MemCpyOptimizer.cpp revision 824b958e6fb1236e92e4d07f3acf18fca107cdc0
1//===- MemCpyOptimizer.cpp - Optimize use of memcpy and friends -----------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This pass performs various transformations related to eliminating memcpy 11// calls, or transforming sets of stores into memset's. 12// 13//===----------------------------------------------------------------------===// 14 15#define DEBUG_TYPE "memcpyopt" 16#include "llvm/Transforms/Scalar.h" 17#include "llvm/IntrinsicInst.h" 18#include "llvm/Instructions.h" 19#include "llvm/ADT/SmallVector.h" 20#include "llvm/ADT/Statistic.h" 21#include "llvm/Analysis/Dominators.h" 22#include "llvm/Analysis/AliasAnalysis.h" 23#include "llvm/Analysis/MemoryDependenceAnalysis.h" 24#include "llvm/Support/Debug.h" 25#include "llvm/Support/GetElementPtrTypeIterator.h" 26#include "llvm/Target/TargetData.h" 27#include <list> 28using namespace llvm; 29 30STATISTIC(NumMemCpyInstr, "Number of memcpy instructions deleted"); 31STATISTIC(NumMemSetInfer, "Number of memsets inferred"); 32 33/// isBytewiseValue - If the specified value can be set by repeating the same 34/// byte in memory, return the i8 value that it is represented with. This is 35/// true for all i8 values obviously, but is also true for i32 0, i32 -1, 36/// i16 0xF0F0, double 0.0 etc. If the value can't be handled with a repeated 37/// byte store (e.g. i16 0x1234), return null. 38static Value *isBytewiseValue(Value *V) { 39 // All byte-wide stores are splatable, even of arbitrary variables. 40 if (V->getType() == Type::Int8Ty) return V; 41 42 // Constant float and double values can be handled as integer values if the 43 // corresponding integer value is "byteable". An important case is 0.0. 44 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) { 45 if (CFP->getType() == Type::FloatTy) 46 V = ConstantExpr::getBitCast(CFP, Type::Int32Ty); 47 if (CFP->getType() == Type::DoubleTy) 48 V = ConstantExpr::getBitCast(CFP, Type::Int64Ty); 49 // Don't handle long double formats, which have strange constraints. 50 } 51 52 // We can handle constant integers that are power of two in size and a 53 // multiple of 8 bits. 54 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { 55 unsigned Width = CI->getBitWidth(); 56 if (isPowerOf2_32(Width) && Width > 8) { 57 // We can handle this value if the recursive binary decomposition is the 58 // same at all levels. 59 APInt Val = CI->getValue(); 60 APInt Val2; 61 while (Val.getBitWidth() != 8) { 62 unsigned NextWidth = Val.getBitWidth()/2; 63 Val2 = Val.lshr(NextWidth); 64 Val2.trunc(Val.getBitWidth()/2); 65 Val.trunc(Val.getBitWidth()/2); 66 67 // If the top/bottom halves aren't the same, reject it. 68 if (Val != Val2) 69 return 0; 70 } 71 return ConstantInt::get(Val); 72 } 73 } 74 75 // Conceptually, we could handle things like: 76 // %a = zext i8 %X to i16 77 // %b = shl i16 %a, 8 78 // %c = or i16 %a, %b 79 // but until there is an example that actually needs this, it doesn't seem 80 // worth worrying about. 81 return 0; 82} 83 84static int64_t GetOffsetFromIndex(const GetElementPtrInst *GEP, unsigned Idx, 85 bool &VariableIdxFound, TargetData &TD) { 86 // Skip over the first indices. 87 gep_type_iterator GTI = gep_type_begin(GEP); 88 for (unsigned i = 1; i != Idx; ++i, ++GTI) 89 /*skip along*/; 90 91 // Compute the offset implied by the rest of the indices. 92 int64_t Offset = 0; 93 for (unsigned i = Idx, e = GEP->getNumOperands(); i != e; ++i, ++GTI) { 94 ConstantInt *OpC = dyn_cast<ConstantInt>(GEP->getOperand(i)); 95 if (OpC == 0) 96 return VariableIdxFound = true; 97 if (OpC->isZero()) continue; // No offset. 98 99 // Handle struct indices, which add their field offset to the pointer. 100 if (const StructType *STy = dyn_cast<StructType>(*GTI)) { 101 Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue()); 102 continue; 103 } 104 105 // Otherwise, we have a sequential type like an array or vector. Multiply 106 // the index by the ElementSize. 107 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType()); 108 Offset += Size*OpC->getSExtValue(); 109 } 110 111 return Offset; 112} 113 114/// IsPointerOffset - Return true if Ptr1 is provably equal to Ptr2 plus a 115/// constant offset, and return that constant offset. For example, Ptr1 might 116/// be &A[42], and Ptr2 might be &A[40]. In this case offset would be -8. 117static bool IsPointerOffset(Value *Ptr1, Value *Ptr2, int64_t &Offset, 118 TargetData &TD) { 119 // Right now we handle the case when Ptr1/Ptr2 are both GEPs with an identical 120 // base. After that base, they may have some number of common (and 121 // potentially variable) indices. After that they handle some constant 122 // offset, which determines their offset from each other. At this point, we 123 // handle no other case. 124 GetElementPtrInst *GEP1 = dyn_cast<GetElementPtrInst>(Ptr1); 125 GetElementPtrInst *GEP2 = dyn_cast<GetElementPtrInst>(Ptr2); 126 if (!GEP1 || !GEP2 || GEP1->getOperand(0) != GEP2->getOperand(0)) 127 return false; 128 129 // Skip any common indices and track the GEP types. 130 unsigned Idx = 1; 131 for (; Idx != GEP1->getNumOperands() && Idx != GEP2->getNumOperands(); ++Idx) 132 if (GEP1->getOperand(Idx) != GEP2->getOperand(Idx)) 133 break; 134 135 bool VariableIdxFound = false; 136 int64_t Offset1 = GetOffsetFromIndex(GEP1, Idx, VariableIdxFound, TD); 137 int64_t Offset2 = GetOffsetFromIndex(GEP2, Idx, VariableIdxFound, TD); 138 if (VariableIdxFound) return false; 139 140 Offset = Offset2-Offset1; 141 return true; 142} 143 144 145/// MemsetRange - Represents a range of memset'd bytes with the ByteVal value. 146/// This allows us to analyze stores like: 147/// store 0 -> P+1 148/// store 0 -> P+0 149/// store 0 -> P+3 150/// store 0 -> P+2 151/// which sometimes happens with stores to arrays of structs etc. When we see 152/// the first store, we make a range [1, 2). The second store extends the range 153/// to [0, 2). The third makes a new range [2, 3). The fourth store joins the 154/// two ranges into [0, 3) which is memset'able. 155namespace { 156struct MemsetRange { 157 // Start/End - A semi range that describes the span that this range covers. 158 // The range is closed at the start and open at the end: [Start, End). 159 int64_t Start, End; 160 161 /// StartPtr - The getelementptr instruction that points to the start of the 162 /// range. 163 Value *StartPtr; 164 165 /// Alignment - The known alignment of the first store. 166 unsigned Alignment; 167 168 /// TheStores - The actual stores that make up this range. 169 SmallVector<StoreInst*, 16> TheStores; 170 171 bool isProfitableToUseMemset(const TargetData &TD) const; 172 173}; 174} // end anon namespace 175 176bool MemsetRange::isProfitableToUseMemset(const TargetData &TD) const { 177 // If we found more than 8 stores to merge or 64 bytes, use memset. 178 if (TheStores.size() >= 8 || End-Start >= 64) return true; 179 180 // Assume that the code generator is capable of merging pairs of stores 181 // together if it wants to. 182 if (TheStores.size() <= 2) return false; 183 184 // If we have fewer than 8 stores, it can still be worthwhile to do this. 185 // For example, merging 4 i8 stores into an i32 store is useful almost always. 186 // However, merging 2 32-bit stores isn't useful on a 32-bit architecture (the 187 // memset will be split into 2 32-bit stores anyway) and doing so can 188 // pessimize the llvm optimizer. 189 // 190 // Since we don't have perfect knowledge here, make some assumptions: assume 191 // the maximum GPR width is the same size as the pointer size and assume that 192 // this width can be stored. If so, check to see whether we will end up 193 // actually reducing the number of stores used. 194 unsigned Bytes = unsigned(End-Start); 195 unsigned NumPointerStores = Bytes/TD.getPointerSize(); 196 197 // Assume the remaining bytes if any are done a byte at a time. 198 unsigned NumByteStores = Bytes - NumPointerStores*TD.getPointerSize(); 199 200 // If we will reduce the # stores (according to this heuristic), do the 201 // transformation. This encourages merging 4 x i8 -> i32 and 2 x i16 -> i32 202 // etc. 203 return TheStores.size() > NumPointerStores+NumByteStores; 204} 205 206 207namespace { 208class MemsetRanges { 209 /// Ranges - A sorted list of the memset ranges. We use std::list here 210 /// because each element is relatively large and expensive to copy. 211 std::list<MemsetRange> Ranges; 212 typedef std::list<MemsetRange>::iterator range_iterator; 213 TargetData &TD; 214public: 215 MemsetRanges(TargetData &td) : TD(td) {} 216 217 typedef std::list<MemsetRange>::const_iterator const_iterator; 218 const_iterator begin() const { return Ranges.begin(); } 219 const_iterator end() const { return Ranges.end(); } 220 bool empty() const { return Ranges.empty(); } 221 222 void addStore(int64_t OffsetFromFirst, StoreInst *SI); 223}; 224 225} // end anon namespace 226 227 228/// addStore - Add a new store to the MemsetRanges data structure. This adds a 229/// new range for the specified store at the specified offset, merging into 230/// existing ranges as appropriate. 231void MemsetRanges::addStore(int64_t Start, StoreInst *SI) { 232 int64_t End = Start+TD.getTypeStoreSize(SI->getOperand(0)->getType()); 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. 238 range_iterator I = Ranges.begin(), E = Ranges.end(); 239 240 while (I != E && Start > I->End) 241 ++I; 242 243 // We now know that I == E, in which case we didn't find anything to merge 244 // with, or that Start <= I->End. If End < I->Start or I == E, then we need 245 // to insert a new range. Handle this now. 246 if (I == E || End < I->Start) { 247 MemsetRange &R = *Ranges.insert(I, MemsetRange()); 248 R.Start = Start; 249 R.End = End; 250 R.StartPtr = SI->getPointerOperand(); 251 R.Alignment = SI->getAlignment(); 252 R.TheStores.push_back(SI); 253 return; 254 } 255 256 // This store overlaps with I, add it. 257 I->TheStores.push_back(SI); 258 259 // At this point, we may have an interval that completely contains our store. 260 // If so, just add it to the interval and return. 261 if (I->Start <= Start && I->End >= End) 262 return; 263 264 // Now we know that Start <= I->End and End >= I->Start so the range overlaps 265 // but is not entirely contained within the range. 266 267 // See if the range extends the start of the range. In this case, it couldn't 268 // possibly cause it to join the prior range, because otherwise we would have 269 // stopped on *it*. 270 if (Start < I->Start) { 271 I->Start = Start; 272 I->StartPtr = SI->getPointerOperand(); 273 } 274 275 // Now we know that Start <= I->End and Start >= I->Start (so the startpoint 276 // is in or right at the end of I), and that End >= I->Start. Extend I out to 277 // End. 278 if (End > I->End) { 279 I->End = End; 280 range_iterator NextI = I;; 281 while (++NextI != E && End >= NextI->Start) { 282 // Merge the range in. 283 I->TheStores.append(NextI->TheStores.begin(), NextI->TheStores.end()); 284 if (NextI->End > I->End) 285 I->End = NextI->End; 286 Ranges.erase(NextI); 287 NextI = I; 288 } 289 } 290} 291 292//===----------------------------------------------------------------------===// 293// MemCpyOpt Pass 294//===----------------------------------------------------------------------===// 295 296namespace { 297 298 class VISIBILITY_HIDDEN MemCpyOpt : public FunctionPass { 299 bool runOnFunction(Function &F); 300 public: 301 static char ID; // Pass identification, replacement for typeid 302 MemCpyOpt() : FunctionPass(&ID) {} 303 304 private: 305 // This transformation requires dominator postdominator info 306 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 307 AU.setPreservesCFG(); 308 AU.addRequired<DominatorTree>(); 309 AU.addRequired<MemoryDependenceAnalysis>(); 310 AU.addRequired<AliasAnalysis>(); 311 AU.addRequired<TargetData>(); 312 AU.addPreserved<AliasAnalysis>(); 313 AU.addPreserved<MemoryDependenceAnalysis>(); 314 AU.addPreserved<TargetData>(); 315 } 316 317 // Helper fuctions 318 bool processStore(StoreInst *SI, BasicBlock::iterator& BBI); 319 bool processMemCpy(MemCpyInst* M); 320 bool performCallSlotOptzn(MemCpyInst* cpy, CallInst* C); 321 bool iterateOnFunction(Function &F); 322 }; 323 324 char MemCpyOpt::ID = 0; 325} 326 327// createMemCpyOptPass - The public interface to this file... 328FunctionPass *llvm::createMemCpyOptPass() { return new MemCpyOpt(); } 329 330static RegisterPass<MemCpyOpt> X("memcpyopt", 331 "MemCpy Optimization"); 332 333 334 335/// processStore - When GVN is scanning forward over instructions, we look for 336/// some other patterns to fold away. In particular, this looks for stores to 337/// neighboring locations of memory. If it sees enough consequtive ones 338/// (currently 4) it attempts to merge them together into a memcpy/memset. 339bool MemCpyOpt::processStore(StoreInst *SI, BasicBlock::iterator& BBI) { 340 if (SI->isVolatile()) return false; 341 342 // There are two cases that are interesting for this code to handle: memcpy 343 // and memset. Right now we only handle memset. 344 345 // Ensure that the value being stored is something that can be memset'able a 346 // byte at a time like "0" or "-1" or any width, as well as things like 347 // 0xA0A0A0A0 and 0.0. 348 Value *ByteVal = isBytewiseValue(SI->getOperand(0)); 349 if (!ByteVal) 350 return false; 351 352 TargetData &TD = getAnalysis<TargetData>(); 353 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 354 355 // Okay, so we now have a single store that can be splatable. Scan to find 356 // all subsequent stores of the same value to offset from the same pointer. 357 // Join these together into ranges, so we can decide whether contiguous blocks 358 // are stored. 359 MemsetRanges Ranges(TD); 360 361 Value *StartPtr = SI->getPointerOperand(); 362 363 BasicBlock::iterator BI = SI; 364 for (++BI; !isa<TerminatorInst>(BI); ++BI) { 365 if (isa<CallInst>(BI) || isa<InvokeInst>(BI)) { 366 // If the call is readnone, ignore it, otherwise bail out. We don't even 367 // allow readonly here because we don't want something like: 368 // A[1] = 2; strlen(A); A[2] = 2; -> memcpy(A, ...); strlen(A). 369 if (AA.getModRefBehavior(CallSite::get(BI)) == 370 AliasAnalysis::DoesNotAccessMemory) 371 continue; 372 373 // TODO: If this is a memset, try to join it in. 374 375 break; 376 } else if (isa<VAArgInst>(BI) || isa<LoadInst>(BI)) 377 break; 378 379 // If this is a non-store instruction it is fine, ignore it. 380 StoreInst *NextStore = dyn_cast<StoreInst>(BI); 381 if (NextStore == 0) continue; 382 383 // If this is a store, see if we can merge it in. 384 if (NextStore->isVolatile()) break; 385 386 // Check to see if this stored value is of the same byte-splattable value. 387 if (ByteVal != isBytewiseValue(NextStore->getOperand(0))) 388 break; 389 390 // Check to see if this store is to a constant offset from the start ptr. 391 int64_t Offset; 392 if (!IsPointerOffset(StartPtr, NextStore->getPointerOperand(), Offset, TD)) 393 break; 394 395 Ranges.addStore(Offset, NextStore); 396 } 397 398 // If we have no ranges, then we just had a single store with nothing that 399 // could be merged in. This is a very common case of course. 400 if (Ranges.empty()) 401 return false; 402 403 // If we had at least one store that could be merged in, add the starting 404 // store as well. We try to avoid this unless there is at least something 405 // interesting as a small compile-time optimization. 406 Ranges.addStore(0, SI); 407 408 409 Function *MemSetF = 0; 410 411 // Now that we have full information about ranges, loop over the ranges and 412 // emit memset's for anything big enough to be worthwhile. 413 bool MadeChange = false; 414 for (MemsetRanges::const_iterator I = Ranges.begin(), E = Ranges.end(); 415 I != E; ++I) { 416 const MemsetRange &Range = *I; 417 418 if (Range.TheStores.size() == 1) continue; 419 420 // If it is profitable to lower this range to memset, do so now. 421 if (!Range.isProfitableToUseMemset(TD)) 422 continue; 423 424 // Otherwise, we do want to transform this! Create a new memset. We put 425 // the memset right before the first instruction that isn't part of this 426 // memset block. This ensure that the memset is dominated by any addressing 427 // instruction needed by the start of the block. 428 BasicBlock::iterator InsertPt = BI; 429 430 if (MemSetF == 0) { 431 const Type *Tys[] = {Type::Int64Ty}; 432 MemSetF = Intrinsic::getDeclaration(SI->getParent()->getParent() 433 ->getParent(), Intrinsic::memset, 434 Tys, 1); 435 } 436 437 // Get the starting pointer of the block. 438 StartPtr = Range.StartPtr; 439 440 // Cast the start ptr to be i8* as memset requires. 441 const Type *i8Ptr = PointerType::getUnqual(Type::Int8Ty); 442 if (StartPtr->getType() != i8Ptr) 443 StartPtr = new BitCastInst(StartPtr, i8Ptr, StartPtr->getNameStart(), 444 InsertPt); 445 446 Value *Ops[] = { 447 StartPtr, ByteVal, // Start, value 448 ConstantInt::get(Type::Int64Ty, Range.End-Range.Start), // size 449 ConstantInt::get(Type::Int32Ty, Range.Alignment) // align 450 }; 451 Value *C = CallInst::Create(MemSetF, Ops, Ops+4, "", InsertPt); 452 DEBUG(cerr << "Replace stores:\n"; 453 for (unsigned i = 0, e = Range.TheStores.size(); i != e; ++i) 454 cerr << *Range.TheStores[i]; 455 cerr << "With: " << *C); C=C; 456 457 // Don't invalidate the iterator 458 BBI = BI; 459 460 // Zap all the stores. 461 for (SmallVector<StoreInst*, 16>::const_iterator SI = Range.TheStores.begin(), 462 SE = Range.TheStores.end(); SI != SE; ++SI) 463 (*SI)->eraseFromParent(); 464 ++NumMemSetInfer; 465 MadeChange = true; 466 } 467 468 return MadeChange; 469} 470 471 472/// performCallSlotOptzn - takes a memcpy and a call that it depends on, 473/// and checks for the possibility of a call slot optimization by having 474/// the call write its result directly into the destination of the memcpy. 475bool MemCpyOpt::performCallSlotOptzn(MemCpyInst *cpy, CallInst *C) { 476 // The general transformation to keep in mind is 477 // 478 // call @func(..., src, ...) 479 // memcpy(dest, src, ...) 480 // 481 // -> 482 // 483 // memcpy(dest, src, ...) 484 // call @func(..., dest, ...) 485 // 486 // Since moving the memcpy is technically awkward, we additionally check that 487 // src only holds uninitialized values at the moment of the call, meaning that 488 // the memcpy can be discarded rather than moved. 489 490 // Deliberately get the source and destination with bitcasts stripped away, 491 // because we'll need to do type comparisons based on the underlying type. 492 Value* cpyDest = cpy->getDest(); 493 Value* cpySrc = cpy->getSource(); 494 CallSite CS = CallSite::get(C); 495 496 // We need to be able to reason about the size of the memcpy, so we require 497 // that it be a constant. 498 ConstantInt* cpyLength = dyn_cast<ConstantInt>(cpy->getLength()); 499 if (!cpyLength) 500 return false; 501 502 // Require that src be an alloca. This simplifies the reasoning considerably. 503 AllocaInst* srcAlloca = dyn_cast<AllocaInst>(cpySrc); 504 if (!srcAlloca) 505 return false; 506 507 // Check that all of src is copied to dest. 508 TargetData& TD = getAnalysis<TargetData>(); 509 510 ConstantInt* srcArraySize = dyn_cast<ConstantInt>(srcAlloca->getArraySize()); 511 if (!srcArraySize) 512 return false; 513 514 uint64_t srcSize = TD.getABITypeSize(srcAlloca->getAllocatedType()) * 515 srcArraySize->getZExtValue(); 516 517 if (cpyLength->getZExtValue() < srcSize) 518 return false; 519 520 // Check that accessing the first srcSize bytes of dest will not cause a 521 // trap. Otherwise the transform is invalid since it might cause a trap 522 // to occur earlier than it otherwise would. 523 if (AllocaInst* A = dyn_cast<AllocaInst>(cpyDest)) { 524 // The destination is an alloca. Check it is larger than srcSize. 525 ConstantInt* destArraySize = dyn_cast<ConstantInt>(A->getArraySize()); 526 if (!destArraySize) 527 return false; 528 529 uint64_t destSize = TD.getABITypeSize(A->getAllocatedType()) * 530 destArraySize->getZExtValue(); 531 532 if (destSize < srcSize) 533 return false; 534 } else if (Argument* A = dyn_cast<Argument>(cpyDest)) { 535 // If the destination is an sret parameter then only accesses that are 536 // outside of the returned struct type can trap. 537 if (!A->hasStructRetAttr()) 538 return false; 539 540 const Type* StructTy = cast<PointerType>(A->getType())->getElementType(); 541 uint64_t destSize = TD.getABITypeSize(StructTy); 542 543 if (destSize < srcSize) 544 return false; 545 } else { 546 return false; 547 } 548 549 // Check that src is not accessed except via the call and the memcpy. This 550 // guarantees that it holds only undefined values when passed in (so the final 551 // memcpy can be dropped), that it is not read or written between the call and 552 // the memcpy, and that writing beyond the end of it is undefined. 553 SmallVector<User*, 8> srcUseList(srcAlloca->use_begin(), 554 srcAlloca->use_end()); 555 while (!srcUseList.empty()) { 556 User* UI = srcUseList.back(); 557 srcUseList.pop_back(); 558 559 if (isa<BitCastInst>(UI)) { 560 for (User::use_iterator I = UI->use_begin(), E = UI->use_end(); 561 I != E; ++I) 562 srcUseList.push_back(*I); 563 } else if (GetElementPtrInst* G = dyn_cast<GetElementPtrInst>(UI)) { 564 if (G->hasAllZeroIndices()) 565 for (User::use_iterator I = UI->use_begin(), E = UI->use_end(); 566 I != E; ++I) 567 srcUseList.push_back(*I); 568 else 569 return false; 570 } else if (UI != C && UI != cpy) { 571 return false; 572 } 573 } 574 575 // Since we're changing the parameter to the callsite, we need to make sure 576 // that what would be the new parameter dominates the callsite. 577 DominatorTree& DT = getAnalysis<DominatorTree>(); 578 if (Instruction* cpyDestInst = dyn_cast<Instruction>(cpyDest)) 579 if (!DT.dominates(cpyDestInst, C)) 580 return false; 581 582 // In addition to knowing that the call does not access src in some 583 // unexpected manner, for example via a global, which we deduce from 584 // the use analysis, we also need to know that it does not sneakily 585 // access dest. We rely on AA to figure this out for us. 586 AliasAnalysis& AA = getAnalysis<AliasAnalysis>(); 587 if (AA.getModRefInfo(C, cpy->getRawDest(), srcSize) != 588 AliasAnalysis::NoModRef) 589 return false; 590 591 // All the checks have passed, so do the transformation. 592 bool changedArgument = false; 593 for (unsigned i = 0; i < CS.arg_size(); ++i) 594 if (CS.getArgument(i)->stripPointerCasts() == cpySrc) { 595 if (cpySrc->getType() != cpyDest->getType()) 596 cpyDest = CastInst::CreatePointerCast(cpyDest, cpySrc->getType(), 597 cpyDest->getName(), C); 598 changedArgument = true; 599 if (CS.getArgument(i)->getType() != cpyDest->getType()) 600 CS.setArgument(i, CastInst::CreatePointerCast(cpyDest, 601 CS.getArgument(i)->getType(), cpyDest->getName(), C)); 602 else 603 CS.setArgument(i, cpyDest); 604 } 605 606 if (!changedArgument) 607 return false; 608 609 // Drop any cached information about the call, because we may have changed 610 // its dependence information by changing its parameter. 611 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>(); 612 MD.dropInstruction(C); 613 614 // Remove the memcpy 615 MD.removeInstruction(cpy); 616 cpy->eraseFromParent(); 617 NumMemCpyInstr++; 618 619 return true; 620} 621 622/// processMemCpy - perform simplication of memcpy's. If we have memcpy A which 623/// copies X to Y, and memcpy B which copies Y to Z, then we can rewrite B to be 624/// a memcpy from X to Z (or potentially a memmove, depending on circumstances). 625/// This allows later passes to remove the first memcpy altogether. 626bool MemCpyOpt::processMemCpy(MemCpyInst* M) { 627 MemoryDependenceAnalysis& MD = getAnalysis<MemoryDependenceAnalysis>(); 628 629 // The are two possible optimizations we can do for memcpy: 630 // a) memcpy-memcpy xform which exposes redundance for DSE 631 // b) call-memcpy xform for return slot optimization 632 Instruction* dep = MD.getDependency(M); 633 if (dep == MemoryDependenceAnalysis::None || 634 dep == MemoryDependenceAnalysis::NonLocal) 635 return false; 636 else if (!isa<MemCpyInst>(dep)) { 637 if (CallInst* C = dyn_cast<CallInst>(dep)) 638 return performCallSlotOptzn(M, C); 639 else 640 return false; 641 } 642 643 MemCpyInst* MDep = cast<MemCpyInst>(dep); 644 645 // We can only transforms memcpy's where the dest of one is the source of the 646 // other 647 if (M->getSource() != MDep->getDest()) 648 return false; 649 650 // Second, the length of the memcpy's must be the same, or the preceeding one 651 // must be larger than the following one. 652 ConstantInt* C1 = dyn_cast<ConstantInt>(MDep->getLength()); 653 ConstantInt* C2 = dyn_cast<ConstantInt>(M->getLength()); 654 if (!C1 || !C2) 655 return false; 656 657 uint64_t DepSize = C1->getValue().getZExtValue(); 658 uint64_t CpySize = C2->getValue().getZExtValue(); 659 660 if (DepSize < CpySize) 661 return false; 662 663 // Finally, we have to make sure that the dest of the second does not 664 // alias the source of the first 665 AliasAnalysis& AA = getAnalysis<AliasAnalysis>(); 666 if (AA.alias(M->getRawDest(), CpySize, MDep->getRawSource(), DepSize) != 667 AliasAnalysis::NoAlias) 668 return false; 669 else if (AA.alias(M->getRawDest(), CpySize, M->getRawSource(), CpySize) != 670 AliasAnalysis::NoAlias) 671 return false; 672 else if (AA.alias(MDep->getRawDest(), DepSize, MDep->getRawSource(), DepSize) 673 != AliasAnalysis::NoAlias) 674 return false; 675 676 // If all checks passed, then we can transform these memcpy's 677 const Type *Tys[1]; 678 Tys[0] = M->getLength()->getType(); 679 Function* MemCpyFun = Intrinsic::getDeclaration( 680 M->getParent()->getParent()->getParent(), 681 M->getIntrinsicID(), Tys, 1); 682 683 std::vector<Value*> args; 684 args.push_back(M->getRawDest()); 685 args.push_back(MDep->getRawSource()); 686 args.push_back(M->getLength()); 687 args.push_back(M->getAlignment()); 688 689 CallInst* C = CallInst::Create(MemCpyFun, args.begin(), args.end(), "", M); 690 691 692 // If C and M don't interfere, then this is a valid transformation. If they 693 // did, this would mean that the two sources overlap, which would be bad. 694 if (MD.getDependency(C) == MDep) { 695 MD.dropInstruction(M); 696 M->eraseFromParent(); 697 698 NumMemCpyInstr++; 699 700 return true; 701 } 702 703 // Otherwise, there was no point in doing this, so we remove the call we 704 // inserted and act like nothing happened. 705 MD.removeInstruction(C); 706 C->eraseFromParent(); 707 708 return false; 709} 710 711// MemCpyOpt::runOnFunction - This is the main transformation entry point for a 712// function. 713// 714bool MemCpyOpt::runOnFunction(Function& F) { 715 716 bool changed = false; 717 bool shouldContinue = true; 718 719 while (shouldContinue) { 720 shouldContinue = iterateOnFunction(F); 721 changed |= shouldContinue; 722 } 723 724 return changed; 725} 726 727 728// MemCpyOpt::iterateOnFunction - Executes one iteration of GVN 729bool MemCpyOpt::iterateOnFunction(Function &F) { 730 bool changed_function = false; 731 732 // Walk all instruction in the function 733 for (Function::iterator BB = F.begin(), BBE = F.end(); BB != BBE; ++BB) { 734 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); 735 BI != BE;) { 736 // Avoid invalidating the iterator 737 Instruction* I = BI++; 738 739 if (StoreInst *SI = dyn_cast<StoreInst>(I)) 740 changed_function |= processStore(SI, BI); 741 else if (MemCpyInst* M = dyn_cast<MemCpyInst>(I)) { 742 changed_function |= processMemCpy(M); 743 } 744 } 745 } 746 747 return changed_function; 748} 749