MemCpyOptimizer.cpp revision 221345
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 "llvm/Target/TargetLibraryInfo.h" 32#include <list> 33using namespace llvm; 34 35STATISTIC(NumMemCpyInstr, "Number of memcpy instructions deleted"); 36STATISTIC(NumMemSetInfer, "Number of memsets inferred"); 37STATISTIC(NumMoveToCpy, "Number of memmoves converted to memcpy"); 38STATISTIC(NumCpyToSet, "Number of memcpys converted to memset"); 39 40static int64_t GetOffsetFromIndex(const GetElementPtrInst *GEP, unsigned Idx, 41 bool &VariableIdxFound, const TargetData &TD){ 42 // Skip over the first indices. 43 gep_type_iterator GTI = gep_type_begin(GEP); 44 for (unsigned i = 1; i != Idx; ++i, ++GTI) 45 /*skip along*/; 46 47 // Compute the offset implied by the rest of the indices. 48 int64_t Offset = 0; 49 for (unsigned i = Idx, e = GEP->getNumOperands(); i != e; ++i, ++GTI) { 50 ConstantInt *OpC = dyn_cast<ConstantInt>(GEP->getOperand(i)); 51 if (OpC == 0) 52 return VariableIdxFound = true; 53 if (OpC->isZero()) continue; // No offset. 54 55 // Handle struct indices, which add their field offset to the pointer. 56 if (const StructType *STy = dyn_cast<StructType>(*GTI)) { 57 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue()); 58 continue; 59 } 60 61 // Otherwise, we have a sequential type like an array or vector. Multiply 62 // the index by the ElementSize. 63 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()); 64 Offset += Size*OpC->getSExtValue(); 65 } 66 67 return Offset; 68} 69 70/// IsPointerOffset - Return true if Ptr1 is provably equal to Ptr2 plus a 71/// constant offset, and return that constant offset. For example, Ptr1 might 72/// be &A[42], and Ptr2 might be &A[40]. In this case offset would be -8. 73static bool IsPointerOffset(Value *Ptr1, Value *Ptr2, int64_t &Offset, 74 const TargetData &TD) { 75 Ptr1 = Ptr1->stripPointerCasts(); 76 Ptr2 = Ptr2->stripPointerCasts(); 77 GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(Ptr1); 78 GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(Ptr2); 79 80 bool VariableIdxFound = false; 81 82 // If one pointer is a GEP and the other isn't, then see if the GEP is a 83 // constant offset from the base, as in "P" and "gep P, 1". 84 if (GEP1 && GEP2 == 0 && GEP1->getOperand(0)->stripPointerCasts() == Ptr2) { 85 Offset = -GetOffsetFromIndex(GEP1, 1, VariableIdxFound, TD); 86 return !VariableIdxFound; 87 } 88 89 if (GEP2 && GEP1 == 0 && GEP2->getOperand(0)->stripPointerCasts() == Ptr1) { 90 Offset = GetOffsetFromIndex(GEP2, 1, VariableIdxFound, TD); 91 return !VariableIdxFound; 92 } 93 94 // Right now we handle the case when Ptr1/Ptr2 are both GEPs with an identical 95 // base. After that base, they may have some number of common (and 96 // potentially variable) indices. After that they handle some constant 97 // offset, which determines their offset from each other. At this point, we 98 // handle no other case. 99 if (!GEP1 || !GEP2 || GEP1->getOperand(0) != GEP2->getOperand(0)) 100 return false; 101 102 // Skip any common indices and track the GEP types. 103 unsigned Idx = 1; 104 for (; Idx != GEP1->getNumOperands() && Idx != GEP2->getNumOperands(); ++Idx) 105 if (GEP1->getOperand(Idx) != GEP2->getOperand(Idx)) 106 break; 107 108 int64_t Offset1 = GetOffsetFromIndex(GEP1, Idx, VariableIdxFound, TD); 109 int64_t Offset2 = GetOffsetFromIndex(GEP2, Idx, VariableIdxFound, TD); 110 if (VariableIdxFound) return false; 111 112 Offset = Offset2-Offset1; 113 return true; 114} 115 116 117/// MemsetRange - Represents a range of memset'd bytes with the ByteVal value. 118/// This allows us to analyze stores like: 119/// store 0 -> P+1 120/// store 0 -> P+0 121/// store 0 -> P+3 122/// store 0 -> P+2 123/// which sometimes happens with stores to arrays of structs etc. When we see 124/// the first store, we make a range [1, 2). The second store extends the range 125/// to [0, 2). The third makes a new range [2, 3). The fourth store joins the 126/// two ranges into [0, 3) which is memset'able. 127namespace { 128struct MemsetRange { 129 // Start/End - A semi range that describes the span that this range covers. 130 // The range is closed at the start and open at the end: [Start, End). 131 int64_t Start, End; 132 133 /// StartPtr - The getelementptr instruction that points to the start of the 134 /// range. 135 Value *StartPtr; 136 137 /// Alignment - The known alignment of the first store. 138 unsigned Alignment; 139 140 /// TheStores - The actual stores that make up this range. 141 SmallVector<Instruction*, 16> TheStores; 142 143 bool isProfitableToUseMemset(const TargetData &TD) const; 144 145}; 146} // end anon namespace 147 148bool MemsetRange::isProfitableToUseMemset(const TargetData &TD) const { 149 // If we found more than 8 stores to merge or 64 bytes, use memset. 150 if (TheStores.size() >= 8 || End-Start >= 64) return true; 151 152 // If there is nothing to merge, don't do anything. 153 if (TheStores.size() < 2) return false; 154 155 // If any of the stores are a memset, then it is always good to extend the 156 // memset. 157 for (unsigned i = 0, e = TheStores.size(); i != e; ++i) 158 if (!isa<StoreInst>(TheStores[i])) 159 return true; 160 161 // Assume that the code generator is capable of merging pairs of stores 162 // together if it wants to. 163 if (TheStores.size() == 2) return false; 164 165 // If we have fewer than 8 stores, it can still be worthwhile to do this. 166 // For example, merging 4 i8 stores into an i32 store is useful almost always. 167 // However, merging 2 32-bit stores isn't useful on a 32-bit architecture (the 168 // memset will be split into 2 32-bit stores anyway) and doing so can 169 // pessimize the llvm optimizer. 170 // 171 // Since we don't have perfect knowledge here, make some assumptions: assume 172 // the maximum GPR width is the same size as the pointer size and assume that 173 // this width can be stored. If so, check to see whether we will end up 174 // actually reducing the number of stores used. 175 unsigned Bytes = unsigned(End-Start); 176 unsigned NumPointerStores = Bytes/TD.getPointerSize(); 177 178 // Assume the remaining bytes if any are done a byte at a time. 179 unsigned NumByteStores = Bytes - NumPointerStores*TD.getPointerSize(); 180 181 // If we will reduce the # stores (according to this heuristic), do the 182 // transformation. This encourages merging 4 x i8 -> i32 and 2 x i16 -> i32 183 // etc. 184 return TheStores.size() > NumPointerStores+NumByteStores; 185} 186 187 188namespace { 189class MemsetRanges { 190 /// Ranges - A sorted list of the memset ranges. We use std::list here 191 /// because each element is relatively large and expensive to copy. 192 std::list<MemsetRange> Ranges; 193 typedef std::list<MemsetRange>::iterator range_iterator; 194 const TargetData &TD; 195public: 196 MemsetRanges(const TargetData &td) : TD(td) {} 197 198 typedef std::list<MemsetRange>::const_iterator const_iterator; 199 const_iterator begin() const { return Ranges.begin(); } 200 const_iterator end() const { return Ranges.end(); } 201 bool empty() const { return Ranges.empty(); } 202 203 void addInst(int64_t OffsetFromFirst, Instruction *Inst) { 204 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) 205 addStore(OffsetFromFirst, SI); 206 else 207 addMemSet(OffsetFromFirst, cast<MemSetInst>(Inst)); 208 } 209 210 void addStore(int64_t OffsetFromFirst, StoreInst *SI) { 211 int64_t StoreSize = TD.getTypeStoreSize(SI->getOperand(0)->getType()); 212 213 addRange(OffsetFromFirst, StoreSize, 214 SI->getPointerOperand(), SI->getAlignment(), SI); 215 } 216 217 void addMemSet(int64_t OffsetFromFirst, MemSetInst *MSI) { 218 int64_t Size = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 219 addRange(OffsetFromFirst, Size, MSI->getDest(), MSI->getAlignment(), MSI); 220 } 221 222 void addRange(int64_t Start, int64_t Size, Value *Ptr, 223 unsigned Alignment, Instruction *Inst); 224 225}; 226 227} // end anon namespace 228 229 230/// addRange - Add a new store to the MemsetRanges data structure. This adds a 231/// new range for the specified store at the specified offset, merging into 232/// existing ranges as appropriate. 233/// 234/// Do a linear search of the ranges to see if this can be joined and/or to 235/// find the insertion point in the list. We keep the ranges sorted for 236/// simplicity here. This is a linear search of a linked list, which is ugly, 237/// however the number of ranges is limited, so this won't get crazy slow. 238void MemsetRanges::addRange(int64_t Start, int64_t Size, Value *Ptr, 239 unsigned Alignment, Instruction *Inst) { 240 int64_t End = Start+Size; 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 = Ptr; 254 R.Alignment = Alignment; 255 R.TheStores.push_back(Inst); 256 return; 257 } 258 259 // This store overlaps with I, add it. 260 I->TheStores.push_back(Inst); 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 = Ptr; 276 I->Alignment = Alignment; 277 } 278 279 // Now we know that Start <= I->End and Start >= I->Start (so the startpoint 280 // is in or right at the end of I), and that End >= I->Start. Extend I out to 281 // End. 282 if (End > I->End) { 283 I->End = End; 284 range_iterator NextI = I; 285 while (++NextI != E && End >= NextI->Start) { 286 // Merge the range in. 287 I->TheStores.append(NextI->TheStores.begin(), NextI->TheStores.end()); 288 if (NextI->End > I->End) 289 I->End = NextI->End; 290 Ranges.erase(NextI); 291 NextI = I; 292 } 293 } 294} 295 296//===----------------------------------------------------------------------===// 297// MemCpyOpt Pass 298//===----------------------------------------------------------------------===// 299 300namespace { 301 class MemCpyOpt : public FunctionPass { 302 MemoryDependenceAnalysis *MD; 303 TargetLibraryInfo *TLI; 304 const TargetData *TD; 305 public: 306 static char ID; // Pass identification, replacement for typeid 307 MemCpyOpt() : FunctionPass(ID) { 308 initializeMemCpyOptPass(*PassRegistry::getPassRegistry()); 309 MD = 0; 310 TLI = 0; 311 TD = 0; 312 } 313 314 bool runOnFunction(Function &F); 315 316 private: 317 // This transformation requires dominator postdominator info 318 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 319 AU.setPreservesCFG(); 320 AU.addRequired<DominatorTree>(); 321 AU.addRequired<MemoryDependenceAnalysis>(); 322 AU.addRequired<AliasAnalysis>(); 323 AU.addRequired<TargetLibraryInfo>(); 324 AU.addPreserved<AliasAnalysis>(); 325 AU.addPreserved<MemoryDependenceAnalysis>(); 326 } 327 328 // Helper fuctions 329 bool processStore(StoreInst *SI, BasicBlock::iterator &BBI); 330 bool processMemSet(MemSetInst *SI, BasicBlock::iterator &BBI); 331 bool processMemCpy(MemCpyInst *M); 332 bool processMemMove(MemMoveInst *M); 333 bool performCallSlotOptzn(Instruction *cpy, Value *cpyDst, Value *cpySrc, 334 uint64_t cpyLen, CallInst *C); 335 bool processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep, 336 uint64_t MSize); 337 bool processByValArgument(CallSite CS, unsigned ArgNo); 338 Instruction *tryMergingIntoMemset(Instruction *I, Value *StartPtr, 339 Value *ByteVal); 340 341 bool iterateOnFunction(Function &F); 342 }; 343 344 char MemCpyOpt::ID = 0; 345} 346 347// createMemCpyOptPass - The public interface to this file... 348FunctionPass *llvm::createMemCpyOptPass() { return new MemCpyOpt(); } 349 350INITIALIZE_PASS_BEGIN(MemCpyOpt, "memcpyopt", "MemCpy Optimization", 351 false, false) 352INITIALIZE_PASS_DEPENDENCY(DominatorTree) 353INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis) 354INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) 355INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 356INITIALIZE_PASS_END(MemCpyOpt, "memcpyopt", "MemCpy Optimization", 357 false, false) 358 359/// tryMergingIntoMemset - When scanning forward over instructions, we look for 360/// some other patterns to fold away. In particular, this looks for stores to 361/// neighboring locations of memory. If it sees enough consecutive ones, it 362/// attempts to merge them together into a memcpy/memset. 363Instruction *MemCpyOpt::tryMergingIntoMemset(Instruction *StartInst, 364 Value *StartPtr, Value *ByteVal) { 365 if (TD == 0) return 0; 366 367 // Okay, so we now have a single store that can be splatable. Scan to find 368 // all subsequent stores of the same value to offset from the same pointer. 369 // Join these together into ranges, so we can decide whether contiguous blocks 370 // are stored. 371 MemsetRanges Ranges(*TD); 372 373 BasicBlock::iterator BI = StartInst; 374 for (++BI; !isa<TerminatorInst>(BI); ++BI) { 375 if (!isa<StoreInst>(BI) && !isa<MemSetInst>(BI)) { 376 // If the instruction is readnone, ignore it, otherwise bail out. We 377 // don't even allow readonly here because we don't want something like: 378 // A[1] = 2; strlen(A); A[2] = 2; -> memcpy(A, ...); strlen(A). 379 if (BI->mayWriteToMemory() || BI->mayReadFromMemory()) 380 break; 381 continue; 382 } 383 384 if (StoreInst *NextStore = dyn_cast<StoreInst>(BI)) { 385 // If this is a store, see if we can merge it in. 386 if (NextStore->isVolatile()) break; 387 388 // Check to see if this stored value is of the same byte-splattable value. 389 if (ByteVal != isBytewiseValue(NextStore->getOperand(0))) 390 break; 391 392 // Check to see if this store is to a constant offset from the start ptr. 393 int64_t Offset; 394 if (!IsPointerOffset(StartPtr, NextStore->getPointerOperand(), 395 Offset, *TD)) 396 break; 397 398 Ranges.addStore(Offset, NextStore); 399 } else { 400 MemSetInst *MSI = cast<MemSetInst>(BI); 401 402 if (MSI->isVolatile() || ByteVal != MSI->getValue() || 403 !isa<ConstantInt>(MSI->getLength())) 404 break; 405 406 // Check to see if this store is to a constant offset from the start ptr. 407 int64_t Offset; 408 if (!IsPointerOffset(StartPtr, MSI->getDest(), Offset, *TD)) 409 break; 410 411 Ranges.addMemSet(Offset, MSI); 412 } 413 } 414 415 // If we have no ranges, then we just had a single store with nothing that 416 // could be merged in. This is a very common case of course. 417 if (Ranges.empty()) 418 return 0; 419 420 // If we had at least one store that could be merged in, add the starting 421 // store as well. We try to avoid this unless there is at least something 422 // interesting as a small compile-time optimization. 423 Ranges.addInst(0, StartInst); 424 425 // If we create any memsets, we put it right before the first instruction that 426 // isn't part of the memset block. This ensure that the memset is dominated 427 // by any addressing instruction needed by the start of the block. 428 IRBuilder<> Builder(BI); 429 430 // Now that we have full information about ranges, loop over the ranges and 431 // emit memset's for anything big enough to be worthwhile. 432 Instruction *AMemSet = 0; 433 for (MemsetRanges::const_iterator I = Ranges.begin(), E = Ranges.end(); 434 I != E; ++I) { 435 const MemsetRange &Range = *I; 436 437 if (Range.TheStores.size() == 1) continue; 438 439 // If it is profitable to lower this range to memset, do so now. 440 if (!Range.isProfitableToUseMemset(*TD)) 441 continue; 442 443 // Otherwise, we do want to transform this! Create a new memset. 444 // Get the starting pointer of the block. 445 StartPtr = Range.StartPtr; 446 447 // Determine alignment 448 unsigned Alignment = Range.Alignment; 449 if (Alignment == 0) { 450 const Type *EltType = 451 cast<PointerType>(StartPtr->getType())->getElementType(); 452 Alignment = TD->getABITypeAlignment(EltType); 453 } 454 455 AMemSet = 456 Builder.CreateMemSet(StartPtr, ByteVal, Range.End-Range.Start, Alignment); 457 458 DEBUG(dbgs() << "Replace stores:\n"; 459 for (unsigned i = 0, e = Range.TheStores.size(); i != e; ++i) 460 dbgs() << *Range.TheStores[i] << '\n'; 461 dbgs() << "With: " << *AMemSet << '\n'); 462 463 // Zap all the stores. 464 for (SmallVector<Instruction*, 16>::const_iterator 465 SI = Range.TheStores.begin(), 466 SE = Range.TheStores.end(); SI != SE; ++SI) { 467 MD->removeInstruction(*SI); 468 (*SI)->eraseFromParent(); 469 } 470 ++NumMemSetInfer; 471 } 472 473 return AMemSet; 474} 475 476 477bool MemCpyOpt::processStore(StoreInst *SI, BasicBlock::iterator &BBI) { 478 if (SI->isVolatile()) return false; 479 480 if (TD == 0) return false; 481 482 // Detect cases where we're performing call slot forwarding, but 483 // happen to be using a load-store pair to implement it, rather than 484 // a memcpy. 485 if (LoadInst *LI = dyn_cast<LoadInst>(SI->getOperand(0))) { 486 if (!LI->isVolatile() && LI->hasOneUse()) { 487 MemDepResult dep = MD->getDependency(LI); 488 CallInst *C = 0; 489 if (dep.isClobber() && !isa<MemCpyInst>(dep.getInst())) 490 C = dyn_cast<CallInst>(dep.getInst()); 491 492 if (C) { 493 bool changed = performCallSlotOptzn(LI, 494 SI->getPointerOperand()->stripPointerCasts(), 495 LI->getPointerOperand()->stripPointerCasts(), 496 TD->getTypeStoreSize(SI->getOperand(0)->getType()), C); 497 if (changed) { 498 MD->removeInstruction(SI); 499 SI->eraseFromParent(); 500 MD->removeInstruction(LI); 501 LI->eraseFromParent(); 502 ++NumMemCpyInstr; 503 return true; 504 } 505 } 506 } 507 } 508 509 // There are two cases that are interesting for this code to handle: memcpy 510 // and memset. Right now we only handle memset. 511 512 // Ensure that the value being stored is something that can be memset'able a 513 // byte at a time like "0" or "-1" or any width, as well as things like 514 // 0xA0A0A0A0 and 0.0. 515 if (Value *ByteVal = isBytewiseValue(SI->getOperand(0))) 516 if (Instruction *I = tryMergingIntoMemset(SI, SI->getPointerOperand(), 517 ByteVal)) { 518 BBI = I; // Don't invalidate iterator. 519 return true; 520 } 521 522 return false; 523} 524 525bool MemCpyOpt::processMemSet(MemSetInst *MSI, BasicBlock::iterator &BBI) { 526 // See if there is another memset or store neighboring this memset which 527 // allows us to widen out the memset to do a single larger store. 528 if (isa<ConstantInt>(MSI->getLength()) && !MSI->isVolatile()) 529 if (Instruction *I = tryMergingIntoMemset(MSI, MSI->getDest(), 530 MSI->getValue())) { 531 BBI = I; // Don't invalidate iterator. 532 return true; 533 } 534 return false; 535} 536 537 538/// performCallSlotOptzn - takes a memcpy and a call that it depends on, 539/// and checks for the possibility of a call slot optimization by having 540/// the call write its result directly into the destination of the memcpy. 541bool MemCpyOpt::performCallSlotOptzn(Instruction *cpy, 542 Value *cpyDest, Value *cpySrc, 543 uint64_t cpyLen, CallInst *C) { 544 // The general transformation to keep in mind is 545 // 546 // call @func(..., src, ...) 547 // memcpy(dest, src, ...) 548 // 549 // -> 550 // 551 // memcpy(dest, src, ...) 552 // call @func(..., dest, ...) 553 // 554 // Since moving the memcpy is technically awkward, we additionally check that 555 // src only holds uninitialized values at the moment of the call, meaning that 556 // the memcpy can be discarded rather than moved. 557 558 // Deliberately get the source and destination with bitcasts stripped away, 559 // because we'll need to do type comparisons based on the underlying type. 560 CallSite CS(C); 561 562 // Require that src be an alloca. This simplifies the reasoning considerably. 563 AllocaInst *srcAlloca = dyn_cast<AllocaInst>(cpySrc); 564 if (!srcAlloca) 565 return false; 566 567 // Check that all of src is copied to dest. 568 if (TD == 0) return false; 569 570 ConstantInt *srcArraySize = dyn_cast<ConstantInt>(srcAlloca->getArraySize()); 571 if (!srcArraySize) 572 return false; 573 574 uint64_t srcSize = TD->getTypeAllocSize(srcAlloca->getAllocatedType()) * 575 srcArraySize->getZExtValue(); 576 577 if (cpyLen < srcSize) 578 return false; 579 580 // Check that accessing the first srcSize bytes of dest will not cause a 581 // trap. Otherwise the transform is invalid since it might cause a trap 582 // to occur earlier than it otherwise would. 583 if (AllocaInst *A = dyn_cast<AllocaInst>(cpyDest)) { 584 // The destination is an alloca. Check it is larger than srcSize. 585 ConstantInt *destArraySize = dyn_cast<ConstantInt>(A->getArraySize()); 586 if (!destArraySize) 587 return false; 588 589 uint64_t destSize = TD->getTypeAllocSize(A->getAllocatedType()) * 590 destArraySize->getZExtValue(); 591 592 if (destSize < srcSize) 593 return false; 594 } else if (Argument *A = dyn_cast<Argument>(cpyDest)) { 595 // If the destination is an sret parameter then only accesses that are 596 // outside of the returned struct type can trap. 597 if (!A->hasStructRetAttr()) 598 return false; 599 600 const Type *StructTy = cast<PointerType>(A->getType())->getElementType(); 601 uint64_t destSize = TD->getTypeAllocSize(StructTy); 602 603 if (destSize < srcSize) 604 return false; 605 } else { 606 return false; 607 } 608 609 // Check that src is not accessed except via the call and the memcpy. This 610 // guarantees that it holds only undefined values when passed in (so the final 611 // memcpy can be dropped), that it is not read or written between the call and 612 // the memcpy, and that writing beyond the end of it is undefined. 613 SmallVector<User*, 8> srcUseList(srcAlloca->use_begin(), 614 srcAlloca->use_end()); 615 while (!srcUseList.empty()) { 616 User *UI = srcUseList.pop_back_val(); 617 618 if (isa<BitCastInst>(UI)) { 619 for (User::use_iterator I = UI->use_begin(), E = UI->use_end(); 620 I != E; ++I) 621 srcUseList.push_back(*I); 622 } else if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(UI)) { 623 if (G->hasAllZeroIndices()) 624 for (User::use_iterator I = UI->use_begin(), E = UI->use_end(); 625 I != E; ++I) 626 srcUseList.push_back(*I); 627 else 628 return false; 629 } else if (UI != C && UI != cpy) { 630 return false; 631 } 632 } 633 634 // Since we're changing the parameter to the callsite, we need to make sure 635 // that what would be the new parameter dominates the callsite. 636 DominatorTree &DT = getAnalysis<DominatorTree>(); 637 if (Instruction *cpyDestInst = dyn_cast<Instruction>(cpyDest)) 638 if (!DT.dominates(cpyDestInst, C)) 639 return false; 640 641 // In addition to knowing that the call does not access src in some 642 // unexpected manner, for example via a global, which we deduce from 643 // the use analysis, we also need to know that it does not sneakily 644 // access dest. We rely on AA to figure this out for us. 645 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 646 if (AA.getModRefInfo(C, cpyDest, srcSize) != AliasAnalysis::NoModRef) 647 return false; 648 649 // All the checks have passed, so do the transformation. 650 bool changedArgument = false; 651 for (unsigned i = 0; i < CS.arg_size(); ++i) 652 if (CS.getArgument(i)->stripPointerCasts() == cpySrc) { 653 if (cpySrc->getType() != cpyDest->getType()) 654 cpyDest = CastInst::CreatePointerCast(cpyDest, cpySrc->getType(), 655 cpyDest->getName(), C); 656 changedArgument = true; 657 if (CS.getArgument(i)->getType() == cpyDest->getType()) 658 CS.setArgument(i, cpyDest); 659 else 660 CS.setArgument(i, CastInst::CreatePointerCast(cpyDest, 661 CS.getArgument(i)->getType(), cpyDest->getName(), C)); 662 } 663 664 if (!changedArgument) 665 return false; 666 667 // Drop any cached information about the call, because we may have changed 668 // its dependence information by changing its parameter. 669 MD->removeInstruction(C); 670 671 // Remove the memcpy. 672 MD->removeInstruction(cpy); 673 ++NumMemCpyInstr; 674 675 return true; 676} 677 678/// processMemCpyMemCpyDependence - We've found that the (upward scanning) 679/// memory dependence of memcpy 'M' is the memcpy 'MDep'. Try to simplify M to 680/// copy from MDep's input if we can. MSize is the size of M's copy. 681/// 682bool MemCpyOpt::processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep, 683 uint64_t MSize) { 684 // We can only transforms memcpy's where the dest of one is the source of the 685 // other. 686 if (M->getSource() != MDep->getDest() || MDep->isVolatile()) 687 return false; 688 689 // If dep instruction is reading from our current input, then it is a noop 690 // transfer and substituting the input won't change this instruction. Just 691 // ignore the input and let someone else zap MDep. This handles cases like: 692 // memcpy(a <- a) 693 // memcpy(b <- a) 694 if (M->getSource() == MDep->getSource()) 695 return false; 696 697 // Second, the length of the memcpy's must be the same, or the preceding one 698 // must be larger than the following one. 699 ConstantInt *MDepLen = dyn_cast<ConstantInt>(MDep->getLength()); 700 ConstantInt *MLen = dyn_cast<ConstantInt>(M->getLength()); 701 if (!MDepLen || !MLen || MDepLen->getZExtValue() < MLen->getZExtValue()) 702 return false; 703 704 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 705 706 // Verify that the copied-from memory doesn't change in between the two 707 // transfers. For example, in: 708 // memcpy(a <- b) 709 // *b = 42; 710 // memcpy(c <- a) 711 // It would be invalid to transform the second memcpy into memcpy(c <- b). 712 // 713 // TODO: If the code between M and MDep is transparent to the destination "c", 714 // then we could still perform the xform by moving M up to the first memcpy. 715 // 716 // NOTE: This is conservative, it will stop on any read from the source loc, 717 // not just the defining memcpy. 718 MemDepResult SourceDep = 719 MD->getPointerDependencyFrom(AA.getLocationForSource(MDep), 720 false, M, M->getParent()); 721 if (!SourceDep.isClobber() || SourceDep.getInst() != MDep) 722 return false; 723 724 // If the dest of the second might alias the source of the first, then the 725 // source and dest might overlap. We still want to eliminate the intermediate 726 // value, but we have to generate a memmove instead of memcpy. 727 bool UseMemMove = false; 728 if (!AA.isNoAlias(AA.getLocationForDest(M), AA.getLocationForSource(MDep))) 729 UseMemMove = true; 730 731 // If all checks passed, then we can transform M. 732 733 // Make sure to use the lesser of the alignment of the source and the dest 734 // since we're changing where we're reading from, but don't want to increase 735 // the alignment past what can be read from or written to. 736 // TODO: Is this worth it if we're creating a less aligned memcpy? For 737 // example we could be moving from movaps -> movq on x86. 738 unsigned Align = std::min(MDep->getAlignment(), M->getAlignment()); 739 740 IRBuilder<> Builder(M); 741 if (UseMemMove) 742 Builder.CreateMemMove(M->getRawDest(), MDep->getRawSource(), M->getLength(), 743 Align, M->isVolatile()); 744 else 745 Builder.CreateMemCpy(M->getRawDest(), MDep->getRawSource(), M->getLength(), 746 Align, M->isVolatile()); 747 748 // Remove the instruction we're replacing. 749 MD->removeInstruction(M); 750 M->eraseFromParent(); 751 ++NumMemCpyInstr; 752 return true; 753} 754 755 756/// processMemCpy - perform simplification of memcpy's. If we have memcpy A 757/// which copies X to Y, and memcpy B which copies Y to Z, then we can rewrite 758/// B to be a memcpy from X to Z (or potentially a memmove, depending on 759/// circumstances). This allows later passes to remove the first memcpy 760/// altogether. 761bool MemCpyOpt::processMemCpy(MemCpyInst *M) { 762 // We can only optimize statically-sized memcpy's that are non-volatile. 763 ConstantInt *CopySize = dyn_cast<ConstantInt>(M->getLength()); 764 if (CopySize == 0 || M->isVolatile()) return false; 765 766 // If the source and destination of the memcpy are the same, then zap it. 767 if (M->getSource() == M->getDest()) { 768 MD->removeInstruction(M); 769 M->eraseFromParent(); 770 return false; 771 } 772 773 // If copying from a constant, try to turn the memcpy into a memset. 774 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(M->getSource())) 775 if (GV->isConstant() && GV->hasDefinitiveInitializer()) 776 if (Value *ByteVal = isBytewiseValue(GV->getInitializer())) { 777 IRBuilder<> Builder(M); 778 Builder.CreateMemSet(M->getRawDest(), ByteVal, CopySize, 779 M->getAlignment(), false); 780 MD->removeInstruction(M); 781 M->eraseFromParent(); 782 ++NumCpyToSet; 783 return true; 784 } 785 786 // The are two possible optimizations we can do for memcpy: 787 // a) memcpy-memcpy xform which exposes redundance for DSE. 788 // b) call-memcpy xform for return slot optimization. 789 MemDepResult DepInfo = MD->getDependency(M); 790 if (!DepInfo.isClobber()) 791 return false; 792 793 if (MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst())) 794 return processMemCpyMemCpyDependence(M, MDep, CopySize->getZExtValue()); 795 796 if (CallInst *C = dyn_cast<CallInst>(DepInfo.getInst())) { 797 if (performCallSlotOptzn(M, M->getDest(), M->getSource(), 798 CopySize->getZExtValue(), C)) { 799 MD->removeInstruction(M); 800 M->eraseFromParent(); 801 return true; 802 } 803 } 804 805 return false; 806} 807 808/// processMemMove - Transforms memmove calls to memcpy calls when the src/dst 809/// are guaranteed not to alias. 810bool MemCpyOpt::processMemMove(MemMoveInst *M) { 811 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 812 813 if (!TLI->has(LibFunc::memmove)) 814 return false; 815 816 // See if the pointers alias. 817 if (!AA.isNoAlias(AA.getLocationForDest(M), AA.getLocationForSource(M))) 818 return false; 819 820 DEBUG(dbgs() << "MemCpyOpt: Optimizing memmove -> memcpy: " << *M << "\n"); 821 822 // If not, then we know we can transform this. 823 Module *Mod = M->getParent()->getParent()->getParent(); 824 const Type *ArgTys[3] = { M->getRawDest()->getType(), 825 M->getRawSource()->getType(), 826 M->getLength()->getType() }; 827 M->setCalledFunction(Intrinsic::getDeclaration(Mod, Intrinsic::memcpy, 828 ArgTys, 3)); 829 830 // MemDep may have over conservative information about this instruction, just 831 // conservatively flush it from the cache. 832 MD->removeInstruction(M); 833 834 ++NumMoveToCpy; 835 return true; 836} 837 838/// processByValArgument - This is called on every byval argument in call sites. 839bool MemCpyOpt::processByValArgument(CallSite CS, unsigned ArgNo) { 840 if (TD == 0) return false; 841 842 // Find out what feeds this byval argument. 843 Value *ByValArg = CS.getArgument(ArgNo); 844 const Type *ByValTy =cast<PointerType>(ByValArg->getType())->getElementType(); 845 uint64_t ByValSize = TD->getTypeAllocSize(ByValTy); 846 MemDepResult DepInfo = 847 MD->getPointerDependencyFrom(AliasAnalysis::Location(ByValArg, ByValSize), 848 true, CS.getInstruction(), 849 CS.getInstruction()->getParent()); 850 if (!DepInfo.isClobber()) 851 return false; 852 853 // If the byval argument isn't fed by a memcpy, ignore it. If it is fed by 854 // a memcpy, see if we can byval from the source of the memcpy instead of the 855 // result. 856 MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst()); 857 if (MDep == 0 || MDep->isVolatile() || 858 ByValArg->stripPointerCasts() != MDep->getDest()) 859 return false; 860 861 // The length of the memcpy must be larger or equal to the size of the byval. 862 ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength()); 863 if (C1 == 0 || C1->getValue().getZExtValue() < ByValSize) 864 return false; 865 866 // Get the alignment of the byval. If it is greater than the memcpy, then we 867 // can't do the substitution. If the call doesn't specify the alignment, then 868 // it is some target specific value that we can't know. 869 unsigned ByValAlign = CS.getParamAlignment(ArgNo+1); 870 if (ByValAlign == 0 || MDep->getAlignment() < ByValAlign) 871 return false; 872 873 // Verify that the copied-from memory doesn't change in between the memcpy and 874 // the byval call. 875 // memcpy(a <- b) 876 // *b = 42; 877 // foo(*a) 878 // It would be invalid to transform the second memcpy into foo(*b). 879 // 880 // NOTE: This is conservative, it will stop on any read from the source loc, 881 // not just the defining memcpy. 882 MemDepResult SourceDep = 883 MD->getPointerDependencyFrom(AliasAnalysis::getLocationForSource(MDep), 884 false, CS.getInstruction(), MDep->getParent()); 885 if (!SourceDep.isClobber() || SourceDep.getInst() != MDep) 886 return false; 887 888 Value *TmpCast = MDep->getSource(); 889 if (MDep->getSource()->getType() != ByValArg->getType()) 890 TmpCast = new BitCastInst(MDep->getSource(), ByValArg->getType(), 891 "tmpcast", CS.getInstruction()); 892 893 DEBUG(dbgs() << "MemCpyOpt: Forwarding memcpy to byval:\n" 894 << " " << *MDep << "\n" 895 << " " << *CS.getInstruction() << "\n"); 896 897 // Otherwise we're good! Update the byval argument. 898 CS.setArgument(ArgNo, TmpCast); 899 ++NumMemCpyInstr; 900 return true; 901} 902 903/// iterateOnFunction - Executes one iteration of MemCpyOpt. 904bool MemCpyOpt::iterateOnFunction(Function &F) { 905 bool MadeChange = false; 906 907 // Walk all instruction in the function. 908 for (Function::iterator BB = F.begin(), BBE = F.end(); BB != BBE; ++BB) { 909 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE;) { 910 // Avoid invalidating the iterator. 911 Instruction *I = BI++; 912 913 bool RepeatInstruction = false; 914 915 if (StoreInst *SI = dyn_cast<StoreInst>(I)) 916 MadeChange |= processStore(SI, BI); 917 else if (MemSetInst *M = dyn_cast<MemSetInst>(I)) 918 RepeatInstruction = processMemSet(M, BI); 919 else if (MemCpyInst *M = dyn_cast<MemCpyInst>(I)) 920 RepeatInstruction = processMemCpy(M); 921 else if (MemMoveInst *M = dyn_cast<MemMoveInst>(I)) 922 RepeatInstruction = processMemMove(M); 923 else if (CallSite CS = (Value*)I) { 924 for (unsigned i = 0, e = CS.arg_size(); i != e; ++i) 925 if (CS.paramHasAttr(i+1, Attribute::ByVal)) 926 MadeChange |= processByValArgument(CS, i); 927 } 928 929 // Reprocess the instruction if desired. 930 if (RepeatInstruction) { 931 if (BI != BB->begin()) --BI; 932 MadeChange = true; 933 } 934 } 935 } 936 937 return MadeChange; 938} 939 940// MemCpyOpt::runOnFunction - This is the main transformation entry point for a 941// function. 942// 943bool MemCpyOpt::runOnFunction(Function &F) { 944 bool MadeChange = false; 945 MD = &getAnalysis<MemoryDependenceAnalysis>(); 946 TD = getAnalysisIfAvailable<TargetData>(); 947 TLI = &getAnalysis<TargetLibraryInfo>(); 948 949 // If we don't have at least memset and memcpy, there is little point of doing 950 // anything here. These are required by a freestanding implementation, so if 951 // even they are disabled, there is no point in trying hard. 952 if (!TLI->has(LibFunc::memset) || !TLI->has(LibFunc::memcpy)) 953 return false; 954 955 while (1) { 956 if (!iterateOnFunction(F)) 957 break; 958 MadeChange = true; 959 } 960 961 MD = 0; 962 return MadeChange; 963} 964