ScalarReplAggregates.cpp revision 202375
150769Sdfr//===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===// 250769Sdfr// 350769Sdfr// The LLVM Compiler Infrastructure 450769Sdfr// 550769Sdfr// This file is distributed under the University of Illinois Open Source 650769Sdfr// License. See LICENSE.TXT for details. 750769Sdfr// 850769Sdfr//===----------------------------------------------------------------------===// 950769Sdfr// 1050769Sdfr// This transformation implements the well known scalar replacement of 1150769Sdfr// aggregates transformation. This xform breaks up alloca instructions of 1250769Sdfr// aggregate type (structure or array) into individual alloca instructions for 1350769Sdfr// each member (if possible). Then, if possible, it transforms the individual 1450769Sdfr// alloca instructions into nice clean scalar SSA form. 1550769Sdfr// 1650769Sdfr// This combines a simple SRoA algorithm with the Mem2Reg algorithm because 1750769Sdfr// often interact, especially for C++ programs. As such, iterating between 1850769Sdfr// SRoA, then Mem2Reg until we run out of things to promote works well. 1950769Sdfr// 2050769Sdfr//===----------------------------------------------------------------------===// 2150769Sdfr 2250769Sdfr#define DEBUG_TYPE "scalarrepl" 2350769Sdfr#include "llvm/Transforms/Scalar.h" 2450769Sdfr#include "llvm/Constants.h" 2550769Sdfr#include "llvm/DerivedTypes.h" 2650769Sdfr#include "llvm/Function.h" 2750769Sdfr#include "llvm/GlobalVariable.h" 2850769Sdfr#include "llvm/Instructions.h" 29116181Sobrien#include "llvm/IntrinsicInst.h" 30116181Sobrien#include "llvm/LLVMContext.h" 31116181Sobrien#include "llvm/Pass.h" 3250769Sdfr#include "llvm/Analysis/Dominators.h" 3350769Sdfr#include "llvm/Target/TargetData.h" 3450769Sdfr#include "llvm/Transforms/Utils/PromoteMemToReg.h" 3550769Sdfr#include "llvm/Transforms/Utils/Local.h" 3650769Sdfr#include "llvm/Support/Debug.h" 3750769Sdfr#include "llvm/Support/ErrorHandling.h" 3850769Sdfr#include "llvm/Support/GetElementPtrTypeIterator.h" 3950769Sdfr#include "llvm/Support/IRBuilder.h" 4050769Sdfr#include "llvm/Support/MathExtras.h" 4165176Sdfr#include "llvm/Support/raw_ostream.h" 4250769Sdfr#include "llvm/ADT/SmallVector.h" 4350769Sdfr#include "llvm/ADT/Statistic.h" 44139268Simpusing namespace llvm; 45139268Simp 4650769SdfrSTATISTIC(NumReplaced, "Number of allocas broken up"); 4750769SdfrSTATISTIC(NumPromoted, "Number of allocas promoted"); 4850769SdfrSTATISTIC(NumConverted, "Number of aggregates converted to scalar"); 4950769SdfrSTATISTIC(NumGlobals, "Number of allocas copied from constant global"); 5050769Sdfr 5150769Sdfrnamespace { 5250769Sdfr struct SROA : public FunctionPass { 5350769Sdfr static char ID; // Pass identification, replacement for typeid 5450769Sdfr explicit SROA(signed T = -1) : FunctionPass(&ID) { 55139268Simp if (T == -1) 56139268Simp SRThreshold = 128; 5750769Sdfr else 5850769Sdfr SRThreshold = T; 5962947Stanimura } 6050769Sdfr 6150769Sdfr bool runOnFunction(Function &F); 6250769Sdfr 6350769Sdfr bool performScalarRepl(Function &F); 6450769Sdfr bool performPromotion(Function &F); 6550769Sdfr 6650769Sdfr // getAnalysisUsage - This pass does not require any passes, but we know it 6750769Sdfr // will not alter the CFG, so say so. 6850769Sdfr virtual void getAnalysisUsage(AnalysisUsage &AU) const { 6950769Sdfr AU.addRequired<DominatorTree>(); 7050769Sdfr AU.addRequired<DominanceFrontier>(); 7150769Sdfr AU.setPreservesCFG(); 7262947Stanimura } 7362947Stanimura 7462947Stanimura private: 7562947Stanimura TargetData *TD; 7662947Stanimura 7762947Stanimura /// DeadInsts - Keep track of instructions we have made dead, so that 7862947Stanimura /// we can remove them after we are done working. 7962947Stanimura SmallVector<Value*, 32> DeadInsts; 8062947Stanimura 8162947Stanimura /// AllocaInfo - When analyzing uses of an alloca instruction, this captures 8262947Stanimura /// information about the uses. All these fields are initialized to false 8362947Stanimura /// and set to true when something is learned. 8462947Stanimura struct AllocaInfo { 8562947Stanimura /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 8662947Stanimura bool isUnsafe : 1; 8762947Stanimura 8890234Stanimura /// needsCleanup - This is set to true if there is some use of the alloca 8990234Stanimura /// that requires cleanup. 9062947Stanimura bool needsCleanup : 1; 9162947Stanimura 9250769Sdfr /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 9350769Sdfr bool isMemCpySrc : 1; 9450769Sdfr 9550769Sdfr /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 96104142Snyan bool isMemCpyDst : 1; 97104142Snyan 98104142Snyan AllocaInfo() 99104142Snyan : isUnsafe(false), needsCleanup(false), 100104142Snyan isMemCpySrc(false), isMemCpyDst(false) {} 101104142Snyan }; 102104142Snyan 103104142Snyan unsigned SRThreshold; 104104142Snyan 105104142Snyan void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; } 106104142Snyan 10750769Sdfr int isSafeAllocaToScalarRepl(AllocaInst *AI); 10850769Sdfr 10950769Sdfr void isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 11050769Sdfr AllocaInfo &Info); 11150769Sdfr void isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t &Offset, 11250769Sdfr AllocaInfo &Info); 11350769Sdfr void isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize, 11452241Sdfr const Type *MemOpType, bool isStore, AllocaInfo &Info); 115139268Simp bool TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size); 11652241Sdfr uint64_t FindElementAndOffset(const Type *&T, uint64_t &Offset, 117139268Simp const Type *&IdxTy); 11852241Sdfr 11952241Sdfr void DoScalarReplacement(AllocaInst *AI, 12052241Sdfr std::vector<AllocaInst*> &WorkList); 12152241Sdfr void DeleteDeadInstructions(); 12252241Sdfr void CleanupAllocaUsers(Value *V); 12352241Sdfr AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocaInst *Base); 12452241Sdfr 12552241Sdfr void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 12652241Sdfr SmallVector<AllocaInst*, 32> &NewElts); 12752241Sdfr void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 12852241Sdfr SmallVector<AllocaInst*, 32> &NewElts); 12952241Sdfr void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 13052241Sdfr SmallVector<AllocaInst*, 32> &NewElts); 13152241Sdfr void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 13250769Sdfr AllocaInst *AI, 13350769Sdfr SmallVector<AllocaInst*, 32> &NewElts); 13450769Sdfr void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 13550769Sdfr SmallVector<AllocaInst*, 32> &NewElts); 13650769Sdfr void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 13750769Sdfr SmallVector<AllocaInst*, 32> &NewElts); 13850769Sdfr 13950769Sdfr bool CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy, 14050769Sdfr bool &SawVec, uint64_t Offset, unsigned AllocaSize); 14150769Sdfr void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); 14250769Sdfr Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType, 14350769Sdfr uint64_t Offset, IRBuilder<> &Builder); 14450769Sdfr Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 14550769Sdfr uint64_t Offset, IRBuilder<> &Builder); 14650769Sdfr static Instruction *isOnlyCopiedFromConstantGlobal(AllocaInst *AI); 14750769Sdfr }; 14850769Sdfr} 14950769Sdfr 15050769Sdfrchar SROA::ID = 0; 15150769Sdfrstatic RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates"); 15250769Sdfr 15350769Sdfr// Public interface to the ScalarReplAggregates pass 15450769SdfrFunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) { 15550769Sdfr return new SROA(Threshold); 15650769Sdfr} 15750769Sdfr 15850769Sdfr 15950769Sdfrbool SROA::runOnFunction(Function &F) { 16050769Sdfr TD = getAnalysisIfAvailable<TargetData>(); 16150769Sdfr 16250769Sdfr bool Changed = performPromotion(F); 16350769Sdfr 16450769Sdfr // FIXME: ScalarRepl currently depends on TargetData more than it 16550769Sdfr // theoretically needs to. It should be refactored in order to support 16650769Sdfr // target-independent IR. Until this is done, just skip the actual 16750769Sdfr // scalar-replacement portion of this pass. 16850769Sdfr if (!TD) return Changed; 16950769Sdfr 17050769Sdfr while (1) { 17150769Sdfr bool LocalChange = performScalarRepl(F); 17250769Sdfr if (!LocalChange) break; // No need to repromote if no scalarrepl 17350769Sdfr Changed = true; 17450769Sdfr LocalChange = performPromotion(F); 17550769Sdfr if (!LocalChange) break; // No need to re-scalarrepl if no promotion 17650769Sdfr } 17750769Sdfr 17850769Sdfr return Changed; 17950769Sdfr} 18050769Sdfr 18150769Sdfr 18250769Sdfrbool SROA::performPromotion(Function &F) { 18350769Sdfr std::vector<AllocaInst*> Allocas; 184139268Simp DominatorTree &DT = getAnalysis<DominatorTree>(); 18550769Sdfr DominanceFrontier &DF = getAnalysis<DominanceFrontier>(); 18650769Sdfr 18750769Sdfr BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 18850769Sdfr 18950769Sdfr bool Changed = false; 190139268Simp 19150769Sdfr while (1) { 19250769Sdfr Allocas.clear(); 19350769Sdfr 19450769Sdfr // Find allocas that are safe to promote, by looking at all instructions in 19552059Sdfr // the entry node 19650769Sdfr for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 19750769Sdfr if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 19850769Sdfr if (isAllocaPromotable(AI)) 19950769Sdfr Allocas.push_back(AI); 20052059Sdfr 20150769Sdfr if (Allocas.empty()) break; 20250769Sdfr 20352059Sdfr PromoteMemToReg(Allocas, DT, DF); 20450769Sdfr NumPromoted += Allocas.size(); 20550769Sdfr Changed = true; 20650769Sdfr } 20750769Sdfr 20850769Sdfr return Changed; 20950769Sdfr} 21050769Sdfr 21150769Sdfr/// getNumSAElements - Return the number of elements in the specific struct or 21250769Sdfr/// array. 213139268Simpstatic uint64_t getNumSAElements(const Type *T) { 21450769Sdfr if (const StructType *ST = dyn_cast<StructType>(T)) 21550769Sdfr return ST->getNumElements(); 21650769Sdfr return cast<ArrayType>(T)->getNumElements(); 21750769Sdfr} 21850769Sdfr 21952059Sdfr// performScalarRepl - This algorithm is a simple worklist driven algorithm, 22050769Sdfr// which runs on all of the malloc/alloca instructions in the function, removing 221139268Simp// them if they are only used by getelementptr instructions. 22250769Sdfr// 22350769Sdfrbool SROA::performScalarRepl(Function &F) { 22450769Sdfr std::vector<AllocaInst*> WorkList; 22550769Sdfr 22650769Sdfr // Scan the entry basic block, adding any alloca's and mallocs to the worklist 22750769Sdfr BasicBlock &BB = F.getEntryBlock(); 22850769Sdfr for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 22950769Sdfr if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 23050769Sdfr WorkList.push_back(A); 23150769Sdfr 23250769Sdfr // Process the worklist 23350769Sdfr bool Changed = false; 23450769Sdfr while (!WorkList.empty()) { 23550769Sdfr AllocaInst *AI = WorkList.back(); 23650769Sdfr WorkList.pop_back(); 23750769Sdfr 23850769Sdfr // Handle dead allocas trivially. These can be formed by SROA'ing arrays 23950769Sdfr // with unused elements. 24050769Sdfr if (AI->use_empty()) { 24150769Sdfr AI->eraseFromParent(); 24250769Sdfr continue; 24350769Sdfr } 24450769Sdfr 24550769Sdfr // If this alloca is impossible for us to promote, reject it early. 24650769Sdfr if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 24750769Sdfr continue; 24866840Smsmith 24966840Smsmith // Check to see if this allocation is only modified by a memcpy/memmove from 25066840Smsmith // a constant global. If this is the case, we can change all users to use 251139268Simp // the constant global instead. This is commonly produced by the CFE by 252139268Simp // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 253139268Simp // is only subsequently read. 25466840Smsmith if (Instruction *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 25566840Smsmith DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); 256139268Simp DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n'); 257139268Simp Constant *TheSrc = cast<Constant>(TheCopy->getOperand(2)); 258139268Simp AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 25966840Smsmith TheCopy->eraseFromParent(); // Don't mutate the global. 26066840Smsmith AI->eraseFromParent(); 261139268Simp ++NumGlobals; 262139268Simp Changed = true; 263139268Simp continue; 26466840Smsmith } 26566840Smsmith 266139268Simp // Check to see if we can perform the core SROA transformation. We cannot 267139268Simp // transform the allocation instruction if it is an array allocation 268139268Simp // (allocations OF arrays are ok though), and an allocation of a scalar 26966840Smsmith // value cannot be decomposed at all. 27066840Smsmith uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 27166840Smsmith 27250769Sdfr // Do not promote [0 x %struct]. 27350769Sdfr if (AllocaSize == 0) continue; 27450769Sdfr 275139268Simp // Do not promote any struct whose size is too big. 276139268Simp if (AllocaSize > SRThreshold) continue; 27783051Syokota 27883051Syokota if ((isa<StructType>(AI->getAllocatedType()) || 27983051Syokota isa<ArrayType>(AI->getAllocatedType())) && 28083051Syokota // Do not promote any struct into more than "32" separate vars. 28183051Syokota getNumSAElements(AI->getAllocatedType()) <= SRThreshold/4) { 28283051Syokota // Check that all of the users of the allocation are capable of being 28383051Syokota // transformed. 28483051Syokota switch (isSafeAllocaToScalarRepl(AI)) { 28583051Syokota default: llvm_unreachable("Unexpected value!"); 28683051Syokota case 0: // Not safe to scalar replace. 28783051Syokota break; 28883051Syokota case 1: // Safe, but requires cleanup/canonicalizations first 28983051Syokota CleanupAllocaUsers(AI); 29083051Syokota // FALL THROUGH. 29183051Syokota case 3: // Safe to scalar replace. 29283051Syokota DoScalarReplacement(AI, WorkList); 29383051Syokota Changed = true; 29450769Sdfr continue; 29566840Smsmith } 29650769Sdfr } 29750769Sdfr 29850769Sdfr // If we can turn this aggregate value (potentially with casts) into a 29950769Sdfr // simple scalar value that can be mem2reg'd into a register value. 30050769Sdfr // IsNotTrivial tracks whether this is something that mem2reg could have 30150769Sdfr // promoted itself. If so, we don't want to transform it needlessly. Note 30250769Sdfr // that we can't just check based on the type: the alloca may be of an i32 303139268Simp // but that has pointer arithmetic to set byte 3 of it or something. 30483051Syokota bool IsNotTrivial = false; 30583051Syokota const Type *VectorTy = 0; 30683051Syokota bool HadAVector = false; 30783051Syokota if (CanConvertToScalar(AI, IsNotTrivial, VectorTy, HadAVector, 30883051Syokota 0, unsigned(AllocaSize)) && IsNotTrivial) { 30983051Syokota AllocaInst *NewAI; 31083051Syokota // If we were able to find a vector type that can handle this with 31183051Syokota // insert/extract elements, and if there was at least one use that had 31283051Syokota // a vector type, promote this to a vector. We don't want to promote 31350769Sdfr // random stuff that doesn't use vectors (e.g. <9 x double>) because then 31466840Smsmith // we just get a lot of insert/extracts. If at least one vector is 31550769Sdfr // involved, then we probably really do have a union of vector/array. 31650769Sdfr if (VectorTy && isa<VectorType>(VectorTy) && HadAVector) { 31750769Sdfr DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = " 31850769Sdfr << *VectorTy << '\n'); 31950769Sdfr 32083051Syokota // Create and insert the vector alloca. 32183051Syokota NewAI = new AllocaInst(VectorTy, 0, "", AI->getParent()->begin()); 32283051Syokota ConvertUsesToScalar(AI, NewAI, 0); 32383051Syokota } else { 32483051Syokota DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"); 32583051Syokota 32683051Syokota // Create and insert the integer alloca. 32783051Syokota const Type *NewTy = IntegerType::get(AI->getContext(), AllocaSize*8); 32883051Syokota NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 32983051Syokota ConvertUsesToScalar(AI, NewAI, 0); 33083051Syokota } 33150769Sdfr NewAI->takeName(AI); 33266840Smsmith AI->eraseFromParent(); 33350769Sdfr ++NumConverted; 33450769Sdfr Changed = true; 33550769Sdfr continue; 33650769Sdfr } 33750769Sdfr 33883051Syokota // Otherwise, couldn't process this alloca. 33950769Sdfr } 34050769Sdfr 34150769Sdfr return Changed; 34283051Syokota} 34383051Syokota 34483051Syokota/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 34583051Syokota/// predicate, do SROA now. 34683051Syokotavoid SROA::DoScalarReplacement(AllocaInst *AI, 34783051Syokota std::vector<AllocaInst*> &WorkList) { 34883051Syokota DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 34983051Syokota SmallVector<AllocaInst*, 32> ElementAllocas; 35050769Sdfr if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 35166840Smsmith ElementAllocas.reserve(ST->getNumContainedTypes()); 35250769Sdfr for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 35350769Sdfr AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 35450769Sdfr AI->getAlignment(), 35550769Sdfr AI->getName() + "." + Twine(i), AI); 35650769Sdfr ElementAllocas.push_back(NA); 35750769Sdfr WorkList.push_back(NA); // Add to worklist for recursive processing 35850769Sdfr } 35950769Sdfr } else { 36050769Sdfr const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 36150769Sdfr ElementAllocas.reserve(AT->getNumElements()); 36250769Sdfr const Type *ElTy = AT->getElementType(); 36350769Sdfr for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 36450769Sdfr AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 36550769Sdfr AI->getName() + "." + Twine(i), AI); 36650769Sdfr ElementAllocas.push_back(NA); 36750769Sdfr WorkList.push_back(NA); // Add to worklist for recursive processing 36850769Sdfr } 36950769Sdfr } 37062947Stanimura 371139268Simp // Now that we have created the new alloca instructions, rewrite all the 372139268Simp // uses of the old alloca. 37350769Sdfr RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 37450769Sdfr 37550769Sdfr // Now erase any instructions that were made dead while rewriting the alloca. 37650769Sdfr DeleteDeadInstructions(); 37750769Sdfr AI->eraseFromParent(); 378139268Simp 37950769Sdfr NumReplaced++; 38050769Sdfr} 38150769Sdfr 38250769Sdfr/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 38350769Sdfr/// recursively including all their operands that become trivially dead. 38462947Stanimuravoid SROA::DeleteDeadInstructions() { 38562947Stanimura while (!DeadInsts.empty()) { 38662947Stanimura Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 38762947Stanimura 38862947Stanimura for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 38962947Stanimura if (Instruction *U = dyn_cast<Instruction>(*OI)) { 39062947Stanimura // Zero out the operand and see if it becomes trivially dead. 39162947Stanimura // (But, don't add allocas to the dead instruction list -- they are 39262947Stanimura // already on the worklist and will be deleted separately.) 39362947Stanimura *OI = 0; 39462947Stanimura if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 39562947Stanimura DeadInsts.push_back(U); 39662947Stanimura } 39762947Stanimura 39862947Stanimura I->eraseFromParent(); 39962947Stanimura } 40050769Sdfr} 40150769Sdfr 40250769Sdfr/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 40350769Sdfr/// performing scalar replacement of alloca AI. The results are flagged in 40450769Sdfr/// the Info parameter. Offset indicates the position within AI that is 40550769Sdfr/// referenced by this instruction. 40650769Sdfrvoid SROA::isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 40750769Sdfr AllocaInfo &Info) { 40850769Sdfr for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 40952059Sdfr Instruction *User = cast<Instruction>(*UI); 41050769Sdfr 41150769Sdfr if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 41250769Sdfr isSafeForScalarRepl(BC, AI, Offset, Info); 41352059Sdfr } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 414139268Simp uint64_t GEPOffset = Offset; 41550769Sdfr isSafeGEP(GEPI, AI, GEPOffset, Info); 41652059Sdfr if (!Info.isUnsafe) 41752059Sdfr isSafeForScalarRepl(GEPI, AI, GEPOffset, Info); 418139268Simp } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) { 41950769Sdfr ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 42050769Sdfr if (Length) 42150769Sdfr isSafeMemAccess(AI, Offset, Length->getZExtValue(), 0, 42252059Sdfr UI.getOperandNo() == 1, Info); 42350769Sdfr else 42450769Sdfr MarkUnsafe(Info); 42552059Sdfr } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 42652059Sdfr if (!LI->isVolatile()) { 42752059Sdfr const Type *LIType = LI->getType(); 42852059Sdfr isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(LIType), 42952059Sdfr LIType, false, Info); 43052059Sdfr } else 43152059Sdfr MarkUnsafe(Info); 43250769Sdfr } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 43350769Sdfr // Store is ok if storing INTO the pointer, not storing the pointer 43450769Sdfr if (!SI->isVolatile() && SI->getOperand(0) != I) { 43552059Sdfr const Type *SIType = SI->getOperand(0)->getType(); 43652059Sdfr isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(SIType), 43750769Sdfr SIType, true, Info); 43852059Sdfr } else 43950769Sdfr MarkUnsafe(Info); 44050769Sdfr } else if (isa<DbgInfoIntrinsic>(UI)) { 44150769Sdfr // If one user is DbgInfoIntrinsic then check if all users are 44252059Sdfr // DbgInfoIntrinsics. 44352059Sdfr if (OnlyUsedByDbgInfoIntrinsics(I)) { 44452059Sdfr Info.needsCleanup = true; 44550769Sdfr return; 44650769Sdfr } 44783504Syokota MarkUnsafe(Info); 44883504Syokota } else { 44983504Syokota DEBUG(errs() << " Transformation preventing inst: " << *User << '\n'); 45083504Syokota MarkUnsafe(Info); 45183504Syokota } 45283504Syokota if (Info.isUnsafe) return; 45383504Syokota } 45483504Syokota} 45583504Syokota 45652059Sdfr/// isSafeGEP - Check if a GEP instruction can be handled for scalar 45752059Sdfr/// replacement. It is safe when all the indices are constant, in-bounds 45852059Sdfr/// references, and when the resulting offset corresponds to an element within 45952059Sdfr/// the alloca type. The results are flagged in the Info parameter. Upon 46050769Sdfr/// return, Offset is adjusted as specified by the GEP indices. 46152059Sdfrvoid SROA::isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, 46250769Sdfr uint64_t &Offset, AllocaInfo &Info) { 46352059Sdfr gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 46452059Sdfr if (GEPIt == E) 46552059Sdfr return; 46652059Sdfr 46750769Sdfr // Walk through the GEP type indices, checking the types that this indexes 46850769Sdfr // into. 46950769Sdfr for (; GEPIt != E; ++GEPIt) { 47052059Sdfr // Ignore struct elements, no extra checking needed for these. 47152059Sdfr if (isa<StructType>(*GEPIt)) 47252059Sdfr continue; 47352059Sdfr 47452059Sdfr ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 47552059Sdfr if (!IdxVal) 47652059Sdfr return MarkUnsafe(Info); 47752059Sdfr } 47852059Sdfr 47952059Sdfr // Compute the offset due to this GEP and check if the alloca has a 48052059Sdfr // component element at that offset. 48152059Sdfr SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 48252059Sdfr Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 48352059Sdfr &Indices[0], Indices.size()); 48452059Sdfr if (!TypeHasComponent(AI->getAllocatedType(), Offset, 0)) 48552059Sdfr MarkUnsafe(Info); 48652059Sdfr} 48752059Sdfr 48852059Sdfr/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 48952059Sdfr/// alloca or has an offset and size that corresponds to a component element 490139268Simp/// within it. The offset checked here may have been formed from a GEP with a 49152059Sdfr/// pointer bitcasted to a different type. 49252059Sdfrvoid SROA::isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize, 49350769Sdfr const Type *MemOpType, bool isStore, 49450769Sdfr AllocaInfo &Info) { 49552059Sdfr // Check if this is a load/store of the entire alloca. 49652059Sdfr if (Offset == 0 && MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) { 49752059Sdfr bool UsesAggregateType = (MemOpType == AI->getAllocatedType()); 49852059Sdfr // This is safe for MemIntrinsics (where MemOpType is 0), integer types 49952059Sdfr // (which are essentially the same as the MemIntrinsics, especially with 50062947Stanimura // regard to copying padding between elements), or references using the 50152059Sdfr // aggregate type of the alloca. 50252059Sdfr if (!MemOpType || isa<IntegerType>(MemOpType) || UsesAggregateType) { 50352059Sdfr if (!UsesAggregateType) { 50483504Syokota if (isStore) 50583504Syokota Info.isMemCpyDst = true; 506139268Simp else 50752059Sdfr Info.isMemCpySrc = true; 50852059Sdfr } 50952059Sdfr return; 51083051Syokota } 511139268Simp } 51252059Sdfr // Check if the offset/size correspond to a component within the alloca type. 51352059Sdfr const Type *T = AI->getAllocatedType(); 514139268Simp if (TypeHasComponent(T, Offset, MemSize)) 51550769Sdfr return; 51652059Sdfr 51750769Sdfr return MarkUnsafe(Info); 51852059Sdfr} 51952059Sdfr 520139268Simp/// TypeHasComponent - Return true if T has a component type with the 521139268Simp/// specified offset and size. If Size is zero, do not check the size. 52252059Sdfrbool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) { 52352059Sdfr const Type *EltTy; 52452059Sdfr uint64_t EltSize; 52552059Sdfr if (const StructType *ST = dyn_cast<StructType>(T)) { 52652059Sdfr const StructLayout *Layout = TD->getStructLayout(ST); 52752059Sdfr unsigned EltIdx = Layout->getElementContainingOffset(Offset); 528139268Simp EltTy = ST->getContainedType(EltIdx); 52950769Sdfr EltSize = TD->getTypeAllocSize(EltTy); 53050769Sdfr Offset -= Layout->getElementOffset(EltIdx); 53150769Sdfr } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 53252059Sdfr EltTy = AT->getElementType(); 533139268Simp EltSize = TD->getTypeAllocSize(EltTy); 53452059Sdfr if (Offset >= AT->getNumElements() * EltSize) 53552059Sdfr return false; 53652059Sdfr Offset %= EltSize; 53752059Sdfr } else { 53850769Sdfr return false; 53952059Sdfr } 54052059Sdfr if (Offset == 0 && (Size == 0 || EltSize == Size)) 54152059Sdfr return true; 54250769Sdfr // Check if the component spans multiple elements. 54350769Sdfr if (Offset + Size > EltSize) 54450769Sdfr return false; 545139268Simp return TypeHasComponent(EltTy, Offset, Size); 54650769Sdfr} 54750769Sdfr 54850769Sdfr/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 54952059Sdfr/// the instruction I, which references it, to use the separate elements. 55052059Sdfr/// Offset indicates the position within AI that is referenced by this 55152059Sdfr/// instruction. 55252059Sdfrvoid SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 55352059Sdfr SmallVector<AllocaInst*, 32> &NewElts) { 55452059Sdfr for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 555105226Sphk Instruction *User = cast<Instruction>(*UI); 556105226Sphk 557105226Sphk if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 558105226Sphk RewriteBitCast(BC, AI, Offset, NewElts); 55952059Sdfr } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 56052059Sdfr RewriteGEP(GEPI, AI, Offset, NewElts); 56152059Sdfr } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 56252059Sdfr ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 56352059Sdfr uint64_t MemSize = Length->getZExtValue(); 56452059Sdfr if (Offset == 0 && 56552059Sdfr MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 56652059Sdfr RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 56752059Sdfr // Otherwise the intrinsic can only touch a single element and the 56852059Sdfr // address operand will be updated, so nothing else needs to be done. 56952059Sdfr } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 57052059Sdfr const Type *LIType = LI->getType(); 57152059Sdfr if (LIType == AI->getAllocatedType()) { 572139268Simp // Replace: 57352059Sdfr // %res = load { i32, i32 }* %alloc 57452059Sdfr // with: 57552059Sdfr // %load.0 = load i32* %alloc.0 57652059Sdfr // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 57752059Sdfr // %load.1 = load i32* %alloc.1 57852059Sdfr // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 57952059Sdfr // (Also works for arrays instead of structs) 580115543Sphk Value *Insert = UndefValue::get(LIType); 581115543Sphk for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 582139268Simp Value *Load = new LoadInst(NewElts[i], "load", LI); 583115543Sphk Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 58452059Sdfr } 58552059Sdfr LI->replaceAllUsesWith(Insert); 58652059Sdfr DeadInsts.push_back(LI); 58752059Sdfr } else if (isa<IntegerType>(LIType) && 58852059Sdfr TD->getTypeAllocSize(LIType) == 58952059Sdfr TD->getTypeAllocSize(AI->getAllocatedType())) { 59052059Sdfr // If this is a load of the entire alloca to an integer, rewrite it. 591139268Simp RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 59252059Sdfr } 59352059Sdfr } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 59452059Sdfr Value *Val = SI->getOperand(0); 59552059Sdfr const Type *SIType = Val->getType(); 59652059Sdfr if (SIType == AI->getAllocatedType()) { 59752059Sdfr // Replace: 598139268Simp // store { i32, i32 } %val, { i32, i32 }* %alloc 59952059Sdfr // with: 60052059Sdfr // %val.0 = extractvalue { i32, i32 } %val, 0 60152059Sdfr // store i32 %val.0, i32* %alloc.0 60252059Sdfr // %val.1 = extractvalue { i32, i32 } %val, 1 60352059Sdfr // store i32 %val.1, i32* %alloc.1 60452059Sdfr // (Also works for arrays instead of structs) 60552059Sdfr for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 60652059Sdfr Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 60752059Sdfr new StoreInst(Extract, NewElts[i], SI); 60852059Sdfr } 60952059Sdfr DeadInsts.push_back(SI); 61052059Sdfr } else if (isa<IntegerType>(SIType) && 61152059Sdfr TD->getTypeAllocSize(SIType) == 61252059Sdfr TD->getTypeAllocSize(AI->getAllocatedType())) { 61352059Sdfr // If this is a store of the entire alloca from an integer, rewrite it. 61452059Sdfr RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 61552059Sdfr } 61652059Sdfr } 61752059Sdfr } 61852059Sdfr} 61952059Sdfr 62052059Sdfr/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 62152059Sdfr/// and recursively continue updating all of its uses. 62252059Sdfrvoid SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 62352059Sdfr SmallVector<AllocaInst*, 32> &NewElts) { 62452059Sdfr RewriteForScalarRepl(BC, AI, Offset, NewElts); 62552059Sdfr if (BC->getOperand(0) != AI) 62652059Sdfr return; 62752059Sdfr 62852059Sdfr // The bitcast references the original alloca. Replace its uses with 629139268Simp // references to the first new element alloca. 63052059Sdfr Instruction *Val = NewElts[0]; 63152059Sdfr if (Val->getType() != BC->getDestTy()) { 63252059Sdfr Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 63352059Sdfr Val->takeName(BC); 63452059Sdfr } 63552059Sdfr BC->replaceAllUsesWith(Val); 63652059Sdfr DeadInsts.push_back(BC); 63752059Sdfr} 63852059Sdfr 63952059Sdfr/// FindElementAndOffset - Return the index of the element containing Offset 64052059Sdfr/// within the specified type, which must be either a struct or an array. 64152059Sdfr/// Sets T to the type of the element and Offset to the offset within that 642139268Simp/// element. IdxTy is set to the type of the index result to be used in a 643139268Simp/// GEP instruction. 64452059Sdfruint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset, 64552059Sdfr const Type *&IdxTy) { 64652059Sdfr uint64_t Idx = 0; 64752059Sdfr if (const StructType *ST = dyn_cast<StructType>(T)) { 64852059Sdfr const StructLayout *Layout = TD->getStructLayout(ST); 64952059Sdfr Idx = Layout->getElementContainingOffset(Offset); 65052059Sdfr T = ST->getContainedType(Idx); 65152059Sdfr Offset -= Layout->getElementOffset(Idx); 65252059Sdfr IdxTy = Type::getInt32Ty(T->getContext()); 653139268Simp return Idx; 65452059Sdfr } 65552059Sdfr const ArrayType *AT = cast<ArrayType>(T); 65652059Sdfr T = AT->getElementType(); 65750769Sdfr uint64_t EltSize = TD->getTypeAllocSize(T); 65850769Sdfr Idx = Offset / EltSize; 65950769Sdfr Offset -= Idx * EltSize; 66050769Sdfr IdxTy = Type::getInt64Ty(T->getContext()); 66150769Sdfr return Idx; 66250769Sdfr} 66350769Sdfr 66450769Sdfr/// RewriteGEP - Check if this GEP instruction moves the pointer across 66550769Sdfr/// elements of the alloca that are being split apart, and if so, rewrite 66650769Sdfr/// the GEP to be relative to the new element. 66750769Sdfrvoid SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 66850769Sdfr SmallVector<AllocaInst*, 32> &NewElts) { 66950769Sdfr uint64_t OldOffset = Offset; 67050769Sdfr SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 67152059Sdfr Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 67252059Sdfr &Indices[0], Indices.size()); 67352059Sdfr 67452059Sdfr RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 675104142Snyan 676104142Snyan const Type *T = AI->getAllocatedType(); 677104142Snyan const Type *IdxTy; 678104142Snyan uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 67950769Sdfr if (GEPI->getOperand(0) == AI) 68050769Sdfr OldIdx = ~0ULL; // Force the GEP to be rewritten. 68150769Sdfr 68250769Sdfr T = AI->getAllocatedType(); 68350769Sdfr uint64_t EltOffset = Offset; 68450769Sdfr uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 68550769Sdfr 68650769Sdfr // If this GEP does not move the pointer across elements of the alloca 68750769Sdfr // being split, then it does not needs to be rewritten. 68850769Sdfr if (Idx == OldIdx) 68950769Sdfr return; 69050769Sdfr 69150769Sdfr const Type *i32Ty = Type::getInt32Ty(AI->getContext()); 69250769Sdfr SmallVector<Value*, 8> NewArgs; 69350769Sdfr NewArgs.push_back(Constant::getNullValue(i32Ty)); 69450769Sdfr while (EltOffset != 0) { 69550769Sdfr uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 69650769Sdfr NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 69750769Sdfr } 69850769Sdfr Instruction *Val = NewElts[Idx]; 69950769Sdfr if (NewArgs.size() > 1) { 70050769Sdfr Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(), 70150769Sdfr NewArgs.end(), "", GEPI); 70250769Sdfr Val->takeName(GEPI); 70350769Sdfr } 70450769Sdfr if (Val->getType() != GEPI->getType()) 70550769Sdfr Val = new BitCastInst(Val, GEPI->getType(), Val->getNameStr(), GEPI); 70650769Sdfr GEPI->replaceAllUsesWith(Val); 70750769Sdfr DeadInsts.push_back(GEPI); 70850769Sdfr} 70950769Sdfr 71050769Sdfr/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 71150769Sdfr/// Rewrite it to copy or set the elements of the scalarized memory. 71250769Sdfrvoid SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 71352059Sdfr AllocaInst *AI, 71450769Sdfr SmallVector<AllocaInst*, 32> &NewElts) { 71550769Sdfr // If this is a memcpy/memmove, construct the other pointer as the 71650769Sdfr // appropriate type. The "Other" pointer is the pointer that goes to memory 71750769Sdfr // that doesn't have anything to do with the alloca that we are promoting. For 71850769Sdfr // memset, this Value* stays null. 71950769Sdfr Value *OtherPtr = 0; 720104142Snyan LLVMContext &Context = MI->getContext(); 721104142Snyan unsigned MemAlignment = MI->getAlignment(); 722104142Snyan if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 723104142Snyan if (Inst == MTI->getRawDest()) 724104142Snyan OtherPtr = MTI->getRawSource(); 725104142Snyan else { 726104142Snyan assert(Inst == MTI->getRawSource()); 727104142Snyan OtherPtr = MTI->getRawDest(); 728104142Snyan } 729104142Snyan } 730104142Snyan 731104142Snyan // If there is an other pointer, we want to convert it to the same pointer 732139268Simp // type as AI has, so we can GEP through it safely. 733139268Simp if (OtherPtr) { 734104142Snyan 735104142Snyan // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 736104142Snyan // optimization, but it's also required to detect the corner case where 737104142Snyan // both pointer operands are referencing the same memory, and where 738139268Simp // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 739139268Simp // function is only called for mem intrinsics that access the whole 740104142Snyan // aggregate, so non-zero GEPs are not an issue here.) 741104142Snyan while (1) { 742104142Snyan if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr)) { 743139268Simp OtherPtr = BC->getOperand(0); 744104142Snyan continue; 745139268Simp } 74652059Sdfr if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr)) { 74752059Sdfr // All zero GEPs are effectively bitcasts. 748139268Simp if (GEP->hasAllZeroIndices()) { 74950769Sdfr OtherPtr = GEP->getOperand(0); 75050769Sdfr continue; 75150769Sdfr } 75250769Sdfr } 75350769Sdfr break; 75450769Sdfr } 75550769Sdfr // If OtherPtr has already been rewritten, this intrinsic will be dead. 75650769Sdfr if (OtherPtr == NewElts[0]) 75750769Sdfr return; 75850769Sdfr 75950769Sdfr if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr)) 76050769Sdfr if (BCE->getOpcode() == Instruction::BitCast) 76150769Sdfr OtherPtr = BCE->getOperand(0); 76250769Sdfr 76350769Sdfr // If the pointer is not the right type, insert a bitcast to the right 76450769Sdfr // type. 76550769Sdfr if (OtherPtr->getType() != AI->getType()) 76650769Sdfr OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(), 76750769Sdfr MI); 76852059Sdfr } 76952059Sdfr 77052059Sdfr // Process each element of the aggregate. 77152059Sdfr Value *TheFn = MI->getOperand(0); 77252059Sdfr const Type *BytePtrTy = MI->getRawDest()->getType(); 77352059Sdfr bool SROADest = MI->getRawDest() == Inst; 774139268Simp 77550769Sdfr Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 77650769Sdfr 77750769Sdfr for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 77850769Sdfr // If this is a memcpy/memmove, emit a GEP of the other element address. 77950769Sdfr Value *OtherElt = 0; 78050769Sdfr unsigned OtherEltAlign = MemAlignment; 78150769Sdfr 78250769Sdfr if (OtherPtr == AI) { 78350769Sdfr OtherElt = NewElts[i]; 78450769Sdfr OtherEltAlign = 0; 78550769Sdfr } else if (OtherPtr) { 78650769Sdfr Value *Idx[2] = { Zero, 78750769Sdfr ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 78850769Sdfr OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2, 78950769Sdfr OtherPtr->getNameStr()+"."+Twine(i), 79050769Sdfr MI); 79150769Sdfr uint64_t EltOffset; 79250769Sdfr const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 79350769Sdfr if (const StructType *ST = 79450769Sdfr dyn_cast<StructType>(OtherPtrTy->getElementType())) { 79550769Sdfr EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 796139271Simp } else { 797139268Simp const Type *EltTy = 79850769Sdfr cast<SequentialType>(OtherPtr->getType())->getElementType(); 79950769Sdfr EltOffset = TD->getTypeAllocSize(EltTy)*i; 80050769Sdfr } 80150769Sdfr 80250769Sdfr // The alignment of the other pointer is the guaranteed alignment of the 803139269Simp // element, which is affected by both the known alignment of the whole 804139269Simp // mem intrinsic and the alignment of the element. If the alignment of 80550769Sdfr // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 80650769Sdfr // known alignment is just 4 bytes. 80750769Sdfr OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 80850769Sdfr } 80950769Sdfr 81050769Sdfr Value *EltPtr = NewElts[i]; 81150769Sdfr const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 81250769Sdfr 81350769Sdfr // If we got down to a scalar, insert a load or store as appropriate. 81450769Sdfr if (EltTy->isSingleValueType()) { 81550769Sdfr if (isa<MemTransferInst>(MI)) { 81650769Sdfr if (SROADest) { 81750769Sdfr // From Other to Alloca. 81850769Sdfr Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 81950769Sdfr new StoreInst(Elt, EltPtr, MI); 82050769Sdfr } else { 82150769Sdfr // From Alloca to Other. 82250769Sdfr Value *Elt = new LoadInst(EltPtr, "tmp", MI); 823 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 824 } 825 continue; 826 } 827 assert(isa<MemSetInst>(MI)); 828 829 // If the stored element is zero (common case), just store a null 830 // constant. 831 Constant *StoreVal; 832 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) { 833 if (CI->isZero()) { 834 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 835 } else { 836 // If EltTy is a vector type, get the element type. 837 const Type *ValTy = EltTy->getScalarType(); 838 839 // Construct an integer with the right value. 840 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 841 APInt OneVal(EltSize, CI->getZExtValue()); 842 APInt TotalVal(OneVal); 843 // Set each byte. 844 for (unsigned i = 0; 8*i < EltSize; ++i) { 845 TotalVal = TotalVal.shl(8); 846 TotalVal |= OneVal; 847 } 848 849 // Convert the integer value to the appropriate type. 850 StoreVal = ConstantInt::get(Context, TotalVal); 851 if (isa<PointerType>(ValTy)) 852 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 853 else if (ValTy->isFloatingPoint()) 854 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 855 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 856 857 // If the requested value was a vector constant, create it. 858 if (EltTy != ValTy) { 859 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 860 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 861 StoreVal = ConstantVector::get(&Elts[0], NumElts); 862 } 863 } 864 new StoreInst(StoreVal, EltPtr, MI); 865 continue; 866 } 867 // Otherwise, if we're storing a byte variable, use a memset call for 868 // this element. 869 } 870 871 // Cast the element pointer to BytePtrTy. 872 if (EltPtr->getType() != BytePtrTy) 873 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI); 874 875 // Cast the other pointer (if we have one) to BytePtrTy. 876 if (OtherElt && OtherElt->getType() != BytePtrTy) 877 OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(), 878 MI); 879 880 unsigned EltSize = TD->getTypeAllocSize(EltTy); 881 882 // Finally, insert the meminst for this element. 883 if (isa<MemTransferInst>(MI)) { 884 Value *Ops[] = { 885 SROADest ? EltPtr : OtherElt, // Dest ptr 886 SROADest ? OtherElt : EltPtr, // Src ptr 887 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 888 // Align 889 ConstantInt::get(Type::getInt32Ty(MI->getContext()), OtherEltAlign) 890 }; 891 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 892 } else { 893 assert(isa<MemSetInst>(MI)); 894 Value *Ops[] = { 895 EltPtr, MI->getOperand(2), // Dest, Value, 896 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 897 Zero // Align 898 }; 899 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 900 } 901 } 902 DeadInsts.push_back(MI); 903} 904 905/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 906/// overwrites the entire allocation. Extract out the pieces of the stored 907/// integer and store them individually. 908void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 909 SmallVector<AllocaInst*, 32> &NewElts){ 910 // Extract each element out of the integer according to its structure offset 911 // and store the element value to the individual alloca. 912 Value *SrcVal = SI->getOperand(0); 913 const Type *AllocaEltTy = AI->getAllocatedType(); 914 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 915 916 // Handle tail padding by extending the operand 917 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 918 SrcVal = new ZExtInst(SrcVal, 919 IntegerType::get(SI->getContext(), AllocaSizeBits), 920 "", SI); 921 922 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 923 << '\n'); 924 925 // There are two forms here: AI could be an array or struct. Both cases 926 // have different ways to compute the element offset. 927 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 928 const StructLayout *Layout = TD->getStructLayout(EltSTy); 929 930 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 931 // Get the number of bits to shift SrcVal to get the value. 932 const Type *FieldTy = EltSTy->getElementType(i); 933 uint64_t Shift = Layout->getElementOffsetInBits(i); 934 935 if (TD->isBigEndian()) 936 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 937 938 Value *EltVal = SrcVal; 939 if (Shift) { 940 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 941 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 942 "sroa.store.elt", SI); 943 } 944 945 // Truncate down to an integer of the right size. 946 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 947 948 // Ignore zero sized fields like {}, they obviously contain no data. 949 if (FieldSizeBits == 0) continue; 950 951 if (FieldSizeBits != AllocaSizeBits) 952 EltVal = new TruncInst(EltVal, 953 IntegerType::get(SI->getContext(), FieldSizeBits), 954 "", SI); 955 Value *DestField = NewElts[i]; 956 if (EltVal->getType() == FieldTy) { 957 // Storing to an integer field of this size, just do it. 958 } else if (FieldTy->isFloatingPoint() || isa<VectorType>(FieldTy)) { 959 // Bitcast to the right element type (for fp/vector values). 960 EltVal = new BitCastInst(EltVal, FieldTy, "", SI); 961 } else { 962 // Otherwise, bitcast the dest pointer (for aggregates). 963 DestField = new BitCastInst(DestField, 964 PointerType::getUnqual(EltVal->getType()), 965 "", SI); 966 } 967 new StoreInst(EltVal, DestField, SI); 968 } 969 970 } else { 971 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 972 const Type *ArrayEltTy = ATy->getElementType(); 973 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 974 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 975 976 uint64_t Shift; 977 978 if (TD->isBigEndian()) 979 Shift = AllocaSizeBits-ElementOffset; 980 else 981 Shift = 0; 982 983 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 984 // Ignore zero sized fields like {}, they obviously contain no data. 985 if (ElementSizeBits == 0) continue; 986 987 Value *EltVal = SrcVal; 988 if (Shift) { 989 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 990 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 991 "sroa.store.elt", SI); 992 } 993 994 // Truncate down to an integer of the right size. 995 if (ElementSizeBits != AllocaSizeBits) 996 EltVal = new TruncInst(EltVal, 997 IntegerType::get(SI->getContext(), 998 ElementSizeBits),"",SI); 999 Value *DestField = NewElts[i]; 1000 if (EltVal->getType() == ArrayEltTy) { 1001 // Storing to an integer field of this size, just do it. 1002 } else if (ArrayEltTy->isFloatingPoint() || isa<VectorType>(ArrayEltTy)) { 1003 // Bitcast to the right element type (for fp/vector values). 1004 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI); 1005 } else { 1006 // Otherwise, bitcast the dest pointer (for aggregates). 1007 DestField = new BitCastInst(DestField, 1008 PointerType::getUnqual(EltVal->getType()), 1009 "", SI); 1010 } 1011 new StoreInst(EltVal, DestField, SI); 1012 1013 if (TD->isBigEndian()) 1014 Shift -= ElementOffset; 1015 else 1016 Shift += ElementOffset; 1017 } 1018 } 1019 1020 DeadInsts.push_back(SI); 1021} 1022 1023/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 1024/// an integer. Load the individual pieces to form the aggregate value. 1025void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 1026 SmallVector<AllocaInst*, 32> &NewElts) { 1027 // Extract each element out of the NewElts according to its structure offset 1028 // and form the result value. 1029 const Type *AllocaEltTy = AI->getAllocatedType(); 1030 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1031 1032 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 1033 << '\n'); 1034 1035 // There are two forms here: AI could be an array or struct. Both cases 1036 // have different ways to compute the element offset. 1037 const StructLayout *Layout = 0; 1038 uint64_t ArrayEltBitOffset = 0; 1039 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1040 Layout = TD->getStructLayout(EltSTy); 1041 } else { 1042 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 1043 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1044 } 1045 1046 Value *ResultVal = 1047 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 1048 1049 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1050 // Load the value from the alloca. If the NewElt is an aggregate, cast 1051 // the pointer to an integer of the same size before doing the load. 1052 Value *SrcField = NewElts[i]; 1053 const Type *FieldTy = 1054 cast<PointerType>(SrcField->getType())->getElementType(); 1055 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1056 1057 // Ignore zero sized fields like {}, they obviously contain no data. 1058 if (FieldSizeBits == 0) continue; 1059 1060 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 1061 FieldSizeBits); 1062 if (!isa<IntegerType>(FieldTy) && !FieldTy->isFloatingPoint() && 1063 !isa<VectorType>(FieldTy)) 1064 SrcField = new BitCastInst(SrcField, 1065 PointerType::getUnqual(FieldIntTy), 1066 "", LI); 1067 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 1068 1069 // If SrcField is a fp or vector of the right size but that isn't an 1070 // integer type, bitcast to an integer so we can shift it. 1071 if (SrcField->getType() != FieldIntTy) 1072 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 1073 1074 // Zero extend the field to be the same size as the final alloca so that 1075 // we can shift and insert it. 1076 if (SrcField->getType() != ResultVal->getType()) 1077 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 1078 1079 // Determine the number of bits to shift SrcField. 1080 uint64_t Shift; 1081 if (Layout) // Struct case. 1082 Shift = Layout->getElementOffsetInBits(i); 1083 else // Array case. 1084 Shift = i*ArrayEltBitOffset; 1085 1086 if (TD->isBigEndian()) 1087 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 1088 1089 if (Shift) { 1090 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 1091 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 1092 } 1093 1094 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 1095 } 1096 1097 // Handle tail padding by truncating the result 1098 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 1099 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 1100 1101 LI->replaceAllUsesWith(ResultVal); 1102 DeadInsts.push_back(LI); 1103} 1104 1105/// HasPadding - Return true if the specified type has any structure or 1106/// alignment padding, false otherwise. 1107static bool HasPadding(const Type *Ty, const TargetData &TD) { 1108 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 1109 const StructLayout *SL = TD.getStructLayout(STy); 1110 unsigned PrevFieldBitOffset = 0; 1111 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1112 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 1113 1114 // Padding in sub-elements? 1115 if (HasPadding(STy->getElementType(i), TD)) 1116 return true; 1117 1118 // Check to see if there is any padding between this element and the 1119 // previous one. 1120 if (i) { 1121 unsigned PrevFieldEnd = 1122 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 1123 if (PrevFieldEnd < FieldBitOffset) 1124 return true; 1125 } 1126 1127 PrevFieldBitOffset = FieldBitOffset; 1128 } 1129 1130 // Check for tail padding. 1131 if (unsigned EltCount = STy->getNumElements()) { 1132 unsigned PrevFieldEnd = PrevFieldBitOffset + 1133 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 1134 if (PrevFieldEnd < SL->getSizeInBits()) 1135 return true; 1136 } 1137 1138 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1139 return HasPadding(ATy->getElementType(), TD); 1140 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) { 1141 return HasPadding(VTy->getElementType(), TD); 1142 } 1143 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 1144} 1145 1146/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 1147/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 1148/// or 1 if safe after canonicalization has been performed. 1149int SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 1150 // Loop over the use list of the alloca. We can only transform it if all of 1151 // the users are safe to transform. 1152 AllocaInfo Info; 1153 1154 isSafeForScalarRepl(AI, AI, 0, Info); 1155 if (Info.isUnsafe) { 1156 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 1157 return 0; 1158 } 1159 1160 // Okay, we know all the users are promotable. If the aggregate is a memcpy 1161 // source and destination, we have to be careful. In particular, the memcpy 1162 // could be moving around elements that live in structure padding of the LLVM 1163 // types, but may actually be used. In these cases, we refuse to promote the 1164 // struct. 1165 if (Info.isMemCpySrc && Info.isMemCpyDst && 1166 HasPadding(AI->getAllocatedType(), *TD)) 1167 return 0; 1168 1169 // If we require cleanup, return 1, otherwise return 3. 1170 return Info.needsCleanup ? 1 : 3; 1171} 1172 1173/// CleanupAllocaUsers - If SROA reported that it can promote the specified 1174/// allocation, but only if cleaned up, perform the cleanups required. 1175void SROA::CleanupAllocaUsers(Value *V) { 1176 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); 1177 UI != E; ) { 1178 User *U = *UI++; 1179 Instruction *I = cast<Instruction>(U); 1180 SmallVector<DbgInfoIntrinsic *, 2> DbgInUses; 1181 if (!isa<StoreInst>(I) && OnlyUsedByDbgInfoIntrinsics(I, &DbgInUses)) { 1182 // Safe to remove debug info uses. 1183 while (!DbgInUses.empty()) { 1184 DbgInfoIntrinsic *DI = DbgInUses.pop_back_val(); 1185 DI->eraseFromParent(); 1186 } 1187 I->eraseFromParent(); 1188 } 1189 } 1190} 1191 1192/// MergeInType - Add the 'In' type to the accumulated type (Accum) so far at 1193/// the offset specified by Offset (which is specified in bytes). 1194/// 1195/// There are two cases we handle here: 1196/// 1) A union of vector types of the same size and potentially its elements. 1197/// Here we turn element accesses into insert/extract element operations. 1198/// This promotes a <4 x float> with a store of float to the third element 1199/// into a <4 x float> that uses insert element. 1200/// 2) A fully general blob of memory, which we turn into some (potentially 1201/// large) integer type with extract and insert operations where the loads 1202/// and stores would mutate the memory. 1203static void MergeInType(const Type *In, uint64_t Offset, const Type *&VecTy, 1204 unsigned AllocaSize, const TargetData &TD, 1205 LLVMContext &Context) { 1206 // If this could be contributing to a vector, analyze it. 1207 if (VecTy != Type::getVoidTy(Context)) { // either null or a vector type. 1208 1209 // If the In type is a vector that is the same size as the alloca, see if it 1210 // matches the existing VecTy. 1211 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 1212 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 1213 // If we're storing/loading a vector of the right size, allow it as a 1214 // vector. If this the first vector we see, remember the type so that 1215 // we know the element size. 1216 if (VecTy == 0) 1217 VecTy = VInTy; 1218 return; 1219 } 1220 } else if (In->isFloatTy() || In->isDoubleTy() || 1221 (isa<IntegerType>(In) && In->getPrimitiveSizeInBits() >= 8 && 1222 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 1223 // If we're accessing something that could be an element of a vector, see 1224 // if the implied vector agrees with what we already have and if Offset is 1225 // compatible with it. 1226 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 1227 if (Offset % EltSize == 0 && 1228 AllocaSize % EltSize == 0 && 1229 (VecTy == 0 || 1230 cast<VectorType>(VecTy)->getElementType() 1231 ->getPrimitiveSizeInBits()/8 == EltSize)) { 1232 if (VecTy == 0) 1233 VecTy = VectorType::get(In, AllocaSize/EltSize); 1234 return; 1235 } 1236 } 1237 } 1238 1239 // Otherwise, we have a case that we can't handle with an optimized vector 1240 // form. We can still turn this into a large integer. 1241 VecTy = Type::getVoidTy(Context); 1242} 1243 1244/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 1245/// its accesses to a single vector type, return true and set VecTy to 1246/// the new type. If we could convert the alloca into a single promotable 1247/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 1248/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 1249/// is the current offset from the base of the alloca being analyzed. 1250/// 1251/// If we see at least one access to the value that is as a vector type, set the 1252/// SawVec flag. 1253bool SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy, 1254 bool &SawVec, uint64_t Offset, 1255 unsigned AllocaSize) { 1256 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1257 Instruction *User = cast<Instruction>(*UI); 1258 1259 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1260 // Don't break volatile loads. 1261 if (LI->isVolatile()) 1262 return false; 1263 MergeInType(LI->getType(), Offset, VecTy, 1264 AllocaSize, *TD, V->getContext()); 1265 SawVec |= isa<VectorType>(LI->getType()); 1266 continue; 1267 } 1268 1269 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1270 // Storing the pointer, not into the value? 1271 if (SI->getOperand(0) == V || SI->isVolatile()) return 0; 1272 MergeInType(SI->getOperand(0)->getType(), Offset, 1273 VecTy, AllocaSize, *TD, V->getContext()); 1274 SawVec |= isa<VectorType>(SI->getOperand(0)->getType()); 1275 continue; 1276 } 1277 1278 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 1279 if (!CanConvertToScalar(BCI, IsNotTrivial, VecTy, SawVec, Offset, 1280 AllocaSize)) 1281 return false; 1282 IsNotTrivial = true; 1283 continue; 1284 } 1285 1286 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1287 // If this is a GEP with a variable indices, we can't handle it. 1288 if (!GEP->hasAllConstantIndices()) 1289 return false; 1290 1291 // Compute the offset that this GEP adds to the pointer. 1292 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1293 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getPointerOperandType(), 1294 &Indices[0], Indices.size()); 1295 // See if all uses can be converted. 1296 if (!CanConvertToScalar(GEP, IsNotTrivial, VecTy, SawVec,Offset+GEPOffset, 1297 AllocaSize)) 1298 return false; 1299 IsNotTrivial = true; 1300 continue; 1301 } 1302 1303 // If this is a constant sized memset of a constant value (e.g. 0) we can 1304 // handle it. 1305 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 1306 // Store of constant value and constant size. 1307 if (isa<ConstantInt>(MSI->getValue()) && 1308 isa<ConstantInt>(MSI->getLength())) { 1309 IsNotTrivial = true; 1310 continue; 1311 } 1312 } 1313 1314 // If this is a memcpy or memmove into or out of the whole allocation, we 1315 // can handle it like a load or store of the scalar type. 1316 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 1317 if (ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength())) 1318 if (Len->getZExtValue() == AllocaSize && Offset == 0) { 1319 IsNotTrivial = true; 1320 continue; 1321 } 1322 } 1323 1324 // Ignore dbg intrinsic. 1325 if (isa<DbgInfoIntrinsic>(User)) 1326 continue; 1327 1328 // Otherwise, we cannot handle this! 1329 return false; 1330 } 1331 1332 return true; 1333} 1334 1335/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 1336/// directly. This happens when we are converting an "integer union" to a 1337/// single integer scalar, or when we are converting a "vector union" to a 1338/// vector with insert/extractelement instructions. 1339/// 1340/// Offset is an offset from the original alloca, in bits that need to be 1341/// shifted to the right. By the end of this, there should be no uses of Ptr. 1342void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset) { 1343 while (!Ptr->use_empty()) { 1344 Instruction *User = cast<Instruction>(Ptr->use_back()); 1345 1346 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 1347 ConvertUsesToScalar(CI, NewAI, Offset); 1348 CI->eraseFromParent(); 1349 continue; 1350 } 1351 1352 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1353 // Compute the offset that this GEP adds to the pointer. 1354 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1355 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getPointerOperandType(), 1356 &Indices[0], Indices.size()); 1357 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 1358 GEP->eraseFromParent(); 1359 continue; 1360 } 1361 1362 IRBuilder<> Builder(User->getParent(), User); 1363 1364 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1365 // The load is a bit extract from NewAI shifted right by Offset bits. 1366 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 1367 Value *NewLoadVal 1368 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 1369 LI->replaceAllUsesWith(NewLoadVal); 1370 LI->eraseFromParent(); 1371 continue; 1372 } 1373 1374 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1375 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 1376 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 1377 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 1378 Builder); 1379 Builder.CreateStore(New, NewAI); 1380 SI->eraseFromParent(); 1381 1382 // If the load we just inserted is now dead, then the inserted store 1383 // overwrote the entire thing. 1384 if (Old->use_empty()) 1385 Old->eraseFromParent(); 1386 continue; 1387 } 1388 1389 // If this is a constant sized memset of a constant value (e.g. 0) we can 1390 // transform it into a store of the expanded constant value. 1391 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 1392 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 1393 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 1394 if (NumBytes != 0) { 1395 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 1396 1397 // Compute the value replicated the right number of times. 1398 APInt APVal(NumBytes*8, Val); 1399 1400 // Splat the value if non-zero. 1401 if (Val) 1402 for (unsigned i = 1; i != NumBytes; ++i) 1403 APVal |= APVal << 8; 1404 1405 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 1406 Value *New = ConvertScalar_InsertValue( 1407 ConstantInt::get(User->getContext(), APVal), 1408 Old, Offset, Builder); 1409 Builder.CreateStore(New, NewAI); 1410 1411 // If the load we just inserted is now dead, then the memset overwrote 1412 // the entire thing. 1413 if (Old->use_empty()) 1414 Old->eraseFromParent(); 1415 } 1416 MSI->eraseFromParent(); 1417 continue; 1418 } 1419 1420 // If this is a memcpy or memmove into or out of the whole allocation, we 1421 // can handle it like a load or store of the scalar type. 1422 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 1423 assert(Offset == 0 && "must be store to start of alloca"); 1424 1425 // If the source and destination are both to the same alloca, then this is 1426 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 1427 // as appropriate. 1428 AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject()); 1429 1430 if (MTI->getSource()->getUnderlyingObject() != OrigAI) { 1431 // Dest must be OrigAI, change this to be a load from the original 1432 // pointer (bitcasted), then a store to our new alloca. 1433 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 1434 Value *SrcPtr = MTI->getSource(); 1435 SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType()); 1436 1437 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 1438 SrcVal->setAlignment(MTI->getAlignment()); 1439 Builder.CreateStore(SrcVal, NewAI); 1440 } else if (MTI->getDest()->getUnderlyingObject() != OrigAI) { 1441 // Src must be OrigAI, change this to be a load from NewAI then a store 1442 // through the original dest pointer (bitcasted). 1443 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 1444 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 1445 1446 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType()); 1447 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 1448 NewStore->setAlignment(MTI->getAlignment()); 1449 } else { 1450 // Noop transfer. Src == Dst 1451 } 1452 1453 1454 MTI->eraseFromParent(); 1455 continue; 1456 } 1457 1458 // If user is a dbg info intrinsic then it is safe to remove it. 1459 if (isa<DbgInfoIntrinsic>(User)) { 1460 User->eraseFromParent(); 1461 continue; 1462 } 1463 1464 llvm_unreachable("Unsupported operation!"); 1465 } 1466} 1467 1468/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 1469/// or vector value FromVal, extracting the bits from the offset specified by 1470/// Offset. This returns the value, which is of type ToType. 1471/// 1472/// This happens when we are converting an "integer union" to a single 1473/// integer scalar, or when we are converting a "vector union" to a vector with 1474/// insert/extractelement instructions. 1475/// 1476/// Offset is an offset from the original alloca, in bits that need to be 1477/// shifted to the right. 1478Value *SROA::ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 1479 uint64_t Offset, IRBuilder<> &Builder) { 1480 // If the load is of the whole new alloca, no conversion is needed. 1481 if (FromVal->getType() == ToType && Offset == 0) 1482 return FromVal; 1483 1484 // If the result alloca is a vector type, this is either an element 1485 // access or a bitcast to another vector type of the same size. 1486 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) { 1487 if (isa<VectorType>(ToType)) 1488 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 1489 1490 // Otherwise it must be an element access. 1491 unsigned Elt = 0; 1492 if (Offset) { 1493 unsigned EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 1494 Elt = Offset/EltSize; 1495 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 1496 } 1497 // Return the element extracted out of it. 1498 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get( 1499 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp"); 1500 if (V->getType() != ToType) 1501 V = Builder.CreateBitCast(V, ToType, "tmp"); 1502 return V; 1503 } 1504 1505 // If ToType is a first class aggregate, extract out each of the pieces and 1506 // use insertvalue's to form the FCA. 1507 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 1508 const StructLayout &Layout = *TD->getStructLayout(ST); 1509 Value *Res = UndefValue::get(ST); 1510 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1511 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 1512 Offset+Layout.getElementOffsetInBits(i), 1513 Builder); 1514 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1515 } 1516 return Res; 1517 } 1518 1519 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 1520 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 1521 Value *Res = UndefValue::get(AT); 1522 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1523 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 1524 Offset+i*EltSize, Builder); 1525 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1526 } 1527 return Res; 1528 } 1529 1530 // Otherwise, this must be a union that was converted to an integer value. 1531 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 1532 1533 // If this is a big-endian system and the load is narrower than the 1534 // full alloca type, we need to do a shift to get the right bits. 1535 int ShAmt = 0; 1536 if (TD->isBigEndian()) { 1537 // On big-endian machines, the lowest bit is stored at the bit offset 1538 // from the pointer given by getTypeStoreSizeInBits. This matters for 1539 // integers with a bitwidth that is not a multiple of 8. 1540 ShAmt = TD->getTypeStoreSizeInBits(NTy) - 1541 TD->getTypeStoreSizeInBits(ToType) - Offset; 1542 } else { 1543 ShAmt = Offset; 1544 } 1545 1546 // Note: we support negative bitwidths (with shl) which are not defined. 1547 // We do this to support (f.e.) loads off the end of a structure where 1548 // only some bits are used. 1549 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 1550 FromVal = Builder.CreateLShr(FromVal, 1551 ConstantInt::get(FromVal->getType(), 1552 ShAmt), "tmp"); 1553 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 1554 FromVal = Builder.CreateShl(FromVal, 1555 ConstantInt::get(FromVal->getType(), 1556 -ShAmt), "tmp"); 1557 1558 // Finally, unconditionally truncate the integer to the right width. 1559 unsigned LIBitWidth = TD->getTypeSizeInBits(ToType); 1560 if (LIBitWidth < NTy->getBitWidth()) 1561 FromVal = 1562 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 1563 LIBitWidth), "tmp"); 1564 else if (LIBitWidth > NTy->getBitWidth()) 1565 FromVal = 1566 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 1567 LIBitWidth), "tmp"); 1568 1569 // If the result is an integer, this is a trunc or bitcast. 1570 if (isa<IntegerType>(ToType)) { 1571 // Should be done. 1572 } else if (ToType->isFloatingPoint() || isa<VectorType>(ToType)) { 1573 // Just do a bitcast, we know the sizes match up. 1574 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 1575 } else { 1576 // Otherwise must be a pointer. 1577 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 1578 } 1579 assert(FromVal->getType() == ToType && "Didn't convert right?"); 1580 return FromVal; 1581} 1582 1583/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 1584/// or vector value "Old" at the offset specified by Offset. 1585/// 1586/// This happens when we are converting an "integer union" to a 1587/// single integer scalar, or when we are converting a "vector union" to a 1588/// vector with insert/extractelement instructions. 1589/// 1590/// Offset is an offset from the original alloca, in bits that need to be 1591/// shifted to the right. 1592Value *SROA::ConvertScalar_InsertValue(Value *SV, Value *Old, 1593 uint64_t Offset, IRBuilder<> &Builder) { 1594 1595 // Convert the stored type to the actual type, shift it left to insert 1596 // then 'or' into place. 1597 const Type *AllocaType = Old->getType(); 1598 LLVMContext &Context = Old->getContext(); 1599 1600 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 1601 uint64_t VecSize = TD->getTypeAllocSizeInBits(VTy); 1602 uint64_t ValSize = TD->getTypeAllocSizeInBits(SV->getType()); 1603 1604 // Changing the whole vector with memset or with an access of a different 1605 // vector type? 1606 if (ValSize == VecSize) 1607 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 1608 1609 uint64_t EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 1610 1611 // Must be an element insertion. 1612 unsigned Elt = Offset/EltSize; 1613 1614 if (SV->getType() != VTy->getElementType()) 1615 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 1616 1617 SV = Builder.CreateInsertElement(Old, SV, 1618 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt), 1619 "tmp"); 1620 return SV; 1621 } 1622 1623 // If SV is a first-class aggregate value, insert each value recursively. 1624 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 1625 const StructLayout &Layout = *TD->getStructLayout(ST); 1626 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1627 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1628 Old = ConvertScalar_InsertValue(Elt, Old, 1629 Offset+Layout.getElementOffsetInBits(i), 1630 Builder); 1631 } 1632 return Old; 1633 } 1634 1635 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 1636 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 1637 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1638 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1639 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 1640 } 1641 return Old; 1642 } 1643 1644 // If SV is a float, convert it to the appropriate integer type. 1645 // If it is a pointer, do the same. 1646 unsigned SrcWidth = TD->getTypeSizeInBits(SV->getType()); 1647 unsigned DestWidth = TD->getTypeSizeInBits(AllocaType); 1648 unsigned SrcStoreWidth = TD->getTypeStoreSizeInBits(SV->getType()); 1649 unsigned DestStoreWidth = TD->getTypeStoreSizeInBits(AllocaType); 1650 if (SV->getType()->isFloatingPoint() || isa<VectorType>(SV->getType())) 1651 SV = Builder.CreateBitCast(SV, 1652 IntegerType::get(SV->getContext(),SrcWidth), "tmp"); 1653 else if (isa<PointerType>(SV->getType())) 1654 SV = Builder.CreatePtrToInt(SV, TD->getIntPtrType(SV->getContext()), "tmp"); 1655 1656 // Zero extend or truncate the value if needed. 1657 if (SV->getType() != AllocaType) { 1658 if (SV->getType()->getPrimitiveSizeInBits() < 1659 AllocaType->getPrimitiveSizeInBits()) 1660 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 1661 else { 1662 // Truncation may be needed if storing more than the alloca can hold 1663 // (undefined behavior). 1664 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 1665 SrcWidth = DestWidth; 1666 SrcStoreWidth = DestStoreWidth; 1667 } 1668 } 1669 1670 // If this is a big-endian system and the store is narrower than the 1671 // full alloca type, we need to do a shift to get the right bits. 1672 int ShAmt = 0; 1673 if (TD->isBigEndian()) { 1674 // On big-endian machines, the lowest bit is stored at the bit offset 1675 // from the pointer given by getTypeStoreSizeInBits. This matters for 1676 // integers with a bitwidth that is not a multiple of 8. 1677 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 1678 } else { 1679 ShAmt = Offset; 1680 } 1681 1682 // Note: we support negative bitwidths (with shr) which are not defined. 1683 // We do this to support (f.e.) stores off the end of a structure where 1684 // only some bits in the structure are set. 1685 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 1686 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 1687 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 1688 ShAmt), "tmp"); 1689 Mask <<= ShAmt; 1690 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1691 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 1692 -ShAmt), "tmp"); 1693 Mask = Mask.lshr(-ShAmt); 1694 } 1695 1696 // Mask out the bits we are about to insert from the old value, and or 1697 // in the new bits. 1698 if (SrcWidth != DestWidth) { 1699 assert(DestWidth > SrcWidth); 1700 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 1701 SV = Builder.CreateOr(Old, SV, "ins"); 1702 } 1703 return SV; 1704} 1705 1706 1707 1708/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1709/// some part of a constant global variable. This intentionally only accepts 1710/// constant expressions because we don't can't rewrite arbitrary instructions. 1711static bool PointsToConstantGlobal(Value *V) { 1712 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1713 return GV->isConstant(); 1714 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1715 if (CE->getOpcode() == Instruction::BitCast || 1716 CE->getOpcode() == Instruction::GetElementPtr) 1717 return PointsToConstantGlobal(CE->getOperand(0)); 1718 return false; 1719} 1720 1721/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1722/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1723/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1724/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1725/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1726/// the alloca, and if the source pointer is a pointer to a constant global, we 1727/// can optimize this. 1728static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy, 1729 bool isOffset) { 1730 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1731 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) 1732 // Ignore non-volatile loads, they are always ok. 1733 if (!LI->isVolatile()) 1734 continue; 1735 1736 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) { 1737 // If uses of the bitcast are ok, we are ok. 1738 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 1739 return false; 1740 continue; 1741 } 1742 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) { 1743 // If the GEP has all zero indices, it doesn't offset the pointer. If it 1744 // doesn't, it does. 1745 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 1746 isOffset || !GEP->hasAllZeroIndices())) 1747 return false; 1748 continue; 1749 } 1750 1751 // If this is isn't our memcpy/memmove, reject it as something we can't 1752 // handle. 1753 if (!isa<MemTransferInst>(*UI)) 1754 return false; 1755 1756 // If we already have seen a copy, reject the second one. 1757 if (TheCopy) return false; 1758 1759 // If the pointer has been offset from the start of the alloca, we can't 1760 // safely handle this. 1761 if (isOffset) return false; 1762 1763 // If the memintrinsic isn't using the alloca as the dest, reject it. 1764 if (UI.getOperandNo() != 1) return false; 1765 1766 MemIntrinsic *MI = cast<MemIntrinsic>(*UI); 1767 1768 // If the source of the memcpy/move is not a constant global, reject it. 1769 if (!PointsToConstantGlobal(MI->getOperand(2))) 1770 return false; 1771 1772 // Otherwise, the transform is safe. Remember the copy instruction. 1773 TheCopy = MI; 1774 } 1775 return true; 1776} 1777 1778/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 1779/// modified by a copy from a constant global. If we can prove this, we can 1780/// replace any uses of the alloca with uses of the global directly. 1781Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) { 1782 Instruction *TheCopy = 0; 1783 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 1784 return TheCopy; 1785 return 0; 1786} 1787