1193323Sed//===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===// 2193323Sed// 3193323Sed// The LLVM Compiler Infrastructure 4193323Sed// 5193323Sed// This file is distributed under the University of Illinois Open Source 6193323Sed// License. See LICENSE.TXT for details. 7193323Sed// 8193323Sed//===----------------------------------------------------------------------===// 9193323Sed// 10193323Sed// This transformation implements the well known scalar replacement of 11193323Sed// aggregates transformation. This xform breaks up alloca instructions of 12193323Sed// aggregate type (structure or array) into individual alloca instructions for 13193323Sed// each member (if possible). Then, if possible, it transforms the individual 14193323Sed// alloca instructions into nice clean scalar SSA form. 15193323Sed// 16235633Sdim// This combines a simple SRoA algorithm with the Mem2Reg algorithm because they 17193323Sed// often interact, especially for C++ programs. As such, iterating between 18193323Sed// SRoA, then Mem2Reg until we run out of things to promote works well. 19193323Sed// 20193323Sed//===----------------------------------------------------------------------===// 21193323Sed 22193323Sed#define DEBUG_TYPE "scalarrepl" 23193323Sed#include "llvm/Transforms/Scalar.h" 24245431Sdim#include "llvm/ADT/SetVector.h" 25245431Sdim#include "llvm/ADT/SmallVector.h" 26245431Sdim#include "llvm/ADT/Statistic.h" 27193323Sed#include "llvm/Analysis/Dominators.h" 28218893Sdim#include "llvm/Analysis/Loads.h" 29218893Sdim#include "llvm/Analysis/ValueTracking.h" 30252723Sdim#include "llvm/DIBuilder.h" 31252723Sdim#include "llvm/DebugInfo.h" 32252723Sdim#include "llvm/IR/Constants.h" 33252723Sdim#include "llvm/IR/DataLayout.h" 34252723Sdim#include "llvm/IR/DerivedTypes.h" 35252723Sdim#include "llvm/IR/Function.h" 36252723Sdim#include "llvm/IR/GlobalVariable.h" 37252723Sdim#include "llvm/IR/IRBuilder.h" 38252723Sdim#include "llvm/IR/Instructions.h" 39252723Sdim#include "llvm/IR/IntrinsicInst.h" 40252723Sdim#include "llvm/IR/LLVMContext.h" 41252723Sdim#include "llvm/IR/Module.h" 42252723Sdim#include "llvm/IR/Operator.h" 43252723Sdim#include "llvm/Pass.h" 44218893Sdim#include "llvm/Support/CallSite.h" 45193323Sed#include "llvm/Support/Debug.h" 46198090Srdivacky#include "llvm/Support/ErrorHandling.h" 47193323Sed#include "llvm/Support/GetElementPtrTypeIterator.h" 48193323Sed#include "llvm/Support/MathExtras.h" 49198090Srdivacky#include "llvm/Support/raw_ostream.h" 50245431Sdim#include "llvm/Transforms/Utils/Local.h" 51245431Sdim#include "llvm/Transforms/Utils/PromoteMemToReg.h" 52245431Sdim#include "llvm/Transforms/Utils/SSAUpdater.h" 53193323Sedusing namespace llvm; 54193323Sed 55193323SedSTATISTIC(NumReplaced, "Number of allocas broken up"); 56193323SedSTATISTIC(NumPromoted, "Number of allocas promoted"); 57218893SdimSTATISTIC(NumAdjusted, "Number of scalar allocas adjusted to allow promotion"); 58193323SedSTATISTIC(NumConverted, "Number of aggregates converted to scalar"); 59193323Sed 60193323Sednamespace { 61198090Srdivacky struct SROA : public FunctionPass { 62245431Sdim SROA(int T, bool hasDT, char &ID, int ST, int AT, int SLT) 63218893Sdim : FunctionPass(ID), HasDomTree(hasDT) { 64193323Sed if (T == -1) 65193323Sed SRThreshold = 128; 66193323Sed else 67193323Sed SRThreshold = T; 68245431Sdim if (ST == -1) 69245431Sdim StructMemberThreshold = 32; 70245431Sdim else 71245431Sdim StructMemberThreshold = ST; 72245431Sdim if (AT == -1) 73245431Sdim ArrayElementThreshold = 8; 74245431Sdim else 75245431Sdim ArrayElementThreshold = AT; 76245431Sdim if (SLT == -1) 77245431Sdim // Do not limit the scalar integer load size if no threshold is given. 78245431Sdim ScalarLoadThreshold = -1; 79245431Sdim else 80245431Sdim ScalarLoadThreshold = SLT; 81193323Sed } 82193323Sed 83193323Sed bool runOnFunction(Function &F); 84193323Sed 85193323Sed bool performScalarRepl(Function &F); 86193323Sed bool performPromotion(Function &F); 87193323Sed 88193323Sed private: 89218893Sdim bool HasDomTree; 90245431Sdim DataLayout *TD; 91218893Sdim 92201360Srdivacky /// DeadInsts - Keep track of instructions we have made dead, so that 93201360Srdivacky /// we can remove them after we are done working. 94201360Srdivacky SmallVector<Value*, 32> DeadInsts; 95201360Srdivacky 96193323Sed /// AllocaInfo - When analyzing uses of an alloca instruction, this captures 97193323Sed /// information about the uses. All these fields are initialized to false 98193323Sed /// and set to true when something is learned. 99193323Sed struct AllocaInfo { 100218893Sdim /// The alloca to promote. 101218893Sdim AllocaInst *AI; 102245431Sdim 103218893Sdim /// CheckedPHIs - This is a set of verified PHI nodes, to prevent infinite 104218893Sdim /// looping and avoid redundant work. 105218893Sdim SmallPtrSet<PHINode*, 8> CheckedPHIs; 106245431Sdim 107193323Sed /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 108193323Sed bool isUnsafe : 1; 109218893Sdim 110193323Sed /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 111193323Sed bool isMemCpySrc : 1; 112193323Sed 113193323Sed /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 114193323Sed bool isMemCpyDst : 1; 115193323Sed 116218893Sdim /// hasSubelementAccess - This is true if a subelement of the alloca is 117218893Sdim /// ever accessed, or false if the alloca is only accessed with mem 118218893Sdim /// intrinsics or load/store that only access the entire alloca at once. 119218893Sdim bool hasSubelementAccess : 1; 120245431Sdim 121218893Sdim /// hasALoadOrStore - This is true if there are any loads or stores to it. 122218893Sdim /// The alloca may just be accessed with memcpy, for example, which would 123218893Sdim /// not set this. 124218893Sdim bool hasALoadOrStore : 1; 125245431Sdim 126218893Sdim explicit AllocaInfo(AllocaInst *ai) 127218893Sdim : AI(ai), isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false), 128218893Sdim hasSubelementAccess(false), hasALoadOrStore(false) {} 129193323Sed }; 130218893Sdim 131245431Sdim /// SRThreshold - The maximum alloca size to considered for SROA. 132193323Sed unsigned SRThreshold; 133193323Sed 134245431Sdim /// StructMemberThreshold - The maximum number of members a struct can 135245431Sdim /// contain to be considered for SROA. 136245431Sdim unsigned StructMemberThreshold; 137245431Sdim 138245431Sdim /// ArrayElementThreshold - The maximum number of elements an array can 139245431Sdim /// have to be considered for SROA. 140245431Sdim unsigned ArrayElementThreshold; 141245431Sdim 142245431Sdim /// ScalarLoadThreshold - The maximum size in bits of scalars to load when 143245431Sdim /// converting to scalar 144245431Sdim unsigned ScalarLoadThreshold; 145245431Sdim 146218893Sdim void MarkUnsafe(AllocaInfo &I, Instruction *User) { 147218893Sdim I.isUnsafe = true; 148218893Sdim DEBUG(dbgs() << " Transformation preventing inst: " << *User << '\n'); 149218893Sdim } 150193323Sed 151202878Srdivacky bool isSafeAllocaToScalarRepl(AllocaInst *AI); 152193323Sed 153218893Sdim void isSafeForScalarRepl(Instruction *I, uint64_t Offset, AllocaInfo &Info); 154218893Sdim void isSafePHISelectUseForScalarRepl(Instruction *User, uint64_t Offset, 155218893Sdim AllocaInfo &Info); 156218893Sdim void isSafeGEP(GetElementPtrInst *GEPI, uint64_t &Offset, AllocaInfo &Info); 157218893Sdim void isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 158226890Sdim Type *MemOpType, bool isStore, AllocaInfo &Info, 159218893Sdim Instruction *TheAccess, bool AllowWholeAccess); 160226890Sdim bool TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size); 161226890Sdim uint64_t FindElementAndOffset(Type *&T, uint64_t &Offset, 162226890Sdim Type *&IdxTy); 163218893Sdim 164218893Sdim void DoScalarReplacement(AllocaInst *AI, 165198892Srdivacky std::vector<AllocaInst*> &WorkList); 166201360Srdivacky void DeleteDeadInstructions(); 167218893Sdim 168201360Srdivacky void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 169263509Sdim SmallVectorImpl<AllocaInst *> &NewElts); 170201360Srdivacky void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 171263509Sdim SmallVectorImpl<AllocaInst *> &NewElts); 172201360Srdivacky void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 173263509Sdim SmallVectorImpl<AllocaInst *> &NewElts); 174226890Sdim void RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI, 175226890Sdim uint64_t Offset, 176263509Sdim SmallVectorImpl<AllocaInst *> &NewElts); 177201360Srdivacky void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 178198892Srdivacky AllocaInst *AI, 179263509Sdim SmallVectorImpl<AllocaInst *> &NewElts); 180198892Srdivacky void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 181263509Sdim SmallVectorImpl<AllocaInst *> &NewElts); 182198892Srdivacky void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 183263509Sdim SmallVectorImpl<AllocaInst *> &NewElts); 184245431Sdim bool ShouldAttemptScalarRepl(AllocaInst *AI); 185245431Sdim }; 186218893Sdim 187218893Sdim // SROA_DT - SROA that uses DominatorTree. 188218893Sdim struct SROA_DT : public SROA { 189218893Sdim static char ID; 190218893Sdim public: 191245431Sdim SROA_DT(int T = -1, int ST = -1, int AT = -1, int SLT = -1) : 192245431Sdim SROA(T, true, ID, ST, AT, SLT) { 193218893Sdim initializeSROA_DTPass(*PassRegistry::getPassRegistry()); 194218893Sdim } 195245431Sdim 196218893Sdim // getAnalysisUsage - This pass does not require any passes, but we know it 197218893Sdim // will not alter the CFG, so say so. 198218893Sdim virtual void getAnalysisUsage(AnalysisUsage &AU) const { 199218893Sdim AU.addRequired<DominatorTree>(); 200218893Sdim AU.setPreservesCFG(); 201218893Sdim } 202218893Sdim }; 203245431Sdim 204218893Sdim // SROA_SSAUp - SROA that uses SSAUpdater. 205218893Sdim struct SROA_SSAUp : public SROA { 206218893Sdim static char ID; 207218893Sdim public: 208245431Sdim SROA_SSAUp(int T = -1, int ST = -1, int AT = -1, int SLT = -1) : 209245431Sdim SROA(T, false, ID, ST, AT, SLT) { 210218893Sdim initializeSROA_SSAUpPass(*PassRegistry::getPassRegistry()); 211218893Sdim } 212245431Sdim 213218893Sdim // getAnalysisUsage - This pass does not require any passes, but we know it 214218893Sdim // will not alter the CFG, so say so. 215218893Sdim virtual void getAnalysisUsage(AnalysisUsage &AU) const { 216218893Sdim AU.setPreservesCFG(); 217218893Sdim } 218218893Sdim }; 219245431Sdim 220193323Sed} 221193323Sed 222218893Sdimchar SROA_DT::ID = 0; 223218893Sdimchar SROA_SSAUp::ID = 0; 224193323Sed 225218893SdimINITIALIZE_PASS_BEGIN(SROA_DT, "scalarrepl", 226218893Sdim "Scalar Replacement of Aggregates (DT)", false, false) 227218893SdimINITIALIZE_PASS_DEPENDENCY(DominatorTree) 228218893SdimINITIALIZE_PASS_END(SROA_DT, "scalarrepl", 229218893Sdim "Scalar Replacement of Aggregates (DT)", false, false) 230218893Sdim 231218893SdimINITIALIZE_PASS_BEGIN(SROA_SSAUp, "scalarrepl-ssa", 232218893Sdim "Scalar Replacement of Aggregates (SSAUp)", false, false) 233218893SdimINITIALIZE_PASS_END(SROA_SSAUp, "scalarrepl-ssa", 234218893Sdim "Scalar Replacement of Aggregates (SSAUp)", false, false) 235218893Sdim 236193323Sed// Public interface to the ScalarReplAggregates pass 237218893SdimFunctionPass *llvm::createScalarReplAggregatesPass(int Threshold, 238245431Sdim bool UseDomTree, 239245431Sdim int StructMemberThreshold, 240245431Sdim int ArrayElementThreshold, 241245431Sdim int ScalarLoadThreshold) { 242218893Sdim if (UseDomTree) 243245431Sdim return new SROA_DT(Threshold, StructMemberThreshold, ArrayElementThreshold, 244245431Sdim ScalarLoadThreshold); 245245431Sdim return new SROA_SSAUp(Threshold, StructMemberThreshold, 246245431Sdim ArrayElementThreshold, ScalarLoadThreshold); 247193323Sed} 248193323Sed 249193323Sed 250207618Srdivacky//===----------------------------------------------------------------------===// 251207618Srdivacky// Convert To Scalar Optimization. 252207618Srdivacky//===----------------------------------------------------------------------===// 253207618Srdivacky 254207618Srdivackynamespace { 255207618Srdivacky/// ConvertToScalarInfo - This class implements the "Convert To Scalar" 256207618Srdivacky/// optimization, which scans the uses of an alloca and determines if it can 257207618Srdivacky/// rewrite it in terms of a single new alloca that can be mem2reg'd. 258207618Srdivackyclass ConvertToScalarInfo { 259221345Sdim /// AllocaSize - The size of the alloca being considered in bytes. 260207618Srdivacky unsigned AllocaSize; 261245431Sdim const DataLayout &TD; 262245431Sdim unsigned ScalarLoadThreshold; 263218893Sdim 264207618Srdivacky /// IsNotTrivial - This is set to true if there is some access to the object 265207618Srdivacky /// which means that mem2reg can't promote it. 266207618Srdivacky bool IsNotTrivial; 267218893Sdim 268224145Sdim /// ScalarKind - Tracks the kind of alloca being considered for promotion, 269224145Sdim /// computed based on the uses of the alloca rather than the LLVM type system. 270224145Sdim enum { 271224145Sdim Unknown, 272224145Sdim 273224145Sdim // Accesses via GEPs that are consistent with element access of a vector 274224145Sdim // type. This will not be converted into a vector unless there is a later 275224145Sdim // access using an actual vector type. 276224145Sdim ImplicitVector, 277224145Sdim 278224145Sdim // Accesses via vector operations and GEPs that are consistent with the 279224145Sdim // layout of a vector type. 280224145Sdim Vector, 281224145Sdim 282224145Sdim // An integer bag-of-bits with bitwise operations for insertion and 283224145Sdim // extraction. Any combination of types can be converted into this kind 284224145Sdim // of scalar. 285224145Sdim Integer 286224145Sdim } ScalarKind; 287224145Sdim 288207618Srdivacky /// VectorTy - This tracks the type that we should promote the vector to if 289207618Srdivacky /// it is possible to turn it into a vector. This starts out null, and if it 290207618Srdivacky /// isn't possible to turn into a vector type, it gets set to VoidTy. 291226890Sdim VectorType *VectorTy; 292218893Sdim 293245431Sdim /// HadNonMemTransferAccess - True if there is at least one access to the 294221345Sdim /// alloca that is not a MemTransferInst. We don't want to turn structs into 295221345Sdim /// large integers unless there is some potential for optimization. 296221345Sdim bool HadNonMemTransferAccess; 297221345Sdim 298245431Sdim /// HadDynamicAccess - True if some element of this alloca was dynamic. 299245431Sdim /// We don't yet have support for turning a dynamic access into a large 300245431Sdim /// integer. 301245431Sdim bool HadDynamicAccess; 302245431Sdim 303207618Srdivackypublic: 304245431Sdim explicit ConvertToScalarInfo(unsigned Size, const DataLayout &td, 305245431Sdim unsigned SLT) 306245431Sdim : AllocaSize(Size), TD(td), ScalarLoadThreshold(SLT), IsNotTrivial(false), 307245431Sdim ScalarKind(Unknown), VectorTy(0), HadNonMemTransferAccess(false), 308245431Sdim HadDynamicAccess(false) { } 309218893Sdim 310207618Srdivacky AllocaInst *TryConvert(AllocaInst *AI); 311218893Sdim 312207618Srdivackyprivate: 313245431Sdim bool CanConvertToScalar(Value *V, uint64_t Offset, Value* NonConstantIdx); 314226890Sdim void MergeInTypeForLoadOrStore(Type *In, uint64_t Offset); 315226890Sdim bool MergeInVectorType(VectorType *VInTy, uint64_t Offset); 316245431Sdim void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset, 317245431Sdim Value *NonConstantIdx); 318218893Sdim 319226890Sdim Value *ConvertScalar_ExtractValue(Value *NV, Type *ToType, 320245431Sdim uint64_t Offset, Value* NonConstantIdx, 321245431Sdim IRBuilder<> &Builder); 322207618Srdivacky Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 323245431Sdim uint64_t Offset, Value* NonConstantIdx, 324245431Sdim IRBuilder<> &Builder); 325207618Srdivacky}; 326207618Srdivacky} // end anonymous namespace. 327207618Srdivacky 328212904Sdim 329207618Srdivacky/// TryConvert - Analyze the specified alloca, and if it is safe to do so, 330207618Srdivacky/// rewrite it to be a new alloca which is mem2reg'able. This returns the new 331207618Srdivacky/// alloca if possible or null if not. 332207618SrdivackyAllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) { 333207618Srdivacky // If we can't convert this scalar, or if mem2reg can trivially do it, bail 334207618Srdivacky // out. 335245431Sdim if (!CanConvertToScalar(AI, 0, 0) || !IsNotTrivial) 336207618Srdivacky return 0; 337218893Sdim 338224145Sdim // If an alloca has only memset / memcpy uses, it may still have an Unknown 339224145Sdim // ScalarKind. Treat it as an Integer below. 340224145Sdim if (ScalarKind == Unknown) 341224145Sdim ScalarKind = Integer; 342224145Sdim 343224145Sdim if (ScalarKind == Vector && VectorTy->getBitWidth() != AllocaSize * 8) 344224145Sdim ScalarKind = Integer; 345224145Sdim 346207618Srdivacky // If we were able to find a vector type that can handle this with 347207618Srdivacky // insert/extract elements, and if there was at least one use that had 348207618Srdivacky // a vector type, promote this to a vector. We don't want to promote 349207618Srdivacky // random stuff that doesn't use vectors (e.g. <9 x double>) because then 350207618Srdivacky // we just get a lot of insert/extracts. If at least one vector is 351207618Srdivacky // involved, then we probably really do have a union of vector/array. 352226890Sdim Type *NewTy; 353224145Sdim if (ScalarKind == Vector) { 354224145Sdim assert(VectorTy && "Missing type for vector scalar."); 355207618Srdivacky DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = " 356207618Srdivacky << *VectorTy << '\n'); 357207618Srdivacky NewTy = VectorTy; // Use the vector type. 358207618Srdivacky } else { 359221345Sdim unsigned BitWidth = AllocaSize * 8; 360245431Sdim 361245431Sdim // Do not convert to scalar integer if the alloca size exceeds the 362245431Sdim // scalar load threshold. 363245431Sdim if (BitWidth > ScalarLoadThreshold) 364245431Sdim return 0; 365245431Sdim 366224145Sdim if ((ScalarKind == ImplicitVector || ScalarKind == Integer) && 367224145Sdim !HadNonMemTransferAccess && !TD.fitsInLegalInteger(BitWidth)) 368221345Sdim return 0; 369245431Sdim // Dynamic accesses on integers aren't yet supported. They need us to shift 370245431Sdim // by a dynamic amount which could be difficult to work out as we might not 371245431Sdim // know whether to use a left or right shift. 372245431Sdim if (ScalarKind == Integer && HadDynamicAccess) 373245431Sdim return 0; 374221345Sdim 375207618Srdivacky DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"); 376207618Srdivacky // Create and insert the integer alloca. 377221345Sdim NewTy = IntegerType::get(AI->getContext(), BitWidth); 378207618Srdivacky } 379207618Srdivacky AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 380245431Sdim ConvertUsesToScalar(AI, NewAI, 0, 0); 381207618Srdivacky return NewAI; 382207618Srdivacky} 383207618Srdivacky 384224145Sdim/// MergeInTypeForLoadOrStore - Add the 'In' type to the accumulated vector type 385224145Sdim/// (VectorTy) so far at the offset specified by Offset (which is specified in 386224145Sdim/// bytes). 387207618Srdivacky/// 388226890Sdim/// There are two cases we handle here: 389207618Srdivacky/// 1) A union of vector types of the same size and potentially its elements. 390207618Srdivacky/// Here we turn element accesses into insert/extract element operations. 391207618Srdivacky/// This promotes a <4 x float> with a store of float to the third element 392207618Srdivacky/// into a <4 x float> that uses insert element. 393226890Sdim/// 2) A fully general blob of memory, which we turn into some (potentially 394207618Srdivacky/// large) integer type with extract and insert operations where the loads 395207618Srdivacky/// and stores would mutate the memory. We mark this by setting VectorTy 396207618Srdivacky/// to VoidTy. 397226890Sdimvoid ConvertToScalarInfo::MergeInTypeForLoadOrStore(Type *In, 398224145Sdim uint64_t Offset) { 399207618Srdivacky // If we already decided to turn this into a blob of integer memory, there is 400207618Srdivacky // nothing to be done. 401224145Sdim if (ScalarKind == Integer) 402207618Srdivacky return; 403218893Sdim 404207618Srdivacky // If this could be contributing to a vector, analyze it. 405207618Srdivacky 406207618Srdivacky // If the In type is a vector that is the same size as the alloca, see if it 407207618Srdivacky // matches the existing VecTy. 408226890Sdim if (VectorType *VInTy = dyn_cast<VectorType>(In)) { 409221345Sdim if (MergeInVectorType(VInTy, Offset)) 410207618Srdivacky return; 411207618Srdivacky } else if (In->isFloatTy() || In->isDoubleTy() || 412207618Srdivacky (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 && 413207618Srdivacky isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 414221345Sdim // Full width accesses can be ignored, because they can always be turned 415221345Sdim // into bitcasts. 416221345Sdim unsigned EltSize = In->getPrimitiveSizeInBits()/8; 417224145Sdim if (EltSize == AllocaSize) 418221345Sdim return; 419221345Sdim 420207618Srdivacky // If we're accessing something that could be an element of a vector, see 421207618Srdivacky // if the implied vector agrees with what we already have and if Offset is 422207618Srdivacky // compatible with it. 423223017Sdim if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 && 424226890Sdim (!VectorTy || EltSize == VectorTy->getElementType() 425226890Sdim ->getPrimitiveSizeInBits()/8)) { 426221345Sdim if (!VectorTy) { 427224145Sdim ScalarKind = ImplicitVector; 428207618Srdivacky VectorTy = VectorType::get(In, AllocaSize/EltSize); 429221345Sdim } 430226890Sdim return; 431207618Srdivacky } 432207618Srdivacky } 433218893Sdim 434207618Srdivacky // Otherwise, we have a case that we can't handle with an optimized vector 435207618Srdivacky // form. We can still turn this into a large integer. 436224145Sdim ScalarKind = Integer; 437207618Srdivacky} 438207618Srdivacky 439224145Sdim/// MergeInVectorType - Handles the vector case of MergeInTypeForLoadOrStore, 440224145Sdim/// returning true if the type was successfully merged and false otherwise. 441226890Sdimbool ConvertToScalarInfo::MergeInVectorType(VectorType *VInTy, 442221345Sdim uint64_t Offset) { 443226890Sdim if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 444226890Sdim // If we're storing/loading a vector of the right size, allow it as a 445226890Sdim // vector. If this the first vector we see, remember the type so that 446226890Sdim // we know the element size. If this is a subsequent access, ignore it 447226890Sdim // even if it is a differing type but the same size. Worst case we can 448226890Sdim // bitcast the resultant vectors. 449226890Sdim if (!VectorTy) 450226890Sdim VectorTy = VInTy; 451224145Sdim ScalarKind = Vector; 452221345Sdim return true; 453221345Sdim } 454221345Sdim 455226890Sdim return false; 456221345Sdim} 457221345Sdim 458207618Srdivacky/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 459207618Srdivacky/// its accesses to a single vector type, return true and set VecTy to 460207618Srdivacky/// the new type. If we could convert the alloca into a single promotable 461207618Srdivacky/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 462207618Srdivacky/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 463207618Srdivacky/// is the current offset from the base of the alloca being analyzed. 464207618Srdivacky/// 465207618Srdivacky/// If we see at least one access to the value that is as a vector type, set the 466207618Srdivacky/// SawVec flag. 467245431Sdimbool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset, 468245431Sdim Value* NonConstantIdx) { 469207618Srdivacky for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 470207618Srdivacky Instruction *User = cast<Instruction>(*UI); 471218893Sdim 472207618Srdivacky if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 473207618Srdivacky // Don't break volatile loads. 474226890Sdim if (!LI->isSimple()) 475207618Srdivacky return false; 476218893Sdim // Don't touch MMX operations. 477218893Sdim if (LI->getType()->isX86_MMXTy()) 478218893Sdim return false; 479221345Sdim HadNonMemTransferAccess = true; 480224145Sdim MergeInTypeForLoadOrStore(LI->getType(), Offset); 481207618Srdivacky continue; 482207618Srdivacky } 483218893Sdim 484207618Srdivacky if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 485207618Srdivacky // Storing the pointer, not into the value? 486226890Sdim if (SI->getOperand(0) == V || !SI->isSimple()) return false; 487218893Sdim // Don't touch MMX operations. 488218893Sdim if (SI->getOperand(0)->getType()->isX86_MMXTy()) 489218893Sdim return false; 490221345Sdim HadNonMemTransferAccess = true; 491224145Sdim MergeInTypeForLoadOrStore(SI->getOperand(0)->getType(), Offset); 492207618Srdivacky continue; 493207618Srdivacky } 494218893Sdim 495207618Srdivacky if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 496226890Sdim if (!onlyUsedByLifetimeMarkers(BCI)) 497226890Sdim IsNotTrivial = true; // Can't be mem2reg'd. 498245431Sdim if (!CanConvertToScalar(BCI, Offset, NonConstantIdx)) 499207618Srdivacky return false; 500207618Srdivacky continue; 501207618Srdivacky } 502207618Srdivacky 503207618Srdivacky if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 504207618Srdivacky // If this is a GEP with a variable indices, we can't handle it. 505245431Sdim PointerType* PtrTy = dyn_cast<PointerType>(GEP->getPointerOperandType()); 506245431Sdim if (!PtrTy) 507207618Srdivacky return false; 508218893Sdim 509207618Srdivacky // Compute the offset that this GEP adds to the pointer. 510207618Srdivacky SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 511245431Sdim Value *GEPNonConstantIdx = 0; 512245431Sdim if (!GEP->hasAllConstantIndices()) { 513245431Sdim if (!isa<VectorType>(PtrTy->getElementType())) 514245431Sdim return false; 515245431Sdim if (NonConstantIdx) 516245431Sdim return false; 517245431Sdim GEPNonConstantIdx = Indices.pop_back_val(); 518245431Sdim if (!GEPNonConstantIdx->getType()->isIntegerTy(32)) 519245431Sdim return false; 520245431Sdim HadDynamicAccess = true; 521245431Sdim } else 522245431Sdim GEPNonConstantIdx = NonConstantIdx; 523245431Sdim uint64_t GEPOffset = TD.getIndexedOffset(PtrTy, 524226890Sdim Indices); 525207618Srdivacky // See if all uses can be converted. 526245431Sdim if (!CanConvertToScalar(GEP, Offset+GEPOffset, GEPNonConstantIdx)) 527207618Srdivacky return false; 528207618Srdivacky IsNotTrivial = true; // Can't be mem2reg'd. 529221345Sdim HadNonMemTransferAccess = true; 530207618Srdivacky continue; 531207618Srdivacky } 532207618Srdivacky 533207618Srdivacky // If this is a constant sized memset of a constant value (e.g. 0) we can 534207618Srdivacky // handle it. 535207618Srdivacky if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 536245431Sdim // Store to dynamic index. 537245431Sdim if (NonConstantIdx) 538245431Sdim return false; 539224145Sdim // Store of constant value. 540224145Sdim if (!isa<ConstantInt>(MSI->getValue())) 541207618Srdivacky return false; 542224145Sdim 543224145Sdim // Store of constant size. 544224145Sdim ConstantInt *Len = dyn_cast<ConstantInt>(MSI->getLength()); 545224145Sdim if (!Len) 546224145Sdim return false; 547224145Sdim 548224145Sdim // If the size differs from the alloca, we can only convert the alloca to 549224145Sdim // an integer bag-of-bits. 550224145Sdim // FIXME: This should handle all of the cases that are currently accepted 551224145Sdim // as vector element insertions. 552224145Sdim if (Len->getZExtValue() != AllocaSize || Offset != 0) 553224145Sdim ScalarKind = Integer; 554224145Sdim 555207618Srdivacky IsNotTrivial = true; // Can't be mem2reg'd. 556221345Sdim HadNonMemTransferAccess = true; 557207618Srdivacky continue; 558207618Srdivacky } 559207618Srdivacky 560207618Srdivacky // If this is a memcpy or memmove into or out of the whole allocation, we 561207618Srdivacky // can handle it like a load or store of the scalar type. 562207618Srdivacky if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 563245431Sdim // Store to dynamic index. 564245431Sdim if (NonConstantIdx) 565245431Sdim return false; 566207618Srdivacky ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()); 567207618Srdivacky if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0) 568207618Srdivacky return false; 569218893Sdim 570207618Srdivacky IsNotTrivial = true; // Can't be mem2reg'd. 571207618Srdivacky continue; 572207618Srdivacky } 573218893Sdim 574226890Sdim // If this is a lifetime intrinsic, we can handle it. 575226890Sdim if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 576226890Sdim if (II->getIntrinsicID() == Intrinsic::lifetime_start || 577226890Sdim II->getIntrinsicID() == Intrinsic::lifetime_end) { 578226890Sdim continue; 579226890Sdim } 580226890Sdim } 581226890Sdim 582207618Srdivacky // Otherwise, we cannot handle this! 583207618Srdivacky return false; 584207618Srdivacky } 585218893Sdim 586207618Srdivacky return true; 587207618Srdivacky} 588207618Srdivacky 589207618Srdivacky/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 590207618Srdivacky/// directly. This happens when we are converting an "integer union" to a 591207618Srdivacky/// single integer scalar, or when we are converting a "vector union" to a 592207618Srdivacky/// vector with insert/extractelement instructions. 593207618Srdivacky/// 594207618Srdivacky/// Offset is an offset from the original alloca, in bits that need to be 595207618Srdivacky/// shifted to the right. By the end of this, there should be no uses of Ptr. 596207618Srdivackyvoid ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, 597245431Sdim uint64_t Offset, 598245431Sdim Value* NonConstantIdx) { 599207618Srdivacky while (!Ptr->use_empty()) { 600207618Srdivacky Instruction *User = cast<Instruction>(Ptr->use_back()); 601207618Srdivacky 602207618Srdivacky if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 603245431Sdim ConvertUsesToScalar(CI, NewAI, Offset, NonConstantIdx); 604207618Srdivacky CI->eraseFromParent(); 605207618Srdivacky continue; 606207618Srdivacky } 607207618Srdivacky 608207618Srdivacky if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 609207618Srdivacky // Compute the offset that this GEP adds to the pointer. 610207618Srdivacky SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 611245431Sdim Value* GEPNonConstantIdx = 0; 612245431Sdim if (!GEP->hasAllConstantIndices()) { 613245431Sdim assert(!NonConstantIdx && 614245431Sdim "Dynamic GEP reading from dynamic GEP unsupported"); 615245431Sdim GEPNonConstantIdx = Indices.pop_back_val(); 616245431Sdim } else 617245431Sdim GEPNonConstantIdx = NonConstantIdx; 618207618Srdivacky uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 619226890Sdim Indices); 620245431Sdim ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8, GEPNonConstantIdx); 621207618Srdivacky GEP->eraseFromParent(); 622207618Srdivacky continue; 623207618Srdivacky } 624218893Sdim 625218893Sdim IRBuilder<> Builder(User); 626218893Sdim 627207618Srdivacky if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 628207618Srdivacky // The load is a bit extract from NewAI shifted right by Offset bits. 629226890Sdim Value *LoadedVal = Builder.CreateLoad(NewAI); 630207618Srdivacky Value *NewLoadVal 631245431Sdim = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, 632245431Sdim NonConstantIdx, Builder); 633207618Srdivacky LI->replaceAllUsesWith(NewLoadVal); 634207618Srdivacky LI->eraseFromParent(); 635207618Srdivacky continue; 636207618Srdivacky } 637218893Sdim 638207618Srdivacky if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 639207618Srdivacky assert(SI->getOperand(0) != Ptr && "Consistency error!"); 640207618Srdivacky Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 641207618Srdivacky Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 642245431Sdim NonConstantIdx, Builder); 643207618Srdivacky Builder.CreateStore(New, NewAI); 644207618Srdivacky SI->eraseFromParent(); 645218893Sdim 646207618Srdivacky // If the load we just inserted is now dead, then the inserted store 647207618Srdivacky // overwrote the entire thing. 648207618Srdivacky if (Old->use_empty()) 649207618Srdivacky Old->eraseFromParent(); 650207618Srdivacky continue; 651207618Srdivacky } 652218893Sdim 653207618Srdivacky // If this is a constant sized memset of a constant value (e.g. 0) we can 654207618Srdivacky // transform it into a store of the expanded constant value. 655207618Srdivacky if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 656207618Srdivacky assert(MSI->getRawDest() == Ptr && "Consistency error!"); 657245431Sdim assert(!NonConstantIdx && "Cannot replace dynamic memset with insert"); 658235633Sdim int64_t SNumBytes = cast<ConstantInt>(MSI->getLength())->getSExtValue(); 659235633Sdim if (SNumBytes > 0 && (SNumBytes >> 32) == 0) { 660235633Sdim unsigned NumBytes = static_cast<unsigned>(SNumBytes); 661207618Srdivacky unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 662218893Sdim 663207618Srdivacky // Compute the value replicated the right number of times. 664207618Srdivacky APInt APVal(NumBytes*8, Val); 665207618Srdivacky 666207618Srdivacky // Splat the value if non-zero. 667207618Srdivacky if (Val) 668207618Srdivacky for (unsigned i = 1; i != NumBytes; ++i) 669207618Srdivacky APVal |= APVal << 8; 670218893Sdim 671207618Srdivacky Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 672207618Srdivacky Value *New = ConvertScalar_InsertValue( 673207618Srdivacky ConstantInt::get(User->getContext(), APVal), 674245431Sdim Old, Offset, 0, Builder); 675207618Srdivacky Builder.CreateStore(New, NewAI); 676218893Sdim 677207618Srdivacky // If the load we just inserted is now dead, then the memset overwrote 678207618Srdivacky // the entire thing. 679207618Srdivacky if (Old->use_empty()) 680218893Sdim Old->eraseFromParent(); 681207618Srdivacky } 682207618Srdivacky MSI->eraseFromParent(); 683207618Srdivacky continue; 684207618Srdivacky } 685207618Srdivacky 686207618Srdivacky // If this is a memcpy or memmove into or out of the whole allocation, we 687207618Srdivacky // can handle it like a load or store of the scalar type. 688207618Srdivacky if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 689207618Srdivacky assert(Offset == 0 && "must be store to start of alloca"); 690245431Sdim assert(!NonConstantIdx && "Cannot replace dynamic transfer with insert"); 691218893Sdim 692207618Srdivacky // If the source and destination are both to the same alloca, then this is 693207618Srdivacky // a noop copy-to-self, just delete it. Otherwise, emit a load and store 694207618Srdivacky // as appropriate. 695218893Sdim AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, &TD, 0)); 696218893Sdim 697218893Sdim if (GetUnderlyingObject(MTI->getSource(), &TD, 0) != OrigAI) { 698207618Srdivacky // Dest must be OrigAI, change this to be a load from the original 699207618Srdivacky // pointer (bitcasted), then a store to our new alloca. 700207618Srdivacky assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 701207618Srdivacky Value *SrcPtr = MTI->getSource(); 702226890Sdim PointerType* SPTy = cast<PointerType>(SrcPtr->getType()); 703226890Sdim PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 704218893Sdim if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 705218893Sdim AIPTy = PointerType::get(AIPTy->getElementType(), 706218893Sdim SPTy->getAddressSpace()); 707218893Sdim } 708218893Sdim SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy); 709218893Sdim 710207618Srdivacky LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 711207618Srdivacky SrcVal->setAlignment(MTI->getAlignment()); 712207618Srdivacky Builder.CreateStore(SrcVal, NewAI); 713218893Sdim } else if (GetUnderlyingObject(MTI->getDest(), &TD, 0) != OrigAI) { 714207618Srdivacky // Src must be OrigAI, change this to be a load from NewAI then a store 715207618Srdivacky // through the original dest pointer (bitcasted). 716207618Srdivacky assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 717207618Srdivacky LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 718207618Srdivacky 719226890Sdim PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType()); 720226890Sdim PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 721218893Sdim if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 722218893Sdim AIPTy = PointerType::get(AIPTy->getElementType(), 723218893Sdim DPTy->getAddressSpace()); 724218893Sdim } 725218893Sdim Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy); 726218893Sdim 727207618Srdivacky StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 728207618Srdivacky NewStore->setAlignment(MTI->getAlignment()); 729207618Srdivacky } else { 730207618Srdivacky // Noop transfer. Src == Dst 731207618Srdivacky } 732207618Srdivacky 733207618Srdivacky MTI->eraseFromParent(); 734207618Srdivacky continue; 735207618Srdivacky } 736218893Sdim 737226890Sdim if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 738226890Sdim if (II->getIntrinsicID() == Intrinsic::lifetime_start || 739226890Sdim II->getIntrinsicID() == Intrinsic::lifetime_end) { 740226890Sdim // There's no need to preserve these, as the resulting alloca will be 741226890Sdim // converted to a register anyways. 742226890Sdim II->eraseFromParent(); 743226890Sdim continue; 744226890Sdim } 745226890Sdim } 746226890Sdim 747207618Srdivacky llvm_unreachable("Unsupported operation!"); 748207618Srdivacky } 749207618Srdivacky} 750207618Srdivacky 751207618Srdivacky/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 752207618Srdivacky/// or vector value FromVal, extracting the bits from the offset specified by 753207618Srdivacky/// Offset. This returns the value, which is of type ToType. 754207618Srdivacky/// 755207618Srdivacky/// This happens when we are converting an "integer union" to a single 756207618Srdivacky/// integer scalar, or when we are converting a "vector union" to a vector with 757207618Srdivacky/// insert/extractelement instructions. 758207618Srdivacky/// 759207618Srdivacky/// Offset is an offset from the original alloca, in bits that need to be 760207618Srdivacky/// shifted to the right. 761207618SrdivackyValue *ConvertToScalarInfo:: 762226890SdimConvertScalar_ExtractValue(Value *FromVal, Type *ToType, 763245431Sdim uint64_t Offset, Value* NonConstantIdx, 764245431Sdim IRBuilder<> &Builder) { 765207618Srdivacky // If the load is of the whole new alloca, no conversion is needed. 766226890Sdim Type *FromType = FromVal->getType(); 767221345Sdim if (FromType == ToType && Offset == 0) 768207618Srdivacky return FromVal; 769207618Srdivacky 770207618Srdivacky // If the result alloca is a vector type, this is either an element 771207618Srdivacky // access or a bitcast to another vector type of the same size. 772226890Sdim if (VectorType *VTy = dyn_cast<VectorType>(FromType)) { 773223017Sdim unsigned FromTypeSize = TD.getTypeAllocSize(FromType); 774221345Sdim unsigned ToTypeSize = TD.getTypeAllocSize(ToType); 775226890Sdim if (FromTypeSize == ToTypeSize) 776226890Sdim return Builder.CreateBitCast(FromVal, ToType); 777207618Srdivacky 778207618Srdivacky // Otherwise it must be an element access. 779207618Srdivacky unsigned Elt = 0; 780207618Srdivacky if (Offset) { 781207618Srdivacky unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 782207618Srdivacky Elt = Offset/EltSize; 783207618Srdivacky assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 784207618Srdivacky } 785207618Srdivacky // Return the element extracted out of it. 786245431Sdim Value *Idx; 787245431Sdim if (NonConstantIdx) { 788245431Sdim if (Elt) 789245431Sdim Idx = Builder.CreateAdd(NonConstantIdx, 790245431Sdim Builder.getInt32(Elt), 791245431Sdim "dyn.offset"); 792245431Sdim else 793245431Sdim Idx = NonConstantIdx; 794245431Sdim } else 795245431Sdim Idx = Builder.getInt32(Elt); 796245431Sdim Value *V = Builder.CreateExtractElement(FromVal, Idx); 797207618Srdivacky if (V->getType() != ToType) 798226890Sdim V = Builder.CreateBitCast(V, ToType); 799207618Srdivacky return V; 800207618Srdivacky } 801218893Sdim 802207618Srdivacky // If ToType is a first class aggregate, extract out each of the pieces and 803207618Srdivacky // use insertvalue's to form the FCA. 804226890Sdim if (StructType *ST = dyn_cast<StructType>(ToType)) { 805245431Sdim assert(!NonConstantIdx && 806245431Sdim "Dynamic indexing into struct types not supported"); 807207618Srdivacky const StructLayout &Layout = *TD.getStructLayout(ST); 808207618Srdivacky Value *Res = UndefValue::get(ST); 809207618Srdivacky for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 810207618Srdivacky Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 811207618Srdivacky Offset+Layout.getElementOffsetInBits(i), 812245431Sdim 0, Builder); 813226890Sdim Res = Builder.CreateInsertValue(Res, Elt, i); 814207618Srdivacky } 815207618Srdivacky return Res; 816207618Srdivacky } 817218893Sdim 818226890Sdim if (ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 819245431Sdim assert(!NonConstantIdx && 820245431Sdim "Dynamic indexing into array types not supported"); 821207618Srdivacky uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 822207618Srdivacky Value *Res = UndefValue::get(AT); 823207618Srdivacky for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 824207618Srdivacky Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 825245431Sdim Offset+i*EltSize, 0, Builder); 826226890Sdim Res = Builder.CreateInsertValue(Res, Elt, i); 827207618Srdivacky } 828207618Srdivacky return Res; 829207618Srdivacky } 830207618Srdivacky 831207618Srdivacky // Otherwise, this must be a union that was converted to an integer value. 832226890Sdim IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 833207618Srdivacky 834207618Srdivacky // If this is a big-endian system and the load is narrower than the 835207618Srdivacky // full alloca type, we need to do a shift to get the right bits. 836207618Srdivacky int ShAmt = 0; 837207618Srdivacky if (TD.isBigEndian()) { 838207618Srdivacky // On big-endian machines, the lowest bit is stored at the bit offset 839207618Srdivacky // from the pointer given by getTypeStoreSizeInBits. This matters for 840207618Srdivacky // integers with a bitwidth that is not a multiple of 8. 841207618Srdivacky ShAmt = TD.getTypeStoreSizeInBits(NTy) - 842207618Srdivacky TD.getTypeStoreSizeInBits(ToType) - Offset; 843207618Srdivacky } else { 844207618Srdivacky ShAmt = Offset; 845207618Srdivacky } 846207618Srdivacky 847207618Srdivacky // Note: we support negative bitwidths (with shl) which are not defined. 848207618Srdivacky // We do this to support (f.e.) loads off the end of a structure where 849207618Srdivacky // only some bits are used. 850207618Srdivacky if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 851207618Srdivacky FromVal = Builder.CreateLShr(FromVal, 852226890Sdim ConstantInt::get(FromVal->getType(), ShAmt)); 853207618Srdivacky else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 854218893Sdim FromVal = Builder.CreateShl(FromVal, 855226890Sdim ConstantInt::get(FromVal->getType(), -ShAmt)); 856207618Srdivacky 857207618Srdivacky // Finally, unconditionally truncate the integer to the right width. 858207618Srdivacky unsigned LIBitWidth = TD.getTypeSizeInBits(ToType); 859207618Srdivacky if (LIBitWidth < NTy->getBitWidth()) 860207618Srdivacky FromVal = 861218893Sdim Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 862226890Sdim LIBitWidth)); 863207618Srdivacky else if (LIBitWidth > NTy->getBitWidth()) 864207618Srdivacky FromVal = 865218893Sdim Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 866226890Sdim LIBitWidth)); 867207618Srdivacky 868207618Srdivacky // If the result is an integer, this is a trunc or bitcast. 869207618Srdivacky if (ToType->isIntegerTy()) { 870207618Srdivacky // Should be done. 871207618Srdivacky } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 872207618Srdivacky // Just do a bitcast, we know the sizes match up. 873226890Sdim FromVal = Builder.CreateBitCast(FromVal, ToType); 874207618Srdivacky } else { 875207618Srdivacky // Otherwise must be a pointer. 876226890Sdim FromVal = Builder.CreateIntToPtr(FromVal, ToType); 877207618Srdivacky } 878207618Srdivacky assert(FromVal->getType() == ToType && "Didn't convert right?"); 879207618Srdivacky return FromVal; 880207618Srdivacky} 881207618Srdivacky 882207618Srdivacky/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 883207618Srdivacky/// or vector value "Old" at the offset specified by Offset. 884207618Srdivacky/// 885207618Srdivacky/// This happens when we are converting an "integer union" to a 886207618Srdivacky/// single integer scalar, or when we are converting a "vector union" to a 887207618Srdivacky/// vector with insert/extractelement instructions. 888207618Srdivacky/// 889207618Srdivacky/// Offset is an offset from the original alloca, in bits that need to be 890207618Srdivacky/// shifted to the right. 891245431Sdim/// 892245431Sdim/// NonConstantIdx is an index value if there was a GEP with a non-constant 893245431Sdim/// index value. If this is 0 then all GEPs used to find this insert address 894245431Sdim/// are constant. 895207618SrdivackyValue *ConvertToScalarInfo:: 896207618SrdivackyConvertScalar_InsertValue(Value *SV, Value *Old, 897245431Sdim uint64_t Offset, Value* NonConstantIdx, 898245431Sdim IRBuilder<> &Builder) { 899207618Srdivacky // Convert the stored type to the actual type, shift it left to insert 900207618Srdivacky // then 'or' into place. 901226890Sdim Type *AllocaType = Old->getType(); 902207618Srdivacky LLVMContext &Context = Old->getContext(); 903207618Srdivacky 904226890Sdim if (VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 905207618Srdivacky uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy); 906207618Srdivacky uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType()); 907218893Sdim 908207618Srdivacky // Changing the whole vector with memset or with an access of a different 909207618Srdivacky // vector type? 910226890Sdim if (ValSize == VecSize) 911226890Sdim return Builder.CreateBitCast(SV, AllocaType); 912207618Srdivacky 913207618Srdivacky // Must be an element insertion. 914235633Sdim Type *EltTy = VTy->getElementType(); 915235633Sdim if (SV->getType() != EltTy) 916235633Sdim SV = Builder.CreateBitCast(SV, EltTy); 917235633Sdim uint64_t EltSize = TD.getTypeAllocSizeInBits(EltTy); 918207618Srdivacky unsigned Elt = Offset/EltSize; 919245431Sdim Value *Idx; 920245431Sdim if (NonConstantIdx) { 921245431Sdim if (Elt) 922245431Sdim Idx = Builder.CreateAdd(NonConstantIdx, 923245431Sdim Builder.getInt32(Elt), 924245431Sdim "dyn.offset"); 925245431Sdim else 926245431Sdim Idx = NonConstantIdx; 927245431Sdim } else 928245431Sdim Idx = Builder.getInt32(Elt); 929245431Sdim return Builder.CreateInsertElement(Old, SV, Idx); 930207618Srdivacky } 931218893Sdim 932207618Srdivacky // If SV is a first-class aggregate value, insert each value recursively. 933226890Sdim if (StructType *ST = dyn_cast<StructType>(SV->getType())) { 934245431Sdim assert(!NonConstantIdx && 935245431Sdim "Dynamic indexing into struct types not supported"); 936207618Srdivacky const StructLayout &Layout = *TD.getStructLayout(ST); 937207618Srdivacky for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 938226890Sdim Value *Elt = Builder.CreateExtractValue(SV, i); 939218893Sdim Old = ConvertScalar_InsertValue(Elt, Old, 940207618Srdivacky Offset+Layout.getElementOffsetInBits(i), 941245431Sdim 0, Builder); 942207618Srdivacky } 943207618Srdivacky return Old; 944207618Srdivacky } 945218893Sdim 946226890Sdim if (ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 947245431Sdim assert(!NonConstantIdx && 948245431Sdim "Dynamic indexing into array types not supported"); 949207618Srdivacky uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 950207618Srdivacky for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 951226890Sdim Value *Elt = Builder.CreateExtractValue(SV, i); 952245431Sdim Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, 0, Builder); 953207618Srdivacky } 954207618Srdivacky return Old; 955207618Srdivacky } 956207618Srdivacky 957207618Srdivacky // If SV is a float, convert it to the appropriate integer type. 958207618Srdivacky // If it is a pointer, do the same. 959207618Srdivacky unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 960207618Srdivacky unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); 961207618Srdivacky unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); 962207618Srdivacky unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); 963207618Srdivacky if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 964226890Sdim SV = Builder.CreateBitCast(SV, IntegerType::get(SV->getContext(),SrcWidth)); 965207618Srdivacky else if (SV->getType()->isPointerTy()) 966263509Sdim SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getType())); 967207618Srdivacky 968207618Srdivacky // Zero extend or truncate the value if needed. 969207618Srdivacky if (SV->getType() != AllocaType) { 970207618Srdivacky if (SV->getType()->getPrimitiveSizeInBits() < 971207618Srdivacky AllocaType->getPrimitiveSizeInBits()) 972226890Sdim SV = Builder.CreateZExt(SV, AllocaType); 973207618Srdivacky else { 974207618Srdivacky // Truncation may be needed if storing more than the alloca can hold 975207618Srdivacky // (undefined behavior). 976226890Sdim SV = Builder.CreateTrunc(SV, AllocaType); 977207618Srdivacky SrcWidth = DestWidth; 978207618Srdivacky SrcStoreWidth = DestStoreWidth; 979207618Srdivacky } 980207618Srdivacky } 981207618Srdivacky 982207618Srdivacky // If this is a big-endian system and the store is narrower than the 983207618Srdivacky // full alloca type, we need to do a shift to get the right bits. 984207618Srdivacky int ShAmt = 0; 985207618Srdivacky if (TD.isBigEndian()) { 986207618Srdivacky // On big-endian machines, the lowest bit is stored at the bit offset 987207618Srdivacky // from the pointer given by getTypeStoreSizeInBits. This matters for 988207618Srdivacky // integers with a bitwidth that is not a multiple of 8. 989207618Srdivacky ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 990207618Srdivacky } else { 991207618Srdivacky ShAmt = Offset; 992207618Srdivacky } 993207618Srdivacky 994207618Srdivacky // Note: we support negative bitwidths (with shr) which are not defined. 995207618Srdivacky // We do this to support (f.e.) stores off the end of a structure where 996207618Srdivacky // only some bits in the structure are set. 997207618Srdivacky APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 998207618Srdivacky if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 999226890Sdim SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), ShAmt)); 1000207618Srdivacky Mask <<= ShAmt; 1001207618Srdivacky } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1002226890Sdim SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), -ShAmt)); 1003207618Srdivacky Mask = Mask.lshr(-ShAmt); 1004207618Srdivacky } 1005207618Srdivacky 1006207618Srdivacky // Mask out the bits we are about to insert from the old value, and or 1007207618Srdivacky // in the new bits. 1008207618Srdivacky if (SrcWidth != DestWidth) { 1009207618Srdivacky assert(DestWidth > SrcWidth); 1010207618Srdivacky Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 1011207618Srdivacky SV = Builder.CreateOr(Old, SV, "ins"); 1012207618Srdivacky } 1013207618Srdivacky return SV; 1014207618Srdivacky} 1015207618Srdivacky 1016207618Srdivacky 1017207618Srdivacky//===----------------------------------------------------------------------===// 1018207618Srdivacky// SRoA Driver 1019207618Srdivacky//===----------------------------------------------------------------------===// 1020207618Srdivacky 1021207618Srdivacky 1022193323Sedbool SROA::runOnFunction(Function &F) { 1023245431Sdim TD = getAnalysisIfAvailable<DataLayout>(); 1024198090Srdivacky 1025193323Sed bool Changed = performPromotion(F); 1026198090Srdivacky 1027245431Sdim // FIXME: ScalarRepl currently depends on DataLayout more than it 1028198090Srdivacky // theoretically needs to. It should be refactored in order to support 1029198090Srdivacky // target-independent IR. Until this is done, just skip the actual 1030198090Srdivacky // scalar-replacement portion of this pass. 1031198090Srdivacky if (!TD) return Changed; 1032198090Srdivacky 1033193323Sed while (1) { 1034193323Sed bool LocalChange = performScalarRepl(F); 1035193323Sed if (!LocalChange) break; // No need to repromote if no scalarrepl 1036193323Sed Changed = true; 1037193323Sed LocalChange = performPromotion(F); 1038193323Sed if (!LocalChange) break; // No need to re-scalarrepl if no promotion 1039193323Sed } 1040193323Sed 1041193323Sed return Changed; 1042193323Sed} 1043193323Sed 1044218893Sdimnamespace { 1045218893Sdimclass AllocaPromoter : public LoadAndStorePromoter { 1046218893Sdim AllocaInst *AI; 1047224145Sdim DIBuilder *DIB; 1048224145Sdim SmallVector<DbgDeclareInst *, 4> DDIs; 1049224145Sdim SmallVector<DbgValueInst *, 4> DVIs; 1050218893Sdimpublic: 1051223017Sdim AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S, 1052224145Sdim DIBuilder *DB) 1053224145Sdim : LoadAndStorePromoter(Insts, S), AI(0), DIB(DB) {} 1054245431Sdim 1055218893Sdim void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) { 1056218893Sdim // Remember which alloca we're promoting (for isInstInList). 1057218893Sdim this->AI = AI; 1058235633Sdim if (MDNode *DebugNode = MDNode::getIfExists(AI->getContext(), AI)) { 1059224145Sdim for (Value::use_iterator UI = DebugNode->use_begin(), 1060224145Sdim E = DebugNode->use_end(); UI != E; ++UI) 1061224145Sdim if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI)) 1062224145Sdim DDIs.push_back(DDI); 1063224145Sdim else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI)) 1064224145Sdim DVIs.push_back(DVI); 1065235633Sdim } 1066224145Sdim 1067218893Sdim LoadAndStorePromoter::run(Insts); 1068218893Sdim AI->eraseFromParent(); 1069263509Sdim for (SmallVectorImpl<DbgDeclareInst *>::iterator I = DDIs.begin(), 1070224145Sdim E = DDIs.end(); I != E; ++I) { 1071224145Sdim DbgDeclareInst *DDI = *I; 1072224145Sdim DDI->eraseFromParent(); 1073224145Sdim } 1074263509Sdim for (SmallVectorImpl<DbgValueInst *>::iterator I = DVIs.begin(), 1075224145Sdim E = DVIs.end(); I != E; ++I) { 1076224145Sdim DbgValueInst *DVI = *I; 1077224145Sdim DVI->eraseFromParent(); 1078224145Sdim } 1079218893Sdim } 1080245431Sdim 1081218893Sdim virtual bool isInstInList(Instruction *I, 1082218893Sdim const SmallVectorImpl<Instruction*> &Insts) const { 1083218893Sdim if (LoadInst *LI = dyn_cast<LoadInst>(I)) 1084218893Sdim return LI->getOperand(0) == AI; 1085218893Sdim return cast<StoreInst>(I)->getPointerOperand() == AI; 1086218893Sdim } 1087224145Sdim 1088224145Sdim virtual void updateDebugInfo(Instruction *Inst) const { 1089263509Sdim for (SmallVectorImpl<DbgDeclareInst *>::const_iterator I = DDIs.begin(), 1090224145Sdim E = DDIs.end(); I != E; ++I) { 1091224145Sdim DbgDeclareInst *DDI = *I; 1092224145Sdim if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) 1093224145Sdim ConvertDebugDeclareToDebugValue(DDI, SI, *DIB); 1094224145Sdim else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) 1095224145Sdim ConvertDebugDeclareToDebugValue(DDI, LI, *DIB); 1096224145Sdim } 1097263509Sdim for (SmallVectorImpl<DbgValueInst *>::const_iterator I = DVIs.begin(), 1098224145Sdim E = DVIs.end(); I != E; ++I) { 1099224145Sdim DbgValueInst *DVI = *I; 1100235633Sdim Value *Arg = NULL; 1101224145Sdim if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 1102224145Sdim // If an argument is zero extended then use argument directly. The ZExt 1103224145Sdim // may be zapped by an optimization pass in future. 1104224145Sdim if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0))) 1105235633Sdim Arg = dyn_cast<Argument>(ZExt->getOperand(0)); 1106224145Sdim if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0))) 1107235633Sdim Arg = dyn_cast<Argument>(SExt->getOperand(0)); 1108235633Sdim if (!Arg) 1109235633Sdim Arg = SI->getOperand(0); 1110224145Sdim } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 1111235633Sdim Arg = LI->getOperand(0); 1112235633Sdim } else { 1113235633Sdim continue; 1114224145Sdim } 1115235633Sdim Instruction *DbgVal = 1116235633Sdim DIB->insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()), 1117235633Sdim Inst); 1118235633Sdim DbgVal->setDebugLoc(DVI->getDebugLoc()); 1119224145Sdim } 1120224145Sdim } 1121218893Sdim}; 1122218893Sdim} // end anon namespace 1123193323Sed 1124218893Sdim/// isSafeSelectToSpeculate - Select instructions that use an alloca and are 1125218893Sdim/// subsequently loaded can be rewritten to load both input pointers and then 1126218893Sdim/// select between the result, allowing the load of the alloca to be promoted. 1127218893Sdim/// From this: 1128218893Sdim/// %P2 = select i1 %cond, i32* %Alloca, i32* %Other 1129218893Sdim/// %V = load i32* %P2 1130218893Sdim/// to: 1131218893Sdim/// %V1 = load i32* %Alloca -> will be mem2reg'd 1132218893Sdim/// %V2 = load i32* %Other 1133218893Sdim/// %V = select i1 %cond, i32 %V1, i32 %V2 1134218893Sdim/// 1135218893Sdim/// We can do this to a select if its only uses are loads and if the operand to 1136218893Sdim/// the select can be loaded unconditionally. 1137245431Sdimstatic bool isSafeSelectToSpeculate(SelectInst *SI, const DataLayout *TD) { 1138218893Sdim bool TDerefable = SI->getTrueValue()->isDereferenceablePointer(); 1139218893Sdim bool FDerefable = SI->getFalseValue()->isDereferenceablePointer(); 1140245431Sdim 1141218893Sdim for (Value::use_iterator UI = SI->use_begin(), UE = SI->use_end(); 1142218893Sdim UI != UE; ++UI) { 1143218893Sdim LoadInst *LI = dyn_cast<LoadInst>(*UI); 1144226890Sdim if (LI == 0 || !LI->isSimple()) return false; 1145245431Sdim 1146218893Sdim // Both operands to the select need to be dereferencable, either absolutely 1147218893Sdim // (e.g. allocas) or at this point because we can see other accesses to it. 1148218893Sdim if (!TDerefable && !isSafeToLoadUnconditionally(SI->getTrueValue(), LI, 1149218893Sdim LI->getAlignment(), TD)) 1150218893Sdim return false; 1151218893Sdim if (!FDerefable && !isSafeToLoadUnconditionally(SI->getFalseValue(), LI, 1152218893Sdim LI->getAlignment(), TD)) 1153218893Sdim return false; 1154218893Sdim } 1155245431Sdim 1156218893Sdim return true; 1157218893Sdim} 1158218893Sdim 1159218893Sdim/// isSafePHIToSpeculate - PHI instructions that use an alloca and are 1160218893Sdim/// subsequently loaded can be rewritten to load both input pointers in the pred 1161218893Sdim/// blocks and then PHI the results, allowing the load of the alloca to be 1162218893Sdim/// promoted. 1163218893Sdim/// From this: 1164218893Sdim/// %P2 = phi [i32* %Alloca, i32* %Other] 1165218893Sdim/// %V = load i32* %P2 1166218893Sdim/// to: 1167218893Sdim/// %V1 = load i32* %Alloca -> will be mem2reg'd 1168218893Sdim/// ... 1169218893Sdim/// %V2 = load i32* %Other 1170218893Sdim/// ... 1171218893Sdim/// %V = phi [i32 %V1, i32 %V2] 1172218893Sdim/// 1173218893Sdim/// We can do this to a select if its only uses are loads and if the operand to 1174218893Sdim/// the select can be loaded unconditionally. 1175245431Sdimstatic bool isSafePHIToSpeculate(PHINode *PN, const DataLayout *TD) { 1176218893Sdim // For now, we can only do this promotion if the load is in the same block as 1177218893Sdim // the PHI, and if there are no stores between the phi and load. 1178218893Sdim // TODO: Allow recursive phi users. 1179218893Sdim // TODO: Allow stores. 1180218893Sdim BasicBlock *BB = PN->getParent(); 1181218893Sdim unsigned MaxAlign = 0; 1182218893Sdim for (Value::use_iterator UI = PN->use_begin(), UE = PN->use_end(); 1183218893Sdim UI != UE; ++UI) { 1184218893Sdim LoadInst *LI = dyn_cast<LoadInst>(*UI); 1185226890Sdim if (LI == 0 || !LI->isSimple()) return false; 1186245431Sdim 1187218893Sdim // For now we only allow loads in the same block as the PHI. This is a 1188218893Sdim // common case that happens when instcombine merges two loads through a PHI. 1189218893Sdim if (LI->getParent() != BB) return false; 1190245431Sdim 1191218893Sdim // Ensure that there are no instructions between the PHI and the load that 1192218893Sdim // could store. 1193218893Sdim for (BasicBlock::iterator BBI = PN; &*BBI != LI; ++BBI) 1194218893Sdim if (BBI->mayWriteToMemory()) 1195218893Sdim return false; 1196245431Sdim 1197218893Sdim MaxAlign = std::max(MaxAlign, LI->getAlignment()); 1198218893Sdim } 1199245431Sdim 1200218893Sdim // Okay, we know that we have one or more loads in the same block as the PHI. 1201218893Sdim // We can transform this if it is safe to push the loads into the predecessor 1202218893Sdim // blocks. The only thing to watch out for is that we can't put a possibly 1203218893Sdim // trapping load in the predecessor if it is a critical edge. 1204218893Sdim for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1205218893Sdim BasicBlock *Pred = PN->getIncomingBlock(i); 1206226890Sdim Value *InVal = PN->getIncomingValue(i); 1207218893Sdim 1208226890Sdim // If the terminator of the predecessor has side-effects (an invoke), 1209226890Sdim // there is no safe place to put a load in the predecessor. 1210226890Sdim if (Pred->getTerminator()->mayHaveSideEffects()) 1211226890Sdim return false; 1212226890Sdim 1213226890Sdim // If the value is produced by the terminator of the predecessor 1214226890Sdim // (an invoke), there is no valid place to put a load in the predecessor. 1215226890Sdim if (Pred->getTerminator() == InVal) 1216226890Sdim return false; 1217226890Sdim 1218218893Sdim // If the predecessor has a single successor, then the edge isn't critical. 1219218893Sdim if (Pred->getTerminator()->getNumSuccessors() == 1) 1220218893Sdim continue; 1221218893Sdim 1222218893Sdim // If this pointer is always safe to load, or if we can prove that there is 1223218893Sdim // already a load in the block, then we can move the load to the pred block. 1224218893Sdim if (InVal->isDereferenceablePointer() || 1225218893Sdim isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign, TD)) 1226218893Sdim continue; 1227245431Sdim 1228218893Sdim return false; 1229218893Sdim } 1230245431Sdim 1231218893Sdim return true; 1232218893Sdim} 1233218893Sdim 1234218893Sdim 1235218893Sdim/// tryToMakeAllocaBePromotable - This returns true if the alloca only has 1236218893Sdim/// direct (non-volatile) loads and stores to it. If the alloca is close but 1237218893Sdim/// not quite there, this will transform the code to allow promotion. As such, 1238218893Sdim/// it is a non-pure predicate. 1239245431Sdimstatic bool tryToMakeAllocaBePromotable(AllocaInst *AI, const DataLayout *TD) { 1240218893Sdim SetVector<Instruction*, SmallVector<Instruction*, 4>, 1241218893Sdim SmallPtrSet<Instruction*, 4> > InstsToRewrite; 1242245431Sdim 1243218893Sdim for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end(); 1244218893Sdim UI != UE; ++UI) { 1245218893Sdim User *U = *UI; 1246218893Sdim if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 1247226890Sdim if (!LI->isSimple()) 1248218893Sdim return false; 1249218893Sdim continue; 1250218893Sdim } 1251245431Sdim 1252218893Sdim if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 1253226890Sdim if (SI->getOperand(0) == AI || !SI->isSimple()) 1254218893Sdim return false; // Don't allow a store OF the AI, only INTO the AI. 1255218893Sdim continue; 1256218893Sdim } 1257218893Sdim 1258218893Sdim if (SelectInst *SI = dyn_cast<SelectInst>(U)) { 1259218893Sdim // If the condition being selected on is a constant, fold the select, yes 1260218893Sdim // this does (rarely) happen early on. 1261218893Sdim if (ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition())) { 1262218893Sdim Value *Result = SI->getOperand(1+CI->isZero()); 1263218893Sdim SI->replaceAllUsesWith(Result); 1264218893Sdim SI->eraseFromParent(); 1265245431Sdim 1266218893Sdim // This is very rare and we just scrambled the use list of AI, start 1267218893Sdim // over completely. 1268218893Sdim return tryToMakeAllocaBePromotable(AI, TD); 1269218893Sdim } 1270218893Sdim 1271218893Sdim // If it is safe to turn "load (select c, AI, ptr)" into a select of two 1272218893Sdim // loads, then we can transform this by rewriting the select. 1273218893Sdim if (!isSafeSelectToSpeculate(SI, TD)) 1274218893Sdim return false; 1275245431Sdim 1276218893Sdim InstsToRewrite.insert(SI); 1277218893Sdim continue; 1278218893Sdim } 1279245431Sdim 1280218893Sdim if (PHINode *PN = dyn_cast<PHINode>(U)) { 1281218893Sdim if (PN->use_empty()) { // Dead PHIs can be stripped. 1282218893Sdim InstsToRewrite.insert(PN); 1283218893Sdim continue; 1284218893Sdim } 1285245431Sdim 1286218893Sdim // If it is safe to turn "load (phi [AI, ptr, ...])" into a PHI of loads 1287218893Sdim // in the pred blocks, then we can transform this by rewriting the PHI. 1288218893Sdim if (!isSafePHIToSpeculate(PN, TD)) 1289218893Sdim return false; 1290245431Sdim 1291218893Sdim InstsToRewrite.insert(PN); 1292218893Sdim continue; 1293218893Sdim } 1294245431Sdim 1295226890Sdim if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 1296226890Sdim if (onlyUsedByLifetimeMarkers(BCI)) { 1297226890Sdim InstsToRewrite.insert(BCI); 1298226890Sdim continue; 1299226890Sdim } 1300226890Sdim } 1301245431Sdim 1302218893Sdim return false; 1303218893Sdim } 1304218893Sdim 1305218893Sdim // If there are no instructions to rewrite, then all uses are load/stores and 1306218893Sdim // we're done! 1307218893Sdim if (InstsToRewrite.empty()) 1308218893Sdim return true; 1309245431Sdim 1310218893Sdim // If we have instructions that need to be rewritten for this to be promotable 1311218893Sdim // take care of it now. 1312218893Sdim for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) { 1313226890Sdim if (BitCastInst *BCI = dyn_cast<BitCastInst>(InstsToRewrite[i])) { 1314226890Sdim // This could only be a bitcast used by nothing but lifetime intrinsics. 1315226890Sdim for (BitCastInst::use_iterator I = BCI->use_begin(), E = BCI->use_end(); 1316226890Sdim I != E;) { 1317226890Sdim Use &U = I.getUse(); 1318226890Sdim ++I; 1319226890Sdim cast<Instruction>(U.getUser())->eraseFromParent(); 1320226890Sdim } 1321226890Sdim BCI->eraseFromParent(); 1322226890Sdim continue; 1323226890Sdim } 1324226890Sdim 1325218893Sdim if (SelectInst *SI = dyn_cast<SelectInst>(InstsToRewrite[i])) { 1326218893Sdim // Selects in InstsToRewrite only have load uses. Rewrite each as two 1327218893Sdim // loads with a new select. 1328218893Sdim while (!SI->use_empty()) { 1329218893Sdim LoadInst *LI = cast<LoadInst>(SI->use_back()); 1330245431Sdim 1331218893Sdim IRBuilder<> Builder(LI); 1332245431Sdim LoadInst *TrueLoad = 1333218893Sdim Builder.CreateLoad(SI->getTrueValue(), LI->getName()+".t"); 1334245431Sdim LoadInst *FalseLoad = 1335224145Sdim Builder.CreateLoad(SI->getFalseValue(), LI->getName()+".f"); 1336245431Sdim 1337218893Sdim // Transfer alignment and TBAA info if present. 1338218893Sdim TrueLoad->setAlignment(LI->getAlignment()); 1339218893Sdim FalseLoad->setAlignment(LI->getAlignment()); 1340218893Sdim if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) { 1341218893Sdim TrueLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1342218893Sdim FalseLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1343218893Sdim } 1344245431Sdim 1345218893Sdim Value *V = Builder.CreateSelect(SI->getCondition(), TrueLoad, FalseLoad); 1346218893Sdim V->takeName(LI); 1347218893Sdim LI->replaceAllUsesWith(V); 1348218893Sdim LI->eraseFromParent(); 1349218893Sdim } 1350245431Sdim 1351218893Sdim // Now that all the loads are gone, the select is gone too. 1352218893Sdim SI->eraseFromParent(); 1353218893Sdim continue; 1354218893Sdim } 1355245431Sdim 1356218893Sdim // Otherwise, we have a PHI node which allows us to push the loads into the 1357218893Sdim // predecessors. 1358218893Sdim PHINode *PN = cast<PHINode>(InstsToRewrite[i]); 1359218893Sdim if (PN->use_empty()) { 1360218893Sdim PN->eraseFromParent(); 1361218893Sdim continue; 1362218893Sdim } 1363245431Sdim 1364226890Sdim Type *LoadTy = cast<PointerType>(PN->getType())->getElementType(); 1365221345Sdim PHINode *NewPN = PHINode::Create(LoadTy, PN->getNumIncomingValues(), 1366221345Sdim PN->getName()+".ld", PN); 1367218893Sdim 1368218893Sdim // Get the TBAA tag and alignment to use from one of the loads. It doesn't 1369218893Sdim // matter which one we get and if any differ, it doesn't matter. 1370218893Sdim LoadInst *SomeLoad = cast<LoadInst>(PN->use_back()); 1371218893Sdim MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa); 1372218893Sdim unsigned Align = SomeLoad->getAlignment(); 1373245431Sdim 1374218893Sdim // Rewrite all loads of the PN to use the new PHI. 1375218893Sdim while (!PN->use_empty()) { 1376218893Sdim LoadInst *LI = cast<LoadInst>(PN->use_back()); 1377218893Sdim LI->replaceAllUsesWith(NewPN); 1378218893Sdim LI->eraseFromParent(); 1379218893Sdim } 1380245431Sdim 1381218893Sdim // Inject loads into all of the pred blocks. Keep track of which blocks we 1382218893Sdim // insert them into in case we have multiple edges from the same block. 1383218893Sdim DenseMap<BasicBlock*, LoadInst*> InsertedLoads; 1384245431Sdim 1385218893Sdim for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1386218893Sdim BasicBlock *Pred = PN->getIncomingBlock(i); 1387218893Sdim LoadInst *&Load = InsertedLoads[Pred]; 1388218893Sdim if (Load == 0) { 1389218893Sdim Load = new LoadInst(PN->getIncomingValue(i), 1390218893Sdim PN->getName() + "." + Pred->getName(), 1391218893Sdim Pred->getTerminator()); 1392218893Sdim Load->setAlignment(Align); 1393218893Sdim if (TBAATag) Load->setMetadata(LLVMContext::MD_tbaa, TBAATag); 1394218893Sdim } 1395245431Sdim 1396218893Sdim NewPN->addIncoming(Load, Pred); 1397218893Sdim } 1398245431Sdim 1399218893Sdim PN->eraseFromParent(); 1400218893Sdim } 1401245431Sdim 1402218893Sdim ++NumAdjusted; 1403218893Sdim return true; 1404218893Sdim} 1405218893Sdim 1406193323Sedbool SROA::performPromotion(Function &F) { 1407193323Sed std::vector<AllocaInst*> Allocas; 1408218893Sdim DominatorTree *DT = 0; 1409218893Sdim if (HasDomTree) 1410218893Sdim DT = &getAnalysis<DominatorTree>(); 1411193323Sed 1412193323Sed BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 1413224145Sdim DIBuilder DIB(*F.getParent()); 1414193323Sed bool Changed = false; 1415218893Sdim SmallVector<Instruction*, 64> Insts; 1416193323Sed while (1) { 1417193323Sed Allocas.clear(); 1418193323Sed 1419193323Sed // Find allocas that are safe to promote, by looking at all instructions in 1420193323Sed // the entry node 1421193323Sed for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 1422193323Sed if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 1423218893Sdim if (tryToMakeAllocaBePromotable(AI, TD)) 1424193323Sed Allocas.push_back(AI); 1425193323Sed 1426193323Sed if (Allocas.empty()) break; 1427193323Sed 1428218893Sdim if (HasDomTree) 1429218893Sdim PromoteMemToReg(Allocas, *DT); 1430218893Sdim else { 1431218893Sdim SSAUpdater SSA; 1432218893Sdim for (unsigned i = 0, e = Allocas.size(); i != e; ++i) { 1433218893Sdim AllocaInst *AI = Allocas[i]; 1434245431Sdim 1435218893Sdim // Build list of instructions to promote. 1436218893Sdim for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 1437218893Sdim UI != E; ++UI) 1438218893Sdim Insts.push_back(cast<Instruction>(*UI)); 1439224145Sdim AllocaPromoter(Insts, SSA, &DIB).run(AI, Insts); 1440218893Sdim Insts.clear(); 1441218893Sdim } 1442218893Sdim } 1443193323Sed NumPromoted += Allocas.size(); 1444193323Sed Changed = true; 1445193323Sed } 1446193323Sed 1447193323Sed return Changed; 1448193323Sed} 1449193323Sed 1450207618Srdivacky 1451203954Srdivacky/// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 1452203954Srdivacky/// SROA. It must be a struct or array type with a small number of elements. 1453245431Sdimbool SROA::ShouldAttemptScalarRepl(AllocaInst *AI) { 1454226890Sdim Type *T = AI->getAllocatedType(); 1455245431Sdim // Do not promote any struct that has too many members. 1456226890Sdim if (StructType *ST = dyn_cast<StructType>(T)) 1457245431Sdim return ST->getNumElements() <= StructMemberThreshold; 1458245431Sdim // Do not promote any array that has too many elements. 1459226890Sdim if (ArrayType *AT = dyn_cast<ArrayType>(T)) 1460245431Sdim return AT->getNumElements() <= ArrayElementThreshold; 1461203954Srdivacky return false; 1462193323Sed} 1463193323Sed 1464193323Sed// performScalarRepl - This algorithm is a simple worklist driven algorithm, 1465252723Sdim// which runs on all of the alloca instructions in the entry block, removing 1466252723Sdim// them if they are only used by getelementptr instructions. 1467193323Sed// 1468193323Sedbool SROA::performScalarRepl(Function &F) { 1469198892Srdivacky std::vector<AllocaInst*> WorkList; 1470193323Sed 1471207618Srdivacky // Scan the entry basic block, adding allocas to the worklist. 1472193323Sed BasicBlock &BB = F.getEntryBlock(); 1473193323Sed for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 1474198892Srdivacky if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 1475193323Sed WorkList.push_back(A); 1476193323Sed 1477193323Sed // Process the worklist 1478193323Sed bool Changed = false; 1479193323Sed while (!WorkList.empty()) { 1480198892Srdivacky AllocaInst *AI = WorkList.back(); 1481193323Sed WorkList.pop_back(); 1482218893Sdim 1483193323Sed // Handle dead allocas trivially. These can be formed by SROA'ing arrays 1484193323Sed // with unused elements. 1485193323Sed if (AI->use_empty()) { 1486193323Sed AI->eraseFromParent(); 1487207618Srdivacky Changed = true; 1488193323Sed continue; 1489193323Sed } 1490193323Sed 1491193323Sed // If this alloca is impossible for us to promote, reject it early. 1492193323Sed if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 1493193323Sed continue; 1494218893Sdim 1495193323Sed // Check to see if we can perform the core SROA transformation. We cannot 1496193323Sed // transform the allocation instruction if it is an array allocation 1497193323Sed // (allocations OF arrays are ok though), and an allocation of a scalar 1498193323Sed // value cannot be decomposed at all. 1499193323Sed uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 1500193323Sed 1501198090Srdivacky // Do not promote [0 x %struct]. 1502198090Srdivacky if (AllocaSize == 0) continue; 1503218893Sdim 1504207618Srdivacky // Do not promote any struct whose size is too big. 1505207618Srdivacky if (AllocaSize > SRThreshold) continue; 1506218893Sdim 1507203954Srdivacky // If the alloca looks like a good candidate for scalar replacement, and if 1508203954Srdivacky // all its users can be transformed, then split up the aggregate into its 1509203954Srdivacky // separate elements. 1510203954Srdivacky if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 1511203954Srdivacky DoScalarReplacement(AI, WorkList); 1512203954Srdivacky Changed = true; 1513203954Srdivacky continue; 1514203954Srdivacky } 1515203954Srdivacky 1516193323Sed // If we can turn this aggregate value (potentially with casts) into a 1517193323Sed // simple scalar value that can be mem2reg'd into a register value. 1518193323Sed // IsNotTrivial tracks whether this is something that mem2reg could have 1519193323Sed // promoted itself. If so, we don't want to transform it needlessly. Note 1520193323Sed // that we can't just check based on the type: the alloca may be of an i32 1521193323Sed // but that has pointer arithmetic to set byte 3 of it or something. 1522245431Sdim if (AllocaInst *NewAI = ConvertToScalarInfo( 1523245431Sdim (unsigned)AllocaSize, *TD, ScalarLoadThreshold).TryConvert(AI)) { 1524193323Sed NewAI->takeName(AI); 1525193323Sed AI->eraseFromParent(); 1526193323Sed ++NumConverted; 1527193323Sed Changed = true; 1528193323Sed continue; 1529218893Sdim } 1530218893Sdim 1531193323Sed // Otherwise, couldn't process this alloca. 1532193323Sed } 1533193323Sed 1534193323Sed return Changed; 1535193323Sed} 1536193323Sed 1537193323Sed/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 1538193323Sed/// predicate, do SROA now. 1539218893Sdimvoid SROA::DoScalarReplacement(AllocaInst *AI, 1540198892Srdivacky std::vector<AllocaInst*> &WorkList) { 1541202375Srdivacky DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 1542193323Sed SmallVector<AllocaInst*, 32> ElementAllocas; 1543226890Sdim if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 1544193323Sed ElementAllocas.reserve(ST->getNumContainedTypes()); 1545193323Sed for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 1546218893Sdim AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 1547193323Sed AI->getAlignment(), 1548198090Srdivacky AI->getName() + "." + Twine(i), AI); 1549193323Sed ElementAllocas.push_back(NA); 1550193323Sed WorkList.push_back(NA); // Add to worklist for recursive processing 1551193323Sed } 1552193323Sed } else { 1553226890Sdim ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 1554193323Sed ElementAllocas.reserve(AT->getNumElements()); 1555226890Sdim Type *ElTy = AT->getElementType(); 1556193323Sed for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1557193323Sed AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 1558198090Srdivacky AI->getName() + "." + Twine(i), AI); 1559193323Sed ElementAllocas.push_back(NA); 1560193323Sed WorkList.push_back(NA); // Add to worklist for recursive processing 1561193323Sed } 1562193323Sed } 1563193323Sed 1564201360Srdivacky // Now that we have created the new alloca instructions, rewrite all the 1565201360Srdivacky // uses of the old alloca. 1566201360Srdivacky RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 1567193323Sed 1568201360Srdivacky // Now erase any instructions that were made dead while rewriting the alloca. 1569201360Srdivacky DeleteDeadInstructions(); 1570201360Srdivacky AI->eraseFromParent(); 1571193323Sed 1572210299Sed ++NumReplaced; 1573201360Srdivacky} 1574193323Sed 1575201360Srdivacky/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 1576201360Srdivacky/// recursively including all their operands that become trivially dead. 1577201360Srdivackyvoid SROA::DeleteDeadInstructions() { 1578201360Srdivacky while (!DeadInsts.empty()) { 1579201360Srdivacky Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 1580193323Sed 1581201360Srdivacky for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 1582201360Srdivacky if (Instruction *U = dyn_cast<Instruction>(*OI)) { 1583201360Srdivacky // Zero out the operand and see if it becomes trivially dead. 1584201360Srdivacky // (But, don't add allocas to the dead instruction list -- they are 1585201360Srdivacky // already on the worklist and will be deleted separately.) 1586201360Srdivacky *OI = 0; 1587201360Srdivacky if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 1588201360Srdivacky DeadInsts.push_back(U); 1589201360Srdivacky } 1590201360Srdivacky 1591201360Srdivacky I->eraseFromParent(); 1592193323Sed } 1593193323Sed} 1594218893Sdim 1595201360Srdivacky/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 1596201360Srdivacky/// performing scalar replacement of alloca AI. The results are flagged in 1597201360Srdivacky/// the Info parameter. Offset indicates the position within AI that is 1598201360Srdivacky/// referenced by this instruction. 1599218893Sdimvoid SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset, 1600201360Srdivacky AllocaInfo &Info) { 1601201360Srdivacky for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1602201360Srdivacky Instruction *User = cast<Instruction>(*UI); 1603193323Sed 1604201360Srdivacky if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1605218893Sdim isSafeForScalarRepl(BC, Offset, Info); 1606201360Srdivacky } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1607201360Srdivacky uint64_t GEPOffset = Offset; 1608218893Sdim isSafeGEP(GEPI, GEPOffset, Info); 1609201360Srdivacky if (!Info.isUnsafe) 1610218893Sdim isSafeForScalarRepl(GEPI, GEPOffset, Info); 1611210299Sed } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1612201360Srdivacky ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1613218893Sdim if (Length == 0) 1614218893Sdim return MarkUnsafe(Info, User); 1615235633Sdim if (Length->isNegative()) 1616235633Sdim return MarkUnsafe(Info, User); 1617235633Sdim 1618218893Sdim isSafeMemAccess(Offset, Length->getZExtValue(), 0, 1619218893Sdim UI.getOperandNo() == 0, Info, MI, 1620218893Sdim true /*AllowWholeAccess*/); 1621201360Srdivacky } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1622226890Sdim if (!LI->isSimple()) 1623218893Sdim return MarkUnsafe(Info, User); 1624226890Sdim Type *LIType = LI->getType(); 1625218893Sdim isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1626218893Sdim LIType, false, Info, LI, true /*AllowWholeAccess*/); 1627218893Sdim Info.hasALoadOrStore = true; 1628245431Sdim 1629201360Srdivacky } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1630193323Sed // Store is ok if storing INTO the pointer, not storing the pointer 1631226890Sdim if (!SI->isSimple() || SI->getOperand(0) == I) 1632218893Sdim return MarkUnsafe(Info, User); 1633245431Sdim 1634226890Sdim Type *SIType = SI->getOperand(0)->getType(); 1635218893Sdim isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1636218893Sdim SIType, true, Info, SI, true /*AllowWholeAccess*/); 1637218893Sdim Info.hasALoadOrStore = true; 1638226890Sdim } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 1639226890Sdim if (II->getIntrinsicID() != Intrinsic::lifetime_start && 1640226890Sdim II->getIntrinsicID() != Intrinsic::lifetime_end) 1641226890Sdim return MarkUnsafe(Info, User); 1642218893Sdim } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1643218893Sdim isSafePHISelectUseForScalarRepl(User, Offset, Info); 1644201360Srdivacky } else { 1645218893Sdim return MarkUnsafe(Info, User); 1646193323Sed } 1647201360Srdivacky if (Info.isUnsafe) return; 1648193323Sed } 1649193323Sed} 1650193323Sed 1651245431Sdim 1652218893Sdim/// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer 1653218893Sdim/// derived from the alloca, we can often still split the alloca into elements. 1654218893Sdim/// This is useful if we have a large alloca where one element is phi'd 1655218893Sdim/// together somewhere: we can SRoA and promote all the other elements even if 1656218893Sdim/// we end up not being able to promote this one. 1657218893Sdim/// 1658218893Sdim/// All we require is that the uses of the PHI do not index into other parts of 1659218893Sdim/// the alloca. The most important use case for this is single load and stores 1660218893Sdim/// that are PHI'd together, which can happen due to code sinking. 1661218893Sdimvoid SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset, 1662218893Sdim AllocaInfo &Info) { 1663218893Sdim // If we've already checked this PHI, don't do it again. 1664218893Sdim if (PHINode *PN = dyn_cast<PHINode>(I)) 1665218893Sdim if (!Info.CheckedPHIs.insert(PN)) 1666218893Sdim return; 1667245431Sdim 1668218893Sdim for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1669218893Sdim Instruction *User = cast<Instruction>(*UI); 1670245431Sdim 1671218893Sdim if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1672218893Sdim isSafePHISelectUseForScalarRepl(BC, Offset, Info); 1673218893Sdim } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1674218893Sdim // Only allow "bitcast" GEPs for simplicity. We could generalize this, 1675218893Sdim // but would have to prove that we're staying inside of an element being 1676218893Sdim // promoted. 1677218893Sdim if (!GEPI->hasAllZeroIndices()) 1678218893Sdim return MarkUnsafe(Info, User); 1679218893Sdim isSafePHISelectUseForScalarRepl(GEPI, Offset, Info); 1680218893Sdim } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1681226890Sdim if (!LI->isSimple()) 1682218893Sdim return MarkUnsafe(Info, User); 1683226890Sdim Type *LIType = LI->getType(); 1684218893Sdim isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1685218893Sdim LIType, false, Info, LI, false /*AllowWholeAccess*/); 1686218893Sdim Info.hasALoadOrStore = true; 1687245431Sdim 1688218893Sdim } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1689218893Sdim // Store is ok if storing INTO the pointer, not storing the pointer 1690226890Sdim if (!SI->isSimple() || SI->getOperand(0) == I) 1691218893Sdim return MarkUnsafe(Info, User); 1692245431Sdim 1693226890Sdim Type *SIType = SI->getOperand(0)->getType(); 1694218893Sdim isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1695218893Sdim SIType, true, Info, SI, false /*AllowWholeAccess*/); 1696218893Sdim Info.hasALoadOrStore = true; 1697218893Sdim } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1698218893Sdim isSafePHISelectUseForScalarRepl(User, Offset, Info); 1699218893Sdim } else { 1700218893Sdim return MarkUnsafe(Info, User); 1701218893Sdim } 1702218893Sdim if (Info.isUnsafe) return; 1703218893Sdim } 1704218893Sdim} 1705218893Sdim 1706201360Srdivacky/// isSafeGEP - Check if a GEP instruction can be handled for scalar 1707201360Srdivacky/// replacement. It is safe when all the indices are constant, in-bounds 1708201360Srdivacky/// references, and when the resulting offset corresponds to an element within 1709201360Srdivacky/// the alloca type. The results are flagged in the Info parameter. Upon 1710201360Srdivacky/// return, Offset is adjusted as specified by the GEP indices. 1711218893Sdimvoid SROA::isSafeGEP(GetElementPtrInst *GEPI, 1712201360Srdivacky uint64_t &Offset, AllocaInfo &Info) { 1713201360Srdivacky gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 1714201360Srdivacky if (GEPIt == E) 1715201360Srdivacky return; 1716245431Sdim bool NonConstant = false; 1717245431Sdim unsigned NonConstantIdxSize = 0; 1718193323Sed 1719201360Srdivacky // Walk through the GEP type indices, checking the types that this indexes 1720201360Srdivacky // into. 1721201360Srdivacky for (; GEPIt != E; ++GEPIt) { 1722201360Srdivacky // Ignore struct elements, no extra checking needed for these. 1723204642Srdivacky if ((*GEPIt)->isStructTy()) 1724201360Srdivacky continue; 1725193323Sed 1726201360Srdivacky ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 1727252723Sdim if (!IdxVal) 1728252723Sdim return MarkUnsafe(Info, GEPI); 1729193323Sed } 1730193323Sed 1731201360Srdivacky // Compute the offset due to this GEP and check if the alloca has a 1732201360Srdivacky // component element at that offset. 1733201360Srdivacky SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1734245431Sdim // If this GEP is non constant then the last operand must have been a 1735245431Sdim // dynamic index into a vector. Pop this now as it has no impact on the 1736245431Sdim // constant part of the offset. 1737245431Sdim if (NonConstant) 1738245431Sdim Indices.pop_back(); 1739226890Sdim Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices); 1740245431Sdim if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 1741245431Sdim NonConstantIdxSize)) 1742218893Sdim MarkUnsafe(Info, GEPI); 1743201360Srdivacky} 1744193323Sed 1745218893Sdim/// isHomogeneousAggregate - Check if type T is a struct or array containing 1746218893Sdim/// elements of the same type (which is always true for arrays). If so, 1747218893Sdim/// return true with NumElts and EltTy set to the number of elements and the 1748218893Sdim/// element type, respectively. 1749226890Sdimstatic bool isHomogeneousAggregate(Type *T, unsigned &NumElts, 1750226890Sdim Type *&EltTy) { 1751226890Sdim if (ArrayType *AT = dyn_cast<ArrayType>(T)) { 1752218893Sdim NumElts = AT->getNumElements(); 1753218893Sdim EltTy = (NumElts == 0 ? 0 : AT->getElementType()); 1754218893Sdim return true; 1755218893Sdim } 1756226890Sdim if (StructType *ST = dyn_cast<StructType>(T)) { 1757218893Sdim NumElts = ST->getNumContainedTypes(); 1758218893Sdim EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0)); 1759218893Sdim for (unsigned n = 1; n < NumElts; ++n) { 1760218893Sdim if (ST->getContainedType(n) != EltTy) 1761218893Sdim return false; 1762218893Sdim } 1763218893Sdim return true; 1764218893Sdim } 1765218893Sdim return false; 1766218893Sdim} 1767218893Sdim 1768218893Sdim/// isCompatibleAggregate - Check if T1 and T2 are either the same type or are 1769218893Sdim/// "homogeneous" aggregates with the same element type and number of elements. 1770226890Sdimstatic bool isCompatibleAggregate(Type *T1, Type *T2) { 1771218893Sdim if (T1 == T2) 1772218893Sdim return true; 1773218893Sdim 1774218893Sdim unsigned NumElts1, NumElts2; 1775226890Sdim Type *EltTy1, *EltTy2; 1776218893Sdim if (isHomogeneousAggregate(T1, NumElts1, EltTy1) && 1777218893Sdim isHomogeneousAggregate(T2, NumElts2, EltTy2) && 1778218893Sdim NumElts1 == NumElts2 && 1779218893Sdim EltTy1 == EltTy2) 1780218893Sdim return true; 1781218893Sdim 1782218893Sdim return false; 1783218893Sdim} 1784218893Sdim 1785201360Srdivacky/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1786201360Srdivacky/// alloca or has an offset and size that corresponds to a component element 1787201360Srdivacky/// within it. The offset checked here may have been formed from a GEP with a 1788201360Srdivacky/// pointer bitcasted to a different type. 1789218893Sdim/// 1790218893Sdim/// If AllowWholeAccess is true, then this allows uses of the entire alloca as a 1791218893Sdim/// unit. If false, it only allows accesses known to be in a single element. 1792218893Sdimvoid SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 1793226890Sdim Type *MemOpType, bool isStore, 1794218893Sdim AllocaInfo &Info, Instruction *TheAccess, 1795218893Sdim bool AllowWholeAccess) { 1796201360Srdivacky // Check if this is a load/store of the entire alloca. 1797218893Sdim if (Offset == 0 && AllowWholeAccess && 1798218893Sdim MemSize == TD->getTypeAllocSize(Info.AI->getAllocatedType())) { 1799218893Sdim // This can be safe for MemIntrinsics (where MemOpType is 0) and integer 1800218893Sdim // loads/stores (which are essentially the same as the MemIntrinsics with 1801218893Sdim // regard to copying padding between elements). But, if an alloca is 1802218893Sdim // flagged as both a source and destination of such operations, we'll need 1803218893Sdim // to check later for padding between elements. 1804218893Sdim if (!MemOpType || MemOpType->isIntegerTy()) { 1805218893Sdim if (isStore) 1806218893Sdim Info.isMemCpyDst = true; 1807218893Sdim else 1808218893Sdim Info.isMemCpySrc = true; 1809201360Srdivacky return; 1810193323Sed } 1811218893Sdim // This is also safe for references using a type that is compatible with 1812218893Sdim // the type of the alloca, so that loads/stores can be rewritten using 1813218893Sdim // insertvalue/extractvalue. 1814218893Sdim if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) { 1815218893Sdim Info.hasSubelementAccess = true; 1816218893Sdim return; 1817218893Sdim } 1818193323Sed } 1819201360Srdivacky // Check if the offset/size correspond to a component within the alloca type. 1820226890Sdim Type *T = Info.AI->getAllocatedType(); 1821218893Sdim if (TypeHasComponent(T, Offset, MemSize)) { 1822218893Sdim Info.hasSubelementAccess = true; 1823201360Srdivacky return; 1824218893Sdim } 1825193323Sed 1826218893Sdim return MarkUnsafe(Info, TheAccess); 1827193323Sed} 1828193323Sed 1829201360Srdivacky/// TypeHasComponent - Return true if T has a component type with the 1830201360Srdivacky/// specified offset and size. If Size is zero, do not check the size. 1831226890Sdimbool SROA::TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size) { 1832226890Sdim Type *EltTy; 1833201360Srdivacky uint64_t EltSize; 1834226890Sdim if (StructType *ST = dyn_cast<StructType>(T)) { 1835201360Srdivacky const StructLayout *Layout = TD->getStructLayout(ST); 1836201360Srdivacky unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1837201360Srdivacky EltTy = ST->getContainedType(EltIdx); 1838201360Srdivacky EltSize = TD->getTypeAllocSize(EltTy); 1839201360Srdivacky Offset -= Layout->getElementOffset(EltIdx); 1840226890Sdim } else if (ArrayType *AT = dyn_cast<ArrayType>(T)) { 1841201360Srdivacky EltTy = AT->getElementType(); 1842201360Srdivacky EltSize = TD->getTypeAllocSize(EltTy); 1843201360Srdivacky if (Offset >= AT->getNumElements() * EltSize) 1844201360Srdivacky return false; 1845201360Srdivacky Offset %= EltSize; 1846245431Sdim } else if (VectorType *VT = dyn_cast<VectorType>(T)) { 1847245431Sdim EltTy = VT->getElementType(); 1848245431Sdim EltSize = TD->getTypeAllocSize(EltTy); 1849245431Sdim if (Offset >= VT->getNumElements() * EltSize) 1850245431Sdim return false; 1851245431Sdim Offset %= EltSize; 1852201360Srdivacky } else { 1853201360Srdivacky return false; 1854193323Sed } 1855201360Srdivacky if (Offset == 0 && (Size == 0 || EltSize == Size)) 1856201360Srdivacky return true; 1857201360Srdivacky // Check if the component spans multiple elements. 1858201360Srdivacky if (Offset + Size > EltSize) 1859201360Srdivacky return false; 1860201360Srdivacky return TypeHasComponent(EltTy, Offset, Size); 1861193323Sed} 1862193323Sed 1863201360Srdivacky/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1864201360Srdivacky/// the instruction I, which references it, to use the separate elements. 1865201360Srdivacky/// Offset indicates the position within AI that is referenced by this 1866201360Srdivacky/// instruction. 1867201360Srdivackyvoid SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1868263509Sdim SmallVectorImpl<AllocaInst *> &NewElts) { 1869218893Sdim for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) { 1870218893Sdim Use &TheUse = UI.getUse(); 1871218893Sdim Instruction *User = cast<Instruction>(*UI++); 1872193323Sed 1873201360Srdivacky if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1874201360Srdivacky RewriteBitCast(BC, AI, Offset, NewElts); 1875218893Sdim continue; 1876218893Sdim } 1877245431Sdim 1878218893Sdim if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1879201360Srdivacky RewriteGEP(GEPI, AI, Offset, NewElts); 1880218893Sdim continue; 1881218893Sdim } 1882245431Sdim 1883218893Sdim if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1884201360Srdivacky ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1885201360Srdivacky uint64_t MemSize = Length->getZExtValue(); 1886201360Srdivacky if (Offset == 0 && 1887201360Srdivacky MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 1888201360Srdivacky RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1889201360Srdivacky // Otherwise the intrinsic can only touch a single element and the 1890201360Srdivacky // address operand will be updated, so nothing else needs to be done. 1891218893Sdim continue; 1892218893Sdim } 1893226890Sdim 1894226890Sdim if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 1895226890Sdim if (II->getIntrinsicID() == Intrinsic::lifetime_start || 1896226890Sdim II->getIntrinsicID() == Intrinsic::lifetime_end) { 1897226890Sdim RewriteLifetimeIntrinsic(II, AI, Offset, NewElts); 1898226890Sdim } 1899226890Sdim continue; 1900226890Sdim } 1901245431Sdim 1902218893Sdim if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1903226890Sdim Type *LIType = LI->getType(); 1904245431Sdim 1905218893Sdim if (isCompatibleAggregate(LIType, AI->getAllocatedType())) { 1906201360Srdivacky // Replace: 1907201360Srdivacky // %res = load { i32, i32 }* %alloc 1908201360Srdivacky // with: 1909201360Srdivacky // %load.0 = load i32* %alloc.0 1910201360Srdivacky // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1911201360Srdivacky // %load.1 = load i32* %alloc.1 1912201360Srdivacky // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1913201360Srdivacky // (Also works for arrays instead of structs) 1914201360Srdivacky Value *Insert = UndefValue::get(LIType); 1915223017Sdim IRBuilder<> Builder(LI); 1916201360Srdivacky for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1917223017Sdim Value *Load = Builder.CreateLoad(NewElts[i], "load"); 1918223017Sdim Insert = Builder.CreateInsertValue(Insert, Load, i, "insert"); 1919201360Srdivacky } 1920201360Srdivacky LI->replaceAllUsesWith(Insert); 1921201360Srdivacky DeadInsts.push_back(LI); 1922204642Srdivacky } else if (LIType->isIntegerTy() && 1923201360Srdivacky TD->getTypeAllocSize(LIType) == 1924201360Srdivacky TD->getTypeAllocSize(AI->getAllocatedType())) { 1925201360Srdivacky // If this is a load of the entire alloca to an integer, rewrite it. 1926201360Srdivacky RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1927193323Sed } 1928218893Sdim continue; 1929218893Sdim } 1930245431Sdim 1931218893Sdim if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1932201360Srdivacky Value *Val = SI->getOperand(0); 1933226890Sdim Type *SIType = Val->getType(); 1934218893Sdim if (isCompatibleAggregate(SIType, AI->getAllocatedType())) { 1935201360Srdivacky // Replace: 1936201360Srdivacky // store { i32, i32 } %val, { i32, i32 }* %alloc 1937201360Srdivacky // with: 1938201360Srdivacky // %val.0 = extractvalue { i32, i32 } %val, 0 1939201360Srdivacky // store i32 %val.0, i32* %alloc.0 1940201360Srdivacky // %val.1 = extractvalue { i32, i32 } %val, 1 1941201360Srdivacky // store i32 %val.1, i32* %alloc.1 1942201360Srdivacky // (Also works for arrays instead of structs) 1943223017Sdim IRBuilder<> Builder(SI); 1944201360Srdivacky for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1945223017Sdim Value *Extract = Builder.CreateExtractValue(Val, i, Val->getName()); 1946223017Sdim Builder.CreateStore(Extract, NewElts[i]); 1947201360Srdivacky } 1948201360Srdivacky DeadInsts.push_back(SI); 1949204642Srdivacky } else if (SIType->isIntegerTy() && 1950201360Srdivacky TD->getTypeAllocSize(SIType) == 1951201360Srdivacky TD->getTypeAllocSize(AI->getAllocatedType())) { 1952201360Srdivacky // If this is a store of the entire alloca from an integer, rewrite it. 1953201360Srdivacky RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 1954193323Sed } 1955218893Sdim continue; 1956193323Sed } 1957245431Sdim 1958218893Sdim if (isa<SelectInst>(User) || isa<PHINode>(User)) { 1959245431Sdim // If we have a PHI user of the alloca itself (as opposed to a GEP or 1960218893Sdim // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to 1961218893Sdim // the new pointer. 1962218893Sdim if (!isa<AllocaInst>(I)) continue; 1963245431Sdim 1964218893Sdim assert(Offset == 0 && NewElts[0] && 1965218893Sdim "Direct alloca use should have a zero offset"); 1966245431Sdim 1967218893Sdim // If we have a use of the alloca, we know the derived uses will be 1968218893Sdim // utilizing just the first element of the scalarized result. Insert a 1969218893Sdim // bitcast of the first alloca before the user as required. 1970218893Sdim AllocaInst *NewAI = NewElts[0]; 1971218893Sdim BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI); 1972218893Sdim NewAI->moveBefore(BCI); 1973218893Sdim TheUse = BCI; 1974218893Sdim continue; 1975218893Sdim } 1976193323Sed } 1977193323Sed} 1978193323Sed 1979201360Srdivacky/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 1980201360Srdivacky/// and recursively continue updating all of its uses. 1981201360Srdivackyvoid SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 1982263509Sdim SmallVectorImpl<AllocaInst *> &NewElts) { 1983201360Srdivacky RewriteForScalarRepl(BC, AI, Offset, NewElts); 1984201360Srdivacky if (BC->getOperand(0) != AI) 1985201360Srdivacky return; 1986193323Sed 1987201360Srdivacky // The bitcast references the original alloca. Replace its uses with 1988235633Sdim // references to the alloca containing offset zero (which is normally at 1989235633Sdim // index zero, but might not be in cases involving structs with elements 1990235633Sdim // of size zero). 1991235633Sdim Type *T = AI->getAllocatedType(); 1992235633Sdim uint64_t EltOffset = 0; 1993235633Sdim Type *IdxTy; 1994235633Sdim uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1995235633Sdim Instruction *Val = NewElts[Idx]; 1996201360Srdivacky if (Val->getType() != BC->getDestTy()) { 1997201360Srdivacky Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 1998201360Srdivacky Val->takeName(BC); 1999201360Srdivacky } 2000201360Srdivacky BC->replaceAllUsesWith(Val); 2001201360Srdivacky DeadInsts.push_back(BC); 2002201360Srdivacky} 2003193323Sed 2004201360Srdivacky/// FindElementAndOffset - Return the index of the element containing Offset 2005201360Srdivacky/// within the specified type, which must be either a struct or an array. 2006201360Srdivacky/// Sets T to the type of the element and Offset to the offset within that 2007201360Srdivacky/// element. IdxTy is set to the type of the index result to be used in a 2008201360Srdivacky/// GEP instruction. 2009226890Sdimuint64_t SROA::FindElementAndOffset(Type *&T, uint64_t &Offset, 2010226890Sdim Type *&IdxTy) { 2011201360Srdivacky uint64_t Idx = 0; 2012226890Sdim if (StructType *ST = dyn_cast<StructType>(T)) { 2013201360Srdivacky const StructLayout *Layout = TD->getStructLayout(ST); 2014201360Srdivacky Idx = Layout->getElementContainingOffset(Offset); 2015201360Srdivacky T = ST->getContainedType(Idx); 2016201360Srdivacky Offset -= Layout->getElementOffset(Idx); 2017201360Srdivacky IdxTy = Type::getInt32Ty(T->getContext()); 2018201360Srdivacky return Idx; 2019245431Sdim } else if (ArrayType *AT = dyn_cast<ArrayType>(T)) { 2020245431Sdim T = AT->getElementType(); 2021245431Sdim uint64_t EltSize = TD->getTypeAllocSize(T); 2022245431Sdim Idx = Offset / EltSize; 2023245431Sdim Offset -= Idx * EltSize; 2024245431Sdim IdxTy = Type::getInt64Ty(T->getContext()); 2025245431Sdim return Idx; 2026193323Sed } 2027245431Sdim VectorType *VT = cast<VectorType>(T); 2028245431Sdim T = VT->getElementType(); 2029201360Srdivacky uint64_t EltSize = TD->getTypeAllocSize(T); 2030201360Srdivacky Idx = Offset / EltSize; 2031201360Srdivacky Offset -= Idx * EltSize; 2032201360Srdivacky IdxTy = Type::getInt64Ty(T->getContext()); 2033201360Srdivacky return Idx; 2034193323Sed} 2035193323Sed 2036201360Srdivacky/// RewriteGEP - Check if this GEP instruction moves the pointer across 2037201360Srdivacky/// elements of the alloca that are being split apart, and if so, rewrite 2038201360Srdivacky/// the GEP to be relative to the new element. 2039201360Srdivackyvoid SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 2040263509Sdim SmallVectorImpl<AllocaInst *> &NewElts) { 2041201360Srdivacky uint64_t OldOffset = Offset; 2042201360Srdivacky SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 2043245431Sdim // If the GEP was dynamic then it must have been a dynamic vector lookup. 2044245431Sdim // In this case, it must be the last GEP operand which is dynamic so keep that 2045245431Sdim // aside until we've found the constant GEP offset then add it back in at the 2046245431Sdim // end. 2047245431Sdim Value* NonConstantIdx = 0; 2048245431Sdim if (!GEPI->hasAllConstantIndices()) 2049245431Sdim NonConstantIdx = Indices.pop_back_val(); 2050226890Sdim Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices); 2051201360Srdivacky 2052201360Srdivacky RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 2053201360Srdivacky 2054226890Sdim Type *T = AI->getAllocatedType(); 2055226890Sdim Type *IdxTy; 2056201360Srdivacky uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 2057201360Srdivacky if (GEPI->getOperand(0) == AI) 2058201360Srdivacky OldIdx = ~0ULL; // Force the GEP to be rewritten. 2059201360Srdivacky 2060201360Srdivacky T = AI->getAllocatedType(); 2061201360Srdivacky uint64_t EltOffset = Offset; 2062201360Srdivacky uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 2063201360Srdivacky 2064201360Srdivacky // If this GEP does not move the pointer across elements of the alloca 2065201360Srdivacky // being split, then it does not needs to be rewritten. 2066201360Srdivacky if (Idx == OldIdx) 2067201360Srdivacky return; 2068201360Srdivacky 2069226890Sdim Type *i32Ty = Type::getInt32Ty(AI->getContext()); 2070201360Srdivacky SmallVector<Value*, 8> NewArgs; 2071201360Srdivacky NewArgs.push_back(Constant::getNullValue(i32Ty)); 2072201360Srdivacky while (EltOffset != 0) { 2073201360Srdivacky uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 2074201360Srdivacky NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 2075201360Srdivacky } 2076245431Sdim if (NonConstantIdx) { 2077245431Sdim Type* GepTy = T; 2078245431Sdim // This GEP has a dynamic index. We need to add "i32 0" to index through 2079245431Sdim // any structs or arrays in the original type until we get to the vector 2080245431Sdim // to index. 2081245431Sdim while (!isa<VectorType>(GepTy)) { 2082245431Sdim NewArgs.push_back(Constant::getNullValue(i32Ty)); 2083245431Sdim GepTy = cast<CompositeType>(GepTy)->getTypeAtIndex(0U); 2084245431Sdim } 2085245431Sdim NewArgs.push_back(NonConstantIdx); 2086245431Sdim } 2087201360Srdivacky Instruction *Val = NewElts[Idx]; 2088201360Srdivacky if (NewArgs.size() > 1) { 2089226890Sdim Val = GetElementPtrInst::CreateInBounds(Val, NewArgs, "", GEPI); 2090201360Srdivacky Val->takeName(GEPI); 2091201360Srdivacky } 2092201360Srdivacky if (Val->getType() != GEPI->getType()) 2093203954Srdivacky Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 2094201360Srdivacky GEPI->replaceAllUsesWith(Val); 2095201360Srdivacky DeadInsts.push_back(GEPI); 2096201360Srdivacky} 2097201360Srdivacky 2098226890Sdim/// RewriteLifetimeIntrinsic - II is a lifetime.start/lifetime.end. Rewrite it 2099226890Sdim/// to mark the lifetime of the scalarized memory. 2100226890Sdimvoid SROA::RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI, 2101226890Sdim uint64_t Offset, 2102263509Sdim SmallVectorImpl<AllocaInst *> &NewElts) { 2103226890Sdim ConstantInt *OldSize = cast<ConstantInt>(II->getArgOperand(0)); 2104226890Sdim // Put matching lifetime markers on everything from Offset up to 2105226890Sdim // Offset+OldSize. 2106226890Sdim Type *AIType = AI->getAllocatedType(); 2107226890Sdim uint64_t NewOffset = Offset; 2108226890Sdim Type *IdxTy; 2109226890Sdim uint64_t Idx = FindElementAndOffset(AIType, NewOffset, IdxTy); 2110226890Sdim 2111226890Sdim IRBuilder<> Builder(II); 2112226890Sdim uint64_t Size = OldSize->getLimitedValue(); 2113226890Sdim 2114226890Sdim if (NewOffset) { 2115226890Sdim // Splice the first element and index 'NewOffset' bytes in. SROA will 2116226890Sdim // split the alloca again later. 2117226890Sdim Value *V = Builder.CreateBitCast(NewElts[Idx], Builder.getInt8PtrTy()); 2118226890Sdim V = Builder.CreateGEP(V, Builder.getInt64(NewOffset)); 2119226890Sdim 2120226890Sdim IdxTy = NewElts[Idx]->getAllocatedType(); 2121226890Sdim uint64_t EltSize = TD->getTypeAllocSize(IdxTy) - NewOffset; 2122226890Sdim if (EltSize > Size) { 2123226890Sdim EltSize = Size; 2124226890Sdim Size = 0; 2125226890Sdim } else { 2126226890Sdim Size -= EltSize; 2127226890Sdim } 2128226890Sdim if (II->getIntrinsicID() == Intrinsic::lifetime_start) 2129226890Sdim Builder.CreateLifetimeStart(V, Builder.getInt64(EltSize)); 2130226890Sdim else 2131226890Sdim Builder.CreateLifetimeEnd(V, Builder.getInt64(EltSize)); 2132226890Sdim ++Idx; 2133226890Sdim } 2134226890Sdim 2135226890Sdim for (; Idx != NewElts.size() && Size; ++Idx) { 2136226890Sdim IdxTy = NewElts[Idx]->getAllocatedType(); 2137226890Sdim uint64_t EltSize = TD->getTypeAllocSize(IdxTy); 2138226890Sdim if (EltSize > Size) { 2139226890Sdim EltSize = Size; 2140226890Sdim Size = 0; 2141226890Sdim } else { 2142226890Sdim Size -= EltSize; 2143226890Sdim } 2144226890Sdim if (II->getIntrinsicID() == Intrinsic::lifetime_start) 2145226890Sdim Builder.CreateLifetimeStart(NewElts[Idx], 2146226890Sdim Builder.getInt64(EltSize)); 2147226890Sdim else 2148226890Sdim Builder.CreateLifetimeEnd(NewElts[Idx], 2149226890Sdim Builder.getInt64(EltSize)); 2150226890Sdim } 2151226890Sdim DeadInsts.push_back(II); 2152226890Sdim} 2153226890Sdim 2154193323Sed/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 2155193323Sed/// Rewrite it to copy or set the elements of the scalarized memory. 2156263509Sdimvoid 2157263509SdimSROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 2158263509Sdim AllocaInst *AI, 2159263509Sdim SmallVectorImpl<AllocaInst *> &NewElts) { 2160193323Sed // If this is a memcpy/memmove, construct the other pointer as the 2161193323Sed // appropriate type. The "Other" pointer is the pointer that goes to memory 2162193323Sed // that doesn't have anything to do with the alloca that we are promoting. For 2163193323Sed // memset, this Value* stays null. 2164193323Sed Value *OtherPtr = 0; 2165193323Sed unsigned MemAlignment = MI->getAlignment(); 2166193323Sed if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 2167201360Srdivacky if (Inst == MTI->getRawDest()) 2168193323Sed OtherPtr = MTI->getRawSource(); 2169193323Sed else { 2170201360Srdivacky assert(Inst == MTI->getRawSource()); 2171193323Sed OtherPtr = MTI->getRawDest(); 2172193323Sed } 2173193323Sed } 2174200581Srdivacky 2175193323Sed // If there is an other pointer, we want to convert it to the same pointer 2176193323Sed // type as AI has, so we can GEP through it safely. 2177193323Sed if (OtherPtr) { 2178210299Sed unsigned AddrSpace = 2179210299Sed cast<PointerType>(OtherPtr->getType())->getAddressSpace(); 2180201360Srdivacky 2181201360Srdivacky // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 2182201360Srdivacky // optimization, but it's also required to detect the corner case where 2183201360Srdivacky // both pointer operands are referencing the same memory, and where 2184201360Srdivacky // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 2185201360Srdivacky // function is only called for mem intrinsics that access the whole 2186201360Srdivacky // aggregate, so non-zero GEPs are not an issue here.) 2187210299Sed OtherPtr = OtherPtr->stripPointerCasts(); 2188218893Sdim 2189202878Srdivacky // Copying the alloca to itself is a no-op: just delete it. 2190202878Srdivacky if (OtherPtr == AI || OtherPtr == NewElts[0]) { 2191202878Srdivacky // This code will run twice for a no-op memcpy -- once for each operand. 2192202878Srdivacky // Put only one reference to MI on the DeadInsts list. 2193263509Sdim for (SmallVectorImpl<Value *>::const_iterator I = DeadInsts.begin(), 2194202878Srdivacky E = DeadInsts.end(); I != E; ++I) 2195202878Srdivacky if (*I == MI) return; 2196202878Srdivacky DeadInsts.push_back(MI); 2197201360Srdivacky return; 2198202878Srdivacky } 2199218893Sdim 2200193323Sed // If the pointer is not the right type, insert a bitcast to the right 2201193323Sed // type. 2202226890Sdim Type *NewTy = 2203210299Sed PointerType::get(AI->getType()->getElementType(), AddrSpace); 2204218893Sdim 2205210299Sed if (OtherPtr->getType() != NewTy) 2206210299Sed OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI); 2207193323Sed } 2208218893Sdim 2209193323Sed // Process each element of the aggregate. 2210201360Srdivacky bool SROADest = MI->getRawDest() == Inst; 2211218893Sdim 2212198090Srdivacky Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 2213193323Sed 2214193323Sed for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2215193323Sed // If this is a memcpy/memmove, emit a GEP of the other element address. 2216193323Sed Value *OtherElt = 0; 2217193323Sed unsigned OtherEltAlign = MemAlignment; 2218218893Sdim 2219202878Srdivacky if (OtherPtr) { 2220198090Srdivacky Value *Idx[2] = { Zero, 2221198090Srdivacky ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 2222226890Sdim OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, 2223203954Srdivacky OtherPtr->getName()+"."+Twine(i), 2224201360Srdivacky MI); 2225193323Sed uint64_t EltOffset; 2226226890Sdim PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 2227226890Sdim Type *OtherTy = OtherPtrTy->getElementType(); 2228226890Sdim if (StructType *ST = dyn_cast<StructType>(OtherTy)) { 2229193323Sed EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 2230193323Sed } else { 2231226890Sdim Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); 2232193323Sed EltOffset = TD->getTypeAllocSize(EltTy)*i; 2233193323Sed } 2234218893Sdim 2235193323Sed // The alignment of the other pointer is the guaranteed alignment of the 2236193323Sed // element, which is affected by both the known alignment of the whole 2237193323Sed // mem intrinsic and the alignment of the element. If the alignment of 2238193323Sed // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 2239193323Sed // known alignment is just 4 bytes. 2240193323Sed OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 2241193323Sed } 2242218893Sdim 2243193323Sed Value *EltPtr = NewElts[i]; 2244226890Sdim Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 2245218893Sdim 2246193323Sed // If we got down to a scalar, insert a load or store as appropriate. 2247193323Sed if (EltTy->isSingleValueType()) { 2248193323Sed if (isa<MemTransferInst>(MI)) { 2249193323Sed if (SROADest) { 2250193323Sed // From Other to Alloca. 2251193323Sed Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 2252193323Sed new StoreInst(Elt, EltPtr, MI); 2253193323Sed } else { 2254193323Sed // From Alloca to Other. 2255193323Sed Value *Elt = new LoadInst(EltPtr, "tmp", MI); 2256193323Sed new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 2257193323Sed } 2258193323Sed continue; 2259193323Sed } 2260193323Sed assert(isa<MemSetInst>(MI)); 2261218893Sdim 2262193323Sed // If the stored element is zero (common case), just store a null 2263193323Sed // constant. 2264193323Sed Constant *StoreVal; 2265210299Sed if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) { 2266193323Sed if (CI->isZero()) { 2267198090Srdivacky StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 2268193323Sed } else { 2269193323Sed // If EltTy is a vector type, get the element type. 2270226890Sdim Type *ValTy = EltTy->getScalarType(); 2271194612Sed 2272193323Sed // Construct an integer with the right value. 2273193323Sed unsigned EltSize = TD->getTypeSizeInBits(ValTy); 2274193323Sed APInt OneVal(EltSize, CI->getZExtValue()); 2275193323Sed APInt TotalVal(OneVal); 2276193323Sed // Set each byte. 2277193323Sed for (unsigned i = 0; 8*i < EltSize; ++i) { 2278193323Sed TotalVal = TotalVal.shl(8); 2279193323Sed TotalVal |= OneVal; 2280193323Sed } 2281218893Sdim 2282193323Sed // Convert the integer value to the appropriate type. 2283207618Srdivacky StoreVal = ConstantInt::get(CI->getContext(), TotalVal); 2284204642Srdivacky if (ValTy->isPointerTy()) 2285198090Srdivacky StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 2286203954Srdivacky else if (ValTy->isFloatingPointTy()) 2287198090Srdivacky StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 2288193323Sed assert(StoreVal->getType() == ValTy && "Type mismatch!"); 2289218893Sdim 2290193323Sed // If the requested value was a vector constant, create it. 2291226890Sdim if (EltTy->isVectorTy()) { 2292226890Sdim unsigned NumElts = cast<VectorType>(EltTy)->getNumElements(); 2293235633Sdim StoreVal = ConstantVector::getSplat(NumElts, StoreVal); 2294193323Sed } 2295193323Sed } 2296193323Sed new StoreInst(StoreVal, EltPtr, MI); 2297193323Sed continue; 2298193323Sed } 2299193323Sed // Otherwise, if we're storing a byte variable, use a memset call for 2300193323Sed // this element. 2301193323Sed } 2302218893Sdim 2303193323Sed unsigned EltSize = TD->getTypeAllocSize(EltTy); 2304235633Sdim if (!EltSize) 2305235633Sdim continue; 2306218893Sdim 2307218893Sdim IRBuilder<> Builder(MI); 2308218893Sdim 2309193323Sed // Finally, insert the meminst for this element. 2310218893Sdim if (isa<MemSetInst>(MI)) { 2311218893Sdim Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize, 2312218893Sdim MI->isVolatile()); 2313193323Sed } else { 2314218893Sdim assert(isa<MemTransferInst>(MI)); 2315218893Sdim Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr 2316218893Sdim Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr 2317218893Sdim 2318218893Sdim if (isa<MemCpyInst>(MI)) 2319218893Sdim Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile()); 2320218893Sdim else 2321218893Sdim Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile()); 2322193323Sed } 2323193323Sed } 2324201360Srdivacky DeadInsts.push_back(MI); 2325193323Sed} 2326193323Sed 2327200581Srdivacky/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 2328193323Sed/// overwrites the entire allocation. Extract out the pieces of the stored 2329193323Sed/// integer and store them individually. 2330263509Sdimvoid 2331263509SdimSROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 2332263509Sdim SmallVectorImpl<AllocaInst *> &NewElts) { 2333193323Sed // Extract each element out of the integer according to its structure offset 2334193323Sed // and store the element value to the individual alloca. 2335193323Sed Value *SrcVal = SI->getOperand(0); 2336226890Sdim Type *AllocaEltTy = AI->getAllocatedType(); 2337193323Sed uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2338218893Sdim 2339218893Sdim IRBuilder<> Builder(SI); 2340245431Sdim 2341193323Sed // Handle tail padding by extending the operand 2342193323Sed if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 2343218893Sdim SrcVal = Builder.CreateZExt(SrcVal, 2344218893Sdim IntegerType::get(SI->getContext(), AllocaSizeBits)); 2345193323Sed 2346202375Srdivacky DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 2347198090Srdivacky << '\n'); 2348193323Sed 2349193323Sed // There are two forms here: AI could be an array or struct. Both cases 2350193323Sed // have different ways to compute the element offset. 2351226890Sdim if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2352193323Sed const StructLayout *Layout = TD->getStructLayout(EltSTy); 2353218893Sdim 2354193323Sed for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2355193323Sed // Get the number of bits to shift SrcVal to get the value. 2356226890Sdim Type *FieldTy = EltSTy->getElementType(i); 2357193323Sed uint64_t Shift = Layout->getElementOffsetInBits(i); 2358218893Sdim 2359193323Sed if (TD->isBigEndian()) 2360193323Sed Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 2361218893Sdim 2362193323Sed Value *EltVal = SrcVal; 2363193323Sed if (Shift) { 2364198090Srdivacky Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2365218893Sdim EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2366193323Sed } 2367218893Sdim 2368193323Sed // Truncate down to an integer of the right size. 2369193323Sed uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2370218893Sdim 2371193323Sed // Ignore zero sized fields like {}, they obviously contain no data. 2372193323Sed if (FieldSizeBits == 0) continue; 2373218893Sdim 2374193323Sed if (FieldSizeBits != AllocaSizeBits) 2375218893Sdim EltVal = Builder.CreateTrunc(EltVal, 2376218893Sdim IntegerType::get(SI->getContext(), FieldSizeBits)); 2377193323Sed Value *DestField = NewElts[i]; 2378193323Sed if (EltVal->getType() == FieldTy) { 2379193323Sed // Storing to an integer field of this size, just do it. 2380204642Srdivacky } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 2381193323Sed // Bitcast to the right element type (for fp/vector values). 2382218893Sdim EltVal = Builder.CreateBitCast(EltVal, FieldTy); 2383193323Sed } else { 2384193323Sed // Otherwise, bitcast the dest pointer (for aggregates). 2385218893Sdim DestField = Builder.CreateBitCast(DestField, 2386218893Sdim PointerType::getUnqual(EltVal->getType())); 2387193323Sed } 2388193323Sed new StoreInst(EltVal, DestField, SI); 2389193323Sed } 2390218893Sdim 2391193323Sed } else { 2392226890Sdim ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 2393226890Sdim Type *ArrayEltTy = ATy->getElementType(); 2394193323Sed uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2395193323Sed uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 2396193323Sed 2397193323Sed uint64_t Shift; 2398218893Sdim 2399193323Sed if (TD->isBigEndian()) 2400193323Sed Shift = AllocaSizeBits-ElementOffset; 2401218893Sdim else 2402193323Sed Shift = 0; 2403218893Sdim 2404193323Sed for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2405193323Sed // Ignore zero sized fields like {}, they obviously contain no data. 2406193323Sed if (ElementSizeBits == 0) continue; 2407218893Sdim 2408193323Sed Value *EltVal = SrcVal; 2409193323Sed if (Shift) { 2410198090Srdivacky Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2411218893Sdim EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2412193323Sed } 2413218893Sdim 2414193323Sed // Truncate down to an integer of the right size. 2415193323Sed if (ElementSizeBits != AllocaSizeBits) 2416218893Sdim EltVal = Builder.CreateTrunc(EltVal, 2417218893Sdim IntegerType::get(SI->getContext(), 2418218893Sdim ElementSizeBits)); 2419193323Sed Value *DestField = NewElts[i]; 2420193323Sed if (EltVal->getType() == ArrayEltTy) { 2421193323Sed // Storing to an integer field of this size, just do it. 2422203954Srdivacky } else if (ArrayEltTy->isFloatingPointTy() || 2423204642Srdivacky ArrayEltTy->isVectorTy()) { 2424193323Sed // Bitcast to the right element type (for fp/vector values). 2425218893Sdim EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy); 2426193323Sed } else { 2427193323Sed // Otherwise, bitcast the dest pointer (for aggregates). 2428218893Sdim DestField = Builder.CreateBitCast(DestField, 2429218893Sdim PointerType::getUnqual(EltVal->getType())); 2430193323Sed } 2431193323Sed new StoreInst(EltVal, DestField, SI); 2432218893Sdim 2433193323Sed if (TD->isBigEndian()) 2434193323Sed Shift -= ElementOffset; 2435218893Sdim else 2436193323Sed Shift += ElementOffset; 2437193323Sed } 2438193323Sed } 2439218893Sdim 2440201360Srdivacky DeadInsts.push_back(SI); 2441193323Sed} 2442193323Sed 2443200581Srdivacky/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 2444193323Sed/// an integer. Load the individual pieces to form the aggregate value. 2445263509Sdimvoid 2446263509SdimSROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 2447263509Sdim SmallVectorImpl<AllocaInst *> &NewElts) { 2448193323Sed // Extract each element out of the NewElts according to its structure offset 2449193323Sed // and form the result value. 2450226890Sdim Type *AllocaEltTy = AI->getAllocatedType(); 2451193323Sed uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2452218893Sdim 2453202375Srdivacky DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 2454198090Srdivacky << '\n'); 2455218893Sdim 2456193323Sed // There are two forms here: AI could be an array or struct. Both cases 2457193323Sed // have different ways to compute the element offset. 2458193323Sed const StructLayout *Layout = 0; 2459193323Sed uint64_t ArrayEltBitOffset = 0; 2460226890Sdim if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2461193323Sed Layout = TD->getStructLayout(EltSTy); 2462193323Sed } else { 2463226890Sdim Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 2464193323Sed ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2465218893Sdim } 2466218893Sdim 2467218893Sdim Value *ResultVal = 2468198090Srdivacky Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 2469218893Sdim 2470193323Sed for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2471193323Sed // Load the value from the alloca. If the NewElt is an aggregate, cast 2472193323Sed // the pointer to an integer of the same size before doing the load. 2473193323Sed Value *SrcField = NewElts[i]; 2474226890Sdim Type *FieldTy = 2475193323Sed cast<PointerType>(SrcField->getType())->getElementType(); 2476193323Sed uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2477218893Sdim 2478193323Sed // Ignore zero sized fields like {}, they obviously contain no data. 2479193323Sed if (FieldSizeBits == 0) continue; 2480218893Sdim 2481226890Sdim IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 2482198090Srdivacky FieldSizeBits); 2483204642Srdivacky if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 2484204642Srdivacky !FieldTy->isVectorTy()) 2485195340Sed SrcField = new BitCastInst(SrcField, 2486198090Srdivacky PointerType::getUnqual(FieldIntTy), 2487193323Sed "", LI); 2488193323Sed SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 2489193323Sed 2490193323Sed // If SrcField is a fp or vector of the right size but that isn't an 2491193323Sed // integer type, bitcast to an integer so we can shift it. 2492193323Sed if (SrcField->getType() != FieldIntTy) 2493193323Sed SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 2494193323Sed 2495193323Sed // Zero extend the field to be the same size as the final alloca so that 2496193323Sed // we can shift and insert it. 2497193323Sed if (SrcField->getType() != ResultVal->getType()) 2498193323Sed SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 2499218893Sdim 2500193323Sed // Determine the number of bits to shift SrcField. 2501193323Sed uint64_t Shift; 2502193323Sed if (Layout) // Struct case. 2503193323Sed Shift = Layout->getElementOffsetInBits(i); 2504193323Sed else // Array case. 2505193323Sed Shift = i*ArrayEltBitOffset; 2506218893Sdim 2507193323Sed if (TD->isBigEndian()) 2508193323Sed Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 2509218893Sdim 2510193323Sed if (Shift) { 2511198090Srdivacky Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 2512193323Sed SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 2513193323Sed } 2514193323Sed 2515210299Sed // Don't create an 'or x, 0' on the first iteration. 2516210299Sed if (!isa<Constant>(ResultVal) || 2517210299Sed !cast<Constant>(ResultVal)->isNullValue()) 2518210299Sed ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 2519210299Sed else 2520210299Sed ResultVal = SrcField; 2521193323Sed } 2522193323Sed 2523193323Sed // Handle tail padding by truncating the result 2524193323Sed if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 2525193323Sed ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 2526193323Sed 2527193323Sed LI->replaceAllUsesWith(ResultVal); 2528201360Srdivacky DeadInsts.push_back(LI); 2529193323Sed} 2530193323Sed 2531193323Sed/// HasPadding - Return true if the specified type has any structure or 2532218893Sdim/// alignment padding in between the elements that would be split apart 2533218893Sdim/// by SROA; return false otherwise. 2534245431Sdimstatic bool HasPadding(Type *Ty, const DataLayout &TD) { 2535226890Sdim if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 2536218893Sdim Ty = ATy->getElementType(); 2537218893Sdim return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 2538218893Sdim } 2539193323Sed 2540218893Sdim // SROA currently handles only Arrays and Structs. 2541226890Sdim StructType *STy = cast<StructType>(Ty); 2542218893Sdim const StructLayout *SL = TD.getStructLayout(STy); 2543218893Sdim unsigned PrevFieldBitOffset = 0; 2544218893Sdim for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 2545218893Sdim unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 2546193323Sed 2547218893Sdim // Check to see if there is any padding between this element and the 2548218893Sdim // previous one. 2549218893Sdim if (i) { 2550218893Sdim unsigned PrevFieldEnd = 2551193323Sed PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 2552218893Sdim if (PrevFieldEnd < FieldBitOffset) 2553193323Sed return true; 2554193323Sed } 2555218893Sdim PrevFieldBitOffset = FieldBitOffset; 2556193323Sed } 2557218893Sdim // Check for tail padding. 2558218893Sdim if (unsigned EltCount = STy->getNumElements()) { 2559218893Sdim unsigned PrevFieldEnd = PrevFieldBitOffset + 2560218893Sdim TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 2561218893Sdim if (PrevFieldEnd < SL->getSizeInBits()) 2562218893Sdim return true; 2563218893Sdim } 2564218893Sdim return false; 2565193323Sed} 2566193323Sed 2567193323Sed/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 2568193323Sed/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 2569193323Sed/// or 1 if safe after canonicalization has been performed. 2570202878Srdivackybool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 2571193323Sed // Loop over the use list of the alloca. We can only transform it if all of 2572193323Sed // the users are safe to transform. 2573218893Sdim AllocaInfo Info(AI); 2574218893Sdim 2575218893Sdim isSafeForScalarRepl(AI, 0, Info); 2576201360Srdivacky if (Info.isUnsafe) { 2577202375Srdivacky DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 2578202878Srdivacky return false; 2579193323Sed } 2580218893Sdim 2581193323Sed // Okay, we know all the users are promotable. If the aggregate is a memcpy 2582193323Sed // source and destination, we have to be careful. In particular, the memcpy 2583193323Sed // could be moving around elements that live in structure padding of the LLVM 2584193323Sed // types, but may actually be used. In these cases, we refuse to promote the 2585193323Sed // struct. 2586193323Sed if (Info.isMemCpySrc && Info.isMemCpyDst && 2587201360Srdivacky HasPadding(AI->getAllocatedType(), *TD)) 2588202878Srdivacky return false; 2589193323Sed 2590218893Sdim // If the alloca never has an access to just *part* of it, but is accessed 2591218893Sdim // via loads and stores, then we should use ConvertToScalarInfo to promote 2592218893Sdim // the alloca instead of promoting each piece at a time and inserting fission 2593218893Sdim // and fusion code. 2594218893Sdim if (!Info.hasSubelementAccess && Info.hasALoadOrStore) { 2595218893Sdim // If the struct/array just has one element, use basic SRoA. 2596226890Sdim if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 2597218893Sdim if (ST->getNumElements() > 1) return false; 2598218893Sdim } else { 2599218893Sdim if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1) 2600218893Sdim return false; 2601218893Sdim } 2602218893Sdim } 2603193323Sed 2604193323Sed return true; 2605193323Sed} 2606