-
-
Notifications
You must be signed in to change notification settings - Fork 5.5k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Implement ReinterpretArray #23750
Merged
Merged
Implement ReinterpretArray #23750
Changes from 1 commit
Commits
Show all changes
3 commits
Select commit
Hold shift + click to select a range
File filter
Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
365 changes: 365 additions & 0 deletions
365
deps/patches/llvm-D37939-Mem2Reg-Also-handle-memcpy.patch
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,365 @@ | ||
| From da4504b2d3c6629fbd58634bf76f1b85939d07cf Mon Sep 17 00:00:00 2001 | ||
| From: Keno Fischer <[email protected]> | ||
| Date: Fri, 15 Sep 2017 18:30:59 -0400 | ||
| Subject: [PATCH] [Mem2Reg] Also handle memcpy | ||
|
|
||
| Summary: | ||
| In julia, when we know we're moving data between two memory locations, | ||
| we always emit that as a memcpy rather than a load/store pair. However, | ||
| this can give worse optimization results in certain cases because some | ||
| optimizations that can handle load/store pairs cannot handle memcpys. | ||
| Mem2reg is one of these optimizations. This patch adds rudamentary | ||
| support for mem2reg for recognizing memcpys that cover the whole alloca | ||
| we're promoting. While several more sophisticated passes (SROA, GVN) | ||
| can get similar optimizations, it is preferable to have these kinds | ||
| of cases caught early to expose optimization opportunities before | ||
| getting to these later passes. The approach taken here is to split | ||
| the memcpy into a load/store pair early (after legality analysis) | ||
| and retain the rest of the analysis only on loads/stores. It would | ||
| be possible of course to leave the memcpy as is and generate the | ||
| left over load or store only on demand. However, that would entail | ||
| a significantly larger patch for unclear benefit. | ||
|
|
||
| Reviewers: chandlerc, dberlin | ||
|
|
||
| Subscribers: llvm-commits | ||
|
|
||
| Differential Revision: https://reviews.llvm.org/D37939 | ||
| --- | ||
| lib/Transforms/Utils/PromoteMemoryToRegister.cpp | 166 ++++++++++++++++++++--- | ||
| test/Transforms/Mem2Reg/memcpy.ll | 101 ++++++++++++++ | ||
| 2 files changed, 251 insertions(+), 16 deletions(-) | ||
| create mode 100644 test/Transforms/Mem2Reg/memcpy.ll | ||
|
|
||
| diff --git a/lib/Transforms/Utils/PromoteMemoryToRegister.cpp b/lib/Transforms/Utils/PromoteMemoryToRegister.cpp | ||
| index ac28f59..b08a0a1 100644 | ||
| --- a/lib/Transforms/Utils/PromoteMemoryToRegister.cpp | ||
| +++ b/lib/Transforms/Utils/PromoteMemoryToRegister.cpp | ||
| @@ -49,6 +49,58 @@ STATISTIC(NumSingleStore, "Number of alloca's promoted with a single store"); | ||
| STATISTIC(NumDeadAlloca, "Number of dead alloca's removed"); | ||
| STATISTIC(NumPHIInsert, "Number of PHI nodes inserted"); | ||
|
|
||
| +static bool isSplittableMemCpy(const MemCpyInst *MCI, const AllocaInst *AI) { | ||
| + // Punt if this alloca is an array allocation | ||
| + if (AI->isArrayAllocation()) | ||
| + return false; | ||
| + if (MCI->isVolatile()) | ||
| + return false; | ||
| + Value *Length = MCI->getLength(); | ||
| + if (!isa<ConstantInt>(Length)) | ||
| + return false; | ||
| + // Anything less than the full alloca, we leave for SROA | ||
| + const DataLayout &DL = AI->getModule()->getDataLayout(); | ||
| + size_t AIElSize = DL.getTypeAllocSize(AI->getAllocatedType()); | ||
| + if (cast<ConstantInt>(Length)->getZExtValue() != AIElSize) | ||
| + return false; | ||
| + // If the other argument is also an alloca, we need to be sure that either | ||
| + // the types are bitcastable, or the other alloca is not eligible for | ||
| + // promotion (e.g. because the memcpy is for less than the whole size of | ||
| + // that alloca), otherwise we risk turning an allocatable alloca into a | ||
| + // non-allocatable one when splitting the memcpy. | ||
| + AllocaInst *OtherAI = dyn_cast<AllocaInst>( | ||
| + AI == MCI->getSource() ? MCI->getDest() : MCI->getSource()); | ||
| + if (OtherAI) { | ||
| + if (!CastInst::isBitCastable(AI->getAllocatedType(), | ||
| + OtherAI->getAllocatedType()) && | ||
| + DL.getTypeAllocSize(OtherAI->getAllocatedType()) == AIElSize) | ||
| + return false; | ||
| + } | ||
| + return true; | ||
| +} | ||
| + | ||
| +/// Look at the result of a bitcast and see if it's only used by lifetime | ||
| +/// intrinsics or splittable memcpys. This is needed, because IRBuilder | ||
| +/// will always insert a bitcast to i8* for these intrinsics. | ||
| +static bool onlyHasCanonicalizableUsers(const AllocaInst *AI, const Value *V) { | ||
| + for (const User *U : V->users()) { | ||
| + const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U); | ||
| + if (!II) | ||
| + return false; | ||
| + | ||
| + if (isa<MemCpyInst>(II)) { | ||
| + if (!isSplittableMemCpy(cast<MemCpyInst>(II), AI)) | ||
| + return false; | ||
| + continue; | ||
| + } | ||
| + | ||
| + if (II->getIntrinsicID() != Intrinsic::lifetime_start && | ||
| + II->getIntrinsicID() != Intrinsic::lifetime_end) | ||
| + return false; | ||
| + } | ||
| + return true; | ||
| +} | ||
| + | ||
| bool llvm::isAllocaPromotable(const AllocaInst *AI) { | ||
| // FIXME: If the memory unit is of pointer or integer type, we can permit | ||
| // assignments to subsections of the memory unit. | ||
| @@ -68,6 +120,9 @@ bool llvm::isAllocaPromotable(const AllocaInst *AI) { | ||
| // not have any meaning for a local alloca. | ||
| if (SI->isVolatile()) | ||
| return false; | ||
| + } else if (const MemCpyInst *MCI = dyn_cast<MemCpyInst>(U)) { | ||
| + if (!isSplittableMemCpy(MCI, AI)) | ||
| + return false; | ||
| } else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) { | ||
| if (II->getIntrinsicID() != Intrinsic::lifetime_start && | ||
| II->getIntrinsicID() != Intrinsic::lifetime_end) | ||
| @@ -75,7 +130,7 @@ bool llvm::isAllocaPromotable(const AllocaInst *AI) { | ||
| } else if (const BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { | ||
| if (BCI->getType() != Type::getInt8PtrTy(U->getContext(), AS)) | ||
| return false; | ||
| - if (!onlyUsedByLifetimeMarkers(BCI)) | ||
| + if (!onlyHasCanonicalizableUsers(AI, BCI)) | ||
| return false; | ||
| } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) { | ||
| if (GEPI->getType() != Type::getInt8PtrTy(U->getContext(), AS)) | ||
| @@ -181,7 +235,13 @@ public: | ||
| /// This code only looks at accesses to allocas. | ||
| static bool isInterestingInstruction(const Instruction *I) { | ||
| + if (isa<MemCpyInst>(I)) { | ||
| + const MemCpyInst *MCI = cast<MemCpyInst>(I); | ||
| + return isa<AllocaInst>(MCI->getSource()) || | ||
| + isa<AllocaInst>(MCI->getDest()); | ||
| + } else { | ||
| return (isa<LoadInst>(I) && isa<AllocaInst>(I->getOperand(0))) || | ||
| (isa<StoreInst>(I) && isa<AllocaInst>(I->getOperand(1))); | ||
| } | ||
| + } | ||
|
|
||
| /// Get or calculate the index of the specified instruction. | ||
| @@ -208,6 +264,25 @@ public: | ||
| return It->second; | ||
| } | ||
|
|
||
| + // When we split a memcpy intrinsic, we need to update the numbering in this | ||
| + // struct. To make sure the relative ordering remains the same, we give both | ||
| + // the LI and the SI the number that the MCI used to have (if they are both | ||
| + // interesting). This means that they will have equal numbers, which usually | ||
| + // can't happen. However, since they can never reference the same alloca | ||
| + // (since memcpy operands may not overlap), this is fine, because we will | ||
| + // never compare instruction indices for instructions that operate on distinct | ||
| + // allocas. | ||
| + void splitMemCpy(MemCpyInst *MCI, LoadInst *LI, StoreInst *SI) { | ||
| + DenseMap<const Instruction *, unsigned>::iterator It = | ||
| + InstNumbers.find(MCI); | ||
| + if (It == InstNumbers.end()) | ||
| + return; | ||
| + unsigned MemCpyNumber = It->second; | ||
| + InstNumbers[LI] = MemCpyNumber; | ||
| + InstNumbers[SI] = MemCpyNumber; | ||
| + deleteValue(MCI); | ||
| + } | ||
| + | ||
| void deleteValue(const Instruction *I) { InstNumbers.erase(I); } | ||
|
|
||
| void clear() { InstNumbers.clear(); } | ||
| @@ -305,9 +380,58 @@ static void addAssumeNonNull(AssumptionCache *AC, LoadInst *LI) { | ||
| AC->registerAssumption(CI); | ||
| } | ||
|
|
||
| -static void removeLifetimeIntrinsicUsers(AllocaInst *AI) { | ||
| - // Knowing that this alloca is promotable, we know that it's safe to kill all | ||
| - // instructions except for load and store. | ||
| +/// Split a memcpy instruction into the corresponding load/store. It is a little | ||
| +/// more complicated than one might imagine, because we need to deal with the | ||
| +/// fact that the side of the copy we're not currently processing might also | ||
| +/// be a promotable alloca. We need to be careful to not break the promotable | ||
| +/// predicate for that other alloca (if any). | ||
| +static void doMemCpySplit(LargeBlockInfo &LBI, MemCpyInst *MCI, | ||
| + AllocaInst *AI) { | ||
| + AAMDNodes AA; | ||
| + MCI->getAAMetadata(AA); | ||
| + Value *MCISrc = MCI->getSource(); | ||
| + Type *LoadType = AI->getAllocatedType(); | ||
| + AllocaInst *SrcAI = dyn_cast<AllocaInst>(MCISrc); | ||
| + if (SrcAI && SrcAI->getType() != AI->getType()) { | ||
| + if (CastInst::isBitCastable(SrcAI->getAllocatedType(), LoadType)) | ||
| + LoadType = SrcAI->getAllocatedType(); | ||
| + } | ||
| + if (cast<PointerType>(MCISrc->getType())->getElementType() != LoadType) | ||
| + MCISrc = CastInst::Create( | ||
| + Instruction::BitCast, MCISrc, | ||
| + LoadType->getPointerTo( | ||
| + cast<PointerType>(MCISrc->getType())->getAddressSpace()), | ||
| + "", MCI); | ||
| + // This might add to the end of the use list, but that's fine. At worst, | ||
| + // we'd not visit the instructions we insert here, but we don't care | ||
| + // about them in this loop anyway. | ||
| + LoadInst *LI = new LoadInst(LoadType, MCISrc, "", MCI->isVolatile(), | ||
| + MCI->getAlignment(), MCI); | ||
| + Value *Val = LI; | ||
| + Value *MCIDest = MCI->getDest(); | ||
| + AllocaInst *DestAI = dyn_cast<AllocaInst>(MCIDest); | ||
| + Type *DestElTy = DestAI ? DestAI->getAllocatedType() : AI->getAllocatedType(); | ||
| + if (LI->getType() != DestElTy && | ||
| + CastInst::isBitCastable(LI->getType(), DestElTy)) | ||
| + Val = CastInst::Create(Instruction::BitCast, Val, DestElTy, "", MCI); | ||
| + if (cast<PointerType>(MCIDest->getType())->getElementType() != Val->getType()) | ||
| + MCIDest = CastInst::Create( | ||
| + Instruction::BitCast, MCIDest, | ||
| + Val->getType()->getPointerTo( | ||
| + cast<PointerType>(MCIDest->getType())->getAddressSpace()), | ||
| + "", MCI); | ||
| + StoreInst *SI = | ||
| + new StoreInst(Val, MCIDest, MCI->isVolatile(), MCI->getAlignment(), MCI); | ||
| + LI->setAAMetadata(AA); | ||
| + SI->setAAMetadata(AA); | ||
| + LBI.splitMemCpy(MCI, LI, SI); | ||
| + MCI->eraseFromParent(); | ||
| +} | ||
| + | ||
| +static void canonicalizeUsers(LargeBlockInfo &LBI, AllocaInst *AI) { | ||
| + // Knowing that this alloca is promotable, we know that it's safe to split | ||
| + // MTIs into load/store and to kill all other instructions except for | ||
| + // load and store. | ||
|
|
||
| for (auto UI = AI->user_begin(), UE = AI->user_end(); UI != UE;) { | ||
| Instruction *I = cast<Instruction>(*UI); | ||
| @@ -315,14 +439,24 @@ static void removeLifetimeIntrinsicUsers(AllocaInst *AI) { | ||
| if (isa<LoadInst>(I) || isa<StoreInst>(I)) | ||
| continue; | ||
|
|
||
| + if (isa<MemCpyInst>(I)) { | ||
| + MemCpyInst *MCI = cast<MemCpyInst>(I); | ||
| + doMemCpySplit(LBI, MCI, AI); | ||
| + continue; | ||
| + } | ||
| + | ||
| if (!I->getType()->isVoidTy()) { | ||
| - // The only users of this bitcast/GEP instruction are lifetime intrinsics. | ||
| - // Follow the use/def chain to erase them now instead of leaving it for | ||
| - // dead code elimination later. | ||
| + // The only users of this bitcast/GEP instruction are lifetime/memcpy | ||
| + // intrinsics. Split memcpys and delete lifetime intrinsics. | ||
| for (auto UUI = I->user_begin(), UUE = I->user_end(); UUI != UUE;) { | ||
| Instruction *Inst = cast<Instruction>(*UUI); | ||
| ++UUI; | ||
| - Inst->eraseFromParent(); | ||
| + if (isa<MemCpyInst>(Inst)) { | ||
| + doMemCpySplit(LBI, cast<MemCpyInst>(Inst), AI); | ||
| + } else { | ||
| + // Must be a lifetime intrinsic | ||
| + Inst->eraseFromParent(); | ||
| + } | ||
| } | ||
| } | ||
| I->eraseFromParent(); | ||
| @@ -542,7 +676,7 @@ void PromoteMem2Reg::run() { | ||
| assert(AI->getParent()->getParent() == &F && | ||
| "All allocas should be in the same function, which is same as DF!"); | ||
|
|
||
| - removeLifetimeIntrinsicUsers(AI); | ||
| + canonicalizeUsers(LBI, AI); | ||
|
|
||
| if (AI->use_empty()) { | ||
| // If there are no uses of the alloca, just delete it now. | ||
| diff --git a/test/Transforms/Mem2Reg/memcpy.ll b/test/Transforms/Mem2Reg/memcpy.ll | ||
| new file mode 100644 | ||
| index 0000000..fbc4096 | ||
| --- /dev/null | ||
| +++ b/test/Transforms/Mem2Reg/memcpy.ll | ||
| @@ -0,0 +1,101 @@ | ||
| +; RUN: opt < %s -mem2reg -S | FileCheck %s | ||
| + | ||
| +target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128" | ||
| + | ||
| +declare void @llvm.memcpy.p0i128.p0i64.i32(i128 *, i64 *, i32, i32, i1) | ||
| +declare void @llvm.memcpy.p0i8.p0i8.i32(i8 *, i8 *, i32, i32, i1) | ||
| +declare void @llvm.memcpy.p0i64.p0i64.i32(i64 *, i64 *, i32, i32, i1) | ||
| +declare void @llvm.memcpy.p0f64.p0i64.i32(double *, i64 *, i32, i32, i1) | ||
| + | ||
| +define i128 @test_cpy_different(i64) { | ||
| +; CHECK-LABEL: @test_cpy_different | ||
| +; CHECK-NOT: alloca i64 | ||
| +; CHECK: store i64 %0 | ||
| + %a = alloca i64 | ||
| + %b = alloca i128 | ||
| + store i128 0, i128 *%b | ||
| + store i64 %0, i64 *%a | ||
| + call void @llvm.memcpy.p0i128.p0i64.i32(i128 *%b, i64 *%a, i32 8, i32 0, i1 0) | ||
| + %loaded = load i128, i128 *%b | ||
| + ret i128 %loaded | ||
| +} | ||
| + | ||
| +define i64 @test_cpy_same(i64) { | ||
| +; CHECK-LABEL: @test_cpy_same | ||
| +; CHECK-NOT: alloca | ||
| +; CHECK: ret i64 %0 | ||
| + %a = alloca i64 | ||
| + %b = alloca i64 | ||
| + store i64 %0, i64 *%a | ||
| + call void @llvm.memcpy.p0i64.p0i64.i32(i64 *%b, i64 *%a, i32 8, i32 0, i1 0) | ||
| + %loaded = load i64, i64 *%b | ||
| + ret i64 %loaded | ||
| +} | ||
| + | ||
| +define double @test_cpy_different_type(i64) { | ||
| +; CHECK-LABEL: @test_cpy_different_type | ||
| +; CHECK-NOT: alloca | ||
| +; CHECK: bitcast i64 %0 to double | ||
| + %a = alloca i64 | ||
| + %b = alloca double | ||
| + store i64 %0, i64 *%a | ||
| + call void @llvm.memcpy.p0f64.p0i64.i32(double *%b, i64 *%a, i32 8, i32 0, i1 0) | ||
| + %loaded = load double, double *%b | ||
| + ret double %loaded | ||
| +} | ||
| + | ||
| +define i128 @test_cpy_differenti8(i64) { | ||
| +; CHECK-LABEL: @test_cpy_differenti8 | ||
| +; CHECK-NOT: alloca i64 | ||
| +; CHECK: store i64 %0 | ||
| + %a = alloca i64 | ||
| + %b = alloca i128 | ||
| + store i128 0, i128 *%b | ||
| + store i64 %0, i64 *%a | ||
| + %acast = bitcast i64* %a to i8* | ||
| + %bcast = bitcast i128* %b to i8* | ||
| + call void @llvm.memcpy.p0i8.p0i8.i32(i8 *%bcast, i8 *%acast, i32 8, i32 0, i1 0) | ||
| + %loaded = load i128, i128 *%b | ||
| + ret i128 %loaded | ||
| +} | ||
| + | ||
| +define i64 @test_cpy_samei8(i64) { | ||
| +; CHECK-LABEL: @test_cpy_samei8 | ||
| +; CHECK-NOT: alloca | ||
| +; CHECK: ret i64 %0 | ||
| + %a = alloca i64 | ||
| + %b = alloca i64 | ||
| + store i64 %0, i64 *%a | ||
| + %acast = bitcast i64* %a to i8* | ||
| + %bcast = bitcast i64* %b to i8* | ||
| + call void @llvm.memcpy.p0i8.p0i8.i32(i8 *%bcast, i8 *%acast, i32 8, i32 0, i1 0) | ||
| + %loaded = load i64, i64 *%b | ||
| + ret i64 %loaded | ||
| +} | ||
| + | ||
| +define double @test_cpy_different_typei8(i64) { | ||
| +; CHECK-LABEL: @test_cpy_different_typei8 | ||
| +; CHECK-NOT: alloca | ||
| +; CHECK: bitcast i64 %0 to double | ||
| + %a = alloca i64 | ||
| + %b = alloca double | ||
| + store i64 %0, i64 *%a | ||
| + %acast = bitcast i64* %a to i8* | ||
| + %bcast = bitcast double* %b to i8* | ||
| + call void @llvm.memcpy.p0i8.p0i8.i32(i8 *%bcast, i8 *%acast, i32 8, i32 0, i1 0) | ||
| + %loaded = load double, double *%b | ||
| + ret double %loaded | ||
| +} | ||
| + | ||
| +define i64 @test_cpy_differenti8_reverse(i128) { | ||
| +; CHECK-LABEL: @test_cpy_differenti8_reverse | ||
| +; CHECK-NOT: alloca i64 | ||
| + %a = alloca i64 | ||
| + %b = alloca i128 | ||
| + store i128 %0, i128 *%b | ||
| + %acast = bitcast i64* %a to i8* | ||
| + %bcast = bitcast i128* %b to i8* | ||
| + call void @llvm.memcpy.p0i8.p0i8.i32(i8 *%acast, i8 *%bcast, i32 8, i32 0, i1 0) | ||
| + %loaded = load i64, i64 *%a | ||
| + ret i64 %loaded | ||
| +} | ||
| -- | ||
| 2.9.3 | ||
|
|
||
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
rudimentary