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 #include "llvm/Transforms/Scalar.h" 16 #include "llvm/ADT/SmallVector.h" 17 #include "llvm/ADT/Statistic.h" 18 #include "llvm/Analysis/AliasAnalysis.h" 19 #include "llvm/Analysis/AssumptionCache.h" 20 #include "llvm/Analysis/MemoryDependenceAnalysis.h" 21 #include "llvm/Analysis/TargetLibraryInfo.h" 22 #include "llvm/Analysis/ValueTracking.h" 23 #include "llvm/IR/DataLayout.h" 24 #include "llvm/IR/Dominators.h" 25 #include "llvm/IR/GetElementPtrTypeIterator.h" 26 #include "llvm/IR/GlobalVariable.h" 27 #include "llvm/IR/IRBuilder.h" 28 #include "llvm/IR/Instructions.h" 29 #include "llvm/IR/IntrinsicInst.h" 30 #include "llvm/Support/Debug.h" 31 #include "llvm/Support/raw_ostream.h" 32 #include "llvm/Transforms/Utils/Local.h" 33 #include <list> 34 using namespace llvm; 35 36 #define DEBUG_TYPE "memcpyopt" 37 38 STATISTIC(NumMemCpyInstr, "Number of memcpy instructions deleted"); 39 STATISTIC(NumMemSetInfer, "Number of memsets inferred"); 40 STATISTIC(NumMoveToCpy, "Number of memmoves converted to memcpy"); 41 STATISTIC(NumCpyToSet, "Number of memcpys converted to memset"); 42 43 static int64_t GetOffsetFromIndex(const GEPOperator *GEP, unsigned Idx, 44 bool &VariableIdxFound, 45 const DataLayout &DL) { 46 // Skip over the first indices. 47 gep_type_iterator GTI = gep_type_begin(GEP); 48 for (unsigned i = 1; i != Idx; ++i, ++GTI) 49 /*skip along*/; 50 51 // Compute the offset implied by the rest of the indices. 52 int64_t Offset = 0; 53 for (unsigned i = Idx, e = GEP->getNumOperands(); i != e; ++i, ++GTI) { 54 ConstantInt *OpC = dyn_cast<ConstantInt>(GEP->getOperand(i)); 55 if (!OpC) 56 return VariableIdxFound = true; 57 if (OpC->isZero()) continue; // No offset. 58 59 // Handle struct indices, which add their field offset to the pointer. 60 if (StructType *STy = dyn_cast<StructType>(*GTI)) { 61 Offset += DL.getStructLayout(STy)->getElementOffset(OpC->getZExtValue()); 62 continue; 63 } 64 65 // Otherwise, we have a sequential type like an array or vector. Multiply 66 // the index by the ElementSize. 67 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType()); 68 Offset += Size*OpC->getSExtValue(); 69 } 70 71 return Offset; 72 } 73 74 /// IsPointerOffset - Return true if Ptr1 is provably equal to Ptr2 plus a 75 /// constant offset, and return that constant offset. For example, Ptr1 might 76 /// be &A[42], and Ptr2 might be &A[40]. In this case offset would be -8. 77 static bool IsPointerOffset(Value *Ptr1, Value *Ptr2, int64_t &Offset, 78 const DataLayout &DL) { 79 Ptr1 = Ptr1->stripPointerCasts(); 80 Ptr2 = Ptr2->stripPointerCasts(); 81 82 // Handle the trivial case first. 83 if (Ptr1 == Ptr2) { 84 Offset = 0; 85 return true; 86 } 87 88 GEPOperator *GEP1 = dyn_cast<GEPOperator>(Ptr1); 89 GEPOperator *GEP2 = dyn_cast<GEPOperator>(Ptr2); 90 91 bool VariableIdxFound = false; 92 93 // If one pointer is a GEP and the other isn't, then see if the GEP is a 94 // constant offset from the base, as in "P" and "gep P, 1". 95 if (GEP1 && !GEP2 && GEP1->getOperand(0)->stripPointerCasts() == Ptr2) { 96 Offset = -GetOffsetFromIndex(GEP1, 1, VariableIdxFound, DL); 97 return !VariableIdxFound; 98 } 99 100 if (GEP2 && !GEP1 && GEP2->getOperand(0)->stripPointerCasts() == Ptr1) { 101 Offset = GetOffsetFromIndex(GEP2, 1, VariableIdxFound, DL); 102 return !VariableIdxFound; 103 } 104 105 // Right now we handle the case when Ptr1/Ptr2 are both GEPs with an identical 106 // base. After that base, they may have some number of common (and 107 // potentially variable) indices. After that they handle some constant 108 // offset, which determines their offset from each other. At this point, we 109 // handle no other case. 110 if (!GEP1 || !GEP2 || GEP1->getOperand(0) != GEP2->getOperand(0)) 111 return false; 112 113 // Skip any common indices and track the GEP types. 114 unsigned Idx = 1; 115 for (; Idx != GEP1->getNumOperands() && Idx != GEP2->getNumOperands(); ++Idx) 116 if (GEP1->getOperand(Idx) != GEP2->getOperand(Idx)) 117 break; 118 119 int64_t Offset1 = GetOffsetFromIndex(GEP1, Idx, VariableIdxFound, DL); 120 int64_t Offset2 = GetOffsetFromIndex(GEP2, Idx, VariableIdxFound, DL); 121 if (VariableIdxFound) return false; 122 123 Offset = Offset2-Offset1; 124 return true; 125 } 126 127 128 /// MemsetRange - Represents a range of memset'd bytes with the ByteVal value. 129 /// This allows us to analyze stores like: 130 /// store 0 -> P+1 131 /// store 0 -> P+0 132 /// store 0 -> P+3 133 /// store 0 -> P+2 134 /// which sometimes happens with stores to arrays of structs etc. When we see 135 /// the first store, we make a range [1, 2). The second store extends the range 136 /// to [0, 2). The third makes a new range [2, 3). The fourth store joins the 137 /// two ranges into [0, 3) which is memset'able. 138 namespace { 139 struct MemsetRange { 140 // Start/End - A semi range that describes the span that this range covers. 141 // The range is closed at the start and open at the end: [Start, End). 142 int64_t Start, End; 143 144 /// StartPtr - The getelementptr instruction that points to the start of the 145 /// range. 146 Value *StartPtr; 147 148 /// Alignment - The known alignment of the first store. 149 unsigned Alignment; 150 151 /// TheStores - The actual stores that make up this range. 152 SmallVector<Instruction*, 16> TheStores; 153 154 bool isProfitableToUseMemset(const DataLayout &DL) const; 155 }; 156 } // end anon namespace 157 158 bool MemsetRange::isProfitableToUseMemset(const DataLayout &DL) const { 159 // If we found more than 4 stores to merge or 16 bytes, use memset. 160 if (TheStores.size() >= 4 || End-Start >= 16) return true; 161 162 // If there is nothing to merge, don't do anything. 163 if (TheStores.size() < 2) return false; 164 165 // If any of the stores are a memset, then it is always good to extend the 166 // memset. 167 for (unsigned i = 0, e = TheStores.size(); i != e; ++i) 168 if (!isa<StoreInst>(TheStores[i])) 169 return true; 170 171 // Assume that the code generator is capable of merging pairs of stores 172 // together if it wants to. 173 if (TheStores.size() == 2) return false; 174 175 // If we have fewer than 8 stores, it can still be worthwhile to do this. 176 // For example, merging 4 i8 stores into an i32 store is useful almost always. 177 // However, merging 2 32-bit stores isn't useful on a 32-bit architecture (the 178 // memset will be split into 2 32-bit stores anyway) and doing so can 179 // pessimize the llvm optimizer. 180 // 181 // Since we don't have perfect knowledge here, make some assumptions: assume 182 // the maximum GPR width is the same size as the largest legal integer 183 // size. If so, check to see whether we will end up actually reducing the 184 // number of stores used. 185 unsigned Bytes = unsigned(End-Start); 186 unsigned MaxIntSize = DL.getLargestLegalIntTypeSize(); 187 if (MaxIntSize == 0) 188 MaxIntSize = 1; 189 unsigned NumPointerStores = Bytes / MaxIntSize; 190 191 // Assume the remaining bytes if any are done a byte at a time. 192 unsigned NumByteStores = Bytes - NumPointerStores * MaxIntSize; 193 194 // If we will reduce the # stores (according to this heuristic), do the 195 // transformation. This encourages merging 4 x i8 -> i32 and 2 x i16 -> i32 196 // etc. 197 return TheStores.size() > NumPointerStores+NumByteStores; 198 } 199 200 201 namespace { 202 class MemsetRanges { 203 /// Ranges - A sorted list of the memset ranges. We use std::list here 204 /// because each element is relatively large and expensive to copy. 205 std::list<MemsetRange> Ranges; 206 typedef std::list<MemsetRange>::iterator range_iterator; 207 const DataLayout &DL; 208 public: 209 MemsetRanges(const DataLayout &DL) : DL(DL) {} 210 211 typedef std::list<MemsetRange>::const_iterator const_iterator; 212 const_iterator begin() const { return Ranges.begin(); } 213 const_iterator end() const { return Ranges.end(); } 214 bool empty() const { return Ranges.empty(); } 215 216 void addInst(int64_t OffsetFromFirst, Instruction *Inst) { 217 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) 218 addStore(OffsetFromFirst, SI); 219 else 220 addMemSet(OffsetFromFirst, cast<MemSetInst>(Inst)); 221 } 222 223 void addStore(int64_t OffsetFromFirst, StoreInst *SI) { 224 int64_t StoreSize = DL.getTypeStoreSize(SI->getOperand(0)->getType()); 225 226 addRange(OffsetFromFirst, StoreSize, 227 SI->getPointerOperand(), SI->getAlignment(), SI); 228 } 229 230 void addMemSet(int64_t OffsetFromFirst, MemSetInst *MSI) { 231 int64_t Size = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 232 addRange(OffsetFromFirst, Size, MSI->getDest(), MSI->getAlignment(), MSI); 233 } 234 235 void addRange(int64_t Start, int64_t Size, Value *Ptr, 236 unsigned Alignment, Instruction *Inst); 237 238 }; 239 240 } // end anon namespace 241 242 243 /// addRange - Add a new store to the MemsetRanges data structure. This adds a 244 /// new range for the specified store at the specified offset, merging into 245 /// existing ranges as appropriate. 246 /// 247 /// Do a linear search of the ranges to see if this can be joined and/or to 248 /// find the insertion point in the list. We keep the ranges sorted for 249 /// simplicity here. This is a linear search of a linked list, which is ugly, 250 /// however the number of ranges is limited, so this won't get crazy slow. 251 void MemsetRanges::addRange(int64_t Start, int64_t Size, Value *Ptr, 252 unsigned Alignment, Instruction *Inst) { 253 int64_t End = Start+Size; 254 range_iterator I = Ranges.begin(), E = Ranges.end(); 255 256 while (I != E && Start > I->End) 257 ++I; 258 259 // We now know that I == E, in which case we didn't find anything to merge 260 // with, or that Start <= I->End. If End < I->Start or I == E, then we need 261 // to insert a new range. Handle this now. 262 if (I == E || End < I->Start) { 263 MemsetRange &R = *Ranges.insert(I, MemsetRange()); 264 R.Start = Start; 265 R.End = End; 266 R.StartPtr = Ptr; 267 R.Alignment = Alignment; 268 R.TheStores.push_back(Inst); 269 return; 270 } 271 272 // This store overlaps with I, add it. 273 I->TheStores.push_back(Inst); 274 275 // At this point, we may have an interval that completely contains our store. 276 // If so, just add it to the interval and return. 277 if (I->Start <= Start && I->End >= End) 278 return; 279 280 // Now we know that Start <= I->End and End >= I->Start so the range overlaps 281 // but is not entirely contained within the range. 282 283 // See if the range extends the start of the range. In this case, it couldn't 284 // possibly cause it to join the prior range, because otherwise we would have 285 // stopped on *it*. 286 if (Start < I->Start) { 287 I->Start = Start; 288 I->StartPtr = Ptr; 289 I->Alignment = Alignment; 290 } 291 292 // Now we know that Start <= I->End and Start >= I->Start (so the startpoint 293 // is in or right at the end of I), and that End >= I->Start. Extend I out to 294 // End. 295 if (End > I->End) { 296 I->End = End; 297 range_iterator NextI = I; 298 while (++NextI != E && End >= NextI->Start) { 299 // Merge the range in. 300 I->TheStores.append(NextI->TheStores.begin(), NextI->TheStores.end()); 301 if (NextI->End > I->End) 302 I->End = NextI->End; 303 Ranges.erase(NextI); 304 NextI = I; 305 } 306 } 307 } 308 309 //===----------------------------------------------------------------------===// 310 // MemCpyOpt Pass 311 //===----------------------------------------------------------------------===// 312 313 namespace { 314 class MemCpyOpt : public FunctionPass { 315 MemoryDependenceAnalysis *MD; 316 TargetLibraryInfo *TLI; 317 public: 318 static char ID; // Pass identification, replacement for typeid 319 MemCpyOpt() : FunctionPass(ID) { 320 initializeMemCpyOptPass(*PassRegistry::getPassRegistry()); 321 MD = nullptr; 322 TLI = nullptr; 323 } 324 325 bool runOnFunction(Function &F) override; 326 327 private: 328 // This transformation requires dominator postdominator info 329 void getAnalysisUsage(AnalysisUsage &AU) const override { 330 AU.setPreservesCFG(); 331 AU.addRequired<AssumptionCacheTracker>(); 332 AU.addRequired<DominatorTreeWrapperPass>(); 333 AU.addRequired<MemoryDependenceAnalysis>(); 334 AU.addRequired<AliasAnalysis>(); 335 AU.addRequired<TargetLibraryInfoWrapperPass>(); 336 AU.addPreserved<AliasAnalysis>(); 337 AU.addPreserved<MemoryDependenceAnalysis>(); 338 } 339 340 // Helper fuctions 341 bool processStore(StoreInst *SI, BasicBlock::iterator &BBI); 342 bool processMemSet(MemSetInst *SI, BasicBlock::iterator &BBI); 343 bool processMemCpy(MemCpyInst *M); 344 bool processMemMove(MemMoveInst *M); 345 bool performCallSlotOptzn(Instruction *cpy, Value *cpyDst, Value *cpySrc, 346 uint64_t cpyLen, unsigned cpyAlign, CallInst *C); 347 bool processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep, 348 uint64_t MSize); 349 bool processByValArgument(CallSite CS, unsigned ArgNo); 350 Instruction *tryMergingIntoMemset(Instruction *I, Value *StartPtr, 351 Value *ByteVal); 352 353 bool iterateOnFunction(Function &F); 354 }; 355 356 char MemCpyOpt::ID = 0; 357 } 358 359 // createMemCpyOptPass - The public interface to this file... 360 FunctionPass *llvm::createMemCpyOptPass() { return new MemCpyOpt(); } 361 362 INITIALIZE_PASS_BEGIN(MemCpyOpt, "memcpyopt", "MemCpy Optimization", 363 false, false) 364 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 365 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 366 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis) 367 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 368 INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 369 INITIALIZE_PASS_END(MemCpyOpt, "memcpyopt", "MemCpy Optimization", 370 false, false) 371 372 /// tryMergingIntoMemset - When scanning forward over instructions, we look for 373 /// some other patterns to fold away. In particular, this looks for stores to 374 /// neighboring locations of memory. If it sees enough consecutive ones, it 375 /// attempts to merge them together into a memcpy/memset. 376 Instruction *MemCpyOpt::tryMergingIntoMemset(Instruction *StartInst, 377 Value *StartPtr, Value *ByteVal) { 378 const DataLayout &DL = StartInst->getModule()->getDataLayout(); 379 380 // Okay, so we now have a single store that can be splatable. Scan to find 381 // all subsequent stores of the same value to offset from the same pointer. 382 // Join these together into ranges, so we can decide whether contiguous blocks 383 // are stored. 384 MemsetRanges Ranges(DL); 385 386 BasicBlock::iterator BI = StartInst; 387 for (++BI; !isa<TerminatorInst>(BI); ++BI) { 388 if (!isa<StoreInst>(BI) && !isa<MemSetInst>(BI)) { 389 // If the instruction is readnone, ignore it, otherwise bail out. We 390 // don't even allow readonly here because we don't want something like: 391 // A[1] = 2; strlen(A); A[2] = 2; -> memcpy(A, ...); strlen(A). 392 if (BI->mayWriteToMemory() || BI->mayReadFromMemory()) 393 break; 394 continue; 395 } 396 397 if (StoreInst *NextStore = dyn_cast<StoreInst>(BI)) { 398 // If this is a store, see if we can merge it in. 399 if (!NextStore->isSimple()) break; 400 401 // Check to see if this stored value is of the same byte-splattable value. 402 if (ByteVal != isBytewiseValue(NextStore->getOperand(0))) 403 break; 404 405 // Check to see if this store is to a constant offset from the start ptr. 406 int64_t Offset; 407 if (!IsPointerOffset(StartPtr, NextStore->getPointerOperand(), Offset, 408 DL)) 409 break; 410 411 Ranges.addStore(Offset, NextStore); 412 } else { 413 MemSetInst *MSI = cast<MemSetInst>(BI); 414 415 if (MSI->isVolatile() || ByteVal != MSI->getValue() || 416 !isa<ConstantInt>(MSI->getLength())) 417 break; 418 419 // Check to see if this store is to a constant offset from the start ptr. 420 int64_t Offset; 421 if (!IsPointerOffset(StartPtr, MSI->getDest(), Offset, DL)) 422 break; 423 424 Ranges.addMemSet(Offset, MSI); 425 } 426 } 427 428 // If we have no ranges, then we just had a single store with nothing that 429 // could be merged in. This is a very common case of course. 430 if (Ranges.empty()) 431 return nullptr; 432 433 // If we had at least one store that could be merged in, add the starting 434 // store as well. We try to avoid this unless there is at least something 435 // interesting as a small compile-time optimization. 436 Ranges.addInst(0, StartInst); 437 438 // If we create any memsets, we put it right before the first instruction that 439 // isn't part of the memset block. This ensure that the memset is dominated 440 // by any addressing instruction needed by the start of the block. 441 IRBuilder<> Builder(BI); 442 443 // Now that we have full information about ranges, loop over the ranges and 444 // emit memset's for anything big enough to be worthwhile. 445 Instruction *AMemSet = nullptr; 446 for (MemsetRanges::const_iterator I = Ranges.begin(), E = Ranges.end(); 447 I != E; ++I) { 448 const MemsetRange &Range = *I; 449 450 if (Range.TheStores.size() == 1) continue; 451 452 // If it is profitable to lower this range to memset, do so now. 453 if (!Range.isProfitableToUseMemset(DL)) 454 continue; 455 456 // Otherwise, we do want to transform this! Create a new memset. 457 // Get the starting pointer of the block. 458 StartPtr = Range.StartPtr; 459 460 // Determine alignment 461 unsigned Alignment = Range.Alignment; 462 if (Alignment == 0) { 463 Type *EltType = 464 cast<PointerType>(StartPtr->getType())->getElementType(); 465 Alignment = DL.getABITypeAlignment(EltType); 466 } 467 468 AMemSet = 469 Builder.CreateMemSet(StartPtr, ByteVal, Range.End-Range.Start, Alignment); 470 471 DEBUG(dbgs() << "Replace stores:\n"; 472 for (unsigned i = 0, e = Range.TheStores.size(); i != e; ++i) 473 dbgs() << *Range.TheStores[i] << '\n'; 474 dbgs() << "With: " << *AMemSet << '\n'); 475 476 if (!Range.TheStores.empty()) 477 AMemSet->setDebugLoc(Range.TheStores[0]->getDebugLoc()); 478 479 // Zap all the stores. 480 for (SmallVectorImpl<Instruction *>::const_iterator 481 SI = Range.TheStores.begin(), 482 SE = Range.TheStores.end(); SI != SE; ++SI) { 483 MD->removeInstruction(*SI); 484 (*SI)->eraseFromParent(); 485 } 486 ++NumMemSetInfer; 487 } 488 489 return AMemSet; 490 } 491 492 493 bool MemCpyOpt::processStore(StoreInst *SI, BasicBlock::iterator &BBI) { 494 if (!SI->isSimple()) return false; 495 const DataLayout &DL = SI->getModule()->getDataLayout(); 496 497 // Detect cases where we're performing call slot forwarding, but 498 // happen to be using a load-store pair to implement it, rather than 499 // a memcpy. 500 if (LoadInst *LI = dyn_cast<LoadInst>(SI->getOperand(0))) { 501 if (LI->isSimple() && LI->hasOneUse() && 502 LI->getParent() == SI->getParent()) { 503 MemDepResult ldep = MD->getDependency(LI); 504 CallInst *C = nullptr; 505 if (ldep.isClobber() && !isa<MemCpyInst>(ldep.getInst())) 506 C = dyn_cast<CallInst>(ldep.getInst()); 507 508 if (C) { 509 // Check that nothing touches the dest of the "copy" between 510 // the call and the store. 511 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 512 AliasAnalysis::Location StoreLoc = AA.getLocation(SI); 513 for (BasicBlock::iterator I = --BasicBlock::iterator(SI), 514 E = C; I != E; --I) { 515 if (AA.getModRefInfo(&*I, StoreLoc) != AliasAnalysis::NoModRef) { 516 C = nullptr; 517 break; 518 } 519 } 520 } 521 522 if (C) { 523 unsigned storeAlign = SI->getAlignment(); 524 if (!storeAlign) 525 storeAlign = DL.getABITypeAlignment(SI->getOperand(0)->getType()); 526 unsigned loadAlign = LI->getAlignment(); 527 if (!loadAlign) 528 loadAlign = DL.getABITypeAlignment(LI->getType()); 529 530 bool changed = performCallSlotOptzn( 531 LI, SI->getPointerOperand()->stripPointerCasts(), 532 LI->getPointerOperand()->stripPointerCasts(), 533 DL.getTypeStoreSize(SI->getOperand(0)->getType()), 534 std::min(storeAlign, loadAlign), C); 535 if (changed) { 536 MD->removeInstruction(SI); 537 SI->eraseFromParent(); 538 MD->removeInstruction(LI); 539 LI->eraseFromParent(); 540 ++NumMemCpyInstr; 541 return true; 542 } 543 } 544 } 545 } 546 547 // There are two cases that are interesting for this code to handle: memcpy 548 // and memset. Right now we only handle memset. 549 550 // Ensure that the value being stored is something that can be memset'able a 551 // byte at a time like "0" or "-1" or any width, as well as things like 552 // 0xA0A0A0A0 and 0.0. 553 if (Value *ByteVal = isBytewiseValue(SI->getOperand(0))) 554 if (Instruction *I = tryMergingIntoMemset(SI, SI->getPointerOperand(), 555 ByteVal)) { 556 BBI = I; // Don't invalidate iterator. 557 return true; 558 } 559 560 return false; 561 } 562 563 bool MemCpyOpt::processMemSet(MemSetInst *MSI, BasicBlock::iterator &BBI) { 564 // See if there is another memset or store neighboring this memset which 565 // allows us to widen out the memset to do a single larger store. 566 if (isa<ConstantInt>(MSI->getLength()) && !MSI->isVolatile()) 567 if (Instruction *I = tryMergingIntoMemset(MSI, MSI->getDest(), 568 MSI->getValue())) { 569 BBI = I; // Don't invalidate iterator. 570 return true; 571 } 572 return false; 573 } 574 575 576 /// performCallSlotOptzn - takes a memcpy and a call that it depends on, 577 /// and checks for the possibility of a call slot optimization by having 578 /// the call write its result directly into the destination of the memcpy. 579 bool MemCpyOpt::performCallSlotOptzn(Instruction *cpy, 580 Value *cpyDest, Value *cpySrc, 581 uint64_t cpyLen, unsigned cpyAlign, 582 CallInst *C) { 583 // The general transformation to keep in mind is 584 // 585 // call @func(..., src, ...) 586 // memcpy(dest, src, ...) 587 // 588 // -> 589 // 590 // memcpy(dest, src, ...) 591 // call @func(..., dest, ...) 592 // 593 // Since moving the memcpy is technically awkward, we additionally check that 594 // src only holds uninitialized values at the moment of the call, meaning that 595 // the memcpy can be discarded rather than moved. 596 597 // Deliberately get the source and destination with bitcasts stripped away, 598 // because we'll need to do type comparisons based on the underlying type. 599 CallSite CS(C); 600 601 // Require that src be an alloca. This simplifies the reasoning considerably. 602 AllocaInst *srcAlloca = dyn_cast<AllocaInst>(cpySrc); 603 if (!srcAlloca) 604 return false; 605 606 ConstantInt *srcArraySize = dyn_cast<ConstantInt>(srcAlloca->getArraySize()); 607 if (!srcArraySize) 608 return false; 609 610 const DataLayout &DL = cpy->getModule()->getDataLayout(); 611 uint64_t srcSize = DL.getTypeAllocSize(srcAlloca->getAllocatedType()) * 612 srcArraySize->getZExtValue(); 613 614 if (cpyLen < srcSize) 615 return false; 616 617 // Check that accessing the first srcSize bytes of dest will not cause a 618 // trap. Otherwise the transform is invalid since it might cause a trap 619 // to occur earlier than it otherwise would. 620 if (AllocaInst *A = dyn_cast<AllocaInst>(cpyDest)) { 621 // The destination is an alloca. Check it is larger than srcSize. 622 ConstantInt *destArraySize = dyn_cast<ConstantInt>(A->getArraySize()); 623 if (!destArraySize) 624 return false; 625 626 uint64_t destSize = DL.getTypeAllocSize(A->getAllocatedType()) * 627 destArraySize->getZExtValue(); 628 629 if (destSize < srcSize) 630 return false; 631 } else if (Argument *A = dyn_cast<Argument>(cpyDest)) { 632 if (A->getDereferenceableBytes() < srcSize) { 633 // If the destination is an sret parameter then only accesses that are 634 // outside of the returned struct type can trap. 635 if (!A->hasStructRetAttr()) 636 return false; 637 638 Type *StructTy = cast<PointerType>(A->getType())->getElementType(); 639 if (!StructTy->isSized()) { 640 // The call may never return and hence the copy-instruction may never 641 // be executed, and therefore it's not safe to say "the destination 642 // has at least <cpyLen> bytes, as implied by the copy-instruction", 643 return false; 644 } 645 646 uint64_t destSize = DL.getTypeAllocSize(StructTy); 647 if (destSize < srcSize) 648 return false; 649 } 650 } else { 651 return false; 652 } 653 654 // Check that dest points to memory that is at least as aligned as src. 655 unsigned srcAlign = srcAlloca->getAlignment(); 656 if (!srcAlign) 657 srcAlign = DL.getABITypeAlignment(srcAlloca->getAllocatedType()); 658 bool isDestSufficientlyAligned = srcAlign <= cpyAlign; 659 // If dest is not aligned enough and we can't increase its alignment then 660 // bail out. 661 if (!isDestSufficientlyAligned && !isa<AllocaInst>(cpyDest)) 662 return false; 663 664 // Check that src is not accessed except via the call and the memcpy. This 665 // guarantees that it holds only undefined values when passed in (so the final 666 // memcpy can be dropped), that it is not read or written between the call and 667 // the memcpy, and that writing beyond the end of it is undefined. 668 SmallVector<User*, 8> srcUseList(srcAlloca->user_begin(), 669 srcAlloca->user_end()); 670 while (!srcUseList.empty()) { 671 User *U = srcUseList.pop_back_val(); 672 673 if (isa<BitCastInst>(U) || isa<AddrSpaceCastInst>(U)) { 674 for (User *UU : U->users()) 675 srcUseList.push_back(UU); 676 continue; 677 } 678 if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(U)) { 679 if (!G->hasAllZeroIndices()) 680 return false; 681 682 for (User *UU : U->users()) 683 srcUseList.push_back(UU); 684 continue; 685 } 686 if (const IntrinsicInst *IT = dyn_cast<IntrinsicInst>(U)) 687 if (IT->getIntrinsicID() == Intrinsic::lifetime_start || 688 IT->getIntrinsicID() == Intrinsic::lifetime_end) 689 continue; 690 691 if (U != C && U != cpy) 692 return false; 693 } 694 695 // Check that src isn't captured by the called function since the 696 // transformation can cause aliasing issues in that case. 697 for (unsigned i = 0, e = CS.arg_size(); i != e; ++i) 698 if (CS.getArgument(i) == cpySrc && !CS.doesNotCapture(i)) 699 return false; 700 701 // Since we're changing the parameter to the callsite, we need to make sure 702 // that what would be the new parameter dominates the callsite. 703 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 704 if (Instruction *cpyDestInst = dyn_cast<Instruction>(cpyDest)) 705 if (!DT.dominates(cpyDestInst, C)) 706 return false; 707 708 // In addition to knowing that the call does not access src in some 709 // unexpected manner, for example via a global, which we deduce from 710 // the use analysis, we also need to know that it does not sneakily 711 // access dest. We rely on AA to figure this out for us. 712 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 713 AliasAnalysis::ModRefResult MR = AA.getModRefInfo(C, cpyDest, srcSize); 714 // If necessary, perform additional analysis. 715 if (MR != AliasAnalysis::NoModRef) 716 MR = AA.callCapturesBefore(C, cpyDest, srcSize, &DT); 717 if (MR != AliasAnalysis::NoModRef) 718 return false; 719 720 // All the checks have passed, so do the transformation. 721 bool changedArgument = false; 722 for (unsigned i = 0; i < CS.arg_size(); ++i) 723 if (CS.getArgument(i)->stripPointerCasts() == cpySrc) { 724 Value *Dest = cpySrc->getType() == cpyDest->getType() ? cpyDest 725 : CastInst::CreatePointerCast(cpyDest, cpySrc->getType(), 726 cpyDest->getName(), C); 727 changedArgument = true; 728 if (CS.getArgument(i)->getType() == Dest->getType()) 729 CS.setArgument(i, Dest); 730 else 731 CS.setArgument(i, CastInst::CreatePointerCast(Dest, 732 CS.getArgument(i)->getType(), Dest->getName(), C)); 733 } 734 735 if (!changedArgument) 736 return false; 737 738 // If the destination wasn't sufficiently aligned then increase its alignment. 739 if (!isDestSufficientlyAligned) { 740 assert(isa<AllocaInst>(cpyDest) && "Can only increase alloca alignment!"); 741 cast<AllocaInst>(cpyDest)->setAlignment(srcAlign); 742 } 743 744 // Drop any cached information about the call, because we may have changed 745 // its dependence information by changing its parameter. 746 MD->removeInstruction(C); 747 748 // Update AA metadata 749 // FIXME: MD_tbaa_struct and MD_mem_parallel_loop_access should also be 750 // handled here, but combineMetadata doesn't support them yet 751 unsigned KnownIDs[] = { 752 LLVMContext::MD_tbaa, 753 LLVMContext::MD_alias_scope, 754 LLVMContext::MD_noalias, 755 }; 756 combineMetadata(C, cpy, KnownIDs); 757 758 // Remove the memcpy. 759 MD->removeInstruction(cpy); 760 ++NumMemCpyInstr; 761 762 return true; 763 } 764 765 /// processMemCpyMemCpyDependence - We've found that the (upward scanning) 766 /// memory dependence of memcpy 'M' is the memcpy 'MDep'. Try to simplify M to 767 /// copy from MDep's input if we can. MSize is the size of M's copy. 768 /// 769 bool MemCpyOpt::processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep, 770 uint64_t MSize) { 771 // We can only transforms memcpy's where the dest of one is the source of the 772 // other. 773 if (M->getSource() != MDep->getDest() || MDep->isVolatile()) 774 return false; 775 776 // If dep instruction is reading from our current input, then it is a noop 777 // transfer and substituting the input won't change this instruction. Just 778 // ignore the input and let someone else zap MDep. This handles cases like: 779 // memcpy(a <- a) 780 // memcpy(b <- a) 781 if (M->getSource() == MDep->getSource()) 782 return false; 783 784 // Second, the length of the memcpy's must be the same, or the preceding one 785 // must be larger than the following one. 786 ConstantInt *MDepLen = dyn_cast<ConstantInt>(MDep->getLength()); 787 ConstantInt *MLen = dyn_cast<ConstantInt>(M->getLength()); 788 if (!MDepLen || !MLen || MDepLen->getZExtValue() < MLen->getZExtValue()) 789 return false; 790 791 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 792 793 // Verify that the copied-from memory doesn't change in between the two 794 // transfers. For example, in: 795 // memcpy(a <- b) 796 // *b = 42; 797 // memcpy(c <- a) 798 // It would be invalid to transform the second memcpy into memcpy(c <- b). 799 // 800 // TODO: If the code between M and MDep is transparent to the destination "c", 801 // then we could still perform the xform by moving M up to the first memcpy. 802 // 803 // NOTE: This is conservative, it will stop on any read from the source loc, 804 // not just the defining memcpy. 805 MemDepResult SourceDep = 806 MD->getPointerDependencyFrom(AA.getLocationForSource(MDep), 807 false, M, M->getParent()); 808 if (!SourceDep.isClobber() || SourceDep.getInst() != MDep) 809 return false; 810 811 // If the dest of the second might alias the source of the first, then the 812 // source and dest might overlap. We still want to eliminate the intermediate 813 // value, but we have to generate a memmove instead of memcpy. 814 bool UseMemMove = false; 815 if (!AA.isNoAlias(AA.getLocationForDest(M), AA.getLocationForSource(MDep))) 816 UseMemMove = true; 817 818 // If all checks passed, then we can transform M. 819 820 // Make sure to use the lesser of the alignment of the source and the dest 821 // since we're changing where we're reading from, but don't want to increase 822 // the alignment past what can be read from or written to. 823 // TODO: Is this worth it if we're creating a less aligned memcpy? For 824 // example we could be moving from movaps -> movq on x86. 825 unsigned Align = std::min(MDep->getAlignment(), M->getAlignment()); 826 827 IRBuilder<> Builder(M); 828 if (UseMemMove) 829 Builder.CreateMemMove(M->getRawDest(), MDep->getRawSource(), M->getLength(), 830 Align, M->isVolatile()); 831 else 832 Builder.CreateMemCpy(M->getRawDest(), MDep->getRawSource(), M->getLength(), 833 Align, M->isVolatile()); 834 835 // Remove the instruction we're replacing. 836 MD->removeInstruction(M); 837 M->eraseFromParent(); 838 ++NumMemCpyInstr; 839 return true; 840 } 841 842 843 /// processMemCpy - perform simplification of memcpy's. If we have memcpy A 844 /// which copies X to Y, and memcpy B which copies Y to Z, then we can rewrite 845 /// B to be a memcpy from X to Z (or potentially a memmove, depending on 846 /// circumstances). This allows later passes to remove the first memcpy 847 /// altogether. 848 bool MemCpyOpt::processMemCpy(MemCpyInst *M) { 849 // We can only optimize non-volatile memcpy's. 850 if (M->isVolatile()) return false; 851 852 // If the source and destination of the memcpy are the same, then zap it. 853 if (M->getSource() == M->getDest()) { 854 MD->removeInstruction(M); 855 M->eraseFromParent(); 856 return false; 857 } 858 859 // If copying from a constant, try to turn the memcpy into a memset. 860 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(M->getSource())) 861 if (GV->isConstant() && GV->hasDefinitiveInitializer()) 862 if (Value *ByteVal = isBytewiseValue(GV->getInitializer())) { 863 IRBuilder<> Builder(M); 864 Builder.CreateMemSet(M->getRawDest(), ByteVal, M->getLength(), 865 M->getAlignment(), false); 866 MD->removeInstruction(M); 867 M->eraseFromParent(); 868 ++NumCpyToSet; 869 return true; 870 } 871 872 // The optimizations after this point require the memcpy size. 873 ConstantInt *CopySize = dyn_cast<ConstantInt>(M->getLength()); 874 if (!CopySize) return false; 875 876 // The are three possible optimizations we can do for memcpy: 877 // a) memcpy-memcpy xform which exposes redundance for DSE. 878 // b) call-memcpy xform for return slot optimization. 879 // c) memcpy from freshly alloca'd space or space that has just started its 880 // lifetime copies undefined data, and we can therefore eliminate the 881 // memcpy in favor of the data that was already at the destination. 882 MemDepResult DepInfo = MD->getDependency(M); 883 if (DepInfo.isClobber()) { 884 if (CallInst *C = dyn_cast<CallInst>(DepInfo.getInst())) { 885 if (performCallSlotOptzn(M, M->getDest(), M->getSource(), 886 CopySize->getZExtValue(), M->getAlignment(), 887 C)) { 888 MD->removeInstruction(M); 889 M->eraseFromParent(); 890 return true; 891 } 892 } 893 } 894 895 AliasAnalysis::Location SrcLoc = AliasAnalysis::getLocationForSource(M); 896 MemDepResult SrcDepInfo = MD->getPointerDependencyFrom(SrcLoc, true, 897 M, M->getParent()); 898 if (SrcDepInfo.isClobber()) { 899 if (MemCpyInst *MDep = dyn_cast<MemCpyInst>(SrcDepInfo.getInst())) 900 return processMemCpyMemCpyDependence(M, MDep, CopySize->getZExtValue()); 901 } else if (SrcDepInfo.isDef()) { 902 Instruction *I = SrcDepInfo.getInst(); 903 bool hasUndefContents = false; 904 905 if (isa<AllocaInst>(I)) { 906 hasUndefContents = true; 907 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 908 if (II->getIntrinsicID() == Intrinsic::lifetime_start) 909 if (ConstantInt *LTSize = dyn_cast<ConstantInt>(II->getArgOperand(0))) 910 if (LTSize->getZExtValue() >= CopySize->getZExtValue()) 911 hasUndefContents = true; 912 } 913 914 if (hasUndefContents) { 915 MD->removeInstruction(M); 916 M->eraseFromParent(); 917 ++NumMemCpyInstr; 918 return true; 919 } 920 } 921 922 return false; 923 } 924 925 /// processMemMove - Transforms memmove calls to memcpy calls when the src/dst 926 /// are guaranteed not to alias. 927 bool MemCpyOpt::processMemMove(MemMoveInst *M) { 928 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 929 930 if (!TLI->has(LibFunc::memmove)) 931 return false; 932 933 // See if the pointers alias. 934 if (!AA.isNoAlias(AA.getLocationForDest(M), AA.getLocationForSource(M))) 935 return false; 936 937 DEBUG(dbgs() << "MemCpyOpt: Optimizing memmove -> memcpy: " << *M << "\n"); 938 939 // If not, then we know we can transform this. 940 Module *Mod = M->getParent()->getParent()->getParent(); 941 Type *ArgTys[3] = { M->getRawDest()->getType(), 942 M->getRawSource()->getType(), 943 M->getLength()->getType() }; 944 M->setCalledFunction(Intrinsic::getDeclaration(Mod, Intrinsic::memcpy, 945 ArgTys)); 946 947 // MemDep may have over conservative information about this instruction, just 948 // conservatively flush it from the cache. 949 MD->removeInstruction(M); 950 951 ++NumMoveToCpy; 952 return true; 953 } 954 955 /// processByValArgument - This is called on every byval argument in call sites. 956 bool MemCpyOpt::processByValArgument(CallSite CS, unsigned ArgNo) { 957 const DataLayout &DL = CS.getCaller()->getParent()->getDataLayout(); 958 // Find out what feeds this byval argument. 959 Value *ByValArg = CS.getArgument(ArgNo); 960 Type *ByValTy = cast<PointerType>(ByValArg->getType())->getElementType(); 961 uint64_t ByValSize = DL.getTypeAllocSize(ByValTy); 962 MemDepResult DepInfo = 963 MD->getPointerDependencyFrom(AliasAnalysis::Location(ByValArg, ByValSize), 964 true, CS.getInstruction(), 965 CS.getInstruction()->getParent()); 966 if (!DepInfo.isClobber()) 967 return false; 968 969 // If the byval argument isn't fed by a memcpy, ignore it. If it is fed by 970 // a memcpy, see if we can byval from the source of the memcpy instead of the 971 // result. 972 MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst()); 973 if (!MDep || MDep->isVolatile() || 974 ByValArg->stripPointerCasts() != MDep->getDest()) 975 return false; 976 977 // The length of the memcpy must be larger or equal to the size of the byval. 978 ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength()); 979 if (!C1 || C1->getValue().getZExtValue() < ByValSize) 980 return false; 981 982 // Get the alignment of the byval. If the call doesn't specify the alignment, 983 // then it is some target specific value that we can't know. 984 unsigned ByValAlign = CS.getParamAlignment(ArgNo+1); 985 if (ByValAlign == 0) return false; 986 987 // If it is greater than the memcpy, then we check to see if we can force the 988 // source of the memcpy to the alignment we need. If we fail, we bail out. 989 AssumptionCache &AC = 990 getAnalysis<AssumptionCacheTracker>().getAssumptionCache( 991 *CS->getParent()->getParent()); 992 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 993 if (MDep->getAlignment() < ByValAlign && 994 getOrEnforceKnownAlignment(MDep->getSource(), ByValAlign, DL, 995 CS.getInstruction(), &AC, &DT) < ByValAlign) 996 return false; 997 998 // Verify that the copied-from memory doesn't change in between the memcpy and 999 // the byval call. 1000 // memcpy(a <- b) 1001 // *b = 42; 1002 // foo(*a) 1003 // It would be invalid to transform the second memcpy into foo(*b). 1004 // 1005 // NOTE: This is conservative, it will stop on any read from the source loc, 1006 // not just the defining memcpy. 1007 MemDepResult SourceDep = 1008 MD->getPointerDependencyFrom(AliasAnalysis::getLocationForSource(MDep), 1009 false, CS.getInstruction(), MDep->getParent()); 1010 if (!SourceDep.isClobber() || SourceDep.getInst() != MDep) 1011 return false; 1012 1013 Value *TmpCast = MDep->getSource(); 1014 if (MDep->getSource()->getType() != ByValArg->getType()) 1015 TmpCast = new BitCastInst(MDep->getSource(), ByValArg->getType(), 1016 "tmpcast", CS.getInstruction()); 1017 1018 DEBUG(dbgs() << "MemCpyOpt: Forwarding memcpy to byval:\n" 1019 << " " << *MDep << "\n" 1020 << " " << *CS.getInstruction() << "\n"); 1021 1022 // Otherwise we're good! Update the byval argument. 1023 CS.setArgument(ArgNo, TmpCast); 1024 ++NumMemCpyInstr; 1025 return true; 1026 } 1027 1028 /// iterateOnFunction - Executes one iteration of MemCpyOpt. 1029 bool MemCpyOpt::iterateOnFunction(Function &F) { 1030 bool MadeChange = false; 1031 1032 // Walk all instruction in the function. 1033 for (Function::iterator BB = F.begin(), BBE = F.end(); BB != BBE; ++BB) { 1034 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE;) { 1035 // Avoid invalidating the iterator. 1036 Instruction *I = BI++; 1037 1038 bool RepeatInstruction = false; 1039 1040 if (StoreInst *SI = dyn_cast<StoreInst>(I)) 1041 MadeChange |= processStore(SI, BI); 1042 else if (MemSetInst *M = dyn_cast<MemSetInst>(I)) 1043 RepeatInstruction = processMemSet(M, BI); 1044 else if (MemCpyInst *M = dyn_cast<MemCpyInst>(I)) 1045 RepeatInstruction = processMemCpy(M); 1046 else if (MemMoveInst *M = dyn_cast<MemMoveInst>(I)) 1047 RepeatInstruction = processMemMove(M); 1048 else if (auto CS = CallSite(I)) { 1049 for (unsigned i = 0, e = CS.arg_size(); i != e; ++i) 1050 if (CS.isByValArgument(i)) 1051 MadeChange |= processByValArgument(CS, i); 1052 } 1053 1054 // Reprocess the instruction if desired. 1055 if (RepeatInstruction) { 1056 if (BI != BB->begin()) --BI; 1057 MadeChange = true; 1058 } 1059 } 1060 } 1061 1062 return MadeChange; 1063 } 1064 1065 // MemCpyOpt::runOnFunction - This is the main transformation entry point for a 1066 // function. 1067 // 1068 bool MemCpyOpt::runOnFunction(Function &F) { 1069 if (skipOptnoneFunction(F)) 1070 return false; 1071 1072 bool MadeChange = false; 1073 MD = &getAnalysis<MemoryDependenceAnalysis>(); 1074 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(); 1075 1076 // If we don't have at least memset and memcpy, there is little point of doing 1077 // anything here. These are required by a freestanding implementation, so if 1078 // even they are disabled, there is no point in trying hard. 1079 if (!TLI->has(LibFunc::memset) || !TLI->has(LibFunc::memcpy)) 1080 return false; 1081 1082 while (1) { 1083 if (!iterateOnFunction(F)) 1084 break; 1085 MadeChange = true; 1086 } 1087 1088 MD = nullptr; 1089 return MadeChange; 1090 } 1091