MemCpyOptimizer.cpp revision 05cd03b33559732f8ed55e5ff7554fd06d59eb6a
1//===- MemCpyOptimizer.cpp - Optimize use of memcpy and friends -----------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This pass performs various transformations related to eliminating memcpy 11// calls, or transforming sets of stores into memset's. 12// 13//===----------------------------------------------------------------------===// 14 15#define DEBUG_TYPE "memcpyopt" 16#include "llvm/Transforms/Scalar.h" 17#include "llvm/IntrinsicInst.h" 18#include "llvm/Instructions.h" 19#include "llvm/LLVMContext.h" 20#include "llvm/ADT/SmallVector.h" 21#include "llvm/ADT/Statistic.h" 22#include "llvm/Analysis/Dominators.h" 23#include "llvm/Analysis/AliasAnalysis.h" 24#include "llvm/Analysis/MemoryDependenceAnalysis.h" 25#include "llvm/Support/Debug.h" 26#include "llvm/Support/GetElementPtrTypeIterator.h" 27#include "llvm/Support/raw_ostream.h" 28#include "llvm/Target/TargetData.h" 29#include <list> 30using namespace llvm; 31 32STATISTIC(NumMemCpyInstr, "Number of memcpy instructions deleted"); 33STATISTIC(NumMemSetInfer, "Number of memsets inferred"); 34STATISTIC(NumMoveToCpy, "Number of memmoves converted to memcpy"); 35 36/// isBytewiseValue - If the specified value can be set by repeating the same 37/// byte in memory, return the i8 value that it is represented with. This is 38/// true for all i8 values obviously, but is also true for i32 0, i32 -1, 39/// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated 40/// byte store (e.g. i16 0x1234), return null. 41static Value *isBytewiseValue(Value *V, LLVMContext &Context) { 42 // All byte-wide stores are splatable, even of arbitrary variables. 43 if (V->getType() == Type::getInt8Ty(Context)) return V; 44 45 // Constant float and double values can be handled as integer values if the 46 // corresponding integer value is "byteable". An important case is 0.0. 47 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) { 48 if (CFP->getType() == Type::getFloatTy(Context)) 49 V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(Context)); 50 if (CFP->getType() == Type::getDoubleTy(Context)) 51 V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(Context)); 52 // Don't handle long double formats, which have strange constraints. 53 } 54 55 // We can handle constant integers that are power of two in size and a 56 // multiple of 8 bits. 57 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 58 unsigned Width = CI->getBitWidth(); 59 if (isPowerOf2_32(Width) && Width > 8) { 60 // We can handle this value if the recursive binary decomposition is the 61 // same at all levels. 62 APInt Val = CI->getValue(); 63 APInt Val2; 64 while (Val.getBitWidth() != 8) { 65 unsigned NextWidth = Val.getBitWidth()/2; 66 Val2 = Val.lshr(NextWidth); 67 Val2.trunc(Val.getBitWidth()/2); 68 Val.trunc(Val.getBitWidth()/2); 69 70 // If the top/bottom halves aren't the same, reject it. 71 if (Val != Val2) 72 return 0; 73 } 74 return ConstantInt::get(Context, Val); 75 } 76 } 77 78 // Conceptually, we could handle things like: 79 // %a = zext i8 %X to i16 80 // %b = shl i16 %a, 8 81 // %c = or i16 %a, %b 82 // but until there is an example that actually needs this, it doesn't seem 83 // worth worrying about. 84 return 0; 85} 86 87static int64_t GetOffsetFromIndex(const GetElementPtrInst *GEP, unsigned Idx, 88 bool &VariableIdxFound, TargetData &TD) { 89 // Skip over the first indices. 90 gep_type_iterator GTI = gep_type_begin(GEP); 91 for (unsigned i = 1; i != Idx; ++i, ++GTI) 92 /*skip along*/; 93 94 // Compute the offset implied by the rest of the indices. 95 int64_t Offset = 0; 96 for (unsigned i = Idx, e = GEP->getNumOperands(); i != e; ++i, ++GTI) { 97 ConstantInt *OpC = dyn_cast<ConstantInt>(GEP->getOperand(i)); 98 if (OpC == 0) 99 return VariableIdxFound = true; 100 if (OpC->isZero()) continue; // No offset. 101 102 // Handle struct indices, which add their field offset to the pointer. 103 if (const StructType *STy = dyn_cast<StructType>(*GTI)) { 104 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue()); 105 continue; 106 } 107 108 // Otherwise, we have a sequential type like an array or vector. Multiply 109 // the index by the ElementSize. 110 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()); 111 Offset += Size*OpC->getSExtValue(); 112 } 113 114 return Offset; 115} 116 117/// IsPointerOffset - Return true if Ptr1 is provably equal to Ptr2 plus a 118/// constant offset, and return that constant offset. For example, Ptr1 might 119/// be &A[42], and Ptr2 might be &A[40]. In this case offset would be -8. 120static bool IsPointerOffset(Value *Ptr1, Value *Ptr2, int64_t &Offset, 121 TargetData &TD) { 122 // Right now we handle the case when Ptr1/Ptr2 are both GEPs with an identical 123 // base. After that base, they may have some number of common (and 124 // potentially variable) indices. After that they handle some constant 125 // offset, which determines their offset from each other. At this point, we 126 // handle no other case. 127 GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(Ptr1); 128 GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(Ptr2); 129 if (!GEP1 || !GEP2 || GEP1->getOperand(0) != GEP2->getOperand(0)) 130 return false; 131 132 // Skip any common indices and track the GEP types. 133 unsigned Idx = 1; 134 for (; Idx != GEP1->getNumOperands() && Idx != GEP2->getNumOperands(); ++Idx) 135 if (GEP1->getOperand(Idx) != GEP2->getOperand(Idx)) 136 break; 137 138 bool VariableIdxFound = false; 139 int64_t Offset1 = GetOffsetFromIndex(GEP1, Idx, VariableIdxFound, TD); 140 int64_t Offset2 = GetOffsetFromIndex(GEP2, Idx, VariableIdxFound, TD); 141 if (VariableIdxFound) return false; 142 143 Offset = Offset2-Offset1; 144 return true; 145} 146 147 148/// MemsetRange - Represents a range of memset'd bytes with the ByteVal value. 149/// This allows us to analyze stores like: 150/// store 0 -> P+1 151/// store 0 -> P+0 152/// store 0 -> P+3 153/// store 0 -> P+2 154/// which sometimes happens with stores to arrays of structs etc. When we see 155/// the first store, we make a range [1, 2). The second store extends the range 156/// to [0, 2). The third makes a new range [2, 3). The fourth store joins the 157/// two ranges into [0, 3) which is memset'able. 158namespace { 159struct MemsetRange { 160 // Start/End - A semi range that describes the span that this range covers. 161 // The range is closed at the start and open at the end: [Start, End). 162 int64_t Start, End; 163 164 /// StartPtr - The getelementptr instruction that points to the start of the 165 /// range. 166 Value *StartPtr; 167 168 /// Alignment - The known alignment of the first store. 169 unsigned Alignment; 170 171 /// TheStores - The actual stores that make up this range. 172 SmallVector<StoreInst*, 16> TheStores; 173 174 bool isProfitableToUseMemset(const TargetData &TD) const; 175 176}; 177} // end anon namespace 178 179bool MemsetRange::isProfitableToUseMemset(const TargetData &TD) const { 180 // If we found more than 8 stores to merge or 64 bytes, use memset. 181 if (TheStores.size() >= 8 || End-Start >= 64) return true; 182 183 // Assume that the code generator is capable of merging pairs of stores 184 // together if it wants to. 185 if (TheStores.size() <= 2) return false; 186 187 // If we have fewer than 8 stores, it can still be worthwhile to do this. 188 // For example, merging 4 i8 stores into an i32 store is useful almost always. 189 // However, merging 2 32-bit stores isn't useful on a 32-bit architecture (the 190 // memset will be split into 2 32-bit stores anyway) and doing so can 191 // pessimize the llvm optimizer. 192 // 193 // Since we don't have perfect knowledge here, make some assumptions: assume 194 // the maximum GPR width is the same size as the pointer size and assume that 195 // this width can be stored. If so, check to see whether we will end up 196 // actually reducing the number of stores used. 197 unsigned Bytes = unsigned(End-Start); 198 unsigned NumPointerStores = Bytes/TD.getPointerSize(); 199 200 // Assume the remaining bytes if any are done a byte at a time. 201 unsigned NumByteStores = Bytes - NumPointerStores*TD.getPointerSize(); 202 203 // If we will reduce the # stores (according to this heuristic), do the 204 // transformation. This encourages merging 4 x i8 -> i32 and 2 x i16 -> i32 205 // etc. 206 return TheStores.size() > NumPointerStores+NumByteStores; 207} 208 209 210namespace { 211class MemsetRanges { 212 /// Ranges - A sorted list of the memset ranges. We use std::list here 213 /// because each element is relatively large and expensive to copy. 214 std::list<MemsetRange> Ranges; 215 typedef std::list<MemsetRange>::iterator range_iterator; 216 TargetData &TD; 217public: 218 MemsetRanges(TargetData &td) : TD(td) {} 219 220 typedef std::list<MemsetRange>::const_iterator const_iterator; 221 const_iterator begin() const { return Ranges.begin(); } 222 const_iterator end() const { return Ranges.end(); } 223 bool empty() const { return Ranges.empty(); } 224 225 void addStore(int64_t OffsetFromFirst, StoreInst *SI); 226}; 227 228} // end anon namespace 229 230 231/// addStore - Add a new store to the MemsetRanges data structure. This adds a 232/// new range for the specified store at the specified offset, merging into 233/// existing ranges as appropriate. 234void MemsetRanges::addStore(int64_t Start, StoreInst *SI) { 235 int64_t End = Start+TD.getTypeStoreSize(SI->getOperand(0)->getType()); 236 237 // Do a linear search of the ranges to see if this can be joined and/or to 238 // find the insertion point in the list. We keep the ranges sorted for 239 // simplicity here. This is a linear search of a linked list, which is ugly, 240 // however the number of ranges is limited, so this won't get crazy slow. 241 range_iterator I = Ranges.begin(), E = Ranges.end(); 242 243 while (I != E && Start > I->End) 244 ++I; 245 246 // We now know that I == E, in which case we didn't find anything to merge 247 // with, or that Start <= I->End. If End < I->Start or I == E, then we need 248 // to insert a new range. Handle this now. 249 if (I == E || End < I->Start) { 250 MemsetRange &R = *Ranges.insert(I, MemsetRange()); 251 R.Start = Start; 252 R.End = End; 253 R.StartPtr = SI->getPointerOperand(); 254 R.Alignment = SI->getAlignment(); 255 R.TheStores.push_back(SI); 256 return; 257 } 258 259 // This store overlaps with I, add it. 260 I->TheStores.push_back(SI); 261 262 // At this point, we may have an interval that completely contains our store. 263 // If so, just add it to the interval and return. 264 if (I->Start <= Start && I->End >= End) 265 return; 266 267 // Now we know that Start <= I->End and End >= I->Start so the range overlaps 268 // but is not entirely contained within the range. 269 270 // See if the range extends the start of the range. In this case, it couldn't 271 // possibly cause it to join the prior range, because otherwise we would have 272 // stopped on *it*. 273 if (Start < I->Start) { 274 I->Start = Start; 275 I->StartPtr = SI->getPointerOperand(); 276 } 277 278 // Now we know that Start <= I->End and Start >= I->Start (so the startpoint 279 // is in or right at the end of I), and that End >= I->Start. Extend I out to 280 // End. 281 if (End > I->End) { 282 I->End = End; 283 range_iterator NextI = I; 284 while (++NextI != E && End >= NextI->Start) { 285 // Merge the range in. 286 I->TheStores.append(NextI->TheStores.begin(), NextI->TheStores.end()); 287 if (NextI->End > I->End) 288 I->End = NextI->End; 289 Ranges.erase(NextI); 290 NextI = I; 291 } 292 } 293} 294 295//===----------------------------------------------------------------------===// 296// MemCpyOpt Pass 297//===----------------------------------------------------------------------===// 298 299namespace { 300 class MemCpyOpt : public FunctionPass { 301 bool runOnFunction(Function &F); 302 public: 303 static char ID; // Pass identification, replacement for typeid 304 MemCpyOpt() : FunctionPass(&ID) {} 305 306 private: 307 // This transformation requires dominator postdominator info 308 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 309 AU.setPreservesCFG(); 310 AU.addRequired<DominatorTree>(); 311 AU.addRequired<MemoryDependenceAnalysis>(); 312 AU.addRequired<AliasAnalysis>(); 313 AU.addPreserved<AliasAnalysis>(); 314 AU.addPreserved<MemoryDependenceAnalysis>(); 315 } 316 317 // Helper fuctions 318 bool processStore(StoreInst *SI, BasicBlock::iterator &BBI); 319 bool processMemCpy(MemCpyInst *M); 320 bool processMemMove(MemMoveInst *M); 321 bool performCallSlotOptzn(MemCpyInst *cpy, CallInst *C); 322 bool iterateOnFunction(Function &F); 323 }; 324 325 char MemCpyOpt::ID = 0; 326} 327 328// createMemCpyOptPass - The public interface to this file... 329FunctionPass *llvm::createMemCpyOptPass() { return new MemCpyOpt(); } 330 331static RegisterPass<MemCpyOpt> X("memcpyopt", 332 "MemCpy Optimization"); 333 334 335 336/// processStore - When GVN is scanning forward over instructions, we look for 337/// some other patterns to fold away. In particular, this looks for stores to 338/// neighboring locations of memory. If it sees enough consequtive ones 339/// (currently 4) it attempts to merge them together into a memcpy/memset. 340bool MemCpyOpt::processStore(StoreInst *SI, BasicBlock::iterator &BBI) { 341 if (SI->isVolatile()) return false; 342 343 // There are two cases that are interesting for this code to handle: memcpy 344 // and memset. Right now we only handle memset. 345 346 // Ensure that the value being stored is something that can be memset'able a 347 // byte at a time like "0" or "-1" or any width, as well as things like 348 // 0xA0A0A0A0 and 0.0. 349 Value *ByteVal = isBytewiseValue(SI->getOperand(0), SI->getContext()); 350 if (!ByteVal) 351 return false; 352 353 TargetData *TD = getAnalysisIfAvailable<TargetData>(); 354 if (!TD) return false; 355 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 356 Module *M = SI->getParent()->getParent()->getParent(); 357 358 // Okay, so we now have a single store that can be splatable. Scan to find 359 // all subsequent stores of the same value to offset from the same pointer. 360 // Join these together into ranges, so we can decide whether contiguous blocks 361 // are stored. 362 MemsetRanges Ranges(*TD); 363 364 Value *StartPtr = SI->getPointerOperand(); 365 366 BasicBlock::iterator BI = SI; 367 for (++BI; !isa<TerminatorInst>(BI); ++BI) { 368 if (isa<CallInst>(BI) || isa<InvokeInst>(BI)) { 369 // If the call is readnone, ignore it, otherwise bail out. We don't even 370 // allow readonly here because we don't want something like: 371 // A[1] = 2; strlen(A); A[2] = 2; -> memcpy(A, ...); strlen(A). 372 if (AA.getModRefBehavior(CallSite::get(BI)) == 373 AliasAnalysis::DoesNotAccessMemory) 374 continue; 375 376 // TODO: If this is a memset, try to join it in. 377 378 break; 379 } else if (isa<VAArgInst>(BI) || isa<LoadInst>(BI)) 380 break; 381 382 // If this is a non-store instruction it is fine, ignore it. 383 StoreInst *NextStore = dyn_cast<StoreInst>(BI); 384 if (NextStore == 0) continue; 385 386 // If this is a store, see if we can merge it in. 387 if (NextStore->isVolatile()) break; 388 389 // Check to see if this stored value is of the same byte-splattable value. 390 if (ByteVal != isBytewiseValue(NextStore->getOperand(0), 391 NextStore->getContext())) 392 break; 393 394 // Check to see if this store is to a constant offset from the start ptr. 395 int64_t Offset; 396 if (!IsPointerOffset(StartPtr, NextStore->getPointerOperand(), Offset, *TD)) 397 break; 398 399 Ranges.addStore(Offset, NextStore); 400 } 401 402 // If we have no ranges, then we just had a single store with nothing that 403 // could be merged in. This is a very common case of course. 404 if (Ranges.empty()) 405 return false; 406 407 // If we had at least one store that could be merged in, add the starting 408 // store as well. We try to avoid this unless there is at least something 409 // interesting as a small compile-time optimization. 410 Ranges.addStore(0, SI); 411 412 413 Function *MemSetF = 0; 414 415 // Now that we have full information about ranges, loop over the ranges and 416 // emit memset's for anything big enough to be worthwhile. 417 bool MadeChange = false; 418 for (MemsetRanges::const_iterator I = Ranges.begin(), E = Ranges.end(); 419 I != E; ++I) { 420 const MemsetRange &Range = *I; 421 422 if (Range.TheStores.size() == 1) continue; 423 424 // If it is profitable to lower this range to memset, do so now. 425 if (!Range.isProfitableToUseMemset(*TD)) 426 continue; 427 428 // Otherwise, we do want to transform this! Create a new memset. We put 429 // the memset right before the first instruction that isn't part of this 430 // memset block. This ensure that the memset is dominated by any addressing 431 // instruction needed by the start of the block. 432 BasicBlock::iterator InsertPt = BI; 433 434 if (MemSetF == 0) { 435 const Type *Ty = Type::getInt64Ty(SI->getContext()); 436 MemSetF = Intrinsic::getDeclaration(M, Intrinsic::memset, &Ty, 1); 437 } 438 439 // Get the starting pointer of the block. 440 StartPtr = Range.StartPtr; 441 442 // Cast the start ptr to be i8* as memset requires. 443 const Type *i8Ptr = 444 PointerType::getUnqual(Type::getInt8Ty(SI->getContext())); 445 if (StartPtr->getType() != i8Ptr) 446 StartPtr = new BitCastInst(StartPtr, i8Ptr, StartPtr->getName(), 447 InsertPt); 448 449 Value *Ops[] = { 450 StartPtr, ByteVal, // Start, value 451 // size 452 ConstantInt::get(Type::getInt64Ty(SI->getContext()), 453 Range.End-Range.Start), 454 // align 455 ConstantInt::get(Type::getInt32Ty(SI->getContext()), Range.Alignment) 456 }; 457 Value *C = CallInst::Create(MemSetF, Ops, Ops+4, "", InsertPt); 458 DEBUG(errs() << "Replace stores:\n"; 459 for (unsigned i = 0, e = Range.TheStores.size(); i != e; ++i) 460 errs() << *Range.TheStores[i]; 461 errs() << "With: " << *C); C=C; 462 463 // Don't invalidate the iterator 464 BBI = BI; 465 466 // Zap all the stores. 467 for (SmallVector<StoreInst*, 16>::const_iterator SI = Range.TheStores.begin(), 468 SE = Range.TheStores.end(); SI != SE; ++SI) 469 (*SI)->eraseFromParent(); 470 ++NumMemSetInfer; 471 MadeChange = true; 472 } 473 474 return MadeChange; 475} 476 477 478/// performCallSlotOptzn - takes a memcpy and a call that it depends on, 479/// and checks for the possibility of a call slot optimization by having 480/// the call write its result directly into the destination of the memcpy. 481bool MemCpyOpt::performCallSlotOptzn(MemCpyInst *cpy, CallInst *C) { 482 // The general transformation to keep in mind is 483 // 484 // call @func(..., src, ...) 485 // memcpy(dest, src, ...) 486 // 487 // -> 488 // 489 // memcpy(dest, src, ...) 490 // call @func(..., dest, ...) 491 // 492 // Since moving the memcpy is technically awkward, we additionally check that 493 // src only holds uninitialized values at the moment of the call, meaning that 494 // the memcpy can be discarded rather than moved. 495 496 // Deliberately get the source and destination with bitcasts stripped away, 497 // because we'll need to do type comparisons based on the underlying type. 498 Value *cpyDest = cpy->getDest(); 499 Value *cpySrc = cpy->getSource(); 500 CallSite CS = CallSite::get(C); 501 502 // We need to be able to reason about the size of the memcpy, so we require 503 // that it be a constant. 504 ConstantInt *cpyLength = dyn_cast<ConstantInt>(cpy->getLength()); 505 if (!cpyLength) 506 return false; 507 508 // Require that src be an alloca. This simplifies the reasoning considerably. 509 AllocaInst *srcAlloca = dyn_cast<AllocaInst>(cpySrc); 510 if (!srcAlloca) 511 return false; 512 513 // Check that all of src is copied to dest. 514 TargetData *TD = getAnalysisIfAvailable<TargetData>(); 515 if (!TD) return false; 516 517 ConstantInt *srcArraySize = dyn_cast<ConstantInt>(srcAlloca->getArraySize()); 518 if (!srcArraySize) 519 return false; 520 521 uint64_t srcSize = TD->getTypeAllocSize(srcAlloca->getAllocatedType()) * 522 srcArraySize->getZExtValue(); 523 524 if (cpyLength->getZExtValue() < srcSize) 525 return false; 526 527 // Check that accessing the first srcSize bytes of dest will not cause a 528 // trap. Otherwise the transform is invalid since it might cause a trap 529 // to occur earlier than it otherwise would. 530 if (AllocaInst *A = dyn_cast<AllocaInst>(cpyDest)) { 531 // The destination is an alloca. Check it is larger than srcSize. 532 ConstantInt *destArraySize = dyn_cast<ConstantInt>(A->getArraySize()); 533 if (!destArraySize) 534 return false; 535 536 uint64_t destSize = TD->getTypeAllocSize(A->getAllocatedType()) * 537 destArraySize->getZExtValue(); 538 539 if (destSize < srcSize) 540 return false; 541 } else if (Argument *A = dyn_cast<Argument>(cpyDest)) { 542 // If the destination is an sret parameter then only accesses that are 543 // outside of the returned struct type can trap. 544 if (!A->hasStructRetAttr()) 545 return false; 546 547 const Type *StructTy = cast<PointerType>(A->getType())->getElementType(); 548 uint64_t destSize = TD->getTypeAllocSize(StructTy); 549 550 if (destSize < srcSize) 551 return false; 552 } else { 553 return false; 554 } 555 556 // Check that src is not accessed except via the call and the memcpy. This 557 // guarantees that it holds only undefined values when passed in (so the final 558 // memcpy can be dropped), that it is not read or written between the call and 559 // the memcpy, and that writing beyond the end of it is undefined. 560 SmallVector<User*, 8> srcUseList(srcAlloca->use_begin(), 561 srcAlloca->use_end()); 562 while (!srcUseList.empty()) { 563 User *UI = srcUseList.back(); 564 srcUseList.pop_back(); 565 566 if (isa<BitCastInst>(UI)) { 567 for (User::use_iterator I = UI->use_begin(), E = UI->use_end(); 568 I != E; ++I) 569 srcUseList.push_back(*I); 570 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(UI)) { 571 if (G->hasAllZeroIndices()) 572 for (User::use_iterator I = UI->use_begin(), E = UI->use_end(); 573 I != E; ++I) 574 srcUseList.push_back(*I); 575 else 576 return false; 577 } else if (UI != C && UI != cpy) { 578 return false; 579 } 580 } 581 582 // Since we're changing the parameter to the callsite, we need to make sure 583 // that what would be the new parameter dominates the callsite. 584 DominatorTree &DT = getAnalysis<DominatorTree>(); 585 if (Instruction *cpyDestInst = dyn_cast<Instruction>(cpyDest)) 586 if (!DT.dominates(cpyDestInst, C)) 587 return false; 588 589 // In addition to knowing that the call does not access src in some 590 // unexpected manner, for example via a global, which we deduce from 591 // the use analysis, we also need to know that it does not sneakily 592 // access dest. We rely on AA to figure this out for us. 593 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 594 if (AA.getModRefInfo(C, cpy->getRawDest(), srcSize) != 595 AliasAnalysis::NoModRef) 596 return false; 597 598 // All the checks have passed, so do the transformation. 599 bool changedArgument = false; 600 for (unsigned i = 0; i < CS.arg_size(); ++i) 601 if (CS.getArgument(i)->stripPointerCasts() == cpySrc) { 602 if (cpySrc->getType() != cpyDest->getType()) 603 cpyDest = CastInst::CreatePointerCast(cpyDest, cpySrc->getType(), 604 cpyDest->getName(), C); 605 changedArgument = true; 606 if (CS.getArgument(i)->getType() == cpyDest->getType()) 607 CS.setArgument(i, cpyDest); 608 else 609 CS.setArgument(i, CastInst::CreatePointerCast(cpyDest, 610 CS.getArgument(i)->getType(), cpyDest->getName(), C)); 611 } 612 613 if (!changedArgument) 614 return false; 615 616 // Drop any cached information about the call, because we may have changed 617 // its dependence information by changing its parameter. 618 MemoryDependenceAnalysis &MD = getAnalysis<MemoryDependenceAnalysis>(); 619 MD.removeInstruction(C); 620 621 // Remove the memcpy 622 MD.removeInstruction(cpy); 623 cpy->eraseFromParent(); 624 NumMemCpyInstr++; 625 626 return true; 627} 628 629/// processMemCpy - perform simplication of memcpy's. If we have memcpy A which 630/// copies X to Y, and memcpy B which copies Y to Z, then we can rewrite B to be 631/// a memcpy from X to Z (or potentially a memmove, depending on circumstances). 632/// This allows later passes to remove the first memcpy altogether. 633bool MemCpyOpt::processMemCpy(MemCpyInst *M) { 634 MemoryDependenceAnalysis &MD = getAnalysis<MemoryDependenceAnalysis>(); 635 636 // The are two possible optimizations we can do for memcpy: 637 // a) memcpy-memcpy xform which exposes redundance for DSE. 638 // b) call-memcpy xform for return slot optimization. 639 MemDepResult dep = MD.getDependency(M); 640 if (!dep.isClobber()) 641 return false; 642 if (!isa<MemCpyInst>(dep.getInst())) { 643 if (CallInst *C = dyn_cast<CallInst>(dep.getInst())) 644 return performCallSlotOptzn(M, C); 645 return false; 646 } 647 648 MemCpyInst *MDep = cast<MemCpyInst>(dep.getInst()); 649 650 // We can only transforms memcpy's where the dest of one is the source of the 651 // other 652 if (M->getSource() != MDep->getDest()) 653 return false; 654 655 // Second, the length of the memcpy's must be the same, or the preceeding one 656 // must be larger than the following one. 657 ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength()); 658 ConstantInt *C2 = dyn_cast<ConstantInt>(M->getLength()); 659 if (!C1 || !C2) 660 return false; 661 662 uint64_t DepSize = C1->getValue().getZExtValue(); 663 uint64_t CpySize = C2->getValue().getZExtValue(); 664 665 if (DepSize < CpySize) 666 return false; 667 668 // Finally, we have to make sure that the dest of the second does not 669 // alias the source of the first 670 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 671 if (AA.alias(M->getRawDest(), CpySize, MDep->getRawSource(), DepSize) != 672 AliasAnalysis::NoAlias) 673 return false; 674 else if (AA.alias(M->getRawDest(), CpySize, M->getRawSource(), CpySize) != 675 AliasAnalysis::NoAlias) 676 return false; 677 else if (AA.alias(MDep->getRawDest(), DepSize, MDep->getRawSource(), DepSize) 678 != AliasAnalysis::NoAlias) 679 return false; 680 681 // If all checks passed, then we can transform these memcpy's 682 const Type *Ty = M->getLength()->getType(); 683 Function *MemCpyFun = Intrinsic::getDeclaration( 684 M->getParent()->getParent()->getParent(), 685 M->getIntrinsicID(), &Ty, 1); 686 687 Value *Args[4] = { 688 M->getRawDest(), MDep->getRawSource(), M->getLength(), M->getAlignmentCst() 689 }; 690 691 CallInst *C = CallInst::Create(MemCpyFun, Args, Args+4, "", M); 692 693 694 // If C and M don't interfere, then this is a valid transformation. If they 695 // did, this would mean that the two sources overlap, which would be bad. 696 if (MD.getDependency(C) == dep) { 697 MD.removeInstruction(M); 698 M->eraseFromParent(); 699 NumMemCpyInstr++; 700 return true; 701 } 702 703 // Otherwise, there was no point in doing this, so we remove the call we 704 // inserted and act like nothing happened. 705 MD.removeInstruction(C); 706 C->eraseFromParent(); 707 return false; 708} 709 710/// processMemMove - Transforms memmove calls to memcpy calls when the src/dst 711/// are guaranteed not to alias. 712bool MemCpyOpt::processMemMove(MemMoveInst *M) { 713 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 714 715 // If the memmove is a constant size, use it for the alias query, this allows 716 // us to optimize things like: memmove(P, P+64, 64); 717 uint64_t MemMoveSize = ~0ULL; 718 if (ConstantInt *Len = dyn_cast<ConstantInt>(M->getLength())) 719 MemMoveSize = Len->getZExtValue(); 720 721 // See if the pointers alias. 722 if (AA.alias(M->getRawDest(), MemMoveSize, M->getRawSource(), MemMoveSize) != 723 AliasAnalysis::NoAlias) 724 return false; 725 726 DEBUG(errs() << "MemCpyOpt: Optimizing memmove -> memcpy: " << *M << "\n"); 727 728 // If not, then we know we can transform this. 729 Module *Mod = M->getParent()->getParent()->getParent(); 730 const Type *Ty = M->getLength()->getType(); 731 M->setOperand(0, Intrinsic::getDeclaration(Mod, Intrinsic::memcpy, &Ty, 1)); 732 733 // MemDep may have over conservative information about this instruction, just 734 // conservatively flush it from the cache. 735 getAnalysis<MemoryDependenceAnalysis>().removeInstruction(M); 736 737 ++NumMoveToCpy; 738 return true; 739} 740 741 742// MemCpyOpt::iterateOnFunction - Executes one iteration of GVN. 743bool MemCpyOpt::iterateOnFunction(Function &F) { 744 bool MadeChange = false; 745 746 // Walk all instruction in the function. 747 for (Function::iterator BB = F.begin(), BBE = F.end(); BB != BBE; ++BB) { 748 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); 749 BI != BE;) { 750 // Avoid invalidating the iterator. 751 Instruction *I = BI++; 752 753 if (StoreInst *SI = dyn_cast<StoreInst>(I)) 754 MadeChange |= processStore(SI, BI); 755 else if (MemCpyInst *M = dyn_cast<MemCpyInst>(I)) 756 MadeChange |= processMemCpy(M); 757 else if (MemMoveInst *M = dyn_cast<MemMoveInst>(I)) { 758 if (processMemMove(M)) { 759 --BI; // Reprocess the new memcpy. 760 MadeChange = true; 761 } 762 } 763 } 764 } 765 766 return MadeChange; 767} 768 769// MemCpyOpt::runOnFunction - This is the main transformation entry point for a 770// function. 771// 772bool MemCpyOpt::runOnFunction(Function &F) { 773 bool MadeChange = false; 774 while (1) { 775 if (!iterateOnFunction(F)) 776 break; 777 MadeChange = true; 778 } 779 780 return MadeChange; 781} 782 783 784 785