MemCpyOptimizer.cpp revision 221345
1193323Sed//===- MemCpyOptimizer.cpp - Optimize use of memcpy and friends -----------===// 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 pass performs various transformations related to eliminating memcpy 11193323Sed// calls, or transforming sets of stores into memset's. 12193323Sed// 13193323Sed//===----------------------------------------------------------------------===// 14193323Sed 15193323Sed#define DEBUG_TYPE "memcpyopt" 16193323Sed#include "llvm/Transforms/Scalar.h" 17218893Sdim#include "llvm/GlobalVariable.h" 18193323Sed#include "llvm/IntrinsicInst.h" 19193323Sed#include "llvm/Instructions.h" 20193323Sed#include "llvm/ADT/SmallVector.h" 21193323Sed#include "llvm/ADT/Statistic.h" 22193323Sed#include "llvm/Analysis/Dominators.h" 23193323Sed#include "llvm/Analysis/AliasAnalysis.h" 24193323Sed#include "llvm/Analysis/MemoryDependenceAnalysis.h" 25218893Sdim#include "llvm/Analysis/ValueTracking.h" 26193323Sed#include "llvm/Support/Debug.h" 27193323Sed#include "llvm/Support/GetElementPtrTypeIterator.h" 28218893Sdim#include "llvm/Support/IRBuilder.h" 29198090Srdivacky#include "llvm/Support/raw_ostream.h" 30193323Sed#include "llvm/Target/TargetData.h" 31221345Sdim#include "llvm/Target/TargetLibraryInfo.h" 32193323Sed#include <list> 33193323Sedusing namespace llvm; 34193323Sed 35193323SedSTATISTIC(NumMemCpyInstr, "Number of memcpy instructions deleted"); 36193323SedSTATISTIC(NumMemSetInfer, "Number of memsets inferred"); 37198090SrdivackySTATISTIC(NumMoveToCpy, "Number of memmoves converted to memcpy"); 38218893SdimSTATISTIC(NumCpyToSet, "Number of memcpys converted to memset"); 39193323Sed 40193323Sedstatic int64_t GetOffsetFromIndex(const GetElementPtrInst *GEP, unsigned Idx, 41218893Sdim bool &VariableIdxFound, const TargetData &TD){ 42193323Sed // Skip over the first indices. 43193323Sed gep_type_iterator GTI = gep_type_begin(GEP); 44193323Sed for (unsigned i = 1; i != Idx; ++i, ++GTI) 45193323Sed /*skip along*/; 46193323Sed 47193323Sed // Compute the offset implied by the rest of the indices. 48193323Sed int64_t Offset = 0; 49193323Sed for (unsigned i = Idx, e = GEP->getNumOperands(); i != e; ++i, ++GTI) { 50193323Sed ConstantInt *OpC = dyn_cast<ConstantInt>(GEP->getOperand(i)); 51193323Sed if (OpC == 0) 52193323Sed return VariableIdxFound = true; 53193323Sed if (OpC->isZero()) continue; // No offset. 54193323Sed 55193323Sed // Handle struct indices, which add their field offset to the pointer. 56193323Sed if (const StructType *STy = dyn_cast<StructType>(*GTI)) { 57193323Sed Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue()); 58193323Sed continue; 59193323Sed } 60193323Sed 61193323Sed // Otherwise, we have a sequential type like an array or vector. Multiply 62193323Sed // the index by the ElementSize. 63193323Sed uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()); 64193323Sed Offset += Size*OpC->getSExtValue(); 65193323Sed } 66193323Sed 67193323Sed return Offset; 68193323Sed} 69193323Sed 70193323Sed/// IsPointerOffset - Return true if Ptr1 is provably equal to Ptr2 plus a 71193323Sed/// constant offset, and return that constant offset. For example, Ptr1 might 72193323Sed/// be &A[42], and Ptr2 might be &A[40]. In this case offset would be -8. 73193323Sedstatic bool IsPointerOffset(Value *Ptr1, Value *Ptr2, int64_t &Offset, 74218893Sdim const TargetData &TD) { 75218893Sdim Ptr1 = Ptr1->stripPointerCasts(); 76218893Sdim Ptr2 = Ptr2->stripPointerCasts(); 77218893Sdim GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(Ptr1); 78218893Sdim GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(Ptr2); 79218893Sdim 80218893Sdim bool VariableIdxFound = false; 81218893Sdim 82218893Sdim // If one pointer is a GEP and the other isn't, then see if the GEP is a 83218893Sdim // constant offset from the base, as in "P" and "gep P, 1". 84218893Sdim if (GEP1 && GEP2 == 0 && GEP1->getOperand(0)->stripPointerCasts() == Ptr2) { 85218893Sdim Offset = -GetOffsetFromIndex(GEP1, 1, VariableIdxFound, TD); 86218893Sdim return !VariableIdxFound; 87218893Sdim } 88218893Sdim 89218893Sdim if (GEP2 && GEP1 == 0 && GEP2->getOperand(0)->stripPointerCasts() == Ptr1) { 90218893Sdim Offset = GetOffsetFromIndex(GEP2, 1, VariableIdxFound, TD); 91218893Sdim return !VariableIdxFound; 92218893Sdim } 93218893Sdim 94193323Sed // Right now we handle the case when Ptr1/Ptr2 are both GEPs with an identical 95193323Sed // base. After that base, they may have some number of common (and 96193323Sed // potentially variable) indices. After that they handle some constant 97193323Sed // offset, which determines their offset from each other. At this point, we 98193323Sed // handle no other case. 99193323Sed if (!GEP1 || !GEP2 || GEP1->getOperand(0) != GEP2->getOperand(0)) 100193323Sed return false; 101193323Sed 102193323Sed // Skip any common indices and track the GEP types. 103193323Sed unsigned Idx = 1; 104193323Sed for (; Idx != GEP1->getNumOperands() && Idx != GEP2->getNumOperands(); ++Idx) 105193323Sed if (GEP1->getOperand(Idx) != GEP2->getOperand(Idx)) 106193323Sed break; 107193323Sed 108193323Sed int64_t Offset1 = GetOffsetFromIndex(GEP1, Idx, VariableIdxFound, TD); 109193323Sed int64_t Offset2 = GetOffsetFromIndex(GEP2, Idx, VariableIdxFound, TD); 110193323Sed if (VariableIdxFound) return false; 111193323Sed 112193323Sed Offset = Offset2-Offset1; 113193323Sed return true; 114193323Sed} 115193323Sed 116193323Sed 117193323Sed/// MemsetRange - Represents a range of memset'd bytes with the ByteVal value. 118193323Sed/// This allows us to analyze stores like: 119193323Sed/// store 0 -> P+1 120193323Sed/// store 0 -> P+0 121193323Sed/// store 0 -> P+3 122193323Sed/// store 0 -> P+2 123193323Sed/// which sometimes happens with stores to arrays of structs etc. When we see 124193323Sed/// the first store, we make a range [1, 2). The second store extends the range 125193323Sed/// to [0, 2). The third makes a new range [2, 3). The fourth store joins the 126193323Sed/// two ranges into [0, 3) which is memset'able. 127193323Sednamespace { 128193323Sedstruct MemsetRange { 129193323Sed // Start/End - A semi range that describes the span that this range covers. 130193323Sed // The range is closed at the start and open at the end: [Start, End). 131193323Sed int64_t Start, End; 132193323Sed 133193323Sed /// StartPtr - The getelementptr instruction that points to the start of the 134193323Sed /// range. 135193323Sed Value *StartPtr; 136193323Sed 137193323Sed /// Alignment - The known alignment of the first store. 138193323Sed unsigned Alignment; 139193323Sed 140193323Sed /// TheStores - The actual stores that make up this range. 141218893Sdim SmallVector<Instruction*, 16> TheStores; 142193323Sed 143193323Sed bool isProfitableToUseMemset(const TargetData &TD) const; 144193323Sed 145193323Sed}; 146193323Sed} // end anon namespace 147193323Sed 148193323Sedbool MemsetRange::isProfitableToUseMemset(const TargetData &TD) const { 149193323Sed // If we found more than 8 stores to merge or 64 bytes, use memset. 150193323Sed if (TheStores.size() >= 8 || End-Start >= 64) return true; 151218893Sdim 152218893Sdim // If there is nothing to merge, don't do anything. 153218893Sdim if (TheStores.size() < 2) return false; 154193323Sed 155218893Sdim // If any of the stores are a memset, then it is always good to extend the 156218893Sdim // memset. 157218893Sdim for (unsigned i = 0, e = TheStores.size(); i != e; ++i) 158218893Sdim if (!isa<StoreInst>(TheStores[i])) 159218893Sdim return true; 160218893Sdim 161193323Sed // Assume that the code generator is capable of merging pairs of stores 162193323Sed // together if it wants to. 163218893Sdim if (TheStores.size() == 2) return false; 164193323Sed 165193323Sed // If we have fewer than 8 stores, it can still be worthwhile to do this. 166193323Sed // For example, merging 4 i8 stores into an i32 store is useful almost always. 167193323Sed // However, merging 2 32-bit stores isn't useful on a 32-bit architecture (the 168193323Sed // memset will be split into 2 32-bit stores anyway) and doing so can 169193323Sed // pessimize the llvm optimizer. 170193323Sed // 171193323Sed // Since we don't have perfect knowledge here, make some assumptions: assume 172193323Sed // the maximum GPR width is the same size as the pointer size and assume that 173193323Sed // this width can be stored. If so, check to see whether we will end up 174193323Sed // actually reducing the number of stores used. 175193323Sed unsigned Bytes = unsigned(End-Start); 176193323Sed unsigned NumPointerStores = Bytes/TD.getPointerSize(); 177193323Sed 178193323Sed // Assume the remaining bytes if any are done a byte at a time. 179193323Sed unsigned NumByteStores = Bytes - NumPointerStores*TD.getPointerSize(); 180193323Sed 181193323Sed // If we will reduce the # stores (according to this heuristic), do the 182193323Sed // transformation. This encourages merging 4 x i8 -> i32 and 2 x i16 -> i32 183193323Sed // etc. 184193323Sed return TheStores.size() > NumPointerStores+NumByteStores; 185193323Sed} 186193323Sed 187193323Sed 188193323Sednamespace { 189193323Sedclass MemsetRanges { 190193323Sed /// Ranges - A sorted list of the memset ranges. We use std::list here 191193323Sed /// because each element is relatively large and expensive to copy. 192193323Sed std::list<MemsetRange> Ranges; 193193323Sed typedef std::list<MemsetRange>::iterator range_iterator; 194218893Sdim const TargetData &TD; 195193323Sedpublic: 196218893Sdim MemsetRanges(const TargetData &td) : TD(td) {} 197193323Sed 198193323Sed typedef std::list<MemsetRange>::const_iterator const_iterator; 199193323Sed const_iterator begin() const { return Ranges.begin(); } 200193323Sed const_iterator end() const { return Ranges.end(); } 201193323Sed bool empty() const { return Ranges.empty(); } 202193323Sed 203218893Sdim void addInst(int64_t OffsetFromFirst, Instruction *Inst) { 204218893Sdim if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) 205218893Sdim addStore(OffsetFromFirst, SI); 206218893Sdim else 207218893Sdim addMemSet(OffsetFromFirst, cast<MemSetInst>(Inst)); 208218893Sdim } 209218893Sdim 210218893Sdim void addStore(int64_t OffsetFromFirst, StoreInst *SI) { 211218893Sdim int64_t StoreSize = TD.getTypeStoreSize(SI->getOperand(0)->getType()); 212218893Sdim 213218893Sdim addRange(OffsetFromFirst, StoreSize, 214218893Sdim SI->getPointerOperand(), SI->getAlignment(), SI); 215218893Sdim } 216218893Sdim 217218893Sdim void addMemSet(int64_t OffsetFromFirst, MemSetInst *MSI) { 218218893Sdim int64_t Size = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 219218893Sdim addRange(OffsetFromFirst, Size, MSI->getDest(), MSI->getAlignment(), MSI); 220218893Sdim } 221218893Sdim 222218893Sdim void addRange(int64_t Start, int64_t Size, Value *Ptr, 223218893Sdim unsigned Alignment, Instruction *Inst); 224218893Sdim 225193323Sed}; 226193323Sed 227193323Sed} // end anon namespace 228193323Sed 229193323Sed 230218893Sdim/// addRange - Add a new store to the MemsetRanges data structure. This adds a 231193323Sed/// new range for the specified store at the specified offset, merging into 232193323Sed/// existing ranges as appropriate. 233218893Sdim/// 234218893Sdim/// Do a linear search of the ranges to see if this can be joined and/or to 235218893Sdim/// find the insertion point in the list. We keep the ranges sorted for 236218893Sdim/// simplicity here. This is a linear search of a linked list, which is ugly, 237218893Sdim/// however the number of ranges is limited, so this won't get crazy slow. 238218893Sdimvoid MemsetRanges::addRange(int64_t Start, int64_t Size, Value *Ptr, 239218893Sdim unsigned Alignment, Instruction *Inst) { 240218893Sdim int64_t End = Start+Size; 241193323Sed range_iterator I = Ranges.begin(), E = Ranges.end(); 242193323Sed 243193323Sed while (I != E && Start > I->End) 244193323Sed ++I; 245193323Sed 246193323Sed // We now know that I == E, in which case we didn't find anything to merge 247193323Sed // with, or that Start <= I->End. If End < I->Start or I == E, then we need 248193323Sed // to insert a new range. Handle this now. 249193323Sed if (I == E || End < I->Start) { 250193323Sed MemsetRange &R = *Ranges.insert(I, MemsetRange()); 251193323Sed R.Start = Start; 252193323Sed R.End = End; 253218893Sdim R.StartPtr = Ptr; 254218893Sdim R.Alignment = Alignment; 255218893Sdim R.TheStores.push_back(Inst); 256193323Sed return; 257193323Sed } 258218893Sdim 259193323Sed // This store overlaps with I, add it. 260218893Sdim I->TheStores.push_back(Inst); 261193323Sed 262193323Sed // At this point, we may have an interval that completely contains our store. 263193323Sed // If so, just add it to the interval and return. 264193323Sed if (I->Start <= Start && I->End >= End) 265193323Sed return; 266193323Sed 267193323Sed // Now we know that Start <= I->End and End >= I->Start so the range overlaps 268193323Sed // but is not entirely contained within the range. 269193323Sed 270193323Sed // See if the range extends the start of the range. In this case, it couldn't 271193323Sed // possibly cause it to join the prior range, because otherwise we would have 272193323Sed // stopped on *it*. 273193323Sed if (Start < I->Start) { 274193323Sed I->Start = Start; 275218893Sdim I->StartPtr = Ptr; 276218893Sdim I->Alignment = Alignment; 277193323Sed } 278193323Sed 279193323Sed // Now we know that Start <= I->End and Start >= I->Start (so the startpoint 280193323Sed // is in or right at the end of I), and that End >= I->Start. Extend I out to 281193323Sed // End. 282193323Sed if (End > I->End) { 283193323Sed I->End = End; 284193323Sed range_iterator NextI = I; 285193323Sed while (++NextI != E && End >= NextI->Start) { 286193323Sed // Merge the range in. 287193323Sed I->TheStores.append(NextI->TheStores.begin(), NextI->TheStores.end()); 288193323Sed if (NextI->End > I->End) 289193323Sed I->End = NextI->End; 290193323Sed Ranges.erase(NextI); 291193323Sed NextI = I; 292193323Sed } 293193323Sed } 294193323Sed} 295193323Sed 296193323Sed//===----------------------------------------------------------------------===// 297193323Sed// MemCpyOpt Pass 298193323Sed//===----------------------------------------------------------------------===// 299193323Sed 300193323Sednamespace { 301198090Srdivacky class MemCpyOpt : public FunctionPass { 302218893Sdim MemoryDependenceAnalysis *MD; 303221345Sdim TargetLibraryInfo *TLI; 304218893Sdim const TargetData *TD; 305193323Sed public: 306193323Sed static char ID; // Pass identification, replacement for typeid 307218893Sdim MemCpyOpt() : FunctionPass(ID) { 308218893Sdim initializeMemCpyOptPass(*PassRegistry::getPassRegistry()); 309218893Sdim MD = 0; 310221345Sdim TLI = 0; 311221345Sdim TD = 0; 312218893Sdim } 313193323Sed 314218893Sdim bool runOnFunction(Function &F); 315218893Sdim 316193323Sed private: 317193323Sed // This transformation requires dominator postdominator info 318193323Sed virtual void getAnalysisUsage(AnalysisUsage &AU) const { 319193323Sed AU.setPreservesCFG(); 320193323Sed AU.addRequired<DominatorTree>(); 321193323Sed AU.addRequired<MemoryDependenceAnalysis>(); 322193323Sed AU.addRequired<AliasAnalysis>(); 323221345Sdim AU.addRequired<TargetLibraryInfo>(); 324193323Sed AU.addPreserved<AliasAnalysis>(); 325193323Sed AU.addPreserved<MemoryDependenceAnalysis>(); 326193323Sed } 327193323Sed 328193323Sed // Helper fuctions 329198090Srdivacky bool processStore(StoreInst *SI, BasicBlock::iterator &BBI); 330218893Sdim bool processMemSet(MemSetInst *SI, BasicBlock::iterator &BBI); 331198090Srdivacky bool processMemCpy(MemCpyInst *M); 332198090Srdivacky bool processMemMove(MemMoveInst *M); 333218893Sdim bool performCallSlotOptzn(Instruction *cpy, Value *cpyDst, Value *cpySrc, 334218893Sdim uint64_t cpyLen, CallInst *C); 335218893Sdim bool processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep, 336218893Sdim uint64_t MSize); 337218893Sdim bool processByValArgument(CallSite CS, unsigned ArgNo); 338218893Sdim Instruction *tryMergingIntoMemset(Instruction *I, Value *StartPtr, 339218893Sdim Value *ByteVal); 340218893Sdim 341193323Sed bool iterateOnFunction(Function &F); 342193323Sed }; 343193323Sed 344193323Sed char MemCpyOpt::ID = 0; 345193323Sed} 346193323Sed 347193323Sed// createMemCpyOptPass - The public interface to this file... 348193323SedFunctionPass *llvm::createMemCpyOptPass() { return new MemCpyOpt(); } 349193323Sed 350218893SdimINITIALIZE_PASS_BEGIN(MemCpyOpt, "memcpyopt", "MemCpy Optimization", 351218893Sdim false, false) 352218893SdimINITIALIZE_PASS_DEPENDENCY(DominatorTree) 353218893SdimINITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis) 354221345SdimINITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) 355218893SdimINITIALIZE_AG_DEPENDENCY(AliasAnalysis) 356218893SdimINITIALIZE_PASS_END(MemCpyOpt, "memcpyopt", "MemCpy Optimization", 357218893Sdim false, false) 358193323Sed 359218893Sdim/// tryMergingIntoMemset - When scanning forward over instructions, we look for 360193323Sed/// some other patterns to fold away. In particular, this looks for stores to 361218893Sdim/// neighboring locations of memory. If it sees enough consecutive ones, it 362218893Sdim/// attempts to merge them together into a memcpy/memset. 363218893SdimInstruction *MemCpyOpt::tryMergingIntoMemset(Instruction *StartInst, 364218893Sdim Value *StartPtr, Value *ByteVal) { 365218893Sdim if (TD == 0) return 0; 366193323Sed 367193323Sed // Okay, so we now have a single store that can be splatable. Scan to find 368193323Sed // all subsequent stores of the same value to offset from the same pointer. 369193323Sed // Join these together into ranges, so we can decide whether contiguous blocks 370193323Sed // are stored. 371198090Srdivacky MemsetRanges Ranges(*TD); 372193323Sed 373218893Sdim BasicBlock::iterator BI = StartInst; 374193323Sed for (++BI; !isa<TerminatorInst>(BI); ++BI) { 375218893Sdim if (!isa<StoreInst>(BI) && !isa<MemSetInst>(BI)) { 376218893Sdim // If the instruction is readnone, ignore it, otherwise bail out. We 377218893Sdim // don't even allow readonly here because we don't want something like: 378193323Sed // A[1] = 2; strlen(A); A[2] = 2; -> memcpy(A, ...); strlen(A). 379218893Sdim if (BI->mayWriteToMemory() || BI->mayReadFromMemory()) 380218893Sdim break; 381218893Sdim continue; 382218893Sdim } 383218893Sdim 384218893Sdim if (StoreInst *NextStore = dyn_cast<StoreInst>(BI)) { 385218893Sdim // If this is a store, see if we can merge it in. 386218893Sdim if (NextStore->isVolatile()) break; 387218893Sdim 388218893Sdim // Check to see if this stored value is of the same byte-splattable value. 389218893Sdim if (ByteVal != isBytewiseValue(NextStore->getOperand(0))) 390218893Sdim break; 391193323Sed 392218893Sdim // Check to see if this store is to a constant offset from the start ptr. 393218893Sdim int64_t Offset; 394218893Sdim if (!IsPointerOffset(StartPtr, NextStore->getPointerOperand(), 395218893Sdim Offset, *TD)) 396218893Sdim break; 397193323Sed 398218893Sdim Ranges.addStore(Offset, NextStore); 399218893Sdim } else { 400218893Sdim MemSetInst *MSI = cast<MemSetInst>(BI); 401218893Sdim 402218893Sdim if (MSI->isVolatile() || ByteVal != MSI->getValue() || 403218893Sdim !isa<ConstantInt>(MSI->getLength())) 404218893Sdim break; 405218893Sdim 406218893Sdim // Check to see if this store is to a constant offset from the start ptr. 407218893Sdim int64_t Offset; 408218893Sdim if (!IsPointerOffset(StartPtr, MSI->getDest(), Offset, *TD)) 409218893Sdim break; 410218893Sdim 411218893Sdim Ranges.addMemSet(Offset, MSI); 412218893Sdim } 413193323Sed } 414218893Sdim 415193323Sed // If we have no ranges, then we just had a single store with nothing that 416193323Sed // could be merged in. This is a very common case of course. 417193323Sed if (Ranges.empty()) 418218893Sdim return 0; 419193323Sed 420193323Sed // If we had at least one store that could be merged in, add the starting 421193323Sed // store as well. We try to avoid this unless there is at least something 422193323Sed // interesting as a small compile-time optimization. 423218893Sdim Ranges.addInst(0, StartInst); 424218893Sdim 425218893Sdim // If we create any memsets, we put it right before the first instruction that 426218893Sdim // isn't part of the memset block. This ensure that the memset is dominated 427218893Sdim // by any addressing instruction needed by the start of the block. 428218893Sdim IRBuilder<> Builder(BI); 429218893Sdim 430193323Sed // Now that we have full information about ranges, loop over the ranges and 431193323Sed // emit memset's for anything big enough to be worthwhile. 432218893Sdim Instruction *AMemSet = 0; 433193323Sed for (MemsetRanges::const_iterator I = Ranges.begin(), E = Ranges.end(); 434193323Sed I != E; ++I) { 435193323Sed const MemsetRange &Range = *I; 436218893Sdim 437193323Sed if (Range.TheStores.size() == 1) continue; 438193323Sed 439193323Sed // If it is profitable to lower this range to memset, do so now. 440198090Srdivacky if (!Range.isProfitableToUseMemset(*TD)) 441193323Sed continue; 442193323Sed 443218893Sdim // Otherwise, we do want to transform this! Create a new memset. 444193323Sed // Get the starting pointer of the block. 445193323Sed StartPtr = Range.StartPtr; 446218893Sdim 447206274Srdivacky // Determine alignment 448206274Srdivacky unsigned Alignment = Range.Alignment; 449206274Srdivacky if (Alignment == 0) { 450206274Srdivacky const Type *EltType = 451218893Sdim cast<PointerType>(StartPtr->getType())->getElementType(); 452206274Srdivacky Alignment = TD->getABITypeAlignment(EltType); 453206274Srdivacky } 454218893Sdim 455218893Sdim AMemSet = 456218893Sdim Builder.CreateMemSet(StartPtr, ByteVal, Range.End-Range.Start, Alignment); 457218893Sdim 458202375Srdivacky DEBUG(dbgs() << "Replace stores:\n"; 459193323Sed for (unsigned i = 0, e = Range.TheStores.size(); i != e; ++i) 460218893Sdim dbgs() << *Range.TheStores[i] << '\n'; 461218893Sdim dbgs() << "With: " << *AMemSet << '\n'); 462218893Sdim 463193323Sed // Zap all the stores. 464218893Sdim for (SmallVector<Instruction*, 16>::const_iterator 465198090Srdivacky SI = Range.TheStores.begin(), 466218893Sdim SE = Range.TheStores.end(); SI != SE; ++SI) { 467218893Sdim MD->removeInstruction(*SI); 468193323Sed (*SI)->eraseFromParent(); 469218893Sdim } 470193323Sed ++NumMemSetInfer; 471193323Sed } 472193323Sed 473218893Sdim return AMemSet; 474193323Sed} 475193323Sed 476193323Sed 477218893Sdimbool MemCpyOpt::processStore(StoreInst *SI, BasicBlock::iterator &BBI) { 478218893Sdim if (SI->isVolatile()) return false; 479218893Sdim 480218893Sdim if (TD == 0) return false; 481218893Sdim 482218893Sdim // Detect cases where we're performing call slot forwarding, but 483218893Sdim // happen to be using a load-store pair to implement it, rather than 484218893Sdim // a memcpy. 485218893Sdim if (LoadInst *LI = dyn_cast<LoadInst>(SI->getOperand(0))) { 486218893Sdim if (!LI->isVolatile() && LI->hasOneUse()) { 487218893Sdim MemDepResult dep = MD->getDependency(LI); 488218893Sdim CallInst *C = 0; 489218893Sdim if (dep.isClobber() && !isa<MemCpyInst>(dep.getInst())) 490218893Sdim C = dyn_cast<CallInst>(dep.getInst()); 491218893Sdim 492218893Sdim if (C) { 493218893Sdim bool changed = performCallSlotOptzn(LI, 494218893Sdim SI->getPointerOperand()->stripPointerCasts(), 495218893Sdim LI->getPointerOperand()->stripPointerCasts(), 496218893Sdim TD->getTypeStoreSize(SI->getOperand(0)->getType()), C); 497218893Sdim if (changed) { 498218893Sdim MD->removeInstruction(SI); 499218893Sdim SI->eraseFromParent(); 500218893Sdim MD->removeInstruction(LI); 501218893Sdim LI->eraseFromParent(); 502218893Sdim ++NumMemCpyInstr; 503218893Sdim return true; 504218893Sdim } 505218893Sdim } 506218893Sdim } 507218893Sdim } 508218893Sdim 509218893Sdim // There are two cases that are interesting for this code to handle: memcpy 510218893Sdim // and memset. Right now we only handle memset. 511218893Sdim 512218893Sdim // Ensure that the value being stored is something that can be memset'able a 513218893Sdim // byte at a time like "0" or "-1" or any width, as well as things like 514218893Sdim // 0xA0A0A0A0 and 0.0. 515218893Sdim if (Value *ByteVal = isBytewiseValue(SI->getOperand(0))) 516218893Sdim if (Instruction *I = tryMergingIntoMemset(SI, SI->getPointerOperand(), 517218893Sdim ByteVal)) { 518218893Sdim BBI = I; // Don't invalidate iterator. 519218893Sdim return true; 520218893Sdim } 521218893Sdim 522218893Sdim return false; 523218893Sdim} 524218893Sdim 525218893Sdimbool MemCpyOpt::processMemSet(MemSetInst *MSI, BasicBlock::iterator &BBI) { 526218893Sdim // See if there is another memset or store neighboring this memset which 527218893Sdim // allows us to widen out the memset to do a single larger store. 528218893Sdim if (isa<ConstantInt>(MSI->getLength()) && !MSI->isVolatile()) 529218893Sdim if (Instruction *I = tryMergingIntoMemset(MSI, MSI->getDest(), 530218893Sdim MSI->getValue())) { 531218893Sdim BBI = I; // Don't invalidate iterator. 532218893Sdim return true; 533218893Sdim } 534218893Sdim return false; 535218893Sdim} 536218893Sdim 537218893Sdim 538193323Sed/// performCallSlotOptzn - takes a memcpy and a call that it depends on, 539193323Sed/// and checks for the possibility of a call slot optimization by having 540193323Sed/// the call write its result directly into the destination of the memcpy. 541218893Sdimbool MemCpyOpt::performCallSlotOptzn(Instruction *cpy, 542218893Sdim Value *cpyDest, Value *cpySrc, 543218893Sdim uint64_t cpyLen, CallInst *C) { 544193323Sed // The general transformation to keep in mind is 545193323Sed // 546193323Sed // call @func(..., src, ...) 547193323Sed // memcpy(dest, src, ...) 548193323Sed // 549193323Sed // -> 550193323Sed // 551193323Sed // memcpy(dest, src, ...) 552193323Sed // call @func(..., dest, ...) 553193323Sed // 554193323Sed // Since moving the memcpy is technically awkward, we additionally check that 555193323Sed // src only holds uninitialized values at the moment of the call, meaning that 556193323Sed // the memcpy can be discarded rather than moved. 557193323Sed 558193323Sed // Deliberately get the source and destination with bitcasts stripped away, 559193323Sed // because we'll need to do type comparisons based on the underlying type. 560212904Sdim CallSite CS(C); 561193323Sed 562193323Sed // Require that src be an alloca. This simplifies the reasoning considerably. 563198090Srdivacky AllocaInst *srcAlloca = dyn_cast<AllocaInst>(cpySrc); 564193323Sed if (!srcAlloca) 565193323Sed return false; 566193323Sed 567193323Sed // Check that all of src is copied to dest. 568218893Sdim if (TD == 0) return false; 569193323Sed 570198090Srdivacky ConstantInt *srcArraySize = dyn_cast<ConstantInt>(srcAlloca->getArraySize()); 571193323Sed if (!srcArraySize) 572193323Sed return false; 573193323Sed 574198090Srdivacky uint64_t srcSize = TD->getTypeAllocSize(srcAlloca->getAllocatedType()) * 575193323Sed srcArraySize->getZExtValue(); 576193323Sed 577218893Sdim if (cpyLen < srcSize) 578193323Sed return false; 579193323Sed 580193323Sed // Check that accessing the first srcSize bytes of dest will not cause a 581193323Sed // trap. Otherwise the transform is invalid since it might cause a trap 582193323Sed // to occur earlier than it otherwise would. 583198090Srdivacky if (AllocaInst *A = dyn_cast<AllocaInst>(cpyDest)) { 584193323Sed // The destination is an alloca. Check it is larger than srcSize. 585198090Srdivacky ConstantInt *destArraySize = dyn_cast<ConstantInt>(A->getArraySize()); 586193323Sed if (!destArraySize) 587193323Sed return false; 588193323Sed 589198090Srdivacky uint64_t destSize = TD->getTypeAllocSize(A->getAllocatedType()) * 590193323Sed destArraySize->getZExtValue(); 591193323Sed 592193323Sed if (destSize < srcSize) 593193323Sed return false; 594198090Srdivacky } else if (Argument *A = dyn_cast<Argument>(cpyDest)) { 595193323Sed // If the destination is an sret parameter then only accesses that are 596193323Sed // outside of the returned struct type can trap. 597193323Sed if (!A->hasStructRetAttr()) 598193323Sed return false; 599193323Sed 600198090Srdivacky const Type *StructTy = cast<PointerType>(A->getType())->getElementType(); 601198090Srdivacky uint64_t destSize = TD->getTypeAllocSize(StructTy); 602193323Sed 603193323Sed if (destSize < srcSize) 604193323Sed return false; 605193323Sed } else { 606193323Sed return false; 607193323Sed } 608193323Sed 609193323Sed // Check that src is not accessed except via the call and the memcpy. This 610193323Sed // guarantees that it holds only undefined values when passed in (so the final 611193323Sed // memcpy can be dropped), that it is not read or written between the call and 612193323Sed // the memcpy, and that writing beyond the end of it is undefined. 613193323Sed SmallVector<User*, 8> srcUseList(srcAlloca->use_begin(), 614193323Sed srcAlloca->use_end()); 615193323Sed while (!srcUseList.empty()) { 616202375Srdivacky User *UI = srcUseList.pop_back_val(); 617193323Sed 618193323Sed if (isa<BitCastInst>(UI)) { 619193323Sed for (User::use_iterator I = UI->use_begin(), E = UI->use_end(); 620193323Sed I != E; ++I) 621193323Sed srcUseList.push_back(*I); 622198090Srdivacky } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(UI)) { 623193323Sed if (G->hasAllZeroIndices()) 624193323Sed for (User::use_iterator I = UI->use_begin(), E = UI->use_end(); 625193323Sed I != E; ++I) 626193323Sed srcUseList.push_back(*I); 627193323Sed else 628193323Sed return false; 629193323Sed } else if (UI != C && UI != cpy) { 630193323Sed return false; 631193323Sed } 632193323Sed } 633193323Sed 634193323Sed // Since we're changing the parameter to the callsite, we need to make sure 635193323Sed // that what would be the new parameter dominates the callsite. 636198090Srdivacky DominatorTree &DT = getAnalysis<DominatorTree>(); 637198090Srdivacky if (Instruction *cpyDestInst = dyn_cast<Instruction>(cpyDest)) 638193323Sed if (!DT.dominates(cpyDestInst, C)) 639193323Sed return false; 640193323Sed 641193323Sed // In addition to knowing that the call does not access src in some 642193323Sed // unexpected manner, for example via a global, which we deduce from 643193323Sed // the use analysis, we also need to know that it does not sneakily 644193323Sed // access dest. We rely on AA to figure this out for us. 645198090Srdivacky AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 646218893Sdim if (AA.getModRefInfo(C, cpyDest, srcSize) != AliasAnalysis::NoModRef) 647193323Sed return false; 648193323Sed 649193323Sed // All the checks have passed, so do the transformation. 650193323Sed bool changedArgument = false; 651193323Sed for (unsigned i = 0; i < CS.arg_size(); ++i) 652193323Sed if (CS.getArgument(i)->stripPointerCasts() == cpySrc) { 653193323Sed if (cpySrc->getType() != cpyDest->getType()) 654193323Sed cpyDest = CastInst::CreatePointerCast(cpyDest, cpySrc->getType(), 655193323Sed cpyDest->getName(), C); 656193323Sed changedArgument = true; 657198090Srdivacky if (CS.getArgument(i)->getType() == cpyDest->getType()) 658198090Srdivacky CS.setArgument(i, cpyDest); 659198090Srdivacky else 660193323Sed CS.setArgument(i, CastInst::CreatePointerCast(cpyDest, 661198090Srdivacky CS.getArgument(i)->getType(), cpyDest->getName(), C)); 662193323Sed } 663193323Sed 664193323Sed if (!changedArgument) 665193323Sed return false; 666193323Sed 667193323Sed // Drop any cached information about the call, because we may have changed 668193323Sed // its dependence information by changing its parameter. 669218893Sdim MD->removeInstruction(C); 670193323Sed 671218893Sdim // Remove the memcpy. 672218893Sdim MD->removeInstruction(cpy); 673210299Sed ++NumMemCpyInstr; 674193323Sed 675193323Sed return true; 676193323Sed} 677193323Sed 678218893Sdim/// processMemCpyMemCpyDependence - We've found that the (upward scanning) 679218893Sdim/// memory dependence of memcpy 'M' is the memcpy 'MDep'. Try to simplify M to 680218893Sdim/// copy from MDep's input if we can. MSize is the size of M's copy. 681218893Sdim/// 682218893Sdimbool MemCpyOpt::processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep, 683218893Sdim uint64_t MSize) { 684218893Sdim // We can only transforms memcpy's where the dest of one is the source of the 685218893Sdim // other. 686218893Sdim if (M->getSource() != MDep->getDest() || MDep->isVolatile()) 687193323Sed return false; 688193323Sed 689218893Sdim // If dep instruction is reading from our current input, then it is a noop 690218893Sdim // transfer and substituting the input won't change this instruction. Just 691218893Sdim // ignore the input and let someone else zap MDep. This handles cases like: 692218893Sdim // memcpy(a <- a) 693218893Sdim // memcpy(b <- a) 694218893Sdim if (M->getSource() == MDep->getSource()) 695193323Sed return false; 696193323Sed 697221345Sdim // Second, the length of the memcpy's must be the same, or the preceding one 698193323Sed // must be larger than the following one. 699218893Sdim ConstantInt *MDepLen = dyn_cast<ConstantInt>(MDep->getLength()); 700218893Sdim ConstantInt *MLen = dyn_cast<ConstantInt>(M->getLength()); 701218893Sdim if (!MDepLen || !MLen || MDepLen->getZExtValue() < MLen->getZExtValue()) 702193323Sed return false; 703193323Sed 704198090Srdivacky AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 705218893Sdim 706218893Sdim // Verify that the copied-from memory doesn't change in between the two 707218893Sdim // transfers. For example, in: 708218893Sdim // memcpy(a <- b) 709218893Sdim // *b = 42; 710218893Sdim // memcpy(c <- a) 711218893Sdim // It would be invalid to transform the second memcpy into memcpy(c <- b). 712218893Sdim // 713218893Sdim // TODO: If the code between M and MDep is transparent to the destination "c", 714218893Sdim // then we could still perform the xform by moving M up to the first memcpy. 715218893Sdim // 716218893Sdim // NOTE: This is conservative, it will stop on any read from the source loc, 717218893Sdim // not just the defining memcpy. 718218893Sdim MemDepResult SourceDep = 719218893Sdim MD->getPointerDependencyFrom(AA.getLocationForSource(MDep), 720218893Sdim false, M, M->getParent()); 721218893Sdim if (!SourceDep.isClobber() || SourceDep.getInst() != MDep) 722193323Sed return false; 723193323Sed 724218893Sdim // If the dest of the second might alias the source of the first, then the 725218893Sdim // source and dest might overlap. We still want to eliminate the intermediate 726218893Sdim // value, but we have to generate a memmove instead of memcpy. 727218893Sdim bool UseMemMove = false; 728218893Sdim if (!AA.isNoAlias(AA.getLocationForDest(M), AA.getLocationForSource(MDep))) 729218893Sdim UseMemMove = true; 730193323Sed 731218893Sdim // If all checks passed, then we can transform M. 732193323Sed 733218893Sdim // Make sure to use the lesser of the alignment of the source and the dest 734218893Sdim // since we're changing where we're reading from, but don't want to increase 735218893Sdim // the alignment past what can be read from or written to. 736218893Sdim // TODO: Is this worth it if we're creating a less aligned memcpy? For 737218893Sdim // example we could be moving from movaps -> movq on x86. 738218893Sdim unsigned Align = std::min(MDep->getAlignment(), M->getAlignment()); 739193323Sed 740218893Sdim IRBuilder<> Builder(M); 741218893Sdim if (UseMemMove) 742218893Sdim Builder.CreateMemMove(M->getRawDest(), MDep->getRawSource(), M->getLength(), 743218893Sdim Align, M->isVolatile()); 744218893Sdim else 745218893Sdim Builder.CreateMemCpy(M->getRawDest(), MDep->getRawSource(), M->getLength(), 746218893Sdim Align, M->isVolatile()); 747218893Sdim 748218893Sdim // Remove the instruction we're replacing. 749218893Sdim MD->removeInstruction(M); 750218893Sdim M->eraseFromParent(); 751218893Sdim ++NumMemCpyInstr; 752218893Sdim return true; 753218893Sdim} 754218893Sdim 755218893Sdim 756218893Sdim/// processMemCpy - perform simplification of memcpy's. If we have memcpy A 757218893Sdim/// which copies X to Y, and memcpy B which copies Y to Z, then we can rewrite 758218893Sdim/// B to be a memcpy from X to Z (or potentially a memmove, depending on 759218893Sdim/// circumstances). This allows later passes to remove the first memcpy 760218893Sdim/// altogether. 761218893Sdimbool MemCpyOpt::processMemCpy(MemCpyInst *M) { 762218893Sdim // We can only optimize statically-sized memcpy's that are non-volatile. 763218893Sdim ConstantInt *CopySize = dyn_cast<ConstantInt>(M->getLength()); 764218893Sdim if (CopySize == 0 || M->isVolatile()) return false; 765218893Sdim 766218893Sdim // If the source and destination of the memcpy are the same, then zap it. 767218893Sdim if (M->getSource() == M->getDest()) { 768218893Sdim MD->removeInstruction(M); 769193323Sed M->eraseFromParent(); 770218893Sdim return false; 771193323Sed } 772218893Sdim 773218893Sdim // If copying from a constant, try to turn the memcpy into a memset. 774218893Sdim if (GlobalVariable *GV = dyn_cast<GlobalVariable>(M->getSource())) 775218893Sdim if (GV->isConstant() && GV->hasDefinitiveInitializer()) 776218893Sdim if (Value *ByteVal = isBytewiseValue(GV->getInitializer())) { 777218893Sdim IRBuilder<> Builder(M); 778218893Sdim Builder.CreateMemSet(M->getRawDest(), ByteVal, CopySize, 779218893Sdim M->getAlignment(), false); 780218893Sdim MD->removeInstruction(M); 781218893Sdim M->eraseFromParent(); 782218893Sdim ++NumCpyToSet; 783218893Sdim return true; 784218893Sdim } 785218893Sdim 786218893Sdim // The are two possible optimizations we can do for memcpy: 787218893Sdim // a) memcpy-memcpy xform which exposes redundance for DSE. 788218893Sdim // b) call-memcpy xform for return slot optimization. 789218893Sdim MemDepResult DepInfo = MD->getDependency(M); 790218893Sdim if (!DepInfo.isClobber()) 791218893Sdim return false; 792193323Sed 793218893Sdim if (MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst())) 794218893Sdim return processMemCpyMemCpyDependence(M, MDep, CopySize->getZExtValue()); 795218893Sdim 796218893Sdim if (CallInst *C = dyn_cast<CallInst>(DepInfo.getInst())) { 797218893Sdim if (performCallSlotOptzn(M, M->getDest(), M->getSource(), 798218893Sdim CopySize->getZExtValue(), C)) { 799218893Sdim MD->removeInstruction(M); 800218893Sdim M->eraseFromParent(); 801218893Sdim return true; 802218893Sdim } 803218893Sdim } 804218893Sdim 805193323Sed return false; 806193323Sed} 807193323Sed 808198090Srdivacky/// processMemMove - Transforms memmove calls to memcpy calls when the src/dst 809198090Srdivacky/// are guaranteed not to alias. 810198090Srdivackybool MemCpyOpt::processMemMove(MemMoveInst *M) { 811198090Srdivacky AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 812198090Srdivacky 813221345Sdim if (!TLI->has(LibFunc::memmove)) 814221345Sdim return false; 815221345Sdim 816198090Srdivacky // See if the pointers alias. 817218893Sdim if (!AA.isNoAlias(AA.getLocationForDest(M), AA.getLocationForSource(M))) 818198090Srdivacky return false; 819193323Sed 820202375Srdivacky DEBUG(dbgs() << "MemCpyOpt: Optimizing memmove -> memcpy: " << *M << "\n"); 821193323Sed 822198090Srdivacky // If not, then we know we can transform this. 823198090Srdivacky Module *Mod = M->getParent()->getParent()->getParent(); 824206274Srdivacky const Type *ArgTys[3] = { M->getRawDest()->getType(), 825206274Srdivacky M->getRawSource()->getType(), 826206274Srdivacky M->getLength()->getType() }; 827212904Sdim M->setCalledFunction(Intrinsic::getDeclaration(Mod, Intrinsic::memcpy, 828212904Sdim ArgTys, 3)); 829198090Srdivacky 830198090Srdivacky // MemDep may have over conservative information about this instruction, just 831198090Srdivacky // conservatively flush it from the cache. 832218893Sdim MD->removeInstruction(M); 833198090Srdivacky 834198090Srdivacky ++NumMoveToCpy; 835198090Srdivacky return true; 836193323Sed} 837198090Srdivacky 838218893Sdim/// processByValArgument - This is called on every byval argument in call sites. 839218893Sdimbool MemCpyOpt::processByValArgument(CallSite CS, unsigned ArgNo) { 840218893Sdim if (TD == 0) return false; 841193323Sed 842218893Sdim // Find out what feeds this byval argument. 843218893Sdim Value *ByValArg = CS.getArgument(ArgNo); 844218893Sdim const Type *ByValTy =cast<PointerType>(ByValArg->getType())->getElementType(); 845218893Sdim uint64_t ByValSize = TD->getTypeAllocSize(ByValTy); 846218893Sdim MemDepResult DepInfo = 847218893Sdim MD->getPointerDependencyFrom(AliasAnalysis::Location(ByValArg, ByValSize), 848218893Sdim true, CS.getInstruction(), 849218893Sdim CS.getInstruction()->getParent()); 850218893Sdim if (!DepInfo.isClobber()) 851218893Sdim return false; 852218893Sdim 853218893Sdim // If the byval argument isn't fed by a memcpy, ignore it. If it is fed by 854218893Sdim // a memcpy, see if we can byval from the source of the memcpy instead of the 855218893Sdim // result. 856218893Sdim MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst()); 857218893Sdim if (MDep == 0 || MDep->isVolatile() || 858218893Sdim ByValArg->stripPointerCasts() != MDep->getDest()) 859218893Sdim return false; 860218893Sdim 861218893Sdim // The length of the memcpy must be larger or equal to the size of the byval. 862218893Sdim ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength()); 863218893Sdim if (C1 == 0 || C1->getValue().getZExtValue() < ByValSize) 864218893Sdim return false; 865218893Sdim 866218893Sdim // Get the alignment of the byval. If it is greater than the memcpy, then we 867218893Sdim // can't do the substitution. If the call doesn't specify the alignment, then 868218893Sdim // it is some target specific value that we can't know. 869218893Sdim unsigned ByValAlign = CS.getParamAlignment(ArgNo+1); 870218893Sdim if (ByValAlign == 0 || MDep->getAlignment() < ByValAlign) 871218893Sdim return false; 872218893Sdim 873218893Sdim // Verify that the copied-from memory doesn't change in between the memcpy and 874218893Sdim // the byval call. 875218893Sdim // memcpy(a <- b) 876218893Sdim // *b = 42; 877218893Sdim // foo(*a) 878218893Sdim // It would be invalid to transform the second memcpy into foo(*b). 879218893Sdim // 880218893Sdim // NOTE: This is conservative, it will stop on any read from the source loc, 881218893Sdim // not just the defining memcpy. 882218893Sdim MemDepResult SourceDep = 883218893Sdim MD->getPointerDependencyFrom(AliasAnalysis::getLocationForSource(MDep), 884218893Sdim false, CS.getInstruction(), MDep->getParent()); 885218893Sdim if (!SourceDep.isClobber() || SourceDep.getInst() != MDep) 886218893Sdim return false; 887218893Sdim 888218893Sdim Value *TmpCast = MDep->getSource(); 889218893Sdim if (MDep->getSource()->getType() != ByValArg->getType()) 890218893Sdim TmpCast = new BitCastInst(MDep->getSource(), ByValArg->getType(), 891218893Sdim "tmpcast", CS.getInstruction()); 892218893Sdim 893218893Sdim DEBUG(dbgs() << "MemCpyOpt: Forwarding memcpy to byval:\n" 894218893Sdim << " " << *MDep << "\n" 895218893Sdim << " " << *CS.getInstruction() << "\n"); 896218893Sdim 897218893Sdim // Otherwise we're good! Update the byval argument. 898218893Sdim CS.setArgument(ArgNo, TmpCast); 899218893Sdim ++NumMemCpyInstr; 900218893Sdim return true; 901218893Sdim} 902218893Sdim 903218893Sdim/// iterateOnFunction - Executes one iteration of MemCpyOpt. 904193323Sedbool MemCpyOpt::iterateOnFunction(Function &F) { 905198090Srdivacky bool MadeChange = false; 906193323Sed 907198090Srdivacky // Walk all instruction in the function. 908193323Sed for (Function::iterator BB = F.begin(), BBE = F.end(); BB != BBE; ++BB) { 909218893Sdim for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE;) { 910198090Srdivacky // Avoid invalidating the iterator. 911198090Srdivacky Instruction *I = BI++; 912193323Sed 913218893Sdim bool RepeatInstruction = false; 914218893Sdim 915193323Sed if (StoreInst *SI = dyn_cast<StoreInst>(I)) 916198090Srdivacky MadeChange |= processStore(SI, BI); 917218893Sdim else if (MemSetInst *M = dyn_cast<MemSetInst>(I)) 918218893Sdim RepeatInstruction = processMemSet(M, BI); 919198090Srdivacky else if (MemCpyInst *M = dyn_cast<MemCpyInst>(I)) 920218893Sdim RepeatInstruction = processMemCpy(M); 921218893Sdim else if (MemMoveInst *M = dyn_cast<MemMoveInst>(I)) 922218893Sdim RepeatInstruction = processMemMove(M); 923218893Sdim else if (CallSite CS = (Value*)I) { 924218893Sdim for (unsigned i = 0, e = CS.arg_size(); i != e; ++i) 925218893Sdim if (CS.paramHasAttr(i+1, Attribute::ByVal)) 926218893Sdim MadeChange |= processByValArgument(CS, i); 927193323Sed } 928218893Sdim 929218893Sdim // Reprocess the instruction if desired. 930218893Sdim if (RepeatInstruction) { 931218893Sdim if (BI != BB->begin()) --BI; 932218893Sdim MadeChange = true; 933218893Sdim } 934193323Sed } 935193323Sed } 936193323Sed 937198090Srdivacky return MadeChange; 938193323Sed} 939198090Srdivacky 940198090Srdivacky// MemCpyOpt::runOnFunction - This is the main transformation entry point for a 941198090Srdivacky// function. 942198090Srdivacky// 943198090Srdivackybool MemCpyOpt::runOnFunction(Function &F) { 944198090Srdivacky bool MadeChange = false; 945218893Sdim MD = &getAnalysis<MemoryDependenceAnalysis>(); 946218893Sdim TD = getAnalysisIfAvailable<TargetData>(); 947221345Sdim TLI = &getAnalysis<TargetLibraryInfo>(); 948221345Sdim 949221345Sdim // If we don't have at least memset and memcpy, there is little point of doing 950221345Sdim // anything here. These are required by a freestanding implementation, so if 951221345Sdim // even they are disabled, there is no point in trying hard. 952221345Sdim if (!TLI->has(LibFunc::memset) || !TLI->has(LibFunc::memcpy)) 953221345Sdim return false; 954221345Sdim 955198090Srdivacky while (1) { 956198090Srdivacky if (!iterateOnFunction(F)) 957198090Srdivacky break; 958198090Srdivacky MadeChange = true; 959198090Srdivacky } 960198090Srdivacky 961218893Sdim MD = 0; 962198090Srdivacky return MadeChange; 963198090Srdivacky} 964