|
| 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 | + expandMemCpyAsLoop(MemCpyI, TTI); |
| 123 | + auto *CopyLoopBB = getBasicBlockByName(F, "load-store-loop"); |
| 124 | + Instruction *LoadInst = |
| 125 | + getInstructionByOpcode(*CopyLoopBB, Instruction::Load, 1); |
| 126 | + EXPECT_NONFATAL_FAILURE( |
| 127 | + EXPECT_NE(LoadInst->getMetadata(LLVMContext::MD_alias_scope), |
| 128 | + nullptr), |
| 129 | + ""); |
| 130 | + Instruction *StoreInst = |
| 131 | + getInstructionByOpcode(*CopyLoopBB, Instruction::Store, 1); |
| 132 | + EXPECT_NONFATAL_FAILURE( |
| 133 | + EXPECT_NE(StoreInst->getMetadata(LLVMContext::MD_noalias), nullptr), |
| 134 | + ""); |
| 135 | + return PreservedAnalyses::none(); |
| 136 | + })); |
| 137 | + MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM))); |
| 138 | + |
| 139 | + MPM.run(*M, MAM); |
| 140 | +} |
| 141 | + |
| 142 | +// This test indirectly checks that loads and stores (generated as a result of |
| 143 | +// llvm.memcpy lowering) doesn't alias by making sure the loop can be |
| 144 | +// successfully vectorized without additional runtime checks. |
| 145 | +TEST_F(MemTransferLowerTest, VecMemCpyKnownLength) { |
| 146 | + ParseAssembly("declare void @llvm.memcpy.p0i8.p0i8.i64(i8*, i8 *, i64, i1)\n" |
| 147 | + "define void @foo(i8* %dst, i8* %src, i64 %n) optsize {\n" |
| 148 | + "entry:\n" |
| 149 | + " %is_not_equal = icmp ne i8* %dst, %src\n" |
| 150 | + " br i1 %is_not_equal, label %memcpy, label %exit\n" |
| 151 | + "memcpy:\n" |
| 152 | + " call void @llvm.memcpy.p0i8.p0i8.i64(i8* %dst, i8* %src, " |
| 153 | + "i64 1024, i1 false)\n" |
| 154 | + " br label %exit\n" |
| 155 | + "exit:\n" |
| 156 | + " ret void\n" |
| 157 | + "}\n"); |
| 158 | + |
| 159 | + FunctionPassManager FPM; |
| 160 | + FPM.addPass(ForwardingPass( |
| 161 | + [=](Function &F, FunctionAnalysisManager &FAM) -> PreservedAnalyses { |
| 162 | + TargetTransformInfo TTI(M->getDataLayout()); |
| 163 | + auto *MemCpyBB = getBasicBlockByName(F, "memcpy"); |
| 164 | + Instruction *Inst = &MemCpyBB->front(); |
| 165 | + MemCpyInst *MemCpyI = cast<MemCpyInst>(Inst); |
| 166 | + expandMemCpyAsLoop(MemCpyI, TTI); |
| 167 | + return PreservedAnalyses::none(); |
| 168 | + })); |
| 169 | + FPM.addPass(LoopVectorizePass(LoopVectorizeOptions())); |
| 170 | + FPM.addPass(ForwardingPass( |
| 171 | + [=](Function &F, FunctionAnalysisManager &FAM) -> PreservedAnalyses { |
| 172 | + auto *TargetBB = getBasicBlockByName(F, "vector.body"); |
| 173 | + EXPECT_NONFATAL_FAILURE(EXPECT_NE(TargetBB, nullptr), ""); |
| 174 | + return PreservedAnalyses::all(); |
| 175 | + })); |
| 176 | + MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM))); |
| 177 | + |
| 178 | + MPM.run(*M, MAM); |
| 179 | +} |
| 180 | +} // namespace |
0 commit comments