);
7211 }
7212 case NEON::BI__builtin_neon_vqdmulhq_laneq_v:
7213 case NEON::BI__builtin_neon_vqdmulh_laneq_v:
7214 case NEON::BI__builtin_neon_vqrdmulhq_laneq_v:
7215 case NEON::BI__builtin_neon_vqrdmulh_laneq_v: {
7216 llvm::Type *Tys[2] = {
7217 Ty, GetNeonType(this, NeonTypeFlags(Type.getEltType(), false,
7218 /*isQuad*/ true))};
7219 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, NameHint);
7220 }
7221 case NEON::BI__builtin_neon_vqshl_n_v:
7222 case NEON::BI__builtin_neon_vqshlq_n_v:
7223 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n",
7224 1, false);
7225 case NEON::BI__builtin_neon_vqshlu_n_v:
7226 case NEON::BI__builtin_neon_vqshluq_n_v:
7227 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshlu_n",
7228 1, false);
7229 case NEON::BI__builtin_neon_vrecpe_v:
7230 case NEON::BI__builtin_neon_vrecpeq_v:
7231 case NEON::BI__builtin_neon_vrsqrte_v:
7232 case NEON::BI__builtin_neon_vrsqrteq_v:
7233 Int = Ty->isFPOrFPVectorTy() ? LLVMIntrinsic : AltLLVMIntrinsic;
7234 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, NameHint);
7235 case NEON::BI__builtin_neon_vrndi_v:
7236 case NEON::BI__builtin_neon_vrndiq_v:
7237 Int = Builder.getIsFPConstrained()
7238 ? Intrinsic::experimental_constrained_nearbyint
7239 : Intrinsic::nearbyint;
7240 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, NameHint);
7241 case NEON::BI__builtin_neon_vrshr_n_v:
7242 case NEON::BI__builtin_neon_vrshrq_n_v:
7243 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n",
7244 1, true);
7245 case NEON::BI__builtin_neon_vsha512hq_u64:
7246 case NEON::BI__builtin_neon_vsha512h2q_u64:
7247 case NEON::BI__builtin_neon_vsha512su0q_u64:
7248 case NEON::BI__builtin_neon_vsha512su1q_u64: {
7249 Function *F = CGM.getIntrinsic(Int);
7250 return EmitNeonCall(F, Ops, "");
7251 }
7252 case NEON::BI__builtin_neon_vshl_n_v:
7253 case NEON::BI__builtin_neon_vshlq_n_v:
7254 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false);
7255 return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1],
7256 "vshl_n");
7257 case NEON::BI__builtin_neon_vshll_n_v: {
7258 llvm::FixedVectorType *SrcTy =
7259 llvm::FixedVectorType::getTruncatedElementVectorType(VTy);
7260 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy);
7261 if (Usgn)
7262 Ops[0] = Builder.CreateZExt(Ops[0], VTy);
7263 else
7264 Ops[0] = Builder.CreateSExt(Ops[0], VTy);
7265 Ops[1] = EmitNeonShiftVector(Ops[1], VTy, false);
7266 return Builder.CreateShl(Ops[0], Ops[1], "vshll_n");
7267 }
7268 case NEON::BI__builtin_neon_vshrn_n_v: {
7269 llvm::FixedVectorType *SrcTy =
7270 llvm::FixedVectorType::getExtendedElementVectorType(VTy);
7271 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy);
7272 Ops[1] = EmitNeonShiftVector(Ops[1], SrcTy, false);
7273 if (Usgn)
7274 Ops[0] = Builder.CreateLShr(Ops[0], Ops[1]);
7275 else
7276 Ops[0] = Builder.CreateAShr(Ops[0], Ops[1]);
7277 return Builder.CreateTrunc(Ops[0], Ty, "vshrn_n");
7278 }
7279 case NEON::BI__builtin_neon_vshr_n_v:
7280 case NEON::BI__builtin_neon_vshrq_n_v:
7281 return EmitNeonRShiftImm(Ops[0], Ops[1], Ty, Usgn, "vshr_n");
7282 case NEON::BI__builtin_neon_vst1_v:
7283 case NEON::BI__builtin_neon_vst1q_v:
7284 case NEON::BI__builtin_neon_vst2_v:
7285 case NEON::BI__builtin_neon_vst2q_v:
7286 case NEON::BI__builtin_neon_vst3_v:
7287 case NEON::BI__builtin_neon_vst3q_v:
7288 case NEON::BI__builtin_neon_vst4_v:
7289 case NEON::BI__builtin_neon_vst4q_v:
7290 case NEON::BI__builtin_neon_vst2_lane_v:
7291 case NEON::BI__builtin_neon_vst2q_lane_v:
7292 case NEON::BI__builtin_neon_vst3_lane_v:
7293 case NEON::BI__builtin_neon_vst3q_lane_v:
7294 case NEON::BI__builtin_neon_vst4_lane_v:
7295 case NEON::BI__builtin_neon_vst4q_lane_v: {
7296 llvm::Type *Tys[] = {Int8PtrTy, Ty};
7297 Ops.push_back(getAlignmentValue32(PtrOp0));
7298 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "");
7299 }
7300 case NEON::BI__builtin_neon_vsm3partw1q_u32:
7301 case NEON::BI__builtin_neon_vsm3partw2q_u32:
7302 case NEON::BI__builtin_neon_vsm3ss1q_u32:
7303 case NEON::BI__builtin_neon_vsm4ekeyq_u32:
7304 case NEON::BI__builtin_neon_vsm4eq_u32: {
7305 Function *F = CGM.getIntrinsic(Int);
7306 return EmitNeonCall(F, Ops, "");
7307 }
7308 case NEON::BI__builtin_neon_vsm3tt1aq_u32:
7309 case NEON::BI__builtin_neon_vsm3tt1bq_u32:
7310 case NEON::BI__builtin_neon_vsm3tt2aq_u32:
7311 case NEON::BI__builtin_neon_vsm3tt2bq_u32: {
7312 Function *F = CGM.getIntrinsic(Int);
7313 Ops[3] = Builder.CreateZExt(Ops[3], Int64Ty);
7314 return EmitNeonCall(F, Ops, "");
7315 }
7316 case NEON::BI__builtin_neon_vst1_x2_v:
7317 case NEON::BI__builtin_neon_vst1q_x2_v:
7318 case NEON::BI__builtin_neon_vst1_x3_v:
7319 case NEON::BI__builtin_neon_vst1q_x3_v:
7320 case NEON::BI__builtin_neon_vst1_x4_v:
7321 case NEON::BI__builtin_neon_vst1q_x4_v: {
7322 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getElementType());
7323 // TODO: Currently in AArch32 mode the pointer operand comes first, whereas
7324 // in AArch64 it comes last. We may want to stick to one or another.
7325 if (Arch == llvm::Triple::aarch64 || Arch == llvm::Triple::aarch64_be ||
7326 Arch == llvm::Triple::aarch64_32) {
7327 llvm::Type *Tys[2] = { VTy, PTy };
7328 std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end());
7329 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, "");
7330 }
7331 llvm::Type *Tys[2] = { PTy, VTy };
7332 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, "");
7333 }
7334 case NEON::BI__builtin_neon_vsubhn_v: {
7335 llvm::FixedVectorType *SrcTy =
7336 llvm::FixedVectorType::getExtendedElementVectorType(VTy);
7337
7338 // %sum = add <4 x i32> %lhs, %rhs
7339 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy);
7340 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy);
7341 Ops[0] = Builder.CreateSub(Ops[0], Ops[1], "vsubhn");
7342
7343 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16>
7344 Constant *ShiftAmt =
7345 ConstantInt::get(SrcTy, SrcTy->getScalarSizeInBits() / 2);
7346 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vsubhn");
7347
7348 // %res = trunc <4 x i32> %high to <4 x i16>
7349 return Builder.CreateTrunc(Ops[0], VTy, "vsubhn");
7350 }
7351 case NEON::BI__builtin_neon_vtrn_v:
7352 case NEON::BI__builtin_neon_vtrnq_v: {
7353 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
7354 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
7355 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
7356 Value *SV = nullptr;
7357
7358 for (unsigned vi = 0; vi != 2; ++vi) {
7359 SmallVector<int, 16> Indices;
7360 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) {
7361 Indices.push_back(i+vi);
7362 Indices.push_back(i+e+vi);
7363 }
7364 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi);
7365 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vtrn");
7366 SV = Builder.CreateDefaultAlignedStore(SV, Addr);
7367 }
7368 return SV;
7369 }
7370 case NEON::BI__builtin_neon_vtst_v:
7371 case NEON::BI__builtin_neon_vtstq_v: {
7372 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
7373 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
7374 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]);
7375 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0],
7376 ConstantAggregateZero::get(Ty));
7377 return Builder.CreateSExt(Ops[0], Ty, "vtst");
7378 }
7379 case NEON::BI__builtin_neon_vuzp_v:
7380 case NEON::BI__builtin_neon_vuzpq_v: {
7381 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
7382 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
7383 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
7384 Value *SV = nullptr;
7385
7386 for (unsigned vi = 0; vi != 2; ++vi) {
7387 SmallVector<int, 16> Indices;
7388 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i)
7389 Indices.push_back(2*i+vi);
7390
7391 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi);
7392 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vuzp");
7393 SV = Builder.CreateDefaultAlignedStore(SV, Addr);
7394 }
7395 return SV;
7396 }
7397 case NEON::BI__builtin_neon_vxarq_u64: {
7398 Function *F = CGM.getIntrinsic(Int);
7399 Ops[2] = Builder.CreateZExt(Ops[2], Int64Ty);
7400 return EmitNeonCall(F, Ops, "");
7401 }
7402 case NEON::BI__builtin_neon_vzip_v:
7403 case NEON::BI__builtin_neon_vzipq_v: {
7404 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
7405 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
7406 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
7407 Value *SV = nullptr;
7408
7409 for (unsigned vi = 0; vi != 2; ++vi) {
7410 SmallVector<int, 16> Indices;
7411 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) {
7412 Indices.push_back((i + vi*e) >> 1);
7413 Indices.push_back(((i + vi*e) >> 1)+e);
7414 }
7415 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi);
7416 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vzip");
7417 SV = Builder.CreateDefaultAlignedStore(SV, Addr);
7418 }
7419 return SV;
7420 }
7421 case NEON::BI__builtin_neon_vdot_s32:
7422 case NEON::BI__builtin_neon_vdot_u32:
7423 case NEON::BI__builtin_neon_vdotq_s32:
7424 case NEON::BI__builtin_neon_vdotq_u32: {
7425 auto *InputTy =
7426 llvm::FixedVectorType::get(Int8Ty, Ty->getPrimitiveSizeInBits() / 8);
7427 llvm::Type *Tys[2] = { Ty, InputTy };
7428 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vdot");
7429 }
7430 case NEON::BI__builtin_neon_vfmlal_low_f16:
7431 case NEON::BI__builtin_neon_vfmlalq_low_f16: {
7432 auto *InputTy =
7433 llvm::FixedVectorType::get(HalfTy, Ty->getPrimitiveSizeInBits() / 16);
7434 llvm::Type *Tys[2] = { Ty, InputTy };
7435 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vfmlal_low");
7436 }
7437 case NEON::BI__builtin_neon_vfmlsl_low_f16:
7438 case NEON::BI__builtin_neon_vfmlslq_low_f16: {
7439 auto *InputTy =
7440 llvm::FixedVectorType::get(HalfTy, Ty->getPrimitiveSizeInBits() / 16);
7441 llvm::Type *Tys[2] = { Ty, InputTy };
7442 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vfmlsl_low");
7443 }
7444 case NEON::BI__builtin_neon_vfmlal_high_f16:
7445 case NEON::BI__builtin_neon_vfmlalq_high_f16: {
7446 auto *InputTy =
7447 llvm::FixedVectorType::get(HalfTy, Ty->getPrimitiveSizeInBits() / 16);
7448 llvm::Type *Tys[2] = { Ty, InputTy };
7449 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vfmlal_high");
7450 }
7451 case NEON::BI__builtin_neon_vfmlsl_high_f16:
7452 case NEON::BI__builtin_neon_vfmlslq_high_f16: {
7453 auto *InputTy =
7454 llvm::FixedVectorType::get(HalfTy, Ty->getPrimitiveSizeInBits() / 16);
7455 llvm::Type *Tys[2] = { Ty, InputTy };
7456 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vfmlsl_high");
7457 }
7458 case NEON::BI__builtin_neon_vmmlaq_s32:
7459 case NEON::BI__builtin_neon_vmmlaq_u32: {
7460 auto *InputTy =
7461 llvm::FixedVectorType::get(Int8Ty, Ty->getPrimitiveSizeInBits() / 8);
7462 llvm::Type *Tys[2] = { Ty, InputTy };
7463 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, "vmmla");
7464 }
7465 case NEON::BI__builtin_neon_vusmmlaq_s32: {
7466 auto *InputTy =
7467 llvm::FixedVectorType::get(Int8Ty, Ty->getPrimitiveSizeInBits() / 8);
7468 llvm::Type *Tys[2] = { Ty, InputTy };
7469 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vusmmla");
7470 }
7471 case NEON::BI__builtin_neon_vusdot_s32:
7472 case NEON::BI__builtin_neon_vusdotq_s32: {
7473 auto *InputTy =
7474 llvm::FixedVectorType::get(Int8Ty, Ty->getPrimitiveSizeInBits() / 8);
7475 llvm::Type *Tys[2] = { Ty, InputTy };
7476 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vusdot");
7477 }
7478 case NEON::BI__builtin_neon_vbfdot_f32:
7479 case NEON::BI__builtin_neon_vbfdotq_f32: {
7480 llvm::Type *InputTy =
7481 llvm::FixedVectorType::get(BFloatTy, Ty->getPrimitiveSizeInBits() / 16);
7482 llvm::Type *Tys[2] = { Ty, InputTy };
7483 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vbfdot");
7484 }
7485 case NEON::BI__builtin_neon___a32_vcvt_bf16_f32: {
7486 llvm::Type *Tys[1] = { Ty };
7487 Function *F = CGM.getIntrinsic(Int, Tys);
7488 return EmitNeonCall(F, Ops, "vcvtfp2bf");
7489 }
7490
7491 }
7492
7493 assert(Int && "Expected valid intrinsic number");
7494
7495 // Determine the type(s) of this overloaded AArch64 intrinsic.
7496 Function *F = LookupNeonLLVMIntrinsic(Int, Modifier, Ty, E);
7497
7498 Value *Result = EmitNeonCall(F, Ops, NameHint);
7499 llvm::Type *ResultType = ConvertType(E->getType());
7500 // AArch64 intrinsic one-element vector type cast to
7501 // scalar type expected by the builtin
7502 return Builder.CreateBitCast(Result, ResultType, NameHint);
7503 }
7504
7505 Value *CodeGenFunction::EmitAArch64CompareBuiltinExpr(
7506 Value *Op, llvm::Type *Ty, const CmpInst::Predicate Fp,
7507 const CmpInst::Predicate Ip, const Twine &Name) {
7508 llvm::Type *OTy = Op->getType();
7509
7510 // FIXME: this is utterly horrific. We should not be looking at previous
7511 // codegen context to find out what needs doing. Unfortunately TableGen
7512 // currently gives us exactly the same calls for vceqz_f32 and vceqz_s32
7513 // (etc).
7514 if (BitCastInst *BI = dyn_cast<BitCastInst>(Op))
7515 OTy = BI->getOperand(0)->getType();
7516
7517 Op = Builder.CreateBitCast(Op, OTy);
7518 if (OTy->getScalarType()->isFloatingPointTy()) {
7519 if (Fp == CmpInst::FCMP_OEQ)
7520 Op = Builder.CreateFCmp(Fp, Op, Constant::getNullValue(OTy));
7521 else
7522 Op = Builder.CreateFCmpS(Fp, Op, Constant::getNullValue(OTy));
7523 } else {
7524 Op = Builder.CreateICmp(Ip, Op, Constant::getNullValue(OTy));
7525 }
7526 return Builder.CreateSExt(Op, Ty, Name);
7527 }
7528
7529 static Value *packTBLDVectorList(CodeGenFunction &CGF, ArrayRef<Value *> Ops,
7530 Value *ExtOp, Value *IndexOp,
7531 llvm::Type *ResTy, unsigned IntID,
7532 const char *Name) {
7533 SmallVector<Value *, 2> TblOps;
7534 if (ExtOp)
7535 TblOps.push_back(ExtOp);
7536
7537 // Build a vector containing sequential number like (0, 1, 2, ..., 15)
7538 SmallVector<int, 16> Indices;
7539 auto *TblTy = cast<llvm::FixedVectorType>(Ops[0]->getType());
7540 for (unsigned i = 0, e = TblTy->getNumElements(); i != e; ++i) {
7541 Indices.push_back(2*i);
7542 Indices.push_back(2*i+1);
7543 }
7544
7545 int PairPos = 0, End = Ops.size() - 1;
7546 while (PairPos < End) {
7547 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos],
7548 Ops[PairPos+1], Indices,
7549 Name));
7550 PairPos += 2;
7551 }
7552
7553 // If there's an odd number of 64-bit lookup table, fill the high 64-bit
7554 // of the 128-bit lookup table with zero.
7555 if (PairPos == End) {
7556 Value *ZeroTbl = ConstantAggregateZero::get(TblTy);
7557 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos],
7558 ZeroTbl, Indices, Name));
7559 }
7560
7561 Function *TblF;
7562 TblOps.push_back(IndexOp);
7563 TblF = CGF.CGM.getIntrinsic(IntID, ResTy);
7564
7565 return CGF.EmitNeonCall(TblF, TblOps, Name);
7566 }
7567
7568 Value *CodeGenFunction::GetValueForARMHint(unsigned BuiltinID) {
7569 unsigned Value;
7570 switch (BuiltinID) {
7571 default:
7572 return nullptr;
7573 case clang::ARM::BI__builtin_arm_nop:
7574 Value = 0;
7575 break;
7576 case clang::ARM::BI__builtin_arm_yield:
7577 case clang::ARM::BI__yield:
7578 Value = 1;
7579 break;
7580 case clang::ARM::BI__builtin_arm_wfe:
7581 case clang::ARM::BI__wfe:
7582 Value = 2;
7583 break;
7584 case clang::ARM::BI__builtin_arm_wfi:
7585 case clang::ARM::BI__wfi:
7586 Value = 3;
7587 break;
7588 case clang::ARM::BI__builtin_arm_sev:
7589 case clang::ARM::BI__sev:
7590 Value = 4;
7591 break;
7592 case clang::ARM::BI__builtin_arm_sevl:
7593 case clang::ARM::BI__sevl:
7594 Value = 5;
7595 break;
7596 }
7597
7598 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint),
7599 llvm::ConstantInt::get(Int32Ty, Value));
7600 }
7601
7602 enum SpecialRegisterAccessKind {
7603 NormalRead,
7604 VolatileRead,
7605 Write,
7606 };
7607
7608 // Generates the IR for the read/write special register builtin,
7609 // ValueType is the type of the value that is to be written or read,
7610 // RegisterType is the type of the register being written to or read from.
7611 static Value *EmitSpecialRegisterBuiltin(CodeGenFunction &CGF,
7612 const CallExpr *E,
7613 llvm::Type *RegisterType,
7614 llvm::Type *ValueType,
7615 SpecialRegisterAccessKind AccessKind,
7616 StringRef SysReg = "") {
7617 // write and register intrinsics only support 32, 64 and 128 bit operations.
7618 assert((RegisterType->isIntegerTy(32) || RegisterType->isIntegerTy(64) ||
7619 RegisterType->isIntegerTy(128)) &&
7620 "Unsupported size for register.");
7621
7622 CodeGen::CGBuilderTy &Builder = CGF.Builder;
7623 CodeGen::CodeGenModule &CGM = CGF.CGM;
7624 LLVMContext &Context = CGM.getLLVMContext();
7625
7626 if (SysReg.empty()) {
7627 const Expr *SysRegStrExpr = E->getArg(0)->IgnoreParenCasts();
7628 SysReg = cast<clang::StringLiteral>(SysRegStrExpr)->getString();
7629 }
7630
7631 llvm::Metadata *Ops[] = { llvm::MDString::get(Context, SysReg) };
7632 llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops);
7633 llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName);
7634
7635 llvm::Type *Types[] = { RegisterType };
7636
7637 bool MixedTypes = RegisterType->isIntegerTy(64) && ValueType->isIntegerTy(32);
7638 assert(!(RegisterType->isIntegerTy(32) && ValueType->isIntegerTy(64))
7639 && "Can't fit 64-bit value in 32-bit register");
7640
7641 if (AccessKind != Write) {
7642 assert(AccessKind == NormalRead || AccessKind == VolatileRead);
7643 llvm::Function *F = CGM.getIntrinsic(
7644 AccessKind == VolatileRead ? llvm::Intrinsic::read_volatile_register
7645 : llvm::Intrinsic::read_register,
7646 Types);
7647 llvm::Value *Call = Builder.CreateCall(F, Metadata);
7648
7649 if (MixedTypes)
7650 // Read into 64 bit register and then truncate result to 32 bit.
7651 return Builder.CreateTrunc(Call, ValueType);
7652
7653 if (ValueType->isPointerTy())
7654 // Have i32/i64 result (Call) but want to return a VoidPtrTy (i8*).
7655 return Builder.CreateIntToPtr(Call, ValueType);
7656
7657 return Call;
7658 }
7659
7660 llvm::Function *F = CGM.getIntrinsic(llvm::Intrinsic::write_register, Types);
7661 llvm::Value *ArgValue = CGF.EmitScalarExpr(E->getArg(1));
7662 if (MixedTypes) {
7663 // Extend 32 bit write value to 64 bit to pass to write.
7664 ArgValue = Builder.CreateZExt(ArgValue, RegisterType);
7665 return Builder.CreateCall(F, { Metadata, ArgValue });
7666 }
7667
7668 if (ValueType->isPointerTy()) {
7669 // Have VoidPtrTy ArgValue but want to return an i32/i64.
7670 ArgValue = Builder.CreatePtrToInt(ArgValue, RegisterType);
7671 return Builder.CreateCall(F, { Metadata, ArgValue });
7672 }
7673
7674 return Builder.CreateCall(F, { Metadata, ArgValue });
7675 }
7676
7677 /// Return true if BuiltinID is an overloaded Neon intrinsic with an extra
7678 /// argument that specifies the vector type.
7679 static bool HasExtraNeonArgument(unsigned BuiltinID) {
7680 switch (BuiltinID) {
7681 default: break;
7682 case NEON::BI__builtin_neon_vget_lane_i8:
7683 case NEON::BI__builtin_neon_vget_lane_i16:
7684 case NEON::BI__builtin_neon_vget_lane_bf16:
7685 case NEON::BI__builtin_neon_vget_lane_i32:
7686 case NEON::BI__builtin_neon_vget_lane_i64:
7687 case NEON::BI__builtin_neon_vget_lane_f32:
7688 case NEON::BI__builtin_neon_vgetq_lane_i8:
7689 case NEON::BI__builtin_neon_vgetq_lane_i16:
7690 case NEON::BI__builtin_neon_vgetq_lane_bf16:
7691 case NEON::BI__builtin_neon_vgetq_lane_i32:
7692 case NEON::BI__builtin_neon_vgetq_lane_i64:
7693 case NEON::BI__builtin_neon_vgetq_lane_f32:
7694 case NEON::BI__builtin_neon_vduph_lane_bf16:
7695 case NEON::BI__builtin_neon_vduph_laneq_bf16:
7696 case NEON::BI__builtin_neon_vset_lane_i8:
7697 case NEON::BI__builtin_neon_vset_lane_i16:
7698 case NEON::BI__builtin_neon_vset_lane_bf16:
7699 case NEON::BI__builtin_neon_vset_lane_i32:
7700 case NEON::BI__builtin_neon_vset_lane_i64:
7701 case NEON::BI__builtin_neon_vset_lane_f32:
7702 case NEON::BI__builtin_neon_vsetq_lane_i8:
7703 case NEON::BI__builtin_neon_vsetq_lane_i16:
7704 case NEON::BI__builtin_neon_vsetq_lane_bf16:
7705 case NEON::BI__builtin_neon_vsetq_lane_i32:
7706 case NEON::BI__builtin_neon_vsetq_lane_i64:
7707 case NEON::BI__builtin_neon_vsetq_lane_f32:
7708 case NEON::BI__builtin_neon_vsha1h_u32:
7709 case NEON::BI__builtin_neon_vsha1cq_u32:
7710 case NEON::BI__builtin_neon_vsha1pq_u32:
7711 case NEON::BI__builtin_neon_vsha1mq_u32:
7712 case NEON::BI__builtin_neon_vcvth_bf16_f32:
7713 case clang::ARM::BI_MoveToCoprocessor:
7714 case clang::ARM::BI_MoveToCoprocessor2:
7715 return false;
7716 }
7717 return true;
7718 }
7719
7720 Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID,
7721 const CallExpr *E,
7722 ReturnValueSlot ReturnValue,
7723 llvm::Triple::ArchType Arch) {
7724 if (auto Hint = GetValueForARMHint(BuiltinID))
7725 return Hint;
7726
7727 if (BuiltinID == clang::ARM::BI__emit) {
7728 bool IsThumb = getTarget().getTriple().getArch() == llvm::Triple::thumb;
7729 llvm::FunctionType *FTy =
7730 llvm::FunctionType::get(VoidTy, /*Variadic=*/false);
7731
7732 Expr::EvalResult Result;
7733 if (!E->getArg(0)->EvaluateAsInt(Result, CGM.getContext()))
7734 llvm_unreachable("Sema will ensure that the parameter is constant");
7735
7736 llvm::APSInt Value = Result.Val.getInt();
7737 uint64_t ZExtValue = Value.zextOrTrunc(IsThumb ? 16 : 32).getZExtValue();
7738
7739 llvm::InlineAsm *Emit =
7740 IsThumb ? InlineAsm::get(FTy, ".inst.n 0x" + utohexstr(ZExtValue), "",
7741 /*hasSideEffects=*/true)
7742 : InlineAsm::get(FTy, ".inst 0x" + utohexstr(ZExtValue), "",
7743 /*hasSideEffects=*/true);
7744
7745 return Builder.CreateCall(Emit);
7746 }
7747
7748 if (BuiltinID == clang::ARM::BI__builtin_arm_dbg) {
7749 Value *Option = EmitScalarExpr(E->getArg(0));
7750 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_dbg), Option);
7751 }
7752
7753 if (BuiltinID == clang::ARM::BI__builtin_arm_prefetch) {
7754 Value *Address = EmitScalarExpr(E->getArg(0));
7755 Value *RW = EmitScalarExpr(E->getArg(1));
7756 Value *IsData = EmitScalarExpr(E->getArg(2));
7757
7758 // Locality is not supported on ARM target
7759 Value *Locality = llvm::ConstantInt::get(Int32Ty, 3);
7760
7761 Function *F = CGM.getIntrinsic(Intrinsic::prefetch, Address->getType());
7762 return Builder.CreateCall(F, {Address, RW, Locality, IsData});
7763 }
7764
7765 if (BuiltinID == clang::ARM::BI__builtin_arm_rbit) {
7766 llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
7767 return Builder.CreateCall(
7768 CGM.getIntrinsic(Intrinsic::bitreverse, Arg->getType()), Arg, "rbit");
7769 }
7770
7771 if (BuiltinID == clang::ARM::BI__builtin_arm_cls) {
7772 llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
7773 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_cls), Arg, "cls");
7774 }
7775 if (BuiltinID == clang::ARM::BI__builtin_arm_cls64) {
7776 llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
7777 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_cls64), Arg,
7778 "cls");
7779 }
7780
7781 if (BuiltinID == clang::ARM::BI__clear_cache) {
7782 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments");
7783 const FunctionDecl *FD = E->getDirectCallee();
7784 Value *Ops[2];
7785 for (unsigned i = 0; i < 2; i++)
7786 Ops[i] = EmitScalarExpr(E->getArg(i));
7787 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType());
7788 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty);
7789 StringRef Name = FD->getName();
7790 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops);
7791 }
7792
7793 if (BuiltinID == clang::ARM::BI__builtin_arm_mcrr ||
7794 BuiltinID == clang::ARM::BI__builtin_arm_mcrr2) {
7795 Function *F;
7796
7797 switch (BuiltinID) {
7798 default: llvm_unreachable("unexpected builtin");
7799 case clang::ARM::BI__builtin_arm_mcrr:
7800 F = CGM.getIntrinsic(Intrinsic::arm_mcrr);
7801 break;
7802 case clang::ARM::BI__builtin_arm_mcrr2:
7803 F = CGM.getIntrinsic(Intrinsic::arm_mcrr2);
7804 break;
7805 }
7806
7807 // MCRR{2} instruction has 5 operands but
7808 // the intrinsic has 4 because Rt and Rt2
7809 // are represented as a single unsigned 64
7810 // bit integer in the intrinsic definition
7811 // but internally it's represented as 2 32
7812 // bit integers.
7813
7814 Value *Coproc = EmitScalarExpr(E->getArg(0));
7815 Value *Opc1 = EmitScalarExpr(E->getArg(1));
7816 Value *RtAndRt2 = EmitScalarExpr(E->getArg(2));
7817 Value *CRm = EmitScalarExpr(E->getArg(3));
7818
7819 Value *C1 = llvm::ConstantInt::get(Int64Ty, 32);
7820 Value *Rt = Builder.CreateTruncOrBitCast(RtAndRt2, Int32Ty);
7821 Value *Rt2 = Builder.CreateLShr(RtAndRt2, C1);
7822 Rt2 = Builder.CreateTruncOrBitCast(Rt2, Int32Ty);
7823
7824 return Builder.CreateCall(F, {Coproc, Opc1, Rt, Rt2, CRm});
7825 }
7826
7827 if (BuiltinID == clang::ARM::BI__builtin_arm_mrrc ||
7828 BuiltinID == clang::ARM::BI__builtin_arm_mrrc2) {
7829 Function *F;
7830
7831 switch (BuiltinID) {
7832 default: llvm_unreachable("unexpected builtin");
7833 case clang::ARM::BI__builtin_arm_mrrc:
7834 F = CGM.getIntrinsic(Intrinsic::arm_mrrc);
7835 break;
7836 case clang::ARM::BI__builtin_arm_mrrc2:
7837 F = CGM.getIntrinsic(Intrinsic::arm_mrrc2);
7838 break;
7839 }
7840
7841 Value *Coproc = EmitScalarExpr(E->getArg(0));
7842 Value *Opc1 = EmitScalarExpr(E->getArg(1));
7843 Value *CRm = EmitScalarExpr(E->getArg(2));
7844 Value *RtAndRt2 = Builder.CreateCall(F, {Coproc, Opc1, CRm});
7845
7846 // Returns an unsigned 64 bit integer, represented
7847 // as two 32 bit integers.
7848
7849 Value *Rt = Builder.CreateExtractValue(RtAndRt2, 1);
7850 Value *Rt1 = Builder.CreateExtractValue(RtAndRt2, 0);
7851 Rt = Builder.CreateZExt(Rt, Int64Ty);
7852 Rt1 = Builder.CreateZExt(Rt1, Int64Ty);
7853
7854 Value *ShiftCast = llvm::ConstantInt::get(Int64Ty, 32);
7855 RtAndRt2 = Builder.CreateShl(Rt, ShiftCast, "shl", true);
7856 RtAndRt2 = Builder.CreateOr(RtAndRt2, Rt1);
7857
7858 return Builder.CreateBitCast(RtAndRt2, ConvertType(E->getType()));
7859 }
7860
7861 if (BuiltinID == clang::ARM::BI__builtin_arm_ldrexd ||
7862 ((BuiltinID == clang::ARM::BI__builtin_arm_ldrex ||
7863 BuiltinID == clang::ARM::BI__builtin_arm_ldaex) &&
7864 getContext().getTypeSize(E->getType()) == 64) ||
7865 BuiltinID == clang::ARM::BI__ldrexd) {
7866 Function *F;
7867
7868 switch (BuiltinID) {
7869 default: llvm_unreachable("unexpected builtin");
7870 case clang::ARM::BI__builtin_arm_ldaex:
7871 F = CGM.getIntrinsic(Intrinsic::arm_ldaexd);
7872 break;
7873 case clang::ARM::BI__builtin_arm_ldrexd:
7874 case clang::ARM::BI__builtin_arm_ldrex:
7875 case clang::ARM::BI__ldrexd:
7876 F = CGM.getIntrinsic(Intrinsic::arm_ldrexd);
7877 break;
7878 }
7879
7880 Value *LdPtr = EmitScalarExpr(E->getArg(0));
7881 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy),
7882 "ldrexd");
7883
7884 Value *Val0 = Builder.CreateExtractValue(Val, 1);
7885 Value *Val1 = Builder.CreateExtractValue(Val, 0);
7886 Val0 = Builder.CreateZExt(Val0, Int64Ty);
7887 Val1 = Builder.CreateZExt(Val1, Int64Ty);
7888
7889 Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32);
7890 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */);
7891 Val = Builder.CreateOr(Val, Val1);
7892 return Builder.CreateBitCast(Val, ConvertType(E->getType()));
7893 }
7894
7895 if (BuiltinID == clang::ARM::BI__builtin_arm_ldrex ||
7896 BuiltinID == clang::ARM::BI__builtin_arm_ldaex) {
7897 Value *LoadAddr = EmitScalarExpr(E->getArg(0));
7898
7899 QualType Ty = E->getType();
7900 llvm::Type *RealResTy = ConvertType(Ty);
7901 llvm::Type *IntTy =
7902 llvm::IntegerType::get(getLLVMContext(), getContext().getTypeSize(Ty));
7903 llvm::Type *PtrTy = IntTy->getPointerTo();
7904 LoadAddr = Builder.CreateBitCast(LoadAddr, PtrTy);
7905
7906 Function *F = CGM.getIntrinsic(
7907 BuiltinID == clang::ARM::BI__builtin_arm_ldaex ? Intrinsic::arm_ldaex
7908 : Intrinsic::arm_ldrex,
7909 PtrTy);
7910 CallInst *Val = Builder.CreateCall(F, LoadAddr, "ldrex");
7911 Val->addParamAttr(
7912 0, Attribute::get(getLLVMContext(), Attribute::ElementType, IntTy));
7913
7914 if (RealResTy->isPointerTy())
7915 return Builder.CreateIntToPtr(Val, RealResTy);
7916 else {
7917 llvm::Type *IntResTy = llvm::IntegerType::get(
7918 getLLVMContext(), CGM.getDataLayout().getTypeSizeInBits(RealResTy));
7919 return Builder.CreateBitCast(Builder.CreateTruncOrBitCast(Val, IntResTy),
7920 RealResTy);
7921 }
7922 }
7923
7924 if (BuiltinID == clang::ARM::BI__builtin_arm_strexd ||
7925 ((BuiltinID == clang::ARM::BI__builtin_arm_stlex ||
7926 BuiltinID == clang::ARM::BI__builtin_arm_strex) &&
7927 getContext().getTypeSize(E->getArg(0)->getType()) == 64)) {
7928 Function *F = CGM.getIntrinsic(
7929 BuiltinID == clang::ARM::BI__builtin_arm_stlex ? Intrinsic::arm_stlexd
7930 : Intrinsic::arm_strexd);
7931 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty);
7932
7933 Address Tmp = CreateMemTemp(E->getArg(0)->getType());
7934 Value *Val = EmitScalarExpr(E->getArg(0));
7935 Builder.CreateStore(Val, Tmp);
7936
7937 Address LdPtr = Builder.CreateElementBitCast(Tmp, STy);
7938 Val = Builder.CreateLoad(LdPtr);
7939
7940 Value *Arg0 = Builder.CreateExtractValue(Val, 0);
7941 Value *Arg1 = Builder.CreateExtractValue(Val, 1);
7942 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), Int8PtrTy);
7943 return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "strexd");
7944 }
7945
7946 if (BuiltinID == clang::ARM::BI__builtin_arm_strex ||
7947 BuiltinID == clang::ARM::BI__builtin_arm_stlex) {
7948 Value *StoreVal = EmitScalarExpr(E->getArg(0));
7949 Value *StoreAddr = EmitScalarExpr(E->getArg(1));
7950
7951 QualType Ty = E->getArg(0)->getType();
7952 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(),
7953 getContext().getTypeSize(Ty));
7954 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo());
7955
7956 if (StoreVal->getType()->isPointerTy())
7957 StoreVal = Builder.CreatePtrToInt(StoreVal, Int32Ty);
7958 else {
7959 llvm::Type *IntTy = llvm::IntegerType::get(
7960 getLLVMContext(),
7961 CGM.getDataLayout().getTypeSizeInBits(StoreVal->getType()));
7962 StoreVal = Builder.CreateBitCast(StoreVal, IntTy);
7963 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int32Ty);
7964 }
7965
7966 Function *F = CGM.getIntrinsic(
7967 BuiltinID == clang::ARM::BI__builtin_arm_stlex ? Intrinsic::arm_stlex
7968 : Intrinsic::arm_strex,
7969 StoreAddr->getType());
7970
7971 CallInst *CI = Builder.CreateCall(F, {StoreVal, StoreAddr}, "strex");
7972 CI->addParamAttr(
7973 1, Attribute::get(getLLVMContext(), Attribute::ElementType, StoreTy));
7974 return CI;
7975 }
7976
7977 if (BuiltinID == clang::ARM::BI__builtin_arm_clrex) {
7978 Function *F = CGM.getIntrinsic(Intrinsic::arm_clrex);
7979 return Builder.CreateCall(F);
7980 }
7981
7982 // CRC32
7983 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic;
7984 switch (BuiltinID) {
7985 case clang::ARM::BI__builtin_arm_crc32b:
7986 CRCIntrinsicID = Intrinsic::arm_crc32b; break;
7987 case clang::ARM::BI__builtin_arm_crc32cb:
7988 CRCIntrinsicID = Intrinsic::arm_crc32cb; break;
7989 case clang::ARM::BI__builtin_arm_crc32h:
7990 CRCIntrinsicID = Intrinsic::arm_crc32h; break;
7991 case clang::ARM::BI__builtin_arm_crc32ch:
7992 CRCIntrinsicID = Intrinsic::arm_crc32ch; break;
7993 case clang::ARM::BI__builtin_arm_crc32w:
7994 case clang::ARM::BI__builtin_arm_crc32d:
7995 CRCIntrinsicID = Intrinsic::arm_crc32w; break;
7996 case clang::ARM::BI__builtin_arm_crc32cw:
7997 case clang::ARM::BI__builtin_arm_crc32cd:
7998 CRCIntrinsicID = Intrinsic::arm_crc32cw; break;
7999 }
8000
8001 if (CRCIntrinsicID != Intrinsic::not_intrinsic) {
8002 Value *Arg0 = EmitScalarExpr(E->getArg(0));
8003 Value *Arg1 = EmitScalarExpr(E->getArg(1));
8004
8005 // crc32{c,}d intrinsics are implemented as two calls to crc32{c,}w
8006 // intrinsics, hence we need different codegen for these cases.
8007 if (BuiltinID == clang::ARM::BI__builtin_arm_crc32d ||
8008 BuiltinID == clang::ARM::BI__builtin_arm_crc32cd) {
8009 Value *C1 = llvm::ConstantInt::get(Int64Ty, 32);
8010 Value *Arg1a = Builder.CreateTruncOrBitCast(Arg1, Int32Ty);
8011 Value *Arg1b = Builder.CreateLShr(Arg1, C1);
8012 Arg1b = Builder.CreateTruncOrBitCast(Arg1b, Int32Ty);
8013
8014 Function *F = CGM.getIntrinsic(CRCIntrinsicID);
8015 Value *Res = Builder.CreateCall(F, {Arg0, Arg1a});
8016 return Builder.CreateCall(F, {Res, Arg1b});
8017 } else {
8018 Arg1 = Builder.CreateZExtOrBitCast(Arg1, Int32Ty);
8019
8020 Function *F = CGM.getIntrinsic(CRCIntrinsicID);
8021 return Builder.CreateCall(F, {Arg0, Arg1});
8022 }
8023 }
8024
8025 if (BuiltinID == clang::ARM::BI__builtin_arm_rsr ||
8026 BuiltinID == clang::ARM::BI__builtin_arm_rsr64 ||
8027 BuiltinID == clang::ARM::BI__builtin_arm_rsrp ||
8028 BuiltinID == clang::ARM::BI__builtin_arm_wsr ||
8029 BuiltinID == clang::ARM::BI__builtin_arm_wsr64 ||
8030 BuiltinID == clang::ARM::BI__builtin_arm_wsrp) {
8031
8032 SpecialRegisterAccessKind AccessKind = Write;
8033 if (BuiltinID == clang::ARM::BI__builtin_arm_rsr ||
8034 BuiltinID == clang::ARM::BI__builtin_arm_rsr64 ||
8035 BuiltinID == clang::ARM::BI__builtin_arm_rsrp)
8036 AccessKind = VolatileRead;
8037
8038 bool IsPointerBuiltin = BuiltinID == clang::ARM::BI__builtin_arm_rsrp ||
8039 BuiltinID == clang::ARM::BI__builtin_arm_wsrp;
8040
8041 bool Is64Bit = BuiltinID == clang::ARM::BI__builtin_arm_rsr64 ||
8042 BuiltinID == clang::ARM::BI__builtin_arm_wsr64;
8043
8044 llvm::Type *ValueType;
8045 llvm::Type *RegisterType;
8046 if (IsPointerBuiltin) {
8047 ValueType = VoidPtrTy;
8048 RegisterType = Int32Ty;
8049 } else if (Is64Bit) {
8050 ValueType = RegisterType = Int64Ty;
8051 } else {
8052 ValueType = RegisterType = Int32Ty;
8053 }
8054
8055 return EmitSpecialRegisterBuiltin(*this, E, RegisterType, ValueType,
8056 AccessKind);
8057 }
8058
8059 if (BuiltinID == ARM::BI__builtin_sponentry) {
8060 llvm::Function *F = CGM.getIntrinsic(Intrinsic::sponentry, AllocaInt8PtrTy);
8061 return Builder.CreateCall(F);
8062 }
8063
8064 // Handle MSVC intrinsics before argument evaluation to prevent double
8065 // evaluation.
8066 if (std::optional<MSVCIntrin> MsvcIntId = translateArmToMsvcIntrin(BuiltinID))
8067 return EmitMSVCBuiltinExpr(*MsvcIntId, E);
8068
8069 // Deal with MVE builtins
8070 if (Value *Result = EmitARMMVEBuiltinExpr(BuiltinID, E, ReturnValue, Arch))
8071 return Result;
8072 // Handle CDE builtins
8073 if (Value *Result = EmitARMCDEBuiltinExpr(BuiltinID, E, ReturnValue, Arch))
8074 return Result;
8075
8076 // Some intrinsics are equivalent - if they are use the base intrinsic ID.
8077 auto It = llvm::find_if(NEONEquivalentIntrinsicMap, [BuiltinID](auto &P) {
8078 return P.first == BuiltinID;
8079 });
8080 if (It != end(NEONEquivalentIntrinsicMap))
8081 BuiltinID = It->second;
8082
8083 // Find out if any arguments are required to be integer constant
8084 // expressions.
8085 unsigned ICEArguments = 0;
8086 ASTContext::GetBuiltinTypeError Error;
8087 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments);
8088 assert(Error == ASTContext::GE_None && "Should not codegen an error");
8089
8090 auto getAlignmentValue32 = [&](Address addr) -> Value* {
8091 return Builder.getInt32(addr.getAlignment().getQuantity());
8092 };
8093
8094 Address PtrOp0 = Address::invalid();
8095 Address PtrOp1 = Address::invalid();
8096 SmallVector<Value*, 4> Ops;
8097 bool HasExtraArg = HasExtraNeonArgument(BuiltinID);
8098 unsigned NumArgs = E->getNumArgs() - (HasExtraArg ? 1 : 0);
8099 for (unsigned i = 0, e = NumArgs; i != e; i++) {
8100 if (i == 0) {
8101 switch (BuiltinID) {
8102 case NEON::BI__builtin_neon_vld1_v:
8103 case NEON::BI__builtin_neon_vld1q_v:
8104 case NEON::BI__builtin_neon_vld1q_lane_v:
8105 case NEON::BI__builtin_neon_vld1_lane_v:
8106 case NEON::BI__builtin_neon_vld1_dup_v:
8107 case NEON::BI__builtin_neon_vld1q_dup_v:
8108 case NEON::BI__builtin_neon_vst1_v:
8109 case NEON::BI__builtin_neon_vst1q_v:
8110 case NEON::BI__builtin_neon_vst1q_lane_v:
8111 case NEON::BI__builtin_neon_vst1_lane_v:
8112 case NEON::BI__builtin_neon_vst2_v:
8113 case NEON::BI__builtin_neon_vst2q_v:
8114 case NEON::BI__builtin_neon_vst2_lane_v:
8115 case NEON::BI__builtin_neon_vst2q_lane_v:
8116 case NEON::BI__builtin_neon_vst3_v:
8117 case NEON::BI__builtin_neon_vst3q_v:
8118 case NEON::BI__builtin_neon_vst3_lane_v:
8119 case NEON::BI__builtin_neon_vst3q_lane_v:
8120 case NEON::BI__builtin_neon_vst4_v:
8121 case NEON::BI__builtin_neon_vst4q_v:
8122 case NEON::BI__builtin_neon_vst4_lane_v:
8123 case NEON::BI__builtin_neon_vst4q_lane_v:
8124 // Get the alignment for the argument in addition to the value;
8125 // we'll use it later.
8126 PtrOp0 = EmitPointerWithAlignment(E->getArg(0));
8127 Ops.push_back(PtrOp0.getPointer());
8128 continue;
8129 }
8130 }
8131 if (i == 1) {
8132 switch (BuiltinID) {
8133 case NEON::BI__builtin_neon_vld2_v:
8134 case NEON::BI__builtin_neon_vld2q_v:
8135 case NEON::BI__builtin_neon_vld3_v:
8136 case NEON::BI__builtin_neon_vld3q_v:
8137 case NEON::BI__builtin_neon_vld4_v:
8138 case NEON::BI__builtin_neon_vld4q_v:
8139 case NEON::BI__builtin_neon_vld2_lane_v:
8140 case NEON::BI__builtin_neon_vld2q_lane_v:
8141 case NEON::BI__builtin_neon_vld3_lane_v:
8142 case NEON::BI__builtin_neon_vld3q_lane_v:
8143 case NEON::BI__builtin_neon_vld4_lane_v:
8144 case NEON::BI__builtin_neon_vld4q_lane_v:
8145 case NEON::BI__builtin_neon_vld2_dup_v:
8146 case NEON::BI__builtin_neon_vld2q_dup_v:
8147 case NEON::BI__builtin_neon_vld3_dup_v:
8148 case NEON::BI__builtin_neon_vld3q_dup_v:
8149 case NEON::BI__builtin_neon_vld4_dup_v:
8150 case NEON::BI__builtin_neon_vld4q_dup_v:
8151 // Get the alignment for the argument in addition to the value;
8152 // we'll use it later.
8153 PtrOp1 = EmitPointerWithAlignment(E->getArg(1));
8154 Ops.push_back(PtrOp1.getPointer());
8155 continue;
8156 }
8157 }
8158
8159 if ((ICEArguments & (1 << i)) == 0) {
8160 Ops.push_back(EmitScalarExpr(E->getArg(i)));
8161 } else {
8162 // If this is required to be a constant, constant fold it so that we know
8163 // that the generated intrinsic gets a ConstantInt.
8164 Ops.push_back(llvm::ConstantInt::get(
8165 getLLVMContext(),
8166 *E->getArg(i)->getIntegerConstantExpr(getContext())));
8167 }
8168 }
8169
8170 switch (BuiltinID) {
8171 default: break;
8172
8173 case NEON::BI__builtin_neon_vget_lane_i8:
8174 case NEON::BI__builtin_neon_vget_lane_i16:
8175 case NEON::BI__builtin_neon_vget_lane_i32:
8176 case NEON::BI__builtin_neon_vget_lane_i64:
8177 case NEON::BI__builtin_neon_vget_lane_bf16:
8178 case NEON::BI__builtin_neon_vget_lane_f32:
8179 case NEON::BI__builtin_neon_vgetq_lane_i8:
8180 case NEON::BI__builtin_neon_vgetq_lane_i16:
8181 case NEON::BI__builtin_neon_vgetq_lane_i32:
8182 case NEON::BI__builtin_neon_vgetq_lane_i64:
8183 case NEON::BI__builtin_neon_vgetq_lane_bf16:
8184 case NEON::BI__builtin_neon_vgetq_lane_f32:
8185 case NEON::BI__builtin_neon_vduph_lane_bf16:
8186 case NEON::BI__builtin_neon_vduph_laneq_bf16:
8187 return Builder.CreateExtractElement(Ops[0], Ops[1], "vget_lane");
8188
8189 case NEON::BI__builtin_neon_vrndns_f32: {
8190 Value *Arg = EmitScalarExpr(E->getArg(0));
8191 llvm::Type *Tys[] = {Arg->getType()};
8192 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vrintn, Tys);
8193 return Builder.CreateCall(F, {Arg}, "vrndn"); }
8194
8195 case NEON::BI__builtin_neon_vset_lane_i8:
8196 case NEON::BI__builtin_neon_vset_lane_i16:
8197 case NEON::BI__builtin_neon_vset_lane_i32:
8198 case NEON::BI__builtin_neon_vset_lane_i64:
8199 case NEON::BI__builtin_neon_vset_lane_bf16:
8200 case NEON::BI__builtin_neon_vset_lane_f32:
8201 case NEON::BI__builtin_neon_vsetq_lane_i8:
8202 case NEON::BI__builtin_neon_vsetq_lane_i16:
8203 case NEON::BI__builtin_neon_vsetq_lane_i32:
8204 case NEON::BI__builtin_neon_vsetq_lane_i64:
8205 case NEON::BI__builtin_neon_vsetq_lane_bf16:
8206 case NEON::BI__builtin_neon_vsetq_lane_f32:
8207 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane");
8208
8209 case NEON::BI__builtin_neon_vsha1h_u32:
8210 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1h), Ops,
8211 "vsha1h");
8212 case NEON::BI__builtin_neon_vsha1cq_u32:
8213 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1c), Ops,
8214 "vsha1h");
8215 case NEON::BI__builtin_neon_vsha1pq_u32:
8216 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1p), Ops,
8217 "vsha1h");
8218 case NEON::BI__builtin_neon_vsha1mq_u32:
8219 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1m), Ops,
8220 "vsha1h");
8221
8222 case NEON::BI__builtin_neon_vcvth_bf16_f32: {
8223 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vcvtbfp2bf), Ops,
8224 "vcvtbfp2bf");
8225 }
8226
8227 // The ARM _MoveToCoprocessor builtins put the input register value as
8228 // the first argument, but the LLVM intrinsic expects it as the third one.
8229 case clang::ARM::BI_MoveToCoprocessor:
8230 case clang::ARM::BI_MoveToCoprocessor2: {
8231 Function *F = CGM.getIntrinsic(BuiltinID == clang::ARM::BI_MoveToCoprocessor
8232 ? Intrinsic::arm_mcr
8233 : Intrinsic::arm_mcr2);
8234 return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0],
8235 Ops[3], Ops[4], Ops[5]});
8236 }
8237 }
8238
8239 // Get the last argument, which specifies the vector type.
8240 assert(HasExtraArg);
8241 const Expr *Arg = E->getArg(E->getNumArgs()-1);
8242 std::optional<llvm::APSInt> Result =
8243 Arg->getIntegerConstantExpr(getContext());
8244 if (!Result)
8245 return nullptr;
8246
8247 if (BuiltinID == clang::ARM::BI__builtin_arm_vcvtr_f ||
8248 BuiltinID == clang::ARM::BI__builtin_arm_vcvtr_d) {
8249 // Determine the overloaded type of this builtin.
8250 llvm::Type *Ty;
8251 if (BuiltinID == clang::ARM::BI__builtin_arm_vcvtr_f)
8252 Ty = FloatTy;
8253 else
8254 Ty = DoubleTy;
8255
8256 // Determine whether this is an unsigned conversion or not.
8257 bool usgn = Result->getZExtValue() == 1;
8258 unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr;
8259
8260 // Call the appropriate intrinsic.
8261 Function *F = CGM.getIntrinsic(Int, Ty);
8262 return Builder.CreateCall(F, Ops, "vcvtr");
8263 }
8264
8265 // Determine the type of this overloaded NEON intrinsic.
8266 NeonTypeFlags Type = Result->getZExtValue();
8267 bool usgn = Type.isUnsigned();
8268 bool rightShift = false;
8269
8270 llvm::FixedVectorType *VTy =
8271 GetNeonType(this, Type, getTarget().hasLegalHalfType(), false,
8272 getTarget().hasBFloat16Type());
8273 llvm::Type *Ty = VTy;
8274 if (!Ty)
8275 return nullptr;
8276
8277 // Many NEON builtins have identical semantics and uses in ARM and
8278 // AArch64. Emit these in a single function.
8279 auto IntrinsicMap = ArrayRef(ARMSIMDIntrinsicMap);
8280 const ARMVectorIntrinsicInfo *Builtin = findARMVectorIntrinsicInMap(
8281 IntrinsicMap, BuiltinID, NEONSIMDIntrinsicsProvenSorted);
8282 if (Builtin)
8283 return EmitCommonNeonBuiltinExpr(
8284 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic,
8285 Builtin->NameHint, Builtin->TypeModifier, E, Ops, PtrOp0, PtrOp1, Arch);
8286
8287 unsigned Int;
8288 switch (BuiltinID) {
8289 default: return nullptr;
8290 case NEON::BI__builtin_neon_vld1q_lane_v:
8291 // Handle 64-bit integer elements as a special case. Use shuffles of
8292 // one-element vectors to avoid poor code for i64 in the backend.
8293 if (VTy->getElementType()->isIntegerTy(64)) {
8294 // Extract the other lane.
8295 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
8296 int Lane = cast<ConstantInt>(Ops[2])->getZExtValue();
8297 Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane));
8298 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV);
8299 // Load the value as a one-element vector.
8300 Ty = llvm::FixedVectorType::get(VTy->getElementType(), 1);
8301 llvm::Type *Tys[] = {Ty, Int8PtrTy};
8302 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Tys);
8303 Value *Align = getAlignmentValue32(PtrOp0);
8304 Value *Ld = Builder.CreateCall(F, {Ops[0], Align});
8305 // Combine them.
8306 int Indices[] = {1 - Lane, Lane};
8307 return Builder.CreateShuffleVector(Ops[1], Ld, Indices, "vld1q_lane");
8308 }
8309 [[fallthrough]];
8310 case NEON::BI__builtin_neon_vld1_lane_v: {
8311 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
8312 PtrOp0 = Builder.CreateElementBitCast(PtrOp0, VTy->getElementType());
8313 Value *Ld = Builder.CreateLoad(PtrOp0);
8314 return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane");
8315 }
8316 case NEON::BI__builtin_neon_vqrshrn_n_v:
8317 Int =
8318 usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns;
8319 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n",
8320 1, true);
8321 case NEON::BI__builtin_neon_vqrshrun_n_v:
8322 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty),
8323 Ops, "vqrshrun_n", 1, true);
8324 case NEON::BI__builtin_neon_vqshrn_n_v:
8325 Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns;
8326 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n",
8327 1, true);
8328 case NEON::BI__builtin_neon_vqshrun_n_v:
8329 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty),
8330 Ops, "vqshrun_n", 1, true);
8331 case NEON::BI__builtin_neon_vrecpe_v:
8332 case NEON::BI__builtin_neon_vrecpeq_v:
8333 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty),
8334 Ops, "vrecpe");
8335 case NEON::BI__builtin_neon_vrshrn_n_v:
8336 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty),
8337 Ops, "vrshrn_n", 1, true);
8338 case NEON::BI__builtin_neon_vrsra_n_v:
8339 case NEON::BI__builtin_neon_vrsraq_n_v:
8340 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
8341 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
8342 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true);
8343 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts;
8344 Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Ty), {Ops[1], Ops[2]});
8345 return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n");
8346 case NEON::BI__builtin_neon_vsri_n_v:
8347 case NEON::BI__builtin_neon_vsriq_n_v:
8348 rightShift = true;
8349 [[fallthrough]];
8350 case NEON::BI__builtin_neon_vsli_n_v:
8351 case NEON::BI__builtin_neon_vsliq_n_v:
8352 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift);
8353 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty),
8354 Ops, "vsli_n");
8355 case NEON::BI__builtin_neon_vsra_n_v:
8356 case NEON::BI__builtin_neon_vsraq_n_v:
8357 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
8358 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n");
8359 return Builder.CreateAdd(Ops[0], Ops[1]);
8360 case NEON::BI__builtin_neon_vst1q_lane_v:
8361 // Handle 64-bit integer elements as a special case. Use a shuffle to get
8362 // a one-element vector and avoid poor code for i64 in the backend.
8363 if (VTy->getElementType()->isIntegerTy(64)) {
8364 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
8365 Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2]));
8366 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV);
8367 Ops[2] = getAlignmentValue32(PtrOp0);
8368 llvm::Type *Tys[] = {Int8PtrTy, Ops[1]->getType()};
8369 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1,
8370 Tys), Ops);
8371 }
8372 [[fallthrough]];
8373 case NEON::BI__builtin_neon_vst1_lane_v: {
8374 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
8375 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]);
8376 auto St = Builder.CreateStore(
8377 Ops[1], Builder.CreateElementBitCast(PtrOp0, Ops[1]->getType()));
8378 return St;
8379 }
8380 case NEON::BI__builtin_neon_vtbl1_v:
8381 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1),
8382 Ops, "vtbl1");
8383 case NEON::BI__builtin_neon_vtbl2_v:
8384 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2),
8385 Ops, "vtbl2");
8386 case NEON::BI__builtin_neon_vtbl3_v:
8387 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3),
8388 Ops, "vtbl3");
8389 case NEON::BI__builtin_neon_vtbl4_v:
8390 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4),
8391 Ops, "vtbl4");
8392 case NEON::BI__builtin_neon_vtbx1_v:
8393 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1),
8394 Ops, "vtbx1");
8395 case NEON::BI__builtin_neon_vtbx2_v:
8396 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2),
8397 Ops, "vtbx2");
8398 case NEON::BI__builtin_neon_vtbx3_v:
8399 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3),
8400 Ops, "vtbx3");
8401 case NEON::BI__builtin_neon_vtbx4_v:
8402 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4),
8403 Ops, "vtbx4");
8404 }
8405 }
8406
8407 template<typename Integer>
8408 static Integer GetIntegerConstantValue(const Expr *E, ASTContext &Context) {
8409 return E->getIntegerConstantExpr(Context)->getExtValue();
8410 }
8411
8412 static llvm::Value *SignOrZeroExtend(CGBuilderTy &Builder, llvm::Value *V,
8413 llvm::Type *T, bool Unsigned) {
8414 // Helper function called by Tablegen-constructed ARM MVE builtin codegen,
8415 // which finds it convenient to specify signed/unsigned as a boolean flag.
8416 return Unsigned ? Builder.CreateZExt(V, T) : Builder.CreateSExt(V, T);
8417 }
8418
8419 static llvm::Value *MVEImmediateShr(CGBuilderTy &Builder, llvm::Value *V,
8420 uint32_t Shift, bool Unsigned) {
8421 // MVE helper function for integer shift right. This must handle signed vs
8422 // unsigned, and also deal specially with the case where the shift count is
8423 // equal to the lane size. In LLVM IR, an LShr with that parameter would be
8424 // undefined behavior, but in MVE it's legal, so we must convert it to code
8425 // that is not undefined in IR.
8426 unsigned LaneBits = cast<llvm::VectorType>(V->getType())
8427 ->getElementType()
8428 ->getPrimitiveSizeInBits();
8429 if (Shift == LaneBits) {
8430 // An unsigned shift of the full lane size always generates zero, so we can
8431 // simply emit a zero vector. A signed shift of the full lane size does the
8432 // same thing as shifting by one bit fewer.
8433 if (Unsigned)
8434 return llvm::Constant::getNullValue(V->getType());
8435 else
8436 --Shift;
8437 }
8438 return Unsigned ? Builder.CreateLShr(V, Shift) : Builder.CreateAShr(V, Shift);
8439 }
8440
8441 static llvm::Value *ARMMVEVectorSplat(CGBuilderTy &Builder, llvm::Value *V) {
8442 // MVE-specific helper function for a vector splat, which infers the element
8443 // count of the output vector by knowing that MVE vectors are all 128 bits
8444 // wide.
8445 unsigned Elements = 128 / V->getType()->getPrimitiveSizeInBits();
8446 return Builder.CreateVectorSplat(Elements, V);
8447 }
8448
8449 static llvm::Value *ARMMVEVectorReinterpret(CGBuilderTy &Builder,
8450 CodeGenFunction *CGF,
8451 llvm::Value *V,
8452 llvm::Type *DestType) {
8453 // Convert one MVE vector type into another by reinterpreting its in-register
8454 // format.
8455 //
8456 // Little-endian, this is identical to a bitcast (which reinterprets the
8457 // memory format). But big-endian, they're not necessarily the same, because
8458 // the register and memory formats map to each other differently depending on
8459 // the lane size.
8460 //
8461 // We generate a bitcast whenever we can (if we're little-endian, or if the
8462 // lane sizes are the same anyway). Otherwise we fall back to an IR intrinsic
8463 // that performs the different kind of reinterpretation.
8464 if (CGF->getTarget().isBigEndian() &&
8465 V->getType()->getScalarSizeInBits() != DestType->getScalarSizeInBits()) {
8466 return Builder.CreateCall(
8467 CGF->CGM.getIntrinsic(Intrinsic::arm_mve_vreinterpretq,
8468 {DestType, V->getType()}),
8469 V);
8470 } else {
8471 return Builder.CreateBitCast(V, DestType);
8472 }
8473 }
8474
8475 static llvm::Value *VectorUnzip(CGBuilderTy &Builder, llvm::Value *V, bool Odd) {
8476 // Make a shufflevector that extracts every other element of a vector (evens
8477 // or odds, as desired).
8478 SmallVector<int, 16> Indices;
8479 unsigned InputElements =
8480 cast<llvm::FixedVectorType>(V->getType())->getNumElements();
8481 for (unsigned i = 0; i < InputElements; i += 2)
8482 Indices.push_back(i + Odd);
8483 return Builder.CreateShuffleVector(V, Indices);
8484 }
8485
8486 static llvm::Value *VectorZip(CGBuilderTy &Builder, llvm::Value *V0,
8487 llvm::Value *V1) {
8488 // Make a shufflevector that interleaves two vectors element by element.
8489 assert(V0->getType() == V1->getType() && "Can't zip different vector types");
8490 SmallVector<int, 16> Indices;
8491 unsigned InputElements =
8492 cast<llvm::FixedVectorType>(V0->getType())->getNumElements();
8493 for (unsigned i = 0; i < InputElements; i++) {
8494 Indices.push_back(i);
8495 Indices.push_back(i + InputElements);
8496 }
8497 return Builder.CreateShuffleVector(V0, V1, Indices);
8498 }
8499
8500 template<unsigned HighBit, unsigned OtherBits>
8501 static llvm::Value *ARMMVEConstantSplat(CGBuilderTy &Builder, llvm::Type *VT) {
8502 // MVE-specific helper function to make a vector splat of a constant such as
8503 // UINT_MAX or INT_MIN, in which all bits below the highest one are equal.
8504 llvm::Type *T = cast<llvm::VectorType>(VT)->getElementType();
8505 unsigned LaneBits = T->getPrimitiveSizeInBits();
8506 uint32_t Value = HighBit << (LaneBits - 1);
8507 if (OtherBits)
8508 Value |= (1UL << (LaneBits - 1)) - 1;
8509 llvm::Value *Lane = llvm::ConstantInt::get(T, Value);
8510 return ARMMVEVectorSplat(Builder, Lane);
8511 }
8512
8513 static llvm::Value *ARMMVEVectorElementReverse(CGBuilderTy &Builder,
8514 llvm::Value *V,
8515 unsigned ReverseWidth) {
8516 // MVE-specific helper function which reverses the elements of a
8517 // vector within every (ReverseWidth)-bit collection of lanes.
8518 SmallVector<int, 16> Indices;
8519 unsigned LaneSize = V->getType()->getScalarSizeInBits();
8520 unsigned Elements = 128 / LaneSize;
8521 unsigned Mask = ReverseWidth / LaneSize - 1;
8522 for (unsigned i = 0; i < Elements; i++)
8523 Indices.push_back(i ^ Mask);
8524 return Builder.CreateShuffleVector(V, Indices);
8525 }
8526
8527 Value *CodeGenFunction::EmitARMMVEBuiltinExpr(unsigned BuiltinID,
8528 const CallExpr *E,
8529 ReturnValueSlot ReturnValue,
8530 llvm::Triple::ArchType Arch) {
8531 enum class CustomCodeGen { VLD24, VST24 } CustomCodeGenType;
8532 Intrinsic::ID IRIntr;
8533 unsigned NumVectors;
8534
8535 // Code autogenerated by Tablegen will handle all the simple builtins.
8536 switch (BuiltinID) {
8537 #include "clang/Basic/arm_mve_builtin_cg.inc"
8538
8539 // If we didn't match an MVE builtin id at all, go back to the
8540 // main EmitARMBuiltinExpr.
8541 default:
8542 return nullptr;
8543 }
8544
8545 // Anything that breaks from that switch is an MVE builtin that
8546 // needs handwritten code to generate.
8547
8548 switch (CustomCodeGenType) {
8549
8550 case CustomCodeGen::VLD24: {
8551 llvm::SmallVector<Value *, 4> Ops;
8552 llvm::SmallVector<llvm::Type *, 4> Tys;
8553
8554 auto MvecCType = E->getType();
8555 auto MvecLType = ConvertType(MvecCType);
8556 assert(MvecLType->isStructTy() &&
8557 "Return type for vld[24]q should be a struct");
8558 assert(MvecLType->getStructNumElements() == 1 &&
8559 "Return-type struct for vld[24]q should have one element");
8560 auto MvecLTypeInner = MvecLType->getStructElementType(0);
8561 assert(MvecLTypeInner->isArrayTy() &&
8562 "Return-type struct for vld[24]q should contain an array");
8563 assert(MvecLTypeInner->getArrayNumElements() == NumVectors &&
8564 "Array member of return-type struct vld[24]q has wrong length");
8565 auto VecLType = MvecLTypeInner->getArrayElementType();
8566
8567 Tys.push_back(VecLType);
8568
8569 auto Addr = E->getArg(0);
8570 Ops.push_back(EmitScalarExpr(Addr));
8571 Tys.push_back(ConvertType(Addr->getType()));
8572
8573 Function *F = CGM.getIntrinsic(IRIntr, ArrayRef(Tys));
8574 Value *LoadResult = Builder.CreateCall(F, Ops);
8575 Value *MvecOut = PoisonValue::get(MvecLType);
8576 for (unsigned i = 0; i < NumVectors; ++i) {
8577 Value *Vec = Builder.CreateExtractValue(LoadResult, i);
8578 MvecOut = Builder.CreateInsertValue(MvecOut, Vec, {0, i});
8579 }
8580
8581 if (ReturnValue.isNull())
8582 return MvecOut;
8583 else
8584 return Builder.CreateStore(MvecOut, ReturnValue.getValue());
8585 }
8586
8587 case CustomCodeGen::VST24: {
8588 llvm::SmallVector<Value *, 4> Ops;
8589 llvm::SmallVector<llvm::Type *, 4> Tys;
8590
8591 auto Addr = E->getArg(0);
8592 Ops.push_back(EmitScalarExpr(Addr));
8593 Tys.push_back(ConvertType(Addr->getType()));
8594
8595 auto MvecCType = E->getArg(1)->getType();
8596 auto MvecLType = ConvertType(MvecCType);
8597 assert(MvecLType->isStructTy() && "Data type for vst2q should be a struct");
8598 assert(MvecLType->getStructNumElements() == 1 &&
8599 "Data-type struct for vst2q should have one element");
8600 auto MvecLTypeInner = MvecLType->getStructElementType(0);
8601 assert(MvecLTypeInner->isArrayTy() &&
8602 "Data-type struct for vst2q should contain an array");
8603 assert(MvecLTypeInner->getArrayNumElements() == NumVectors &&
8604 "Array member of return-type struct vld[24]q has wrong length");
8605 auto VecLType = MvecLTypeInner->getArrayElementType();
8606
8607 Tys.push_back(VecLType);
8608
8609 AggValueSlot MvecSlot = CreateAggTemp(MvecCType);
8610 EmitAggExpr(E->getArg(1), MvecSlot);
8611 auto Mvec = Builder.CreateLoad(MvecSlot.getAddress());
8612 for (unsigned i = 0; i < NumVectors; i++)
8613 Ops.push_back(Builder.CreateExtractValue(Mvec, {0, i}));
8614
8615 Function *F = CGM.getIntrinsic(IRIntr, ArrayRef(Tys));
8616 Value *ToReturn = nullptr;
8617 for (unsigned i = 0; i < NumVectors; i++) {
8618 Ops.push_back(llvm::ConstantInt::get(Int32Ty, i));
8619 ToReturn = Builder.CreateCall(F, Ops);
8620 Ops.pop_back();
8621 }
8622 return ToReturn;
8623 }
8624 }
8625 llvm_unreachable("unknown custom codegen type.");
8626 }
8627
8628 Value *CodeGenFunction::EmitARMCDEBuiltinExpr(unsigned BuiltinID,
8629 const CallExpr *E,
8630 ReturnValueSlot ReturnValue,
8631 llvm::Triple::ArchType Arch) {
8632 switch (BuiltinID) {
8633 default:
8634 return nullptr;
8635 #include "clang/Basic/arm_cde_builtin_cg.inc"
8636 }
8637 }
8638
8639 static Value *EmitAArch64TblBuiltinExpr(CodeGenFunction &CGF, unsigned BuiltinID,
8640 const CallExpr *E,
8641 SmallVectorImpl<Value *> &Ops,
8642 llvm::Triple::ArchType Arch) {
8643 unsigned int Int = 0;
8644 const char *s = nullptr;
8645
8646 switch (BuiltinID) {
8647 default:
8648 return nullptr;
8649 case NEON::BI__builtin_neon_vtbl1_v:
8650 case NEON::BI__builtin_neon_vqtbl1_v:
8651 case NEON::BI__builtin_neon_vqtbl1q_v:
8652 case NEON::BI__builtin_neon_vtbl2_v:
8653 case NEON::BI__builtin_neon_vqtbl2_v:
8654 case NEON::BI__builtin_neon_vqtbl2q_v:
8655 case NEON::BI__builtin_neon_vtbl3_v:
8656 case NEON::BI__builtin_neon_vqtbl3_v:
8657 case NEON::BI__builtin_neon_vqtbl3q_v:
8658 case NEON::BI__builtin_neon_vtbl4_v:
8659 case NEON::BI__builtin_neon_vqtbl4_v:
8660 case NEON::BI__builtin_neon_vqtbl4q_v:
8661 break;
8662 case NEON::BI__builtin_neon_vtbx1_v:
8663 case NEON::BI__builtin_neon_vqtbx1_v:
8664 case NEON::BI__builtin_neon_vqtbx1q_v:
8665 case NEON::BI__builtin_neon_vtbx2_v:
8666 case NEON::BI__builtin_neon_vqtbx2_v:
8667 case NEON::BI__builtin_neon_vqtbx2q_v:
8668 case NEON::BI__builtin_neon_vtbx3_v:
8669 case NEON::BI__builtin_neon_vqtbx3_v:
8670 case NEON::BI__builtin_neon_vqtbx3q_v:
8671 case NEON::BI__builtin_neon_vtbx4_v:
8672 case NEON::BI__builtin_neon_vqtbx4_v:
8673 case NEON::BI__builtin_neon_vqtbx4q_v:
8674 break;
8675 }
8676
8677 assert(E->getNumArgs() >= 3);
8678
8679 // Get the last argument, which specifies the vector type.
8680 const Expr *Arg = E->getArg(E->getNumArgs() - 1);
8681 std::optional<llvm::APSInt> Result =
8682 Arg->getIntegerConstantExpr(CGF.getContext());
8683 if (!Result)
8684 return nullptr;
8685
8686 // Determine the type of this overloaded NEON intrinsic.
8687 NeonTypeFlags Type = Result->getZExtValue();
8688 llvm::FixedVectorType *Ty = GetNeonType(&CGF, Type);
8689 if (!Ty)
8690 return nullptr;
8691
8692 CodeGen::CGBuilderTy &Builder = CGF.Builder;
8693
8694 // AArch64 scalar builtins are not overloaded, they do not have an extra
8695 // argument that specifies the vector type, need to handle each case.
8696 switch (BuiltinID) {
8697 case NEON::BI__builtin_neon_vtbl1_v: {
8698 return packTBLDVectorList(CGF, ArrayRef(Ops).slice(0, 1), nullptr, Ops[1],
8699 Ty, Intrinsic::aarch64_neon_tbl1, "vtbl1");
8700 }
8701 case NEON::BI__builtin_neon_vtbl2_v: {
8702 return packTBLDVectorList(CGF, ArrayRef(Ops).slice(0, 2), nullptr, Ops[2],
8703 Ty, Intrinsic::aarch64_neon_tbl1, "vtbl1");
8704 }
8705 case NEON::BI__builtin_neon_vtbl3_v: {
8706 return packTBLDVectorList(CGF, ArrayRef(Ops).slice(0, 3), nullptr, Ops[3],
8707 Ty, Intrinsic::aarch64_neon_tbl2, "vtbl2");
8708 }
8709 case NEON::BI__builtin_neon_vtbl4_v: {
8710 return packTBLDVectorList(CGF, ArrayRef(Ops).slice(0, 4), nullptr, Ops[4],
8711 Ty, Intrinsic::aarch64_neon_tbl2, "vtbl2");
8712 }
8713 case NEON::BI__builtin_neon_vtbx1_v: {
8714 Value *TblRes =
8715 packTBLDVectorList(CGF, ArrayRef(Ops).slice(1, 1), nullptr, Ops[2], Ty,
8716 Intrinsic::aarch64_neon_tbl1, "vtbl1");
8717
8718 llvm::Constant *EightV = ConstantInt::get(Ty, 8);
8719 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[2], EightV);
8720 CmpRes = Builder.CreateSExt(CmpRes, Ty);
8721
8722 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]);
8723 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes);
8724 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx");
8725 }
8726 case NEON::BI__builtin_neon_vtbx2_v: {
8727 return packTBLDVectorList(CGF, ArrayRef(Ops).slice(1, 2), Ops[0], Ops[3],
8728 Ty, Intrinsic::aarch64_neon_tbx1, "vtbx1");
8729 }
8730 case NEON::BI__builtin_neon_vtbx3_v: {
8731 Value *TblRes =
8732 packTBLDVectorList(CGF, ArrayRef(Ops).slice(1, 3), nullptr, Ops[4], Ty,
8733 Intrinsic::aarch64_neon_tbl2, "vtbl2");
8734
8735 llvm::Constant *TwentyFourV = ConstantInt::get(Ty, 24);
8736 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[4],
8737 TwentyFourV);
8738 CmpRes = Builder.CreateSExt(CmpRes, Ty);
8739
8740 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]);
8741 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes);
8742 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx");
8743 }
8744 case NEON::BI__builtin_neon_vtbx4_v: {
8745 return packTBLDVectorList(CGF, ArrayRef(Ops).slice(1, 4), Ops[0], Ops[5],
8746 Ty, Intrinsic::aarch64_neon_tbx2, "vtbx2");
8747 }
8748 case NEON::BI__builtin_neon_vqtbl1_v:
8749 case NEON::BI__builtin_neon_vqtbl1q_v:
8750 Int = Intrinsic::aarch64_neon_tbl1; s = "vtbl1"; break;
8751 case NEON::BI__builtin_neon_vqtbl2_v:
8752 case NEON::BI__builtin_neon_vqtbl2q_v: {
8753 Int = Intrinsic::aarch64_neon_tbl2; s = "vtbl2"; break;
8754 case NEON::BI__builtin_neon_vqtbl3_v:
8755 case NEON::BI__builtin_neon_vqtbl3q_v:
8756 Int = Intrinsic::aarch64_neon_tbl3; s = "vtbl3"; break;
8757 case NEON::BI__builtin_neon_vqtbl4_v:
8758 case NEON::BI__builtin_neon_vqtbl4q_v:
8759 Int = Intrinsic::aarch64_neon_tbl4; s = "vtbl4"; break;
8760 case NEON::BI__builtin_neon_vqtbx1_v:
8761 case NEON::BI__builtin_neon_vqtbx1q_v:
8762 Int = Intrinsic::aarch64_neon_tbx1; s = "vtbx1"; break;
8763 case NEON::BI__builtin_neon_vqtbx2_v:
8764 case NEON::BI__builtin_neon_vqtbx2q_v:
8765 Int = Intrinsic::aarch64_neon_tbx2; s = "vtbx2"; break;
8766 case NEON::BI__builtin_neon_vqtbx3_v:
8767 case NEON::BI__builtin_neon_vqtbx3q_v:
8768 Int = Intrinsic::aarch64_neon_tbx3; s = "vtbx3"; break;
8769 case NEON::BI__builtin_neon_vqtbx4_v:
8770 case NEON::BI__builtin_neon_vqtbx4q_v:
8771 Int = Intrinsic::aarch64_neon_tbx4; s = "vtbx4"; break;
8772 }
8773 }
8774
8775 if (!Int)
8776 return nullptr;
8777
8778 Function *F = CGF.CGM.getIntrinsic(Int, Ty);
8779 return CGF.EmitNeonCall(F, Ops, s);
8780 }
8781
8782 Value *CodeGenFunction::vectorWrapScalar16(Value *Op) {
8783 auto *VTy = llvm::FixedVectorType::get(Int16Ty, 4);
8784 Op = Builder.CreateBitCast(Op, Int16Ty);
8785 Value *V = PoisonValue::get(VTy);
8786 llvm::Constant *CI = ConstantInt::get(SizeTy, 0);
8787 Op = Builder.CreateInsertElement(V, Op, CI);
8788 return Op;
8789 }
8790
8791 /// SVEBuiltinMemEltTy - Returns the memory element type for this memory
8792 /// access builtin. Only required if it can't be inferred from the base pointer
8793 /// operand.
8794 llvm::Type *CodeGenFunction::SVEBuiltinMemEltTy(const SVETypeFlags &TypeFlags) {
8795 switch (TypeFlags.getMemEltType()) {
8796 case SVETypeFlags::MemEltTyDefault:
8797 return getEltType(TypeFlags);
8798 case SVETypeFlags::MemEltTyInt8:
8799 return Builder.getInt8Ty();
8800 case SVETypeFlags::MemEltTyInt16:
8801 return Builder.getInt16Ty();
8802 case SVETypeFlags::MemEltTyInt32:
8803 return Builder.getInt32Ty();
8804 case SVETypeFlags::MemEltTyInt64:
8805 return Builder.getInt64Ty();
8806 }
8807 llvm_unreachable("Unknown MemEltType");
8808 }
8809
8810 llvm::Type *CodeGenFunction::getEltType(const SVETypeFlags &TypeFlags) {
8811 switch (TypeFlags.getEltType()) {
8812 default:
8813 llvm_unreachable("Invalid SVETypeFlag!");
8814
8815 case SVETypeFlags::EltTyInt8:
8816 return Builder.getInt8Ty();
8817 case SVETypeFlags::EltTyInt16:
8818 return Builder.getInt16Ty();
8819 case SVETypeFlags::EltTyInt32:
8820 return Builder.getInt32Ty();
8821 case SVETypeFlags::EltTyInt64:
8822 return Builder.getInt64Ty();
8823
8824 case SVETypeFlags::EltTyFloat16:
8825 return Builder.getHalfTy();
8826 case SVETypeFlags::EltTyFloat32:
8827 return Builder.getFloatTy();
8828 case SVETypeFlags::EltTyFloat64:
8829 return Builder.getDoubleTy();
8830
8831 case SVETypeFlags::EltTyBFloat16:
8832 return Builder.getBFloatTy();
8833
8834 case SVETypeFlags::EltTyBool8:
8835 case SVETypeFlags::EltTyBool16:
8836 case SVETypeFlags::EltTyBool32:
8837 case SVETypeFlags::EltTyBool64:
8838 return Builder.getInt1Ty();
8839 }
8840 }
8841
8842 // Return the llvm predicate vector type corresponding to the specified element
8843 // TypeFlags.
8844 llvm::ScalableVectorType *
8845 CodeGenFunction::getSVEPredType(const SVETypeFlags &TypeFlags) {
8846 switch (TypeFlags.getEltType()) {
8847 default: llvm_unreachable("Unhandled SVETypeFlag!");
8848
8849 case SVETypeFlags::EltTyInt8:
8850 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 16);
8851 case SVETypeFlags::EltTyInt16:
8852 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 8);
8853 case SVETypeFlags::EltTyInt32:
8854 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 4);
8855 case SVETypeFlags::EltTyInt64:
8856 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 2);
8857
8858 case SVETypeFlags::EltTyBFloat16:
8859 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 8);
8860 case SVETypeFlags::EltTyFloat16:
8861 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 8);
8862 case SVETypeFlags::EltTyFloat32:
8863 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 4);
8864 case SVETypeFlags::EltTyFloat64:
8865 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 2);
8866
8867 case SVETypeFlags::EltTyBool8:
8868 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 16);
8869 case SVETypeFlags::EltTyBool16:
8870 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 8);
8871 case SVETypeFlags::EltTyBool32:
8872 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 4);
8873 case SVETypeFlags::EltTyBool64:
8874 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 2);
8875 }
8876 }
8877
8878 // Return the llvm vector type corresponding to the specified element TypeFlags.
8879 llvm::ScalableVectorType *
8880 CodeGenFunction::getSVEType(const SVETypeFlags &TypeFlags) {
8881 switch (TypeFlags.getEltType()) {
8882 default:
8883 llvm_unreachable("Invalid SVETypeFlag!");
8884
8885 case SVETypeFlags::EltTyInt8:
8886 return llvm::ScalableVectorType::get(Builder.getInt8Ty(), 16);
8887 case SVETypeFlags::EltTyInt16:
8888 return llvm::ScalableVectorType::get(Builder.getInt16Ty(), 8);
8889 case SVETypeFlags::EltTyInt32:
8890 return llvm::ScalableVectorType::get(Builder.getInt32Ty(), 4);
8891 case SVETypeFlags::EltTyInt64:
8892 return llvm::ScalableVectorType::get(Builder.getInt64Ty(), 2);
8893
8894 case SVETypeFlags::EltTyFloat16:
8895 return llvm::ScalableVectorType::get(Builder.getHalfTy(), 8);
8896 case SVETypeFlags::EltTyBFloat16:
8897 return llvm::ScalableVectorType::get(Builder.getBFloatTy(), 8);
8898 case SVETypeFlags::EltTyFloat32:
8899 return llvm::ScalableVectorType::get(Builder.getFloatTy(), 4);
8900 case SVETypeFlags::EltTyFloat64:
8901 return llvm::ScalableVectorType::get(Builder.getDoubleTy(), 2);
8902
8903 case SVETypeFlags::EltTyBool8:
8904 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 16);
8905 case SVETypeFlags::EltTyBool16:
8906 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 8);
8907 case SVETypeFlags::EltTyBool32:
8908 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 4);
8909 case SVETypeFlags::EltTyBool64:
8910 return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 2);
8911 }
8912 }
8913
8914 llvm::Value *
8915 CodeGenFunction::EmitSVEAllTruePred(const SVETypeFlags &TypeFlags) {
8916 Function *Ptrue =
8917 CGM.getIntrinsic(Intrinsic::aarch64_sve_ptrue, getSVEPredType(TypeFlags));
8918 return Builder.CreateCall(Ptrue, {Builder.getInt32(/*SV_ALL*/ 31)});
8919 }
8920
8921 constexpr unsigned SVEBitsPerBlock = 128;
8922
8923 static llvm::ScalableVectorType *getSVEVectorForElementType(llvm::Type *EltTy) {
8924 unsigned NumElts = SVEBitsPerBlock / EltTy->getScalarSizeInBits();
8925 return llvm::ScalableVectorType::get(EltTy, NumElts);
8926 }
8927
8928 // Reinterpret the input predicate so that it can be used to correctly isolate
8929 // the elements of the specified datatype.
8930 Value *CodeGenFunction::EmitSVEPredicateCast(Value *Pred,
8931 llvm::ScalableVectorType *VTy) {
8932 auto *RTy = llvm::VectorType::get(IntegerType::get(getLLVMContext(), 1), VTy);
8933 if (Pred->getType() == RTy)
8934 return Pred;
8935
8936 unsigned IntID;
8937 llvm::Type *IntrinsicTy;
8938 switch (VTy->getMinNumElements()) {
8939 default:
8940 llvm_unreachable("unsupported element count!");
8941 case 2:
8942 case 4:
8943 case 8:
8944 IntID = Intrinsic::aarch64_sve_convert_from_svbool;
8945 IntrinsicTy = RTy;
8946 break;
8947 case 16:
8948 IntID = Intrinsic::aarch64_sve_convert_to_svbool;
8949 IntrinsicTy = Pred->getType();
8950 break;
8951 }
8952
8953 Function *F = CGM.getIntrinsic(IntID, IntrinsicTy);
8954 Value *C = Builder.CreateCall(F, Pred);
8955 assert(C->getType() == RTy && "Unexpected return type!");
8956 return C;
8957 }
8958
8959 Value *CodeGenFunction::EmitSVEGatherLoad(const SVETypeFlags &TypeFlags,
8960 SmallVectorImpl<Value *> &Ops,
8961 unsigned IntID) {
8962 auto *ResultTy = getSVEType(TypeFlags);
8963 auto *OverloadedTy =
8964 llvm::ScalableVectorType::get(SVEBuiltinMemEltTy(TypeFlags), ResultTy);
8965
8966 // At the ACLE level there's only one predicate type, svbool_t, which is
8967 // mapped to <n x 16 x i1>. However, this might be incompatible with the
8968 // actual type being loaded. For example, when loading doubles (i64) the
8969 // predicated should be <n x 2 x i1> instead. At the IR level the type of
8970 // the predicate and the data being loaded must match. Cast accordingly.
8971 Ops[0] = EmitSVEPredicateCast(Ops[0], OverloadedTy);
8972
8973 Function *F = nullptr;
8974 if (Ops[1]->getType()->isVectorTy())
8975 // This is the "vector base, scalar offset" case. In order to uniquely
8976 // map this built-in to an LLVM IR intrinsic, we need both the return type
8977 // and the type of the vector base.
8978 F = CGM.getIntrinsic(IntID, {OverloadedTy, Ops[1]->getType()});
8979 else
8980 // This is the "scalar base, vector offset case". The type of the offset
8981 // is encoded in the name of the intrinsic. We only need to specify the
8982 // return type in order to uniquely map this built-in to an LLVM IR
8983 // intrinsic.
8984 F = CGM.getIntrinsic(IntID, OverloadedTy);
8985
8986 // Pass 0 when the offset is missing. This can only be applied when using
8987 // the "vector base" addressing mode for which ACLE allows no offset. The
8988 // corresponding LLVM IR always requires an offset.
8989 if (Ops.size() == 2) {
8990 assert(Ops[1]->getType()->isVectorTy() && "Scalar base requires an offset");
8991 Ops.push_back(ConstantInt::get(Int64Ty, 0));
8992 }
8993
8994 // For "vector base, scalar index" scale the index so that it becomes a
8995 // scalar offset.
8996 if (!TypeFlags.isByteIndexed() && Ops[1]->getType()->isVectorTy()) {
8997 unsigned BytesPerElt =
8998 OverloadedTy->getElementType()->getScalarSizeInBits() / 8;
8999 Ops[2] = Builder.CreateShl(Ops[2], Log2_32(BytesPerElt));
9000 }
9001
9002 Value *Call = Builder.CreateCall(F, Ops);
9003
9004 // The following sext/zext is only needed when ResultTy != OverloadedTy. In
9005 // other cases it's folded into a nop.
9006 return TypeFlags.isZExtReturn() ? Builder.CreateZExt(Call, ResultTy)
9007 : Builder.CreateSExt(Call, ResultTy);
9008 }
9009
9010 Value *CodeGenFunction::EmitSVEScatterStore(const SVETypeFlags &TypeFlags,
9011 SmallVectorImpl<Value *> &Ops,
9012 unsigned IntID) {
9013 auto *SrcDataTy = getSVEType(TypeFlags);
9014 auto *OverloadedTy =
9015 llvm::ScalableVectorType::get(SVEBuiltinMemEltTy(TypeFlags), SrcDataTy);
9016
9017 // In ACLE the source data is passed in the last argument, whereas in LLVM IR
9018 // it's the first argument. Move it accordingly.
9019 Ops.insert(Ops.begin(), Ops.pop_back_val());
9020
9021 Function *F = nullptr;
9022 if (Ops[2]->getType()->isVectorTy())
9023 // This is the "vector base, scalar offset" case. In order to uniquely
9024 // map this built-in to an LLVM IR intrinsic, we need both the return type
9025 // and the type of the vector base.
9026 F = CGM.getIntrinsic(IntID, {OverloadedTy, Ops[2]->getType()});
9027 else
9028 // This is the "scalar base, vector offset case". The type of the offset
9029 // is encoded in the name of the intrinsic. We only need to specify the
9030 // return type in order to uniquely map this built-in to an LLVM IR
9031 // intrinsic.
9032 F = CGM.getIntrinsic(IntID, OverloadedTy);
9033
9034 // Pass 0 when the offset is missing. This can only be applied when using
9035 // the "vector base" addressing mode for which ACLE allows no offset. The
9036 // corresponding LLVM IR always requires an offset.
9037 if (Ops.size() == 3) {
9038 assert(Ops[1]->getType()->isVectorTy() && "Scalar base requires an offset");
9039 Ops.push_back(ConstantInt::get(Int64Ty, 0));
9040 }
9041
9042 // Truncation is needed when SrcDataTy != OverloadedTy. In other cases it's
9043 // folded into a nop.
9044 Ops[0] = Builder.CreateTrunc(Ops[0], OverloadedTy);
9045
9046 // At the ACLE level there's only one predicate type, svbool_t, which is
9047 // mapped to <n x 16 x i1>. However, this might be incompatible with the
9048 // actual type being stored. For example, when storing doubles (i64) the
9049 // predicated should be <n x 2 x i1> instead. At the IR level the type of
9050 // the predicate and the data being stored must match. Cast accordingly.
9051 Ops[1] = EmitSVEPredicateCast(Ops[1], OverloadedTy);
9052
9053 // For "vector base, scalar index" scale the index so that it becomes a
9054 // scalar offset.
9055 if (!TypeFlags.isByteIndexed() && Ops[2]->getType()->isVectorTy()) {
9056 unsigned BytesPerElt =
9057 OverloadedTy->getElementType()->getScalarSizeInBits() / 8;
9058 Ops[3] = Builder.CreateShl(Ops[3], Log2_32(BytesPerElt));
9059 }
9060
9061 return Builder.CreateCall(F, Ops);
9062 }
9063
9064 Value *CodeGenFunction::EmitSVEGatherPrefetch(const SVETypeFlags &TypeFlags,
9065 SmallVectorImpl<Value *> &Ops,
9066 unsigned IntID) {
9067 // The gather prefetches are overloaded on the vector input - this can either
9068 // be the vector of base addresses or vector of offsets.
9069 auto *OverloadedTy = dyn_cast<llvm::ScalableVectorType>(Ops[1]->getType());
9070 if (!OverloadedTy)
9071 OverloadedTy = cast<llvm::ScalableVectorType>(Ops[2]->getType());
9072
9073 // Cast the predicate from svbool_t to the right number of elements.
9074 Ops[0] = EmitSVEPredicateCast(Ops[0], OverloadedTy);
9075
9076 // vector + imm addressing modes
9077 if (Ops[1]->getType()->isVectorTy()) {
9078 if (Ops.size() == 3) {
9079 // Pass 0 for 'vector+imm' when the index is omitted.
9080 Ops.push_back(ConstantInt::get(Int64Ty, 0));
9081
9082 // The sv_prfop is the last operand in the builtin and IR intrinsic.
9083 std::swap(Ops[2], Ops[3]);
9084 } else {
9085 // Index needs to be passed as scaled offset.
9086 llvm::Type *MemEltTy = SVEBuiltinMemEltTy(TypeFlags);
9087 unsigned BytesPerElt = MemEltTy->getPrimitiveSizeInBits() / 8;
9088 if (BytesPerElt > 1)
9089 Ops[2] = Builder.CreateShl(Ops[2], Log2_32(BytesPerElt));
9090 }
9091 }
9092
9093 Function *F = CGM.getIntrinsic(IntID, OverloadedTy);
9094 return Builder.CreateCall(F, Ops);
9095 }
9096
9097 Value *CodeGenFunction::EmitSVEStructLoad(const SVETypeFlags &TypeFlags,
9098 SmallVectorImpl<Value*> &Ops,
9099 unsigned IntID) {
9100 llvm::ScalableVectorType *VTy = getSVEType(TypeFlags);
9101 auto VecPtrTy = llvm::PointerType::getUnqual(VTy);
9102 auto EltPtrTy = llvm::PointerType::getUnqual(VTy->getElementType());
9103
9104 unsigned N;
9105 switch (IntID) {
9106 case Intrinsic::aarch64_sve_ld2_sret:
9107 N = 2;
9108 break;
9109 case Intrinsic::aarch64_sve_ld3_sret:
9110 N = 3;
9111 break;
9112 case Intrinsic::aarch64_sve_ld4_sret:
9113 N = 4;
9114 break;
9115 default:
9116 llvm_unreachable("unknown intrinsic!");
9117 }
9118 auto RetTy = llvm::VectorType::get(VTy->getElementType(),
9119 VTy->getElementCount() * N);
9120
9121 Value *Predicate = EmitSVEPredicateCast(Ops[0], VTy);
9122 Value *BasePtr= Builder.CreateBitCast(Ops[1], VecPtrTy);
9123
9124 // Does the load have an offset?
9125 if (Ops.size() > 2)
9126 BasePtr = Builder.CreateGEP(VTy, BasePtr, Ops[2]);
9127
9128 BasePtr = Builder.CreateBitCast(BasePtr, EltPtrTy);
9129 Function *F = CGM.getIntrinsic(IntID, {VTy});
9130 Value *Call = Builder.CreateCall(F, {Predicate, BasePtr});
9131 unsigned MinElts = VTy->getMinNumElements();
9132 Value *Ret = llvm::PoisonValue::get(RetTy);
9133 for (unsigned I = 0; I < N; I++) {
9134 Value *Idx = ConstantInt::get(CGM.Int64Ty, I * MinElts);
9135 Value *SRet = Builder.CreateExtractValue(Call, I);
9136 Ret = Builder.CreateInsertVector(RetTy, Ret, SRet, Idx);
9137 }
9138 return Ret;
9139 }
9140
9141 Value *CodeGenFunction::EmitSVEStructStore(const SVETypeFlags &TypeFlags,
9142 SmallVectorImpl<Value*> &Ops,
9143 unsigned IntID) {
9144 llvm::ScalableVectorType *VTy = getSVEType(TypeFlags);
9145 auto VecPtrTy = llvm::PointerType::getUnqual(VTy);
9146 auto EltPtrTy = llvm::PointerType::getUnqual(VTy->getElementType());
9147
9148 unsigned N;
9149 switch (IntID) {
9150 case Intrinsic::aarch64_sve_st2:
9151 N = 2;
9152 break;
9153 case Intrinsic::aarch64_sve_st3:
9154 N = 3;
9155 break;
9156 case Intrinsic::aarch64_sve_st4:
9157 N = 4;
9158 break;
9159 default:
9160 llvm_unreachable("unknown intrinsic!");
9161 }
9162
9163 Value *Predicate = EmitSVEPredicateCast(Ops[0], VTy);
9164 Value *BasePtr = Builder.CreateBitCast(Ops[1], VecPtrTy);
9165
9166 // Does the store have an offset?
9167 if (Ops.size() > 3)
9168 BasePtr = Builder.CreateGEP(VTy, BasePtr, Ops[2]);
9169
9170 BasePtr = Builder.CreateBitCast(BasePtr, EltPtrTy);
9171 Value *Val = Ops.back();
9172
9173 // The llvm.aarch64.sve.st2/3/4 intrinsics take legal part vectors, so we
9174 // need to break up the tuple vector.
9175 SmallVector<llvm::Value*, 5> Operands;
9176 unsigned MinElts = VTy->getElementCount().getKnownMinValue();
9177 for (unsigned I = 0; I < N; ++I) {
9178 Value *Idx = ConstantInt::get(CGM.Int64Ty, I * MinElts);
9179 Operands.push_back(Builder.CreateExtractVector(VTy, Val, Idx));
9180 }
9181 Operands.append({Predicate, BasePtr});
9182
9183 Function *F = CGM.getIntrinsic(IntID, { VTy });
9184 return Builder.CreateCall(F, Operands);
9185 }
9186
9187 // SVE2's svpmullb and svpmullt builtins are similar to the svpmullb_pair and
9188 // svpmullt_pair intrinsics, with the exception that their results are bitcast
9189 // to a wider type.
9190 Value *CodeGenFunction::EmitSVEPMull(const SVETypeFlags &TypeFlags,
9191 SmallVectorImpl<Value *> &Ops,
9192 unsigned BuiltinID) {
9193 // Splat scalar operand to vector (intrinsics with _n infix)
9194 if (TypeFlags.hasSplatOperand()) {
9195 unsigned OpNo = TypeFlags.getSplatOperand();
9196 Ops[OpNo] = EmitSVEDupX(Ops[OpNo]);
9197 }
9198
9199 // The pair-wise function has a narrower overloaded type.
9200 Function *F = CGM.getIntrinsic(BuiltinID, Ops[0]->getType());
9201 Value *Call = Builder.CreateCall(F, {Ops[0], Ops[1]});
9202
9203 // Now bitcast to the wider result type.
9204 llvm::ScalableVectorType *Ty = getSVEType(TypeFlags);
9205 return EmitSVEReinterpret(Call, Ty);
9206 }
9207
9208 Value *CodeGenFunction::EmitSVEMovl(const SVETypeFlags &TypeFlags,
9209 ArrayRef<Value *> Ops, unsigned BuiltinID) {
9210 llvm::Type *OverloadedTy = getSVEType(TypeFlags);
9211 Function *F = CGM.getIntrinsic(BuiltinID, OverloadedTy);
9212 return Builder.CreateCall(F, {Ops[0], Builder.getInt32(0)});
9213 }
9214
9215 Value *CodeGenFunction::EmitSVEPrefetchLoad(const SVETypeFlags &TypeFlags,
9216 SmallVectorImpl<Value *> &Ops,
9217 unsigned BuiltinID) {
9218 auto *MemEltTy = SVEBuiltinMemEltTy(TypeFlags);
9219 auto *VectorTy = getSVEVectorForElementType(MemEltTy);
9220 auto *MemoryTy = llvm::ScalableVectorType::get(MemEltTy, VectorTy);
9221
9222 Value *Predicate = EmitSVEPredicateCast(Ops[0], MemoryTy);
9223 Value *BasePtr = Ops[1];
9224
9225 // Implement the index operand if not omitted.
9226 if (Ops.size() > 3) {
9227 BasePtr = Builder.CreateBitCast(BasePtr, MemoryTy->getPointerTo());
9228 BasePtr = Builder.CreateGEP(MemoryTy, BasePtr, Ops[2]);
9229 }
9230
9231 // Prefetch intriniscs always expect an i8*
9232 BasePtr = Builder.CreateBitCast(BasePtr, llvm::PointerType::getUnqual(Int8Ty));
9233 Value *PrfOp = Ops.back();
9234
9235 Function *F = CGM.getIntrinsic(BuiltinID, Predicate->getType());
9236 return Builder.CreateCall(F, {Predicate, BasePtr, PrfOp});
9237 }
9238
9239 Value *CodeGenFunction::EmitSVEMaskedLoad(const CallExpr *E,
9240 llvm::Type *ReturnTy,
9241 SmallVectorImpl<Value *> &Ops,
9242 unsigned BuiltinID,
9243 bool IsZExtReturn) {
9244 QualType LangPTy = E->getArg(1)->getType();
9245 llvm::Type *MemEltTy = CGM.getTypes().ConvertType(
9246 LangPTy->castAs<PointerType>()->getPointeeType());
9247
9248 // The vector type that is returned may be different from the
9249 // eventual type loaded from memory.
9250 auto VectorTy = cast<llvm::ScalableVectorType>(ReturnTy);
9251 auto MemoryTy = llvm::ScalableVectorType::get(MemEltTy, VectorTy);
9252
9253 Value *Predicate = EmitSVEPredicateCast(Ops[0], MemoryTy);
9254 Value *BasePtr = Builder.CreateBitCast(Ops[1], MemoryTy->getPointerTo());
9255
9256 // Does the load have an offset?
9257 if (Ops.size() > 2)
9258 BasePtr = Builder.CreateGEP(MemoryTy, BasePtr, Ops[2]);
9259
9260 BasePtr = Builder.CreateBitCast(BasePtr, MemEltTy->getPointerTo());
9261 Function *F = CGM.getIntrinsic(BuiltinID, MemoryTy);
9262 auto *Load =
9263 cast<llvm::Instruction>(Builder.CreateCall(F, {Predicate, BasePtr}));
9264 auto TBAAInfo = CGM.getTBAAAccessInfo(LangPTy->getPointeeType());
9265 CGM.DecorateInstructionWithTBAA(Load, TBAAInfo);
9266
9267 return IsZExtReturn ? Builder.CreateZExt(Load, VectorTy)
9268 : Builder.CreateSExt(Load, VectorTy);
9269 }
9270
9271 Value *CodeGenFunction::EmitSVEMaskedStore(const CallExpr *E,
9272 SmallVectorImpl<Value *> &Ops,
9273 unsigned BuiltinID) {
9274 QualType LangPTy = E->getArg(1)->getType();
9275 llvm::Type *MemEltTy = CGM.getTypes().ConvertType(
9276 LangPTy->castAs<PointerType>()->getPointeeType());
9277
9278 // The vector type that is stored may be different from the
9279 // eventual type stored to memory.
9280 auto VectorTy = cast<llvm::ScalableVectorType>(Ops.back()->getType());
9281 auto MemoryTy = llvm::ScalableVectorType::get(MemEltTy, VectorTy);
9282
9283 Value *Predicate = EmitSVEPredicateCast(Ops[0], MemoryTy);
9284 Value *BasePtr = Builder.CreateBitCast(Ops[1], MemoryTy->getPointerTo());
9285
9286 // Does the store have an offset?
9287 if (Ops.size() == 4)
9288 BasePtr = Builder.CreateGEP(MemoryTy, BasePtr, Ops[2]);
9289
9290 // Last value is always the data
9291 llvm::Value *Val = Builder.CreateTrunc(Ops.back(), MemoryTy);
9292
9293 BasePtr = Builder.CreateBitCast(BasePtr, MemEltTy->getPointerTo());
9294 Function *F = CGM.getIntrinsic(BuiltinID, MemoryTy);
9295 auto *Store =
9296 cast<llvm::Instruction>(Builder.CreateCall(F, {Val, Predicate, BasePtr}));
9297 auto TBAAInfo = CGM.getTBAAAccessInfo(LangPTy->getPointeeType());
9298 CGM.DecorateInstructionWithTBAA(Store, TBAAInfo);
9299 return Store;
9300 }
9301
9302 // Limit the usage of scalable llvm IR generated by the ACLE by using the
9303 // sve dup.x intrinsic instead of IRBuilder::CreateVectorSplat.
9304 Value *CodeGenFunction::EmitSVEDupX(Value *Scalar, llvm::Type *Ty) {
9305 return Builder.CreateVectorSplat(
9306 cast<llvm::VectorType>(Ty)->getElementCount(), Scalar);
9307 }
9308
9309 Value *CodeGenFunction::EmitSVEDupX(Value* Scalar) {
9310 return EmitSVEDupX(Scalar, getSVEVectorForElementType(Scalar->getType()));
9311 }
9312
9313 Value *CodeGenFunction::EmitSVEReinterpret(Value *Val, llvm::Type *Ty) {
9314 // FIXME: For big endian this needs an additional REV, or needs a separate
9315 // intrinsic that is code-generated as a no-op, because the LLVM bitcast
9316 // instruction is defined as 'bitwise' equivalent from memory point of
9317 // view (when storing/reloading), whereas the svreinterpret builtin
9318 // implements bitwise equivalent cast from register point of view.
9319 // LLVM CodeGen for a bitcast must add an explicit REV for big-endian.
9320 return Builder.CreateBitCast(Val, Ty);
9321 }
9322
9323 static void InsertExplicitZeroOperand(CGBuilderTy &Builder, llvm::Type *Ty,
9324 SmallVectorImpl<Value *> &Ops) {
9325 auto *SplatZero = Constant::getNullValue(Ty);
9326 Ops.insert(Ops.begin(), SplatZero);
9327 }
9328
9329 static void InsertExplicitUndefOperand(CGBuilderTy &Builder, llvm::Type *Ty,
9330 SmallVectorImpl<Value *> &Ops) {
9331 auto *SplatUndef = UndefValue::get(Ty);
9332 Ops.insert(Ops.begin(), SplatUndef);
9333 }
9334
9335 SmallVector<llvm::Type *, 2>
9336 CodeGenFunction::getSVEOverloadTypes(const SVETypeFlags &TypeFlags,
9337 llvm::Type *ResultType,
9338 ArrayRef<Value *> Ops) {
9339 if (TypeFlags.isOverloadNone())
9340 return {};
9341
9342 llvm::Type *DefaultType = getSVEType(TypeFlags);
9343
9344 if (TypeFlags.isOverloadWhile())
9345 return {DefaultType, Ops[1]->getType()};
9346
9347 if (TypeFlags.isOverloadWhileRW())
9348 return {getSVEPredType(TypeFlags), Ops[0]->getType()};
9349
9350 if (TypeFlags.isOverloadCvt())
9351 return {Ops[0]->getType(), Ops.back()->getType()};
9352
9353 assert(TypeFlags.isOverloadDefault() && "Unexpected value for overloads");
9354 return {DefaultType};
9355 }
9356
9357 Value *CodeGenFunction::EmitSVETupleSetOrGet(const SVETypeFlags &TypeFlags,
9358 llvm::Type *Ty,
9359 ArrayRef<Value *> Ops) {
9360 assert((TypeFlags.isTupleSet() || TypeFlags.isTupleGet()) &&
9361 "Expects TypleFlag isTupleSet or TypeFlags.isTupleSet()");
9362
9363 unsigned I = cast<ConstantInt>(Ops[1])->getSExtValue();
9364 auto *SingleVecTy = dyn_cast<llvm::ScalableVectorType>(
9365 TypeFlags.isTupleSet() ? Ops[2]->getType() : Ty);
9366 Value *Idx = ConstantInt::get(CGM.Int64Ty,
9367 I * SingleVecTy->getMinNumElements());
9368
9369 if (TypeFlags.isTupleSet())
9370 return Builder.CreateInsertVector(Ty, Ops[0], Ops[2], Idx);
9371 return Builder.CreateExtractVector(Ty, Ops[0], Idx);
9372 }
9373
9374 Value *CodeGenFunction::EmitSVETupleCreate(const SVETypeFlags &TypeFlags,
9375 llvm::Type *Ty,
9376 ArrayRef<Value *> Ops) {
9377 assert(TypeFlags.isTupleCreate() && "Expects TypleFlag isTupleCreate");
9378
9379 auto *SrcTy = dyn_cast<llvm::ScalableVectorType>(Ops[0]->getType());
9380 unsigned MinElts = SrcTy->getMinNumElements();
9381 Value *Call = llvm::PoisonValue::get(Ty);
9382 for (unsigned I = 0; I < Ops.size(); I++) {
9383 Value *Idx = ConstantInt::get(CGM.Int64Ty, I * MinElts);
9384 Call = Builder.CreateInsertVector(Ty, Call, Ops[I], Idx);
9385 }
9386
9387 return Call;
9388 }
9389
9390 Value *CodeGenFunction::EmitAArch64SVEBuiltinExpr(unsigned BuiltinID,
9391 const CallExpr *E) {
9392 // Find out if any arguments are required to be integer constant expressions.
9393 unsigned ICEArguments = 0;
9394 ASTContext::GetBuiltinTypeError Error;
9395 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments);
9396 assert(Error == ASTContext::GE_None && "Should not codegen an error");
9397
9398 llvm::Type *Ty = ConvertType(E->getType());
9399 if (BuiltinID >= SVE::BI__builtin_sve_reinterpret_s8_s8 &&
9400 BuiltinID <= SVE::BI__builtin_sve_reinterpret_f64_f64) {
9401 Value *Val = EmitScalarExpr(E->getArg(0));
9402 return EmitSVEReinterpret(Val, Ty);
9403 }
9404
9405 llvm::SmallVector<Value *, 4> Ops;
9406 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) {
9407 if ((ICEArguments & (1 << i)) == 0)
9408 Ops.push_back(EmitScalarExpr(E->getArg(i)));
9409 else {
9410 // If this is required to be a constant, constant fold it so that we know
9411 // that the generated intrinsic gets a ConstantInt.
9412 std::optional<llvm::APSInt> Result =
9413 E->getArg(i)->getIntegerConstantExpr(getContext());
9414 assert(Result && "Expected argument to be a constant");
9415
9416 // Immediates for SVE llvm intrinsics are always 32bit. We can safely
9417 // truncate because the immediate has been range checked and no valid
9418 // immediate requires more than a handful of bits.
9419 *Result = Result->extOrTrunc(32);
9420 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), *Result));
9421 }
9422 }
9423
9424 auto *Builtin = findARMVectorIntrinsicInMap(AArch64SVEIntrinsicMap, BuiltinID,
9425 AArch64SVEIntrinsicsProvenSorted);
9426 SVETypeFlags TypeFlags(Builtin->TypeModifier);
9427 if (TypeFlags.isLoad())
9428 return EmitSVEMaskedLoad(E, Ty, Ops, Builtin->LLVMIntrinsic,
9429 TypeFlags.isZExtReturn());
9430 else if (TypeFlags.isStore())
9431 return EmitSVEMaskedStore(E, Ops, Builtin->LLVMIntrinsic);
9432 else if (TypeFlags.isGatherLoad())
9433 return EmitSVEGatherLoad(TypeFlags, Ops, Builtin->LLVMIntrinsic);
9434 else if (TypeFlags.isScatterStore())
9435 return EmitSVEScatterStore(TypeFlags, Ops, Builtin->LLVMIntrinsic);
9436 else if (TypeFlags.isPrefetch())
9437 return EmitSVEPrefetchLoad(TypeFlags, Ops, Builtin->LLVMIntrinsic);
9438 else if (TypeFlags.isGatherPrefetch())
9439 return EmitSVEGatherPrefetch(TypeFlags, Ops, Builtin->LLVMIntrinsic);
9440 else if (TypeFlags.isStructLoad())
9441 return EmitSVEStructLoad(TypeFlags, Ops, Builtin->LLVMIntrinsic);
9442 else if (TypeFlags.isStructStore())
9443 return EmitSVEStructStore(TypeFlags, Ops, Builtin->LLVMIntrinsic);
9444 else if (TypeFlags.isTupleSet() || TypeFlags.isTupleGet())
9445 return EmitSVETupleSetOrGet(TypeFlags, Ty, Ops);
9446 else if (TypeFlags.isTupleCreate())
9447 return EmitSVETupleCreate(TypeFlags, Ty, Ops);
9448 else if (TypeFlags.isUndef())
9449 return UndefValue::get(Ty);
9450 else if (Builtin->LLVMIntrinsic != 0) {
9451 if (TypeFlags.getMergeType() == SVETypeFlags::MergeZeroExp)
9452 InsertExplicitZeroOperand(Builder, Ty, Ops);
9453
9454 if (TypeFlags.getMergeType() == SVETypeFlags::MergeAnyExp)
9455 InsertExplicitUndefOperand(Builder, Ty, Ops);
9456
9457 // Some ACLE builtins leave out the argument to specify the predicate
9458 // pattern, which is expected to be expanded to an SV_ALL pattern.
9459 if (TypeFlags.isAppendSVALL())
9460 Ops.push_back(Builder.getInt32(/*SV_ALL*/ 31));
9461 if (TypeFlags.isInsertOp1SVALL())
9462 Ops.insert(&Ops[1], Builder.getInt32(/*SV_ALL*/ 31));
9463
9464 // Predicates must match the main datatype.
9465 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
9466 if (auto PredTy = dyn_cast<llvm::VectorType>(Ops[i]->getType()))
9467 if (PredTy->getElementType()->isIntegerTy(1))
9468 Ops[i] = EmitSVEPredicateCast(Ops[i], getSVEType(TypeFlags));
9469
9470 // Splat scalar operand to vector (intrinsics with _n infix)
9471 if (TypeFlags.hasSplatOperand()) {
9472 unsigned OpNo = TypeFlags.getSplatOperand();
9473 Ops[OpNo] = EmitSVEDupX(Ops[OpNo]);
9474 }
9475
9476 if (TypeFlags.isReverseCompare())
9477 std::swap(Ops[1], Ops[2]);
9478
9479 if (TypeFlags.isReverseUSDOT())
9480 std::swap(Ops[1], Ops[2]);
9481
9482 // Predicated intrinsics with _z suffix need a select w/ zeroinitializer.
9483 if (TypeFlags.getMergeType() == SVETypeFlags::MergeZero) {
9484 llvm::Type *OpndTy = Ops[1]->getType();
9485 auto *SplatZero = Constant::getNullValue(OpndTy);
9486 Ops[1] = Builder.CreateSelect(Ops[0], Ops[1], SplatZero);
9487 }
9488
9489 Function *F = CGM.getIntrinsic(Builtin->LLVMIntrinsic,
9490 getSVEOverloadTypes(TypeFlags, Ty, Ops));
9491 Value *Call = Builder.CreateCall(F, Ops);
9492
9493 // Predicate results must be converted to svbool_t.
9494 if (auto PredTy = dyn_cast<llvm::VectorType>(Call->getType()))
9495 if (PredTy->getScalarType()->isIntegerTy(1))
9496 Call = EmitSVEPredicateCast(Call, cast<llvm::ScalableVectorType>(Ty));
9497
9498 return Call;
9499 }
9500
9501 switch (BuiltinID) {
9502 default:
9503 return nullptr;
9504
9505 case SVE::BI__builtin_sve_svmov_b_z: {
9506 // svmov_b_z(pg, op) <=> svand_b_z(pg, op, op)
9507 SVETypeFlags TypeFlags(Builtin->TypeModifier);
9508 llvm::Type* OverloadedTy = getSVEType(TypeFlags);
9509 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_sve_and_z, OverloadedTy);
9510 return Builder.CreateCall(F, {Ops[0], Ops[1], Ops[1]});
9511 }
9512
9513 case SVE::BI__builtin_sve_svnot_b_z: {
9514 // svnot_b_z(pg, op) <=> sveor_b_z(pg, op, pg)
9515 SVETypeFlags TypeFlags(Builtin->TypeModifier);
9516 llvm::Type* OverloadedTy = getSVEType(TypeFlags);
9517 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_sve_eor_z, OverloadedTy);
9518 return Builder.CreateCall(F, {Ops[0], Ops[1], Ops[0]});
9519 }
9520
9521 case SVE::BI__builtin_sve_svmovlb_u16:
9522 case SVE::BI__builtin_sve_svmovlb_u32:
9523 case SVE::BI__builtin_sve_svmovlb_u64:
9524 return EmitSVEMovl(TypeFlags, Ops, Intrinsic::aarch64_sve_ushllb);
9525
9526 case SVE::BI__builtin_sve_svmovlb_s16:
9527 case SVE::BI__builtin_sve_svmovlb_s32:
9528 case SVE::BI__builtin_sve_svmovlb_s64:
9529 return EmitSVEMovl(TypeFlags, Ops, Intrinsic::aarch64_sve_sshllb);
9530
9531 case SVE::BI__builtin_sve_svmovlt_u16:
9532 case SVE::BI__builtin_sve_svmovlt_u32:
9533 case SVE::BI__builtin_sve_svmovlt_u64:
9534 return EmitSVEMovl(TypeFlags, Ops, Intrinsic::aarch64_sve_ushllt);
9535
9536 case SVE::BI__builtin_sve_svmovlt_s16:
9537 case SVE::BI__builtin_sve_svmovlt_s32:
9538 case SVE::BI__builtin_sve_svmovlt_s64:
9539 return EmitSVEMovl(TypeFlags, Ops, Intrinsic::aarch64_sve_sshllt);
9540
9541 case SVE::BI__builtin_sve_svpmullt_u16:
9542 case SVE::BI__builtin_sve_svpmullt_u64:
9543 case SVE::BI__builtin_sve_svpmullt_n_u16:
9544 case SVE::BI__builtin_sve_svpmullt_n_u64:
9545 return EmitSVEPMull(TypeFlags, Ops, Intrinsic::aarch64_sve_pmullt_pair);
9546
9547 case SVE::BI__builtin_sve_svpmullb_u16:
9548 case SVE::BI__builtin_sve_svpmullb_u64:
9549 case SVE::BI__builtin_sve_svpmullb_n_u16:
9550 case SVE::BI__builtin_sve_svpmullb_n_u64:
9551 return EmitSVEPMull(TypeFlags, Ops, Intrinsic::aarch64_sve_pmullb_pair);
9552
9553 case SVE::BI__builtin_sve_svdup_n_b8:
9554 case SVE::BI__builtin_sve_svdup_n_b16:
9555 case SVE::BI__builtin_sve_svdup_n_b32:
9556 case SVE::BI__builtin_sve_svdup_n_b64: {
9557 Value *CmpNE =
9558 Builder.CreateICmpNE(Ops[0], Constant::getNullValue(Ops[0]->getType()));
9559 llvm::ScalableVectorType *OverloadedTy = getSVEType(TypeFlags);
9560 Value *Dup = EmitSVEDupX(CmpNE, OverloadedTy);
9561 return EmitSVEPredicateCast(Dup, cast<llvm::ScalableVectorType>(Ty));
9562 }
9563
9564 case SVE::BI__builtin_sve_svdupq_n_b8:
9565 case SVE::BI__builtin_sve_svdupq_n_b16:
9566 case SVE::BI__builtin_sve_svdupq_n_b32:
9567 case SVE::BI__builtin_sve_svdupq_n_b64:
9568 case SVE::BI__builtin_sve_svdupq_n_u8:
9569 case SVE::BI__builtin_sve_svdupq_n_s8:
9570 case SVE::BI__builtin_sve_svdupq_n_u64:
9571 case SVE::BI__builtin_sve_svdupq_n_f64:
9572 case SVE::BI__builtin_sve_svdupq_n_s64:
9573 case SVE::BI__builtin_sve_svdupq_n_u16:
9574 case SVE::BI__builtin_sve_svdupq_n_f16:
9575 case SVE::BI__builtin_sve_svdupq_n_bf16:
9576 case SVE::BI__builtin_sve_svdupq_n_s16:
9577 case SVE::BI__builtin_sve_svdupq_n_u32:
9578 case SVE::BI__builtin_sve_svdupq_n_f32:
9579 case SVE::BI__builtin_sve_svdupq_n_s32: {
9580 // These builtins are implemented by storing each element to an array and using
9581 // ld1rq to materialize a vector.
9582 unsigned NumOpnds = Ops.size();
9583
9584 bool IsBoolTy =
9585 cast<llvm::VectorType>(Ty)->getElementType()->isIntegerTy(1);
9586
9587 // For svdupq_n_b* the element type of is an integer of type 128/numelts,
9588 // so that the compare can use the width that is natural for the expected
9589 // number of predicate lanes.
9590 llvm::Type *EltTy = Ops[0]->getType();
9591 if (IsBoolTy)
9592 EltTy = IntegerType::get(getLLVMContext(), SVEBitsPerBlock / NumOpnds);
9593
9594 SmallVector<llvm::Value *, 16> VecOps;
9595 for (unsigned I = 0; I < NumOpnds; ++I)
9596 VecOps.push_back(Builder.CreateZExt(Ops[I], EltTy));
9597 Value *Vec = BuildVector(VecOps);
9598
9599 llvm::Type *OverloadedTy = getSVEVectorForElementType(EltTy);
9600 Value *InsertSubVec = Builder.CreateInsertVector(
9601 OverloadedTy, PoisonValue::get(OverloadedTy), Vec, Builder.getInt64(0));
9602
9603 Function *F =
9604 CGM.getIntrinsic(Intrinsic::aarch64_sve_dupq_lane, OverloadedTy);
9605 Value *DupQLane =
9606 Builder.CreateCall(F, {InsertSubVec, Builder.getInt64(0)});
9607
9608 if (!IsBoolTy)
9609 return DupQLane;
9610
9611 SVETypeFlags TypeFlags(Builtin->TypeModifier);
9612 Value *Pred = EmitSVEAllTruePred(TypeFlags);
9613
9614 // For svdupq_n_b* we need to add an additional 'cmpne' with '0'.
9615 F = CGM.getIntrinsic(NumOpnds == 2 ? Intrinsic::aarch64_sve_cmpne
9616 : Intrinsic::aarch64_sve_cmpne_wide,
9617 OverloadedTy);
9618 Value *Call = Builder.CreateCall(
9619 F, {Pred, DupQLane, EmitSVEDupX(Builder.getInt64(0))});
9620 return EmitSVEPredicateCast(Call, cast<llvm::ScalableVectorType>(Ty));
9621 }
9622
9623 case SVE::BI__builtin_sve_svpfalse_b:
9624 return ConstantInt::getFalse(Ty);
9625
9626 case SVE::BI__builtin_sve_svlen_bf16:
9627 case SVE::BI__builtin_sve_svlen_f16:
9628 case SVE::BI__builtin_sve_svlen_f32:
9629 case SVE::BI__builtin_sve_svlen_f64:
9630 case SVE::BI__builtin_sve_svlen_s8:
9631 case SVE::BI__builtin_sve_svlen_s16:
9632 case SVE::BI__builtin_sve_svlen_s32:
9633 case SVE::BI__builtin_sve_svlen_s64:
9634 case SVE::BI__builtin_sve_svlen_u8:
9635 case SVE::BI__builtin_sve_svlen_u16:
9636 case SVE::BI__builtin_sve_svlen_u32:
9637 case SVE::BI__builtin_sve_svlen_u64: {
9638 SVETypeFlags TF(Builtin->TypeModifier);
9639 auto VTy = cast<llvm::VectorType>(getSVEType(TF));
9640 auto *NumEls =
9641 llvm::ConstantInt::get(Ty, VTy->getElementCount().getKnownMinValue());
9642
9643 Function *F = CGM.getIntrinsic(Intrinsic::vscale, Ty);
9644 return Builder.CreateMul(NumEls, Builder.CreateCall(F));
9645 }
9646
9647 case SVE::BI__builtin_sve_svtbl2_u8:
9648 case SVE::BI__builtin_sve_svtbl2_s8:
9649 case SVE::BI__builtin_sve_svtbl2_u16:
9650 case SVE::BI__builtin_sve_svtbl2_s16:
9651 case SVE::BI__builtin_sve_svtbl2_u32:
9652 case SVE::BI__builtin_sve_svtbl2_s32:
9653 case SVE::BI__builtin_sve_svtbl2_u64:
9654 case SVE::BI__builtin_sve_svtbl2_s64:
9655 case SVE::BI__builtin_sve_svtbl2_f16:
9656 case SVE::BI__builtin_sve_svtbl2_bf16:
9657 case SVE::BI__builtin_sve_svtbl2_f32:
9658 case SVE::BI__builtin_sve_svtbl2_f64: {
9659 SVETypeFlags TF(Builtin->TypeModifier);
9660 auto VTy = cast<llvm::ScalableVectorType>(getSVEType(TF));
9661 Value *V0 = Builder.CreateExtractVector(VTy, Ops[0],
9662 ConstantInt::get(CGM.Int64Ty, 0));
9663 unsigned MinElts = VTy->getMinNumElements();
9664 Value *V1 = Builder.CreateExtractVector(
9665 VTy, Ops[0], ConstantInt::get(CGM.Int64Ty, MinElts));
9666 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_sve_tbl2, VTy);
9667 return Builder.CreateCall(F, {V0, V1, Ops[1]});
9668 }
9669
9670 case SVE::BI__builtin_sve_svset_neonq_s8:
9671 case SVE::BI__builtin_sve_svset_neonq_s16:
9672 case SVE::BI__builtin_sve_svset_neonq_s32:
9673 case SVE::BI__builtin_sve_svset_neonq_s64:
9674 case SVE::BI__builtin_sve_svset_neonq_u8:
9675 case SVE::BI__builtin_sve_svset_neonq_u16:
9676 case SVE::BI__builtin_sve_svset_neonq_u32:
9677 case SVE::BI__builtin_sve_svset_neonq_u64:
9678 case SVE::BI__builtin_sve_svset_neonq_f16:
9679 case SVE::BI__builtin_sve_svset_neonq_f32:
9680 case SVE::BI__builtin_sve_svset_neonq_f64:
9681 case SVE::BI__builtin_sve_svset_neonq_bf16: {
9682 return Builder.CreateInsertVector(Ty, Ops[0], Ops[1], Builder.getInt64(0));
9683 }
9684
9685 case SVE::BI__builtin_sve_svget_neonq_s8:
9686 case SVE::BI__builtin_sve_svget_neonq_s16:
9687 case SVE::BI__builtin_sve_svget_neonq_s32:
9688 case SVE::BI__builtin_sve_svget_neonq_s64:
9689 case SVE::BI__builtin_sve_svget_neonq_u8:
9690 case SVE::BI__builtin_sve_svget_neonq_u16:
9691 case SVE::BI__builtin_sve_svget_neonq_u32:
9692 case SVE::BI__builtin_sve_svget_neonq_u64:
9693 case SVE::BI__builtin_sve_svget_neonq_f16:
9694 case SVE::BI__builtin_sve_svget_neonq_f32:
9695 case SVE::BI__builtin_sve_svget_neonq_f64:
9696 case SVE::BI__builtin_sve_svget_neonq_bf16: {
9697 return Builder.CreateExtractVector(Ty, Ops[0], Builder.getInt64(0));
9698 }
9699
9700 case SVE::BI__builtin_sve_svdup_neonq_s8:
9701 case SVE::BI__builtin_sve_svdup_neonq_s16:
9702 case SVE::BI__builtin_sve_svdup_neonq_s32:
9703 case SVE::BI__builtin_sve_svdup_neonq_s64:
9704 case SVE::BI__builtin_sve_svdup_neonq_u8:
9705 case SVE::BI__builtin_sve_svdup_neonq_u16:
9706 case SVE::BI__builtin_sve_svdup_neonq_u32:
9707 case SVE::BI__builtin_sve_svdup_neonq_u64:
9708 case SVE::BI__builtin_sve_svdup_neonq_f16:
9709 case SVE::BI__builtin_sve_svdup_neonq_f32:
9710 case SVE::BI__builtin_sve_svdup_neonq_f64:
9711 case SVE::BI__builtin_sve_svdup_neonq_bf16: {
9712 Value *Insert = Builder.CreateInsertVector(Ty, PoisonValue::get(Ty), Ops[0],
9713 Builder.getInt64(0));
9714 return Builder.CreateIntrinsic(Intrinsic::aarch64_sve_dupq_lane, {Ty},
9715 {Insert, Builder.getInt64(0)});
9716 }
9717 }
9718
9719 /// Should not happen
9720 return nullptr;
9721 }
9722
9723 Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID,
9724 const CallExpr *E,
9725 llvm::Triple::ArchType Arch) {
9726 if (BuiltinID >= clang::AArch64::FirstSVEBuiltin &&
9727 BuiltinID <= clang::AArch64::LastSVEBuiltin)
9728 return EmitAArch64SVEBuiltinExpr(BuiltinID, E);
9729
9730 unsigned HintID = static_cast<unsigned>(-1);
9731 switch (BuiltinID) {
9732 default: break;
9733 case clang::AArch64::BI__builtin_arm_nop:
9734 HintID = 0;
9735 break;
9736 case clang::AArch64::BI__builtin_arm_yield:
9737 case clang::AArch64::BI__yield:
9738 HintID = 1;
9739 break;
9740 case clang::AArch64::BI__builtin_arm_wfe:
9741 case clang::AArch64::BI__wfe:
9742 HintID = 2;
9743 break;
9744 case clang::AArch64::BI__builtin_arm_wfi:
9745 case clang::AArch64::BI__wfi:
9746 HintID = 3;
9747 break;
9748 case clang::AArch64::BI__builtin_arm_sev:
9749 case clang::AArch64::BI__sev:
9750 HintID = 4;
9751 break;
9752 case clang::AArch64::BI__builtin_arm_sevl:
9753 case clang::AArch64::BI__sevl:
9754 HintID = 5;
9755 break;
9756 }
9757
9758 if (HintID != static_cast<unsigned>(-1)) {
9759 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_hint);
9760 return Builder.CreateCall(F, llvm::ConstantInt::get(Int32Ty, HintID));
9761 }
9762
9763 if (BuiltinID == clang::AArch64::BI__builtin_arm_rbit) {
9764 assert((getContext().getTypeSize(E->getType()) == 32) &&
9765 "rbit of unusual size!");
9766 llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
9767 return Builder.CreateCall(
9768 CGM.getIntrinsic(Intrinsic::bitreverse, Arg->getType()), Arg, "rbit");
9769 }
9770 if (BuiltinID == clang::AArch64::BI__builtin_arm_rbit64) {
9771 assert((getContext().getTypeSize(E->getType()) == 64) &&
9772 "rbit of unusual size!");
9773 llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
9774 return Builder.CreateCall(
9775 CGM.getIntrinsic(Intrinsic::bitreverse, Arg->getType()), Arg, "rbit");
9776 }
9777
9778 if (BuiltinID == clang::AArch64::BI__builtin_arm_cls) {
9779 llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
9780 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::aarch64_cls), Arg,
9781 "cls");
9782 }
9783 if (BuiltinID == clang::AArch64::BI__builtin_arm_cls64) {
9784 llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
9785 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::aarch64_cls64), Arg,
9786 "cls");
9787 }
9788
9789 if (BuiltinID == clang::AArch64::BI__builtin_arm_rint32zf ||
9790 BuiltinID == clang::AArch64::BI__builtin_arm_rint32z) {
9791 llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
9792 llvm::Type *Ty = Arg->getType();
9793 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::aarch64_frint32z, Ty),
9794 Arg, "frint32z");
9795 }
9796
9797 if (BuiltinID == clang::AArch64::BI__builtin_arm_rint64zf ||
9798 BuiltinID == clang::AArch64::BI__builtin_arm_rint64z) {
9799 llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
9800 llvm::Type *Ty = Arg->getType();
9801 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::aarch64_frint64z, Ty),
9802 Arg, "frint64z");
9803 }
9804
9805 if (BuiltinID == clang::AArch64::BI__builtin_arm_rint32xf ||
9806 BuiltinID == clang::AArch64::BI__builtin_arm_rint32x) {
9807 llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
9808 llvm::Type *Ty = Arg->getType();
9809 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::aarch64_frint32x, Ty),
9810 Arg, "frint32x");
9811 }
9812
9813 if (BuiltinID == clang::AArch64::BI__builtin_arm_rint64xf ||
9814 BuiltinID == clang::AArch64::BI__builtin_arm_rint64x) {
9815 llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
9816 llvm::Type *Ty = Arg->getType();
9817 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::aarch64_frint64x, Ty),
9818 Arg, "frint64x");
9819 }
9820
9821 if (BuiltinID == clang::AArch64::BI__builtin_arm_jcvt) {
9822 assert((getContext().getTypeSize(E->getType()) == 32) &&
9823 "__jcvt of unusual size!");
9824 llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
9825 return Builder.CreateCall(
9826 CGM.getIntrinsic(Intrinsic::aarch64_fjcvtzs), Arg);
9827 }
9828
9829 if (BuiltinID == clang::AArch64::BI__builtin_arm_ld64b ||
9830 BuiltinID == clang::AArch64::BI__builtin_arm_st64b ||
9831 BuiltinID == clang::AArch64::BI__builtin_arm_st64bv ||
9832 BuiltinID == clang::AArch64::BI__builtin_arm_st64bv0) {
9833 llvm::Value *MemAddr = EmitScalarExpr(E->getArg(0));
9834 llvm::Value *ValPtr = EmitScalarExpr(E->getArg(1));
9835
9836 if (BuiltinID == clang::AArch64::BI__builtin_arm_ld64b) {
9837 // Load from the address via an LLVM intrinsic, receiving a
9838 // tuple of 8 i64 words, and store each one to ValPtr.
9839 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_ld64b);
9840 llvm::Value *Val = Builder.CreateCall(F, MemAddr);
9841 llvm::Value *ToRet;
9842 for (size_t i = 0; i < 8; i++) {
9843 llvm::Value *ValOffsetPtr =
9844 Builder.CreateGEP(Int64Ty, ValPtr, Builder.getInt32(i));
9845 Address Addr =
9846 Address(ValOffsetPtr, Int64Ty, CharUnits::fromQuantity(8));
9847 ToRet = Builder.CreateStore(Builder.CreateExtractValue(Val, i), Addr);
9848 }
9849 return ToRet;
9850 } else {
9851 // Load 8 i64 words from ValPtr, and store them to the address
9852 // via an LLVM intrinsic.
9853 SmallVector<llvm::Value *, 9> Args;
9854 Args.push_back(MemAddr);
9855 for (size_t i = 0; i < 8; i++) {
9856 llvm::Value *ValOffsetPtr =
9857 Builder.CreateGEP(Int64Ty, ValPtr, Builder.getInt32(i));
9858 Address Addr =
9859 Address(ValOffsetPtr, Int64Ty, CharUnits::fromQuantity(8));
9860 Args.push_back(Builder.CreateLoad(Addr));
9861 }
9862
9863 auto Intr = (BuiltinID == clang::AArch64::BI__builtin_arm_st64b
9864 ? Intrinsic::aarch64_st64b
9865 : BuiltinID == clang::AArch64::BI__builtin_arm_st64bv
9866 ? Intrinsic::aarch64_st64bv
9867 : Intrinsic::aarch64_st64bv0);
9868 Function *F = CGM.getIntrinsic(Intr);
9869 return Builder.CreateCall(F, Args);
9870 }
9871 }
9872
9873 if (BuiltinID == clang::AArch64::BI__builtin_arm_rndr ||
9874 BuiltinID == clang::AArch64::BI__builtin_arm_rndrrs) {
9875
9876 auto Intr = (BuiltinID == clang::AArch64::BI__builtin_arm_rndr
9877 ? Intrinsic::aarch64_rndr
9878 : Intrinsic::aarch64_rndrrs);
9879 Function *F = CGM.getIntrinsic(Intr);
9880 llvm::Value *Val = Builder.CreateCall(F);
9881 Value *RandomValue = Builder.CreateExtractValue(Val, 0);
9882 Value *Status = Builder.CreateExtractValue(Val, 1);
9883
9884 Address MemAddress = EmitPointerWithAlignment(E->getArg(0));
9885 Builder.CreateStore(RandomValue, MemAddress);
9886 Status = Builder.CreateZExt(Status, Int32Ty);
9887 return Status;
9888 }
9889
9890 if (BuiltinID == clang::AArch64::BI__clear_cache) {
9891 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments");
9892 const FunctionDecl *FD = E->getDirectCallee();
9893 Value *Ops[2];
9894 for (unsigned i = 0; i < 2; i++)
9895 Ops[i] = EmitScalarExpr(E->getArg(i));
9896 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType());
9897 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty);
9898 StringRef Name = FD->getName();
9899 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops);
9900 }
9901
9902 if ((BuiltinID == clang::AArch64::BI__builtin_arm_ldrex ||
9903 BuiltinID == clang::AArch64::BI__builtin_arm_ldaex) &&
9904 getContext().getTypeSize(E->getType()) == 128) {
9905 Function *F =
9906 CGM.getIntrinsic(BuiltinID == clang::AArch64::BI__builtin_arm_ldaex
9907 ? Intrinsic::aarch64_ldaxp
9908 : Intrinsic::aarch64_ldxp);
9909
9910 Value *LdPtr = EmitScalarExpr(E->getArg(0));
9911 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy),
9912 "ldxp");
9913
9914 Value *Val0 = Builder.CreateExtractValue(Val, 1);
9915 Value *Val1 = Builder.CreateExtractValue(Val, 0);
9916 llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128);
9917 Val0 = Builder.CreateZExt(Val0, Int128Ty);
9918 Val1 = Builder.CreateZExt(Val1, Int128Ty);
9919
9920 Value *ShiftCst = llvm::ConstantInt::get(Int128Ty, 64);
9921 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */);
9922 Val = Builder.CreateOr(Val, Val1);
9923 return Builder.CreateBitCast(Val, ConvertType(E->getType()));
9924 } else if (BuiltinID == clang::AArch64::BI__builtin_arm_ldrex ||
9925 BuiltinID == clang::AArch64::BI__builtin_arm_ldaex) {
9926 Value *LoadAddr = EmitScalarExpr(E->getArg(0));
9927
9928 QualType Ty = E->getType();
9929 llvm::Type *RealResTy = ConvertType(Ty);
9930 llvm::Type *IntTy =
9931 llvm::IntegerType::get(getLLVMContext(), getContext().getTypeSize(Ty));
9932 llvm::Type *PtrTy = IntTy->getPointerTo();
9933 LoadAddr = Builder.CreateBitCast(LoadAddr, PtrTy);
9934
9935 Function *F =
9936 CGM.getIntrinsic(BuiltinID == clang::AArch64::BI__builtin_arm_ldaex
9937 ? Intrinsic::aarch64_ldaxr
9938 : Intrinsic::aarch64_ldxr,
9939 PtrTy);
9940 CallInst *Val = Builder.CreateCall(F, LoadAddr, "ldxr");
9941 Val->addParamAttr(
9942 0, Attribute::get(getLLVMContext(), Attribute::ElementType, IntTy));
9943
9944 if (RealResTy->isPointerTy())
9945 return Builder.CreateIntToPtr(Val, RealResTy);
9946
9947 llvm::Type *IntResTy = llvm::IntegerType::get(
9948 getLLVMContext(), CGM.getDataLayout().getTypeSizeInBits(RealResTy));
9949 return Builder.CreateBitCast(Builder.CreateTruncOrBitCast(Val, IntResTy),
9950 RealResTy);
9951 }
9952
9953 if ((BuiltinID == clang::AArch64::BI__builtin_arm_strex ||
9954 BuiltinID == clang::AArch64::BI__builtin_arm_stlex) &&
9955 getContext().getTypeSize(E->getArg(0)->getType()) == 128) {
9956 Function *F =
9957 CGM.getIntrinsic(BuiltinID == clang::AArch64::BI__builtin_arm_stlex
9958 ? Intrinsic::aarch64_stlxp
9959 : Intrinsic::aarch64_stxp);
9960 llvm::Type *STy = llvm::StructType::get(Int64Ty, Int64Ty);
9961
9962 Address Tmp = CreateMemTemp(E->getArg(0)->getType());
9963 EmitAnyExprToMem(E->getArg(0), Tmp, Qualifiers(), /*init*/ true);
9964
9965 Tmp = Builder.CreateElementBitCast(Tmp, STy);
9966 llvm::Value *Val = Builder.CreateLoad(Tmp);
9967
9968 Value *Arg0 = Builder.CreateExtractValue(Val, 0);
9969 Value *Arg1 = Builder.CreateExtractValue(Val, 1);
9970 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)),
9971 Int8PtrTy);
9972 return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "stxp");
9973 }
9974
9975 if (BuiltinID == clang::AArch64::BI__builtin_arm_strex ||
9976 BuiltinID == clang::AArch64::BI__builtin_arm_stlex) {
9977 Value *StoreVal = EmitScalarExpr(E->getArg(0));
9978 Value *StoreAddr = EmitScalarExpr(E->getArg(1));
9979
9980 QualType Ty = E->getArg(0)->getType();
9981 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(),
9982 getContext().getTypeSize(Ty));
9983 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo());
9984
9985 if (StoreVal->getType()->isPointerTy())
9986 StoreVal = Builder.CreatePtrToInt(StoreVal, Int64Ty);
9987 else {
9988 llvm::Type *IntTy = llvm::IntegerType::get(
9989 getLLVMContext(),
9990 CGM.getDataLayout().getTypeSizeInBits(StoreVal->getType()));
9991 StoreVal = Builder.CreateBitCast(StoreVal, IntTy);
9992 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int64Ty);
9993 }
9994
9995 Function *F =
9996 CGM.getIntrinsic(BuiltinID == clang::AArch64::BI__builtin_arm_stlex
9997 ? Intrinsic::aarch64_stlxr
9998 : Intrinsic::aarch64_stxr,
9999 StoreAddr->getType());
10000 CallInst *CI = Builder.CreateCall(F, {StoreVal, StoreAddr}, "stxr");
10001 CI->addParamAttr(
10002 1, Attribute::get(getLLVMContext(), Attribute::ElementType, StoreTy));
10003 return CI;
10004 }
10005
10006 if (BuiltinID == clang::AArch64::BI__getReg) {
10007 Expr::EvalResult Result;
10008 if (!E->getArg(0)->EvaluateAsInt(Result, CGM.getContext()))
10009 llvm_unreachable("Sema will ensure that the parameter is constant");
10010
10011 llvm::APSInt Value = Result.Val.getInt();
10012 LLVMContext &Context = CGM.getLLVMContext();
10013 std::string Reg = Value == 31 ? "sp" : "x" + toString(Value, 10);
10014
10015 llvm::Metadata *Ops[] = {llvm::MDString::get(Context, Reg)};
10016 llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops);
10017 llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName);
10018
10019 llvm::Function *F =
10020 CGM.getIntrinsic(llvm::Intrinsic::read_register, {Int64Ty});
10021 return Builder.CreateCall(F, Metadata);
10022 }
10023
10024 if (BuiltinID == clang::AArch64::BI__break) {
10025 Expr::EvalResult Result;
10026 if (!E->getArg(0)->EvaluateAsInt(Result, CGM.getContext()))
10027 llvm_unreachable("Sema will ensure that the parameter is constant");
10028
10029 llvm::Function *F = CGM.getIntrinsic(llvm::Intrinsic::aarch64_break);
10030 return Builder.CreateCall(F, {EmitScalarExpr(E->getArg(0))});
10031 }
10032
10033 if (BuiltinID == clang::AArch64::BI__builtin_arm_clrex) {
10034 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_clrex);
10035 return Builder.CreateCall(F);
10036 }
10037
10038 if (BuiltinID == clang::AArch64::BI_ReadWriteBarrier)
10039 return Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent,
10040 llvm::SyncScope::SingleThread);
10041
10042 // CRC32
10043 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic;
10044 switch (BuiltinID) {
10045 case clang::AArch64::BI__builtin_arm_crc32b:
10046 CRCIntrinsicID = Intrinsic::aarch64_crc32b; break;
10047 case clang::AArch64::BI__builtin_arm_crc32cb:
10048 CRCIntrinsicID = Intrinsic::aarch64_crc32cb; break;
10049 case clang::AArch64::BI__builtin_arm_crc32h:
10050 CRCIntrinsicID = Intrinsic::aarch64_crc32h; break;
10051 case clang::AArch64::BI__builtin_arm_crc32ch:
10052 CRCIntrinsicID = Intrinsic::aarch64_crc32ch; break;
10053 case clang::AArch64::BI__builtin_arm_crc32w:
10054 CRCIntrinsicID = Intrinsic::aarch64_crc32w; break;
10055 case clang::AArch64::BI__builtin_arm_crc32cw:
10056 CRCIntrinsicID = Intrinsic::aarch64_crc32cw; break;
10057 case clang::AArch64::BI__builtin_arm_crc32d:
10058 CRCIntrinsicID = Intrinsic::aarch64_crc32x; break;
10059 case clang::AArch64::BI__builtin_arm_crc32cd:
10060 CRCIntrinsicID = Intrinsic::aarch64_crc32cx; break;
10061 }
10062
10063 if (CRCIntrinsicID != Intrinsic::not_intrinsic) {
10064 Value *Arg0 = EmitScalarExpr(E->getArg(0));
10065 Value *Arg1 = EmitScalarExpr(E->getArg(1));
10066 Function *F = CGM.getIntrinsic(CRCIntrinsicID);
10067
10068 llvm::Type *DataTy = F->getFunctionType()->getParamType(1);
10069 Arg1 = Builder.CreateZExtOrBitCast(Arg1, DataTy);
10070
10071 return Builder.CreateCall(F, {Arg0, Arg1});
10072 }
10073
10074 // Memory Operations (MOPS)
10075 if (BuiltinID == AArch64::BI__builtin_arm_mops_memset_tag) {
10076 Value *Dst = EmitScalarExpr(E->getArg(0));
10077 Value *Val = EmitScalarExpr(E->getArg(1));
10078 Value *Size = EmitScalarExpr(E->getArg(2));
10079 Dst = Builder.CreatePointerCast(Dst, Int8PtrTy);
10080 Val = Builder.CreateTrunc(Val, Int8Ty);
10081 Size = Builder.CreateIntCast(Size, Int64Ty, false);
10082 return Builder.CreateCall(
10083 CGM.getIntrinsic(Intrinsic::aarch64_mops_memset_tag), {Dst, Val, Size});
10084 }
10085
10086 // Memory Tagging Extensions (MTE) Intrinsics
10087 Intrinsic::ID MTEIntrinsicID = Intrinsic::not_intrinsic;
10088 switch (BuiltinID) {
10089 case clang::AArch64::BI__builtin_arm_irg:
10090 MTEIntrinsicID = Intrinsic::aarch64_irg; break;
10091 case clang::AArch64::BI__builtin_arm_addg:
10092 MTEIntrinsicID = Intrinsic::aarch64_addg; break;
10093 case clang::AArch64::BI__builtin_arm_gmi:
10094 MTEIntrinsicID = Intrinsic::aarch64_gmi; break;
10095 case clang::AArch64::BI__builtin_arm_ldg:
10096 MTEIntrinsicID = Intrinsic::aarch64_ldg; break;
10097 case clang::AArch64::BI__builtin_arm_stg:
10098 MTEIntrinsicID = Intrinsic::aarch64_stg; break;
10099 case clang::AArch64::BI__builtin_arm_subp:
10100 MTEIntrinsicID = Intrinsic::aarch64_subp; break;
10101 }
10102
10103 if (MTEIntrinsicID != Intrinsic::not_intrinsic) {
10104 llvm::Type *T = ConvertType(E->getType());
10105
10106 if (MTEIntrinsicID == Intrinsic::aarch64_irg) {
10107 Value *Pointer = EmitScalarExpr(E->getArg(0));
10108 Value *Mask = EmitScalarExpr(E->getArg(1));
10109
10110 Pointer = Builder.CreatePointerCast(Pointer, Int8PtrTy);
10111 Mask = Builder.CreateZExt(Mask, Int64Ty);
10112 Value *RV = Builder.CreateCall(
10113 CGM.getIntrinsic(MTEIntrinsicID), {Pointer, Mask});
10114 return Builder.CreatePointerCast(RV, T);
10115 }
10116 if (MTEIntrinsicID == Intrinsic::aarch64_addg) {
10117 Value *Pointer = EmitScalarExpr(E->getArg(0));
10118 Value *TagOffset = EmitScalarExpr(E->getArg(1));
10119
10120 Pointer = Builder.CreatePointerCast(Pointer, Int8PtrTy);
10121 TagOffset = Builder.CreateZExt(TagOffset, Int64Ty);
10122 Value *RV = Builder.CreateCall(
10123 CGM.getIntrinsic(MTEIntrinsicID), {Pointer, TagOffset});
10124 return Builder.CreatePointerCast(RV, T);
10125 }
10126 if (MTEIntrinsicID == Intrinsic::aarch64_gmi) {
10127 Value *Pointer = EmitScalarExpr(E->getArg(0));
10128 Value *ExcludedMask = EmitScalarExpr(E->getArg(1));
10129
10130 ExcludedMask = Builder.CreateZExt(ExcludedMask, Int64Ty);
10131 Pointer = Builder.CreatePointerCast(Pointer, Int8PtrTy);
10132 return Builder.CreateCall(
10133 CGM.getIntrinsic(MTEIntrinsicID), {Pointer, ExcludedMask});
10134 }
10135 // Although it is possible to supply a different return
10136 // address (first arg) to this intrinsic, for now we set
10137 // return address same as input address.
10138 if (MTEIntrinsicID == Intrinsic::aarch64_ldg) {
10139 Value *TagAddress = EmitScalarExpr(E->getArg(0));
10140 TagAddress = Builder.CreatePointerCast(TagAddress, Int8PtrTy);
10141 Value *RV = Builder.CreateCall(
10142 CGM.getIntrinsic(MTEIntrinsicID), {TagAddress, TagAddress});
10143 return Builder.CreatePointerCast(RV, T);
10144 }
10145 // Although it is possible to supply a different tag (to set)
10146 // to this intrinsic (as first arg), for now we supply
10147 // the tag that is in input address arg (common use case).
10148 if (MTEIntrinsicID == Intrinsic::aarch64_stg) {
10149 Value *TagAddress = EmitScalarExpr(E->getArg(0));
10150 TagAddress = Builder.CreatePointerCast(TagAddress, Int8PtrTy);
10151 return Builder.CreateCall(
10152 CGM.getIntrinsic(MTEIntrinsicID), {TagAddress, TagAddress});
10153 }
10154 if (MTEIntrinsicID == Intrinsic::aarch64_subp) {
10155 Value *PointerA = EmitScalarExpr(E->getArg(0));
10156 Value *PointerB = EmitScalarExpr(E->getArg(1));
10157 PointerA = Builder.CreatePointerCast(PointerA, Int8PtrTy);
10158 PointerB = Builder.CreatePointerCast(PointerB, Int8PtrTy);
10159 return Builder.CreateCall(
10160 CGM.getIntrinsic(MTEIntrinsicID), {PointerA, PointerB});
10161 }
10162 }
10163
10164 if (BuiltinID == clang::AArch64::BI__builtin_arm_rsr ||
10165 BuiltinID == clang::AArch64::BI__builtin_arm_rsr64 ||
10166 BuiltinID == clang::AArch64::BI__builtin_arm_rsr128 ||
10167 BuiltinID == clang::AArch64::BI__builtin_arm_rsrp ||
10168 BuiltinID == clang::AArch64::BI__builtin_arm_wsr ||
10169 BuiltinID == clang::AArch64::BI__builtin_arm_wsr64 ||
10170 BuiltinID == clang::AArch64::BI__builtin_arm_wsr128 ||
10171 BuiltinID == clang::AArch64::BI__builtin_arm_wsrp) {
10172
10173 SpecialRegisterAccessKind AccessKind = Write;
10174 if (BuiltinID == clang::AArch64::BI__builtin_arm_rsr ||
10175 BuiltinID == clang::AArch64::BI__builtin_arm_rsr64 ||
10176 BuiltinID == clang::AArch64::BI__builtin_arm_rsr128 ||
10177 BuiltinID == clang::AArch64::BI__builtin_arm_rsrp)
10178 AccessKind = VolatileRead;
10179
10180 bool IsPointerBuiltin = BuiltinID == clang::AArch64::BI__builtin_arm_rsrp ||
10181 BuiltinID == clang::AArch64::BI__builtin_arm_wsrp;
10182
10183 bool Is32Bit = BuiltinID == clang::AArch64::BI__builtin_arm_rsr ||
10184 BuiltinID == clang::AArch64::BI__builtin_arm_wsr;
10185
10186 bool Is128Bit = BuiltinID == clang::AArch64::BI__builtin_arm_rsr128 ||
10187 BuiltinID == clang::AArch64::BI__builtin_arm_wsr128;
10188
10189 llvm::Type *ValueType;
10190 llvm::Type *RegisterType = Int64Ty;
10191 if (Is32Bit) {
10192 ValueType = Int32Ty;
10193 } else if (Is128Bit) {
10194 llvm::Type *Int128Ty =
10195 llvm::IntegerType::getInt128Ty(CGM.getLLVMContext());
10196 ValueType = Int128Ty;
10197 RegisterType = Int128Ty;
10198 } else if (IsPointerBuiltin) {
10199 ValueType = VoidPtrTy;
10200 } else {
10201 ValueType = Int64Ty;
10202 };
10203
10204 return EmitSpecialRegisterBuiltin(*this, E, RegisterType, ValueType,
10205 AccessKind);
10206 }
10207
10208 if (BuiltinID == clang::AArch64::BI_ReadStatusReg ||
10209 BuiltinID == clang::AArch64::BI_WriteStatusReg) {
10210 LLVMContext &Context = CGM.getLLVMContext();
10211
10212 unsigned SysReg =
10213 E->getArg(0)->EvaluateKnownConstInt(getContext()).getZExtValue();
10214
10215 std::string SysRegStr;
10216 llvm::raw_string_ostream(SysRegStr) <<
10217 ((1 << 1) | ((SysReg >> 14) & 1)) << ":" <<
10218 ((SysReg >> 11) & 7) << ":" <<
10219 ((SysReg >> 7) & 15) << ":" <<
10220 ((SysReg >> 3) & 15) << ":" <<
10221 ( SysReg & 7);
10222
10223 llvm::Metadata *Ops[] = { llvm::MDString::get(Context, SysRegStr) };
10224 llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops);
10225 llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName);
10226
10227 llvm::Type *RegisterType = Int64Ty;
10228 llvm::Type *Types[] = { RegisterType };
10229
10230 if (BuiltinID == clang::AArch64::BI_ReadStatusReg) {
10231 llvm::Function *F = CGM.getIntrinsic(llvm::Intrinsic::read_register, Types);
10232
10233 return Builder.CreateCall(F, Metadata);
10234 }
10235
10236 llvm::Function *F = CGM.getIntrinsic(llvm::Intrinsic::write_register, Types);
10237 llvm::Value *ArgValue = EmitScalarExpr(E->getArg(1));
10238
10239 return Builder.CreateCall(F, { Metadata, ArgValue });
10240 }
10241
10242 if (BuiltinID == clang::AArch64::BI_AddressOfReturnAddress) {
10243 llvm::Function *F =
10244 CGM.getIntrinsic(Intrinsic::addressofreturnaddress, AllocaInt8PtrTy);
10245 return Builder.CreateCall(F);
10246 }
10247
10248 if (BuiltinID == clang::AArch64::BI__builtin_sponentry) {
10249 llvm::Function *F = CGM.getIntrinsic(Intrinsic::sponentry, AllocaInt8PtrTy);
10250 return Builder.CreateCall(F);
10251 }
10252
10253 if (BuiltinID == clang::AArch64::BI__mulh ||
10254 BuiltinID == clang::AArch64::BI__umulh) {
10255 llvm::Type *ResType = ConvertType(E->getType());
10256 llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128);
10257
10258 bool IsSigned = BuiltinID == clang::AArch64::BI__mulh;
10259 Value *LHS =
10260 Builder.CreateIntCast(EmitScalarExpr(E->getArg(0)), Int128Ty, IsSigned);
10261 Value *RHS =
10262 Builder.CreateIntCast(EmitScalarExpr(E->getArg(1)), Int128Ty, IsSigned);
10263
10264 Value *MulResult, *HigherBits;
10265 if (IsSigned) {
10266 MulResult = Builder.CreateNSWMul(LHS, RHS);
10267 HigherBits = Builder.CreateAShr(MulResult, 64);
10268 } else {
10269 MulResult = Builder.CreateNUWMul(LHS, RHS);
10270 HigherBits = Builder.CreateLShr(MulResult, 64);
10271 }
10272 HigherBits = Builder.CreateIntCast(HigherBits, ResType, IsSigned);
10273
10274 return HigherBits;
10275 }
10276
10277 if (BuiltinID == AArch64::BI__writex18byte ||
10278 BuiltinID == AArch64::BI__writex18word ||
10279 BuiltinID == AArch64::BI__writex18dword ||
10280 BuiltinID == AArch64::BI__writex18qword) {
10281 llvm::Type *IntTy = ConvertType(E->getArg(1)->getType());
10282
10283 // Read x18 as i8*
10284 LLVMContext &Context = CGM.getLLVMContext();
10285 llvm::Metadata *Ops[] = {llvm::MDString::get(Context, "x18")};
10286 llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops);
10287 llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName);
10288 llvm::Function *F =
10289 CGM.getIntrinsic(llvm::Intrinsic::read_register, {Int64Ty});
10290 llvm::Value *X18 = Builder.CreateCall(F, Metadata);
10291 X18 = Builder.CreateIntToPtr(X18, llvm::PointerType::get(Int8Ty, 0));
10292
10293 // Store val at x18 + offset
10294 Value *Offset = Builder.CreateZExt(EmitScalarExpr(E->getArg(0)), Int64Ty);
10295 Value *Ptr = Builder.CreateGEP(Int8Ty, X18, Offset);
10296 Ptr = Builder.CreatePointerCast(Ptr, llvm::PointerType::get(IntTy, 0));
10297 Value *Val = EmitScalarExpr(E->getArg(1));
10298 StoreInst *Store = Builder.CreateAlignedStore(Val, Ptr, CharUnits::One());
10299 return Store;
10300 }
10301
10302 if (BuiltinID == AArch64::BI__readx18byte ||
10303 BuiltinID == AArch64::BI__readx18word ||
10304 BuiltinID == AArch64::BI__readx18dword ||
10305 BuiltinID == AArch64::BI__readx18qword) {
10306 llvm::Type *IntTy = ConvertType(E->getType());
10307
10308 // Read x18 as i8*
10309 LLVMContext &Context = CGM.getLLVMContext();
10310 llvm::Metadata *Ops[] = {llvm::MDString::get(Context, "x18")};
10311 llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops);
10312 llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName);
10313 llvm::Function *F =
10314 CGM.getIntrinsic(llvm::Intrinsic::read_register, {Int64Ty});
10315 llvm::Value *X18 = Builder.CreateCall(F, Metadata);
10316 X18 = Builder.CreateIntToPtr(X18, llvm::PointerType::get(Int8Ty, 0));
10317
10318 // Load x18 + offset
10319 Value *Offset = Builder.CreateZExt(EmitScalarExpr(E->getArg(0)), Int64Ty);
10320 Value *Ptr = Builder.CreateGEP(Int8Ty, X18, Offset);
10321 Ptr = Builder.CreatePointerCast(Ptr, llvm::PointerType::get(IntTy, 0));
10322 LoadInst *Load = Builder.CreateAlignedLoad(IntTy, Ptr, CharUnits::One());
10323 return Load;
10324 }
10325
10326 // Handle MSVC intrinsics before argument evaluation to prevent double
10327 // evaluation.
10328 if (std::optional<MSVCIntrin> MsvcIntId =
10329 translateAarch64ToMsvcIntrin(BuiltinID))
10330 return EmitMSVCBuiltinExpr(*MsvcIntId, E);
10331
10332 // Some intrinsics are equivalent - if they are use the base intrinsic ID.
10333 auto It = llvm::find_if(NEONEquivalentIntrinsicMap, [BuiltinID](auto &P) {
10334 return P.first == BuiltinID;
10335 });
10336 if (It != end(NEONEquivalentIntrinsicMap))
10337 BuiltinID = It->second;
10338
10339 // Find out if any arguments are required to be integer constant
10340 // expressions.
10341 unsigned ICEArguments = 0;
10342 ASTContext::GetBuiltinTypeError Error;
10343 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments);
10344 assert(Error == ASTContext::GE_None && "Should not codegen an error");
10345
10346 llvm::SmallVector<Value*, 4> Ops;
10347 Address PtrOp0 = Address::invalid();
10348 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) {
10349 if (i == 0) {
10350 switch (BuiltinID) {
10351 case NEON::BI__builtin_neon_vld1_v:
10352 case NEON::BI__builtin_neon_vld1q_v:
10353 case NEON::BI__builtin_neon_vld1_dup_v:
10354 case NEON::BI__builtin_neon_vld1q_dup_v:
10355 case NEON::BI__builtin_neon_vld1_lane_v:
10356 case NEON::BI__builtin_neon_vld1q_lane_v:
10357 case NEON::BI__builtin_neon_vst1_v:
10358 case NEON::BI__builtin_neon_vst1q_v:
10359 case NEON::BI__builtin_neon_vst1_lane_v:
10360 case NEON::BI__builtin_neon_vst1q_lane_v:
10361 // Get the alignment for the argument in addition to the value;
10362 // we'll use it later.
10363 PtrOp0 = EmitPointerWithAlignment(E->getArg(0));
10364 Ops.push_back(PtrOp0.getPointer());
10365 continue;
10366 }
10367 }
10368 if ((ICEArguments & (1 << i)) == 0) {
10369 Ops.push_back(EmitScalarExpr(E->getArg(i)));
10370 } else {
10371 // If this is required to be a constant, constant fold it so that we know
10372 // that the generated intrinsic gets a ConstantInt.
10373 Ops.push_back(llvm::ConstantInt::get(
10374 getLLVMContext(),
10375 *E->getArg(i)->getIntegerConstantExpr(getContext())));
10376 }
10377 }
10378
10379 auto SISDMap = ArrayRef(AArch64SISDIntrinsicMap);
10380 const ARMVectorIntrinsicInfo *Builtin = findARMVectorIntrinsicInMap(
10381 SISDMap, BuiltinID, AArch64SISDIntrinsicsProvenSorted);
10382
10383 if (Builtin) {
10384 Ops.push_back(EmitScalarExpr(E->getArg(E->getNumArgs() - 1)));
10385 Value *Result = EmitCommonNeonSISDBuiltinExpr(*this, *Builtin, Ops, E);
10386 assert(Result && "SISD intrinsic should have been handled");
10387 return Result;
10388 }
10389
10390 const Expr *Arg = E->getArg(E->getNumArgs()-1);
10391 NeonTypeFlags Type(0);
10392 if (std::optional<llvm::APSInt> Result =
10393 Arg->getIntegerConstantExpr(getContext()))
10394 // Determine the type of this overloaded NEON intrinsic.
10395 Type = NeonTypeFlags(Result->getZExtValue());
10396
10397 bool usgn = Type.isUnsigned();
10398 bool quad = Type.isQuad();
10399
10400 // Handle non-overloaded intrinsics first.
10401 switch (BuiltinID) {
10402 default: break;
10403 case NEON::BI__builtin_neon_vabsh_f16:
10404 Ops.push_back(EmitScalarExpr(E->getArg(0)));
10405 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::fabs, HalfTy), Ops, "vabs");
10406 case NEON::BI__builtin_neon_vaddq_p128: {
10407 llvm::Type *Ty = GetNeonType(this, NeonTypeFlags::Poly128);
10408 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10409 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
10410 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
10411 Ops[0] = Builder.CreateXor(Ops[0], Ops[1]);
10412 llvm::Type *Int128Ty = llvm::Type::getIntNTy(getLLVMContext(), 128);
10413 return Builder.CreateBitCast(Ops[0], Int128Ty);
10414 }
10415 case NEON::BI__builtin_neon_vldrq_p128: {
10416 llvm::Type *Int128Ty = llvm::Type::getIntNTy(getLLVMContext(), 128);
10417 llvm::Type *Int128PTy = llvm::PointerType::get(Int128Ty, 0);
10418 Value *Ptr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int128PTy);
10419 return Builder.CreateAlignedLoad(Int128Ty, Ptr,
10420 CharUnits::fromQuantity(16));
10421 }
10422 case NEON::BI__builtin_neon_vstrq_p128: {
10423 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128);
10424 Value *Ptr = Builder.CreateBitCast(Ops[0], Int128PTy);
10425 return Builder.CreateDefaultAlignedStore(EmitScalarExpr(E->getArg(1)), Ptr);
10426 }
10427 case NEON::BI__builtin_neon_vcvts_f32_u32:
10428 case NEON::BI__builtin_neon_vcvtd_f64_u64:
10429 usgn = true;
10430 [[fallthrough]];
10431 case NEON::BI__builtin_neon_vcvts_f32_s32:
10432 case NEON::BI__builtin_neon_vcvtd_f64_s64: {
10433 Ops.push_back(EmitScalarExpr(E->getArg(0)));
10434 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64;
10435 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty;
10436 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy;
10437 Ops[0] = Builder.CreateBitCast(Ops[0], InTy);
10438 if (usgn)
10439 return Builder.CreateUIToFP(Ops[0], FTy);
10440 return Builder.CreateSIToFP(Ops[0], FTy);
10441 }
10442 case NEON::BI__builtin_neon_vcvth_f16_u16:
10443 case NEON::BI__builtin_neon_vcvth_f16_u32:
10444 case NEON::BI__builtin_neon_vcvth_f16_u64:
10445 usgn = true;
10446 [[fallthrough]];
10447 case NEON::BI__builtin_neon_vcvth_f16_s16:
10448 case NEON::BI__builtin_neon_vcvth_f16_s32:
10449 case NEON::BI__builtin_neon_vcvth_f16_s64: {
10450 Ops.push_back(EmitScalarExpr(E->getArg(0)));
10451 llvm::Type *FTy = HalfTy;
10452 llvm::Type *InTy;
10453 if (Ops[0]->getType()->getPrimitiveSizeInBits() == 64)
10454 InTy = Int64Ty;
10455 else if (Ops[0]->getType()->getPrimitiveSizeInBits() == 32)
10456 InTy = Int32Ty;
10457 else
10458 InTy = Int16Ty;
10459 Ops[0] = Builder.CreateBitCast(Ops[0], InTy);
10460 if (usgn)
10461 return Builder.CreateUIToFP(Ops[0], FTy);
10462 return Builder.CreateSIToFP(Ops[0], FTy);
10463 }
10464 case NEON::BI__builtin_neon_vcvtah_u16_f16:
10465 case NEON::BI__builtin_neon_vcvtmh_u16_f16:
10466 case NEON::BI__builtin_neon_vcvtnh_u16_f16:
10467 case NEON::BI__builtin_neon_vcvtph_u16_f16:
10468 case NEON::BI__builtin_neon_vcvth_u16_f16:
10469 case NEON::BI__builtin_neon_vcvtah_s16_f16:
10470 case NEON::BI__builtin_neon_vcvtmh_s16_f16:
10471 case NEON::BI__builtin_neon_vcvtnh_s16_f16:
10472 case NEON::BI__builtin_neon_vcvtph_s16_f16:
10473 case NEON::BI__builtin_neon_vcvth_s16_f16: {
10474 unsigned Int;
10475 llvm::Type* InTy = Int32Ty;
10476 llvm::Type* FTy = HalfTy;
10477 llvm::Type *Tys[2] = {InTy, FTy};
10478 Ops.push_back(EmitScalarExpr(E->getArg(0)));
10479 switch (BuiltinID) {
10480 default: llvm_unreachable("missing builtin ID in switch!");
10481 case NEON::BI__builtin_neon_vcvtah_u16_f16:
10482 Int = Intrinsic::aarch64_neon_fcvtau; break;
10483 case NEON::BI__builtin_neon_vcvtmh_u16_f16:
10484 Int = Intrinsic::aarch64_neon_fcvtmu; break;
10485 case NEON::BI__builtin_neon_vcvtnh_u16_f16:
10486 Int = Intrinsic::aarch64_neon_fcvtnu; break;
10487 case NEON::BI__builtin_neon_vcvtph_u16_f16:
10488 Int = Intrinsic::aarch64_neon_fcvtpu; break;
10489 case NEON::BI__builtin_neon_vcvth_u16_f16:
10490 Int = Intrinsic::aarch64_neon_fcvtzu; break;
10491 case NEON::BI__builtin_neon_vcvtah_s16_f16:
10492 Int = Intrinsic::aarch64_neon_fcvtas; break;
10493 case NEON::BI__builtin_neon_vcvtmh_s16_f16:
10494 Int = Intrinsic::aarch64_neon_fcvtms; break;
10495 case NEON::BI__builtin_neon_vcvtnh_s16_f16:
10496 Int = Intrinsic::aarch64_neon_fcvtns; break;
10497 case NEON::BI__builtin_neon_vcvtph_s16_f16:
10498 Int = Intrinsic::aarch64_neon_fcvtps; break;
10499 case NEON::BI__builtin_neon_vcvth_s16_f16:
10500 Int = Intrinsic::aarch64_neon_fcvtzs; break;
10501 }
10502 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "fcvt");
10503 return Builder.CreateTrunc(Ops[0], Int16Ty);
10504 }
10505 case NEON::BI__builtin_neon_vcaleh_f16:
10506 case NEON::BI__builtin_neon_vcalth_f16:
10507 case NEON::BI__builtin_neon_vcageh_f16:
10508 case NEON::BI__builtin_neon_vcagth_f16: {
10509 unsigned Int;
10510 llvm::Type* InTy = Int32Ty;
10511 llvm::Type* FTy = HalfTy;
10512 llvm::Type *Tys[2] = {InTy, FTy};
10513 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10514 switch (BuiltinID) {
10515 default: llvm_unreachable("missing builtin ID in switch!");
10516 case NEON::BI__builtin_neon_vcageh_f16:
10517 Int = Intrinsic::aarch64_neon_facge; break;
10518 case NEON::BI__builtin_neon_vcagth_f16:
10519 Int = Intrinsic::aarch64_neon_facgt; break;
10520 case NEON::BI__builtin_neon_vcaleh_f16:
10521 Int = Intrinsic::aarch64_neon_facge; std::swap(Ops[0], Ops[1]); break;
10522 case NEON::BI__builtin_neon_vcalth_f16:
10523 Int = Intrinsic::aarch64_neon_facgt; std::swap(Ops[0], Ops[1]); break;
10524 }
10525 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "facg");
10526 return Builder.CreateTrunc(Ops[0], Int16Ty);
10527 }
10528 case NEON::BI__builtin_neon_vcvth_n_s16_f16:
10529 case NEON::BI__builtin_neon_vcvth_n_u16_f16: {
10530 unsigned Int;
10531 llvm::Type* InTy = Int32Ty;
10532 llvm::Type* FTy = HalfTy;
10533 llvm::Type *Tys[2] = {InTy, FTy};
10534 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10535 switch (BuiltinID) {
10536 default: llvm_unreachable("missing builtin ID in switch!");
10537 case NEON::BI__builtin_neon_vcvth_n_s16_f16:
10538 Int = Intrinsic::aarch64_neon_vcvtfp2fxs; break;
10539 case NEON::BI__builtin_neon_vcvth_n_u16_f16:
10540 Int = Intrinsic::aarch64_neon_vcvtfp2fxu; break;
10541 }
10542 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "fcvth_n");
10543 return Builder.CreateTrunc(Ops[0], Int16Ty);
10544 }
10545 case NEON::BI__builtin_neon_vcvth_n_f16_s16:
10546 case NEON::BI__builtin_neon_vcvth_n_f16_u16: {
10547 unsigned Int;
10548 llvm::Type* FTy = HalfTy;
10549 llvm::Type* InTy = Int32Ty;
10550 llvm::Type *Tys[2] = {FTy, InTy};
10551 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10552 switch (BuiltinID) {
10553 default: llvm_unreachable("missing builtin ID in switch!");
10554 case NEON::BI__builtin_neon_vcvth_n_f16_s16:
10555 Int = Intrinsic::aarch64_neon_vcvtfxs2fp;
10556 Ops[0] = Builder.CreateSExt(Ops[0], InTy, "sext");
10557 break;
10558 case NEON::BI__builtin_neon_vcvth_n_f16_u16:
10559 Int = Intrinsic::aarch64_neon_vcvtfxu2fp;
10560 Ops[0] = Builder.CreateZExt(Ops[0], InTy);
10561 break;
10562 }
10563 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "fcvth_n");
10564 }
10565 case NEON::BI__builtin_neon_vpaddd_s64: {
10566 auto *Ty = llvm::FixedVectorType::get(Int64Ty, 2);
10567 Value *Vec = EmitScalarExpr(E->getArg(0));
10568 // The vector is v2f64, so make sure it's bitcast to that.
10569 Vec = Builder.CreateBitCast(Vec, Ty, "v2i64");
10570 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0);
10571 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1);
10572 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0");
10573 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1");
10574 // Pairwise addition of a v2f64 into a scalar f64.
10575 return Builder.CreateAdd(Op0, Op1, "vpaddd");
10576 }
10577 case NEON::BI__builtin_neon_vpaddd_f64: {
10578 auto *Ty = llvm::FixedVectorType::get(DoubleTy, 2);
10579 Value *Vec = EmitScalarExpr(E->getArg(0));
10580 // The vector is v2f64, so make sure it's bitcast to that.
10581 Vec = Builder.CreateBitCast(Vec, Ty, "v2f64");
10582 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0);
10583 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1);
10584 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0");
10585 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1");
10586 // Pairwise addition of a v2f64 into a scalar f64.
10587 return Builder.CreateFAdd(Op0, Op1, "vpaddd");
10588 }
10589 case NEON::BI__builtin_neon_vpadds_f32: {
10590 auto *Ty = llvm::FixedVectorType::get(FloatTy, 2);
10591 Value *Vec = EmitScalarExpr(E->getArg(0));
10592 // The vector is v2f32, so make sure it's bitcast to that.
10593 Vec = Builder.CreateBitCast(Vec, Ty, "v2f32");
10594 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0);
10595 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1);
10596 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0");
10597 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1");
10598 // Pairwise addition of a v2f32 into a scalar f32.
10599 return Builder.CreateFAdd(Op0, Op1, "vpaddd");
10600 }
10601 case NEON::BI__builtin_neon_vceqzd_s64:
10602 case NEON::BI__builtin_neon_vceqzd_f64:
10603 case NEON::BI__builtin_neon_vceqzs_f32:
10604 case NEON::BI__builtin_neon_vceqzh_f16:
10605 Ops.push_back(EmitScalarExpr(E->getArg(0)));
10606 return EmitAArch64CompareBuiltinExpr(
10607 Ops[0], ConvertType(E->getCallReturnType(getContext())),
10608 ICmpInst::FCMP_OEQ, ICmpInst::ICMP_EQ, "vceqz");
10609 case NEON::BI__builtin_neon_vcgezd_s64:
10610 case NEON::BI__builtin_neon_vcgezd_f64:
10611 case NEON::BI__builtin_neon_vcgezs_f32:
10612 case NEON::BI__builtin_neon_vcgezh_f16:
10613 Ops.push_back(EmitScalarExpr(E->getArg(0)));
10614 return EmitAArch64CompareBuiltinExpr(
10615 Ops[0], ConvertType(E->getCallReturnType(getContext())),
10616 ICmpInst::FCMP_OGE, ICmpInst::ICMP_SGE, "vcgez");
10617 case NEON::BI__builtin_neon_vclezd_s64:
10618 case NEON::BI__builtin_neon_vclezd_f64:
10619 case NEON::BI__builtin_neon_vclezs_f32:
10620 case NEON::BI__builtin_neon_vclezh_f16:
10621 Ops.push_back(EmitScalarExpr(E->getArg(0)));
10622 return EmitAArch64CompareBuiltinExpr(
10623 Ops[0], ConvertType(E->getCallReturnType(getContext())),
10624 ICmpInst::FCMP_OLE, ICmpInst::ICMP_SLE, "vclez");
10625 case NEON::BI__builtin_neon_vcgtzd_s64:
10626 case NEON::BI__builtin_neon_vcgtzd_f64:
10627 case NEON::BI__builtin_neon_vcgtzs_f32:
10628 case NEON::BI__builtin_neon_vcgtzh_f16:
10629 Ops.push_back(EmitScalarExpr(E->getArg(0)));
10630 return EmitAArch64CompareBuiltinExpr(
10631 Ops[0], ConvertType(E->getCallReturnType(getContext())),
10632 ICmpInst::FCMP_OGT, ICmpInst::ICMP_SGT, "vcgtz");
10633 case NEON::BI__builtin_neon_vcltzd_s64:
10634 case NEON::BI__builtin_neon_vcltzd_f64:
10635 case NEON::BI__builtin_neon_vcltzs_f32:
10636 case NEON::BI__builtin_neon_vcltzh_f16:
10637 Ops.push_back(EmitScalarExpr(E->getArg(0)));
10638 return EmitAArch64CompareBuiltinExpr(
10639 Ops[0], ConvertType(E->getCallReturnType(getContext())),
10640 ICmpInst::FCMP_OLT, ICmpInst::ICMP_SLT, "vcltz");
10641
10642 case NEON::BI__builtin_neon_vceqzd_u64: {
10643 Ops.push_back(EmitScalarExpr(E->getArg(0)));
10644 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty);
10645 Ops[0] =
10646 Builder.CreateICmpEQ(Ops[0], llvm::Constant::getNullValue(Int64Ty));
10647 return Builder.CreateSExt(Ops[0], Int64Ty, "vceqzd");
10648 }
10649 case NEON::BI__builtin_neon_vceqd_f64:
10650 case NEON::BI__builtin_neon_vcled_f64:
10651 case NEON::BI__builtin_neon_vcltd_f64:
10652 case NEON::BI__builtin_neon_vcged_f64:
10653 case NEON::BI__builtin_neon_vcgtd_f64: {
10654 llvm::CmpInst::Predicate P;
10655 switch (BuiltinID) {
10656 default: llvm_unreachable("missing builtin ID in switch!");
10657 case NEON::BI__builtin_neon_vceqd_f64: P = llvm::FCmpInst::FCMP_OEQ; break;
10658 case NEON::BI__builtin_neon_vcled_f64: P = llvm::FCmpInst::FCMP_OLE; break;
10659 case NEON::BI__builtin_neon_vcltd_f64: P = llvm::FCmpInst::FCMP_OLT; break;
10660 case NEON::BI__builtin_neon_vcged_f64: P = llvm::FCmpInst::FCMP_OGE; break;
10661 case NEON::BI__builtin_neon_vcgtd_f64: P = llvm::FCmpInst::FCMP_OGT; break;
10662 }
10663 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10664 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy);
10665 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy);
10666 if (P == llvm::FCmpInst::FCMP_OEQ)
10667 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]);
10668 else
10669 Ops[0] = Builder.CreateFCmpS(P, Ops[0], Ops[1]);
10670 return Builder.CreateSExt(Ops[0], Int64Ty, "vcmpd");
10671 }
10672 case NEON::BI__builtin_neon_vceqs_f32:
10673 case NEON::BI__builtin_neon_vcles_f32:
10674 case NEON::BI__builtin_neon_vclts_f32:
10675 case NEON::BI__builtin_neon_vcges_f32:
10676 case NEON::BI__builtin_neon_vcgts_f32: {
10677 llvm::CmpInst::Predicate P;
10678 switch (BuiltinID) {
10679 default: llvm_unreachable("missing builtin ID in switch!");
10680 case NEON::BI__builtin_neon_vceqs_f32: P = llvm::FCmpInst::FCMP_OEQ; break;
10681 case NEON::BI__builtin_neon_vcles_f32: P = llvm::FCmpInst::FCMP_OLE; break;
10682 case NEON::BI__builtin_neon_vclts_f32: P = llvm::FCmpInst::FCMP_OLT; break;
10683 case NEON::BI__builtin_neon_vcges_f32: P = llvm::FCmpInst::FCMP_OGE; break;
10684 case NEON::BI__builtin_neon_vcgts_f32: P = llvm::FCmpInst::FCMP_OGT; break;
10685 }
10686 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10687 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy);
10688 Ops[1] = Builder.CreateBitCast(Ops[1], FloatTy);
10689 if (P == llvm::FCmpInst::FCMP_OEQ)
10690 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]);
10691 else
10692 Ops[0] = Builder.CreateFCmpS(P, Ops[0], Ops[1]);
10693 return Builder.CreateSExt(Ops[0], Int32Ty, "vcmpd");
10694 }
10695 case NEON::BI__builtin_neon_vceqh_f16:
10696 case NEON::BI__builtin_neon_vcleh_f16:
10697 case NEON::BI__builtin_neon_vclth_f16:
10698 case NEON::BI__builtin_neon_vcgeh_f16:
10699 case NEON::BI__builtin_neon_vcgth_f16: {
10700 llvm::CmpInst::Predicate P;
10701 switch (BuiltinID) {
10702 default: llvm_unreachable("missing builtin ID in switch!");
10703 case NEON::BI__builtin_neon_vceqh_f16: P = llvm::FCmpInst::FCMP_OEQ; break;
10704 case NEON::BI__builtin_neon_vcleh_f16: P = llvm::FCmpInst::FCMP_OLE; break;
10705 case NEON::BI__builtin_neon_vclth_f16: P = llvm::FCmpInst::FCMP_OLT; break;
10706 case NEON::BI__builtin_neon_vcgeh_f16: P = llvm::FCmpInst::FCMP_OGE; break;
10707 case NEON::BI__builtin_neon_vcgth_f16: P = llvm::FCmpInst::FCMP_OGT; break;
10708 }
10709 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10710 Ops[0] = Builder.CreateBitCast(Ops[0], HalfTy);
10711 Ops[1] = Builder.CreateBitCast(Ops[1], HalfTy);
10712 if (P == llvm::FCmpInst::FCMP_OEQ)
10713 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]);
10714 else
10715 Ops[0] = Builder.CreateFCmpS(P, Ops[0], Ops[1]);
10716 return Builder.CreateSExt(Ops[0], Int16Ty, "vcmpd");
10717 }
10718 case NEON::BI__builtin_neon_vceqd_s64:
10719 case NEON::BI__builtin_neon_vceqd_u64:
10720 case NEON::BI__builtin_neon_vcgtd_s64:
10721 case NEON::BI__builtin_neon_vcgtd_u64:
10722 case NEON::BI__builtin_neon_vcltd_s64:
10723 case NEON::BI__builtin_neon_vcltd_u64:
10724 case NEON::BI__builtin_neon_vcged_u64:
10725 case NEON::BI__builtin_neon_vcged_s64:
10726 case NEON::BI__builtin_neon_vcled_u64:
10727 case NEON::BI__builtin_neon_vcled_s64: {
10728 llvm::CmpInst::Predicate P;
10729 switch (BuiltinID) {
10730 default: llvm_unreachable("missing builtin ID in switch!");
10731 case NEON::BI__builtin_neon_vceqd_s64:
10732 case NEON::BI__builtin_neon_vceqd_u64:P = llvm::ICmpInst::ICMP_EQ;break;
10733 case NEON::BI__builtin_neon_vcgtd_s64:P = llvm::ICmpInst::ICMP_SGT;break;
10734 case NEON::BI__builtin_neon_vcgtd_u64:P = llvm::ICmpInst::ICMP_UGT;break;
10735 case NEON::BI__builtin_neon_vcltd_s64:P = llvm::ICmpInst::ICMP_SLT;break;
10736 case NEON::BI__builtin_neon_vcltd_u64:P = llvm::ICmpInst::ICMP_ULT;break;
10737 case NEON::BI__builtin_neon_vcged_u64:P = llvm::ICmpInst::ICMP_UGE;break;
10738 case NEON::BI__builtin_neon_vcged_s64:P = llvm::ICmpInst::ICMP_SGE;break;
10739 case NEON::BI__builtin_neon_vcled_u64:P = llvm::ICmpInst::ICMP_ULE;break;
10740 case NEON::BI__builtin_neon_vcled_s64:P = llvm::ICmpInst::ICMP_SLE;break;
10741 }
10742 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10743 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty);
10744 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty);
10745 Ops[0] = Builder.CreateICmp(P, Ops[0], Ops[1]);
10746 return Builder.CreateSExt(Ops[0], Int64Ty, "vceqd");
10747 }
10748 case NEON::BI__builtin_neon_vtstd_s64:
10749 case NEON::BI__builtin_neon_vtstd_u64: {
10750 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10751 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty);
10752 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty);
10753 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]);
10754 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0],
10755 llvm::Constant::getNullValue(Int64Ty));
10756 return Builder.CreateSExt(Ops[0], Int64Ty, "vtstd");
10757 }
10758 case NEON::BI__builtin_neon_vset_lane_i8:
10759 case NEON::BI__builtin_neon_vset_lane_i16:
10760 case NEON::BI__builtin_neon_vset_lane_i32:
10761 case NEON::BI__builtin_neon_vset_lane_i64:
10762 case NEON::BI__builtin_neon_vset_lane_bf16:
10763 case NEON::BI__builtin_neon_vset_lane_f32:
10764 case NEON::BI__builtin_neon_vsetq_lane_i8:
10765 case NEON::BI__builtin_neon_vsetq_lane_i16:
10766 case NEON::BI__builtin_neon_vsetq_lane_i32:
10767 case NEON::BI__builtin_neon_vsetq_lane_i64:
10768 case NEON::BI__builtin_neon_vsetq_lane_bf16:
10769 case NEON::BI__builtin_neon_vsetq_lane_f32:
10770 Ops.push_back(EmitScalarExpr(E->getArg(2)));
10771 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane");
10772 case NEON::BI__builtin_neon_vset_lane_f64:
10773 // The vector type needs a cast for the v1f64 variant.
10774 Ops[1] =
10775 Builder.CreateBitCast(Ops[1], llvm::FixedVectorType::get(DoubleTy, 1));
10776 Ops.push_back(EmitScalarExpr(E->getArg(2)));
10777 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane");
10778 case NEON::BI__builtin_neon_vsetq_lane_f64:
10779 // The vector type needs a cast for the v2f64 variant.
10780 Ops[1] =
10781 Builder.CreateBitCast(Ops[1], llvm::FixedVectorType::get(DoubleTy, 2));
10782 Ops.push_back(EmitScalarExpr(E->getArg(2)));
10783 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane");
10784
10785 case NEON::BI__builtin_neon_vget_lane_i8:
10786 case NEON::BI__builtin_neon_vdupb_lane_i8:
10787 Ops[0] =
10788 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(Int8Ty, 8));
10789 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
10790 "vget_lane");
10791 case NEON::BI__builtin_neon_vgetq_lane_i8:
10792 case NEON::BI__builtin_neon_vdupb_laneq_i8:
10793 Ops[0] =
10794 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(Int8Ty, 16));
10795 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
10796 "vgetq_lane");
10797 case NEON::BI__builtin_neon_vget_lane_i16:
10798 case NEON::BI__builtin_neon_vduph_lane_i16:
10799 Ops[0] =
10800 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(Int16Ty, 4));
10801 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
10802 "vget_lane");
10803 case NEON::BI__builtin_neon_vgetq_lane_i16:
10804 case NEON::BI__builtin_neon_vduph_laneq_i16:
10805 Ops[0] =
10806 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(Int16Ty, 8));
10807 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
10808 "vgetq_lane");
10809 case NEON::BI__builtin_neon_vget_lane_i32:
10810 case NEON::BI__builtin_neon_vdups_lane_i32:
10811 Ops[0] =
10812 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(Int32Ty, 2));
10813 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
10814 "vget_lane");
10815 case NEON::BI__builtin_neon_vdups_lane_f32:
10816 Ops[0] =
10817 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(FloatTy, 2));
10818 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
10819 "vdups_lane");
10820 case NEON::BI__builtin_neon_vgetq_lane_i32:
10821 case NEON::BI__builtin_neon_vdups_laneq_i32:
10822 Ops[0] =
10823 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(Int32Ty, 4));
10824 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
10825 "vgetq_lane");
10826 case NEON::BI__builtin_neon_vget_lane_i64:
10827 case NEON::BI__builtin_neon_vdupd_lane_i64:
10828 Ops[0] =
10829 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(Int64Ty, 1));
10830 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
10831 "vget_lane");
10832 case NEON::BI__builtin_neon_vdupd_lane_f64:
10833 Ops[0] =
10834 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(DoubleTy, 1));
10835 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
10836 "vdupd_lane");
10837 case NEON::BI__builtin_neon_vgetq_lane_i64:
10838 case NEON::BI__builtin_neon_vdupd_laneq_i64:
10839 Ops[0] =
10840 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(Int64Ty, 2));
10841 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
10842 "vgetq_lane");
10843 case NEON::BI__builtin_neon_vget_lane_f32:
10844 Ops[0] =
10845 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(FloatTy, 2));
10846 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
10847 "vget_lane");
10848 case NEON::BI__builtin_neon_vget_lane_f64:
10849 Ops[0] =
10850 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(DoubleTy, 1));
10851 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
10852 "vget_lane");
10853 case NEON::BI__builtin_neon_vgetq_lane_f32:
10854 case NEON::BI__builtin_neon_vdups_laneq_f32:
10855 Ops[0] =
10856 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(FloatTy, 4));
10857 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
10858 "vgetq_lane");
10859 case NEON::BI__builtin_neon_vgetq_lane_f64:
10860 case NEON::BI__builtin_neon_vdupd_laneq_f64:
10861 Ops[0] =
10862 Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(DoubleTy, 2));
10863 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
10864 "vgetq_lane");
10865 case NEON::BI__builtin_neon_vaddh_f16:
10866 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10867 return Builder.CreateFAdd(Ops[0], Ops[1], "vaddh");
10868 case NEON::BI__builtin_neon_vsubh_f16:
10869 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10870 return Builder.CreateFSub(Ops[0], Ops[1], "vsubh");
10871 case NEON::BI__builtin_neon_vmulh_f16:
10872 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10873 return Builder.CreateFMul(Ops[0], Ops[1], "vmulh");
10874 case NEON::BI__builtin_neon_vdivh_f16:
10875 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10876 return Builder.CreateFDiv(Ops[0], Ops[1], "vdivh");
10877 case NEON::BI__builtin_neon_vfmah_f16:
10878 // NEON intrinsic puts accumulator first, unlike the LLVM fma.
10879 return emitCallMaybeConstrainedFPBuiltin(
10880 *this, Intrinsic::fma, Intrinsic::experimental_constrained_fma, HalfTy,
10881 {EmitScalarExpr(E->getArg(1)), EmitScalarExpr(E->getArg(2)), Ops[0]});
10882 case NEON::BI__builtin_neon_vfmsh_f16: {
10883 // FIXME: This should be an fneg instruction:
10884 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(HalfTy);
10885 Value* Sub = Builder.CreateFSub(Zero, EmitScalarExpr(E->getArg(1)), "vsubh");
10886
10887 // NEON intrinsic puts accumulator first, unlike the LLVM fma.
10888 return emitCallMaybeConstrainedFPBuiltin(
10889 *this, Intrinsic::fma, Intrinsic::experimental_constrained_fma, HalfTy,
10890 {Sub, EmitScalarExpr(E->getArg(2)), Ops[0]});
10891 }
10892 case NEON::BI__builtin_neon_vaddd_s64:
10893 case NEON::BI__builtin_neon_vaddd_u64:
10894 return Builder.CreateAdd(Ops[0], EmitScalarExpr(E->getArg(1)), "vaddd");
10895 case NEON::BI__builtin_neon_vsubd_s64:
10896 case NEON::BI__builtin_neon_vsubd_u64:
10897 return Builder.CreateSub(Ops[0], EmitScalarExpr(E->getArg(1)), "vsubd");
10898 case NEON::BI__builtin_neon_vqdmlalh_s16:
10899 case NEON::BI__builtin_neon_vqdmlslh_s16: {
10900 SmallVector<Value *, 2> ProductOps;
10901 ProductOps.push_back(vectorWrapScalar16(Ops[1]));
10902 ProductOps.push_back(vectorWrapScalar16(EmitScalarExpr(E->getArg(2))));
10903 auto *VTy = llvm::FixedVectorType::get(Int32Ty, 4);
10904 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy),
10905 ProductOps, "vqdmlXl");
10906 Constant *CI = ConstantInt::get(SizeTy, 0);
10907 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0");
10908
10909 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlalh_s16
10910 ? Intrinsic::aarch64_neon_sqadd
10911 : Intrinsic::aarch64_neon_sqsub;
10912 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int32Ty), Ops, "vqdmlXl");
10913 }
10914 case NEON::BI__builtin_neon_vqshlud_n_s64: {
10915 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10916 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty);
10917 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqshlu, Int64Ty),
10918 Ops, "vqshlu_n");
10919 }
10920 case NEON::BI__builtin_neon_vqshld_n_u64:
10921 case NEON::BI__builtin_neon_vqshld_n_s64: {
10922 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vqshld_n_u64
10923 ? Intrinsic::aarch64_neon_uqshl
10924 : Intrinsic::aarch64_neon_sqshl;
10925 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10926 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty);
10927 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vqshl_n");
10928 }
10929 case NEON::BI__builtin_neon_vrshrd_n_u64:
10930 case NEON::BI__builtin_neon_vrshrd_n_s64: {
10931 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrshrd_n_u64
10932 ? Intrinsic::aarch64_neon_urshl
10933 : Intrinsic::aarch64_neon_srshl;
10934 Ops.push_back(EmitScalarExpr(E->getArg(1)));
10935 int SV = cast<ConstantInt>(Ops[1])->getSExtValue();
10936 Ops[1] = ConstantInt::get(Int64Ty, -SV);
10937 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vrshr_n");
10938 }
10939 case NEON::BI__builtin_neon_vrsrad_n_u64:
10940 case NEON::BI__builtin_neon_vrsrad_n_s64: {
10941 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrsrad_n_u64
10942 ? Intrinsic::aarch64_neon_urshl
10943 : Intrinsic::aarch64_neon_srshl;
10944 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty);
10945 Ops.push_back(Builder.CreateNeg(EmitScalarExpr(E->getArg(2))));
10946 Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Int64Ty),
10947 {Ops[1], Builder.CreateSExt(Ops[2], Int64Ty)});
10948 return Builder.CreateAdd(Ops[0], Builder.CreateBitCast(Ops[1], Int64Ty));
10949 }
10950 case NEON::BI__builtin_neon_vshld_n_s64:
10951 case NEON::BI__builtin_neon_vshld_n_u64: {
10952 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1)));
10953 return Builder.CreateShl(
10954 Ops[0], ConstantInt::get(Int64Ty, Amt->getZExtValue()), "shld_n");
10955 }
10956 case NEON::BI__builtin_neon_vshrd_n_s64: {
10957 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1)));
10958 return Builder.CreateAShr(
10959 Ops[0], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63),
10960 Amt->getZExtValue())),
10961 "shrd_n");
10962 }
10963 case NEON::BI__builtin_neon_vshrd_n_u64: {
10964 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1)));
10965 uint64_t ShiftAmt = Amt->getZExtValue();
10966 // Right-shifting an unsigned value by its size yields 0.
10967 if (ShiftAmt == 64)
10968 return ConstantInt::get(Int64Ty, 0);
10969 return Builder.CreateLShr(Ops[0], ConstantInt::get(Int64Ty, ShiftAmt),
10970 "shrd_n");
10971 }
10972 case NEON::BI__builtin_neon_vsrad_n_s64: {
10973 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2)));
10974 Ops[1] = Builder.CreateAShr(
10975 Ops[1], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63),
10976 Amt->getZExtValue())),
10977 "shrd_n");
10978 return Builder.CreateAdd(Ops[0], Ops[1]);
10979 }
10980 case NEON::BI__builtin_neon_vsrad_n_u64: {
10981 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2)));
10982 uint64_t ShiftAmt = Amt->getZExtValue();
10983 // Right-shifting an unsigned value by its size yields 0.
10984 // As Op + 0 = Op, return Ops[0] directly.
10985 if (ShiftAmt == 64)
10986 return Ops[0];
10987 Ops[1] = Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, ShiftAmt),
10988 "shrd_n");
10989 return Builder.CreateAdd(Ops[0], Ops[1]);
10990 }
10991 case NEON::BI__builtin_neon_vqdmlalh_lane_s16:
10992 case NEON::BI__builtin_neon_vqdmlalh_laneq_s16:
10993 case NEON::BI__builtin_neon_vqdmlslh_lane_s16:
10994 case NEON::BI__builtin_neon_vqdmlslh_laneq_s16: {
10995 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)),
10996 "lane");
10997 SmallVector<Value *, 2> ProductOps;
10998 ProductOps.push_back(vectorWrapScalar16(Ops[1]));
10999 ProductOps.push_back(vectorWrapScalar16(Ops[2]));
11000 auto *VTy = llvm::FixedVectorType::get(Int32Ty, 4);
11001 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy),
11002 ProductOps, "vqdmlXl");
11003 Constant *CI = ConstantInt::get(SizeTy, 0);
11004 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0");
11005 Ops.pop_back();
11006
11007 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlalh_lane_s16 ||
11008 BuiltinID == NEON::BI__builtin_neon_vqdmlalh_laneq_s16)
11009 ? Intrinsic::aarch64_neon_sqadd
11010 : Intrinsic::aarch64_neon_sqsub;
11011 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int32Ty), Ops, "vqdmlXl");
11012 }
11013 case NEON::BI__builtin_neon_vqdmlals_s32:
11014 case NEON::BI__builtin_neon_vqdmlsls_s32: {
11015 SmallVector<Value *, 2> ProductOps;
11016 ProductOps.push_back(Ops[1]);
11017 ProductOps.push_back(EmitScalarExpr(E->getArg(2)));
11018 Ops[1] =
11019 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar),
11020 ProductOps, "vqdmlXl");
11021
11022 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlals_s32
11023 ? Intrinsic::aarch64_neon_sqadd
11024 : Intrinsic::aarch64_neon_sqsub;
11025 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int64Ty), Ops, "vqdmlXl");
11026 }
11027 case NEON::BI__builtin_neon_vqdmlals_lane_s32:
11028 case NEON::BI__builtin_neon_vqdmlals_laneq_s32:
11029 case NEON::BI__builtin_neon_vqdmlsls_lane_s32:
11030 case NEON::BI__builtin_neon_vqdmlsls_laneq_s32: {
11031 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)),
11032 "lane");
11033 SmallVector<Value *, 2> ProductOps;
11034 ProductOps.push_back(Ops[1]);
11035 ProductOps.push_back(Ops[2]);
11036 Ops[1] =
11037 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar),
11038 ProductOps, "vqdmlXl");
11039 Ops.pop_back();
11040
11041 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlals_lane_s32 ||
11042 BuiltinID == NEON::BI__builtin_neon_vqdmlals_laneq_s32)
11043 ? Intrinsic::aarch64_neon_sqadd
11044 : Intrinsic::aarch64_neon_sqsub;
11045 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int64Ty), Ops, "vqdmlXl");
11046 }
11047 case NEON::BI__builtin_neon_vget_lane_bf16:
11048 case NEON::BI__builtin_neon_vduph_lane_bf16:
11049 case NEON::BI__builtin_neon_vduph_lane_f16: {
11050 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
11051 "vget_lane");
11052 }
11053 case NEON::BI__builtin_neon_vgetq_lane_bf16:
11054 case NEON::BI__builtin_neon_vduph_laneq_bf16:
11055 case NEON::BI__builtin_neon_vduph_laneq_f16: {
11056 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)),
11057 "vgetq_lane");
11058 }
11059
11060 case clang::AArch64::BI_InterlockedAdd: {
11061 Value *Arg0 = EmitScalarExpr(E->getArg(0));
11062 Value *Arg1 = EmitScalarExpr(E->getArg(1));
11063 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW(
11064 AtomicRMWInst::Add, Arg0, Arg1,
11065 llvm::AtomicOrdering::SequentiallyConsistent);
11066 return Builder.CreateAdd(RMWI, Arg1);
11067 }
11068 }
11069
11070 llvm::FixedVectorType *VTy = GetNeonType(this, Type);
11071 llvm::Type *Ty = VTy;
11072 if (!Ty)
11073 return nullptr;
11074
11075 // Not all intrinsics handled by the common case work for AArch64 yet, so only
11076 // defer to common code if it's been added to our special map.
11077 Builtin = findARMVectorIntrinsicInMap(AArch64SIMDIntrinsicMap, BuiltinID,
11078 AArch64SIMDIntrinsicsProvenSorted);
11079
11080 if (Builtin)
11081 return EmitCommonNeonBuiltinExpr(
11082 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic,
11083 Builtin->NameHint, Builtin->TypeModifier, E, Ops,
11084 /*never use addresses*/ Address::invalid(), Address::invalid(), Arch);
11085
11086 if (Value *V = EmitAArch64TblBuiltinExpr(*this, BuiltinID, E, Ops, Arch))
11087 return V;
11088
11089 unsigned Int;
11090 switch (BuiltinID) {
11091 default: return nullptr;
11092 case NEON::BI__builtin_neon_vbsl_v:
11093 case NEON::BI__builtin_neon_vbslq_v: {
11094 llvm::Type *BitTy = llvm::VectorType::getInteger(VTy);
11095 Ops[0] = Builder.CreateBitCast(Ops[0], BitTy, "vbsl");
11096 Ops[1] = Builder.CreateBitCast(Ops[1], BitTy, "vbsl");
11097 Ops[2] = Builder.CreateBitCast(Ops[2], BitTy, "vbsl");
11098
11099 Ops[1] = Builder.CreateAnd(Ops[0], Ops[1], "vbsl");
11100 Ops[2] = Builder.CreateAnd(Builder.CreateNot(Ops[0]), Ops[2], "vbsl");
11101 Ops[0] = Builder.CreateOr(Ops[1], Ops[2], "vbsl");
11102 return Builder.CreateBitCast(Ops[0], Ty);
11103 }
11104 case NEON::BI__builtin_neon_vfma_lane_v:
11105 case NEON::BI__builtin_neon_vfmaq_lane_v: { // Only used for FP types
11106 // The ARM builtins (and instructions) have the addend as the first
11107 // operand, but the 'fma' intrinsics have it last. Swap it around here.
11108 Value *Addend = Ops[0];
11109 Value *Multiplicand = Ops[1];
11110 Value *LaneSource = Ops[2];
11111 Ops[0] = Multiplicand;
11112 Ops[1] = LaneSource;
11113 Ops[2] = Addend;
11114
11115 // Now adjust things to handle the lane access.
11116 auto *SourceTy = BuiltinID == NEON::BI__builtin_neon_vfmaq_lane_v
11117 ? llvm::FixedVectorType::get(VTy->getElementType(),
11118 VTy->getNumElements() / 2)
11119 : VTy;
11120 llvm::Constant *cst = cast<Constant>(Ops[3]);
11121 Value *SV = llvm::ConstantVector::getSplat(VTy->getElementCount(), cst);
11122 Ops[1] = Builder.CreateBitCast(Ops[1], SourceTy);
11123 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV, "lane");
11124
11125 Ops.pop_back();
11126 Int = Builder.getIsFPConstrained() ? Intrinsic::experimental_constrained_fma
11127 : Intrinsic::fma;
11128 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmla");
11129 }
11130 case NEON::BI__builtin_neon_vfma_laneq_v: {
11131 auto *VTy = cast<llvm::FixedVectorType>(Ty);
11132 // v1f64 fma should be mapped to Neon scalar f64 fma
11133 if (VTy && VTy->getElementType() == DoubleTy) {
11134 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy);
11135 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy);
11136 llvm::FixedVectorType *VTy =
11137 GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, true));
11138 Ops[2] = Builder.CreateBitCast(Ops[2], VTy);
11139 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract");
11140 Value *Result;
11141 Result = emitCallMaybeConstrainedFPBuiltin(
11142 *this, Intrinsic::fma, Intrinsic::experimental_constrained_fma,
11143 DoubleTy, {Ops[1], Ops[2], Ops[0]});
11144 return Builder.CreateBitCast(Result, Ty);
11145 }
11146 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
11147 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
11148
11149 auto *STy = llvm::FixedVectorType::get(VTy->getElementType(),
11150 VTy->getNumElements() * 2);
11151 Ops[2] = Builder.CreateBitCast(Ops[2], STy);
11152 Value *SV = llvm::ConstantVector::getSplat(VTy->getElementCount(),
11153 cast<ConstantInt>(Ops[3]));
11154 Ops[2] = Builder.CreateShuffleVector(Ops[2], Ops[2], SV, "lane");
11155
11156 return emitCallMaybeConstrainedFPBuiltin(
11157 *this, Intrinsic::fma, Intrinsic::experimental_constrained_fma, Ty,
11158 {Ops[2], Ops[1], Ops[0]});
11159 }
11160 case NEON::BI__builtin_neon_vfmaq_laneq_v: {
11161 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
11162 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
11163
11164 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
11165 Ops[2] = EmitNeonSplat(Ops[2], cast<ConstantInt>(Ops[3]));
11166 return emitCallMaybeConstrainedFPBuiltin(
11167 *this, Intrinsic::fma, Intrinsic::experimental_constrained_fma, Ty,
11168 {Ops[2], Ops[1], Ops[0]});
11169 }
11170 case NEON::BI__builtin_neon_vfmah_lane_f16:
11171 case NEON::BI__builtin_neon_vfmas_lane_f32:
11172 case NEON::BI__builtin_neon_vfmah_laneq_f16:
11173 case NEON::BI__builtin_neon_vfmas_laneq_f32:
11174 case NEON::BI__builtin_neon_vfmad_lane_f64:
11175 case NEON::BI__builtin_neon_vfmad_laneq_f64: {
11176 Ops.push_back(EmitScalarExpr(E->getArg(3)));
11177 llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext()));
11178 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract");
11179 return emitCallMaybeConstrainedFPBuiltin(
11180 *this, Intrinsic::fma, Intrinsic::experimental_constrained_fma, Ty,
11181 {Ops[1], Ops[2], Ops[0]});
11182 }
11183 case NEON::BI__builtin_neon_vmull_v:
11184 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics.
11185 Int = usgn ? Intrinsic::aarch64_neon_umull : Intrinsic::aarch64_neon_smull;
11186 if (Type.isPoly()) Int = Intrinsic::aarch64_neon_pmull;
11187 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull");
11188 case NEON::BI__builtin_neon_vmax_v:
11189 case NEON::BI__builtin_neon_vmaxq_v:
11190 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics.
11191 Int = usgn ? Intrinsic::aarch64_neon_umax : Intrinsic::aarch64_neon_smax;
11192 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmax;
11193 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax");
11194 case NEON::BI__builtin_neon_vmaxh_f16: {
11195 Ops.push_back(EmitScalarExpr(E->getArg(1)));
11196 Int = Intrinsic::aarch64_neon_fmax;
11197 return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vmax");
11198 }
11199 case NEON::BI__builtin_neon_vmin_v:
11200 case NEON::BI__builtin_neon_vminq_v:
11201 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics.
11202 Int = usgn ? Intrinsic::aarch64_neon_umin : Intrinsic::aarch64_neon_smin;
11203 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmin;
11204 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin");
11205 case NEON::BI__builtin_neon_vminh_f16: {
11206 Ops.push_back(EmitScalarExpr(E->getArg(1)));
11207 Int = Intrinsic::aarch64_neon_fmin;
11208 return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vmin");
11209 }
11210 case NEON::BI__builtin_neon_vabd_v:
11211 case NEON::BI__builtin_neon_vabdq_v:
11212 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics.
11213 Int = usgn ? Intrinsic::aarch64_neon_uabd : Intrinsic::aarch64_neon_sabd;
11214 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fabd;
11215 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd");
11216 case NEON::BI__builtin_neon_vpadal_v:
11217 case NEON::BI__builtin_neon_vpadalq_v: {
11218 unsigned ArgElts = VTy->getNumElements();
11219 llvm::IntegerType *EltTy = cast<IntegerType>(VTy->getElementType());
11220 unsigned BitWidth = EltTy->getBitWidth();
11221 auto *ArgTy = llvm::FixedVectorType::get(
11222 llvm::IntegerType::get(getLLVMContext(), BitWidth / 2), 2 * ArgElts);
11223 llvm::Type* Tys[2] = { VTy, ArgTy };
11224 Int = usgn ? Intrinsic::aarch64_neon_uaddlp : Intrinsic::aarch64_neon_saddlp;
11225 SmallVector<llvm::Value*, 1> TmpOps;
11226 TmpOps.push_back(Ops[1]);
11227 Function *F = CGM.getIntrinsic(Int, Tys);
11228 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vpadal");
11229 llvm::Value *addend = Builder.CreateBitCast(Ops[0], tmp->getType());
11230 return Builder.CreateAdd(tmp, addend);
11231 }
11232 case NEON::BI__builtin_neon_vpmin_v:
11233 case NEON::BI__builtin_neon_vpminq_v:
11234 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics.
11235 Int = usgn ? Intrinsic::aarch64_neon_uminp : Intrinsic::aarch64_neon_sminp;
11236 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fminp;
11237 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin");
11238 case NEON::BI__builtin_neon_vpmax_v:
11239 case NEON::BI__builtin_neon_vpmaxq_v:
11240 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics.
11241 Int = usgn ? Intrinsic::aarch64_neon_umaxp : Intrinsic::aarch64_neon_smaxp;
11242 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmaxp;
11243 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax");
11244 case NEON::BI__builtin_neon_vminnm_v:
11245 case NEON::BI__builtin_neon_vminnmq_v:
11246 Int = Intrinsic::aarch64_neon_fminnm;
11247 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vminnm");
11248 case NEON::BI__builtin_neon_vminnmh_f16:
11249 Ops.push_back(EmitScalarExpr(E->getArg(1)));
11250 Int = Intrinsic::aarch64_neon_fminnm;
11251 return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vminnm");
11252 case NEON::BI__builtin_neon_vmaxnm_v:
11253 case NEON::BI__builtin_neon_vmaxnmq_v:
11254 Int = Intrinsic::aarch64_neon_fmaxnm;
11255 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmaxnm");
11256 case NEON::BI__builtin_neon_vmaxnmh_f16:
11257 Ops.push_back(EmitScalarExpr(E->getArg(1)));
11258 Int = Intrinsic::aarch64_neon_fmaxnm;
11259 return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vmaxnm");
11260 case NEON::BI__builtin_neon_vrecpss_f32: {
11261 Ops.push_back(EmitScalarExpr(E->getArg(1)));
11262 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, FloatTy),
11263 Ops, "vrecps");
11264 }
11265 case NEON::BI__builtin_neon_vrecpsd_f64:
11266 Ops.push_back(EmitScalarExpr(E->getArg(1)));
11267 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, DoubleTy),
11268 Ops, "vrecps");
11269 case NEON::BI__builtin_neon_vrecpsh_f16:
11270 Ops.push_back(EmitScalarExpr(E->getArg(1)));
11271 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, HalfTy),
11272 Ops, "vrecps");
11273 case NEON::BI__builtin_neon_vqshrun_n_v:
11274 Int = Intrinsic::aarch64_neon_sqshrun;
11275 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrun_n");
11276 case NEON::BI__builtin_neon_vqrshrun_n_v:
11277 Int = Intrinsic::aarch64_neon_sqrshrun;
11278 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrun_n");
11279 case NEON::BI__builtin_neon_vqshrn_n_v:
11280 Int = usgn ? Intrinsic::aarch64_neon_uqshrn : Intrinsic::aarch64_neon_sqshrn;
11281 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n");
11282 case NEON::BI__builtin_neon_vrshrn_n_v:
11283 Int = Intrinsic::aarch64_neon_rshrn;
11284 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshrn_n");
11285 case NEON::BI__builtin_neon_vqrshrn_n_v:
11286 Int = usgn ? Intrinsic::aarch64_neon_uqrshrn : Intrinsic::aarch64_neon_sqrshrn;
11287 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n");
11288 case NEON::BI__builtin_neon_vrndah_f16: {
11289 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11290 Int = Builder.getIsFPConstrained()
11291 ? Intrinsic::experimental_constrained_round
11292 : Intrinsic::round;
11293 return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrnda");
11294 }
11295 case NEON::BI__builtin_neon_vrnda_v:
11296 case NEON::BI__builtin_neon_vrndaq_v: {
11297 Int = Builder.getIsFPConstrained()
11298 ? Intrinsic::experimental_constrained_round
11299 : Intrinsic::round;
11300 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnda");
11301 }
11302 case NEON::BI__builtin_neon_vrndih_f16: {
11303 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11304 Int = Builder.getIsFPConstrained()
11305 ? Intrinsic::experimental_constrained_nearbyint
11306 : Intrinsic::nearbyint;
11307 return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrndi");
11308 }
11309 case NEON::BI__builtin_neon_vrndmh_f16: {
11310 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11311 Int = Builder.getIsFPConstrained()
11312 ? Intrinsic::experimental_constrained_floor
11313 : Intrinsic::floor;
11314 return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrndm");
11315 }
11316 case NEON::BI__builtin_neon_vrndm_v:
11317 case NEON::BI__builtin_neon_vrndmq_v: {
11318 Int = Builder.getIsFPConstrained()
11319 ? Intrinsic::experimental_constrained_floor
11320 : Intrinsic::floor;
11321 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndm");
11322 }
11323 case NEON::BI__builtin_neon_vrndnh_f16: {
11324 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11325 Int = Builder.getIsFPConstrained()
11326 ? Intrinsic::experimental_constrained_roundeven
11327 : Intrinsic::roundeven;
11328 return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrndn");
11329 }
11330 case NEON::BI__builtin_neon_vrndn_v:
11331 case NEON::BI__builtin_neon_vrndnq_v: {
11332 Int = Builder.getIsFPConstrained()
11333 ? Intrinsic::experimental_constrained_roundeven
11334 : Intrinsic::roundeven;
11335 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndn");
11336 }
11337 case NEON::BI__builtin_neon_vrndns_f32: {
11338 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11339 Int = Builder.getIsFPConstrained()
11340 ? Intrinsic::experimental_constrained_roundeven
11341 : Intrinsic::roundeven;
11342 return EmitNeonCall(CGM.getIntrinsic(Int, FloatTy), Ops, "vrndn");
11343 }
11344 case NEON::BI__builtin_neon_vrndph_f16: {
11345 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11346 Int = Builder.getIsFPConstrained()
11347 ? Intrinsic::experimental_constrained_ceil
11348 : Intrinsic::ceil;
11349 return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrndp");
11350 }
11351 case NEON::BI__builtin_neon_vrndp_v:
11352 case NEON::BI__builtin_neon_vrndpq_v: {
11353 Int = Builder.getIsFPConstrained()
11354 ? Intrinsic::experimental_constrained_ceil
11355 : Intrinsic::ceil;
11356 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndp");
11357 }
11358 case NEON::BI__builtin_neon_vrndxh_f16: {
11359 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11360 Int = Builder.getIsFPConstrained()
11361 ? Intrinsic::experimental_constrained_rint
11362 : Intrinsic::rint;
11363 return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrndx");
11364 }
11365 case NEON::BI__builtin_neon_vrndx_v:
11366 case NEON::BI__builtin_neon_vrndxq_v: {
11367 Int = Builder.getIsFPConstrained()
11368 ? Intrinsic::experimental_constrained_rint
11369 : Intrinsic::rint;
11370 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndx");
11371 }
11372 case NEON::BI__builtin_neon_vrndh_f16: {
11373 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11374 Int = Builder.getIsFPConstrained()
11375 ? Intrinsic::experimental_constrained_trunc
11376 : Intrinsic::trunc;
11377 return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrndz");
11378 }
11379 case NEON::BI__builtin_neon_vrnd32x_f32:
11380 case NEON::BI__builtin_neon_vrnd32xq_f32: {
11381 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11382 Int = Intrinsic::aarch64_neon_frint32x;
11383 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnd32x");
11384 }
11385 case NEON::BI__builtin_neon_vrnd32z_f32:
11386 case NEON::BI__builtin_neon_vrnd32zq_f32: {
11387 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11388 Int = Intrinsic::aarch64_neon_frint32z;
11389 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnd32z");
11390 }
11391 case NEON::BI__builtin_neon_vrnd64x_f32:
11392 case NEON::BI__builtin_neon_vrnd64xq_f32: {
11393 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11394 Int = Intrinsic::aarch64_neon_frint64x;
11395 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnd64x");
11396 }
11397 case NEON::BI__builtin_neon_vrnd64z_f32:
11398 case NEON::BI__builtin_neon_vrnd64zq_f32: {
11399 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11400 Int = Intrinsic::aarch64_neon_frint64z;
11401 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnd64z");
11402 }
11403 case NEON::BI__builtin_neon_vrnd_v:
11404 case NEON::BI__builtin_neon_vrndq_v: {
11405 Int = Builder.getIsFPConstrained()
11406 ? Intrinsic::experimental_constrained_trunc
11407 : Intrinsic::trunc;
11408 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndz");
11409 }
11410 case NEON::BI__builtin_neon_vcvt_f64_v:
11411 case NEON::BI__builtin_neon_vcvtq_f64_v:
11412 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
11413 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, quad));
11414 return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt")
11415 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt");
11416 case NEON::BI__builtin_neon_vcvt_f64_f32: {
11417 assert(Type.getEltType() == NeonTypeFlags::Float64 && quad &&
11418 "unexpected vcvt_f64_f32 builtin");
11419 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float32, false, false);
11420 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag));
11421
11422 return Builder.CreateFPExt(Ops[0], Ty, "vcvt");
11423 }
11424 case NEON::BI__builtin_neon_vcvt_f32_f64: {
11425 assert(Type.getEltType() == NeonTypeFlags::Float32 &&
11426 "unexpected vcvt_f32_f64 builtin");
11427 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float64, false, true);
11428 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag));
11429
11430 return Builder.CreateFPTrunc(Ops[0], Ty, "vcvt");
11431 }
11432 case NEON::BI__builtin_neon_vcvt_s32_v:
11433 case NEON::BI__builtin_neon_vcvt_u32_v:
11434 case NEON::BI__builtin_neon_vcvt_s64_v:
11435 case NEON::BI__builtin_neon_vcvt_u64_v:
11436 case NEON::BI__builtin_neon_vcvt_s16_f16:
11437 case NEON::BI__builtin_neon_vcvt_u16_f16:
11438 case NEON::BI__builtin_neon_vcvtq_s32_v:
11439 case NEON::BI__builtin_neon_vcvtq_u32_v:
11440 case NEON::BI__builtin_neon_vcvtq_s64_v:
11441 case NEON::BI__builtin_neon_vcvtq_u64_v:
11442 case NEON::BI__builtin_neon_vcvtq_s16_f16:
11443 case NEON::BI__builtin_neon_vcvtq_u16_f16: {
11444 Int =
11445 usgn ? Intrinsic::aarch64_neon_fcvtzu : Intrinsic::aarch64_neon_fcvtzs;
11446 llvm::Type *Tys[2] = {Ty, GetFloatNeonType(this, Type)};
11447 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtz");
11448 }
11449 case NEON::BI__builtin_neon_vcvta_s16_f16:
11450 case NEON::BI__builtin_neon_vcvta_u16_f16:
11451 case NEON::BI__builtin_neon_vcvta_s32_v:
11452 case NEON::BI__builtin_neon_vcvtaq_s16_f16:
11453 case NEON::BI__builtin_neon_vcvtaq_s32_v:
11454 case NEON::BI__builtin_neon_vcvta_u32_v:
11455 case NEON::BI__builtin_neon_vcvtaq_u16_f16:
11456 case NEON::BI__builtin_neon_vcvtaq_u32_v:
11457 case NEON::BI__builtin_neon_vcvta_s64_v:
11458 case NEON::BI__builtin_neon_vcvtaq_s64_v:
11459 case NEON::BI__builtin_neon_vcvta_u64_v:
11460 case NEON::BI__builtin_neon_vcvtaq_u64_v: {
11461 Int = usgn ? Intrinsic::aarch64_neon_fcvtau : Intrinsic::aarch64_neon_fcvtas;
11462 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) };
11463 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvta");
11464 }
11465 case NEON::BI__builtin_neon_vcvtm_s16_f16:
11466 case NEON::BI__builtin_neon_vcvtm_s32_v:
11467 case NEON::BI__builtin_neon_vcvtmq_s16_f16:
11468 case NEON::BI__builtin_neon_vcvtmq_s32_v:
11469 case NEON::BI__builtin_neon_vcvtm_u16_f16:
11470 case NEON::BI__builtin_neon_vcvtm_u32_v:
11471 case NEON::BI__builtin_neon_vcvtmq_u16_f16:
11472 case NEON::BI__builtin_neon_vcvtmq_u32_v:
11473 case NEON::BI__builtin_neon_vcvtm_s64_v:
11474 case NEON::BI__builtin_neon_vcvtmq_s64_v:
11475 case NEON::BI__builtin_neon_vcvtm_u64_v:
11476 case NEON::BI__builtin_neon_vcvtmq_u64_v: {
11477 Int = usgn ? Intrinsic::aarch64_neon_fcvtmu : Intrinsic::aarch64_neon_fcvtms;
11478 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) };
11479 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtm");
11480 }
11481 case NEON::BI__builtin_neon_vcvtn_s16_f16:
11482 case NEON::BI__builtin_neon_vcvtn_s32_v:
11483 case NEON::BI__builtin_neon_vcvtnq_s16_f16:
11484 case NEON::BI__builtin_neon_vcvtnq_s32_v:
11485 case NEON::BI__builtin_neon_vcvtn_u16_f16:
11486 case NEON::BI__builtin_neon_vcvtn_u32_v:
11487 case NEON::BI__builtin_neon_vcvtnq_u16_f16:
11488 case NEON::BI__builtin_neon_vcvtnq_u32_v:
11489 case NEON::BI__builtin_neon_vcvtn_s64_v:
11490 case NEON::BI__builtin_neon_vcvtnq_s64_v:
11491 case NEON::BI__builtin_neon_vcvtn_u64_v:
11492 case NEON::BI__builtin_neon_vcvtnq_u64_v: {
11493 Int = usgn ? Intrinsic::aarch64_neon_fcvtnu : Intrinsic::aarch64_neon_fcvtns;
11494 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) };
11495 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtn");
11496 }
11497 case NEON::BI__builtin_neon_vcvtp_s16_f16:
11498 case NEON::BI__builtin_neon_vcvtp_s32_v:
11499 case NEON::BI__builtin_neon_vcvtpq_s16_f16:
11500 case NEON::BI__builtin_neon_vcvtpq_s32_v:
11501 case NEON::BI__builtin_neon_vcvtp_u16_f16:
11502 case NEON::BI__builtin_neon_vcvtp_u32_v:
11503 case NEON::BI__builtin_neon_vcvtpq_u16_f16:
11504 case NEON::BI__builtin_neon_vcvtpq_u32_v:
11505 case NEON::BI__builtin_neon_vcvtp_s64_v:
11506 case NEON::BI__builtin_neon_vcvtpq_s64_v:
11507 case NEON::BI__builtin_neon_vcvtp_u64_v:
11508 case NEON::BI__builtin_neon_vcvtpq_u64_v: {
11509 Int = usgn ? Intrinsic::aarch64_neon_fcvtpu : Intrinsic::aarch64_neon_fcvtps;
11510 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) };
11511 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtp");
11512 }
11513 case NEON::BI__builtin_neon_vmulx_v:
11514 case NEON::BI__builtin_neon_vmulxq_v: {
11515 Int = Intrinsic::aarch64_neon_fmulx;
11516 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmulx");
11517 }
11518 case NEON::BI__builtin_neon_vmulxh_lane_f16:
11519 case NEON::BI__builtin_neon_vmulxh_laneq_f16: {
11520 // vmulx_lane should be mapped to Neon scalar mulx after
11521 // extracting the scalar element
11522 Ops.push_back(EmitScalarExpr(E->getArg(2)));
11523 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract");
11524 Ops.pop_back();
11525 Int = Intrinsic::aarch64_neon_fmulx;
11526 return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vmulx");
11527 }
11528 case NEON::BI__builtin_neon_vmul_lane_v:
11529 case NEON::BI__builtin_neon_vmul_laneq_v: {
11530 // v1f64 vmul_lane should be mapped to Neon scalar mul lane
11531 bool Quad = false;
11532 if (BuiltinID == NEON::BI__builtin_neon_vmul_laneq_v)
11533 Quad = true;
11534 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy);
11535 llvm::FixedVectorType *VTy =
11536 GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, Quad));
11537 Ops[1] = Builder.CreateBitCast(Ops[1], VTy);
11538 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract");
11539 Value *Result = Builder.CreateFMul(Ops[0], Ops[1]);
11540 return Builder.CreateBitCast(Result, Ty);
11541 }
11542 case NEON::BI__builtin_neon_vnegd_s64:
11543 return Builder.CreateNeg(EmitScalarExpr(E->getArg(0)), "vnegd");
11544 case NEON::BI__builtin_neon_vnegh_f16:
11545 return Builder.CreateFNeg(EmitScalarExpr(E->getArg(0)), "vnegh");
11546 case NEON::BI__builtin_neon_vpmaxnm_v:
11547 case NEON::BI__builtin_neon_vpmaxnmq_v: {
11548 Int = Intrinsic::aarch64_neon_fmaxnmp;
11549 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmaxnm");
11550 }
11551 case NEON::BI__builtin_neon_vpminnm_v:
11552 case NEON::BI__builtin_neon_vpminnmq_v: {
11553 Int = Intrinsic::aarch64_neon_fminnmp;
11554 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpminnm");
11555 }
11556 case NEON::BI__builtin_neon_vsqrth_f16: {
11557 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11558 Int = Builder.getIsFPConstrained()
11559 ? Intrinsic::experimental_constrained_sqrt
11560 : Intrinsic::sqrt;
11561 return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vsqrt");
11562 }
11563 case NEON::BI__builtin_neon_vsqrt_v:
11564 case NEON::BI__builtin_neon_vsqrtq_v: {
11565 Int = Builder.getIsFPConstrained()
11566 ? Intrinsic::experimental_constrained_sqrt
11567 : Intrinsic::sqrt;
11568 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
11569 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqrt");
11570 }
11571 case NEON::BI__builtin_neon_vrbit_v:
11572 case NEON::BI__builtin_neon_vrbitq_v: {
11573 Int = Intrinsic::bitreverse;
11574 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrbit");
11575 }
11576 case NEON::BI__builtin_neon_vaddv_u8:
11577 // FIXME: These are handled by the AArch64 scalar code.
11578 usgn = true;
11579 [[fallthrough]];
11580 case NEON::BI__builtin_neon_vaddv_s8: {
11581 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv;
11582 Ty = Int32Ty;
11583 VTy = llvm::FixedVectorType::get(Int8Ty, 8);
11584 llvm::Type *Tys[2] = { Ty, VTy };
11585 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11586 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv");
11587 return Builder.CreateTrunc(Ops[0], Int8Ty);
11588 }
11589 case NEON::BI__builtin_neon_vaddv_u16:
11590 usgn = true;
11591 [[fallthrough]];
11592 case NEON::BI__builtin_neon_vaddv_s16: {
11593 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv;
11594 Ty = Int32Ty;
11595 VTy = llvm::FixedVectorType::get(Int16Ty, 4);
11596 llvm::Type *Tys[2] = { Ty, VTy };
11597 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11598 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv");
11599 return Builder.CreateTrunc(Ops[0], Int16Ty);
11600 }
11601 case NEON::BI__builtin_neon_vaddvq_u8:
11602 usgn = true;
11603 [[fallthrough]];
11604 case NEON::BI__builtin_neon_vaddvq_s8: {
11605 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv;
11606 Ty = Int32Ty;
11607 VTy = llvm::FixedVectorType::get(Int8Ty, 16);
11608 llvm::Type *Tys[2] = { Ty, VTy };
11609 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11610 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv");
11611 return Builder.CreateTrunc(Ops[0], Int8Ty);
11612 }
11613 case NEON::BI__builtin_neon_vaddvq_u16:
11614 usgn = true;
11615 [[fallthrough]];
11616 case NEON::BI__builtin_neon_vaddvq_s16: {
11617 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv;
11618 Ty = Int32Ty;
11619 VTy = llvm::FixedVectorType::get(Int16Ty, 8);
11620 llvm::Type *Tys[2] = { Ty, VTy };
11621 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11622 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv");
11623 return Builder.CreateTrunc(Ops[0], Int16Ty);
11624 }
11625 case NEON::BI__builtin_neon_vmaxv_u8: {
11626 Int = Intrinsic::aarch64_neon_umaxv;
11627 Ty = Int32Ty;
11628 VTy = llvm::FixedVectorType::get(Int8Ty, 8);
11629 llvm::Type *Tys[2] = { Ty, VTy };
11630 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11631 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
11632 return Builder.CreateTrunc(Ops[0], Int8Ty);
11633 }
11634 case NEON::BI__builtin_neon_vmaxv_u16: {
11635 Int = Intrinsic::aarch64_neon_umaxv;
11636 Ty = Int32Ty;
11637 VTy = llvm::FixedVectorType::get(Int16Ty, 4);
11638 llvm::Type *Tys[2] = { Ty, VTy };
11639 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11640 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
11641 return Builder.CreateTrunc(Ops[0], Int16Ty);
11642 }
11643 case NEON::BI__builtin_neon_vmaxvq_u8: {
11644 Int = Intrinsic::aarch64_neon_umaxv;
11645 Ty = Int32Ty;
11646 VTy = llvm::FixedVectorType::get(Int8Ty, 16);
11647 llvm::Type *Tys[2] = { Ty, VTy };
11648 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11649 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
11650 return Builder.CreateTrunc(Ops[0], Int8Ty);
11651 }
11652 case NEON::BI__builtin_neon_vmaxvq_u16: {
11653 Int = Intrinsic::aarch64_neon_umaxv;
11654 Ty = Int32Ty;
11655 VTy = llvm::FixedVectorType::get(Int16Ty, 8);
11656 llvm::Type *Tys[2] = { Ty, VTy };
11657 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11658 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
11659 return Builder.CreateTrunc(Ops[0], Int16Ty);
11660 }
11661 case NEON::BI__builtin_neon_vmaxv_s8: {
11662 Int = Intrinsic::aarch64_neon_smaxv;
11663 Ty = Int32Ty;
11664 VTy = llvm::FixedVectorType::get(Int8Ty, 8);
11665 llvm::Type *Tys[2] = { Ty, VTy };
11666 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11667 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
11668 return Builder.CreateTrunc(Ops[0], Int8Ty);
11669 }
11670 case NEON::BI__builtin_neon_vmaxv_s16: {
11671 Int = Intrinsic::aarch64_neon_smaxv;
11672 Ty = Int32Ty;
11673 VTy = llvm::FixedVectorType::get(Int16Ty, 4);
11674 llvm::Type *Tys[2] = { Ty, VTy };
11675 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11676 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
11677 return Builder.CreateTrunc(Ops[0], Int16Ty);
11678 }
11679 case NEON::BI__builtin_neon_vmaxvq_s8: {
11680 Int = Intrinsic::aarch64_neon_smaxv;
11681 Ty = Int32Ty;
11682 VTy = llvm::FixedVectorType::get(Int8Ty, 16);
11683 llvm::Type *Tys[2] = { Ty, VTy };
11684 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11685 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
11686 return Builder.CreateTrunc(Ops[0], Int8Ty);
11687 }
11688 case NEON::BI__builtin_neon_vmaxvq_s16: {
11689 Int = Intrinsic::aarch64_neon_smaxv;
11690 Ty = Int32Ty;
11691 VTy = llvm::FixedVectorType::get(Int16Ty, 8);
11692 llvm::Type *Tys[2] = { Ty, VTy };
11693 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11694 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
11695 return Builder.CreateTrunc(Ops[0], Int16Ty);
11696 }
11697 case NEON::BI__builtin_neon_vmaxv_f16: {
11698 Int = Intrinsic::aarch64_neon_fmaxv;
11699 Ty = HalfTy;
11700 VTy = llvm::FixedVectorType::get(HalfTy, 4);
11701 llvm::Type *Tys[2] = { Ty, VTy };
11702 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11703 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
11704 return Builder.CreateTrunc(Ops[0], HalfTy);
11705 }
11706 case NEON::BI__builtin_neon_vmaxvq_f16: {
11707 Int = Intrinsic::aarch64_neon_fmaxv;
11708 Ty = HalfTy;
11709 VTy = llvm::FixedVectorType::get(HalfTy, 8);
11710 llvm::Type *Tys[2] = { Ty, VTy };
11711 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11712 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv");
11713 return Builder.CreateTrunc(Ops[0], HalfTy);
11714 }
11715 case NEON::BI__builtin_neon_vminv_u8: {
11716 Int = Intrinsic::aarch64_neon_uminv;
11717 Ty = Int32Ty;
11718 VTy = llvm::FixedVectorType::get(Int8Ty, 8);
11719 llvm::Type *Tys[2] = { Ty, VTy };
11720 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11721 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
11722 return Builder.CreateTrunc(Ops[0], Int8Ty);
11723 }
11724 case NEON::BI__builtin_neon_vminv_u16: {
11725 Int = Intrinsic::aarch64_neon_uminv;
11726 Ty = Int32Ty;
11727 VTy = llvm::FixedVectorType::get(Int16Ty, 4);
11728 llvm::Type *Tys[2] = { Ty, VTy };
11729 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11730 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
11731 return Builder.CreateTrunc(Ops[0], Int16Ty);
11732 }
11733 case NEON::BI__builtin_neon_vminvq_u8: {
11734 Int = Intrinsic::aarch64_neon_uminv;
11735 Ty = Int32Ty;
11736 VTy = llvm::FixedVectorType::get(Int8Ty, 16);
11737 llvm::Type *Tys[2] = { Ty, VTy };
11738 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11739 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
11740 return Builder.CreateTrunc(Ops[0], Int8Ty);
11741 }
11742 case NEON::BI__builtin_neon_vminvq_u16: {
11743 Int = Intrinsic::aarch64_neon_uminv;
11744 Ty = Int32Ty;
11745 VTy = llvm::FixedVectorType::get(Int16Ty, 8);
11746 llvm::Type *Tys[2] = { Ty, VTy };
11747 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11748 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
11749 return Builder.CreateTrunc(Ops[0], Int16Ty);
11750 }
11751 case NEON::BI__builtin_neon_vminv_s8: {
11752 Int = Intrinsic::aarch64_neon_sminv;
11753 Ty = Int32Ty;
11754 VTy = llvm::FixedVectorType::get(Int8Ty, 8);
11755 llvm::Type *Tys[2] = { Ty, VTy };
11756 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11757 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
11758 return Builder.CreateTrunc(Ops[0], Int8Ty);
11759 }
11760 case NEON::BI__builtin_neon_vminv_s16: {
11761 Int = Intrinsic::aarch64_neon_sminv;
11762 Ty = Int32Ty;
11763 VTy = llvm::FixedVectorType::get(Int16Ty, 4);
11764 llvm::Type *Tys[2] = { Ty, VTy };
11765 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11766 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
11767 return Builder.CreateTrunc(Ops[0], Int16Ty);
11768 }
11769 case NEON::BI__builtin_neon_vminvq_s8: {
11770 Int = Intrinsic::aarch64_neon_sminv;
11771 Ty = Int32Ty;
11772 VTy = llvm::FixedVectorType::get(Int8Ty, 16);
11773 llvm::Type *Tys[2] = { Ty, VTy };
11774 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11775 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
11776 return Builder.CreateTrunc(Ops[0], Int8Ty);
11777 }
11778 case NEON::BI__builtin_neon_vminvq_s16: {
11779 Int = Intrinsic::aarch64_neon_sminv;
11780 Ty = Int32Ty;
11781 VTy = llvm::FixedVectorType::get(Int16Ty, 8);
11782 llvm::Type *Tys[2] = { Ty, VTy };
11783 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11784 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
11785 return Builder.CreateTrunc(Ops[0], Int16Ty);
11786 }
11787 case NEON::BI__builtin_neon_vminv_f16: {
11788 Int = Intrinsic::aarch64_neon_fminv;
11789 Ty = HalfTy;
11790 VTy = llvm::FixedVectorType::get(HalfTy, 4);
11791 llvm::Type *Tys[2] = { Ty, VTy };
11792 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11793 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
11794 return Builder.CreateTrunc(Ops[0], HalfTy);
11795 }
11796 case NEON::BI__builtin_neon_vminvq_f16: {
11797 Int = Intrinsic::aarch64_neon_fminv;
11798 Ty = HalfTy;
11799 VTy = llvm::FixedVectorType::get(HalfTy, 8);
11800 llvm::Type *Tys[2] = { Ty, VTy };
11801 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11802 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv");
11803 return Builder.CreateTrunc(Ops[0], HalfTy);
11804 }
11805 case NEON::BI__builtin_neon_vmaxnmv_f16: {
11806 Int = Intrinsic::aarch64_neon_fmaxnmv;
11807 Ty = HalfTy;
11808 VTy = llvm::FixedVectorType::get(HalfTy, 4);
11809 llvm::Type *Tys[2] = { Ty, VTy };
11810 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11811 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxnmv");
11812 return Builder.CreateTrunc(Ops[0], HalfTy);
11813 }
11814 case NEON::BI__builtin_neon_vmaxnmvq_f16: {
11815 Int = Intrinsic::aarch64_neon_fmaxnmv;
11816 Ty = HalfTy;
11817 VTy = llvm::FixedVectorType::get(HalfTy, 8);
11818 llvm::Type *Tys[2] = { Ty, VTy };
11819 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11820 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxnmv");
11821 return Builder.CreateTrunc(Ops[0], HalfTy);
11822 }
11823 case NEON::BI__builtin_neon_vminnmv_f16: {
11824 Int = Intrinsic::aarch64_neon_fminnmv;
11825 Ty = HalfTy;
11826 VTy = llvm::FixedVectorType::get(HalfTy, 4);
11827 llvm::Type *Tys[2] = { Ty, VTy };
11828 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11829 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminnmv");
11830 return Builder.CreateTrunc(Ops[0], HalfTy);
11831 }
11832 case NEON::BI__builtin_neon_vminnmvq_f16: {
11833 Int = Intrinsic::aarch64_neon_fminnmv;
11834 Ty = HalfTy;
11835 VTy = llvm::FixedVectorType::get(HalfTy, 8);
11836 llvm::Type *Tys[2] = { Ty, VTy };
11837 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11838 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminnmv");
11839 return Builder.CreateTrunc(Ops[0], HalfTy);
11840 }
11841 case NEON::BI__builtin_neon_vmul_n_f64: {
11842 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy);
11843 Value *RHS = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), DoubleTy);
11844 return Builder.CreateFMul(Ops[0], RHS);
11845 }
11846 case NEON::BI__builtin_neon_vaddlv_u8: {
11847 Int = Intrinsic::aarch64_neon_uaddlv;
11848 Ty = Int32Ty;
11849 VTy = llvm::FixedVectorType::get(Int8Ty, 8);
11850 llvm::Type *Tys[2] = { Ty, VTy };
11851 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11852 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
11853 return Builder.CreateTrunc(Ops[0], Int16Ty);
11854 }
11855 case NEON::BI__builtin_neon_vaddlv_u16: {
11856 Int = Intrinsic::aarch64_neon_uaddlv;
11857 Ty = Int32Ty;
11858 VTy = llvm::FixedVectorType::get(Int16Ty, 4);
11859 llvm::Type *Tys[2] = { Ty, VTy };
11860 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11861 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
11862 }
11863 case NEON::BI__builtin_neon_vaddlvq_u8: {
11864 Int = Intrinsic::aarch64_neon_uaddlv;
11865 Ty = Int32Ty;
11866 VTy = llvm::FixedVectorType::get(Int8Ty, 16);
11867 llvm::Type *Tys[2] = { Ty, VTy };
11868 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11869 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
11870 return Builder.CreateTrunc(Ops[0], Int16Ty);
11871 }
11872 case NEON::BI__builtin_neon_vaddlvq_u16: {
11873 Int = Intrinsic::aarch64_neon_uaddlv;
11874 Ty = Int32Ty;
11875 VTy = llvm::FixedVectorType::get(Int16Ty, 8);
11876 llvm::Type *Tys[2] = { Ty, VTy };
11877 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11878 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
11879 }
11880 case NEON::BI__builtin_neon_vaddlv_s8: {
11881 Int = Intrinsic::aarch64_neon_saddlv;
11882 Ty = Int32Ty;
11883 VTy = llvm::FixedVectorType::get(Int8Ty, 8);
11884 llvm::Type *Tys[2] = { Ty, VTy };
11885 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11886 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
11887 return Builder.CreateTrunc(Ops[0], Int16Ty);
11888 }
11889 case NEON::BI__builtin_neon_vaddlv_s16: {
11890 Int = Intrinsic::aarch64_neon_saddlv;
11891 Ty = Int32Ty;
11892 VTy = llvm::FixedVectorType::get(Int16Ty, 4);
11893 llvm::Type *Tys[2] = { Ty, VTy };
11894 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11895 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
11896 }
11897 case NEON::BI__builtin_neon_vaddlvq_s8: {
11898 Int = Intrinsic::aarch64_neon_saddlv;
11899 Ty = Int32Ty;
11900 VTy = llvm::FixedVectorType::get(Int8Ty, 16);
11901 llvm::Type *Tys[2] = { Ty, VTy };
11902 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11903 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
11904 return Builder.CreateTrunc(Ops[0], Int16Ty);
11905 }
11906 case NEON::BI__builtin_neon_vaddlvq_s16: {
11907 Int = Intrinsic::aarch64_neon_saddlv;
11908 Ty = Int32Ty;
11909 VTy = llvm::FixedVectorType::get(Int16Ty, 8);
11910 llvm::Type *Tys[2] = { Ty, VTy };
11911 Ops.push_back(EmitScalarExpr(E->getArg(0)));
11912 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv");
11913 }
11914 case NEON::BI__builtin_neon_vsri_n_v:
11915 case NEON::BI__builtin_neon_vsriq_n_v: {
11916 Int = Intrinsic::aarch64_neon_vsri;
11917 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty);
11918 return EmitNeonCall(Intrin, Ops, "vsri_n");
11919 }
11920 case NEON::BI__builtin_neon_vsli_n_v:
11921 case NEON::BI__builtin_neon_vsliq_n_v: {
11922 Int = Intrinsic::aarch64_neon_vsli;
11923 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty);
11924 return EmitNeonCall(Intrin, Ops, "vsli_n");
11925 }
11926 case NEON::BI__builtin_neon_vsra_n_v:
11927 case NEON::BI__builtin_neon_vsraq_n_v:
11928 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
11929 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n");
11930 return Builder.CreateAdd(Ops[0], Ops[1]);
11931 case NEON::BI__builtin_neon_vrsra_n_v:
11932 case NEON::BI__builtin_neon_vrsraq_n_v: {
11933 Int = usgn ? Intrinsic::aarch64_neon_urshl : Intrinsic::aarch64_neon_srshl;
11934 SmallVector<llvm::Value*,2> TmpOps;
11935 TmpOps.push_back(Ops[1]);
11936 TmpOps.push_back(Ops[2]);
11937 Function* F = CGM.getIntrinsic(Int, Ty);
11938 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vrshr_n", 1, true);
11939 Ops[0] = Builder.CreateBitCast(Ops[0], VTy);
11940 return Builder.CreateAdd(Ops[0], tmp);
11941 }
11942 case NEON::BI__builtin_neon_vld1_v:
11943 case NEON::BI__builtin_neon_vld1q_v: {
11944 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy));
11945 return Builder.CreateAlignedLoad(VTy, Ops[0], PtrOp0.getAlignment());
11946 }
11947 case NEON::BI__builtin_neon_vst1_v:
11948 case NEON::BI__builtin_neon_vst1q_v:
11949 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy));
11950 Ops[1] = Builder.CreateBitCast(Ops[1], VTy);
11951 return Builder.CreateAlignedStore(Ops[1], Ops[0], PtrOp0.getAlignment());
11952 case NEON::BI__builtin_neon_vld1_lane_v:
11953 case NEON::BI__builtin_neon_vld1q_lane_v: {
11954 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
11955 Ty = llvm::PointerType::getUnqual(VTy->getElementType());
11956 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
11957 Ops[0] = Builder.CreateAlignedLoad(VTy->getElementType(), Ops[0],
11958 PtrOp0.getAlignment());
11959 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vld1_lane");
11960 }
11961 case NEON::BI__builtin_neon_vld1_dup_v:
11962 case NEON::BI__builtin_neon_vld1q_dup_v: {
11963 Value *V = PoisonValue::get(Ty);
11964 Ty = llvm::PointerType::getUnqual(VTy->getElementType());
11965 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
11966 Ops[0] = Builder.CreateAlignedLoad(VTy->getElementType(), Ops[0],
11967 PtrOp0.getAlignment());
11968 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0);
11969 Ops[0] = Builder.CreateInsertElement(V, Ops[0], CI);
11970 return EmitNeonSplat(Ops[0], CI);
11971 }
11972 case NEON::BI__builtin_neon_vst1_lane_v:
11973 case NEON::BI__builtin_neon_vst1q_lane_v:
11974 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
11975 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]);
11976 Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
11977 return Builder.CreateAlignedStore(Ops[1], Builder.CreateBitCast(Ops[0], Ty),
11978 PtrOp0.getAlignment());
11979 case NEON::BI__builtin_neon_vld2_v:
11980 case NEON::BI__builtin_neon_vld2q_v: {
11981 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy);
11982 Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
11983 llvm::Type *Tys[2] = { VTy, PTy };
11984 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2, Tys);
11985 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2");
11986 Ops[0] = Builder.CreateBitCast(Ops[0],
11987 llvm::PointerType::getUnqual(Ops[1]->getType()));
11988 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
11989 }
11990 case NEON::BI__builtin_neon_vld3_v:
11991 case NEON::BI__builtin_neon_vld3q_v: {
11992 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy);
11993 Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
11994 llvm::Type *Tys[2] = { VTy, PTy };
11995 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3, Tys);
11996 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3");
11997 Ops[0] = Builder.CreateBitCast(Ops[0],
11998 llvm::PointerType::getUnqual(Ops[1]->getType()));
11999 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
12000 }
12001 case NEON::BI__builtin_neon_vld4_v:
12002 case NEON::BI__builtin_neon_vld4q_v: {
12003 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy);
12004 Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
12005 llvm::Type *Tys[2] = { VTy, PTy };
12006 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4, Tys);
12007 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4");
12008 Ops[0] = Builder.CreateBitCast(Ops[0],
12009 llvm::PointerType::getUnqual(Ops[1]->getType()));
12010 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
12011 }
12012 case NEON::BI__builtin_neon_vld2_dup_v:
12013 case NEON::BI__builtin_neon_vld2q_dup_v: {
12014 llvm::Type *PTy =
12015 llvm::PointerType::getUnqual(VTy->getElementType());
12016 Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
12017 llvm::Type *Tys[2] = { VTy, PTy };
12018 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2r, Tys);
12019 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2");
12020 Ops[0] = Builder.CreateBitCast(Ops[0],
12021 llvm::PointerType::getUnqual(Ops[1]->getType()));
12022 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
12023 }
12024 case NEON::BI__builtin_neon_vld3_dup_v:
12025 case NEON::BI__builtin_neon_vld3q_dup_v: {
12026 llvm::Type *PTy =
12027 llvm::PointerType::getUnqual(VTy->getElementType());
12028 Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
12029 llvm::Type *Tys[2] = { VTy, PTy };
12030 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3r, Tys);
12031 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3");
12032 Ops[0] = Builder.CreateBitCast(Ops[0],
12033 llvm::PointerType::getUnqual(Ops[1]->getType()));
12034 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
12035 }
12036 case NEON::BI__builtin_neon_vld4_dup_v:
12037 case NEON::BI__builtin_neon_vld4q_dup_v: {
12038 llvm::Type *PTy =
12039 llvm::PointerType::getUnqual(VTy->getElementType());
12040 Ops[1] = Builder.CreateBitCast(Ops[1], PTy);
12041 llvm::Type *Tys[2] = { VTy, PTy };
12042 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4r, Tys);
12043 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4");
12044 Ops[0] = Builder.CreateBitCast(Ops[0],
12045 llvm::PointerType::getUnqual(Ops[1]->getType()));
12046 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
12047 }
12048 case NEON::BI__builtin_neon_vld2_lane_v:
12049 case NEON::BI__builtin_neon_vld2q_lane_v: {
12050 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() };
12051 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2lane, Tys);
12052 std::rotate(Ops.begin() + 1, Ops.begin() + 2, Ops.end());
12053 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
12054 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
12055 Ops[3] = Builder.CreateZExt(Ops[3], Int64Ty);
12056 Ops[1] = Builder.CreateCall(F, ArrayRef(Ops).slice(1), "vld2_lane");
12057 Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
12058 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
12059 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
12060 }
12061 case NEON::BI__builtin_neon_vld3_lane_v:
12062 case NEON::BI__builtin_neon_vld3q_lane_v: {
12063 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() };
12064 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3lane, Tys);
12065 std::rotate(Ops.begin() + 1, Ops.begin() + 2, Ops.end());
12066 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
12067 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
12068 Ops[3] = Builder.CreateBitCast(Ops[3], Ty);
12069 Ops[4] = Builder.CreateZExt(Ops[4], Int64Ty);
12070 Ops[1] = Builder.CreateCall(F, ArrayRef(Ops).slice(1), "vld3_lane");
12071 Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
12072 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
12073 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
12074 }
12075 case NEON::BI__builtin_neon_vld4_lane_v:
12076 case NEON::BI__builtin_neon_vld4q_lane_v: {
12077 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() };
12078 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4lane, Tys);
12079 std::rotate(Ops.begin() + 1, Ops.begin() + 2, Ops.end());
12080 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
12081 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
12082 Ops[3] = Builder.CreateBitCast(Ops[3], Ty);
12083 Ops[4] = Builder.CreateBitCast(Ops[4], Ty);
12084 Ops[5] = Builder.CreateZExt(Ops[5], Int64Ty);
12085 Ops[1] = Builder.CreateCall(F, ArrayRef(Ops).slice(1), "vld4_lane");
12086 Ty = llvm::PointerType::getUnqual(Ops[1]->getType());
12087 Ops[0] = Builder.CreateBitCast(Ops[0], Ty);
12088 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]);
12089 }
12090 case NEON::BI__builtin_neon_vst2_v:
12091 case NEON::BI__builtin_neon_vst2q_v: {
12092 std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end());
12093 llvm::Type *Tys[2] = { VTy, Ops[2]->getType() };
12094 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2, Tys),
12095 Ops, "");
12096 }
12097 case NEON::BI__builtin_neon_vst2_lane_v:
12098 case NEON::BI__builtin_neon_vst2q_lane_v: {
12099 std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end());
12100 Ops[2] = Builder.CreateZExt(Ops[2], Int64Ty);
12101 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() };
12102 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2lane, Tys),
12103 Ops, "");
12104 }
12105 case NEON::BI__builtin_neon_vst3_v:
12106 case NEON::BI__builtin_neon_vst3q_v: {
12107 std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end());
12108 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() };
12109 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3, Tys),
12110 Ops, "");
12111 }
12112 case NEON::BI__builtin_neon_vst3_lane_v:
12113 case NEON::BI__builtin_neon_vst3q_lane_v: {
12114 std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end());
12115 Ops[3] = Builder.CreateZExt(Ops[3], Int64Ty);
12116 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() };
12117 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3lane, Tys),
12118 Ops, "");
12119 }
12120 case NEON::BI__builtin_neon_vst4_v:
12121 case NEON::BI__builtin_neon_vst4q_v: {
12122 std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end());
12123 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() };
12124 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4, Tys),
12125 Ops, "");
12126 }
12127 case NEON::BI__builtin_neon_vst4_lane_v:
12128 case NEON::BI__builtin_neon_vst4q_lane_v: {
12129 std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end());
12130 Ops[4] = Builder.CreateZExt(Ops[4], Int64Ty);
12131 llvm::Type *Tys[2] = { VTy, Ops[5]->getType() };
12132 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4lane, Tys),
12133 Ops, "");
12134 }
12135 case NEON::BI__builtin_neon_vtrn_v:
12136 case NEON::BI__builtin_neon_vtrnq_v: {
12137 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
12138 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
12139 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
12140 Value *SV = nullptr;
12141
12142 for (unsigned vi = 0; vi != 2; ++vi) {
12143 SmallVector<int, 16> Indices;
12144 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) {
12145 Indices.push_back(i+vi);
12146 Indices.push_back(i+e+vi);
12147 }
12148 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi);
12149 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vtrn");
12150 SV = Builder.CreateDefaultAlignedStore(SV, Addr);
12151 }
12152 return SV;
12153 }
12154 case NEON::BI__builtin_neon_vuzp_v:
12155 case NEON::BI__builtin_neon_vuzpq_v: {
12156 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
12157 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
12158 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
12159 Value *SV = nullptr;
12160
12161 for (unsigned vi = 0; vi != 2; ++vi) {
12162 SmallVector<int, 16> Indices;
12163 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i)
12164 Indices.push_back(2*i+vi);
12165
12166 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi);
12167 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vuzp");
12168 SV = Builder.CreateDefaultAlignedStore(SV, Addr);
12169 }
12170 return SV;
12171 }
12172 case NEON::BI__builtin_neon_vzip_v:
12173 case NEON::BI__builtin_neon_vzipq_v: {
12174 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
12175 Ops[1] = Builder.CreateBitCast(Ops[1], Ty);
12176 Ops[2] = Builder.CreateBitCast(Ops[2], Ty);
12177 Value *SV = nullptr;
12178
12179 for (unsigned vi = 0; vi != 2; ++vi) {
12180 SmallVector<int, 16> Indices;
12181 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) {
12182 Indices.push_back((i + vi*e) >> 1);
12183 Indices.push_back(((i + vi*e) >> 1)+e);
12184 }
12185 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi);
12186 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vzip");
12187 SV = Builder.CreateDefaultAlignedStore(SV, Addr);
12188 }
12189 return SV;
12190 }
12191 case NEON::BI__builtin_neon_vqtbl1q_v: {
12192 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl1, Ty),
12193 Ops, "vtbl1");
12194 }
12195 case NEON::BI__builtin_neon_vqtbl2q_v: {
12196 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl2, Ty),
12197 Ops, "vtbl2");
12198 }
12199 case NEON::BI__builtin_neon_vqtbl3q_v: {
12200 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl3, Ty),
12201 Ops, "vtbl3");
12202 }
12203 case NEON::BI__builtin_neon_vqtbl4q_v: {
12204 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl4, Ty),
12205 Ops, "vtbl4");
12206 }
12207 case NEON::BI__builtin_neon_vqtbx1q_v: {
12208 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx1, Ty),
12209 Ops, "vtbx1");
12210 }
12211 case NEON::BI__builtin_neon_vqtbx2q_v: {
12212 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx2, Ty),
12213 Ops, "vtbx2");
12214 }
12215 case NEON::BI__builtin_neon_vqtbx3q_v: {
12216 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx3, Ty),
12217 Ops, "vtbx3");
12218 }
12219 case NEON::BI__builtin_neon_vqtbx4q_v: {
12220 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx4, Ty),
12221 Ops, "vtbx4");
12222 }
12223 case NEON::BI__builtin_neon_vsqadd_v:
12224 case NEON::BI__builtin_neon_vsqaddq_v: {
12225 Int = Intrinsic::aarch64_neon_usqadd;
12226 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqadd");
12227 }
12228 case NEON::BI__builtin_neon_vuqadd_v:
12229 case NEON::BI__builtin_neon_vuqaddq_v: {
12230 Int = Intrinsic::aarch64_neon_suqadd;
12231 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vuqadd");
12232 }
12233 }
12234 }
12235
12236 Value *CodeGenFunction::EmitBPFBuiltinExpr(unsigned BuiltinID,
12237 const CallExpr *E) {
12238 assert((BuiltinID == BPF::BI__builtin_preserve_field_info ||
12239 BuiltinID == BPF::BI__builtin_btf_type_id ||
12240 BuiltinID == BPF::BI__builtin_preserve_type_info ||
12241 BuiltinID == BPF::BI__builtin_preserve_enum_value) &&
12242 "unexpected BPF builtin");
12243
12244 // A sequence number, injected into IR builtin functions, to
12245 // prevent CSE given the only difference of the function
12246 // may just be the debuginfo metadata.
12247 static uint32_t BuiltinSeqNum;
12248
12249 switch (BuiltinID) {
12250 default:
12251 llvm_unreachable("Unexpected BPF builtin");
12252 case BPF::BI__builtin_preserve_field_info: {
12253 const Expr *Arg = E->getArg(0);
12254 bool IsBitField = Arg->IgnoreParens()->getObjectKind() == OK_BitField;
12255
12256 if (!getDebugInfo()) {
12257 CGM.Error(E->getExprLoc(),
12258 "using __builtin_preserve_field_info() without -g");
12259 return IsBitField ? EmitLValue(Arg).getBitFieldPointer()
12260 : EmitLValue(Arg).getPointer(*this);
12261 }
12262
12263 // Enable underlying preserve_*_access_index() generation.
12264 bool OldIsInPreservedAIRegion = IsInPreservedAIRegion;
12265 IsInPreservedAIRegion = true;
12266 Value *FieldAddr = IsBitField ? EmitLValue(Arg).getBitFieldPointer()
12267 : EmitLValue(Arg).getPointer(*this);
12268 IsInPreservedAIRegion = OldIsInPreservedAIRegion;
12269
12270 ConstantInt *C = cast<ConstantInt>(EmitScalarExpr(E->getArg(1)));
12271 Value *InfoKind = ConstantInt::get(Int64Ty, C->getSExtValue());
12272
12273 // Built the IR for the preserve_field_info intrinsic.
12274 llvm::Function *FnGetFieldInfo = llvm::Intrinsic::getDeclaration(
12275 &CGM.getModule(), llvm::Intrinsic::bpf_preserve_field_info,
12276 {FieldAddr->getType()});
12277 return Builder.CreateCall(FnGetFieldInfo, {FieldAddr, InfoKind});
12278 }
12279 case BPF::BI__builtin_btf_type_id:
12280 case BPF::BI__builtin_preserve_type_info: {
12281 if (!getDebugInfo()) {
12282 CGM.Error(E->getExprLoc(), "using builtin function without -g");
12283 return nullptr;
12284 }
12285
12286 const Expr *Arg0 = E->getArg(0);
12287 llvm::DIType *DbgInfo = getDebugInfo()->getOrCreateStandaloneType(
12288 Arg0->getType(), Arg0->getExprLoc());
12289
12290 ConstantInt *Flag = cast<ConstantInt>(EmitScalarExpr(E->getArg(1)));
12291 Value *FlagValue = ConstantInt::get(Int64Ty, Flag->getSExtValue());
12292 Value *SeqNumVal = ConstantInt::get(Int32Ty, BuiltinSeqNum++);
12293
12294 llvm::Function *FnDecl;
12295 if (BuiltinID == BPF::BI__builtin_btf_type_id)
12296 FnDecl = llvm::Intrinsic::getDeclaration(
12297 &CGM.getModule(), llvm::Intrinsic::bpf_btf_type_id, {});
12298 else
12299 FnDecl = llvm::Intrinsic::getDeclaration(
12300 &CGM.getModule(), llvm::Intrinsic::bpf_preserve_type_info, {});
12301 CallInst *Fn = Builder.CreateCall(FnDecl, {SeqNumVal, FlagValue});
12302 Fn->setMetadata(LLVMContext::MD_preserve_access_index, DbgInfo);
12303 return Fn;
12304 }
12305 case BPF::BI__builtin_preserve_enum_value: {
12306 if (!getDebugInfo()) {
12307 CGM.Error(E->getExprLoc(), "using builtin function without -g");
12308 return nullptr;
12309 }
12310
12311 const Expr *Arg0 = E->getArg(0);
12312 llvm::DIType *DbgInfo = getDebugInfo()->getOrCreateStandaloneType(
12313 Arg0->getType(), Arg0->getExprLoc());
12314
12315 // Find enumerator
12316 const auto *UO = cast<UnaryOperator>(Arg0->IgnoreParens());
12317 const auto *CE = cast<CStyleCastExpr>(UO->getSubExpr());
12318 const auto *DR = cast<DeclRefExpr>(CE->getSubExpr());
12319 const auto *Enumerator = cast<EnumConstantDecl>(DR->getDecl());
12320
12321 auto &InitVal = Enumerator->getInitVal();
12322 std::string InitValStr;
12323 if (InitVal.isNegative() || InitVal > uint64_t(INT64_MAX))
12324 InitValStr = std::to_string(InitVal.getSExtValue());
12325 else
12326 InitValStr = std::to_string(InitVal.getZExtValue());
12327 std::string EnumStr = Enumerator->getNameAsString() + ":" + InitValStr;
12328 Value *EnumStrVal = Builder.CreateGlobalStringPtr(EnumStr);
12329
12330 ConstantInt *Flag = cast<ConstantInt>(EmitScalarExpr(E->getArg(1)));
12331 Value *FlagValue = ConstantInt::get(Int64Ty, Flag->getSExtValue());
12332 Value *SeqNumVal = ConstantInt::get(Int32Ty, BuiltinSeqNum++);
12333
12334 llvm::Function *IntrinsicFn = llvm::Intrinsic::getDeclaration(
12335 &CGM.getModule(), llvm::Intrinsic::bpf_preserve_enum_value, {});
12336 CallInst *Fn =
12337 Builder.CreateCall(IntrinsicFn, {SeqNumVal, EnumStrVal, FlagValue});
12338 Fn->setMetadata(LLVMContext::MD_preserve_access_index, DbgInfo);
12339 return Fn;
12340 }
12341 }
12342 }
12343
12344 llvm::Value *CodeGenFunction::
12345 BuildVector(ArrayRef<llvm::Value*> Ops) {
12346 assert((Ops.size() & (Ops.size() - 1)) == 0 &&
12347 "Not a power-of-two sized vector!");
12348 bool AllConstants = true;
12349 for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i)
12350 AllConstants &= isa<Constant>(Ops[i]);
12351
12352 // If this is a constant vector, create a ConstantVector.
12353 if (AllConstants) {
12354 SmallVector<llvm::Constant*, 16> CstOps;
12355 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
12356 CstOps.push_back(cast<Constant>(Ops[i]));
12357 return llvm::ConstantVector::get(CstOps);
12358 }
12359
12360 // Otherwise, insertelement the values to build the vector.
12361 Value *Result = llvm::PoisonValue::get(
12362 llvm::FixedVectorType::get(Ops[0]->getType(), Ops.size()));
12363
12364 for (unsigned i = 0, e = Ops.size(); i != e; ++i)
12365 Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt64(i));
12366
12367 return Result;
12368 }
12369
12370 // Convert the mask from an integer type to a vector of i1.
12371 static Value *getMaskVecValue(CodeGenFunction &CGF, Value *Mask,
12372 unsigned NumElts) {
12373
12374 auto *MaskTy = llvm::FixedVectorType::get(
12375 CGF.Builder.getInt1Ty(),
12376 cast<IntegerType>(Mask->getType())->getBitWidth());
12377 Value *MaskVec = CGF.Builder.CreateBitCast(Mask, MaskTy);
12378
12379 // If we have less than 8 elements, then the starting mask was an i8 and
12380 // we need to extract down to the right number of elements.
12381 if (NumElts < 8) {
12382 int Indices[4];
12383 for (unsigned i = 0; i != NumElts; ++i)
12384 Indices[i] = i;
12385 MaskVec = CGF.Builder.CreateShuffleVector(
12386 MaskVec, MaskVec, ArrayRef(Indices, NumElts), "extract");
12387 }
12388 return MaskVec;
12389 }
12390
12391 static Value *EmitX86MaskedStore(CodeGenFunction &CGF, ArrayRef<Value *> Ops,
12392 Align Alignment) {
12393 // Cast the pointer to right type.
12394 Value *Ptr = CGF.Builder.CreateBitCast(Ops[0],
12395 llvm::PointerType::getUnqual(Ops[1]->getType()));
12396
12397 Value *MaskVec = getMaskVecValue(
12398 CGF, Ops[2],
12399 cast<llvm::FixedVectorType>(Ops[1]->getType())->getNumElements());
12400
12401 return CGF.Builder.CreateMaskedStore(Ops[1], Ptr, Alignment, MaskVec);
12402 }
12403
12404 static Value *EmitX86MaskedLoad(CodeGenFunction &CGF, ArrayRef<Value *> Ops,
12405 Align Alignment) {
12406 // Cast the pointer to right type.
12407 llvm::Type *Ty = Ops[1]->getType();
12408 Value *Ptr =
12409 CGF.Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty));
12410
12411 Value *MaskVec = getMaskVecValue(
12412 CGF, Ops[2], cast<llvm::FixedVectorType>(Ty)->getNumElements());
12413
12414 return CGF.Builder.CreateMaskedLoad(Ty, Ptr, Alignment, MaskVec, Ops[1]);
12415 }
12416
12417 static Value *EmitX86ExpandLoad(CodeGenFunction &CGF,
12418 ArrayRef<Value *> Ops) {
12419 auto *ResultTy = cast<llvm::VectorType>(Ops[1]->getType());
12420 llvm::Type *PtrTy = ResultTy->getElementType();
12421
12422 // Cast the pointer to element type.
12423 Value *Ptr = CGF.Builder.CreateBitCast(Ops[0],
12424 llvm::PointerType::getUnqual(PtrTy));
12425
12426 Value *MaskVec = getMaskVecValue(
12427 CGF, Ops[2], cast<FixedVectorType>(ResultTy)->getNumElements());
12428
12429 llvm::Function *F = CGF.CGM.getIntrinsic(Intrinsic::masked_expandload,
12430 ResultTy);
12431 return CGF.Builder.CreateCall(F, { Ptr, MaskVec, Ops[1] });
12432 }
12433
12434 static Value *EmitX86CompressExpand(CodeGenFunction &CGF,
12435 ArrayRef<Value *> Ops,
12436 bool IsCompress) {
12437 auto *ResultTy = cast<llvm::FixedVectorType>(Ops[1]->getType());
12438
12439 Value *MaskVec = getMaskVecValue(CGF, Ops[2], ResultTy->getNumElements());
12440
12441 Intrinsic::ID IID = IsCompress ? Intrinsic::x86_avx512_mask_compress
12442 : Intrinsic::x86_avx512_mask_expand;
12443 llvm::Function *F = CGF.CGM.getIntrinsic(IID, ResultTy);
12444 return CGF.Builder.CreateCall(F, { Ops[0], Ops[1], MaskVec });
12445 }
12446
12447 static Value *EmitX86CompressStore(CodeGenFunction &CGF,
12448 ArrayRef<Value *> Ops) {
12449 auto *ResultTy = cast<llvm::FixedVectorType>(Ops[1]->getType());
12450 llvm::Type *PtrTy = ResultTy->getElementType();
12451
12452 // Cast the pointer to element type.
12453 Value *Ptr = CGF.Builder.CreateBitCast(Ops[0],
12454 llvm::PointerType::getUnqual(PtrTy));
12455
12456 Value *MaskVec = getMaskVecValue(CGF, Ops[2], ResultTy->getNumElements());
12457
12458 llvm::Function *F = CGF.CGM.getIntrinsic(Intrinsic::masked_compressstore,
12459 ResultTy);
12460 return CGF.Builder.CreateCall(F, { Ops[1], Ptr, MaskVec });
12461 }
12462
12463 static Value *EmitX86MaskLogic(CodeGenFunction &CGF, Instruction::BinaryOps Opc,
12464 ArrayRef<Value *> Ops,
12465 bool InvertLHS = false) {
12466 unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth();
12467 Value *LHS = getMaskVecValue(CGF, Ops[0], NumElts);
12468 Value *RHS = getMaskVecValue(CGF, Ops[1], NumElts);
12469
12470 if (InvertLHS)
12471 LHS = CGF.Builder.CreateNot(LHS);
12472
12473 return CGF.Builder.CreateBitCast(CGF.Builder.CreateBinOp(Opc, LHS, RHS),
12474 Ops[0]->getType());
12475 }
12476
12477 static Value *EmitX86FunnelShift(CodeGenFunction &CGF, Value *Op0, Value *Op1,
12478 Value *Amt, bool IsRight) {
12479 llvm::Type *Ty = Op0->getType();
12480
12481 // Amount may be scalar immediate, in which case create a splat vector.
12482 // Funnel shifts amounts are treated as modulo and types are all power-of-2 so
12483 // we only care about the lowest log2 bits anyway.
12484 if (Amt->getType() != Ty) {
12485 unsigned NumElts = cast<llvm::FixedVectorType>(Ty)->getNumElements();
12486 Amt = CGF.Builder.CreateIntCast(Amt, Ty->getScalarType(), false);
12487 Amt = CGF.Builder.CreateVectorSplat(NumElts, Amt);
12488 }
12489
12490 unsigned IID = IsRight ? Intrinsic::fshr : Intrinsic::fshl;
12491 Function *F = CGF.CGM.getIntrinsic(IID, Ty);
12492 return CGF.Builder.CreateCall(F, {Op0, Op1, Amt});
12493 }
12494
12495 static Value *EmitX86vpcom(CodeGenFunction &CGF, ArrayRef<Value *> Ops,
12496 bool IsSigned) {
12497 Value *Op0 = Ops[0];
12498 Value *Op1 = Ops[1];
12499 llvm::Type *Ty = Op0->getType();
12500 uint64_t Imm = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0x7;
12501
12502 CmpInst::Predicate Pred;
12503 switch (Imm) {
12504 case 0x0:
12505 Pred = IsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
12506 break;
12507 case 0x1:
12508 Pred = IsSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
12509 break;
12510 case 0x2:
12511 Pred = IsSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
12512 break;
12513 case 0x3:
12514 Pred = IsSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE;
12515 break;
12516 case 0x4:
12517 Pred = ICmpInst::ICMP_EQ;
12518 break;
12519 case 0x5:
12520 Pred = ICmpInst::ICMP_NE;
12521 break;
12522 case 0x6:
12523 return llvm::Constant::getNullValue(Ty); // FALSE
12524 case 0x7:
12525 return llvm::Constant::getAllOnesValue(Ty); // TRUE
12526 default:
12527 llvm_unreachable("Unexpected XOP vpcom/vpcomu predicate");
12528 }
12529
12530 Value *Cmp = CGF.Builder.CreateICmp(Pred, Op0, Op1);
12531 Value *Res = CGF.Builder.CreateSExt(Cmp, Ty);
12532 return Res;
12533 }
12534
12535 static Value *EmitX86Select(CodeGenFunction &CGF,
12536 Value *Mask, Value *Op0, Value *Op1) {
12537
12538 // If the mask is all ones just return first argument.
12539 if (const auto *C = dyn_cast<Constant>(Mask))
12540 if (C->isAllOnesValue())
12541 return Op0;
12542
12543 Mask = getMaskVecValue(
12544 CGF, Mask, cast<llvm::FixedVectorType>(Op0->getType())->getNumElements());
12545
12546 return CGF.Builder.CreateSelect(Mask, Op0, Op1);
12547 }
12548
12549 static Value *EmitX86ScalarSelect(CodeGenFunction &CGF,
12550 Value *Mask, Value *Op0, Value *Op1) {
12551 // If the mask is all ones just return first argument.
12552 if (const auto *C = dyn_cast<Constant>(Mask))
12553 if (C->isAllOnesValue())
12554 return Op0;
12555
12556 auto *MaskTy = llvm::FixedVectorType::get(
12557 CGF.Builder.getInt1Ty(), Mask->getType()->getIntegerBitWidth());
12558 Mask = CGF.Builder.CreateBitCast(Mask, MaskTy);
12559 Mask = CGF.Builder.CreateExtractElement(Mask, (uint64_t)0);
12560 return CGF.Builder.CreateSelect(Mask, Op0, Op1);
12561 }
12562
12563 static Value *EmitX86MaskedCompareResult(CodeGenFunction &CGF, Value *Cmp,
12564 unsigned NumElts, Value *MaskIn) {
12565 if (MaskIn) {
12566 const auto *C = dyn_cast<Constant>(MaskIn);
12567 if (!C || !C->isAllOnesValue())
12568 Cmp = CGF.Builder.CreateAnd(Cmp, getMaskVecValue(CGF, MaskIn, NumElts));
12569 }
12570
12571 if (NumElts < 8) {
12572 int Indices[8];
12573 for (unsigned i = 0; i != NumElts; ++i)
12574 Indices[i] = i;
12575 for (unsigned i = NumElts; i != 8; ++i)
12576 Indices[i] = i % NumElts + NumElts;
12577 Cmp = CGF.Builder.CreateShuffleVector(
12578 Cmp, llvm::Constant::getNullValue(Cmp->getType()), Indices);
12579 }
12580
12581 return CGF.Builder.CreateBitCast(Cmp,
12582 IntegerType::get(CGF.getLLVMContext(),
12583 std::max(NumElts, 8U)));
12584 }
12585
12586 static Value *EmitX86MaskedCompare(CodeGenFunction &CGF, unsigned CC,
12587 bool Signed, ArrayRef<Value *> Ops) {
12588 assert((Ops.size() == 2 || Ops.size() == 4) &&
12589 "Unexpected number of arguments");
12590 unsigned NumElts =
12591 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements();
12592 Value *Cmp;
12593
12594 if (CC == 3) {
12595 Cmp = Constant::getNullValue(
12596 llvm::FixedVectorType::get(CGF.Builder.getInt1Ty(), NumElts));
12597 } else if (CC == 7) {
12598 Cmp = Constant::getAllOnesValue(
12599 llvm::FixedVectorType::get(CGF.Builder.getInt1Ty(), NumElts));
12600 } else {
12601 ICmpInst::Predicate Pred;
12602 switch (CC) {
12603 default: llvm_unreachable("Unknown condition code");
12604 case 0: Pred = ICmpInst::ICMP_EQ; break;
12605 case 1: Pred = Signed ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; break;
12606 case 2: Pred = Signed ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; break;
12607 case 4: Pred = ICmpInst::ICMP_NE; break;
12608 case 5: Pred = Signed ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; break;
12609 case 6: Pred = Signed ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; break;
12610 }
12611 Cmp = CGF.Builder.CreateICmp(Pred, Ops[0], Ops[1]);
12612 }
12613
12614 Value *MaskIn = nullptr;
12615 if (Ops.size() == 4)
12616 MaskIn = Ops[3];
12617
12618 return EmitX86MaskedCompareResult(CGF, Cmp, NumElts, MaskIn);
12619 }
12620
12621 static Value *EmitX86ConvertToMask(CodeGenFunction &CGF, Value *In) {
12622 Value *Zero = Constant::getNullValue(In->getType());
12623 return EmitX86MaskedCompare(CGF, 1, true, { In, Zero });
12624 }
12625
12626 static Value *EmitX86ConvertIntToFp(CodeGenFunction &CGF, const CallExpr *E,
12627 ArrayRef<Value *> Ops, bool IsSigned) {
12628 unsigned Rnd = cast<llvm::ConstantInt>(Ops[3])->getZExtValue();
12629 llvm::Type *Ty = Ops[1]->getType();
12630
12631 Value *Res;
12632 if (Rnd != 4) {
12633 Intrinsic::ID IID = IsSigned ? Intrinsic::x86_avx512_sitofp_round
12634 : Intrinsic::x86_avx512_uitofp_round;
12635 Function *F = CGF.CGM.getIntrinsic(IID, { Ty, Ops[0]->getType() });
12636 Res = CGF.Builder.CreateCall(F, { Ops[0], Ops[3] });
12637 } else {
12638 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(CGF, E);
12639 Res = IsSigned ? CGF.Builder.CreateSIToFP(Ops[0], Ty)
12640 : CGF.Builder.CreateUIToFP(Ops[0], Ty);
12641 }
12642
12643 return EmitX86Select(CGF, Ops[2], Res, Ops[1]);
12644 }
12645
12646 // Lowers X86 FMA intrinsics to IR.
12647 static Value *EmitX86FMAExpr(CodeGenFunction &CGF, const CallExpr *E,
12648 ArrayRef<Value *> Ops, unsigned BuiltinID,
12649 bool IsAddSub) {
12650
12651 bool Subtract = false;
12652 Intrinsic::ID IID = Intrinsic::not_intrinsic;
12653 switch (BuiltinID) {
12654 default: break;
12655 case clang::X86::BI__builtin_ia32_vfmsubph512_mask3:
12656 Subtract = true;
12657 [[fallthrough]];
12658 case clang::X86::BI__builtin_ia32_vfmaddph512_mask:
12659 case clang::X86::BI__builtin_ia32_vfmaddph512_maskz:
12660 case clang::X86::BI__builtin_ia32_vfmaddph512_mask3:
12661 IID = llvm::Intrinsic::x86_avx512fp16_vfmadd_ph_512;
12662 break;
12663 case clang::X86::BI__builtin_ia32_vfmsubaddph512_mask3:
12664 Subtract = true;
12665 [[fallthrough]];
12666 case clang::X86::BI__builtin_ia32_vfmaddsubph512_mask:
12667 case clang::X86::BI__builtin_ia32_vfmaddsubph512_maskz:
12668 case clang::X86::BI__builtin_ia32_vfmaddsubph512_mask3:
12669 IID = llvm::Intrinsic::x86_avx512fp16_vfmaddsub_ph_512;
12670 break;
12671 case clang::X86::BI__builtin_ia32_vfmsubps512_mask3:
12672 Subtract = true;
12673 [[fallthrough]];
12674 case clang::X86::BI__builtin_ia32_vfmaddps512_mask:
12675 case clang::X86::BI__builtin_ia32_vfmaddps512_maskz:
12676 case clang::X86::BI__builtin_ia32_vfmaddps512_mask3:
12677 IID = llvm::Intrinsic::x86_avx512_vfmadd_ps_512; break;
12678 case clang::X86::BI__builtin_ia32_vfmsubpd512_mask3:
12679 Subtract = true;
12680 [[fallthrough]];
12681 case clang::X86::BI__builtin_ia32_vfmaddpd512_mask:
12682 case clang::X86::BI__builtin_ia32_vfmaddpd512_maskz:
12683 case clang::X86::BI__builtin_ia32_vfmaddpd512_mask3:
12684 IID = llvm::Intrinsic::x86_avx512_vfmadd_pd_512; break;
12685 case clang::X86::BI__builtin_ia32_vfmsubaddps512_mask3:
12686 Subtract = true;
12687 [[fallthrough]];
12688 case clang::X86::BI__builtin_ia32_vfmaddsubps512_mask:
12689 case clang::X86::BI__builtin_ia32_vfmaddsubps512_maskz:
12690 case clang::X86::BI__builtin_ia32_vfmaddsubps512_mask3:
12691 IID = llvm::Intrinsic::x86_avx512_vfmaddsub_ps_512;
12692 break;
12693 case clang::X86::BI__builtin_ia32_vfmsubaddpd512_mask3:
12694 Subtract = true;
12695 [[fallthrough]];
12696 case clang::X86::BI__builtin_ia32_vfmaddsubpd512_mask:
12697 case clang::X86::BI__builtin_ia32_vfmaddsubpd512_maskz:
12698 case clang::X86::BI__builtin_ia32_vfmaddsubpd512_mask3:
12699 IID = llvm::Intrinsic::x86_avx512_vfmaddsub_pd_512;
12700 break;
12701 }
12702
12703 Value *A = Ops[0];
12704 Value *B = Ops[1];
12705 Value *C = Ops[2];
12706
12707 if (Subtract)
12708 C = CGF.Builder.CreateFNeg(C);
12709
12710 Value *Res;
12711
12712 // Only handle in case of _MM_FROUND_CUR_DIRECTION/4 (no rounding).
12713 if (IID != Intrinsic::not_intrinsic &&
12714 (cast<llvm::ConstantInt>(Ops.back())->getZExtValue() != (uint64_t)4 ||
12715 IsAddSub)) {
12716 Function *Intr = CGF.CGM.getIntrinsic(IID);
12717 Res = CGF.Builder.CreateCall(Intr, {A, B, C, Ops.back() });
12718 } else {
12719 llvm::Type *Ty = A->getType();
12720 Function *FMA;
12721 if (CGF.Builder.getIsFPConstrained()) {
12722 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(CGF, E);
12723 FMA = CGF.CGM.getIntrinsic(Intrinsic::experimental_constrained_fma, Ty);
12724 Res = CGF.Builder.CreateConstrainedFPCall(FMA, {A, B, C});
12725 } else {
12726 FMA = CGF.CGM.getIntrinsic(Intrinsic::fma, Ty);
12727 Res = CGF.Builder.CreateCall(FMA, {A, B, C});
12728 }
12729 }
12730
12731 // Handle any required masking.
12732 Value *MaskFalseVal = nullptr;
12733 switch (BuiltinID) {
12734 case clang::X86::BI__builtin_ia32_vfmaddph512_mask:
12735 case clang::X86::BI__builtin_ia32_vfmaddps512_mask:
12736 case clang::X86::BI__builtin_ia32_vfmaddpd512_mask:
12737 case clang::X86::BI__builtin_ia32_vfmaddsubph512_mask:
12738 case clang::X86::BI__builtin_ia32_vfmaddsubps512_mask:
12739 case clang::X86::BI__builtin_ia32_vfmaddsubpd512_mask:
12740 MaskFalseVal = Ops[0];
12741 break;
12742 case clang::X86::BI__builtin_ia32_vfmaddph512_maskz:
12743 case clang::X86::BI__builtin_ia32_vfmaddps512_maskz:
12744 case clang::X86::BI__builtin_ia32_vfmaddpd512_maskz:
12745 case clang::X86::BI__builtin_ia32_vfmaddsubph512_maskz:
12746 case clang::X86::BI__builtin_ia32_vfmaddsubps512_maskz:
12747 case clang::X86::BI__builtin_ia32_vfmaddsubpd512_maskz:
12748 MaskFalseVal = Constant::getNullValue(Ops[0]->getType());
12749 break;
12750 case clang::X86::BI__builtin_ia32_vfmsubph512_mask3:
12751 case clang::X86::BI__builtin_ia32_vfmaddph512_mask3:
12752 case clang::X86::BI__builtin_ia32_vfmsubps512_mask3:
12753 case clang::X86::BI__builtin_ia32_vfmaddps512_mask3:
12754 case clang::X86::BI__builtin_ia32_vfmsubpd512_mask3:
12755 case clang::X86::BI__builtin_ia32_vfmaddpd512_mask3:
12756 case clang::X86::BI__builtin_ia32_vfmsubaddph512_mask3:
12757 case clang::X86::BI__builtin_ia32_vfmaddsubph512_mask3:
12758 case clang::X86::BI__builtin_ia32_vfmsubaddps512_mask3:
12759 case clang::X86::BI__builtin_ia32_vfmaddsubps512_mask3:
12760 case clang::X86::BI__builtin_ia32_vfmsubaddpd512_mask3:
12761 case clang::X86::BI__builtin_ia32_vfmaddsubpd512_mask3:
12762 MaskFalseVal = Ops[2];
12763 break;
12764 }
12765
12766 if (MaskFalseVal)
12767 return EmitX86Select(CGF, Ops[3], Res, MaskFalseVal);
12768
12769 return Res;
12770 }
12771
12772 static Value *EmitScalarFMAExpr(CodeGenFunction &CGF, const CallExpr *E,
12773 MutableArrayRef<Value *> Ops, Value *Upper,
12774 bool ZeroMask = false, unsigned PTIdx = 0,
12775 bool NegAcc = false) {
12776 unsigned Rnd = 4;
12777 if (Ops.size() > 4)
12778 Rnd = cast<llvm::ConstantInt>(Ops[4])->getZExtValue();
12779
12780 if (NegAcc)
12781 Ops[2] = CGF.Builder.CreateFNeg(Ops[2]);
12782
12783 Ops[0] = CGF.Builder.CreateExtractElement(Ops[0], (uint64_t)0);
12784 Ops[1] = CGF.Builder.CreateExtractElement(Ops[1], (uint64_t)0);
12785 Ops[2] = CGF.Builder.CreateExtractElement(Ops[2], (uint64_t)0);
12786 Value *Res;
12787 if (Rnd != 4) {
12788 Intrinsic::ID IID;
12789
12790 switch (Ops[0]->getType()->getPrimitiveSizeInBits()) {
12791 case 16:
12792 IID = Intrinsic::x86_avx512fp16_vfmadd_f16;
12793 break;
12794 case 32:
12795 IID = Intrinsic::x86_avx512_vfmadd_f32;
12796 break;
12797 case 64:
12798 IID = Intrinsic::x86_avx512_vfmadd_f64;
12799 break;
12800 default:
12801 llvm_unreachable("Unexpected size");
12802 }
12803 Res = CGF.Builder.CreateCall(CGF.CGM.getIntrinsic(IID),
12804 {Ops[0], Ops[1], Ops[2], Ops[4]});
12805 } else if (CGF.Builder.getIsFPConstrained()) {
12806 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(CGF, E);
12807 Function *FMA = CGF.CGM.getIntrinsic(
12808 Intrinsic::experimental_constrained_fma, Ops[0]->getType());
12809 Res = CGF.Builder.CreateConstrainedFPCall(FMA, Ops.slice(0, 3));
12810 } else {
12811 Function *FMA = CGF.CGM.getIntrinsic(Intrinsic::fma, Ops[0]->getType());
12812 Res = CGF.Builder.CreateCall(FMA, Ops.slice(0, 3));
12813 }
12814 // If we have more than 3 arguments, we need to do masking.
12815 if (Ops.size() > 3) {
12816 Value *PassThru = ZeroMask ? Constant::getNullValue(Res->getType())
12817 : Ops[PTIdx];
12818
12819 // If we negated the accumulator and the its the PassThru value we need to
12820 // bypass the negate. Conveniently Upper should be the same thing in this
12821 // case.
12822 if (NegAcc && PTIdx == 2)
12823 PassThru = CGF.Builder.CreateExtractElement(Upper, (uint64_t)0);
12824
12825 Res = EmitX86ScalarSelect(CGF, Ops[3], Res, PassThru);
12826 }
12827 return CGF.Builder.CreateInsertElement(Upper, Res, (uint64_t)0);
12828 }
12829
12830 static Value *EmitX86Muldq(CodeGenFunction &CGF, bool IsSigned,
12831 ArrayRef<Value *> Ops) {
12832 llvm::Type *Ty = Ops[0]->getType();
12833 // Arguments have a vXi32 type so cast to vXi64.
12834 Ty = llvm::FixedVectorType::get(CGF.Int64Ty,
12835 Ty->getPrimitiveSizeInBits() / 64);
12836 Value *LHS = CGF.Builder.CreateBitCast(Ops[0], Ty);
12837 Value *RHS = CGF.Builder.CreateBitCast(Ops[1], Ty);
12838
12839 if (IsSigned) {
12840 // Shift left then arithmetic shift right.
12841 Constant *ShiftAmt = ConstantInt::get(Ty, 32);
12842 LHS = CGF.Builder.CreateShl(LHS, ShiftAmt);
12843 LHS = CGF.Builder.CreateAShr(LHS, ShiftAmt);
12844 RHS = CGF.Builder.CreateShl(RHS, ShiftAmt);
12845 RHS = CGF.Builder.CreateAShr(RHS, ShiftAmt);
12846 } else {
12847 // Clear the upper bits.
12848 Constant *Mask = ConstantInt::get(Ty, 0xffffffff);
12849 LHS = CGF.Builder.CreateAnd(LHS, Mask);
12850 RHS = CGF.Builder.CreateAnd(RHS, Mask);
12851 }
12852
12853 return CGF.Builder.CreateMul(LHS, RHS);
12854 }
12855
12856 // Emit a masked pternlog intrinsic. This only exists because the header has to
12857 // use a macro and we aren't able to pass the input argument to a pternlog
12858 // builtin and a select builtin without evaluating it twice.
12859 static Value *EmitX86Ternlog(CodeGenFunction &CGF, bool ZeroMask,
12860 ArrayRef<Value *> Ops) {
12861 llvm::Type *Ty = Ops[0]->getType();
12862
12863 unsigned VecWidth = Ty->getPrimitiveSizeInBits();
12864 unsigned EltWidth = Ty->getScalarSizeInBits();
12865 Intrinsic::ID IID;
12866 if (VecWidth == 128 && EltWidth == 32)
12867 IID = Intrinsic::x86_avx512_pternlog_d_128;
12868 else if (VecWidth == 256 && EltWidth == 32)
12869 IID = Intrinsic::x86_avx512_pternlog_d_256;
12870 else if (VecWidth == 512 && EltWidth == 32)
12871 IID = Intrinsic::x86_avx512_pternlog_d_512;
12872 else if (VecWidth == 128 && EltWidth == 64)
12873 IID = Intrinsic::x86_avx512_pternlog_q_128;
12874 else if (VecWidth == 256 && EltWidth == 64)
12875 IID = Intrinsic::x86_avx512_pternlog_q_256;
12876 else if (VecWidth == 512 && EltWidth == 64)
12877 IID = Intrinsic::x86_avx512_pternlog_q_512;
12878 else
12879 llvm_unreachable("Unexpected intrinsic");
12880
12881 Value *Ternlog = CGF.Builder.CreateCall(CGF.CGM.getIntrinsic(IID),
12882 Ops.drop_back());
12883 Value *PassThru = ZeroMask ? ConstantAggregateZero::get(Ty) : Ops[0];
12884 return EmitX86Select(CGF, Ops[4], Ternlog, PassThru);
12885 }
12886
12887 static Value *EmitX86SExtMask(CodeGenFunction &CGF, Value *Op,
12888 llvm::Type *DstTy) {
12889 unsigned NumberOfElements =
12890 cast<llvm::FixedVectorType>(DstTy)->getNumElements();
12891 Value *Mask = getMaskVecValue(CGF, Op, NumberOfElements);
12892 return CGF.Builder.CreateSExt(Mask, DstTy, "vpmovm2");
12893 }
12894
12895 Value *CodeGenFunction::EmitX86CpuIs(const CallExpr *E) {
12896 const Expr *CPUExpr = E->getArg(0)->IgnoreParenCasts();
12897 StringRef CPUStr = cast<clang::StringLiteral>(CPUExpr)->getString();
12898 return EmitX86CpuIs(CPUStr);
12899 }
12900
12901 // Convert F16 halfs to floats.
12902 static Value *EmitX86CvtF16ToFloatExpr(CodeGenFunction &CGF,
12903 ArrayRef<Value *> Ops,
12904 llvm::Type *DstTy) {
12905 assert((Ops.size() == 1 || Ops.size() == 3 || Ops.size() == 4) &&
12906 "Unknown cvtph2ps intrinsic");
12907
12908 // If the SAE intrinsic doesn't use default rounding then we can't upgrade.
12909 if (Ops.size() == 4 && cast<llvm::ConstantInt>(Ops[3])->getZExtValue() != 4) {
12910 Function *F =
12911 CGF.CGM.getIntrinsic(Intrinsic::x86_avx512_mask_vcvtph2ps_512);
12912 return CGF.Builder.CreateCall(F, {Ops[0], Ops[1], Ops[2], Ops[3]});
12913 }
12914
12915 unsigned NumDstElts = cast<llvm::FixedVectorType>(DstTy)->getNumElements();
12916 Value *Src = Ops[0];
12917
12918 // Extract the subvector.
12919 if (NumDstElts !=
12920 cast<llvm::FixedVectorType>(Src->getType())->getNumElements()) {
12921 assert(NumDstElts == 4 && "Unexpected vector size");
12922 Src = CGF.Builder.CreateShuffleVector(Src, ArrayRef<int>{0, 1, 2, 3});
12923 }
12924
12925 // Bitcast from vXi16 to vXf16.
12926 auto *HalfTy = llvm::FixedVectorType::get(
12927 llvm::Type::getHalfTy(CGF.getLLVMContext()), NumDstElts);
12928 Src = CGF.Builder.CreateBitCast(Src, HalfTy);
12929
12930 // Perform the fp-extension.
12931 Value *Res = CGF.Builder.CreateFPExt(Src, DstTy, "cvtph2ps");
12932
12933 if (Ops.size() >= 3)
12934 Res = EmitX86Select(CGF, Ops[2], Res, Ops[1]);
12935 return Res;
12936 }
12937
12938 Value *CodeGenFunction::EmitX86CpuIs(StringRef CPUStr) {
12939
12940 llvm::Type *Int32Ty = Builder.getInt32Ty();
12941
12942 // Matching the struct layout from the compiler-rt/libgcc structure that is
12943 // filled in:
12944 // unsigned int __cpu_vendor;
12945 // unsigned int __cpu_type;
12946 // unsigned int __cpu_subtype;
12947 // unsigned int __cpu_features[1];
12948 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, Int32Ty,
12949 llvm::ArrayType::get(Int32Ty, 1));
12950
12951 // Grab the global __cpu_model.
12952 llvm::Constant *CpuModel = CGM.CreateRuntimeVariable(STy, "__cpu_model");
12953 cast<llvm::GlobalValue>(CpuModel)->setDSOLocal(true);
12954
12955 // Calculate the index needed to access the correct field based on the
12956 // range. Also adjust the expected value.
12957 unsigned Index;
12958 unsigned Value;
12959 std::tie(Index, Value) = StringSwitch<std::pair<unsigned, unsigned>>(CPUStr)
12960 #define X86_VENDOR(ENUM, STRING) \
12961 .Case(STRING, {0u, static_cast<unsigned>(llvm::X86::ENUM)})
12962 #define X86_CPU_TYPE_ALIAS(ENUM, ALIAS) \
12963 .Case(ALIAS, {1u, static_cast<unsigned>(llvm::X86::ENUM)})
12964 #define X86_CPU_TYPE(ENUM, STR) \
12965 .Case(STR, {1u, static_cast<unsigned>(llvm::X86::ENUM)})
12966 #define X86_CPU_SUBTYPE_ALIAS(ENUM, ALIAS) \
12967 .Case(ALIAS, {2u, static_cast<unsigned>(llvm::X86::ENUM)})
12968 #define X86_CPU_SUBTYPE(ENUM, STR) \
12969 .Case(STR, {2u, static_cast<unsigned>(llvm::X86::ENUM)})
12970 #include "llvm/TargetParser/X86TargetParser.def"
12971 .Default({0, 0});
12972 assert(Value != 0 && "Invalid CPUStr passed to CpuIs");
12973
12974 // Grab the appropriate field from __cpu_model.
12975 llvm::Value *Idxs[] = {ConstantInt::get(Int32Ty, 0),
12976 ConstantInt::get(Int32Ty, Index)};
12977 llvm::Value *CpuValue = Builder.CreateGEP(STy, CpuModel, Idxs);
12978 CpuValue = Builder.CreateAlignedLoad(Int32Ty, CpuValue,
12979 CharUnits::fromQuantity(4));
12980
12981 // Check the value of the field against the requested value.
12982 return Builder.CreateICmpEQ(CpuValue,
12983 llvm::ConstantInt::get(Int32Ty, Value));
12984 }
12985
12986 Value *CodeGenFunction::EmitX86CpuSupports(const CallExpr *E) {
12987 const Expr *FeatureExpr = E->getArg(0)->IgnoreParenCasts();
12988 StringRef FeatureStr = cast<StringLiteral>(FeatureExpr)->getString();
12989 return EmitX86CpuSupports(FeatureStr);
12990 }
12991
12992 Value *CodeGenFunction::EmitX86CpuSupports(ArrayRef<StringRef> FeatureStrs) {
12993 return EmitX86CpuSupports(llvm::X86::getCpuSupportsMask(FeatureStrs));
12994 }
12995
12996 llvm::Value *CodeGenFunction::EmitX86CpuSupports(uint64_t FeaturesMask) {
12997 uint32_t Features1 = Lo_32(FeaturesMask);
12998 uint32_t Features2 = Hi_32(FeaturesMask);
12999
13000 Value *Result = Builder.getTrue();
13001
13002 if (Features1 != 0) {
13003 // Matching the struct layout from the compiler-rt/libgcc structure that is
13004 // filled in:
13005 // unsigned int __cpu_vendor;
13006 // unsigned int __cpu_type;
13007 // unsigned int __cpu_subtype;
13008 // unsigned int __cpu_features[1];
13009 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, Int32Ty,
13010 llvm::ArrayType::get(Int32Ty, 1));
13011
13012 // Grab the global __cpu_model.
13013 llvm::Constant *CpuModel = CGM.CreateRuntimeVariable(STy, "__cpu_model");
13014 cast<llvm::GlobalValue>(CpuModel)->setDSOLocal(true);
13015
13016 // Grab the first (0th) element from the field __cpu_features off of the
13017 // global in the struct STy.
13018 Value *Idxs[] = {Builder.getInt32(0), Builder.getInt32(3),
13019 Builder.getInt32(0)};
13020 Value *CpuFeatures = Builder.CreateGEP(STy, CpuModel, Idxs);
13021 Value *Features = Builder.CreateAlignedLoad(Int32Ty, CpuFeatures,
13022 CharUnits::fromQuantity(4));
13023
13024 // Check the value of the bit corresponding to the feature requested.
13025 Value *Mask = Builder.getInt32(Features1);
13026 Value *Bitset = Builder.CreateAnd(Features, Mask);
13027 Value *Cmp = Builder.CreateICmpEQ(Bitset, Mask);
13028 Result = Builder.CreateAnd(Result, Cmp);
13029 }
13030
13031 if (Features2 != 0) {
13032 llvm::Constant *CpuFeatures2 = CGM.CreateRuntimeVariable(Int32Ty,
13033 "__cpu_features2");
13034 cast<llvm::GlobalValue>(CpuFeatures2)->setDSOLocal(true);
13035
13036 Value *Features = Builder.CreateAlignedLoad(Int32Ty, CpuFeatures2,
13037 CharUnits::fromQuantity(4));
13038
13039 // Check the value of the bit corresponding to the feature requested.
13040 Value *Mask = Builder.getInt32(Features2);
13041 Value *Bitset = Builder.CreateAnd(Features, Mask);
13042 Value *Cmp = Builder.CreateICmpEQ(Bitset, Mask);
13043 Result = Builder.CreateAnd(Result, Cmp);
13044 }
13045
13046 return Result;
13047 }
13048
13049 Value *CodeGenFunction::EmitAArch64CpuInit() {
13050 llvm::FunctionType *FTy = llvm::FunctionType::get(VoidTy, false);
13051 llvm::FunctionCallee Func =
13052 CGM.CreateRuntimeFunction(FTy, "init_cpu_features_resolver");
13053 cast<llvm::GlobalValue>(Func.getCallee())->setDSOLocal(true);
13054 cast<llvm::GlobalValue>(Func.getCallee())
13055 ->setDLLStorageClass(llvm::GlobalValue::DefaultStorageClass);
13056 return Builder.CreateCall(Func);
13057 }
13058
13059 Value *CodeGenFunction::EmitX86CpuInit() {
13060 llvm::FunctionType *FTy = llvm::FunctionType::get(VoidTy,
13061 /*Variadic*/ false);
13062 llvm::FunctionCallee Func =
13063 CGM.CreateRuntimeFunction(FTy, "__cpu_indicator_init");
13064 cast<llvm::GlobalValue>(Func.getCallee())->setDSOLocal(true);
13065 cast<llvm::GlobalValue>(Func.getCallee())
13066 ->setDLLStorageClass(llvm::GlobalValue::DefaultStorageClass);
13067 return Builder.CreateCall(Func);
13068 }
13069
13070 llvm::Value *
13071 CodeGenFunction::EmitAArch64CpuSupports(ArrayRef<StringRef> FeaturesStrs) {
13072 uint64_t FeaturesMask = llvm::AArch64::getCpuSupportsMask(FeaturesStrs);
13073 Value *Result = Builder.getTrue();
13074 if (FeaturesMask != 0) {
13075 // Get features from structure in runtime library
13076 // struct {
13077 // unsigned long long features;
13078 // } __aarch64_cpu_features;
13079 llvm::Type *STy = llvm::StructType::get(Int64Ty);
13080 llvm::Constant *AArch64CPUFeatures =
13081 CGM.CreateRuntimeVariable(STy, "__aarch64_cpu_features");
13082 cast<llvm::GlobalValue>(AArch64CPUFeatures)->setDSOLocal(true);
13083 llvm::Value *CpuFeatures = Builder.CreateGEP(
13084 STy, AArch64CPUFeatures,
13085 {ConstantInt::get(Int32Ty, 0), ConstantInt::get(Int32Ty, 0)});
13086 Value *Features = Builder.CreateAlignedLoad(Int64Ty, CpuFeatures,
13087 CharUnits::fromQuantity(8));
13088 Value *Mask = Builder.getInt64(FeaturesMask);
13089 Value *Bitset = Builder.CreateAnd(Features, Mask);
13090 Value *Cmp = Builder.CreateICmpEQ(Bitset, Mask);
13091 Result = Builder.CreateAnd(Result, Cmp);
13092 }
13093 return Result;
13094 }
13095
13096 Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID,
13097 const CallExpr *E) {
13098 if (BuiltinID == X86::BI__builtin_cpu_is)
13099 return EmitX86CpuIs(E);
13100 if (BuiltinID == X86::BI__builtin_cpu_supports)
13101 return EmitX86CpuSupports(E);
13102 if (BuiltinID == X86::BI__builtin_cpu_init)
13103 return EmitX86CpuInit();
13104
13105 // Handle MSVC intrinsics before argument evaluation to prevent double
13106 // evaluation.
13107 if (std::optional<MSVCIntrin> MsvcIntId = translateX86ToMsvcIntrin(BuiltinID))
13108 return EmitMSVCBuiltinExpr(*MsvcIntId, E);
13109
13110 SmallVector<Value*, 4> Ops;
13111 bool IsMaskFCmp = false;
13112 bool IsConjFMA = false;
13113
13114 // Find out if any arguments are required to be integer constant expressions.
13115 unsigned ICEArguments = 0;
13116 ASTContext::GetBuiltinTypeError Error;
13117 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments);
13118 assert(Error == ASTContext::GE_None && "Should not codegen an error");
13119
13120 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) {
13121 // If this is a normal argument, just emit it as a scalar.
13122 if ((ICEArguments & (1 << i)) == 0) {
13123 Ops.push_back(EmitScalarExpr(E->getArg(i)));
13124 continue;
13125 }
13126
13127 // If this is required to be a constant, constant fold it so that we know
13128 // that the generated intrinsic gets a ConstantInt.
13129 Ops.push_back(llvm::ConstantInt::get(
13130 getLLVMContext(), *E->getArg(i)->getIntegerConstantExpr(getContext())));
13131 }
13132
13133 // These exist so that the builtin that takes an immediate can be bounds
13134 // checked by clang to avoid passing bad immediates to the backend. Since
13135 // AVX has a larger immediate than SSE we would need separate builtins to
13136 // do the different bounds checking. Rather than create a clang specific
13137 // SSE only builtin, this implements eight separate builtins to match gcc
13138 // implementation.
13139 auto getCmpIntrinsicCall = [this, &Ops](Intrinsic::ID ID, unsigned Imm) {
13140 Ops.push_back(llvm::ConstantInt::get(Int8Ty, Imm));
13141 llvm::Function *F = CGM.getIntrinsic(ID);
13142 return Builder.CreateCall(F, Ops);
13143 };
13144
13145 // For the vector forms of FP comparisons, translate the builtins directly to
13146 // IR.
13147 // TODO: The builtins could be removed if the SSE header files used vector
13148 // extension comparisons directly (vector ordered/unordered may need
13149 // additional support via __builtin_isnan()).
13150 auto getVectorFCmpIR = [this, &Ops, E](CmpInst::Predicate Pred,
13151 bool IsSignaling) {
13152 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E);
13153 Value *Cmp;
13154 if (IsSignaling)
13155 Cmp = Builder.CreateFCmpS(Pred, Ops[0], Ops[1]);
13156 else
13157 Cmp = Builder.CreateFCmp(Pred, Ops[0], Ops[1]);
13158 llvm::VectorType *FPVecTy = cast<llvm::VectorType>(Ops[0]->getType());
13159 llvm::VectorType *IntVecTy = llvm::VectorType::getInteger(FPVecTy);
13160 Value *Sext = Builder.CreateSExt(Cmp, IntVecTy);
13161 return Builder.CreateBitCast(Sext, FPVecTy);
13162 };
13163
13164 switch (BuiltinID) {
13165 default: return nullptr;
13166 case X86::BI_mm_prefetch: {
13167 Value *Address = Ops[0];
13168 ConstantInt *C = cast<ConstantInt>(Ops[1]);
13169 Value *RW = ConstantInt::get(Int32Ty, (C->getZExtValue() >> 2) & 0x1);
13170 Value *Locality = ConstantInt::get(Int32Ty, C->getZExtValue() & 0x3);
13171 Value *Data = ConstantInt::get(Int32Ty, 1);
13172 Function *F = CGM.getIntrinsic(Intrinsic::prefetch, Address->getType());
13173 return Builder.CreateCall(F, {Address, RW, Locality, Data});
13174 }
13175 case X86::BI_mm_clflush: {
13176 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse2_clflush),
13177 Ops[0]);
13178 }
13179 case X86::BI_mm_lfence: {
13180 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse2_lfence));
13181 }
13182 case X86::BI_mm_mfence: {
13183 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse2_mfence));
13184 }
13185 case X86::BI_mm_sfence: {
13186 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_sfence));
13187 }
13188 case X86::BI_mm_pause: {
13189 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse2_pause));
13190 }
13191 case X86::BI__rdtsc: {
13192 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_rdtsc));
13193 }
13194 case X86::BI__builtin_ia32_rdtscp: {
13195 Value *Call = Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_rdtscp));
13196 Builder.CreateDefaultAlignedStore(Builder.CreateExtractValue(Call, 1),
13197 Ops[0]);
13198 return Builder.CreateExtractValue(Call, 0);
13199 }
13200 case X86::BI__builtin_ia32_lzcnt_u16:
13201 case X86::BI__builtin_ia32_lzcnt_u32:
13202 case X86::BI__builtin_ia32_lzcnt_u64: {
13203 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, Ops[0]->getType());
13204 return Builder.CreateCall(F, {Ops[0], Builder.getInt1(false)});
13205 }
13206 case X86::BI__builtin_ia32_tzcnt_u16:
13207 case X86::BI__builtin_ia32_tzcnt_u32:
13208 case X86::BI__builtin_ia32_tzcnt_u64: {
13209 Function *F = CGM.getIntrinsic(Intrinsic::cttz, Ops[0]->getType());
13210 return Builder.CreateCall(F, {Ops[0], Builder.getInt1(false)});
13211 }
13212 case X86::BI__builtin_ia32_undef128:
13213 case X86::BI__builtin_ia32_undef256:
13214 case X86::BI__builtin_ia32_undef512:
13215 // The x86 definition of "undef" is not the same as the LLVM definition
13216 // (PR32176). We leave optimizing away an unnecessary zero constant to the
13217 // IR optimizer and backend.
13218 // TODO: If we had a "freeze" IR instruction to generate a fixed undef
13219 // value, we should use that here instead of a zero.
13220 return llvm::Constant::getNullValue(ConvertType(E->getType()));
13221 case X86::BI__builtin_ia32_vec_init_v8qi:
13222 case X86::BI__builtin_ia32_vec_init_v4hi:
13223 case X86::BI__builtin_ia32_vec_init_v2si:
13224 return Builder.CreateBitCast(BuildVector(Ops),
13225 llvm::Type::getX86_MMXTy(getLLVMContext()));
13226 case X86::BI__builtin_ia32_vec_ext_v2si:
13227 case X86::BI__builtin_ia32_vec_ext_v16qi:
13228 case X86::BI__builtin_ia32_vec_ext_v8hi:
13229 case X86::BI__builtin_ia32_vec_ext_v4si:
13230 case X86::BI__builtin_ia32_vec_ext_v4sf:
13231 case X86::BI__builtin_ia32_vec_ext_v2di:
13232 case X86::BI__builtin_ia32_vec_ext_v32qi:
13233 case X86::BI__builtin_ia32_vec_ext_v16hi:
13234 case X86::BI__builtin_ia32_vec_ext_v8si:
13235 case X86::BI__builtin_ia32_vec_ext_v4di: {
13236 unsigned NumElts =
13237 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements();
13238 uint64_t Index = cast<ConstantInt>(Ops[1])->getZExtValue();
13239 Index &= NumElts - 1;
13240 // These builtins exist so we can ensure the index is an ICE and in range.
13241 // Otherwise we could just do this in the header file.
13242 return Builder.CreateExtractElement(Ops[0], Index);
13243 }
13244 case X86::BI__builtin_ia32_vec_set_v16qi:
13245 case X86::BI__builtin_ia32_vec_set_v8hi:
13246 case X86::BI__builtin_ia32_vec_set_v4si:
13247 case X86::BI__builtin_ia32_vec_set_v2di:
13248 case X86::BI__builtin_ia32_vec_set_v32qi:
13249 case X86::BI__builtin_ia32_vec_set_v16hi:
13250 case X86::BI__builtin_ia32_vec_set_v8si:
13251 case X86::BI__builtin_ia32_vec_set_v4di: {
13252 unsigned NumElts =
13253 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements();
13254 unsigned Index = cast<ConstantInt>(Ops[2])->getZExtValue();
13255 Index &= NumElts - 1;
13256 // These builtins exist so we can ensure the index is an ICE and in range.
13257 // Otherwise we could just do this in the header file.
13258 return Builder.CreateInsertElement(Ops[0], Ops[1], Index);
13259 }
13260 case X86::BI_mm_setcsr:
13261 case X86::BI__builtin_ia32_ldmxcsr: {
13262 Address Tmp = CreateMemTemp(E->getArg(0)->getType());
13263 Builder.CreateStore(Ops[0], Tmp);
13264 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr),
13265 Builder.CreateBitCast(Tmp.getPointer(), Int8PtrTy));
13266 }
13267 case X86::BI_mm_getcsr:
13268 case X86::BI__builtin_ia32_stmxcsr: {
13269 Address Tmp = CreateMemTemp(E->getType());
13270 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr),
13271 Builder.CreateBitCast(Tmp.getPointer(), Int8PtrTy));
13272 return Builder.CreateLoad(Tmp, "stmxcsr");
13273 }
13274 case X86::BI__builtin_ia32_xsave:
13275 case X86::BI__builtin_ia32_xsave64:
13276 case X86::BI__builtin_ia32_xrstor:
13277 case X86::BI__builtin_ia32_xrstor64:
13278 case X86::BI__builtin_ia32_xsaveopt:
13279 case X86::BI__builtin_ia32_xsaveopt64:
13280 case X86::BI__builtin_ia32_xrstors:
13281 case X86::BI__builtin_ia32_xrstors64:
13282 case X86::BI__builtin_ia32_xsavec:
13283 case X86::BI__builtin_ia32_xsavec64:
13284 case X86::BI__builtin_ia32_xsaves:
13285 case X86::BI__builtin_ia32_xsaves64:
13286 case X86::BI__builtin_ia32_xsetbv:
13287 case X86::BI_xsetbv: {
13288 Intrinsic::ID ID;
13289 #define INTRINSIC_X86_XSAVE_ID(NAME) \
13290 case X86::BI__builtin_ia32_##NAME: \
13291 ID = Intrinsic::x86_##NAME; \
13292 break
13293 switch (BuiltinID) {
13294 default: llvm_unreachable("Unsupported intrinsic!");
13295 INTRINSIC_X86_XSAVE_ID(xsave);
13296 INTRINSIC_X86_XSAVE_ID(xsave64);
13297 INTRINSIC_X86_XSAVE_ID(xrstor);
13298 INTRINSIC_X86_XSAVE_ID(xrstor64);
13299 INTRINSIC_X86_XSAVE_ID(xsaveopt);
13300 INTRINSIC_X86_XSAVE_ID(xsaveopt64);
13301 INTRINSIC_X86_XSAVE_ID(xrstors);
13302 INTRINSIC_X86_XSAVE_ID(xrstors64);
13303 INTRINSIC_X86_XSAVE_ID(xsavec);
13304 INTRINSIC_X86_XSAVE_ID(xsavec64);
13305 INTRINSIC_X86_XSAVE_ID(xsaves);
13306 INTRINSIC_X86_XSAVE_ID(xsaves64);
13307 INTRINSIC_X86_XSAVE_ID(xsetbv);
13308 case X86::BI_xsetbv:
13309 ID = Intrinsic::x86_xsetbv;
13310 break;
13311 }
13312 #undef INTRINSIC_X86_XSAVE_ID
13313 Value *Mhi = Builder.CreateTrunc(
13314 Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, 32)), Int32Ty);
13315 Value *Mlo = Builder.CreateTrunc(Ops[1], Int32Ty);
13316 Ops[1] = Mhi;
13317 Ops.push_back(Mlo);
13318 return Builder.CreateCall(CGM.getIntrinsic(ID), Ops);
13319 }
13320 case X86::BI__builtin_ia32_xgetbv:
13321 case X86::BI_xgetbv:
13322 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_xgetbv), Ops);
13323 case X86::BI__builtin_ia32_storedqudi128_mask:
13324 case X86::BI__builtin_ia32_storedqusi128_mask:
13325 case X86::BI__builtin_ia32_storedquhi128_mask:
13326 case X86::BI__builtin_ia32_storedquqi128_mask:
13327 case X86::BI__builtin_ia32_storeupd128_mask:
13328 case X86::BI__builtin_ia32_storeups128_mask:
13329 case X86::BI__builtin_ia32_storedqudi256_mask:
13330 case X86::BI__builtin_ia32_storedqusi256_mask:
13331 case X86::BI__builtin_ia32_storedquhi256_mask:
13332 case X86::BI__builtin_ia32_storedquqi256_mask:
13333 case X86::BI__builtin_ia32_storeupd256_mask:
13334 case X86::BI__builtin_ia32_storeups256_mask:
13335 case X86::BI__builtin_ia32_storedqudi512_mask:
13336 case X86::BI__builtin_ia32_storedqusi512_mask:
13337 case X86::BI__builtin_ia32_storedquhi512_mask:
13338 case X86::BI__builtin_ia32_storedquqi512_mask:
13339 case X86::BI__builtin_ia32_storeupd512_mask:
13340 case X86::BI__builtin_ia32_storeups512_mask:
13341 return EmitX86MaskedStore(*this, Ops, Align(1));
13342
13343 case X86::BI__builtin_ia32_storesh128_mask:
13344 case X86::BI__builtin_ia32_storess128_mask:
13345 case X86::BI__builtin_ia32_storesd128_mask:
13346 return EmitX86MaskedStore(*this, Ops, Align(1));
13347
13348 case X86::BI__builtin_ia32_vpopcntb_128:
13349 case X86::BI__builtin_ia32_vpopcntd_128:
13350 case X86::BI__builtin_ia32_vpopcntq_128:
13351 case X86::BI__builtin_ia32_vpopcntw_128:
13352 case X86::BI__builtin_ia32_vpopcntb_256:
13353 case X86::BI__builtin_ia32_vpopcntd_256:
13354 case X86::BI__builtin_ia32_vpopcntq_256:
13355 case X86::BI__builtin_ia32_vpopcntw_256:
13356 case X86::BI__builtin_ia32_vpopcntb_512:
13357 case X86::BI__builtin_ia32_vpopcntd_512:
13358 case X86::BI__builtin_ia32_vpopcntq_512:
13359 case X86::BI__builtin_ia32_vpopcntw_512: {
13360 llvm::Type *ResultType = ConvertType(E->getType());
13361 llvm::Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ResultType);
13362 return Builder.CreateCall(F, Ops);
13363 }
13364 case X86::BI__builtin_ia32_cvtmask2b128:
13365 case X86::BI__builtin_ia32_cvtmask2b256:
13366 case X86::BI__builtin_ia32_cvtmask2b512:
13367 case X86::BI__builtin_ia32_cvtmask2w128:
13368 case X86::BI__builtin_ia32_cvtmask2w256:
13369 case X86::BI__builtin_ia32_cvtmask2w512:
13370 case X86::BI__builtin_ia32_cvtmask2d128:
13371 case X86::BI__builtin_ia32_cvtmask2d256:
13372 case X86::BI__builtin_ia32_cvtmask2d512:
13373 case X86::BI__builtin_ia32_cvtmask2q128:
13374 case X86::BI__builtin_ia32_cvtmask2q256:
13375 case X86::BI__builtin_ia32_cvtmask2q512:
13376 return EmitX86SExtMask(*this, Ops[0], ConvertType(E->getType()));
13377
13378 case X86::BI__builtin_ia32_cvtb2mask128:
13379 case X86::BI__builtin_ia32_cvtb2mask256:
13380 case X86::BI__builtin_ia32_cvtb2mask512:
13381 case X86::BI__builtin_ia32_cvtw2mask128:
13382 case X86::BI__builtin_ia32_cvtw2mask256:
13383 case X86::BI__builtin_ia32_cvtw2mask512:
13384 case X86::BI__builtin_ia32_cvtd2mask128:
13385 case X86::BI__builtin_ia32_cvtd2mask256:
13386 case X86::BI__builtin_ia32_cvtd2mask512:
13387 case X86::BI__builtin_ia32_cvtq2mask128:
13388 case X86::BI__builtin_ia32_cvtq2mask256:
13389 case X86::BI__builtin_ia32_cvtq2mask512:
13390 return EmitX86ConvertToMask(*this, Ops[0]);
13391
13392 case X86::BI__builtin_ia32_cvtdq2ps512_mask:
13393 case X86::BI__builtin_ia32_cvtqq2ps512_mask:
13394 case X86::BI__builtin_ia32_cvtqq2pd512_mask:
13395 case X86::BI__builtin_ia32_vcvtw2ph512_mask:
13396 case X86::BI__builtin_ia32_vcvtdq2ph512_mask:
13397 case X86::BI__builtin_ia32_vcvtqq2ph512_mask:
13398 return EmitX86ConvertIntToFp(*this, E, Ops, /*IsSigned*/ true);
13399 case X86::BI__builtin_ia32_cvtudq2ps512_mask:
13400 case X86::BI__builtin_ia32_cvtuqq2ps512_mask:
13401 case X86::BI__builtin_ia32_cvtuqq2pd512_mask:
13402 case X86::BI__builtin_ia32_vcvtuw2ph512_mask:
13403 case X86::BI__builtin_ia32_vcvtudq2ph512_mask:
13404 case X86::BI__builtin_ia32_vcvtuqq2ph512_mask:
13405 return EmitX86ConvertIntToFp(*this, E, Ops, /*IsSigned*/ false);
13406
13407 case X86::BI__builtin_ia32_vfmaddss3:
13408 case X86::BI__builtin_ia32_vfmaddsd3:
13409 case X86::BI__builtin_ia32_vfmaddsh3_mask:
13410 case X86::BI__builtin_ia32_vfmaddss3_mask:
13411 case X86::BI__builtin_ia32_vfmaddsd3_mask:
13412 return EmitScalarFMAExpr(*this, E, Ops, Ops[0]);
13413 case X86::BI__builtin_ia32_vfmaddss:
13414 case X86::BI__builtin_ia32_vfmaddsd:
13415 return EmitScalarFMAExpr(*this, E, Ops,
13416 Constant::getNullValue(Ops[0]->getType()));
13417 case X86::BI__builtin_ia32_vfmaddsh3_maskz:
13418 case X86::BI__builtin_ia32_vfmaddss3_maskz:
13419 case X86::BI__builtin_ia32_vfmaddsd3_maskz:
13420 return EmitScalarFMAExpr(*this, E, Ops, Ops[0], /*ZeroMask*/ true);
13421 case X86::BI__builtin_ia32_vfmaddsh3_mask3:
13422 case X86::BI__builtin_ia32_vfmaddss3_mask3:
13423 case X86::BI__builtin_ia32_vfmaddsd3_mask3:
13424 return EmitScalarFMAExpr(*this, E, Ops, Ops[2], /*ZeroMask*/ false, 2);
13425 case X86::BI__builtin_ia32_vfmsubsh3_mask3:
13426 case X86::BI__builtin_ia32_vfmsubss3_mask3:
13427 case X86::BI__builtin_ia32_vfmsubsd3_mask3:
13428 return EmitScalarFMAExpr(*this, E, Ops, Ops[2], /*ZeroMask*/ false, 2,
13429 /*NegAcc*/ true);
13430 case X86::BI__builtin_ia32_vfmaddph:
13431 case X86::BI__builtin_ia32_vfmaddps:
13432 case X86::BI__builtin_ia32_vfmaddpd:
13433 case X86::BI__builtin_ia32_vfmaddph256:
13434 case X86::BI__builtin_ia32_vfmaddps256:
13435 case X86::BI__builtin_ia32_vfmaddpd256:
13436 case X86::BI__builtin_ia32_vfmaddph512_mask:
13437 case X86::BI__builtin_ia32_vfmaddph512_maskz:
13438 case X86::BI__builtin_ia32_vfmaddph512_mask3:
13439 case X86::BI__builtin_ia32_vfmaddps512_mask:
13440 case X86::BI__builtin_ia32_vfmaddps512_maskz:
13441 case X86::BI__builtin_ia32_vfmaddps512_mask3:
13442 case X86::BI__builtin_ia32_vfmsubps512_mask3:
13443 case X86::BI__builtin_ia32_vfmaddpd512_mask:
13444 case X86::BI__builtin_ia32_vfmaddpd512_maskz:
13445 case X86::BI__builtin_ia32_vfmaddpd512_mask3:
13446 case X86::BI__builtin_ia32_vfmsubpd512_mask3:
13447 case X86::BI__builtin_ia32_vfmsubph512_mask3:
13448 return EmitX86FMAExpr(*this, E, Ops, BuiltinID, /*IsAddSub*/ false);
13449 case X86::BI__builtin_ia32_vfmaddsubph512_mask:
13450 case X86::BI__builtin_ia32_vfmaddsubph512_maskz:
13451 case X86::BI__builtin_ia32_vfmaddsubph512_mask3:
13452 case X86::BI__builtin_ia32_vfmsubaddph512_mask3:
13453 case X86::BI__builtin_ia32_vfmaddsubps512_mask:
13454 case X86::BI__builtin_ia32_vfmaddsubps512_maskz:
13455 case X86::BI__builtin_ia32_vfmaddsubps512_mask3:
13456 case X86::BI__builtin_ia32_vfmsubaddps512_mask3:
13457 case X86::BI__builtin_ia32_vfmaddsubpd512_mask:
13458 case X86::BI__builtin_ia32_vfmaddsubpd512_maskz:
13459 case X86::BI__builtin_ia32_vfmaddsubpd512_mask3:
13460 case X86::BI__builtin_ia32_vfmsubaddpd512_mask3:
13461 return EmitX86FMAExpr(*this, E, Ops, BuiltinID, /*IsAddSub*/ true);
13462
13463 case X86::BI__builtin_ia32_movdqa32store128_mask:
13464 case X86::BI__builtin_ia32_movdqa64store128_mask:
13465 case X86::BI__builtin_ia32_storeaps128_mask:
13466 case X86::BI__builtin_ia32_storeapd128_mask:
13467 case X86::BI__builtin_ia32_movdqa32store256_mask:
13468 case X86::BI__builtin_ia32_movdqa64store256_mask:
13469 case X86::BI__builtin_ia32_storeaps256_mask:
13470 case X86::BI__builtin_ia32_storeapd256_mask:
13471 case X86::BI__builtin_ia32_movdqa32store512_mask:
13472 case X86::BI__builtin_ia32_movdqa64store512_mask:
13473 case X86::BI__builtin_ia32_storeaps512_mask:
13474 case X86::BI__builtin_ia32_storeapd512_mask:
13475 return EmitX86MaskedStore(
13476 *this, Ops,
13477 getContext().getTypeAlignInChars(E->getArg(1)->getType()).getAsAlign());
13478
13479 case X86::BI__builtin_ia32_loadups128_mask:
13480 case X86::BI__builtin_ia32_loadups256_mask:
13481 case X86::BI__builtin_ia32_loadups512_mask:
13482 case X86::BI__builtin_ia32_loadupd128_mask:
13483 case X86::BI__builtin_ia32_loadupd256_mask:
13484 case X86::BI__builtin_ia32_loadupd512_mask:
13485 case X86::BI__builtin_ia32_loaddquqi128_mask:
13486 case X86::BI__builtin_ia32_loaddquqi256_mask:
13487 case X86::BI__builtin_ia32_loaddquqi512_mask:
13488 case X86::BI__builtin_ia32_loaddquhi128_mask:
13489 case X86::BI__builtin_ia32_loaddquhi256_mask:
13490 case X86::BI__builtin_ia32_loaddquhi512_mask:
13491 case X86::BI__builtin_ia32_loaddqusi128_mask:
13492 case X86::BI__builtin_ia32_loaddqusi256_mask:
13493 case X86::BI__builtin_ia32_loaddqusi512_mask:
13494 case X86::BI__builtin_ia32_loaddqudi128_mask:
13495 case X86::BI__builtin_ia32_loaddqudi256_mask:
13496 case X86::BI__builtin_ia32_loaddqudi512_mask:
13497 return EmitX86MaskedLoad(*this, Ops, Align(1));
13498
13499 case X86::BI__builtin_ia32_loadsh128_mask:
13500 case X86::BI__builtin_ia32_loadss128_mask:
13501 case X86::BI__builtin_ia32_loadsd128_mask:
13502 return EmitX86MaskedLoad(*this, Ops, Align(1));
13503
13504 case X86::BI__builtin_ia32_loadaps128_mask:
13505 case X86::BI__builtin_ia32_loadaps256_mask:
13506 case X86::BI__builtin_ia32_loadaps512_mask:
13507 case X86::BI__builtin_ia32_loadapd128_mask:
13508 case X86::BI__builtin_ia32_loadapd256_mask:
13509 case X86::BI__builtin_ia32_loadapd512_mask:
13510 case X86::BI__builtin_ia32_movdqa32load128_mask:
13511 case X86::BI__builtin_ia32_movdqa32load256_mask:
13512 case X86::BI__builtin_ia32_movdqa32load512_mask:
13513 case X86::BI__builtin_ia32_movdqa64load128_mask:
13514 case X86::BI__builtin_ia32_movdqa64load256_mask:
13515 case X86::BI__builtin_ia32_movdqa64load512_mask:
13516 return EmitX86MaskedLoad(
13517 *this, Ops,
13518 getContext().getTypeAlignInChars(E->getArg(1)->getType()).getAsAlign());
13519
13520 case X86::BI__builtin_ia32_expandloaddf128_mask:
13521 case X86::BI__builtin_ia32_expandloaddf256_mask:
13522 case X86::BI__builtin_ia32_expandloaddf512_mask:
13523 case X86::BI__builtin_ia32_expandloadsf128_mask:
13524 case X86::BI__builtin_ia32_expandloadsf256_mask:
13525 case X86::BI__builtin_ia32_expandloadsf512_mask:
13526 case X86::BI__builtin_ia32_expandloaddi128_mask:
13527 case X86::BI__builtin_ia32_expandloaddi256_mask:
13528 case X86::BI__builtin_ia32_expandloaddi512_mask:
13529 case X86::BI__builtin_ia32_expandloadsi128_mask:
13530 case X86::BI__builtin_ia32_expandloadsi256_mask:
13531 case X86::BI__builtin_ia32_expandloadsi512_mask:
13532 case X86::BI__builtin_ia32_expandloadhi128_mask:
13533 case X86::BI__builtin_ia32_expandloadhi256_mask:
13534 case X86::BI__builtin_ia32_expandloadhi512_mask:
13535 case X86::BI__builtin_ia32_expandloadqi128_mask:
13536 case X86::BI__builtin_ia32_expandloadqi256_mask:
13537 case X86::BI__builtin_ia32_expandloadqi512_mask:
13538 return EmitX86ExpandLoad(*this, Ops);
13539
13540 case X86::BI__builtin_ia32_compressstoredf128_mask:
13541 case X86::BI__builtin_ia32_compressstoredf256_mask:
13542 case X86::BI__builtin_ia32_compressstoredf512_mask:
13543 case X86::BI__builtin_ia32_compressstoresf128_mask:
13544 case X86::BI__builtin_ia32_compressstoresf256_mask:
13545 case X86::BI__builtin_ia32_compressstoresf512_mask:
13546 case X86::BI__builtin_ia32_compressstoredi128_mask:
13547 case X86::BI__builtin_ia32_compressstoredi256_mask:
13548 case X86::BI__builtin_ia32_compressstoredi512_mask:
13549 case X86::BI__builtin_ia32_compressstoresi128_mask:
13550 case X86::BI__builtin_ia32_compressstoresi256_mask:
13551 case X86::BI__builtin_ia32_compressstoresi512_mask:
13552 case X86::BI__builtin_ia32_compressstorehi128_mask:
13553 case X86::BI__builtin_ia32_compressstorehi256_mask:
13554 case X86::BI__builtin_ia32_compressstorehi512_mask:
13555 case X86::BI__builtin_ia32_compressstoreqi128_mask:
13556 case X86::BI__builtin_ia32_compressstoreqi256_mask:
13557 case X86::BI__builtin_ia32_compressstoreqi512_mask:
13558 return EmitX86CompressStore(*this, Ops);
13559
13560 case X86::BI__builtin_ia32_expanddf128_mask:
13561 case X86::BI__builtin_ia32_expanddf256_mask:
13562 case X86::BI__builtin_ia32_expanddf512_mask:
13563 case X86::BI__builtin_ia32_expandsf128_mask:
13564 case X86::BI__builtin_ia32_expandsf256_mask:
13565 case X86::BI__builtin_ia32_expandsf512_mask:
13566 case X86::BI__builtin_ia32_expanddi128_mask:
13567 case X86::BI__builtin_ia32_expanddi256_mask:
13568 case X86::BI__builtin_ia32_expanddi512_mask:
13569 case X86::BI__builtin_ia32_expandsi128_mask:
13570 case X86::BI__builtin_ia32_expandsi256_mask:
13571 case X86::BI__builtin_ia32_expandsi512_mask:
13572 case X86::BI__builtin_ia32_expandhi128_mask:
13573 case X86::BI__builtin_ia32_expandhi256_mask:
13574 case X86::BI__builtin_ia32_expandhi512_mask:
13575 case X86::BI__builtin_ia32_expandqi128_mask:
13576 case X86::BI__builtin_ia32_expandqi256_mask:
13577 case X86::BI__builtin_ia32_expandqi512_mask:
13578 return EmitX86CompressExpand(*this, Ops, /*IsCompress*/false);
13579
13580 case X86::BI__builtin_ia32_compressdf128_mask:
13581 case X86::BI__builtin_ia32_compressdf256_mask:
13582 case X86::BI__builtin_ia32_compressdf512_mask:
13583 case X86::BI__builtin_ia32_compresssf128_mask:
13584 case X86::BI__builtin_ia32_compresssf256_mask:
13585 case X86::BI__builtin_ia32_compresssf512_mask:
13586 case X86::BI__builtin_ia32_compressdi128_mask:
13587 case X86::BI__builtin_ia32_compressdi256_mask:
13588 case X86::BI__builtin_ia32_compressdi512_mask:
13589 case X86::BI__builtin_ia32_compresssi128_mask:
13590 case X86::BI__builtin_ia32_compresssi256_mask:
13591 case X86::BI__builtin_ia32_compresssi512_mask:
13592 case X86::BI__builtin_ia32_compresshi128_mask:
13593 case X86::BI__builtin_ia32_compresshi256_mask:
13594 case X86::BI__builtin_ia32_compresshi512_mask:
13595 case X86::BI__builtin_ia32_compressqi128_mask:
13596 case X86::BI__builtin_ia32_compressqi256_mask:
13597 case X86::BI__builtin_ia32_compressqi512_mask:
13598 return EmitX86CompressExpand(*this, Ops, /*IsCompress*/true);
13599
13600 case X86::BI__builtin_ia32_gather3div2df:
13601 case X86::BI__builtin_ia32_gather3div2di:
13602 case X86::BI__builtin_ia32_gather3div4df:
13603 case X86::BI__builtin_ia32_gather3div4di:
13604 case X86::BI__builtin_ia32_gather3div4sf:
13605 case X86::BI__builtin_ia32_gather3div4si:
13606 case X86::BI__builtin_ia32_gather3div8sf:
13607 case X86::BI__builtin_ia32_gather3div8si:
13608 case X86::BI__builtin_ia32_gather3siv2df:
13609 case X86::BI__builtin_ia32_gather3siv2di:
13610 case X86::BI__builtin_ia32_gather3siv4df:
13611 case X86::BI__builtin_ia32_gather3siv4di:
13612 case X86::BI__builtin_ia32_gather3siv4sf:
13613 case X86::BI__builtin_ia32_gather3siv4si:
13614 case X86::BI__builtin_ia32_gather3siv8sf:
13615 case X86::BI__builtin_ia32_gather3siv8si:
13616 case X86::BI__builtin_ia32_gathersiv8df:
13617 case X86::BI__builtin_ia32_gathersiv16sf:
13618 case X86::BI__builtin_ia32_gatherdiv8df:
13619 case X86::BI__builtin_ia32_gatherdiv16sf:
13620 case X86::BI__builtin_ia32_gathersiv8di:
13621 case X86::BI__builtin_ia32_gathersiv16si:
13622 case X86::BI__builtin_ia32_gatherdiv8di:
13623 case X86::BI__builtin_ia32_gatherdiv16si: {
13624 Intrinsic::ID IID;
13625 switch (BuiltinID) {
13626 default: llvm_unreachable("Unexpected builtin");
13627 case X86::BI__builtin_ia32_gather3div2df:
13628 IID = Intrinsic::x86_avx512_mask_gather3div2_df;
13629 break;
13630 case X86::BI__builtin_ia32_gather3div2di:
13631 IID = Intrinsic::x86_avx512_mask_gather3div2_di;
13632 break;
13633 case X86::BI__builtin_ia32_gather3div4df:
13634 IID = Intrinsic::x86_avx512_mask_gather3div4_df;
13635 break;
13636 case X86::BI__builtin_ia32_gather3div4di:
13637 IID = Intrinsic::x86_avx512_mask_gather3div4_di;
13638 break;
13639 case X86::BI__builtin_ia32_gather3div4sf:
13640 IID = Intrinsic::x86_avx512_mask_gather3div4_sf;
13641 break;
13642 case X86::BI__builtin_ia32_gather3div4si:
13643 IID = Intrinsic::x86_avx512_mask_gather3div4_si;
13644 break;
13645 case X86::BI__builtin_ia32_gather3div8sf:
13646 IID = Intrinsic::x86_avx512_mask_gather3div8_sf;
13647 break;
13648 case X86::BI__builtin_ia32_gather3div8si:
13649 IID = Intrinsic::x86_avx512_mask_gather3div8_si;
13650 break;
13651 case X86::BI__builtin_ia32_gather3siv2df:
13652 IID = Intrinsic::x86_avx512_mask_gather3siv2_df;
13653 break;
13654 case X86::BI__builtin_ia32_gather3siv2di:
13655 IID = Intrinsic::x86_avx512_mask_gather3siv2_di;
13656 break;
13657 case X86::BI__builtin_ia32_gather3siv4df:
13658 IID = Intrinsic::x86_avx512_mask_gather3siv4_df;
13659 break;
13660 case X86::BI__builtin_ia32_gather3siv4di:
13661 IID = Intrinsic::x86_avx512_mask_gather3siv4_di;
13662 break;
13663 case X86::BI__builtin_ia32_gather3siv4sf:
13664 IID = Intrinsic::x86_avx512_mask_gather3siv4_sf;
13665 break;
13666 case X86::BI__builtin_ia32_gather3siv4si:
13667 IID = Intrinsic::x86_avx512_mask_gather3siv4_si;
13668 break;
13669 case X86::BI__builtin_ia32_gather3siv8sf:
13670 IID = Intrinsic::x86_avx512_mask_gather3siv8_sf;
13671 break;
13672 case X86::BI__builtin_ia32_gather3siv8si:
13673 IID = Intrinsic::x86_avx512_mask_gather3siv8_si;
13674 break;
13675 case X86::BI__builtin_ia32_gathersiv8df:
13676 IID = Intrinsic::x86_avx512_mask_gather_dpd_512;
13677 break;
13678 case X86::BI__builtin_ia32_gathersiv16sf:
13679 IID = Intrinsic::x86_avx512_mask_gather_dps_512;
13680 break;
13681 case X86::BI__builtin_ia32_gatherdiv8df:
13682 IID = Intrinsic::x86_avx512_mask_gather_qpd_512;
13683 break;
13684 case X86::BI__builtin_ia32_gatherdiv16sf:
13685 IID = Intrinsic::x86_avx512_mask_gather_qps_512;
13686 break;
13687 case X86::BI__builtin_ia32_gathersiv8di:
13688 IID = Intrinsic::x86_avx512_mask_gather_dpq_512;
13689 break;
13690 case X86::BI__builtin_ia32_gathersiv16si:
13691 IID = Intrinsic::x86_avx512_mask_gather_dpi_512;
13692 break;
13693 case X86::BI__builtin_ia32_gatherdiv8di:
13694 IID = Intrinsic::x86_avx512_mask_gather_qpq_512;
13695 break;
13696 case X86::BI__builtin_ia32_gatherdiv16si:
13697 IID = Intrinsic::x86_avx512_mask_gather_qpi_512;
13698 break;
13699 }
13700
13701 unsigned MinElts = std::min(
13702 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements(),
13703 cast<llvm::FixedVectorType>(Ops[2]->getType())->getNumElements());
13704 Ops[3] = getMaskVecValue(*this, Ops[3], MinElts);
13705 Function *Intr = CGM.getIntrinsic(IID);
13706 return Builder.CreateCall(Intr, Ops);
13707 }
13708
13709 case X86::BI__builtin_ia32_scattersiv8df:
13710 case X86::BI__builtin_ia32_scattersiv16sf:
13711 case X86::BI__builtin_ia32_scatterdiv8df:
13712 case X86::BI__builtin_ia32_scatterdiv16sf:
13713 case X86::BI__builtin_ia32_scattersiv8di:
13714 case X86::BI__builtin_ia32_scattersiv16si:
13715 case X86::BI__builtin_ia32_scatterdiv8di:
13716 case X86::BI__builtin_ia32_scatterdiv16si:
13717 case X86::BI__builtin_ia32_scatterdiv2df:
13718 case X86::BI__builtin_ia32_scatterdiv2di:
13719 case X86::BI__builtin_ia32_scatterdiv4df:
13720 case X86::BI__builtin_ia32_scatterdiv4di:
13721 case X86::BI__builtin_ia32_scatterdiv4sf:
13722 case X86::BI__builtin_ia32_scatterdiv4si:
13723 case X86::BI__builtin_ia32_scatterdiv8sf:
13724 case X86::BI__builtin_ia32_scatterdiv8si:
13725 case X86::BI__builtin_ia32_scattersiv2df:
13726 case X86::BI__builtin_ia32_scattersiv2di:
13727 case X86::BI__builtin_ia32_scattersiv4df:
13728 case X86::BI__builtin_ia32_scattersiv4di:
13729 case X86::BI__builtin_ia32_scattersiv4sf:
13730 case X86::BI__builtin_ia32_scattersiv4si:
13731 case X86::BI__builtin_ia32_scattersiv8sf:
13732 case X86::BI__builtin_ia32_scattersiv8si: {
13733 Intrinsic::ID IID;
13734 switch (BuiltinID) {
13735 default: llvm_unreachable("Unexpected builtin");
13736 case X86::BI__builtin_ia32_scattersiv8df:
13737 IID = Intrinsic::x86_avx512_mask_scatter_dpd_512;
13738 break;
13739 case X86::BI__builtin_ia32_scattersiv16sf:
13740 IID = Intrinsic::x86_avx512_mask_scatter_dps_512;
13741 break;
13742 case X86::BI__builtin_ia32_scatterdiv8df:
13743 IID = Intrinsic::x86_avx512_mask_scatter_qpd_512;
13744 break;
13745 case X86::BI__builtin_ia32_scatterdiv16sf:
13746 IID = Intrinsic::x86_avx512_mask_scatter_qps_512;
13747 break;
13748 case X86::BI__builtin_ia32_scattersiv8di:
13749 IID = Intrinsic::x86_avx512_mask_scatter_dpq_512;
13750 break;
13751 case X86::BI__builtin_ia32_scattersiv16si:
13752 IID = Intrinsic::x86_avx512_mask_scatter_dpi_512;
13753 break;
13754 case X86::BI__builtin_ia32_scatterdiv8di:
13755 IID = Intrinsic::x86_avx512_mask_scatter_qpq_512;
13756 break;
13757 case X86::BI__builtin_ia32_scatterdiv16si:
13758 IID = Intrinsic::x86_avx512_mask_scatter_qpi_512;
13759 break;
13760 case X86::BI__builtin_ia32_scatterdiv2df:
13761 IID = Intrinsic::x86_avx512_mask_scatterdiv2_df;
13762 break;
13763 case X86::BI__builtin_ia32_scatterdiv2di:
13764 IID = Intrinsic::x86_avx512_mask_scatterdiv2_di;
13765 break;
13766 case X86::BI__builtin_ia32_scatterdiv4df:
13767 IID = Intrinsic::x86_avx512_mask_scatterdiv4_df;
13768 break;
13769 case X86::BI__builtin_ia32_scatterdiv4di:
13770 IID = Intrinsic::x86_avx512_mask_scatterdiv4_di;
13771 break;
13772 case X86::BI__builtin_ia32_scatterdiv4sf:
13773 IID = Intrinsic::x86_avx512_mask_scatterdiv4_sf;
13774 break;
13775 case X86::BI__builtin_ia32_scatterdiv4si:
13776 IID = Intrinsic::x86_avx512_mask_scatterdiv4_si;
13777 break;
13778 case X86::BI__builtin_ia32_scatterdiv8sf:
13779 IID = Intrinsic::x86_avx512_mask_scatterdiv8_sf;
13780 break;
13781 case X86::BI__builtin_ia32_scatterdiv8si:
13782 IID = Intrinsic::x86_avx512_mask_scatterdiv8_si;
13783 break;
13784 case X86::BI__builtin_ia32_scattersiv2df:
13785 IID = Intrinsic::x86_avx512_mask_scattersiv2_df;
13786 break;
13787 case X86::BI__builtin_ia32_scattersiv2di:
13788 IID = Intrinsic::x86_avx512_mask_scattersiv2_di;
13789 break;
13790 case X86::BI__builtin_ia32_scattersiv4df:
13791 IID = Intrinsic::x86_avx512_mask_scattersiv4_df;
13792 break;
13793 case X86::BI__builtin_ia32_scattersiv4di:
13794 IID = Intrinsic::x86_avx512_mask_scattersiv4_di;
13795 break;
13796 case X86::BI__builtin_ia32_scattersiv4sf:
13797 IID = Intrinsic::x86_avx512_mask_scattersiv4_sf;
13798 break;
13799 case X86::BI__builtin_ia32_scattersiv4si:
13800 IID = Intrinsic::x86_avx512_mask_scattersiv4_si;
13801 break;
13802 case X86::BI__builtin_ia32_scattersiv8sf:
13803 IID = Intrinsic::x86_avx512_mask_scattersiv8_sf;
13804 break;
13805 case X86::BI__builtin_ia32_scattersiv8si:
13806 IID = Intrinsic::x86_avx512_mask_scattersiv8_si;
13807 break;
13808 }
13809
13810 unsigned MinElts = std::min(
13811 cast<llvm::FixedVectorType>(Ops[2]->getType())->getNumElements(),
13812 cast<llvm::FixedVectorType>(Ops[3]->getType())->getNumElements());
13813 Ops[1] = getMaskVecValue(*this, Ops[1], MinElts);
13814 Function *Intr = CGM.getIntrinsic(IID);
13815 return Builder.CreateCall(Intr, Ops);
13816 }
13817
13818 case X86::BI__builtin_ia32_vextractf128_pd256:
13819 case X86::BI__builtin_ia32_vextractf128_ps256:
13820 case X86::BI__builtin_ia32_vextractf128_si256:
13821 case X86::BI__builtin_ia32_extract128i256:
13822 case X86::BI__builtin_ia32_extractf64x4_mask:
13823 case X86::BI__builtin_ia32_extractf32x4_mask:
13824 case X86::BI__builtin_ia32_extracti64x4_mask:
13825 case X86::BI__builtin_ia32_extracti32x4_mask:
13826 case X86::BI__builtin_ia32_extractf32x8_mask:
13827 case X86::BI__builtin_ia32_extracti32x8_mask:
13828 case X86::BI__builtin_ia32_extractf32x4_256_mask:
13829 case X86::BI__builtin_ia32_extracti32x4_256_mask:
13830 case X86::BI__builtin_ia32_extractf64x2_256_mask:
13831 case X86::BI__builtin_ia32_extracti64x2_256_mask:
13832 case X86::BI__builtin_ia32_extractf64x2_512_mask:
13833 case X86::BI__builtin_ia32_extracti64x2_512_mask: {
13834 auto *DstTy = cast<llvm::FixedVectorType>(ConvertType(E->getType()));
13835 unsigned NumElts = DstTy->getNumElements();
13836 unsigned SrcNumElts =
13837 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements();
13838 unsigned SubVectors = SrcNumElts / NumElts;
13839 unsigned Index = cast<ConstantInt>(Ops[1])->getZExtValue();
13840 assert(llvm::isPowerOf2_32(SubVectors) && "Expected power of 2 subvectors");
13841 Index &= SubVectors - 1; // Remove any extra bits.
13842 Index *= NumElts;
13843
13844 int Indices[16];
13845 for (unsigned i = 0; i != NumElts; ++i)
13846 Indices[i] = i + Index;
13847
13848 Value *Res = Builder.CreateShuffleVector(Ops[0], ArrayRef(Indices, NumElts),
13849 "extract");
13850
13851 if (Ops.size() == 4)
13852 Res = EmitX86Select(*this, Ops[3], Res, Ops[2]);
13853
13854 return Res;
13855 }
13856 case X86::BI__builtin_ia32_vinsertf128_pd256:
13857 case X86::BI__builtin_ia32_vinsertf128_ps256:
13858 case X86::BI__builtin_ia32_vinsertf128_si256:
13859 case X86::BI__builtin_ia32_insert128i256:
13860 case X86::BI__builtin_ia32_insertf64x4:
13861 case X86::BI__builtin_ia32_insertf32x4:
13862 case X86::BI__builtin_ia32_inserti64x4:
13863 case X86::BI__builtin_ia32_inserti32x4:
13864 case X86::BI__builtin_ia32_insertf32x8:
13865 case X86::BI__builtin_ia32_inserti32x8:
13866 case X86::BI__builtin_ia32_insertf32x4_256:
13867 case X86::BI__builtin_ia32_inserti32x4_256:
13868 case X86::BI__builtin_ia32_insertf64x2_256:
13869 case X86::BI__builtin_ia32_inserti64x2_256:
13870 case X86::BI__builtin_ia32_insertf64x2_512:
13871 case X86::BI__builtin_ia32_inserti64x2_512: {
13872 unsigned DstNumElts =
13873 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements();
13874 unsigned SrcNumElts =
13875 cast<llvm::FixedVectorType>(Ops[1]->getType())->getNumElements();
13876 unsigned SubVectors = DstNumElts / SrcNumElts;
13877 unsigned Index = cast<ConstantInt>(Ops[2])->getZExtValue();
13878 assert(llvm::isPowerOf2_32(SubVectors) && "Expected power of 2 subvectors");
13879 Index &= SubVectors - 1; // Remove any extra bits.
13880 Index *= SrcNumElts;
13881
13882 int Indices[16];
13883 for (unsigned i = 0; i != DstNumElts; ++i)
13884 Indices[i] = (i >= SrcNumElts) ? SrcNumElts + (i % SrcNumElts) : i;
13885
13886 Value *Op1 = Builder.CreateShuffleVector(
13887 Ops[1], ArrayRef(Indices, DstNumElts), "widen");
13888
13889 for (unsigned i = 0; i != DstNumElts; ++i) {
13890 if (i >= Index && i < (Index + SrcNumElts))
13891 Indices[i] = (i - Index) + DstNumElts;
13892 else
13893 Indices[i] = i;
13894 }
13895
13896 return Builder.CreateShuffleVector(Ops[0], Op1,
13897 ArrayRef(Indices, DstNumElts), "insert");
13898 }
13899 case X86::BI__builtin_ia32_pmovqd512_mask:
13900 case X86::BI__builtin_ia32_pmovwb512_mask: {
13901 Value *Res = Builder.CreateTrunc(Ops[0], Ops[1]->getType());
13902 return EmitX86Select(*this, Ops[2], Res, Ops[1]);
13903 }
13904 case X86::BI__builtin_ia32_pmovdb512_mask:
13905 case X86::BI__builtin_ia32_pmovdw512_mask:
13906 case X86::BI__builtin_ia32_pmovqw512_mask: {
13907 if (const auto *C = dyn_cast<Constant>(Ops[2]))
13908 if (C->isAllOnesValue())
13909 return Builder.CreateTrunc(Ops[0], Ops[1]->getType());
13910
13911 Intrinsic::ID IID;
13912 switch (BuiltinID) {
13913 default: llvm_unreachable("Unsupported intrinsic!");
13914 case X86::BI__builtin_ia32_pmovdb512_mask:
13915 IID = Intrinsic::x86_avx512_mask_pmov_db_512;
13916 break;
13917 case X86::BI__builtin_ia32_pmovdw512_mask:
13918 IID = Intrinsic::x86_avx512_mask_pmov_dw_512;
13919 break;
13920 case X86::BI__builtin_ia32_pmovqw512_mask:
13921 IID = Intrinsic::x86_avx512_mask_pmov_qw_512;
13922 break;
13923 }
13924
13925 Function *Intr = CGM.getIntrinsic(IID);
13926 return Builder.CreateCall(Intr, Ops);
13927 }
13928 case X86::BI__builtin_ia32_pblendw128:
13929 case X86::BI__builtin_ia32_blendpd:
13930 case X86::BI__builtin_ia32_blendps:
13931 case X86::BI__builtin_ia32_blendpd256:
13932 case X86::BI__builtin_ia32_blendps256:
13933 case X86::BI__builtin_ia32_pblendw256:
13934 case X86::BI__builtin_ia32_pblendd128:
13935 case X86::BI__builtin_ia32_pblendd256: {
13936 unsigned NumElts =
13937 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements();
13938 unsigned Imm = cast<llvm::ConstantInt>(Ops[2])->getZExtValue();
13939
13940 int Indices[16];
13941 // If there are more than 8 elements, the immediate is used twice so make
13942 // sure we handle that.
13943 for (unsigned i = 0; i != NumElts; ++i)
13944 Indices[i] = ((Imm >> (i % 8)) & 0x1) ? NumElts + i : i;
13945
13946 return Builder.CreateShuffleVector(Ops[0], Ops[1],
13947 ArrayRef(Indices, NumElts), "blend");
13948 }
13949 case X86::BI__builtin_ia32_pshuflw:
13950 case X86::BI__builtin_ia32_pshuflw256:
13951 case X86::BI__builtin_ia32_pshuflw512: {
13952 uint32_t Imm = cast<llvm::ConstantInt>(Ops[1])->getZExtValue();
13953 auto *Ty = cast<llvm::FixedVectorType>(Ops[0]->getType());
13954 unsigned NumElts = Ty->getNumElements();
13955
13956 // Splat the 8-bits of immediate 4 times to help the loop wrap around.
13957 Imm = (Imm & 0xff) * 0x01010101;
13958
13959 int Indices[32];
13960 for (unsigned l = 0; l != NumElts; l += 8) {
13961 for (unsigned i = 0; i != 4; ++i) {
13962 Indices[l + i] = l + (Imm & 3);
13963 Imm >>= 2;
13964 }
13965 for (unsigned i = 4; i != 8; ++i)
13966 Indices[l + i] = l + i;
13967 }
13968
13969 return Builder.CreateShuffleVector(Ops[0], ArrayRef(Indices, NumElts),
13970 "pshuflw");
13971 }
13972 case X86::BI__builtin_ia32_pshufhw:
13973 case X86::BI__builtin_ia32_pshufhw256:
13974 case X86::BI__builtin_ia32_pshufhw512: {
13975 uint32_t Imm = cast<llvm::ConstantInt>(Ops[1])->getZExtValue();
13976 auto *Ty = cast<llvm::FixedVectorType>(Ops[0]->getType());
13977 unsigned NumElts = Ty->getNumElements();
13978
13979 // Splat the 8-bits of immediate 4 times to help the loop wrap around.
13980 Imm = (Imm & 0xff) * 0x01010101;
13981
13982 int Indices[32];
13983 for (unsigned l = 0; l != NumElts; l += 8) {
13984 for (unsigned i = 0; i != 4; ++i)
13985 Indices[l + i] = l + i;
13986 for (unsigned i = 4; i != 8; ++i) {
13987 Indices[l + i] = l + 4 + (Imm & 3);
13988 Imm >>= 2;
13989 }
13990 }
13991
13992 return Builder.CreateShuffleVector(Ops[0], ArrayRef(Indices, NumElts),
13993 "pshufhw");
13994 }
13995 case X86::BI__builtin_ia32_pshufd:
13996 case X86::BI__builtin_ia32_pshufd256:
13997 case X86::BI__builtin_ia32_pshufd512:
13998 case X86::BI__builtin_ia32_vpermilpd:
13999 case X86::BI__builtin_ia32_vpermilps:
14000 case X86::BI__builtin_ia32_vpermilpd256:
14001 case X86::BI__builtin_ia32_vpermilps256:
14002 case X86::BI__builtin_ia32_vpermilpd512:
14003 case X86::BI__builtin_ia32_vpermilps512: {
14004 uint32_t Imm = cast<llvm::ConstantInt>(Ops[1])->getZExtValue();
14005 auto *Ty = cast<llvm::FixedVectorType>(Ops[0]->getType());
14006 unsigned NumElts = Ty->getNumElements();
14007 unsigned NumLanes = Ty->getPrimitiveSizeInBits() / 128;
14008 unsigned NumLaneElts = NumElts / NumLanes;
14009
14010 // Splat the 8-bits of immediate 4 times to help the loop wrap around.
14011 Imm = (Imm & 0xff) * 0x01010101;
14012
14013 int Indices[16];
14014 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
14015 for (unsigned i = 0; i != NumLaneElts; ++i) {
14016 Indices[i + l] = (Imm % NumLaneElts) + l;
14017 Imm /= NumLaneElts;
14018 }
14019 }
14020
14021 return Builder.CreateShuffleVector(Ops[0], ArrayRef(Indices, NumElts),
14022 "permil");
14023 }
14024 case X86::BI__builtin_ia32_shufpd:
14025 case X86::BI__builtin_ia32_shufpd256:
14026 case X86::BI__builtin_ia32_shufpd512:
14027 case X86::BI__builtin_ia32_shufps:
14028 case X86::BI__builtin_ia32_shufps256:
14029 case X86::BI__builtin_ia32_shufps512: {
14030 uint32_t Imm = cast<llvm::ConstantInt>(Ops[2])->getZExtValue();
14031 auto *Ty = cast<llvm::FixedVectorType>(Ops[0]->getType());
14032 unsigned NumElts = Ty->getNumElements();
14033 unsigned NumLanes = Ty->getPrimitiveSizeInBits() / 128;
14034 unsigned NumLaneElts = NumElts / NumLanes;
14035
14036 // Splat the 8-bits of immediate 4 times to help the loop wrap around.
14037 Imm = (Imm & 0xff) * 0x01010101;
14038
14039 int Indices[16];
14040 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
14041 for (unsigned i = 0; i != NumLaneElts; ++i) {
14042 unsigned Index = Imm % NumLaneElts;
14043 Imm /= NumLaneElts;
14044 if (i >= (NumLaneElts / 2))
14045 Index += NumElts;
14046 Indices[l + i] = l + Index;
14047 }
14048 }
14049
14050 return Builder.CreateShuffleVector(Ops[0], Ops[1],
14051 ArrayRef(Indices, NumElts), "shufp");
14052 }
14053 case X86::BI__builtin_ia32_permdi256:
14054 case X86::BI__builtin_ia32_permdf256:
14055 case X86::BI__builtin_ia32_permdi512:
14056 case X86::BI__builtin_ia32_permdf512: {
14057 unsigned Imm = cast<llvm::ConstantInt>(Ops[1])->getZExtValue();
14058 auto *Ty = cast<llvm::FixedVectorType>(Ops[0]->getType());
14059 unsigned NumElts = Ty->getNumElements();
14060
14061 // These intrinsics operate on 256-bit lanes of four 64-bit elements.
14062 int Indices[8];
14063 for (unsigned l = 0; l != NumElts; l += 4)
14064 for (unsigned i = 0; i != 4; ++i)
14065 Indices[l + i] = l + ((Imm >> (2 * i)) & 0x3);
14066
14067 return Builder.CreateShuffleVector(Ops[0], ArrayRef(Indices, NumElts),
14068 "perm");
14069 }
14070 case X86::BI__builtin_ia32_palignr128:
14071 case X86::BI__builtin_ia32_palignr256:
14072 case X86::BI__builtin_ia32_palignr512: {
14073 unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0xff;
14074
14075 unsigned NumElts =
14076 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements();
14077 assert(NumElts % 16 == 0);
14078
14079 // If palignr is shifting the pair of vectors more than the size of two
14080 // lanes, emit zero.
14081 if (ShiftVal >= 32)
14082 return llvm::Constant::getNullValue(ConvertType(E->getType()));
14083
14084 // If palignr is shifting the pair of input vectors more than one lane,
14085 // but less than two lanes, convert to shifting in zeroes.
14086 if (ShiftVal > 16) {
14087 ShiftVal -= 16;
14088 Ops[1] = Ops[0];
14089 Ops[0] = llvm::Constant::getNullValue(Ops[0]->getType());
14090 }
14091
14092 int Indices[64];
14093 // 256-bit palignr operates on 128-bit lanes so we need to handle that
14094 for (unsigned l = 0; l != NumElts; l += 16) {
14095 for (unsigned i = 0; i != 16; ++i) {
14096 unsigned Idx = ShiftVal + i;
14097 if (Idx >= 16)
14098 Idx += NumElts - 16; // End of lane, switch operand.
14099 Indices[l + i] = Idx + l;
14100 }
14101 }
14102
14103 return Builder.CreateShuffleVector(Ops[1], Ops[0],
14104 ArrayRef(Indices, NumElts), "palignr");
14105 }
14106 case X86::BI__builtin_ia32_alignd128:
14107 case X86::BI__builtin_ia32_alignd256:
14108 case X86::BI__builtin_ia32_alignd512:
14109 case X86::BI__builtin_ia32_alignq128:
14110 case X86::BI__builtin_ia32_alignq256:
14111 case X86::BI__builtin_ia32_alignq512: {
14112 unsigned NumElts =
14113 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements();
14114 unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0xff;
14115
14116 // Mask the shift amount to width of a vector.
14117 ShiftVal &= NumElts - 1;
14118
14119 int Indices[16];
14120 for (unsigned i = 0; i != NumElts; ++i)
14121 Indices[i] = i + ShiftVal;
14122
14123 return Builder.CreateShuffleVector(Ops[1], Ops[0],
14124 ArrayRef(Indices, NumElts), "valign");
14125 }
14126 case X86::BI__builtin_ia32_shuf_f32x4_256:
14127 case X86::BI__builtin_ia32_shuf_f64x2_256:
14128 case X86::BI__builtin_ia32_shuf_i32x4_256:
14129 case X86::BI__builtin_ia32_shuf_i64x2_256:
14130 case X86::BI__builtin_ia32_shuf_f32x4:
14131 case X86::BI__builtin_ia32_shuf_f64x2:
14132 case X86::BI__builtin_ia32_shuf_i32x4:
14133 case X86::BI__builtin_ia32_shuf_i64x2: {
14134 unsigned Imm = cast<llvm::ConstantInt>(Ops[2])->getZExtValue();
14135 auto *Ty = cast<llvm::FixedVectorType>(Ops[0]->getType());
14136 unsigned NumElts = Ty->getNumElements();
14137 unsigned NumLanes = Ty->getPrimitiveSizeInBits() == 512 ? 4 : 2;
14138 unsigned NumLaneElts = NumElts / NumLanes;
14139
14140 int Indices[16];
14141 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
14142 unsigned Index = (Imm % NumLanes) * NumLaneElts;
14143 Imm /= NumLanes; // Discard the bits we just used.
14144 if (l >= (NumElts / 2))
14145 Index += NumElts; // Switch to other source.
14146 for (unsigned i = 0; i != NumLaneElts; ++i) {
14147 Indices[l + i] = Index + i;
14148 }
14149 }
14150
14151 return Builder.CreateShuffleVector(Ops[0], Ops[1],
14152 ArrayRef(Indices, NumElts), "shuf");
14153 }
14154
14155 case X86::BI__builtin_ia32_vperm2f128_pd256:
14156 case X86::BI__builtin_ia32_vperm2f128_ps256:
14157 case X86::BI__builtin_ia32_vperm2f128_si256:
14158 case X86::BI__builtin_ia32_permti256: {
14159 unsigned Imm = cast<llvm::ConstantInt>(Ops[2])->getZExtValue();
14160 unsigned NumElts =
14161 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements();
14162
14163 // This takes a very simple approach since there are two lanes and a
14164 // shuffle can have 2 inputs. So we reserve the first input for the first
14165 // lane and the second input for the second lane. This may result in
14166 // duplicate sources, but this can be dealt with in the backend.
14167
14168 Value *OutOps[2];
14169 int Indices[8];
14170 for (unsigned l = 0; l != 2; ++l) {
14171 // Determine the source for this lane.
14172 if (Imm & (1 << ((l * 4) + 3)))
14173 OutOps[l] = llvm::ConstantAggregateZero::get(Ops[0]->getType());
14174 else if (Imm & (1 << ((l * 4) + 1)))
14175 OutOps[l] = Ops[1];
14176 else
14177 OutOps[l] = Ops[0];
14178
14179 for (unsigned i = 0; i != NumElts/2; ++i) {
14180 // Start with ith element of the source for this lane.
14181 unsigned Idx = (l * NumElts) + i;
14182 // If bit 0 of the immediate half is set, switch to the high half of
14183 // the source.
14184 if (Imm & (1 << (l * 4)))
14185 Idx += NumElts/2;
14186 Indices[(l * (NumElts/2)) + i] = Idx;
14187 }
14188 }
14189
14190 return Builder.CreateShuffleVector(OutOps[0], OutOps[1],
14191 ArrayRef(Indices, NumElts), "vperm");
14192 }
14193
14194 case X86::BI__builtin_ia32_pslldqi128_byteshift:
14195 case X86::BI__builtin_ia32_pslldqi256_byteshift:
14196 case X86::BI__builtin_ia32_pslldqi512_byteshift: {
14197 unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() & 0xff;
14198 auto *ResultType = cast<llvm::FixedVectorType>(Ops[0]->getType());
14199 // Builtin type is vXi64 so multiply by 8 to get bytes.
14200 unsigned NumElts = ResultType->getNumElements() * 8;
14201
14202 // If pslldq is shifting the vector more than 15 bytes, emit zero.
14203 if (ShiftVal >= 16)
14204 return llvm::Constant::getNullValue(ResultType);
14205
14206 int Indices[64];
14207 // 256/512-bit pslldq operates on 128-bit lanes so we need to handle that
14208 for (unsigned l = 0; l != NumElts; l += 16) {
14209 for (unsigned i = 0; i != 16; ++i) {
14210 unsigned Idx = NumElts + i - ShiftVal;
14211 if (Idx < NumElts) Idx -= NumElts - 16; // end of lane, switch operand.
14212 Indices[l + i] = Idx + l;
14213 }
14214 }
14215
14216 auto *VecTy = llvm::FixedVectorType::get(Int8Ty, NumElts);
14217 Value *Cast = Builder.CreateBitCast(Ops[0], VecTy, "cast");
14218 Value *Zero = llvm::Constant::getNullValue(VecTy);
14219 Value *SV = Builder.CreateShuffleVector(
14220 Zero, Cast, ArrayRef(Indices, NumElts), "pslldq");
14221 return Builder.CreateBitCast(SV, Ops[0]->getType(), "cast");
14222 }
14223 case X86::BI__builtin_ia32_psrldqi128_byteshift:
14224 case X86::BI__builtin_ia32_psrldqi256_byteshift:
14225 case X86::BI__builtin_ia32_psrldqi512_byteshift: {
14226 unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() & 0xff;
14227 auto *ResultType = cast<llvm::FixedVectorType>(Ops[0]->getType());
14228 // Builtin type is vXi64 so multiply by 8 to get bytes.
14229 unsigned NumElts = ResultType->getNumElements() * 8;
14230
14231 // If psrldq is shifting the vector more than 15 bytes, emit zero.
14232 if (ShiftVal >= 16)
14233 return llvm::Constant::getNullValue(ResultType);
14234
14235 int Indices[64];
14236 // 256/512-bit psrldq operates on 128-bit lanes so we need to handle that
14237 for (unsigned l = 0; l != NumElts; l += 16) {
14238 for (unsigned i = 0; i != 16; ++i) {
14239 unsigned Idx = i + ShiftVal;
14240 if (Idx >= 16) Idx += NumElts - 16; // end of lane, switch operand.
14241 Indices[l + i] = Idx + l;
14242 }
14243 }
14244
14245 auto *VecTy = llvm::FixedVectorType::get(Int8Ty, NumElts);
14246 Value *Cast = Builder.CreateBitCast(Ops[0], VecTy, "cast");
14247 Value *Zero = llvm::Constant::getNullValue(VecTy);
14248 Value *SV = Builder.CreateShuffleVector(
14249 Cast, Zero, ArrayRef(Indices, NumElts), "psrldq");
14250 return Builder.CreateBitCast(SV, ResultType, "cast");
14251 }
14252 case X86::BI__builtin_ia32_kshiftliqi:
14253 case X86::BI__builtin_ia32_kshiftlihi:
14254 case X86::BI__builtin_ia32_kshiftlisi:
14255 case X86::BI__builtin_ia32_kshiftlidi: {
14256 unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() & 0xff;
14257 unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth();
14258
14259 if (ShiftVal >= NumElts)
14260 return llvm::Constant::getNullValue(Ops[0]->getType());
14261
14262 Value *In = getMaskVecValue(*this, Ops[0], NumElts);
14263
14264 int Indices[64];
14265 for (unsigned i = 0; i != NumElts; ++i)
14266 Indices[i] = NumElts + i - ShiftVal;
14267
14268 Value *Zero = llvm::Constant::getNullValue(In->getType());
14269 Value *SV = Builder.CreateShuffleVector(
14270 Zero, In, ArrayRef(Indices, NumElts), "kshiftl");
14271 return Builder.CreateBitCast(SV, Ops[0]->getType());
14272 }
14273 case X86::BI__builtin_ia32_kshiftriqi:
14274 case X86::BI__builtin_ia32_kshiftrihi:
14275 case X86::BI__builtin_ia32_kshiftrisi:
14276 case X86::BI__builtin_ia32_kshiftridi: {
14277 unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() & 0xff;
14278 unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth();
14279
14280 if (ShiftVal >= NumElts)
14281 return llvm::Constant::getNullValue(Ops[0]->getType());
14282
14283 Value *In = getMaskVecValue(*this, Ops[0], NumElts);
14284
14285 int Indices[64];
14286 for (unsigned i = 0; i != NumElts; ++i)
14287 Indices[i] = i + ShiftVal;
14288
14289 Value *Zero = llvm::Constant::getNullValue(In->getType());
14290 Value *SV = Builder.CreateShuffleVector(
14291 In, Zero, ArrayRef(Indices, NumElts), "kshiftr");
14292 return Builder.CreateBitCast(SV, Ops[0]->getType());
14293 }
14294 case X86::BI__builtin_ia32_movnti:
14295 case X86::BI__builtin_ia32_movnti64:
14296 case X86::BI__builtin_ia32_movntsd:
14297 case X86::BI__builtin_ia32_movntss: {
14298 llvm::MDNode *Node = llvm::MDNode::get(
14299 getLLVMContext(), llvm::ConstantAsMetadata::get(Builder.getInt32(1)));
14300
14301 Value *Ptr = Ops[0];
14302 Value *Src = Ops[1];
14303
14304 // Extract the 0'th element of the source vector.
14305 if (BuiltinID == X86::BI__builtin_ia32_movntsd ||
14306 BuiltinID == X86::BI__builtin_ia32_movntss)
14307 Src = Builder.CreateExtractElement(Src, (uint64_t)0, "extract");
14308
14309 // Convert the type of the pointer to a pointer to the stored type.
14310 Value *BC = Builder.CreateBitCast(
14311 Ptr, llvm::PointerType::getUnqual(Src->getType()), "cast");
14312
14313 // Unaligned nontemporal store of the scalar value.
14314 StoreInst *SI = Builder.CreateDefaultAlignedStore(Src, BC);
14315 SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node);
14316 SI->setAlignment(llvm::Align(1));
14317 return SI;
14318 }
14319 // Rotate is a special case of funnel shift - 1st 2 args are the same.
14320 case X86::BI__builtin_ia32_vprotb:
14321 case X86::BI__builtin_ia32_vprotw:
14322 case X86::BI__builtin_ia32_vprotd:
14323 case X86::BI__builtin_ia32_vprotq:
14324 case X86::BI__builtin_ia32_vprotbi:
14325 case X86::BI__builtin_ia32_vprotwi:
14326 case X86::BI__builtin_ia32_vprotdi:
14327 case X86::BI__builtin_ia32_vprotqi:
14328 case X86::BI__builtin_ia32_prold128:
14329 case X86::BI__builtin_ia32_prold256:
14330 case X86::BI__builtin_ia32_prold512:
14331 case X86::BI__builtin_ia32_prolq128:
14332 case X86::BI__builtin_ia32_prolq256:
14333 case X86::BI__builtin_ia32_prolq512:
14334 case X86::BI__builtin_ia32_prolvd128:
14335 case X86::BI__builtin_ia32_prolvd256:
14336 case X86::BI__builtin_ia32_prolvd512:
14337 case X86::BI__builtin_ia32_prolvq128:
14338 case X86::BI__builtin_ia32_prolvq256:
14339 case X86::BI__builtin_ia32_prolvq512:
14340 return EmitX86FunnelShift(*this, Ops[0], Ops[0], Ops[1], false);
14341 case X86::BI__builtin_ia32_prord128:
14342 case X86::BI__builtin_ia32_prord256:
14343 case X86::BI__builtin_ia32_prord512:
14344 case X86::BI__builtin_ia32_prorq128:
14345 case X86::BI__builtin_ia32_prorq256:
14346 case X86::BI__builtin_ia32_prorq512:
14347 case X86::BI__builtin_ia32_prorvd128:
14348 case X86::BI__builtin_ia32_prorvd256:
14349 case X86::BI__builtin_ia32_prorvd512:
14350 case X86::BI__builtin_ia32_prorvq128:
14351 case X86::BI__builtin_ia32_prorvq256:
14352 case X86::BI__builtin_ia32_prorvq512:
14353 return EmitX86FunnelShift(*this, Ops[0], Ops[0], Ops[1], true);
14354 case X86::BI__builtin_ia32_selectb_128:
14355 case X86::BI__builtin_ia32_selectb_256:
14356 case X86::BI__builtin_ia32_selectb_512:
14357 case X86::BI__builtin_ia32_selectw_128:
14358 case X86::BI__builtin_ia32_selectw_256:
14359 case X86::BI__builtin_ia32_selectw_512:
14360 case X86::BI__builtin_ia32_selectd_128:
14361 case X86::BI__builtin_ia32_selectd_256:
14362 case X86::BI__builtin_ia32_selectd_512:
14363 case X86::BI__builtin_ia32_selectq_128:
14364 case X86::BI__builtin_ia32_selectq_256:
14365 case X86::BI__builtin_ia32_selectq_512:
14366 case X86::BI__builtin_ia32_selectph_128:
14367 case X86::BI__builtin_ia32_selectph_256:
14368 case X86::BI__builtin_ia32_selectph_512:
14369 case X86::BI__builtin_ia32_selectpbf_128:
14370 case X86::BI__builtin_ia32_selectpbf_256:
14371 case X86::BI__builtin_ia32_selectpbf_512:
14372 case X86::BI__builtin_ia32_selectps_128:
14373 case X86::BI__builtin_ia32_selectps_256:
14374 case X86::BI__builtin_ia32_selectps_512:
14375 case X86::BI__builtin_ia32_selectpd_128:
14376 case X86::BI__builtin_ia32_selectpd_256:
14377 case X86::BI__builtin_ia32_selectpd_512:
14378 return EmitX86Select(*this, Ops[0], Ops[1], Ops[2]);
14379 case X86::BI__builtin_ia32_selectsh_128:
14380 case X86::BI__builtin_ia32_selectsbf_128:
14381 case X86::BI__builtin_ia32_selectss_128:
14382 case X86::BI__builtin_ia32_selectsd_128: {
14383 Value *A = Builder.CreateExtractElement(Ops[1], (uint64_t)0);
14384 Value *B = Builder.CreateExtractElement(Ops[2], (uint64_t)0);
14385 A = EmitX86ScalarSelect(*this, Ops[0], A, B);
14386 return Builder.CreateInsertElement(Ops[1], A, (uint64_t)0);
14387 }
14388 case X86::BI__builtin_ia32_cmpb128_mask:
14389 case X86::BI__builtin_ia32_cmpb256_mask:
14390 case X86::BI__builtin_ia32_cmpb512_mask:
14391 case X86::BI__builtin_ia32_cmpw128_mask:
14392 case X86::BI__builtin_ia32_cmpw256_mask:
14393 case X86::BI__builtin_ia32_cmpw512_mask:
14394 case X86::BI__builtin_ia32_cmpd128_mask:
14395 case X86::BI__builtin_ia32_cmpd256_mask:
14396 case X86::BI__builtin_ia32_cmpd512_mask:
14397 case X86::BI__builtin_ia32_cmpq128_mask:
14398 case X86::BI__builtin_ia32_cmpq256_mask:
14399 case X86::BI__builtin_ia32_cmpq512_mask: {
14400 unsigned CC = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0x7;
14401 return EmitX86MaskedCompare(*this, CC, true, Ops);
14402 }
14403 case X86::BI__builtin_ia32_ucmpb128_mask:
14404 case X86::BI__builtin_ia32_ucmpb256_mask:
14405 case X86::BI__builtin_ia32_ucmpb512_mask:
14406 case X86::BI__builtin_ia32_ucmpw128_mask:
14407 case X86::BI__builtin_ia32_ucmpw256_mask:
14408 case X86::BI__builtin_ia32_ucmpw512_mask:
14409 case X86::BI__builtin_ia32_ucmpd128_mask:
14410 case X86::BI__builtin_ia32_ucmpd256_mask:
14411 case X86::BI__builtin_ia32_ucmpd512_mask:
14412 case X86::BI__builtin_ia32_ucmpq128_mask:
14413 case X86::BI__builtin_ia32_ucmpq256_mask:
14414 case X86::BI__builtin_ia32_ucmpq512_mask: {
14415 unsigned CC = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0x7;
14416 return EmitX86MaskedCompare(*this, CC, false, Ops);
14417 }
14418 case X86::BI__builtin_ia32_vpcomb:
14419 case X86::BI__builtin_ia32_vpcomw:
14420 case X86::BI__builtin_ia32_vpcomd:
14421 case X86::BI__builtin_ia32_vpcomq:
14422 return EmitX86vpcom(*this, Ops, true);
14423 case X86::BI__builtin_ia32_vpcomub:
14424 case X86::BI__builtin_ia32_vpcomuw:
14425 case X86::BI__builtin_ia32_vpcomud:
14426 case X86::BI__builtin_ia32_vpcomuq:
14427 return EmitX86vpcom(*this, Ops, false);
14428
14429 case X86::BI__builtin_ia32_kortestcqi:
14430 case X86::BI__builtin_ia32_kortestchi:
14431 case X86::BI__builtin_ia32_kortestcsi:
14432 case X86::BI__builtin_ia32_kortestcdi: {
14433 Value *Or = EmitX86MaskLogic(*this, Instruction::Or, Ops);
14434 Value *C = llvm::Constant::getAllOnesValue(Ops[0]->getType());
14435 Value *Cmp = Builder.CreateICmpEQ(Or, C);
14436 return Builder.CreateZExt(Cmp, ConvertType(E->getType()));
14437 }
14438 case X86::BI__builtin_ia32_kortestzqi:
14439 case X86::BI__builtin_ia32_kortestzhi:
14440 case X86::BI__builtin_ia32_kortestzsi:
14441 case X86::BI__builtin_ia32_kortestzdi: {
14442 Value *Or = EmitX86MaskLogic(*this, Instruction::Or, Ops);
14443 Value *C = llvm::Constant::getNullValue(Ops[0]->getType());
14444 Value *Cmp = Builder.CreateICmpEQ(Or, C);
14445 return Builder.CreateZExt(Cmp, ConvertType(E->getType()));
14446 }
14447
14448 case X86::BI__builtin_ia32_ktestcqi:
14449 case X86::BI__builtin_ia32_ktestzqi:
14450 case X86::BI__builtin_ia32_ktestchi:
14451 case X86::BI__builtin_ia32_ktestzhi:
14452 case X86::BI__builtin_ia32_ktestcsi:
14453 case X86::BI__builtin_ia32_ktestzsi:
14454 case X86::BI__builtin_ia32_ktestcdi:
14455 case X86::BI__builtin_ia32_ktestzdi: {
14456 Intrinsic::ID IID;
14457 switch (BuiltinID) {
14458 default: llvm_unreachable("Unsupported intrinsic!");
14459 case X86::BI__builtin_ia32_ktestcqi:
14460 IID = Intrinsic::x86_avx512_ktestc_b;
14461 break;
14462 case X86::BI__builtin_ia32_ktestzqi:
14463 IID = Intrinsic::x86_avx512_ktestz_b;
14464 break;
14465 case X86::BI__builtin_ia32_ktestchi:
14466 IID = Intrinsic::x86_avx512_ktestc_w;
14467 break;
14468 case X86::BI__builtin_ia32_ktestzhi:
14469 IID = Intrinsic::x86_avx512_ktestz_w;
14470 break;
14471 case X86::BI__builtin_ia32_ktestcsi:
14472 IID = Intrinsic::x86_avx512_ktestc_d;
14473 break;
14474 case X86::BI__builtin_ia32_ktestzsi:
14475 IID = Intrinsic::x86_avx512_ktestz_d;
14476 break;
14477 case X86::BI__builtin_ia32_ktestcdi:
14478 IID = Intrinsic::x86_avx512_ktestc_q;
14479 break;
14480 case X86::BI__builtin_ia32_ktestzdi:
14481 IID = Intrinsic::x86_avx512_ktestz_q;
14482 break;
14483 }
14484
14485 unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth();
14486 Value *LHS = getMaskVecValue(*this, Ops[0], NumElts);
14487 Value *RHS = getMaskVecValue(*this, Ops[1], NumElts);
14488 Function *Intr = CGM.getIntrinsic(IID);
14489 return Builder.CreateCall(Intr, {LHS, RHS});
14490 }
14491
14492 case X86::BI__builtin_ia32_kaddqi:
14493 case X86::BI__builtin_ia32_kaddhi:
14494 case X86::BI__builtin_ia32_kaddsi:
14495 case X86::BI__builtin_ia32_kadddi: {
14496 Intrinsic::ID IID;
14497 switch (BuiltinID) {
14498 default: llvm_unreachable("Unsupported intrinsic!");
14499 case X86::BI__builtin_ia32_kaddqi:
14500 IID = Intrinsic::x86_avx512_kadd_b;
14501 break;
14502 case X86::BI__builtin_ia32_kaddhi:
14503 IID = Intrinsic::x86_avx512_kadd_w;
14504 break;
14505 case X86::BI__builtin_ia32_kaddsi:
14506 IID = Intrinsic::x86_avx512_kadd_d;
14507 break;
14508 case X86::BI__builtin_ia32_kadddi:
14509 IID = Intrinsic::x86_avx512_kadd_q;
14510 break;
14511 }
14512
14513 unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth();
14514 Value *LHS = getMaskVecValue(*this, Ops[0], NumElts);
14515 Value *RHS = getMaskVecValue(*this, Ops[1], NumElts);
14516 Function *Intr = CGM.getIntrinsic(IID);
14517 Value *Res = Builder.CreateCall(Intr, {LHS, RHS});
14518 return Builder.CreateBitCast(Res, Ops[0]->getType());
14519 }
14520 case X86::BI__builtin_ia32_kandqi:
14521 case X86::BI__builtin_ia32_kandhi:
14522 case X86::BI__builtin_ia32_kandsi:
14523 case X86::BI__builtin_ia32_kanddi:
14524 return EmitX86MaskLogic(*this, Instruction::And, Ops);
14525 case X86::BI__builtin_ia32_kandnqi:
14526 case X86::BI__builtin_ia32_kandnhi:
14527 case X86::BI__builtin_ia32_kandnsi:
14528 case X86::BI__builtin_ia32_kandndi:
14529 return EmitX86MaskLogic(*this, Instruction::And, Ops, true);
14530 case X86::BI__builtin_ia32_korqi:
14531 case X86::BI__builtin_ia32_korhi:
14532 case X86::BI__builtin_ia32_korsi:
14533 case X86::BI__builtin_ia32_kordi:
14534 return EmitX86MaskLogic(*this, Instruction::Or, Ops);
14535 case X86::BI__builtin_ia32_kxnorqi:
14536 case X86::BI__builtin_ia32_kxnorhi:
14537 case X86::BI__builtin_ia32_kxnorsi:
14538 case X86::BI__builtin_ia32_kxnordi:
14539 return EmitX86MaskLogic(*this, Instruction::Xor, Ops, true);
14540 case X86::BI__builtin_ia32_kxorqi:
14541 case X86::BI__builtin_ia32_kxorhi:
14542 case X86::BI__builtin_ia32_kxorsi:
14543 case X86::BI__builtin_ia32_kxordi:
14544 return EmitX86MaskLogic(*this, Instruction::Xor, Ops);
14545 case X86::BI__builtin_ia32_knotqi:
14546 case X86::BI__builtin_ia32_knothi:
14547 case X86::BI__builtin_ia32_knotsi:
14548 case X86::BI__builtin_ia32_knotdi: {
14549 unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth();
14550 Value *Res = getMaskVecValue(*this, Ops[0], NumElts);
14551 return Builder.CreateBitCast(Builder.CreateNot(Res),
14552 Ops[0]->getType());
14553 }
14554 case X86::BI__builtin_ia32_kmovb:
14555 case X86::BI__builtin_ia32_kmovw:
14556 case X86::BI__builtin_ia32_kmovd:
14557 case X86::BI__builtin_ia32_kmovq: {
14558 // Bitcast to vXi1 type and then back to integer. This gets the mask
14559 // register type into the IR, but might be optimized out depending on
14560 // what's around it.
14561 unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth();
14562 Value *Res = getMaskVecValue(*this, Ops[0], NumElts);
14563 return Builder.CreateBitCast(Res, Ops[0]->getType());
14564 }
14565
14566 case X86::BI__builtin_ia32_kunpckdi:
14567 case X86::BI__builtin_ia32_kunpcksi:
14568 case X86::BI__builtin_ia32_kunpckhi: {
14569 unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth();
14570 Value *LHS = getMaskVecValue(*this, Ops[0], NumElts);
14571 Value *RHS = getMaskVecValue(*this, Ops[1], NumElts);
14572 int Indices[64];
14573 for (unsigned i = 0; i != NumElts; ++i)
14574 Indices[i] = i;
14575
14576 // First extract half of each vector. This gives better codegen than
14577 // doing it in a single shuffle.
14578 LHS = Builder.CreateShuffleVector(LHS, LHS, ArrayRef(Indices, NumElts / 2));
14579 RHS = Builder.CreateShuffleVector(RHS, RHS, ArrayRef(Indices, NumElts / 2));
14580 // Concat the vectors.
14581 // NOTE: Operands are swapped to match the intrinsic definition.
14582 Value *Res =
14583 Builder.CreateShuffleVector(RHS, LHS, ArrayRef(Indices, NumElts));
14584 return Builder.CreateBitCast(Res, Ops[0]->getType());
14585 }
14586
14587 case X86::BI__builtin_ia32_vplzcntd_128:
14588 case X86::BI__builtin_ia32_vplzcntd_256:
14589 case X86::BI__builtin_ia32_vplzcntd_512:
14590 case X86::BI__builtin_ia32_vplzcntq_128:
14591 case X86::BI__builtin_ia32_vplzcntq_256:
14592 case X86::BI__builtin_ia32_vplzcntq_512: {
14593 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, Ops[0]->getType());
14594 return Builder.CreateCall(F, {Ops[0],Builder.getInt1(false)});
14595 }
14596 case X86::BI__builtin_ia32_sqrtss:
14597 case X86::BI__builtin_ia32_sqrtsd: {
14598 Value *A = Builder.CreateExtractElement(Ops[0], (uint64_t)0);
14599 Function *F;
14600 if (Builder.getIsFPConstrained()) {
14601 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E);
14602 F = CGM.getIntrinsic(Intrinsic::experimental_constrained_sqrt,
14603 A->getType());
14604 A = Builder.CreateConstrainedFPCall(F, {A});
14605 } else {
14606 F = CGM.getIntrinsic(Intrinsic::sqrt, A->getType());
14607 A = Builder.CreateCall(F, {A});
14608 }
14609 return Builder.CreateInsertElement(Ops[0], A, (uint64_t)0);
14610 }
14611 case X86::BI__builtin_ia32_sqrtsh_round_mask:
14612 case X86::BI__builtin_ia32_sqrtsd_round_mask:
14613 case X86::BI__builtin_ia32_sqrtss_round_mask: {
14614 unsigned CC = cast<llvm::ConstantInt>(Ops[4])->getZExtValue();
14615 // Support only if the rounding mode is 4 (AKA CUR_DIRECTION),
14616 // otherwise keep the intrinsic.
14617 if (CC != 4) {
14618 Intrinsic::ID IID;
14619
14620 switch (BuiltinID) {
14621 default:
14622 llvm_unreachable("Unsupported intrinsic!");
14623 case X86::BI__builtin_ia32_sqrtsh_round_mask:
14624 IID = Intrinsic::x86_avx512fp16_mask_sqrt_sh;
14625 break;
14626 case X86::BI__builtin_ia32_sqrtsd_round_mask:
14627 IID = Intrinsic::x86_avx512_mask_sqrt_sd;
14628 break;
14629 case X86::BI__builtin_ia32_sqrtss_round_mask:
14630 IID = Intrinsic::x86_avx512_mask_sqrt_ss;
14631 break;
14632 }
14633 return Builder.CreateCall(CGM.getIntrinsic(IID), Ops);
14634 }
14635 Value *A = Builder.CreateExtractElement(Ops[1], (uint64_t)0);
14636 Function *F;
14637 if (Builder.getIsFPConstrained()) {
14638 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E);
14639 F = CGM.getIntrinsic(Intrinsic::experimental_constrained_sqrt,
14640 A->getType());
14641 A = Builder.CreateConstrainedFPCall(F, A);
14642 } else {
14643 F = CGM.getIntrinsic(Intrinsic::sqrt, A->getType());
14644 A = Builder.CreateCall(F, A);
14645 }
14646 Value *Src = Builder.CreateExtractElement(Ops[2], (uint64_t)0);
14647 A = EmitX86ScalarSelect(*this, Ops[3], A, Src);
14648 return Builder.CreateInsertElement(Ops[0], A, (uint64_t)0);
14649 }
14650 case X86::BI__builtin_ia32_sqrtpd256:
14651 case X86::BI__builtin_ia32_sqrtpd:
14652 case X86::BI__builtin_ia32_sqrtps256:
14653 case X86::BI__builtin_ia32_sqrtps:
14654 case X86::BI__builtin_ia32_sqrtph256:
14655 case X86::BI__builtin_ia32_sqrtph:
14656 case X86::BI__builtin_ia32_sqrtph512:
14657 case X86::BI__builtin_ia32_sqrtps512:
14658 case X86::BI__builtin_ia32_sqrtpd512: {
14659 if (Ops.size() == 2) {
14660 unsigned CC = cast<llvm::ConstantInt>(Ops[1])->getZExtValue();
14661 // Support only if the rounding mode is 4 (AKA CUR_DIRECTION),
14662 // otherwise keep the intrinsic.
14663 if (CC != 4) {
14664 Intrinsic::ID IID;
14665
14666 switch (BuiltinID) {
14667 default:
14668 llvm_unreachable("Unsupported intrinsic!");
14669 case X86::BI__builtin_ia32_sqrtph512:
14670 IID = Intrinsic::x86_avx512fp16_sqrt_ph_512;
14671 break;
14672 case X86::BI__builtin_ia32_sqrtps512:
14673 IID = Intrinsic::x86_avx512_sqrt_ps_512;
14674 break;
14675 case X86::BI__builtin_ia32_sqrtpd512:
14676 IID = Intrinsic::x86_avx512_sqrt_pd_512;
14677 break;
14678 }
14679 return Builder.CreateCall(CGM.getIntrinsic(IID), Ops);
14680 }
14681 }
14682 if (Builder.getIsFPConstrained()) {
14683 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E);
14684 Function *F = CGM.getIntrinsic(Intrinsic::experimental_constrained_sqrt,
14685 Ops[0]->getType());
14686 return Builder.CreateConstrainedFPCall(F, Ops[0]);
14687 } else {
14688 Function *F = CGM.getIntrinsic(Intrinsic::sqrt, Ops[0]->getType());
14689 return Builder.CreateCall(F, Ops[0]);
14690 }
14691 }
14692
14693 case X86::BI__builtin_ia32_pmuludq128:
14694 case X86::BI__builtin_ia32_pmuludq256:
14695 case X86::BI__builtin_ia32_pmuludq512:
14696 return EmitX86Muldq(*this, /*IsSigned*/false, Ops);
14697
14698 case X86::BI__builtin_ia32_pmuldq128:
14699 case X86::BI__builtin_ia32_pmuldq256:
14700 case X86::BI__builtin_ia32_pmuldq512:
14701 return EmitX86Muldq(*this, /*IsSigned*/true, Ops);
14702
14703 case X86::BI__builtin_ia32_pternlogd512_mask:
14704 case X86::BI__builtin_ia32_pternlogq512_mask:
14705 case X86::BI__builtin_ia32_pternlogd128_mask:
14706 case X86::BI__builtin_ia32_pternlogd256_mask:
14707 case X86::BI__builtin_ia32_pternlogq128_mask:
14708 case X86::BI__builtin_ia32_pternlogq256_mask:
14709 return EmitX86Ternlog(*this, /*ZeroMask*/false, Ops);
14710
14711 case X86::BI__builtin_ia32_pternlogd512_maskz:
14712 case X86::BI__builtin_ia32_pternlogq512_maskz:
14713 case X86::BI__builtin_ia32_pternlogd128_maskz:
14714 case X86::BI__builtin_ia32_pternlogd256_maskz:
14715 case X86::BI__builtin_ia32_pternlogq128_maskz:
14716 case X86::BI__builtin_ia32_pternlogq256_maskz:
14717 return EmitX86Ternlog(*this, /*ZeroMask*/true, Ops);
14718
14719 case X86::BI__builtin_ia32_vpshldd128:
14720 case X86::BI__builtin_ia32_vpshldd256:
14721 case X86::BI__builtin_ia32_vpshldd512:
14722 case X86::BI__builtin_ia32_vpshldq128:
14723 case X86::BI__builtin_ia32_vpshldq256:
14724 case X86::BI__builtin_ia32_vpshldq512:
14725 case X86::BI__builtin_ia32_vpshldw128:
14726 case X86::BI__builtin_ia32_vpshldw256:
14727 case X86::BI__builtin_ia32_vpshldw512:
14728 return EmitX86FunnelShift(*this, Ops[0], Ops[1], Ops[2], false);
14729
14730 case X86::BI__builtin_ia32_vpshrdd128:
14731 case X86::BI__builtin_ia32_vpshrdd256:
14732 case X86::BI__builtin_ia32_vpshrdd512:
14733 case X86::BI__builtin_ia32_vpshrdq128:
14734 case X86::BI__builtin_ia32_vpshrdq256:
14735 case X86::BI__builtin_ia32_vpshrdq512:
14736 case X86::BI__builtin_ia32_vpshrdw128:
14737 case X86::BI__builtin_ia32_vpshrdw256:
14738 case X86::BI__builtin_ia32_vpshrdw512:
14739 // Ops 0 and 1 are swapped.
14740 return EmitX86FunnelShift(*this, Ops[1], Ops[0], Ops[2], true);
14741
14742 case X86::BI__builtin_ia32_vpshldvd128:
14743 case X86::BI__builtin_ia32_vpshldvd256:
14744 case X86::BI__builtin_ia32_vpshldvd512:
14745 case X86::BI__builtin_ia32_vpshldvq128:
14746 case X86::BI__builtin_ia32_vpshldvq256:
14747 case X86::BI__builtin_ia32_vpshldvq512:
14748 case X86::BI__builtin_ia32_vpshldvw128:
14749 case X86::BI__builtin_ia32_vpshldvw256:
14750 case X86::BI__builtin_ia32_vpshldvw512:
14751 return EmitX86FunnelShift(*this, Ops[0], Ops[1], Ops[2], false);
14752
14753 case X86::BI__builtin_ia32_vpshrdvd128:
14754 case X86::BI__builtin_ia32_vpshrdvd256:
14755 case X86::BI__builtin_ia32_vpshrdvd512:
14756 case X86::BI__builtin_ia32_vpshrdvq128:
14757 case X86::BI__builtin_ia32_vpshrdvq256:
14758 case X86::BI__builtin_ia32_vpshrdvq512:
14759 case X86::BI__builtin_ia32_vpshrdvw128:
14760 case X86::BI__builtin_ia32_vpshrdvw256:
14761 case X86::BI__builtin_ia32_vpshrdvw512:
14762 // Ops 0 and 1 are swapped.
14763 return EmitX86FunnelShift(*this, Ops[1], Ops[0], Ops[2], true);
14764
14765 // Reductions
14766 case X86::BI__builtin_ia32_reduce_fadd_pd512:
14767 case X86::BI__builtin_ia32_reduce_fadd_ps512:
14768 case X86::BI__builtin_ia32_reduce_fadd_ph512:
14769 case X86::BI__builtin_ia32_reduce_fadd_ph256:
14770 case X86::BI__builtin_ia32_reduce_fadd_ph128: {
14771 Function *F =
14772 CGM.getIntrinsic(Intrinsic::vector_reduce_fadd, Ops[1]->getType());
14773 IRBuilder<>::FastMathFlagGuard FMFGuard(Builder);
14774 Builder.getFastMathFlags().setAllowReassoc();
14775 return Builder.CreateCall(F, {Ops[0], Ops[1]});
14776 }
14777 case X86::BI__builtin_ia32_reduce_fmul_pd512:
14778 case X86::BI__builtin_ia32_reduce_fmul_ps512:
14779 case X86::BI__builtin_ia32_reduce_fmul_ph512:
14780 case X86::BI__builtin_ia32_reduce_fmul_ph256:
14781 case X86::BI__builtin_ia32_reduce_fmul_ph128: {
14782 Function *F =
14783 CGM.getIntrinsic(Intrinsic::vector_reduce_fmul, Ops[1]->getType());
14784 IRBuilder<>::FastMathFlagGuard FMFGuard(Builder);
14785 Builder.getFastMathFlags().setAllowReassoc();
14786 return Builder.CreateCall(F, {Ops[0], Ops[1]});
14787 }
14788 case X86::BI__builtin_ia32_reduce_fmax_pd512:
14789 case X86::BI__builtin_ia32_reduce_fmax_ps512:
14790 case X86::BI__builtin_ia32_reduce_fmax_ph512:
14791 case X86::BI__builtin_ia32_reduce_fmax_ph256:
14792 case X86::BI__builtin_ia32_reduce_fmax_ph128: {
14793 Function *F =
14794 CGM.getIntrinsic(Intrinsic::vector_reduce_fmax, Ops[0]->getType());
14795 IRBuilder<>::FastMathFlagGuard FMFGuard(Builder);
14796 Builder.getFastMathFlags().setNoNaNs();
14797 return Builder.CreateCall(F, {Ops[0]});
14798 }
14799 case X86::BI__builtin_ia32_reduce_fmin_pd512:
14800 case X86::BI__builtin_ia32_reduce_fmin_ps512:
14801 case X86::BI__builtin_ia32_reduce_fmin_ph512:
14802 case X86::BI__builtin_ia32_reduce_fmin_ph256:
14803 case X86::BI__builtin_ia32_reduce_fmin_ph128: {
14804 Function *F =
14805 CGM.getIntrinsic(Intrinsic::vector_reduce_fmin, Ops[0]->getType());
14806 IRBuilder<>::FastMathFlagGuard FMFGuard(Builder);
14807 Builder.getFastMathFlags().setNoNaNs();
14808 return Builder.CreateCall(F, {Ops[0]});
14809 }
14810
14811 // 3DNow!
14812 case X86::BI__builtin_ia32_pswapdsf:
14813 case X86::BI__builtin_ia32_pswapdsi: {
14814 llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext());
14815 Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast");
14816 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_3dnowa_pswapd);
14817 return Builder.CreateCall(F, Ops, "pswapd");
14818 }
14819 case X86::BI__builtin_ia32_rdrand16_step:
14820 case X86::BI__builtin_ia32_rdrand32_step:
14821 case X86::BI__builtin_ia32_rdrand64_step:
14822 case X86::BI__builtin_ia32_rdseed16_step:
14823 case X86::BI__builtin_ia32_rdseed32_step:
14824 case X86::BI__builtin_ia32_rdseed64_step: {
14825 Intrinsic::ID ID;
14826 switch (BuiltinID) {
14827 default: llvm_unreachable("Unsupported intrinsic!");
14828 case X86::BI__builtin_ia32_rdrand16_step:
14829 ID = Intrinsic::x86_rdrand_16;
14830 break;
14831 case X86::BI__builtin_ia32_rdrand32_step:
14832 ID = Intrinsic::x86_rdrand_32;
14833 break;
14834 case X86::BI__builtin_ia32_rdrand64_step:
14835 ID = Intrinsic::x86_rdrand_64;
14836 break;
14837 case X86::BI__builtin_ia32_rdseed16_step:
14838 ID = Intrinsic::x86_rdseed_16;
14839 break;
14840 case X86::BI__builtin_ia32_rdseed32_step:
14841 ID = Intrinsic::x86_rdseed_32;
14842 break;
14843 case X86::BI__builtin_ia32_rdseed64_step:
14844 ID = Intrinsic::x86_rdseed_64;
14845 break;
14846 }
14847
14848 Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID));
14849 Builder.CreateDefaultAlignedStore(Builder.CreateExtractValue(Call, 0),
14850 Ops[0]);
14851 return Builder.CreateExtractValue(Call, 1);
14852 }
14853 case X86::BI__builtin_ia32_addcarryx_u32:
14854 case X86::BI__builtin_ia32_addcarryx_u64:
14855 case X86::BI__builtin_ia32_subborrow_u32:
14856 case X86::BI__builtin_ia32_subborrow_u64: {
14857 Intrinsic::ID IID;
14858 switch (BuiltinID) {
14859 default: llvm_unreachable("Unsupported intrinsic!");
14860 case X86::BI__builtin_ia32_addcarryx_u32:
14861 IID = Intrinsic::x86_addcarry_32;
14862 break;
14863 case X86::BI__builtin_ia32_addcarryx_u64:
14864 IID = Intrinsic::x86_addcarry_64;
14865 break;
14866 case X86::BI__builtin_ia32_subborrow_u32:
14867 IID = Intrinsic::x86_subborrow_32;
14868 break;
14869 case X86::BI__builtin_ia32_subborrow_u64:
14870 IID = Intrinsic::x86_subborrow_64;
14871 break;
14872 }
14873
14874 Value *Call = Builder.CreateCall(CGM.getIntrinsic(IID),
14875 { Ops[0], Ops[1], Ops[2] });
14876 Builder.CreateDefaultAlignedStore(Builder.CreateExtractValue(Call, 1),
14877 Ops[3]);
14878 return Builder.CreateExtractValue(Call, 0);
14879 }
14880
14881 case X86::BI__builtin_ia32_fpclassps128_mask:
14882 case X86::BI__builtin_ia32_fpclassps256_mask:
14883 case X86::BI__builtin_ia32_fpclassps512_mask:
14884 case X86::BI__builtin_ia32_fpclassph128_mask:
14885 case X86::BI__builtin_ia32_fpclassph256_mask:
14886 case X86::BI__builtin_ia32_fpclassph512_mask:
14887 case X86::BI__builtin_ia32_fpclasspd128_mask:
14888 case X86::BI__builtin_ia32_fpclasspd256_mask:
14889 case X86::BI__builtin_ia32_fpclasspd512_mask: {
14890 unsigned NumElts =
14891 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements();
14892 Value *MaskIn = Ops[2];
14893 Ops.erase(&Ops[2]);
14894
14895 Intrinsic::ID ID;
14896 switch (BuiltinID) {
14897 default: llvm_unreachable("Unsupported intrinsic!");
14898 case X86::BI__builtin_ia32_fpclassph128_mask:
14899 ID = Intrinsic::x86_avx512fp16_fpclass_ph_128;
14900 break;
14901 case X86::BI__builtin_ia32_fpclassph256_mask:
14902 ID = Intrinsic::x86_avx512fp16_fpclass_ph_256;
14903 break;
14904 case X86::BI__builtin_ia32_fpclassph512_mask:
14905 ID = Intrinsic::x86_avx512fp16_fpclass_ph_512;
14906 break;
14907 case X86::BI__builtin_ia32_fpclassps128_mask:
14908 ID = Intrinsic::x86_avx512_fpclass_ps_128;
14909 break;
14910 case X86::BI__builtin_ia32_fpclassps256_mask:
14911 ID = Intrinsic::x86_avx512_fpclass_ps_256;
14912 break;
14913 case X86::BI__builtin_ia32_fpclassps512_mask:
14914 ID = Intrinsic::x86_avx512_fpclass_ps_512;
14915 break;
14916 case X86::BI__builtin_ia32_fpclasspd128_mask:
14917 ID = Intrinsic::x86_avx512_fpclass_pd_128;
14918 break;
14919 case X86::BI__builtin_ia32_fpclasspd256_mask:
14920 ID = Intrinsic::x86_avx512_fpclass_pd_256;
14921 break;
14922 case X86::BI__builtin_ia32_fpclasspd512_mask:
14923 ID = Intrinsic::x86_avx512_fpclass_pd_512;
14924 break;
14925 }
14926
14927 Value *Fpclass = Builder.CreateCall(CGM.getIntrinsic(ID), Ops);
14928 return EmitX86MaskedCompareResult(*this, Fpclass, NumElts, MaskIn);
14929 }
14930
14931 case X86::BI__builtin_ia32_vp2intersect_q_512:
14932 case X86::BI__builtin_ia32_vp2intersect_q_256:
14933 case X86::BI__builtin_ia32_vp2intersect_q_128:
14934 case X86::BI__builtin_ia32_vp2intersect_d_512:
14935 case X86::BI__builtin_ia32_vp2intersect_d_256:
14936 case X86::BI__builtin_ia32_vp2intersect_d_128: {
14937 unsigned NumElts =
14938 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements();
14939 Intrinsic::ID ID;
14940
14941 switch (BuiltinID) {
14942 default: llvm_unreachable("Unsupported intrinsic!");
14943 case X86::BI__builtin_ia32_vp2intersect_q_512:
14944 ID = Intrinsic::x86_avx512_vp2intersect_q_512;
14945 break;
14946 case X86::BI__builtin_ia32_vp2intersect_q_256:
14947 ID = Intrinsic::x86_avx512_vp2intersect_q_256;
14948 break;
14949 case X86::BI__builtin_ia32_vp2intersect_q_128:
14950 ID = Intrinsic::x86_avx512_vp2intersect_q_128;
14951 break;
14952 case X86::BI__builtin_ia32_vp2intersect_d_512:
14953 ID = Intrinsic::x86_avx512_vp2intersect_d_512;
14954 break;
14955 case X86::BI__builtin_ia32_vp2intersect_d_256:
14956 ID = Intrinsic::x86_avx512_vp2intersect_d_256;
14957 break;
14958 case X86::BI__builtin_ia32_vp2intersect_d_128:
14959 ID = Intrinsic::x86_avx512_vp2intersect_d_128;
14960 break;
14961 }
14962
14963 Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID), {Ops[0], Ops[1]});
14964 Value *Result = Builder.CreateExtractValue(Call, 0);
14965 Result = EmitX86MaskedCompareResult(*this, Result, NumElts, nullptr);
14966 Builder.CreateDefaultAlignedStore(Result, Ops[2]);
14967
14968 Result = Builder.CreateExtractValue(Call, 1);
14969 Result = EmitX86MaskedCompareResult(*this, Result, NumElts, nullptr);
14970 return Builder.CreateDefaultAlignedStore(Result, Ops[3]);
14971 }
14972
14973 case X86::BI__builtin_ia32_vpmultishiftqb128:
14974 case X86::BI__builtin_ia32_vpmultishiftqb256:
14975 case X86::BI__builtin_ia32_vpmultishiftqb512: {
14976 Intrinsic::ID ID;
14977 switch (BuiltinID) {
14978 default: llvm_unreachable("Unsupported intrinsic!");
14979 case X86::BI__builtin_ia32_vpmultishiftqb128:
14980 ID = Intrinsic::x86_avx512_pmultishift_qb_128;
14981 break;
14982 case X86::BI__builtin_ia32_vpmultishiftqb256:
14983 ID = Intrinsic::x86_avx512_pmultishift_qb_256;
14984 break;
14985 case X86::BI__builtin_ia32_vpmultishiftqb512:
14986 ID = Intrinsic::x86_avx512_pmultishift_qb_512;
14987 break;
14988 }
14989
14990 return Builder.CreateCall(CGM.getIntrinsic(ID), Ops);
14991 }
14992
14993 case X86::BI__builtin_ia32_vpshufbitqmb128_mask:
14994 case X86::BI__builtin_ia32_vpshufbitqmb256_mask:
14995 case X86::BI__builtin_ia32_vpshufbitqmb512_mask: {
14996 unsigned NumElts =
14997 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements();
14998 Value *MaskIn = Ops[2];
14999 Ops.erase(&Ops[2]);
15000
15001 Intrinsic::ID ID;
15002 switch (BuiltinID) {
15003 default: llvm_unreachable("Unsupported intrinsic!");
15004 case X86::BI__builtin_ia32_vpshufbitqmb128_mask:
15005 ID = Intrinsic::x86_avx512_vpshufbitqmb_128;
15006 break;
15007 case X86::BI__builtin_ia32_vpshufbitqmb256_mask:
15008 ID = Intrinsic::x86_avx512_vpshufbitqmb_256;
15009 break;
15010 case X86::BI__builtin_ia32_vpshufbitqmb512_mask:
15011 ID = Intrinsic::x86_avx512_vpshufbitqmb_512;
15012 break;
15013 }
15014
15015 Value *Shufbit = Builder.CreateCall(CGM.getIntrinsic(ID), Ops);
15016 return EmitX86MaskedCompareResult(*this, Shufbit, NumElts, MaskIn);
15017 }
15018
15019 // packed comparison intrinsics
15020 case X86::BI__builtin_ia32_cmpeqps:
15021 case X86::BI__builtin_ia32_cmpeqpd:
15022 return getVectorFCmpIR(CmpInst::FCMP_OEQ, /*IsSignaling*/false);
15023 case X86::BI__builtin_ia32_cmpltps:
15024 case X86::BI__builtin_ia32_cmpltpd:
15025 return getVectorFCmpIR(CmpInst::FCMP_OLT, /*IsSignaling*/true);
15026 case X86::BI__builtin_ia32_cmpleps:
15027 case X86::BI__builtin_ia32_cmplepd:
15028 return getVectorFCmpIR(CmpInst::FCMP_OLE, /*IsSignaling*/true);
15029 case X86::BI__builtin_ia32_cmpunordps:
15030 case X86::BI__builtin_ia32_cmpunordpd:
15031 return getVectorFCmpIR(CmpInst::FCMP_UNO, /*IsSignaling*/false);
15032 case X86::BI__builtin_ia32_cmpneqps:
15033 case X86::BI__builtin_ia32_cmpneqpd:
15034 return getVectorFCmpIR(CmpInst::FCMP_UNE, /*IsSignaling*/false);
15035 case X86::BI__builtin_ia32_cmpnltps:
15036 case X86::BI__builtin_ia32_cmpnltpd:
15037 return getVectorFCmpIR(CmpInst::FCMP_UGE, /*IsSignaling*/true);
15038 case X86::BI__builtin_ia32_cmpnleps:
15039 case X86::BI__builtin_ia32_cmpnlepd:
15040 return getVectorFCmpIR(CmpInst::FCMP_UGT, /*IsSignaling*/true);
15041 case X86::BI__builtin_ia32_cmpordps:
15042 case X86::BI__builtin_ia32_cmpordpd:
15043 return getVectorFCmpIR(CmpInst::FCMP_ORD, /*IsSignaling*/false);
15044 case X86::BI__builtin_ia32_cmpph128_mask:
15045 case X86::BI__builtin_ia32_cmpph256_mask:
15046 case X86::BI__builtin_ia32_cmpph512_mask:
15047 case X86::BI__builtin_ia32_cmpps128_mask:
15048 case X86::BI__builtin_ia32_cmpps256_mask:
15049 case X86::BI__builtin_ia32_cmpps512_mask:
15050 case X86::BI__builtin_ia32_cmppd128_mask:
15051 case X86::BI__builtin_ia32_cmppd256_mask:
15052 case X86::BI__builtin_ia32_cmppd512_mask:
15053 IsMaskFCmp = true;
15054 [[fallthrough]];
15055 case X86::BI__builtin_ia32_cmpps:
15056 case X86::BI__builtin_ia32_cmpps256:
15057 case X86::BI__builtin_ia32_cmppd:
15058 case X86::BI__builtin_ia32_cmppd256: {
15059 // Lowering vector comparisons to fcmp instructions, while
15060 // ignoring signalling behaviour requested
15061 // ignoring rounding mode requested
15062 // This is only possible if fp-model is not strict and FENV_ACCESS is off.
15063
15064 // The third argument is the comparison condition, and integer in the
15065 // range [0, 31]
15066 unsigned CC = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0x1f;
15067
15068 // Lowering to IR fcmp instruction.
15069 // Ignoring requested signaling behaviour,
15070 // e.g. both _CMP_GT_OS & _CMP_GT_OQ are translated to FCMP_OGT.
15071 FCmpInst::Predicate Pred;
15072 bool IsSignaling;
15073 // Predicates for 16-31 repeat the 0-15 predicates. Only the signalling
15074 // behavior is inverted. We'll handle that after the switch.
15075 switch (CC & 0xf) {
15076 case 0x00: Pred = FCmpInst::FCMP_OEQ; IsSignaling = false; break;
15077 case 0x01: Pred = FCmpInst::FCMP_OLT; IsSignaling = true; break;
15078 case 0x02: Pred = FCmpInst::FCMP_OLE; IsSignaling = true; break;
15079 case 0x03: Pred = FCmpInst::FCMP_UNO; IsSignaling = false; break;
15080 case 0x04: Pred = FCmpInst::FCMP_UNE; IsSignaling = false; break;
15081 case 0x05: Pred = FCmpInst::FCMP_UGE; IsSignaling = true; break;
15082 case 0x06: Pred = FCmpInst::FCMP_UGT; IsSignaling = true; break;
15083 case 0x07: Pred = FCmpInst::FCMP_ORD; IsSignaling = false; break;
15084 case 0x08: Pred = FCmpInst::FCMP_UEQ; IsSignaling = false; break;
15085 case 0x09: Pred = FCmpInst::FCMP_ULT; IsSignaling = true; break;
15086 case 0x0a: Pred = FCmpInst::FCMP_ULE; IsSignaling = true; break;
15087 case 0x0b: Pred = FCmpInst::FCMP_FALSE; IsSignaling = false; break;
15088 case 0x0c: Pred = FCmpInst::FCMP_ONE; IsSignaling = false; break;
15089 case 0x0d: Pred = FCmpInst::FCMP_OGE; IsSignaling = true; break;
15090 case 0x0e: Pred = FCmpInst::FCMP_OGT; IsSignaling = true; break;
15091 case 0x0f: Pred = FCmpInst::FCMP_TRUE; IsSignaling = false; break;
15092 default: llvm_unreachable("Unhandled CC");
15093 }
15094
15095 // Invert the signalling behavior for 16-31.
15096 if (CC & 0x10)
15097 IsSignaling = !IsSignaling;
15098
15099 // If the predicate is true or false and we're using constrained intrinsics,
15100 // we don't have a compare intrinsic we can use. Just use the legacy X86
15101 // specific intrinsic.
15102 // If the intrinsic is mask enabled and we're using constrained intrinsics,
15103 // use the legacy X86 specific intrinsic.
15104 if (Builder.getIsFPConstrained() &&
15105 (Pred == FCmpInst::FCMP_TRUE || Pred == FCmpInst::FCMP_FALSE ||
15106 IsMaskFCmp)) {
15107
15108 Intrinsic::ID IID;
15109 switch (BuiltinID) {
15110 default: llvm_unreachable("Unexpected builtin");
15111 case X86::BI__builtin_ia32_cmpps:
15112 IID = Intrinsic::x86_sse_cmp_ps;
15113 break;
15114 case X86::BI__builtin_ia32_cmpps256:
15115 IID = Intrinsic::x86_avx_cmp_ps_256;
15116 break;
15117 case X86::BI__builtin_ia32_cmppd:
15118 IID = Intrinsic::x86_sse2_cmp_pd;
15119 break;
15120 case X86::BI__builtin_ia32_cmppd256:
15121 IID = Intrinsic::x86_avx_cmp_pd_256;
15122 break;
15123 case X86::BI__builtin_ia32_cmpps512_mask:
15124 IID = Intrinsic::x86_avx512_mask_cmp_ps_512;
15125 break;
15126 case X86::BI__builtin_ia32_cmppd512_mask:
15127 IID = Intrinsic::x86_avx512_mask_cmp_pd_512;
15128 break;
15129 case X86::BI__builtin_ia32_cmpps128_mask:
15130 IID = Intrinsic::x86_avx512_mask_cmp_ps_128;
15131 break;
15132 case X86::BI__builtin_ia32_cmpps256_mask:
15133 IID = Intrinsic::x86_avx512_mask_cmp_ps_256;
15134 break;
15135 case X86::BI__builtin_ia32_cmppd128_mask:
15136 IID = Intrinsic::x86_avx512_mask_cmp_pd_128;
15137 break;
15138 case X86::BI__builtin_ia32_cmppd256_mask:
15139 IID = Intrinsic::x86_avx512_mask_cmp_pd_256;
15140 break;
15141 }
15142
15143 Function *Intr = CGM.getIntrinsic(IID);
15144 if (IsMaskFCmp) {
15145 unsigned NumElts =
15146 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements();
15147 Ops[3] = getMaskVecValue(*this, Ops[3], NumElts);
15148 Value *Cmp = Builder.CreateCall(Intr, Ops);
15149 return EmitX86MaskedCompareResult(*this, Cmp, NumElts, nullptr);
15150 }
15151
15152 return Builder.CreateCall(Intr, Ops);
15153 }
15154
15155 // Builtins without the _mask suffix return a vector of integers
15156 // of the same width as the input vectors
15157 if (IsMaskFCmp) {
15158 // We ignore SAE if strict FP is disabled. We only keep precise
15159 // exception behavior under strict FP.
15160 // NOTE: If strict FP does ever go through here a CGFPOptionsRAII
15161 // object will be required.
15162 unsigned NumElts =
15163 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements();
15164 Value *Cmp;
15165 if (IsSignaling)
15166 Cmp = Builder.CreateFCmpS(Pred, Ops[0], Ops[1]);
15167 else
15168 Cmp = Builder.CreateFCmp(Pred, Ops[0], Ops[1]);
15169 return EmitX86MaskedCompareResult(*this, Cmp, NumElts, Ops[3]);
15170 }
15171
15172 return getVectorFCmpIR(Pred, IsSignaling);
15173 }
15174
15175 // SSE scalar comparison intrinsics
15176 case X86::BI__builtin_ia32_cmpeqss:
15177 return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 0);
15178 case X86::BI__builtin_ia32_cmpltss:
15179 return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 1);
15180 case X86::BI__builtin_ia32_cmpless:
15181 return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 2);
15182 case X86::BI__builtin_ia32_cmpunordss:
15183 return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 3);
15184 case X86::BI__builtin_ia32_cmpneqss:
15185 return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 4);
15186 case X86::BI__builtin_ia32_cmpnltss:
15187 return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 5);
15188 case X86::BI__builtin_ia32_cmpnless:
15189 return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 6);
15190 case X86::BI__builtin_ia32_cmpordss:
15191 return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 7);
15192 case X86::BI__builtin_ia32_cmpeqsd:
15193 return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 0);
15194 case X86::BI__builtin_ia32_cmpltsd:
15195 return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 1);
15196 case X86::BI__builtin_ia32_cmplesd:
15197 return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 2);
15198 case X86::BI__builtin_ia32_cmpunordsd:
15199 return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 3);
15200 case X86::BI__builtin_ia32_cmpneqsd:
15201 return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 4);
15202 case X86::BI__builtin_ia32_cmpnltsd:
15203 return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 5);
15204 case X86::BI__builtin_ia32_cmpnlesd:
15205 return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 6);
15206 case X86::BI__builtin_ia32_cmpordsd:
15207 return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 7);
15208
15209 // f16c half2float intrinsics
15210 case X86::BI__builtin_ia32_vcvtph2ps:
15211 case X86::BI__builtin_ia32_vcvtph2ps256:
15212 case X86::BI__builtin_ia32_vcvtph2ps_mask:
15213 case X86::BI__builtin_ia32_vcvtph2ps256_mask:
15214 case X86::BI__builtin_ia32_vcvtph2ps512_mask: {
15215 CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E);
15216 return EmitX86CvtF16ToFloatExpr(*this, Ops, ConvertType(E->getType()));
15217 }
15218
15219 // AVX512 bf16 intrinsics
15220 case X86::BI__builtin_ia32_cvtneps2bf16_128_mask: {
15221 Ops[2] = getMaskVecValue(
15222 *this, Ops[2],
15223 cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements());
15224 Intrinsic::ID IID = Intrinsic::x86_avx512bf16_mask_cvtneps2bf16_128;
15225 return Builder.CreateCall(CGM.getIntrinsic(IID), Ops);
15226 }
15227 case X86::BI__builtin_ia32_cvtsbf162ss_32:
15228 return Builder.CreateFPExt(Ops[0], Builder.getFloatTy());
15229
15230 case X86::BI__builtin_ia32_cvtneps2bf16_256_mask:
15231 case X86::BI__builtin_ia32_cvtneps2bf16_512_mask: {
15232 Intrinsic::ID IID;
15233 switch (BuiltinID) {
15234 default: llvm_unreachable("Unsupported intrinsic!");
15235 case X86::BI__builtin_ia32_cvtneps2bf16_256_mask:
15236 IID = Intrinsic::x86_avx512bf16_cvtneps2bf16_256;
15237 break;
15238 case X86::BI__builtin_ia32_cvtneps2bf16_512_mask:
15239 IID = Intrinsic::x86_avx512bf16_cvtneps2bf16_512;
15240 break;
15241 }
15242 Value *Res = Builder.CreateCall(CGM.getIntrinsic(IID), Ops[0]);
15243 return EmitX86Select(*this, Ops[2], Res, Ops[1]);
15244 }
15245
15246 case X86::BI__cpuid:
15247 case X86::BI__cpuidex: {
15248 Value *FuncId = EmitScalarExpr(E->getArg(1));
15249 Value *SubFuncId = BuiltinID == X86::BI__cpuidex
15250 ? EmitScalarExpr(E->getArg(2))
15251 : llvm::ConstantInt::get(Int32Ty, 0);
15252
15253 llvm::StructType *CpuidRetTy =
15254 llvm::StructType::get(Int32Ty, Int32Ty, Int32Ty, Int32Ty);
15255 llvm::FunctionType *FTy =
15256 llvm::FunctionType::get(CpuidRetTy, {Int32Ty, Int32Ty}, false);
15257
15258 StringRef Asm, Constraints;
15259 if (getTarget().getTriple().getArch() == llvm::Triple::x86) {
15260 Asm = "cpuid";
15261 Constraints = "={ax},={bx},={cx},={dx},{ax},{cx}";
15262 } else {
15263 // x86-64 uses %rbx as the base register, so preserve it.
15264 Asm = "xchgq %rbx, ${1:q}\n"
15265 "cpuid\n"
15266 "xchgq %rbx, ${1:q}";
15267 Constraints = "={ax},=r,={cx},={dx},0,2";
15268 }
15269
15270 llvm::InlineAsm *IA = llvm::InlineAsm::get(FTy, Asm, Constraints,
15271 /*hasSideEffects=*/false);
15272 Value *IACall = Builder.CreateCall(IA, {FuncId, SubFuncId});
15273 Value *BasePtr = EmitScalarExpr(E->getArg(0));
15274 Value *Store = nullptr;
15275 for (unsigned i = 0; i < 4; i++) {
15276 Value *Extracted = Builder.CreateExtractValue(IACall, i);
15277 Value *StorePtr = Builder.CreateConstInBoundsGEP1_32(Int32Ty, BasePtr, i);
15278 Store = Builder.CreateAlignedStore(Extracted, StorePtr, getIntAlign());
15279 }
15280
15281 // Return the last store instruction to signal that we have emitted the
15282 // the intrinsic.
15283 return Store;
15284 }
15285
15286 case X86::BI__emul:
15287 case X86::BI__emulu: {
15288 llvm::Type *Int64Ty = llvm::IntegerType::get(getLLVMContext(), 64);
15289 bool isSigned = (BuiltinID == X86::BI__emul);
15290 Value *LHS = Builder.CreateIntCast(Ops[0], Int64Ty, isSigned);
15291 Value *RHS = Builder.CreateIntCast(Ops[1], Int64Ty, isSigned);
15292 return Builder.CreateMul(LHS, RHS, "", !isSigned, isSigned);
15293 }
15294 case X86::BI__mulh:
15295 case X86::BI__umulh:
15296 case X86::BI_mul128:
15297 case X86::BI_umul128: {
15298 llvm::Type *ResType = ConvertType(E->getType());
15299 llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128);
15300
15301 bool IsSigned = (BuiltinID == X86::BI__mulh || BuiltinID == X86::BI_mul128);
15302 Value *LHS = Builder.CreateIntCast(Ops[0], Int128Ty, IsSigned);
15303 Value *RHS = Builder.CreateIntCast(Ops[1], Int128Ty, IsSigned);
15304
15305 Value *MulResult, *HigherBits;
15306 if (IsSigned) {
15307 MulResult = Builder.CreateNSWMul(LHS, RHS);
15308 HigherBits = Builder.CreateAShr(MulResult, 64);
15309 } else {
15310 MulResult = Builder.CreateNUWMul(LHS, RHS);
15311 HigherBits = Builder.CreateLShr(MulResult, 64);
15312 }
15313 HigherBits = Builder.CreateIntCast(HigherBits, ResType, IsSigned);
15314
15315 if (BuiltinID == X86::BI__mulh || BuiltinID == X86::BI__umulh)
15316 return HigherBits;
15317
15318 Address HighBitsAddress = EmitPointerWithAlignment(E->getArg(2));
15319 Builder.CreateStore(HigherBits, HighBitsAddress);
15320 return Builder.CreateIntCast(MulResult, ResType, IsSigned);
15321 }
15322
15323 case X86::BI__faststorefence: {
15324 return Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent,
15325 llvm::SyncScope::System);
15326 }
15327 case X86::BI__shiftleft128:
15328 case X86::BI__shiftright128: {
15329 llvm::Function *F = CGM.getIntrinsic(
15330 BuiltinID == X86::BI__shiftleft128 ? Intrinsic::fshl : Intrinsic::fshr,
15331 Int64Ty);
15332 // Flip low/high ops and zero-extend amount to matching type.
15333 // shiftleft128(Low, High, Amt) -> fshl(High, Low, Amt)
15334 // shiftright128(Low, High, Amt) -> fshr(High, Low, Amt)
15335 std::swap(Ops[0], Ops[1]);
15336 Ops[2] = Builder.CreateZExt(Ops[2], Int64Ty);
15337 return Builder.CreateCall(F, Ops);
15338 }
15339 case X86::BI_ReadWriteBarrier:
15340 case X86::BI_ReadBarrier:
15341 case X86::BI_WriteBarrier: {
15342 return Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent,
15343 llvm::SyncScope::SingleThread);
15344 }
15345
15346 case X86::BI_AddressOfReturnAddress: {
15347 Function *F =
15348 CGM.getIntrinsic(Intrinsic::addressofreturnaddress, AllocaInt8PtrTy);
15349 return Builder.CreateCall(F);
15350 }
15351 case X86::BI__stosb: {
15352 // We treat __stosb as a volatile memset - it may not generate "rep stosb"
15353 // instruction, but it will create a memset that won't be optimized away.
15354 return Builder.CreateMemSet(Ops[0], Ops[1], Ops[2], Align(1), true);
15355 }
15356 case X86::BI__ud2:
15357 // llvm.trap makes a ud2a instruction on x86.
15358 return EmitTrapCall(Intrinsic::trap);
15359 case X86::BI__int2c: {
15360 // This syscall signals a driver assertion failure in x86 NT kernels.
15361 llvm::FunctionType *FTy = llvm::FunctionType::get(VoidTy, false);
15362 llvm::InlineAsm *IA =
15363 llvm::InlineAsm::get(FTy, "int $$0x2c", "", /*hasSideEffects=*/true);
15364 llvm::AttributeList NoReturnAttr = llvm::AttributeList::get(
15365 getLLVMContext(), llvm::AttributeList::FunctionIndex,
15366 llvm::Attribute::NoReturn);
15367 llvm::CallInst *CI = Builder.CreateCall(IA);
15368 CI->setAttributes(NoReturnAttr);
15369 return CI;
15370 }
15371 case X86::BI__readfsbyte:
15372 case X86::BI__readfsword:
15373 case X86::BI__readfsdword:
15374 case X86::BI__readfsqword: {
15375 llvm::Type *IntTy = ConvertType(E->getType());
15376 Value *Ptr =
15377 Builder.CreateIntToPtr(Ops[0], llvm::PointerType::get(IntTy, 257));
15378 LoadInst *Load = Builder.CreateAlignedLoad(
15379 IntTy, Ptr, getContext().getTypeAlignInChars(E->getType()));
15380 Load->setVolatile(true);
15381 return Load;
15382 }
15383 case X86::BI__readgsbyte:
15384 case X86::BI__readgsword:
15385 case X86::BI__readgsdword:
15386 case X86::BI__readgsqword: {
15387 llvm::Type *IntTy = ConvertType(E->getType());
15388 Value *Ptr =
15389 Builder.CreateIntToPtr(Ops[0], llvm::PointerType::get(IntTy, 256));
15390 LoadInst *Load = Builder.CreateAlignedLoad(
15391 IntTy, Ptr, getContext().getTypeAlignInChars(E->getType()));
15392 Load->setVolatile(true);
15393 return Load;
15394 }
15395 case X86::BI__builtin_ia32_encodekey128_u32: {
15396 Intrinsic::ID IID = Intrinsic::x86_encodekey128;
15397
15398 Value *Call = Builder.CreateCall(CGM.getIntrinsic(IID), {Ops[0], Ops[1]});
15399
15400 for (int i = 0; i < 3; ++i) {
15401 Value *Extract = Builder.CreateExtractValue(Call, i + 1);
15402 Value *Ptr = Builder.CreateConstGEP1_32(Int8Ty, Ops[2], i * 16);
15403 Ptr = Builder.CreateBitCast(
15404 Ptr, llvm::PointerType::getUnqual(Extract->getType()));
15405 Builder.CreateAlignedStore(Extract, Ptr, Align(1));
15406 }
15407
15408 return Builder.CreateExtractValue(Call, 0);
15409 }
15410 case X86::BI__builtin_ia32_encodekey256_u32: {
15411 Intrinsic::ID IID = Intrinsic::x86_encodekey256;
15412
15413 Value *Call =
15414 Builder.CreateCall(CGM.getIntrinsic(IID), {Ops[0], Ops[1], Ops[2]});
15415
15416 for (int i = 0; i < 4; ++i) {
15417 Value *Extract = Builder.CreateExtractValue(Call, i + 1);
15418 Value *Ptr = Builder.CreateConstGEP1_32(Int8Ty, Ops[3], i * 16);
15419 Ptr = Builder.CreateBitCast(
15420 Ptr, llvm::PointerType::getUnqual(Extract->getType()));
15421 Builder.CreateAlignedStore(Extract, Ptr, Align(1));
15422 }
15423
15424 return Builder.CreateExtractValue(Call, 0);
15425 }
15426 case X86::BI__builtin_ia32_aesenc128kl_u8:
15427 case X86::BI__builtin_ia32_aesdec128kl_u8:
15428 case X86::BI__builtin_ia32_aesenc256kl_u8:
15429 case X86::BI__builtin_ia32_aesdec256kl_u8: {
15430 Intrinsic::ID IID;
15431 StringRef BlockName;
15432 switch (BuiltinID) {
15433 default:
15434 llvm_unreachable("Unexpected builtin");
15435 case X86::BI__builtin_ia32_aesenc128kl_u8:
15436 IID = Intrinsic::x86_aesenc128kl;
15437 BlockName = "aesenc128kl";
15438 break;
15439 case X86::BI__builtin_ia32_aesdec128kl_u8:
15440 IID = Intrinsic::x86_aesdec128kl;
15441 BlockName = "aesdec128kl";
15442 break;
15443 case X86::BI__builtin_ia32_aesenc256kl_u8:
15444 IID = Intrinsic::x86_aesenc256kl;
15445 BlockName = "aesenc256kl";
15446 break;
15447 case X86::BI__builtin_ia32_aesdec256kl_u8:
15448 IID = Intrinsic::x86_aesdec256kl;
15449 BlockName = "aesdec256kl";
15450 break;
15451 }
15452
15453 Value *Call = Builder.CreateCall(CGM.getIntrinsic(IID), {Ops[1], Ops[2]});
15454
15455 BasicBlock *NoError =
15456 createBasicBlock(BlockName + "_no_error", this->CurFn);
15457 BasicBlock *Error = createBasicBlock(BlockName + "_error", this->CurFn);
15458 BasicBlock *End = createBasicBlock(BlockName + "_end", this->CurFn);
15459
15460 Value *Ret = Builder.CreateExtractValue(Call, 0);
15461 Value *Succ = Builder.CreateTrunc(Ret, Builder.getInt1Ty());
15462 Value *Out = Builder.CreateExtractValue(Call, 1);
15463 Builder.CreateCondBr(Succ, NoError, Error);
15464
15465 Builder.SetInsertPoint(NoError);
15466 Builder.CreateDefaultAlignedStore(Out, Ops[0]);
15467 Builder.CreateBr(End);
15468
15469 Builder.SetInsertPoint(Error);
15470 Constant *Zero = llvm::Constant::getNullValue(Out->getType());
15471 Builder.CreateDefaultAlignedStore(Zero, Ops[0]);
15472 Builder.CreateBr(End);
15473
15474 Builder.SetInsertPoint(End);
15475 return Builder.CreateExtractValue(Call, 0);
15476 }
15477 case X86::BI__builtin_ia32_aesencwide128kl_u8:
15478 case X86::BI__builtin_ia32_aesdecwide128kl_u8:
15479 case X86::BI__builtin_ia32_aesencwide256kl_u8:
15480 case X86::BI__builtin_ia32_aesdecwide256kl_u8: {
15481 Intrinsic::ID IID;
15482 StringRef BlockName;
15483 switch (BuiltinID) {
15484 case X86::BI__builtin_ia32_aesencwide128kl_u8:
15485 IID = Intrinsic::x86_aesencwide128kl;
15486 BlockName = "aesencwide128kl";
15487 break;
15488 case X86::BI__builtin_ia32_aesdecwide128kl_u8:
15489 IID = Intrinsic::x86_aesdecwide128kl;
15490 BlockName = "aesdecwide128kl";
15491 break;
15492 case X86::BI__builtin_ia32_aesencwide256kl_u8:
15493 IID = Intrinsic::x86_aesencwide256kl;
15494 BlockName = "aesencwide256kl";
15495 break;
15496 case X86::BI__builtin_ia32_aesdecwide256kl_u8:
15497 IID = Intrinsic::x86_aesdecwide256kl;
15498 BlockName = "aesdecwide256kl";
15499 break;
15500 }
15501
15502 llvm::Type *Ty = FixedVectorType::get(Builder.getInt64Ty(), 2);
15503 Value *InOps[9];
15504 InOps[0] = Ops[2];
15505 for (int i = 0; i != 8; ++i) {
15506 Value *Ptr = Builder.CreateConstGEP1_32(Ty, Ops[1], i);
15507 InOps[i + 1] = Builder.CreateAlignedLoad(Ty, Ptr, Align(16));
15508 }
15509
15510 Value *Call = Builder.CreateCall(CGM.getIntrinsic(IID), InOps);
15511
15512 BasicBlock *NoError =
15513 createBasicBlock(BlockName + "_no_error", this->CurFn);
15514 BasicBlock *Error = createBasicBlock(BlockName + "_error", this->CurFn);
15515 BasicBlock *End = createBasicBlock(BlockName + "_end", this->CurFn);
15516
15517 Value *Ret = Builder.CreateExtractValue(Call, 0);
15518 Value *Succ = Builder.CreateTrunc(Ret, Builder.getInt1Ty());
15519 Builder.CreateCondBr(Succ, NoError, Error);
15520
15521 Builder.SetInsertPoint(NoError);
15522 for (int i = 0; i != 8; ++i) {
15523 Value *Extract = Builder.CreateExtractValue(Call, i + 1);
15524 Value *Ptr = Builder.CreateConstGEP1_32(Extract->getType(), Ops[0], i);
15525 Builder.CreateAlignedStore(Extract, Ptr, Align(16));
15526 }
15527 Builder.CreateBr(End);
15528
15529 Builder.SetInsertPoint(Error);
15530 for (int i = 0; i != 8; ++i) {
15531 Value *Out = Builder.CreateExtractValue(Call, i + 1);
15532 Constant *Zero = llvm::Constant::getNullValue(Out->getType());
15533 Value *Ptr = Builder.CreateConstGEP1_32(Out->getType(), Ops[0], i);
15534 Builder.CreateAlignedStore(Zero, Ptr, Align(16));
15535 }
15536 Builder.CreateBr(End);
15537
15538 Builder.SetInsertPoint(End);
15539 return Builder.CreateExtractValue(Call, 0);
15540 }
15541 case X86::BI__builtin_ia32_vfcmaddcph512_mask:
15542 IsConjFMA = true;
15543 [[fallthrough]];
15544 case X86::BI__builtin_ia32_vfmaddcph512_mask: {
15545 Intrinsic::ID IID = IsConjFMA
15546 ? Intrinsic::x86_avx512fp16_mask_vfcmadd_cph_512
15547 : Intrinsic::x86_avx512fp16_mask_vfmadd_cph_512;
15548 Value *Call = Builder.CreateCall(CGM.getIntrinsic(IID), Ops);
15549 return EmitX86Select(*this, Ops[3], Call, Ops[0]);
15550 }
15551 case X86::BI__builtin_ia32_vfcmaddcsh_round_mask:
15552 IsConjFMA = true;
15553 [[fallthrough]];
15554 case X86::BI__builtin_ia32_vfmaddcsh_round_mask: {
15555 Intrinsic::ID IID = IsConjFMA ? Intrinsic::x86_avx512fp16_mask_vfcmadd_csh
15556 : Intrinsic::x86_avx512fp16_mask_vfmadd_csh;
15557 Value *Call = Builder.CreateCall(CGM.getIntrinsic(IID), Ops);
15558 Value *And = Builder.CreateAnd(Ops[3], llvm::ConstantInt::get(Int8Ty, 1));
15559 return EmitX86Select(*this, And, Call, Ops[0]);
15560 }
15561 case X86::BI__builtin_ia32_vfcmaddcsh_round_mask3:
15562 IsConjFMA = true;
15563 [[fallthrough]];
15564 case X86::BI__builtin_ia32_vfmaddcsh_round_mask3: {
15565 Intrinsic::ID IID = IsConjFMA ? Intrinsic::x86_avx512fp16_mask_vfcmadd_csh
15566 : Intrinsic::x86_avx512fp16_mask_vfmadd_csh;
15567 Value *Call = Builder.CreateCall(CGM.getIntrinsic(IID), Ops);
15568 static constexpr int Mask[] = {0, 5, 6, 7};
15569 return Builder.CreateShuffleVector(Call, Ops[2], Mask);
15570 }
15571 case X86::BI__builtin_ia32_prefetchi:
15572 return Builder.CreateCall(
15573 CGM.getIntrinsic(Intrinsic::prefetch, Ops[0]->getType()),
15574 {Ops[0], llvm::ConstantInt::get(Int32Ty, 0), Ops[1],
15575 llvm::ConstantInt::get(Int32Ty, 0)});
15576 }
15577 }
15578
15579 Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID,
15580 const CallExpr *E) {
15581 // Do not emit the builtin arguments in the arguments of a function call,
15582 // because the evaluation order of function arguments is not specified in C++.
15583 // This is important when testing to ensure the arguments are emitted in the
15584 // same order every time. Eg:
15585 // Instead of:
15586 // return Builder.CreateFDiv(EmitScalarExpr(E->getArg(0)),
15587 // EmitScalarExpr(E->getArg(1)), "swdiv");
15588 // Use:
15589 // Value *Op0 = EmitScalarExpr(E->getArg(0));
15590 // Value *Op1 = EmitScalarExpr(E->getArg(1));
15591 // return Builder.CreateFDiv(Op0, Op1, "swdiv")
15592
15593 Intrinsic::ID ID = Intrinsic::not_intrinsic;
15594
15595 switch (BuiltinID) {
15596 default: return nullptr;
15597
15598 // __builtin_ppc_get_timebase is GCC 4.8+'s PowerPC-specific name for what we
15599 // call __builtin_readcyclecounter.
15600 case PPC::BI__builtin_ppc_get_timebase:
15601 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::readcyclecounter));
15602
15603 // vec_ld, vec_xl_be, vec_lvsl, vec_lvsr
15604 case PPC::BI__builtin_altivec_lvx:
15605 case PPC::BI__builtin_altivec_lvxl:
15606 case PPC::BI__builtin_altivec_lvebx:
15607 case PPC::BI__builtin_altivec_lvehx:
15608 case PPC::BI__builtin_altivec_lvewx:
15609 case PPC::BI__builtin_altivec_lvsl:
15610 case PPC::BI__builtin_altivec_lvsr:
15611 case PPC::BI__builtin_vsx_lxvd2x:
15612 case PPC::BI__builtin_vsx_lxvw4x:
15613 case PPC::BI__builtin_vsx_lxvd2x_be:
15614 case PPC::BI__builtin_vsx_lxvw4x_be:
15615 case PPC::BI__builtin_vsx_lxvl:
15616 case PPC::BI__builtin_vsx_lxvll:
15617 {
15618 SmallVector<Value *, 2> Ops;
15619 Ops.push_back(EmitScalarExpr(E->getArg(0)));
15620 Ops.push_back(EmitScalarExpr(E->getArg(1)));
15621 if(BuiltinID == PPC::BI__builtin_vsx_lxvl ||
15622 BuiltinID == PPC::BI__builtin_vsx_lxvll){
15623 Ops[0] = Builder.CreateBitCast(Ops[0], Int8PtrTy);
15624 }else {
15625 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy);
15626 Ops[0] = Builder.CreateGEP(Int8Ty, Ops[1], Ops[0]);
15627 Ops.pop_back();
15628 }
15629
15630 switch (BuiltinID) {
15631 default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!");
15632 case PPC::BI__builtin_altivec_lvx:
15633 ID = Intrinsic::ppc_altivec_lvx;
15634 break;
15635 case PPC::BI__builtin_altivec_lvxl:
15636 ID = Intrinsic::ppc_altivec_lvxl;
15637 break;
15638 case PPC::BI__builtin_altivec_lvebx:
15639 ID = Intrinsic::ppc_altivec_lvebx;
15640 break;
15641 case PPC::BI__builtin_altivec_lvehx:
15642 ID = Intrinsic::ppc_altivec_lvehx;
15643 break;
15644 case PPC::BI__builtin_altivec_lvewx:
15645 ID = Intrinsic::ppc_altivec_lvewx;
15646 break;
15647 case PPC::BI__builtin_altivec_lvsl:
15648 ID = Intrinsic::ppc_altivec_lvsl;
15649 break;
15650 case PPC::BI__builtin_altivec_lvsr:
15651 ID = Intrinsic::ppc_altivec_lvsr;
15652 break;
15653 case PPC::BI__builtin_vsx_lxvd2x:
15654 ID = Intrinsic::ppc_vsx_lxvd2x;
15655 break;
15656 case PPC::BI__builtin_vsx_lxvw4x:
15657 ID = Intrinsic::ppc_vsx_lxvw4x;
15658 break;
15659 case PPC::BI__builtin_vsx_lxvd2x_be:
15660 ID = Intrinsic::ppc_vsx_lxvd2x_be;
15661 break;
15662 case PPC::BI__builtin_vsx_lxvw4x_be:
15663 ID = Intrinsic::ppc_vsx_lxvw4x_be;
15664 break;
15665 case PPC::BI__builtin_vsx_lxvl:
15666 ID = Intrinsic::ppc_vsx_lxvl;
15667 break;
15668 case PPC::BI__builtin_vsx_lxvll:
15669 ID = Intrinsic::ppc_vsx_lxvll;
15670 break;
15671 }
15672 llvm::Function *F = CGM.getIntrinsic(ID);
15673 return Builder.CreateCall(F, Ops, "");
15674 }
15675
15676 // vec_st, vec_xst_be
15677 case PPC::BI__builtin_altivec_stvx:
15678 case PPC::BI__builtin_altivec_stvxl:
15679 case PPC::BI__builtin_altivec_stvebx:
15680 case PPC::BI__builtin_altivec_stvehx:
15681 case PPC::BI__builtin_altivec_stvewx:
15682 case PPC::BI__builtin_vsx_stxvd2x:
15683 case PPC::BI__builtin_vsx_stxvw4x:
15684 case PPC::BI__builtin_vsx_stxvd2x_be:
15685 case PPC::BI__builtin_vsx_stxvw4x_be:
15686 case PPC::BI__builtin_vsx_stxvl:
15687 case PPC::BI__builtin_vsx_stxvll:
15688 {
15689 SmallVector<Value *, 3> Ops;
15690 Ops.push_back(EmitScalarExpr(E->getArg(0)));
15691 Ops.push_back(EmitScalarExpr(E->getArg(1)));
15692 Ops.push_back(EmitScalarExpr(E->getArg(2)));
15693 if(BuiltinID == PPC::BI__builtin_vsx_stxvl ||
15694 BuiltinID == PPC::BI__builtin_vsx_stxvll ){
15695 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy);
15696 }else {
15697 Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy);
15698 Ops[1] = Builder.CreateGEP(Int8Ty, Ops[2], Ops[1]);
15699 Ops.pop_back();
15700 }
15701
15702 switch (BuiltinID) {
15703 default: llvm_unreachable("Unsupported st intrinsic!");
15704 case PPC::BI__builtin_altivec_stvx:
15705 ID = Intrinsic::ppc_altivec_stvx;
15706 break;
15707 case PPC::BI__builtin_altivec_stvxl:
15708 ID = Intrinsic::ppc_altivec_stvxl;
15709 break;
15710 case PPC::BI__builtin_altivec_stvebx:
15711 ID = Intrinsic::ppc_altivec_stvebx;
15712 break;
15713 case PPC::BI__builtin_altivec_stvehx:
15714 ID = Intrinsic::ppc_altivec_stvehx;
15715 break;
15716 case PPC::BI__builtin_altivec_stvewx:
15717 ID = Intrinsic::ppc_altivec_stvewx;
15718 break;
15719 case PPC::BI__builtin_vsx_stxvd2x:
15720 ID = Intrinsic::ppc_vsx_stxvd2x;
15721 break;
15722 case PPC::BI__builtin_vsx_stxvw4x:
15723 ID = Intrinsic::ppc_vsx_stxvw4x;
15724 break;
15725 case PPC::BI__builtin_vsx_stxvd2x_be:
15726 ID = Intrinsic::ppc_vsx_stxvd2x_be;
15727 break;
15728 case PPC::BI__builtin_vsx_stxvw4x_be:
15729 ID = Intrinsic::ppc_vsx_stxvw4x_be;
15730 break;
15731 case PPC::BI__builtin_vsx_stxvl:
15732 ID = Intrinsic::ppc_vsx_stxvl;
15733 break;
15734 case PPC::BI__builtin_vsx_stxvll:
15735 ID = Intrinsic::ppc_vsx_stxvll;
15736 break;
15737 }
15738 llvm::Function *F = CGM.getIntrinsic(ID);
15739 return Builder.CreateCall(F, Ops, "");
15740 }
15741 case PPC::BI__builtin_vsx_ldrmb: {
15742 // Essentially boils down to performing an unaligned VMX load sequence so
15743 // as to avoid crossing a page boundary and then shuffling the elements
15744 // into the right side of the vector register.
15745 Value *Op0 = EmitScalarExpr(E->getArg(0));
15746 Value *Op1 = EmitScalarExpr(E->getArg(1));
15747 int64_t NumBytes = cast<ConstantInt>(Op1)->getZExtValue();
15748 llvm::Type *ResTy = ConvertType(E->getType());
15749 bool IsLE = getTarget().isLittleEndian();
15750
15751 // If the user wants the entire vector, just load the entire vector.
15752 if (NumBytes == 16) {
15753 Value *BC = Builder.CreateBitCast(Op0, ResTy->getPointerTo());
15754 Value *LD =
15755 Builder.CreateLoad(Address(BC, ResTy, CharUnits::fromQuantity(1)));
15756 if (!IsLE)
15757 return LD;
15758
15759 // Reverse the bytes on LE.
15760 SmallVector<int, 16> RevMask;
15761 for (int Idx = 0; Idx < 16; Idx++)
15762 RevMask.push_back(15 - Idx);
15763 return Builder.CreateShuffleVector(LD, LD, RevMask);
15764 }
15765
15766 llvm::Function *Lvx = CGM.getIntrinsic(Intrinsic::ppc_altivec_lvx);
15767 llvm::Function *Lvs = CGM.getIntrinsic(IsLE ? Intrinsic::ppc_altivec_lvsr
15768 : Intrinsic::ppc_altivec_lvsl);
15769 llvm::Function *Vperm = CGM.getIntrinsic(Intrinsic::ppc_altivec_vperm);
15770 Value *HiMem = Builder.CreateGEP(
15771 Int8Ty, Op0, ConstantInt::get(Op1->getType(), NumBytes - 1));
15772 Value *LoLd = Builder.CreateCall(Lvx, Op0, "ld.lo");
15773 Value *HiLd = Builder.CreateCall(Lvx, HiMem, "ld.hi");
15774 Value *Mask1 = Builder.CreateCall(Lvs, Op0, "mask1");
15775
15776 Op0 = IsLE ? HiLd : LoLd;
15777 Op1 = IsLE ? LoLd : HiLd;
15778 Value *AllElts = Builder.CreateCall(Vperm, {Op0, Op1, Mask1}, "shuffle1");
15779 Constant *Zero = llvm::Constant::getNullValue(IsLE ? ResTy : AllElts->getType());
15780
15781 if (IsLE) {
15782 SmallVector<int, 16> Consts;
15783 for (int Idx = 0; Idx < 16; Idx++) {
15784 int Val = (NumBytes - Idx - 1 >= 0) ? (NumBytes - Idx - 1)
15785 : 16 - (NumBytes - Idx);
15786 Consts.push_back(Val);
15787 }
15788 return Builder.CreateShuffleVector(Builder.CreateBitCast(AllElts, ResTy),
15789 Zero, Consts);
15790 }
15791 SmallVector<Constant *, 16> Consts;
15792 for (int Idx = 0; Idx < 16; Idx++)
15793 Consts.push_back(Builder.getInt8(NumBytes + Idx));
15794 Value *Mask2 = ConstantVector::get(Consts);
15795 return Builder.CreateBitCast(
15796 Builder.CreateCall(Vperm, {Zero, AllElts, Mask2}, "shuffle2"), ResTy);
15797 }
15798 case PPC::BI__builtin_vsx_strmb: {
15799 Value *Op0 = EmitScalarExpr(E->getArg(0));
15800 Value *Op1 = EmitScalarExpr(E->getArg(1));
15801 Value *Op2 = EmitScalarExpr(E->getArg(2));
15802 int64_t NumBytes = cast<ConstantInt>(Op1)->getZExtValue();
15803 bool IsLE = getTarget().isLittleEndian();
15804 auto StoreSubVec = [&](unsigned Width, unsigned Offset, unsigned EltNo) {
15805 // Storing the whole vector, simply store it on BE and reverse bytes and
15806 // store on LE.
15807 if (Width == 16) {
15808 Value *BC = Builder.CreateBitCast(Op0, Op2->getType()->getPointerTo());
15809 Value *StVec = Op2;
15810 if (IsLE) {
15811 SmallVector<int, 16> RevMask;
15812 for (int Idx = 0; Idx < 16; Idx++)
15813 RevMask.push_back(15 - Idx);
15814 StVec = Builder.CreateShuffleVector(Op2, Op2, RevMask);
15815 }
15816 return Builder.CreateStore(
15817 StVec, Address(BC, Op2->getType(), CharUnits::fromQuantity(1)));
15818 }
15819 auto *ConvTy = Int64Ty;
15820 unsigned NumElts = 0;
15821 switch (Width) {
15822 default:
15823 llvm_unreachable("width for stores must be a power of 2");
15824 case 8:
15825 ConvTy = Int64Ty;
15826 NumElts = 2;
15827 break;
15828 case 4:
15829 ConvTy = Int32Ty;
15830 NumElts = 4;
15831 break;
15832 case 2:
15833 ConvTy = Int16Ty;
15834 NumElts = 8;
15835 break;
15836 case 1:
15837 ConvTy = Int8Ty;
15838 NumElts = 16;
15839 break;
15840 }
15841 Value *Vec = Builder.CreateBitCast(
15842 Op2, llvm::FixedVectorType::get(ConvTy, NumElts));
15843 Value *Ptr =
15844 Builder.CreateGEP(Int8Ty, Op0, ConstantInt::get(Int64Ty, Offset));
15845 Value *PtrBC = Builder.CreateBitCast(Ptr, ConvTy->getPointerTo());
15846 Value *Elt = Builder.CreateExtractElement(Vec, EltNo);
15847 if (IsLE && Width > 1) {
15848 Function *F = CGM.getIntrinsic(Intrinsic::bswap, ConvTy);
15849 Elt = Builder.CreateCall(F, Elt);
15850 }
15851 return Builder.CreateStore(
15852 Elt, Address(PtrBC, ConvTy, CharUnits::fromQuantity(1)));
15853 };
15854 unsigned Stored = 0;
15855 unsigned RemainingBytes = NumBytes;
15856 Value *Result;
15857 if (NumBytes == 16)
15858 return StoreSubVec(16, 0, 0);
15859 if (NumBytes >= 8) {
15860 Result = StoreSubVec(8, NumBytes - 8, IsLE ? 0 : 1);
15861 RemainingBytes -= 8;
15862 Stored += 8;
15863 }
15864 if (RemainingBytes >= 4) {
15865 Result = StoreSubVec(4, NumBytes - Stored - 4,
15866 IsLE ? (Stored >> 2) : 3 - (Stored >> 2));
15867 RemainingBytes -= 4;
15868 Stored += 4;
15869 }
15870 if (RemainingBytes >= 2) {
15871 Result = StoreSubVec(2, NumBytes - Stored - 2,
15872 IsLE ? (Stored >> 1) : 7 - (Stored >> 1));
15873 RemainingBytes -= 2;
15874 Stored += 2;
15875 }
15876 if (RemainingBytes)
15877 Result =
15878 StoreSubVec(1, NumBytes - Stored - 1, IsLE ? Stored : 15 - Stored);
15879 return Result;
15880 }
15881 // Square root
15882 case PPC::BI__builtin_vsx_xvsqrtsp:
15883 case PPC::BI__builtin_vsx_xvsqrtdp: {
15884 llvm::Type *ResultType = ConvertType(E->getType());
15885 Value *X = EmitScalarExpr(E->getArg(0));
15886 if (Builder.getIsFPConstrained()) {
15887 llvm::Function *F = CGM.getIntrinsic(
15888 Intrinsic::experimental_constrained_sqrt, ResultType);
15889 return Builder.CreateConstrainedFPCall(F, X);
15890 } else {
15891 llvm::Function *F = CGM.getIntrinsic(Intrinsic::sqrt, ResultType);
15892 return Builder.CreateCall(F, X);
15893 }
15894 }
15895 // Count leading zeros
15896 case PPC::BI__builtin_altivec_vclzb:
15897 case PPC::BI__builtin_altivec_vclzh:
15898 case PPC::BI__builtin_altivec_vclzw:
15899 case PPC::BI__builtin_altivec_vclzd: {
15900 llvm::Type *ResultType = ConvertType(E->getType());
15901 Value *X = EmitScalarExpr(E->getArg(0));
15902 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false);
15903 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ResultType);
15904 return Builder.CreateCall(F, {X, Undef});
15905 }
15906 case PPC::BI__builtin_altivec_vctzb:
15907 case PPC::BI__builtin_altivec_vctzh:
15908 case PPC::BI__builtin_altivec_vctzw:
15909 case PPC::BI__builtin_altivec_vctzd: {
15910 llvm::Type *ResultType = ConvertType(E->getType());
15911 Value *X = EmitScalarExpr(E->getArg(0));
15912 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false);
15913 Function *F = CGM.getIntrinsic(Intrinsic::cttz, ResultType);
15914 return Builder.CreateCall(F, {X, Undef});
15915 }
15916 case PPC::BI__builtin_altivec_vinsd:
15917 case PPC::BI__builtin_altivec_vinsw:
15918 case PPC::BI__builtin_altivec_vinsd_elt:
15919 case PPC::BI__builtin_altivec_vinsw_elt: {
15920 llvm::Type *ResultType = ConvertType(E->getType());
15921 Value *Op0 = EmitScalarExpr(E->getArg(0));
15922 Value *Op1 = EmitScalarExpr(E->getArg(1));
15923 Value *Op2 = EmitScalarExpr(E->getArg(2));
15924
15925 bool IsUnaligned = (BuiltinID == PPC::BI__builtin_altivec_vinsw ||
15926 BuiltinID == PPC::BI__builtin_altivec_vinsd);
15927
15928 bool Is32bit = (BuiltinID == PPC::BI__builtin_altivec_vinsw ||
15929 BuiltinID == PPC::BI__builtin_altivec_vinsw_elt);
15930
15931 // The third argument must be a compile time constant.
15932 ConstantInt *ArgCI = dyn_cast<ConstantInt>(Op2);
15933 assert(ArgCI &&
15934 "Third Arg to vinsw/vinsd intrinsic must be a constant integer!");
15935
15936 // Valid value for the third argument is dependent on the input type and
15937 // builtin called.
15938 int ValidMaxValue = 0;
15939 if (IsUnaligned)
15940 ValidMaxValue = (Is32bit) ? 12 : 8;
15941 else
15942 ValidMaxValue = (Is32bit) ? 3 : 1;
15943
15944 // Get value of third argument.
15945 int64_t ConstArg = ArgCI->getSExtValue();
15946
15947 // Compose range checking error message.
15948 std::string RangeErrMsg = IsUnaligned ? "byte" : "element";
15949 RangeErrMsg += " number " + llvm::to_string(ConstArg);
15950 RangeErrMsg += " is outside of the valid range [0, ";
15951 RangeErrMsg += llvm::to_string(ValidMaxValue) + "]";
15952
15953 // Issue error if third argument is not within the valid range.
15954 if (ConstArg < 0 || ConstArg > ValidMaxValue)
15955 CGM.Error(E->getExprLoc(), RangeErrMsg);
15956
15957 // Input to vec_replace_elt is an element index, convert to byte index.
15958 if (!IsUnaligned) {
15959 ConstArg *= Is32bit ? 4 : 8;
15960 // Fix the constant according to endianess.
15961 if (getTarget().isLittleEndian())
15962 ConstArg = (Is32bit ? 12 : 8) - ConstArg;
15963 }
15964
15965 ID = Is32bit ? Intrinsic::ppc_altivec_vinsw : Intrinsic::ppc_altivec_vinsd;
15966 Op2 = ConstantInt::getSigned(Int32Ty, ConstArg);
15967 // Casting input to vector int as per intrinsic definition.
15968 Op0 =
15969 Is32bit
15970 ? Builder.CreateBitCast(Op0, llvm::FixedVectorType::get(Int32Ty, 4))
15971 : Builder.CreateBitCast(Op0,
15972 llvm::FixedVectorType::get(Int64Ty, 2));
15973 return Builder.CreateBitCast(
15974 Builder.CreateCall(CGM.getIntrinsic(ID), {Op0, Op1, Op2}), ResultType);
15975 }
15976 case PPC::BI__builtin_altivec_vpopcntb:
15977 case PPC::BI__builtin_altivec_vpopcnth:
15978 case PPC::BI__builtin_altivec_vpopcntw:
15979 case PPC::BI__builtin_altivec_vpopcntd: {
15980 llvm::Type *ResultType = ConvertType(E->getType());
15981 Value *X = EmitScalarExpr(E->getArg(0));
15982 llvm::Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ResultType);
15983 return Builder.CreateCall(F, X);
15984 }
15985 case PPC::BI__builtin_altivec_vadduqm:
15986 case PPC::BI__builtin_altivec_vsubuqm: {
15987 Value *Op0 = EmitScalarExpr(E->getArg(0));
15988 Value *Op1 = EmitScalarExpr(E->getArg(1));
15989 llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128);
15990 Op0 = Builder.CreateBitCast(Op0, llvm::FixedVectorType::get(Int128Ty, 1));
15991 Op1 = Builder.CreateBitCast(Op1, llvm::FixedVectorType::get(Int128Ty, 1));
15992 if (BuiltinID == PPC::BI__builtin_altivec_vadduqm)
15993 return Builder.CreateAdd(Op0, Op1, "vadduqm");
15994 else
15995 return Builder.CreateSub(Op0, Op1, "vsubuqm");
15996 }
15997 case PPC::BI__builtin_altivec_vaddcuq_c:
15998 case PPC::BI__builtin_altivec_vsubcuq_c: {
15999 SmallVector<Value *, 2> Ops;
16000 Value *Op0 = EmitScalarExpr(E->getArg(0));
16001 Value *Op1 = EmitScalarExpr(E->getArg(1));
16002 llvm::Type *V1I128Ty = llvm::FixedVectorType::get(
16003 llvm::IntegerType::get(getLLVMContext(), 128), 1);
16004 Ops.push_back(Builder.CreateBitCast(Op0, V1I128Ty));
16005 Ops.push_back(Builder.CreateBitCast(Op1, V1I128Ty));
16006 ID = (BuiltinID == PPC::BI__builtin_altivec_vaddcuq_c)
16007 ? Intrinsic::ppc_altivec_vaddcuq
16008 : Intrinsic::ppc_altivec_vsubcuq;
16009 return Builder.CreateCall(CGM.getIntrinsic(ID), Ops, "");
16010 }
16011 case PPC::BI__builtin_altivec_vaddeuqm_c:
16012 case PPC::BI__builtin_altivec_vaddecuq_c:
16013 case PPC::BI__builtin_altivec_vsubeuqm_c:
16014 case PPC::BI__builtin_altivec_vsubecuq_c: {
16015 SmallVector<Value *, 3> Ops;
16016 Value *Op0 = EmitScalarExpr(E->getArg(0));
16017 Value *Op1 = EmitScalarExpr(E->getArg(1));
16018 Value *Op2 = EmitScalarExpr(E->getArg(2));
16019 llvm::Type *V1I128Ty = llvm::FixedVectorType::get(
16020 llvm::IntegerType::get(getLLVMContext(), 128), 1);
16021 Ops.push_back(Builder.CreateBitCast(Op0, V1I128Ty));
16022 Ops.push_back(Builder.CreateBitCast(Op1, V1I128Ty));
16023 Ops.push_back(Builder.CreateBitCast(Op2, V1I128Ty));
16024 switch (BuiltinID) {
16025 default:
16026 llvm_unreachable("Unsupported intrinsic!");
16027 case PPC::BI__builtin_altivec_vaddeuqm_c:
16028 ID = Intrinsic::ppc_altivec_vaddeuqm;
16029 break;
16030 case PPC::BI__builtin_altivec_vaddecuq_c:
16031 ID = Intrinsic::ppc_altivec_vaddecuq;
16032 break;
16033 case PPC::BI__builtin_altivec_vsubeuqm_c:
16034 ID = Intrinsic::ppc_altivec_vsubeuqm;
16035 break;
16036 case PPC::BI__builtin_altivec_vsubecuq_c:
16037 ID = Intrinsic::ppc_altivec_vsubecuq;
16038 break;
16039 }
16040 return Builder.CreateCall(CGM.getIntrinsic(ID), Ops, "");
16041 }
16042 // Rotate and insert under mask operation.
16043 // __rldimi(rs, is, shift, mask)
16044 // (rotl64(rs, shift) & mask) | (is & ~mask)
16045 // __rlwimi(rs, is, shift, mask)
16046 // (rotl(rs, shift) & mask) | (is & ~mask)
16047 case PPC::BI__builtin_ppc_rldimi:
16048 case PPC::BI__builtin_ppc_rlwimi: {
16049 Value *Op0 = EmitScalarExpr(E->getArg(0));
16050 Value *Op1 = EmitScalarExpr(E->getArg(1));
16051 Value *Op2 = EmitScalarExpr(E->getArg(2));
16052 Value *Op3 = EmitScalarExpr(E->getArg(3));
16053 llvm::Type *Ty = Op0->getType();
16054 Function *F = CGM.getIntrinsic(Intrinsic::fshl, Ty);
16055 if (BuiltinID == PPC::BI__builtin_ppc_rldimi)
16056 Op2 = Builder.CreateZExt(Op2, Int64Ty);
16057 Value *Shift = Builder.CreateCall(F, {Op0, Op0, Op2});
16058 Value *X = Builder.CreateAnd(Shift, Op3);
16059 Value *Y = Builder.CreateAnd(Op1, Builder.CreateNot(Op3));
16060 return Builder.CreateOr(X, Y);
16061 }
16062 // Rotate and insert under mask operation.
16063 // __rlwnm(rs, shift, mask)
16064 // rotl(rs, shift) & mask
16065 case PPC::BI__builtin_ppc_rlwnm: {
16066 Value *Op0 = EmitScalarExpr(E->getArg(0));
16067 Value *Op1 = EmitScalarExpr(E->getArg(1));
16068 Value *Op2 = EmitScalarExpr(E->getArg(2));
16069 llvm::Type *Ty = Op0->getType();
16070 Function *F = CGM.getIntrinsic(Intrinsic::fshl, Ty);
16071 Value *Shift = Builder.CreateCall(F, {Op0, Op0, Op1});
16072 return Builder.CreateAnd(Shift, Op2);
16073 }
16074 case PPC::BI__builtin_ppc_poppar4:
16075 case PPC::BI__builtin_ppc_poppar8: {
16076 Value *Op0 = EmitScalarExpr(E->getArg(0));
16077 llvm::Type *ArgType = Op0->getType();
16078 Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType);
16079 Value *Tmp = Builder.CreateCall(F, Op0);
16080
16081 llvm::Type *ResultType = ConvertType(E->getType());
16082 Value *Result = Builder.CreateAnd(Tmp, llvm::ConstantInt::get(ArgType, 1));
16083 if (Result->getType() != ResultType)
16084 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true,
16085 "cast");
16086 return Result;
16087 }
16088 case PPC::BI__builtin_ppc_cmpb: {
16089 Value *Op0 = EmitScalarExpr(E->getArg(0));
16090 Value *Op1 = EmitScalarExpr(E->getArg(1));
16091 if (getTarget().getTriple().isPPC64()) {
16092 Function *F =
16093 CGM.getIntrinsic(Intrinsic::ppc_cmpb, {Int64Ty, Int64Ty, Int64Ty});
16094 return Builder.CreateCall(F, {Op0, Op1}, "cmpb");
16095 }
16096 // For 32 bit, emit the code as below:
16097 // %conv = trunc i64 %a to i32
16098 // %conv1 = trunc i64 %b to i32
16099 // %shr = lshr i64 %a, 32
16100 // %conv2 = trunc i64 %shr to i32
16101 // %shr3 = lshr i64 %b, 32
16102 // %conv4 = trunc i64 %shr3 to i32
16103 // %0 = tail call i32 @llvm.ppc.cmpb32(i32 %conv, i32 %conv1)
16104 // %conv5 = zext i32 %0 to i64
16105 // %1 = tail call i32 @llvm.ppc.cmpb32(i32 %conv2, i32 %conv4)
16106 // %conv614 = zext i32 %1 to i64
16107 // %shl = shl nuw i64 %conv614, 32
16108 // %or = or i64 %shl, %conv5
16109 // ret i64 %or
16110 Function *F =
16111 CGM.getIntrinsic(Intrinsic::ppc_cmpb, {Int32Ty, Int32Ty, Int32Ty});
16112 Value *ArgOneLo = Builder.CreateTrunc(Op0, Int32Ty);
16113 Value *ArgTwoLo = Builder.CreateTrunc(Op1, Int32Ty);
16114 Constant *ShiftAmt = ConstantInt::get(Int64Ty, 32);
16115 Value *ArgOneHi =
16116 Builder.CreateTrunc(Builder.CreateLShr(Op0, ShiftAmt), Int32Ty);
16117 Value *ArgTwoHi =
16118 Builder.CreateTrunc(Builder.CreateLShr(Op1, ShiftAmt), Int32Ty);
16119 Value *ResLo = Builder.CreateZExt(
16120 Builder.CreateCall(F, {ArgOneLo, ArgTwoLo}, "cmpb"), Int64Ty);
16121 Value *ResHiShift = Builder.CreateZExt(
16122 Builder.CreateCall(F, {ArgOneHi, ArgTwoHi}, "cmpb"), Int64Ty);
16123 Value *ResHi = Builder.CreateShl(ResHiShift, ShiftAmt);
16124 return Builder.CreateOr(ResLo, ResHi);
16125 }
16126 // Copy sign
16127 case PPC::BI__builtin_vsx_xvcpsgnsp:
16128 case PPC::BI__builtin_vsx_xvcpsgndp: {
16129 llvm::Type *ResultType = ConvertType(E->getType());
16130 Value *X = EmitScalarExpr(E->getArg(0));
16131 Value *Y = EmitScalarExpr(E->getArg(1));
16132 ID = Intrinsic::copysign;
16133 llvm::Function *F = CGM.getIntrinsic(ID, ResultType);
16134 return Builder.CreateCall(F, {X, Y});
16135 }
16136 // Rounding/truncation
16137 case PPC::BI__builtin_vsx_xvrspip:
16138 case PPC::BI__builtin_vsx_xvrdpip:
16139 case PPC::BI__builtin_vsx_xvrdpim:
16140 case PPC::BI__builtin_vsx_xvrspim:
16141 case PPC::BI__builtin_vsx_xvrdpi:
16142 case PPC::BI__builtin_vsx_xvrspi:
16143 case PPC::BI__builtin_vsx_xvrdpic:
16144 case PPC::BI__builtin_vsx_xvrspic:
16145 case PPC::BI__builtin_vsx_xvrdpiz:
16146 case PPC::BI__builtin_vsx_xvrspiz: {
16147 llvm::Type *ResultType = ConvertType(E->getType());
16148 Value *X = EmitScalarExpr(E->getArg(0));
16149 if (BuiltinID == PPC::BI__builtin_vsx_xvrdpim ||
16150 BuiltinID == PPC::BI__builtin_vsx_xvrspim)
16151 ID = Builder.getIsFPConstrained()
16152 ? Intrinsic::experimental_constrained_floor
16153 : Intrinsic::floor;
16154 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpi ||
16155 BuiltinID == PPC::BI__builtin_vsx_xvrspi)
16156 ID = Builder.getIsFPConstrained()
16157 ? Intrinsic::experimental_constrained_round
16158 : Intrinsic::round;
16159 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpic ||
16160 BuiltinID == PPC::BI__builtin_vsx_xvrspic)
16161 ID = Builder.getIsFPConstrained()
16162 ? Intrinsic::experimental_constrained_rint
16163 : Intrinsic::rint;
16164 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpip ||
16165 BuiltinID == PPC::BI__builtin_vsx_xvrspip)
16166 ID = Builder.getIsFPConstrained()
16167 ? Intrinsic::experimental_constrained_ceil
16168 : Intrinsic::ceil;
16169 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpiz ||
16170 BuiltinID == PPC::BI__builtin_vsx_xvrspiz)
16171 ID = Builder.getIsFPConstrained()
16172 ? Intrinsic::experimental_constrained_trunc
16173 : Intrinsic::trunc;
16174 llvm::Function *F = CGM.getIntrinsic(ID, ResultType);
16175 return Builder.getIsFPConstrained() ? Builder.CreateConstrainedFPCall(F, X)
16176 : Builder.CreateCall(F, X);
16177 }
16178
16179 // Absolute value
16180 case PPC::BI__builtin_vsx_xvabsdp:
16181 case PPC::BI__builtin_vsx_xvabssp: {
16182 llvm::Type *ResultType = ConvertType(E->getType());
16183 Value *X = EmitScalarExpr(E->getArg(0));
16184 llvm::Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType);
16185 return Builder.CreateCall(F, X);
16186 }
16187
16188 // Fastmath by default
16189 case PPC::BI__builtin_ppc_recipdivf:
16190 case PPC::BI__builtin_ppc_recipdivd:
16191 case PPC::BI__builtin_ppc_rsqrtf:
16192 case PPC::BI__builtin_ppc_rsqrtd: {
16193 FastMathFlags FMF = Builder.getFastMathFlags();
16194 Builder.getFastMathFlags().setFast();
16195 llvm::Type *ResultType = ConvertType(E->getType());
16196 Value *X = EmitScalarExpr(E->getArg(0));
16197
16198 if (BuiltinID == PPC::BI__builtin_ppc_recipdivf ||
16199 BuiltinID == PPC::BI__builtin_ppc_recipdivd) {
16200 Value *Y = EmitScalarExpr(E->getArg(1));
16201 Value *FDiv = Builder.CreateFDiv(X, Y, "recipdiv");
16202 Builder.getFastMathFlags() &= (FMF);
16203 return FDiv;
16204 }
16205 auto *One = ConstantFP::get(ResultType, 1.0);
16206 llvm::Function *F = CGM.getIntrinsic(Intrinsic::sqrt, ResultType);
16207 Value *FDiv = Builder.CreateFDiv(One, Builder.CreateCall(F, X), "rsqrt");
16208 Builder.getFastMathFlags() &= (FMF);
16209 return FDiv;
16210 }
16211 case PPC::BI__builtin_ppc_alignx: {
16212 Value *Op0 = EmitScalarExpr(E->getArg(0));
16213 Value *Op1 = EmitScalarExpr(E->getArg(1));
16214 ConstantInt *AlignmentCI = cast<ConstantInt>(Op0);
16215 if (AlignmentCI->getValue().ugt(llvm::Value::MaximumAlignment))
16216 AlignmentCI = ConstantInt::get(AlignmentCI->getType(),
16217 llvm::Value::MaximumAlignment);
16218
16219 emitAlignmentAssumption(Op1, E->getArg(1),
16220 /*The expr loc is sufficient.*/ SourceLocation(),
16221 AlignmentCI, nullptr);
16222 return Op1;
16223 }
16224 case PPC::BI__builtin_ppc_rdlam: {
16225 Value *Op0 = EmitScalarExpr(E->getArg(0));
16226 Value *Op1 = EmitScalarExpr(E->getArg(1));
16227 Value *Op2 = EmitScalarExpr(E->getArg(2));
16228 llvm::Type *Ty = Op0->getType();
16229 Value *ShiftAmt = Builder.CreateIntCast(Op1, Ty, false);
16230 Function *F = CGM.getIntrinsic(Intrinsic::fshl, Ty);
16231 Value *Rotate = Builder.CreateCall(F, {Op0, Op0, ShiftAmt});
16232 return Builder.CreateAnd(Rotate, Op2);
16233 }
16234 case PPC::BI__builtin_ppc_load2r: {
16235 Function *F = CGM.getIntrinsic(Intrinsic::ppc_load2r);
16236 Value *Op0 = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int8PtrTy);
16237 Value *LoadIntrinsic = Builder.CreateCall(F, {Op0});
16238 return Builder.CreateTrunc(LoadIntrinsic, Int16Ty);
16239 }
16240 // FMA variations
16241 case PPC::BI__builtin_ppc_fnmsub:
16242 case PPC::BI__builtin_ppc_fnmsubs:
16243 case PPC::BI__builtin_vsx_xvmaddadp:
16244 case PPC::BI__builtin_vsx_xvmaddasp:
16245 case PPC::BI__builtin_vsx_xvnmaddadp:
16246 case PPC::BI__builtin_vsx_xvnmaddasp:
16247 case PPC::BI__builtin_vsx_xvmsubadp:
16248 case PPC::BI__builtin_vsx_xvmsubasp:
16249 case PPC::BI__builtin_vsx_xvnmsubadp:
16250 case PPC::BI__builtin_vsx_xvnmsubasp: {
16251 llvm::Type *ResultType = ConvertType(E->getType());
16252 Value *X = EmitScalarExpr(E->getArg(0));
16253 Value *Y = EmitScalarExpr(E->getArg(1));
16254 Value *Z = EmitScalarExpr(E->getArg(2));
16255 llvm::Function *F;
16256 if (Builder.getIsFPConstrained())
16257 F = CGM.getIntrinsic(Intrinsic::experimental_constrained_fma, ResultType);
16258 else
16259 F = CGM.getIntrinsic(Intrinsic::fma, ResultType);
16260 switch (BuiltinID) {
16261 case PPC::BI__builtin_vsx_xvmaddadp:
16262 case PPC::BI__builtin_vsx_xvmaddasp:
16263 if (Builder.getIsFPConstrained())
16264 return Builder.CreateConstrainedFPCall(F, {X, Y, Z});
16265 else
16266 return Builder.CreateCall(F, {X, Y, Z});
16267 case PPC::BI__builtin_vsx_xvnmaddadp:
16268 case PPC::BI__builtin_vsx_xvnmaddasp:
16269 if (Builder.getIsFPConstrained())
16270 return Builder.CreateFNeg(
16271 Builder.CreateConstrainedFPCall(F, {X, Y, Z}), "neg");
16272 else
16273 return Builder.CreateFNeg(Builder.CreateCall(F, {X, Y, Z}), "neg");
16274 case PPC::BI__builtin_vsx_xvmsubadp:
16275 case PPC::BI__builtin_vsx_xvmsubasp:
16276 if (Builder.getIsFPConstrained())
16277 return Builder.CreateConstrainedFPCall(
16278 F, {X, Y, Builder.CreateFNeg(Z, "neg")});
16279 else
16280 return Builder.CreateCall(F, {X, Y, Builder.CreateFNeg(Z, "neg")});
16281 case PPC::BI__builtin_ppc_fnmsub:
16282 case PPC::BI__builtin_ppc_fnmsubs:
16283 case PPC::BI__builtin_vsx_xvnmsubadp:
16284 case PPC::BI__builtin_vsx_xvnmsubasp:
16285 if (Builder.getIsFPConstrained())
16286 return Builder.CreateFNeg(
16287 Builder.CreateConstrainedFPCall(
16288 F, {X, Y, Builder.CreateFNeg(Z, "neg")}),
16289 "neg");
16290 else
16291 return Builder.CreateCall(
16292 CGM.getIntrinsic(Intrinsic::ppc_fnmsub, ResultType), {X, Y, Z});
16293 }
16294 llvm_unreachable("Unknown FMA operation");
16295 return nullptr; // Suppress no-return warning
16296 }
16297
16298 case PPC::BI__builtin_vsx_insertword: {
16299 Value *Op0 = EmitScalarExpr(E->getArg(0));
16300 Value *Op1 = EmitScalarExpr(E->getArg(1));
16301 Value *Op2 = EmitScalarExpr(E->getArg(2));
16302 llvm::Function *F = CGM.getIntrinsic(Intrinsic::ppc_vsx_xxinsertw);
16303
16304 // Third argument is a compile time constant int. It must be clamped to
16305 // to the range [0, 12].
16306 ConstantInt *ArgCI = dyn_cast<ConstantInt>(Op2);
16307 assert(ArgCI &&
16308 "Third arg to xxinsertw intrinsic must be constant integer");
16309 const int64_t MaxIndex = 12;
16310 int64_t Index = std::clamp(ArgCI->getSExtValue(), (int64_t)0, MaxIndex);
16311
16312 // The builtin semantics don't exactly match the xxinsertw instructions
16313 // semantics (which ppc_vsx_xxinsertw follows). The builtin extracts the
16314 // word from the first argument, and inserts it in the second argument. The
16315 // instruction extracts the word from its second input register and inserts
16316 // it into its first input register, so swap the first and second arguments.
16317 std::swap(Op0, Op1);
16318
16319 // Need to cast the second argument from a vector of unsigned int to a
16320 // vector of long long.
16321 Op1 = Builder.CreateBitCast(Op1, llvm::FixedVectorType::get(Int64Ty, 2));
16322
16323 if (getTarget().isLittleEndian()) {
16324 // Reverse the double words in the vector we will extract from.
16325 Op0 = Builder.CreateBitCast(Op0, llvm::FixedVectorType::get(Int64Ty, 2));
16326 Op0 = Builder.CreateShuffleVector(Op0, Op0, ArrayRef<int>{1, 0});
16327
16328 // Reverse the index.
16329 Index = MaxIndex - Index;
16330 }
16331
16332 // Intrinsic expects the first arg to be a vector of int.
16333 Op0 = Builder.CreateBitCast(Op0, llvm::FixedVectorType::get(Int32Ty, 4));
16334 Op2 = ConstantInt::getSigned(Int32Ty, Index);
16335 return Builder.CreateCall(F, {Op0, Op1, Op2});
16336 }
16337
16338 case PPC::BI__builtin_vsx_extractuword: {
16339 Value *Op0 = EmitScalarExpr(E->getArg(0));
16340 Value *Op1 = EmitScalarExpr(E->getArg(1));
16341 llvm::Function *F = CGM.getIntrinsic(Intrinsic::ppc_vsx_xxextractuw);
16342
16343 // Intrinsic expects the first argument to be a vector of doublewords.
16344 Op0 = Builder.CreateBitCast(Op0, llvm::FixedVectorType::get(Int64Ty, 2));
16345
16346 // The second argument is a compile time constant int that needs to
16347 // be clamped to the range [0, 12].
16348 ConstantInt *ArgCI = dyn_cast<ConstantInt>(Op1);
16349 assert(ArgCI &&
16350 "Second Arg to xxextractuw intrinsic must be a constant integer!");
16351 const int64_t MaxIndex = 12;
16352 int64_t Index = std::clamp(ArgCI->getSExtValue(), (int64_t)0, MaxIndex);
16353
16354 if (getTarget().isLittleEndian()) {
16355 // Reverse the index.
16356 Index = MaxIndex - Index;
16357 Op1 = ConstantInt::getSigned(Int32Ty, Index);
16358
16359 // Emit the call, then reverse the double words of the results vector.
16360 Value *Call = Builder.CreateCall(F, {Op0, Op1});
16361
16362 Value *ShuffleCall =
16363 Builder.CreateShuffleVector(Call, Call, ArrayRef<int>{1, 0});
16364 return ShuffleCall;
16365 } else {
16366 Op1 = ConstantInt::getSigned(Int32Ty, Index);
16367 return Builder.CreateCall(F, {Op0, Op1});
16368 }
16369 }
16370
16371 case PPC::BI__builtin_vsx_xxpermdi: {
16372 Value *Op0 = EmitScalarExpr(E->getArg(0));
16373 Value *Op1 = EmitScalarExpr(E->getArg(1));
16374 Value *Op2 = EmitScalarExpr(E->getArg(2));
16375 ConstantInt *ArgCI = dyn_cast<ConstantInt>(Op2);
16376 assert(ArgCI && "Third arg must be constant integer!");
16377
16378 unsigned Index = ArgCI->getZExtValue();
16379 Op0 = Builder.CreateBitCast(Op0, llvm::FixedVectorType::get(Int64Ty, 2));
16380 Op1 = Builder.CreateBitCast(Op1, llvm::FixedVectorType::get(Int64Ty, 2));
16381
16382 // Account for endianness by treating this as just a shuffle. So we use the
16383 // same indices for both LE and BE in order to produce expected results in
16384 // both cases.
16385 int ElemIdx0 = (Index & 2) >> 1;
16386 int ElemIdx1 = 2 + (Index & 1);
16387
16388 int ShuffleElts[2] = {ElemIdx0, ElemIdx1};
16389 Value *ShuffleCall = Builder.CreateShuffleVector(Op0, Op1, ShuffleElts);
16390 QualType BIRetType = E->getType();
16391 auto RetTy = ConvertType(BIRetType);
16392 return Builder.CreateBitCast(ShuffleCall, RetTy);
16393 }
16394
16395 case PPC::BI__builtin_vsx_xxsldwi: {
16396 Value *Op0 = EmitScalarExpr(E->getArg(0));
16397 Value *Op1 = EmitScalarExpr(E->getArg(1));
16398 Value *Op2 = EmitScalarExpr(E->getArg(2));
16399 ConstantInt *ArgCI = dyn_cast<ConstantInt>(Op2);
16400 assert(ArgCI && "Third argument must be a compile time constant");
16401 unsigned Index = ArgCI->getZExtValue() & 0x3;
16402 Op0 = Builder.CreateBitCast(Op0, llvm::FixedVectorType::get(Int32Ty, 4));
16403 Op1 = Builder.CreateBitCast(Op1, llvm::FixedVectorType::get(Int32Ty, 4));
16404
16405 // Create a shuffle mask
16406 int ElemIdx0;
16407 int ElemIdx1;
16408 int ElemIdx2;
16409 int ElemIdx3;
16410 if (getTarget().isLittleEndian()) {
16411 // Little endian element N comes from element 8+N-Index of the
16412 // concatenated wide vector (of course, using modulo arithmetic on
16413 // the total number of elements).
16414 ElemIdx0 = (8 - Index) % 8;
16415 ElemIdx1 = (9 - Index) % 8;
16416 ElemIdx2 = (10 - Index) % 8;
16417 ElemIdx3 = (11 - Index) % 8;
16418 } else {
16419 // Big endian ElemIdx<N> = Index + N
16420 ElemIdx0 = Index;
16421 ElemIdx1 = Index + 1;
16422 ElemIdx2 = Index + 2;
16423 ElemIdx3 = Index + 3;
16424 }
16425
16426 int ShuffleElts[4] = {ElemIdx0, ElemIdx1, ElemIdx2, ElemIdx3};
16427 Value *ShuffleCall = Builder.CreateShuffleVector(Op0, Op1, ShuffleElts);
16428 QualType BIRetType = E->getType();
16429 auto RetTy = ConvertType(BIRetType);
16430 return Builder.CreateBitCast(ShuffleCall, RetTy);
16431 }
16432
16433 case PPC::BI__builtin_pack_vector_int128: {
16434 Value *Op0 = EmitScalarExpr(E->getArg(0));
16435 Value *Op1 = EmitScalarExpr(E->getArg(1));
16436 bool isLittleEndian = getTarget().isLittleEndian();
16437 Value *PoisonValue =
16438 llvm::PoisonValue::get(llvm::FixedVectorType::get(Op0->getType(), 2));
16439 Value *Res = Builder.CreateInsertElement(
16440 PoisonValue, Op0, (uint64_t)(isLittleEndian ? 1 : 0));
16441 Res = Builder.CreateInsertElement(Res, Op1,
16442 (uint64_t)(isLittleEndian ? 0 : 1));
16443 return Builder.CreateBitCast(Res, ConvertType(E->getType()));
16444 }
16445
16446 case PPC::BI__builtin_unpack_vector_int128: {
16447 Value *Op0 = EmitScalarExpr(E->getArg(0));
16448 Value *Op1 = EmitScalarExpr(E->getArg(1));
16449 ConstantInt *Index = cast<ConstantInt>(Op1);
16450 Value *Unpacked = Builder.CreateBitCast(
16451 Op0, llvm::FixedVectorType::get(ConvertType(E->getType()), 2));
16452
16453 if (getTarget().isLittleEndian())
16454 Index = ConstantInt::get(Index->getType(), 1 - Index->getZExtValue());
16455
16456 return Builder.CreateExtractElement(Unpacked, Index);
16457 }
16458
16459 case PPC::BI__builtin_ppc_sthcx: {
16460 llvm::Function *F = CGM.getIntrinsic(Intrinsic::ppc_sthcx);
16461 Value *Op0 = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int8PtrTy);
16462 Value *Op1 = Builder.CreateSExt(EmitScalarExpr(E->getArg(1)), Int32Ty);
16463 return Builder.CreateCall(F, {Op0, Op1});
16464 }
16465
16466 // The PPC MMA builtins take a pointer to a __vector_quad as an argument.
16467 // Some of the MMA instructions accumulate their result into an existing
16468 // accumulator whereas the others generate a new accumulator. So we need to
16469 // use custom code generation to expand a builtin call with a pointer to a
16470 // load (if the corresponding instruction accumulates its result) followed by
16471 // the call to the intrinsic and a store of the result.
16472 #define CUSTOM_BUILTIN(Name, Intr, Types, Accumulate) \
16473 case PPC::BI__builtin_##Name:
16474 #include "clang/Basic/BuiltinsPPC.def"
16475 {
16476 SmallVector<Value *, 4> Ops;
16477 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++)
16478 if (E->getArg(i)->getType()->isArrayType())
16479 Ops.push_back(EmitArrayToPointerDecay(E->getArg(i)).getPointer());
16480 else
16481 Ops.push_back(EmitScalarExpr(E->getArg(i)));
16482 // The first argument of these two builtins is a pointer used to store their
16483 // result. However, the llvm intrinsics return their result in multiple
16484 // return values. So, here we emit code extracting these values from the
16485 // intrinsic results and storing them using that pointer.
16486 if (BuiltinID == PPC::BI__builtin_mma_disassemble_acc ||
16487 BuiltinID == PPC::BI__builtin_vsx_disassemble_pair ||
16488 BuiltinID == PPC::BI__builtin_mma_disassemble_pair) {
16489 unsigned NumVecs = 2;
16490 auto Intrinsic = Intrinsic::ppc_vsx_disassemble_pair;
16491 if (BuiltinID == PPC::BI__builtin_mma_disassemble_acc) {
16492 NumVecs = 4;
16493 Intrinsic = Intrinsic::ppc_mma_disassemble_acc;
16494 }
16495 llvm::Function *F = CGM.getIntrinsic(Intrinsic);
16496 Address Addr = EmitPointerWithAlignment(E->getArg(1));
16497 Value *Vec = Builder.CreateLoad(Addr);
16498 Value *Call = Builder.CreateCall(F, {Vec});
16499 llvm::Type *VTy = llvm::FixedVectorType::get(Int8Ty, 16);
16500 Value *Ptr = Builder.CreateBitCast(Ops[0], VTy->getPointerTo());
16501 for (unsigned i=0; i<NumVecs; i++) {
16502 Value *Vec = Builder.CreateExtractValue(Call, i);
16503 llvm::ConstantInt* Index = llvm::ConstantInt::get(IntTy, i);
16504 Value *GEP = Builder.CreateInBoundsGEP(VTy, Ptr, Index);
16505 Builder.CreateAlignedStore(Vec, GEP, MaybeAlign(16));
16506 }
16507 return Call;
16508 }
16509 if (BuiltinID == PPC::BI__builtin_vsx_build_pair ||
16510 BuiltinID == PPC::BI__builtin_mma_build_acc) {
16511 // Reverse the order of the operands for LE, so the
16512 // same builtin call can be used on both LE and BE
16513 // without the need for the programmer to swap operands.
16514 // The operands are reversed starting from the second argument,
16515 // the first operand is the pointer to the pair/accumulator
16516 // that is being built.
16517 if (getTarget().isLittleEndian())
16518 std::reverse(Ops.begin() + 1, Ops.end());
16519 }
16520 bool Accumulate;
16521 switch (BuiltinID) {
16522 #define CUSTOM_BUILTIN(Name, Intr, Types, Acc) \
16523 case PPC::BI__builtin_##Name: \
16524 ID = Intrinsic::ppc_##Intr; \
16525 Accumulate = Acc; \
16526 break;
16527 #include "clang/Basic/BuiltinsPPC.def"
16528 }
16529 if (BuiltinID == PPC::BI__builtin_vsx_lxvp ||
16530 BuiltinID == PPC::BI__builtin_vsx_stxvp ||
16531 BuiltinID == PPC::BI__builtin_mma_lxvp ||
16532 BuiltinID == PPC::BI__builtin_mma_stxvp) {
16533 if (BuiltinID == PPC::BI__builtin_vsx_lxvp ||
16534 BuiltinID == PPC::BI__builtin_mma_lxvp) {
16535 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy);
16536 Ops[0] = Builder.CreateGEP(Int8Ty, Ops[1], Ops[0]);
16537 } else {
16538 Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy);
16539 Ops[1] = Builder.CreateGEP(Int8Ty, Ops[2], Ops[1]);
16540 }
16541 Ops.pop_back();
16542 llvm::Function *F = CGM.getIntrinsic(ID);
16543 return Builder.CreateCall(F, Ops, "");
16544 }
16545 SmallVector<Value*, 4> CallOps;
16546 if (Accumulate) {
16547 Address Addr = EmitPointerWithAlignment(E->getArg(0));
16548 Value *Acc = Builder.CreateLoad(Addr);
16549 CallOps.push_back(Acc);
16550 }
16551 for (unsigned i=1; i<Ops.size(); i++)
16552 CallOps.push_back(Ops[i]);
16553 llvm::Function *F = CGM.getIntrinsic(ID);
16554 Value *Call = Builder.CreateCall(F, CallOps);
16555 return Builder.CreateAlignedStore(Call, Ops[0], MaybeAlign(64));
16556 }
16557
16558 case PPC::BI__builtin_ppc_compare_and_swap:
16559 case PPC::BI__builtin_ppc_compare_and_swaplp: {
16560 Address Addr = EmitPointerWithAlignment(E->getArg(0));
16561 Address OldValAddr = EmitPointerWithAlignment(E->getArg(1));
16562 Value *OldVal = Builder.CreateLoad(OldValAddr);
16563 QualType AtomicTy = E->getArg(0)->getType()->getPointeeType();
16564 LValue LV = MakeAddrLValue(Addr, AtomicTy);
16565 Value *Op2 = EmitScalarExpr(E->getArg(2));
16566 auto Pair = EmitAtomicCompareExchange(
16567 LV, RValue::get(OldVal), RValue::get(Op2), E->getExprLoc(),
16568 llvm::AtomicOrdering::Monotonic, llvm::AtomicOrdering::Monotonic, true);
16569 // Unlike c11's atomic_compare_exchange, according to
16570 // https://www.ibm.com/docs/en/xl-c-and-cpp-aix/16.1?topic=functions-compare-swap-compare-swaplp
16571 // > In either case, the contents of the memory location specified by addr
16572 // > are copied into the memory location specified by old_val_addr.
16573 // But it hasn't specified storing to OldValAddr is atomic or not and
16574 // which order to use. Now following XL's codegen, treat it as a normal
16575 // store.
16576 Value *LoadedVal = Pair.first.getScalarVal();
16577 Builder.CreateStore(LoadedVal, OldValAddr);
16578 return Builder.CreateZExt(Pair.second, Builder.getInt32Ty());
16579 }
16580 case PPC::BI__builtin_ppc_fetch_and_add:
16581 case PPC::BI__builtin_ppc_fetch_and_addlp: {
16582 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Add, E,
16583 llvm::AtomicOrdering::Monotonic);
16584 }
16585 case PPC::BI__builtin_ppc_fetch_and_and:
16586 case PPC::BI__builtin_ppc_fetch_and_andlp: {
16587 return MakeBinaryAtomicValue(*this, AtomicRMWInst::And, E,
16588 llvm::AtomicOrdering::Monotonic);
16589 }
16590
16591 case PPC::BI__builtin_ppc_fetch_and_or:
16592 case PPC::BI__builtin_ppc_fetch_and_orlp: {
16593 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Or, E,
16594 llvm::AtomicOrdering::Monotonic);
16595 }
16596 case PPC::BI__builtin_ppc_fetch_and_swap:
16597 case PPC::BI__builtin_ppc_fetch_and_swaplp: {
16598 return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xchg, E,
16599 llvm::AtomicOrdering::Monotonic);
16600 }
16601 case PPC::BI__builtin_ppc_ldarx:
16602 case PPC::BI__builtin_ppc_lwarx:
16603 case PPC::BI__builtin_ppc_lharx:
16604 case PPC::BI__builtin_ppc_lbarx:
16605 return emitPPCLoadReserveIntrinsic(*this, BuiltinID, E);
16606 case PPC::BI__builtin_ppc_mfspr: {
16607 Value *Op0 = EmitScalarExpr(E->getArg(0));
16608 llvm::Type *RetType = CGM.getDataLayout().getTypeSizeInBits(VoidPtrTy) == 32
16609 ? Int32Ty
16610 : Int64Ty;
16611 Function *F = CGM.getIntrinsic(Intrinsic::ppc_mfspr, RetType);
16612 return Builder.CreateCall(F, {Op0});
16613 }
16614 case PPC::BI__builtin_ppc_mtspr: {
16615 Value *Op0 = EmitScalarExpr(E->getArg(0));
16616 Value *Op1 = EmitScalarExpr(E->getArg(1));
16617 llvm::Type *RetType = CGM.getDataLayout().getTypeSizeInBits(VoidPtrTy) == 32
16618 ? Int32Ty
16619 : Int64Ty;
16620 Function *F = CGM.getIntrinsic(Intrinsic::ppc_mtspr, RetType);
16621 return Builder.CreateCall(F, {Op0, Op1});
16622 }
16623 case PPC::BI__builtin_ppc_popcntb: {
16624 Value *ArgValue = EmitScalarExpr(E->getArg(0));
16625 llvm::Type *ArgType = ArgValue->getType();
16626 Function *F = CGM.getIntrinsic(Intrinsic::ppc_popcntb, {ArgType, ArgType});
16627 return Builder.CreateCall(F, {ArgValue}, "popcntb");
16628 }
16629 case PPC::BI__builtin_ppc_mtfsf: {
16630 // The builtin takes a uint32 that needs to be cast to an
16631 // f64 to be passed to the intrinsic.
16632 Value *Op0 = EmitScalarExpr(E->getArg(0));
16633 Value *Op1 = EmitScalarExpr(E->getArg(1));
16634 Value *Cast = Builder.CreateUIToFP(Op1, DoubleTy);
16635 llvm::Function *F = CGM.getIntrinsic(Intrinsic::ppc_mtfsf);
16636 return Builder.CreateCall(F, {Op0, Cast}, "");
16637 }
16638
16639 case PPC::BI__builtin_ppc_swdiv_nochk:
16640 case PPC::BI__builtin_ppc_swdivs_nochk: {
16641 Value *Op0 = EmitScalarExpr(E->getArg(0));
16642 Value *Op1 = EmitScalarExpr(E->getArg(1));
16643 FastMathFlags FMF = Builder.getFastMathFlags();
16644 Builder.getFastMathFlags().setFast();
16645 Value *FDiv = Builder.CreateFDiv(Op0, Op1, "swdiv_nochk");
16646 Builder.getFastMathFlags() &= (FMF);
16647 return FDiv;
16648 }
16649 case PPC::BI__builtin_ppc_fric:
16650 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(
16651 *this, E, Intrinsic::rint,
16652 Intrinsic::experimental_constrained_rint))
16653 .getScalarVal();
16654 case PPC::BI__builtin_ppc_frim:
16655 case PPC::BI__builtin_ppc_frims:
16656 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(
16657 *this, E, Intrinsic::floor,
16658 Intrinsic::experimental_constrained_floor))
16659 .getScalarVal();
16660 case PPC::BI__builtin_ppc_frin:
16661 case PPC::BI__builtin_ppc_frins:
16662 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(
16663 *this, E, Intrinsic::round,
16664 Intrinsic::experimental_constrained_round))
16665 .getScalarVal();
16666 case PPC::BI__builtin_ppc_frip:
16667 case PPC::BI__builtin_ppc_frips:
16668 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(
16669 *this, E, Intrinsic::ceil,
16670 Intrinsic::experimental_constrained_ceil))
16671 .getScalarVal();
16672 case PPC::BI__builtin_ppc_friz:
16673 case PPC::BI__builtin_ppc_frizs:
16674 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(
16675 *this, E, Intrinsic::trunc,
16676 Intrinsic::experimental_constrained_trunc))
16677 .getScalarVal();
16678 case PPC::BI__builtin_ppc_fsqrt:
16679 case PPC::BI__builtin_ppc_fsqrts:
16680 return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(
16681 *this, E, Intrinsic::sqrt,
16682 Intrinsic::experimental_constrained_sqrt))
16683 .getScalarVal();
16684 case PPC::BI__builtin_ppc_test_data_class: {
16685 Value *Op0 = EmitScalarExpr(E->getArg(0));
16686 Value *Op1 = EmitScalarExpr(E->getArg(1));
16687 return Builder.CreateCall(
16688 CGM.getIntrinsic(Intrinsic::ppc_test_data_class, Op0->getType()),
16689 {Op0, Op1}, "test_data_class");
16690 }
16691 case PPC::BI__builtin_ppc_maxfe: {
16692 Value *Op0 = EmitScalarExpr(E->getArg(0));
16693 Value *Op1 = EmitScalarExpr(E->getArg(1));
16694 Value *Op2 = EmitScalarExpr(E->getArg(2));
16695 Value *Op3 = EmitScalarExpr(E->getArg(3));
16696 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::ppc_maxfe),
16697 {Op0, Op1, Op2, Op3});
16698 }
16699 case PPC::BI__builtin_ppc_maxfl: {
16700 Value *Op0 = EmitScalarExpr(E->getArg(0));
16701 Value *Op1 = EmitScalarExpr(E->getArg(1));
16702 Value *Op2 = EmitScalarExpr(E->getArg(2));
16703 Value *Op3 = EmitScalarExpr(E->getArg(3));
16704 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::ppc_maxfl),
16705 {Op0, Op1, Op2, Op3});
16706 }
16707 case PPC::BI__builtin_ppc_maxfs: {
16708 Value *Op0 = EmitScalarExpr(E->getArg(0));
16709 Value *Op1 = EmitScalarExpr(E->getArg(1));
16710 Value *Op2 = EmitScalarExpr(E->getArg(2));
16711 Value *Op3 = EmitScalarExpr(E->getArg(3));
16712 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::ppc_maxfs),
16713 {Op0, Op1, Op2, Op3});
16714 }
16715 case PPC::BI__builtin_ppc_minfe: {
16716 Value *Op0 = EmitScalarExpr(E->getArg(0));
16717 Value *Op1 = EmitScalarExpr(E->getArg(1));
16718 Value *Op2 = EmitScalarExpr(E->getArg(2));
16719 Value *Op3 = EmitScalarExpr(E->getArg(3));
16720 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::ppc_minfe),
16721 {Op0, Op1, Op2, Op3});
16722 }
16723 case PPC::BI__builtin_ppc_minfl: {
16724 Value *Op0 = EmitScalarExpr(E->getArg(0));
16725 Value *Op1 = EmitScalarExpr(E->getArg(1));
16726 Value *Op2 = EmitScalarExpr(E->getArg(2));
16727 Value *Op3 = EmitScalarExpr(E->getArg(3));
16728 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::ppc_minfl),
16729 {Op0, Op1, Op2, Op3});
16730 }
16731 case PPC::BI__builtin_ppc_minfs: {
16732 Value *Op0 = EmitScalarExpr(E->getArg(0));
16733 Value *Op1 = EmitScalarExpr(E->getArg(1));
16734 Value *Op2 = EmitScalarExpr(E->getArg(2));
16735 Value *Op3 = EmitScalarExpr(E->getArg(3));
16736 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::ppc_minfs),
16737 {Op0, Op1, Op2, Op3});
16738 }
16739 case PPC::BI__builtin_ppc_swdiv:
16740 case PPC::BI__builtin_ppc_swdivs: {
16741 Value *Op0 = EmitScalarExpr(E->getArg(0));
16742 Value *Op1 = EmitScalarExpr(E->getArg(1));
16743 return Builder.CreateFDiv(Op0, Op1, "swdiv");
16744 }
16745 }
16746 }
16747
16748 namespace {
16749 // If \p E is not null pointer, insert address space cast to match return
16750 // type of \p E if necessary.
16751 Value *EmitAMDGPUDispatchPtr(CodeGenFunction &CGF,
16752 const CallExpr *E = nullptr) {
16753 auto *F = CGF.CGM.getIntrinsic(Intrinsic::amdgcn_dispatch_ptr);
16754 auto *Call = CGF.Builder.CreateCall(F);
16755 Call->addRetAttr(
16756 Attribute::getWithDereferenceableBytes(Call->getContext(), 64));
16757 Call->addRetAttr(Attribute::getWithAlignment(Call->getContext(), Align(4)));
16758 if (!E)
16759 return Call;
16760 QualType BuiltinRetType = E->getType();
16761 auto *RetTy = cast<llvm::PointerType>(CGF.ConvertType(BuiltinRetType));
16762 if (RetTy == Call->getType())
16763 return Call;
16764 return CGF.Builder.CreateAddrSpaceCast(Call, RetTy);
16765 }
16766
16767 Value *EmitAMDGPUImplicitArgPtr(CodeGenFunction &CGF) {
16768 auto *F = CGF.CGM.getIntrinsic(Intrinsic::amdgcn_implicitarg_ptr);
16769 auto *Call = CGF.Builder.CreateCall(F);
16770 Call->addRetAttr(
16771 Attribute::getWithDereferenceableBytes(Call->getContext(), 256));
16772 Call->addRetAttr(Attribute::getWithAlignment(Call->getContext(), Align(8)));
16773 return Call;
16774 }
16775
16776 // \p Index is 0, 1, and 2 for x, y, and z dimension, respectively.
16777 Value *EmitAMDGPUWorkGroupSize(CodeGenFunction &CGF, unsigned Index) {
16778 bool IsCOV_5 = CGF.getTarget().getTargetOpts().CodeObjectVersion ==
16779 clang::TargetOptions::COV_5;
16780 Constant *Offset;
16781 Value *DP;
16782 if (IsCOV_5) {
16783 // Indexing the implicit kernarg segment.
16784 Offset = llvm::ConstantInt::get(CGF.Int32Ty, 12 + Index * 2);
16785 DP = EmitAMDGPUImplicitArgPtr(CGF);
16786 } else {
16787 // Indexing the HSA kernel_dispatch_packet struct.
16788 Offset = llvm::ConstantInt::get(CGF.Int32Ty, 4 + Index * 2);
16789 DP = EmitAMDGPUDispatchPtr(CGF);
16790 }
16791
16792 auto *GEP = CGF.Builder.CreateGEP(CGF.Int8Ty, DP, Offset);
16793 auto *DstTy =
16794 CGF.Int16Ty->getPointerTo(GEP->getType()->getPointerAddressSpace());
16795 auto *Cast = CGF.Builder.CreateBitCast(GEP, DstTy);
16796 auto *LD = CGF.Builder.CreateLoad(
16797 Address(Cast, CGF.Int16Ty, CharUnits::fromQuantity(2)));
16798 llvm::MDBuilder MDHelper(CGF.getLLVMContext());
16799 llvm::MDNode *RNode = MDHelper.createRange(APInt(16, 1),
16800 APInt(16, CGF.getTarget().getMaxOpenCLWorkGroupSize() + 1));
16801 LD->setMetadata(llvm::LLVMContext::MD_range, RNode);
16802 LD->setMetadata(llvm::LLVMContext::MD_noundef,
16803 llvm::MDNode::get(CGF.getLLVMContext(), std::nullopt));
16804 LD->setMetadata(llvm::LLVMContext::MD_invariant_load,
16805 llvm::MDNode::get(CGF.getLLVMContext(), std::nullopt));
16806 return LD;
16807 }
16808
16809 // \p Index is 0, 1, and 2 for x, y, and z dimension, respectively.
16810 Value *EmitAMDGPUGridSize(CodeGenFunction &CGF, unsigned Index) {
16811 const unsigned XOffset = 12;
16812 auto *DP = EmitAMDGPUDispatchPtr(CGF);
16813 // Indexing the HSA kernel_dispatch_packet struct.
16814 auto *Offset = llvm::ConstantInt::get(CGF.Int32Ty, XOffset + Index * 4);
16815 auto *GEP = CGF.Builder.CreateGEP(CGF.Int8Ty, DP, Offset);
16816 auto *DstTy =
16817 CGF.Int32Ty->getPointerTo(GEP->getType()->getPointerAddressSpace());
16818 auto *Cast = CGF.Builder.CreateBitCast(GEP, DstTy);
16819 auto *LD = CGF.Builder.CreateLoad(
16820 Address(Cast, CGF.Int32Ty, CharUnits::fromQuantity(4)));
16821 LD->setMetadata(llvm::LLVMContext::MD_invariant_load,
16822 llvm::MDNode::get(CGF.getLLVMContext(), std::nullopt));
16823 return LD;
16824 }
16825 } // namespace
16826
16827 // For processing memory ordering and memory scope arguments of various
16828 // amdgcn builtins.
16829 // \p Order takes a C++11 comptabile memory-ordering specifier and converts
16830 // it into LLVM's memory ordering specifier using atomic C ABI, and writes
16831 // to \p AO. \p Scope takes a const char * and converts it into AMDGCN
16832 // specific SyncScopeID and writes it to \p SSID.
16833 void CodeGenFunction::ProcessOrderScopeAMDGCN(Value *Order, Value *Scope,
16834 llvm::AtomicOrdering &AO,
16835 llvm::SyncScope::ID &SSID) {
16836 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue();
16837
16838 // Map C11/C++11 memory ordering to LLVM memory ordering
16839 assert(llvm::isValidAtomicOrderingCABI(ord));
16840 switch (static_cast<llvm::AtomicOrderingCABI>(ord)) {
16841 case llvm::AtomicOrderingCABI::acquire:
16842 case llvm::AtomicOrderingCABI::consume:
16843 AO = llvm::AtomicOrdering::Acquire;
16844 break;
16845 case llvm::AtomicOrderingCABI::release:
16846 AO = llvm::AtomicOrdering::Release;
16847 break;
16848 case llvm::AtomicOrderingCABI::acq_rel:
16849 AO = llvm::AtomicOrdering::AcquireRelease;
16850 break;
16851 case llvm::AtomicOrderingCABI::seq_cst:
16852 AO = llvm::AtomicOrdering::SequentiallyConsistent;
16853 break;
16854 case llvm::AtomicOrderingCABI::relaxed:
16855 AO = llvm::AtomicOrdering::Monotonic;
16856 break;
16857 }
16858
16859 StringRef scp;
16860 llvm::getConstantStringInfo(Scope, scp);
16861 SSID = getLLVMContext().getOrInsertSyncScopeID(scp);
16862 }
16863
16864 Value *CodeGenFunction::EmitAMDGPUBuiltinExpr(unsigned BuiltinID,
16865 const CallExpr *E) {
16866 llvm::AtomicOrdering AO = llvm::AtomicOrdering::SequentiallyConsistent;
16867 llvm::SyncScope::ID SSID;
16868 switch (BuiltinID) {
16869 case AMDGPU::BI__builtin_amdgcn_div_scale:
16870 case AMDGPU::BI__builtin_amdgcn_div_scalef: {
16871 // Translate from the intrinsics's struct return to the builtin's out
16872 // argument.
16873
16874 Address FlagOutPtr = EmitPointerWithAlignment(E->getArg(3));
16875
16876 llvm::Value *X = EmitScalarExpr(E->getArg(0));
16877 llvm::Value *Y = EmitScalarExpr(E->getArg(1));
16878 llvm::Value *Z = EmitScalarExpr(E->getArg(2));
16879
16880 llvm::Function *Callee = CGM.getIntrinsic(Intrinsic::amdgcn_div_scale,
16881 X->getType());
16882
16883 llvm::Value *Tmp = Builder.CreateCall(Callee, {X, Y, Z});
16884
16885 llvm::Value *Result = Builder.CreateExtractValue(Tmp, 0);
16886 llvm::Value *Flag = Builder.CreateExtractValue(Tmp, 1);
16887
16888 llvm::Type *RealFlagType = FlagOutPtr.getElementType();
16889
16890 llvm::Value *FlagExt = Builder.CreateZExt(Flag, RealFlagType);
16891 Builder.CreateStore(FlagExt, FlagOutPtr);
16892 return Result;
16893 }
16894 case AMDGPU::BI__builtin_amdgcn_div_fmas:
16895 case AMDGPU::BI__builtin_amdgcn_div_fmasf: {
16896 llvm::Value *Src0 = EmitScalarExpr(E->getArg(0));
16897 llvm::Value *Src1 = EmitScalarExpr(E->getArg(1));
16898 llvm::Value *Src2 = EmitScalarExpr(E->getArg(2));
16899 llvm::Value *Src3 = EmitScalarExpr(E->getArg(3));
16900
16901 llvm::Function *F = CGM.getIntrinsic(Intrinsic::amdgcn_div_fmas,
16902 Src0->getType());
16903 llvm::Value *Src3ToBool = Builder.CreateIsNotNull(Src3);
16904 return Builder.CreateCall(F, {Src0, Src1, Src2, Src3ToBool});
16905 }
16906
16907 case AMDGPU::BI__builtin_amdgcn_ds_swizzle:
16908 return emitBinaryBuiltin(*this, E, Intrinsic::amdgcn_ds_swizzle);
16909 case AMDGPU::BI__builtin_amdgcn_mov_dpp8:
16910 return emitBinaryBuiltin(*this, E, Intrinsic::amdgcn_mov_dpp8);
16911 case AMDGPU::BI__builtin_amdgcn_mov_dpp:
16912 case AMDGPU::BI__builtin_amdgcn_update_dpp: {
16913 llvm::SmallVector<llvm::Value *, 6> Args;
16914 for (unsigned I = 0; I != E->getNumArgs(); ++I)
16915 Args.push_back(EmitScalarExpr(E->getArg(I)));
16916 assert(Args.size() == 5 || Args.size() == 6);
16917 if (Args.size() == 5)
16918 Args.insert(Args.begin(), llvm::PoisonValue::get(Args[0]->getType()));
16919 Function *F =
16920 CGM.getIntrinsic(Intrinsic::amdgcn_update_dpp, Args[0]->getType());
16921 return Builder.CreateCall(F, Args);
16922 }
16923 case AMDGPU::BI__builtin_amdgcn_div_fixup:
16924 case AMDGPU::BI__builtin_amdgcn_div_fixupf:
16925 case AMDGPU::BI__builtin_amdgcn_div_fixuph:
16926 return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_div_fixup);
16927 case AMDGPU::BI__builtin_amdgcn_trig_preop:
16928 case AMDGPU::BI__builtin_amdgcn_trig_preopf:
16929 return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_trig_preop);
16930 case AMDGPU::BI__builtin_amdgcn_rcp:
16931 case AMDGPU::BI__builtin_amdgcn_rcpf:
16932 case AMDGPU::BI__builtin_amdgcn_rcph:
16933 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rcp);
16934 case AMDGPU::BI__builtin_amdgcn_sqrt:
16935 case AMDGPU::BI__builtin_amdgcn_sqrtf:
16936 case AMDGPU::BI__builtin_amdgcn_sqrth:
16937 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_sqrt);
16938 case AMDGPU::BI__builtin_amdgcn_rsq:
16939 case AMDGPU::BI__builtin_amdgcn_rsqf:
16940 case AMDGPU::BI__builtin_amdgcn_rsqh:
16941 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rsq);
16942 case AMDGPU::BI__builtin_amdgcn_rsq_clamp:
16943 case AMDGPU::BI__builtin_amdgcn_rsq_clampf:
16944 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rsq_clamp);
16945 case AMDGPU::BI__builtin_amdgcn_sinf:
16946 case AMDGPU::BI__builtin_amdgcn_sinh:
16947 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_sin);
16948 case AMDGPU::BI__builtin_amdgcn_cosf:
16949 case AMDGPU::BI__builtin_amdgcn_cosh:
16950 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_cos);
16951 case AMDGPU::BI__builtin_amdgcn_dispatch_ptr:
16952 return EmitAMDGPUDispatchPtr(*this, E);
16953 case AMDGPU::BI__builtin_amdgcn_log_clampf:
16954 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_log_clamp);
16955 case AMDGPU::BI__builtin_amdgcn_ldexp:
16956 case AMDGPU::BI__builtin_amdgcn_ldexpf:
16957 case AMDGPU::BI__builtin_amdgcn_ldexph:
16958 return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_ldexp);
16959 case AMDGPU::BI__builtin_amdgcn_frexp_mant:
16960 case AMDGPU::BI__builtin_amdgcn_frexp_mantf:
16961 case AMDGPU::BI__builtin_amdgcn_frexp_manth:
16962 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_frexp_mant);
16963 case AMDGPU::BI__builtin_amdgcn_frexp_exp:
16964 case AMDGPU::BI__builtin_amdgcn_frexp_expf: {
16965 Value *Src0 = EmitScalarExpr(E->getArg(0));
16966 Function *F = CGM.getIntrinsic(Intrinsic::amdgcn_frexp_exp,
16967 { Builder.getInt32Ty(), Src0->getType() });
16968 return Builder.CreateCall(F, Src0);
16969 }
16970 case AMDGPU::BI__builtin_amdgcn_frexp_exph: {
16971 Value *Src0 = EmitScalarExpr(E->getArg(0));
16972 Function *F = CGM.getIntrinsic(Intrinsic::amdgcn_frexp_exp,
16973 { Builder.getInt16Ty(), Src0->getType() });
16974 return Builder.CreateCall(F, Src0);
16975 }
16976 case AMDGPU::BI__builtin_amdgcn_fract:
16977 case AMDGPU::BI__builtin_amdgcn_fractf:
16978 case AMDGPU::BI__builtin_amdgcn_fracth:
16979 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_fract);
16980 case AMDGPU::BI__builtin_amdgcn_lerp:
16981 return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_lerp);
16982 case AMDGPU::BI__builtin_amdgcn_ubfe:
16983 return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_ubfe);
16984 case AMDGPU::BI__builtin_amdgcn_sbfe:
16985 return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_sbfe);
16986 case AMDGPU::BI__builtin_amdgcn_ballot_w32:
16987 case AMDGPU::BI__builtin_amdgcn_ballot_w64: {
16988 llvm::Type *ResultType = ConvertType(E->getType());
16989 llvm::Value *Src = EmitScalarExpr(E->getArg(0));
16990 Function *F = CGM.getIntrinsic(Intrinsic::amdgcn_ballot, { ResultType });
16991 return Builder.CreateCall(F, { Src });
16992 }
16993 case AMDGPU::BI__builtin_amdgcn_uicmp:
16994 case AMDGPU::BI__builtin_amdgcn_uicmpl:
16995 case AMDGPU::BI__builtin_amdgcn_sicmp:
16996 case AMDGPU::BI__builtin_amdgcn_sicmpl: {
16997 llvm::Value *Src0 = EmitScalarExpr(E->getArg(0));
16998 llvm::Value *Src1 = EmitScalarExpr(E->getArg(1));
16999 llvm::Value *Src2 = EmitScalarExpr(E->getArg(2));
17000
17001 // FIXME-GFX10: How should 32 bit mask be handled?
17002 Function *F = CGM.getIntrinsic(Intrinsic::amdgcn_icmp,
17003 { Builder.getInt64Ty(), Src0->getType() });
17004 return Builder.CreateCall(F, { Src0, Src1, Src2 });
17005 }
17006 case AMDGPU::BI__builtin_amdgcn_fcmp:
17007 case AMDGPU::BI__builtin_amdgcn_fcmpf: {
17008 llvm::Value *Src0 = EmitScalarExpr(E->getArg(0));
17009 llvm::Value *Src1 = EmitScalarExpr(E->getArg(1));
17010 llvm::Value *Src2 = EmitScalarExpr(E->getArg(2));
17011
17012 // FIXME-GFX10: How should 32 bit mask be handled?
17013 Function *F = CGM.getIntrinsic(Intrinsic::amdgcn_fcmp,
17014 { Builder.getInt64Ty(), Src0->getType() });
17015 return Builder.CreateCall(F, { Src0, Src1, Src2 });
17016 }
17017 case AMDGPU::BI__builtin_amdgcn_class:
17018 case AMDGPU::BI__builtin_amdgcn_classf:
17019 case AMDGPU::BI__builtin_amdgcn_classh:
17020 return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_class);
17021 case AMDGPU::BI__builtin_amdgcn_fmed3f:
17022 case AMDGPU::BI__builtin_amdgcn_fmed3h:
17023 return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_fmed3);
17024 case AMDGPU::BI__builtin_amdgcn_ds_append:
17025 case AMDGPU::BI__builtin_amdgcn_ds_consume: {
17026 Intrinsic::ID Intrin = BuiltinID == AMDGPU::BI__builtin_amdgcn_ds_append ?
17027 Intrinsic::amdgcn_ds_append : Intrinsic::amdgcn_ds_consume;
17028 Value *Src0 = EmitScalarExpr(E->getArg(0));
17029 Function *F = CGM.getIntrinsic(Intrin, { Src0->getType() });
17030 return Builder.CreateCall(F, { Src0, Builder.getFalse() });
17031 }
17032 case AMDGPU::BI__builtin_amdgcn_ds_faddf:
17033 case AMDGPU::BI__builtin_amdgcn_ds_fminf:
17034 case AMDGPU::BI__builtin_amdgcn_ds_fmaxf: {
17035 Intrinsic::ID Intrin;
17036 switch (BuiltinID) {
17037 case AMDGPU::BI__builtin_amdgcn_ds_faddf:
17038 Intrin = Intrinsic::amdgcn_ds_fadd;
17039 break;
17040 case AMDGPU::BI__builtin_amdgcn_ds_fminf:
17041 Intrin = Intrinsic::amdgcn_ds_fmin;
17042 break;
17043 case AMDGPU::BI__builtin_amdgcn_ds_fmaxf:
17044 Intrin = Intrinsic::amdgcn_ds_fmax;
17045 break;
17046 }
17047 llvm::Value *Src0 = EmitScalarExpr(E->getArg(0));
17048 llvm::Value *Src1 = EmitScalarExpr(E->getArg(1));
17049 llvm::Value *Src2 = EmitScalarExpr(E->getArg(2));
17050 llvm::Value *Src3 = EmitScalarExpr(E->getArg(3));
17051 llvm::Value *Src4 = EmitScalarExpr(E->getArg(4));
17052 llvm::Function *F = CGM.getIntrinsic(Intrin, { Src1->getType() });
17053 llvm::FunctionType *FTy = F->getFunctionType();
17054 llvm::Type *PTy = FTy->getParamType(0);
17055 Src0 = Builder.CreatePointerBitCastOrAddrSpaceCast(Src0, PTy);
17056 return Builder.CreateCall(F, { Src0, Src1, Src2, Src3, Src4 });
17057 }
17058 case AMDGPU::BI__builtin_amdgcn_global_atomic_fadd_f64:
17059 case AMDGPU::BI__builtin_amdgcn_global_atomic_fadd_f32:
17060 case AMDGPU::BI__builtin_amdgcn_global_atomic_fadd_v2f16:
17061 case AMDGPU::BI__builtin_amdgcn_global_atomic_fmin_f64:
17062 case AMDGPU::BI__builtin_amdgcn_global_atomic_fmax_f64:
17063 case AMDGPU::BI__builtin_amdgcn_flat_atomic_fadd_f64:
17064 case AMDGPU::BI__builtin_amdgcn_flat_atomic_fmin_f64:
17065 case AMDGPU::BI__builtin_amdgcn_flat_atomic_fmax_f64:
17066 case AMDGPU::BI__builtin_amdgcn_flat_atomic_fadd_f32:
17067 case AMDGPU::BI__builtin_amdgcn_flat_atomic_fadd_v2f16: {
17068 Intrinsic::ID IID;
17069 llvm::Type *ArgTy = llvm::Type::getDoubleTy(getLLVMContext());
17070 switch (BuiltinID) {
17071 case AMDGPU::BI__builtin_amdgcn_global_atomic_fadd_f32:
17072 ArgTy = llvm::Type::getFloatTy(getLLVMContext());
17073 IID = Intrinsic::amdgcn_global_atomic_fadd;
17074 break;
17075 case AMDGPU::BI__builtin_amdgcn_global_atomic_fadd_v2f16:
17076 ArgTy = llvm::FixedVectorType::get(
17077 llvm::Type::getHalfTy(getLLVMContext()), 2);
17078 IID = Intrinsic::amdgcn_global_atomic_fadd;
17079 break;
17080 case AMDGPU::BI__builtin_amdgcn_global_atomic_fadd_f64:
17081 IID = Intrinsic::amdgcn_global_atomic_fadd;
17082 break;
17083 case AMDGPU::BI__builtin_amdgcn_global_atomic_fmin_f64:
17084 IID = Intrinsic::amdgcn_global_atomic_fmin;
17085 break;
17086 case AMDGPU::BI__builtin_amdgcn_global_atomic_fmax_f64:
17087 IID = Intrinsic::amdgcn_global_atomic_fmax;
17088 break;
17089 case AMDGPU::BI__builtin_amdgcn_flat_atomic_fadd_f64:
17090 IID = Intrinsic::amdgcn_flat_atomic_fadd;
17091 break;
17092 case AMDGPU::BI__builtin_amdgcn_flat_atomic_fmin_f64:
17093 IID = Intrinsic::amdgcn_flat_atomic_fmin;
17094 break;
17095 case AMDGPU::BI__builtin_amdgcn_flat_atomic_fmax_f64:
17096 IID = Intrinsic::amdgcn_flat_atomic_fmax;
17097 break;
17098 case AMDGPU::BI__builtin_amdgcn_flat_atomic_fadd_f32:
17099 ArgTy = llvm::Type::getFloatTy(getLLVMContext());
17100 IID = Intrinsic::amdgcn_flat_atomic_fadd;
17101 break;
17102 case AMDGPU::BI__builtin_amdgcn_flat_atomic_fadd_v2f16:
17103 ArgTy = llvm::FixedVectorType::get(
17104 llvm::Type::getHalfTy(getLLVMContext()), 2);
17105 IID = Intrinsic::amdgcn_flat_atomic_fadd;
17106 break;
17107 }
17108 llvm::Value *Addr = EmitScalarExpr(E->getArg(0));
17109 llvm::Value *Val = EmitScalarExpr(E->getArg(1));
17110 llvm::Function *F =
17111 CGM.getIntrinsic(IID, {ArgTy, Addr->getType(), Val->getType()});
17112 return Builder.CreateCall(F, {Addr, Val});
17113 }
17114 case AMDGPU::BI__builtin_amdgcn_global_atomic_fadd_v2bf16:
17115 case AMDGPU::BI__builtin_amdgcn_flat_atomic_fadd_v2bf16: {
17116 Intrinsic::ID IID;
17117 switch (BuiltinID) {
17118 case AMDGPU::BI__builtin_amdgcn_global_atomic_fadd_v2bf16:
17119 IID = Intrinsic::amdgcn_global_atomic_fadd_v2bf16;
17120 break;
17121 case AMDGPU::BI__builtin_amdgcn_flat_atomic_fadd_v2bf16:
17122 IID = Intrinsic::amdgcn_flat_atomic_fadd_v2bf16;
17123 break;
17124 }
17125 llvm::Value *Addr = EmitScalarExpr(E->getArg(0));
17126 llvm::Value *Val = EmitScalarExpr(E->getArg(1));
17127 llvm::Function *F = CGM.getIntrinsic(IID, {Addr->getType()});
17128 return Builder.CreateCall(F, {Addr, Val});
17129 }
17130 case AMDGPU::BI__builtin_amdgcn_ds_atomic_fadd_f64:
17131 case AMDGPU::BI__builtin_amdgcn_ds_atomic_fadd_f32: {
17132 Intrinsic::ID IID;
17133 llvm::Type *ArgTy;
17134 switch (BuiltinID) {
17135 case AMDGPU::BI__builtin_amdgcn_ds_atomic_fadd_f32:
17136 ArgTy = llvm::Type::getFloatTy(getLLVMContext());
17137 IID = Intrinsic::amdgcn_ds_fadd;
17138 break;
17139 case AMDGPU::BI__builtin_amdgcn_ds_atomic_fadd_f64:
17140 ArgTy = llvm::Type::getDoubleTy(getLLVMContext());
17141 IID = Intrinsic::amdgcn_ds_fadd;
17142 break;
17143 }
17144 llvm::Value *Addr = EmitScalarExpr(E->getArg(0));
17145 llvm::Value *Val = EmitScalarExpr(E->getArg(1));
17146 llvm::Constant *ZeroI32 = llvm::ConstantInt::getIntegerValue(
17147 llvm::Type::getInt32Ty(getLLVMContext()), APInt(32, 0, true));
17148 llvm::Constant *ZeroI1 = llvm::ConstantInt::getIntegerValue(
17149 llvm::Type::getInt1Ty(getLLVMContext()), APInt(1, 0));
17150 llvm::Function *F = CGM.getIntrinsic(IID, {ArgTy});
17151 return Builder.CreateCall(F, {Addr, Val, ZeroI32, ZeroI32, ZeroI1});
17152 }
17153 case AMDGPU::BI__builtin_amdgcn_read_exec: {
17154 CallInst *CI = cast<CallInst>(
17155 EmitSpecialRegisterBuiltin(*this, E, Int64Ty, Int64Ty, NormalRead, "exec"));
17156 CI->setConvergent();
17157 return CI;
17158 }
17159 case AMDGPU::BI__builtin_amdgcn_read_exec_lo:
17160 case AMDGPU::BI__builtin_amdgcn_read_exec_hi: {
17161 StringRef RegName = BuiltinID == AMDGPU::BI__builtin_amdgcn_read_exec_lo ?
17162 "exec_lo" : "exec_hi";
17163 CallInst *CI = cast<CallInst>(
17164 EmitSpecialRegisterBuiltin(*this, E, Int32Ty, Int32Ty, NormalRead, RegName));
17165 CI->setConvergent();
17166 return CI;
17167 }
17168 case AMDGPU::BI__builtin_amdgcn_image_bvh_intersect_ray:
17169 case AMDGPU::BI__builtin_amdgcn_image_bvh_intersect_ray_h:
17170 case AMDGPU::BI__builtin_amdgcn_image_bvh_intersect_ray_l:
17171 case AMDGPU::BI__builtin_amdgcn_image_bvh_intersect_ray_lh: {
17172 llvm::Value *NodePtr = EmitScalarExpr(E->getArg(0));
17173 llvm::Value *RayExtent = EmitScalarExpr(E->getArg(1));
17174 llvm::Value *RayOrigin = EmitScalarExpr(E->getArg(2));
17175 llvm::Value *RayDir = EmitScalarExpr(E->getArg(3));
17176 llvm::Value *RayInverseDir = EmitScalarExpr(E->getArg(4));
17177 llvm::Value *TextureDescr = EmitScalarExpr(E->getArg(5));
17178
17179 // The builtins take these arguments as vec4 where the last element is
17180 // ignored. The intrinsic takes them as vec3.
17181 RayOrigin = Builder.CreateShuffleVector(RayOrigin, RayOrigin,
17182 ArrayRef<int>{0, 1, 2});
17183 RayDir =
17184 Builder.CreateShuffleVector(RayDir, RayDir, ArrayRef<int>{0, 1, 2});
17185 RayInverseDir = Builder.CreateShuffleVector(RayInverseDir, RayInverseDir,
17186 ArrayRef<int>{0, 1, 2});
17187
17188 Function *F = CGM.getIntrinsic(Intrinsic::amdgcn_image_bvh_intersect_ray,
17189 {NodePtr->getType(), RayDir->getType()});
17190 return Builder.CreateCall(F, {NodePtr, RayExtent, RayOrigin, RayDir,
17191 RayInverseDir, TextureDescr});
17192 }
17193
17194 case AMDGPU::BI__builtin_amdgcn_ds_bvh_stack_rtn: {
17195 SmallVector<Value *, 4> Args;
17196 for (int i = 0, e = E->getNumArgs(); i != e; ++i)
17197 Args.push_back(EmitScalarExpr(E->getArg(i)));
17198
17199 Function *F = CGM.getIntrinsic(Intrinsic::amdgcn_ds_bvh_stack_rtn);
17200 Value *Call = Builder.CreateCall(F, Args);
17201 Value *Rtn = Builder.CreateExtractValue(Call, 0);
17202 Value *A = Builder.CreateExtractValue(Call, 1);
17203 llvm::Type *RetTy = ConvertType(E->getType());
17204 Value *I0 = Builder.CreateInsertElement(PoisonValue::get(RetTy), Rtn,
17205 (uint64_t)0);
17206 return Builder.CreateInsertElement(I0, A, 1);
17207 }
17208
17209 case AMDGPU::BI__builtin_amdgcn_wmma_bf16_16x16x16_bf16_w32:
17210 case AMDGPU::BI__builtin_amdgcn_wmma_bf16_16x16x16_bf16_w64:
17211 case AMDGPU::BI__builtin_amdgcn_wmma_f16_16x16x16_f16_w32:
17212 case AMDGPU::BI__builtin_amdgcn_wmma_f16_16x16x16_f16_w64:
17213 case AMDGPU::BI__builtin_amdgcn_wmma_f32_16x16x16_bf16_w32:
17214 case AMDGPU::BI__builtin_amdgcn_wmma_f32_16x16x16_bf16_w64:
17215 case AMDGPU::BI__builtin_amdgcn_wmma_f32_16x16x16_f16_w32:
17216 case AMDGPU::BI__builtin_amdgcn_wmma_f32_16x16x16_f16_w64:
17217 case AMDGPU::BI__builtin_amdgcn_wmma_i32_16x16x16_iu4_w32:
17218 case AMDGPU::BI__builtin_amdgcn_wmma_i32_16x16x16_iu4_w64:
17219 case AMDGPU::BI__builtin_amdgcn_wmma_i32_16x16x16_iu8_w32:
17220 case AMDGPU::BI__builtin_amdgcn_wmma_i32_16x16x16_iu8_w64: {
17221
17222 // These operations perform a matrix multiplication and accumulation of
17223 // the form:
17224 // D = A * B + C
17225 // The return type always matches the type of matrix C.
17226 unsigned ArgForMatchingRetType;
17227 unsigned BuiltinWMMAOp;
17228
17229 switch (BuiltinID) {
17230 case AMDGPU::BI__builtin_amdgcn_wmma_f32_16x16x16_f16_w32:
17231 case AMDGPU::BI__builtin_amdgcn_wmma_f32_16x16x16_f16_w64:
17232 ArgForMatchingRetType = 2;
17233 BuiltinWMMAOp = Intrinsic::amdgcn_wmma_f32_16x16x16_f16;
17234 break;
17235 case AMDGPU::BI__builtin_amdgcn_wmma_f32_16x16x16_bf16_w32:
17236 case AMDGPU::BI__builtin_amdgcn_wmma_f32_16x16x16_bf16_w64:
17237 ArgForMatchingRetType = 2;
17238 BuiltinWMMAOp = Intrinsic::amdgcn_wmma_f32_16x16x16_bf16;
17239 break;
17240 case AMDGPU::BI__builtin_amdgcn_wmma_f16_16x16x16_f16_w32:
17241 case AMDGPU::BI__builtin_amdgcn_wmma_f16_16x16x16_f16_w64:
17242 ArgForMatchingRetType = 2;
17243 BuiltinWMMAOp = Intrinsic::amdgcn_wmma_f16_16x16x16_f16;
17244 break;
17245 case AMDGPU::BI__builtin_amdgcn_wmma_bf16_16x16x16_bf16_w32:
17246 case AMDGPU::BI__builtin_amdgcn_wmma_bf16_16x16x16_bf16_w64:
17247 ArgForMatchingRetType = 2;
17248 BuiltinWMMAOp = Intrinsic::amdgcn_wmma_bf16_16x16x16_bf16;
17249 break;
17250 case AMDGPU::BI__builtin_amdgcn_wmma_i32_16x16x16_iu8_w32:
17251 case AMDGPU::BI__builtin_amdgcn_wmma_i32_16x16x16_iu8_w64:
17252 ArgForMatchingRetType = 4;
17253 BuiltinWMMAOp = Intrinsic::amdgcn_wmma_i32_16x16x16_iu8;
17254 break;
17255 case AMDGPU::BI__builtin_amdgcn_wmma_i32_16x16x16_iu4_w32:
17256 case AMDGPU::BI__builtin_amdgcn_wmma_i32_16x16x16_iu4_w64:
17257 ArgForMatchingRetType = 4;
17258 BuiltinWMMAOp = Intrinsic::amdgcn_wmma_i32_16x16x16_iu4;
17259 break;
17260 }
17261
17262 SmallVector<Value *, 6> Args;
17263 for (int i = 0, e = E->getNumArgs(); i != e; ++i)
17264 Args.push_back(EmitScalarExpr(E->getArg(i)));
17265
17266 Function *F = CGM.getIntrinsic(BuiltinWMMAOp,
17267 {Args[ArgForMatchingRetType]->getType()});
17268
17269 return Builder.CreateCall(F, Args);
17270 }
17271
17272 // amdgcn workitem
17273 case AMDGPU::BI__builtin_amdgcn_workitem_id_x:
17274 return emitRangedBuiltin(*this, Intrinsic::amdgcn_workitem_id_x, 0, 1024);
17275 case AMDGPU::BI__builtin_amdgcn_workitem_id_y:
17276 return emitRangedBuiltin(*this, Intrinsic::amdgcn_workitem_id_y, 0, 1024);
17277 case AMDGPU::BI__builtin_amdgcn_workitem_id_z:
17278 return emitRangedBuiltin(*this, Intrinsic::amdgcn_workitem_id_z, 0, 1024);
17279
17280 // amdgcn workgroup size
17281 case AMDGPU::BI__builtin_amdgcn_workgroup_size_x:
17282 return EmitAMDGPUWorkGroupSize(*this, 0);
17283 case AMDGPU::BI__builtin_amdgcn_workgroup_size_y:
17284 return EmitAMDGPUWorkGroupSize(*this, 1);
17285 case AMDGPU::BI__builtin_amdgcn_workgroup_size_z:
17286 return EmitAMDGPUWorkGroupSize(*this, 2);
17287
17288 // amdgcn grid size
17289 case AMDGPU::BI__builtin_amdgcn_grid_size_x:
17290 return EmitAMDGPUGridSize(*this, 0);
17291 case AMDGPU::BI__builtin_amdgcn_grid_size_y:
17292 return EmitAMDGPUGridSize(*this, 1);
17293 case AMDGPU::BI__builtin_amdgcn_grid_size_z:
17294 return EmitAMDGPUGridSize(*this, 2);
17295
17296 // r600 intrinsics
17297 case AMDGPU::BI__builtin_r600_recipsqrt_ieee:
17298 case AMDGPU::BI__builtin_r600_recipsqrt_ieeef:
17299 return emitUnaryBuiltin(*this, E, Intrinsic::r600_recipsqrt_ieee);
17300 case AMDGPU::BI__builtin_r600_read_tidig_x:
17301 return emitRangedBuiltin(*this, Intrinsic::r600_read_tidig_x, 0, 1024);
17302 case AMDGPU::BI__builtin_r600_read_tidig_y:
17303 return emitRangedBuiltin(*this, Intrinsic::r600_read_tidig_y, 0, 1024);
17304 case AMDGPU::BI__builtin_r600_read_tidig_z:
17305 return emitRangedBuiltin(*this, Intrinsic::r600_read_tidig_z, 0, 1024);
17306 case AMDGPU::BI__builtin_amdgcn_alignbit: {
17307 llvm::Value *Src0 = EmitScalarExpr(E->getArg(0));
17308 llvm::Value *Src1 = EmitScalarExpr(E->getArg(1));
17309 llvm::Value *Src2 = EmitScalarExpr(E->getArg(2));
17310 Function *F = CGM.getIntrinsic(Intrinsic::fshr, Src0->getType());
17311 return Builder.CreateCall(F, { Src0, Src1, Src2 });
17312 }
17313 case AMDGPU::BI__builtin_amdgcn_fence: {
17314 ProcessOrderScopeAMDGCN(EmitScalarExpr(E->getArg(0)),
17315 EmitScalarExpr(E->getArg(1)), AO, SSID);
17316 return Builder.CreateFence(AO, SSID);
17317 }
17318 case AMDGPU::BI__builtin_amdgcn_atomic_inc32:
17319 case AMDGPU::BI__builtin_amdgcn_atomic_inc64:
17320 case AMDGPU::BI__builtin_amdgcn_atomic_dec32:
17321 case AMDGPU::BI__builtin_amdgcn_atomic_dec64: {
17322 unsigned BuiltinAtomicOp;
17323 llvm::Type *ResultType = ConvertType(E->getType());
17324
17325 switch (BuiltinID) {
17326 case AMDGPU::BI__builtin_amdgcn_atomic_inc32:
17327 case AMDGPU::BI__builtin_amdgcn_atomic_inc64:
17328 BuiltinAtomicOp = Intrinsic::amdgcn_atomic_inc;
17329 break;
17330 case AMDGPU::BI__builtin_amdgcn_atomic_dec32:
17331 case AMDGPU::BI__builtin_amdgcn_atomic_dec64:
17332 BuiltinAtomicOp = Intrinsic::amdgcn_atomic_dec;
17333 break;
17334 }
17335
17336 Value *Ptr = EmitScalarExpr(E->getArg(0));
17337 Value *Val = EmitScalarExpr(E->getArg(1));
17338
17339 llvm::Function *F =
17340 CGM.getIntrinsic(BuiltinAtomicOp, {ResultType, Ptr->getType()});
17341
17342 ProcessOrderScopeAMDGCN(EmitScalarExpr(E->getArg(2)),
17343 EmitScalarExpr(E->getArg(3)), AO, SSID);
17344
17345 // llvm.amdgcn.atomic.inc and llvm.amdgcn.atomic.dec expects ordering and
17346 // scope as unsigned values
17347 Value *MemOrder = Builder.getInt32(static_cast<int>(AO));
17348 Value *MemScope = Builder.getInt32(static_cast<int>(SSID));
17349
17350 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType();
17351 bool Volatile =
17352 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified();
17353 Value *IsVolatile = Builder.getInt1(static_cast<bool>(Volatile));
17354
17355 return Builder.CreateCall(F, {Ptr, Val, MemOrder, MemScope, IsVolatile});
17356 }
17357 case AMDGPU::BI__builtin_amdgcn_s_sendmsg_rtn:
17358 case AMDGPU::BI__builtin_amdgcn_s_sendmsg_rtnl: {
17359 llvm::Value *Arg = EmitScalarExpr(E->getArg(0));
17360 llvm::Type *ResultType = ConvertType(E->getType());
17361 // s_sendmsg_rtn is mangled using return type only.
17362 Function *F =
17363 CGM.getIntrinsic(Intrinsic::amdgcn_s_sendmsg_rtn, {ResultType});
17364 return Builder.CreateCall(F, {Arg});
17365 }
17366 default:
17367 return nullptr;
17368 }
17369 }
17370
17371 /// Handle a SystemZ function in which the final argument is a pointer
17372 /// to an int that receives the post-instruction CC value. At the LLVM level
17373 /// this is represented as a function that returns a {result, cc} pair.
17374 static Value *EmitSystemZIntrinsicWithCC(CodeGenFunction &CGF,
17375 unsigned IntrinsicID,
17376 const CallExpr *E) {
17377 unsigned NumArgs = E->getNumArgs() - 1;
17378 SmallVector<Value *, 8> Args(NumArgs);
17379 for (unsigned I = 0; I < NumArgs; ++I)
17380 Args[I] = CGF.EmitScalarExpr(E->getArg(I));
17381 Address CCPtr = CGF.EmitPointerWithAlignment(E->getArg(NumArgs));
17382 Function *F = CGF.CGM.getIntrinsic(IntrinsicID);
17383 Value *Call = CGF.Builder.CreateCall(F, Args);
17384 Value *CC = CGF.Builder.CreateExtractValue(Call, 1);
17385 CGF.Builder.CreateStore(CC, CCPtr);
17386 return CGF.Builder.CreateExtractValue(Call, 0);
17387 }
17388
17389 Value *CodeGenFunction::EmitSystemZBuiltinExpr(unsigned BuiltinID,
17390 const CallExpr *E) {
17391 switch (BuiltinID) {
17392 case SystemZ::BI__builtin_tbegin: {
17393 Value *TDB = EmitScalarExpr(E->getArg(0));
17394 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c);
17395 Function *F = CGM.getIntrinsic(Intrinsic::s390_tbegin);
17396 return Builder.CreateCall(F, {TDB, Control});
17397 }
17398 case SystemZ::BI__builtin_tbegin_nofloat: {
17399 Value *TDB = EmitScalarExpr(E->getArg(0));
17400 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c);
17401 Function *F = CGM.getIntrinsic(Intrinsic::s390_tbegin_nofloat);
17402 return Builder.CreateCall(F, {TDB, Control});
17403 }
17404 case SystemZ::BI__builtin_tbeginc: {
17405 Value *TDB = llvm::ConstantPointerNull::get(Int8PtrTy);
17406 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff08);
17407 Function *F = CGM.getIntrinsic(Intrinsic::s390_tbeginc);
17408 return Builder.CreateCall(F, {TDB, Control});
17409 }
17410 case SystemZ::BI__builtin_tabort: {
17411 Value *Data = EmitScalarExpr(E->getArg(0));
17412 Function *F = CGM.getIntrinsic(Intrinsic::s390_tabort);
17413 return Builder.CreateCall(F, Builder.CreateSExt(Data, Int64Ty, "tabort"));
17414 }
17415 case SystemZ::BI__builtin_non_tx_store: {
17416 Value *Address = EmitScalarExpr(E->getArg(0));
17417 Value *Data = EmitScalarExpr(E->getArg(1));
17418 Function *F = CGM.getIntrinsic(Intrinsic::s390_ntstg);
17419 return Builder.CreateCall(F, {Data, Address});
17420 }
17421
17422 // Vector builtins. Note that most vector builtins are mapped automatically
17423 // to target-specific LLVM intrinsics. The ones handled specially here can
17424 // be represented via standard LLVM IR, which is preferable to enable common
17425 // LLVM optimizations.
17426
17427 case SystemZ::BI__builtin_s390_vpopctb:
17428 case SystemZ::BI__builtin_s390_vpopcth:
17429 case SystemZ::BI__builtin_s390_vpopctf:
17430 case SystemZ::BI__builtin_s390_vpopctg: {
17431 llvm::Type *ResultType = ConvertType(E->getType());
17432 Value *X = EmitScalarExpr(E->getArg(0));
17433 Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ResultType);
17434 return Builder.CreateCall(F, X);
17435 }
17436
17437 case SystemZ::BI__builtin_s390_vclzb:
17438 case SystemZ::BI__builtin_s390_vclzh:
17439 case SystemZ::BI__builtin_s390_vclzf:
17440 case SystemZ::BI__builtin_s390_vclzg: {
17441 llvm::Type *ResultType = ConvertType(E->getType());
17442 Value *X = EmitScalarExpr(E->getArg(0));
17443 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false);
17444 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ResultType);
17445 return Builder.CreateCall(F, {X, Undef});
17446 }
17447
17448 case SystemZ::BI__builtin_s390_vctzb:
17449 case SystemZ::BI__builtin_s390_vctzh:
17450 case SystemZ::BI__builtin_s390_vctzf:
17451 case SystemZ::BI__builtin_s390_vctzg: {
17452 llvm::Type *ResultType = ConvertType(E->getType());
17453 Value *X = EmitScalarExpr(E->getArg(0));
17454 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false);
17455 Function *F = CGM.getIntrinsic(Intrinsic::cttz, ResultType);
17456 return Builder.CreateCall(F, {X, Undef});
17457 }
17458
17459 case SystemZ::BI__builtin_s390_vfsqsb:
17460 case SystemZ::BI__builtin_s390_vfsqdb: {
17461 llvm::Type *ResultType = ConvertType(E->getType());
17462 Value *X = EmitScalarExpr(E->getArg(0));
17463 if (Builder.getIsFPConstrained()) {
17464 Function *F = CGM.getIntrinsic(Intrinsic::experimental_constrained_sqrt, ResultType);
17465 return Builder.CreateConstrainedFPCall(F, { X });
17466 } else {
17467 Function *F = CGM.getIntrinsic(Intrinsic::sqrt, ResultType);
17468 return Builder.CreateCall(F, X);
17469 }
17470 }
17471 case SystemZ::BI__builtin_s390_vfmasb:
17472 case SystemZ::BI__builtin_s390_vfmadb: {
17473 llvm::Type *ResultType = ConvertType(E->getType());
17474 Value *X = EmitScalarExpr(E->getArg(0));
17475 Value *Y = EmitScalarExpr(E->getArg(1));
17476 Value *Z = EmitScalarExpr(E->getArg(2));
17477 if (Builder.getIsFPConstrained()) {
17478 Function *F = CGM.getIntrinsic(Intrinsic::experimental_constrained_fma, ResultType);
17479 return Builder.CreateConstrainedFPCall(F, {X, Y, Z});
17480 } else {
17481 Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType);
17482 return Builder.CreateCall(F, {X, Y, Z});
17483 }
17484 }
17485 case SystemZ::BI__builtin_s390_vfmssb:
17486 case SystemZ::BI__builtin_s390_vfmsdb: {
17487 llvm::Type *ResultType = ConvertType(E->getType());
17488 Value *X = EmitScalarExpr(E->getArg(0));
17489 Value *Y = EmitScalarExpr(E->getArg(1));
17490 Value *Z = EmitScalarExpr(E->getArg(2));
17491 if (Builder.getIsFPConstrained()) {
17492 Function *F = CGM.getIntrinsic(Intrinsic::experimental_constrained_fma, ResultType);
17493 return Builder.CreateConstrainedFPCall(F, {X, Y, Builder.CreateFNeg(Z, "neg")});
17494 } else {
17495 Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType);
17496 return Builder.CreateCall(F, {X, Y, Builder.CreateFNeg(Z, "neg")});
17497 }
17498 }
17499 case SystemZ::BI__builtin_s390_vfnmasb:
17500 case SystemZ::BI__builtin_s390_vfnmadb: {
17501 llvm::Type *ResultType = ConvertType(E->getType());
17502 Value *X = EmitScalarExpr(E->getArg(0));
17503 Value *Y = EmitScalarExpr(E->getArg(1));
17504 Value *Z = EmitScalarExpr(E->getArg(2));
17505 if (Builder.getIsFPConstrained()) {
17506 Function *F = CGM.getIntrinsic(Intrinsic::experimental_constrained_fma, ResultType);
17507 return Builder.CreateFNeg(Builder.CreateConstrainedFPCall(F, {X, Y, Z}), "neg");
17508 } else {
17509 Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType);
17510 return Builder.CreateFNeg(Builder.CreateCall(F, {X, Y, Z}), "neg");
17511 }
17512 }
17513 case SystemZ::BI__builtin_s390_vfnmssb:
17514 case SystemZ::BI__builtin_s390_vfnmsdb: {
17515 llvm::Type *ResultType = ConvertType(E->getType());
17516 Value *X = EmitScalarExpr(E->getArg(0));
17517 Value *Y = EmitScalarExpr(E->getArg(1));
17518 Value *Z = EmitScalarExpr(E->getArg(2));
17519 if (Builder.getIsFPConstrained()) {
17520 Function *F = CGM.getIntrinsic(Intrinsic::experimental_constrained_fma, ResultType);
17521 Value *NegZ = Builder.CreateFNeg(Z, "sub");
17522 return Builder.CreateFNeg(Builder.CreateConstrainedFPCall(F, {X, Y, NegZ}));
17523 } else {
17524 Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType);
17525 Value *NegZ = Builder.CreateFNeg(Z, "neg");
17526 return Builder.CreateFNeg(Builder.CreateCall(F, {X, Y, NegZ}));
17527 }
17528 }
17529 case SystemZ::BI__builtin_s390_vflpsb:
17530 case SystemZ::BI__builtin_s390_vflpdb: {
17531 llvm::Type *ResultType = ConvertType(E->getType());
17532 Value *X = EmitScalarExpr(E->getArg(0));
17533 Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType);
17534 return Builder.CreateCall(F, X);
17535 }
17536 case SystemZ::BI__builtin_s390_vflnsb:
17537 case SystemZ::BI__builtin_s390_vflndb: {
17538 llvm::Type *ResultType = ConvertType(E->getType());
17539 Value *X = EmitScalarExpr(E->getArg(0));
17540 Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType);
17541 return Builder.CreateFNeg(Builder.CreateCall(F, X), "neg");
17542 }
17543 case SystemZ::BI__builtin_s390_vfisb:
17544 case SystemZ::BI__builtin_s390_vfidb: {
17545 llvm::Type *ResultType = ConvertType(E->getType());
17546 Value *X = EmitScalarExpr(E->getArg(0));
17547 // Constant-fold the M4 and M5 mask arguments.
17548 llvm::APSInt M4 = *E->getArg(1)->getIntegerConstantExpr(getContext());
17549 llvm::APSInt M5 = *E->getArg(2)->getIntegerConstantExpr(getContext());
17550 // Check whether this instance can be represented via a LLVM standard
17551 // intrinsic. We only support some combinations of M4 and M5.
17552 Intrinsic::ID ID = Intrinsic::not_intrinsic;
17553 Intrinsic::ID CI;
17554 switch (M4.getZExtValue()) {
17555 default: break;
17556 case 0: // IEEE-inexact exception allowed
17557 switch (M5.getZExtValue()) {
17558 default: break;
17559 case 0: ID = Intrinsic::rint;
17560 CI = Intrinsic::experimental_constrained_rint; break;
17561 }
17562 break;
17563 case 4: // IEEE-inexact exception suppressed
17564 switch (M5.getZExtValue()) {
17565 default: break;
17566 case 0: ID = Intrinsic::nearbyint;
17567 CI = Intrinsic::experimental_constrained_nearbyint; break;
17568 case 1: ID = Intrinsic::round;
17569 CI = Intrinsic::experimental_constrained_round; break;
17570 case 5: ID = Intrinsic::trunc;
17571 CI = Intrinsic::experimental_constrained_trunc; break;
17572 case 6: ID = Intrinsic::ceil;
17573 CI = Intrinsic::experimental_constrained_ceil; break;
17574 case 7: ID = Intrinsic::floor;
17575 CI = Intrinsic::experimental_constrained_floor; break;
17576 }
17577 break;
17578 }
17579 if (ID != Intrinsic::not_intrinsic) {
17580 if (Builder.getIsFPConstrained()) {
17581 Function *F = CGM.getIntrinsic(CI, ResultType);
17582 return Builder.CreateConstrainedFPCall(F, X);
17583 } else {
17584 Function *F = CGM.getIntrinsic(ID, ResultType);
17585 return Builder.CreateCall(F, X);
17586 }
17587 }
17588 switch (BuiltinID) { // FIXME: constrained version?
17589 case SystemZ::BI__builtin_s390_vfisb: ID = Intrinsic::s390_vfisb; break;
17590 case SystemZ::BI__builtin_s390_vfidb: ID = Intrinsic::s390_vfidb; break;
17591 default: llvm_unreachable("Unknown BuiltinID");
17592 }
17593 Function *F = CGM.getIntrinsic(ID);
17594 Value *M4Value = llvm::ConstantInt::get(getLLVMContext(), M4);
17595 Value *M5Value = llvm::ConstantInt::get(getLLVMContext(), M5);
17596 return Builder.CreateCall(F, {X, M4Value, M5Value});
17597 }
17598 case SystemZ::BI__builtin_s390_vfmaxsb:
17599 case SystemZ::BI__builtin_s390_vfmaxdb: {
17600 llvm::Type *ResultType = ConvertType(E->getType());
17601 Value *X = EmitScalarExpr(E->getArg(0));
17602 Value *Y = EmitScalarExpr(E->getArg(1));
17603 // Constant-fold the M4 mask argument.
17604 llvm::APSInt M4 = *E->getArg(2)->getIntegerConstantExpr(getContext());
17605 // Check whether this instance can be represented via a LLVM standard
17606 // intrinsic. We only support some values of M4.
17607 Intrinsic::ID ID = Intrinsic::not_intrinsic;
17608 Intrinsic::ID CI;
17609 switch (M4.getZExtValue()) {
17610 default: break;
17611 case 4: ID = Intrinsic::maxnum;
17612 CI = Intrinsic::experimental_constrained_maxnum; break;
17613 }
17614 if (ID != Intrinsic::not_intrinsic) {
17615 if (Builder.getIsFPConstrained()) {
17616 Function *F = CGM.getIntrinsic(CI, ResultType);
17617 return Builder.CreateConstrainedFPCall(F, {X, Y});
17618 } else {
17619 Function *F = CGM.getIntrinsic(ID, ResultType);
17620 return Builder.CreateCall(F, {X, Y});
17621 }
17622 }
17623 switch (BuiltinID) {
17624 case SystemZ::BI__builtin_s390_vfmaxsb: ID = Intrinsic::s390_vfmaxsb; break;
17625 case SystemZ::BI__builtin_s390_vfmaxdb: ID = Intrinsic::s390_vfmaxdb; break;
17626 default: llvm_unreachable("Unknown BuiltinID");
17627 }
17628 Function *F = CGM.getIntrinsic(ID);
17629 Value *M4Value = llvm::ConstantInt::get(getLLVMContext(), M4);
17630 return Builder.CreateCall(F, {X, Y, M4Value});
17631 }
17632 case SystemZ::BI__builtin_s390_vfminsb:
17633 case SystemZ::BI__builtin_s390_vfmindb: {
17634 llvm::Type *ResultType = ConvertType(E->getType());
17635 Value *X = EmitScalarExpr(E->getArg(0));
17636 Value *Y = EmitScalarExpr(E->getArg(1));
17637 // Constant-fold the M4 mask argument.
17638 llvm::APSInt M4 = *E->getArg(2)->getIntegerConstantExpr(getContext());
17639 // Check whether this instance can be represented via a LLVM standard
17640 // intrinsic. We only support some values of M4.
17641 Intrinsic::ID ID = Intrinsic::not_intrinsic;
17642 Intrinsic::ID CI;
17643 switch (M4.getZExtValue()) {
17644 default: break;
17645 case 4: ID = Intrinsic::minnum;
17646 CI = Intrinsic::experimental_constrained_minnum; break;
17647 }
17648 if (ID != Intrinsic::not_intrinsic) {
17649 if (Builder.getIsFPConstrained()) {
17650 Function *F = CGM.getIntrinsic(CI, ResultType);
17651 return Builder.CreateConstrainedFPCall(F, {X, Y});
17652 } else {
17653 Function *F = CGM.getIntrinsic(ID, ResultType);
17654 return Builder.CreateCall(F, {X, Y});
17655 }
17656 }
17657 switch (BuiltinID) {
17658 case SystemZ::BI__builtin_s390_vfminsb: ID = Intrinsic::s390_vfminsb; break;
17659 case SystemZ::BI__builtin_s390_vfmindb: ID = Intrinsic::s390_vfmindb; break;
17660 default: llvm_unreachable("Unknown BuiltinID");
17661 }
17662 Function *F = CGM.getIntrinsic(ID);
17663 Value *M4Value = llvm::ConstantInt::get(getLLVMContext(), M4);
17664 return Builder.CreateCall(F, {X, Y, M4Value});
17665 }
17666
17667 case SystemZ::BI__builtin_s390_vlbrh:
17668 case SystemZ::BI__builtin_s390_vlbrf:
17669 case SystemZ::BI__builtin_s390_vlbrg: {
17670 llvm::Type *ResultType = ConvertType(E->getType());
17671 Value *X = EmitScalarExpr(E->getArg(0));
17672 Function *F = CGM.getIntrinsic(Intrinsic::bswap, ResultType);
17673 return Builder.CreateCall(F, X);
17674 }
17675
17676 // Vector intrinsics that output the post-instruction CC value.
17677
17678 #define INTRINSIC_WITH_CC(NAME) \
17679 case SystemZ::BI__builtin_##NAME: \
17680 return EmitSystemZIntrinsicWithCC(*this, Intrinsic::NAME, E)
17681
17682 INTRINSIC_WITH_CC(s390_vpkshs);
17683 INTRINSIC_WITH_CC(s390_vpksfs);
17684 INTRINSIC_WITH_CC(s390_vpksgs);
17685
17686 INTRINSIC_WITH_CC(s390_vpklshs);
17687 INTRINSIC_WITH_CC(s390_vpklsfs);
17688 INTRINSIC_WITH_CC(s390_vpklsgs);
17689
17690 INTRINSIC_WITH_CC(s390_vceqbs);
17691 INTRINSIC_WITH_CC(s390_vceqhs);
17692 INTRINSIC_WITH_CC(s390_vceqfs);
17693 INTRINSIC_WITH_CC(s390_vceqgs);
17694
17695 INTRINSIC_WITH_CC(s390_vchbs);
17696 INTRINSIC_WITH_CC(s390_vchhs);
17697 INTRINSIC_WITH_CC(s390_vchfs);
17698 INTRINSIC_WITH_CC(s390_vchgs);
17699
17700 INTRINSIC_WITH_CC(s390_vchlbs);
17701 INTRINSIC_WITH_CC(s390_vchlhs);
17702 INTRINSIC_WITH_CC(s390_vchlfs);
17703 INTRINSIC_WITH_CC(s390_vchlgs);
17704
17705 INTRINSIC_WITH_CC(s390_vfaebs);
17706 INTRINSIC_WITH_CC(s390_vfaehs);
17707 INTRINSIC_WITH_CC(s390_vfaefs);
17708
17709 INTRINSIC_WITH_CC(s390_vfaezbs);
17710 INTRINSIC_WITH_CC(s390_vfaezhs);
17711 INTRINSIC_WITH_CC(s390_vfaezfs);
17712
17713 INTRINSIC_WITH_CC(s390_vfeebs);
17714 INTRINSIC_WITH_CC(s390_vfeehs);
17715 INTRINSIC_WITH_CC(s390_vfeefs);
17716
17717 INTRINSIC_WITH_CC(s390_vfeezbs);
17718 INTRINSIC_WITH_CC(s390_vfeezhs);
17719 INTRINSIC_WITH_CC(s390_vfeezfs);
17720
17721 INTRINSIC_WITH_CC(s390_vfenebs);
17722 INTRINSIC_WITH_CC(s390_vfenehs);
17723 INTRINSIC_WITH_CC(s390_vfenefs);
17724
17725 INTRINSIC_WITH_CC(s390_vfenezbs);
17726 INTRINSIC_WITH_CC(s390_vfenezhs);
17727 INTRINSIC_WITH_CC(s390_vfenezfs);
17728
17729 INTRINSIC_WITH_CC(s390_vistrbs);
17730 INTRINSIC_WITH_CC(s390_vistrhs);
17731 INTRINSIC_WITH_CC(s390_vistrfs);
17732
17733 INTRINSIC_WITH_CC(s390_vstrcbs);
17734 INTRINSIC_WITH_CC(s390_vstrchs);
17735 INTRINSIC_WITH_CC(s390_vstrcfs);
17736
17737 INTRINSIC_WITH_CC(s390_vstrczbs);
17738 INTRINSIC_WITH_CC(s390_vstrczhs);
17739 INTRINSIC_WITH_CC(s390_vstrczfs);
17740
17741 INTRINSIC_WITH_CC(s390_vfcesbs);
17742 INTRINSIC_WITH_CC(s390_vfcedbs);
17743 INTRINSIC_WITH_CC(s390_vfchsbs);
17744 INTRINSIC_WITH_CC(s390_vfchdbs);
17745 INTRINSIC_WITH_CC(s390_vfchesbs);
17746 INTRINSIC_WITH_CC(s390_vfchedbs);
17747
17748 INTRINSIC_WITH_CC(s390_vftcisb);
17749 INTRINSIC_WITH_CC(s390_vftcidb);
17750
17751 INTRINSIC_WITH_CC(s390_vstrsb);
17752 INTRINSIC_WITH_CC(s390_vstrsh);
17753 INTRINSIC_WITH_CC(s390_vstrsf);
17754
17755 INTRINSIC_WITH_CC(s390_vstrszb);
17756 INTRINSIC_WITH_CC(s390_vstrszh);
17757 INTRINSIC_WITH_CC(s390_vstrszf);
17758
17759 #undef INTRINSIC_WITH_CC
17760
17761 default:
17762 return nullptr;
17763 }
17764 }
17765
17766 namespace {
17767 // Helper classes for mapping MMA builtins to particular LLVM intrinsic variant.
17768 struct NVPTXMmaLdstInfo {
17769 unsigned NumResults; // Number of elements to load/store
17770 // Intrinsic IDs for row/col variants. 0 if particular layout is unsupported.
17771 unsigned IID_col;
17772 unsigned IID_row;
17773 };
17774
17775 #define MMA_INTR(geom_op_type, layout) \
17776 Intrinsic::nvvm_wmma_##geom_op_type##_##layout##_stride
17777 #define MMA_LDST(n, geom_op_type) \
17778 { n, MMA_INTR(geom_op_type, col), MMA_INTR(geom_op_type, row) }
17779
17780 static NVPTXMmaLdstInfo getNVPTXMmaLdstInfo(unsigned BuiltinID) {
17781 switch (BuiltinID) {
17782 // FP MMA loads
17783 case NVPTX::BI__hmma_m16n16k16_ld_a:
17784 return MMA_LDST(8, m16n16k16_load_a_f16);
17785 case NVPTX::BI__hmma_m16n16k16_ld_b:
17786 return MMA_LDST(8, m16n16k16_load_b_f16);
17787 case NVPTX::BI__hmma_m16n16k16_ld_c_f16:
17788 return MMA_LDST(4, m16n16k16_load_c_f16);
17789 case NVPTX::BI__hmma_m16n16k16_ld_c_f32:
17790 return MMA_LDST(8, m16n16k16_load_c_f32);
17791 case NVPTX::BI__hmma_m32n8k16_ld_a:
17792 return MMA_LDST(8, m32n8k16_load_a_f16);
17793 case NVPTX::BI__hmma_m32n8k16_ld_b:
17794 return MMA_LDST(8, m32n8k16_load_b_f16);
17795 case NVPTX::BI__hmma_m32n8k16_ld_c_f16:
17796 return MMA_LDST(4, m32n8k16_load_c_f16);
17797 case NVPTX::BI__hmma_m32n8k16_ld_c_f32:
17798 return MMA_LDST(8, m32n8k16_load_c_f32);
17799 case NVPTX::BI__hmma_m8n32k16_ld_a:
17800 return MMA_LDST(8, m8n32k16_load_a_f16);
17801 case NVPTX::BI__hmma_m8n32k16_ld_b:
17802 return MMA_LDST(8, m8n32k16_load_b_f16);
17803 case NVPTX::BI__hmma_m8n32k16_ld_c_f16:
17804 return MMA_LDST(4, m8n32k16_load_c_f16);
17805 case NVPTX::BI__hmma_m8n32k16_ld_c_f32:
17806 return MMA_LDST(8, m8n32k16_load_c_f32);
17807
17808 // Integer MMA loads
17809 case NVPTX::BI__imma_m16n16k16_ld_a_s8:
17810 return MMA_LDST(2, m16n16k16_load_a_s8);
17811 case NVPTX::BI__imma_m16n16k16_ld_a_u8:
17812 return MMA_LDST(2, m16n16k16_load_a_u8);
17813 case NVPTX::BI__imma_m16n16k16_ld_b_s8:
17814 return MMA_LDST(2, m16n16k16_load_b_s8);
17815 case NVPTX::BI__imma_m16n16k16_ld_b_u8:
17816 return MMA_LDST(2, m16n16k16_load_b_u8);
17817 case NVPTX::BI__imma_m16n16k16_ld_c:
17818 return MMA_LDST(8, m16n16k16_load_c_s32);
17819 case NVPTX::BI__imma_m32n8k16_ld_a_s8:
17820 return MMA_LDST(4, m32n8k16_load_a_s8);
17821 case NVPTX::BI__imma_m32n8k16_ld_a_u8:
17822 return MMA_LDST(4, m32n8k16_load_a_u8);
17823 case NVPTX::BI__imma_m32n8k16_ld_b_s8:
17824 return MMA_LDST(1, m32n8k16_load_b_s8);
17825 case NVPTX::BI__imma_m32n8k16_ld_b_u8:
17826 return MMA_LDST(1, m32n8k16_load_b_u8);
17827 case NVPTX::BI__imma_m32n8k16_ld_c:
17828 return MMA_LDST(8, m32n8k16_load_c_s32);
17829 case NVPTX::BI__imma_m8n32k16_ld_a_s8:
17830 return MMA_LDST(1, m8n32k16_load_a_s8);
17831 case NVPTX::BI__imma_m8n32k16_ld_a_u8:
17832 return MMA_LDST(1, m8n32k16_load_a_u8);
17833 case NVPTX::BI__imma_m8n32k16_ld_b_s8:
17834 return MMA_LDST(4, m8n32k16_load_b_s8);
17835 case NVPTX::BI__imma_m8n32k16_ld_b_u8:
17836 return MMA_LDST(4, m8n32k16_load_b_u8);
17837 case NVPTX::BI__imma_m8n32k16_ld_c:
17838 return MMA_LDST(8, m8n32k16_load_c_s32);
17839
17840 // Sub-integer MMA loads.
17841 // Only row/col layout is supported by A/B fragments.
17842 case NVPTX::BI__imma_m8n8k32_ld_a_s4:
17843 return {1, 0, MMA_INTR(m8n8k32_load_a_s4, row)};
17844 case NVPTX::BI__imma_m8n8k32_ld_a_u4:
17845 return {1, 0, MMA_INTR(m8n8k32_load_a_u4, row)};
17846 case NVPTX::BI__imma_m8n8k32_ld_b_s4:
17847 return {1, MMA_INTR(m8n8k32_load_b_s4, col), 0};
17848 case NVPTX::BI__imma_m8n8k32_ld_b_u4:
17849 return {1, MMA_INTR(m8n8k32_load_b_u4, col), 0};
17850 case NVPTX::BI__imma_m8n8k32_ld_c:
17851 return MMA_LDST(2, m8n8k32_load_c_s32);
17852 case NVPTX::BI__bmma_m8n8k128_ld_a_b1:
17853 return {1, 0, MMA_INTR(m8n8k128_load_a_b1, row)};
17854 case NVPTX::BI__bmma_m8n8k128_ld_b_b1:
17855 return {1, MMA_INTR(m8n8k128_load_b_b1, col), 0};
17856 case NVPTX::BI__bmma_m8n8k128_ld_c:
17857 return MMA_LDST(2, m8n8k128_load_c_s32);
17858
17859 // Double MMA loads
17860 case NVPTX::BI__dmma_m8n8k4_ld_a:
17861 return MMA_LDST(1, m8n8k4_load_a_f64);
17862 case NVPTX::BI__dmma_m8n8k4_ld_b:
17863 return MMA_LDST(1, m8n8k4_load_b_f64);
17864 case NVPTX::BI__dmma_m8n8k4_ld_c:
17865 return MMA_LDST(2, m8n8k4_load_c_f64);
17866
17867 // Alternate float MMA loads
17868 case NVPTX::BI__mma_bf16_m16n16k16_ld_a:
17869 return MMA_LDST(4, m16n16k16_load_a_bf16);
17870 case NVPTX::BI__mma_bf16_m16n16k16_ld_b:
17871 return MMA_LDST(4, m16n16k16_load_b_bf16);
17872 case NVPTX::BI__mma_bf16_m8n32k16_ld_a:
17873 return MMA_LDST(2, m8n32k16_load_a_bf16);
17874 case NVPTX::BI__mma_bf16_m8n32k16_ld_b:
17875 return MMA_LDST(8, m8n32k16_load_b_bf16);
17876 case NVPTX::BI__mma_bf16_m32n8k16_ld_a:
17877 return MMA_LDST(8, m32n8k16_load_a_bf16);
17878 case NVPTX::BI__mma_bf16_m32n8k16_ld_b:
17879 return MMA_LDST(2, m32n8k16_load_b_bf16);
17880 case NVPTX::BI__mma_tf32_m16n16k8_ld_a:
17881 return MMA_LDST(4, m16n16k8_load_a_tf32);
17882 case NVPTX::BI__mma_tf32_m16n16k8_ld_b:
17883 return MMA_LDST(4, m16n16k8_load_b_tf32);
17884 case NVPTX::BI__mma_tf32_m16n16k8_ld_c:
17885 return MMA_LDST(8, m16n16k8_load_c_f32);
17886
17887 // NOTE: We need to follow inconsitent naming scheme used by NVCC. Unlike
17888 // PTX and LLVM IR where stores always use fragment D, NVCC builtins always
17889 // use fragment C for both loads and stores.
17890 // FP MMA stores.
17891 case NVPTX::BI__hmma_m16n16k16_st_c_f16:
17892 return MMA_LDST(4, m16n16k16_store_d_f16);
17893 case NVPTX::BI__hmma_m16n16k16_st_c_f32:
17894 return MMA_LDST(8, m16n16k16_store_d_f32);
17895 case NVPTX::BI__hmma_m32n8k16_st_c_f16:
17896 return MMA_LDST(4, m32n8k16_store_d_f16);
17897 case NVPTX::BI__hmma_m32n8k16_st_c_f32:
17898 return MMA_LDST(8, m32n8k16_store_d_f32);
17899 case NVPTX::BI__hmma_m8n32k16_st_c_f16:
17900 return MMA_LDST(4, m8n32k16_store_d_f16);
17901 case NVPTX::BI__hmma_m8n32k16_st_c_f32:
17902 return MMA_LDST(8, m8n32k16_store_d_f32);
17903
17904 // Integer and sub-integer MMA stores.
17905 // Another naming quirk. Unlike other MMA builtins that use PTX types in the
17906 // name, integer loads/stores use LLVM's i32.
17907 case NVPTX::BI__imma_m16n16k16_st_c_i32:
17908 return MMA_LDST(8, m16n16k16_store_d_s32);
17909 case NVPTX::BI__imma_m32n8k16_st_c_i32:
17910 return MMA_LDST(8, m32n8k16_store_d_s32);
17911 case NVPTX::BI__imma_m8n32k16_st_c_i32:
17912 return MMA_LDST(8, m8n32k16_store_d_s32);
17913 case NVPTX::BI__imma_m8n8k32_st_c_i32:
17914 return MMA_LDST(2, m8n8k32_store_d_s32);
17915 case NVPTX::BI__bmma_m8n8k128_st_c_i32:
17916 return MMA_LDST(2, m8n8k128_store_d_s32);
17917
17918 // Double MMA store
17919 case NVPTX::BI__dmma_m8n8k4_st_c_f64:
17920 return MMA_LDST(2, m8n8k4_store_d_f64);
17921
17922 // Alternate float MMA store
17923 case NVPTX::BI__mma_m16n16k8_st_c_f32:
17924 return MMA_LDST(8, m16n16k8_store_d_f32);
17925
17926 default:
17927 llvm_unreachable("Unknown MMA builtin");
17928 }
17929 }
17930 #undef MMA_LDST
17931 #undef MMA_INTR
17932
17933
17934 struct NVPTXMmaInfo {
17935 unsigned NumEltsA;
17936 unsigned NumEltsB;
17937 unsigned NumEltsC;
17938 unsigned NumEltsD;
17939
17940 // Variants are ordered by layout-A/layout-B/satf, where 'row' has priority
17941 // over 'col' for layout. The index of non-satf variants is expected to match
17942 // the undocumented layout constants used by CUDA's mma.hpp.
17943 std::array<unsigned, 8> Variants;
17944
17945 unsigned getMMAIntrinsic(int Layout, bool Satf) {
17946 unsigned Index = Layout + 4 * Satf;
17947 if (Index >= Variants.size())
17948 return 0;
17949 return Variants[Index];
17950 }
17951 };
17952
17953 // Returns an intrinsic that matches Layout and Satf for valid combinations of
17954 // Layout and Satf, 0 otherwise.
17955 static NVPTXMmaInfo getNVPTXMmaInfo(unsigned BuiltinID) {
17956 // clang-format off
17957 #define MMA_VARIANTS(geom, type) \
17958 Intrinsic::nvvm_wmma_##geom##_mma_row_row_##type, \
17959 Intrinsic::nvvm_wmma_##geom##_mma_row_col_##type, \
17960 Intrinsic::nvvm_wmma_##geom##_mma_col_row_##type, \
17961 Intrinsic::nvvm_wmma_##geom##_mma_col_col_##type
17962 #define MMA_SATF_VARIANTS(geom, type) \
17963 MMA_VARIANTS(geom, type), \
17964 Intrinsic::nvvm_wmma_##geom##_mma_row_row_##type##_satfinite, \
17965 Intrinsic::nvvm_wmma_##geom##_mma_row_col_##type##_satfinite, \
17966 Intrinsic::nvvm_wmma_##geom##_mma_col_row_##type##_satfinite, \
17967 Intrinsic::nvvm_wmma_##geom##_mma_col_col_##type##_satfinite
17968 // Sub-integer MMA only supports row.col layout.
17969 #define MMA_VARIANTS_I4(geom, type) \
17970 0, \
17971 Intrinsic::nvvm_wmma_##geom##_mma_row_col_##type, \
17972 0, \
17973 0, \
17974 0, \
17975 Intrinsic::nvvm_wmma_##geom##_mma_row_col_##type##_satfinite, \
17976 0, \
17977 0
17978 // b1 MMA does not support .satfinite.
17979 #define MMA_VARIANTS_B1_XOR(geom, type) \
17980 0, \
17981 Intrinsic::nvvm_wmma_##geom##_mma_xor_popc_row_col_##type, \
17982 0, \
17983 0, \
17984 0, \
17985 0, \
17986 0, \
17987 0
17988 #define MMA_VARIANTS_B1_AND(geom, type) \
17989 0, \
17990 Intrinsic::nvvm_wmma_##geom##_mma_and_popc_row_col_##type, \
17991 0, \
17992 0, \
17993 0, \
17994 0, \
17995 0, \
17996 0
17997 // clang-format on
17998 switch (BuiltinID) {
17999 // FP MMA
18000 // Note that 'type' argument of MMA_SATF_VARIANTS uses D_C notation, while
18001 // NumEltsN of return value are ordered as A,B,C,D.
18002 case NVPTX::BI__hmma_m16n16k16_mma_f16f16:
18003 return {8, 8, 4, 4, {{MMA_SATF_VARIANTS(m16n16k16, f16_f16)}}};
18004 case NVPTX::BI__hmma_m16n16k16_mma_f32f16:
18005 return {8, 8, 4, 8, {{MMA_SATF_VARIANTS(m16n16k16, f32_f16)}}};
18006 case NVPTX::BI__hmma_m16n16k16_mma_f16f32:
18007 return {8, 8, 8, 4, {{MMA_SATF_VARIANTS(m16n16k16, f16_f32)}}};
18008 case NVPTX::BI__hmma_m16n16k16_mma_f32f32:
18009 return {8, 8, 8, 8, {{MMA_SATF_VARIANTS(m16n16k16, f32_f32)}}};
18010 case NVPTX::BI__hmma_m32n8k16_mma_f16f16:
18011 return {8, 8, 4, 4, {{MMA_SATF_VARIANTS(m32n8k16, f16_f16)}}};
18012 case NVPTX::BI__hmma_m32n8k16_mma_f32f16:
18013 return {8, 8, 4, 8, {{MMA_SATF_VARIANTS(m32n8k16, f32_f16)}}};
18014 case NVPTX::BI__hmma_m32n8k16_mma_f16f32:
18015 return {8, 8, 8, 4, {{MMA_SATF_VARIANTS(m32n8k16, f16_f32)}}};
18016 case NVPTX::BI__hmma_m32n8k16_mma_f32f32:
18017 return {8, 8, 8, 8, {{MMA_SATF_VARIANTS(m32n8k16, f32_f32)}}};
18018 case NVPTX::BI__hmma_m8n32k16_mma_f16f16:
18019 return {8, 8, 4, 4, {{MMA_SATF_VARIANTS(m8n32k16, f16_f16)}}};
18020 case NVPTX::BI__hmma_m8n32k16_mma_f32f16:
18021 return {8, 8, 4, 8, {{MMA_SATF_VARIANTS(m8n32k16, f32_f16)}}};
18022 case NVPTX::BI__hmma_m8n32k16_mma_f16f32:
18023 return {8, 8, 8, 4, {{MMA_SATF_VARIANTS(m8n32k16, f16_f32)}}};
18024 case NVPTX::BI__hmma_m8n32k16_mma_f32f32:
18025 return {8, 8, 8, 8, {{MMA_SATF_VARIANTS(m8n32k16, f32_f32)}}};
18026
18027 // Integer MMA
18028 case NVPTX::BI__imma_m16n16k16_mma_s8:
18029 return {2, 2, 8, 8, {{MMA_SATF_VARIANTS(m16n16k16, s8)}}};
18030 case NVPTX::BI__imma_m16n16k16_mma_u8:
18031 return {2, 2, 8, 8, {{MMA_SATF_VARIANTS(m16n16k16, u8)}}};
18032 case NVPTX::BI__imma_m32n8k16_mma_s8:
18033 return {4, 1, 8, 8, {{MMA_SATF_VARIANTS(m32n8k16, s8)}}};
18034 case NVPTX::BI__imma_m32n8k16_mma_u8:
18035 return {4, 1, 8, 8, {{MMA_SATF_VARIANTS(m32n8k16, u8)}}};
18036 case NVPTX::BI__imma_m8n32k16_mma_s8:
18037 return {1, 4, 8, 8, {{MMA_SATF_VARIANTS(m8n32k16, s8)}}};
18038 case NVPTX::BI__imma_m8n32k16_mma_u8:
18039 return {1, 4, 8, 8, {{MMA_SATF_VARIANTS(m8n32k16, u8)}}};
18040
18041 // Sub-integer MMA
18042 case NVPTX::BI__imma_m8n8k32_mma_s4:
18043 return {1, 1, 2, 2, {{MMA_VARIANTS_I4(m8n8k32, s4)}}};
18044 case NVPTX::BI__imma_m8n8k32_mma_u4:
18045 return {1, 1, 2, 2, {{MMA_VARIANTS_I4(m8n8k32, u4)}}};
18046 case NVPTX::BI__bmma_m8n8k128_mma_xor_popc_b1:
18047 return {1, 1, 2, 2, {{MMA_VARIANTS_B1_XOR(m8n8k128, b1)}}};
18048 case NVPTX::BI__bmma_m8n8k128_mma_and_popc_b1:
18049 return {1, 1, 2, 2, {{MMA_VARIANTS_B1_AND(m8n8k128, b1)}}};
18050
18051 // Double MMA
18052 case NVPTX::BI__dmma_m8n8k4_mma_f64:
18053 return {1, 1, 2, 2, {{MMA_VARIANTS(m8n8k4, f64)}}};
18054
18055 // Alternate FP MMA
18056 case NVPTX::BI__mma_bf16_m16n16k16_mma_f32:
18057 return {4, 4, 8, 8, {{MMA_VARIANTS(m16n16k16, bf16)}}};
18058 case NVPTX::BI__mma_bf16_m8n32k16_mma_f32:
18059 return {2, 8, 8, 8, {{MMA_VARIANTS(m8n32k16, bf16)}}};
18060 case NVPTX::BI__mma_bf16_m32n8k16_mma_f32:
18061 return {8, 2, 8, 8, {{MMA_VARIANTS(m32n8k16, bf16)}}};
18062 case NVPTX::BI__mma_tf32_m16n16k8_mma_f32:
18063 return {4, 4, 8, 8, {{MMA_VARIANTS(m16n16k8, tf32)}}};
18064 default:
18065 llvm_unreachable("Unexpected builtin ID.");
18066 }
18067 #undef MMA_VARIANTS
18068 #undef MMA_SATF_VARIANTS
18069 #undef MMA_VARIANTS_I4
18070 #undef MMA_VARIANTS_B1_AND
18071 #undef MMA_VARIANTS_B1_XOR
18072 }
18073
18074 } // namespace
18075
18076 Value *
18077 CodeGenFunction::EmitNVPTXBuiltinExpr(unsigned BuiltinID, const CallExpr *E) {
18078 auto MakeLdg = [&](unsigned IntrinsicID) {
18079 Value *Ptr = EmitScalarExpr(E->getArg(0));
18080 QualType ArgType = E->getArg(0)->getType();
18081 clang::CharUnits Align = CGM.getNaturalPointeeTypeAlignment(ArgType);
18082 llvm::Type *ElemTy = ConvertTypeForMem(ArgType->getPointeeType());
18083 return Builder.CreateCall(
18084 CGM.getIntrinsic(IntrinsicID, {ElemTy, Ptr->getType()}),
18085 {Ptr, ConstantInt::get(Builder.getInt32Ty(), Align.getQuantity())});
18086 };
18087 auto MakeScopedAtomic = [&](unsigned IntrinsicID) {
18088 Value *Ptr = EmitScalarExpr(E->getArg(0));
18089 llvm::Type *ElemTy =
18090 ConvertTypeForMem(E->getArg(0)->getType()->getPointeeType());
18091 return Builder.CreateCall(
18092 CGM.getIntrinsic(IntrinsicID, {ElemTy, Ptr->getType()}),
18093 {Ptr, EmitScalarExpr(E->getArg(1))});
18094 };
18095 switch (BuiltinID) {
18096 case NVPTX::BI__nvvm_atom_add_gen_i:
18097 case NVPTX::BI__nvvm_atom_add_gen_l:
18098 case NVPTX::BI__nvvm_atom_add_gen_ll:
18099 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Add, E);
18100
18101 case NVPTX::BI__nvvm_atom_sub_gen_i:
18102 case NVPTX::BI__nvvm_atom_sub_gen_l:
18103 case NVPTX::BI__nvvm_atom_sub_gen_ll:
18104 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Sub, E);
18105
18106 case NVPTX::BI__nvvm_atom_and_gen_i:
18107 case NVPTX::BI__nvvm_atom_and_gen_l:
18108 case NVPTX::BI__nvvm_atom_and_gen_ll:
18109 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::And, E);
18110
18111 case NVPTX::BI__nvvm_atom_or_gen_i:
18112 case NVPTX::BI__nvvm_atom_or_gen_l:
18113 case NVPTX::BI__nvvm_atom_or_gen_ll:
18114 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Or, E);
18115
18116 case NVPTX::BI__nvvm_atom_xor_gen_i:
18117 case NVPTX::BI__nvvm_atom_xor_gen_l:
18118 case NVPTX::BI__nvvm_atom_xor_gen_ll:
18119 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Xor, E);
18120
18121 case NVPTX::BI__nvvm_atom_xchg_gen_i:
18122 case NVPTX::BI__nvvm_atom_xchg_gen_l:
18123 case NVPTX::BI__nvvm_atom_xchg_gen_ll:
18124 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Xchg, E);
18125
18126 case NVPTX::BI__nvvm_atom_max_gen_i:
18127 case NVPTX::BI__nvvm_atom_max_gen_l:
18128 case NVPTX::BI__nvvm_atom_max_gen_ll:
18129 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Max, E);
18130
18131 case NVPTX::BI__nvvm_atom_max_gen_ui:
18132 case NVPTX::BI__nvvm_atom_max_gen_ul:
18133 case NVPTX::BI__nvvm_atom_max_gen_ull:
18134 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::UMax, E);
18135
18136 case NVPTX::BI__nvvm_atom_min_gen_i:
18137 case NVPTX::BI__nvvm_atom_min_gen_l:
18138 case NVPTX::BI__nvvm_atom_min_gen_ll:
18139 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Min, E);
18140
18141 case NVPTX::BI__nvvm_atom_min_gen_ui:
18142 case NVPTX::BI__nvvm_atom_min_gen_ul:
18143 case NVPTX::BI__nvvm_atom_min_gen_ull:
18144 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::UMin, E);
18145
18146 case NVPTX::BI__nvvm_atom_cas_gen_i:
18147 case NVPTX::BI__nvvm_atom_cas_gen_l:
18148 case NVPTX::BI__nvvm_atom_cas_gen_ll:
18149 // __nvvm_atom_cas_gen_* should return the old value rather than the
18150 // success flag.
18151 return MakeAtomicCmpXchgValue(*this, E, /*ReturnBool=*/false);
18152
18153 case NVPTX::BI__nvvm_atom_add_gen_f:
18154 case NVPTX::BI__nvvm_atom_add_gen_d: {
18155 Value *Ptr = EmitScalarExpr(E->getArg(0));
18156 Value *Val = EmitScalarExpr(E->getArg(1));
18157 return Builder.CreateAtomicRMW(llvm::AtomicRMWInst::FAdd, Ptr, Val,
18158 AtomicOrdering::SequentiallyConsistent);
18159 }
18160
18161 case NVPTX::BI__nvvm_atom_inc_gen_ui: {
18162 Value *Ptr = EmitScalarExpr(E->getArg(0));
18163 Value *Val = EmitScalarExpr(E->getArg(1));
18164 Function *FnALI32 =
18165 CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_inc_32, Ptr->getType());
18166 return Builder.CreateCall(FnALI32, {Ptr, Val});
18167 }
18168
18169 case NVPTX::BI__nvvm_atom_dec_gen_ui: {
18170 Value *Ptr = EmitScalarExpr(E->getArg(0));
18171 Value *Val = EmitScalarExpr(E->getArg(1));
18172 Function *FnALD32 =
18173 CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_dec_32, Ptr->getType());
18174 return Builder.CreateCall(FnALD32, {Ptr, Val});
18175 }
18176
18177 case NVPTX::BI__nvvm_ldg_c:
18178 case NVPTX::BI__nvvm_ldg_c2:
18179 case NVPTX::BI__nvvm_ldg_c4:
18180 case NVPTX::BI__nvvm_ldg_s:
18181 case NVPTX::BI__nvvm_ldg_s2:
18182 case NVPTX::BI__nvvm_ldg_s4:
18183 case NVPTX::BI__nvvm_ldg_i:
18184 case NVPTX::BI__nvvm_ldg_i2:
18185 case NVPTX::BI__nvvm_ldg_i4:
18186 case NVPTX::BI__nvvm_ldg_l:
18187 case NVPTX::BI__nvvm_ldg_ll:
18188 case NVPTX::BI__nvvm_ldg_ll2:
18189 case NVPTX::BI__nvvm_ldg_uc:
18190 case NVPTX::BI__nvvm_ldg_uc2:
18191 case NVPTX::BI__nvvm_ldg_uc4:
18192 case NVPTX::BI__nvvm_ldg_us:
18193 case NVPTX::BI__nvvm_ldg_us2:
18194 case NVPTX::BI__nvvm_ldg_us4:
18195 case NVPTX::BI__nvvm_ldg_ui:
18196 case NVPTX::BI__nvvm_ldg_ui2:
18197 case NVPTX::BI__nvvm_ldg_ui4:
18198 case NVPTX::BI__nvvm_ldg_ul:
18199 case NVPTX::BI__nvvm_ldg_ull:
18200 case NVPTX::BI__nvvm_ldg_ull2:
18201 // PTX Interoperability section 2.2: "For a vector with an even number of
18202 // elements, its alignment is set to number of elements times the alignment
18203 // of its member: n*alignof(t)."
18204 return MakeLdg(Intrinsic::nvvm_ldg_global_i);
18205 case NVPTX::BI__nvvm_ldg_f:
18206 case NVPTX::BI__nvvm_ldg_f2:
18207 case NVPTX::BI__nvvm_ldg_f4:
18208 case NVPTX::BI__nvvm_ldg_d:
18209 case NVPTX::BI__nvvm_ldg_d2:
18210 return MakeLdg(Intrinsic::nvvm_ldg_global_f);
18211
18212 case NVPTX::BI__nvvm_atom_cta_add_gen_i:
18213 case NVPTX::BI__nvvm_atom_cta_add_gen_l:
18214 case NVPTX::BI__nvvm_atom_cta_add_gen_ll:
18215 return MakeScopedAtomic(Intrinsic::nvvm_atomic_add_gen_i_cta);
18216 case NVPTX::BI__nvvm_atom_sys_add_gen_i:
18217 case NVPTX::BI__nvvm_atom_sys_add_gen_l:
18218 case NVPTX::BI__nvvm_atom_sys_add_gen_ll:
18219 return MakeScopedAtomic(Intrinsic::nvvm_atomic_add_gen_i_sys);
18220 case NVPTX::BI__nvvm_atom_cta_add_gen_f:
18221 case NVPTX::BI__nvvm_atom_cta_add_gen_d:
18222 return MakeScopedAtomic(Intrinsic::nvvm_atomic_add_gen_f_cta);
18223 case NVPTX::BI__nvvm_atom_sys_add_gen_f:
18224 case NVPTX::BI__nvvm_atom_sys_add_gen_d:
18225 return MakeScopedAtomic(Intrinsic::nvvm_atomic_add_gen_f_sys);
18226 case NVPTX::BI__nvvm_atom_cta_xchg_gen_i:
18227 case NVPTX::BI__nvvm_atom_cta_xchg_gen_l:
18228 case NVPTX::BI__nvvm_atom_cta_xchg_gen_ll:
18229 return MakeScopedAtomic(Intrinsic::nvvm_atomic_exch_gen_i_cta);
18230 case NVPTX::BI__nvvm_atom_sys_xchg_gen_i:
18231 case NVPTX::BI__nvvm_atom_sys_xchg_gen_l:
18232 case NVPTX::BI__nvvm_atom_sys_xchg_gen_ll:
18233 return MakeScopedAtomic(Intrinsic::nvvm_atomic_exch_gen_i_sys);
18234 case NVPTX::BI__nvvm_atom_cta_max_gen_i:
18235 case NVPTX::BI__nvvm_atom_cta_max_gen_ui:
18236 case NVPTX::BI__nvvm_atom_cta_max_gen_l:
18237 case NVPTX::BI__nvvm_atom_cta_max_gen_ul:
18238 case NVPTX::BI__nvvm_atom_cta_max_gen_ll:
18239 case NVPTX::BI__nvvm_atom_cta_max_gen_ull:
18240 return MakeScopedAtomic(Intrinsic::nvvm_atomic_max_gen_i_cta);
18241 case NVPTX::BI__nvvm_atom_sys_max_gen_i:
18242 case NVPTX::BI__nvvm_atom_sys_max_gen_ui:
18243 case NVPTX::BI__nvvm_atom_sys_max_gen_l:
18244 case NVPTX::BI__nvvm_atom_sys_max_gen_ul:
18245 case NVPTX::BI__nvvm_atom_sys_max_gen_ll:
18246 case NVPTX::BI__nvvm_atom_sys_max_gen_ull:
18247 return MakeScopedAtomic(Intrinsic::nvvm_atomic_max_gen_i_sys);
18248 case NVPTX::BI__nvvm_atom_cta_min_gen_i:
18249 case NVPTX::BI__nvvm_atom_cta_min_gen_ui:
18250 case NVPTX::BI__nvvm_atom_cta_min_gen_l:
18251 case NVPTX::BI__nvvm_atom_cta_min_gen_ul:
18252 case NVPTX::BI__nvvm_atom_cta_min_gen_ll:
18253 case NVPTX::BI__nvvm_atom_cta_min_gen_ull:
18254 return MakeScopedAtomic(Intrinsic::nvvm_atomic_min_gen_i_cta);
18255 case NVPTX::BI__nvvm_atom_sys_min_gen_i:
18256 case NVPTX::BI__nvvm_atom_sys_min_gen_ui:
18257 case NVPTX::BI__nvvm_atom_sys_min_gen_l:
18258 case NVPTX::BI__nvvm_atom_sys_min_gen_ul:
18259 case NVPTX::BI__nvvm_atom_sys_min_gen_ll:
18260 case NVPTX::BI__nvvm_atom_sys_min_gen_ull:
18261 return MakeScopedAtomic(Intrinsic::nvvm_atomic_min_gen_i_sys);
18262 case NVPTX::BI__nvvm_atom_cta_inc_gen_ui:
18263 return MakeScopedAtomic(Intrinsic::nvvm_atomic_inc_gen_i_cta);
18264 case NVPTX::BI__nvvm_atom_cta_dec_gen_ui:
18265 return MakeScopedAtomic(Intrinsic::nvvm_atomic_dec_gen_i_cta);
18266 case NVPTX::BI__nvvm_atom_sys_inc_gen_ui:
18267 return MakeScopedAtomic(Intrinsic::nvvm_atomic_inc_gen_i_sys);
18268 case NVPTX::BI__nvvm_atom_sys_dec_gen_ui:
18269 return MakeScopedAtomic(Intrinsic::nvvm_atomic_dec_gen_i_sys);
18270 case NVPTX::BI__nvvm_atom_cta_and_gen_i:
18271 case NVPTX::BI__nvvm_atom_cta_and_gen_l:
18272 case NVPTX::BI__nvvm_atom_cta_and_gen_ll:
18273 return MakeScopedAtomic(Intrinsic::nvvm_atomic_and_gen_i_cta);
18274 case NVPTX::BI__nvvm_atom_sys_and_gen_i:
18275 case NVPTX::BI__nvvm_atom_sys_and_gen_l:
18276 case NVPTX::BI__nvvm_atom_sys_and_gen_ll:
18277 return MakeScopedAtomic(Intrinsic::nvvm_atomic_and_gen_i_sys);
18278 case NVPTX::BI__nvvm_atom_cta_or_gen_i:
18279 case NVPTX::BI__nvvm_atom_cta_or_gen_l:
18280 case NVPTX::BI__nvvm_atom_cta_or_gen_ll:
18281 return MakeScopedAtomic(Intrinsic::nvvm_atomic_or_gen_i_cta);
18282 case NVPTX::BI__nvvm_atom_sys_or_gen_i:
18283 case NVPTX::BI__nvvm_atom_sys_or_gen_l:
18284 case NVPTX::BI__nvvm_atom_sys_or_gen_ll:
18285 return MakeScopedAtomic(Intrinsic::nvvm_atomic_or_gen_i_sys);
18286 case NVPTX::BI__nvvm_atom_cta_xor_gen_i:
18287 case NVPTX::BI__nvvm_atom_cta_xor_gen_l:
18288 case NVPTX::BI__nvvm_atom_cta_xor_gen_ll:
18289 return MakeScopedAtomic(Intrinsic::nvvm_atomic_xor_gen_i_cta);
18290 case NVPTX::BI__nvvm_atom_sys_xor_gen_i:
18291 case NVPTX::BI__nvvm_atom_sys_xor_gen_l:
18292 case NVPTX::BI__nvvm_atom_sys_xor_gen_ll:
18293 return MakeScopedAtomic(Intrinsic::nvvm_atomic_xor_gen_i_sys);
18294 case NVPTX::BI__nvvm_atom_cta_cas_gen_i:
18295 case NVPTX::BI__nvvm_atom_cta_cas_gen_l:
18296 case NVPTX::BI__nvvm_atom_cta_cas_gen_ll: {
18297 Value *Ptr = EmitScalarExpr(E->getArg(0));
18298 llvm::Type *ElemTy =
18299 ConvertTypeForMem(E->getArg(0)->getType()->getPointeeType());
18300 return Builder.CreateCall(
18301 CGM.getIntrinsic(
18302 Intrinsic::nvvm_atomic_cas_gen_i_cta, {ElemTy, Ptr->getType()}),
18303 {Ptr, EmitScalarExpr(E->getArg(1)), EmitScalarExpr(E->getArg(2))});
18304 }
18305 case NVPTX::BI__nvvm_atom_sys_cas_gen_i:
18306 case NVPTX::BI__nvvm_atom_sys_cas_gen_l:
18307 case NVPTX::BI__nvvm_atom_sys_cas_gen_ll: {
18308 Value *Ptr = EmitScalarExpr(E->getArg(0));
18309 llvm::Type *ElemTy =
18310 ConvertTypeForMem(E->getArg(0)->getType()->getPointeeType());
18311 return Builder.CreateCall(
18312 CGM.getIntrinsic(
18313 Intrinsic::nvvm_atomic_cas_gen_i_sys, {ElemTy, Ptr->getType()}),
18314 {Ptr, EmitScalarExpr(E->getArg(1)), EmitScalarExpr(E->getArg(2))});
18315 }
18316 case NVPTX::BI__nvvm_match_all_sync_i32p:
18317 case NVPTX::BI__nvvm_match_all_sync_i64p: {
18318 Value *Mask = EmitScalarExpr(E->getArg(0));
18319 Value *Val = EmitScalarExpr(E->getArg(1));
18320 Address PredOutPtr = EmitPointerWithAlignment(E->getArg(2));
18321 Value *ResultPair = Builder.CreateCall(
18322 CGM.getIntrinsic(BuiltinID == NVPTX::BI__nvvm_match_all_sync_i32p
18323 ? Intrinsic::nvvm_match_all_sync_i32p
18324 : Intrinsic::nvvm_match_all_sync_i64p),
18325 {Mask, Val});
18326 Value *Pred = Builder.CreateZExt(Builder.CreateExtractValue(ResultPair, 1),
18327 PredOutPtr.getElementType());
18328 Builder.CreateStore(Pred, PredOutPtr);
18329 return Builder.CreateExtractValue(ResultPair, 0);
18330 }
18331
18332 // FP MMA loads
18333 case NVPTX::BI__hmma_m16n16k16_ld_a:
18334 case NVPTX::BI__hmma_m16n16k16_ld_b:
18335 case NVPTX::BI__hmma_m16n16k16_ld_c_f16:
18336 case NVPTX::BI__hmma_m16n16k16_ld_c_f32:
18337 case NVPTX::BI__hmma_m32n8k16_ld_a:
18338 case NVPTX::BI__hmma_m32n8k16_ld_b:
18339 case NVPTX::BI__hmma_m32n8k16_ld_c_f16:
18340 case NVPTX::BI__hmma_m32n8k16_ld_c_f32:
18341 case NVPTX::BI__hmma_m8n32k16_ld_a:
18342 case NVPTX::BI__hmma_m8n32k16_ld_b:
18343 case NVPTX::BI__hmma_m8n32k16_ld_c_f16:
18344 case NVPTX::BI__hmma_m8n32k16_ld_c_f32:
18345 // Integer MMA loads.
18346 case NVPTX::BI__imma_m16n16k16_ld_a_s8:
18347 case NVPTX::BI__imma_m16n16k16_ld_a_u8:
18348 case NVPTX::BI__imma_m16n16k16_ld_b_s8:
18349 case NVPTX::BI__imma_m16n16k16_ld_b_u8:
18350 case NVPTX::BI__imma_m16n16k16_ld_c:
18351 case NVPTX::BI__imma_m32n8k16_ld_a_s8:
18352 case NVPTX::BI__imma_m32n8k16_ld_a_u8:
18353 case NVPTX::BI__imma_m32n8k16_ld_b_s8:
18354 case NVPTX::BI__imma_m32n8k16_ld_b_u8:
18355 case NVPTX::BI__imma_m32n8k16_ld_c:
18356 case NVPTX::BI__imma_m8n32k16_ld_a_s8:
18357 case NVPTX::BI__imma_m8n32k16_ld_a_u8:
18358 case NVPTX::BI__imma_m8n32k16_ld_b_s8:
18359 case NVPTX::BI__imma_m8n32k16_ld_b_u8:
18360 case NVPTX::BI__imma_m8n32k16_ld_c:
18361 // Sub-integer MMA loads.
18362 case NVPTX::BI__imma_m8n8k32_ld_a_s4:
18363 case NVPTX::BI__imma_m8n8k32_ld_a_u4:
18364 case NVPTX::BI__imma_m8n8k32_ld_b_s4:
18365 case NVPTX::BI__imma_m8n8k32_ld_b_u4:
18366 case NVPTX::BI__imma_m8n8k32_ld_c:
18367 case NVPTX::BI__bmma_m8n8k128_ld_a_b1:
18368 case NVPTX::BI__bmma_m8n8k128_ld_b_b1:
18369 case NVPTX::BI__bmma_m8n8k128_ld_c:
18370 // Double MMA loads.
18371 case NVPTX::BI__dmma_m8n8k4_ld_a:
18372 case NVPTX::BI__dmma_m8n8k4_ld_b:
18373 case NVPTX::BI__dmma_m8n8k4_ld_c:
18374 // Alternate float MMA loads.
18375 case NVPTX::BI__mma_bf16_m16n16k16_ld_a:
18376 case NVPTX::BI__mma_bf16_m16n16k16_ld_b:
18377 case NVPTX::BI__mma_bf16_m8n32k16_ld_a:
18378 case NVPTX::BI__mma_bf16_m8n32k16_ld_b:
18379 case NVPTX::BI__mma_bf16_m32n8k16_ld_a:
18380 case NVPTX::BI__mma_bf16_m32n8k16_ld_b:
18381 case NVPTX::BI__mma_tf32_m16n16k8_ld_a:
18382 case NVPTX::BI__mma_tf32_m16n16k8_ld_b:
18383 case NVPTX::BI__mma_tf32_m16n16k8_ld_c: {
18384 Address Dst = EmitPointerWithAlignment(E->getArg(0));
18385 Value *Src = EmitScalarExpr(E->getArg(1));
18386 Value *Ldm = EmitScalarExpr(E->getArg(2));
18387 std::optional<llvm::APSInt> isColMajorArg =
18388 E->getArg(3)->getIntegerConstantExpr(getContext());
18389 if (!isColMajorArg)
18390 return nullptr;
18391 bool isColMajor = isColMajorArg->getSExtValue();
18392 NVPTXMmaLdstInfo II = getNVPTXMmaLdstInfo(BuiltinID);
18393 unsigned IID = isColMajor ? II.IID_col : II.IID_row;
18394 if (IID == 0)
18395 return nullptr;
18396
18397 Value *Result =
18398 Builder.CreateCall(CGM.getIntrinsic(IID, Src->getType()), {Src, Ldm});
18399
18400 // Save returned values.
18401 assert(II.NumResults);
18402 if (II.NumResults == 1) {
18403 Builder.CreateAlignedStore(Result, Dst.getPointer(),
18404 CharUnits::fromQuantity(4));
18405 } else {
18406 for (unsigned i = 0; i < II.NumResults; ++i) {
18407 Builder.CreateAlignedStore(
18408 Builder.CreateBitCast(Builder.CreateExtractValue(Result, i),
18409 Dst.getElementType()),
18410 Builder.CreateGEP(Dst.getElementType(), Dst.getPointer(),
18411 llvm::ConstantInt::get(IntTy, i)),
18412 CharUnits::fromQuantity(4));
18413 }
18414 }
18415 return Result;
18416 }
18417
18418 case NVPTX::BI__hmma_m16n16k16_st_c_f16:
18419 case NVPTX::BI__hmma_m16n16k16_st_c_f32:
18420 case NVPTX::BI__hmma_m32n8k16_st_c_f16:
18421 case NVPTX::BI__hmma_m32n8k16_st_c_f32:
18422 case NVPTX::BI__hmma_m8n32k16_st_c_f16:
18423 case NVPTX::BI__hmma_m8n32k16_st_c_f32:
18424 case NVPTX::BI__imma_m16n16k16_st_c_i32:
18425 case NVPTX::BI__imma_m32n8k16_st_c_i32:
18426 case NVPTX::BI__imma_m8n32k16_st_c_i32:
18427 case NVPTX::BI__imma_m8n8k32_st_c_i32:
18428 case NVPTX::BI__bmma_m8n8k128_st_c_i32:
18429 case NVPTX::BI__dmma_m8n8k4_st_c_f64:
18430 case NVPTX::BI__mma_m16n16k8_st_c_f32: {
18431 Value *Dst = EmitScalarExpr(E->getArg(0));
18432 Address Src = EmitPointerWithAlignment(E->getArg(1));
18433 Value *Ldm = EmitScalarExpr(E->getArg(2));
18434 std::optional<llvm::APSInt> isColMajorArg =
18435 E->getArg(3)->getIntegerConstantExpr(getContext());
18436 if (!isColMajorArg)
18437 return nullptr;
18438 bool isColMajor = isColMajorArg->getSExtValue();
18439 NVPTXMmaLdstInfo II = getNVPTXMmaLdstInfo(BuiltinID);
18440 unsigned IID = isColMajor ? II.IID_col : II.IID_row;
18441 if (IID == 0)
18442 return nullptr;
18443 Function *Intrinsic =
18444 CGM.getIntrinsic(IID, Dst->getType());
18445 llvm::Type *ParamType = Intrinsic->getFunctionType()->getParamType(1);
18446 SmallVector<Value *, 10> Values = {Dst};
18447 for (unsigned i = 0; i < II.NumResults; ++i) {
18448 Value *V = Builder.CreateAlignedLoad(
18449 Src.getElementType(),
18450 Builder.CreateGEP(Src.getElementType(), Src.getPointer(),
18451 llvm::ConstantInt::get(IntTy, i)),
18452 CharUnits::fromQuantity(4));
18453 Values.push_back(Builder.CreateBitCast(V, ParamType));
18454 }
18455 Values.push_back(Ldm);
18456 Value *Result = Builder.CreateCall(Intrinsic, Values);
18457 return Result;
18458 }
18459
18460 // BI__hmma_m16n16k16_mma_<Dtype><CType>(d, a, b, c, layout, satf) -->
18461 // Intrinsic::nvvm_wmma_m16n16k16_mma_sync<layout A,B><DType><CType><Satf>
18462 case NVPTX::BI__hmma_m16n16k16_mma_f16f16:
18463 case NVPTX::BI__hmma_m16n16k16_mma_f32f16:
18464 case NVPTX::BI__hmma_m16n16k16_mma_f32f32:
18465 case NVPTX::BI__hmma_m16n16k16_mma_f16f32:
18466 case NVPTX::BI__hmma_m32n8k16_mma_f16f16:
18467 case NVPTX::BI__hmma_m32n8k16_mma_f32f16:
18468 case NVPTX::BI__hmma_m32n8k16_mma_f32f32:
18469 case NVPTX::BI__hmma_m32n8k16_mma_f16f32:
18470 case NVPTX::BI__hmma_m8n32k16_mma_f16f16:
18471 case NVPTX::BI__hmma_m8n32k16_mma_f32f16:
18472 case NVPTX::BI__hmma_m8n32k16_mma_f32f32:
18473 case NVPTX::BI__hmma_m8n32k16_mma_f16f32:
18474 case NVPTX::BI__imma_m16n16k16_mma_s8:
18475 case NVPTX::BI__imma_m16n16k16_mma_u8:
18476 case NVPTX::BI__imma_m32n8k16_mma_s8:
18477 case NVPTX::BI__imma_m32n8k16_mma_u8:
18478 case NVPTX::BI__imma_m8n32k16_mma_s8:
18479 case NVPTX::BI__imma_m8n32k16_mma_u8:
18480 case NVPTX::BI__imma_m8n8k32_mma_s4:
18481 case NVPTX::BI__imma_m8n8k32_mma_u4:
18482 case NVPTX::BI__bmma_m8n8k128_mma_xor_popc_b1:
18483 case NVPTX::BI__bmma_m8n8k128_mma_and_popc_b1:
18484 case NVPTX::BI__dmma_m8n8k4_mma_f64:
18485 case NVPTX::BI__mma_bf16_m16n16k16_mma_f32:
18486 case NVPTX::BI__mma_bf16_m8n32k16_mma_f32:
18487 case NVPTX::BI__mma_bf16_m32n8k16_mma_f32:
18488 case NVPTX::BI__mma_tf32_m16n16k8_mma_f32: {
18489 Address Dst = EmitPointerWithAlignment(E->getArg(0));
18490 Address SrcA = EmitPointerWithAlignment(E->getArg(1));
18491 Address SrcB = EmitPointerWithAlignment(E->getArg(2));
18492 Address SrcC = EmitPointerWithAlignment(E->getArg(3));
18493 std::optional<llvm::APSInt> LayoutArg =
18494 E->getArg(4)->getIntegerConstantExpr(getContext());
18495 if (!LayoutArg)
18496 return nullptr;
18497 int Layout = LayoutArg->getSExtValue();
18498 if (Layout < 0 || Layout > 3)
18499 return nullptr;
18500 llvm::APSInt SatfArg;
18501 if (BuiltinID == NVPTX::BI__bmma_m8n8k128_mma_xor_popc_b1 ||
18502 BuiltinID == NVPTX::BI__bmma_m8n8k128_mma_and_popc_b1)
18503 SatfArg = 0; // .b1 does not have satf argument.
18504 else if (std::optional<llvm::APSInt> OptSatfArg =
18505 E->getArg(5)->getIntegerConstantExpr(getContext()))
18506 SatfArg = *OptSatfArg;
18507 else
18508 return nullptr;
18509 bool Satf = SatfArg.getSExtValue();
18510 NVPTXMmaInfo MI = getNVPTXMmaInfo(BuiltinID);
18511 unsigned IID = MI.getMMAIntrinsic(Layout, Satf);
18512 if (IID == 0) // Unsupported combination of Layout/Satf.
18513 return nullptr;
18514
18515 SmallVector<Value *, 24> Values;
18516 Function *Intrinsic = CGM.getIntrinsic(IID);
18517 llvm::Type *AType = Intrinsic->getFunctionType()->getParamType(0);
18518 // Load A
18519 for (unsigned i = 0; i < MI.NumEltsA; ++i) {
18520 Value *V = Builder.CreateAlignedLoad(
18521 SrcA.getElementType(),
18522 Builder.CreateGEP(SrcA.getElementType(), SrcA.getPointer(),
18523 llvm::ConstantInt::get(IntTy, i)),
18524 CharUnits::fromQuantity(4));
18525 Values.push_back(Builder.CreateBitCast(V, AType));
18526 }
18527 // Load B
18528 llvm::Type *BType = Intrinsic->getFunctionType()->getParamType(MI.NumEltsA);
18529 for (unsigned i = 0; i < MI.NumEltsB; ++i) {
18530 Value *V = Builder.CreateAlignedLoad(
18531 SrcB.getElementType(),
18532 Builder.CreateGEP(SrcB.getElementType(), SrcB.getPointer(),
18533 llvm::ConstantInt::get(IntTy, i)),
18534 CharUnits::fromQuantity(4));
18535 Values.push_back(Builder.CreateBitCast(V, BType));
18536 }
18537 // Load C
18538 llvm::Type *CType =
18539 Intrinsic->getFunctionType()->getParamType(MI.NumEltsA + MI.NumEltsB);
18540 for (unsigned i = 0; i < MI.NumEltsC; ++i) {
18541 Value *V = Builder.CreateAlignedLoad(
18542 SrcC.getElementType(),
18543 Builder.CreateGEP(SrcC.getElementType(), SrcC.getPointer(),
18544 llvm::ConstantInt::get(IntTy, i)),
18545 CharUnits::fromQuantity(4));
18546 Values.push_back(Builder.CreateBitCast(V, CType));
18547 }
18548 Value *Result = Builder.CreateCall(Intrinsic, Values);
18549 llvm::Type *DType = Dst.getElementType();
18550 for (unsigned i = 0; i < MI.NumEltsD; ++i)
18551 Builder.CreateAlignedStore(
18552 Builder.CreateBitCast(Builder.CreateExtractValue(Result, i), DType),
18553 Builder.CreateGEP(Dst.getElementType(), Dst.getPointer(),
18554 llvm::ConstantInt::get(IntTy, i)),
18555 CharUnits::fromQuantity(4));
18556 return Result;
18557 }
18558 default:
18559 return nullptr;
18560 }
18561 }
18562
18563 namespace {
18564 struct BuiltinAlignArgs {
18565 llvm::Value *Src = nullptr;
18566 llvm::Type *SrcType = nullptr;
18567 llvm::Value *Alignment = nullptr;
18568 llvm::Value *Mask = nullptr;
18569 llvm::IntegerType *IntType = nullptr;
18570
18571 BuiltinAlignArgs(const CallExpr *E, CodeGenFunction &CGF) {
18572 QualType AstType = E->getArg(0)->getType();
18573 if (AstType->isArrayType())
18574 Src = CGF.EmitArrayToPointerDecay(E->getArg(0)).getPointer();
18575 else
18576 Src = CGF.EmitScalarExpr(E->getArg(0));
18577 SrcType = Src->getType();
18578 if (SrcType->isPointerTy()) {
18579 IntType = IntegerType::get(
18580 CGF.getLLVMContext(),
18581 CGF.CGM.getDataLayout().getIndexTypeSizeInBits(SrcType));
18582 } else {
18583 assert(SrcType->isIntegerTy());
18584 IntType = cast<llvm::IntegerType>(SrcType);
18585 }
18586 Alignment = CGF.EmitScalarExpr(E->getArg(1));
18587 Alignment = CGF.Builder.CreateZExtOrTrunc(Alignment, IntType, "alignment");
18588 auto *One = llvm::ConstantInt::get(IntType, 1);
18589 Mask = CGF.Builder.CreateSub(Alignment, One, "mask");
18590 }
18591 };
18592 } // namespace
18593
18594 /// Generate (x & (y-1)) == 0.
18595 RValue CodeGenFunction::EmitBuiltinIsAligned(const CallExpr *E) {
18596 BuiltinAlignArgs Args(E, *this);
18597 llvm::Value *SrcAddress = Args.Src;
18598 if (Args.SrcType->isPointerTy())
18599 SrcAddress =
18600 Builder.CreateBitOrPointerCast(Args.Src, Args.IntType, "src_addr");
18601 return RValue::get(Builder.CreateICmpEQ(
18602 Builder.CreateAnd(SrcAddress, Args.Mask, "set_bits"),
18603 llvm::Constant::getNullValue(Args.IntType), "is_aligned"));
18604 }
18605
18606 /// Generate (x & ~(y-1)) to align down or ((x+(y-1)) & ~(y-1)) to align up.
18607 /// Note: For pointer types we can avoid ptrtoint/inttoptr pairs by using the
18608 /// llvm.ptrmask intrinsic (with a GEP before in the align_up case).
18609 /// TODO: actually use ptrmask once most optimization passes know about it.
18610 RValue CodeGenFunction::EmitBuiltinAlignTo(const CallExpr *E, bool AlignUp) {
18611 BuiltinAlignArgs Args(E, *this);
18612 llvm::Value *SrcAddr = Args.Src;
18613 if (Args.Src->getType()->isPointerTy())
18614 SrcAddr = Builder.CreatePtrToInt(Args.Src, Args.IntType, "intptr");
18615 llvm::Value *SrcForMask = SrcAddr;
18616 if (AlignUp) {
18617 // When aligning up we have to first add the mask to ensure we go over the
18618 // next alignment value and then align down to the next valid multiple.
18619 // By adding the mask, we ensure that align_up on an already aligned
18620 // value will not change the value.
18621 SrcForMask = Builder.CreateAdd(SrcForMask, Args.Mask, "over_boundary");
18622 }
18623 // Invert the mask to only clear the lower bits.
18624 llvm::Value *InvertedMask = Builder.CreateNot(Args.Mask, "inverted_mask");
18625 llvm::Value *Result =
18626 Builder.CreateAnd(SrcForMask, InvertedMask, "aligned_result");
18627 if (Args.Src->getType()->isPointerTy()) {
18628 /// TODO: Use ptrmask instead of ptrtoint+gep once it is optimized well.
18629 // Result = Builder.CreateIntrinsic(
18630 // Intrinsic::ptrmask, {Args.SrcType, SrcForMask->getType(), Args.IntType},
18631 // {SrcForMask, NegatedMask}, nullptr, "aligned_result");
18632 Result->setName("aligned_intptr");
18633 llvm::Value *Difference = Builder.CreateSub(Result, SrcAddr, "diff");
18634 // The result must point to the same underlying allocation. This means we
18635 // can use an inbounds GEP to enable better optimization.
18636 Value *Base = EmitCastToVoidPtr(Args.Src);
18637 if (getLangOpts().isSignedOverflowDefined())
18638 Result = Builder.CreateGEP(Int8Ty, Base, Difference, "aligned_result");
18639 else
18640 Result = EmitCheckedInBoundsGEP(Int8Ty, Base, Difference,
18641 /*SignedIndices=*/true,
18642 /*isSubtraction=*/!AlignUp,
18643 E->getExprLoc(), "aligned_result");
18644 Result = Builder.CreatePointerCast(Result, Args.SrcType);
18645 // Emit an alignment assumption to ensure that the new alignment is
18646 // propagated to loads/stores, etc.
18647 emitAlignmentAssumption(Result, E, E->getExprLoc(), Args.Alignment);
18648 }
18649 assert(Result->getType() == Args.SrcType);
18650 return RValue::get(Result);
18651 }
18652
18653 Value *CodeGenFunction::EmitWebAssemblyBuiltinExpr(unsigned BuiltinID,
18654 const CallExpr *E) {
18655 switch (BuiltinID) {
18656 case WebAssembly::BI__builtin_wasm_memory_size: {
18657 llvm::Type *ResultType = ConvertType(E->getType());
18658 Value *I = EmitScalarExpr(E->getArg(0));
18659 Function *Callee =
18660 CGM.getIntrinsic(Intrinsic::wasm_memory_size, ResultType);
18661 return Builder.CreateCall(Callee, I);
18662 }
18663 case WebAssembly::BI__builtin_wasm_memory_grow: {
18664 llvm::Type *ResultType = ConvertType(E->getType());
18665 Value *Args[] = {EmitScalarExpr(E->getArg(0)),
18666 EmitScalarExpr(E->getArg(1))};
18667 Function *Callee =
18668 CGM.getIntrinsic(Intrinsic::wasm_memory_grow, ResultType);
18669 return Builder.CreateCall(Callee, Args);
18670 }
18671 case WebAssembly::BI__builtin_wasm_tls_size: {
18672 llvm::Type *ResultType = ConvertType(E->getType());
18673 Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_tls_size, ResultType);
18674 return Builder.CreateCall(Callee);
18675 }
18676 case WebAssembly::BI__builtin_wasm_tls_align: {
18677 llvm::Type *ResultType = ConvertType(E->getType());
18678 Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_tls_align, ResultType);
18679 return Builder.CreateCall(Callee);
18680 }
18681 case WebAssembly::BI__builtin_wasm_tls_base: {
18682 Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_tls_base);
18683 return Builder.CreateCall(Callee);
18684 }
18685 case WebAssembly::BI__builtin_wasm_throw: {
18686 Value *Tag = EmitScalarExpr(E->getArg(0));
18687 Value *Obj = EmitScalarExpr(E->getArg(1));
18688 Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_throw);
18689 return Builder.CreateCall(Callee, {Tag, Obj});
18690 }
18691 case WebAssembly::BI__builtin_wasm_rethrow: {
18692 Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_rethrow);
18693 return Builder.CreateCall(Callee);
18694 }
18695 case WebAssembly::BI__builtin_wasm_memory_atomic_wait32: {
18696 Value *Addr = EmitScalarExpr(E->getArg(0));
18697 Value *Expected = EmitScalarExpr(E->getArg(1));
18698 Value *Timeout = EmitScalarExpr(E->getArg(2));
18699 Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_memory_atomic_wait32);
18700 return Builder.CreateCall(Callee, {Addr, Expected, Timeout});
18701 }
18702 case WebAssembly::BI__builtin_wasm_memory_atomic_wait64: {
18703 Value *Addr = EmitScalarExpr(E->getArg(0));
18704 Value *Expected = EmitScalarExpr(E->getArg(1));
18705 Value *Timeout = EmitScalarExpr(E->getArg(2));
18706 Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_memory_atomic_wait64);
18707 return Builder.CreateCall(Callee, {Addr, Expected, Timeout});
18708 }
18709 case WebAssembly::BI__builtin_wasm_memory_atomic_notify: {
18710 Value *Addr = EmitScalarExpr(E->getArg(0));
18711 Value *Count = EmitScalarExpr(E->getArg(1));
18712 Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_memory_atomic_notify);
18713 return Builder.CreateCall(Callee, {Addr, Count});
18714 }
18715 case WebAssembly::BI__builtin_wasm_trunc_s_i32_f32:
18716 case WebAssembly::BI__builtin_wasm_trunc_s_i32_f64:
18717 case WebAssembly::BI__builtin_wasm_trunc_s_i64_f32:
18718 case WebAssembly::BI__builtin_wasm_trunc_s_i64_f64: {
18719 Value *Src = EmitScalarExpr(E->getArg(0));
18720 llvm::Type *ResT = ConvertType(E->getType());
18721 Function *Callee =
18722 CGM.getIntrinsic(Intrinsic::wasm_trunc_signed, {ResT, Src->getType()});
18723 return Builder.CreateCall(Callee, {Src});
18724 }
18725 case WebAssembly::BI__builtin_wasm_trunc_u_i32_f32:
18726 case WebAssembly::BI__builtin_wasm_trunc_u_i32_f64:
18727 case WebAssembly::BI__builtin_wasm_trunc_u_i64_f32:
18728 case WebAssembly::BI__builtin_wasm_trunc_u_i64_f64: {
18729 Value *Src = EmitScalarExpr(E->getArg(0));
18730 llvm::Type *ResT = ConvertType(E->getType());
18731 Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_trunc_unsigned,
18732 {ResT, Src->getType()});
18733 return Builder.CreateCall(Callee, {Src});
18734 }
18735 case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i32_f32:
18736 case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i32_f64:
18737 case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i64_f32:
18738 case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i64_f64:
18739 case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i32x4_f32x4: {
18740 Value *Src = EmitScalarExpr(E->getArg(0));
18741 llvm::Type *ResT = ConvertType(E->getType());
18742 Function *Callee =
18743 CGM.getIntrinsic(Intrinsic::fptosi_sat, {ResT, Src->getType()});
18744 return Builder.CreateCall(Callee, {Src});
18745 }
18746 case WebAssembly::BI__builtin_wasm_trunc_saturate_u_i32_f32:
18747 case WebAssembly::BI__builtin_wasm_trunc_saturate_u_i32_f64:
18748 case WebAssembly::BI__builtin_wasm_trunc_saturate_u_i64_f32:
18749 case WebAssembly::BI__builtin_wasm_trunc_saturate_u_i64_f64:
18750 case WebAssembly::BI__builtin_wasm_trunc_saturate_u_i32x4_f32x4: {
18751 Value *Src = EmitScalarExpr(E->getArg(0));
18752 llvm::Type *ResT = ConvertType(E->getType());
18753 Function *Callee =
18754 CGM.getIntrinsic(Intrinsic::fptoui_sat, {ResT, Src->getType()});
18755 return Builder.CreateCall(Callee, {Src});
18756 }
18757 case WebAssembly::BI__builtin_wasm_min_f32:
18758 case WebAssembly::BI__builtin_wasm_min_f64:
18759 case WebAssembly::BI__builtin_wasm_min_f32x4:
18760 case WebAssembly::BI__builtin_wasm_min_f64x2: {
18761 Value *LHS = EmitScalarExpr(E->getArg(0));
18762 Value *RHS = EmitScalarExpr(E->getArg(1));
18763 Function *Callee =
18764 CGM.getIntrinsic(Intrinsic::minimum, ConvertType(E->getType()));
18765 return Builder.CreateCall(Callee, {LHS, RHS});
18766 }
18767 case WebAssembly::BI__builtin_wasm_max_f32:
18768 case WebAssembly::BI__builtin_wasm_max_f64:
18769 case WebAssembly::BI__builtin_wasm_max_f32x4:
18770 case WebAssembly::BI__builtin_wasm_max_f64x2: {
18771 Value *LHS = EmitScalarExpr(E->getArg(0));
18772 Value *RHS = EmitScalarExpr(E->getArg(1));
18773 Function *Callee =
18774 CGM.getIntrinsic(Intrinsic::maximum, ConvertType(E->getType()));
18775 return Builder.CreateCall(Callee, {LHS, RHS});
18776 }
18777 case WebAssembly::BI__builtin_wasm_pmin_f32x4:
18778 case WebAssembly::BI__builtin_wasm_pmin_f64x2: {
18779 Value *LHS = EmitScalarExpr(E->getArg(0));
18780 Value *RHS = EmitScalarExpr(E->getArg(1));
18781 Function *Callee =
18782 CGM.getIntrinsic(Intrinsic::wasm_pmin, ConvertType(E->getType()));
18783 return Builder.CreateCall(Callee, {LHS, RHS});
18784 }
18785 case WebAssembly::BI__builtin_wasm_pmax_f32x4:
18786 case WebAssembly::BI__builtin_wasm_pmax_f64x2: {
18787 Value *LHS = EmitScalarExpr(E->getArg(0));
18788 Value *RHS = EmitScalarExpr(E->getArg(1));
18789 Function *Callee =
18790 CGM.getIntrinsic(Intrinsic::wasm_pmax, ConvertType(E->getType()));
18791 return Builder.CreateCall(Callee, {LHS, RHS});
18792 }
18793 case WebAssembly::BI__builtin_wasm_ceil_f32x4:
18794 case WebAssembly::BI__builtin_wasm_floor_f32x4:
18795 case WebAssembly::BI__builtin_wasm_trunc_f32x4:
18796 case WebAssembly::BI__builtin_wasm_nearest_f32x4:
18797 case WebAssembly::BI__builtin_wasm_ceil_f64x2:
18798 case WebAssembly::BI__builtin_wasm_floor_f64x2:
18799 case WebAssembly::BI__builtin_wasm_trunc_f64x2:
18800 case WebAssembly::BI__builtin_wasm_nearest_f64x2: {
18801 unsigned IntNo;
18802 switch (BuiltinID) {
18803 case WebAssembly::BI__builtin_wasm_ceil_f32x4:
18804 case WebAssembly::BI__builtin_wasm_ceil_f64x2:
18805 IntNo = Intrinsic::ceil;
18806 break;
18807 case WebAssembly::BI__builtin_wasm_floor_f32x4:
18808 case WebAssembly::BI__builtin_wasm_floor_f64x2:
18809 IntNo = Intrinsic::floor;
18810 break;
18811 case WebAssembly::BI__builtin_wasm_trunc_f32x4:
18812 case WebAssembly::BI__builtin_wasm_trunc_f64x2:
18813 IntNo = Intrinsic::trunc;
18814 break;
18815 case WebAssembly::BI__builtin_wasm_nearest_f32x4:
18816 case WebAssembly::BI__builtin_wasm_nearest_f64x2:
18817 IntNo = Intrinsic::nearbyint;
18818 break;
18819 default:
18820 llvm_unreachable("unexpected builtin ID");
18821 }
18822 Value *Value = EmitScalarExpr(E->getArg(0));
18823 Function *Callee = CGM.getIntrinsic(IntNo, ConvertType(E->getType()));
18824 return Builder.CreateCall(Callee, Value);
18825 }
18826 case WebAssembly::BI__builtin_wasm_swizzle_i8x16: {
18827 Value *Src = EmitScalarExpr(E->getArg(0));
18828 Value *Indices = EmitScalarExpr(E->getArg(1));
18829 Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_swizzle);
18830 return Builder.CreateCall(Callee, {Src, Indices});
18831 }
18832 case WebAssembly::BI__builtin_wasm_add_sat_s_i8x16:
18833 case WebAssembly::BI__builtin_wasm_add_sat_u_i8x16:
18834 case WebAssembly::BI__builtin_wasm_add_sat_s_i16x8:
18835 case WebAssembly::BI__builtin_wasm_add_sat_u_i16x8:
18836 case WebAssembly::BI__builtin_wasm_sub_sat_s_i8x16:
18837 case WebAssembly::BI__builtin_wasm_sub_sat_u_i8x16:
18838 case WebAssembly::BI__builtin_wasm_sub_sat_s_i16x8:
18839 case WebAssembly::BI__builtin_wasm_sub_sat_u_i16x8: {
18840 unsigned IntNo;
18841 switch (BuiltinID) {
18842 case WebAssembly::BI__builtin_wasm_add_sat_s_i8x16:
18843 case WebAssembly::BI__builtin_wasm_add_sat_s_i16x8:
18844 IntNo = Intrinsic::sadd_sat;
18845 break;
18846 case WebAssembly::BI__builtin_wasm_add_sat_u_i8x16:
18847 case WebAssembly::BI__builtin_wasm_add_sat_u_i16x8:
18848 IntNo = Intrinsic::uadd_sat;
18849 break;
18850 case WebAssembly::BI__builtin_wasm_sub_sat_s_i8x16:
18851 case WebAssembly::BI__builtin_wasm_sub_sat_s_i16x8:
18852 IntNo = Intrinsic::wasm_sub_sat_signed;
18853 break;
18854 case WebAssembly::BI__builtin_wasm_sub_sat_u_i8x16:
18855 case WebAssembly::BI__builtin_wasm_sub_sat_u_i16x8:
18856 IntNo = Intrinsic::wasm_sub_sat_unsigned;
18857 break;
18858 default:
18859 llvm_unreachable("unexpected builtin ID");
18860 }
18861 Value *LHS = EmitScalarExpr(E->getArg(0));
18862 Value *RHS = EmitScalarExpr(E->getArg(1));
18863 Function *Callee = CGM.getIntrinsic(IntNo, ConvertType(E->getType()));
18864 return Builder.CreateCall(Callee, {LHS, RHS});
18865 }
18866 case WebAssembly::BI__builtin_wasm_abs_i8x16:
18867 case WebAssembly::BI__builtin_wasm_abs_i16x8:
18868 case WebAssembly::BI__builtin_wasm_abs_i32x4:
18869 case WebAssembly::BI__builtin_wasm_abs_i64x2: {
18870 Value *Vec = EmitScalarExpr(E->getArg(0));
18871 Value *Neg = Builder.CreateNeg(Vec, "neg");
18872 Constant *Zero = llvm::Constant::getNullValue(Vec->getType());
18873 Value *ICmp = Builder.CreateICmpSLT(Vec, Zero, "abscond");
18874 return Builder.CreateSelect(ICmp, Neg, Vec, "abs");
18875 }
18876 case WebAssembly::BI__builtin_wasm_min_s_i8x16:
18877 case WebAssembly::BI__builtin_wasm_min_u_i8x16:
18878 case WebAssembly::BI__builtin_wasm_max_s_i8x16:
18879 case WebAssembly::BI__builtin_wasm_max_u_i8x16:
18880 case WebAssembly::BI__builtin_wasm_min_s_i16x8:
18881 case WebAssembly::BI__builtin_wasm_min_u_i16x8:
18882 case WebAssembly::BI__builtin_wasm_max_s_i16x8:
18883 case WebAssembly::BI__builtin_wasm_max_u_i16x8:
18884 case WebAssembly::BI__builtin_wasm_min_s_i32x4:
18885 case WebAssembly::BI__builtin_wasm_min_u_i32x4:
18886 case WebAssembly::BI__builtin_wasm_max_s_i32x4:
18887 case WebAssembly::BI__builtin_wasm_max_u_i32x4: {
18888 Value *LHS = EmitScalarExpr(E->getArg(0));
18889 Value *RHS = EmitScalarExpr(E->getArg(1));
18890 Value *ICmp;
18891 switch (BuiltinID) {
18892 case WebAssembly::BI__builtin_wasm_min_s_i8x16:
18893 case WebAssembly::BI__builtin_wasm_min_s_i16x8:
18894 case WebAssembly::BI__builtin_wasm_min_s_i32x4:
18895 ICmp = Builder.CreateICmpSLT(LHS, RHS);
18896 break;
18897 case WebAssembly::BI__builtin_wasm_min_u_i8x16:
18898 case WebAssembly::BI__builtin_wasm_min_u_i16x8:
18899 case WebAssembly::BI__builtin_wasm_min_u_i32x4:
18900 ICmp = Builder.CreateICmpULT(LHS, RHS);
18901 break;
18902 case WebAssembly::BI__builtin_wasm_max_s_i8x16:
18903 case WebAssembly::BI__builtin_wasm_max_s_i16x8:
18904 case WebAssembly::BI__builtin_wasm_max_s_i32x4:
18905 ICmp = Builder.CreateICmpSGT(LHS, RHS);
18906 break;
18907 case WebAssembly::BI__builtin_wasm_max_u_i8x16:
18908 case WebAssembly::BI__builtin_wasm_max_u_i16x8:
18909 case WebAssembly::BI__builtin_wasm_max_u_i32x4:
18910 ICmp = Builder.CreateICmpUGT(LHS, RHS);
18911 break;
18912 default:
18913 llvm_unreachable("unexpected builtin ID");
18914 }
18915 return Builder.CreateSelect(ICmp, LHS, RHS);
18916 }
18917 case WebAssembly::BI__builtin_wasm_avgr_u_i8x16:
18918 case WebAssembly::BI__builtin_wasm_avgr_u_i16x8: {
18919 Value *LHS = EmitScalarExpr(E->getArg(0));
18920 Value *RHS = EmitScalarExpr(E->getArg(1));
18921 Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_avgr_unsigned,
18922 ConvertType(E->getType()));
18923 return Builder.CreateCall(Callee, {LHS, RHS});
18924 }
18925 case WebAssembly::BI__builtin_wasm_q15mulr_sat_s_i16x8: {
18926 Value *LHS = EmitScalarExpr(E->getArg(0));
18927 Value *RHS = EmitScalarExpr(E->getArg(1));
18928 Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_q15mulr_sat_signed);
18929 return Builder.CreateCall(Callee, {LHS, RHS});
18930 }
18931 case WebAssembly::BI__builtin_wasm_extadd_pairwise_i8x16_s_i16x8:
18932 case WebAssembly::BI__builtin_wasm_extadd_pairwise_i8x16_u_i16x8:
18933 case WebAssembly::BI__builtin_wasm_extadd_pairwise_i16x8_s_i32x4:
18934 case WebAssembly::BI__builtin_wasm_extadd_pairwise_i16x8_u_i32x4: {
18935 Value *Vec = EmitScalarExpr(E->getArg(0));
18936 unsigned IntNo;
18937 switch (BuiltinID) {
18938 case WebAssembly::BI__builtin_wasm_extadd_pairwise_i8x16_s_i16x8:
18939 case WebAssembly::BI__builtin_wasm_extadd_pairwise_i16x8_s_i32x4:
18940 IntNo = Intrinsic::wasm_extadd_pairwise_signed;
18941 break;
18942 case WebAssembly::BI__builtin_wasm_extadd_pairwise_i8x16_u_i16x8:
18943 case WebAssembly::BI__builtin_wasm_extadd_pairwise_i16x8_u_i32x4:
18944 IntNo = Intrinsic::wasm_extadd_pairwise_unsigned;
18945 break;
18946 default:
18947 llvm_unreachable("unexpected builtin ID");
18948 }
18949
18950 Function *Callee = CGM.getIntrinsic(IntNo, ConvertType(E->getType()));
18951 return Builder.CreateCall(Callee, Vec);
18952 }
18953 case WebAssembly::BI__builtin_wasm_bitselect: {
18954 Value *V1 = EmitScalarExpr(E->getArg(0));
18955 Value *V2 = EmitScalarExpr(E->getArg(1));
18956 Value *C = EmitScalarExpr(E->getArg(2));
18957 Function *Callee =
18958 CGM.getIntrinsic(Intrinsic::wasm_bitselect, ConvertType(E->getType()));
18959 return Builder.CreateCall(Callee, {V1, V2, C});
18960 }
18961 case WebAssembly::BI__builtin_wasm_dot_s_i32x4_i16x8: {
18962 Value *LHS = EmitScalarExpr(E->getArg(0));
18963 Value *RHS = EmitScalarExpr(E->getArg(1));
18964 Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_dot);
18965 return Builder.CreateCall(Callee, {LHS, RHS});
18966 }
18967 case WebAssembly::BI__builtin_wasm_popcnt_i8x16: {
18968 Value *Vec = EmitScalarExpr(E->getArg(0));
18969 Function *Callee =
18970 CGM.getIntrinsic(Intrinsic::ctpop, ConvertType(E->getType()));
18971 return Builder.CreateCall(Callee, {Vec});
18972 }
18973 case WebAssembly::BI__builtin_wasm_any_true_v128:
18974 case WebAssembly::BI__builtin_wasm_all_true_i8x16:
18975 case WebAssembly::BI__builtin_wasm_all_true_i16x8:
18976 case WebAssembly::BI__builtin_wasm_all_true_i32x4:
18977 case WebAssembly::BI__builtin_wasm_all_true_i64x2: {
18978 unsigned IntNo;
18979 switch (BuiltinID) {
18980 case WebAssembly::BI__builtin_wasm_any_true_v128:
18981 IntNo = Intrinsic::wasm_anytrue;
18982 break;
18983 case WebAssembly::BI__builtin_wasm_all_true_i8x16:
18984 case WebAssembly::BI__builtin_wasm_all_true_i16x8:
18985 case WebAssembly::BI__builtin_wasm_all_true_i32x4:
18986 case WebAssembly::BI__builtin_wasm_all_true_i64x2:
18987 IntNo = Intrinsic::wasm_alltrue;
18988 break;
18989 default:
18990 llvm_unreachable("unexpected builtin ID");
18991 }
18992 Value *Vec = EmitScalarExpr(E->getArg(0));
18993 Function *Callee = CGM.getIntrinsic(IntNo, Vec->getType());
18994 return Builder.CreateCall(Callee, {Vec});
18995 }
18996 case WebAssembly::BI__builtin_wasm_bitmask_i8x16:
18997 case WebAssembly::BI__builtin_wasm_bitmask_i16x8:
18998 case WebAssembly::BI__builtin_wasm_bitmask_i32x4:
18999 case WebAssembly::BI__builtin_wasm_bitmask_i64x2: {
19000 Value *Vec = EmitScalarExpr(E->getArg(0));
19001 Function *Callee =
19002 CGM.getIntrinsic(Intrinsic::wasm_bitmask, Vec->getType());
19003 return Builder.CreateCall(Callee, {Vec});
19004 }
19005 case WebAssembly::BI__builtin_wasm_abs_f32x4:
19006 case WebAssembly::BI__builtin_wasm_abs_f64x2: {
19007 Value *Vec = EmitScalarExpr(E->getArg(0));
19008 Function *Callee = CGM.getIntrinsic(Intrinsic::fabs, Vec->getType());
19009 return Builder.CreateCall(Callee, {Vec});
19010 }
19011 case WebAssembly::BI__builtin_wasm_sqrt_f32x4:
19012 case WebAssembly::BI__builtin_wasm_sqrt_f64x2: {
19013 Value *Vec = EmitScalarExpr(E->getArg(0));
19014 Function *Callee = CGM.getIntrinsic(Intrinsic::sqrt, Vec->getType());
19015 return Builder.CreateCall(Callee, {Vec});
19016 }
19017 case WebAssembly::BI__builtin_wasm_narrow_s_i8x16_i16x8:
19018 case WebAssembly::BI__builtin_wasm_narrow_u_i8x16_i16x8:
19019 case WebAssembly::BI__builtin_wasm_narrow_s_i16x8_i32x4:
19020 case WebAssembly::BI__builtin_wasm_narrow_u_i16x8_i32x4: {
19021 Value *Low = EmitScalarExpr(E->getArg(0));
19022 Value *High = EmitScalarExpr(E->getArg(1));
19023 unsigned IntNo;
19024 switch (BuiltinID) {
19025 case WebAssembly::BI__builtin_wasm_narrow_s_i8x16_i16x8:
19026 case WebAssembly::BI__builtin_wasm_narrow_s_i16x8_i32x4:
19027 IntNo = Intrinsic::wasm_narrow_signed;
19028 break;
19029 case WebAssembly::BI__builtin_wasm_narrow_u_i8x16_i16x8:
19030 case WebAssembly::BI__builtin_wasm_narrow_u_i16x8_i32x4:
19031 IntNo = Intrinsic::wasm_narrow_unsigned;
19032 break;
19033 default:
19034 llvm_unreachable("unexpected builtin ID");
19035 }
19036 Function *Callee =
19037 CGM.getIntrinsic(IntNo, {ConvertType(E->getType()), Low->getType()});
19038 return Builder.CreateCall(Callee, {Low, High});
19039 }
19040 case WebAssembly::BI__builtin_wasm_trunc_sat_s_zero_f64x2_i32x4:
19041 case WebAssembly::BI__builtin_wasm_trunc_sat_u_zero_f64x2_i32x4: {
19042 Value *Vec = EmitScalarExpr(E->getArg(0));
19043 unsigned IntNo;
19044 switch (BuiltinID) {
19045 case WebAssembly::BI__builtin_wasm_trunc_sat_s_zero_f64x2_i32x4:
19046 IntNo = Intrinsic::fptosi_sat;
19047 break;
19048 case WebAssembly::BI__builtin_wasm_trunc_sat_u_zero_f64x2_i32x4:
19049 IntNo = Intrinsic::fptoui_sat;
19050 break;
19051 default:
19052 llvm_unreachable("unexpected builtin ID");
19053 }
19054 llvm::Type *SrcT = Vec->getType();
19055 llvm::Type *TruncT = SrcT->getWithNewType(Builder.getInt32Ty());
19056 Function *Callee = CGM.getIntrinsic(IntNo, {TruncT, SrcT});
19057 Value *Trunc = Builder.CreateCall(Callee, Vec);
19058 Value *Splat = Constant::getNullValue(TruncT);
19059 return Builder.CreateShuffleVector(Trunc, Splat, ArrayRef<int>{0, 1, 2, 3});
19060 }
19061 case WebAssembly::BI__builtin_wasm_shuffle_i8x16: {
19062 Value *Ops[18];
19063 size_t OpIdx = 0;
19064 Ops[OpIdx++] = EmitScalarExpr(E->getArg(0));
19065 Ops[OpIdx++] = EmitScalarExpr(E->getArg(1));
19066 while (OpIdx < 18) {
19067 std::optional<llvm::APSInt> LaneConst =
19068 E->getArg(OpIdx)->getIntegerConstantExpr(getContext());
19069 assert(LaneConst && "Constant arg isn't actually constant?");
19070 Ops[OpIdx++] = llvm::ConstantInt::get(getLLVMContext(), *LaneConst);
19071 }
19072 Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_shuffle);
19073 return Builder.CreateCall(Callee, Ops);
19074 }
19075 case WebAssembly::BI__builtin_wasm_relaxed_madd_f32x4:
19076 case WebAssembly::BI__builtin_wasm_relaxed_nmadd_f32x4:
19077 case WebAssembly::BI__builtin_wasm_relaxed_madd_f64x2:
19078 case WebAssembly::BI__builtin_wasm_relaxed_nmadd_f64x2: {
19079 Value *A = EmitScalarExpr(E->getArg(0));
19080 Value *B = EmitScalarExpr(E->getArg(1));
19081 Value *C = EmitScalarExpr(E->getArg(2));
19082 unsigned IntNo;
19083 switch (BuiltinID) {
19084 case WebAssembly::BI__builtin_wasm_relaxed_madd_f32x4:
19085 case WebAssembly::BI__builtin_wasm_relaxed_madd_f64x2:
19086 IntNo = Intrinsic::wasm_relaxed_madd;
19087 break;
19088 case WebAssembly::BI__builtin_wasm_relaxed_nmadd_f32x4:
19089 case WebAssembly::BI__builtin_wasm_relaxed_nmadd_f64x2:
19090 IntNo = Intrinsic::wasm_relaxed_nmadd;
19091 break;
19092 default:
19093 llvm_unreachable("unexpected builtin ID");
19094 }
19095 Function *Callee = CGM.getIntrinsic(IntNo, A->getType());
19096 return Builder.CreateCall(Callee, {A, B, C});
19097 }
19098 case WebAssembly::BI__builtin_wasm_relaxed_laneselect_i8x16:
19099 case WebAssembly::BI__builtin_wasm_relaxed_laneselect_i16x8:
19100 case WebAssembly::BI__builtin_wasm_relaxed_laneselect_i32x4:
19101 case WebAssembly::BI__builtin_wasm_relaxed_laneselect_i64x2: {
19102 Value *A = EmitScalarExpr(E->getArg(0));
19103 Value *B = EmitScalarExpr(E->getArg(1));
19104 Value *C = EmitScalarExpr(E->getArg(2));
19105 Function *Callee =
19106 CGM.getIntrinsic(Intrinsic::wasm_relaxed_laneselect, A->getType());
19107 return Builder.CreateCall(Callee, {A, B, C});
19108 }
19109 case WebAssembly::BI__builtin_wasm_relaxed_swizzle_i8x16: {
19110 Value *Src = EmitScalarExpr(E->getArg(0));
19111 Value *Indices = EmitScalarExpr(E->getArg(1));
19112 Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_relaxed_swizzle);
19113 return Builder.CreateCall(Callee, {Src, Indices});
19114 }
19115 case WebAssembly::BI__builtin_wasm_relaxed_min_f32x4:
19116 case WebAssembly::BI__builtin_wasm_relaxed_max_f32x4:
19117 case WebAssembly::BI__builtin_wasm_relaxed_min_f64x2:
19118 case WebAssembly::BI__builtin_wasm_relaxed_max_f64x2: {
19119 Value *LHS = EmitScalarExpr(E->getArg(0));
19120 Value *RHS = EmitScalarExpr(E->getArg(1));
19121 unsigned IntNo;
19122 switch (BuiltinID) {
19123 case WebAssembly::BI__builtin_wasm_relaxed_min_f32x4:
19124 case WebAssembly::BI__builtin_wasm_relaxed_min_f64x2:
19125 IntNo = Intrinsic::wasm_relaxed_min;
19126 break;
19127 case WebAssembly::BI__builtin_wasm_relaxed_max_f32x4:
19128 case WebAssembly::BI__builtin_wasm_relaxed_max_f64x2:
19129 IntNo = Intrinsic::wasm_relaxed_max;
19130 break;
19131 default:
19132 llvm_unreachable("unexpected builtin ID");
19133 }
19134 Function *Callee = CGM.getIntrinsic(IntNo, LHS->getType());
19135 return Builder.CreateCall(Callee, {LHS, RHS});
19136 }
19137 case WebAssembly::BI__builtin_wasm_relaxed_trunc_s_i32x4_f32x4:
19138 case WebAssembly::BI__builtin_wasm_relaxed_trunc_u_i32x4_f32x4:
19139 case WebAssembly::BI__builtin_wasm_relaxed_trunc_s_zero_i32x4_f64x2:
19140 case WebAssembly::BI__builtin_wasm_relaxed_trunc_u_zero_i32x4_f64x2: {
19141 Value *Vec = EmitScalarExpr(E->getArg(0));
19142 unsigned IntNo;
19143 switch (BuiltinID) {
19144 case WebAssembly::BI__builtin_wasm_relaxed_trunc_s_i32x4_f32x4:
19145 IntNo = Intrinsic::wasm_relaxed_trunc_signed;
19146 break;
19147 case WebAssembly::BI__builtin_wasm_relaxed_trunc_u_i32x4_f32x4:
19148 IntNo = Intrinsic::wasm_relaxed_trunc_unsigned;
19149 break;
19150 case WebAssembly::BI__builtin_wasm_relaxed_trunc_s_zero_i32x4_f64x2:
19151 IntNo = Intrinsic::wasm_relaxed_trunc_signed_zero;
19152 break;
19153 case WebAssembly::BI__builtin_wasm_relaxed_trunc_u_zero_i32x4_f64x2:
19154 IntNo = Intrinsic::wasm_relaxed_trunc_unsigned_zero;
19155 break;
19156 default:
19157 llvm_unreachable("unexpected builtin ID");
19158 }
19159 Function *Callee = CGM.getIntrinsic(IntNo);
19160 return Builder.CreateCall(Callee, {Vec});
19161 }
19162 case WebAssembly::BI__builtin_wasm_relaxed_q15mulr_s_i16x8: {
19163 Value *LHS = EmitScalarExpr(E->getArg(0));
19164 Value *RHS = EmitScalarExpr(E->getArg(1));
19165 Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_relaxed_q15mulr_signed);
19166 return Builder.CreateCall(Callee, {LHS, RHS});
19167 }
19168 case WebAssembly::BI__builtin_wasm_relaxed_dot_i8x16_i7x16_s_i16x8: {
19169 Value *LHS = EmitScalarExpr(E->getArg(0));
19170 Value *RHS = EmitScalarExpr(E->getArg(1));
19171 Function *Callee =
19172 CGM.getIntrinsic(Intrinsic::wasm_relaxed_dot_i8x16_i7x16_signed);
19173 return Builder.CreateCall(Callee, {LHS, RHS});
19174 }
19175 case WebAssembly::BI__builtin_wasm_relaxed_dot_i8x16_i7x16_add_s_i32x4: {
19176 Value *LHS = EmitScalarExpr(E->getArg(0));
19177 Value *RHS = EmitScalarExpr(E->getArg(1));
19178 Value *Acc = EmitScalarExpr(E->getArg(2));
19179 Function *Callee =
19180 CGM.getIntrinsic(Intrinsic::wasm_relaxed_dot_i8x16_i7x16_add_signed);
19181 return Builder.CreateCall(Callee, {LHS, RHS, Acc});
19182 }
19183 case WebAssembly::BI__builtin_wasm_relaxed_dot_bf16x8_add_f32_f32x4: {
19184 Value *LHS = EmitScalarExpr(E->getArg(0));
19185 Value *RHS = EmitScalarExpr(E->getArg(1));
19186 Value *Acc = EmitScalarExpr(E->getArg(2));
19187 Function *Callee =
19188 CGM.getIntrinsic(Intrinsic::wasm_relaxed_dot_bf16x8_add_f32);
19189 return Builder.CreateCall(Callee, {LHS, RHS, Acc});
19190 }
19191 default:
19192 return nullptr;
19193 }
19194 }
19195
19196 static std::pair<Intrinsic::ID, unsigned>
19197 getIntrinsicForHexagonNonClangBuiltin(unsigned BuiltinID) {
19198 struct Info {
19199 unsigned BuiltinID;
19200 Intrinsic::ID IntrinsicID;
19201 unsigned VecLen;
19202 };
19203 static Info Infos[] = {
19204 #define CUSTOM_BUILTIN_MAPPING(x,s) \
19205 { Hexagon::BI__builtin_HEXAGON_##x, Intrinsic::hexagon_##x, s },
19206 CUSTOM_BUILTIN_MAPPING(L2_loadrub_pci, 0)
19207 CUSTOM_BUILTIN_MAPPING(L2_loadrb_pci, 0)
19208 CUSTOM_BUILTIN_MAPPING(L2_loadruh_pci, 0)
19209 CUSTOM_BUILTIN_MAPPING(L2_loadrh_pci, 0)
19210 CUSTOM_BUILTIN_MAPPING(L2_loadri_pci, 0)
19211 CUSTOM_BUILTIN_MAPPING(L2_loadrd_pci, 0)
19212 CUSTOM_BUILTIN_MAPPING(L2_loadrub_pcr, 0)
19213 CUSTOM_BUILTIN_MAPPING(L2_loadrb_pcr, 0)
19214 CUSTOM_BUILTIN_MAPPING(L2_loadruh_pcr, 0)
19215 CUSTOM_BUILTIN_MAPPING(L2_loadrh_pcr, 0)
19216 CUSTOM_BUILTIN_MAPPING(L2_loadri_pcr, 0)
19217 CUSTOM_BUILTIN_MAPPING(L2_loadrd_pcr, 0)
19218 CUSTOM_BUILTIN_MAPPING(S2_storerb_pci, 0)
19219 CUSTOM_BUILTIN_MAPPING(S2_storerh_pci, 0)
19220 CUSTOM_BUILTIN_MAPPING(S2_storerf_pci, 0)
19221 CUSTOM_BUILTIN_MAPPING(S2_storeri_pci, 0)
19222 CUSTOM_BUILTIN_MAPPING(S2_storerd_pci, 0)
19223 CUSTOM_BUILTIN_MAPPING(S2_storerb_pcr, 0)
19224 CUSTOM_BUILTIN_MAPPING(S2_storerh_pcr, 0)
19225 CUSTOM_BUILTIN_MAPPING(S2_storerf_pcr, 0)
19226 CUSTOM_BUILTIN_MAPPING(S2_storeri_pcr, 0)
19227 CUSTOM_BUILTIN_MAPPING(S2_storerd_pcr, 0)
19228 // Legacy builtins that take a vector in place of a vector predicate.
19229 CUSTOM_BUILTIN_MAPPING(V6_vmaskedstoreq, 64)
19230 CUSTOM_BUILTIN_MAPPING(V6_vmaskedstorenq, 64)
19231 CUSTOM_BUILTIN_MAPPING(V6_vmaskedstorentq, 64)
19232 CUSTOM_BUILTIN_MAPPING(V6_vmaskedstorentnq, 64)
19233 CUSTOM_BUILTIN_MAPPING(V6_vmaskedstoreq_128B, 128)
19234 CUSTOM_BUILTIN_MAPPING(V6_vmaskedstorenq_128B, 128)
19235 CUSTOM_BUILTIN_MAPPING(V6_vmaskedstorentq_128B, 128)
19236 CUSTOM_BUILTIN_MAPPING(V6_vmaskedstorentnq_128B, 128)
19237 #include "clang/Basic/BuiltinsHexagonMapCustomDep.def"
19238 #undef CUSTOM_BUILTIN_MAPPING
19239 };
19240
19241 auto CmpInfo = [] (Info A, Info B) { return A.BuiltinID < B.BuiltinID; };
19242 static const bool SortOnce = (llvm::sort(Infos, CmpInfo), true);
19243 (void)SortOnce;
19244
19245 const Info *F = llvm::lower_bound(Infos, Info{BuiltinID, 0, 0}, CmpInfo);
19246 if (F == std::end(Infos) || F->BuiltinID != BuiltinID)
19247 return {Intrinsic::not_intrinsic, 0};
19248
19249 return {F->IntrinsicID, F->VecLen};
19250 }
19251
19252 Value *CodeGenFunction::EmitHexagonBuiltinExpr(unsigned BuiltinID,
19253 const CallExpr *E) {
19254 Intrinsic::ID ID;
19255 unsigned VecLen;
19256 std::tie(ID, VecLen) = getIntrinsicForHexagonNonClangBuiltin(BuiltinID);
19257
19258 auto MakeCircOp = [this, E](unsigned IntID, bool IsLoad) {
19259 // The base pointer is passed by address, so it needs to be loaded.
19260 Address A = EmitPointerWithAlignment(E->getArg(0));
19261 Address BP = Address(Builder.CreateBitCast(
19262 A.getPointer(), Int8PtrPtrTy), Int8PtrTy, A.getAlignment());
19263 llvm::Value *Base = Builder.CreateLoad(BP);
19264 // The treatment of both loads and stores is the same: the arguments for
19265 // the builtin are the same as the arguments for the intrinsic.
19266 // Load:
19267 // builtin(Base, Inc, Mod, Start) -> intr(Base, Inc, Mod, Start)
19268 // builtin(Base, Mod, Start) -> intr(Base, Mod, Start)
19269 // Store:
19270 // builtin(Base, Inc, Mod, Val, Start) -> intr(Base, Inc, Mod, Val, Start)
19271 // builtin(Base, Mod, Val, Start) -> intr(Base, Mod, Val, Start)
19272 SmallVector<llvm::Value*,5> Ops = { Base };
19273 for (unsigned i = 1, e = E->getNumArgs(); i != e; ++i)
19274 Ops.push_back(EmitScalarExpr(E->getArg(i)));
19275
19276 llvm::Value *Result = Builder.CreateCall(CGM.getIntrinsic(IntID), Ops);
19277 // The load intrinsics generate two results (Value, NewBase), stores
19278 // generate one (NewBase). The new base address needs to be stored.
19279 llvm::Value *NewBase = IsLoad ? Builder.CreateExtractValue(Result, 1)
19280 : Result;
19281 llvm::Value *LV = Builder.CreateBitCast(
19282 EmitScalarExpr(E->getArg(0)), NewBase->getType()->getPointerTo());
19283 Address Dest = EmitPointerWithAlignment(E->getArg(0));
19284 llvm::Value *RetVal =
19285 Builder.CreateAlignedStore(NewBase, LV, Dest.getAlignment());
19286 if (IsLoad)
19287 RetVal = Builder.CreateExtractValue(Result, 0);
19288 return RetVal;
19289 };
19290
19291 // Handle the conversion of bit-reverse load intrinsics to bit code.
19292 // The intrinsic call after this function only reads from memory and the
19293 // write to memory is dealt by the store instruction.
19294 auto MakeBrevLd = [this, E](unsigned IntID, llvm::Type *DestTy) {
19295 // The intrinsic generates one result, which is the new value for the base
19296 // pointer. It needs to be returned. The result of the load instruction is
19297 // passed to intrinsic by address, so the value needs to be stored.
19298 llvm::Value *BaseAddress =
19299 Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int8PtrTy);
19300
19301 // Expressions like &(*pt++) will be incremented per evaluation.
19302 // EmitPointerWithAlignment and EmitScalarExpr evaluates the expression
19303 // per call.
19304 Address DestAddr = EmitPointerWithAlignment(E->getArg(1));
19305 DestAddr = Address(Builder.CreateBitCast(DestAddr.getPointer(), Int8PtrTy),
19306 Int8Ty, DestAddr.getAlignment());
19307 llvm::Value *DestAddress = DestAddr.getPointer();
19308
19309 // Operands are Base, Dest, Modifier.
19310 // The intrinsic format in LLVM IR is defined as
19311 // { ValueType, i8* } (i8*, i32).
19312 llvm::Value *Result = Builder.CreateCall(
19313 CGM.getIntrinsic(IntID), {BaseAddress, EmitScalarExpr(E->getArg(2))});
19314
19315 // The value needs to be stored as the variable is passed by reference.
19316 llvm::Value *DestVal = Builder.CreateExtractValue(Result, 0);
19317
19318 // The store needs to be truncated to fit the destination type.
19319 // While i32 and i64 are natively supported on Hexagon, i8 and i16 needs
19320 // to be handled with stores of respective destination type.
19321 DestVal = Builder.CreateTrunc(DestVal, DestTy);
19322
19323 llvm::Value *DestForStore =
19324 Builder.CreateBitCast(DestAddress, DestVal->getType()->getPointerTo());
19325 Builder.CreateAlignedStore(DestVal, DestForStore, DestAddr.getAlignment());
19326 // The updated value of the base pointer is returned.
19327 return Builder.CreateExtractValue(Result, 1);
19328 };
19329
19330 auto V2Q = [this, VecLen] (llvm::Value *Vec) {
19331 Intrinsic::ID ID = VecLen == 128 ? Intrinsic::hexagon_V6_vandvrt_128B
19332 : Intrinsic::hexagon_V6_vandvrt;
19333 return Builder.CreateCall(CGM.getIntrinsic(ID),
19334 {Vec, Builder.getInt32(-1)});
19335 };
19336 auto Q2V = [this, VecLen] (llvm::Value *Pred) {
19337 Intrinsic::ID ID = VecLen == 128 ? Intrinsic::hexagon_V6_vandqrt_128B
19338 : Intrinsic::hexagon_V6_vandqrt;
19339 return Builder.CreateCall(CGM.getIntrinsic(ID),
19340 {Pred, Builder.getInt32(-1)});
19341 };
19342
19343 switch (BuiltinID) {
19344 // These intrinsics return a tuple {Vector, VectorPred} in LLVM IR,
19345 // and the corresponding C/C++ builtins use loads/stores to update
19346 // the predicate.
19347 case Hexagon::BI__builtin_HEXAGON_V6_vaddcarry:
19348 case Hexagon::BI__builtin_HEXAGON_V6_vaddcarry_128B:
19349 case Hexagon::BI__builtin_HEXAGON_V6_vsubcarry:
19350 case Hexagon::BI__builtin_HEXAGON_V6_vsubcarry_128B: {
19351 // Get the type from the 0-th argument.
19352 llvm::Type *VecType = ConvertType(E->getArg(0)->getType());
19353 Address PredAddr = Builder.CreateElementBitCast(
19354 EmitPointerWithAlignment(E->getArg(2)), VecType);
19355 llvm::Value *PredIn = V2Q(Builder.CreateLoad(PredAddr));
19356 llvm::Value *Result = Builder.CreateCall(CGM.getIntrinsic(ID),
19357 {EmitScalarExpr(E->getArg(0)), EmitScalarExpr(E->getArg(1)), PredIn});
19358
19359 llvm::Value *PredOut = Builder.CreateExtractValue(Result, 1);
19360 Builder.CreateAlignedStore(Q2V(PredOut), PredAddr.getPointer(),
19361 PredAddr.getAlignment());
19362 return Builder.CreateExtractValue(Result, 0);
19363 }
19364 // These are identical to the builtins above, except they don't consume
19365 // input carry, only generate carry-out. Since they still produce two
19366 // outputs, generate the store of the predicate, but no load.
19367 case Hexagon::BI__builtin_HEXAGON_V6_vaddcarryo:
19368 case Hexagon::BI__builtin_HEXAGON_V6_vaddcarryo_128B:
19369 case Hexagon::BI__builtin_HEXAGON_V6_vsubcarryo:
19370 case Hexagon::BI__builtin_HEXAGON_V6_vsubcarryo_128B: {
19371 // Get the type from the 0-th argument.
19372 llvm::Type *VecType = ConvertType(E->getArg(0)->getType());
19373 Address PredAddr = Builder.CreateElementBitCast(
19374 EmitPointerWithAlignment(E->getArg(2)), VecType);
19375 llvm::Value *Result = Builder.CreateCall(CGM.getIntrinsic(ID),
19376 {EmitScalarExpr(E->getArg(0)), EmitScalarExpr(E->getArg(1))});
19377
19378 llvm::Value *PredOut = Builder.CreateExtractValue(Result, 1);
19379 Builder.CreateAlignedStore(Q2V(PredOut), PredAddr.getPointer(),
19380 PredAddr.getAlignment());
19381 return Builder.CreateExtractValue(Result, 0);
19382 }
19383
19384 case Hexagon::BI__builtin_HEXAGON_V6_vmaskedstoreq:
19385 case Hexagon::BI__builtin_HEXAGON_V6_vmaskedstorenq:
19386 case Hexagon::BI__builtin_HEXAGON_V6_vmaskedstorentq:
19387 case Hexagon::BI__builtin_HEXAGON_V6_vmaskedstorentnq:
19388 case Hexagon::BI__builtin_HEXAGON_V6_vmaskedstoreq_128B:
19389 case Hexagon::BI__builtin_HEXAGON_V6_vmaskedstorenq_128B:
19390 case Hexagon::BI__builtin_HEXAGON_V6_vmaskedstorentq_128B:
19391 case Hexagon::BI__builtin_HEXAGON_V6_vmaskedstorentnq_128B: {
19392 SmallVector<llvm::Value*,4> Ops;
19393 const Expr *PredOp = E->getArg(0);
19394 // There will be an implicit cast to a boolean vector. Strip it.
19395 if (auto *Cast = dyn_cast<ImplicitCastExpr>(PredOp)) {
19396 if (Cast->getCastKind() == CK_BitCast)
19397 PredOp = Cast->getSubExpr();
19398 Ops.push_back(V2Q(EmitScalarExpr(PredOp)));
19399 }
19400 for (int i = 1, e = E->getNumArgs(); i != e; ++i)
19401 Ops.push_back(EmitScalarExpr(E->getArg(i)));
19402 return Builder.CreateCall(CGM.getIntrinsic(ID), Ops);
19403 }
19404
19405 case Hexagon::BI__builtin_HEXAGON_L2_loadrub_pci:
19406 case Hexagon::BI__builtin_HEXAGON_L2_loadrb_pci:
19407 case Hexagon::BI__builtin_HEXAGON_L2_loadruh_pci:
19408 case Hexagon::BI__builtin_HEXAGON_L2_loadrh_pci:
19409 case Hexagon::BI__builtin_HEXAGON_L2_loadri_pci:
19410 case Hexagon::BI__builtin_HEXAGON_L2_loadrd_pci:
19411 case Hexagon::BI__builtin_HEXAGON_L2_loadrub_pcr:
19412 case Hexagon::BI__builtin_HEXAGON_L2_loadrb_pcr:
19413 case Hexagon::BI__builtin_HEXAGON_L2_loadruh_pcr:
19414 case Hexagon::BI__builtin_HEXAGON_L2_loadrh_pcr:
19415 case Hexagon::BI__builtin_HEXAGON_L2_loadri_pcr:
19416 case Hexagon::BI__builtin_HEXAGON_L2_loadrd_pcr:
19417 return MakeCircOp(ID, /*IsLoad=*/true);
19418 case Hexagon::BI__builtin_HEXAGON_S2_storerb_pci:
19419 case Hexagon::BI__builtin_HEXAGON_S2_storerh_pci:
19420 case Hexagon::BI__builtin_HEXAGON_S2_storerf_pci:
19421 case Hexagon::BI__builtin_HEXAGON_S2_storeri_pci:
19422 case Hexagon::BI__builtin_HEXAGON_S2_storerd_pci:
19423 case Hexagon::BI__builtin_HEXAGON_S2_storerb_pcr:
19424 case Hexagon::BI__builtin_HEXAGON_S2_storerh_pcr:
19425 case Hexagon::BI__builtin_HEXAGON_S2_storerf_pcr:
19426 case Hexagon::BI__builtin_HEXAGON_S2_storeri_pcr:
19427 case Hexagon::BI__builtin_HEXAGON_S2_storerd_pcr:
19428 return MakeCircOp(ID, /*IsLoad=*/false);
19429 case Hexagon::BI__builtin_brev_ldub:
19430 return MakeBrevLd(Intrinsic::hexagon_L2_loadrub_pbr, Int8Ty);
19431 case Hexagon::BI__builtin_brev_ldb:
19432 return MakeBrevLd(Intrinsic::hexagon_L2_loadrb_pbr, Int8Ty);
19433 case Hexagon::BI__builtin_brev_lduh:
19434 return MakeBrevLd(Intrinsic::hexagon_L2_loadruh_pbr, Int16Ty);
19435 case Hexagon::BI__builtin_brev_ldh:
19436 return MakeBrevLd(Intrinsic::hexagon_L2_loadrh_pbr, Int16Ty);
19437 case Hexagon::BI__builtin_brev_ldw:
19438 return MakeBrevLd(Intrinsic::hexagon_L2_loadri_pbr, Int32Ty);
19439 case Hexagon::BI__builtin_brev_ldd:
19440 return MakeBrevLd(Intrinsic::hexagon_L2_loadrd_pbr, Int64Ty);
19441 } // switch
19442
19443 return nullptr;
19444 }
19445
19446 Value *CodeGenFunction::EmitRISCVBuiltinExpr(unsigned BuiltinID,
19447 const CallExpr *E,
19448 ReturnValueSlot ReturnValue) {
19449 SmallVector<Value *, 4> Ops;
19450 llvm::Type *ResultType = ConvertType(E->getType());
19451
19452 // Find out if any arguments are required to be integer constant expressions.
19453 unsigned ICEArguments = 0;
19454 ASTContext::GetBuiltinTypeError Error;
19455 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments);
19456 if (Error == ASTContext::GE_Missing_type) {
19457 // Vector intrinsics don't have a type string.
19458 assert(BuiltinID >= clang::RISCV::FirstRVVBuiltin &&
19459 BuiltinID <= clang::RISCV::LastRVVBuiltin);
19460 ICEArguments = 0;
19461 if (BuiltinID == RISCVVector::BI__builtin_rvv_vget_v ||
19462 BuiltinID == RISCVVector::BI__builtin_rvv_vset_v)
19463 ICEArguments = 1 << 1;
19464 } else {
19465 assert(Error == ASTContext::GE_None && "Unexpected error");
19466 }
19467
19468 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) {
19469 // If this is a normal argument, just emit it as a scalar.
19470 if ((ICEArguments & (1 << i)) == 0) {
19471 Ops.push_back(EmitScalarExpr(E->getArg(i)));
19472 continue;
19473 }
19474
19475 // If this is required to be a constant, constant fold it so that we know
19476 // that the generated intrinsic gets a ConstantInt.
19477 Ops.push_back(llvm::ConstantInt::get(
19478 getLLVMContext(), *E->getArg(i)->getIntegerConstantExpr(getContext())));
19479 }
19480
19481 Intrinsic::ID ID = Intrinsic::not_intrinsic;
19482 unsigned NF = 1;
19483 // The 0th bit simulates the `vta` of RVV
19484 // The 1st bit simulates the `vma` of RVV
19485 constexpr unsigned RVV_VTA = 0x1;
19486 constexpr unsigned RVV_VMA = 0x2;
19487 int PolicyAttrs = 0;
19488 bool IsMasked = false;
19489
19490 // Required for overloaded intrinsics.
19491 llvm::SmallVector<llvm::Type *, 2> IntrinsicTypes;
19492 switch (BuiltinID) {
19493 default: llvm_unreachable("unexpected builtin ID");
19494 case RISCV::BI__builtin_riscv_orc_b_32:
19495 case RISCV::BI__builtin_riscv_orc_b_64:
19496 case RISCV::BI__builtin_riscv_clz_32:
19497 case RISCV::BI__builtin_riscv_clz_64:
19498 case RISCV::BI__builtin_riscv_ctz_32:
19499 case RISCV::BI__builtin_riscv_ctz_64:
19500 case RISCV::BI__builtin_riscv_clmul:
19501 case RISCV::BI__builtin_riscv_clmulh:
19502 case RISCV::BI__builtin_riscv_clmulr:
19503 case RISCV::BI__builtin_riscv_xperm4:
19504 case RISCV::BI__builtin_riscv_xperm8:
19505 case RISCV::BI__builtin_riscv_brev8:
19506 case RISCV::BI__builtin_riscv_zip_32:
19507 case RISCV::BI__builtin_riscv_unzip_32: {
19508 switch (BuiltinID) {
19509 default: llvm_unreachable("unexpected builtin ID");
19510 // Zbb
19511 case RISCV::BI__builtin_riscv_orc_b_32:
19512 case RISCV::BI__builtin_riscv_orc_b_64:
19513 ID = Intrinsic::riscv_orc_b;
19514 break;
19515 case RISCV::BI__builtin_riscv_clz_32:
19516 case RISCV::BI__builtin_riscv_clz_64: {
19517 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, Ops[0]->getType());
19518 return Builder.CreateCall(F, {Ops[0], Builder.getInt1(false)});
19519 }
19520 case RISCV::BI__builtin_riscv_ctz_32:
19521 case RISCV::BI__builtin_riscv_ctz_64: {
19522 Function *F = CGM.getIntrinsic(Intrinsic::cttz, Ops[0]->getType());
19523 return Builder.CreateCall(F, {Ops[0], Builder.getInt1(false)});
19524 }
19525
19526 // Zbc
19527 case RISCV::BI__builtin_riscv_clmul:
19528 ID = Intrinsic::riscv_clmul;
19529 break;
19530 case RISCV::BI__builtin_riscv_clmulh:
19531 ID = Intrinsic::riscv_clmulh;
19532 break;
19533 case RISCV::BI__builtin_riscv_clmulr:
19534 ID = Intrinsic::riscv_clmulr;
19535 break;
19536
19537 // Zbkx
19538 case RISCV::BI__builtin_riscv_xperm8:
19539 ID = Intrinsic::riscv_xperm8;
19540 break;
19541 case RISCV::BI__builtin_riscv_xperm4:
19542 ID = Intrinsic::riscv_xperm4;
19543 break;
19544
19545 // Zbkb
19546 case RISCV::BI__builtin_riscv_brev8:
19547 ID = Intrinsic::riscv_brev8;
19548 break;
19549 case RISCV::BI__builtin_riscv_zip_32:
19550 ID = Intrinsic::riscv_zip;
19551 break;
19552 case RISCV::BI__builtin_riscv_unzip_32:
19553 ID = Intrinsic::riscv_unzip;
19554 break;
19555 }
19556
19557 IntrinsicTypes = {ResultType};
19558 break;
19559 }
19560
19561 // Zk builtins
19562
19563 // Zknd
19564 case RISCV::BI__builtin_riscv_aes32dsi_32:
19565 ID = Intrinsic::riscv_aes32dsi;
19566 break;
19567 case RISCV::BI__builtin_riscv_aes32dsmi_32:
19568 ID = Intrinsic::riscv_aes32dsmi;
19569 break;
19570 case RISCV::BI__builtin_riscv_aes64ds_64:
19571 ID = Intrinsic::riscv_aes64ds;
19572 break;
19573 case RISCV::BI__builtin_riscv_aes64dsm_64:
19574 ID = Intrinsic::riscv_aes64dsm;
19575 break;
19576 case RISCV::BI__builtin_riscv_aes64im_64:
19577 ID = Intrinsic::riscv_aes64im;
19578 break;
19579
19580 // Zkne
19581 case RISCV::BI__builtin_riscv_aes32esi_32:
19582 ID = Intrinsic::riscv_aes32esi;
19583 break;
19584 case RISCV::BI__builtin_riscv_aes32esmi_32:
19585 ID = Intrinsic::riscv_aes32esmi;
19586 break;
19587 case RISCV::BI__builtin_riscv_aes64es_64:
19588 ID = Intrinsic::riscv_aes64es;
19589 break;
19590 case RISCV::BI__builtin_riscv_aes64esm_64:
19591 ID = Intrinsic::riscv_aes64esm;
19592 break;
19593
19594 // Zknd & Zkne
19595 case RISCV::BI__builtin_riscv_aes64ks1i_64:
19596 ID = Intrinsic::riscv_aes64ks1i;
19597 break;
19598 case RISCV::BI__builtin_riscv_aes64ks2_64:
19599 ID = Intrinsic::riscv_aes64ks2;
19600 break;
19601
19602 // Zknh
19603 case RISCV::BI__builtin_riscv_sha256sig0:
19604 ID = Intrinsic::riscv_sha256sig0;
19605 IntrinsicTypes = {ResultType};
19606 break;
19607 case RISCV::BI__builtin_riscv_sha256sig1:
19608 ID = Intrinsic::riscv_sha256sig1;
19609 IntrinsicTypes = {ResultType};
19610 break;
19611 case RISCV::BI__builtin_riscv_sha256sum0:
19612 ID = Intrinsic::riscv_sha256sum0;
19613 IntrinsicTypes = {ResultType};
19614 break;
19615 case RISCV::BI__builtin_riscv_sha256sum1:
19616 ID = Intrinsic::riscv_sha256sum1;
19617 IntrinsicTypes = {ResultType};
19618 break;
19619 case RISCV::BI__builtin_riscv_sha512sig0_64:
19620 ID = Intrinsic::riscv_sha512sig0;
19621 break;
19622 case RISCV::BI__builtin_riscv_sha512sig0h_32:
19623 ID = Intrinsic::riscv_sha512sig0h;
19624 break;
19625 case RISCV::BI__builtin_riscv_sha512sig0l_32:
19626 ID = Intrinsic::riscv_sha512sig0l;
19627 break;
19628 case RISCV::BI__builtin_riscv_sha512sig1_64:
19629 ID = Intrinsic::riscv_sha512sig1;
19630 break;
19631 case RISCV::BI__builtin_riscv_sha512sig1h_32:
19632 ID = Intrinsic::riscv_sha512sig1h;
19633 break;
19634 case RISCV::BI__builtin_riscv_sha512sig1l_32:
19635 ID = Intrinsic::riscv_sha512sig1l;
19636 break;
19637 case RISCV::BI__builtin_riscv_sha512sum0_64:
19638 ID = Intrinsic::riscv_sha512sum0;
19639 break;
19640 case RISCV::BI__builtin_riscv_sha512sum0r_32:
19641 ID = Intrinsic::riscv_sha512sum0r;
19642 break;
19643 case RISCV::BI__builtin_riscv_sha512sum1_64:
19644 ID = Intrinsic::riscv_sha512sum1;
19645 break;
19646 case RISCV::BI__builtin_riscv_sha512sum1r_32:
19647 ID = Intrinsic::riscv_sha512sum1r;
19648 break;
19649
19650 // Zksed
19651 case RISCV::BI__builtin_riscv_sm4ks:
19652 ID = Intrinsic::riscv_sm4ks;
19653 IntrinsicTypes = {ResultType};
19654 break;
19655 case RISCV::BI__builtin_riscv_sm4ed:
19656 ID = Intrinsic::riscv_sm4ed;
19657 IntrinsicTypes = {ResultType};
19658 break;
19659
19660 // Zksh
19661 case RISCV::BI__builtin_riscv_sm3p0:
19662 ID = Intrinsic::riscv_sm3p0;
19663 IntrinsicTypes = {ResultType};
19664 break;
19665 case RISCV::BI__builtin_riscv_sm3p1:
19666 ID = Intrinsic::riscv_sm3p1;
19667 IntrinsicTypes = {ResultType};
19668 break;
19669
19670 // Vector builtins are handled from here.
19671 #include "clang/Basic/riscv_vector_builtin_cg.inc"
19672 }
19673
19674 assert(ID != Intrinsic::not_intrinsic);
19675
19676 llvm::Function *F = CGM.getIntrinsic(ID, IntrinsicTypes);
19677 return Builder.CreateCall(F, Ops, "");
19678 }
19679
19680 Value *CodeGenFunction::EmitLoongArchBuiltinExpr(unsigned BuiltinID,
19681 const CallExpr *E) {
19682 SmallVector<Value *, 4> Ops;
19683
19684 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++)
19685 Ops.push_back(EmitScalarExpr(E->getArg(i)));
19686
19687 Intrinsic::ID ID = Intrinsic::not_intrinsic;
19688
19689 switch (BuiltinID) {
19690 default:
19691 llvm_unreachable("unexpected builtin ID.");
19692 case LoongArch::BI__builtin_loongarch_cacop_d:
19693 ID = Intrinsic::loongarch_cacop_d;
19694 break;
19695 case LoongArch::BI__builtin_loongarch_cacop_w:
19696 ID = Intrinsic::loongarch_cacop_w;
19697 break;
19698 case LoongArch::BI__builtin_loongarch_dbar:
19699 ID = Intrinsic::loongarch_dbar;
19700 break;
19701 case LoongArch::BI__builtin_loongarch_break:
19702 ID = Intrinsic::loongarch_break;
19703 break;
19704 case LoongArch::BI__builtin_loongarch_ibar:
19705 ID = Intrinsic::loongarch_ibar;
19706 break;
19707 case LoongArch::BI__builtin_loongarch_movfcsr2gr:
19708 ID = Intrinsic::loongarch_movfcsr2gr;
19709 break;
19710 case LoongArch::BI__builtin_loongarch_movgr2fcsr:
19711 ID = Intrinsic::loongarch_movgr2fcsr;
19712 break;
19713 case LoongArch::BI__builtin_loongarch_syscall:
19714 ID = Intrinsic::loongarch_syscall;
19715 break;
19716 case LoongArch::BI__builtin_loongarch_crc_w_b_w:
19717 ID = Intrinsic::loongarch_crc_w_b_w;
19718 break;
19719 case LoongArch::BI__builtin_loongarch_crc_w_h_w:
19720 ID = Intrinsic::loongarch_crc_w_h_w;
19721 break;
19722 case LoongArch::BI__builtin_loongarch_crc_w_w_w:
19723 ID = Intrinsic::loongarch_crc_w_w_w;
19724 break;
19725 case LoongArch::BI__builtin_loongarch_crc_w_d_w:
19726 ID = Intrinsic::loongarch_crc_w_d_w;
19727 break;
19728 case LoongArch::BI__builtin_loongarch_crcc_w_b_w:
19729 ID = Intrinsic::loongarch_crcc_w_b_w;
19730 break;
19731 case LoongArch::BI__builtin_loongarch_crcc_w_h_w:
19732 ID = Intrinsic::loongarch_crcc_w_h_w;
19733 break;
19734 case LoongArch::BI__builtin_loongarch_crcc_w_w_w:
19735 ID = Intrinsic::loongarch_crcc_w_w_w;
19736 break;
19737 case LoongArch::BI__builtin_loongarch_crcc_w_d_w:
19738 ID = Intrinsic::loongarch_crcc_w_d_w;
19739 break;
19740 case LoongArch::BI__builtin_loongarch_csrrd_w:
19741 ID = Intrinsic::loongarch_csrrd_w;
19742 break;
19743 case LoongArch::BI__builtin_loongarch_csrwr_w:
19744 ID = Intrinsic::loongarch_csrwr_w;
19745 break;
19746 case LoongArch::BI__builtin_loongarch_csrxchg_w:
19747 ID = Intrinsic::loongarch_csrxchg_w;
19748 break;
19749 case LoongArch::BI__builtin_loongarch_csrrd_d:
19750 ID = Intrinsic::loongarch_csrrd_d;
19751 break;
19752 case LoongArch::BI__builtin_loongarch_csrwr_d:
19753 ID = Intrinsic::loongarch_csrwr_d;
19754 break;
19755 case LoongArch::BI__builtin_loongarch_csrxchg_d:
19756 ID = Intrinsic::loongarch_csrxchg_d;
19757 break;
19758 case LoongArch::BI__builtin_loongarch_iocsrrd_b:
19759 ID = Intrinsic::loongarch_iocsrrd_b;
19760 break;
19761 case LoongArch::BI__builtin_loongarch_iocsrrd_h:
19762 ID = Intrinsic::loongarch_iocsrrd_h;
19763 break;
19764 case LoongArch::BI__builtin_loongarch_iocsrrd_w:
19765 ID = Intrinsic::loongarch_iocsrrd_w;
19766 break;
19767 case LoongArch::BI__builtin_loongarch_iocsrrd_d:
19768 ID = Intrinsic::loongarch_iocsrrd_d;
19769 break;
19770 case LoongArch::BI__builtin_loongarch_iocsrwr_b:
19771 ID = Intrinsic::loongarch_iocsrwr_b;
19772 break;
19773 case LoongArch::BI__builtin_loongarch_iocsrwr_h:
19774 ID = Intrinsic::loongarch_iocsrwr_h;
19775 break;
19776 case LoongArch::BI__builtin_loongarch_iocsrwr_w:
19777 ID = Intrinsic::loongarch_iocsrwr_w;
19778 break;
19779 case LoongArch::BI__builtin_loongarch_iocsrwr_d:
19780 ID = Intrinsic::loongarch_iocsrwr_d;
19781 break;
19782 case LoongArch::BI__builtin_loongarch_cpucfg:
19783 ID = Intrinsic::loongarch_cpucfg;
19784 break;
19785 case LoongArch::BI__builtin_loongarch_asrtle_d:
19786 ID = Intrinsic::loongarch_asrtle_d;
19787 break;
19788 case LoongArch::BI__builtin_loongarch_asrtgt_d:
19789 ID = Intrinsic::loongarch_asrtgt_d;
19790 break;
19791 case LoongArch::BI__builtin_loongarch_lddir_d:
19792 ID = Intrinsic::loongarch_lddir_d;
19793 break;
19794 case LoongArch::BI__builtin_loongarch_ldpte_d:
19795 ID = Intrinsic::loongarch_ldpte_d;
19796 break;
19797 // TODO: Support more Intrinsics.
19798 }
19799
19800 assert(ID != Intrinsic::not_intrinsic);
19801
19802 llvm::Function *F = CGM.getIntrinsic(ID);
19803 return Builder.CreateCall(F, Ops);
19804 }