MemCpyOptimizer.cpp revision 9fa11e94b5a7709cf05396420b3b3eaad6fb8e37
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 Ptr1 = Ptr1->stripPointerCasts(); 75 Ptr2 = Ptr2->stripPointerCasts(); 76 GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(Ptr1); 77 GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(Ptr2); 78 79 bool VariableIdxFound = false; 80 81 // If one pointer is a GEP and the other isn't, then see if the GEP is a 82 // constant offset from the base, as in "P" and "gep P, 1". 83 if (GEP1 && GEP2 == 0 && GEP1->getOperand(0)->stripPointerCasts() == Ptr2) { 84 Offset = -GetOffsetFromIndex(GEP1, 1, VariableIdxFound, TD); 85 return !VariableIdxFound; 86 } 87 88 if (GEP2 && GEP1 == 0 && GEP2->getOperand(0)->stripPointerCasts() == Ptr1) { 89 Offset = GetOffsetFromIndex(GEP2, 1, VariableIdxFound, TD); 90 return !VariableIdxFound; 91 } 92 93 // Right now we handle the case when Ptr1/Ptr2 are both GEPs with an identical 94 // base. After that base, they may have some number of common (and 95 // potentially variable) indices. After that they handle some constant 96 // offset, which determines their offset from each other. At this point, we 97 // handle no other case. 98 if (!GEP1 || !GEP2 || GEP1->getOperand(0) != GEP2->getOperand(0)) 99 return false; 100 101 // Skip any common indices and track the GEP types. 102 unsigned Idx = 1; 103 for (; Idx != GEP1->getNumOperands() && Idx != GEP2->getNumOperands(); ++Idx) 104 if (GEP1->getOperand(Idx) != GEP2->getOperand(Idx)) 105 break; 106 107 int64_t Offset1 = GetOffsetFromIndex(GEP1, Idx, VariableIdxFound, TD); 108 int64_t Offset2 = GetOffsetFromIndex(GEP2, Idx, VariableIdxFound, TD); 109 if (VariableIdxFound) return false; 110 111 Offset = Offset2-Offset1; 112 return true; 113} 114 115 116/// MemsetRange - Represents a range of memset'd bytes with the ByteVal value. 117/// This allows us to analyze stores like: 118/// store 0 -> P+1 119/// store 0 -> P+0 120/// store 0 -> P+3 121/// store 0 -> P+2 122/// which sometimes happens with stores to arrays of structs etc. When we see 123/// the first store, we make a range [1, 2). The second store extends the range 124/// to [0, 2). The third makes a new range [2, 3). The fourth store joins the 125/// two ranges into [0, 3) which is memset'able. 126namespace { 127struct MemsetRange { 128 // Start/End - A semi range that describes the span that this range covers. 129 // The range is closed at the start and open at the end: [Start, End). 130 int64_t Start, End; 131 132 /// StartPtr - The getelementptr instruction that points to the start of the 133 /// range. 134 Value *StartPtr; 135 136 /// Alignment - The known alignment of the first store. 137 unsigned Alignment; 138 139 /// TheStores - The actual stores that make up this range. 140 SmallVector<Instruction*, 16> TheStores; 141 142 bool isProfitableToUseMemset(const TargetData &TD) const; 143 144}; 145} // end anon namespace 146 147bool MemsetRange::isProfitableToUseMemset(const TargetData &TD) const { 148 // If we found more than 8 stores to merge or 64 bytes, use memset. 149 if (TheStores.size() >= 8 || End-Start >= 64) return true; 150 151 // If there is nothing to merge, don't do anything. 152 if (TheStores.size() < 2) return false; 153 154 // If any of the stores are a memset, then it is always good to extend the 155 // memset. 156 for (unsigned i = 0, e = TheStores.size(); i != e; ++i) 157 if (!isa<StoreInst>(TheStores[i])) 158 return true; 159 160 // Assume that the code generator is capable of merging pairs of stores 161 // together if it wants to. 162 if (TheStores.size() == 2) return false; 163 164 // If we have fewer than 8 stores, it can still be worthwhile to do this. 165 // For example, merging 4 i8 stores into an i32 store is useful almost always. 166 // However, merging 2 32-bit stores isn't useful on a 32-bit architecture (the 167 // memset will be split into 2 32-bit stores anyway) and doing so can 168 // pessimize the llvm optimizer. 169 // 170 // Since we don't have perfect knowledge here, make some assumptions: assume 171 // the maximum GPR width is the same size as the pointer size and assume that 172 // this width can be stored. If so, check to see whether we will end up 173 // actually reducing the number of stores used. 174 unsigned Bytes = unsigned(End-Start); 175 unsigned NumPointerStores = Bytes/TD.getPointerSize(); 176 177 // Assume the remaining bytes if any are done a byte at a time. 178 unsigned NumByteStores = Bytes - NumPointerStores*TD.getPointerSize(); 179 180 // If we will reduce the # stores (according to this heuristic), do the 181 // transformation. This encourages merging 4 x i8 -> i32 and 2 x i16 -> i32 182 // etc. 183 return TheStores.size() > NumPointerStores+NumByteStores; 184} 185 186 187namespace { 188class MemsetRanges { 189 /// Ranges - A sorted list of the memset ranges. We use std::list here 190 /// because each element is relatively large and expensive to copy. 191 std::list<MemsetRange> Ranges; 192 typedef std::list<MemsetRange>::iterator range_iterator; 193 const TargetData &TD; 194public: 195 MemsetRanges(const TargetData &td) : TD(td) {} 196 197 typedef std::list<MemsetRange>::const_iterator const_iterator; 198 const_iterator begin() const { return Ranges.begin(); } 199 const_iterator end() const { return Ranges.end(); } 200 bool empty() const { return Ranges.empty(); } 201 202 void addInst(int64_t OffsetFromFirst, Instruction *Inst) { 203 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) 204 addStore(OffsetFromFirst, SI); 205 else 206 addMemSet(OffsetFromFirst, cast<MemSetInst>(Inst)); 207 } 208 209 void addStore(int64_t OffsetFromFirst, StoreInst *SI) { 210 int64_t StoreSize = TD.getTypeStoreSize(SI->getOperand(0)->getType()); 211 212 addRange(OffsetFromFirst, StoreSize, 213 SI->getPointerOperand(), SI->getAlignment(), SI); 214 } 215 216 void addMemSet(int64_t OffsetFromFirst, MemSetInst *MSI) { 217 int64_t Size = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 218 addRange(OffsetFromFirst, Size, MSI->getDest(), MSI->getAlignment(), MSI); 219 } 220 221 void addRange(int64_t Start, int64_t Size, Value *Ptr, 222 unsigned Alignment, Instruction *Inst); 223 224}; 225 226} // end anon namespace 227 228 229/// addRange - Add a new store to the MemsetRanges data structure. This adds a 230/// new range for the specified store at the specified offset, merging into 231/// existing ranges as appropriate. 232/// 233/// Do a linear search of the ranges to see if this can be joined and/or to 234/// find the insertion point in the list. We keep the ranges sorted for 235/// simplicity here. This is a linear search of a linked list, which is ugly, 236/// however the number of ranges is limited, so this won't get crazy slow. 237void MemsetRanges::addRange(int64_t Start, int64_t Size, Value *Ptr, 238 unsigned Alignment, Instruction *Inst) { 239 int64_t End = Start+Size; 240 range_iterator I = Ranges.begin(), E = Ranges.end(); 241 242 while (I != E && Start > I->End) 243 ++I; 244 245 // We now know that I == E, in which case we didn't find anything to merge 246 // with, or that Start <= I->End. If End < I->Start or I == E, then we need 247 // to insert a new range. Handle this now. 248 if (I == E || End < I->Start) { 249 MemsetRange &R = *Ranges.insert(I, MemsetRange()); 250 R.Start = Start; 251 R.End = End; 252 R.StartPtr = Ptr; 253 R.Alignment = Alignment; 254 R.TheStores.push_back(Inst); 255 return; 256 } 257 258 // This store overlaps with I, add it. 259 I->TheStores.push_back(Inst); 260 261 // At this point, we may have an interval that completely contains our store. 262 // If so, just add it to the interval and return. 263 if (I->Start <= Start && I->End >= End) 264 return; 265 266 // Now we know that Start <= I->End and End >= I->Start so the range overlaps 267 // but is not entirely contained within the range. 268 269 // See if the range extends the start of the range. In this case, it couldn't 270 // possibly cause it to join the prior range, because otherwise we would have 271 // stopped on *it*. 272 if (Start < I->Start) { 273 I->Start = Start; 274 I->StartPtr = Ptr; 275 I->Alignment = Alignment; 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 MemoryDependenceAnalysis *MD; 302 const TargetData *TD; 303 public: 304 static char ID; // Pass identification, replacement for typeid 305 MemCpyOpt() : FunctionPass(ID) { 306 initializeMemCpyOptPass(*PassRegistry::getPassRegistry()); 307 MD = 0; 308 } 309 310 bool runOnFunction(Function &F); 311 312 private: 313 // This transformation requires dominator postdominator info 314 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 315 AU.setPreservesCFG(); 316 AU.addRequired<DominatorTree>(); 317 AU.addRequired<MemoryDependenceAnalysis>(); 318 AU.addRequired<AliasAnalysis>(); 319 AU.addPreserved<AliasAnalysis>(); 320 AU.addPreserved<MemoryDependenceAnalysis>(); 321 } 322 323 // Helper fuctions 324 bool processStore(StoreInst *SI, BasicBlock::iterator &BBI); 325 bool processMemCpy(MemCpyInst *M); 326 bool processMemMove(MemMoveInst *M); 327 bool performCallSlotOptzn(Instruction *cpy, Value *cpyDst, Value *cpySrc, 328 uint64_t cpyLen, CallInst *C); 329 bool processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep, 330 uint64_t MSize); 331 bool processByValArgument(CallSite CS, unsigned ArgNo); 332 Instruction *tryMergingIntoMemset(Instruction *I, Value *StartPtr, 333 Value *ByteVal); 334 335 bool iterateOnFunction(Function &F); 336 }; 337 338 char MemCpyOpt::ID = 0; 339} 340 341// createMemCpyOptPass - The public interface to this file... 342FunctionPass *llvm::createMemCpyOptPass() { return new MemCpyOpt(); } 343 344INITIALIZE_PASS_BEGIN(MemCpyOpt, "memcpyopt", "MemCpy Optimization", 345 false, false) 346INITIALIZE_PASS_DEPENDENCY(DominatorTree) 347INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis) 348INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 349INITIALIZE_PASS_END(MemCpyOpt, "memcpyopt", "MemCpy Optimization", 350 false, false) 351 352/// tryMergingIntoMemset - When scanning forward over instructions, we look for 353/// some other patterns to fold away. In particular, this looks for stores to 354/// neighboring locations of memory. If it sees enough consequtive ones, it 355/// attempts to merge them together into a memcpy/memset. 356Instruction *MemCpyOpt::tryMergingIntoMemset(Instruction *StartInst, 357 Value *StartPtr, Value *ByteVal) { 358 if (TD == 0) return 0; 359 360 // Okay, so we now have a single store that can be splatable. Scan to find 361 // all subsequent stores of the same value to offset from the same pointer. 362 // Join these together into ranges, so we can decide whether contiguous blocks 363 // are stored. 364 MemsetRanges Ranges(*TD); 365 366 BasicBlock::iterator BI = StartInst; 367 for (++BI; !isa<TerminatorInst>(BI); ++BI) { 368 if (!isa<StoreInst>(BI) && !isa<MemSetInst>(BI)) { 369 // If the instruction is readnone, ignore it, otherwise bail out. We 370 // don't even 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 (BI->mayWriteToMemory() || BI->mayReadFromMemory()) 373 break; 374 continue; 375 } 376 377 if (StoreInst *NextStore = dyn_cast<StoreInst>(BI)) { 378 // If this is a store, see if we can merge it in. 379 if (NextStore->isVolatile()) break; 380 381 // Check to see if this stored value is of the same byte-splattable value. 382 if (ByteVal != isBytewiseValue(NextStore->getOperand(0))) 383 break; 384 385 // Check to see if this store is to a constant offset from the start ptr. 386 int64_t Offset; 387 if (!IsPointerOffset(StartPtr, NextStore->getPointerOperand(), Offset, *TD)) 388 break; 389 390 Ranges.addStore(Offset, NextStore); 391 } else { 392 MemSetInst *MSI = cast<MemSetInst>(BI); 393 394 if (MSI->isVolatile() || ByteVal != MSI->getValue() || 395 !isa<ConstantInt>(MSI->getLength())) 396 break; 397 398 // Check to see if this store is to a constant offset from the start ptr. 399 int64_t Offset; 400 if (!IsPointerOffset(StartPtr, MSI->getDest(), Offset, *TD)) 401 break; 402 403 Ranges.addMemSet(Offset, MSI); 404 } 405 } 406 407 // If we have no ranges, then we just had a single store with nothing that 408 // could be merged in. This is a very common case of course. 409 if (Ranges.empty()) 410 return 0; 411 412 // If we had at least one store that could be merged in, add the starting 413 // store as well. We try to avoid this unless there is at least something 414 // interesting as a small compile-time optimization. 415 Ranges.addInst(0, StartInst); 416 417 // If we create any memsets, we put it right before the first instruction that 418 // isn't part of the memset block. This ensure that the memset is dominated 419 // by any addressing instruction needed by the start of the block. 420 IRBuilder<> Builder(BI); 421 422 // Now that we have full information about ranges, loop over the ranges and 423 // emit memset's for anything big enough to be worthwhile. 424 Instruction *AMemSet = 0; 425 for (MemsetRanges::const_iterator I = Ranges.begin(), E = Ranges.end(); 426 I != E; ++I) { 427 const MemsetRange &Range = *I; 428 429 if (Range.TheStores.size() == 1) continue; 430 431 // If it is profitable to lower this range to memset, do so now. 432 if (!Range.isProfitableToUseMemset(*TD)) 433 continue; 434 435 // Otherwise, we do want to transform this! Create a new memset. 436 // Get the starting pointer of the block. 437 StartPtr = Range.StartPtr; 438 439 // Determine alignment 440 unsigned Alignment = Range.Alignment; 441 if (Alignment == 0) { 442 const Type *EltType = 443 cast<PointerType>(StartPtr->getType())->getElementType(); 444 Alignment = TD->getABITypeAlignment(EltType); 445 } 446 447 AMemSet = 448 Builder.CreateMemSet(StartPtr, ByteVal, Range.End-Range.Start, Alignment); 449 450 DEBUG(dbgs() << "Replace stores:\n"; 451 for (unsigned i = 0, e = Range.TheStores.size(); i != e; ++i) 452 dbgs() << *Range.TheStores[i] << '\n'; 453 dbgs() << "With: " << *AMemSet << '\n'); 454 455 // Zap all the stores. 456 for (SmallVector<Instruction*, 16>::const_iterator 457 SI = Range.TheStores.begin(), 458 SE = Range.TheStores.end(); SI != SE; ++SI) 459 (*SI)->eraseFromParent(); 460 ++NumMemSetInfer; 461 } 462 463 return AMemSet; 464} 465 466 467bool MemCpyOpt::processStore(StoreInst *SI, BasicBlock::iterator &BBI) { 468 if (SI->isVolatile()) return false; 469 470 if (TD == 0) return false; 471 472 // Detect cases where we're performing call slot forwarding, but 473 // happen to be using a load-store pair to implement it, rather than 474 // a memcpy. 475 if (LoadInst *LI = dyn_cast<LoadInst>(SI->getOperand(0))) { 476 if (!LI->isVolatile() && LI->hasOneUse()) { 477 MemDepResult dep = MD->getDependency(LI); 478 CallInst *C = 0; 479 if (dep.isClobber() && !isa<MemCpyInst>(dep.getInst())) 480 C = dyn_cast<CallInst>(dep.getInst()); 481 482 if (C) { 483 bool changed = performCallSlotOptzn(LI, 484 SI->getPointerOperand()->stripPointerCasts(), 485 LI->getPointerOperand()->stripPointerCasts(), 486 TD->getTypeStoreSize(SI->getOperand(0)->getType()), C); 487 if (changed) { 488 MD->removeInstruction(SI); 489 SI->eraseFromParent(); 490 LI->eraseFromParent(); 491 ++NumMemCpyInstr; 492 return true; 493 } 494 } 495 } 496 } 497 498 // There are two cases that are interesting for this code to handle: memcpy 499 // and memset. Right now we only handle memset. 500 501 // Ensure that the value being stored is something that can be memset'able a 502 // byte at a time like "0" or "-1" or any width, as well as things like 503 // 0xA0A0A0A0 and 0.0. 504 if (Value *ByteVal = isBytewiseValue(SI->getOperand(0))) 505 if (Instruction *I = tryMergingIntoMemset(SI, SI->getPointerOperand(), 506 ByteVal)) { 507 BBI = I; // Don't invalidate iterator. 508 return true; 509 } 510 511 return false; 512} 513 514 515/// performCallSlotOptzn - takes a memcpy and a call that it depends on, 516/// and checks for the possibility of a call slot optimization by having 517/// the call write its result directly into the destination of the memcpy. 518bool MemCpyOpt::performCallSlotOptzn(Instruction *cpy, 519 Value *cpyDest, Value *cpySrc, 520 uint64_t cpyLen, CallInst *C) { 521 // The general transformation to keep in mind is 522 // 523 // call @func(..., src, ...) 524 // memcpy(dest, src, ...) 525 // 526 // -> 527 // 528 // memcpy(dest, src, ...) 529 // call @func(..., dest, ...) 530 // 531 // Since moving the memcpy is technically awkward, we additionally check that 532 // src only holds uninitialized values at the moment of the call, meaning that 533 // the memcpy can be discarded rather than moved. 534 535 // Deliberately get the source and destination with bitcasts stripped away, 536 // because we'll need to do type comparisons based on the underlying type. 537 CallSite CS(C); 538 539 // Require that src be an alloca. This simplifies the reasoning considerably. 540 AllocaInst *srcAlloca = dyn_cast<AllocaInst>(cpySrc); 541 if (!srcAlloca) 542 return false; 543 544 // Check that all of src is copied to dest. 545 if (TD == 0) return false; 546 547 ConstantInt *srcArraySize = dyn_cast<ConstantInt>(srcAlloca->getArraySize()); 548 if (!srcArraySize) 549 return false; 550 551 uint64_t srcSize = TD->getTypeAllocSize(srcAlloca->getAllocatedType()) * 552 srcArraySize->getZExtValue(); 553 554 if (cpyLen < srcSize) 555 return false; 556 557 // Check that accessing the first srcSize bytes of dest will not cause a 558 // trap. Otherwise the transform is invalid since it might cause a trap 559 // to occur earlier than it otherwise would. 560 if (AllocaInst *A = dyn_cast<AllocaInst>(cpyDest)) { 561 // The destination is an alloca. Check it is larger than srcSize. 562 ConstantInt *destArraySize = dyn_cast<ConstantInt>(A->getArraySize()); 563 if (!destArraySize) 564 return false; 565 566 uint64_t destSize = TD->getTypeAllocSize(A->getAllocatedType()) * 567 destArraySize->getZExtValue(); 568 569 if (destSize < srcSize) 570 return false; 571 } else if (Argument *A = dyn_cast<Argument>(cpyDest)) { 572 // If the destination is an sret parameter then only accesses that are 573 // outside of the returned struct type can trap. 574 if (!A->hasStructRetAttr()) 575 return false; 576 577 const Type *StructTy = cast<PointerType>(A->getType())->getElementType(); 578 uint64_t destSize = TD->getTypeAllocSize(StructTy); 579 580 if (destSize < srcSize) 581 return false; 582 } else { 583 return false; 584 } 585 586 // Check that src is not accessed except via the call and the memcpy. This 587 // guarantees that it holds only undefined values when passed in (so the final 588 // memcpy can be dropped), that it is not read or written between the call and 589 // the memcpy, and that writing beyond the end of it is undefined. 590 SmallVector<User*, 8> srcUseList(srcAlloca->use_begin(), 591 srcAlloca->use_end()); 592 while (!srcUseList.empty()) { 593 User *UI = srcUseList.pop_back_val(); 594 595 if (isa<BitCastInst>(UI)) { 596 for (User::use_iterator I = UI->use_begin(), E = UI->use_end(); 597 I != E; ++I) 598 srcUseList.push_back(*I); 599 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(UI)) { 600 if (G->hasAllZeroIndices()) 601 for (User::use_iterator I = UI->use_begin(), E = UI->use_end(); 602 I != E; ++I) 603 srcUseList.push_back(*I); 604 else 605 return false; 606 } else if (UI != C && UI != cpy) { 607 return false; 608 } 609 } 610 611 // Since we're changing the parameter to the callsite, we need to make sure 612 // that what would be the new parameter dominates the callsite. 613 DominatorTree &DT = getAnalysis<DominatorTree>(); 614 if (Instruction *cpyDestInst = dyn_cast<Instruction>(cpyDest)) 615 if (!DT.dominates(cpyDestInst, C)) 616 return false; 617 618 // In addition to knowing that the call does not access src in some 619 // unexpected manner, for example via a global, which we deduce from 620 // the use analysis, we also need to know that it does not sneakily 621 // access dest. We rely on AA to figure this out for us. 622 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 623 if (AA.getModRefInfo(C, cpyDest, srcSize) != AliasAnalysis::NoModRef) 624 return false; 625 626 // All the checks have passed, so do the transformation. 627 bool changedArgument = false; 628 for (unsigned i = 0; i < CS.arg_size(); ++i) 629 if (CS.getArgument(i)->stripPointerCasts() == cpySrc) { 630 if (cpySrc->getType() != cpyDest->getType()) 631 cpyDest = CastInst::CreatePointerCast(cpyDest, cpySrc->getType(), 632 cpyDest->getName(), C); 633 changedArgument = true; 634 if (CS.getArgument(i)->getType() == cpyDest->getType()) 635 CS.setArgument(i, cpyDest); 636 else 637 CS.setArgument(i, CastInst::CreatePointerCast(cpyDest, 638 CS.getArgument(i)->getType(), cpyDest->getName(), C)); 639 } 640 641 if (!changedArgument) 642 return false; 643 644 // Drop any cached information about the call, because we may have changed 645 // its dependence information by changing its parameter. 646 MD->removeInstruction(C); 647 648 // Remove the memcpy. 649 MD->removeInstruction(cpy); 650 ++NumMemCpyInstr; 651 652 return true; 653} 654 655/// processMemCpyMemCpyDependence - We've found that the (upward scanning) 656/// memory dependence of memcpy 'M' is the memcpy 'MDep'. Try to simplify M to 657/// copy from MDep's input if we can. MSize is the size of M's copy. 658/// 659bool MemCpyOpt::processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep, 660 uint64_t MSize) { 661 // We can only transforms memcpy's where the dest of one is the source of the 662 // other. 663 if (M->getSource() != MDep->getDest() || MDep->isVolatile()) 664 return false; 665 666 // If dep instruction is reading from our current input, then it is a noop 667 // transfer and substituting the input won't change this instruction. Just 668 // ignore the input and let someone else zap MDep. This handles cases like: 669 // memcpy(a <- a) 670 // memcpy(b <- a) 671 if (M->getSource() == MDep->getSource()) 672 return false; 673 674 // Second, the length of the memcpy's must be the same, or the preceeding one 675 // must be larger than the following one. 676 ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength()); 677 if (!C1) return false; 678 679 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 680 681 // Verify that the copied-from memory doesn't change in between the two 682 // transfers. For example, in: 683 // memcpy(a <- b) 684 // *b = 42; 685 // memcpy(c <- a) 686 // It would be invalid to transform the second memcpy into memcpy(c <- b). 687 // 688 // TODO: If the code between M and MDep is transparent to the destination "c", 689 // then we could still perform the xform by moving M up to the first memcpy. 690 // 691 // NOTE: This is conservative, it will stop on any read from the source loc, 692 // not just the defining memcpy. 693 MemDepResult SourceDep = 694 MD->getPointerDependencyFrom(AA.getLocationForSource(MDep), 695 false, M, M->getParent()); 696 if (!SourceDep.isClobber() || SourceDep.getInst() != MDep) 697 return false; 698 699 // If the dest of the second might alias the source of the first, then the 700 // source and dest might overlap. We still want to eliminate the intermediate 701 // value, but we have to generate a memmove instead of memcpy. 702 bool UseMemMove = false; 703 if (!AA.isNoAlias(AA.getLocationForDest(M), AA.getLocationForSource(MDep))) 704 UseMemMove = true; 705 706 // If all checks passed, then we can transform M. 707 708 // Make sure to use the lesser of the alignment of the source and the dest 709 // since we're changing where we're reading from, but don't want to increase 710 // the alignment past what can be read from or written to. 711 // TODO: Is this worth it if we're creating a less aligned memcpy? For 712 // example we could be moving from movaps -> movq on x86. 713 unsigned Align = std::min(MDep->getAlignment(), M->getAlignment()); 714 715 IRBuilder<> Builder(M); 716 if (UseMemMove) 717 Builder.CreateMemMove(M->getRawDest(), MDep->getRawSource(), M->getLength(), 718 Align, M->isVolatile()); 719 else 720 Builder.CreateMemCpy(M->getRawDest(), MDep->getRawSource(), M->getLength(), 721 Align, M->isVolatile()); 722 723 // Remove the instruction we're replacing. 724 MD->removeInstruction(M); 725 M->eraseFromParent(); 726 ++NumMemCpyInstr; 727 return true; 728} 729 730 731/// processMemCpy - perform simplification of memcpy's. If we have memcpy A 732/// which copies X to Y, and memcpy B which copies Y to Z, then we can rewrite 733/// B to be a memcpy from X to Z (or potentially a memmove, depending on 734/// circumstances). This allows later passes to remove the first memcpy 735/// altogether. 736bool MemCpyOpt::processMemCpy(MemCpyInst *M) { 737 // We can only optimize statically-sized memcpy's that are non-volatile. 738 ConstantInt *CopySize = dyn_cast<ConstantInt>(M->getLength()); 739 if (CopySize == 0 || M->isVolatile()) return false; 740 741 // If the source and destination of the memcpy are the same, then zap it. 742 if (M->getSource() == M->getDest()) { 743 MD->removeInstruction(M); 744 M->eraseFromParent(); 745 return false; 746 } 747 748 // If copying from a constant, try to turn the memcpy into a memset. 749 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(M->getSource())) 750 if (GV->isConstant() && GV->hasDefinitiveInitializer()) 751 if (Value *ByteVal = isBytewiseValue(GV->getInitializer())) { 752 IRBuilder<> Builder(M); 753 Builder.CreateMemSet(M->getRawDest(), ByteVal, CopySize, 754 M->getAlignment(), false); 755 MD->removeInstruction(M); 756 M->eraseFromParent(); 757 ++NumCpyToSet; 758 return true; 759 } 760 761 // The are two possible optimizations we can do for memcpy: 762 // a) memcpy-memcpy xform which exposes redundance for DSE. 763 // b) call-memcpy xform for return slot optimization. 764 MemDepResult DepInfo = MD->getDependency(M); 765 if (!DepInfo.isClobber()) 766 return false; 767 768 if (MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst())) 769 return processMemCpyMemCpyDependence(M, MDep, CopySize->getZExtValue()); 770 771 if (CallInst *C = dyn_cast<CallInst>(DepInfo.getInst())) { 772 if (performCallSlotOptzn(M, M->getDest(), M->getSource(), 773 CopySize->getZExtValue(), C)) { 774 M->eraseFromParent(); 775 return true; 776 } 777 } 778 return false; 779} 780 781/// processMemMove - Transforms memmove calls to memcpy calls when the src/dst 782/// are guaranteed not to alias. 783bool MemCpyOpt::processMemMove(MemMoveInst *M) { 784 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 785 786 // See if the pointers alias. 787 if (!AA.isNoAlias(AA.getLocationForDest(M), AA.getLocationForSource(M))) 788 return false; 789 790 DEBUG(dbgs() << "MemCpyOpt: Optimizing memmove -> memcpy: " << *M << "\n"); 791 792 // If not, then we know we can transform this. 793 Module *Mod = M->getParent()->getParent()->getParent(); 794 const Type *ArgTys[3] = { M->getRawDest()->getType(), 795 M->getRawSource()->getType(), 796 M->getLength()->getType() }; 797 M->setCalledFunction(Intrinsic::getDeclaration(Mod, Intrinsic::memcpy, 798 ArgTys, 3)); 799 800 // MemDep may have over conservative information about this instruction, just 801 // conservatively flush it from the cache. 802 MD->removeInstruction(M); 803 804 ++NumMoveToCpy; 805 return true; 806} 807 808/// processByValArgument - This is called on every byval argument in call sites. 809bool MemCpyOpt::processByValArgument(CallSite CS, unsigned ArgNo) { 810 if (TD == 0) return false; 811 812 // Find out what feeds this byval argument. 813 Value *ByValArg = CS.getArgument(ArgNo); 814 const Type *ByValTy =cast<PointerType>(ByValArg->getType())->getElementType(); 815 uint64_t ByValSize = TD->getTypeAllocSize(ByValTy); 816 MemDepResult DepInfo = 817 MD->getPointerDependencyFrom(AliasAnalysis::Location(ByValArg, ByValSize), 818 true, CS.getInstruction(), 819 CS.getInstruction()->getParent()); 820 if (!DepInfo.isClobber()) 821 return false; 822 823 // If the byval argument isn't fed by a memcpy, ignore it. If it is fed by 824 // a memcpy, see if we can byval from the source of the memcpy instead of the 825 // result. 826 MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst()); 827 if (MDep == 0 || MDep->isVolatile() || 828 ByValArg->stripPointerCasts() != MDep->getDest()) 829 return false; 830 831 // The length of the memcpy must be larger or equal to the size of the byval. 832 ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength()); 833 if (C1 == 0 || C1->getValue().getZExtValue() < ByValSize) 834 return false; 835 836 // Get the alignment of the byval. If it is greater than the memcpy, then we 837 // can't do the substitution. If the call doesn't specify the alignment, then 838 // it is some target specific value that we can't know. 839 unsigned ByValAlign = CS.getParamAlignment(ArgNo+1); 840 if (ByValAlign == 0 || MDep->getAlignment() < ByValAlign) 841 return false; 842 843 // Verify that the copied-from memory doesn't change in between the memcpy and 844 // the byval call. 845 // memcpy(a <- b) 846 // *b = 42; 847 // foo(*a) 848 // It would be invalid to transform the second memcpy into foo(*b). 849 // 850 // NOTE: This is conservative, it will stop on any read from the source loc, 851 // not just the defining memcpy. 852 MemDepResult SourceDep = 853 MD->getPointerDependencyFrom(AliasAnalysis::getLocationForSource(MDep), 854 false, CS.getInstruction(), MDep->getParent()); 855 if (!SourceDep.isClobber() || SourceDep.getInst() != MDep) 856 return false; 857 858 Value *TmpCast = MDep->getSource(); 859 if (MDep->getSource()->getType() != ByValArg->getType()) 860 TmpCast = new BitCastInst(MDep->getSource(), ByValArg->getType(), 861 "tmpcast", CS.getInstruction()); 862 863 DEBUG(dbgs() << "MemCpyOpt: Forwarding memcpy to byval:\n" 864 << " " << *MDep << "\n" 865 << " " << *CS.getInstruction() << "\n"); 866 867 // Otherwise we're good! Update the byval argument. 868 CS.setArgument(ArgNo, TmpCast); 869 ++NumMemCpyInstr; 870 return true; 871} 872 873/// iterateOnFunction - Executes one iteration of MemCpyOpt. 874bool MemCpyOpt::iterateOnFunction(Function &F) { 875 bool MadeChange = false; 876 877 // Walk all instruction in the function. 878 for (Function::iterator BB = F.begin(), BBE = F.end(); BB != BBE; ++BB) { 879 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE;) { 880 // Avoid invalidating the iterator. 881 Instruction *I = BI++; 882 883 bool RepeatInstruction = false; 884 885 if (StoreInst *SI = dyn_cast<StoreInst>(I)) 886 MadeChange |= processStore(SI, BI); 887 else if (MemCpyInst *M = dyn_cast<MemCpyInst>(I)) { 888 RepeatInstruction = processMemCpy(M); 889 } else if (MemMoveInst *M = dyn_cast<MemMoveInst>(I)) { 890 RepeatInstruction = processMemMove(M); 891 } else if (CallSite CS = (Value*)I) { 892 for (unsigned i = 0, e = CS.arg_size(); i != e; ++i) 893 if (CS.paramHasAttr(i+1, Attribute::ByVal)) 894 MadeChange |= processByValArgument(CS, i); 895 } 896 897 // Reprocess the instruction if desired. 898 if (RepeatInstruction) { 899 --BI; 900 MadeChange = true; 901 } 902 } 903 } 904 905 return MadeChange; 906} 907 908// MemCpyOpt::runOnFunction - This is the main transformation entry point for a 909// function. 910// 911bool MemCpyOpt::runOnFunction(Function &F) { 912 bool MadeChange = false; 913 MD = &getAnalysis<MemoryDependenceAnalysis>(); 914 TD = getAnalysisIfAvailable<TargetData>(); 915 while (1) { 916 if (!iterateOnFunction(F)) 917 break; 918 MadeChange = true; 919 } 920 921 MD = 0; 922 return MadeChange; 923} 924