| 1 |
//=========- MemTransferLowerTest.cpp - MemTransferLower unit tests -=========// |
| 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 |
#include "llvm/Analysis/CGSCCPassManager.h" |
| 10 |
#include "llvm/Analysis/ScalarEvolution.h" |
| 11 |
#include "llvm/Analysis/TargetTransformInfo.h" |
| 12 |
#include "llvm/AsmParser/Parser.h" |
| 13 |
#include "llvm/IR/BasicBlock.h" |
| 14 |
#include "llvm/IR/Function.h" |
| 15 |
#include "llvm/IR/Instructions.h" |
| 16 |
#include "llvm/IR/IntrinsicInst.h" |
| 17 |
#include "llvm/IR/LLVMContext.h" |
| 18 |
#include "llvm/IR/Module.h" |
| 19 |
#include "llvm/IR/PassManager.h" |
| 20 |
#include "llvm/InitializePasses.h" |
| 21 |
#include "llvm/Passes/PassBuilder.h" |
| 22 |
#include "llvm/Support/Debug.h" |
| 23 |
#include "llvm/Support/SourceMgr.h" |
| 24 |
#include "llvm/Testing/Support/Error.h" |
| 25 |
#include "llvm/Transforms/Utils/LowerMemIntrinsics.h" |
| 26 |
#include "llvm/Transforms/Vectorize/LoopVectorize.h" |
| 27 |
|
| 28 |
#include "gtest/gtest-spi.h" |
| 29 |
#include "gtest/gtest.h" |
| 30 |
|
| 31 |
using namespace llvm; |
| 32 |
|
| 33 |
namespace { |
| 34 |
struct ForwardingPass : public PassInfoMixin<ForwardingPass> { |
| 35 |
template <typename T> ForwardingPass(T &&Arg) : Func(std::forward<T>(Arg)) {} |
| 36 |
|
| 37 |
PreservedAnalyses run(Function &F, FunctionAnalysisManager &FAM) { |
| 38 |
return Func(F, FAM); |
| 39 |
} |
| 40 |
|
| 41 |
std::function<PreservedAnalyses(Function &, FunctionAnalysisManager &)> Func; |
| 42 |
}; |
| 43 |
|
| 44 |
struct MemTransferLowerTest : public testing::Test { |
| 45 |
PassBuilder PB; |
| 46 |
LoopAnalysisManager LAM; |
| 47 |
FunctionAnalysisManager FAM; |
| 48 |
CGSCCAnalysisManager CGAM; |
| 49 |
ModuleAnalysisManager MAM; |
| 50 |
ModulePassManager MPM; |
| 51 |
LLVMContext Context; |
| 52 |
std::unique_ptr<Module> M; |
| 53 |
|
| 54 |
MemTransferLowerTest() { |
| 55 |
// Register all the basic analyses with the managers. |
| 56 |
PB.registerModuleAnalyses(MAM); |
| 57 |
PB.registerCGSCCAnalyses(CGAM); |
| 58 |
PB.registerFunctionAnalyses(FAM); |
| 59 |
PB.registerLoopAnalyses(LAM); |
| 60 |
PB.crossRegisterProxies(LAM, FAM, CGAM, MAM); |
| 61 |
} |
| 62 |
|
| 63 |
BasicBlock *getBasicBlockByName(Function &F, StringRef Name) const { |
| 64 |
for (BasicBlock &BB : F) { |
| 65 |
if (BB.getName() == Name) |
| 66 |
return &BB; |
| 67 |
} |
| 68 |
return nullptr; |
| 69 |
} |
| 70 |
|
| 71 |
Instruction *getInstructionByOpcode(BasicBlock &BB, unsigned Opcode, |
| 72 |
unsigned Number) const { |
| 73 |
unsigned CurrNumber = 0; |
| 74 |
for (Instruction &I : BB) |
| 75 |
if (I.getOpcode() == Opcode) { |
| 76 |
++CurrNumber; |
| 77 |
if (CurrNumber == Number) |
| 78 |
return &I; |
| 79 |
} |
| 80 |
return nullptr; |
| 81 |
} |
| 82 |
|
| 83 |
void ParseAssembly(const char *IR) { |
| 84 |
SMDiagnostic Error; |
| 85 |
M = parseAssemblyString(IR, Error, Context); |
| 86 |
std::string errMsg; |
| 87 |
raw_string_ostream os(errMsg); |
| 88 |
Error.print("", os); |
| 89 |
|
| 90 |
// A failure here means that the test itself is buggy. |
| 91 |
if (!M) |
| 92 |
report_fatal_error(os.str().c_str()); |
| 93 |
} |
| 94 |
}; |
| 95 |
|
| 96 |
// By semantics source and destination of llvm.memcpy.* intrinsic |
| 97 |
// are either equal or don't overlap. Once the intrinsic is lowered |
| 98 |
// to a loop it can be hard or impossible to reason about these facts. |
| 99 |
// For that reason expandMemCpyAsLoop is expected to explicitly mark |
| 100 |
// loads from source and stores to destination as not aliasing. |
| 101 |
TEST_F(MemTransferLowerTest, MemCpyKnownLength) { |
| 102 |
ParseAssembly("declare void @llvm.memcpy.p0i8.p0i8.i64(i8*, i8 *, i64, i1)\n" |
| 103 |
"define void @foo(i8* %dst, i8* %src, i64 %n) optsize {\n" |
| 104 |
"entry:\n" |
| 105 |
" %is_not_equal = icmp ne i8* %dst, %src\n" |
| 106 |
" br i1 %is_not_equal, label %memcpy, label %exit\n" |
| 107 |
"memcpy:\n" |
| 108 |
" call void @llvm.memcpy.p0i8.p0i8.i64(i8* %dst, i8* %src, " |
| 109 |
"i64 1024, i1 false)\n" |
| 110 |
" br label %exit\n" |
| 111 |
"exit:\n" |
| 112 |
" ret void\n" |
| 113 |
"}\n"); |
| 114 |
|
| 115 |
FunctionPassManager FPM; |
| 116 |
FPM.addPass(ForwardingPass( |
| 117 |
[=](Function &F, FunctionAnalysisManager &FAM) -> PreservedAnalyses { |
| 118 |
TargetTransformInfo TTI(M->getDataLayout()); |
| 119 |
auto *MemCpyBB = getBasicBlockByName(F, "memcpy"); |
| 120 |
Instruction *Inst = &MemCpyBB->front(); |
| 121 |
MemCpyInst *MemCpyI = cast<MemCpyInst>(Inst); |
| 122 |
auto &SE = FAM.getResult<ScalarEvolutionAnalysis>(F); |
| 123 |
expandMemCpyAsLoop(MemCpyI, TTI, &SE); |
| 124 |
auto *CopyLoopBB = getBasicBlockByName(F, "load-store-loop"); |
| 125 |
Instruction *LoadInst = |
| 126 |
getInstructionByOpcode(*CopyLoopBB, Instruction::Load, 1); |
| 127 |
EXPECT_NE(nullptr, LoadInst->getMetadata(LLVMContext::MD_alias_scope)); |
| 128 |
Instruction *StoreInst = |
| 129 |
getInstructionByOpcode(*CopyLoopBB, Instruction::Store, 1); |
| 130 |
EXPECT_NE(nullptr, StoreInst->getMetadata(LLVMContext::MD_noalias)); |
| 131 |
return PreservedAnalyses::none(); |
| 132 |
})); |
| 133 |
MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM))); |
| 134 |
|
| 135 |
MPM.run(*M, MAM); |
| 136 |
} |
| 137 |
|
| 138 |
// This test indirectly checks that loads and stores (generated as a result of |
| 139 |
// llvm.memcpy lowering) doesn't alias by making sure the loop can be |
| 140 |
// successfully vectorized without additional runtime checks. |
| 141 |
TEST_F(MemTransferLowerTest, VecMemCpyKnownLength) { |
| 142 |
ParseAssembly("declare void @llvm.memcpy.p0i8.p0i8.i64(i8*, i8 *, i64, i1)\n" |
| 143 |
"define void @foo(i8* %dst, i8* %src, i64 %n) optsize {\n" |
| 144 |
"entry:\n" |
| 145 |
" %is_not_equal = icmp ne i8* %dst, %src\n" |
| 146 |
" br i1 %is_not_equal, label %memcpy, label %exit\n" |
| 147 |
"memcpy:\n" |
| 148 |
" call void @llvm.memcpy.p0i8.p0i8.i64(i8* %dst, i8* %src, " |
| 149 |
"i64 1024, i1 false)\n" |
| 150 |
" br label %exit\n" |
| 151 |
"exit:\n" |
| 152 |
" ret void\n" |
| 153 |
"}\n"); |
| 154 |
|
| 155 |
FunctionPassManager FPM; |
| 156 |
FPM.addPass(ForwardingPass( |
| 157 |
[=](Function &F, FunctionAnalysisManager &FAM) -> PreservedAnalyses { |
| 158 |
TargetTransformInfo TTI(M->getDataLayout()); |
| 159 |
auto *MemCpyBB = getBasicBlockByName(F, "memcpy"); |
| 160 |
Instruction *Inst = &MemCpyBB->front(); |
| 161 |
MemCpyInst *MemCpyI = cast<MemCpyInst>(Inst); |
| 162 |
auto &SE = FAM.getResult<ScalarEvolutionAnalysis>(F); |
| 163 |
expandMemCpyAsLoop(MemCpyI, TTI, &SE); |
| 164 |
return PreservedAnalyses::none(); |
| 165 |
})); |
| 166 |
FPM.addPass(LoopVectorizePass(LoopVectorizeOptions())); |
| 167 |
FPM.addPass(ForwardingPass( |
| 168 |
[=](Function &F, FunctionAnalysisManager &FAM) -> PreservedAnalyses { |
| 169 |
auto *TargetBB = getBasicBlockByName(F, "vector.body"); |
| 170 |
EXPECT_NE(nullptr, TargetBB); |
| 171 |
return PreservedAnalyses::all(); |
| 172 |
})); |
| 173 |
MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM))); |
| 174 |
|
| 175 |
MPM.run(*M, MAM); |
| 176 |
} |
| 177 |
|
| 178 |
TEST_F(MemTransferLowerTest, AtomicMemCpyKnownLength) { |
| 179 |
ParseAssembly("declare void " |
| 180 |
"@llvm.memcpy.element.unordered.atomic.p0i32.p0i32.i64(i32*, " |
| 181 |
"i32 *, i64, i32)\n" |
| 182 |
"define void @foo(i32* %dst, i32* %src, i64 %n) optsize {\n" |
| 183 |
"entry:\n" |
| 184 |
" %is_not_equal = icmp ne i32* %dst, %src\n" |
| 185 |
" br i1 %is_not_equal, label %memcpy, label %exit\n" |
| 186 |
"memcpy:\n" |
| 187 |
" call void " |
| 188 |
"@llvm.memcpy.element.unordered.atomic.p0i32.p0i32.i64(i32* " |
| 189 |
"%dst, i32* %src, " |
| 190 |
"i64 1024, i32 4)\n" |
| 191 |
" br label %exit\n" |
| 192 |
"exit:\n" |
| 193 |
" ret void\n" |
| 194 |
"}\n"); |
| 195 |
|
| 196 |
FunctionPassManager FPM; |
| 197 |
FPM.addPass(ForwardingPass( |
| 198 |
[=](Function &F, FunctionAnalysisManager &FAM) -> PreservedAnalyses { |
| 199 |
TargetTransformInfo TTI(M->getDataLayout()); |
| 200 |
auto *MemCpyBB = getBasicBlockByName(F, "memcpy"); |
| 201 |
Instruction *Inst = &MemCpyBB->front(); |
| 202 |
assert(isa<AtomicMemCpyInst>(Inst) && |
| 203 |
"Expecting llvm.memcpy.p0i8.i64 instructon"); |
| 204 |
AtomicMemCpyInst *MemCpyI = cast<AtomicMemCpyInst>(Inst); |
| 205 |
auto &SE = FAM.getResult<ScalarEvolutionAnalysis>(F); |
| 206 |
expandAtomicMemCpyAsLoop(MemCpyI, TTI, &SE); |
| 207 |
auto *CopyLoopBB = getBasicBlockByName(F, "load-store-loop"); |
| 208 |
Instruction *LoadInst = |
| 209 |
getInstructionByOpcode(*CopyLoopBB, Instruction::Load, 1); |
| 210 |
EXPECT_TRUE(LoadInst->isAtomic()); |
| 211 |
EXPECT_NE(LoadInst->getMetadata(LLVMContext::MD_alias_scope), nullptr); |
| 212 |
Instruction *StoreInst = |
| 213 |
getInstructionByOpcode(*CopyLoopBB, Instruction::Store, 1); |
| 214 |
EXPECT_TRUE(StoreInst->isAtomic()); |
| 215 |
EXPECT_NE(StoreInst->getMetadata(LLVMContext::MD_noalias), nullptr); |
| 216 |
return PreservedAnalyses::none(); |
| 217 |
})); |
| 218 |
MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM))); |
| 219 |
|
| 220 |
MPM.run(*M, MAM); |
| 221 |
} |
| 222 |
|
| 223 |
TEST_F(MemTransferLowerTest, AtomicMemCpyUnKnownLength) { |
| 224 |
ParseAssembly("declare void " |
| 225 |
"@llvm.memcpy.element.unordered.atomic.p0i32.p0i32.i64(i32*, " |
| 226 |
"i32 *, i64, i32)\n" |
| 227 |
"define void @foo(i32* %dst, i32* %src, i64 %n) optsize {\n" |
| 228 |
"entry:\n" |
| 229 |
" %is_not_equal = icmp ne i32* %dst, %src\n" |
| 230 |
" br i1 %is_not_equal, label %memcpy, label %exit\n" |
| 231 |
"memcpy:\n" |
| 232 |
" call void " |
| 233 |
"@llvm.memcpy.element.unordered.atomic.p0i32.p0i32.i64(i32* " |
| 234 |
"%dst, i32* %src, " |
| 235 |
"i64 %n, i32 4)\n" |
| 236 |
" br label %exit\n" |
| 237 |
"exit:\n" |
| 238 |
" ret void\n" |
| 239 |
"}\n"); |
| 240 |
|
| 241 |
FunctionPassManager FPM; |
| 242 |
FPM.addPass(ForwardingPass( |
| 243 |
[=](Function &F, FunctionAnalysisManager &FAM) -> PreservedAnalyses { |
| 244 |
TargetTransformInfo TTI(M->getDataLayout()); |
| 245 |
auto *MemCpyBB = getBasicBlockByName(F, "memcpy"); |
| 246 |
Instruction *Inst = &MemCpyBB->front(); |
| 247 |
assert(isa<AtomicMemCpyInst>(Inst) && |
| 248 |
"Expecting llvm.memcpy.p0i8.i64 instructon"); |
| 249 |
AtomicMemCpyInst *MemCpyI = cast<AtomicMemCpyInst>(Inst); |
| 250 |
auto &SE = FAM.getResult<ScalarEvolutionAnalysis>(F); |
| 251 |
expandAtomicMemCpyAsLoop(MemCpyI, TTI, &SE); |
| 252 |
auto *CopyLoopBB = getBasicBlockByName(F, "loop-memcpy-expansion"); |
| 253 |
Instruction *LoadInst = |
| 254 |
getInstructionByOpcode(*CopyLoopBB, Instruction::Load, 1); |
| 255 |
EXPECT_TRUE(LoadInst->isAtomic()); |
| 256 |
EXPECT_NE(LoadInst->getMetadata(LLVMContext::MD_alias_scope), nullptr); |
| 257 |
Instruction *StoreInst = |
| 258 |
getInstructionByOpcode(*CopyLoopBB, Instruction::Store, 1); |
| 259 |
EXPECT_TRUE(StoreInst->isAtomic()); |
| 260 |
EXPECT_NE(StoreInst->getMetadata(LLVMContext::MD_noalias), nullptr); |
| 261 |
return PreservedAnalyses::none(); |
| 262 |
})); |
| 263 |
MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM))); |
| 264 |
|
| 265 |
MPM.run(*M, MAM); |
| 266 |
} |
| 267 |
} // namespace |