LLVM API Documentation
00001 //===- MemCpyOptimizer.cpp - Optimize use of memcpy and friends -----------===// 00002 // 00003 // The LLVM Compiler Infrastructure 00004 // 00005 // This file is distributed under the University of Illinois Open Source 00006 // License. See LICENSE.TXT for details. 00007 // 00008 //===----------------------------------------------------------------------===// 00009 // 00010 // This pass performs various transformations related to eliminating memcpy 00011 // calls, or transforming sets of stores into memset's. 00012 // 00013 //===----------------------------------------------------------------------===// 00014 00015 #define DEBUG_TYPE "memcpyopt" 00016 #include "llvm/Transforms/Scalar.h" 00017 #include "llvm/IntrinsicInst.h" 00018 #include "llvm/Instructions.h" 00019 #include "llvm/LLVMContext.h" 00020 #include "llvm/ADT/SmallVector.h" 00021 #include "llvm/ADT/Statistic.h" 00022 #include "llvm/Analysis/Dominators.h" 00023 #include "llvm/Analysis/AliasAnalysis.h" 00024 #include "llvm/Analysis/MemoryDependenceAnalysis.h" 00025 #include "llvm/Support/Debug.h" 00026 #include "llvm/Support/GetElementPtrTypeIterator.h" 00027 #include "llvm/Target/TargetData.h" 00028 #include <list> 00029 using namespace llvm; 00030 00031 STATISTIC(NumMemCpyInstr, "Number of memcpy instructions deleted"); 00032 STATISTIC(NumMemSetInfer, "Number of memsets inferred"); 00033 00034 /// isBytewiseValue - If the specified value can be set by repeating the same 00035 /// byte in memory, return the i8 value that it is represented with. This is 00036 /// true for all i8 values obviously, but is also true for i32 0, i32 -1, 00037 /// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated 00038 /// byte store (e.g. i16 0x1234), return null. 00039 static Value *isBytewiseValue(Value *V, LLVMContext& Context) { 00040 // All byte-wide stores are splatable, even of arbitrary variables. 00041 if (V->getType() == Type::getInt8Ty(Context)) return V; 00042 00043 // Constant float and double values can be handled as integer values if the 00044 // corresponding integer value is "byteable". An important case is 0.0. 00045 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) { 00046 if (CFP->getType() == Type::getFloatTy(Context)) 00047 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(Context)); 00048 if (CFP->getType() == Type::getDoubleTy(Context)) 00049 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(Context)); 00050 // Don't handle long double formats, which have strange constraints. 00051 } 00052 00053 // We can handle constant integers that are power of two in size and a 00054 // multiple of 8 bits. 00055 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 00056 unsigned Width = CI->getBitWidth(); 00057 if (isPowerOf2_32(Width) && Width > 8) { 00058 // We can handle this value if the recursive binary decomposition is the 00059 // same at all levels. 00060 APInt Val = CI->getValue(); 00061 APInt Val2; 00062 while (Val.getBitWidth() != 8) { 00063 unsigned NextWidth = Val.getBitWidth()/2; 00064 Val2 = Val.lshr(NextWidth); 00065 Val2.trunc(Val.getBitWidth()/2); 00066 Val.trunc(Val.getBitWidth()/2); 00067 00068 // If the top/bottom halves aren't the same, reject it. 00069 if (Val != Val2) 00070 return 0; 00071 } 00072 return ConstantInt::get(Context, Val); 00073 } 00074 } 00075 00076 // Conceptually, we could handle things like: 00077 // %a = zext i8 %X to i16 00078 // %b = shl i16 %a, 8 00079 // %c = or i16 %a, %b 00080 // but until there is an example that actually needs this, it doesn't seem 00081 // worth worrying about. 00082 return 0; 00083 } 00084 00085 static int64_t GetOffsetFromIndex(const GetElementPtrInst *GEP, unsigned Idx, 00086 bool &VariableIdxFound, TargetData &TD) { 00087 // Skip over the first indices. 00088 gep_type_iterator GTI = gep_type_begin(GEP); 00089 for (unsigned i = 1; i != Idx; ++i, ++GTI) 00090 /*skip along*/; 00091 00092 // Compute the offset implied by the rest of the indices. 00093 int64_t Offset = 0; 00094 for (unsigned i = Idx, e = GEP->getNumOperands(); i != e; ++i, ++GTI) { 00095 ConstantInt *OpC = dyn_cast<ConstantInt>(GEP->getOperand(i)); 00096 if (OpC == 0) 00097 return VariableIdxFound = true; 00098 if (OpC->isZero()) continue; // No offset. 00099 00100 // Handle struct indices, which add their field offset to the pointer. 00101 if (const StructType *STy = dyn_cast<StructType>(*GTI)) { 00102 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue()); 00103 continue; 00104 } 00105 00106 // Otherwise, we have a sequential type like an array or vector. Multiply 00107 // the index by the ElementSize. 00108 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()); 00109 Offset += Size*OpC->getSExtValue(); 00110 } 00111 00112 return Offset; 00113 } 00114 00115 /// IsPointerOffset - Return true if Ptr1 is provably equal to Ptr2 plus a 00116 /// constant offset, and return that constant offset. For example, Ptr1 might 00117 /// be &A[42], and Ptr2 might be &A[40]. In this case offset would be -8. 00118 static bool IsPointerOffset(Value *Ptr1, Value *Ptr2, int64_t &Offset, 00119 TargetData &TD) { 00120 // Right now we handle the case when Ptr1/Ptr2 are both GEPs with an identical 00121 // base. After that base, they may have some number of common (and 00122 // potentially variable) indices. After that they handle some constant 00123 // offset, which determines their offset from each other. At this point, we 00124 // handle no other case. 00125 GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(Ptr1); 00126 GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(Ptr2); 00127 if (!GEP1 || !GEP2 || GEP1->getOperand(0) != GEP2->getOperand(0)) 00128 return false; 00129 00130 // Skip any common indices and track the GEP types. 00131 unsigned Idx = 1; 00132 for (; Idx != GEP1->getNumOperands() && Idx != GEP2->getNumOperands(); ++Idx) 00133 if (GEP1->getOperand(Idx) != GEP2->getOperand(Idx)) 00134 break; 00135 00136 bool VariableIdxFound = false; 00137 int64_t Offset1 = GetOffsetFromIndex(GEP1, Idx, VariableIdxFound, TD); 00138 int64_t Offset2 = GetOffsetFromIndex(GEP2, Idx, VariableIdxFound, TD); 00139 if (VariableIdxFound) return false; 00140 00141 Offset = Offset2-Offset1; 00142 return true; 00143 } 00144 00145 00146 /// MemsetRange - Represents a range of memset'd bytes with the ByteVal value. 00147 /// This allows us to analyze stores like: 00148 /// store 0 -> P+1 00149 /// store 0 -> P+0 00150 /// store 0 -> P+3 00151 /// store 0 -> P+2 00152 /// which sometimes happens with stores to arrays of structs etc. When we see 00153 /// the first store, we make a range [1, 2). The second store extends the range 00154 /// to [0, 2). The third makes a new range [2, 3). The fourth store joins the 00155 /// two ranges into [0, 3) which is memset'able. 00156 namespace { 00157 struct MemsetRange { 00158 // Start/End - A semi range that describes the span that this range covers. 00159 // The range is closed at the start and open at the end: [Start, End). 00160 int64_t Start, End; 00161 00162 /// StartPtr - The getelementptr instruction that points to the start of the 00163 /// range. 00164 Value *StartPtr; 00165 00166 /// Alignment - The known alignment of the first store. 00167 unsigned Alignment; 00168 00169 /// TheStores - The actual stores that make up this range. 00170 SmallVector<StoreInst*, 16> TheStores; 00171 00172 bool isProfitableToUseMemset(const TargetData &TD) const; 00173 00174 }; 00175 } // end anon namespace 00176 00177 bool MemsetRange::isProfitableToUseMemset(const TargetData &TD) const { 00178 // If we found more than 8 stores to merge or 64 bytes, use memset. 00179 if (TheStores.size() >= 8 || End-Start >= 64) return true; 00180 00181 // Assume that the code generator is capable of merging pairs of stores 00182 // together if it wants to. 00183 if (TheStores.size() <= 2) return false; 00184 00185 // If we have fewer than 8 stores, it can still be worthwhile to do this. 00186 // For example, merging 4 i8 stores into an i32 store is useful almost always. 00187 // However, merging 2 32-bit stores isn't useful on a 32-bit architecture (the 00188 // memset will be split into 2 32-bit stores anyway) and doing so can 00189 // pessimize the llvm optimizer. 00190 // 00191 // Since we don't have perfect knowledge here, make some assumptions: assume 00192 // the maximum GPR width is the same size as the pointer size and assume that 00193 // this width can be stored. If so, check to see whether we will end up 00194 // actually reducing the number of stores used. 00195 unsigned Bytes = unsigned(End-Start); 00196 unsigned NumPointerStores = Bytes/TD.getPointerSize(); 00197 00198 // Assume the remaining bytes if any are done a byte at a time. 00199 unsigned NumByteStores = Bytes - NumPointerStores*TD.getPointerSize(); 00200 00201 // If we will reduce the # stores (according to this heuristic), do the 00202 // transformation. This encourages merging 4 x i8 -> i32 and 2 x i16 -> i32 00203 // etc. 00204 return TheStores.size() > NumPointerStores+NumByteStores; 00205 } 00206 00207 00208 namespace { 00209 class MemsetRanges { 00210 /// Ranges - A sorted list of the memset ranges. We use std::list here 00211 /// because each element is relatively large and expensive to copy. 00212 std::list<MemsetRange> Ranges; 00213 typedef std::list<MemsetRange>::iterator range_iterator; 00214 TargetData &TD; 00215 public: 00216 MemsetRanges(TargetData &td) : TD(td) {} 00217 00218 typedef std::list<MemsetRange>::const_iterator const_iterator; 00219 const_iterator begin() const { return Ranges.begin(); } 00220 const_iterator end() const { return Ranges.end(); } 00221 bool empty() const { return Ranges.empty(); } 00222 00223 void addStore(int64_t OffsetFromFirst, StoreInst *SI); 00224 }; 00225 00226 } // end anon namespace 00227 00228 00229 /// addStore - Add a new store to the MemsetRanges data structure. This adds a 00230 /// new range for the specified store at the specified offset, merging into 00231 /// existing ranges as appropriate. 00232 void MemsetRanges::addStore(int64_t Start, StoreInst *SI) { 00233 int64_t End = Start+TD.getTypeStoreSize(SI->getOperand(0)->getType()); 00234 00235 // Do a linear search of the ranges to see if this can be joined and/or to 00236 // find the insertion point in the list. We keep the ranges sorted for 00237 // simplicity here. This is a linear search of a linked list, which is ugly, 00238 // however the number of ranges is limited, so this won't get crazy slow. 00239 range_iterator I = Ranges.begin(), E = Ranges.end(); 00240 00241 while (I != E && Start > I->End) 00242 ++I; 00243 00244 // We now know that I == E, in which case we didn't find anything to merge 00245 // with, or that Start <= I->End. If End < I->Start or I == E, then we need 00246 // to insert a new range. Handle this now. 00247 if (I == E || End < I->Start) { 00248 MemsetRange &R = *Ranges.insert(I, MemsetRange()); 00249 R.Start = Start; 00250 R.End = End; 00251 R.StartPtr = SI->getPointerOperand(); 00252 R.Alignment = SI->getAlignment(); 00253 R.TheStores.push_back(SI); 00254 return; 00255 } 00256 00257 // This store overlaps with I, add it. 00258 I->TheStores.push_back(SI); 00259 00260 // At this point, we may have an interval that completely contains our store. 00261 // If so, just add it to the interval and return. 00262 if (I->Start <= Start && I->End >= End) 00263 return; 00264 00265 // Now we know that Start <= I->End and End >= I->Start so the range overlaps 00266 // but is not entirely contained within the range. 00267 00268 // See if the range extends the start of the range. In this case, it couldn't 00269 // possibly cause it to join the prior range, because otherwise we would have 00270 // stopped on *it*. 00271 if (Start < I->Start) { 00272 I->Start = Start; 00273 I->StartPtr = SI->getPointerOperand(); 00274 I->Alignment = SI->getAlignment(); 00275 } 00276 00277 // Now we know that Start <= I->End and Start >= I->Start (so the startpoint 00278 // is in or right at the end of I), and that End >= I->Start. Extend I out to 00279 // End. 00280 if (End > I->End) { 00281 I->End = End; 00282 range_iterator NextI = I; 00283 while (++NextI != E && End >= NextI->Start) { 00284 // Merge the range in. 00285 I->TheStores.append(NextI->TheStores.begin(), NextI->TheStores.end()); 00286 if (NextI->End > I->End) 00287 I->End = NextI->End; 00288 Ranges.erase(NextI); 00289 NextI = I; 00290 } 00291 } 00292 } 00293 00294 //===----------------------------------------------------------------------===// 00295 // MemCpyOpt Pass 00296 //===----------------------------------------------------------------------===// 00297 00298 namespace { 00299 00300 class VISIBILITY_HIDDEN MemCpyOpt : public FunctionPass { 00301 bool runOnFunction(Function &F); 00302 public: 00303 static char ID; // Pass identification, replacement for typeid 00304 MemCpyOpt() : FunctionPass(&ID) {} 00305 00306 private: 00307 // This transformation requires dominator postdominator info 00308 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 00309 AU.setPreservesCFG(); 00310 AU.addRequired<DominatorTree>(); 00311 AU.addRequired<MemoryDependenceAnalysis>(); 00312 AU.addRequired<AliasAnalysis>(); 00313 AU.addPreserved<AliasAnalysis>(); 00314 AU.addPreserved<MemoryDependenceAnalysis>(); 00315 } 00316 00317 // Helper fuctions 00318 bool processStore(StoreInst *SI, BasicBlock::iterator& BBI); 00319 bool processMemCpy(MemCpyInst* M); 00320 bool performCallSlotOptzn(MemCpyInst* cpy, CallInst* C); 00321 bool iterateOnFunction(Function &F); 00322 }; 00323 00324 char MemCpyOpt::ID = 0; 00325 } 00326 00327 // createMemCpyOptPass - The public interface to this file... 00328 FunctionPass *llvm::createMemCpyOptPass() { return new MemCpyOpt(); } 00329 00330 static RegisterPass<MemCpyOpt> X("memcpyopt", 00331 "MemCpy Optimization"); 00332 00333 00334 00335 /// processStore - When GVN is scanning forward over instructions, we look for 00336 /// some other patterns to fold away. In particular, this looks for stores to 00337 /// neighboring locations of memory. If it sees enough consequtive ones 00338 /// (currently 4) it attempts to merge them together into a memcpy/memset. 00339 bool MemCpyOpt::processStore(StoreInst *SI, BasicBlock::iterator& BBI) { 00340 if (SI->isVolatile()) return false; 00341 00342 LLVMContext &Context = SI->getContext(); 00343 00344 // There are two cases that are interesting for this code to handle: memcpy 00345 // and memset. Right now we only handle memset. 00346 00347 // Ensure that the value being stored is something that can be memset'able a 00348 // byte at a time like "0" or "-1" or any width, as well as things like 00349 // 0xA0A0A0A0 and 0.0. 00350 Value *ByteVal = isBytewiseValue(SI->getOperand(0), Context); 00351 if (!ByteVal) 00352 return false; 00353 00354 TargetData *TD = getAnalysisIfAvailable<TargetData>(); 00355 if (!TD) return false; 00356 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 00357 Module *M = SI->getParent()->getParent()->getParent(); 00358 00359 // Okay, so we now have a single store that can be splatable. Scan to find 00360 // all subsequent stores of the same value to offset from the same pointer. 00361 // Join these together into ranges, so we can decide whether contiguous blocks 00362 // are stored. 00363 MemsetRanges Ranges(*TD); 00364 00365 Value *StartPtr = SI->getPointerOperand(); 00366 00367 BasicBlock::iterator BI = SI; 00368 for (++BI; !isa<TerminatorInst>(BI); ++BI) { 00369 if (isa<CallInst>(BI) || isa<InvokeInst>(BI)) { 00370 // If the call is readnone, ignore it, otherwise bail out. We don't even 00371 // allow readonly here because we don't want something like: 00372 // A[1] = 2; strlen(A); A[2] = 2; -> memcpy(A, ...); strlen(A). 00373 if (AA.getModRefBehavior(CallSite::get(BI)) == 00374 AliasAnalysis::DoesNotAccessMemory) 00375 continue; 00376 00377 // TODO: If this is a memset, try to join it in. 00378 00379 break; 00380 } else if (isa<VAArgInst>(BI) || isa<LoadInst>(BI)) 00381 break; 00382 00383 // If this is a non-store instruction it is fine, ignore it. 00384 StoreInst *NextStore = dyn_cast<StoreInst>(BI); 00385 if (NextStore == 0) continue; 00386 00387 // If this is a store, see if we can merge it in. 00388 if (NextStore->isVolatile()) break; 00389 00390 // Check to see if this stored value is of the same byte-splattable value. 00391 if (ByteVal != isBytewiseValue(NextStore->getOperand(0), Context)) 00392 break; 00393 00394 // Check to see if this store is to a constant offset from the start ptr. 00395 int64_t Offset; 00396 if (!IsPointerOffset(StartPtr, NextStore->getPointerOperand(), Offset, *TD)) 00397 break; 00398 00399 Ranges.addStore(Offset, NextStore); 00400 } 00401 00402 // If we have no ranges, then we just had a single store with nothing that 00403 // could be merged in. This is a very common case of course. 00404 if (Ranges.empty()) 00405 return false; 00406 00407 // If we had at least one store that could be merged in, add the starting 00408 // store as well. We try to avoid this unless there is at least something 00409 // interesting as a small compile-time optimization. 00410 Ranges.addStore(0, SI); 00411 00412 Function *MemSetF = 0; 00413 00414 // Now that we have full information about ranges, loop over the ranges and 00415 // emit memset's for anything big enough to be worthwhile. 00416 bool MadeChange = false; 00417 for (MemsetRanges::const_iterator I = Ranges.begin(), E = Ranges.end(); 00418 I != E; ++I) { 00419 const MemsetRange &Range = *I; 00420 00421 if (Range.TheStores.size() == 1) continue; 00422 00423 // If it is profitable to lower this range to memset, do so now. 00424 if (!Range.isProfitableToUseMemset(*TD)) 00425 continue; 00426 00427 // Otherwise, we do want to transform this! Create a new memset. We put 00428 // the memset right before the first instruction that isn't part of this 00429 // memset block. This ensure that the memset is dominated by any addressing 00430 // instruction needed by the start of the block. 00431 BasicBlock::iterator InsertPt = BI; 00432 00433 if (MemSetF == 0) { 00434 const Type *Ty = Type::getInt64Ty(Context); 00435 MemSetF = Intrinsic::getDeclaration(M, Intrinsic::memset, &Ty, 1); 00436 } 00437 00438 // Get the starting pointer of the block. 00439 StartPtr = Range.StartPtr; 00440 00441 // Cast the start ptr to be i8* as memset requires. 00442 const Type *i8Ptr = PointerType::getUnqual(Type::getInt8Ty(Context)); 00443 if (StartPtr->getType() != i8Ptr) 00444 StartPtr = new BitCastInst(StartPtr, i8Ptr, StartPtr->getName(), 00445 InsertPt); 00446 00447 Value *Ops[] = { 00448 StartPtr, ByteVal, // Start, value 00449 // size 00450 ConstantInt::get(Type::getInt64Ty(Context), Range.End-Range.Start), 00451 // align 00452 ConstantInt::get(Type::getInt32Ty(Context), Range.Alignment) 00453 }; 00454 Value *C = CallInst::Create(MemSetF, Ops, Ops+4, "", InsertPt); 00455 DEBUG(cerr << "Replace stores:\n"; 00456 for (unsigned i = 0, e = Range.TheStores.size(); i != e; ++i) 00457 cerr << *Range.TheStores[i]; 00458 cerr << "With: " << *C); C=C; 00459 00460 // Don't invalidate the iterator 00461 BBI = BI; 00462 00463 // Zap all the stores. 00464 for (SmallVector<StoreInst*, 16>::const_iterator 00465 SI = Range.TheStores.begin(), 00466 SE = Range.TheStores.end(); SI != SE; ++SI) 00467 (*SI)->eraseFromParent(); 00468 ++NumMemSetInfer; 00469 MadeChange = true; 00470 } 00471 00472 return MadeChange; 00473 } 00474 00475 00476 /// performCallSlotOptzn - takes a memcpy and a call that it depends on, 00477 /// and checks for the possibility of a call slot optimization by having 00478 /// the call write its result directly into the destination of the memcpy. 00479 bool MemCpyOpt::performCallSlotOptzn(MemCpyInst *cpy, CallInst *C) { 00480 // The general transformation to keep in mind is 00481 // 00482 // call @func(..., src, ...) 00483 // memcpy(dest, src, ...) 00484 // 00485 // -> 00486 // 00487 // memcpy(dest, src, ...) 00488 // call @func(..., dest, ...) 00489 // 00490 // Since moving the memcpy is technically awkward, we additionally check that 00491 // src only holds uninitialized values at the moment of the call, meaning that 00492 // the memcpy can be discarded rather than moved. 00493 00494 // Deliberately get the source and destination with bitcasts stripped away, 00495 // because we'll need to do type comparisons based on the underlying type. 00496 Value* cpyDest = cpy->getDest(); 00497 Value* cpySrc = cpy->getSource(); 00498 CallSite CS = CallSite::get(C); 00499 00500 // We need to be able to reason about the size of the memcpy, so we require 00501 // that it be a constant. 00502 ConstantInt* cpyLength = dyn_cast<ConstantInt>(cpy->getLength()); 00503 if (!cpyLength) 00504 return false; 00505 00506 // Require that src be an alloca. This simplifies the reasoning considerably. 00507 AllocaInst* srcAlloca = dyn_cast<AllocaInst>(cpySrc); 00508 if (!srcAlloca) 00509 return false; 00510 00511 // Check that all of src is copied to dest. 00512 TargetData* TD = getAnalysisIfAvailable<TargetData>(); 00513 if (!TD) return false; 00514 00515 ConstantInt* srcArraySize = dyn_cast<ConstantInt>(srcAlloca->getArraySize()); 00516 if (!srcArraySize) 00517 return false; 00518 00519 uint64_t srcSize = TD->getTypeAllocSize(srcAlloca->getAllocatedType()) * 00520 srcArraySize->getZExtValue(); 00521 00522 if (cpyLength->getZExtValue() < srcSize) 00523 return false; 00524 00525 // Check that accessing the first srcSize bytes of dest will not cause a 00526 // trap. Otherwise the transform is invalid since it might cause a trap 00527 // to occur earlier than it otherwise would. 00528 if (AllocaInst* A = dyn_cast<AllocaInst>(cpyDest)) { 00529 // The destination is an alloca. Check it is larger than srcSize. 00530 ConstantInt* destArraySize = dyn_cast<ConstantInt>(A->getArraySize()); 00531 if (!destArraySize) 00532 return false; 00533 00534 uint64_t destSize = TD->getTypeAllocSize(A->getAllocatedType()) * 00535 destArraySize->getZExtValue(); 00536 00537 if (destSize < srcSize) 00538 return false; 00539 } else if (Argument* A = dyn_cast<Argument>(cpyDest)) { 00540 // If the destination is an sret parameter then only accesses that are 00541 // outside of the returned struct type can trap. 00542 if (!A->hasStructRetAttr()) 00543 return false; 00544 00545 const Type* StructTy = cast<PointerType>(A->getType())->getElementType(); 00546 uint64_t destSize = TD->getTypeAllocSize(StructTy); 00547 00548 if (destSize < srcSize) 00549 return false; 00550 } else { 00551 return false; 00552 } 00553 00554 // Check that src is not accessed except via the call and the memcpy. This 00555 // guarantees that it holds only undefined values when passed in (so the final 00556 // memcpy can be dropped), that it is not read or written between the call and 00557 // the memcpy, and that writing beyond the end of it is undefined. 00558 SmallVector<User*, 8> srcUseList(srcAlloca->use_begin(), 00559 srcAlloca->use_end()); 00560 while (!srcUseList.empty()) { 00561 User* UI = srcUseList.back(); 00562 srcUseList.pop_back(); 00563 00564 if (isa<BitCastInst>(UI)) { 00565 for (User::use_iterator I = UI->use_begin(), E = UI->use_end(); 00566 I != E; ++I) 00567 srcUseList.push_back(*I); 00568 } else if (GetElementPtrInst* G = dyn_cast<GetElementPtrInst>(UI)) { 00569 if (G->hasAllZeroIndices()) 00570 for (User::use_iterator I = UI->use_begin(), E = UI->use_end(); 00571 I != E; ++I) 00572 srcUseList.push_back(*I); 00573 else 00574 return false; 00575 } else if (UI != C && UI != cpy) { 00576 return false; 00577 } 00578 } 00579 00580 // Since we're changing the parameter to the callsite, we need to make sure 00581 // that what would be the new parameter dominates the callsite. 00582 DominatorTree& DT = getAnalysis<DominatorTree>(); 00583 if (Instruction* cpyDestInst = dyn_cast<Instruction>(cpyDest)) 00584 if (!DT.dominates(cpyDestInst, C)) 00585 return false; 00586 00587 // In addition to knowing that the call does not access src in some 00588 // unexpected manner, for example via a global, which we deduce from 00589 // the use analysis, we also need to know that it does not sneakily 00590 // access dest. We rely on AA to figure this out for us. 00591 AliasAnalysis& AA = getAnalysis<AliasAnalysis>(); 00592 if (AA.getModRefInfo(C, cpy->getRawDest(), srcSize) != 00593 AliasAnalysis::NoModRef) 00594 return false; 00595 00596 // All the checks have passed, so do the transformation. 00597 bool changedArgument = false; 00598 for (unsigned i = 0; i < CS.arg_size(); ++i) 00599 if (CS.getArgument(i)->stripPointerCasts() == cpySrc) { 00600 if (cpySrc->getType() != cpyDest->getType()) 00601 cpyDest = CastInst::CreatePointerCast(cpyDest, cpySrc->getType(), 00602 cpyDest->getName(), C); 00603 changedArgument = true; 00604 if (CS.getArgument(i)->getType() != cpyDest->getType()) 00605 CS.setArgument(i, CastInst::CreatePointerCast(cpyDest, 00606 CS.getArgument(i)->getType(), cpyDest->getName(), C)); 00607 else 00608 CS.setArgument(i, cpyDest); 00609 } 00610 00611 if (!changedArgument) 00612 return false; 00613 00614 // Drop any cached information about the call, because we may have changed 00615 // its dependence information by changing its parameter. 00616 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>(); 00617 MD.removeInstruction(C); 00618 00619 // Remove the memcpy 00620 MD.removeInstruction(cpy); 00621 cpy->eraseFromParent(); 00622 NumMemCpyInstr++; 00623 00624 return true; 00625 } 00626 00627 /// processMemCpy - perform simplication of memcpy's. If we have memcpy A which 00628 /// copies X to Y, and memcpy B which copies Y to Z, then we can rewrite B to be 00629 /// a memcpy from X to Z (or potentially a memmove, depending on circumstances). 00630 /// This allows later passes to remove the first memcpy altogether. 00631 bool MemCpyOpt::processMemCpy(MemCpyInst* M) { 00632 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>(); 00633 00634 // The are two possible optimizations we can do for memcpy: 00635 // a) memcpy-memcpy xform which exposes redundance for DSE 00636 // b) call-memcpy xform for return slot optimization 00637 MemDepResult dep = MD.getDependency(M); 00638 if (!dep.isClobber()) 00639 return false; 00640 if (!isa<MemCpyInst>(dep.getInst())) { 00641 if (CallInst* C = dyn_cast<CallInst>(dep.getInst())) 00642 return performCallSlotOptzn(M, C); 00643 return false; 00644 } 00645 00646 MemCpyInst* MDep = cast<MemCpyInst>(dep.getInst()); 00647 00648 // We can only transforms memcpy's where the dest of one is the source of the 00649 // other 00650 if (M->getSource() != MDep->getDest()) 00651 return false; 00652 00653 // Second, the length of the memcpy's must be the same, or the preceeding one 00654 // must be larger than the following one. 00655 ConstantInt* C1 = dyn_cast<ConstantInt>(MDep->getLength()); 00656 ConstantInt* C2 = dyn_cast<ConstantInt>(M->getLength()); 00657 if (!C1 || !C2) 00658 return false; 00659 00660 uint64_t DepSize = C1->getValue().getZExtValue(); 00661 uint64_t CpySize = C2->getValue().getZExtValue(); 00662 00663 if (DepSize < CpySize) 00664 return false; 00665 00666 // Finally, we have to make sure that the dest of the second does not 00667 // alias the source of the first 00668 AliasAnalysis& AA = getAnalysis<AliasAnalysis>(); 00669 if (AA.alias(M->getRawDest(), CpySize, MDep->getRawSource(), DepSize) != 00670 AliasAnalysis::NoAlias) 00671 return false; 00672 else if (AA.alias(M->getRawDest(), CpySize, M->getRawSource(), CpySize) != 00673 AliasAnalysis::NoAlias) 00674 return false; 00675 else if (AA.alias(MDep->getRawDest(), DepSize, MDep->getRawSource(), DepSize) 00676 != AliasAnalysis::NoAlias) 00677 return false; 00678 00679 // If all checks passed, then we can transform these memcpy's 00680 const Type *Tys[1]; 00681 Tys[0] = M->getLength()->getType(); 00682 Function* MemCpyFun = Intrinsic::getDeclaration( 00683 M->getParent()->getParent()->getParent(), 00684 M->getIntrinsicID(), Tys, 1); 00685 00686 Value *Args[4] = { 00687 M->getRawDest(), MDep->getRawSource(), M->getLength(), M->getAlignmentCst() 00688 }; 00689 00690 CallInst* C = CallInst::Create(MemCpyFun, Args, Args+4, "", M); 00691 00692 00693 // If C and M don't interfere, then this is a valid transformation. If they 00694 // did, this would mean that the two sources overlap, which would be bad. 00695 if (MD.getDependency(C) == dep) { 00696 MD.removeInstruction(M); 00697 M->eraseFromParent(); 00698 NumMemCpyInstr++; 00699 return true; 00700 } 00701 00702 // Otherwise, there was no point in doing this, so we remove the call we 00703 // inserted and act like nothing happened. 00704 MD.removeInstruction(C); 00705 C->eraseFromParent(); 00706 return false; 00707 } 00708 00709 // MemCpyOpt::runOnFunction - This is the main transformation entry point for a 00710 // function. 00711 // 00712 bool MemCpyOpt::runOnFunction(Function& F) { 00713 00714 bool changed = false; 00715 bool shouldContinue = true; 00716 00717 while (shouldContinue) { 00718 shouldContinue = iterateOnFunction(F); 00719 changed |= shouldContinue; 00720 } 00721 00722 return changed; 00723 } 00724 00725 00726 // MemCpyOpt::iterateOnFunction - Executes one iteration of GVN 00727 bool MemCpyOpt::iterateOnFunction(Function &F) { 00728 bool changed_function = false; 00729 00730 // Walk all instruction in the function 00731 for (Function::iterator BB = F.begin(), BBE = F.end(); BB != BBE; ++BB) { 00732 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); 00733 BI != BE;) { 00734 // Avoid invalidating the iterator 00735 Instruction* I = BI++; 00736 00737 if (StoreInst *SI = dyn_cast<StoreInst>(I)) 00738 changed_function |= processStore(SI, BI); 00739 else if (MemCpyInst* M = dyn_cast<MemCpyInst>(I)) { 00740 changed_function |= processMemCpy(M); 00741 } 00742 } 00743 } 00744 00745 return changed_function; 00746 }