| 1 |
//===--- CGStmt.cpp - Emit LLVM Code from Statements ----------------------===// |
| 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 Stmt nodes as LLVM code. |
| 10 |
// |
| 11 |
//===----------------------------------------------------------------------===// |
| 12 |
|
| 13 |
#include "CGDebugInfo.h" |
| 14 |
#include "CGOpenMPRuntime.h" |
| 15 |
#include "CodeGenFunction.h" |
| 16 |
#include "CodeGenModule.h" |
| 17 |
#include "TargetInfo.h" |
| 18 |
#include "clang/AST/Attr.h" |
| 19 |
#include "clang/AST/Expr.h" |
| 20 |
#include "clang/AST/Stmt.h" |
| 21 |
#include "clang/AST/StmtVisitor.h" |
| 22 |
#include "clang/Basic/Builtins.h" |
| 23 |
#include "clang/Basic/DiagnosticSema.h" |
| 24 |
#include "clang/Basic/PrettyStackTrace.h" |
| 25 |
#include "clang/Basic/SourceManager.h" |
| 26 |
#include "clang/Basic/TargetInfo.h" |
| 27 |
#include "llvm/ADT/SmallSet.h" |
| 28 |
#include "llvm/ADT/StringExtras.h" |
| 29 |
#include "llvm/IR/Assumptions.h" |
| 30 |
#include "llvm/IR/DataLayout.h" |
| 31 |
#include "llvm/IR/InlineAsm.h" |
| 32 |
#include "llvm/IR/Intrinsics.h" |
| 33 |
#include "llvm/IR/MDBuilder.h" |
| 34 |
#include "llvm/Support/SaveAndRestore.h" |
| 35 |
#include <optional> |
| 36 |
|
| 37 |
using namespace clang; |
| 38 |
using namespace CodeGen; |
| 39 |
|
| 40 |
//===----------------------------------------------------------------------===// |
| 41 |
// Statement Emission |
| 42 |
//===----------------------------------------------------------------------===// |
| 43 |
|
| 44 |
void CodeGenFunction::EmitStopPoint(const Stmt *S) { |
| 45 |
if (CGDebugInfo *DI = getDebugInfo()) { |
| 46 |
SourceLocation Loc; |
| 47 |
Loc = S->getBeginLoc(); |
| 48 |
DI->EmitLocation(Builder, Loc); |
| 49 |
|
| 50 |
LastStopPoint = Loc; |
| 51 |
} |
| 52 |
} |
| 53 |
|
| 54 |
void CodeGenFunction::EmitStmt(const Stmt *S, ArrayRef<const Attr *> Attrs) { |
| 55 |
assert(S && "Null statement?"); |
| 56 |
PGO.setCurrentStmt(S); |
| 57 |
|
| 58 |
// These statements have their own debug info handling. |
| 59 |
if (EmitSimpleStmt(S, Attrs)) |
| 60 |
return; |
| 61 |
|
| 62 |
// Check if we are generating unreachable code. |
| 63 |
if (!HaveInsertPoint()) { |
| 64 |
// If so, and the statement doesn't contain a label, then we do not need to |
| 65 |
// generate actual code. This is safe because (1) the current point is |
| 66 |
// unreachable, so we don't need to execute the code, and (2) we've already |
| 67 |
// handled the statements which update internal data structures (like the |
| 68 |
// local variable map) which could be used by subsequent statements. |
| 69 |
if (!ContainsLabel(S)) { |
| 70 |
// Verify that any decl statements were handled as simple, they may be in |
| 71 |
// scope of subsequent reachable statements. |
| 72 |
assert(!isa<DeclStmt>(*S) && "Unexpected DeclStmt!"); |
| 73 |
return; |
| 74 |
} |
| 75 |
|
| 76 |
// Otherwise, make a new block to hold the code. |
| 77 |
EnsureInsertPoint(); |
| 78 |
} |
| 79 |
|
| 80 |
// Generate a stoppoint if we are emitting debug info. |
| 81 |
EmitStopPoint(S); |
| 82 |
|
| 83 |
// Ignore all OpenMP directives except for simd if OpenMP with Simd is |
| 84 |
// enabled. |
| 85 |
if (getLangOpts().OpenMP && getLangOpts().OpenMPSimd) { |
| 86 |
if (const auto *D = dyn_cast<OMPExecutableDirective>(S)) { |
| 87 |
EmitSimpleOMPExecutableDirective(*D); |
| 88 |
return; |
| 89 |
} |
| 90 |
} |
| 91 |
|
| 92 |
switch (S->getStmtClass()) { |
| 93 |
case Stmt::NoStmtClass: |
| 94 |
case Stmt::CXXCatchStmtClass: |
| 95 |
case Stmt::SEHExceptStmtClass: |
| 96 |
case Stmt::SEHFinallyStmtClass: |
| 97 |
case Stmt::MSDependentExistsStmtClass: |
| 98 |
llvm_unreachable("invalid statement class to emit generically"); |
| 99 |
case Stmt::NullStmtClass: |
| 100 |
case Stmt::CompoundStmtClass: |
| 101 |
case Stmt::DeclStmtClass: |
| 102 |
case Stmt::LabelStmtClass: |
| 103 |
case Stmt::AttributedStmtClass: |
| 104 |
case Stmt::GotoStmtClass: |
| 105 |
case Stmt::BreakStmtClass: |
| 106 |
case Stmt::ContinueStmtClass: |
| 107 |
case Stmt::DefaultStmtClass: |
| 108 |
case Stmt::CaseStmtClass: |
| 109 |
case Stmt::SEHLeaveStmtClass: |
| 110 |
llvm_unreachable("should have emitted these statements as simple"); |
| 111 |
|
| 112 |
#define STMT(Type, Base) |
| 113 |
#define ABSTRACT_STMT(Op) |
| 114 |
#define EXPR(Type, Base) \ |
| 115 |
case Stmt::Type##Class: |
| 116 |
#include "clang/AST/StmtNodes.inc" |
| 117 |
{ |
| 118 |
// Remember the block we came in on. |
| 119 |
llvm::BasicBlock *incoming = Builder.GetInsertBlock(); |
| 120 |
assert(incoming && "expression emission must have an insertion point"); |
| 121 |
|
| 122 |
EmitIgnoredExpr(cast<Expr>(S)); |
| 123 |
|
| 124 |
llvm::BasicBlock *outgoing = Builder.GetInsertBlock(); |
| 125 |
assert(outgoing && "expression emission cleared block!"); |
| 126 |
|
| 127 |
// The expression emitters assume (reasonably!) that the insertion |
| 128 |
// point is always set. To maintain that, the call-emission code |
| 129 |
// for noreturn functions has to enter a new block with no |
| 130 |
// predecessors. We want to kill that block and mark the current |
| 131 |
// insertion point unreachable in the common case of a call like |
| 132 |
// "exit();". Since expression emission doesn't otherwise create |
| 133 |
// blocks with no predecessors, we can just test for that. |
| 134 |
// However, we must be careful not to do this to our incoming |
| 135 |
// block, because *statement* emission does sometimes create |
| 136 |
// reachable blocks which will have no predecessors until later in |
| 137 |
// the function. This occurs with, e.g., labels that are not |
| 138 |
// reachable by fallthrough. |
| 139 |
if (incoming != outgoing && outgoing->use_empty()) { |
| 140 |
outgoing->eraseFromParent(); |
| 141 |
Builder.ClearInsertionPoint(); |
| 142 |
} |
| 143 |
break; |
| 144 |
} |
| 145 |
|
| 146 |
case Stmt::IndirectGotoStmtClass: |
| 147 |
EmitIndirectGotoStmt(cast<IndirectGotoStmt>(*S)); break; |
| 148 |
|
| 149 |
case Stmt::IfStmtClass: EmitIfStmt(cast<IfStmt>(*S)); break; |
| 150 |
case Stmt::WhileStmtClass: EmitWhileStmt(cast<WhileStmt>(*S), Attrs); break; |
| 151 |
case Stmt::DoStmtClass: EmitDoStmt(cast<DoStmt>(*S), Attrs); break; |
| 152 |
case Stmt::ForStmtClass: EmitForStmt(cast<ForStmt>(*S), Attrs); break; |
| 153 |
|
| 154 |
case Stmt::ReturnStmtClass: EmitReturnStmt(cast<ReturnStmt>(*S)); break; |
| 155 |
|
| 156 |
case Stmt::SwitchStmtClass: EmitSwitchStmt(cast<SwitchStmt>(*S)); break; |
| 157 |
case Stmt::GCCAsmStmtClass: // Intentional fall-through. |
| 158 |
case Stmt::MSAsmStmtClass: EmitAsmStmt(cast<AsmStmt>(*S)); break; |
| 159 |
case Stmt::CoroutineBodyStmtClass: |
| 160 |
EmitCoroutineBody(cast<CoroutineBodyStmt>(*S)); |
| 161 |
break; |
| 162 |
case Stmt::CoreturnStmtClass: |
| 163 |
EmitCoreturnStmt(cast<CoreturnStmt>(*S)); |
| 164 |
break; |
| 165 |
case Stmt::CapturedStmtClass: { |
| 166 |
const CapturedStmt *CS = cast<CapturedStmt>(S); |
| 167 |
EmitCapturedStmt(*CS, CS->getCapturedRegionKind()); |
| 168 |
} |
| 169 |
break; |
| 170 |
case Stmt::ObjCAtTryStmtClass: |
| 171 |
EmitObjCAtTryStmt(cast<ObjCAtTryStmt>(*S)); |
| 172 |
break; |
| 173 |
case Stmt::ObjCAtCatchStmtClass: |
| 174 |
llvm_unreachable( |
| 175 |
"@catch statements should be handled by EmitObjCAtTryStmt"); |
| 176 |
case Stmt::ObjCAtFinallyStmtClass: |
| 177 |
llvm_unreachable( |
| 178 |
"@finally statements should be handled by EmitObjCAtTryStmt"); |
| 179 |
case Stmt::ObjCAtThrowStmtClass: |
| 180 |
EmitObjCAtThrowStmt(cast<ObjCAtThrowStmt>(*S)); |
| 181 |
break; |
| 182 |
case Stmt::ObjCAtSynchronizedStmtClass: |
| 183 |
EmitObjCAtSynchronizedStmt(cast<ObjCAtSynchronizedStmt>(*S)); |
| 184 |
break; |
| 185 |
case Stmt::ObjCForCollectionStmtClass: |
| 186 |
EmitObjCForCollectionStmt(cast<ObjCForCollectionStmt>(*S)); |
| 187 |
break; |
| 188 |
case Stmt::ObjCAutoreleasePoolStmtClass: |
| 189 |
EmitObjCAutoreleasePoolStmt(cast<ObjCAutoreleasePoolStmt>(*S)); |
| 190 |
break; |
| 191 |
|
| 192 |
case Stmt::CXXTryStmtClass: |
| 193 |
EmitCXXTryStmt(cast<CXXTryStmt>(*S)); |
| 194 |
break; |
| 195 |
case Stmt::CXXForRangeStmtClass: |
| 196 |
EmitCXXForRangeStmt(cast<CXXForRangeStmt>(*S), Attrs); |
| 197 |
break; |
| 198 |
case Stmt::SEHTryStmtClass: |
| 199 |
EmitSEHTryStmt(cast<SEHTryStmt>(*S)); |
| 200 |
break; |
| 201 |
case Stmt::OMPMetaDirectiveClass: |
| 202 |
EmitOMPMetaDirective(cast<OMPMetaDirective>(*S)); |
| 203 |
break; |
| 204 |
case Stmt::OMPCanonicalLoopClass: |
| 205 |
EmitOMPCanonicalLoop(cast<OMPCanonicalLoop>(S)); |
| 206 |
break; |
| 207 |
case Stmt::OMPParallelDirectiveClass: |
| 208 |
EmitOMPParallelDirective(cast<OMPParallelDirective>(*S)); |
| 209 |
break; |
| 210 |
case Stmt::OMPSimdDirectiveClass: |
| 211 |
EmitOMPSimdDirective(cast<OMPSimdDirective>(*S)); |
| 212 |
break; |
| 213 |
case Stmt::OMPTileDirectiveClass: |
| 214 |
EmitOMPTileDirective(cast<OMPTileDirective>(*S)); |
| 215 |
break; |
| 216 |
case Stmt::OMPUnrollDirectiveClass: |
| 217 |
EmitOMPUnrollDirective(cast<OMPUnrollDirective>(*S)); |
| 218 |
break; |
| 219 |
case Stmt::OMPForDirectiveClass: |
| 220 |
EmitOMPForDirective(cast<OMPForDirective>(*S)); |
| 221 |
break; |
| 222 |
case Stmt::OMPForSimdDirectiveClass: |
| 223 |
EmitOMPForSimdDirective(cast<OMPForSimdDirective>(*S)); |
| 224 |
break; |
| 225 |
case Stmt::OMPSectionsDirectiveClass: |
| 226 |
EmitOMPSectionsDirective(cast<OMPSectionsDirective>(*S)); |
| 227 |
break; |
| 228 |
case Stmt::OMPSectionDirectiveClass: |
| 229 |
EmitOMPSectionDirective(cast<OMPSectionDirective>(*S)); |
| 230 |
break; |
| 231 |
case Stmt::OMPSingleDirectiveClass: |
| 232 |
EmitOMPSingleDirective(cast<OMPSingleDirective>(*S)); |
| 233 |
break; |
| 234 |
case Stmt::OMPMasterDirectiveClass: |
| 235 |
EmitOMPMasterDirective(cast<OMPMasterDirective>(*S)); |
| 236 |
break; |
| 237 |
case Stmt::OMPCriticalDirectiveClass: |
| 238 |
EmitOMPCriticalDirective(cast<OMPCriticalDirective>(*S)); |
| 239 |
break; |
| 240 |
case Stmt::OMPParallelForDirectiveClass: |
| 241 |
EmitOMPParallelForDirective(cast<OMPParallelForDirective>(*S)); |
| 242 |
break; |
| 243 |
case Stmt::OMPParallelForSimdDirectiveClass: |
| 244 |
EmitOMPParallelForSimdDirective(cast<OMPParallelForSimdDirective>(*S)); |
| 245 |
break; |
| 246 |
case Stmt::OMPParallelMasterDirectiveClass: |
| 247 |
EmitOMPParallelMasterDirective(cast<OMPParallelMasterDirective>(*S)); |
| 248 |
break; |
| 249 |
case Stmt::OMPParallelSectionsDirectiveClass: |
| 250 |
EmitOMPParallelSectionsDirective(cast<OMPParallelSectionsDirective>(*S)); |
| 251 |
break; |
| 252 |
case Stmt::OMPTaskDirectiveClass: |
| 253 |
EmitOMPTaskDirective(cast<OMPTaskDirective>(*S)); |
| 254 |
break; |
| 255 |
case Stmt::OMPTaskyieldDirectiveClass: |
| 256 |
EmitOMPTaskyieldDirective(cast<OMPTaskyieldDirective>(*S)); |
| 257 |
break; |
| 258 |
case Stmt::OMPErrorDirectiveClass: |
| 259 |
EmitOMPErrorDirective(cast<OMPErrorDirective>(*S)); |
| 260 |
break; |
| 261 |
case Stmt::OMPBarrierDirectiveClass: |
| 262 |
EmitOMPBarrierDirective(cast<OMPBarrierDirective>(*S)); |
| 263 |
break; |
| 264 |
case Stmt::OMPTaskwaitDirectiveClass: |
| 265 |
EmitOMPTaskwaitDirective(cast<OMPTaskwaitDirective>(*S)); |
| 266 |
break; |
| 267 |
case Stmt::OMPTaskgroupDirectiveClass: |
| 268 |
EmitOMPTaskgroupDirective(cast<OMPTaskgroupDirective>(*S)); |
| 269 |
break; |
| 270 |
case Stmt::OMPFlushDirectiveClass: |
| 271 |
EmitOMPFlushDirective(cast<OMPFlushDirective>(*S)); |
| 272 |
break; |
| 273 |
case Stmt::OMPDepobjDirectiveClass: |
| 274 |
EmitOMPDepobjDirective(cast<OMPDepobjDirective>(*S)); |
| 275 |
break; |
| 276 |
case Stmt::OMPScanDirectiveClass: |
| 277 |
EmitOMPScanDirective(cast<OMPScanDirective>(*S)); |
| 278 |
break; |
| 279 |
case Stmt::OMPOrderedDirectiveClass: |
| 280 |
EmitOMPOrderedDirective(cast<OMPOrderedDirective>(*S)); |
| 281 |
break; |
| 282 |
case Stmt::OMPAtomicDirectiveClass: |
| 283 |
EmitOMPAtomicDirective(cast<OMPAtomicDirective>(*S)); |
| 284 |
break; |
| 285 |
case Stmt::OMPTargetDirectiveClass: |
| 286 |
EmitOMPTargetDirective(cast<OMPTargetDirective>(*S)); |
| 287 |
break; |
| 288 |
case Stmt::OMPTeamsDirectiveClass: |
| 289 |
EmitOMPTeamsDirective(cast<OMPTeamsDirective>(*S)); |
| 290 |
break; |
| 291 |
case Stmt::OMPCancellationPointDirectiveClass: |
| 292 |
EmitOMPCancellationPointDirective(cast<OMPCancellationPointDirective>(*S)); |
| 293 |
break; |
| 294 |
case Stmt::OMPCancelDirectiveClass: |
| 295 |
EmitOMPCancelDirective(cast<OMPCancelDirective>(*S)); |
| 296 |
break; |
| 297 |
case Stmt::OMPTargetDataDirectiveClass: |
| 298 |
EmitOMPTargetDataDirective(cast<OMPTargetDataDirective>(*S)); |
| 299 |
break; |
| 300 |
case Stmt::OMPTargetEnterDataDirectiveClass: |
| 301 |
EmitOMPTargetEnterDataDirective(cast<OMPTargetEnterDataDirective>(*S)); |
| 302 |
break; |
| 303 |
case Stmt::OMPTargetExitDataDirectiveClass: |
| 304 |
EmitOMPTargetExitDataDirective(cast<OMPTargetExitDataDirective>(*S)); |
| 305 |
break; |
| 306 |
case Stmt::OMPTargetParallelDirectiveClass: |
| 307 |
EmitOMPTargetParallelDirective(cast<OMPTargetParallelDirective>(*S)); |
| 308 |
break; |
| 309 |
case Stmt::OMPTargetParallelForDirectiveClass: |
| 310 |
EmitOMPTargetParallelForDirective(cast<OMPTargetParallelForDirective>(*S)); |
| 311 |
break; |
| 312 |
case Stmt::OMPTaskLoopDirectiveClass: |
| 313 |
EmitOMPTaskLoopDirective(cast<OMPTaskLoopDirective>(*S)); |
| 314 |
break; |
| 315 |
case Stmt::OMPTaskLoopSimdDirectiveClass: |
| 316 |
EmitOMPTaskLoopSimdDirective(cast<OMPTaskLoopSimdDirective>(*S)); |
| 317 |
break; |
| 318 |
case Stmt::OMPMasterTaskLoopDirectiveClass: |
| 319 |
EmitOMPMasterTaskLoopDirective(cast<OMPMasterTaskLoopDirective>(*S)); |
| 320 |
break; |
| 321 |
case Stmt::OMPMaskedTaskLoopDirectiveClass: |
| 322 |
llvm_unreachable("masked taskloop directive not supported yet."); |
| 323 |
break; |
| 324 |
case Stmt::OMPMasterTaskLoopSimdDirectiveClass: |
| 325 |
EmitOMPMasterTaskLoopSimdDirective( |
| 326 |
cast<OMPMasterTaskLoopSimdDirective>(*S)); |
| 327 |
break; |
| 328 |
case Stmt::OMPMaskedTaskLoopSimdDirectiveClass: |
| 329 |
llvm_unreachable("masked taskloop simd directive not supported yet."); |
| 330 |
break; |
| 331 |
case Stmt::OMPParallelMasterTaskLoopDirectiveClass: |
| 332 |
EmitOMPParallelMasterTaskLoopDirective( |
| 333 |
cast<OMPParallelMasterTaskLoopDirective>(*S)); |
| 334 |
break; |
| 335 |
case Stmt::OMPParallelMaskedTaskLoopDirectiveClass: |
| 336 |
llvm_unreachable("parallel masked taskloop directive not supported yet."); |
| 337 |
break; |
| 338 |
case Stmt::OMPParallelMasterTaskLoopSimdDirectiveClass: |
| 339 |
EmitOMPParallelMasterTaskLoopSimdDirective( |
| 340 |
cast<OMPParallelMasterTaskLoopSimdDirective>(*S)); |
| 341 |
break; |
| 342 |
case Stmt::OMPParallelMaskedTaskLoopSimdDirectiveClass: |
| 343 |
llvm_unreachable( |
| 344 |
"parallel masked taskloop simd directive not supported yet."); |
| 345 |
break; |
| 346 |
case Stmt::OMPDistributeDirectiveClass: |
| 347 |
EmitOMPDistributeDirective(cast<OMPDistributeDirective>(*S)); |
| 348 |
break; |
| 349 |
case Stmt::OMPTargetUpdateDirectiveClass: |
| 350 |
EmitOMPTargetUpdateDirective(cast<OMPTargetUpdateDirective>(*S)); |
| 351 |
break; |
| 352 |
case Stmt::OMPDistributeParallelForDirectiveClass: |
| 353 |
EmitOMPDistributeParallelForDirective( |
| 354 |
cast<OMPDistributeParallelForDirective>(*S)); |
| 355 |
break; |
| 356 |
case Stmt::OMPDistributeParallelForSimdDirectiveClass: |
| 357 |
EmitOMPDistributeParallelForSimdDirective( |
| 358 |
cast<OMPDistributeParallelForSimdDirective>(*S)); |
| 359 |
break; |
| 360 |
case Stmt::OMPDistributeSimdDirectiveClass: |
| 361 |
EmitOMPDistributeSimdDirective(cast<OMPDistributeSimdDirective>(*S)); |
| 362 |
break; |
| 363 |
case Stmt::OMPTargetParallelForSimdDirectiveClass: |
| 364 |
EmitOMPTargetParallelForSimdDirective( |
| 365 |
cast<OMPTargetParallelForSimdDirective>(*S)); |
| 366 |
break; |
| 367 |
case Stmt::OMPTargetSimdDirectiveClass: |
| 368 |
EmitOMPTargetSimdDirective(cast<OMPTargetSimdDirective>(*S)); |
| 369 |
break; |
| 370 |
case Stmt::OMPTeamsDistributeDirectiveClass: |
| 371 |
EmitOMPTeamsDistributeDirective(cast<OMPTeamsDistributeDirective>(*S)); |
| 372 |
break; |
| 373 |
case Stmt::OMPTeamsDistributeSimdDirectiveClass: |
| 374 |
EmitOMPTeamsDistributeSimdDirective( |
| 375 |
cast<OMPTeamsDistributeSimdDirective>(*S)); |
| 376 |
break; |
| 377 |
case Stmt::OMPTeamsDistributeParallelForSimdDirectiveClass: |
| 378 |
EmitOMPTeamsDistributeParallelForSimdDirective( |
| 379 |
cast<OMPTeamsDistributeParallelForSimdDirective>(*S)); |
| 380 |
break; |
| 381 |
case Stmt::OMPTeamsDistributeParallelForDirectiveClass: |
| 382 |
EmitOMPTeamsDistributeParallelForDirective( |
| 383 |
cast<OMPTeamsDistributeParallelForDirective>(*S)); |
| 384 |
break; |
| 385 |
case Stmt::OMPTargetTeamsDirectiveClass: |
| 386 |
EmitOMPTargetTeamsDirective(cast<OMPTargetTeamsDirective>(*S)); |
| 387 |
break; |
| 388 |
case Stmt::OMPTargetTeamsDistributeDirectiveClass: |
| 389 |
EmitOMPTargetTeamsDistributeDirective( |
| 390 |
cast<OMPTargetTeamsDistributeDirective>(*S)); |
| 391 |
break; |
| 392 |
case Stmt::OMPTargetTeamsDistributeParallelForDirectiveClass: |
| 393 |
EmitOMPTargetTeamsDistributeParallelForDirective( |
| 394 |
cast<OMPTargetTeamsDistributeParallelForDirective>(*S)); |
| 395 |
break; |
| 396 |
case Stmt::OMPTargetTeamsDistributeParallelForSimdDirectiveClass: |
| 397 |
EmitOMPTargetTeamsDistributeParallelForSimdDirective( |
| 398 |
cast<OMPTargetTeamsDistributeParallelForSimdDirective>(*S)); |
| 399 |
break; |
| 400 |
case Stmt::OMPTargetTeamsDistributeSimdDirectiveClass: |
| 401 |
EmitOMPTargetTeamsDistributeSimdDirective( |
| 402 |
cast<OMPTargetTeamsDistributeSimdDirective>(*S)); |
| 403 |
break; |
| 404 |
case Stmt::OMPInteropDirectiveClass: |
| 405 |
EmitOMPInteropDirective(cast<OMPInteropDirective>(*S)); |
| 406 |
break; |
| 407 |
case Stmt::OMPDispatchDirectiveClass: |
| 408 |
llvm_unreachable("Dispatch directive not supported yet."); |
| 409 |
break; |
| 410 |
case Stmt::OMPMaskedDirectiveClass: |
| 411 |
EmitOMPMaskedDirective(cast<OMPMaskedDirective>(*S)); |
| 412 |
break; |
| 413 |
case Stmt::OMPGenericLoopDirectiveClass: |
| 414 |
EmitOMPGenericLoopDirective(cast<OMPGenericLoopDirective>(*S)); |
| 415 |
break; |
| 416 |
case Stmt::OMPTeamsGenericLoopDirectiveClass: |
| 417 |
llvm_unreachable("teams loop directive not supported yet."); |
| 418 |
break; |
| 419 |
case Stmt::OMPTargetTeamsGenericLoopDirectiveClass: |
| 420 |
llvm_unreachable("target teams loop directive not supported yet."); |
| 421 |
break; |
| 422 |
case Stmt::OMPParallelGenericLoopDirectiveClass: |
| 423 |
llvm_unreachable("parallel loop directive not supported yet."); |
| 424 |
break; |
| 425 |
case Stmt::OMPTargetParallelGenericLoopDirectiveClass: |
| 426 |
llvm_unreachable("target parallel loop directive not supported yet."); |
| 427 |
break; |
| 428 |
case Stmt::OMPParallelMaskedDirectiveClass: |
| 429 |
llvm_unreachable("parallel masked directive not supported yet."); |
| 430 |
break; |
| 431 |
} |
| 432 |
} |
| 433 |
|
| 434 |
bool CodeGenFunction::EmitSimpleStmt(const Stmt *S, |
| 435 |
ArrayRef<const Attr *> Attrs) { |
| 436 |
switch (S->getStmtClass()) { |
| 437 |
default: |
| 438 |
return false; |
| 439 |
case Stmt::NullStmtClass: |
| 440 |
break; |
| 441 |
case Stmt::CompoundStmtClass: |
| 442 |
EmitCompoundStmt(cast<CompoundStmt>(*S)); |
| 443 |
break; |
| 444 |
case Stmt::DeclStmtClass: |
| 445 |
EmitDeclStmt(cast<DeclStmt>(*S)); |
| 446 |
break; |
| 447 |
case Stmt::LabelStmtClass: |
| 448 |
EmitLabelStmt(cast<LabelStmt>(*S)); |
| 449 |
break; |
| 450 |
case Stmt::AttributedStmtClass: |
| 451 |
EmitAttributedStmt(cast<AttributedStmt>(*S)); |
| 452 |
break; |
| 453 |
case Stmt::GotoStmtClass: |
| 454 |
EmitGotoStmt(cast<GotoStmt>(*S)); |
| 455 |
break; |
| 456 |
case Stmt::BreakStmtClass: |
| 457 |
EmitBreakStmt(cast<BreakStmt>(*S)); |
| 458 |
break; |
| 459 |
case Stmt::ContinueStmtClass: |
| 460 |
EmitContinueStmt(cast<ContinueStmt>(*S)); |
| 461 |
break; |
| 462 |
case Stmt::DefaultStmtClass: |
| 463 |
EmitDefaultStmt(cast<DefaultStmt>(*S), Attrs); |
| 464 |
break; |
| 465 |
case Stmt::CaseStmtClass: |
| 466 |
EmitCaseStmt(cast<CaseStmt>(*S), Attrs); |
| 467 |
break; |
| 468 |
case Stmt::SEHLeaveStmtClass: |
| 469 |
EmitSEHLeaveStmt(cast<SEHLeaveStmt>(*S)); |
| 470 |
break; |
| 471 |
} |
| 472 |
return true; |
| 473 |
} |
| 474 |
|
| 475 |
/// EmitCompoundStmt - Emit a compound statement {..} node. If GetLast is true, |
| 476 |
/// this captures the expression result of the last sub-statement and returns it |
| 477 |
/// (for use by the statement expression extension). |
| 478 |
Address CodeGenFunction::EmitCompoundStmt(const CompoundStmt &S, bool GetLast, |
| 479 |
AggValueSlot AggSlot) { |
| 480 |
PrettyStackTraceLoc CrashInfo(getContext().getSourceManager(),S.getLBracLoc(), |
| 481 |
"LLVM IR generation of compound statement ('{}')"); |
| 482 |
|
| 483 |
// Keep track of the current cleanup stack depth, including debug scopes. |
| 484 |
LexicalScope Scope(*this, S.getSourceRange()); |
| 485 |
|
| 486 |
return EmitCompoundStmtWithoutScope(S, GetLast, AggSlot); |
| 487 |
} |
| 488 |
|
| 489 |
Address |
| 490 |
CodeGenFunction::EmitCompoundStmtWithoutScope(const CompoundStmt &S, |
| 491 |
bool GetLast, |
| 492 |
AggValueSlot AggSlot) { |
| 493 |
|
| 494 |
const Stmt *ExprResult = S.getStmtExprResult(); |
| 495 |
assert((!GetLast || (GetLast && ExprResult)) && |
| 496 |
"If GetLast is true then the CompoundStmt must have a StmtExprResult"); |
| 497 |
|
| 498 |
Address RetAlloca = Address::invalid(); |
| 499 |
|
| 500 |
for (auto *CurStmt : S.body()) { |
| 501 |
if (GetLast && ExprResult == CurStmt) { |
| 502 |
// We have to special case labels here. They are statements, but when put |
| 503 |
// at the end of a statement expression, they yield the value of their |
| 504 |
// subexpression. Handle this by walking through all labels we encounter, |
| 505 |
// emitting them before we evaluate the subexpr. |
| 506 |
// Similar issues arise for attributed statements. |
| 507 |
while (!isa<Expr>(ExprResult)) { |
| 508 |
if (const auto *LS = dyn_cast<LabelStmt>(ExprResult)) { |
| 509 |
EmitLabel(LS->getDecl()); |
| 510 |
ExprResult = LS->getSubStmt(); |
| 511 |
} else if (const auto *AS = dyn_cast<AttributedStmt>(ExprResult)) { |
| 512 |
// FIXME: Update this if we ever have attributes that affect the |
| 513 |
// semantics of an expression. |
| 514 |
ExprResult = AS->getSubStmt(); |
| 515 |
} else { |
| 516 |
llvm_unreachable("unknown value statement"); |
| 517 |
} |
| 518 |
} |
| 519 |
|
| 520 |
EnsureInsertPoint(); |
| 521 |
|
| 522 |
const Expr *E = cast<Expr>(ExprResult); |
| 523 |
QualType ExprTy = E->getType(); |
| 524 |
if (hasAggregateEvaluationKind(ExprTy)) { |
| 525 |
EmitAggExpr(E, AggSlot); |
| 526 |
} else { |
| 527 |
// We can't return an RValue here because there might be cleanups at |
| 528 |
// the end of the StmtExpr. Because of that, we have to emit the result |
| 529 |
// here into a temporary alloca. |
| 530 |
RetAlloca = CreateMemTemp(ExprTy); |
| 531 |
EmitAnyExprToMem(E, RetAlloca, Qualifiers(), |
| 532 |
/*IsInit*/ false); |
| 533 |
} |
| 534 |
} else { |
| 535 |
EmitStmt(CurStmt); |
| 536 |
} |
| 537 |
} |
| 538 |
|
| 539 |
return RetAlloca; |
| 540 |
} |
| 541 |
|
| 542 |
void CodeGenFunction::SimplifyForwardingBlocks(llvm::BasicBlock *BB) { |
| 543 |
llvm::BranchInst *BI = dyn_cast<llvm::BranchInst>(BB->getTerminator()); |
| 544 |
|
| 545 |
// If there is a cleanup stack, then we it isn't worth trying to |
| 546 |
// simplify this block (we would need to remove it from the scope map |
| 547 |
// and cleanup entry). |
| 548 |
if (!EHStack.empty()) |
| 549 |
return; |
| 550 |
|
| 551 |
// Can only simplify direct branches. |
| 552 |
if (!BI || !BI->isUnconditional()) |
| 553 |
return; |
| 554 |
|
| 555 |
// Can only simplify empty blocks. |
| 556 |
if (BI->getIterator() != BB->begin()) |
| 557 |
return; |
| 558 |
|
| 559 |
BB->replaceAllUsesWith(BI->getSuccessor(0)); |
| 560 |
BI->eraseFromParent(); |
| 561 |
BB->eraseFromParent(); |
| 562 |
} |
| 563 |
|
| 564 |
void CodeGenFunction::EmitBlock(llvm::BasicBlock *BB, bool IsFinished) { |
| 565 |
llvm::BasicBlock *CurBB = Builder.GetInsertBlock(); |
| 566 |
|
| 567 |
// Fall out of the current block (if necessary). |
| 568 |
EmitBranch(BB); |
| 569 |
|
| 570 |
if (IsFinished && BB->use_empty()) { |
| 571 |
delete BB; |
| 572 |
return; |
| 573 |
} |
| 574 |
|
| 575 |
// Place the block after the current block, if possible, or else at |
| 576 |
// the end of the function. |
| 577 |
if (CurBB && CurBB->getParent()) |
| 578 |
CurFn->insert(std::next(CurBB->getIterator()), BB); |
| 579 |
else |
| 580 |
CurFn->insert(CurFn->end(), BB); |
| 581 |
Builder.SetInsertPoint(BB); |
| 582 |
} |
| 583 |
|
| 584 |
void CodeGenFunction::EmitBranch(llvm::BasicBlock *Target) { |
| 585 |
// Emit a branch from the current block to the target one if this |
| 586 |
// was a real block. If this was just a fall-through block after a |
| 587 |
// terminator, don't emit it. |
| 588 |
llvm::BasicBlock *CurBB = Builder.GetInsertBlock(); |
| 589 |
|
| 590 |
if (!CurBB || CurBB->getTerminator()) { |
| 591 |
// If there is no insert point or the previous block is already |
| 592 |
// terminated, don't touch it. |
| 593 |
} else { |
| 594 |
// Otherwise, create a fall-through branch. |
| 595 |
Builder.CreateBr(Target); |
| 596 |
} |
| 597 |
|
| 598 |
Builder.ClearInsertionPoint(); |
| 599 |
} |
| 600 |
|
| 601 |
void CodeGenFunction::EmitBlockAfterUses(llvm::BasicBlock *block) { |
| 602 |
bool inserted = false; |
| 603 |
for (llvm::User *u : block->users()) { |
| 604 |
if (llvm::Instruction *insn = dyn_cast<llvm::Instruction>(u)) { |
| 605 |
CurFn->insert(std::next(insn->getParent()->getIterator()), block); |
| 606 |
inserted = true; |
| 607 |
break; |
| 608 |
} |
| 609 |
} |
| 610 |
|
| 611 |
if (!inserted) |
| 612 |
CurFn->insert(CurFn->end(), block); |
| 613 |
|
| 614 |
Builder.SetInsertPoint(block); |
| 615 |
} |
| 616 |
|
| 617 |
CodeGenFunction::JumpDest |
| 618 |
CodeGenFunction::getJumpDestForLabel(const LabelDecl *D) { |
| 619 |
JumpDest &Dest = LabelMap[D]; |
| 620 |
if (Dest.isValid()) return Dest; |
| 621 |
|
| 622 |
// Create, but don't insert, the new block. |
| 623 |
Dest = JumpDest(createBasicBlock(D->getName()), |
| 624 |
EHScopeStack::stable_iterator::invalid(), |
| 625 |
NextCleanupDestIndex++); |
| 626 |
return Dest; |
| 627 |
} |
| 628 |
|
| 629 |
void CodeGenFunction::EmitLabel(const LabelDecl *D) { |
| 630 |
// Add this label to the current lexical scope if we're within any |
| 631 |
// normal cleanups. Jumps "in" to this label --- when permitted by |
| 632 |
// the language --- may need to be routed around such cleanups. |
| 633 |
if (EHStack.hasNormalCleanups() && CurLexicalScope) |
| 634 |
CurLexicalScope->addLabel(D); |
| 635 |
|
| 636 |
JumpDest &Dest = LabelMap[D]; |
| 637 |
|
| 638 |
// If we didn't need a forward reference to this label, just go |
| 639 |
// ahead and create a destination at the current scope. |
| 640 |
if (!Dest.isValid()) { |
| 641 |
Dest = getJumpDestInCurrentScope(D->getName()); |
| 642 |
|
| 643 |
// Otherwise, we need to give this label a target depth and remove |
| 644 |
// it from the branch-fixups list. |
| 645 |
} else { |
| 646 |
assert(!Dest.getScopeDepth().isValid() && "already emitted label!"); |
| 647 |
Dest.setScopeDepth(EHStack.stable_begin()); |
| 648 |
ResolveBranchFixups(Dest.getBlock()); |
| 649 |
} |
| 650 |
|
| 651 |
EmitBlock(Dest.getBlock()); |
| 652 |
|
| 653 |
// Emit debug info for labels. |
| 654 |
if (CGDebugInfo *DI = getDebugInfo()) { |
| 655 |
if (CGM.getCodeGenOpts().hasReducedDebugInfo()) { |
| 656 |
DI->setLocation(D->getLocation()); |
| 657 |
DI->EmitLabel(D, Builder); |
| 658 |
} |
| 659 |
} |
| 660 |
|
| 661 |
incrementProfileCounter(D->getStmt()); |
| 662 |
} |
| 663 |
|
| 664 |
/// Change the cleanup scope of the labels in this lexical scope to |
| 665 |
/// match the scope of the enclosing context. |
| 666 |
void CodeGenFunction::LexicalScope::rescopeLabels() { |
| 667 |
assert(!Labels.empty()); |
| 668 |
EHScopeStack::stable_iterator innermostScope |
| 669 |
= CGF.EHStack.getInnermostNormalCleanup(); |
| 670 |
|
| 671 |
// Change the scope depth of all the labels. |
| 672 |
for (SmallVectorImpl<const LabelDecl*>::const_iterator |
| 673 |
i = Labels.begin(), e = Labels.end(); i != e; ++i) { |
| 674 |
assert(CGF.LabelMap.count(*i)); |
| 675 |
JumpDest &dest = CGF.LabelMap.find(*i)->second; |
| 676 |
assert(dest.getScopeDepth().isValid()); |
| 677 |
assert(innermostScope.encloses(dest.getScopeDepth())); |
| 678 |
dest.setScopeDepth(innermostScope); |
| 679 |
} |
| 680 |
|
| 681 |
// Reparent the labels if the new scope also has cleanups. |
| 682 |
if (innermostScope != EHScopeStack::stable_end() && ParentScope) { |
| 683 |
ParentScope->Labels.append(Labels.begin(), Labels.end()); |
| 684 |
} |
| 685 |
} |
| 686 |
|
| 687 |
|
| 688 |
void CodeGenFunction::EmitLabelStmt(const LabelStmt &S) { |
| 689 |
EmitLabel(S.getDecl()); |
| 690 |
|
| 691 |
// IsEHa - emit eha.scope.begin if it's a side entry of a scope |
| 692 |
if (getLangOpts().EHAsynch && S.isSideEntry()) |
| 693 |
EmitSehCppScopeBegin(); |
| 694 |
|
| 695 |
EmitStmt(S.getSubStmt()); |
| 696 |
} |
| 697 |
|
| 698 |
void CodeGenFunction::EmitAttributedStmt(const AttributedStmt &S) { |
| 699 |
bool nomerge = false; |
| 700 |
bool noinline = false; |
| 701 |
bool alwaysinline = false; |
| 702 |
const CallExpr *musttail = nullptr; |
| 703 |
|
| 704 |
for (const auto *A : S.getAttrs()) { |
| 705 |
switch (A->getKind()) { |
| 706 |
default: |
| 707 |
break; |
| 708 |
case attr::NoMerge: |
| 709 |
nomerge = true; |
| 710 |
break; |
| 711 |
case attr::NoInline: |
| 712 |
noinline = true; |
| 713 |
break; |
| 714 |
case attr::AlwaysInline: |
| 715 |
alwaysinline = true; |
| 716 |
break; |
| 717 |
case attr::MustTail: |
| 718 |
const Stmt *Sub = S.getSubStmt(); |
| 719 |
const ReturnStmt *R = cast<ReturnStmt>(Sub); |
| 720 |
musttail = cast<CallExpr>(R->getRetValue()->IgnoreParens()); |
| 721 |
break; |
| 722 |
} |
| 723 |
} |
| 724 |
SaveAndRestore save_nomerge(InNoMergeAttributedStmt, nomerge); |
| 725 |
SaveAndRestore save_noinline(InNoInlineAttributedStmt, noinline); |
| 726 |
SaveAndRestore save_alwaysinline(InAlwaysInlineAttributedStmt, alwaysinline); |
| 727 |
SaveAndRestore save_musttail(MustTailCall, musttail); |
| 728 |
EmitStmt(S.getSubStmt(), S.getAttrs()); |
| 729 |
} |
| 730 |
|
| 731 |
void CodeGenFunction::EmitGotoStmt(const GotoStmt &S) { |
| 732 |
// If this code is reachable then emit a stop point (if generating |
| 733 |
// debug info). We have to do this ourselves because we are on the |
| 734 |
// "simple" statement path. |
| 735 |
if (HaveInsertPoint()) |
| 736 |
EmitStopPoint(&S); |
| 737 |
|
| 738 |
EmitBranchThroughCleanup(getJumpDestForLabel(S.getLabel())); |
| 739 |
} |
| 740 |
|
| 741 |
|
| 742 |
void CodeGenFunction::EmitIndirectGotoStmt(const IndirectGotoStmt &S) { |
| 743 |
if (const LabelDecl *Target = S.getConstantTarget()) { |
| 744 |
EmitBranchThroughCleanup(getJumpDestForLabel(Target)); |
| 745 |
return; |
| 746 |
} |
| 747 |
|
| 748 |
// Ensure that we have an i8* for our PHI node. |
| 749 |
llvm::Value *V = Builder.CreateBitCast(EmitScalarExpr(S.getTarget()), |
| 750 |
Int8PtrTy, "addr"); |
| 751 |
llvm::BasicBlock *CurBB = Builder.GetInsertBlock(); |
| 752 |
|
| 753 |
// Get the basic block for the indirect goto. |
| 754 |
llvm::BasicBlock *IndGotoBB = GetIndirectGotoBlock(); |
| 755 |
|
| 756 |
// The first instruction in the block has to be the PHI for the switch dest, |
| 757 |
// add an entry for this branch. |
| 758 |
cast<llvm::PHINode>(IndGotoBB->begin())->addIncoming(V, CurBB); |
| 759 |
|
| 760 |
EmitBranch(IndGotoBB); |
| 761 |
} |
| 762 |
|
| 763 |
void CodeGenFunction::EmitIfStmt(const IfStmt &S) { |
| 764 |
// The else branch of a consteval if statement is always the only branch that |
| 765 |
// can be runtime evaluated. |
| 766 |
if (S.isConsteval()) { |
| 767 |
const Stmt *Executed = S.isNegatedConsteval() ? S.getThen() : S.getElse(); |
| 768 |
if (Executed) { |
| 769 |
RunCleanupsScope ExecutedScope(*this); |
| 770 |
EmitStmt(Executed); |
| 771 |
} |
| 772 |
return; |
| 773 |
} |
| 774 |
|
| 775 |
// C99 6.8.4.1: The first substatement is executed if the expression compares |
| 776 |
// unequal to 0. The condition must be a scalar type. |
| 777 |
LexicalScope ConditionScope(*this, S.getCond()->getSourceRange()); |
| 778 |
|
| 779 |
if (S.getInit()) |
| 780 |
EmitStmt(S.getInit()); |
| 781 |
|
| 782 |
if (S.getConditionVariable()) |
| 783 |
EmitDecl(*S.getConditionVariable()); |
| 784 |
|
| 785 |
// If the condition constant folds and can be elided, try to avoid emitting |
| 786 |
// the condition and the dead arm of the if/else. |
| 787 |
bool CondConstant; |
| 788 |
if (ConstantFoldsToSimpleInteger(S.getCond(), CondConstant, |
| 789 |
S.isConstexpr())) { |
| 790 |
// Figure out which block (then or else) is executed. |
| 791 |
const Stmt *Executed = S.getThen(); |
| 792 |
const Stmt *Skipped = S.getElse(); |
| 793 |
if (!CondConstant) // Condition false? |
| 794 |
std::swap(Executed, Skipped); |
| 795 |
|
| 796 |
// If the skipped block has no labels in it, just emit the executed block. |
| 797 |
// This avoids emitting dead code and simplifies the CFG substantially. |
| 798 |
if (S.isConstexpr() || !ContainsLabel(Skipped)) { |
| 799 |
if (CondConstant) |
| 800 |
incrementProfileCounter(&S); |
| 801 |
if (Executed) { |
| 802 |
RunCleanupsScope ExecutedScope(*this); |
| 803 |
EmitStmt(Executed); |
| 804 |
} |
| 805 |
return; |
| 806 |
} |
| 807 |
} |
| 808 |
|
| 809 |
// Otherwise, the condition did not fold, or we couldn't elide it. Just emit |
| 810 |
// the conditional branch. |
| 811 |
llvm::BasicBlock *ThenBlock = createBasicBlock("if.then"); |
| 812 |
llvm::BasicBlock *ContBlock = createBasicBlock("if.end"); |
| 813 |
llvm::BasicBlock *ElseBlock = ContBlock; |
| 814 |
if (S.getElse()) |
| 815 |
ElseBlock = createBasicBlock("if.else"); |
| 816 |
|
| 817 |
// Prefer the PGO based weights over the likelihood attribute. |
| 818 |
// When the build isn't optimized the metadata isn't used, so don't generate |
| 819 |
// it. |
| 820 |
// Also, differentiate between disabled PGO and a never executed branch with |
| 821 |
// PGO. Assuming PGO is in use: |
| 822 |
// - we want to ignore the [[likely]] attribute if the branch is never |
| 823 |
// executed, |
| 824 |
// - assuming the profile is poor, preserving the attribute may still be |
| 825 |
// beneficial. |
| 826 |
// As an approximation, preserve the attribute only if both the branch and the |
| 827 |
// parent context were not executed. |
| 828 |
Stmt::Likelihood LH = Stmt::LH_None; |
| 829 |
uint64_t ThenCount = getProfileCount(S.getThen()); |
| 830 |
if (!ThenCount && !getCurrentProfileCount() && |
| 831 |
CGM.getCodeGenOpts().OptimizationLevel) |
| 832 |
LH = Stmt::getLikelihood(S.getThen(), S.getElse()); |
| 833 |
EmitBranchOnBoolExpr(S.getCond(), ThenBlock, ElseBlock, ThenCount, LH); |
| 834 |
|
| 835 |
// Emit the 'then' code. |
| 836 |
EmitBlock(ThenBlock); |
| 837 |
incrementProfileCounter(&S); |
| 838 |
{ |
| 839 |
RunCleanupsScope ThenScope(*this); |
| 840 |
EmitStmt(S.getThen()); |
| 841 |
} |
| 842 |
EmitBranch(ContBlock); |
| 843 |
|
| 844 |
// Emit the 'else' code if present. |
| 845 |
if (const Stmt *Else = S.getElse()) { |
| 846 |
{ |
| 847 |
// There is no need to emit line number for an unconditional branch. |
| 848 |
auto NL = ApplyDebugLocation::CreateEmpty(*this); |
| 849 |
EmitBlock(ElseBlock); |
| 850 |
} |
| 851 |
{ |
| 852 |
RunCleanupsScope ElseScope(*this); |
| 853 |
EmitStmt(Else); |
| 854 |
} |
| 855 |
{ |
| 856 |
// There is no need to emit line number for an unconditional branch. |
| 857 |
auto NL = ApplyDebugLocation::CreateEmpty(*this); |
| 858 |
EmitBranch(ContBlock); |
| 859 |
} |
| 860 |
} |
| 861 |
|
| 862 |
// Emit the continuation block for code after the if. |
| 863 |
EmitBlock(ContBlock, true); |
| 864 |
} |
| 865 |
|
| 866 |
void CodeGenFunction::EmitWhileStmt(const WhileStmt &S, |
| 867 |
ArrayRef<const Attr *> WhileAttrs) { |
| 868 |
// Emit the header for the loop, which will also become |
| 869 |
// the continue target. |
| 870 |
JumpDest LoopHeader = getJumpDestInCurrentScope("while.cond"); |
| 871 |
EmitBlock(LoopHeader.getBlock()); |
| 872 |
|
| 873 |
// Create an exit block for when the condition fails, which will |
| 874 |
// also become the break target. |
| 875 |
JumpDest LoopExit = getJumpDestInCurrentScope("while.end"); |
| 876 |
|
| 877 |
// Store the blocks to use for break and continue. |
| 878 |
BreakContinueStack.push_back(BreakContinue(LoopExit, LoopHeader)); |
| 879 |
|
| 880 |
// C++ [stmt.while]p2: |
| 881 |
// When the condition of a while statement is a declaration, the |
| 882 |
// scope of the variable that is declared extends from its point |
| 883 |
// of declaration (3.3.2) to the end of the while statement. |
| 884 |
// [...] |
| 885 |
// The object created in a condition is destroyed and created |
| 886 |
// with each iteration of the loop. |
| 887 |
RunCleanupsScope ConditionScope(*this); |
| 888 |
|
| 889 |
if (S.getConditionVariable()) |
| 890 |
EmitDecl(*S.getConditionVariable()); |
| 891 |
|
| 892 |
// Evaluate the conditional in the while header. C99 6.8.5.1: The |
| 893 |
// evaluation of the controlling expression takes place before each |
| 894 |
// execution of the loop body. |
| 895 |
llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond()); |
| 896 |
|
| 897 |
// while(1) is common, avoid extra exit blocks. Be sure |
| 898 |
// to correctly handle break/continue though. |
| 899 |
llvm::ConstantInt *C = dyn_cast<llvm::ConstantInt>(BoolCondVal); |
| 900 |
bool CondIsConstInt = C != nullptr; |
| 901 |
bool EmitBoolCondBranch = !CondIsConstInt || !C->isOne(); |
| 902 |
const SourceRange &R = S.getSourceRange(); |
| 903 |
LoopStack.push(LoopHeader.getBlock(), CGM.getContext(), CGM.getCodeGenOpts(), |
| 904 |
WhileAttrs, SourceLocToDebugLoc(R.getBegin()), |
| 905 |
SourceLocToDebugLoc(R.getEnd()), |
| 906 |
checkIfLoopMustProgress(CondIsConstInt)); |
| 907 |
|
| 908 |
// As long as the condition is true, go to the loop body. |
| 909 |
llvm::BasicBlock *LoopBody = createBasicBlock("while.body"); |
| 910 |
if (EmitBoolCondBranch) { |
| 911 |
llvm::BasicBlock *ExitBlock = LoopExit.getBlock(); |
| 912 |
if (ConditionScope.requiresCleanups()) |
| 913 |
ExitBlock = createBasicBlock("while.exit"); |
| 914 |
llvm::MDNode *Weights = |
| 915 |
createProfileWeightsForLoop(S.getCond(), getProfileCount(S.getBody())); |
| 916 |
if (!Weights && CGM.getCodeGenOpts().OptimizationLevel) |
| 917 |
BoolCondVal = emitCondLikelihoodViaExpectIntrinsic( |
| 918 |
BoolCondVal, Stmt::getLikelihood(S.getBody())); |
| 919 |
Builder.CreateCondBr(BoolCondVal, LoopBody, ExitBlock, Weights); |
| 920 |
|
| 921 |
if (ExitBlock != LoopExit.getBlock()) { |
| 922 |
EmitBlock(ExitBlock); |
| 923 |
EmitBranchThroughCleanup(LoopExit); |
| 924 |
} |
| 925 |
} else if (const Attr *A = Stmt::getLikelihoodAttr(S.getBody())) { |
| 926 |
CGM.getDiags().Report(A->getLocation(), |
| 927 |
diag::warn_attribute_has_no_effect_on_infinite_loop) |
| 928 |
<< A << A->getRange(); |
| 929 |
CGM.getDiags().Report( |
| 930 |
S.getWhileLoc(), |
| 931 |
diag::note_attribute_has_no_effect_on_infinite_loop_here) |
| 932 |
<< SourceRange(S.getWhileLoc(), S.getRParenLoc()); |
| 933 |
} |
| 934 |
|
| 935 |
// Emit the loop body. We have to emit this in a cleanup scope |
| 936 |
// because it might be a singleton DeclStmt. |
| 937 |
{ |
| 938 |
RunCleanupsScope BodyScope(*this); |
| 939 |
EmitBlock(LoopBody); |
| 940 |
incrementProfileCounter(&S); |
| 941 |
EmitStmt(S.getBody()); |
| 942 |
} |
| 943 |
|
| 944 |
BreakContinueStack.pop_back(); |
| 945 |
|
| 946 |
// Immediately force cleanup. |
| 947 |
ConditionScope.ForceCleanup(); |
| 948 |
|
| 949 |
EmitStopPoint(&S); |
| 950 |
// Branch to the loop header again. |
| 951 |
EmitBranch(LoopHeader.getBlock()); |
| 952 |
|
| 953 |
LoopStack.pop(); |
| 954 |
|
| 955 |
// Emit the exit block. |
| 956 |
EmitBlock(LoopExit.getBlock(), true); |
| 957 |
|
| 958 |
// The LoopHeader typically is just a branch if we skipped emitting |
| 959 |
// a branch, try to erase it. |
| 960 |
if (!EmitBoolCondBranch) |
| 961 |
SimplifyForwardingBlocks(LoopHeader.getBlock()); |
| 962 |
} |
| 963 |
|
| 964 |
void CodeGenFunction::EmitDoStmt(const DoStmt &S, |
| 965 |
ArrayRef<const Attr *> DoAttrs) { |
| 966 |
JumpDest LoopExit = getJumpDestInCurrentScope("do.end"); |
| 967 |
JumpDest LoopCond = getJumpDestInCurrentScope("do.cond"); |
| 968 |
|
| 969 |
uint64_t ParentCount = getCurrentProfileCount(); |
| 970 |
|
| 971 |
// Store the blocks to use for break and continue. |
| 972 |
BreakContinueStack.push_back(BreakContinue(LoopExit, LoopCond)); |
| 973 |
|
| 974 |
// Emit the body of the loop. |
| 975 |
llvm::BasicBlock *LoopBody = createBasicBlock("do.body"); |
| 976 |
|
| 977 |
EmitBlockWithFallThrough(LoopBody, &S); |
| 978 |
{ |
| 979 |
RunCleanupsScope BodyScope(*this); |
| 980 |
EmitStmt(S.getBody()); |
| 981 |
} |
| 982 |
|
| 983 |
EmitBlock(LoopCond.getBlock()); |
| 984 |
|
| 985 |
// C99 6.8.5.2: "The evaluation of the controlling expression takes place |
| 986 |
// after each execution of the loop body." |
| 987 |
|
| 988 |
// Evaluate the conditional in the while header. |
| 989 |
// C99 6.8.5p2/p4: The first substatement is executed if the expression |
| 990 |
// compares unequal to 0. The condition must be a scalar type. |
| 991 |
llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond()); |
| 992 |
|
| 993 |
BreakContinueStack.pop_back(); |
| 994 |
|
| 995 |
// "do {} while (0)" is common in macros, avoid extra blocks. Be sure |
| 996 |
// to correctly handle break/continue though. |
| 997 |
llvm::ConstantInt *C = dyn_cast<llvm::ConstantInt>(BoolCondVal); |
| 998 |
bool CondIsConstInt = C; |
| 999 |
bool EmitBoolCondBranch = !C || !C->isZero(); |
| 1000 |
|
| 1001 |
const SourceRange &R = S.getSourceRange(); |
| 1002 |
LoopStack.push(LoopBody, CGM.getContext(), CGM.getCodeGenOpts(), DoAttrs, |
| 1003 |
SourceLocToDebugLoc(R.getBegin()), |
| 1004 |
SourceLocToDebugLoc(R.getEnd()), |
| 1005 |
checkIfLoopMustProgress(CondIsConstInt)); |
| 1006 |
|
| 1007 |
// As long as the condition is true, iterate the loop. |
| 1008 |
if (EmitBoolCondBranch) { |
| 1009 |
uint64_t BackedgeCount = getProfileCount(S.getBody()) - ParentCount; |
| 1010 |
Builder.CreateCondBr( |
| 1011 |
BoolCondVal, LoopBody, LoopExit.getBlock(), |
| 1012 |
createProfileWeightsForLoop(S.getCond(), BackedgeCount)); |
| 1013 |
} |
| 1014 |
|
| 1015 |
LoopStack.pop(); |
| 1016 |
|
| 1017 |
// Emit the exit block. |
| 1018 |
EmitBlock(LoopExit.getBlock()); |
| 1019 |
|
| 1020 |
// The DoCond block typically is just a branch if we skipped |
| 1021 |
// emitting a branch, try to erase it. |
| 1022 |
if (!EmitBoolCondBranch) |
| 1023 |
SimplifyForwardingBlocks(LoopCond.getBlock()); |
| 1024 |
} |
| 1025 |
|
| 1026 |
void CodeGenFunction::EmitForStmt(const ForStmt &S, |
| 1027 |
ArrayRef<const Attr *> ForAttrs) { |
| 1028 |
JumpDest LoopExit = getJumpDestInCurrentScope("for.end"); |
| 1029 |
|
| 1030 |
LexicalScope ForScope(*this, S.getSourceRange()); |
| 1031 |
|
| 1032 |
// Evaluate the first part before the loop. |
| 1033 |
if (S.getInit()) |
| 1034 |
EmitStmt(S.getInit()); |
| 1035 |
|
| 1036 |
// Start the loop with a block that tests the condition. |
| 1037 |
// If there's an increment, the continue scope will be overwritten |
| 1038 |
// later. |
| 1039 |
JumpDest CondDest = getJumpDestInCurrentScope("for.cond"); |
| 1040 |
llvm::BasicBlock *CondBlock = CondDest.getBlock(); |
| 1041 |
EmitBlock(CondBlock); |
| 1042 |
|
| 1043 |
Expr::EvalResult Result; |
| 1044 |
bool CondIsConstInt = |
| 1045 |
!S.getCond() || S.getCond()->EvaluateAsInt(Result, getContext()); |
| 1046 |
|
| 1047 |
const SourceRange &R = S.getSourceRange(); |
| 1048 |
LoopStack.push(CondBlock, CGM.getContext(), CGM.getCodeGenOpts(), ForAttrs, |
| 1049 |
SourceLocToDebugLoc(R.getBegin()), |
| 1050 |
SourceLocToDebugLoc(R.getEnd()), |
| 1051 |
checkIfLoopMustProgress(CondIsConstInt)); |
| 1052 |
|
| 1053 |
// Create a cleanup scope for the condition variable cleanups. |
| 1054 |
LexicalScope ConditionScope(*this, S.getSourceRange()); |
| 1055 |
|
| 1056 |
// If the for loop doesn't have an increment we can just use the condition as |
| 1057 |
// the continue block. Otherwise, if there is no condition variable, we can |
| 1058 |
// form the continue block now. If there is a condition variable, we can't |
| 1059 |
// form the continue block until after we've emitted the condition, because |
| 1060 |
// the condition is in scope in the increment, but Sema's jump diagnostics |
| 1061 |
// ensure that there are no continues from the condition variable that jump |
| 1062 |
// to the loop increment. |
| 1063 |
JumpDest Continue; |
| 1064 |
if (!S.getInc()) |
| 1065 |
Continue = CondDest; |
| 1066 |
else if (!S.getConditionVariable()) |
| 1067 |
Continue = getJumpDestInCurrentScope("for.inc"); |
| 1068 |
BreakContinueStack.push_back(BreakContinue(LoopExit, Continue)); |
| 1069 |
|
| 1070 |
if (S.getCond()) { |
| 1071 |
// If the for statement has a condition scope, emit the local variable |
| 1072 |
// declaration. |
| 1073 |
if (S.getConditionVariable()) { |
| 1074 |
EmitDecl(*S.getConditionVariable()); |
| 1075 |
|
| 1076 |
// We have entered the condition variable's scope, so we're now able to |
| 1077 |
// jump to the continue block. |
| 1078 |
Continue = S.getInc() ? getJumpDestInCurrentScope("for.inc") : CondDest; |
| 1079 |
BreakContinueStack.back().ContinueBlock = Continue; |
| 1080 |
} |
| 1081 |
|
| 1082 |
llvm::BasicBlock *ExitBlock = LoopExit.getBlock(); |
| 1083 |
// If there are any cleanups between here and the loop-exit scope, |
| 1084 |
// create a block to stage a loop exit along. |
| 1085 |
if (ForScope.requiresCleanups()) |
| 1086 |
ExitBlock = createBasicBlock("for.cond.cleanup"); |
| 1087 |
|
| 1088 |
// As long as the condition is true, iterate the loop. |
| 1089 |
llvm::BasicBlock *ForBody = createBasicBlock("for.body"); |
| 1090 |
|
| 1091 |
// C99 6.8.5p2/p4: The first substatement is executed if the expression |
| 1092 |
// compares unequal to 0. The condition must be a scalar type. |
| 1093 |
llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond()); |
| 1094 |
llvm::MDNode *Weights = |
| 1095 |
createProfileWeightsForLoop(S.getCond(), getProfileCount(S.getBody())); |
| 1096 |
if (!Weights && CGM.getCodeGenOpts().OptimizationLevel) |
| 1097 |
BoolCondVal = emitCondLikelihoodViaExpectIntrinsic( |
| 1098 |
BoolCondVal, Stmt::getLikelihood(S.getBody())); |
| 1099 |
|
| 1100 |
Builder.CreateCondBr(BoolCondVal, ForBody, ExitBlock, Weights); |
| 1101 |
|
| 1102 |
if (ExitBlock != LoopExit.getBlock()) { |
| 1103 |
EmitBlock(ExitBlock); |
| 1104 |
EmitBranchThroughCleanup(LoopExit); |
| 1105 |
} |
| 1106 |
|
| 1107 |
EmitBlock(ForBody); |
| 1108 |
} else { |
| 1109 |
// Treat it as a non-zero constant. Don't even create a new block for the |
| 1110 |
// body, just fall into it. |
| 1111 |
} |
| 1112 |
incrementProfileCounter(&S); |
| 1113 |
|
| 1114 |
{ |
| 1115 |
// Create a separate cleanup scope for the body, in case it is not |
| 1116 |
// a compound statement. |
| 1117 |
RunCleanupsScope BodyScope(*this); |
| 1118 |
EmitStmt(S.getBody()); |
| 1119 |
} |
| 1120 |
|
| 1121 |
// If there is an increment, emit it next. |
| 1122 |
if (S.getInc()) { |
| 1123 |
EmitBlock(Continue.getBlock()); |
| 1124 |
EmitStmt(S.getInc()); |
| 1125 |
} |
| 1126 |
|
| 1127 |
BreakContinueStack.pop_back(); |
| 1128 |
|
| 1129 |
ConditionScope.ForceCleanup(); |
| 1130 |
|
| 1131 |
EmitStopPoint(&S); |
| 1132 |
EmitBranch(CondBlock); |
| 1133 |
|
| 1134 |
ForScope.ForceCleanup(); |
| 1135 |
|
| 1136 |
LoopStack.pop(); |
| 1137 |
|
| 1138 |
// Emit the fall-through block. |
| 1139 |
EmitBlock(LoopExit.getBlock(), true); |
| 1140 |
} |
| 1141 |
|
| 1142 |
void |
| 1143 |
CodeGenFunction::EmitCXXForRangeStmt(const CXXForRangeStmt &S, |
| 1144 |
ArrayRef<const Attr *> ForAttrs) { |
| 1145 |
JumpDest LoopExit = getJumpDestInCurrentScope("for.end"); |
| 1146 |
|
| 1147 |
LexicalScope ForScope(*this, S.getSourceRange()); |
| 1148 |
|
| 1149 |
// Evaluate the first pieces before the loop. |
| 1150 |
if (S.getInit()) |
| 1151 |
EmitStmt(S.getInit()); |
| 1152 |
EmitStmt(S.getRangeStmt()); |
| 1153 |
EmitStmt(S.getBeginStmt()); |
| 1154 |
EmitStmt(S.getEndStmt()); |
| 1155 |
|
| 1156 |
// Start the loop with a block that tests the condition. |
| 1157 |
// If there's an increment, the continue scope will be overwritten |
| 1158 |
// later. |
| 1159 |
llvm::BasicBlock *CondBlock = createBasicBlock("for.cond"); |
| 1160 |
EmitBlock(CondBlock); |
| 1161 |
|
| 1162 |
const SourceRange &R = S.getSourceRange(); |
| 1163 |
LoopStack.push(CondBlock, CGM.getContext(), CGM.getCodeGenOpts(), ForAttrs, |
| 1164 |
SourceLocToDebugLoc(R.getBegin()), |
| 1165 |
SourceLocToDebugLoc(R.getEnd())); |
| 1166 |
|
| 1167 |
// If there are any cleanups between here and the loop-exit scope, |
| 1168 |
// create a block to stage a loop exit along. |
| 1169 |
llvm::BasicBlock *ExitBlock = LoopExit.getBlock(); |
| 1170 |
if (ForScope.requiresCleanups()) |
| 1171 |
ExitBlock = createBasicBlock("for.cond.cleanup"); |
| 1172 |
|
| 1173 |
// The loop body, consisting of the specified body and the loop variable. |
| 1174 |
llvm::BasicBlock *ForBody = createBasicBlock("for.body"); |
| 1175 |
|
| 1176 |
// The body is executed if the expression, contextually converted |
| 1177 |
// to bool, is true. |
| 1178 |
llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond()); |
| 1179 |
llvm::MDNode *Weights = |
| 1180 |
createProfileWeightsForLoop(S.getCond(), getProfileCount(S.getBody())); |
| 1181 |
if (!Weights && CGM.getCodeGenOpts().OptimizationLevel) |
| 1182 |
BoolCondVal = emitCondLikelihoodViaExpectIntrinsic( |
| 1183 |
BoolCondVal, Stmt::getLikelihood(S.getBody())); |
| 1184 |
Builder.CreateCondBr(BoolCondVal, ForBody, ExitBlock, Weights); |
| 1185 |
|
| 1186 |
if (ExitBlock != LoopExit.getBlock()) { |
| 1187 |
EmitBlock(ExitBlock); |
| 1188 |
EmitBranchThroughCleanup(LoopExit); |
| 1189 |
} |
| 1190 |
|
| 1191 |
EmitBlock(ForBody); |
| 1192 |
incrementProfileCounter(&S); |
| 1193 |
|
| 1194 |
// Create a block for the increment. In case of a 'continue', we jump there. |
| 1195 |
JumpDest Continue = getJumpDestInCurrentScope("for.inc"); |
| 1196 |
|
| 1197 |
// Store the blocks to use for break and continue. |
| 1198 |
BreakContinueStack.push_back(BreakContinue(LoopExit, Continue)); |
| 1199 |
|
| 1200 |
{ |
| 1201 |
// Create a separate cleanup scope for the loop variable and body. |
| 1202 |
LexicalScope BodyScope(*this, S.getSourceRange()); |
| 1203 |
EmitStmt(S.getLoopVarStmt()); |
| 1204 |
EmitStmt(S.getBody()); |
| 1205 |
} |
| 1206 |
|
| 1207 |
EmitStopPoint(&S); |
| 1208 |
// If there is an increment, emit it next. |
| 1209 |
EmitBlock(Continue.getBlock()); |
| 1210 |
EmitStmt(S.getInc()); |
| 1211 |
|
| 1212 |
BreakContinueStack.pop_back(); |
| 1213 |
|
| 1214 |
EmitBranch(CondBlock); |
| 1215 |
|
| 1216 |
ForScope.ForceCleanup(); |
| 1217 |
|
| 1218 |
LoopStack.pop(); |
| 1219 |
|
| 1220 |
// Emit the fall-through block. |
| 1221 |
EmitBlock(LoopExit.getBlock(), true); |
| 1222 |
} |
| 1223 |
|
| 1224 |
void CodeGenFunction::EmitReturnOfRValue(RValue RV, QualType Ty) { |
| 1225 |
if (RV.isScalar()) { |
| 1226 |
Builder.CreateStore(RV.getScalarVal(), ReturnValue); |
| 1227 |
} else if (RV.isAggregate()) { |
| 1228 |
LValue Dest = MakeAddrLValue(ReturnValue, Ty); |
| 1229 |
LValue Src = MakeAddrLValue(RV.getAggregateAddress(), Ty); |
| 1230 |
EmitAggregateCopy(Dest, Src, Ty, getOverlapForReturnValue()); |
| 1231 |
} else { |
| 1232 |
EmitStoreOfComplex(RV.getComplexVal(), MakeAddrLValue(ReturnValue, Ty), |
| 1233 |
/*init*/ true); |
| 1234 |
} |
| 1235 |
EmitBranchThroughCleanup(ReturnBlock); |
| 1236 |
} |
| 1237 |
|
| 1238 |
namespace { |
| 1239 |
// RAII struct used to save and restore a return statment's result expression. |
| 1240 |
struct SaveRetExprRAII { |
| 1241 |
SaveRetExprRAII(const Expr *RetExpr, CodeGenFunction &CGF) |
| 1242 |
: OldRetExpr(CGF.RetExpr), CGF(CGF) { |
| 1243 |
CGF.RetExpr = RetExpr; |
| 1244 |
} |
| 1245 |
~SaveRetExprRAII() { CGF.RetExpr = OldRetExpr; } |
| 1246 |
const Expr *OldRetExpr; |
| 1247 |
CodeGenFunction &CGF; |
| 1248 |
}; |
| 1249 |
} // namespace |
| 1250 |
|
| 1251 |
/// If we have 'return f(...);', where both caller and callee are SwiftAsync, |
| 1252 |
/// codegen it as 'tail call ...; ret void;'. |
| 1253 |
static void makeTailCallIfSwiftAsync(const CallExpr *CE, CGBuilderTy &Builder, |
| 1254 |
const CGFunctionInfo *CurFnInfo) { |
| 1255 |
auto calleeQualType = CE->getCallee()->getType(); |
| 1256 |
const FunctionType *calleeType = nullptr; |
| 1257 |
if (calleeQualType->isFunctionPointerType() || |
| 1258 |
calleeQualType->isFunctionReferenceType() || |
| 1259 |
calleeQualType->isBlockPointerType() || |
| 1260 |
calleeQualType->isMemberFunctionPointerType()) { |
| 1261 |
calleeType = calleeQualType->getPointeeType()->castAs<FunctionType>(); |
| 1262 |
} else if (auto *ty = dyn_cast<FunctionType>(calleeQualType)) { |
| 1263 |
calleeType = ty; |
| 1264 |
} else if (auto CMCE = dyn_cast<CXXMemberCallExpr>(CE)) { |
| 1265 |
if (auto methodDecl = CMCE->getMethodDecl()) { |
| 1266 |
// getMethodDecl() doesn't handle member pointers at the moment. |
| 1267 |
calleeType = methodDecl->getType()->castAs<FunctionType>(); |
| 1268 |
} else { |
| 1269 |
return; |
| 1270 |
} |
| 1271 |
} else { |
| 1272 |
return; |
| 1273 |
} |
| 1274 |
if (calleeType->getCallConv() == CallingConv::CC_SwiftAsync && |
| 1275 |
(CurFnInfo->getASTCallingConvention() == CallingConv::CC_SwiftAsync)) { |
| 1276 |
auto CI = cast<llvm::CallInst>(&Builder.GetInsertBlock()->back()); |
| 1277 |
CI->setTailCallKind(llvm::CallInst::TCK_MustTail); |
| 1278 |
Builder.CreateRetVoid(); |
| 1279 |
Builder.ClearInsertionPoint(); |
| 1280 |
} |
| 1281 |
} |
| 1282 |
|
| 1283 |
/// EmitReturnStmt - Note that due to GCC extensions, this can have an operand |
| 1284 |
/// if the function returns void, or may be missing one if the function returns |
| 1285 |
/// non-void. Fun stuff :). |
| 1286 |
void CodeGenFunction::EmitReturnStmt(const ReturnStmt &S) { |
| 1287 |
if (requiresReturnValueCheck()) { |
| 1288 |
llvm::Constant *SLoc = EmitCheckSourceLocation(S.getBeginLoc()); |
| 1289 |
auto *SLocPtr = |
| 1290 |
new llvm::GlobalVariable(CGM.getModule(), SLoc->getType(), false, |
| 1291 |
llvm::GlobalVariable::PrivateLinkage, SLoc); |
| 1292 |
SLocPtr->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global); |
| 1293 |
CGM.getSanitizerMetadata()->disableSanitizerForGlobal(SLocPtr); |
| 1294 |
assert(ReturnLocation.isValid() && "No valid return location"); |
| 1295 |
Builder.CreateStore(Builder.CreateBitCast(SLocPtr, Int8PtrTy), |
| 1296 |
ReturnLocation); |
| 1297 |
} |
| 1298 |
|
| 1299 |
// Returning from an outlined SEH helper is UB, and we already warn on it. |
| 1300 |
if (IsOutlinedSEHHelper) { |
| 1301 |
Builder.CreateUnreachable(); |
| 1302 |
Builder.ClearInsertionPoint(); |
| 1303 |
} |
| 1304 |
|
| 1305 |
// Emit the result value, even if unused, to evaluate the side effects. |
| 1306 |
const Expr *RV = S.getRetValue(); |
| 1307 |
|
| 1308 |
// Record the result expression of the return statement. The recorded |
| 1309 |
// expression is used to determine whether a block capture's lifetime should |
| 1310 |
// end at the end of the full expression as opposed to the end of the scope |
| 1311 |
// enclosing the block expression. |
| 1312 |
// |
| 1313 |
// This permits a small, easily-implemented exception to our over-conservative |
| 1314 |
// rules about not jumping to statements following block literals with |
| 1315 |
// non-trivial cleanups. |
| 1316 |
SaveRetExprRAII SaveRetExpr(RV, *this); |
| 1317 |
|
| 1318 |
RunCleanupsScope cleanupScope(*this); |
| 1319 |
if (const auto *EWC = dyn_cast_or_null<ExprWithCleanups>(RV)) |
| 1320 |
RV = EWC->getSubExpr(); |
| 1321 |
// FIXME: Clean this up by using an LValue for ReturnTemp, |
| 1322 |
// EmitStoreThroughLValue, and EmitAnyExpr. |
| 1323 |
// Check if the NRVO candidate was not globalized in OpenMP mode. |
| 1324 |
if (getLangOpts().ElideConstructors && S.getNRVOCandidate() && |
| 1325 |
S.getNRVOCandidate()->isNRVOVariable() && |
| 1326 |
(!getLangOpts().OpenMP || |
| 1327 |
!CGM.getOpenMPRuntime() |
| 1328 |
.getAddressOfLocalVariable(*this, S.getNRVOCandidate()) |
| 1329 |
.isValid())) { |
| 1330 |
// Apply the named return value optimization for this return statement, |
| 1331 |
// which means doing nothing: the appropriate result has already been |
| 1332 |
// constructed into the NRVO variable. |
| 1333 |
|
| 1334 |
// If there is an NRVO flag for this variable, set it to 1 into indicate |
| 1335 |
// that the cleanup code should not destroy the variable. |
| 1336 |
if (llvm::Value *NRVOFlag = NRVOFlags[S.getNRVOCandidate()]) |
| 1337 |
Builder.CreateFlagStore(Builder.getTrue(), NRVOFlag); |
| 1338 |
} else if (!ReturnValue.isValid() || (RV && RV->getType()->isVoidType())) { |
| 1339 |
// Make sure not to return anything, but evaluate the expression |
| 1340 |
// for side effects. |
| 1341 |
if (RV) { |
| 1342 |
EmitAnyExpr(RV); |
| 1343 |
if (auto *CE = dyn_cast<CallExpr>(RV)) |
| 1344 |
makeTailCallIfSwiftAsync(CE, Builder, CurFnInfo); |
| 1345 |
} |
| 1346 |
} else if (!RV) { |
| 1347 |
// Do nothing (return value is left uninitialized) |
| 1348 |
} else if (FnRetTy->isReferenceType()) { |
| 1349 |
// If this function returns a reference, take the address of the expression |
| 1350 |
// rather than the value. |
| 1351 |
RValue Result = EmitReferenceBindingToExpr(RV); |
| 1352 |
Builder.CreateStore(Result.getScalarVal(), ReturnValue); |
| 1353 |
} else { |
| 1354 |
switch (getEvaluationKind(RV->getType())) { |
| 1355 |
case TEK_Scalar: |
| 1356 |
Builder.CreateStore(EmitScalarExpr(RV), ReturnValue); |
| 1357 |
break; |
| 1358 |
case TEK_Complex: |
| 1359 |
EmitComplexExprIntoLValue(RV, MakeAddrLValue(ReturnValue, RV->getType()), |
| 1360 |
/*isInit*/ true); |
| 1361 |
break; |
| 1362 |
case TEK_Aggregate: |
| 1363 |
EmitAggExpr(RV, AggValueSlot::forAddr( |
| 1364 |
ReturnValue, Qualifiers(), |
| 1365 |
AggValueSlot::IsDestructed, |
| 1366 |
AggValueSlot::DoesNotNeedGCBarriers, |
| 1367 |
AggValueSlot::IsNotAliased, |
| 1368 |
getOverlapForReturnValue())); |
| 1369 |
break; |
| 1370 |
} |
| 1371 |
} |
| 1372 |
|
| 1373 |
++NumReturnExprs; |
| 1374 |
if (!RV || RV->isEvaluatable(getContext())) |
| 1375 |
++NumSimpleReturnExprs; |
| 1376 |
|
| 1377 |
cleanupScope.ForceCleanup(); |
| 1378 |
EmitBranchThroughCleanup(ReturnBlock); |
| 1379 |
} |
| 1380 |
|
| 1381 |
void CodeGenFunction::EmitDeclStmt(const DeclStmt &S) { |
| 1382 |
// As long as debug info is modeled with instructions, we have to ensure we |
| 1383 |
// have a place to insert here and write the stop point here. |
| 1384 |
if (HaveInsertPoint()) |
| 1385 |
EmitStopPoint(&S); |
| 1386 |
|
| 1387 |
for (const auto *I : S.decls()) |
| 1388 |
EmitDecl(*I); |
| 1389 |
} |
| 1390 |
|
| 1391 |
void CodeGenFunction::EmitBreakStmt(const BreakStmt &S) { |
| 1392 |
assert(!BreakContinueStack.empty() && "break stmt not in a loop or switch!"); |
| 1393 |
|
| 1394 |
// If this code is reachable then emit a stop point (if generating |
| 1395 |
// debug info). We have to do this ourselves because we are on the |
| 1396 |
// "simple" statement path. |
| 1397 |
if (HaveInsertPoint()) |
| 1398 |
EmitStopPoint(&S); |
| 1399 |
|
| 1400 |
EmitBranchThroughCleanup(BreakContinueStack.back().BreakBlock); |
| 1401 |
} |
| 1402 |
|
| 1403 |
void CodeGenFunction::EmitContinueStmt(const ContinueStmt &S) { |
| 1404 |
assert(!BreakContinueStack.empty() && "continue stmt not in a loop!"); |
| 1405 |
|
| 1406 |
// If this code is reachable then emit a stop point (if generating |
| 1407 |
// debug info). We have to do this ourselves because we are on the |
| 1408 |
// "simple" statement path. |
| 1409 |
if (HaveInsertPoint()) |
| 1410 |
EmitStopPoint(&S); |
| 1411 |
|
| 1412 |
EmitBranchThroughCleanup(BreakContinueStack.back().ContinueBlock); |
| 1413 |
} |
| 1414 |
|
| 1415 |
/// EmitCaseStmtRange - If case statement range is not too big then |
| 1416 |
/// add multiple cases to switch instruction, one for each value within |
| 1417 |
/// the range. If range is too big then emit "if" condition check. |
| 1418 |
void CodeGenFunction::EmitCaseStmtRange(const CaseStmt &S, |
| 1419 |
ArrayRef<const Attr *> Attrs) { |
| 1420 |
assert(S.getRHS() && "Expected RHS value in CaseStmt"); |
| 1421 |
|
| 1422 |
llvm::APSInt LHS = S.getLHS()->EvaluateKnownConstInt(getContext()); |
| 1423 |
llvm::APSInt RHS = S.getRHS()->EvaluateKnownConstInt(getContext()); |
| 1424 |
|
| 1425 |
// Emit the code for this case. We do this first to make sure it is |
| 1426 |
// properly chained from our predecessor before generating the |
| 1427 |
// switch machinery to enter this block. |
| 1428 |
llvm::BasicBlock *CaseDest = createBasicBlock("sw.bb"); |
| 1429 |
EmitBlockWithFallThrough(CaseDest, &S); |
| 1430 |
EmitStmt(S.getSubStmt()); |
| 1431 |
|
| 1432 |
// If range is empty, do nothing. |
| 1433 |
if (LHS.isSigned() ? RHS.slt(LHS) : RHS.ult(LHS)) |
| 1434 |
return; |
| 1435 |
|
| 1436 |
Stmt::Likelihood LH = Stmt::getLikelihood(Attrs); |
| 1437 |
llvm::APInt Range = RHS - LHS; |
| 1438 |
// FIXME: parameters such as this should not be hardcoded. |
| 1439 |
if (Range.ult(llvm::APInt(Range.getBitWidth(), 64))) { |
| 1440 |
// Range is small enough to add multiple switch instruction cases. |
| 1441 |
uint64_t Total = getProfileCount(&S); |
| 1442 |
unsigned NCases = Range.getZExtValue() + 1; |
| 1443 |
// We only have one region counter for the entire set of cases here, so we |
| 1444 |
// need to divide the weights evenly between the generated cases, ensuring |
| 1445 |
// that the total weight is preserved. E.g., a weight of 5 over three cases |
| 1446 |
// will be distributed as weights of 2, 2, and 1. |
| 1447 |
uint64_t Weight = Total / NCases, Rem = Total % NCases; |
| 1448 |
for (unsigned I = 0; I != NCases; ++I) { |
| 1449 |
if (SwitchWeights) |
| 1450 |
SwitchWeights->push_back(Weight + (Rem ? 1 : 0)); |
| 1451 |
else if (SwitchLikelihood) |
| 1452 |
SwitchLikelihood->push_back(LH); |
| 1453 |
|
| 1454 |
if (Rem) |
| 1455 |
Rem--; |
| 1456 |
SwitchInsn->addCase(Builder.getInt(LHS), CaseDest); |
| 1457 |
++LHS; |
| 1458 |
} |
| 1459 |
return; |
| 1460 |
} |
| 1461 |
|
| 1462 |
// The range is too big. Emit "if" condition into a new block, |
| 1463 |
// making sure to save and restore the current insertion point. |
| 1464 |
llvm::BasicBlock *RestoreBB = Builder.GetInsertBlock(); |
| 1465 |
|
| 1466 |
// Push this test onto the chain of range checks (which terminates |
| 1467 |
// in the default basic block). The switch's default will be changed |
| 1468 |
// to the top of this chain after switch emission is complete. |
| 1469 |
llvm::BasicBlock *FalseDest = CaseRangeBlock; |
| 1470 |
CaseRangeBlock = createBasicBlock("sw.caserange"); |
| 1471 |
|
| 1472 |
CurFn->insert(CurFn->end(), CaseRangeBlock); |
| 1473 |
Builder.SetInsertPoint(CaseRangeBlock); |
| 1474 |
|
| 1475 |
// Emit range check. |
| 1476 |
llvm::Value *Diff = |
| 1477 |
Builder.CreateSub(SwitchInsn->getCondition(), Builder.getInt(LHS)); |
| 1478 |
llvm::Value *Cond = |
| 1479 |
Builder.CreateICmpULE(Diff, Builder.getInt(Range), "inbounds"); |
| 1480 |
|
| 1481 |
llvm::MDNode *Weights = nullptr; |
| 1482 |
if (SwitchWeights) { |
| 1483 |
uint64_t ThisCount = getProfileCount(&S); |
| 1484 |
uint64_t DefaultCount = (*SwitchWeights)[0]; |
| 1485 |
Weights = createProfileWeights(ThisCount, DefaultCount); |
| 1486 |
|
| 1487 |
// Since we're chaining the switch default through each large case range, we |
| 1488 |
// need to update the weight for the default, ie, the first case, to include |
| 1489 |
// this case. |
| 1490 |
(*SwitchWeights)[0] += ThisCount; |
| 1491 |
} else if (SwitchLikelihood) |
| 1492 |
Cond = emitCondLikelihoodViaExpectIntrinsic(Cond, LH); |
| 1493 |
|
| 1494 |
Builder.CreateCondBr(Cond, CaseDest, FalseDest, Weights); |
| 1495 |
|
| 1496 |
// Restore the appropriate insertion point. |
| 1497 |
if (RestoreBB) |
| 1498 |
Builder.SetInsertPoint(RestoreBB); |
| 1499 |
else |
| 1500 |
Builder.ClearInsertionPoint(); |
| 1501 |
} |
| 1502 |
|
| 1503 |
void CodeGenFunction::EmitCaseStmt(const CaseStmt &S, |
| 1504 |
ArrayRef<const Attr *> Attrs) { |
| 1505 |
// If there is no enclosing switch instance that we're aware of, then this |
| 1506 |
// case statement and its block can be elided. This situation only happens |
| 1507 |
// when we've constant-folded the switch, are emitting the constant case, |
| 1508 |
// and part of the constant case includes another case statement. For |
| 1509 |
// instance: switch (4) { case 4: do { case 5: } while (1); } |
| 1510 |
if (!SwitchInsn) { |
| 1511 |
EmitStmt(S.getSubStmt()); |
| 1512 |
return; |
| 1513 |
} |
| 1514 |
|
| 1515 |
// Handle case ranges. |
| 1516 |
if (S.getRHS()) { |
| 1517 |
EmitCaseStmtRange(S, Attrs); |
| 1518 |
return; |
| 1519 |
} |
| 1520 |
|
| 1521 |
llvm::ConstantInt *CaseVal = |
| 1522 |
Builder.getInt(S.getLHS()->EvaluateKnownConstInt(getContext())); |
| 1523 |
|
| 1524 |
// Emit debuginfo for the case value if it is an enum value. |
| 1525 |
const ConstantExpr *CE; |
| 1526 |
if (auto ICE = dyn_cast<ImplicitCastExpr>(S.getLHS())) |
| 1527 |
CE = dyn_cast<ConstantExpr>(ICE->getSubExpr()); |
| 1528 |
else |
| 1529 |
CE = dyn_cast<ConstantExpr>(S.getLHS()); |
| 1530 |
if (CE) { |
| 1531 |
if (auto DE = dyn_cast<DeclRefExpr>(CE->getSubExpr())) |
| 1532 |
if (CGDebugInfo *Dbg = getDebugInfo()) |
| 1533 |
if (CGM.getCodeGenOpts().hasReducedDebugInfo()) |
| 1534 |
Dbg->EmitGlobalVariable(DE->getDecl(), |
| 1535 |
APValue(llvm::APSInt(CaseVal->getValue()))); |
| 1536 |
} |
| 1537 |
|
| 1538 |
if (SwitchLikelihood) |
| 1539 |
SwitchLikelihood->push_back(Stmt::getLikelihood(Attrs)); |
| 1540 |
|
| 1541 |
// If the body of the case is just a 'break', try to not emit an empty block. |
| 1542 |
// If we're profiling or we're not optimizing, leave the block in for better |
| 1543 |
// debug and coverage analysis. |
| 1544 |
if (!CGM.getCodeGenOpts().hasProfileClangInstr() && |
| 1545 |
CGM.getCodeGenOpts().OptimizationLevel > 0 && |
| 1546 |
isa<BreakStmt>(S.getSubStmt())) { |
| 1547 |
JumpDest Block = BreakContinueStack.back().BreakBlock; |
| 1548 |
|
| 1549 |
// Only do this optimization if there are no cleanups that need emitting. |
| 1550 |
if (isObviouslyBranchWithoutCleanups(Block)) { |
| 1551 |
if (SwitchWeights) |
| 1552 |
SwitchWeights->push_back(getProfileCount(&S)); |
| 1553 |
SwitchInsn->addCase(CaseVal, Block.getBlock()); |
| 1554 |
|
| 1555 |
// If there was a fallthrough into this case, make sure to redirect it to |
| 1556 |
// the end of the switch as well. |
| 1557 |
if (Builder.GetInsertBlock()) { |
| 1558 |
Builder.CreateBr(Block.getBlock()); |
| 1559 |
Builder.ClearInsertionPoint(); |
| 1560 |
} |
| 1561 |
return; |
| 1562 |
} |
| 1563 |
} |
| 1564 |
|
| 1565 |
llvm::BasicBlock *CaseDest = createBasicBlock("sw.bb"); |
| 1566 |
EmitBlockWithFallThrough(CaseDest, &S); |
| 1567 |
if (SwitchWeights) |
| 1568 |
SwitchWeights->push_back(getProfileCount(&S)); |
| 1569 |
SwitchInsn->addCase(CaseVal, CaseDest); |
| 1570 |
|
| 1571 |
// Recursively emitting the statement is acceptable, but is not wonderful for |
| 1572 |
// code where we have many case statements nested together, i.e.: |
| 1573 |
// case 1: |
| 1574 |
// case 2: |
| 1575 |
// case 3: etc. |
| 1576 |
// Handling this recursively will create a new block for each case statement |
| 1577 |
// that falls through to the next case which is IR intensive. It also causes |
| 1578 |
// deep recursion which can run into stack depth limitations. Handle |
| 1579 |
// sequential non-range case statements specially. |
| 1580 |
// |
| 1581 |
// TODO When the next case has a likelihood attribute the code returns to the |
| 1582 |
// recursive algorithm. Maybe improve this case if it becomes common practice |
| 1583 |
// to use a lot of attributes. |
| 1584 |
const CaseStmt *CurCase = &S; |
| 1585 |
const CaseStmt *NextCase = dyn_cast<CaseStmt>(S.getSubStmt()); |
| 1586 |
|
| 1587 |
// Otherwise, iteratively add consecutive cases to this switch stmt. |
| 1588 |
while (NextCase && NextCase->getRHS() == nullptr) { |
| 1589 |
CurCase = NextCase; |
| 1590 |
llvm::ConstantInt *CaseVal = |
| 1591 |
Builder.getInt(CurCase->getLHS()->EvaluateKnownConstInt(getContext())); |
| 1592 |
|
| 1593 |
if (SwitchWeights) |
| 1594 |
SwitchWeights->push_back(getProfileCount(NextCase)); |
| 1595 |
if (CGM.getCodeGenOpts().hasProfileClangInstr()) { |
| 1596 |
CaseDest = createBasicBlock("sw.bb"); |
| 1597 |
EmitBlockWithFallThrough(CaseDest, CurCase); |
| 1598 |
} |
| 1599 |
// Since this loop is only executed when the CaseStmt has no attributes |
| 1600 |
// use a hard-coded value. |
| 1601 |
if (SwitchLikelihood) |
| 1602 |
SwitchLikelihood->push_back(Stmt::LH_None); |
| 1603 |
|
| 1604 |
SwitchInsn->addCase(CaseVal, CaseDest); |
| 1605 |
NextCase = dyn_cast<CaseStmt>(CurCase->getSubStmt()); |
| 1606 |
} |
| 1607 |
|
| 1608 |
// Generate a stop point for debug info if the case statement is |
| 1609 |
// followed by a default statement. A fallthrough case before a |
| 1610 |
// default case gets its own branch target. |
| 1611 |
if (CurCase->getSubStmt()->getStmtClass() == Stmt::DefaultStmtClass) |
| 1612 |
EmitStopPoint(CurCase); |
| 1613 |
|
| 1614 |
// Normal default recursion for non-cases. |
| 1615 |
EmitStmt(CurCase->getSubStmt()); |
| 1616 |
} |
| 1617 |
|
| 1618 |
void CodeGenFunction::EmitDefaultStmt(const DefaultStmt &S, |
| 1619 |
ArrayRef<const Attr *> Attrs) { |
| 1620 |
// If there is no enclosing switch instance that we're aware of, then this |
| 1621 |
// default statement can be elided. This situation only happens when we've |
| 1622 |
// constant-folded the switch. |
| 1623 |
if (!SwitchInsn) { |
| 1624 |
EmitStmt(S.getSubStmt()); |
| 1625 |
return; |
| 1626 |
} |
| 1627 |
|
| 1628 |
llvm::BasicBlock *DefaultBlock = SwitchInsn->getDefaultDest(); |
| 1629 |
assert(DefaultBlock->empty() && |
| 1630 |
"EmitDefaultStmt: Default block already defined?"); |
| 1631 |
|
| 1632 |
if (SwitchLikelihood) |
| 1633 |
SwitchLikelihood->front() = Stmt::getLikelihood(Attrs); |
| 1634 |
|
| 1635 |
EmitBlockWithFallThrough(DefaultBlock, &S); |
| 1636 |
|
| 1637 |
EmitStmt(S.getSubStmt()); |
| 1638 |
} |
| 1639 |
|
| 1640 |
/// CollectStatementsForCase - Given the body of a 'switch' statement and a |
| 1641 |
/// constant value that is being switched on, see if we can dead code eliminate |
| 1642 |
/// the body of the switch to a simple series of statements to emit. Basically, |
| 1643 |
/// on a switch (5) we want to find these statements: |
| 1644 |
/// case 5: |
| 1645 |
/// printf(...); <-- |
| 1646 |
/// ++i; <-- |
| 1647 |
/// break; |
| 1648 |
/// |
| 1649 |
/// and add them to the ResultStmts vector. If it is unsafe to do this |
| 1650 |
/// transformation (for example, one of the elided statements contains a label |
| 1651 |
/// that might be jumped to), return CSFC_Failure. If we handled it and 'S' |
| 1652 |
/// should include statements after it (e.g. the printf() line is a substmt of |
| 1653 |
/// the case) then return CSFC_FallThrough. If we handled it and found a break |
| 1654 |
/// statement, then return CSFC_Success. |
| 1655 |
/// |
| 1656 |
/// If Case is non-null, then we are looking for the specified case, checking |
| 1657 |
/// that nothing we jump over contains labels. If Case is null, then we found |
| 1658 |
/// the case and are looking for the break. |
| 1659 |
/// |
| 1660 |
/// If the recursive walk actually finds our Case, then we set FoundCase to |
| 1661 |
/// true. |
| 1662 |
/// |
| 1663 |
enum CSFC_Result { CSFC_Failure, CSFC_FallThrough, CSFC_Success }; |
| 1664 |
static CSFC_Result CollectStatementsForCase(const Stmt *S, |
| 1665 |
const SwitchCase *Case, |
| 1666 |
bool &FoundCase, |
| 1667 |
SmallVectorImpl<const Stmt*> &ResultStmts) { |
| 1668 |
// If this is a null statement, just succeed. |
| 1669 |
if (!S) |
| 1670 |
return Case ? CSFC_Success : CSFC_FallThrough; |
| 1671 |
|
| 1672 |
// If this is the switchcase (case 4: or default) that we're looking for, then |
| 1673 |
// we're in business. Just add the substatement. |
| 1674 |
if (const SwitchCase *SC = dyn_cast<SwitchCase>(S)) { |
| 1675 |
if (S == Case) { |
| 1676 |
FoundCase = true; |
| 1677 |
return CollectStatementsForCase(SC->getSubStmt(), nullptr, FoundCase, |
| 1678 |
ResultStmts); |
| 1679 |
} |
| 1680 |
|
| 1681 |
// Otherwise, this is some other case or default statement, just ignore it. |
| 1682 |
return CollectStatementsForCase(SC->getSubStmt(), Case, FoundCase, |
| 1683 |
ResultStmts); |
| 1684 |
} |
| 1685 |
|
| 1686 |
// If we are in the live part of the code and we found our break statement, |
| 1687 |
// return a success! |
| 1688 |
if (!Case && isa<BreakStmt>(S)) |
| 1689 |
return CSFC_Success; |
| 1690 |
|
| 1691 |
// If this is a switch statement, then it might contain the SwitchCase, the |
| 1692 |
// break, or neither. |
| 1693 |
if (const CompoundStmt *CS = dyn_cast<CompoundStmt>(S)) { |
| 1694 |
// Handle this as two cases: we might be looking for the SwitchCase (if so |
| 1695 |
// the skipped statements must be skippable) or we might already have it. |
| 1696 |
CompoundStmt::const_body_iterator I = CS->body_begin(), E = CS->body_end(); |
| 1697 |
bool StartedInLiveCode = FoundCase; |
| 1698 |
unsigned StartSize = ResultStmts.size(); |
| 1699 |
|
| 1700 |
// If we've not found the case yet, scan through looking for it. |
| 1701 |
if (Case) { |
| 1702 |
// Keep track of whether we see a skipped declaration. The code could be |
| 1703 |
// using the declaration even if it is skipped, so we can't optimize out |
| 1704 |
// the decl if the kept statements might refer to it. |
| 1705 |
bool HadSkippedDecl = false; |
| 1706 |
|
| 1707 |
// If we're looking for the case, just see if we can skip each of the |
| 1708 |
// substatements. |
| 1709 |
for (; Case && I != E; ++I) { |
| 1710 |
HadSkippedDecl |= CodeGenFunction::mightAddDeclToScope(*I); |
| 1711 |
|
| 1712 |
switch (CollectStatementsForCase(*I, Case, FoundCase, ResultStmts)) { |
| 1713 |
case CSFC_Failure: return CSFC_Failure; |
| 1714 |
case CSFC_Success: |
| 1715 |
// A successful result means that either 1) that the statement doesn't |
| 1716 |
// have the case and is skippable, or 2) does contain the case value |
| 1717 |
// and also contains the break to exit the switch. In the later case, |
| 1718 |
// we just verify the rest of the statements are elidable. |
| 1719 |
if (FoundCase) { |
| 1720 |
// If we found the case and skipped declarations, we can't do the |
| 1721 |
// optimization. |
| 1722 |
if (HadSkippedDecl) |
| 1723 |
return CSFC_Failure; |
| 1724 |
|
| 1725 |
for (++I; I != E; ++I) |
| 1726 |
if (CodeGenFunction::ContainsLabel(*I, true)) |
| 1727 |
return CSFC_Failure; |
| 1728 |
return CSFC_Success; |
| 1729 |
} |
| 1730 |
break; |
| 1731 |
case CSFC_FallThrough: |
| 1732 |
// If we have a fallthrough condition, then we must have found the |
| 1733 |
// case started to include statements. Consider the rest of the |
| 1734 |
// statements in the compound statement as candidates for inclusion. |
| 1735 |
assert(FoundCase && "Didn't find case but returned fallthrough?"); |
| 1736 |
// We recursively found Case, so we're not looking for it anymore. |
| 1737 |
Case = nullptr; |
| 1738 |
|
| 1739 |
// If we found the case and skipped declarations, we can't do the |
| 1740 |
// optimization. |
| 1741 |
if (HadSkippedDecl) |
| 1742 |
return CSFC_Failure; |
| 1743 |
break; |
| 1744 |
} |
| 1745 |
} |
| 1746 |
|
| 1747 |
if (!FoundCase) |
| 1748 |
return CSFC_Success; |
| 1749 |
|
| 1750 |
assert(!HadSkippedDecl && "fallthrough after skipping decl"); |
| 1751 |
} |
| 1752 |
|
| 1753 |
// If we have statements in our range, then we know that the statements are |
| 1754 |
// live and need to be added to the set of statements we're tracking. |
| 1755 |
bool AnyDecls = false; |
| 1756 |
for (; I != E; ++I) { |
| 1757 |
AnyDecls |= CodeGenFunction::mightAddDeclToScope(*I); |
| 1758 |
|
| 1759 |
switch (CollectStatementsForCase(*I, nullptr, FoundCase, ResultStmts)) { |
| 1760 |
case CSFC_Failure: return CSFC_Failure; |
| 1761 |
case CSFC_FallThrough: |
| 1762 |
// A fallthrough result means that the statement was simple and just |
| 1763 |
// included in ResultStmt, keep adding them afterwards. |
| 1764 |
break; |
| 1765 |
case CSFC_Success: |
| 1766 |
// A successful result means that we found the break statement and |
| 1767 |
// stopped statement inclusion. We just ensure that any leftover stmts |
| 1768 |
// are skippable and return success ourselves. |
| 1769 |
for (++I; I != E; ++I) |
| 1770 |
if (CodeGenFunction::ContainsLabel(*I, true)) |
| 1771 |
return CSFC_Failure; |
| 1772 |
return CSFC_Success; |
| 1773 |
} |
| 1774 |
} |
| 1775 |
|
| 1776 |
// If we're about to fall out of a scope without hitting a 'break;', we |
| 1777 |
// can't perform the optimization if there were any decls in that scope |
| 1778 |
// (we'd lose their end-of-lifetime). |
| 1779 |
if (AnyDecls) { |
| 1780 |
// If the entire compound statement was live, there's one more thing we |
| 1781 |
// can try before giving up: emit the whole thing as a single statement. |
| 1782 |
// We can do that unless the statement contains a 'break;'. |
| 1783 |
// FIXME: Such a break must be at the end of a construct within this one. |
| 1784 |
// We could emit this by just ignoring the BreakStmts entirely. |
| 1785 |
if (StartedInLiveCode && !CodeGenFunction::containsBreak(S)) { |
| 1786 |
ResultStmts.resize(StartSize); |
| 1787 |
ResultStmts.push_back(S); |
| 1788 |
} else { |
| 1789 |
return CSFC_Failure; |
| 1790 |
} |
| 1791 |
} |
| 1792 |
|
| 1793 |
return CSFC_FallThrough; |
| 1794 |
} |
| 1795 |
|
| 1796 |
// Okay, this is some other statement that we don't handle explicitly, like a |
| 1797 |
// for statement or increment etc. If we are skipping over this statement, |
| 1798 |
// just verify it doesn't have labels, which would make it invalid to elide. |
| 1799 |
if (Case) { |
| 1800 |
if (CodeGenFunction::ContainsLabel(S, true)) |
| 1801 |
return CSFC_Failure; |
| 1802 |
return CSFC_Success; |
| 1803 |
} |
| 1804 |
|
| 1805 |
// Otherwise, we want to include this statement. Everything is cool with that |
| 1806 |
// so long as it doesn't contain a break out of the switch we're in. |
| 1807 |
if (CodeGenFunction::containsBreak(S)) return CSFC_Failure; |
| 1808 |
|
| 1809 |
// Otherwise, everything is great. Include the statement and tell the caller |
| 1810 |
// that we fall through and include the next statement as well. |
| 1811 |
ResultStmts.push_back(S); |
| 1812 |
return CSFC_FallThrough; |
| 1813 |
} |
| 1814 |
|
| 1815 |
/// FindCaseStatementsForValue - Find the case statement being jumped to and |
| 1816 |
/// then invoke CollectStatementsForCase to find the list of statements to emit |
| 1817 |
/// for a switch on constant. See the comment above CollectStatementsForCase |
| 1818 |
/// for more details. |
| 1819 |
static bool FindCaseStatementsForValue(const SwitchStmt &S, |
| 1820 |
const llvm::APSInt &ConstantCondValue, |
| 1821 |
SmallVectorImpl<const Stmt*> &ResultStmts, |
| 1822 |
ASTContext &C, |
| 1823 |
const SwitchCase *&ResultCase) { |
| 1824 |
// First step, find the switch case that is being branched to. We can do this |
| 1825 |
// efficiently by scanning the SwitchCase list. |
| 1826 |
const SwitchCase *Case = S.getSwitchCaseList(); |
| 1827 |
const DefaultStmt *DefaultCase = nullptr; |
| 1828 |
|
| 1829 |
for (; Case; Case = Case->getNextSwitchCase()) { |
| 1830 |
// It's either a default or case. Just remember the default statement in |
| 1831 |
// case we're not jumping to any numbered cases. |
| 1832 |
if (const DefaultStmt *DS = dyn_cast<DefaultStmt>(Case)) { |
| 1833 |
DefaultCase = DS; |
| 1834 |
continue; |
| 1835 |
} |
| 1836 |
|
| 1837 |
// Check to see if this case is the one we're looking for. |
| 1838 |
const CaseStmt *CS = cast<CaseStmt>(Case); |
| 1839 |
// Don't handle case ranges yet. |
| 1840 |
if (CS->getRHS()) return false; |
| 1841 |
|
| 1842 |
// If we found our case, remember it as 'case'. |
| 1843 |
if (CS->getLHS()->EvaluateKnownConstInt(C) == ConstantCondValue) |
| 1844 |
break; |
| 1845 |
} |
| 1846 |
|
| 1847 |
// If we didn't find a matching case, we use a default if it exists, or we |
| 1848 |
// elide the whole switch body! |
| 1849 |
if (!Case) { |
| 1850 |
// It is safe to elide the body of the switch if it doesn't contain labels |
| 1851 |
// etc. If it is safe, return successfully with an empty ResultStmts list. |
| 1852 |
if (!DefaultCase) |
| 1853 |
return !CodeGenFunction::ContainsLabel(&S); |
| 1854 |
Case = DefaultCase; |
| 1855 |
} |
| 1856 |
|
| 1857 |
// Ok, we know which case is being jumped to, try to collect all the |
| 1858 |
// statements that follow it. This can fail for a variety of reasons. Also, |
| 1859 |
// check to see that the recursive walk actually found our case statement. |
| 1860 |
// Insane cases like this can fail to find it in the recursive walk since we |
| 1861 |
// don't handle every stmt kind: |
| 1862 |
// switch (4) { |
| 1863 |
// while (1) { |
| 1864 |
// case 4: ... |
| 1865 |
bool FoundCase = false; |
| 1866 |
ResultCase = Case; |
| 1867 |
return CollectStatementsForCase(S.getBody(), Case, FoundCase, |
| 1868 |
ResultStmts) != CSFC_Failure && |
| 1869 |
FoundCase; |
| 1870 |
} |
| 1871 |
|
| 1872 |
static std::optional<SmallVector<uint64_t, 16>> |
| 1873 |
getLikelihoodWeights(ArrayRef<Stmt::Likelihood> Likelihoods) { |
| 1874 |
// Are there enough branches to weight them? |
| 1875 |
if (Likelihoods.size() <= 1) |
| 1876 |
return std::nullopt; |
| 1877 |
|
| 1878 |
uint64_t NumUnlikely = 0; |
| 1879 |
uint64_t NumNone = 0; |
| 1880 |
uint64_t NumLikely = 0; |
| 1881 |
for (const auto LH : Likelihoods) { |
| 1882 |
switch (LH) { |
| 1883 |
case Stmt::LH_Unlikely: |
| 1884 |
++NumUnlikely; |
| 1885 |
break; |
| 1886 |
case Stmt::LH_None: |
| 1887 |
++NumNone; |
| 1888 |
break; |
| 1889 |
case Stmt::LH_Likely: |
| 1890 |
++NumLikely; |
| 1891 |
break; |
| 1892 |
} |
| 1893 |
} |
| 1894 |
|
| 1895 |
// Is there a likelihood attribute used? |
| 1896 |
if (NumUnlikely == 0 && NumLikely == 0) |
| 1897 |
return std::nullopt; |
| 1898 |
|
| 1899 |
// When multiple cases share the same code they can be combined during |
| 1900 |
// optimization. In that case the weights of the branch will be the sum of |
| 1901 |
// the individual weights. Make sure the combined sum of all neutral cases |
| 1902 |
// doesn't exceed the value of a single likely attribute. |
| 1903 |
// The additions both avoid divisions by 0 and make sure the weights of None |
| 1904 |
// don't exceed the weight of Likely. |
| 1905 |
const uint64_t Likely = INT32_MAX / (NumLikely + 2); |
| 1906 |
const uint64_t None = Likely / (NumNone + 1); |
| 1907 |
const uint64_t Unlikely = 0; |
| 1908 |
|
| 1909 |
SmallVector<uint64_t, 16> Result; |
| 1910 |
Result.reserve(Likelihoods.size()); |
| 1911 |
for (const auto LH : Likelihoods) { |
| 1912 |
switch (LH) { |
| 1913 |
case Stmt::LH_Unlikely: |
| 1914 |
Result.push_back(Unlikely); |
| 1915 |
break; |
| 1916 |
case Stmt::LH_None: |
| 1917 |
Result.push_back(None); |
| 1918 |
break; |
| 1919 |
case Stmt::LH_Likely: |
| 1920 |
Result.push_back(Likely); |
| 1921 |
break; |
| 1922 |
} |
| 1923 |
} |
| 1924 |
|
| 1925 |
return Result; |
| 1926 |
} |
| 1927 |
|
| 1928 |
void CodeGenFunction::EmitSwitchStmt(const SwitchStmt &S) { |
| 1929 |
// Handle nested switch statements. |
| 1930 |
llvm::SwitchInst *SavedSwitchInsn = SwitchInsn; |
| 1931 |
SmallVector<uint64_t, 16> *SavedSwitchWeights = SwitchWeights; |
| 1932 |
SmallVector<Stmt::Likelihood, 16> *SavedSwitchLikelihood = SwitchLikelihood; |
| 1933 |
llvm::BasicBlock *SavedCRBlock = CaseRangeBlock; |
| 1934 |
|
| 1935 |
// See if we can constant fold the condition of the switch and therefore only |
| 1936 |
// emit the live case statement (if any) of the switch. |
| 1937 |
llvm::APSInt ConstantCondValue; |
| 1938 |
if (ConstantFoldsToSimpleInteger(S.getCond(), ConstantCondValue)) { |
| 1939 |
SmallVector<const Stmt*, 4> CaseStmts; |
| 1940 |
const SwitchCase *Case = nullptr; |
| 1941 |
if (FindCaseStatementsForValue(S, ConstantCondValue, CaseStmts, |
| 1942 |
getContext(), Case)) { |
| 1943 |
if (Case) |
| 1944 |
incrementProfileCounter(Case); |
| 1945 |
RunCleanupsScope ExecutedScope(*this); |
| 1946 |
|
| 1947 |
if (S.getInit()) |
| 1948 |
EmitStmt(S.getInit()); |
| 1949 |
|
| 1950 |
// Emit the condition variable if needed inside the entire cleanup scope |
| 1951 |
// used by this special case for constant folded switches. |
| 1952 |
if (S.getConditionVariable()) |
| 1953 |
EmitDecl(*S.getConditionVariable()); |
| 1954 |
|
| 1955 |
// At this point, we are no longer "within" a switch instance, so |
| 1956 |
// we can temporarily enforce this to ensure that any embedded case |
| 1957 |
// statements are not emitted. |
| 1958 |
SwitchInsn = nullptr; |
| 1959 |
|
| 1960 |
// Okay, we can dead code eliminate everything except this case. Emit the |
| 1961 |
// specified series of statements and we're good. |
| 1962 |
for (unsigned i = 0, e = CaseStmts.size(); i != e; ++i) |
| 1963 |
EmitStmt(CaseStmts[i]); |
| 1964 |
incrementProfileCounter(&S); |
| 1965 |
|
| 1966 |
// Now we want to restore the saved switch instance so that nested |
| 1967 |
// switches continue to function properly |
| 1968 |
SwitchInsn = SavedSwitchInsn; |
| 1969 |
|
| 1970 |
return; |
| 1971 |
} |
| 1972 |
} |
| 1973 |
|
| 1974 |
JumpDest SwitchExit = getJumpDestInCurrentScope("sw.epilog"); |
| 1975 |
|
| 1976 |
RunCleanupsScope ConditionScope(*this); |
| 1977 |
|
| 1978 |
if (S.getInit()) |
| 1979 |
EmitStmt(S.getInit()); |
| 1980 |
|
| 1981 |
if (S.getConditionVariable()) |
| 1982 |
EmitDecl(*S.getConditionVariable()); |
| 1983 |
llvm::Value *CondV = EmitScalarExpr(S.getCond()); |
| 1984 |
|
| 1985 |
// Create basic block to hold stuff that comes after switch |
| 1986 |
// statement. We also need to create a default block now so that |
| 1987 |
// explicit case ranges tests can have a place to jump to on |
| 1988 |
// failure. |
| 1989 |
llvm::BasicBlock *DefaultBlock = createBasicBlock("sw.default"); |
| 1990 |
SwitchInsn = Builder.CreateSwitch(CondV, DefaultBlock); |
| 1991 |
if (PGO.haveRegionCounts()) { |
| 1992 |
// Walk the SwitchCase list to find how many there are. |
| 1993 |
uint64_t DefaultCount = 0; |
| 1994 |
unsigned NumCases = 0; |
| 1995 |
for (const SwitchCase *Case = S.getSwitchCaseList(); |
| 1996 |
Case; |
| 1997 |
Case = Case->getNextSwitchCase()) { |
| 1998 |
if (isa<DefaultStmt>(Case)) |
| 1999 |
DefaultCount = getProfileCount(Case); |
| 2000 |
NumCases += 1; |
| 2001 |
} |
| 2002 |
SwitchWeights = new SmallVector<uint64_t, 16>(); |
| 2003 |
SwitchWeights->reserve(NumCases); |
| 2004 |
// The default needs to be first. We store the edge count, so we already |
| 2005 |
// know the right weight. |
| 2006 |
SwitchWeights->push_back(DefaultCount); |
| 2007 |
} else if (CGM.getCodeGenOpts().OptimizationLevel) { |
| 2008 |
SwitchLikelihood = new SmallVector<Stmt::Likelihood, 16>(); |
| 2009 |
// Initialize the default case. |
| 2010 |
SwitchLikelihood->push_back(Stmt::LH_None); |
| 2011 |
} |
| 2012 |
|
| 2013 |
CaseRangeBlock = DefaultBlock; |
| 2014 |
|
| 2015 |
// Clear the insertion point to indicate we are in unreachable code. |
| 2016 |
Builder.ClearInsertionPoint(); |
| 2017 |
|
| 2018 |
// All break statements jump to NextBlock. If BreakContinueStack is non-empty |
| 2019 |
// then reuse last ContinueBlock. |
| 2020 |
JumpDest OuterContinue; |
| 2021 |
if (!BreakContinueStack.empty()) |
| 2022 |
OuterContinue = BreakContinueStack.back().ContinueBlock; |
| 2023 |
|
| 2024 |
BreakContinueStack.push_back(BreakContinue(SwitchExit, OuterContinue)); |
| 2025 |
|
| 2026 |
// Emit switch body. |
| 2027 |
EmitStmt(S.getBody()); |
| 2028 |
|
| 2029 |
BreakContinueStack.pop_back(); |
| 2030 |
|
| 2031 |
// Update the default block in case explicit case range tests have |
| 2032 |
// been chained on top. |
| 2033 |
SwitchInsn->setDefaultDest(CaseRangeBlock); |
| 2034 |
|
| 2035 |
// If a default was never emitted: |
| 2036 |
if (!DefaultBlock->getParent()) { |
| 2037 |
// If we have cleanups, emit the default block so that there's a |
| 2038 |
// place to jump through the cleanups from. |
| 2039 |
if (ConditionScope.requiresCleanups()) { |
| 2040 |
EmitBlock(DefaultBlock); |
| 2041 |
|
| 2042 |
// Otherwise, just forward the default block to the switch end. |
| 2043 |
} else { |
| 2044 |
DefaultBlock->replaceAllUsesWith(SwitchExit.getBlock()); |
| 2045 |
delete DefaultBlock; |
| 2046 |
} |
| 2047 |
} |
| 2048 |
|
| 2049 |
ConditionScope.ForceCleanup(); |
| 2050 |
|
| 2051 |
// Emit continuation. |
| 2052 |
EmitBlock(SwitchExit.getBlock(), true); |
| 2053 |
incrementProfileCounter(&S); |
| 2054 |
|
| 2055 |
// If the switch has a condition wrapped by __builtin_unpredictable, |
| 2056 |
// create metadata that specifies that the switch is unpredictable. |
| 2057 |
// Don't bother if not optimizing because that metadata would not be used. |
| 2058 |
auto *Call = dyn_cast<CallExpr>(S.getCond()); |
| 2059 |
if (Call && CGM.getCodeGenOpts().OptimizationLevel != 0) { |
| 2060 |
auto *FD = dyn_cast_or_null<FunctionDecl>(Call->getCalleeDecl()); |
| 2061 |
if (FD && FD->getBuiltinID() == Builtin::BI__builtin_unpredictable) { |
| 2062 |
llvm::MDBuilder MDHelper(getLLVMContext()); |
| 2063 |
SwitchInsn->setMetadata(llvm::LLVMContext::MD_unpredictable, |
| 2064 |
MDHelper.createUnpredictable()); |
| 2065 |
} |
| 2066 |
} |
| 2067 |
|
| 2068 |
if (SwitchWeights) { |
| 2069 |
assert(SwitchWeights->size() == 1 + SwitchInsn->getNumCases() && |
| 2070 |
"switch weights do not match switch cases"); |
| 2071 |
// If there's only one jump destination there's no sense weighting it. |
| 2072 |
if (SwitchWeights->size() > 1) |
| 2073 |
SwitchInsn->setMetadata(llvm::LLVMContext::MD_prof, |
| 2074 |
createProfileWeights(*SwitchWeights)); |
| 2075 |
delete SwitchWeights; |
| 2076 |
} else if (SwitchLikelihood) { |
| 2077 |
assert(SwitchLikelihood->size() == 1 + SwitchInsn->getNumCases() && |
| 2078 |
"switch likelihoods do not match switch cases"); |
| 2079 |
std::optional<SmallVector<uint64_t, 16>> LHW = |
| 2080 |
getLikelihoodWeights(*SwitchLikelihood); |
| 2081 |
if (LHW) { |
| 2082 |
llvm::MDBuilder MDHelper(CGM.getLLVMContext()); |
| 2083 |
SwitchInsn->setMetadata(llvm::LLVMContext::MD_prof, |
| 2084 |
createProfileWeights(*LHW)); |
| 2085 |
} |
| 2086 |
delete SwitchLikelihood; |
| 2087 |
} |
| 2088 |
SwitchInsn = SavedSwitchInsn; |
| 2089 |
SwitchWeights = SavedSwitchWeights; |
| 2090 |
SwitchLikelihood = SavedSwitchLikelihood; |
| 2091 |
CaseRangeBlock = SavedCRBlock; |
| 2092 |
} |
| 2093 |
|
| 2094 |
static std::string |
| 2095 |
SimplifyConstraint(const char *Constraint, const TargetInfo &Target, |
| 2096 |
SmallVectorImpl<TargetInfo::ConstraintInfo> *OutCons=nullptr) { |
| 2097 |
std::string Result; |
| 2098 |
|
| 2099 |
while (*Constraint) { |
| 2100 |
switch (*Constraint) { |
| 2101 |
default: |
| 2102 |
Result += Target.convertConstraint(Constraint); |
| 2103 |
break; |
| 2104 |
// Ignore these |
| 2105 |
case '*': |
| 2106 |
case '?': |
| 2107 |
case '!': |
| 2108 |
case '=': // Will see this and the following in mult-alt constraints. |
| 2109 |
case '+': |
| 2110 |
break; |
| 2111 |
case '#': // Ignore the rest of the constraint alternative. |
| 2112 |
while (Constraint[1] && Constraint[1] != ',') |
| 2113 |
Constraint++; |
| 2114 |
break; |
| 2115 |
case '&': |
| 2116 |
case '%': |
| 2117 |
Result += *Constraint; |
| 2118 |
while (Constraint[1] && Constraint[1] == *Constraint) |
| 2119 |
Constraint++; |
| 2120 |
break; |
| 2121 |
case ',': |
| 2122 |
Result += "|"; |
| 2123 |
break; |
| 2124 |
case 'g': |
| 2125 |
Result += "imr"; |
| 2126 |
break; |
| 2127 |
case '[': { |
| 2128 |
assert(OutCons && |
| 2129 |
"Must pass output names to constraints with a symbolic name"); |
| 2130 |
unsigned Index; |
| 2131 |
bool result = Target.resolveSymbolicName(Constraint, *OutCons, Index); |
| 2132 |
assert(result && "Could not resolve symbolic name"); (void)result; |
| 2133 |
Result += llvm::utostr(Index); |
| 2134 |
break; |
| 2135 |
} |
| 2136 |
} |
| 2137 |
|
| 2138 |
Constraint++; |
| 2139 |
} |
| 2140 |
|
| 2141 |
return Result; |
| 2142 |
} |
| 2143 |
|
| 2144 |
/// AddVariableConstraints - Look at AsmExpr and if it is a variable declared |
| 2145 |
/// as using a particular register add that as a constraint that will be used |
| 2146 |
/// in this asm stmt. |
| 2147 |
static std::string |
| 2148 |
AddVariableConstraints(const std::string &Constraint, const Expr &AsmExpr, |
| 2149 |
const TargetInfo &Target, CodeGenModule &CGM, |
| 2150 |
const AsmStmt &Stmt, const bool EarlyClobber, |
| 2151 |
std::string *GCCReg = nullptr) { |
| 2152 |
const DeclRefExpr *AsmDeclRef = dyn_cast<DeclRefExpr>(&AsmExpr); |
| 2153 |
if (!AsmDeclRef) |
| 2154 |
return Constraint; |
| 2155 |
const ValueDecl &Value = *AsmDeclRef->getDecl(); |
| 2156 |
const VarDecl *Variable = dyn_cast<VarDecl>(&Value); |
| 2157 |
if (!Variable) |
| 2158 |
return Constraint; |
| 2159 |
if (Variable->getStorageClass() != SC_Register) |
| 2160 |
return Constraint; |
| 2161 |
AsmLabelAttr *Attr = Variable->getAttr<AsmLabelAttr>(); |
| 2162 |
if (!Attr) |
| 2163 |
return Constraint; |
| 2164 |
StringRef Register = Attr->getLabel(); |
| 2165 |
assert(Target.isValidGCCRegisterName(Register)); |
| 2166 |
// We're using validateOutputConstraint here because we only care if |
| 2167 |
// this is a register constraint. |
| 2168 |
TargetInfo::ConstraintInfo Info(Constraint, ""); |
| 2169 |
if (Target.validateOutputConstraint(Info) && |
| 2170 |
!Info.allowsRegister()) { |
| 2171 |
CGM.ErrorUnsupported(&Stmt, "__asm__"); |
| 2172 |
return Constraint; |
| 2173 |
} |
| 2174 |
// Canonicalize the register here before returning it. |
| 2175 |
Register = Target.getNormalizedGCCRegisterName(Register); |
| 2176 |
if (GCCReg != nullptr) |
| 2177 |
*GCCReg = Register.str(); |
| 2178 |
return (EarlyClobber ? "&{" : "{") + Register.str() + "}"; |
| 2179 |
} |
| 2180 |
|
| 2181 |
std::pair<llvm::Value*, llvm::Type *> CodeGenFunction::EmitAsmInputLValue( |
| 2182 |
const TargetInfo::ConstraintInfo &Info, LValue InputValue, |
| 2183 |
QualType InputType, std::string &ConstraintStr, SourceLocation Loc) { |
| 2184 |
if (Info.allowsRegister() || !Info.allowsMemory()) { |
| 2185 |
if (CodeGenFunction::hasScalarEvaluationKind(InputType)) |
| 2186 |
return {EmitLoadOfLValue(InputValue, Loc).getScalarVal(), nullptr}; |
| 2187 |
|
| 2188 |
llvm::Type *Ty = ConvertType(InputType); |
| 2189 |
uint64_t Size = CGM.getDataLayout().getTypeSizeInBits(Ty); |
| 2190 |
if ((Size <= 64 && llvm::isPowerOf2_64(Size)) || |
| 2191 |
getTargetHooks().isScalarizableAsmOperand(*this, Ty)) { |
| 2192 |
Ty = llvm::IntegerType::get(getLLVMContext(), Size); |
| 2193 |
|
| 2194 |
return {Builder.CreateLoad(Builder.CreateElementBitCast( |
| 2195 |
InputValue.getAddress(*this), Ty)), |
| 2196 |
nullptr}; |
| 2197 |
} |
| 2198 |
} |
| 2199 |
|
| 2200 |
Address Addr = InputValue.getAddress(*this); |
| 2201 |
ConstraintStr += '*'; |
| 2202 |
return {Addr.getPointer(), Addr.getElementType()}; |
| 2203 |
} |
| 2204 |
|
| 2205 |
std::pair<llvm::Value *, llvm::Type *> |
| 2206 |
CodeGenFunction::EmitAsmInput(const TargetInfo::ConstraintInfo &Info, |
| 2207 |
const Expr *InputExpr, |
| 2208 |
std::string &ConstraintStr) { |
| 2209 |
// If this can't be a register or memory, i.e., has to be a constant |
| 2210 |
// (immediate or symbolic), try to emit it as such. |
| 2211 |
if (!Info.allowsRegister() && !Info.allowsMemory()) { |
| 2212 |
if (Info.requiresImmediateConstant()) { |
| 2213 |
Expr::EvalResult EVResult; |
| 2214 |
InputExpr->EvaluateAsRValue(EVResult, getContext(), true); |
| 2215 |
|
| 2216 |
llvm::APSInt IntResult; |
| 2217 |
if (EVResult.Val.toIntegralConstant(IntResult, InputExpr->getType(), |
| 2218 |
getContext())) |
| 2219 |
return {llvm::ConstantInt::get(getLLVMContext(), IntResult), nullptr}; |
| 2220 |
} |
| 2221 |
|
| 2222 |
Expr::EvalResult Result; |
| 2223 |
if (InputExpr->EvaluateAsInt(Result, getContext())) |
| 2224 |
return {llvm::ConstantInt::get(getLLVMContext(), Result.Val.getInt()), |
| 2225 |
nullptr}; |
| 2226 |
} |
| 2227 |
|
| 2228 |
if (Info.allowsRegister() || !Info.allowsMemory()) |
| 2229 |
if (CodeGenFunction::hasScalarEvaluationKind(InputExpr->getType())) |
| 2230 |
return {EmitScalarExpr(InputExpr), nullptr}; |
| 2231 |
if (InputExpr->getStmtClass() == Expr::CXXThisExprClass) |
| 2232 |
return {EmitScalarExpr(InputExpr), nullptr}; |
| 2233 |
InputExpr = InputExpr->IgnoreParenNoopCasts(getContext()); |
| 2234 |
LValue Dest = EmitLValue(InputExpr); |
| 2235 |
return EmitAsmInputLValue(Info, Dest, InputExpr->getType(), ConstraintStr, |
| 2236 |
InputExpr->getExprLoc()); |
| 2237 |
} |
| 2238 |
|
| 2239 |
/// getAsmSrcLocInfo - Return the !srcloc metadata node to attach to an inline |
| 2240 |
/// asm call instruction. The !srcloc MDNode contains a list of constant |
| 2241 |
/// integers which are the source locations of the start of each line in the |
| 2242 |
/// asm. |
| 2243 |
static llvm::MDNode *getAsmSrcLocInfo(const StringLiteral *Str, |
| 2244 |
CodeGenFunction &CGF) { |
| 2245 |
SmallVector<llvm::Metadata *, 8> Locs; |
| 2246 |
// Add the location of the first line to the MDNode. |
| 2247 |
Locs.push_back(llvm::ConstantAsMetadata::get(llvm::ConstantInt::get( |
| 2248 |
CGF.Int64Ty, Str->getBeginLoc().getRawEncoding()))); |
| 2249 |
StringRef StrVal = Str->getString(); |
| 2250 |
if (!StrVal.empty()) { |
| 2251 |
const SourceManager &SM = CGF.CGM.getContext().getSourceManager(); |
| 2252 |
const LangOptions &LangOpts = CGF.CGM.getLangOpts(); |
| 2253 |
unsigned StartToken = 0; |
| 2254 |
unsigned ByteOffset = 0; |
| 2255 |
|
| 2256 |
// Add the location of the start of each subsequent line of the asm to the |
| 2257 |
// MDNode. |
| 2258 |
for (unsigned i = 0, e = StrVal.size() - 1; i != e; ++i) { |
| 2259 |
if (StrVal[i] != '\n') continue; |
| 2260 |
SourceLocation LineLoc = Str->getLocationOfByte( |
| 2261 |
i + 1, SM, LangOpts, CGF.getTarget(), &StartToken, &ByteOffset); |
| 2262 |
Locs.push_back(llvm::ConstantAsMetadata::get( |
| 2263 |
llvm::ConstantInt::get(CGF.Int64Ty, LineLoc.getRawEncoding()))); |
| 2264 |
} |
| 2265 |
} |
| 2266 |
|
| 2267 |
return llvm::MDNode::get(CGF.getLLVMContext(), Locs); |
| 2268 |
} |
| 2269 |
|
| 2270 |
static void UpdateAsmCallInst(llvm::CallBase &Result, bool HasSideEffect, |
| 2271 |
bool HasUnwindClobber, bool ReadOnly, |
| 2272 |
bool ReadNone, bool NoMerge, const AsmStmt &S, |
| 2273 |
const std::vector<llvm::Type *> &ResultRegTypes, |
| 2274 |
const std::vector<llvm::Type *> &ArgElemTypes, |
| 2275 |
CodeGenFunction &CGF, |
| 2276 |
std::vector<llvm::Value *> &RegResults) { |
| 2277 |
if (!HasUnwindClobber) |
| 2278 |
Result.addFnAttr(llvm::Attribute::NoUnwind); |
| 2279 |
|
| 2280 |
if (NoMerge) |
| 2281 |
Result.addFnAttr(llvm::Attribute::NoMerge); |
| 2282 |
// Attach readnone and readonly attributes. |
| 2283 |
if (!HasSideEffect) { |
| 2284 |
if (ReadNone) |
| 2285 |
Result.setDoesNotAccessMemory(); |
| 2286 |
else if (ReadOnly) |
| 2287 |
Result.setOnlyReadsMemory(); |
| 2288 |
} |
| 2289 |
|
| 2290 |
// Add elementtype attribute for indirect constraints. |
| 2291 |
for (auto Pair : llvm::enumerate(ArgElemTypes)) { |
| 2292 |
if (Pair.value()) { |
| 2293 |
auto Attr = llvm::Attribute::get( |
| 2294 |
CGF.getLLVMContext(), llvm::Attribute::ElementType, Pair.value()); |
| 2295 |
Result.addParamAttr(Pair.index(), Attr); |
| 2296 |
} |
| 2297 |
} |
| 2298 |
|
| 2299 |
// Slap the source location of the inline asm into a !srcloc metadata on the |
| 2300 |
// call. |
| 2301 |
if (const auto *gccAsmStmt = dyn_cast<GCCAsmStmt>(&S)) |
| 2302 |
Result.setMetadata("srcloc", |
| 2303 |
getAsmSrcLocInfo(gccAsmStmt->getAsmString(), CGF)); |
| 2304 |
else { |
| 2305 |
// At least put the line number on MS inline asm blobs. |
| 2306 |
llvm::Constant *Loc = |
| 2307 |
llvm::ConstantInt::get(CGF.Int64Ty, S.getAsmLoc().getRawEncoding()); |
| 2308 |
Result.setMetadata("srcloc", |
| 2309 |
llvm::MDNode::get(CGF.getLLVMContext(), |
| 2310 |
llvm::ConstantAsMetadata::get(Loc))); |
| 2311 |
} |
| 2312 |
|
| 2313 |
if (CGF.getLangOpts().assumeFunctionsAreConvergent()) |
| 2314 |
// Conservatively, mark all inline asm blocks in CUDA or OpenCL as |
| 2315 |
// convergent (meaning, they may call an intrinsically convergent op, such |
| 2316 |
// as bar.sync, and so can't have certain optimizations applied around |
| 2317 |
// them). |
| 2318 |
Result.addFnAttr(llvm::Attribute::Convergent); |
| 2319 |
// Extract all of the register value results from the asm. |
| 2320 |
if (ResultRegTypes.size() == 1) { |
| 2321 |
RegResults.push_back(&Result); |
| 2322 |
} else { |
| 2323 |
for (unsigned i = 0, e = ResultRegTypes.size(); i != e; ++i) { |
| 2324 |
llvm::Value *Tmp = CGF.Builder.CreateExtractValue(&Result, i, "asmresult"); |
| 2325 |
RegResults.push_back(Tmp); |
| 2326 |
} |
| 2327 |
} |
| 2328 |
} |
| 2329 |
|
| 2330 |
void CodeGenFunction::EmitAsmStmt(const AsmStmt &S) { |
| 2331 |
// Pop all cleanup blocks at the end of the asm statement. |
| 2332 |
CodeGenFunction::RunCleanupsScope Cleanups(*this); |
| 2333 |
|
| 2334 |
// Assemble the final asm string. |
| 2335 |
std::string AsmString = S.generateAsmString(getContext()); |
| 2336 |
|
| 2337 |
// Get all the output and input constraints together. |
| 2338 |
SmallVector<TargetInfo::ConstraintInfo, 4> OutputConstraintInfos; |
| 2339 |
SmallVector<TargetInfo::ConstraintInfo, 4> InputConstraintInfos; |
| 2340 |
|
| 2341 |
for (unsigned i = 0, e = S.getNumOutputs(); i != e; i++) { |
| 2342 |
StringRef Name; |
| 2343 |
if (const GCCAsmStmt *GAS = dyn_cast<GCCAsmStmt>(&S)) |
| 2344 |
Name = GAS->getOutputName(i); |
| 2345 |
TargetInfo::ConstraintInfo Info(S.getOutputConstraint(i), Name); |
| 2346 |
bool IsValid = getTarget().validateOutputConstraint(Info); (void)IsValid; |
| 2347 |
assert(IsValid && "Failed to parse output constraint"); |
| 2348 |
OutputConstraintInfos.push_back(Info); |
| 2349 |
} |
| 2350 |
|
| 2351 |
for (unsigned i = 0, e = S.getNumInputs(); i != e; i++) { |
| 2352 |
StringRef Name; |
| 2353 |
if (const GCCAsmStmt *GAS = dyn_cast<GCCAsmStmt>(&S)) |
| 2354 |
Name = GAS->getInputName(i); |
| 2355 |
TargetInfo::ConstraintInfo Info(S.getInputConstraint(i), Name); |
| 2356 |
bool IsValid = |
| 2357 |
getTarget().validateInputConstraint(OutputConstraintInfos, Info); |
| 2358 |
assert(IsValid && "Failed to parse input constraint"); (void)IsValid; |
| 2359 |
InputConstraintInfos.push_back(Info); |
| 2360 |
} |
| 2361 |
|
| 2362 |
std::string Constraints; |
| 2363 |
|
| 2364 |
std::vector<LValue> ResultRegDests; |
| 2365 |
std::vector<QualType> ResultRegQualTys; |
| 2366 |
std::vector<llvm::Type *> ResultRegTypes; |
| 2367 |
std::vector<llvm::Type *> ResultTruncRegTypes; |
| 2368 |
std::vector<llvm::Type *> ArgTypes; |
| 2369 |
std::vector<llvm::Type *> ArgElemTypes; |
| 2370 |
std::vector<llvm::Value*> Args; |
| 2371 |
llvm::BitVector ResultTypeRequiresCast; |
| 2372 |
llvm::BitVector ResultRegIsFlagReg; |
| 2373 |
|
| 2374 |
// Keep track of inout constraints. |
| 2375 |
std::string InOutConstraints; |
| 2376 |
std::vector<llvm::Value*> InOutArgs; |
| 2377 |
std::vector<llvm::Type*> InOutArgTypes; |
| 2378 |
std::vector<llvm::Type*> InOutArgElemTypes; |
| 2379 |
|
| 2380 |
// Keep track of out constraints for tied input operand. |
| 2381 |
std::vector<std::string> OutputConstraints; |
| 2382 |
|
| 2383 |
// Keep track of defined physregs. |
| 2384 |
llvm::SmallSet<std::string, 8> PhysRegOutputs; |
| 2385 |
|
| 2386 |
// An inline asm can be marked readonly if it meets the following conditions: |
| 2387 |
// - it doesn't have any sideeffects |
| 2388 |
// - it doesn't clobber memory |
| 2389 |
// - it doesn't return a value by-reference |
| 2390 |
// It can be marked readnone if it doesn't have any input memory constraints |
| 2391 |
// in addition to meeting the conditions listed above. |
| 2392 |
bool ReadOnly = true, ReadNone = true; |
| 2393 |
|
| 2394 |
for (unsigned i = 0, e = S.getNumOutputs(); i != e; i++) { |
| 2395 |
TargetInfo::ConstraintInfo &Info = OutputConstraintInfos[i]; |
| 2396 |
|
| 2397 |
// Simplify the output constraint. |
| 2398 |
std::string OutputConstraint(S.getOutputConstraint(i)); |
| 2399 |
OutputConstraint = SimplifyConstraint(OutputConstraint.c_str() + 1, |
| 2400 |
getTarget(), &OutputConstraintInfos); |
| 2401 |
|
| 2402 |
const Expr *OutExpr = S.getOutputExpr(i); |
| 2403 |
OutExpr = OutExpr->IgnoreParenNoopCasts(getContext()); |
| 2404 |
|
| 2405 |
std::string GCCReg; |
| 2406 |
OutputConstraint = AddVariableConstraints(OutputConstraint, *OutExpr, |
| 2407 |
getTarget(), CGM, S, |
| 2408 |
Info.earlyClobber(), |
| 2409 |
&GCCReg); |
| 2410 |
// Give an error on multiple outputs to same physreg. |
| 2411 |
if (!GCCReg.empty() && !PhysRegOutputs.insert(GCCReg).second) |
| 2412 |
CGM.Error(S.getAsmLoc(), "multiple outputs to hard register: " + GCCReg); |
| 2413 |
|
| 2414 |
OutputConstraints.push_back(OutputConstraint); |
| 2415 |
LValue Dest = EmitLValue(OutExpr); |
| 2416 |
if (!Constraints.empty()) |
| 2417 |
Constraints += ','; |
| 2418 |
|
| 2419 |
// If this is a register output, then make the inline asm return it |
| 2420 |
// by-value. If this is a memory result, return the value by-reference. |
| 2421 |
QualType QTy = OutExpr->getType(); |
| 2422 |
const bool IsScalarOrAggregate = hasScalarEvaluationKind(QTy) || |
| 2423 |
hasAggregateEvaluationKind(QTy); |
| 2424 |
if (!Info.allowsMemory() && IsScalarOrAggregate) { |
| 2425 |
|
| 2426 |
Constraints += "=" + OutputConstraint; |
| 2427 |
ResultRegQualTys.push_back(QTy); |
| 2428 |
ResultRegDests.push_back(Dest); |
| 2429 |
|
| 2430 |
bool IsFlagReg = llvm::StringRef(OutputConstraint).startswith("{@cc"); |
| 2431 |
ResultRegIsFlagReg.push_back(IsFlagReg); |
| 2432 |
|
| 2433 |
llvm::Type *Ty = ConvertTypeForMem(QTy); |
| 2434 |
const bool RequiresCast = Info.allowsRegister() && |
| 2435 |
(getTargetHooks().isScalarizableAsmOperand(*this, Ty) || |
| 2436 |
Ty->isAggregateType()); |
| 2437 |
|
| 2438 |
ResultTruncRegTypes.push_back(Ty); |
| 2439 |
ResultTypeRequiresCast.push_back(RequiresCast); |
| 2440 |
|
| 2441 |
if (RequiresCast) { |
| 2442 |
unsigned Size = getContext().getTypeSize(QTy); |
| 2443 |
Ty = llvm::IntegerType::get(getLLVMContext(), Size); |
| 2444 |
} |
| 2445 |
ResultRegTypes.push_back(Ty); |
| 2446 |
// If this output is tied to an input, and if the input is larger, then |
| 2447 |
// we need to set the actual result type of the inline asm node to be the |
| 2448 |
// same as the input type. |
| 2449 |
if (Info.hasMatchingInput()) { |
| 2450 |
unsigned InputNo; |
| 2451 |
for (InputNo = 0; InputNo != S.getNumInputs(); ++InputNo) { |
| 2452 |
TargetInfo::ConstraintInfo &Input = InputConstraintInfos[InputNo]; |
| 2453 |
if (Input.hasTiedOperand() && Input.getTiedOperand() == i) |
| 2454 |
break; |
| 2455 |
} |
| 2456 |
assert(InputNo != S.getNumInputs() && "Didn't find matching input!"); |
| 2457 |
|
| 2458 |
QualType InputTy = S.getInputExpr(InputNo)->getType(); |
| 2459 |
QualType OutputType = OutExpr->getType(); |
| 2460 |
|
| 2461 |
uint64_t InputSize = getContext().getTypeSize(InputTy); |
| 2462 |
if (getContext().getTypeSize(OutputType) < InputSize) { |
| 2463 |
// Form the asm to return the value as a larger integer or fp type. |
| 2464 |
ResultRegTypes.back() = ConvertType(InputTy); |
| 2465 |
} |
| 2466 |
} |
| 2467 |
if (llvm::Type* AdjTy = |
| 2468 |
getTargetHooks().adjustInlineAsmType(*this, OutputConstraint, |
| 2469 |
ResultRegTypes.back())) |
| 2470 |
ResultRegTypes.back() = AdjTy; |
| 2471 |
else { |
| 2472 |
CGM.getDiags().Report(S.getAsmLoc(), |
| 2473 |
diag::err_asm_invalid_type_in_input) |
| 2474 |
<< OutExpr->getType() << OutputConstraint; |
| 2475 |
} |
| 2476 |
|
| 2477 |
// Update largest vector width for any vector types. |
| 2478 |
if (auto *VT = dyn_cast<llvm::VectorType>(ResultRegTypes.back())) |
| 2479 |
LargestVectorWidth = |
| 2480 |
std::max((uint64_t)LargestVectorWidth, |
| 2481 |
VT->getPrimitiveSizeInBits().getKnownMinValue()); |
| 2482 |
} else { |
| 2483 |
Address DestAddr = Dest.getAddress(*this); |
| 2484 |
// Matrix types in memory are represented by arrays, but accessed through |
| 2485 |
// vector pointers, with the alignment specified on the access operation. |
| 2486 |
// For inline assembly, update pointer arguments to use vector pointers. |
| 2487 |
// Otherwise there will be a mis-match if the matrix is also an |
| 2488 |
// input-argument which is represented as vector. |
| 2489 |
if (isa<MatrixType>(OutExpr->getType().getCanonicalType())) |
| 2490 |
DestAddr = Builder.CreateElementBitCast( |
| 2491 |
DestAddr, ConvertType(OutExpr->getType())); |
| 2492 |
|
| 2493 |
ArgTypes.push_back(DestAddr.getType()); |
| 2494 |
ArgElemTypes.push_back(DestAddr.getElementType()); |
| 2495 |
Args.push_back(DestAddr.getPointer()); |
| 2496 |
Constraints += "=*"; |
| 2497 |
Constraints += OutputConstraint; |
| 2498 |
ReadOnly = ReadNone = false; |
| 2499 |
} |
| 2500 |
|
| 2501 |
if (Info.isReadWrite()) { |
| 2502 |
InOutConstraints += ','; |
| 2503 |
|
| 2504 |
const Expr *InputExpr = S.getOutputExpr(i); |
| 2505 |
llvm::Value *Arg; |
| 2506 |
llvm::Type *ArgElemType; |
| 2507 |
std::tie(Arg, ArgElemType) = EmitAsmInputLValue( |
| 2508 |
Info, Dest, InputExpr->getType(), InOutConstraints, |
| 2509 |
InputExpr->getExprLoc()); |
| 2510 |
|
| 2511 |
if (llvm::Type* AdjTy = |
| 2512 |
getTargetHooks().adjustInlineAsmType(*this, OutputConstraint, |
| 2513 |
Arg->getType())) |
| 2514 |
Arg = Builder.CreateBitCast(Arg, AdjTy); |
| 2515 |
|
| 2516 |
// Update largest vector width for any vector types. |
| 2517 |
if (auto *VT = dyn_cast<llvm::VectorType>(Arg->getType())) |
| 2518 |
LargestVectorWidth = |
| 2519 |
std::max((uint64_t)LargestVectorWidth, |
| 2520 |
VT->getPrimitiveSizeInBits().getKnownMinValue()); |
| 2521 |
// Only tie earlyclobber physregs. |
| 2522 |
if (Info.allowsRegister() && (GCCReg.empty() || Info.earlyClobber())) |
| 2523 |
InOutConstraints += llvm::utostr(i); |
| 2524 |
else |
| 2525 |
InOutConstraints += OutputConstraint; |
| 2526 |
|
| 2527 |
InOutArgTypes.push_back(Arg->getType()); |
| 2528 |
InOutArgElemTypes.push_back(ArgElemType); |
| 2529 |
InOutArgs.push_back(Arg); |
| 2530 |
} |
| 2531 |
} |
| 2532 |
|
| 2533 |
// If this is a Microsoft-style asm blob, store the return registers (EAX:EDX) |
| 2534 |
// to the return value slot. Only do this when returning in registers. |
| 2535 |
if (isa<MSAsmStmt>(&S)) { |
| 2536 |
const ABIArgInfo &RetAI = CurFnInfo->getReturnInfo(); |
| 2537 |
if (RetAI.isDirect() || RetAI.isExtend()) { |
| 2538 |
// Make a fake lvalue for the return value slot. |
| 2539 |
LValue ReturnSlot = MakeAddrLValueWithoutTBAA(ReturnValue, FnRetTy); |
| 2540 |
CGM.getTargetCodeGenInfo().addReturnRegisterOutputs( |
| 2541 |
*this, ReturnSlot, Constraints, ResultRegTypes, ResultTruncRegTypes, |
| 2542 |
ResultRegDests, AsmString, S.getNumOutputs()); |
| 2543 |
SawAsmBlock = true; |
| 2544 |
} |
| 2545 |
} |
| 2546 |
|
| 2547 |
for (unsigned i = 0, e = S.getNumInputs(); i != e; i++) { |
| 2548 |
const Expr *InputExpr = S.getInputExpr(i); |
| 2549 |
|
| 2550 |
TargetInfo::ConstraintInfo &Info = InputConstraintInfos[i]; |
| 2551 |
|
| 2552 |
if (Info.allowsMemory()) |
| 2553 |
ReadNone = false; |
| 2554 |
|
| 2555 |
if (!Constraints.empty()) |
| 2556 |
Constraints += ','; |
| 2557 |
|
| 2558 |
// Simplify the input constraint. |
| 2559 |
std::string InputConstraint(S.getInputConstraint(i)); |
| 2560 |
InputConstraint = SimplifyConstraint(InputConstraint.c_str(), getTarget(), |
| 2561 |
&OutputConstraintInfos); |
| 2562 |
|
| 2563 |
InputConstraint = AddVariableConstraints( |
| 2564 |
InputConstraint, *InputExpr->IgnoreParenNoopCasts(getContext()), |
| 2565 |
getTarget(), CGM, S, false /* No EarlyClobber */); |
| 2566 |
|
| 2567 |
std::string ReplaceConstraint (InputConstraint); |
| 2568 |
llvm::Value *Arg; |
| 2569 |
llvm::Type *ArgElemType; |
| 2570 |
std::tie(Arg, ArgElemType) = EmitAsmInput(Info, InputExpr, Constraints); |
| 2571 |
|
| 2572 |
// If this input argument is tied to a larger output result, extend the |
| 2573 |
// input to be the same size as the output. The LLVM backend wants to see |
| 2574 |
// the input and output of a matching constraint be the same size. Note |
| 2575 |
// that GCC does not define what the top bits are here. We use zext because |
| 2576 |
// that is usually cheaper, but LLVM IR should really get an anyext someday. |
| 2577 |
if (Info.hasTiedOperand()) { |
| 2578 |
unsigned Output = Info.getTiedOperand(); |
| 2579 |
QualType OutputType = S.getOutputExpr(Output)->getType(); |
| 2580 |
QualType InputTy = InputExpr->getType(); |
| 2581 |
|
| 2582 |
if (getContext().getTypeSize(OutputType) > |
| 2583 |
getContext().getTypeSize(InputTy)) { |
| 2584 |
// Use ptrtoint as appropriate so that we can do our extension. |
| 2585 |
if (isa<llvm::PointerType>(Arg->getType())) |
| 2586 |
Arg = Builder.CreatePtrToInt(Arg, IntPtrTy); |
| 2587 |
llvm::Type *OutputTy = ConvertType(OutputType); |
| 2588 |
if (isa<llvm::IntegerType>(OutputTy)) |
| 2589 |
Arg = Builder.CreateZExt(Arg, OutputTy); |
| 2590 |
else if (isa<llvm::PointerType>(OutputTy)) |
| 2591 |
Arg = Builder.CreateZExt(Arg, IntPtrTy); |
| 2592 |
else if (OutputTy->isFloatingPointTy()) |
| 2593 |
Arg = Builder.CreateFPExt(Arg, OutputTy); |
| 2594 |
} |
| 2595 |
// Deal with the tied operands' constraint code in adjustInlineAsmType. |
| 2596 |
ReplaceConstraint = OutputConstraints[Output]; |
| 2597 |
} |
| 2598 |
if (llvm::Type* AdjTy = |
| 2599 |
getTargetHooks().adjustInlineAsmType(*this, ReplaceConstraint, |
| 2600 |
Arg->getType())) |
| 2601 |
Arg = Builder.CreateBitCast(Arg, AdjTy); |
| 2602 |
else |
| 2603 |
CGM.getDiags().Report(S.getAsmLoc(), diag::err_asm_invalid_type_in_input) |
| 2604 |
<< InputExpr->getType() << InputConstraint; |
| 2605 |
|
| 2606 |
// Update largest vector width for any vector types. |
| 2607 |
if (auto *VT = dyn_cast<llvm::VectorType>(Arg->getType())) |
| 2608 |
LargestVectorWidth = |
| 2609 |
std::max((uint64_t)LargestVectorWidth, |
| 2610 |
VT->getPrimitiveSizeInBits().getKnownMinValue()); |
| 2611 |
|
| 2612 |
ArgTypes.push_back(Arg->getType()); |
| 2613 |
ArgElemTypes.push_back(ArgElemType); |
| 2614 |
Args.push_back(Arg); |
| 2615 |
Constraints += InputConstraint; |
| 2616 |
} |
| 2617 |
|
| 2618 |
// Append the "input" part of inout constraints. |
| 2619 |
for (unsigned i = 0, e = InOutArgs.size(); i != e; i++) { |
| 2620 |
ArgTypes.push_back(InOutArgTypes[i]); |
| 2621 |
ArgElemTypes.push_back(InOutArgElemTypes[i]); |
| 2622 |
Args.push_back(InOutArgs[i]); |
| 2623 |
} |
| 2624 |
Constraints += InOutConstraints; |
| 2625 |
|
| 2626 |
// Labels |
| 2627 |
SmallVector<llvm::BasicBlock *, 16> Transfer; |
| 2628 |
llvm::BasicBlock *Fallthrough = nullptr; |
| 2629 |
bool IsGCCAsmGoto = false; |
| 2630 |
if (const auto *GS = dyn_cast<GCCAsmStmt>(&S)) { |
| 2631 |
IsGCCAsmGoto = GS->isAsmGoto(); |
| 2632 |
if (IsGCCAsmGoto) { |
| 2633 |
for (const auto *E : GS->labels()) { |
| 2634 |
JumpDest Dest = getJumpDestForLabel(E->getLabel()); |
| 2635 |
Transfer.push_back(Dest.getBlock()); |
| 2636 |
if (!Constraints.empty()) |
| 2637 |
Constraints += ','; |
| 2638 |
Constraints += "!i"; |
| 2639 |
} |
| 2640 |
Fallthrough = createBasicBlock("asm.fallthrough"); |
| 2641 |
} |
| 2642 |
} |
| 2643 |
|
| 2644 |
bool HasUnwindClobber = false; |
| 2645 |
|
| 2646 |
// Clobbers |
| 2647 |
for (unsigned i = 0, e = S.getNumClobbers(); i != e; i++) { |
| 2648 |
StringRef Clobber = S.getClobber(i); |
| 2649 |
|
| 2650 |
if (Clobber == "memory") |
| 2651 |
ReadOnly = ReadNone = false; |
| 2652 |
else if (Clobber == "unwind") { |
| 2653 |
HasUnwindClobber = true; |
| 2654 |
continue; |
| 2655 |
} else if (Clobber != "cc") { |
| 2656 |
Clobber = getTarget().getNormalizedGCCRegisterName(Clobber); |
| 2657 |
if (CGM.getCodeGenOpts().StackClashProtector && |
| 2658 |
getTarget().isSPRegName(Clobber)) { |
| 2659 |
CGM.getDiags().Report(S.getAsmLoc(), |
| 2660 |
diag::warn_stack_clash_protection_inline_asm); |
| 2661 |
} |
| 2662 |
} |
| 2663 |
|
| 2664 |
if (isa<MSAsmStmt>(&S)) { |
| 2665 |
if (Clobber == "eax" || Clobber == "edx") { |
| 2666 |
if (Constraints.find("=&A") != std::string::npos) |
| 2667 |
continue; |
| 2668 |
std::string::size_type position1 = |
| 2669 |
Constraints.find("={" + Clobber.str() + "}"); |
| 2670 |
if (position1 != std::string::npos) { |
| 2671 |
Constraints.insert(position1 + 1, "&"); |
| 2672 |
continue; |
| 2673 |
} |
| 2674 |
std::string::size_type position2 = Constraints.find("=A"); |
| 2675 |
if (position2 != std::string::npos) { |
| 2676 |
Constraints.insert(position2 + 1, "&"); |
| 2677 |
continue; |
| 2678 |
} |
| 2679 |
} |
| 2680 |
} |
| 2681 |
if (!Constraints.empty()) |
| 2682 |
Constraints += ','; |
| 2683 |
|
| 2684 |
Constraints += "~{"; |
| 2685 |
Constraints += Clobber; |
| 2686 |
Constraints += '}'; |
| 2687 |
} |
| 2688 |
|
| 2689 |
assert(!(HasUnwindClobber && IsGCCAsmGoto) && |
| 2690 |
"unwind clobber can't be used with asm goto"); |
| 2691 |
|
| 2692 |
// Add machine specific clobbers |
| 2693 |
std::string MachineClobbers = getTarget().getClobbers(); |
| 2694 |
if (!MachineClobbers.empty()) { |
| 2695 |
if (!Constraints.empty()) |
| 2696 |
Constraints += ','; |
| 2697 |
Constraints += MachineClobbers; |
| 2698 |
} |
| 2699 |
|
| 2700 |
llvm::Type *ResultType; |
| 2701 |
if (ResultRegTypes.empty()) |
| 2702 |
ResultType = VoidTy; |
| 2703 |
else if (ResultRegTypes.size() == 1) |
| 2704 |
ResultType = ResultRegTypes[0]; |
| 2705 |
else |
| 2706 |
ResultType = llvm::StructType::get(getLLVMContext(), ResultRegTypes); |
| 2707 |
|
| 2708 |
llvm::FunctionType *FTy = |
| 2709 |
llvm::FunctionType::get(ResultType, ArgTypes, false); |
| 2710 |
|
| 2711 |
bool HasSideEffect = S.isVolatile() || S.getNumOutputs() == 0; |
| 2712 |
|
| 2713 |
llvm::InlineAsm::AsmDialect GnuAsmDialect = |
| 2714 |
CGM.getCodeGenOpts().getInlineAsmDialect() == CodeGenOptions::IAD_ATT |
| 2715 |
? llvm::InlineAsm::AD_ATT |
| 2716 |
: llvm::InlineAsm::AD_Intel; |
| 2717 |
llvm::InlineAsm::AsmDialect AsmDialect = isa<MSAsmStmt>(&S) ? |
| 2718 |
llvm::InlineAsm::AD_Intel : GnuAsmDialect; |
| 2719 |
|
| 2720 |
llvm::InlineAsm *IA = llvm::InlineAsm::get( |
| 2721 |
FTy, AsmString, Constraints, HasSideEffect, |
| 2722 |
/* IsAlignStack */ false, AsmDialect, HasUnwindClobber); |
| 2723 |
std::vector<llvm::Value*> RegResults; |
| 2724 |
if (IsGCCAsmGoto) { |
| 2725 |
llvm::CallBrInst *Result = |
| 2726 |
Builder.CreateCallBr(IA, Fallthrough, Transfer, Args); |
| 2727 |
EmitBlock(Fallthrough); |
| 2728 |
UpdateAsmCallInst(cast<llvm::CallBase>(*Result), HasSideEffect, false, |
| 2729 |
ReadOnly, ReadNone, InNoMergeAttributedStmt, S, |
| 2730 |
ResultRegTypes, ArgElemTypes, *this, RegResults); |
| 2731 |
} else if (HasUnwindClobber) { |
| 2732 |
llvm::CallBase *Result = EmitCallOrInvoke(IA, Args, ""); |
| 2733 |
UpdateAsmCallInst(*Result, HasSideEffect, true, ReadOnly, ReadNone, |
| 2734 |
InNoMergeAttributedStmt, S, ResultRegTypes, ArgElemTypes, |
| 2735 |
*this, RegResults); |
| 2736 |
} else { |
| 2737 |
llvm::CallInst *Result = |
| 2738 |
Builder.CreateCall(IA, Args, getBundlesForFunclet(IA)); |
| 2739 |
UpdateAsmCallInst(cast<llvm::CallBase>(*Result), HasSideEffect, false, |
| 2740 |
ReadOnly, ReadNone, InNoMergeAttributedStmt, S, |
| 2741 |
ResultRegTypes, ArgElemTypes, *this, RegResults); |
| 2742 |
} |
| 2743 |
|
| 2744 |
assert(RegResults.size() == ResultRegTypes.size()); |
| 2745 |
assert(RegResults.size() == ResultTruncRegTypes.size()); |
| 2746 |
assert(RegResults.size() == ResultRegDests.size()); |
| 2747 |
// ResultRegDests can be also populated by addReturnRegisterOutputs() above, |
| 2748 |
// in which case its size may grow. |
| 2749 |
assert(ResultTypeRequiresCast.size() <= ResultRegDests.size()); |
| 2750 |
assert(ResultRegIsFlagReg.size() <= ResultRegDests.size()); |
| 2751 |
for (unsigned i = 0, e = RegResults.size(); i != e; ++i) { |
| 2752 |
llvm::Value *Tmp = RegResults[i]; |
| 2753 |
llvm::Type *TruncTy = ResultTruncRegTypes[i]; |
| 2754 |
|
| 2755 |
if ((i < ResultRegIsFlagReg.size()) && ResultRegIsFlagReg[i]) { |
| 2756 |
// Target must guarantee the Value `Tmp` here is lowered to a boolean |
| 2757 |
// value. |
| 2758 |
llvm::Constant *Two = llvm::ConstantInt::get(Tmp->getType(), 2); |
| 2759 |
llvm::Value *IsBooleanValue = |
| 2760 |
Builder.CreateCmp(llvm::CmpInst::ICMP_ULT, Tmp, Two); |
| 2761 |
llvm::Function *FnAssume = CGM.getIntrinsic(llvm::Intrinsic::assume); |
| 2762 |
Builder.CreateCall(FnAssume, IsBooleanValue); |
| 2763 |
} |
| 2764 |
|
| 2765 |
// If the result type of the LLVM IR asm doesn't match the result type of |
| 2766 |
// the expression, do the conversion. |
| 2767 |
if (ResultRegTypes[i] != ResultTruncRegTypes[i]) { |
| 2768 |
|
| 2769 |
// Truncate the integer result to the right size, note that TruncTy can be |
| 2770 |
// a pointer. |
| 2771 |
if (TruncTy->isFloatingPointTy()) |
| 2772 |
Tmp = Builder.CreateFPTrunc(Tmp, TruncTy); |
| 2773 |
else if (TruncTy->isPointerTy() && Tmp->getType()->isIntegerTy()) { |
| 2774 |
uint64_t ResSize = CGM.getDataLayout().getTypeSizeInBits(TruncTy); |
| 2775 |
Tmp = Builder.CreateTrunc(Tmp, |
| 2776 |
llvm::IntegerType::get(getLLVMContext(), (unsigned)ResSize)); |
| 2777 |
Tmp = Builder.CreateIntToPtr(Tmp, TruncTy); |
| 2778 |
} else if (Tmp->getType()->isPointerTy() && TruncTy->isIntegerTy()) { |
| 2779 |
uint64_t TmpSize =CGM.getDataLayout().getTypeSizeInBits(Tmp->getType()); |
| 2780 |
Tmp = Builder.CreatePtrToInt(Tmp, |
| 2781 |
llvm::IntegerType::get(getLLVMContext(), (unsigned)TmpSize)); |
| 2782 |
Tmp = Builder.CreateTrunc(Tmp, TruncTy); |
| 2783 |
} else if (TruncTy->isIntegerTy()) { |
| 2784 |
Tmp = Builder.CreateZExtOrTrunc(Tmp, TruncTy); |
| 2785 |
} else if (TruncTy->isVectorTy()) { |
| 2786 |
Tmp = Builder.CreateBitCast(Tmp, TruncTy); |
| 2787 |
} |
| 2788 |
} |
| 2789 |
|
| 2790 |
LValue Dest = ResultRegDests[i]; |
| 2791 |
// ResultTypeRequiresCast elements correspond to the first |
| 2792 |
// ResultTypeRequiresCast.size() elements of RegResults. |
| 2793 |
if ((i < ResultTypeRequiresCast.size()) && ResultTypeRequiresCast[i]) { |
| 2794 |
unsigned Size = getContext().getTypeSize(ResultRegQualTys[i]); |
| 2795 |
Address A = Builder.CreateElementBitCast(Dest.getAddress(*this), |
| 2796 |
ResultRegTypes[i]); |
| 2797 |
if (getTargetHooks().isScalarizableAsmOperand(*this, TruncTy)) { |
| 2798 |
Builder.CreateStore(Tmp, A); |
| 2799 |
continue; |
| 2800 |
} |
| 2801 |
|
| 2802 |
QualType Ty = getContext().getIntTypeForBitwidth(Size, /*Signed*/ false); |
| 2803 |
if (Ty.isNull()) { |
| 2804 |
const Expr *OutExpr = S.getOutputExpr(i); |
| 2805 |
CGM.getDiags().Report(OutExpr->getExprLoc(), |
| 2806 |
diag::err_store_value_to_reg); |
| 2807 |
return; |
| 2808 |
} |
| 2809 |
Dest = MakeAddrLValue(A, Ty); |
| 2810 |
} |
| 2811 |
EmitStoreThroughLValue(RValue::get(Tmp), Dest); |
| 2812 |
} |
| 2813 |
} |
| 2814 |
|
| 2815 |
LValue CodeGenFunction::InitCapturedStruct(const CapturedStmt &S) { |
| 2816 |
const RecordDecl *RD = S.getCapturedRecordDecl(); |
| 2817 |
QualType RecordTy = getContext().getRecordType(RD); |
| 2818 |
|
| 2819 |
// Initialize the captured struct. |
| 2820 |
LValue SlotLV = |
| 2821 |
MakeAddrLValue(CreateMemTemp(RecordTy, "agg.captured"), RecordTy); |
| 2822 |
|
| 2823 |
RecordDecl::field_iterator CurField = RD->field_begin(); |
| 2824 |
for (CapturedStmt::const_capture_init_iterator I = S.capture_init_begin(), |
| 2825 |
E = S.capture_init_end(); |
| 2826 |
I != E; ++I, ++CurField) { |
| 2827 |
LValue LV = EmitLValueForFieldInitialization(SlotLV, *CurField); |
| 2828 |
if (CurField->hasCapturedVLAType()) { |
| 2829 |
EmitLambdaVLACapture(CurField->getCapturedVLAType(), LV); |
| 2830 |
} else { |
| 2831 |
EmitInitializerForField(*CurField, LV, *I); |
| 2832 |
} |
| 2833 |
} |
| 2834 |
|
| 2835 |
return SlotLV; |
| 2836 |
} |
| 2837 |
|
| 2838 |
/// Generate an outlined function for the body of a CapturedStmt, store any |
| 2839 |
/// captured variables into the captured struct, and call the outlined function. |
| 2840 |
llvm::Function * |
| 2841 |
CodeGenFunction::EmitCapturedStmt(const CapturedStmt &S, CapturedRegionKind K) { |
| 2842 |
LValue CapStruct = InitCapturedStruct(S); |
| 2843 |
|
| 2844 |
// Emit the CapturedDecl |
| 2845 |
CodeGenFunction CGF(CGM, true); |
| 2846 |
CGCapturedStmtRAII CapInfoRAII(CGF, new CGCapturedStmtInfo(S, K)); |
| 2847 |
llvm::Function *F = CGF.GenerateCapturedStmtFunction(S); |
| 2848 |
delete CGF.CapturedStmtInfo; |
| 2849 |
|
| 2850 |
// Emit call to the helper function. |
| 2851 |
EmitCallOrInvoke(F, CapStruct.getPointer(*this)); |
| 2852 |
|
| 2853 |
return F; |
| 2854 |
} |
| 2855 |
|
| 2856 |
Address CodeGenFunction::GenerateCapturedStmtArgument(const CapturedStmt &S) { |
| 2857 |
LValue CapStruct = InitCapturedStruct(S); |
| 2858 |
return CapStruct.getAddress(*this); |
| 2859 |
} |
| 2860 |
|
| 2861 |
/// Creates the outlined function for a CapturedStmt. |
| 2862 |
llvm::Function * |
| 2863 |
CodeGenFunction::GenerateCapturedStmtFunction(const CapturedStmt &S) { |
| 2864 |
assert(CapturedStmtInfo && |
| 2865 |
"CapturedStmtInfo should be set when generating the captured function"); |
| 2866 |
const CapturedDecl *CD = S.getCapturedDecl(); |
| 2867 |
const RecordDecl *RD = S.getCapturedRecordDecl(); |
| 2868 |
SourceLocation Loc = S.getBeginLoc(); |
| 2869 |
assert(CD->hasBody() && "missing CapturedDecl body"); |
| 2870 |
|
| 2871 |
// Build the argument list. |
| 2872 |
ASTContext &Ctx = CGM.getContext(); |
| 2873 |
FunctionArgList Args; |
| 2874 |
Args.append(CD->param_begin(), CD->param_end()); |
| 2875 |
|
| 2876 |
// Create the function declaration. |
| 2877 |
const CGFunctionInfo &FuncInfo = |
| 2878 |
CGM.getTypes().arrangeBuiltinFunctionDeclaration(Ctx.VoidTy, Args); |
| 2879 |
llvm::FunctionType *FuncLLVMTy = CGM.getTypes().GetFunctionType(FuncInfo); |
| 2880 |
|
| 2881 |
llvm::Function *F = |
| 2882 |
llvm::Function::Create(FuncLLVMTy, llvm::GlobalValue::InternalLinkage, |
| 2883 |
CapturedStmtInfo->getHelperName(), &CGM.getModule()); |
| 2884 |
CGM.SetInternalFunctionAttributes(CD, F, FuncInfo); |
| 2885 |
if (CD->isNothrow()) |
| 2886 |
F->addFnAttr(llvm::Attribute::NoUnwind); |
| 2887 |
|
| 2888 |
// Generate the function. |
| 2889 |
StartFunction(CD, Ctx.VoidTy, F, FuncInfo, Args, CD->getLocation(), |
| 2890 |
CD->getBody()->getBeginLoc()); |
| 2891 |
// Set the context parameter in CapturedStmtInfo. |
| 2892 |
Address DeclPtr = GetAddrOfLocalVar(CD->getContextParam()); |
| 2893 |
CapturedStmtInfo->setContextValue(Builder.CreateLoad(DeclPtr)); |
| 2894 |
|
| 2895 |
// Initialize variable-length arrays. |
| 2896 |
LValue Base = MakeNaturalAlignAddrLValue(CapturedStmtInfo->getContextValue(), |
| 2897 |
Ctx.getTagDeclType(RD)); |
| 2898 |
for (auto *FD : RD->fields()) { |
| 2899 |
if (FD->hasCapturedVLAType()) { |
| 2900 |
auto *ExprArg = |
| 2901 |
EmitLoadOfLValue(EmitLValueForField(Base, FD), S.getBeginLoc()) |
| 2902 |
.getScalarVal(); |
| 2903 |
auto VAT = FD->getCapturedVLAType(); |
| 2904 |
VLASizeMap[VAT->getSizeExpr()] = ExprArg; |
| 2905 |
} |
| 2906 |
} |
| 2907 |
|
| 2908 |
// If 'this' is captured, load it into CXXThisValue. |
| 2909 |
if (CapturedStmtInfo->isCXXThisExprCaptured()) { |
| 2910 |
FieldDecl *FD = CapturedStmtInfo->getThisFieldDecl(); |
| 2911 |
LValue ThisLValue = EmitLValueForField(Base, FD); |
| 2912 |
CXXThisValue = EmitLoadOfLValue(ThisLValue, Loc).getScalarVal(); |
| 2913 |
} |
| 2914 |
|
| 2915 |
PGO.assignRegionCounters(GlobalDecl(CD), F); |
| 2916 |
CapturedStmtInfo->EmitBody(*this, CD->getBody()); |
| 2917 |
FinishFunction(CD->getBodyRBrace()); |
| 2918 |
|
| 2919 |
return F; |
| 2920 |
} |