X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTransforms%2FScalar%2FScalarReplAggregates.cpp;h=d955da7ce75d2741ec5917b16d0caa67f9a13481;hb=3de535e566ce5b86c1b484efed899c5a78a4519b;hp=fcc5f1985b81030affd253441422bf57b980d0d0;hpb=5a1cb644c903da49dc612a0ba5044505d066259e;p=oota-llvm.git diff --git a/lib/Transforms/Scalar/ScalarReplAggregates.cpp b/lib/Transforms/Scalar/ScalarReplAggregates.cpp index fcc5f1985b8..d955da7ce75 100644 --- a/lib/Transforms/Scalar/ScalarReplAggregates.cpp +++ b/lib/Transforms/Scalar/ScalarReplAggregates.cpp @@ -13,68 +13,82 @@ // each member (if possible). Then, if possible, it transforms the individual // alloca instructions into nice clean scalar SSA form. // -// This combines a simple SRoA algorithm with the Mem2Reg algorithm because +// This combines a simple SRoA algorithm with the Mem2Reg algorithm because they // often interact, especially for C++ programs. As such, iterating between // SRoA, then Mem2Reg until we run out of things to promote works well. // //===----------------------------------------------------------------------===// -#define DEBUG_TYPE "scalarrepl" #include "llvm/Transforms/Scalar.h" -#include "llvm/Constants.h" -#include "llvm/DerivedTypes.h" -#include "llvm/Function.h" -#include "llvm/GlobalVariable.h" -#include "llvm/Instructions.h" -#include "llvm/IntrinsicInst.h" -#include "llvm/LLVMContext.h" -#include "llvm/Module.h" -#include "llvm/Pass.h" -#include "llvm/Analysis/DebugInfo.h" -#include "llvm/Analysis/DIBuilder.h" -#include "llvm/Analysis/Dominators.h" +#include "llvm/ADT/SetVector.h" +#include "llvm/ADT/SmallVector.h" +#include "llvm/ADT/Statistic.h" +#include "llvm/Analysis/AssumptionCache.h" #include "llvm/Analysis/Loads.h" #include "llvm/Analysis/ValueTracking.h" -#include "llvm/Target/TargetData.h" -#include "llvm/Transforms/Utils/PromoteMemToReg.h" -#include "llvm/Transforms/Utils/Local.h" -#include "llvm/Transforms/Utils/SSAUpdater.h" -#include "llvm/Support/CallSite.h" +#include "llvm/IR/CallSite.h" +#include "llvm/IR/Constants.h" +#include "llvm/IR/DIBuilder.h" +#include "llvm/IR/DataLayout.h" +#include "llvm/IR/DebugInfo.h" +#include "llvm/IR/DerivedTypes.h" +#include "llvm/IR/Dominators.h" +#include "llvm/IR/Function.h" +#include "llvm/IR/GetElementPtrTypeIterator.h" +#include "llvm/IR/GlobalVariable.h" +#include "llvm/IR/IRBuilder.h" +#include "llvm/IR/Instructions.h" +#include "llvm/IR/IntrinsicInst.h" +#include "llvm/IR/LLVMContext.h" +#include "llvm/IR/Module.h" +#include "llvm/IR/Operator.h" +#include "llvm/Pass.h" #include "llvm/Support/Debug.h" #include "llvm/Support/ErrorHandling.h" -#include "llvm/Support/GetElementPtrTypeIterator.h" -#include "llvm/Support/IRBuilder.h" #include "llvm/Support/MathExtras.h" #include "llvm/Support/raw_ostream.h" -#include "llvm/ADT/SetVector.h" -#include "llvm/ADT/SmallVector.h" -#include "llvm/ADT/Statistic.h" +#include "llvm/Transforms/Utils/Local.h" +#include "llvm/Transforms/Utils/PromoteMemToReg.h" +#include "llvm/Transforms/Utils/SSAUpdater.h" using namespace llvm; +#define DEBUG_TYPE "scalarrepl" + STATISTIC(NumReplaced, "Number of allocas broken up"); STATISTIC(NumPromoted, "Number of allocas promoted"); STATISTIC(NumAdjusted, "Number of scalar allocas adjusted to allow promotion"); STATISTIC(NumConverted, "Number of aggregates converted to scalar"); -STATISTIC(NumGlobals, "Number of allocas copied from constant global"); namespace { struct SROA : public FunctionPass { - SROA(int T, bool hasDT, char &ID) + SROA(int T, bool hasDT, char &ID, int ST, int AT, int SLT) : FunctionPass(ID), HasDomTree(hasDT) { if (T == -1) SRThreshold = 128; else SRThreshold = T; + if (ST == -1) + StructMemberThreshold = 32; + else + StructMemberThreshold = ST; + if (AT == -1) + ArrayElementThreshold = 8; + else + ArrayElementThreshold = AT; + if (SLT == -1) + // Do not limit the scalar integer load size if no threshold is given. + ScalarLoadThreshold = -1; + else + ScalarLoadThreshold = SLT; } - bool runOnFunction(Function &F); + bool runOnFunction(Function &F) override; bool performScalarRepl(Function &F); bool performPromotion(Function &F); private: bool HasDomTree; - TargetData *TD; /// DeadInsts - Keep track of instructions we have made dead, so that /// we can remove them after we are done working. @@ -86,11 +100,11 @@ namespace { struct AllocaInfo { /// The alloca to promote. AllocaInst *AI; - + /// CheckedPHIs - This is a set of verified PHI nodes, to prevent infinite /// looping and avoid redundant work. SmallPtrSet CheckedPHIs; - + /// isUnsafe - This is set to true if the alloca cannot be SROA'd. bool isUnsafe : 1; @@ -104,19 +118,32 @@ namespace { /// ever accessed, or false if the alloca is only accessed with mem /// intrinsics or load/store that only access the entire alloca at once. bool hasSubelementAccess : 1; - + /// hasALoadOrStore - This is true if there are any loads or stores to it. /// The alloca may just be accessed with memcpy, for example, which would /// not set this. bool hasALoadOrStore : 1; - + explicit AllocaInfo(AllocaInst *ai) : AI(ai), isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false), hasSubelementAccess(false), hasALoadOrStore(false) {} }; + /// SRThreshold - The maximum alloca size to considered for SROA. unsigned SRThreshold; + /// StructMemberThreshold - The maximum number of members a struct can + /// contain to be considered for SROA. + unsigned StructMemberThreshold; + + /// ArrayElementThreshold - The maximum number of elements an array can + /// have to be considered for SROA. + unsigned ArrayElementThreshold; + + /// ScalarLoadThreshold - The maximum size in bits of scalars to load when + /// converting to scalar + unsigned ScalarLoadThreshold; + void MarkUnsafe(AllocaInfo &I, Instruction *User) { I.isUnsafe = true; DEBUG(dbgs() << " Transformation preventing inst: " << *User << '\n'); @@ -131,66 +158,69 @@ namespace { void isSafeMemAccess(uint64_t Offset, uint64_t MemSize, Type *MemOpType, bool isStore, AllocaInfo &Info, Instruction *TheAccess, bool AllowWholeAccess); - bool TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size); - uint64_t FindElementAndOffset(Type *&T, uint64_t &Offset, - Type *&IdxTy); + bool TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size, + const DataLayout &DL); + uint64_t FindElementAndOffset(Type *&T, uint64_t &Offset, Type *&IdxTy, + const DataLayout &DL); void DoScalarReplacement(AllocaInst *AI, std::vector &WorkList); void DeleteDeadInstructions(); void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, - SmallVector &NewElts); + SmallVectorImpl &NewElts); void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, - SmallVector &NewElts); + SmallVectorImpl &NewElts); void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, - SmallVector &NewElts); + SmallVectorImpl &NewElts); void RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI, uint64_t Offset, - SmallVector &NewElts); + SmallVectorImpl &NewElts); void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, AllocaInst *AI, - SmallVector &NewElts); + SmallVectorImpl &NewElts); void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, - SmallVector &NewElts); + SmallVectorImpl &NewElts); void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, - SmallVector &NewElts); - - static MemTransferInst *isOnlyCopiedFromConstantGlobal( - AllocaInst *AI, SmallVector &ToDelete); + SmallVectorImpl &NewElts); + bool ShouldAttemptScalarRepl(AllocaInst *AI); }; - + // SROA_DT - SROA that uses DominatorTree. struct SROA_DT : public SROA { static char ID; public: - SROA_DT(int T = -1) : SROA(T, true, ID) { + SROA_DT(int T = -1, int ST = -1, int AT = -1, int SLT = -1) : + SROA(T, true, ID, ST, AT, SLT) { initializeSROA_DTPass(*PassRegistry::getPassRegistry()); } - + // getAnalysisUsage - This pass does not require any passes, but we know it // will not alter the CFG, so say so. - virtual void getAnalysisUsage(AnalysisUsage &AU) const { - AU.addRequired(); + void getAnalysisUsage(AnalysisUsage &AU) const override { + AU.addRequired(); + AU.addRequired(); AU.setPreservesCFG(); } }; - + // SROA_SSAUp - SROA that uses SSAUpdater. struct SROA_SSAUp : public SROA { static char ID; public: - SROA_SSAUp(int T = -1) : SROA(T, false, ID) { + SROA_SSAUp(int T = -1, int ST = -1, int AT = -1, int SLT = -1) : + SROA(T, false, ID, ST, AT, SLT) { initializeSROA_SSAUpPass(*PassRegistry::getPassRegistry()); } - + // getAnalysisUsage - This pass does not require any passes, but we know it // will not alter the CFG, so say so. - virtual void getAnalysisUsage(AnalysisUsage &AU) const { + void getAnalysisUsage(AnalysisUsage &AU) const override { + AU.addRequired(); AU.setPreservesCFG(); } }; - + } char SROA_DT::ID = 0; @@ -198,21 +228,28 @@ char SROA_SSAUp::ID = 0; INITIALIZE_PASS_BEGIN(SROA_DT, "scalarrepl", "Scalar Replacement of Aggregates (DT)", false, false) -INITIALIZE_PASS_DEPENDENCY(DominatorTree) +INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) +INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) INITIALIZE_PASS_END(SROA_DT, "scalarrepl", "Scalar Replacement of Aggregates (DT)", false, false) INITIALIZE_PASS_BEGIN(SROA_SSAUp, "scalarrepl-ssa", "Scalar Replacement of Aggregates (SSAUp)", false, false) +INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) INITIALIZE_PASS_END(SROA_SSAUp, "scalarrepl-ssa", "Scalar Replacement of Aggregates (SSAUp)", false, false) // Public interface to the ScalarReplAggregates pass FunctionPass *llvm::createScalarReplAggregatesPass(int Threshold, - bool UseDomTree) { + bool UseDomTree, + int StructMemberThreshold, + int ArrayElementThreshold, + int ScalarLoadThreshold) { if (UseDomTree) - return new SROA_DT(Threshold); - return new SROA_SSAUp(Threshold); + return new SROA_DT(Threshold, StructMemberThreshold, ArrayElementThreshold, + ScalarLoadThreshold); + return new SROA_SSAUp(Threshold, StructMemberThreshold, + ArrayElementThreshold, ScalarLoadThreshold); } @@ -227,7 +264,8 @@ namespace { class ConvertToScalarInfo { /// AllocaSize - The size of the alloca being considered in bytes. unsigned AllocaSize; - const TargetData &TD; + const DataLayout &DL; + unsigned ScalarLoadThreshold; /// IsNotTrivial - This is set to true if there is some access to the object /// which means that mem2reg can't promote it. @@ -258,28 +296,38 @@ class ConvertToScalarInfo { /// isn't possible to turn into a vector type, it gets set to VoidTy. VectorType *VectorTy; - /// HadNonMemTransferAccess - True if there is at least one access to the + /// HadNonMemTransferAccess - True if there is at least one access to the /// alloca that is not a MemTransferInst. We don't want to turn structs into /// large integers unless there is some potential for optimization. bool HadNonMemTransferAccess; + /// HadDynamicAccess - True if some element of this alloca was dynamic. + /// We don't yet have support for turning a dynamic access into a large + /// integer. + bool HadDynamicAccess; + public: - explicit ConvertToScalarInfo(unsigned Size, const TargetData &td) - : AllocaSize(Size), TD(td), IsNotTrivial(false), ScalarKind(Unknown), - VectorTy(0), HadNonMemTransferAccess(false) { } + explicit ConvertToScalarInfo(unsigned Size, const DataLayout &DL, + unsigned SLT) + : AllocaSize(Size), DL(DL), ScalarLoadThreshold(SLT), IsNotTrivial(false), + ScalarKind(Unknown), VectorTy(nullptr), HadNonMemTransferAccess(false), + HadDynamicAccess(false) { } AllocaInst *TryConvert(AllocaInst *AI); private: - bool CanConvertToScalar(Value *V, uint64_t Offset); + bool CanConvertToScalar(Value *V, uint64_t Offset, Value* NonConstantIdx); void MergeInTypeForLoadOrStore(Type *In, uint64_t Offset); bool MergeInVectorType(VectorType *VInTy, uint64_t Offset); - void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); + void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset, + Value *NonConstantIdx); Value *ConvertScalar_ExtractValue(Value *NV, Type *ToType, - uint64_t Offset, IRBuilder<> &Builder); + uint64_t Offset, Value* NonConstantIdx, + IRBuilder<> &Builder); Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, - uint64_t Offset, IRBuilder<> &Builder); + uint64_t Offset, Value* NonConstantIdx, + IRBuilder<> &Builder); }; } // end anonymous namespace. @@ -290,16 +338,14 @@ private: AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) { // If we can't convert this scalar, or if mem2reg can trivially do it, bail // out. - if (!CanConvertToScalar(AI, 0) || !IsNotTrivial) - return 0; + if (!CanConvertToScalar(AI, 0, nullptr) || !IsNotTrivial) + return nullptr; // If an alloca has only memset / memcpy uses, it may still have an Unknown // ScalarKind. Treat it as an Integer below. if (ScalarKind == Unknown) ScalarKind = Integer; - // FIXME: It should be possible to promote the vector type up to the alloca's - // size. if (ScalarKind == Vector && VectorTy->getBitWidth() != AllocaSize * 8) ScalarKind = Integer; @@ -317,16 +363,28 @@ AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) { NewTy = VectorTy; // Use the vector type. } else { unsigned BitWidth = AllocaSize * 8; + + // Do not convert to scalar integer if the alloca size exceeds the + // scalar load threshold. + if (BitWidth > ScalarLoadThreshold) + return nullptr; + if ((ScalarKind == ImplicitVector || ScalarKind == Integer) && - !HadNonMemTransferAccess && !TD.fitsInLegalInteger(BitWidth)) - return 0; + !HadNonMemTransferAccess && !DL.fitsInLegalInteger(BitWidth)) + return nullptr; + // Dynamic accesses on integers aren't yet supported. They need us to shift + // by a dynamic amount which could be difficult to work out as we might not + // know whether to use a left or right shift. + if (ScalarKind == Integer && HadDynamicAccess) + return nullptr; DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"); // Create and insert the integer alloca. NewTy = IntegerType::get(AI->getContext(), BitWidth); } - AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); - ConvertUsesToScalar(AI, NewAI, 0); + AllocaInst *NewAI = new AllocaInst(NewTy, nullptr, "", + AI->getParent()->begin()); + ConvertUsesToScalar(AI, NewAI, 0, nullptr); return NewAI; } @@ -334,16 +392,12 @@ AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) { /// (VectorTy) so far at the offset specified by Offset (which is specified in /// bytes). /// -/// There are three cases we handle here: +/// There are two cases we handle here: /// 1) A union of vector types of the same size and potentially its elements. /// Here we turn element accesses into insert/extract element operations. /// This promotes a <4 x float> with a store of float to the third element /// into a <4 x float> that uses insert element. -/// 2) A union of vector types with power-of-2 size differences, e.g. a float, -/// <2 x float> and <4 x float>. Here we turn element accesses into insert -/// and extract element operations, and <2 x float> accesses into a cast to -/// <2 x double>, an extract, and a cast back to <2 x float>. -/// 3) A fully general blob of memory, which we turn into some (potentially +/// 2) A fully general blob of memory, which we turn into some (potentially /// large) integer type with extract and insert operations where the loads /// and stores would mutate the memory. We mark this by setting VectorTy /// to VoidTy. @@ -374,20 +428,13 @@ void ConvertToScalarInfo::MergeInTypeForLoadOrStore(Type *In, // if the implied vector agrees with what we already have and if Offset is // compatible with it. if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 && - (!VectorTy || Offset * 8 < VectorTy->getPrimitiveSizeInBits())) { + (!VectorTy || EltSize == VectorTy->getElementType() + ->getPrimitiveSizeInBits()/8)) { if (!VectorTy) { ScalarKind = ImplicitVector; VectorTy = VectorType::get(In, AllocaSize/EltSize); - return; } - - unsigned CurrentEltSize = VectorTy->getElementType() - ->getPrimitiveSizeInBits()/8; - if (EltSize == CurrentEltSize) - return; - - if (In->isIntegerTy() && isPowerOf2_32(AllocaSize / EltSize)) - return; + return; } } @@ -400,78 +447,19 @@ void ConvertToScalarInfo::MergeInTypeForLoadOrStore(Type *In, /// returning true if the type was successfully merged and false otherwise. bool ConvertToScalarInfo::MergeInVectorType(VectorType *VInTy, uint64_t Offset) { - // TODO: Support nonzero offsets? - if (Offset != 0) - return false; - - // Only allow vectors that are a power-of-2 away from the size of the alloca. - if (!isPowerOf2_64(AllocaSize / (VInTy->getBitWidth() / 8))) - return false; - - // If this the first vector we see, remember the type so that we know the - // element size. - if (!VectorTy) { - ScalarKind = Vector; - VectorTy = VInTy; - return true; - } - - unsigned BitWidth = VectorTy->getBitWidth(); - unsigned InBitWidth = VInTy->getBitWidth(); - - // Vectors of the same size can be converted using a simple bitcast. - if (InBitWidth == BitWidth && AllocaSize == (InBitWidth / 8)) { + if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { + // If we're storing/loading a vector of the right size, allow it as a + // vector. If this the first vector we see, remember the type so that + // we know the element size. If this is a subsequent access, ignore it + // even if it is a differing type but the same size. Worst case we can + // bitcast the resultant vectors. + if (!VectorTy) + VectorTy = VInTy; ScalarKind = Vector; return true; } - Type *ElementTy = VectorTy->getElementType(); - Type *InElementTy = VInTy->getElementType(); - - // If they're the same alloc size, we'll be attempting to convert between - // them with a vector shuffle, which requires the element types to match. - if (TD.getTypeAllocSize(VectorTy) == TD.getTypeAllocSize(VInTy) && - ElementTy != InElementTy) - return false; - - // Do not allow mixed integer and floating-point accesses from vectors of - // different sizes. - if (ElementTy->isFloatingPointTy() != InElementTy->isFloatingPointTy()) - return false; - - if (ElementTy->isFloatingPointTy()) { - // Only allow floating-point vectors of different sizes if they have the - // same element type. - // TODO: This could be loosened a bit, but would anything benefit? - if (ElementTy != InElementTy) - return false; - - // There are no arbitrary-precision floating-point types, which limits the - // number of legal vector types with larger element types that we can form - // to bitcast and extract a subvector. - // TODO: We could support some more cases with mixed fp128 and double here. - if (!(BitWidth == 64 || BitWidth == 128) || - !(InBitWidth == 64 || InBitWidth == 128)) - return false; - } else { - assert(ElementTy->isIntegerTy() && "Vector elements must be either integer " - "or floating-point."); - unsigned BitWidth = ElementTy->getPrimitiveSizeInBits(); - unsigned InBitWidth = InElementTy->getPrimitiveSizeInBits(); - - // Do not allow integer types smaller than a byte or types whose widths are - // not a multiple of a byte. - if (BitWidth < 8 || InBitWidth < 8 || - BitWidth % 8 != 0 || InBitWidth % 8 != 0) - return false; - } - - // Pick the largest of the two vector types. - ScalarKind = Vector; - if (InBitWidth > BitWidth) - VectorTy = VInTy; - - return true; + return false; } /// CanConvertToScalar - V is a pointer. If we can convert the pointee and all @@ -483,13 +471,14 @@ bool ConvertToScalarInfo::MergeInVectorType(VectorType *VInTy, /// /// If we see at least one access to the value that is as a vector type, set the /// SawVec flag. -bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) { - for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { - Instruction *User = cast(*UI); +bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset, + Value* NonConstantIdx) { + for (User *U : V->users()) { + Instruction *UI = cast(U); - if (LoadInst *LI = dyn_cast(User)) { + if (LoadInst *LI = dyn_cast(UI)) { // Don't break volatile loads. - if (LI->isVolatile()) + if (!LI->isSimple()) return false; // Don't touch MMX operations. if (LI->getType()->isX86_MMXTy()) @@ -499,9 +488,9 @@ bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) { continue; } - if (StoreInst *SI = dyn_cast(User)) { + if (StoreInst *SI = dyn_cast(UI)) { // Storing the pointer, not into the value? - if (SI->getOperand(0) == V || SI->isVolatile()) return false; + if (SI->getOperand(0) == V || !SI->isSimple()) return false; // Don't touch MMX operations. if (SI->getOperand(0)->getType()->isX86_MMXTy()) return false; @@ -510,25 +499,38 @@ bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) { continue; } - if (BitCastInst *BCI = dyn_cast(User)) { + if (BitCastInst *BCI = dyn_cast(UI)) { if (!onlyUsedByLifetimeMarkers(BCI)) IsNotTrivial = true; // Can't be mem2reg'd. - if (!CanConvertToScalar(BCI, Offset)) + if (!CanConvertToScalar(BCI, Offset, NonConstantIdx)) return false; continue; } - if (GetElementPtrInst *GEP = dyn_cast(User)) { + if (GetElementPtrInst *GEP = dyn_cast(UI)) { // If this is a GEP with a variable indices, we can't handle it. - if (!GEP->hasAllConstantIndices()) + PointerType* PtrTy = dyn_cast(GEP->getPointerOperandType()); + if (!PtrTy) return false; // Compute the offset that this GEP adds to the pointer. SmallVector Indices(GEP->op_begin()+1, GEP->op_end()); - uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), + Value *GEPNonConstantIdx = nullptr; + if (!GEP->hasAllConstantIndices()) { + if (!isa(PtrTy->getElementType())) + return false; + if (NonConstantIdx) + return false; + GEPNonConstantIdx = Indices.pop_back_val(); + if (!GEPNonConstantIdx->getType()->isIntegerTy(32)) + return false; + HadDynamicAccess = true; + } else + GEPNonConstantIdx = NonConstantIdx; + uint64_t GEPOffset = DL.getIndexedOffset(PtrTy, Indices); // See if all uses can be converted. - if (!CanConvertToScalar(GEP, Offset+GEPOffset)) + if (!CanConvertToScalar(GEP, Offset+GEPOffset, GEPNonConstantIdx)) return false; IsNotTrivial = true; // Can't be mem2reg'd. HadNonMemTransferAccess = true; @@ -537,7 +539,10 @@ bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) { // If this is a constant sized memset of a constant value (e.g. 0) we can // handle it. - if (MemSetInst *MSI = dyn_cast(User)) { + if (MemSetInst *MSI = dyn_cast(UI)) { + // Store to dynamic index. + if (NonConstantIdx) + return false; // Store of constant value. if (!isa(MSI->getValue())) return false; @@ -561,9 +566,12 @@ bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) { // If this is a memcpy or memmove into or out of the whole allocation, we // can handle it like a load or store of the scalar type. - if (MemTransferInst *MTI = dyn_cast(User)) { + if (MemTransferInst *MTI = dyn_cast(UI)) { + // Store to dynamic index. + if (NonConstantIdx) + return false; ConstantInt *Len = dyn_cast(MTI->getLength()); - if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0) + if (!Len || Len->getZExtValue() != AllocaSize || Offset != 0) return false; IsNotTrivial = true; // Can't be mem2reg'd. @@ -571,7 +579,7 @@ bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) { } // If this is a lifetime intrinsic, we can handle it. - if (IntrinsicInst *II = dyn_cast(User)) { + if (IntrinsicInst *II = dyn_cast(UI)) { if (II->getIntrinsicID() == Intrinsic::lifetime_start || II->getIntrinsicID() == Intrinsic::lifetime_end) { continue; @@ -593,12 +601,13 @@ bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) { /// Offset is an offset from the original alloca, in bits that need to be /// shifted to the right. By the end of this, there should be no uses of Ptr. void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, - uint64_t Offset) { + uint64_t Offset, + Value* NonConstantIdx) { while (!Ptr->use_empty()) { - Instruction *User = cast(Ptr->use_back()); + Instruction *User = cast(Ptr->user_back()); if (BitCastInst *CI = dyn_cast(User)) { - ConvertUsesToScalar(CI, NewAI, Offset); + ConvertUsesToScalar(CI, NewAI, Offset, NonConstantIdx); CI->eraseFromParent(); continue; } @@ -606,9 +615,16 @@ void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, if (GetElementPtrInst *GEP = dyn_cast(User)) { // Compute the offset that this GEP adds to the pointer. SmallVector Indices(GEP->op_begin()+1, GEP->op_end()); - uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), + Value* GEPNonConstantIdx = nullptr; + if (!GEP->hasAllConstantIndices()) { + assert(!NonConstantIdx && + "Dynamic GEP reading from dynamic GEP unsupported"); + GEPNonConstantIdx = Indices.pop_back_val(); + } else + GEPNonConstantIdx = NonConstantIdx; + uint64_t GEPOffset = DL.getIndexedOffset(GEP->getPointerOperandType(), Indices); - ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); + ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8, GEPNonConstantIdx); GEP->eraseFromParent(); continue; } @@ -617,9 +633,10 @@ void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, if (LoadInst *LI = dyn_cast(User)) { // The load is a bit extract from NewAI shifted right by Offset bits. - Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); + Value *LoadedVal = Builder.CreateLoad(NewAI); Value *NewLoadVal - = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); + = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, + NonConstantIdx, Builder); LI->replaceAllUsesWith(NewLoadVal); LI->eraseFromParent(); continue; @@ -629,7 +646,7 @@ void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, assert(SI->getOperand(0) != Ptr && "Consistency error!"); Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, - Builder); + NonConstantIdx, Builder); Builder.CreateStore(New, NewAI); SI->eraseFromParent(); @@ -644,8 +661,10 @@ void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, // transform it into a store of the expanded constant value. if (MemSetInst *MSI = dyn_cast(User)) { assert(MSI->getRawDest() == Ptr && "Consistency error!"); - unsigned NumBytes = cast(MSI->getLength())->getZExtValue(); - if (NumBytes != 0) { + assert(!NonConstantIdx && "Cannot replace dynamic memset with insert"); + int64_t SNumBytes = cast(MSI->getLength())->getSExtValue(); + if (SNumBytes > 0 && (SNumBytes >> 32) == 0) { + unsigned NumBytes = static_cast(SNumBytes); unsigned Val = cast(MSI->getValue())->getZExtValue(); // Compute the value replicated the right number of times. @@ -659,7 +678,7 @@ void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); Value *New = ConvertScalar_InsertValue( ConstantInt::get(User->getContext(), APVal), - Old, Offset, Builder); + Old, Offset, nullptr, Builder); Builder.CreateStore(New, NewAI); // If the load we just inserted is now dead, then the memset overwrote @@ -675,13 +694,14 @@ void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, // can handle it like a load or store of the scalar type. if (MemTransferInst *MTI = dyn_cast(User)) { assert(Offset == 0 && "must be store to start of alloca"); + assert(!NonConstantIdx && "Cannot replace dynamic transfer with insert"); // If the source and destination are both to the same alloca, then this is // a noop copy-to-self, just delete it. Otherwise, emit a load and store // as appropriate. - AllocaInst *OrigAI = cast(GetUnderlyingObject(Ptr, &TD, 0)); + AllocaInst *OrigAI = cast(GetUnderlyingObject(Ptr, DL, 0)); - if (GetUnderlyingObject(MTI->getSource(), &TD, 0) != OrigAI) { + if (GetUnderlyingObject(MTI->getSource(), DL, 0) != OrigAI) { // Dest must be OrigAI, change this to be a load from the original // pointer (bitcasted), then a store to our new alloca. assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); @@ -697,7 +717,7 @@ void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); SrcVal->setAlignment(MTI->getAlignment()); Builder.CreateStore(SrcVal, NewAI); - } else if (GetUnderlyingObject(MTI->getDest(), &TD, 0) != OrigAI) { + } else if (GetUnderlyingObject(MTI->getDest(), DL, 0) != OrigAI) { // Src must be OrigAI, change this to be a load from NewAI then a store // through the original dest pointer (bitcasted). assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); @@ -735,63 +755,6 @@ void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, } } -/// getScaledElementType - Gets a scaled element type for a partial vector -/// access of an alloca. The input types must be integer or floating-point -/// scalar or vector types, and the resulting type is an integer, float or -/// double. -static Type *getScaledElementType(Type *Ty1, Type *Ty2, - unsigned NewBitWidth) { - bool IsFP1 = Ty1->isFloatingPointTy() || - (Ty1->isVectorTy() && - cast(Ty1)->getElementType()->isFloatingPointTy()); - bool IsFP2 = Ty2->isFloatingPointTy() || - (Ty2->isVectorTy() && - cast(Ty2)->getElementType()->isFloatingPointTy()); - - LLVMContext &Context = Ty1->getContext(); - - // Prefer floating-point types over integer types, as integer types may have - // been created by earlier scalar replacement. - if (IsFP1 || IsFP2) { - if (NewBitWidth == 32) - return Type::getFloatTy(Context); - if (NewBitWidth == 64) - return Type::getDoubleTy(Context); - } - - return Type::getIntNTy(Context, NewBitWidth); -} - -/// CreateShuffleVectorCast - Creates a shuffle vector to convert one vector -/// to another vector of the same element type which has the same allocation -/// size but different primitive sizes (e.g. <3 x i32> and <4 x i32>). -static Value *CreateShuffleVectorCast(Value *FromVal, Type *ToType, - IRBuilder<> &Builder) { - Type *FromType = FromVal->getType(); - VectorType *FromVTy = cast(FromType); - VectorType *ToVTy = cast(ToType); - assert((ToVTy->getElementType() == FromVTy->getElementType()) && - "Vectors must have the same element type"); - Value *UnV = UndefValue::get(FromType); - unsigned numEltsFrom = FromVTy->getNumElements(); - unsigned numEltsTo = ToVTy->getNumElements(); - - SmallVector Args; - Type* Int32Ty = Builder.getInt32Ty(); - unsigned minNumElts = std::min(numEltsFrom, numEltsTo); - unsigned i; - for (i=0; i != minNumElts; ++i) - Args.push_back(ConstantInt::get(Int32Ty, i)); - - if (i < numEltsTo) { - Constant* UnC = UndefValue::get(Int32Ty); - for (; i != numEltsTo; ++i) - Args.push_back(UnC); - } - Constant *Mask = ConstantVector::get(Args); - return Builder.CreateShuffleVector(FromVal, UnV, Mask, "tmpV"); -} - /// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer /// or vector value FromVal, extracting the bits from the offset specified by /// Offset. This returns the value, which is of type ToType. @@ -804,7 +767,8 @@ static Value *CreateShuffleVectorCast(Value *FromVal, Type *ToType, /// shifted to the right. Value *ConvertToScalarInfo:: ConvertScalar_ExtractValue(Value *FromVal, Type *ToType, - uint64_t Offset, IRBuilder<> &Builder) { + uint64_t Offset, Value* NonConstantIdx, + IRBuilder<> &Builder) { // If the load is of the whole new alloca, no conversion is needed. Type *FromType = FromVal->getType(); if (FromType == ToType && Offset == 0) @@ -813,77 +777,61 @@ ConvertScalar_ExtractValue(Value *FromVal, Type *ToType, // If the result alloca is a vector type, this is either an element // access or a bitcast to another vector type of the same size. if (VectorType *VTy = dyn_cast(FromType)) { - unsigned FromTypeSize = TD.getTypeAllocSize(FromType); - unsigned ToTypeSize = TD.getTypeAllocSize(ToType); - if (FromTypeSize == ToTypeSize) { - // If the two types have the same primitive size, use a bit cast. - // Otherwise, it is two vectors with the same element type that has - // the same allocation size but different number of elements so use - // a shuffle vector. - if (FromType->getPrimitiveSizeInBits() == - ToType->getPrimitiveSizeInBits()) - return Builder.CreateBitCast(FromVal, ToType, "tmp"); - else - return CreateShuffleVectorCast(FromVal, ToType, Builder); - } - - if (isPowerOf2_64(FromTypeSize / ToTypeSize)) { - assert(!(ToType->isVectorTy() && Offset != 0) && "Can't extract a value " - "of a smaller vector type at a nonzero offset."); - - Type *CastElementTy = getScaledElementType(FromType, ToType, - ToTypeSize * 8); - unsigned NumCastVectorElements = FromTypeSize / ToTypeSize; - - LLVMContext &Context = FromVal->getContext(); - Type *CastTy = VectorType::get(CastElementTy, - NumCastVectorElements); - Value *Cast = Builder.CreateBitCast(FromVal, CastTy, "tmp"); - - unsigned EltSize = TD.getTypeAllocSizeInBits(CastElementTy); - unsigned Elt = Offset/EltSize; - assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); - Value *Extract = Builder.CreateExtractElement(Cast, ConstantInt::get( - Type::getInt32Ty(Context), Elt), "tmp"); - return Builder.CreateBitCast(Extract, ToType, "tmp"); - } + unsigned FromTypeSize = DL.getTypeAllocSize(FromType); + unsigned ToTypeSize = DL.getTypeAllocSize(ToType); + if (FromTypeSize == ToTypeSize) + return Builder.CreateBitCast(FromVal, ToType); // Otherwise it must be an element access. unsigned Elt = 0; if (Offset) { - unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); + unsigned EltSize = DL.getTypeAllocSizeInBits(VTy->getElementType()); Elt = Offset/EltSize; assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); } // Return the element extracted out of it. - Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get( - Type::getInt32Ty(FromVal->getContext()), Elt), "tmp"); + Value *Idx; + if (NonConstantIdx) { + if (Elt) + Idx = Builder.CreateAdd(NonConstantIdx, + Builder.getInt32(Elt), + "dyn.offset"); + else + Idx = NonConstantIdx; + } else + Idx = Builder.getInt32(Elt); + Value *V = Builder.CreateExtractElement(FromVal, Idx); if (V->getType() != ToType) - V = Builder.CreateBitCast(V, ToType, "tmp"); + V = Builder.CreateBitCast(V, ToType); return V; } // If ToType is a first class aggregate, extract out each of the pieces and // use insertvalue's to form the FCA. if (StructType *ST = dyn_cast(ToType)) { - const StructLayout &Layout = *TD.getStructLayout(ST); + assert(!NonConstantIdx && + "Dynamic indexing into struct types not supported"); + const StructLayout &Layout = *DL.getStructLayout(ST); Value *Res = UndefValue::get(ST); for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), Offset+Layout.getElementOffsetInBits(i), - Builder); - Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); + nullptr, Builder); + Res = Builder.CreateInsertValue(Res, Elt, i); } return Res; } if (ArrayType *AT = dyn_cast(ToType)) { - uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); + assert(!NonConstantIdx && + "Dynamic indexing into array types not supported"); + uint64_t EltSize = DL.getTypeAllocSizeInBits(AT->getElementType()); Value *Res = UndefValue::get(AT); for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), - Offset+i*EltSize, Builder); - Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); + Offset+i*EltSize, nullptr, + Builder); + Res = Builder.CreateInsertValue(Res, Elt, i); } return Res; } @@ -894,12 +842,12 @@ ConvertScalar_ExtractValue(Value *FromVal, Type *ToType, // If this is a big-endian system and the load is narrower than the // full alloca type, we need to do a shift to get the right bits. int ShAmt = 0; - if (TD.isBigEndian()) { + if (DL.isBigEndian()) { // On big-endian machines, the lowest bit is stored at the bit offset // from the pointer given by getTypeStoreSizeInBits. This matters for // integers with a bitwidth that is not a multiple of 8. - ShAmt = TD.getTypeStoreSizeInBits(NTy) - - TD.getTypeStoreSizeInBits(ToType) - Offset; + ShAmt = DL.getTypeStoreSizeInBits(NTy) - + DL.getTypeStoreSizeInBits(ToType) - Offset; } else { ShAmt = Offset; } @@ -909,33 +857,31 @@ ConvertScalar_ExtractValue(Value *FromVal, Type *ToType, // only some bits are used. if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) FromVal = Builder.CreateLShr(FromVal, - ConstantInt::get(FromVal->getType(), - ShAmt), "tmp"); + ConstantInt::get(FromVal->getType(), ShAmt)); else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) FromVal = Builder.CreateShl(FromVal, - ConstantInt::get(FromVal->getType(), - -ShAmt), "tmp"); + ConstantInt::get(FromVal->getType(), -ShAmt)); // Finally, unconditionally truncate the integer to the right width. - unsigned LIBitWidth = TD.getTypeSizeInBits(ToType); + unsigned LIBitWidth = DL.getTypeSizeInBits(ToType); if (LIBitWidth < NTy->getBitWidth()) FromVal = Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), - LIBitWidth), "tmp"); + LIBitWidth)); else if (LIBitWidth > NTy->getBitWidth()) FromVal = Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), - LIBitWidth), "tmp"); + LIBitWidth)); // If the result is an integer, this is a trunc or bitcast. if (ToType->isIntegerTy()) { // Should be done. } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { // Just do a bitcast, we know the sizes match up. - FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); + FromVal = Builder.CreateBitCast(FromVal, ToType); } else { // Otherwise must be a pointer. - FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); + FromVal = Builder.CreateIntToPtr(FromVal, ToType); } assert(FromVal->getType() == ToType && "Didn't convert right?"); return FromVal; @@ -950,107 +896,93 @@ ConvertScalar_ExtractValue(Value *FromVal, Type *ToType, /// /// Offset is an offset from the original alloca, in bits that need to be /// shifted to the right. +/// +/// NonConstantIdx is an index value if there was a GEP with a non-constant +/// index value. If this is 0 then all GEPs used to find this insert address +/// are constant. Value *ConvertToScalarInfo:: ConvertScalar_InsertValue(Value *SV, Value *Old, - uint64_t Offset, IRBuilder<> &Builder) { + uint64_t Offset, Value* NonConstantIdx, + IRBuilder<> &Builder) { // Convert the stored type to the actual type, shift it left to insert // then 'or' into place. Type *AllocaType = Old->getType(); LLVMContext &Context = Old->getContext(); if (VectorType *VTy = dyn_cast(AllocaType)) { - uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy); - uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType()); + uint64_t VecSize = DL.getTypeAllocSizeInBits(VTy); + uint64_t ValSize = DL.getTypeAllocSizeInBits(SV->getType()); // Changing the whole vector with memset or with an access of a different // vector type? - if (ValSize == VecSize) { - // If the two types have the same primitive size, use a bit cast. - // Otherwise, it is two vectors with the same element type that has - // the same allocation size but different number of elements so use - // a shuffle vector. - if (VTy->getPrimitiveSizeInBits() == - SV->getType()->getPrimitiveSizeInBits()) - return Builder.CreateBitCast(SV, AllocaType, "tmp"); - else - return CreateShuffleVectorCast(SV, VTy, Builder); - } - - if (isPowerOf2_64(VecSize / ValSize)) { - assert(!(SV->getType()->isVectorTy() && Offset != 0) && "Can't insert a " - "value of a smaller vector type at a nonzero offset."); - - Type *CastElementTy = getScaledElementType(VTy, SV->getType(), - ValSize); - unsigned NumCastVectorElements = VecSize / ValSize; - - LLVMContext &Context = SV->getContext(); - Type *OldCastTy = VectorType::get(CastElementTy, - NumCastVectorElements); - Value *OldCast = Builder.CreateBitCast(Old, OldCastTy, "tmp"); - - Value *SVCast = Builder.CreateBitCast(SV, CastElementTy, "tmp"); - - unsigned EltSize = TD.getTypeAllocSizeInBits(CastElementTy); - unsigned Elt = Offset/EltSize; - assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); - Value *Insert = - Builder.CreateInsertElement(OldCast, SVCast, ConstantInt::get( - Type::getInt32Ty(Context), Elt), "tmp"); - return Builder.CreateBitCast(Insert, AllocaType, "tmp"); - } + if (ValSize == VecSize) + return Builder.CreateBitCast(SV, AllocaType); // Must be an element insertion. - assert(SV->getType() == VTy->getElementType()); - uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); + Type *EltTy = VTy->getElementType(); + if (SV->getType() != EltTy) + SV = Builder.CreateBitCast(SV, EltTy); + uint64_t EltSize = DL.getTypeAllocSizeInBits(EltTy); unsigned Elt = Offset/EltSize; - return Builder.CreateInsertElement(Old, SV, - ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt), - "tmp"); + Value *Idx; + if (NonConstantIdx) { + if (Elt) + Idx = Builder.CreateAdd(NonConstantIdx, + Builder.getInt32(Elt), + "dyn.offset"); + else + Idx = NonConstantIdx; + } else + Idx = Builder.getInt32(Elt); + return Builder.CreateInsertElement(Old, SV, Idx); } // If SV is a first-class aggregate value, insert each value recursively. if (StructType *ST = dyn_cast(SV->getType())) { - const StructLayout &Layout = *TD.getStructLayout(ST); + assert(!NonConstantIdx && + "Dynamic indexing into struct types not supported"); + const StructLayout &Layout = *DL.getStructLayout(ST); for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { - Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); + Value *Elt = Builder.CreateExtractValue(SV, i); Old = ConvertScalar_InsertValue(Elt, Old, Offset+Layout.getElementOffsetInBits(i), - Builder); + nullptr, Builder); } return Old; } if (ArrayType *AT = dyn_cast(SV->getType())) { - uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); + assert(!NonConstantIdx && + "Dynamic indexing into array types not supported"); + uint64_t EltSize = DL.getTypeAllocSizeInBits(AT->getElementType()); for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { - Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); - Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); + Value *Elt = Builder.CreateExtractValue(SV, i); + Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, nullptr, + Builder); } return Old; } // If SV is a float, convert it to the appropriate integer type. // If it is a pointer, do the same. - unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); - unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); - unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); - unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); + unsigned SrcWidth = DL.getTypeSizeInBits(SV->getType()); + unsigned DestWidth = DL.getTypeSizeInBits(AllocaType); + unsigned SrcStoreWidth = DL.getTypeStoreSizeInBits(SV->getType()); + unsigned DestStoreWidth = DL.getTypeStoreSizeInBits(AllocaType); if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) - SV = Builder.CreateBitCast(SV, - IntegerType::get(SV->getContext(),SrcWidth), "tmp"); + SV = Builder.CreateBitCast(SV, IntegerType::get(SV->getContext(),SrcWidth)); else if (SV->getType()->isPointerTy()) - SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()), "tmp"); + SV = Builder.CreatePtrToInt(SV, DL.getIntPtrType(SV->getType())); // Zero extend or truncate the value if needed. if (SV->getType() != AllocaType) { if (SV->getType()->getPrimitiveSizeInBits() < AllocaType->getPrimitiveSizeInBits()) - SV = Builder.CreateZExt(SV, AllocaType, "tmp"); + SV = Builder.CreateZExt(SV, AllocaType); else { // Truncation may be needed if storing more than the alloca can hold // (undefined behavior). - SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); + SV = Builder.CreateTrunc(SV, AllocaType); SrcWidth = DestWidth; SrcStoreWidth = DestStoreWidth; } @@ -1059,7 +991,7 @@ ConvertScalar_InsertValue(Value *SV, Value *Old, // If this is a big-endian system and the store is narrower than the // full alloca type, we need to do a shift to get the right bits. int ShAmt = 0; - if (TD.isBigEndian()) { + if (DL.isBigEndian()) { // On big-endian machines, the lowest bit is stored at the bit offset // from the pointer given by getTypeStoreSizeInBits. This matters for // integers with a bitwidth that is not a multiple of 8. @@ -1073,12 +1005,10 @@ ConvertScalar_InsertValue(Value *SV, Value *Old, // only some bits in the structure are set. APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { - SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), - ShAmt), "tmp"); + SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), ShAmt)); Mask <<= ShAmt; } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { - SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), - -ShAmt), "tmp"); + SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), -ShAmt)); Mask = Mask.lshr(-ShAmt); } @@ -1099,16 +1029,11 @@ ConvertScalar_InsertValue(Value *SV, Value *Old, bool SROA::runOnFunction(Function &F) { - TD = getAnalysisIfAvailable(); + if (skipOptnoneFunction(F)) + return false; bool Changed = performPromotion(F); - // FIXME: ScalarRepl currently depends on TargetData more than it - // theoretically needs to. It should be refactored in order to support - // target-independent IR. Until this is done, just skip the actual - // scalar-replacement portion of this pass. - if (!TD) return Changed; - while (1) { bool LocalChange = performScalarRepl(F); if (!LocalChange) break; // No need to repromote if no scalarrepl @@ -1127,44 +1052,46 @@ class AllocaPromoter : public LoadAndStorePromoter { SmallVector DDIs; SmallVector DVIs; public: - AllocaPromoter(const SmallVectorImpl &Insts, SSAUpdater &S, + AllocaPromoter(ArrayRef Insts, SSAUpdater &S, DIBuilder *DB) - : LoadAndStorePromoter(Insts, S), AI(0), DIB(DB) {} - + : LoadAndStorePromoter(Insts, S), AI(nullptr), DIB(DB) {} + void run(AllocaInst *AI, const SmallVectorImpl &Insts) { // Remember which alloca we're promoting (for isInstInList). this->AI = AI; - if (MDNode *DebugNode = MDNode::getIfExists(AI->getContext(), AI)) - for (Value::use_iterator UI = DebugNode->use_begin(), - E = DebugNode->use_end(); UI != E; ++UI) - if (DbgDeclareInst *DDI = dyn_cast(*UI)) - DDIs.push_back(DDI); - else if (DbgValueInst *DVI = dyn_cast(*UI)) - DVIs.push_back(DVI); + if (auto *L = LocalAsMetadata::getIfExists(AI)) { + if (auto *DINode = MetadataAsValue::getIfExists(AI->getContext(), L)) { + for (User *U : DINode->users()) + if (DbgDeclareInst *DDI = dyn_cast(U)) + DDIs.push_back(DDI); + else if (DbgValueInst *DVI = dyn_cast(U)) + DVIs.push_back(DVI); + } + } LoadAndStorePromoter::run(Insts); AI->eraseFromParent(); - for (SmallVector::iterator I = DDIs.begin(), + for (SmallVectorImpl::iterator I = DDIs.begin(), E = DDIs.end(); I != E; ++I) { DbgDeclareInst *DDI = *I; DDI->eraseFromParent(); } - for (SmallVector::iterator I = DVIs.begin(), + for (SmallVectorImpl::iterator I = DVIs.begin(), E = DVIs.end(); I != E; ++I) { DbgValueInst *DVI = *I; DVI->eraseFromParent(); } } - - virtual bool isInstInList(Instruction *I, - const SmallVectorImpl &Insts) const { + + bool isInstInList(Instruction *I, + const SmallVectorImpl &Insts) const override { if (LoadInst *LI = dyn_cast(I)) return LI->getOperand(0) == AI; return cast(I)->getPointerOperand() == AI; } - virtual void updateDebugInfo(Instruction *Inst) const { - for (SmallVector::const_iterator I = DDIs.begin(), + void updateDebugInfo(Instruction *Inst) const override { + for (SmallVectorImpl::const_iterator I = DDIs.begin(), E = DDIs.end(); I != E; ++I) { DbgDeclareInst *DDI = *I; if (StoreInst *SI = dyn_cast(Inst)) @@ -1172,33 +1099,27 @@ public: else if (LoadInst *LI = dyn_cast(Inst)) ConvertDebugDeclareToDebugValue(DDI, LI, *DIB); } - for (SmallVector::const_iterator I = DVIs.begin(), + for (SmallVectorImpl::const_iterator I = DVIs.begin(), E = DVIs.end(); I != E; ++I) { DbgValueInst *DVI = *I; + Value *Arg = nullptr; if (StoreInst *SI = dyn_cast(Inst)) { - Instruction *DbgVal = NULL; // If an argument is zero extended then use argument directly. The ZExt // may be zapped by an optimization pass in future. - Argument *ExtendedArg = NULL; if (ZExtInst *ZExt = dyn_cast(SI->getOperand(0))) - ExtendedArg = dyn_cast(ZExt->getOperand(0)); + Arg = dyn_cast(ZExt->getOperand(0)); if (SExtInst *SExt = dyn_cast(SI->getOperand(0))) - ExtendedArg = dyn_cast(SExt->getOperand(0)); - if (ExtendedArg) - DbgVal = DIB->insertDbgValueIntrinsic(ExtendedArg, 0, - DIVariable(DVI->getVariable()), - SI); - else - DbgVal = DIB->insertDbgValueIntrinsic(SI->getOperand(0), 0, - DIVariable(DVI->getVariable()), - SI); - DbgVal->setDebugLoc(DVI->getDebugLoc()); + Arg = dyn_cast(SExt->getOperand(0)); + if (!Arg) + Arg = SI->getOperand(0); } else if (LoadInst *LI = dyn_cast(Inst)) { - Instruction *DbgVal = - DIB->insertDbgValueIntrinsic(LI->getOperand(0), 0, - DIVariable(DVI->getVariable()), LI); - DbgVal->setDebugLoc(DVI->getDebugLoc()); + Arg = LI->getOperand(0); + } else { + continue; } + DIB->insertDbgValueIntrinsic(Arg, 0, DVI->getVariable(), + DVI->getExpression(), DVI->getDebugLoc(), + Inst); } } }; @@ -1217,25 +1138,27 @@ public: /// /// We can do this to a select if its only uses are loads and if the operand to /// the select can be loaded unconditionally. -static bool isSafeSelectToSpeculate(SelectInst *SI, const TargetData *TD) { - bool TDerefable = SI->getTrueValue()->isDereferenceablePointer(); - bool FDerefable = SI->getFalseValue()->isDereferenceablePointer(); - - for (Value::use_iterator UI = SI->use_begin(), UE = SI->use_end(); - UI != UE; ++UI) { - LoadInst *LI = dyn_cast(*UI); - if (LI == 0 || LI->isVolatile()) return false; - +static bool isSafeSelectToSpeculate(SelectInst *SI) { + const DataLayout &DL = SI->getModule()->getDataLayout(); + bool TDerefable = isDereferenceablePointer(SI->getTrueValue(), DL); + bool FDerefable = isDereferenceablePointer(SI->getFalseValue(), DL); + + for (User *U : SI->users()) { + LoadInst *LI = dyn_cast(U); + if (!LI || !LI->isSimple()) return false; + // Both operands to the select need to be dereferencable, either absolutely // (e.g. allocas) or at this point because we can see other accesses to it. - if (!TDerefable && !isSafeToLoadUnconditionally(SI->getTrueValue(), LI, - LI->getAlignment(), TD)) + if (!TDerefable && + !isSafeToLoadUnconditionally(SI->getTrueValue(), LI, + LI->getAlignment())) return false; - if (!FDerefable && !isSafeToLoadUnconditionally(SI->getFalseValue(), LI, - LI->getAlignment(), TD)) + if (!FDerefable && + !isSafeToLoadUnconditionally(SI->getFalseValue(), LI, + LI->getAlignment())) return false; } - + return true; } @@ -1255,58 +1178,63 @@ static bool isSafeSelectToSpeculate(SelectInst *SI, const TargetData *TD) { /// /// We can do this to a select if its only uses are loads and if the operand to /// the select can be loaded unconditionally. -static bool isSafePHIToSpeculate(PHINode *PN, const TargetData *TD) { +static bool isSafePHIToSpeculate(PHINode *PN) { // For now, we can only do this promotion if the load is in the same block as // the PHI, and if there are no stores between the phi and load. // TODO: Allow recursive phi users. // TODO: Allow stores. BasicBlock *BB = PN->getParent(); unsigned MaxAlign = 0; - for (Value::use_iterator UI = PN->use_begin(), UE = PN->use_end(); - UI != UE; ++UI) { - LoadInst *LI = dyn_cast(*UI); - if (LI == 0 || LI->isVolatile()) return false; - + for (User *U : PN->users()) { + LoadInst *LI = dyn_cast(U); + if (!LI || !LI->isSimple()) return false; + // For now we only allow loads in the same block as the PHI. This is a // common case that happens when instcombine merges two loads through a PHI. if (LI->getParent() != BB) return false; - + // Ensure that there are no instructions between the PHI and the load that // could store. for (BasicBlock::iterator BBI = PN; &*BBI != LI; ++BBI) if (BBI->mayWriteToMemory()) return false; - + MaxAlign = std::max(MaxAlign, LI->getAlignment()); } - + + const DataLayout &DL = PN->getModule()->getDataLayout(); + // Okay, we know that we have one or more loads in the same block as the PHI. // We can transform this if it is safe to push the loads into the predecessor // blocks. The only thing to watch out for is that we can't put a possibly // trapping load in the predecessor if it is a critical edge. for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { BasicBlock *Pred = PN->getIncomingBlock(i); + Value *InVal = PN->getIncomingValue(i); + + // If the terminator of the predecessor has side-effects (an invoke), + // there is no safe place to put a load in the predecessor. + if (Pred->getTerminator()->mayHaveSideEffects()) + return false; + + // If the value is produced by the terminator of the predecessor + // (an invoke), there is no valid place to put a load in the predecessor. + if (Pred->getTerminator() == InVal) + return false; // If the predecessor has a single successor, then the edge isn't critical. if (Pred->getTerminator()->getNumSuccessors() == 1) continue; - - Value *InVal = PN->getIncomingValue(i); - - // If the InVal is an invoke in the pred, we can't put a load on the edge. - if (InvokeInst *II = dyn_cast(InVal)) - if (II->getParent() == Pred) - return false; // If this pointer is always safe to load, or if we can prove that there is // already a load in the block, then we can move the load to the pred block. - if (InVal->isDereferenceablePointer() || - isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign, TD)) + if (isDereferenceablePointer(InVal, DL) || + isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign)) continue; - + return false; } - + return true; } @@ -1315,21 +1243,18 @@ static bool isSafePHIToSpeculate(PHINode *PN, const TargetData *TD) { /// direct (non-volatile) loads and stores to it. If the alloca is close but /// not quite there, this will transform the code to allow promotion. As such, /// it is a non-pure predicate. -static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const TargetData *TD) { +static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const DataLayout &DL) { SetVector, SmallPtrSet > InstsToRewrite; - - for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end(); - UI != UE; ++UI) { - User *U = *UI; + for (User *U : AI->users()) { if (LoadInst *LI = dyn_cast(U)) { - if (LI->isVolatile()) + if (!LI->isSimple()) return false; continue; } - + if (StoreInst *SI = dyn_cast(U)) { - if (SI->getOperand(0) == AI || SI->isVolatile()) + if (SI->getOperand(0) == AI || !SI->isSimple()) return false; // Don't allow a store OF the AI, only INTO the AI. continue; } @@ -1341,43 +1266,43 @@ static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const TargetData *TD) { Value *Result = SI->getOperand(1+CI->isZero()); SI->replaceAllUsesWith(Result); SI->eraseFromParent(); - + // This is very rare and we just scrambled the use list of AI, start // over completely. - return tryToMakeAllocaBePromotable(AI, TD); + return tryToMakeAllocaBePromotable(AI, DL); } // If it is safe to turn "load (select c, AI, ptr)" into a select of two // loads, then we can transform this by rewriting the select. - if (!isSafeSelectToSpeculate(SI, TD)) + if (!isSafeSelectToSpeculate(SI)) return false; - + InstsToRewrite.insert(SI); continue; } - + if (PHINode *PN = dyn_cast(U)) { if (PN->use_empty()) { // Dead PHIs can be stripped. InstsToRewrite.insert(PN); continue; } - + // If it is safe to turn "load (phi [AI, ptr, ...])" into a PHI of loads // in the pred blocks, then we can transform this by rewriting the PHI. - if (!isSafePHIToSpeculate(PN, TD)) + if (!isSafePHIToSpeculate(PN)) return false; - + InstsToRewrite.insert(PN); continue; } - + if (BitCastInst *BCI = dyn_cast(U)) { if (onlyUsedByLifetimeMarkers(BCI)) { InstsToRewrite.insert(BCI); continue; } } - + return false; } @@ -1385,18 +1310,15 @@ static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const TargetData *TD) { // we're done! if (InstsToRewrite.empty()) return true; - + // If we have instructions that need to be rewritten for this to be promotable // take care of it now. for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) { if (BitCastInst *BCI = dyn_cast(InstsToRewrite[i])) { // This could only be a bitcast used by nothing but lifetime intrinsics. - for (BitCastInst::use_iterator I = BCI->use_begin(), E = BCI->use_end(); - I != E;) { - Use &U = I.getUse(); - ++I; - cast(U.getUser())->eraseFromParent(); - } + for (BitCastInst::user_iterator I = BCI->user_begin(), E = BCI->user_end(); + I != E;) + cast(*I++)->eraseFromParent(); BCI->eraseFromParent(); continue; } @@ -1405,33 +1327,36 @@ static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const TargetData *TD) { // Selects in InstsToRewrite only have load uses. Rewrite each as two // loads with a new select. while (!SI->use_empty()) { - LoadInst *LI = cast(SI->use_back()); - + LoadInst *LI = cast(SI->user_back()); + IRBuilder<> Builder(LI); - LoadInst *TrueLoad = + LoadInst *TrueLoad = Builder.CreateLoad(SI->getTrueValue(), LI->getName()+".t"); - LoadInst *FalseLoad = + LoadInst *FalseLoad = Builder.CreateLoad(SI->getFalseValue(), LI->getName()+".f"); - - // Transfer alignment and TBAA info if present. + + // Transfer alignment and AA info if present. TrueLoad->setAlignment(LI->getAlignment()); FalseLoad->setAlignment(LI->getAlignment()); - if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) { - TrueLoad->setMetadata(LLVMContext::MD_tbaa, Tag); - FalseLoad->setMetadata(LLVMContext::MD_tbaa, Tag); + + AAMDNodes Tags; + LI->getAAMetadata(Tags); + if (Tags) { + TrueLoad->setAAMetadata(Tags); + FalseLoad->setAAMetadata(Tags); } - + Value *V = Builder.CreateSelect(SI->getCondition(), TrueLoad, FalseLoad); V->takeName(LI); LI->replaceAllUsesWith(V); LI->eraseFromParent(); } - + // Now that all the loads are gone, the select is gone too. SI->eraseFromParent(); continue; } - + // Otherwise, we have a PHI node which allows us to push the loads into the // predecessors. PHINode *PN = cast(InstsToRewrite[i]); @@ -1439,57 +1364,62 @@ static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const TargetData *TD) { PN->eraseFromParent(); continue; } - + Type *LoadTy = cast(PN->getType())->getElementType(); PHINode *NewPN = PHINode::Create(LoadTy, PN->getNumIncomingValues(), PN->getName()+".ld", PN); - // Get the TBAA tag and alignment to use from one of the loads. It doesn't + // Get the AA tags and alignment to use from one of the loads. It doesn't // matter which one we get and if any differ, it doesn't matter. - LoadInst *SomeLoad = cast(PN->use_back()); - MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa); + LoadInst *SomeLoad = cast(PN->user_back()); + + AAMDNodes AATags; + SomeLoad->getAAMetadata(AATags); unsigned Align = SomeLoad->getAlignment(); - + // Rewrite all loads of the PN to use the new PHI. while (!PN->use_empty()) { - LoadInst *LI = cast(PN->use_back()); + LoadInst *LI = cast(PN->user_back()); LI->replaceAllUsesWith(NewPN); LI->eraseFromParent(); } - + // Inject loads into all of the pred blocks. Keep track of which blocks we // insert them into in case we have multiple edges from the same block. DenseMap InsertedLoads; - + for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { BasicBlock *Pred = PN->getIncomingBlock(i); LoadInst *&Load = InsertedLoads[Pred]; - if (Load == 0) { + if (!Load) { Load = new LoadInst(PN->getIncomingValue(i), PN->getName() + "." + Pred->getName(), Pred->getTerminator()); Load->setAlignment(Align); - if (TBAATag) Load->setMetadata(LLVMContext::MD_tbaa, TBAATag); + if (AATags) Load->setAAMetadata(AATags); } - + NewPN->addIncoming(Load, Pred); } - + PN->eraseFromParent(); } - + ++NumAdjusted; return true; } bool SROA::performPromotion(Function &F) { std::vector Allocas; - DominatorTree *DT = 0; + const DataLayout &DL = F.getParent()->getDataLayout(); + DominatorTree *DT = nullptr; if (HasDomTree) - DT = &getAnalysis(); + DT = &getAnalysis().getDomTree(); + AssumptionCache &AC = + getAnalysis().getAssumptionCache(F); BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function - DIBuilder DIB(*F.getParent()); + DIBuilder DIB(*F.getParent(), /*AllowUnresolved*/ false); bool Changed = false; SmallVector Insts; while (1) { @@ -1499,22 +1429,21 @@ bool SROA::performPromotion(Function &F) { // the entry node for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) if (AllocaInst *AI = dyn_cast(I)) // Is it an alloca? - if (tryToMakeAllocaBePromotable(AI, TD)) + if (tryToMakeAllocaBePromotable(AI, DL)) Allocas.push_back(AI); if (Allocas.empty()) break; if (HasDomTree) - PromoteMemToReg(Allocas, *DT); + PromoteMemToReg(Allocas, *DT, nullptr, &AC); else { SSAUpdater SSA; for (unsigned i = 0, e = Allocas.size(); i != e; ++i) { AllocaInst *AI = Allocas[i]; - + // Build list of instructions to promote. - for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); - UI != E; ++UI) - Insts.push_back(cast(*UI)); + for (User *U : AI->users()) + Insts.push_back(cast(U)); AllocaPromoter(Insts, SSA, &DIB).run(AI, Insts); Insts.clear(); } @@ -1529,25 +1458,24 @@ bool SROA::performPromotion(Function &F) { /// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for /// SROA. It must be a struct or array type with a small number of elements. -static bool ShouldAttemptScalarRepl(AllocaInst *AI) { +bool SROA::ShouldAttemptScalarRepl(AllocaInst *AI) { Type *T = AI->getAllocatedType(); - // Do not promote any struct into more than 32 separate vars. + // Do not promote any struct that has too many members. if (StructType *ST = dyn_cast(T)) - return ST->getNumElements() <= 32; - // Arrays are much less likely to be safe for SROA; only consider - // them if they are very small. + return ST->getNumElements() <= StructMemberThreshold; + // Do not promote any array that has too many elements. if (ArrayType *AT = dyn_cast(T)) - return AT->getNumElements() <= 8; + return AT->getNumElements() <= ArrayElementThreshold; return false; } - // performScalarRepl - This algorithm is a simple worklist driven algorithm, -// which runs on all of the alloca instructions in the function, removing them -// if they are only used by getelementptr instructions. +// which runs on all of the alloca instructions in the entry block, removing +// them if they are only used by getelementptr instructions. // bool SROA::performScalarRepl(Function &F) { std::vector WorkList; + const DataLayout &DL = F.getParent()->getDataLayout(); // Scan the entry basic block, adding allocas to the worklist. BasicBlock &BB = F.getEntryBlock(); @@ -1573,31 +1501,11 @@ bool SROA::performScalarRepl(Function &F) { if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) continue; - // Check to see if this allocation is only modified by a memcpy/memmove from - // a constant global. If this is the case, we can change all users to use - // the constant global instead. This is commonly produced by the CFE by - // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' - // is only subsequently read. - SmallVector ToDelete; - if (MemTransferInst *Copy = isOnlyCopiedFromConstantGlobal(AI, ToDelete)) { - DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); - DEBUG(dbgs() << " memcpy = " << *Copy << '\n'); - for (unsigned i = 0, e = ToDelete.size(); i != e; ++i) - ToDelete[i]->eraseFromParent(); - Constant *TheSrc = cast(Copy->getSource()); - AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); - Copy->eraseFromParent(); // Don't mutate the global. - AI->eraseFromParent(); - ++NumGlobals; - Changed = true; - continue; - } - // Check to see if we can perform the core SROA transformation. We cannot // transform the allocation instruction if it is an array allocation // (allocations OF arrays are ok though), and an allocation of a scalar // value cannot be decomposed at all. - uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); + uint64_t AllocaSize = DL.getTypeAllocSize(AI->getAllocatedType()); // Do not promote [0 x %struct]. if (AllocaSize == 0) continue; @@ -1621,7 +1529,8 @@ bool SROA::performScalarRepl(Function &F) { // that we can't just check based on the type: the alloca may be of an i32 // but that has pointer arithmetic to set byte 3 of it or something. if (AllocaInst *NewAI = - ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) { + ConvertToScalarInfo((unsigned)AllocaSize, DL, ScalarLoadThreshold) + .TryConvert(AI)) { NewAI->takeName(AI); AI->eraseFromParent(); ++NumConverted; @@ -1644,7 +1553,7 @@ void SROA::DoScalarReplacement(AllocaInst *AI, if (StructType *ST = dyn_cast(AI->getAllocatedType())) { ElementAllocas.reserve(ST->getNumContainedTypes()); for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { - AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, + AllocaInst *NA = new AllocaInst(ST->getContainedType(i), nullptr, AI->getAlignment(), AI->getName() + "." + Twine(i), AI); ElementAllocas.push_back(NA); @@ -1655,7 +1564,7 @@ void SROA::DoScalarReplacement(AllocaInst *AI, ElementAllocas.reserve(AT->getNumElements()); Type *ElTy = AT->getElementType(); for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { - AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), + AllocaInst *NA = new AllocaInst(ElTy, nullptr, AI->getAlignment(), AI->getName() + "." + Twine(i), AI); ElementAllocas.push_back(NA); WorkList.push_back(NA); // Add to worklist for recursive processing @@ -1684,7 +1593,7 @@ void SROA::DeleteDeadInstructions() { // Zero out the operand and see if it becomes trivially dead. // (But, don't add allocas to the dead instruction list -- they are // already on the worklist and will be deleted separately.) - *OI = 0; + *OI = nullptr; if (isInstructionTriviallyDead(U) && !isa(U)) DeadInsts.push_back(U); } @@ -1699,8 +1608,9 @@ void SROA::DeleteDeadInstructions() { /// referenced by this instruction. void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset, AllocaInfo &Info) { - for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { - Instruction *User = cast(*UI); + const DataLayout &DL = I->getModule()->getDataLayout(); + for (Use &U : I->uses()) { + Instruction *User = cast(U.getUser()); if (BitCastInst *BC = dyn_cast(User)) { isSafeForScalarRepl(BC, Offset, Info); @@ -1711,27 +1621,28 @@ void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset, isSafeForScalarRepl(GEPI, GEPOffset, Info); } else if (MemIntrinsic *MI = dyn_cast(User)) { ConstantInt *Length = dyn_cast(MI->getLength()); - if (Length == 0) + if (!Length || Length->isNegative()) return MarkUnsafe(Info, User); - isSafeMemAccess(Offset, Length->getZExtValue(), 0, - UI.getOperandNo() == 0, Info, MI, + + isSafeMemAccess(Offset, Length->getZExtValue(), nullptr, + U.getOperandNo() == 0, Info, MI, true /*AllowWholeAccess*/); } else if (LoadInst *LI = dyn_cast(User)) { - if (LI->isVolatile()) + if (!LI->isSimple()) return MarkUnsafe(Info, User); Type *LIType = LI->getType(); - isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), - LIType, false, Info, LI, true /*AllowWholeAccess*/); + isSafeMemAccess(Offset, DL.getTypeAllocSize(LIType), LIType, false, Info, + LI, true /*AllowWholeAccess*/); Info.hasALoadOrStore = true; - + } else if (StoreInst *SI = dyn_cast(User)) { // Store is ok if storing INTO the pointer, not storing the pointer - if (SI->isVolatile() || SI->getOperand(0) == I) + if (!SI->isSimple() || SI->getOperand(0) == I) return MarkUnsafe(Info, User); - + Type *SIType = SI->getOperand(0)->getType(); - isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), - SIType, true, Info, SI, true /*AllowWholeAccess*/); + isSafeMemAccess(Offset, DL.getTypeAllocSize(SIType), SIType, true, Info, + SI, true /*AllowWholeAccess*/); Info.hasALoadOrStore = true; } else if (IntrinsicInst *II = dyn_cast(User)) { if (II->getIntrinsicID() != Intrinsic::lifetime_start && @@ -1745,7 +1656,7 @@ void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset, if (Info.isUnsafe) return; } } - + /// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer /// derived from the alloca, we can often still split the alloca into elements. @@ -1760,42 +1671,43 @@ void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset, AllocaInfo &Info) { // If we've already checked this PHI, don't do it again. if (PHINode *PN = dyn_cast(I)) - if (!Info.CheckedPHIs.insert(PN)) + if (!Info.CheckedPHIs.insert(PN).second) return; - - for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { - Instruction *User = cast(*UI); - - if (BitCastInst *BC = dyn_cast(User)) { + + const DataLayout &DL = I->getModule()->getDataLayout(); + for (User *U : I->users()) { + Instruction *UI = cast(U); + + if (BitCastInst *BC = dyn_cast(UI)) { isSafePHISelectUseForScalarRepl(BC, Offset, Info); - } else if (GetElementPtrInst *GEPI = dyn_cast(User)) { + } else if (GetElementPtrInst *GEPI = dyn_cast(UI)) { // Only allow "bitcast" GEPs for simplicity. We could generalize this, // but would have to prove that we're staying inside of an element being // promoted. if (!GEPI->hasAllZeroIndices()) - return MarkUnsafe(Info, User); + return MarkUnsafe(Info, UI); isSafePHISelectUseForScalarRepl(GEPI, Offset, Info); - } else if (LoadInst *LI = dyn_cast(User)) { - if (LI->isVolatile()) - return MarkUnsafe(Info, User); + } else if (LoadInst *LI = dyn_cast(UI)) { + if (!LI->isSimple()) + return MarkUnsafe(Info, UI); Type *LIType = LI->getType(); - isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), - LIType, false, Info, LI, false /*AllowWholeAccess*/); + isSafeMemAccess(Offset, DL.getTypeAllocSize(LIType), LIType, false, Info, + LI, false /*AllowWholeAccess*/); Info.hasALoadOrStore = true; - - } else if (StoreInst *SI = dyn_cast(User)) { + + } else if (StoreInst *SI = dyn_cast(UI)) { // Store is ok if storing INTO the pointer, not storing the pointer - if (SI->isVolatile() || SI->getOperand(0) == I) - return MarkUnsafe(Info, User); - + if (!SI->isSimple() || SI->getOperand(0) == I) + return MarkUnsafe(Info, UI); + Type *SIType = SI->getOperand(0)->getType(); - isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), - SIType, true, Info, SI, false /*AllowWholeAccess*/); + isSafeMemAccess(Offset, DL.getTypeAllocSize(SIType), SIType, true, Info, + SI, false /*AllowWholeAccess*/); Info.hasALoadOrStore = true; - } else if (isa(User) || isa(User)) { - isSafePHISelectUseForScalarRepl(User, Offset, Info); + } else if (isa(UI) || isa(UI)) { + isSafePHISelectUseForScalarRepl(UI, Offset, Info); } else { - return MarkUnsafe(Info, User); + return MarkUnsafe(Info, UI); } if (Info.isUnsafe) return; } @@ -1811,6 +1723,8 @@ void SROA::isSafeGEP(GetElementPtrInst *GEPI, gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); if (GEPIt == E) return; + bool NonConstant = false; + unsigned NonConstantIdxSize = 0; // Walk through the GEP type indices, checking the types that this indexes // into. @@ -1827,8 +1741,16 @@ void SROA::isSafeGEP(GetElementPtrInst *GEPI, // Compute the offset due to this GEP and check if the alloca has a // component element at that offset. SmallVector Indices(GEPI->op_begin() + 1, GEPI->op_end()); - Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices); - if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 0)) + // If this GEP is non-constant then the last operand must have been a + // dynamic index into a vector. Pop this now as it has no impact on the + // constant part of the offset. + if (NonConstant) + Indices.pop_back(); + + const DataLayout &DL = GEPI->getModule()->getDataLayout(); + Offset += DL.getIndexedOffset(GEPI->getPointerOperandType(), Indices); + if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, NonConstantIdxSize, + DL)) MarkUnsafe(Info, GEPI); } @@ -1840,12 +1762,12 @@ static bool isHomogeneousAggregate(Type *T, unsigned &NumElts, Type *&EltTy) { if (ArrayType *AT = dyn_cast(T)) { NumElts = AT->getNumElements(); - EltTy = (NumElts == 0 ? 0 : AT->getElementType()); + EltTy = (NumElts == 0 ? nullptr : AT->getElementType()); return true; } if (StructType *ST = dyn_cast(T)) { NumElts = ST->getNumContainedTypes(); - EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0)); + EltTy = (NumElts == 0 ? nullptr : ST->getContainedType(0)); for (unsigned n = 1; n < NumElts; ++n) { if (ST->getContainedType(n) != EltTy) return false; @@ -1883,9 +1805,10 @@ void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize, Type *MemOpType, bool isStore, AllocaInfo &Info, Instruction *TheAccess, bool AllowWholeAccess) { + const DataLayout &DL = TheAccess->getModule()->getDataLayout(); // Check if this is a load/store of the entire alloca. if (Offset == 0 && AllowWholeAccess && - MemSize == TD->getTypeAllocSize(Info.AI->getAllocatedType())) { + MemSize == DL.getTypeAllocSize(Info.AI->getAllocatedType())) { // This can be safe for MemIntrinsics (where MemOpType is 0) and integer // loads/stores (which are essentially the same as the MemIntrinsics with // regard to copying padding between elements). But, if an alloca is @@ -1908,7 +1831,7 @@ void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize, } // Check if the offset/size correspond to a component within the alloca type. Type *T = Info.AI->getAllocatedType(); - if (TypeHasComponent(T, Offset, MemSize)) { + if (TypeHasComponent(T, Offset, MemSize, DL)) { Info.hasSubelementAccess = true; return; } @@ -1918,21 +1841,28 @@ void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize, /// TypeHasComponent - Return true if T has a component type with the /// specified offset and size. If Size is zero, do not check the size. -bool SROA::TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size) { +bool SROA::TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size, + const DataLayout &DL) { Type *EltTy; uint64_t EltSize; if (StructType *ST = dyn_cast(T)) { - const StructLayout *Layout = TD->getStructLayout(ST); + const StructLayout *Layout = DL.getStructLayout(ST); unsigned EltIdx = Layout->getElementContainingOffset(Offset); EltTy = ST->getContainedType(EltIdx); - EltSize = TD->getTypeAllocSize(EltTy); + EltSize = DL.getTypeAllocSize(EltTy); Offset -= Layout->getElementOffset(EltIdx); } else if (ArrayType *AT = dyn_cast(T)) { EltTy = AT->getElementType(); - EltSize = TD->getTypeAllocSize(EltTy); + EltSize = DL.getTypeAllocSize(EltTy); if (Offset >= AT->getNumElements() * EltSize) return false; Offset %= EltSize; + } else if (VectorType *VT = dyn_cast(T)) { + EltTy = VT->getElementType(); + EltSize = DL.getTypeAllocSize(EltTy); + if (Offset >= VT->getNumElements() * EltSize) + return false; + Offset %= EltSize; } else { return false; } @@ -1941,7 +1871,7 @@ bool SROA::TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size) { // Check if the component spans multiple elements. if (Offset + Size > EltSize) return false; - return TypeHasComponent(EltTy, Offset, Size); + return TypeHasComponent(EltTy, Offset, Size, DL); } /// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite @@ -1949,26 +1879,26 @@ bool SROA::TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size) { /// Offset indicates the position within AI that is referenced by this /// instruction. void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, - SmallVector &NewElts) { + SmallVectorImpl &NewElts) { + const DataLayout &DL = I->getModule()->getDataLayout(); for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) { - Use &TheUse = UI.getUse(); - Instruction *User = cast(*UI++); + Use &TheUse = *UI++; + Instruction *User = cast(TheUse.getUser()); if (BitCastInst *BC = dyn_cast(User)) { RewriteBitCast(BC, AI, Offset, NewElts); continue; } - + if (GetElementPtrInst *GEPI = dyn_cast(User)) { RewriteGEP(GEPI, AI, Offset, NewElts); continue; } - + if (MemIntrinsic *MI = dyn_cast(User)) { ConstantInt *Length = dyn_cast(MI->getLength()); uint64_t MemSize = Length->getZExtValue(); - if (Offset == 0 && - MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) + if (Offset == 0 && MemSize == DL.getTypeAllocSize(AI->getAllocatedType())) RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); // Otherwise the intrinsic can only touch a single element and the // address operand will be updated, so nothing else needs to be done. @@ -1982,10 +1912,10 @@ void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, } continue; } - + if (LoadInst *LI = dyn_cast(User)) { Type *LIType = LI->getType(); - + if (isCompatibleAggregate(LIType, AI->getAllocatedType())) { // Replace: // %res = load { i32, i32 }* %alloc @@ -2004,14 +1934,14 @@ void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, LI->replaceAllUsesWith(Insert); DeadInsts.push_back(LI); } else if (LIType->isIntegerTy() && - TD->getTypeAllocSize(LIType) == - TD->getTypeAllocSize(AI->getAllocatedType())) { + DL.getTypeAllocSize(LIType) == + DL.getTypeAllocSize(AI->getAllocatedType())) { // If this is a load of the entire alloca to an integer, rewrite it. RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); } continue; } - + if (StoreInst *SI = dyn_cast(User)) { Value *Val = SI->getOperand(0); Type *SIType = Val->getType(); @@ -2031,23 +1961,23 @@ void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, } DeadInsts.push_back(SI); } else if (SIType->isIntegerTy() && - TD->getTypeAllocSize(SIType) == - TD->getTypeAllocSize(AI->getAllocatedType())) { + DL.getTypeAllocSize(SIType) == + DL.getTypeAllocSize(AI->getAllocatedType())) { // If this is a store of the entire alloca from an integer, rewrite it. RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); } continue; } - + if (isa(User) || isa(User)) { - // If we have a PHI user of the alloca itself (as opposed to a GEP or + // If we have a PHI user of the alloca itself (as opposed to a GEP or // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to // the new pointer. if (!isa(I)) continue; - + assert(Offset == 0 && NewElts[0] && "Direct alloca use should have a zero offset"); - + // If we have a use of the alloca, we know the derived uses will be // utilizing just the first element of the scalarized result. Insert a // bitcast of the first alloca before the user as required. @@ -2063,14 +1993,21 @@ void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, /// RewriteBitCast - Update a bitcast reference to the alloca being replaced /// and recursively continue updating all of its uses. void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, - SmallVector &NewElts) { + SmallVectorImpl &NewElts) { RewriteForScalarRepl(BC, AI, Offset, NewElts); if (BC->getOperand(0) != AI) return; // The bitcast references the original alloca. Replace its uses with - // references to the first new element alloca. - Instruction *Val = NewElts[0]; + // references to the alloca containing offset zero (which is normally at + // index zero, but might not be in cases involving structs with elements + // of size zero). + Type *T = AI->getAllocatedType(); + uint64_t EltOffset = 0; + Type *IdxTy; + uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy, + BC->getModule()->getDataLayout()); + Instruction *Val = NewElts[Idx]; if (Val->getType() != BC->getDestTy()) { Val = new BitCastInst(Val, BC->getDestTy(), "", BC); Val->takeName(BC); @@ -2084,20 +2021,28 @@ void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, /// Sets T to the type of the element and Offset to the offset within that /// element. IdxTy is set to the type of the index result to be used in a /// GEP instruction. -uint64_t SROA::FindElementAndOffset(Type *&T, uint64_t &Offset, - Type *&IdxTy) { +uint64_t SROA::FindElementAndOffset(Type *&T, uint64_t &Offset, Type *&IdxTy, + const DataLayout &DL) { uint64_t Idx = 0; + if (StructType *ST = dyn_cast(T)) { - const StructLayout *Layout = TD->getStructLayout(ST); + const StructLayout *Layout = DL.getStructLayout(ST); Idx = Layout->getElementContainingOffset(Offset); T = ST->getContainedType(Idx); Offset -= Layout->getElementOffset(Idx); IdxTy = Type::getInt32Ty(T->getContext()); return Idx; + } else if (ArrayType *AT = dyn_cast(T)) { + T = AT->getElementType(); + uint64_t EltSize = DL.getTypeAllocSize(T); + Idx = Offset / EltSize; + Offset -= Idx * EltSize; + IdxTy = Type::getInt64Ty(T->getContext()); + return Idx; } - ArrayType *AT = cast(T); - T = AT->getElementType(); - uint64_t EltSize = TD->getTypeAllocSize(T); + VectorType *VT = cast(T); + T = VT->getElementType(); + uint64_t EltSize = DL.getTypeAllocSize(T); Idx = Offset / EltSize; Offset -= Idx * EltSize; IdxTy = Type::getInt64Ty(T->getContext()); @@ -2108,22 +2053,30 @@ uint64_t SROA::FindElementAndOffset(Type *&T, uint64_t &Offset, /// elements of the alloca that are being split apart, and if so, rewrite /// the GEP to be relative to the new element. void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, - SmallVector &NewElts) { + SmallVectorImpl &NewElts) { uint64_t OldOffset = Offset; + const DataLayout &DL = GEPI->getModule()->getDataLayout(); SmallVector Indices(GEPI->op_begin() + 1, GEPI->op_end()); - Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices); + // If the GEP was dynamic then it must have been a dynamic vector lookup. + // In this case, it must be the last GEP operand which is dynamic so keep that + // aside until we've found the constant GEP offset then add it back in at the + // end. + Value* NonConstantIdx = nullptr; + if (!GEPI->hasAllConstantIndices()) + NonConstantIdx = Indices.pop_back_val(); + Offset += DL.getIndexedOffset(GEPI->getPointerOperandType(), Indices); RewriteForScalarRepl(GEPI, AI, Offset, NewElts); Type *T = AI->getAllocatedType(); Type *IdxTy; - uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); + uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy, DL); if (GEPI->getOperand(0) == AI) OldIdx = ~0ULL; // Force the GEP to be rewritten. T = AI->getAllocatedType(); uint64_t EltOffset = Offset; - uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); + uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy, DL); // If this GEP does not move the pointer across elements of the alloca // being split, then it does not needs to be rewritten. @@ -2134,9 +2087,20 @@ void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, SmallVector NewArgs; NewArgs.push_back(Constant::getNullValue(i32Ty)); while (EltOffset != 0) { - uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); + uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy, DL); NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); } + if (NonConstantIdx) { + Type* GepTy = T; + // This GEP has a dynamic index. We need to add "i32 0" to index through + // any structs or arrays in the original type until we get to the vector + // to index. + while (!isa(GepTy)) { + NewArgs.push_back(Constant::getNullValue(i32Ty)); + GepTy = cast(GepTy)->getTypeAtIndex(0U); + } + NewArgs.push_back(NonConstantIdx); + } Instruction *Val = NewElts[Idx]; if (NewArgs.size() > 1) { Val = GetElementPtrInst::CreateInBounds(Val, NewArgs, "", GEPI); @@ -2152,14 +2116,15 @@ void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, /// to mark the lifetime of the scalarized memory. void SROA::RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI, uint64_t Offset, - SmallVector &NewElts) { + SmallVectorImpl &NewElts) { ConstantInt *OldSize = cast(II->getArgOperand(0)); // Put matching lifetime markers on everything from Offset up to // Offset+OldSize. Type *AIType = AI->getAllocatedType(); + const DataLayout &DL = II->getModule()->getDataLayout(); uint64_t NewOffset = Offset; Type *IdxTy; - uint64_t Idx = FindElementAndOffset(AIType, NewOffset, IdxTy); + uint64_t Idx = FindElementAndOffset(AIType, NewOffset, IdxTy, DL); IRBuilder<> Builder(II); uint64_t Size = OldSize->getLimitedValue(); @@ -2167,11 +2132,12 @@ void SROA::RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI, if (NewOffset) { // Splice the first element and index 'NewOffset' bytes in. SROA will // split the alloca again later. - Value *V = Builder.CreateBitCast(NewElts[Idx], Builder.getInt8PtrTy()); - V = Builder.CreateGEP(V, Builder.getInt64(NewOffset)); + unsigned AS = AI->getType()->getAddressSpace(); + Value *V = Builder.CreateBitCast(NewElts[Idx], Builder.getInt8PtrTy(AS)); + V = Builder.CreateGEP(Builder.getInt8Ty(), V, Builder.getInt64(NewOffset)); IdxTy = NewElts[Idx]->getAllocatedType(); - uint64_t EltSize = TD->getTypeAllocSize(IdxTy) - NewOffset; + uint64_t EltSize = DL.getTypeAllocSize(IdxTy) - NewOffset; if (EltSize > Size) { EltSize = Size; Size = 0; @@ -2187,7 +2153,7 @@ void SROA::RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI, for (; Idx != NewElts.size() && Size; ++Idx) { IdxTy = NewElts[Idx]->getAllocatedType(); - uint64_t EltSize = TD->getTypeAllocSize(IdxTy); + uint64_t EltSize = DL.getTypeAllocSize(IdxTy); if (EltSize > Size) { EltSize = Size; Size = 0; @@ -2206,14 +2172,15 @@ void SROA::RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI, /// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. /// Rewrite it to copy or set the elements of the scalarized memory. -void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, - AllocaInst *AI, - SmallVector &NewElts) { +void +SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, + AllocaInst *AI, + SmallVectorImpl &NewElts) { // If this is a memcpy/memmove, construct the other pointer as the // appropriate type. The "Other" pointer is the pointer that goes to memory // that doesn't have anything to do with the alloca that we are promoting. For // memset, this Value* stays null. - Value *OtherPtr = 0; + Value *OtherPtr = nullptr; unsigned MemAlignment = MI->getAlignment(); if (MemTransferInst *MTI = dyn_cast(MI)) { // memmove/memcopy if (Inst == MTI->getRawDest()) @@ -2242,7 +2209,7 @@ void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, if (OtherPtr == AI || OtherPtr == NewElts[0]) { // This code will run twice for a no-op memcpy -- once for each operand. // Put only one reference to MI on the DeadInsts list. - for (SmallVector::const_iterator I = DeadInsts.begin(), + for (SmallVectorImpl::const_iterator I = DeadInsts.begin(), E = DeadInsts.end(); I != E; ++I) if (*I == MI) return; DeadInsts.push_back(MI); @@ -2262,10 +2229,11 @@ void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, bool SROADest = MI->getRawDest() == Inst; Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); + const DataLayout &DL = MI->getModule()->getDataLayout(); for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { // If this is a memcpy/memmove, emit a GEP of the other element address. - Value *OtherElt = 0; + Value *OtherElt = nullptr; unsigned OtherEltAlign = MemAlignment; if (OtherPtr) { @@ -2278,10 +2246,10 @@ void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, PointerType *OtherPtrTy = cast(OtherPtr->getType()); Type *OtherTy = OtherPtrTy->getElementType(); if (StructType *ST = dyn_cast(OtherTy)) { - EltOffset = TD->getStructLayout(ST)->getElementOffset(i); + EltOffset = DL.getStructLayout(ST)->getElementOffset(i); } else { Type *EltTy = cast(OtherTy)->getElementType(); - EltOffset = TD->getTypeAllocSize(EltTy)*i; + EltOffset = DL.getTypeAllocSize(EltTy) * i; } // The alignment of the other pointer is the guaranteed alignment of the @@ -2322,7 +2290,7 @@ void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, Type *ValTy = EltTy->getScalarType(); // Construct an integer with the right value. - unsigned EltSize = TD->getTypeSizeInBits(ValTy); + unsigned EltSize = DL.getTypeSizeInBits(ValTy); APInt OneVal(EltSize, CI->getZExtValue()); APInt TotalVal(OneVal); // Set each byte. @@ -2340,10 +2308,9 @@ void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, assert(StoreVal->getType() == ValTy && "Type mismatch!"); // If the requested value was a vector constant, create it. - if (EltTy != ValTy) { - unsigned NumElts = cast(ValTy)->getNumElements(); - SmallVector Elts(NumElts, StoreVal); - StoreVal = ConstantVector::get(Elts); + if (EltTy->isVectorTy()) { + unsigned NumElts = cast(EltTy)->getNumElements(); + StoreVal = ConstantVector::getSplat(NumElts, StoreVal); } } new StoreInst(StoreVal, EltPtr, MI); @@ -2353,7 +2320,9 @@ void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, // this element. } - unsigned EltSize = TD->getTypeAllocSize(EltTy); + unsigned EltSize = DL.getTypeAllocSize(EltTy); + if (!EltSize) + continue; IRBuilder<> Builder(MI); @@ -2378,18 +2347,20 @@ void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, /// RewriteStoreUserOfWholeAlloca - We found a store of an integer that /// overwrites the entire allocation. Extract out the pieces of the stored /// integer and store them individually. -void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, - SmallVector &NewElts){ +void +SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, + SmallVectorImpl &NewElts) { // Extract each element out of the integer according to its structure offset // and store the element value to the individual alloca. Value *SrcVal = SI->getOperand(0); Type *AllocaEltTy = AI->getAllocatedType(); - uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); + const DataLayout &DL = SI->getModule()->getDataLayout(); + uint64_t AllocaSizeBits = DL.getTypeAllocSizeInBits(AllocaEltTy); IRBuilder<> Builder(SI); - + // Handle tail padding by extending the operand - if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) + if (DL.getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) SrcVal = Builder.CreateZExt(SrcVal, IntegerType::get(SI->getContext(), AllocaSizeBits)); @@ -2399,15 +2370,15 @@ void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, // There are two forms here: AI could be an array or struct. Both cases // have different ways to compute the element offset. if (StructType *EltSTy = dyn_cast(AllocaEltTy)) { - const StructLayout *Layout = TD->getStructLayout(EltSTy); + const StructLayout *Layout = DL.getStructLayout(EltSTy); for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { // Get the number of bits to shift SrcVal to get the value. Type *FieldTy = EltSTy->getElementType(i); uint64_t Shift = Layout->getElementOffsetInBits(i); - if (TD->isBigEndian()) - Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); + if (DL.isBigEndian()) + Shift = AllocaSizeBits - Shift - DL.getTypeAllocSizeInBits(FieldTy); Value *EltVal = SrcVal; if (Shift) { @@ -2416,7 +2387,7 @@ void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, } // Truncate down to an integer of the right size. - uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); + uint64_t FieldSizeBits = DL.getTypeSizeInBits(FieldTy); // Ignore zero sized fields like {}, they obviously contain no data. if (FieldSizeBits == 0) continue; @@ -2441,12 +2412,12 @@ void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, } else { ArrayType *ATy = cast(AllocaEltTy); Type *ArrayEltTy = ATy->getElementType(); - uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); - uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); + uint64_t ElementOffset = DL.getTypeAllocSizeInBits(ArrayEltTy); + uint64_t ElementSizeBits = DL.getTypeSizeInBits(ArrayEltTy); uint64_t Shift; - if (TD->isBigEndian()) + if (DL.isBigEndian()) Shift = AllocaSizeBits-ElementOffset; else Shift = 0; @@ -2480,7 +2451,7 @@ void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, } new StoreInst(EltVal, DestField, SI); - if (TD->isBigEndian()) + if (DL.isBigEndian()) Shift -= ElementOffset; else Shift += ElementOffset; @@ -2492,25 +2463,27 @@ void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, /// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to /// an integer. Load the individual pieces to form the aggregate value. -void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, - SmallVector &NewElts) { +void +SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, + SmallVectorImpl &NewElts) { // Extract each element out of the NewElts according to its structure offset // and form the result value. Type *AllocaEltTy = AI->getAllocatedType(); - uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); + const DataLayout &DL = LI->getModule()->getDataLayout(); + uint64_t AllocaSizeBits = DL.getTypeAllocSizeInBits(AllocaEltTy); DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI << '\n'); // There are two forms here: AI could be an array or struct. Both cases // have different ways to compute the element offset. - const StructLayout *Layout = 0; + const StructLayout *Layout = nullptr; uint64_t ArrayEltBitOffset = 0; if (StructType *EltSTy = dyn_cast(AllocaEltTy)) { - Layout = TD->getStructLayout(EltSTy); + Layout = DL.getStructLayout(EltSTy); } else { Type *ArrayEltTy = cast(AllocaEltTy)->getElementType(); - ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); + ArrayEltBitOffset = DL.getTypeAllocSizeInBits(ArrayEltTy); } Value *ResultVal = @@ -2522,7 +2495,7 @@ void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, Value *SrcField = NewElts[i]; Type *FieldTy = cast(SrcField->getType())->getElementType(); - uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); + uint64_t FieldSizeBits = DL.getTypeSizeInBits(FieldTy); // Ignore zero sized fields like {}, they obviously contain no data. if (FieldSizeBits == 0) continue; @@ -2553,7 +2526,7 @@ void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, else // Array case. Shift = i*ArrayEltBitOffset; - if (TD->isBigEndian()) + if (DL.isBigEndian()) Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); if (Shift) { @@ -2570,7 +2543,7 @@ void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, } // Handle tail padding by truncating the result - if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) + if (DL.getTypeSizeInBits(LI->getType()) != AllocaSizeBits) ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); LI->replaceAllUsesWith(ResultVal); @@ -2580,15 +2553,15 @@ void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, /// HasPadding - Return true if the specified type has any structure or /// alignment padding in between the elements that would be split apart /// by SROA; return false otherwise. -static bool HasPadding(Type *Ty, const TargetData &TD) { +static bool HasPadding(Type *Ty, const DataLayout &DL) { if (ArrayType *ATy = dyn_cast(Ty)) { Ty = ATy->getElementType(); - return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); + return DL.getTypeSizeInBits(Ty) != DL.getTypeAllocSizeInBits(Ty); } // SROA currently handles only Arrays and Structs. StructType *STy = cast(Ty); - const StructLayout *SL = TD.getStructLayout(STy); + const StructLayout *SL = DL.getStructLayout(STy); unsigned PrevFieldBitOffset = 0; for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { unsigned FieldBitOffset = SL->getElementOffsetInBits(i); @@ -2597,7 +2570,7 @@ static bool HasPadding(Type *Ty, const TargetData &TD) { // previous one. if (i) { unsigned PrevFieldEnd = - PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); + PrevFieldBitOffset+DL.getTypeSizeInBits(STy->getElementType(i-1)); if (PrevFieldEnd < FieldBitOffset) return true; } @@ -2606,7 +2579,7 @@ static bool HasPadding(Type *Ty, const TargetData &TD) { // Check for tail padding. if (unsigned EltCount = STy->getNumElements()) { unsigned PrevFieldEnd = PrevFieldBitOffset + - TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); + DL.getTypeSizeInBits(STy->getElementType(EltCount-1)); if (PrevFieldEnd < SL->getSizeInBits()) return true; } @@ -2627,13 +2600,15 @@ bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { return false; } + const DataLayout &DL = AI->getModule()->getDataLayout(); + // Okay, we know all the users are promotable. If the aggregate is a memcpy // source and destination, we have to be careful. In particular, the memcpy // could be moving around elements that live in structure padding of the LLVM // types, but may actually be used. In these cases, we refuse to promote the // struct. if (Info.isMemCpySrc && Info.isMemCpyDst && - HasPadding(AI->getAllocatedType(), *TD)) + HasPadding(AI->getAllocatedType(), DL)) return false; // If the alloca never has an access to just *part* of it, but is accessed @@ -2649,138 +2624,6 @@ bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { return false; } } - - return true; -} - - - -/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to -/// some part of a constant global variable. This intentionally only accepts -/// constant expressions because we don't can't rewrite arbitrary instructions. -static bool PointsToConstantGlobal(Value *V) { - if (GlobalVariable *GV = dyn_cast(V)) - return GV->isConstant(); - if (ConstantExpr *CE = dyn_cast(V)) - if (CE->getOpcode() == Instruction::BitCast || - CE->getOpcode() == Instruction::GetElementPtr) - return PointsToConstantGlobal(CE->getOperand(0)); - return false; -} - -/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) -/// pointer to an alloca. Ignore any reads of the pointer, return false if we -/// see any stores or other unknown uses. If we see pointer arithmetic, keep -/// track of whether it moves the pointer (with isOffset) but otherwise traverse -/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to -/// the alloca, and if the source pointer is a pointer to a constant global, we -/// can optimize this. -static bool -isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, - bool isOffset, - SmallVector &LifetimeMarkers) { - // We track lifetime intrinsics as we encounter them. If we decide to go - // ahead and replace the value with the global, this lets the caller quickly - // eliminate the markers. - - for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { - User *U = cast(*UI); - - if (LoadInst *LI = dyn_cast(U)) { - // Ignore non-volatile loads, they are always ok. - if (LI->isVolatile()) return false; - continue; - } - - if (BitCastInst *BCI = dyn_cast(U)) { - // If uses of the bitcast are ok, we are ok. - if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset, - LifetimeMarkers)) - return false; - continue; - } - if (GetElementPtrInst *GEP = dyn_cast(U)) { - // If the GEP has all zero indices, it doesn't offset the pointer. If it - // doesn't, it does. - if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, - isOffset || !GEP->hasAllZeroIndices(), - LifetimeMarkers)) - return false; - continue; - } - - if (CallSite CS = U) { - // If this is the function being called then we treat it like a load and - // ignore it. - if (CS.isCallee(UI)) - continue; - - // If this is a readonly/readnone call site, then we know it is just a - // load (but one that potentially returns the value itself), so we can - // ignore it if we know that the value isn't captured. - unsigned ArgNo = CS.getArgumentNo(UI); - if (CS.onlyReadsMemory() && - (CS.getInstruction()->use_empty() || - CS.paramHasAttr(ArgNo+1, Attribute::NoCapture))) - continue; - - // If this is being passed as a byval argument, the caller is making a - // copy, so it is only a read of the alloca. - if (CS.paramHasAttr(ArgNo+1, Attribute::ByVal)) - continue; - } - - // Lifetime intrinsics can be handled by the caller. - if (IntrinsicInst *II = dyn_cast(U)) { - if (II->getIntrinsicID() == Intrinsic::lifetime_start || - II->getIntrinsicID() == Intrinsic::lifetime_end) { - assert(II->use_empty() && "Lifetime markers have no result to use!"); - LifetimeMarkers.push_back(II); - continue; - } - } - - // If this is isn't our memcpy/memmove, reject it as something we can't - // handle. - MemTransferInst *MI = dyn_cast(U); - if (MI == 0) - return false; - - // If the transfer is using the alloca as a source of the transfer, then - // ignore it since it is a load (unless the transfer is volatile). - if (UI.getOperandNo() == 1) { - if (MI->isVolatile()) return false; - continue; - } - - // If we already have seen a copy, reject the second one. - if (TheCopy) return false; - // If the pointer has been offset from the start of the alloca, we can't - // safely handle this. - if (isOffset) return false; - - // If the memintrinsic isn't using the alloca as the dest, reject it. - if (UI.getOperandNo() != 0) return false; - - // If the source of the memcpy/move is not a constant global, reject it. - if (!PointsToConstantGlobal(MI->getSource())) - return false; - - // Otherwise, the transform is safe. Remember the copy instruction. - TheCopy = MI; - } return true; } - -/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only -/// modified by a copy from a constant global. If we can prove this, we can -/// replace any uses of the alloca with uses of the global directly. -MemTransferInst * -SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI, - SmallVector &ToDelete) { - MemTransferInst *TheCopy = 0; - if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false, ToDelete)) - return TheCopy; - return 0; -}