InstCombineLoadStoreAlloca.cpp revision 36b56886974eae4f9c5ebc96befd3e7bfe5de338
1//===- InstCombineLoadStoreAlloca.cpp -------------------------------------===// 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 file implements the visit functions for load, store and alloca. 11// 12//===----------------------------------------------------------------------===// 13 14#include "InstCombine.h" 15#include "llvm/ADT/Statistic.h" 16#include "llvm/Analysis/Loads.h" 17#include "llvm/IR/DataLayout.h" 18#include "llvm/IR/IntrinsicInst.h" 19#include "llvm/Transforms/Utils/BasicBlockUtils.h" 20#include "llvm/Transforms/Utils/Local.h" 21using namespace llvm; 22 23STATISTIC(NumDeadStore, "Number of dead stores eliminated"); 24STATISTIC(NumGlobalCopies, "Number of allocas copied from constant global"); 25 26/// pointsToConstantGlobal - Return true if V (possibly indirectly) points to 27/// some part of a constant global variable. This intentionally only accepts 28/// constant expressions because we can't rewrite arbitrary instructions. 29static bool pointsToConstantGlobal(Value *V) { 30 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 31 return GV->isConstant(); 32 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 33 if (CE->getOpcode() == Instruction::BitCast || 34 CE->getOpcode() == Instruction::GetElementPtr) 35 return pointsToConstantGlobal(CE->getOperand(0)); 36 return false; 37} 38 39/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 40/// pointer to an alloca. Ignore any reads of the pointer, return false if we 41/// see any stores or other unknown uses. If we see pointer arithmetic, keep 42/// track of whether it moves the pointer (with IsOffset) but otherwise traverse 43/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 44/// the alloca, and if the source pointer is a pointer to a constant global, we 45/// can optimize this. 46static bool 47isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, 48 SmallVectorImpl<Instruction *> &ToDelete, 49 bool IsOffset = false) { 50 // We track lifetime intrinsics as we encounter them. If we decide to go 51 // ahead and replace the value with the global, this lets the caller quickly 52 // eliminate the markers. 53 54 for (Use &U : V->uses()) { 55 Instruction *I = cast<Instruction>(U.getUser()); 56 57 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 58 // Ignore non-volatile loads, they are always ok. 59 if (!LI->isSimple()) return false; 60 continue; 61 } 62 63 if (BitCastInst *BCI = dyn_cast<BitCastInst>(I)) { 64 // If uses of the bitcast are ok, we are ok. 65 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, ToDelete, IsOffset)) 66 return false; 67 continue; 68 } 69 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) { 70 // If the GEP has all zero indices, it doesn't offset the pointer. If it 71 // doesn't, it does. 72 if (!isOnlyCopiedFromConstantGlobal( 73 GEP, TheCopy, ToDelete, IsOffset || !GEP->hasAllZeroIndices())) 74 return false; 75 continue; 76 } 77 78 if (CallSite CS = I) { 79 // If this is the function being called then we treat it like a load and 80 // ignore it. 81 if (CS.isCallee(&U)) 82 continue; 83 84 // Inalloca arguments are clobbered by the call. 85 unsigned ArgNo = CS.getArgumentNo(&U); 86 if (CS.isInAllocaArgument(ArgNo)) 87 return false; 88 89 // If this is a readonly/readnone call site, then we know it is just a 90 // load (but one that potentially returns the value itself), so we can 91 // ignore it if we know that the value isn't captured. 92 if (CS.onlyReadsMemory() && 93 (CS.getInstruction()->use_empty() || CS.doesNotCapture(ArgNo))) 94 continue; 95 96 // If this is being passed as a byval argument, the caller is making a 97 // copy, so it is only a read of the alloca. 98 if (CS.isByValArgument(ArgNo)) 99 continue; 100 } 101 102 // Lifetime intrinsics can be handled by the caller. 103 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 104 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 105 II->getIntrinsicID() == Intrinsic::lifetime_end) { 106 assert(II->use_empty() && "Lifetime markers have no result to use!"); 107 ToDelete.push_back(II); 108 continue; 109 } 110 } 111 112 // If this is isn't our memcpy/memmove, reject it as something we can't 113 // handle. 114 MemTransferInst *MI = dyn_cast<MemTransferInst>(I); 115 if (MI == 0) 116 return false; 117 118 // If the transfer is using the alloca as a source of the transfer, then 119 // ignore it since it is a load (unless the transfer is volatile). 120 if (U.getOperandNo() == 1) { 121 if (MI->isVolatile()) return false; 122 continue; 123 } 124 125 // If we already have seen a copy, reject the second one. 126 if (TheCopy) return false; 127 128 // If the pointer has been offset from the start of the alloca, we can't 129 // safely handle this. 130 if (IsOffset) return false; 131 132 // If the memintrinsic isn't using the alloca as the dest, reject it. 133 if (U.getOperandNo() != 0) return false; 134 135 // If the source of the memcpy/move is not a constant global, reject it. 136 if (!pointsToConstantGlobal(MI->getSource())) 137 return false; 138 139 // Otherwise, the transform is safe. Remember the copy instruction. 140 TheCopy = MI; 141 } 142 return true; 143} 144 145/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 146/// modified by a copy from a constant global. If we can prove this, we can 147/// replace any uses of the alloca with uses of the global directly. 148static MemTransferInst * 149isOnlyCopiedFromConstantGlobal(AllocaInst *AI, 150 SmallVectorImpl<Instruction *> &ToDelete) { 151 MemTransferInst *TheCopy = 0; 152 if (isOnlyCopiedFromConstantGlobal(AI, TheCopy, ToDelete)) 153 return TheCopy; 154 return 0; 155} 156 157Instruction *InstCombiner::visitAllocaInst(AllocaInst &AI) { 158 // Ensure that the alloca array size argument has type intptr_t, so that 159 // any casting is exposed early. 160 if (DL) { 161 Type *IntPtrTy = DL->getIntPtrType(AI.getType()); 162 if (AI.getArraySize()->getType() != IntPtrTy) { 163 Value *V = Builder->CreateIntCast(AI.getArraySize(), 164 IntPtrTy, false); 165 AI.setOperand(0, V); 166 return &AI; 167 } 168 } 169 170 // Convert: alloca Ty, C - where C is a constant != 1 into: alloca [C x Ty], 1 171 if (AI.isArrayAllocation()) { // Check C != 1 172 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) { 173 Type *NewTy = 174 ArrayType::get(AI.getAllocatedType(), C->getZExtValue()); 175 AllocaInst *New = Builder->CreateAlloca(NewTy, 0, AI.getName()); 176 New->setAlignment(AI.getAlignment()); 177 178 // Scan to the end of the allocation instructions, to skip over a block of 179 // allocas if possible...also skip interleaved debug info 180 // 181 BasicBlock::iterator It = New; 182 while (isa<AllocaInst>(*It) || isa<DbgInfoIntrinsic>(*It)) ++It; 183 184 // Now that I is pointing to the first non-allocation-inst in the block, 185 // insert our getelementptr instruction... 186 // 187 Type *IdxTy = DL 188 ? DL->getIntPtrType(AI.getType()) 189 : Type::getInt64Ty(AI.getContext()); 190 Value *NullIdx = Constant::getNullValue(IdxTy); 191 Value *Idx[2] = { NullIdx, NullIdx }; 192 Instruction *GEP = 193 GetElementPtrInst::CreateInBounds(New, Idx, New->getName() + ".sub"); 194 InsertNewInstBefore(GEP, *It); 195 196 // Now make everything use the getelementptr instead of the original 197 // allocation. 198 return ReplaceInstUsesWith(AI, GEP); 199 } else if (isa<UndefValue>(AI.getArraySize())) { 200 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType())); 201 } 202 } 203 204 if (DL && AI.getAllocatedType()->isSized()) { 205 // If the alignment is 0 (unspecified), assign it the preferred alignment. 206 if (AI.getAlignment() == 0) 207 AI.setAlignment(DL->getPrefTypeAlignment(AI.getAllocatedType())); 208 209 // Move all alloca's of zero byte objects to the entry block and merge them 210 // together. Note that we only do this for alloca's, because malloc should 211 // allocate and return a unique pointer, even for a zero byte allocation. 212 if (DL->getTypeAllocSize(AI.getAllocatedType()) == 0) { 213 // For a zero sized alloca there is no point in doing an array allocation. 214 // This is helpful if the array size is a complicated expression not used 215 // elsewhere. 216 if (AI.isArrayAllocation()) { 217 AI.setOperand(0, ConstantInt::get(AI.getArraySize()->getType(), 1)); 218 return &AI; 219 } 220 221 // Get the first instruction in the entry block. 222 BasicBlock &EntryBlock = AI.getParent()->getParent()->getEntryBlock(); 223 Instruction *FirstInst = EntryBlock.getFirstNonPHIOrDbg(); 224 if (FirstInst != &AI) { 225 // If the entry block doesn't start with a zero-size alloca then move 226 // this one to the start of the entry block. There is no problem with 227 // dominance as the array size was forced to a constant earlier already. 228 AllocaInst *EntryAI = dyn_cast<AllocaInst>(FirstInst); 229 if (!EntryAI || !EntryAI->getAllocatedType()->isSized() || 230 DL->getTypeAllocSize(EntryAI->getAllocatedType()) != 0) { 231 AI.moveBefore(FirstInst); 232 return &AI; 233 } 234 235 // If the alignment of the entry block alloca is 0 (unspecified), 236 // assign it the preferred alignment. 237 if (EntryAI->getAlignment() == 0) 238 EntryAI->setAlignment( 239 DL->getPrefTypeAlignment(EntryAI->getAllocatedType())); 240 // Replace this zero-sized alloca with the one at the start of the entry 241 // block after ensuring that the address will be aligned enough for both 242 // types. 243 unsigned MaxAlign = std::max(EntryAI->getAlignment(), 244 AI.getAlignment()); 245 EntryAI->setAlignment(MaxAlign); 246 if (AI.getType() != EntryAI->getType()) 247 return new BitCastInst(EntryAI, AI.getType()); 248 return ReplaceInstUsesWith(AI, EntryAI); 249 } 250 } 251 } 252 253 if (AI.getAlignment()) { 254 // Check to see if this allocation is only modified by a memcpy/memmove from 255 // a constant global whose alignment is equal to or exceeds that of the 256 // allocation. If this is the case, we can change all users to use 257 // the constant global instead. This is commonly produced by the CFE by 258 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 259 // is only subsequently read. 260 SmallVector<Instruction *, 4> ToDelete; 261 if (MemTransferInst *Copy = isOnlyCopiedFromConstantGlobal(&AI, ToDelete)) { 262 unsigned SourceAlign = getOrEnforceKnownAlignment(Copy->getSource(), 263 AI.getAlignment(), DL); 264 if (AI.getAlignment() <= SourceAlign) { 265 DEBUG(dbgs() << "Found alloca equal to global: " << AI << '\n'); 266 DEBUG(dbgs() << " memcpy = " << *Copy << '\n'); 267 for (unsigned i = 0, e = ToDelete.size(); i != e; ++i) 268 EraseInstFromFunction(*ToDelete[i]); 269 Constant *TheSrc = cast<Constant>(Copy->getSource()); 270 Constant *Cast 271 = ConstantExpr::getPointerBitCastOrAddrSpaceCast(TheSrc, AI.getType()); 272 Instruction *NewI = ReplaceInstUsesWith(AI, Cast); 273 EraseInstFromFunction(*Copy); 274 ++NumGlobalCopies; 275 return NewI; 276 } 277 } 278 } 279 280 // At last, use the generic allocation site handler to aggressively remove 281 // unused allocas. 282 return visitAllocSite(AI); 283} 284 285 286/// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible. 287static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI, 288 const DataLayout *DL) { 289 User *CI = cast<User>(LI.getOperand(0)); 290 Value *CastOp = CI->getOperand(0); 291 292 PointerType *DestTy = cast<PointerType>(CI->getType()); 293 Type *DestPTy = DestTy->getElementType(); 294 if (PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) { 295 296 // If the address spaces don't match, don't eliminate the cast. 297 if (DestTy->getAddressSpace() != SrcTy->getAddressSpace()) 298 return 0; 299 300 Type *SrcPTy = SrcTy->getElementType(); 301 302 if (DestPTy->isIntegerTy() || DestPTy->isPointerTy() || 303 DestPTy->isVectorTy()) { 304 // If the source is an array, the code below will not succeed. Check to 305 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for 306 // constants. 307 if (ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy)) 308 if (Constant *CSrc = dyn_cast<Constant>(CastOp)) 309 if (ASrcTy->getNumElements() != 0) { 310 Type *IdxTy = DL 311 ? DL->getIntPtrType(SrcTy) 312 : Type::getInt64Ty(SrcTy->getContext()); 313 Value *Idx = Constant::getNullValue(IdxTy); 314 Value *Idxs[2] = { Idx, Idx }; 315 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs); 316 SrcTy = cast<PointerType>(CastOp->getType()); 317 SrcPTy = SrcTy->getElementType(); 318 } 319 320 if (IC.getDataLayout() && 321 (SrcPTy->isIntegerTy() || SrcPTy->isPointerTy() || 322 SrcPTy->isVectorTy()) && 323 // Do not allow turning this into a load of an integer, which is then 324 // casted to a pointer, this pessimizes pointer analysis a lot. 325 (SrcPTy->isPtrOrPtrVectorTy() == 326 LI.getType()->isPtrOrPtrVectorTy()) && 327 IC.getDataLayout()->getTypeSizeInBits(SrcPTy) == 328 IC.getDataLayout()->getTypeSizeInBits(DestPTy)) { 329 330 // Okay, we are casting from one integer or pointer type to another of 331 // the same size. Instead of casting the pointer before the load, cast 332 // the result of the loaded value. 333 LoadInst *NewLoad = 334 IC.Builder->CreateLoad(CastOp, LI.isVolatile(), CI->getName()); 335 NewLoad->setAlignment(LI.getAlignment()); 336 NewLoad->setAtomic(LI.getOrdering(), LI.getSynchScope()); 337 // Now cast the result of the load. 338 PointerType *OldTy = dyn_cast<PointerType>(NewLoad->getType()); 339 PointerType *NewTy = dyn_cast<PointerType>(LI.getType()); 340 if (OldTy && NewTy && 341 OldTy->getAddressSpace() != NewTy->getAddressSpace()) { 342 return new AddrSpaceCastInst(NewLoad, LI.getType()); 343 } 344 345 return new BitCastInst(NewLoad, LI.getType()); 346 } 347 } 348 } 349 return 0; 350} 351 352Instruction *InstCombiner::visitLoadInst(LoadInst &LI) { 353 Value *Op = LI.getOperand(0); 354 355 // Attempt to improve the alignment. 356 if (DL) { 357 unsigned KnownAlign = 358 getOrEnforceKnownAlignment(Op, DL->getPrefTypeAlignment(LI.getType()),DL); 359 unsigned LoadAlign = LI.getAlignment(); 360 unsigned EffectiveLoadAlign = LoadAlign != 0 ? LoadAlign : 361 DL->getABITypeAlignment(LI.getType()); 362 363 if (KnownAlign > EffectiveLoadAlign) 364 LI.setAlignment(KnownAlign); 365 else if (LoadAlign == 0) 366 LI.setAlignment(EffectiveLoadAlign); 367 } 368 369 // load (cast X) --> cast (load X) iff safe. 370 if (isa<CastInst>(Op)) 371 if (Instruction *Res = InstCombineLoadCast(*this, LI, DL)) 372 return Res; 373 374 // None of the following transforms are legal for volatile/atomic loads. 375 // FIXME: Some of it is okay for atomic loads; needs refactoring. 376 if (!LI.isSimple()) return 0; 377 378 // Do really simple store-to-load forwarding and load CSE, to catch cases 379 // where there are several consecutive memory accesses to the same location, 380 // separated by a few arithmetic operations. 381 BasicBlock::iterator BBI = &LI; 382 if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6)) 383 return ReplaceInstUsesWith(LI, AvailableVal); 384 385 // load(gep null, ...) -> unreachable 386 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) { 387 const Value *GEPI0 = GEPI->getOperand(0); 388 // TODO: Consider a target hook for valid address spaces for this xform. 389 if (isa<ConstantPointerNull>(GEPI0) && GEPI->getPointerAddressSpace() == 0){ 390 // Insert a new store to null instruction before the load to indicate 391 // that this code is not reachable. We do this instead of inserting 392 // an unreachable instruction directly because we cannot modify the 393 // CFG. 394 new StoreInst(UndefValue::get(LI.getType()), 395 Constant::getNullValue(Op->getType()), &LI); 396 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType())); 397 } 398 } 399 400 // load null/undef -> unreachable 401 // TODO: Consider a target hook for valid address spaces for this xform. 402 if (isa<UndefValue>(Op) || 403 (isa<ConstantPointerNull>(Op) && LI.getPointerAddressSpace() == 0)) { 404 // Insert a new store to null instruction before the load to indicate that 405 // this code is not reachable. We do this instead of inserting an 406 // unreachable instruction directly because we cannot modify the CFG. 407 new StoreInst(UndefValue::get(LI.getType()), 408 Constant::getNullValue(Op->getType()), &LI); 409 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType())); 410 } 411 412 // Instcombine load (constantexpr_cast global) -> cast (load global) 413 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) 414 if (CE->isCast()) 415 if (Instruction *Res = InstCombineLoadCast(*this, LI, DL)) 416 return Res; 417 418 if (Op->hasOneUse()) { 419 // Change select and PHI nodes to select values instead of addresses: this 420 // helps alias analysis out a lot, allows many others simplifications, and 421 // exposes redundancy in the code. 422 // 423 // Note that we cannot do the transformation unless we know that the 424 // introduced loads cannot trap! Something like this is valid as long as 425 // the condition is always false: load (select bool %C, int* null, int* %G), 426 // but it would not be valid if we transformed it to load from null 427 // unconditionally. 428 // 429 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) { 430 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2). 431 unsigned Align = LI.getAlignment(); 432 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI, Align, DL) && 433 isSafeToLoadUnconditionally(SI->getOperand(2), SI, Align, DL)) { 434 LoadInst *V1 = Builder->CreateLoad(SI->getOperand(1), 435 SI->getOperand(1)->getName()+".val"); 436 LoadInst *V2 = Builder->CreateLoad(SI->getOperand(2), 437 SI->getOperand(2)->getName()+".val"); 438 V1->setAlignment(Align); 439 V2->setAlignment(Align); 440 return SelectInst::Create(SI->getCondition(), V1, V2); 441 } 442 443 // load (select (cond, null, P)) -> load P 444 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1))) 445 if (C->isNullValue()) { 446 LI.setOperand(0, SI->getOperand(2)); 447 return &LI; 448 } 449 450 // load (select (cond, P, null)) -> load P 451 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2))) 452 if (C->isNullValue()) { 453 LI.setOperand(0, SI->getOperand(1)); 454 return &LI; 455 } 456 } 457 } 458 return 0; 459} 460 461/// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P 462/// when possible. This makes it generally easy to do alias analysis and/or 463/// SROA/mem2reg of the memory object. 464static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) { 465 User *CI = cast<User>(SI.getOperand(1)); 466 Value *CastOp = CI->getOperand(0); 467 468 Type *DestPTy = cast<PointerType>(CI->getType())->getElementType(); 469 PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType()); 470 if (SrcTy == 0) return 0; 471 472 Type *SrcPTy = SrcTy->getElementType(); 473 474 if (!DestPTy->isIntegerTy() && !DestPTy->isPointerTy()) 475 return 0; 476 477 /// NewGEPIndices - If SrcPTy is an aggregate type, we can emit a "noop gep" 478 /// to its first element. This allows us to handle things like: 479 /// store i32 xxx, (bitcast {foo*, float}* %P to i32*) 480 /// on 32-bit hosts. 481 SmallVector<Value*, 4> NewGEPIndices; 482 483 // If the source is an array, the code below will not succeed. Check to 484 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for 485 // constants. 486 if (SrcPTy->isArrayTy() || SrcPTy->isStructTy()) { 487 // Index through pointer. 488 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(SI.getContext())); 489 NewGEPIndices.push_back(Zero); 490 491 while (1) { 492 if (StructType *STy = dyn_cast<StructType>(SrcPTy)) { 493 if (!STy->getNumElements()) /* Struct can be empty {} */ 494 break; 495 NewGEPIndices.push_back(Zero); 496 SrcPTy = STy->getElementType(0); 497 } else if (ArrayType *ATy = dyn_cast<ArrayType>(SrcPTy)) { 498 NewGEPIndices.push_back(Zero); 499 SrcPTy = ATy->getElementType(); 500 } else { 501 break; 502 } 503 } 504 505 SrcTy = PointerType::get(SrcPTy, SrcTy->getAddressSpace()); 506 } 507 508 if (!SrcPTy->isIntegerTy() && !SrcPTy->isPointerTy()) 509 return 0; 510 511 // If the pointers point into different address spaces don't do the 512 // transformation. 513 if (SrcTy->getAddressSpace() != 514 cast<PointerType>(CI->getType())->getAddressSpace()) 515 return 0; 516 517 // If the pointers point to values of different sizes don't do the 518 // transformation. 519 if (!IC.getDataLayout() || 520 IC.getDataLayout()->getTypeSizeInBits(SrcPTy) != 521 IC.getDataLayout()->getTypeSizeInBits(DestPTy)) 522 return 0; 523 524 // If the pointers point to pointers to different address spaces don't do the 525 // transformation. It is not safe to introduce an addrspacecast instruction in 526 // this case since, depending on the target, addrspacecast may not be a no-op 527 // cast. 528 if (SrcPTy->isPointerTy() && DestPTy->isPointerTy() && 529 SrcPTy->getPointerAddressSpace() != DestPTy->getPointerAddressSpace()) 530 return 0; 531 532 // Okay, we are casting from one integer or pointer type to another of 533 // the same size. Instead of casting the pointer before 534 // the store, cast the value to be stored. 535 Value *NewCast; 536 Instruction::CastOps opcode = Instruction::BitCast; 537 Type* CastSrcTy = DestPTy; 538 Type* CastDstTy = SrcPTy; 539 if (CastDstTy->isPointerTy()) { 540 if (CastSrcTy->isIntegerTy()) 541 opcode = Instruction::IntToPtr; 542 } else if (CastDstTy->isIntegerTy()) { 543 if (CastSrcTy->isPointerTy()) 544 opcode = Instruction::PtrToInt; 545 } 546 547 // SIOp0 is a pointer to aggregate and this is a store to the first field, 548 // emit a GEP to index into its first field. 549 if (!NewGEPIndices.empty()) 550 CastOp = IC.Builder->CreateInBoundsGEP(CastOp, NewGEPIndices); 551 552 Value *SIOp0 = SI.getOperand(0); 553 NewCast = IC.Builder->CreateCast(opcode, SIOp0, CastDstTy, 554 SIOp0->getName()+".c"); 555 SI.setOperand(0, NewCast); 556 SI.setOperand(1, CastOp); 557 return &SI; 558} 559 560/// equivalentAddressValues - Test if A and B will obviously have the same 561/// value. This includes recognizing that %t0 and %t1 will have the same 562/// value in code like this: 563/// %t0 = getelementptr \@a, 0, 3 564/// store i32 0, i32* %t0 565/// %t1 = getelementptr \@a, 0, 3 566/// %t2 = load i32* %t1 567/// 568static bool equivalentAddressValues(Value *A, Value *B) { 569 // Test if the values are trivially equivalent. 570 if (A == B) return true; 571 572 // Test if the values come form identical arithmetic instructions. 573 // This uses isIdenticalToWhenDefined instead of isIdenticalTo because 574 // its only used to compare two uses within the same basic block, which 575 // means that they'll always either have the same value or one of them 576 // will have an undefined value. 577 if (isa<BinaryOperator>(A) || 578 isa<CastInst>(A) || 579 isa<PHINode>(A) || 580 isa<GetElementPtrInst>(A)) 581 if (Instruction *BI = dyn_cast<Instruction>(B)) 582 if (cast<Instruction>(A)->isIdenticalToWhenDefined(BI)) 583 return true; 584 585 // Otherwise they may not be equivalent. 586 return false; 587} 588 589Instruction *InstCombiner::visitStoreInst(StoreInst &SI) { 590 Value *Val = SI.getOperand(0); 591 Value *Ptr = SI.getOperand(1); 592 593 // Attempt to improve the alignment. 594 if (DL) { 595 unsigned KnownAlign = 596 getOrEnforceKnownAlignment(Ptr, DL->getPrefTypeAlignment(Val->getType()), 597 DL); 598 unsigned StoreAlign = SI.getAlignment(); 599 unsigned EffectiveStoreAlign = StoreAlign != 0 ? StoreAlign : 600 DL->getABITypeAlignment(Val->getType()); 601 602 if (KnownAlign > EffectiveStoreAlign) 603 SI.setAlignment(KnownAlign); 604 else if (StoreAlign == 0) 605 SI.setAlignment(EffectiveStoreAlign); 606 } 607 608 // Don't hack volatile/atomic stores. 609 // FIXME: Some bits are legal for atomic stores; needs refactoring. 610 if (!SI.isSimple()) return 0; 611 612 // If the RHS is an alloca with a single use, zapify the store, making the 613 // alloca dead. 614 if (Ptr->hasOneUse()) { 615 if (isa<AllocaInst>(Ptr)) 616 return EraseInstFromFunction(SI); 617 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) { 618 if (isa<AllocaInst>(GEP->getOperand(0))) { 619 if (GEP->getOperand(0)->hasOneUse()) 620 return EraseInstFromFunction(SI); 621 } 622 } 623 } 624 625 // Do really simple DSE, to catch cases where there are several consecutive 626 // stores to the same location, separated by a few arithmetic operations. This 627 // situation often occurs with bitfield accesses. 628 BasicBlock::iterator BBI = &SI; 629 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts; 630 --ScanInsts) { 631 --BBI; 632 // Don't count debug info directives, lest they affect codegen, 633 // and we skip pointer-to-pointer bitcasts, which are NOPs. 634 if (isa<DbgInfoIntrinsic>(BBI) || 635 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy())) { 636 ScanInsts++; 637 continue; 638 } 639 640 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) { 641 // Prev store isn't volatile, and stores to the same location? 642 if (PrevSI->isSimple() && equivalentAddressValues(PrevSI->getOperand(1), 643 SI.getOperand(1))) { 644 ++NumDeadStore; 645 ++BBI; 646 EraseInstFromFunction(*PrevSI); 647 continue; 648 } 649 break; 650 } 651 652 // If this is a load, we have to stop. However, if the loaded value is from 653 // the pointer we're loading and is producing the pointer we're storing, 654 // then *this* store is dead (X = load P; store X -> P). 655 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) { 656 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) && 657 LI->isSimple()) 658 return EraseInstFromFunction(SI); 659 660 // Otherwise, this is a load from some other location. Stores before it 661 // may not be dead. 662 break; 663 } 664 665 // Don't skip over loads or things that can modify memory. 666 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory()) 667 break; 668 } 669 670 // store X, null -> turns into 'unreachable' in SimplifyCFG 671 if (isa<ConstantPointerNull>(Ptr) && SI.getPointerAddressSpace() == 0) { 672 if (!isa<UndefValue>(Val)) { 673 SI.setOperand(0, UndefValue::get(Val->getType())); 674 if (Instruction *U = dyn_cast<Instruction>(Val)) 675 Worklist.Add(U); // Dropped a use. 676 } 677 return 0; // Do not modify these! 678 } 679 680 // store undef, Ptr -> noop 681 if (isa<UndefValue>(Val)) 682 return EraseInstFromFunction(SI); 683 684 // If the pointer destination is a cast, see if we can fold the cast into the 685 // source instead. 686 if (isa<CastInst>(Ptr)) 687 if (Instruction *Res = InstCombineStoreToCast(*this, SI)) 688 return Res; 689 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) 690 if (CE->isCast()) 691 if (Instruction *Res = InstCombineStoreToCast(*this, SI)) 692 return Res; 693 694 695 // If this store is the last instruction in the basic block (possibly 696 // excepting debug info instructions), and if the block ends with an 697 // unconditional branch, try to move it to the successor block. 698 BBI = &SI; 699 do { 700 ++BBI; 701 } while (isa<DbgInfoIntrinsic>(BBI) || 702 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy())); 703 if (BranchInst *BI = dyn_cast<BranchInst>(BBI)) 704 if (BI->isUnconditional()) 705 if (SimplifyStoreAtEndOfBlock(SI)) 706 return 0; // xform done! 707 708 return 0; 709} 710 711/// SimplifyStoreAtEndOfBlock - Turn things like: 712/// if () { *P = v1; } else { *P = v2 } 713/// into a phi node with a store in the successor. 714/// 715/// Simplify things like: 716/// *P = v1; if () { *P = v2; } 717/// into a phi node with a store in the successor. 718/// 719bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) { 720 BasicBlock *StoreBB = SI.getParent(); 721 722 // Check to see if the successor block has exactly two incoming edges. If 723 // so, see if the other predecessor contains a store to the same location. 724 // if so, insert a PHI node (if needed) and move the stores down. 725 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0); 726 727 // Determine whether Dest has exactly two predecessors and, if so, compute 728 // the other predecessor. 729 pred_iterator PI = pred_begin(DestBB); 730 BasicBlock *P = *PI; 731 BasicBlock *OtherBB = 0; 732 733 if (P != StoreBB) 734 OtherBB = P; 735 736 if (++PI == pred_end(DestBB)) 737 return false; 738 739 P = *PI; 740 if (P != StoreBB) { 741 if (OtherBB) 742 return false; 743 OtherBB = P; 744 } 745 if (++PI != pred_end(DestBB)) 746 return false; 747 748 // Bail out if all the relevant blocks aren't distinct (this can happen, 749 // for example, if SI is in an infinite loop) 750 if (StoreBB == DestBB || OtherBB == DestBB) 751 return false; 752 753 // Verify that the other block ends in a branch and is not otherwise empty. 754 BasicBlock::iterator BBI = OtherBB->getTerminator(); 755 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI); 756 if (!OtherBr || BBI == OtherBB->begin()) 757 return false; 758 759 // If the other block ends in an unconditional branch, check for the 'if then 760 // else' case. there is an instruction before the branch. 761 StoreInst *OtherStore = 0; 762 if (OtherBr->isUnconditional()) { 763 --BBI; 764 // Skip over debugging info. 765 while (isa<DbgInfoIntrinsic>(BBI) || 766 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy())) { 767 if (BBI==OtherBB->begin()) 768 return false; 769 --BBI; 770 } 771 // If this isn't a store, isn't a store to the same location, or is not the 772 // right kind of store, bail out. 773 OtherStore = dyn_cast<StoreInst>(BBI); 774 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1) || 775 !SI.isSameOperationAs(OtherStore)) 776 return false; 777 } else { 778 // Otherwise, the other block ended with a conditional branch. If one of the 779 // destinations is StoreBB, then we have the if/then case. 780 if (OtherBr->getSuccessor(0) != StoreBB && 781 OtherBr->getSuccessor(1) != StoreBB) 782 return false; 783 784 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an 785 // if/then triangle. See if there is a store to the same ptr as SI that 786 // lives in OtherBB. 787 for (;; --BBI) { 788 // Check to see if we find the matching store. 789 if ((OtherStore = dyn_cast<StoreInst>(BBI))) { 790 if (OtherStore->getOperand(1) != SI.getOperand(1) || 791 !SI.isSameOperationAs(OtherStore)) 792 return false; 793 break; 794 } 795 // If we find something that may be using or overwriting the stored 796 // value, or if we run out of instructions, we can't do the xform. 797 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() || 798 BBI == OtherBB->begin()) 799 return false; 800 } 801 802 // In order to eliminate the store in OtherBr, we have to 803 // make sure nothing reads or overwrites the stored value in 804 // StoreBB. 805 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) { 806 // FIXME: This should really be AA driven. 807 if (I->mayReadFromMemory() || I->mayWriteToMemory()) 808 return false; 809 } 810 } 811 812 // Insert a PHI node now if we need it. 813 Value *MergedVal = OtherStore->getOperand(0); 814 if (MergedVal != SI.getOperand(0)) { 815 PHINode *PN = PHINode::Create(MergedVal->getType(), 2, "storemerge"); 816 PN->addIncoming(SI.getOperand(0), SI.getParent()); 817 PN->addIncoming(OtherStore->getOperand(0), OtherBB); 818 MergedVal = InsertNewInstBefore(PN, DestBB->front()); 819 } 820 821 // Advance to a place where it is safe to insert the new store and 822 // insert it. 823 BBI = DestBB->getFirstInsertionPt(); 824 StoreInst *NewSI = new StoreInst(MergedVal, SI.getOperand(1), 825 SI.isVolatile(), 826 SI.getAlignment(), 827 SI.getOrdering(), 828 SI.getSynchScope()); 829 InsertNewInstBefore(NewSI, *BBI); 830 NewSI->setDebugLoc(OtherStore->getDebugLoc()); 831 832 // If the two stores had the same TBAA tag, preserve it. 833 if (MDNode *TBAATag = SI.getMetadata(LLVMContext::MD_tbaa)) 834 if ((TBAATag = MDNode::getMostGenericTBAA(TBAATag, 835 OtherStore->getMetadata(LLVMContext::MD_tbaa)))) 836 NewSI->setMetadata(LLVMContext::MD_tbaa, TBAATag); 837 838 839 // Nuke the old stores. 840 EraseInstFromFunction(SI); 841 EraseInstFromFunction(*OtherStore); 842 return true; 843} 844