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