MemCpyOptimizer.cpp revision 67a716ab818301fe28068754c879e568c44e62f8
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/GlobalVariable.h" 18#include "llvm/IntrinsicInst.h" 19#include "llvm/Instructions.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/Analysis/ValueTracking.h" 26#include "llvm/Support/Debug.h" 27#include "llvm/Support/GetElementPtrTypeIterator.h" 28#include "llvm/Support/IRBuilder.h" 29#include "llvm/Support/raw_ostream.h" 30#include "llvm/Target/TargetData.h" 31#include <list> 32using namespace llvm; 33 34STATISTIC(NumMemCpyInstr, "Number of memcpy instructions deleted"); 35STATISTIC(NumMemSetInfer, "Number of memsets inferred"); 36STATISTIC(NumMoveToCpy, "Number of memmoves converted to memcpy"); 37STATISTIC(NumCpyToSet, "Number of memcpys converted to memset"); 38 39static int64_t GetOffsetFromIndex(const GetElementPtrInst *GEP, unsigned Idx, 40 bool &VariableIdxFound, const TargetData &TD){ 41 // Skip over the first indices. 42 gep_type_iterator GTI = gep_type_begin(GEP); 43 for (unsigned i = 1; i != Idx; ++i, ++GTI) 44 /*skip along*/; 45 46 // Compute the offset implied by the rest of the indices. 47 int64_t Offset = 0; 48 for (unsigned i = Idx, e = GEP->getNumOperands(); i != e; ++i, ++GTI) { 49 ConstantInt *OpC = dyn_cast<ConstantInt>(GEP->getOperand(i)); 50 if (OpC == 0) 51 return VariableIdxFound = true; 52 if (OpC->isZero()) continue; // No offset. 53 54 // Handle struct indices, which add their field offset to the pointer. 55 if (const StructType *STy = dyn_cast<StructType>(*GTI)) { 56 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue()); 57 continue; 58 } 59 60 // Otherwise, we have a sequential type like an array or vector. Multiply 61 // the index by the ElementSize. 62 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()); 63 Offset += Size*OpC->getSExtValue(); 64 } 65 66 return Offset; 67} 68 69/// IsPointerOffset - Return true if Ptr1 is provably equal to Ptr2 plus a 70/// constant offset, and return that constant offset. For example, Ptr1 might 71/// be &A[42], and Ptr2 might be &A[40]. In this case offset would be -8. 72static bool IsPointerOffset(Value *Ptr1, Value *Ptr2, int64_t &Offset, 73 const TargetData &TD) { 74 // Right now we handle the case when Ptr1/Ptr2 are both GEPs with an identical 75 // base. After that base, they may have some number of common (and 76 // potentially variable) indices. After that they handle some constant 77 // offset, which determines their offset from each other. At this point, we 78 // handle no other case. 79 GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(Ptr1); 80 GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(Ptr2); 81 if (!GEP1 || !GEP2 || GEP1->getOperand(0) != GEP2->getOperand(0)) 82 return false; 83 84 // Skip any common indices and track the GEP types. 85 unsigned Idx = 1; 86 for (; Idx != GEP1->getNumOperands() && Idx != GEP2->getNumOperands(); ++Idx) 87 if (GEP1->getOperand(Idx) != GEP2->getOperand(Idx)) 88 break; 89 90 bool VariableIdxFound = false; 91 int64_t Offset1 = GetOffsetFromIndex(GEP1, Idx, VariableIdxFound, TD); 92 int64_t Offset2 = GetOffsetFromIndex(GEP2, Idx, VariableIdxFound, TD); 93 if (VariableIdxFound) return false; 94 95 Offset = Offset2-Offset1; 96 return true; 97} 98 99 100/// MemsetRange - Represents a range of memset'd bytes with the ByteVal value. 101/// This allows us to analyze stores like: 102/// store 0 -> P+1 103/// store 0 -> P+0 104/// store 0 -> P+3 105/// store 0 -> P+2 106/// which sometimes happens with stores to arrays of structs etc. When we see 107/// the first store, we make a range [1, 2). The second store extends the range 108/// to [0, 2). The third makes a new range [2, 3). The fourth store joins the 109/// two ranges into [0, 3) which is memset'able. 110namespace { 111struct MemsetRange { 112 // Start/End - A semi range that describes the span that this range covers. 113 // The range is closed at the start and open at the end: [Start, End). 114 int64_t Start, End; 115 116 /// StartPtr - The getelementptr instruction that points to the start of the 117 /// range. 118 Value *StartPtr; 119 120 /// Alignment - The known alignment of the first store. 121 unsigned Alignment; 122 123 /// TheStores - The actual stores that make up this range. 124 SmallVector<StoreInst*, 16> TheStores; 125 126 bool isProfitableToUseMemset(const TargetData &TD) const; 127 128}; 129} // end anon namespace 130 131bool MemsetRange::isProfitableToUseMemset(const TargetData &TD) const { 132 // If we found more than 8 stores to merge or 64 bytes, use memset. 133 if (TheStores.size() >= 8 || End-Start >= 64) return true; 134 135 // Assume that the code generator is capable of merging pairs of stores 136 // together if it wants to. 137 if (TheStores.size() <= 2) return false; 138 139 // If we have fewer than 8 stores, it can still be worthwhile to do this. 140 // For example, merging 4 i8 stores into an i32 store is useful almost always. 141 // However, merging 2 32-bit stores isn't useful on a 32-bit architecture (the 142 // memset will be split into 2 32-bit stores anyway) and doing so can 143 // pessimize the llvm optimizer. 144 // 145 // Since we don't have perfect knowledge here, make some assumptions: assume 146 // the maximum GPR width is the same size as the pointer size and assume that 147 // this width can be stored. If so, check to see whether we will end up 148 // actually reducing the number of stores used. 149 unsigned Bytes = unsigned(End-Start); 150 unsigned NumPointerStores = Bytes/TD.getPointerSize(); 151 152 // Assume the remaining bytes if any are done a byte at a time. 153 unsigned NumByteStores = Bytes - NumPointerStores*TD.getPointerSize(); 154 155 // If we will reduce the # stores (according to this heuristic), do the 156 // transformation. This encourages merging 4 x i8 -> i32 and 2 x i16 -> i32 157 // etc. 158 return TheStores.size() > NumPointerStores+NumByteStores; 159} 160 161 162namespace { 163class MemsetRanges { 164 /// Ranges - A sorted list of the memset ranges. We use std::list here 165 /// because each element is relatively large and expensive to copy. 166 std::list<MemsetRange> Ranges; 167 typedef std::list<MemsetRange>::iterator range_iterator; 168 const TargetData &TD; 169public: 170 MemsetRanges(const TargetData &td) : TD(td) {} 171 172 typedef std::list<MemsetRange>::const_iterator const_iterator; 173 const_iterator begin() const { return Ranges.begin(); } 174 const_iterator end() const { return Ranges.end(); } 175 bool empty() const { return Ranges.empty(); } 176 177 void addStore(int64_t OffsetFromFirst, StoreInst *SI); 178 179 void addInst(int64_t OffsetFromFirst, Instruction *Inst) { 180 addStore(OffsetFromFirst, cast<StoreInst>(Inst)); 181 } 182}; 183 184} // end anon namespace 185 186 187/// addStore - Add a new store to the MemsetRanges data structure. This adds a 188/// new range for the specified store at the specified offset, merging into 189/// existing ranges as appropriate. 190void MemsetRanges::addStore(int64_t Start, StoreInst *SI) { 191 int64_t End = Start+TD.getTypeStoreSize(SI->getOperand(0)->getType()); 192 193 // Do a linear search of the ranges to see if this can be joined and/or to 194 // find the insertion point in the list. We keep the ranges sorted for 195 // simplicity here. This is a linear search of a linked list, which is ugly, 196 // however the number of ranges is limited, so this won't get crazy slow. 197 range_iterator I = Ranges.begin(), E = Ranges.end(); 198 199 while (I != E && Start > I->End) 200 ++I; 201 202 // We now know that I == E, in which case we didn't find anything to merge 203 // with, or that Start <= I->End. If End < I->Start or I == E, then we need 204 // to insert a new range. Handle this now. 205 if (I == E || End < I->Start) { 206 MemsetRange &R = *Ranges.insert(I, MemsetRange()); 207 R.Start = Start; 208 R.End = End; 209 R.StartPtr = SI->getPointerOperand(); 210 R.Alignment = SI->getAlignment(); 211 R.TheStores.push_back(SI); 212 return; 213 } 214 215 // This store overlaps with I, add it. 216 I->TheStores.push_back(SI); 217 218 // At this point, we may have an interval that completely contains our store. 219 // If so, just add it to the interval and return. 220 if (I->Start <= Start && I->End >= End) 221 return; 222 223 // Now we know that Start <= I->End and End >= I->Start so the range overlaps 224 // but is not entirely contained within the range. 225 226 // See if the range extends the start of the range. In this case, it couldn't 227 // possibly cause it to join the prior range, because otherwise we would have 228 // stopped on *it*. 229 if (Start < I->Start) { 230 I->Start = Start; 231 I->StartPtr = SI->getPointerOperand(); 232 I->Alignment = SI->getAlignment(); 233 } 234 235 // Now we know that Start <= I->End and Start >= I->Start (so the startpoint 236 // is in or right at the end of I), and that End >= I->Start. Extend I out to 237 // End. 238 if (End > I->End) { 239 I->End = End; 240 range_iterator NextI = I; 241 while (++NextI != E && End >= NextI->Start) { 242 // Merge the range in. 243 I->TheStores.append(NextI->TheStores.begin(), NextI->TheStores.end()); 244 if (NextI->End > I->End) 245 I->End = NextI->End; 246 Ranges.erase(NextI); 247 NextI = I; 248 } 249 } 250} 251 252//===----------------------------------------------------------------------===// 253// MemCpyOpt Pass 254//===----------------------------------------------------------------------===// 255 256namespace { 257 class MemCpyOpt : public FunctionPass { 258 MemoryDependenceAnalysis *MD; 259 const TargetData *TD; 260 public: 261 static char ID; // Pass identification, replacement for typeid 262 MemCpyOpt() : FunctionPass(ID) { 263 initializeMemCpyOptPass(*PassRegistry::getPassRegistry()); 264 MD = 0; 265 } 266 267 bool runOnFunction(Function &F); 268 269 private: 270 // This transformation requires dominator postdominator info 271 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 272 AU.setPreservesCFG(); 273 AU.addRequired<DominatorTree>(); 274 AU.addRequired<MemoryDependenceAnalysis>(); 275 AU.addRequired<AliasAnalysis>(); 276 AU.addPreserved<AliasAnalysis>(); 277 AU.addPreserved<MemoryDependenceAnalysis>(); 278 } 279 280 // Helper fuctions 281 bool processStore(StoreInst *SI, BasicBlock::iterator &BBI); 282 bool processMemCpy(MemCpyInst *M); 283 bool processMemMove(MemMoveInst *M); 284 bool performCallSlotOptzn(Instruction *cpy, Value *cpyDst, Value *cpySrc, 285 uint64_t cpyLen, CallInst *C); 286 bool processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep, 287 uint64_t MSize); 288 bool processByValArgument(CallSite CS, unsigned ArgNo); 289 Instruction *tryMergingIntoMemset(Instruction *I, Value *StartPtr, 290 Value *ByteVal); 291 292 bool iterateOnFunction(Function &F); 293 }; 294 295 char MemCpyOpt::ID = 0; 296} 297 298// createMemCpyOptPass - The public interface to this file... 299FunctionPass *llvm::createMemCpyOptPass() { return new MemCpyOpt(); } 300 301INITIALIZE_PASS_BEGIN(MemCpyOpt, "memcpyopt", "MemCpy Optimization", 302 false, false) 303INITIALIZE_PASS_DEPENDENCY(DominatorTree) 304INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis) 305INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 306INITIALIZE_PASS_END(MemCpyOpt, "memcpyopt", "MemCpy Optimization", 307 false, false) 308 309/// tryMergingIntoMemset - When scanning forward over instructions, we look for 310/// some other patterns to fold away. In particular, this looks for stores to 311/// neighboring locations of memory. If it sees enough consequtive ones, it 312/// attempts to merge them together into a memcpy/memset. 313Instruction *MemCpyOpt::tryMergingIntoMemset(Instruction *StartInst, 314 Value *StartPtr, Value *ByteVal) { 315 if (TD == 0) return 0; 316 317 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 318 319 // Okay, so we now have a single store that can be splatable. Scan to find 320 // all subsequent stores of the same value to offset from the same pointer. 321 // Join these together into ranges, so we can decide whether contiguous blocks 322 // are stored. 323 MemsetRanges Ranges(*TD); 324 325 BasicBlock::iterator BI = StartInst; 326 for (++BI; !isa<TerminatorInst>(BI); ++BI) { 327 if (isa<CallInst>(BI) || isa<InvokeInst>(BI)) { 328 // If the call is readnone, ignore it, otherwise bail out. We don't even 329 // allow readonly here because we don't want something like: 330 // A[1] = 2; strlen(A); A[2] = 2; -> memcpy(A, ...); strlen(A). 331 if (AA.getModRefBehavior(CallSite(BI)) == 332 AliasAnalysis::DoesNotAccessMemory) 333 continue; 334 335 // TODO: If this is a memset, try to join it in. 336 337 break; 338 } else if (isa<VAArgInst>(BI) || isa<LoadInst>(BI)) 339 break; 340 341 // If this is a non-store instruction it is fine, ignore it. 342 StoreInst *NextStore = dyn_cast<StoreInst>(BI); 343 if (NextStore == 0) continue; 344 345 // If this is a store, see if we can merge it in. 346 if (NextStore->isVolatile()) break; 347 348 // Check to see if this stored value is of the same byte-splattable value. 349 if (ByteVal != isBytewiseValue(NextStore->getOperand(0))) 350 break; 351 352 // Check to see if this store is to a constant offset from the start ptr. 353 int64_t Offset; 354 if (!IsPointerOffset(StartPtr, NextStore->getPointerOperand(), Offset, *TD)) 355 break; 356 357 Ranges.addStore(Offset, NextStore); 358 } 359 360 // If we have no ranges, then we just had a single store with nothing that 361 // could be merged in. This is a very common case of course. 362 if (Ranges.empty()) 363 return 0; 364 365 // If we had at least one store that could be merged in, add the starting 366 // store as well. We try to avoid this unless there is at least something 367 // interesting as a small compile-time optimization. 368 Ranges.addInst(0, StartInst); 369 370 // If we create any memsets, we put it right before the first instruction that 371 // isn't part of the memset block. This ensure that the memset is dominated 372 // by any addressing instruction needed by the start of the block. 373 IRBuilder<> Builder(BI); 374 375 // Now that we have full information about ranges, loop over the ranges and 376 // emit memset's for anything big enough to be worthwhile. 377 Instruction *AMemSet = 0; 378 for (MemsetRanges::const_iterator I = Ranges.begin(), E = Ranges.end(); 379 I != E; ++I) { 380 const MemsetRange &Range = *I; 381 382 if (Range.TheStores.size() == 1) continue; 383 384 // If it is profitable to lower this range to memset, do so now. 385 if (!Range.isProfitableToUseMemset(*TD)) 386 continue; 387 388 // Otherwise, we do want to transform this! Create a new memset. 389 // Get the starting pointer of the block. 390 StartPtr = Range.StartPtr; 391 392 // Determine alignment 393 unsigned Alignment = Range.Alignment; 394 if (Alignment == 0) { 395 const Type *EltType = 396 cast<PointerType>(StartPtr->getType())->getElementType(); 397 Alignment = TD->getABITypeAlignment(EltType); 398 } 399 400 AMemSet = 401 Builder.CreateMemSet(StartPtr, ByteVal, Range.End-Range.Start, Alignment); 402 403 DEBUG(dbgs() << "Replace stores:\n"; 404 for (unsigned i = 0, e = Range.TheStores.size(); i != e; ++i) 405 dbgs() << *Range.TheStores[i] << '\n'; 406 dbgs() << "With: " << *AMemSet << '\n'); 407 408 // Zap all the stores. 409 for (SmallVector<StoreInst*, 16>::const_iterator 410 SI = Range.TheStores.begin(), 411 SE = Range.TheStores.end(); SI != SE; ++SI) 412 (*SI)->eraseFromParent(); 413 ++NumMemSetInfer; 414 } 415 416 return AMemSet; 417} 418 419 420bool MemCpyOpt::processStore(StoreInst *SI, BasicBlock::iterator &BBI) { 421 if (SI->isVolatile()) return false; 422 423 if (TD == 0) return false; 424 425 // Detect cases where we're performing call slot forwarding, but 426 // happen to be using a load-store pair to implement it, rather than 427 // a memcpy. 428 if (LoadInst *LI = dyn_cast<LoadInst>(SI->getOperand(0))) { 429 if (!LI->isVolatile() && LI->hasOneUse()) { 430 MemDepResult dep = MD->getDependency(LI); 431 CallInst *C = 0; 432 if (dep.isClobber() && !isa<MemCpyInst>(dep.getInst())) 433 C = dyn_cast<CallInst>(dep.getInst()); 434 435 if (C) { 436 bool changed = performCallSlotOptzn(LI, 437 SI->getPointerOperand()->stripPointerCasts(), 438 LI->getPointerOperand()->stripPointerCasts(), 439 TD->getTypeStoreSize(SI->getOperand(0)->getType()), C); 440 if (changed) { 441 MD->removeInstruction(SI); 442 SI->eraseFromParent(); 443 LI->eraseFromParent(); 444 ++NumMemCpyInstr; 445 return true; 446 } 447 } 448 } 449 } 450 451 // There are two cases that are interesting for this code to handle: memcpy 452 // and memset. Right now we only handle memset. 453 454 // Ensure that the value being stored is something that can be memset'able a 455 // byte at a time like "0" or "-1" or any width, as well as things like 456 // 0xA0A0A0A0 and 0.0. 457 if (Value *ByteVal = isBytewiseValue(SI->getOperand(0))) 458 if (Instruction *I = tryMergingIntoMemset(SI, SI->getPointerOperand(), 459 ByteVal)) { 460 BBI = I; // Don't invalidate iterator. 461 return true; 462 } 463 464 return false; 465} 466 467 468/// performCallSlotOptzn - takes a memcpy and a call that it depends on, 469/// and checks for the possibility of a call slot optimization by having 470/// the call write its result directly into the destination of the memcpy. 471bool MemCpyOpt::performCallSlotOptzn(Instruction *cpy, 472 Value *cpyDest, Value *cpySrc, 473 uint64_t cpyLen, CallInst *C) { 474 // The general transformation to keep in mind is 475 // 476 // call @func(..., src, ...) 477 // memcpy(dest, src, ...) 478 // 479 // -> 480 // 481 // memcpy(dest, src, ...) 482 // call @func(..., dest, ...) 483 // 484 // Since moving the memcpy is technically awkward, we additionally check that 485 // src only holds uninitialized values at the moment of the call, meaning that 486 // the memcpy can be discarded rather than moved. 487 488 // Deliberately get the source and destination with bitcasts stripped away, 489 // because we'll need to do type comparisons based on the underlying type. 490 CallSite CS(C); 491 492 // Require that src be an alloca. This simplifies the reasoning considerably. 493 AllocaInst *srcAlloca = dyn_cast<AllocaInst>(cpySrc); 494 if (!srcAlloca) 495 return false; 496 497 // Check that all of src is copied to dest. 498 if (TD == 0) return false; 499 500 ConstantInt *srcArraySize = dyn_cast<ConstantInt>(srcAlloca->getArraySize()); 501 if (!srcArraySize) 502 return false; 503 504 uint64_t srcSize = TD->getTypeAllocSize(srcAlloca->getAllocatedType()) * 505 srcArraySize->getZExtValue(); 506 507 if (cpyLen < srcSize) 508 return false; 509 510 // Check that accessing the first srcSize bytes of dest will not cause a 511 // trap. Otherwise the transform is invalid since it might cause a trap 512 // to occur earlier than it otherwise would. 513 if (AllocaInst *A = dyn_cast<AllocaInst>(cpyDest)) { 514 // The destination is an alloca. Check it is larger than srcSize. 515 ConstantInt *destArraySize = dyn_cast<ConstantInt>(A->getArraySize()); 516 if (!destArraySize) 517 return false; 518 519 uint64_t destSize = TD->getTypeAllocSize(A->getAllocatedType()) * 520 destArraySize->getZExtValue(); 521 522 if (destSize < srcSize) 523 return false; 524 } else if (Argument *A = dyn_cast<Argument>(cpyDest)) { 525 // If the destination is an sret parameter then only accesses that are 526 // outside of the returned struct type can trap. 527 if (!A->hasStructRetAttr()) 528 return false; 529 530 const Type *StructTy = cast<PointerType>(A->getType())->getElementType(); 531 uint64_t destSize = TD->getTypeAllocSize(StructTy); 532 533 if (destSize < srcSize) 534 return false; 535 } else { 536 return false; 537 } 538 539 // Check that src is not accessed except via the call and the memcpy. This 540 // guarantees that it holds only undefined values when passed in (so the final 541 // memcpy can be dropped), that it is not read or written between the call and 542 // the memcpy, and that writing beyond the end of it is undefined. 543 SmallVector<User*, 8> srcUseList(srcAlloca->use_begin(), 544 srcAlloca->use_end()); 545 while (!srcUseList.empty()) { 546 User *UI = srcUseList.pop_back_val(); 547 548 if (isa<BitCastInst>(UI)) { 549 for (User::use_iterator I = UI->use_begin(), E = UI->use_end(); 550 I != E; ++I) 551 srcUseList.push_back(*I); 552 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(UI)) { 553 if (G->hasAllZeroIndices()) 554 for (User::use_iterator I = UI->use_begin(), E = UI->use_end(); 555 I != E; ++I) 556 srcUseList.push_back(*I); 557 else 558 return false; 559 } else if (UI != C && UI != cpy) { 560 return false; 561 } 562 } 563 564 // Since we're changing the parameter to the callsite, we need to make sure 565 // that what would be the new parameter dominates the callsite. 566 DominatorTree &DT = getAnalysis<DominatorTree>(); 567 if (Instruction *cpyDestInst = dyn_cast<Instruction>(cpyDest)) 568 if (!DT.dominates(cpyDestInst, C)) 569 return false; 570 571 // In addition to knowing that the call does not access src in some 572 // unexpected manner, for example via a global, which we deduce from 573 // the use analysis, we also need to know that it does not sneakily 574 // access dest. We rely on AA to figure this out for us. 575 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 576 if (AA.getModRefInfo(C, cpyDest, srcSize) != 577 AliasAnalysis::NoModRef) 578 return false; 579 580 // All the checks have passed, so do the transformation. 581 bool changedArgument = false; 582 for (unsigned i = 0; i < CS.arg_size(); ++i) 583 if (CS.getArgument(i)->stripPointerCasts() == cpySrc) { 584 if (cpySrc->getType() != cpyDest->getType()) 585 cpyDest = CastInst::CreatePointerCast(cpyDest, cpySrc->getType(), 586 cpyDest->getName(), C); 587 changedArgument = true; 588 if (CS.getArgument(i)->getType() == cpyDest->getType()) 589 CS.setArgument(i, cpyDest); 590 else 591 CS.setArgument(i, CastInst::CreatePointerCast(cpyDest, 592 CS.getArgument(i)->getType(), cpyDest->getName(), C)); 593 } 594 595 if (!changedArgument) 596 return false; 597 598 // Drop any cached information about the call, because we may have changed 599 // its dependence information by changing its parameter. 600 MD->removeInstruction(C); 601 602 // Remove the memcpy. 603 MD->removeInstruction(cpy); 604 ++NumMemCpyInstr; 605 606 return true; 607} 608 609/// processMemCpyMemCpyDependence - We've found that the (upward scanning) 610/// memory dependence of memcpy 'M' is the memcpy 'MDep'. Try to simplify M to 611/// copy from MDep's input if we can. MSize is the size of M's copy. 612/// 613bool MemCpyOpt::processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep, 614 uint64_t MSize) { 615 // We can only transforms memcpy's where the dest of one is the source of the 616 // other. 617 if (M->getSource() != MDep->getDest() || MDep->isVolatile()) 618 return false; 619 620 // If dep instruction is reading from our current input, then it is a noop 621 // transfer and substituting the input won't change this instruction. Just 622 // ignore the input and let someone else zap MDep. This handles cases like: 623 // memcpy(a <- a) 624 // memcpy(b <- a) 625 if (M->getSource() == MDep->getSource()) 626 return false; 627 628 // Second, the length of the memcpy's must be the same, or the preceeding one 629 // must be larger than the following one. 630 ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength()); 631 if (!C1) return false; 632 633 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 634 635 // Verify that the copied-from memory doesn't change in between the two 636 // transfers. For example, in: 637 // memcpy(a <- b) 638 // *b = 42; 639 // memcpy(c <- a) 640 // It would be invalid to transform the second memcpy into memcpy(c <- b). 641 // 642 // TODO: If the code between M and MDep is transparent to the destination "c", 643 // then we could still perform the xform by moving M up to the first memcpy. 644 // 645 // NOTE: This is conservative, it will stop on any read from the source loc, 646 // not just the defining memcpy. 647 MemDepResult SourceDep = 648 MD->getPointerDependencyFrom(AA.getLocationForSource(MDep), 649 false, M, M->getParent()); 650 if (!SourceDep.isClobber() || SourceDep.getInst() != MDep) 651 return false; 652 653 // If the dest of the second might alias the source of the first, then the 654 // source and dest might overlap. We still want to eliminate the intermediate 655 // value, but we have to generate a memmove instead of memcpy. 656 bool UseMemMove = false; 657 if (!AA.isNoAlias(AA.getLocationForDest(M), AA.getLocationForSource(MDep))) 658 UseMemMove = true; 659 660 // If all checks passed, then we can transform M. 661 662 // Make sure to use the lesser of the alignment of the source and the dest 663 // since we're changing where we're reading from, but don't want to increase 664 // the alignment past what can be read from or written to. 665 // TODO: Is this worth it if we're creating a less aligned memcpy? For 666 // example we could be moving from movaps -> movq on x86. 667 unsigned Align = std::min(MDep->getAlignment(), M->getAlignment()); 668 669 IRBuilder<> Builder(M); 670 if (UseMemMove) 671 Builder.CreateMemMove(M->getRawDest(), MDep->getRawSource(), M->getLength(), 672 Align, M->isVolatile()); 673 else 674 Builder.CreateMemCpy(M->getRawDest(), MDep->getRawSource(), M->getLength(), 675 Align, M->isVolatile()); 676 677 // Remove the instruction we're replacing. 678 MD->removeInstruction(M); 679 M->eraseFromParent(); 680 ++NumMemCpyInstr; 681 return true; 682} 683 684 685/// processMemCpy - perform simplification of memcpy's. If we have memcpy A 686/// which copies X to Y, and memcpy B which copies Y to Z, then we can rewrite 687/// B to be a memcpy from X to Z (or potentially a memmove, depending on 688/// circumstances). This allows later passes to remove the first memcpy 689/// altogether. 690bool MemCpyOpt::processMemCpy(MemCpyInst *M) { 691 // We can only optimize statically-sized memcpy's that are non-volatile. 692 ConstantInt *CopySize = dyn_cast<ConstantInt>(M->getLength()); 693 if (CopySize == 0 || M->isVolatile()) return false; 694 695 // If the source and destination of the memcpy are the same, then zap it. 696 if (M->getSource() == M->getDest()) { 697 MD->removeInstruction(M); 698 M->eraseFromParent(); 699 return false; 700 } 701 702 // If copying from a constant, try to turn the memcpy into a memset. 703 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(M->getSource())) 704 if (GV->isConstant() && GV->hasDefinitiveInitializer()) 705 if (Value *ByteVal = isBytewiseValue(GV->getInitializer())) { 706 IRBuilder<> Builder(M); 707 Builder.CreateMemSet(M->getRawDest(), ByteVal, CopySize, 708 M->getAlignment(), false); 709 MD->removeInstruction(M); 710 M->eraseFromParent(); 711 ++NumCpyToSet; 712 return true; 713 } 714 715 // The are two possible optimizations we can do for memcpy: 716 // a) memcpy-memcpy xform which exposes redundance for DSE. 717 // b) call-memcpy xform for return slot optimization. 718 MemDepResult DepInfo = MD->getDependency(M); 719 if (!DepInfo.isClobber()) 720 return false; 721 722 if (MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst())) 723 return processMemCpyMemCpyDependence(M, MDep, CopySize->getZExtValue()); 724 725 if (CallInst *C = dyn_cast<CallInst>(DepInfo.getInst())) { 726 if (performCallSlotOptzn(M, M->getDest(), M->getSource(), 727 CopySize->getZExtValue(), C)) { 728 M->eraseFromParent(); 729 return true; 730 } 731 } 732 return false; 733} 734 735/// processMemMove - Transforms memmove calls to memcpy calls when the src/dst 736/// are guaranteed not to alias. 737bool MemCpyOpt::processMemMove(MemMoveInst *M) { 738 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 739 740 // See if the pointers alias. 741 if (!AA.isNoAlias(AA.getLocationForDest(M), AA.getLocationForSource(M))) 742 return false; 743 744 DEBUG(dbgs() << "MemCpyOpt: Optimizing memmove -> memcpy: " << *M << "\n"); 745 746 // If not, then we know we can transform this. 747 Module *Mod = M->getParent()->getParent()->getParent(); 748 const Type *ArgTys[3] = { M->getRawDest()->getType(), 749 M->getRawSource()->getType(), 750 M->getLength()->getType() }; 751 M->setCalledFunction(Intrinsic::getDeclaration(Mod, Intrinsic::memcpy, 752 ArgTys, 3)); 753 754 // MemDep may have over conservative information about this instruction, just 755 // conservatively flush it from the cache. 756 MD->removeInstruction(M); 757 758 ++NumMoveToCpy; 759 return true; 760} 761 762/// processByValArgument - This is called on every byval argument in call sites. 763bool MemCpyOpt::processByValArgument(CallSite CS, unsigned ArgNo) { 764 if (TD == 0) return false; 765 766 // Find out what feeds this byval argument. 767 Value *ByValArg = CS.getArgument(ArgNo); 768 const Type *ByValTy =cast<PointerType>(ByValArg->getType())->getElementType(); 769 uint64_t ByValSize = TD->getTypeAllocSize(ByValTy); 770 MemDepResult DepInfo = 771 MD->getPointerDependencyFrom(AliasAnalysis::Location(ByValArg, ByValSize), 772 true, CS.getInstruction(), 773 CS.getInstruction()->getParent()); 774 if (!DepInfo.isClobber()) 775 return false; 776 777 // If the byval argument isn't fed by a memcpy, ignore it. If it is fed by 778 // a memcpy, see if we can byval from the source of the memcpy instead of the 779 // result. 780 MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst()); 781 if (MDep == 0 || MDep->isVolatile() || 782 ByValArg->stripPointerCasts() != MDep->getDest()) 783 return false; 784 785 // The length of the memcpy must be larger or equal to the size of the byval. 786 ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength()); 787 if (C1 == 0 || C1->getValue().getZExtValue() < ByValSize) 788 return false; 789 790 // Get the alignment of the byval. If it is greater than the memcpy, then we 791 // can't do the substitution. If the call doesn't specify the alignment, then 792 // it is some target specific value that we can't know. 793 unsigned ByValAlign = CS.getParamAlignment(ArgNo+1); 794 if (ByValAlign == 0 || MDep->getAlignment() < ByValAlign) 795 return false; 796 797 // Verify that the copied-from memory doesn't change in between the memcpy and 798 // the byval call. 799 // memcpy(a <- b) 800 // *b = 42; 801 // foo(*a) 802 // It would be invalid to transform the second memcpy into foo(*b). 803 // 804 // NOTE: This is conservative, it will stop on any read from the source loc, 805 // not just the defining memcpy. 806 MemDepResult SourceDep = 807 MD->getPointerDependencyFrom(AliasAnalysis::getLocationForSource(MDep), 808 false, CS.getInstruction(), MDep->getParent()); 809 if (!SourceDep.isClobber() || SourceDep.getInst() != MDep) 810 return false; 811 812 Value *TmpCast = MDep->getSource(); 813 if (MDep->getSource()->getType() != ByValArg->getType()) 814 TmpCast = new BitCastInst(MDep->getSource(), ByValArg->getType(), 815 "tmpcast", CS.getInstruction()); 816 817 DEBUG(dbgs() << "MemCpyOpt: Forwarding memcpy to byval:\n" 818 << " " << *MDep << "\n" 819 << " " << *CS.getInstruction() << "\n"); 820 821 // Otherwise we're good! Update the byval argument. 822 CS.setArgument(ArgNo, TmpCast); 823 ++NumMemCpyInstr; 824 return true; 825} 826 827/// iterateOnFunction - Executes one iteration of MemCpyOpt. 828bool MemCpyOpt::iterateOnFunction(Function &F) { 829 bool MadeChange = false; 830 831 // Walk all instruction in the function. 832 for (Function::iterator BB = F.begin(), BBE = F.end(); BB != BBE; ++BB) { 833 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE;) { 834 // Avoid invalidating the iterator. 835 Instruction *I = BI++; 836 837 bool RepeatInstruction = false; 838 839 if (StoreInst *SI = dyn_cast<StoreInst>(I)) 840 MadeChange |= processStore(SI, BI); 841 else if (MemCpyInst *M = dyn_cast<MemCpyInst>(I)) { 842 RepeatInstruction = processMemCpy(M); 843 } else if (MemMoveInst *M = dyn_cast<MemMoveInst>(I)) { 844 RepeatInstruction = processMemMove(M); 845 } else if (CallSite CS = (Value*)I) { 846 for (unsigned i = 0, e = CS.arg_size(); i != e; ++i) 847 if (CS.paramHasAttr(i+1, Attribute::ByVal)) 848 MadeChange |= processByValArgument(CS, i); 849 } 850 851 // Reprocess the instruction if desired. 852 if (RepeatInstruction) { 853 --BI; 854 MadeChange = true; 855 } 856 } 857 } 858 859 return MadeChange; 860} 861 862// MemCpyOpt::runOnFunction - This is the main transformation entry point for a 863// function. 864// 865bool MemCpyOpt::runOnFunction(Function &F) { 866 bool MadeChange = false; 867 MD = &getAnalysis<MemoryDependenceAnalysis>(); 868 TD = getAnalysisIfAvailable<TargetData>(); 869 while (1) { 870 if (!iterateOnFunction(F)) 871 break; 872 MadeChange = true; 873 } 874 875 MD = 0; 876 return MadeChange; 877} 878