1 //===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===// 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 transformation implements the well known scalar replacement of 11 // aggregates transformation. This xform breaks up alloca instructions of 12 // aggregate type (structure or array) into individual alloca instructions for 13 // each member (if possible). Then, if possible, it transforms the individual 14 // alloca instructions into nice clean scalar SSA form. 15 // 16 // This combines a simple SRoA algorithm with the Mem2Reg algorithm because 17 // often interact, especially for C++ programs. As such, iterating between 18 // SRoA, then Mem2Reg until we run out of things to promote works well. 19 // 20 //===----------------------------------------------------------------------===// 21 22 #define DEBUG_TYPE "scalarrepl" 23 #include "llvm/Transforms/Scalar.h" 24 #include "llvm/Constants.h" 25 #include "llvm/DerivedTypes.h" 26 #include "llvm/Function.h" 27 #include "llvm/GlobalVariable.h" 28 #include "llvm/Instructions.h" 29 #include "llvm/IntrinsicInst.h" 30 #include "llvm/LLVMContext.h" 31 #include "llvm/Module.h" 32 #include "llvm/Pass.h" 33 #include "llvm/Analysis/DebugInfo.h" 34 #include "llvm/Analysis/DIBuilder.h" 35 #include "llvm/Analysis/Dominators.h" 36 #include "llvm/Analysis/Loads.h" 37 #include "llvm/Analysis/ValueTracking.h" 38 #include "llvm/Target/TargetData.h" 39 #include "llvm/Transforms/Utils/PromoteMemToReg.h" 40 #include "llvm/Transforms/Utils/Local.h" 41 #include "llvm/Transforms/Utils/SSAUpdater.h" 42 #include "llvm/Support/CallSite.h" 43 #include "llvm/Support/Debug.h" 44 #include "llvm/Support/ErrorHandling.h" 45 #include "llvm/Support/GetElementPtrTypeIterator.h" 46 #include "llvm/Support/IRBuilder.h" 47 #include "llvm/Support/MathExtras.h" 48 #include "llvm/Support/raw_ostream.h" 49 #include "llvm/ADT/SetVector.h" 50 #include "llvm/ADT/SmallVector.h" 51 #include "llvm/ADT/Statistic.h" 52 using namespace llvm; 53 54 STATISTIC(NumReplaced, "Number of allocas broken up"); 55 STATISTIC(NumPromoted, "Number of allocas promoted"); 56 STATISTIC(NumAdjusted, "Number of scalar allocas adjusted to allow promotion"); 57 STATISTIC(NumConverted, "Number of aggregates converted to scalar"); 58 STATISTIC(NumGlobals, "Number of allocas copied from constant global"); 59 60 namespace { 61 struct SROA : public FunctionPass { 62 SROA(int T, bool hasDT, char &ID) 63 : FunctionPass(ID), HasDomTree(hasDT) { 64 if (T == -1) 65 SRThreshold = 128; 66 else 67 SRThreshold = T; 68 } 69 70 bool runOnFunction(Function &F); 71 72 bool performScalarRepl(Function &F); 73 bool performPromotion(Function &F); 74 75 private: 76 bool HasDomTree; 77 TargetData *TD; 78 79 /// DeadInsts - Keep track of instructions we have made dead, so that 80 /// we can remove them after we are done working. 81 SmallVector<Value*, 32> DeadInsts; 82 83 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures 84 /// information about the uses. All these fields are initialized to false 85 /// and set to true when something is learned. 86 struct AllocaInfo { 87 /// The alloca to promote. 88 AllocaInst *AI; 89 90 /// CheckedPHIs - This is a set of verified PHI nodes, to prevent infinite 91 /// looping and avoid redundant work. 92 SmallPtrSet<PHINode*, 8> CheckedPHIs; 93 94 /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 95 bool isUnsafe : 1; 96 97 /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 98 bool isMemCpySrc : 1; 99 100 /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 101 bool isMemCpyDst : 1; 102 103 /// hasSubelementAccess - This is true if a subelement of the alloca is 104 /// ever accessed, or false if the alloca is only accessed with mem 105 /// intrinsics or load/store that only access the entire alloca at once. 106 bool hasSubelementAccess : 1; 107 108 /// hasALoadOrStore - This is true if there are any loads or stores to it. 109 /// The alloca may just be accessed with memcpy, for example, which would 110 /// not set this. 111 bool hasALoadOrStore : 1; 112 113 explicit AllocaInfo(AllocaInst *ai) 114 : AI(ai), isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false), 115 hasSubelementAccess(false), hasALoadOrStore(false) {} 116 }; 117 118 unsigned SRThreshold; 119 120 void MarkUnsafe(AllocaInfo &I, Instruction *User) { 121 I.isUnsafe = true; 122 DEBUG(dbgs() << " Transformation preventing inst: " << *User << '\n'); 123 } 124 125 bool isSafeAllocaToScalarRepl(AllocaInst *AI); 126 127 void isSafeForScalarRepl(Instruction *I, uint64_t Offset, AllocaInfo &Info); 128 void isSafePHISelectUseForScalarRepl(Instruction *User, uint64_t Offset, 129 AllocaInfo &Info); 130 void isSafeGEP(GetElementPtrInst *GEPI, uint64_t &Offset, AllocaInfo &Info); 131 void isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 132 Type *MemOpType, bool isStore, AllocaInfo &Info, 133 Instruction *TheAccess, bool AllowWholeAccess); 134 bool TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size); 135 uint64_t FindElementAndOffset(Type *&T, uint64_t &Offset, 136 Type *&IdxTy); 137 138 void DoScalarReplacement(AllocaInst *AI, 139 std::vector<AllocaInst*> &WorkList); 140 void DeleteDeadInstructions(); 141 142 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 143 SmallVector<AllocaInst*, 32> &NewElts); 144 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 145 SmallVector<AllocaInst*, 32> &NewElts); 146 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 147 SmallVector<AllocaInst*, 32> &NewElts); 148 void RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI, 149 uint64_t Offset, 150 SmallVector<AllocaInst*, 32> &NewElts); 151 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 152 AllocaInst *AI, 153 SmallVector<AllocaInst*, 32> &NewElts); 154 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 155 SmallVector<AllocaInst*, 32> &NewElts); 156 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 157 SmallVector<AllocaInst*, 32> &NewElts); 158 159 static MemTransferInst *isOnlyCopiedFromConstantGlobal( 160 AllocaInst *AI, SmallVector<Instruction*, 4> &ToDelete); 161 }; 162 163 // SROA_DT - SROA that uses DominatorTree. 164 struct SROA_DT : public SROA { 165 static char ID; 166 public: 167 SROA_DT(int T = -1) : SROA(T, true, ID) { 168 initializeSROA_DTPass(*PassRegistry::getPassRegistry()); 169 } 170 171 // getAnalysisUsage - This pass does not require any passes, but we know it 172 // will not alter the CFG, so say so. 173 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 174 AU.addRequired<DominatorTree>(); 175 AU.setPreservesCFG(); 176 } 177 }; 178 179 // SROA_SSAUp - SROA that uses SSAUpdater. 180 struct SROA_SSAUp : public SROA { 181 static char ID; 182 public: 183 SROA_SSAUp(int T = -1) : SROA(T, false, ID) { 184 initializeSROA_SSAUpPass(*PassRegistry::getPassRegistry()); 185 } 186 187 // getAnalysisUsage - This pass does not require any passes, but we know it 188 // will not alter the CFG, so say so. 189 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 190 AU.setPreservesCFG(); 191 } 192 }; 193 194 } 195 196 char SROA_DT::ID = 0; 197 char SROA_SSAUp::ID = 0; 198 199 INITIALIZE_PASS_BEGIN(SROA_DT, "scalarrepl", 200 "Scalar Replacement of Aggregates (DT)", false, false) 201 INITIALIZE_PASS_DEPENDENCY(DominatorTree) 202 INITIALIZE_PASS_END(SROA_DT, "scalarrepl", 203 "Scalar Replacement of Aggregates (DT)", false, false) 204 205 INITIALIZE_PASS_BEGIN(SROA_SSAUp, "scalarrepl-ssa", 206 "Scalar Replacement of Aggregates (SSAUp)", false, false) 207 INITIALIZE_PASS_END(SROA_SSAUp, "scalarrepl-ssa", 208 "Scalar Replacement of Aggregates (SSAUp)", false, false) 209 210 // Public interface to the ScalarReplAggregates pass 211 FunctionPass *llvm::createScalarReplAggregatesPass(int Threshold, 212 bool UseDomTree) { 213 if (UseDomTree) 214 return new SROA_DT(Threshold); 215 return new SROA_SSAUp(Threshold); 216 } 217 218 219 //===----------------------------------------------------------------------===// 220 // Convert To Scalar Optimization. 221 //===----------------------------------------------------------------------===// 222 223 namespace { 224 /// ConvertToScalarInfo - This class implements the "Convert To Scalar" 225 /// optimization, which scans the uses of an alloca and determines if it can 226 /// rewrite it in terms of a single new alloca that can be mem2reg'd. 227 class ConvertToScalarInfo { 228 /// AllocaSize - The size of the alloca being considered in bytes. 229 unsigned AllocaSize; 230 const TargetData &TD; 231 232 /// IsNotTrivial - This is set to true if there is some access to the object 233 /// which means that mem2reg can't promote it. 234 bool IsNotTrivial; 235 236 /// ScalarKind - Tracks the kind of alloca being considered for promotion, 237 /// computed based on the uses of the alloca rather than the LLVM type system. 238 enum { 239 Unknown, 240 241 // Accesses via GEPs that are consistent with element access of a vector 242 // type. This will not be converted into a vector unless there is a later 243 // access using an actual vector type. 244 ImplicitVector, 245 246 // Accesses via vector operations and GEPs that are consistent with the 247 // layout of a vector type. 248 Vector, 249 250 // An integer bag-of-bits with bitwise operations for insertion and 251 // extraction. Any combination of types can be converted into this kind 252 // of scalar. 253 Integer 254 } ScalarKind; 255 256 /// VectorTy - This tracks the type that we should promote the vector to if 257 /// it is possible to turn it into a vector. This starts out null, and if it 258 /// isn't possible to turn into a vector type, it gets set to VoidTy. 259 VectorType *VectorTy; 260 261 /// HadNonMemTransferAccess - True if there is at least one access to the 262 /// alloca that is not a MemTransferInst. We don't want to turn structs into 263 /// large integers unless there is some potential for optimization. 264 bool HadNonMemTransferAccess; 265 266 public: 267 explicit ConvertToScalarInfo(unsigned Size, const TargetData &td) 268 : AllocaSize(Size), TD(td), IsNotTrivial(false), ScalarKind(Unknown), 269 VectorTy(0), HadNonMemTransferAccess(false) { } 270 271 AllocaInst *TryConvert(AllocaInst *AI); 272 273 private: 274 bool CanConvertToScalar(Value *V, uint64_t Offset); 275 void MergeInTypeForLoadOrStore(Type *In, uint64_t Offset); 276 bool MergeInVectorType(VectorType *VInTy, uint64_t Offset); 277 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); 278 279 Value *ConvertScalar_ExtractValue(Value *NV, Type *ToType, 280 uint64_t Offset, IRBuilder<> &Builder); 281 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 282 uint64_t Offset, IRBuilder<> &Builder); 283 }; 284 } // end anonymous namespace. 285 286 287 /// TryConvert - Analyze the specified alloca, and if it is safe to do so, 288 /// rewrite it to be a new alloca which is mem2reg'able. This returns the new 289 /// alloca if possible or null if not. 290 AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) { 291 // If we can't convert this scalar, or if mem2reg can trivially do it, bail 292 // out. 293 if (!CanConvertToScalar(AI, 0) || !IsNotTrivial) 294 return 0; 295 296 // If an alloca has only memset / memcpy uses, it may still have an Unknown 297 // ScalarKind. Treat it as an Integer below. 298 if (ScalarKind == Unknown) 299 ScalarKind = Integer; 300 301 if (ScalarKind == Vector && VectorTy->getBitWidth() != AllocaSize * 8) 302 ScalarKind = Integer; 303 304 // If we were able to find a vector type that can handle this with 305 // insert/extract elements, and if there was at least one use that had 306 // a vector type, promote this to a vector. We don't want to promote 307 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 308 // we just get a lot of insert/extracts. If at least one vector is 309 // involved, then we probably really do have a union of vector/array. 310 Type *NewTy; 311 if (ScalarKind == Vector) { 312 assert(VectorTy && "Missing type for vector scalar."); 313 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = " 314 << *VectorTy << '\n'); 315 NewTy = VectorTy; // Use the vector type. 316 } else { 317 unsigned BitWidth = AllocaSize * 8; 318 if ((ScalarKind == ImplicitVector || ScalarKind == Integer) && 319 !HadNonMemTransferAccess && !TD.fitsInLegalInteger(BitWidth)) 320 return 0; 321 322 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"); 323 // Create and insert the integer alloca. 324 NewTy = IntegerType::get(AI->getContext(), BitWidth); 325 } 326 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 327 ConvertUsesToScalar(AI, NewAI, 0); 328 return NewAI; 329 } 330 331 /// MergeInTypeForLoadOrStore - Add the 'In' type to the accumulated vector type 332 /// (VectorTy) so far at the offset specified by Offset (which is specified in 333 /// bytes). 334 /// 335 /// There are two cases we handle here: 336 /// 1) A union of vector types of the same size and potentially its elements. 337 /// Here we turn element accesses into insert/extract element operations. 338 /// This promotes a <4 x float> with a store of float to the third element 339 /// into a <4 x float> that uses insert element. 340 /// 2) A fully general blob of memory, which we turn into some (potentially 341 /// large) integer type with extract and insert operations where the loads 342 /// and stores would mutate the memory. We mark this by setting VectorTy 343 /// to VoidTy. 344 void ConvertToScalarInfo::MergeInTypeForLoadOrStore(Type *In, 345 uint64_t Offset) { 346 // If we already decided to turn this into a blob of integer memory, there is 347 // nothing to be done. 348 if (ScalarKind == Integer) 349 return; 350 351 // If this could be contributing to a vector, analyze it. 352 353 // If the In type is a vector that is the same size as the alloca, see if it 354 // matches the existing VecTy. 355 if (VectorType *VInTy = dyn_cast<VectorType>(In)) { 356 if (MergeInVectorType(VInTy, Offset)) 357 return; 358 } else if (In->isFloatTy() || In->isDoubleTy() || 359 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 && 360 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 361 // Full width accesses can be ignored, because they can always be turned 362 // into bitcasts. 363 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 364 if (EltSize == AllocaSize) 365 return; 366 367 // If we're accessing something that could be an element of a vector, see 368 // if the implied vector agrees with what we already have and if Offset is 369 // compatible with it. 370 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 && 371 (!VectorTy || EltSize == VectorTy->getElementType() 372 ->getPrimitiveSizeInBits()/8)) { 373 if (!VectorTy) { 374 ScalarKind = ImplicitVector; 375 VectorTy = VectorType::get(In, AllocaSize/EltSize); 376 } 377 return; 378 } 379 } 380 381 // Otherwise, we have a case that we can't handle with an optimized vector 382 // form. We can still turn this into a large integer. 383 ScalarKind = Integer; 384 } 385 386 /// MergeInVectorType - Handles the vector case of MergeInTypeForLoadOrStore, 387 /// returning true if the type was successfully merged and false otherwise. 388 bool ConvertToScalarInfo::MergeInVectorType(VectorType *VInTy, 389 uint64_t Offset) { 390 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 391 // If we're storing/loading a vector of the right size, allow it as a 392 // vector. If this the first vector we see, remember the type so that 393 // we know the element size. If this is a subsequent access, ignore it 394 // even if it is a differing type but the same size. Worst case we can 395 // bitcast the resultant vectors. 396 if (!VectorTy) 397 VectorTy = VInTy; 398 ScalarKind = Vector; 399 return true; 400 } 401 402 return false; 403 } 404 405 /// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 406 /// its accesses to a single vector type, return true and set VecTy to 407 /// the new type. If we could convert the alloca into a single promotable 408 /// integer, return true but set VecTy to VoidTy. Further, if the use is not a 409 /// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 410 /// is the current offset from the base of the alloca being analyzed. 411 /// 412 /// If we see at least one access to the value that is as a vector type, set the 413 /// SawVec flag. 414 bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) { 415 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 416 Instruction *User = cast<Instruction>(*UI); 417 418 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 419 // Don't break volatile loads. 420 if (!LI->isSimple()) 421 return false; 422 // Don't touch MMX operations. 423 if (LI->getType()->isX86_MMXTy()) 424 return false; 425 HadNonMemTransferAccess = true; 426 MergeInTypeForLoadOrStore(LI->getType(), Offset); 427 continue; 428 } 429 430 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 431 // Storing the pointer, not into the value? 432 if (SI->getOperand(0) == V || !SI->isSimple()) return false; 433 // Don't touch MMX operations. 434 if (SI->getOperand(0)->getType()->isX86_MMXTy()) 435 return false; 436 HadNonMemTransferAccess = true; 437 MergeInTypeForLoadOrStore(SI->getOperand(0)->getType(), Offset); 438 continue; 439 } 440 441 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 442 if (!onlyUsedByLifetimeMarkers(BCI)) 443 IsNotTrivial = true; // Can't be mem2reg'd. 444 if (!CanConvertToScalar(BCI, Offset)) 445 return false; 446 continue; 447 } 448 449 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 450 // If this is a GEP with a variable indices, we can't handle it. 451 if (!GEP->hasAllConstantIndices()) 452 return false; 453 454 // Compute the offset that this GEP adds to the pointer. 455 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 456 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 457 Indices); 458 // See if all uses can be converted. 459 if (!CanConvertToScalar(GEP, Offset+GEPOffset)) 460 return false; 461 IsNotTrivial = true; // Can't be mem2reg'd. 462 HadNonMemTransferAccess = true; 463 continue; 464 } 465 466 // If this is a constant sized memset of a constant value (e.g. 0) we can 467 // handle it. 468 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 469 // Store of constant value. 470 if (!isa<ConstantInt>(MSI->getValue())) 471 return false; 472 473 // Store of constant size. 474 ConstantInt *Len = dyn_cast<ConstantInt>(MSI->getLength()); 475 if (!Len) 476 return false; 477 478 // If the size differs from the alloca, we can only convert the alloca to 479 // an integer bag-of-bits. 480 // FIXME: This should handle all of the cases that are currently accepted 481 // as vector element insertions. 482 if (Len->getZExtValue() != AllocaSize || Offset != 0) 483 ScalarKind = Integer; 484 485 IsNotTrivial = true; // Can't be mem2reg'd. 486 HadNonMemTransferAccess = true; 487 continue; 488 } 489 490 // If this is a memcpy or memmove into or out of the whole allocation, we 491 // can handle it like a load or store of the scalar type. 492 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 493 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()); 494 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0) 495 return false; 496 497 IsNotTrivial = true; // Can't be mem2reg'd. 498 continue; 499 } 500 501 // If this is a lifetime intrinsic, we can handle it. 502 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 503 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 504 II->getIntrinsicID() == Intrinsic::lifetime_end) { 505 continue; 506 } 507 } 508 509 // Otherwise, we cannot handle this! 510 return false; 511 } 512 513 return true; 514 } 515 516 /// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 517 /// directly. This happens when we are converting an "integer union" to a 518 /// single integer scalar, or when we are converting a "vector union" to a 519 /// vector with insert/extractelement instructions. 520 /// 521 /// Offset is an offset from the original alloca, in bits that need to be 522 /// shifted to the right. By the end of this, there should be no uses of Ptr. 523 void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, 524 uint64_t Offset) { 525 while (!Ptr->use_empty()) { 526 Instruction *User = cast<Instruction>(Ptr->use_back()); 527 528 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 529 ConvertUsesToScalar(CI, NewAI, Offset); 530 CI->eraseFromParent(); 531 continue; 532 } 533 534 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 535 // Compute the offset that this GEP adds to the pointer. 536 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 537 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 538 Indices); 539 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 540 GEP->eraseFromParent(); 541 continue; 542 } 543 544 IRBuilder<> Builder(User); 545 546 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 547 // The load is a bit extract from NewAI shifted right by Offset bits. 548 Value *LoadedVal = Builder.CreateLoad(NewAI); 549 Value *NewLoadVal 550 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 551 LI->replaceAllUsesWith(NewLoadVal); 552 LI->eraseFromParent(); 553 continue; 554 } 555 556 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 557 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 558 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 559 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 560 Builder); 561 Builder.CreateStore(New, NewAI); 562 SI->eraseFromParent(); 563 564 // If the load we just inserted is now dead, then the inserted store 565 // overwrote the entire thing. 566 if (Old->use_empty()) 567 Old->eraseFromParent(); 568 continue; 569 } 570 571 // If this is a constant sized memset of a constant value (e.g. 0) we can 572 // transform it into a store of the expanded constant value. 573 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 574 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 575 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 576 if (NumBytes != 0) { 577 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 578 579 // Compute the value replicated the right number of times. 580 APInt APVal(NumBytes*8, Val); 581 582 // Splat the value if non-zero. 583 if (Val) 584 for (unsigned i = 1; i != NumBytes; ++i) 585 APVal |= APVal << 8; 586 587 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 588 Value *New = ConvertScalar_InsertValue( 589 ConstantInt::get(User->getContext(), APVal), 590 Old, Offset, Builder); 591 Builder.CreateStore(New, NewAI); 592 593 // If the load we just inserted is now dead, then the memset overwrote 594 // the entire thing. 595 if (Old->use_empty()) 596 Old->eraseFromParent(); 597 } 598 MSI->eraseFromParent(); 599 continue; 600 } 601 602 // If this is a memcpy or memmove into or out of the whole allocation, we 603 // can handle it like a load or store of the scalar type. 604 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 605 assert(Offset == 0 && "must be store to start of alloca"); 606 607 // If the source and destination are both to the same alloca, then this is 608 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 609 // as appropriate. 610 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, &TD, 0)); 611 612 if (GetUnderlyingObject(MTI->getSource(), &TD, 0) != OrigAI) { 613 // Dest must be OrigAI, change this to be a load from the original 614 // pointer (bitcasted), then a store to our new alloca. 615 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 616 Value *SrcPtr = MTI->getSource(); 617 PointerType* SPTy = cast<PointerType>(SrcPtr->getType()); 618 PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 619 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 620 AIPTy = PointerType::get(AIPTy->getElementType(), 621 SPTy->getAddressSpace()); 622 } 623 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy); 624 625 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 626 SrcVal->setAlignment(MTI->getAlignment()); 627 Builder.CreateStore(SrcVal, NewAI); 628 } else if (GetUnderlyingObject(MTI->getDest(), &TD, 0) != OrigAI) { 629 // Src must be OrigAI, change this to be a load from NewAI then a store 630 // through the original dest pointer (bitcasted). 631 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 632 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 633 634 PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType()); 635 PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 636 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 637 AIPTy = PointerType::get(AIPTy->getElementType(), 638 DPTy->getAddressSpace()); 639 } 640 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy); 641 642 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 643 NewStore->setAlignment(MTI->getAlignment()); 644 } else { 645 // Noop transfer. Src == Dst 646 } 647 648 MTI->eraseFromParent(); 649 continue; 650 } 651 652 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 653 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 654 II->getIntrinsicID() == Intrinsic::lifetime_end) { 655 // There's no need to preserve these, as the resulting alloca will be 656 // converted to a register anyways. 657 II->eraseFromParent(); 658 continue; 659 } 660 } 661 662 llvm_unreachable("Unsupported operation!"); 663 } 664 } 665 666 /// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 667 /// or vector value FromVal, extracting the bits from the offset specified by 668 /// Offset. This returns the value, which is of type ToType. 669 /// 670 /// This happens when we are converting an "integer union" to a single 671 /// integer scalar, or when we are converting a "vector union" to a vector with 672 /// insert/extractelement instructions. 673 /// 674 /// Offset is an offset from the original alloca, in bits that need to be 675 /// shifted to the right. 676 Value *ConvertToScalarInfo:: 677 ConvertScalar_ExtractValue(Value *FromVal, Type *ToType, 678 uint64_t Offset, IRBuilder<> &Builder) { 679 // If the load is of the whole new alloca, no conversion is needed. 680 Type *FromType = FromVal->getType(); 681 if (FromType == ToType && Offset == 0) 682 return FromVal; 683 684 // If the result alloca is a vector type, this is either an element 685 // access or a bitcast to another vector type of the same size. 686 if (VectorType *VTy = dyn_cast<VectorType>(FromType)) { 687 unsigned FromTypeSize = TD.getTypeAllocSize(FromType); 688 unsigned ToTypeSize = TD.getTypeAllocSize(ToType); 689 if (FromTypeSize == ToTypeSize) 690 return Builder.CreateBitCast(FromVal, ToType); 691 692 // Otherwise it must be an element access. 693 unsigned Elt = 0; 694 if (Offset) { 695 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 696 Elt = Offset/EltSize; 697 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 698 } 699 // Return the element extracted out of it. 700 Value *V = Builder.CreateExtractElement(FromVal, Builder.getInt32(Elt)); 701 if (V->getType() != ToType) 702 V = Builder.CreateBitCast(V, ToType); 703 return V; 704 } 705 706 // If ToType is a first class aggregate, extract out each of the pieces and 707 // use insertvalue's to form the FCA. 708 if (StructType *ST = dyn_cast<StructType>(ToType)) { 709 const StructLayout &Layout = *TD.getStructLayout(ST); 710 Value *Res = UndefValue::get(ST); 711 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 712 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 713 Offset+Layout.getElementOffsetInBits(i), 714 Builder); 715 Res = Builder.CreateInsertValue(Res, Elt, i); 716 } 717 return Res; 718 } 719 720 if (ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 721 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 722 Value *Res = UndefValue::get(AT); 723 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 724 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 725 Offset+i*EltSize, Builder); 726 Res = Builder.CreateInsertValue(Res, Elt, i); 727 } 728 return Res; 729 } 730 731 // Otherwise, this must be a union that was converted to an integer value. 732 IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 733 734 // If this is a big-endian system and the load is narrower than the 735 // full alloca type, we need to do a shift to get the right bits. 736 int ShAmt = 0; 737 if (TD.isBigEndian()) { 738 // On big-endian machines, the lowest bit is stored at the bit offset 739 // from the pointer given by getTypeStoreSizeInBits. This matters for 740 // integers with a bitwidth that is not a multiple of 8. 741 ShAmt = TD.getTypeStoreSizeInBits(NTy) - 742 TD.getTypeStoreSizeInBits(ToType) - Offset; 743 } else { 744 ShAmt = Offset; 745 } 746 747 // Note: we support negative bitwidths (with shl) which are not defined. 748 // We do this to support (f.e.) loads off the end of a structure where 749 // only some bits are used. 750 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 751 FromVal = Builder.CreateLShr(FromVal, 752 ConstantInt::get(FromVal->getType(), ShAmt)); 753 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 754 FromVal = Builder.CreateShl(FromVal, 755 ConstantInt::get(FromVal->getType(), -ShAmt)); 756 757 // Finally, unconditionally truncate the integer to the right width. 758 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType); 759 if (LIBitWidth < NTy->getBitWidth()) 760 FromVal = 761 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 762 LIBitWidth)); 763 else if (LIBitWidth > NTy->getBitWidth()) 764 FromVal = 765 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 766 LIBitWidth)); 767 768 // If the result is an integer, this is a trunc or bitcast. 769 if (ToType->isIntegerTy()) { 770 // Should be done. 771 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 772 // Just do a bitcast, we know the sizes match up. 773 FromVal = Builder.CreateBitCast(FromVal, ToType); 774 } else { 775 // Otherwise must be a pointer. 776 FromVal = Builder.CreateIntToPtr(FromVal, ToType); 777 } 778 assert(FromVal->getType() == ToType && "Didn't convert right?"); 779 return FromVal; 780 } 781 782 /// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 783 /// or vector value "Old" at the offset specified by Offset. 784 /// 785 /// This happens when we are converting an "integer union" to a 786 /// single integer scalar, or when we are converting a "vector union" to a 787 /// vector with insert/extractelement instructions. 788 /// 789 /// Offset is an offset from the original alloca, in bits that need to be 790 /// shifted to the right. 791 Value *ConvertToScalarInfo:: 792 ConvertScalar_InsertValue(Value *SV, Value *Old, 793 uint64_t Offset, IRBuilder<> &Builder) { 794 // Convert the stored type to the actual type, shift it left to insert 795 // then 'or' into place. 796 Type *AllocaType = Old->getType(); 797 LLVMContext &Context = Old->getContext(); 798 799 if (VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 800 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy); 801 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType()); 802 803 // Changing the whole vector with memset or with an access of a different 804 // vector type? 805 if (ValSize == VecSize) 806 return Builder.CreateBitCast(SV, AllocaType); 807 808 // Must be an element insertion. 809 assert(SV->getType() == VTy->getElementType()); 810 uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 811 unsigned Elt = Offset/EltSize; 812 return Builder.CreateInsertElement(Old, SV, Builder.getInt32(Elt)); 813 } 814 815 // If SV is a first-class aggregate value, insert each value recursively. 816 if (StructType *ST = dyn_cast<StructType>(SV->getType())) { 817 const StructLayout &Layout = *TD.getStructLayout(ST); 818 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 819 Value *Elt = Builder.CreateExtractValue(SV, i); 820 Old = ConvertScalar_InsertValue(Elt, Old, 821 Offset+Layout.getElementOffsetInBits(i), 822 Builder); 823 } 824 return Old; 825 } 826 827 if (ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 828 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 829 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 830 Value *Elt = Builder.CreateExtractValue(SV, i); 831 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 832 } 833 return Old; 834 } 835 836 // If SV is a float, convert it to the appropriate integer type. 837 // If it is a pointer, do the same. 838 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 839 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); 840 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); 841 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); 842 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 843 SV = Builder.CreateBitCast(SV, IntegerType::get(SV->getContext(),SrcWidth)); 844 else if (SV->getType()->isPointerTy()) 845 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext())); 846 847 // Zero extend or truncate the value if needed. 848 if (SV->getType() != AllocaType) { 849 if (SV->getType()->getPrimitiveSizeInBits() < 850 AllocaType->getPrimitiveSizeInBits()) 851 SV = Builder.CreateZExt(SV, AllocaType); 852 else { 853 // Truncation may be needed if storing more than the alloca can hold 854 // (undefined behavior). 855 SV = Builder.CreateTrunc(SV, AllocaType); 856 SrcWidth = DestWidth; 857 SrcStoreWidth = DestStoreWidth; 858 } 859 } 860 861 // If this is a big-endian system and the store is narrower than the 862 // full alloca type, we need to do a shift to get the right bits. 863 int ShAmt = 0; 864 if (TD.isBigEndian()) { 865 // On big-endian machines, the lowest bit is stored at the bit offset 866 // from the pointer given by getTypeStoreSizeInBits. This matters for 867 // integers with a bitwidth that is not a multiple of 8. 868 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 869 } else { 870 ShAmt = Offset; 871 } 872 873 // Note: we support negative bitwidths (with shr) which are not defined. 874 // We do this to support (f.e.) stores off the end of a structure where 875 // only some bits in the structure are set. 876 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 877 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 878 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), ShAmt)); 879 Mask <<= ShAmt; 880 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 881 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), -ShAmt)); 882 Mask = Mask.lshr(-ShAmt); 883 } 884 885 // Mask out the bits we are about to insert from the old value, and or 886 // in the new bits. 887 if (SrcWidth != DestWidth) { 888 assert(DestWidth > SrcWidth); 889 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 890 SV = Builder.CreateOr(Old, SV, "ins"); 891 } 892 return SV; 893 } 894 895 896 //===----------------------------------------------------------------------===// 897 // SRoA Driver 898 //===----------------------------------------------------------------------===// 899 900 901 bool SROA::runOnFunction(Function &F) { 902 TD = getAnalysisIfAvailable<TargetData>(); 903 904 bool Changed = performPromotion(F); 905 906 // FIXME: ScalarRepl currently depends on TargetData more than it 907 // theoretically needs to. It should be refactored in order to support 908 // target-independent IR. Until this is done, just skip the actual 909 // scalar-replacement portion of this pass. 910 if (!TD) return Changed; 911 912 while (1) { 913 bool LocalChange = performScalarRepl(F); 914 if (!LocalChange) break; // No need to repromote if no scalarrepl 915 Changed = true; 916 LocalChange = performPromotion(F); 917 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 918 } 919 920 return Changed; 921 } 922 923 namespace { 924 class AllocaPromoter : public LoadAndStorePromoter { 925 AllocaInst *AI; 926 DIBuilder *DIB; 927 SmallVector<DbgDeclareInst *, 4> DDIs; 928 SmallVector<DbgValueInst *, 4> DVIs; 929 public: 930 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S, 931 DIBuilder *DB) 932 : LoadAndStorePromoter(Insts, S), AI(0), DIB(DB) {} 933 934 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) { 935 // Remember which alloca we're promoting (for isInstInList). 936 this->AI = AI; 937 if (MDNode *DebugNode = MDNode::getIfExists(AI->getContext(), AI)) 938 for (Value::use_iterator UI = DebugNode->use_begin(), 939 E = DebugNode->use_end(); UI != E; ++UI) 940 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI)) 941 DDIs.push_back(DDI); 942 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI)) 943 DVIs.push_back(DVI); 944 945 LoadAndStorePromoter::run(Insts); 946 AI->eraseFromParent(); 947 for (SmallVector<DbgDeclareInst *, 4>::iterator I = DDIs.begin(), 948 E = DDIs.end(); I != E; ++I) { 949 DbgDeclareInst *DDI = *I; 950 DDI->eraseFromParent(); 951 } 952 for (SmallVector<DbgValueInst *, 4>::iterator I = DVIs.begin(), 953 E = DVIs.end(); I != E; ++I) { 954 DbgValueInst *DVI = *I; 955 DVI->eraseFromParent(); 956 } 957 } 958 959 virtual bool isInstInList(Instruction *I, 960 const SmallVectorImpl<Instruction*> &Insts) const { 961 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 962 return LI->getOperand(0) == AI; 963 return cast<StoreInst>(I)->getPointerOperand() == AI; 964 } 965 966 virtual void updateDebugInfo(Instruction *Inst) const { 967 for (SmallVector<DbgDeclareInst *, 4>::const_iterator I = DDIs.begin(), 968 E = DDIs.end(); I != E; ++I) { 969 DbgDeclareInst *DDI = *I; 970 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) 971 ConvertDebugDeclareToDebugValue(DDI, SI, *DIB); 972 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) 973 ConvertDebugDeclareToDebugValue(DDI, LI, *DIB); 974 } 975 for (SmallVector<DbgValueInst *, 4>::const_iterator I = DVIs.begin(), 976 E = DVIs.end(); I != E; ++I) { 977 DbgValueInst *DVI = *I; 978 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 979 Instruction *DbgVal = NULL; 980 // If an argument is zero extended then use argument directly. The ZExt 981 // may be zapped by an optimization pass in future. 982 Argument *ExtendedArg = NULL; 983 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0))) 984 ExtendedArg = dyn_cast<Argument>(ZExt->getOperand(0)); 985 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0))) 986 ExtendedArg = dyn_cast<Argument>(SExt->getOperand(0)); 987 if (ExtendedArg) 988 DbgVal = DIB->insertDbgValueIntrinsic(ExtendedArg, 0, 989 DIVariable(DVI->getVariable()), 990 SI); 991 else 992 DbgVal = DIB->insertDbgValueIntrinsic(SI->getOperand(0), 0, 993 DIVariable(DVI->getVariable()), 994 SI); 995 DbgVal->setDebugLoc(DVI->getDebugLoc()); 996 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 997 Instruction *DbgVal = 998 DIB->insertDbgValueIntrinsic(LI->getOperand(0), 0, 999 DIVariable(DVI->getVariable()), LI); 1000 DbgVal->setDebugLoc(DVI->getDebugLoc()); 1001 } 1002 } 1003 } 1004 }; 1005 } // end anon namespace 1006 1007 /// isSafeSelectToSpeculate - Select instructions that use an alloca and are 1008 /// subsequently loaded can be rewritten to load both input pointers and then 1009 /// select between the result, allowing the load of the alloca to be promoted. 1010 /// From this: 1011 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other 1012 /// %V = load i32* %P2 1013 /// to: 1014 /// %V1 = load i32* %Alloca -> will be mem2reg'd 1015 /// %V2 = load i32* %Other 1016 /// %V = select i1 %cond, i32 %V1, i32 %V2 1017 /// 1018 /// We can do this to a select if its only uses are loads and if the operand to 1019 /// the select can be loaded unconditionally. 1020 static bool isSafeSelectToSpeculate(SelectInst *SI, const TargetData *TD) { 1021 bool TDerefable = SI->getTrueValue()->isDereferenceablePointer(); 1022 bool FDerefable = SI->getFalseValue()->isDereferenceablePointer(); 1023 1024 for (Value::use_iterator UI = SI->use_begin(), UE = SI->use_end(); 1025 UI != UE; ++UI) { 1026 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1027 if (LI == 0 || !LI->isSimple()) return false; 1028 1029 // Both operands to the select need to be dereferencable, either absolutely 1030 // (e.g. allocas) or at this point because we can see other accesses to it. 1031 if (!TDerefable && !isSafeToLoadUnconditionally(SI->getTrueValue(), LI, 1032 LI->getAlignment(), TD)) 1033 return false; 1034 if (!FDerefable && !isSafeToLoadUnconditionally(SI->getFalseValue(), LI, 1035 LI->getAlignment(), TD)) 1036 return false; 1037 } 1038 1039 return true; 1040 } 1041 1042 /// isSafePHIToSpeculate - PHI instructions that use an alloca and are 1043 /// subsequently loaded can be rewritten to load both input pointers in the pred 1044 /// blocks and then PHI the results, allowing the load of the alloca to be 1045 /// promoted. 1046 /// From this: 1047 /// %P2 = phi [i32* %Alloca, i32* %Other] 1048 /// %V = load i32* %P2 1049 /// to: 1050 /// %V1 = load i32* %Alloca -> will be mem2reg'd 1051 /// ... 1052 /// %V2 = load i32* %Other 1053 /// ... 1054 /// %V = phi [i32 %V1, i32 %V2] 1055 /// 1056 /// We can do this to a select if its only uses are loads and if the operand to 1057 /// the select can be loaded unconditionally. 1058 static bool isSafePHIToSpeculate(PHINode *PN, const TargetData *TD) { 1059 // For now, we can only do this promotion if the load is in the same block as 1060 // the PHI, and if there are no stores between the phi and load. 1061 // TODO: Allow recursive phi users. 1062 // TODO: Allow stores. 1063 BasicBlock *BB = PN->getParent(); 1064 unsigned MaxAlign = 0; 1065 for (Value::use_iterator UI = PN->use_begin(), UE = PN->use_end(); 1066 UI != UE; ++UI) { 1067 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1068 if (LI == 0 || !LI->isSimple()) return false; 1069 1070 // For now we only allow loads in the same block as the PHI. This is a 1071 // common case that happens when instcombine merges two loads through a PHI. 1072 if (LI->getParent() != BB) return false; 1073 1074 // Ensure that there are no instructions between the PHI and the load that 1075 // could store. 1076 for (BasicBlock::iterator BBI = PN; &*BBI != LI; ++BBI) 1077 if (BBI->mayWriteToMemory()) 1078 return false; 1079 1080 MaxAlign = std::max(MaxAlign, LI->getAlignment()); 1081 } 1082 1083 // Okay, we know that we have one or more loads in the same block as the PHI. 1084 // We can transform this if it is safe to push the loads into the predecessor 1085 // blocks. The only thing to watch out for is that we can't put a possibly 1086 // trapping load in the predecessor if it is a critical edge. 1087 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1088 BasicBlock *Pred = PN->getIncomingBlock(i); 1089 Value *InVal = PN->getIncomingValue(i); 1090 1091 // If the terminator of the predecessor has side-effects (an invoke), 1092 // there is no safe place to put a load in the predecessor. 1093 if (Pred->getTerminator()->mayHaveSideEffects()) 1094 return false; 1095 1096 // If the value is produced by the terminator of the predecessor 1097 // (an invoke), there is no valid place to put a load in the predecessor. 1098 if (Pred->getTerminator() == InVal) 1099 return false; 1100 1101 // If the predecessor has a single successor, then the edge isn't critical. 1102 if (Pred->getTerminator()->getNumSuccessors() == 1) 1103 continue; 1104 1105 // If this pointer is always safe to load, or if we can prove that there is 1106 // already a load in the block, then we can move the load to the pred block. 1107 if (InVal->isDereferenceablePointer() || 1108 isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign, TD)) 1109 continue; 1110 1111 return false; 1112 } 1113 1114 return true; 1115 } 1116 1117 1118 /// tryToMakeAllocaBePromotable - This returns true if the alloca only has 1119 /// direct (non-volatile) loads and stores to it. If the alloca is close but 1120 /// not quite there, this will transform the code to allow promotion. As such, 1121 /// it is a non-pure predicate. 1122 static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const TargetData *TD) { 1123 SetVector<Instruction*, SmallVector<Instruction*, 4>, 1124 SmallPtrSet<Instruction*, 4> > InstsToRewrite; 1125 1126 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end(); 1127 UI != UE; ++UI) { 1128 User *U = *UI; 1129 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 1130 if (!LI->isSimple()) 1131 return false; 1132 continue; 1133 } 1134 1135 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 1136 if (SI->getOperand(0) == AI || !SI->isSimple()) 1137 return false; // Don't allow a store OF the AI, only INTO the AI. 1138 continue; 1139 } 1140 1141 if (SelectInst *SI = dyn_cast<SelectInst>(U)) { 1142 // If the condition being selected on is a constant, fold the select, yes 1143 // this does (rarely) happen early on. 1144 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition())) { 1145 Value *Result = SI->getOperand(1+CI->isZero()); 1146 SI->replaceAllUsesWith(Result); 1147 SI->eraseFromParent(); 1148 1149 // This is very rare and we just scrambled the use list of AI, start 1150 // over completely. 1151 return tryToMakeAllocaBePromotable(AI, TD); 1152 } 1153 1154 // If it is safe to turn "load (select c, AI, ptr)" into a select of two 1155 // loads, then we can transform this by rewriting the select. 1156 if (!isSafeSelectToSpeculate(SI, TD)) 1157 return false; 1158 1159 InstsToRewrite.insert(SI); 1160 continue; 1161 } 1162 1163 if (PHINode *PN = dyn_cast<PHINode>(U)) { 1164 if (PN->use_empty()) { // Dead PHIs can be stripped. 1165 InstsToRewrite.insert(PN); 1166 continue; 1167 } 1168 1169 // If it is safe to turn "load (phi [AI, ptr, ...])" into a PHI of loads 1170 // in the pred blocks, then we can transform this by rewriting the PHI. 1171 if (!isSafePHIToSpeculate(PN, TD)) 1172 return false; 1173 1174 InstsToRewrite.insert(PN); 1175 continue; 1176 } 1177 1178 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 1179 if (onlyUsedByLifetimeMarkers(BCI)) { 1180 InstsToRewrite.insert(BCI); 1181 continue; 1182 } 1183 } 1184 1185 return false; 1186 } 1187 1188 // If there are no instructions to rewrite, then all uses are load/stores and 1189 // we're done! 1190 if (InstsToRewrite.empty()) 1191 return true; 1192 1193 // If we have instructions that need to be rewritten for this to be promotable 1194 // take care of it now. 1195 for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) { 1196 if (BitCastInst *BCI = dyn_cast<BitCastInst>(InstsToRewrite[i])) { 1197 // This could only be a bitcast used by nothing but lifetime intrinsics. 1198 for (BitCastInst::use_iterator I = BCI->use_begin(), E = BCI->use_end(); 1199 I != E;) { 1200 Use &U = I.getUse(); 1201 ++I; 1202 cast<Instruction>(U.getUser())->eraseFromParent(); 1203 } 1204 BCI->eraseFromParent(); 1205 continue; 1206 } 1207 1208 if (SelectInst *SI = dyn_cast<SelectInst>(InstsToRewrite[i])) { 1209 // Selects in InstsToRewrite only have load uses. Rewrite each as two 1210 // loads with a new select. 1211 while (!SI->use_empty()) { 1212 LoadInst *LI = cast<LoadInst>(SI->use_back()); 1213 1214 IRBuilder<> Builder(LI); 1215 LoadInst *TrueLoad = 1216 Builder.CreateLoad(SI->getTrueValue(), LI->getName()+".t"); 1217 LoadInst *FalseLoad = 1218 Builder.CreateLoad(SI->getFalseValue(), LI->getName()+".f"); 1219 1220 // Transfer alignment and TBAA info if present. 1221 TrueLoad->setAlignment(LI->getAlignment()); 1222 FalseLoad->setAlignment(LI->getAlignment()); 1223 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) { 1224 TrueLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1225 FalseLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1226 } 1227 1228 Value *V = Builder.CreateSelect(SI->getCondition(), TrueLoad, FalseLoad); 1229 V->takeName(LI); 1230 LI->replaceAllUsesWith(V); 1231 LI->eraseFromParent(); 1232 } 1233 1234 // Now that all the loads are gone, the select is gone too. 1235 SI->eraseFromParent(); 1236 continue; 1237 } 1238 1239 // Otherwise, we have a PHI node which allows us to push the loads into the 1240 // predecessors. 1241 PHINode *PN = cast<PHINode>(InstsToRewrite[i]); 1242 if (PN->use_empty()) { 1243 PN->eraseFromParent(); 1244 continue; 1245 } 1246 1247 Type *LoadTy = cast<PointerType>(PN->getType())->getElementType(); 1248 PHINode *NewPN = PHINode::Create(LoadTy, PN->getNumIncomingValues(), 1249 PN->getName()+".ld", PN); 1250 1251 // Get the TBAA tag and alignment to use from one of the loads. It doesn't 1252 // matter which one we get and if any differ, it doesn't matter. 1253 LoadInst *SomeLoad = cast<LoadInst>(PN->use_back()); 1254 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa); 1255 unsigned Align = SomeLoad->getAlignment(); 1256 1257 // Rewrite all loads of the PN to use the new PHI. 1258 while (!PN->use_empty()) { 1259 LoadInst *LI = cast<LoadInst>(PN->use_back()); 1260 LI->replaceAllUsesWith(NewPN); 1261 LI->eraseFromParent(); 1262 } 1263 1264 // Inject loads into all of the pred blocks. Keep track of which blocks we 1265 // insert them into in case we have multiple edges from the same block. 1266 DenseMap<BasicBlock*, LoadInst*> InsertedLoads; 1267 1268 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1269 BasicBlock *Pred = PN->getIncomingBlock(i); 1270 LoadInst *&Load = InsertedLoads[Pred]; 1271 if (Load == 0) { 1272 Load = new LoadInst(PN->getIncomingValue(i), 1273 PN->getName() + "." + Pred->getName(), 1274 Pred->getTerminator()); 1275 Load->setAlignment(Align); 1276 if (TBAATag) Load->setMetadata(LLVMContext::MD_tbaa, TBAATag); 1277 } 1278 1279 NewPN->addIncoming(Load, Pred); 1280 } 1281 1282 PN->eraseFromParent(); 1283 } 1284 1285 ++NumAdjusted; 1286 return true; 1287 } 1288 1289 bool SROA::performPromotion(Function &F) { 1290 std::vector<AllocaInst*> Allocas; 1291 DominatorTree *DT = 0; 1292 if (HasDomTree) 1293 DT = &getAnalysis<DominatorTree>(); 1294 1295 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 1296 DIBuilder DIB(*F.getParent()); 1297 bool Changed = false; 1298 SmallVector<Instruction*, 64> Insts; 1299 while (1) { 1300 Allocas.clear(); 1301 1302 // Find allocas that are safe to promote, by looking at all instructions in 1303 // the entry node 1304 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 1305 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 1306 if (tryToMakeAllocaBePromotable(AI, TD)) 1307 Allocas.push_back(AI); 1308 1309 if (Allocas.empty()) break; 1310 1311 if (HasDomTree) 1312 PromoteMemToReg(Allocas, *DT); 1313 else { 1314 SSAUpdater SSA; 1315 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) { 1316 AllocaInst *AI = Allocas[i]; 1317 1318 // Build list of instructions to promote. 1319 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 1320 UI != E; ++UI) 1321 Insts.push_back(cast<Instruction>(*UI)); 1322 AllocaPromoter(Insts, SSA, &DIB).run(AI, Insts); 1323 Insts.clear(); 1324 } 1325 } 1326 NumPromoted += Allocas.size(); 1327 Changed = true; 1328 } 1329 1330 return Changed; 1331 } 1332 1333 1334 /// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 1335 /// SROA. It must be a struct or array type with a small number of elements. 1336 static bool ShouldAttemptScalarRepl(AllocaInst *AI) { 1337 Type *T = AI->getAllocatedType(); 1338 // Do not promote any struct into more than 32 separate vars. 1339 if (StructType *ST = dyn_cast<StructType>(T)) 1340 return ST->getNumElements() <= 32; 1341 // Arrays are much less likely to be safe for SROA; only consider 1342 // them if they are very small. 1343 if (ArrayType *AT = dyn_cast<ArrayType>(T)) 1344 return AT->getNumElements() <= 8; 1345 return false; 1346 } 1347 1348 1349 // performScalarRepl - This algorithm is a simple worklist driven algorithm, 1350 // which runs on all of the alloca instructions in the function, removing them 1351 // if they are only used by getelementptr instructions. 1352 // 1353 bool SROA::performScalarRepl(Function &F) { 1354 std::vector<AllocaInst*> WorkList; 1355 1356 // Scan the entry basic block, adding allocas to the worklist. 1357 BasicBlock &BB = F.getEntryBlock(); 1358 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 1359 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 1360 WorkList.push_back(A); 1361 1362 // Process the worklist 1363 bool Changed = false; 1364 while (!WorkList.empty()) { 1365 AllocaInst *AI = WorkList.back(); 1366 WorkList.pop_back(); 1367 1368 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 1369 // with unused elements. 1370 if (AI->use_empty()) { 1371 AI->eraseFromParent(); 1372 Changed = true; 1373 continue; 1374 } 1375 1376 // If this alloca is impossible for us to promote, reject it early. 1377 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 1378 continue; 1379 1380 // Check to see if this allocation is only modified by a memcpy/memmove from 1381 // a constant global. If this is the case, we can change all users to use 1382 // the constant global instead. This is commonly produced by the CFE by 1383 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 1384 // is only subsequently read. 1385 SmallVector<Instruction *, 4> ToDelete; 1386 if (MemTransferInst *Copy = isOnlyCopiedFromConstantGlobal(AI, ToDelete)) { 1387 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); 1388 DEBUG(dbgs() << " memcpy = " << *Copy << '\n'); 1389 for (unsigned i = 0, e = ToDelete.size(); i != e; ++i) 1390 ToDelete[i]->eraseFromParent(); 1391 Constant *TheSrc = cast<Constant>(Copy->getSource()); 1392 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 1393 Copy->eraseFromParent(); // Don't mutate the global. 1394 AI->eraseFromParent(); 1395 ++NumGlobals; 1396 Changed = true; 1397 continue; 1398 } 1399 1400 // Check to see if we can perform the core SROA transformation. We cannot 1401 // transform the allocation instruction if it is an array allocation 1402 // (allocations OF arrays are ok though), and an allocation of a scalar 1403 // value cannot be decomposed at all. 1404 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 1405 1406 // Do not promote [0 x %struct]. 1407 if (AllocaSize == 0) continue; 1408 1409 // Do not promote any struct whose size is too big. 1410 if (AllocaSize > SRThreshold) continue; 1411 1412 // If the alloca looks like a good candidate for scalar replacement, and if 1413 // all its users can be transformed, then split up the aggregate into its 1414 // separate elements. 1415 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 1416 DoScalarReplacement(AI, WorkList); 1417 Changed = true; 1418 continue; 1419 } 1420 1421 // If we can turn this aggregate value (potentially with casts) into a 1422 // simple scalar value that can be mem2reg'd into a register value. 1423 // IsNotTrivial tracks whether this is something that mem2reg could have 1424 // promoted itself. If so, we don't want to transform it needlessly. Note 1425 // that we can't just check based on the type: the alloca may be of an i32 1426 // but that has pointer arithmetic to set byte 3 of it or something. 1427 if (AllocaInst *NewAI = 1428 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) { 1429 NewAI->takeName(AI); 1430 AI->eraseFromParent(); 1431 ++NumConverted; 1432 Changed = true; 1433 continue; 1434 } 1435 1436 // Otherwise, couldn't process this alloca. 1437 } 1438 1439 return Changed; 1440 } 1441 1442 /// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 1443 /// predicate, do SROA now. 1444 void SROA::DoScalarReplacement(AllocaInst *AI, 1445 std::vector<AllocaInst*> &WorkList) { 1446 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 1447 SmallVector<AllocaInst*, 32> ElementAllocas; 1448 if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 1449 ElementAllocas.reserve(ST->getNumContainedTypes()); 1450 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 1451 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 1452 AI->getAlignment(), 1453 AI->getName() + "." + Twine(i), AI); 1454 ElementAllocas.push_back(NA); 1455 WorkList.push_back(NA); // Add to worklist for recursive processing 1456 } 1457 } else { 1458 ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 1459 ElementAllocas.reserve(AT->getNumElements()); 1460 Type *ElTy = AT->getElementType(); 1461 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1462 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 1463 AI->getName() + "." + Twine(i), AI); 1464 ElementAllocas.push_back(NA); 1465 WorkList.push_back(NA); // Add to worklist for recursive processing 1466 } 1467 } 1468 1469 // Now that we have created the new alloca instructions, rewrite all the 1470 // uses of the old alloca. 1471 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 1472 1473 // Now erase any instructions that were made dead while rewriting the alloca. 1474 DeleteDeadInstructions(); 1475 AI->eraseFromParent(); 1476 1477 ++NumReplaced; 1478 } 1479 1480 /// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 1481 /// recursively including all their operands that become trivially dead. 1482 void SROA::DeleteDeadInstructions() { 1483 while (!DeadInsts.empty()) { 1484 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 1485 1486 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 1487 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 1488 // Zero out the operand and see if it becomes trivially dead. 1489 // (But, don't add allocas to the dead instruction list -- they are 1490 // already on the worklist and will be deleted separately.) 1491 *OI = 0; 1492 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 1493 DeadInsts.push_back(U); 1494 } 1495 1496 I->eraseFromParent(); 1497 } 1498 } 1499 1500 /// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 1501 /// performing scalar replacement of alloca AI. The results are flagged in 1502 /// the Info parameter. Offset indicates the position within AI that is 1503 /// referenced by this instruction. 1504 void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset, 1505 AllocaInfo &Info) { 1506 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1507 Instruction *User = cast<Instruction>(*UI); 1508 1509 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1510 isSafeForScalarRepl(BC, Offset, Info); 1511 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1512 uint64_t GEPOffset = Offset; 1513 isSafeGEP(GEPI, GEPOffset, Info); 1514 if (!Info.isUnsafe) 1515 isSafeForScalarRepl(GEPI, GEPOffset, Info); 1516 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1517 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1518 if (Length == 0) 1519 return MarkUnsafe(Info, User); 1520 isSafeMemAccess(Offset, Length->getZExtValue(), 0, 1521 UI.getOperandNo() == 0, Info, MI, 1522 true /*AllowWholeAccess*/); 1523 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1524 if (!LI->isSimple()) 1525 return MarkUnsafe(Info, User); 1526 Type *LIType = LI->getType(); 1527 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1528 LIType, false, Info, LI, true /*AllowWholeAccess*/); 1529 Info.hasALoadOrStore = true; 1530 1531 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1532 // Store is ok if storing INTO the pointer, not storing the pointer 1533 if (!SI->isSimple() || SI->getOperand(0) == I) 1534 return MarkUnsafe(Info, User); 1535 1536 Type *SIType = SI->getOperand(0)->getType(); 1537 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1538 SIType, true, Info, SI, true /*AllowWholeAccess*/); 1539 Info.hasALoadOrStore = true; 1540 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 1541 if (II->getIntrinsicID() != Intrinsic::lifetime_start && 1542 II->getIntrinsicID() != Intrinsic::lifetime_end) 1543 return MarkUnsafe(Info, User); 1544 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1545 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1546 } else { 1547 return MarkUnsafe(Info, User); 1548 } 1549 if (Info.isUnsafe) return; 1550 } 1551 } 1552 1553 1554 /// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer 1555 /// derived from the alloca, we can often still split the alloca into elements. 1556 /// This is useful if we have a large alloca where one element is phi'd 1557 /// together somewhere: we can SRoA and promote all the other elements even if 1558 /// we end up not being able to promote this one. 1559 /// 1560 /// All we require is that the uses of the PHI do not index into other parts of 1561 /// the alloca. The most important use case for this is single load and stores 1562 /// that are PHI'd together, which can happen due to code sinking. 1563 void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset, 1564 AllocaInfo &Info) { 1565 // If we've already checked this PHI, don't do it again. 1566 if (PHINode *PN = dyn_cast<PHINode>(I)) 1567 if (!Info.CheckedPHIs.insert(PN)) 1568 return; 1569 1570 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1571 Instruction *User = cast<Instruction>(*UI); 1572 1573 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1574 isSafePHISelectUseForScalarRepl(BC, Offset, Info); 1575 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1576 // Only allow "bitcast" GEPs for simplicity. We could generalize this, 1577 // but would have to prove that we're staying inside of an element being 1578 // promoted. 1579 if (!GEPI->hasAllZeroIndices()) 1580 return MarkUnsafe(Info, User); 1581 isSafePHISelectUseForScalarRepl(GEPI, Offset, Info); 1582 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1583 if (!LI->isSimple()) 1584 return MarkUnsafe(Info, User); 1585 Type *LIType = LI->getType(); 1586 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1587 LIType, false, Info, LI, false /*AllowWholeAccess*/); 1588 Info.hasALoadOrStore = true; 1589 1590 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1591 // Store is ok if storing INTO the pointer, not storing the pointer 1592 if (!SI->isSimple() || SI->getOperand(0) == I) 1593 return MarkUnsafe(Info, User); 1594 1595 Type *SIType = SI->getOperand(0)->getType(); 1596 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1597 SIType, true, Info, SI, false /*AllowWholeAccess*/); 1598 Info.hasALoadOrStore = true; 1599 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1600 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1601 } else { 1602 return MarkUnsafe(Info, User); 1603 } 1604 if (Info.isUnsafe) return; 1605 } 1606 } 1607 1608 /// isSafeGEP - Check if a GEP instruction can be handled for scalar 1609 /// replacement. It is safe when all the indices are constant, in-bounds 1610 /// references, and when the resulting offset corresponds to an element within 1611 /// the alloca type. The results are flagged in the Info parameter. Upon 1612 /// return, Offset is adjusted as specified by the GEP indices. 1613 void SROA::isSafeGEP(GetElementPtrInst *GEPI, 1614 uint64_t &Offset, AllocaInfo &Info) { 1615 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 1616 if (GEPIt == E) 1617 return; 1618 1619 // Walk through the GEP type indices, checking the types that this indexes 1620 // into. 1621 for (; GEPIt != E; ++GEPIt) { 1622 // Ignore struct elements, no extra checking needed for these. 1623 if ((*GEPIt)->isStructTy()) 1624 continue; 1625 1626 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 1627 if (!IdxVal) 1628 return MarkUnsafe(Info, GEPI); 1629 } 1630 1631 // Compute the offset due to this GEP and check if the alloca has a 1632 // component element at that offset. 1633 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1634 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices); 1635 if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 0)) 1636 MarkUnsafe(Info, GEPI); 1637 } 1638 1639 /// isHomogeneousAggregate - Check if type T is a struct or array containing 1640 /// elements of the same type (which is always true for arrays). If so, 1641 /// return true with NumElts and EltTy set to the number of elements and the 1642 /// element type, respectively. 1643 static bool isHomogeneousAggregate(Type *T, unsigned &NumElts, 1644 Type *&EltTy) { 1645 if (ArrayType *AT = dyn_cast<ArrayType>(T)) { 1646 NumElts = AT->getNumElements(); 1647 EltTy = (NumElts == 0 ? 0 : AT->getElementType()); 1648 return true; 1649 } 1650 if (StructType *ST = dyn_cast<StructType>(T)) { 1651 NumElts = ST->getNumContainedTypes(); 1652 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0)); 1653 for (unsigned n = 1; n < NumElts; ++n) { 1654 if (ST->getContainedType(n) != EltTy) 1655 return false; 1656 } 1657 return true; 1658 } 1659 return false; 1660 } 1661 1662 /// isCompatibleAggregate - Check if T1 and T2 are either the same type or are 1663 /// "homogeneous" aggregates with the same element type and number of elements. 1664 static bool isCompatibleAggregate(Type *T1, Type *T2) { 1665 if (T1 == T2) 1666 return true; 1667 1668 unsigned NumElts1, NumElts2; 1669 Type *EltTy1, *EltTy2; 1670 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) && 1671 isHomogeneousAggregate(T2, NumElts2, EltTy2) && 1672 NumElts1 == NumElts2 && 1673 EltTy1 == EltTy2) 1674 return true; 1675 1676 return false; 1677 } 1678 1679 /// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1680 /// alloca or has an offset and size that corresponds to a component element 1681 /// within it. The offset checked here may have been formed from a GEP with a 1682 /// pointer bitcasted to a different type. 1683 /// 1684 /// If AllowWholeAccess is true, then this allows uses of the entire alloca as a 1685 /// unit. If false, it only allows accesses known to be in a single element. 1686 void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 1687 Type *MemOpType, bool isStore, 1688 AllocaInfo &Info, Instruction *TheAccess, 1689 bool AllowWholeAccess) { 1690 // Check if this is a load/store of the entire alloca. 1691 if (Offset == 0 && AllowWholeAccess && 1692 MemSize == TD->getTypeAllocSize(Info.AI->getAllocatedType())) { 1693 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer 1694 // loads/stores (which are essentially the same as the MemIntrinsics with 1695 // regard to copying padding between elements). But, if an alloca is 1696 // flagged as both a source and destination of such operations, we'll need 1697 // to check later for padding between elements. 1698 if (!MemOpType || MemOpType->isIntegerTy()) { 1699 if (isStore) 1700 Info.isMemCpyDst = true; 1701 else 1702 Info.isMemCpySrc = true; 1703 return; 1704 } 1705 // This is also safe for references using a type that is compatible with 1706 // the type of the alloca, so that loads/stores can be rewritten using 1707 // insertvalue/extractvalue. 1708 if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) { 1709 Info.hasSubelementAccess = true; 1710 return; 1711 } 1712 } 1713 // Check if the offset/size correspond to a component within the alloca type. 1714 Type *T = Info.AI->getAllocatedType(); 1715 if (TypeHasComponent(T, Offset, MemSize)) { 1716 Info.hasSubelementAccess = true; 1717 return; 1718 } 1719 1720 return MarkUnsafe(Info, TheAccess); 1721 } 1722 1723 /// TypeHasComponent - Return true if T has a component type with the 1724 /// specified offset and size. If Size is zero, do not check the size. 1725 bool SROA::TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size) { 1726 Type *EltTy; 1727 uint64_t EltSize; 1728 if (StructType *ST = dyn_cast<StructType>(T)) { 1729 const StructLayout *Layout = TD->getStructLayout(ST); 1730 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1731 EltTy = ST->getContainedType(EltIdx); 1732 EltSize = TD->getTypeAllocSize(EltTy); 1733 Offset -= Layout->getElementOffset(EltIdx); 1734 } else if (ArrayType *AT = dyn_cast<ArrayType>(T)) { 1735 EltTy = AT->getElementType(); 1736 EltSize = TD->getTypeAllocSize(EltTy); 1737 if (Offset >= AT->getNumElements() * EltSize) 1738 return false; 1739 Offset %= EltSize; 1740 } else { 1741 return false; 1742 } 1743 if (Offset == 0 && (Size == 0 || EltSize == Size)) 1744 return true; 1745 // Check if the component spans multiple elements. 1746 if (Offset + Size > EltSize) 1747 return false; 1748 return TypeHasComponent(EltTy, Offset, Size); 1749 } 1750 1751 /// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1752 /// the instruction I, which references it, to use the separate elements. 1753 /// Offset indicates the position within AI that is referenced by this 1754 /// instruction. 1755 void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1756 SmallVector<AllocaInst*, 32> &NewElts) { 1757 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) { 1758 Use &TheUse = UI.getUse(); 1759 Instruction *User = cast<Instruction>(*UI++); 1760 1761 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1762 RewriteBitCast(BC, AI, Offset, NewElts); 1763 continue; 1764 } 1765 1766 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1767 RewriteGEP(GEPI, AI, Offset, NewElts); 1768 continue; 1769 } 1770 1771 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1772 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1773 uint64_t MemSize = Length->getZExtValue(); 1774 if (Offset == 0 && 1775 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 1776 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1777 // Otherwise the intrinsic can only touch a single element and the 1778 // address operand will be updated, so nothing else needs to be done. 1779 continue; 1780 } 1781 1782 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 1783 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 1784 II->getIntrinsicID() == Intrinsic::lifetime_end) { 1785 RewriteLifetimeIntrinsic(II, AI, Offset, NewElts); 1786 } 1787 continue; 1788 } 1789 1790 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1791 Type *LIType = LI->getType(); 1792 1793 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) { 1794 // Replace: 1795 // %res = load { i32, i32 }* %alloc 1796 // with: 1797 // %load.0 = load i32* %alloc.0 1798 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1799 // %load.1 = load i32* %alloc.1 1800 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1801 // (Also works for arrays instead of structs) 1802 Value *Insert = UndefValue::get(LIType); 1803 IRBuilder<> Builder(LI); 1804 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1805 Value *Load = Builder.CreateLoad(NewElts[i], "load"); 1806 Insert = Builder.CreateInsertValue(Insert, Load, i, "insert"); 1807 } 1808 LI->replaceAllUsesWith(Insert); 1809 DeadInsts.push_back(LI); 1810 } else if (LIType->isIntegerTy() && 1811 TD->getTypeAllocSize(LIType) == 1812 TD->getTypeAllocSize(AI->getAllocatedType())) { 1813 // If this is a load of the entire alloca to an integer, rewrite it. 1814 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1815 } 1816 continue; 1817 } 1818 1819 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1820 Value *Val = SI->getOperand(0); 1821 Type *SIType = Val->getType(); 1822 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) { 1823 // Replace: 1824 // store { i32, i32 } %val, { i32, i32 }* %alloc 1825 // with: 1826 // %val.0 = extractvalue { i32, i32 } %val, 0 1827 // store i32 %val.0, i32* %alloc.0 1828 // %val.1 = extractvalue { i32, i32 } %val, 1 1829 // store i32 %val.1, i32* %alloc.1 1830 // (Also works for arrays instead of structs) 1831 IRBuilder<> Builder(SI); 1832 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1833 Value *Extract = Builder.CreateExtractValue(Val, i, Val->getName()); 1834 Builder.CreateStore(Extract, NewElts[i]); 1835 } 1836 DeadInsts.push_back(SI); 1837 } else if (SIType->isIntegerTy() && 1838 TD->getTypeAllocSize(SIType) == 1839 TD->getTypeAllocSize(AI->getAllocatedType())) { 1840 // If this is a store of the entire alloca from an integer, rewrite it. 1841 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 1842 } 1843 continue; 1844 } 1845 1846 if (isa<SelectInst>(User) || isa<PHINode>(User)) { 1847 // If we have a PHI user of the alloca itself (as opposed to a GEP or 1848 // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to 1849 // the new pointer. 1850 if (!isa<AllocaInst>(I)) continue; 1851 1852 assert(Offset == 0 && NewElts[0] && 1853 "Direct alloca use should have a zero offset"); 1854 1855 // If we have a use of the alloca, we know the derived uses will be 1856 // utilizing just the first element of the scalarized result. Insert a 1857 // bitcast of the first alloca before the user as required. 1858 AllocaInst *NewAI = NewElts[0]; 1859 BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI); 1860 NewAI->moveBefore(BCI); 1861 TheUse = BCI; 1862 continue; 1863 } 1864 } 1865 } 1866 1867 /// RewriteBitCast - Update a bitcast reference to the alloca being replaced 1868 /// and recursively continue updating all of its uses. 1869 void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 1870 SmallVector<AllocaInst*, 32> &NewElts) { 1871 RewriteForScalarRepl(BC, AI, Offset, NewElts); 1872 if (BC->getOperand(0) != AI) 1873 return; 1874 1875 // The bitcast references the original alloca. Replace its uses with 1876 // references to the first new element alloca. 1877 Instruction *Val = NewElts[0]; 1878 if (Val->getType() != BC->getDestTy()) { 1879 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 1880 Val->takeName(BC); 1881 } 1882 BC->replaceAllUsesWith(Val); 1883 DeadInsts.push_back(BC); 1884 } 1885 1886 /// FindElementAndOffset - Return the index of the element containing Offset 1887 /// within the specified type, which must be either a struct or an array. 1888 /// Sets T to the type of the element and Offset to the offset within that 1889 /// element. IdxTy is set to the type of the index result to be used in a 1890 /// GEP instruction. 1891 uint64_t SROA::FindElementAndOffset(Type *&T, uint64_t &Offset, 1892 Type *&IdxTy) { 1893 uint64_t Idx = 0; 1894 if (StructType *ST = dyn_cast<StructType>(T)) { 1895 const StructLayout *Layout = TD->getStructLayout(ST); 1896 Idx = Layout->getElementContainingOffset(Offset); 1897 T = ST->getContainedType(Idx); 1898 Offset -= Layout->getElementOffset(Idx); 1899 IdxTy = Type::getInt32Ty(T->getContext()); 1900 return Idx; 1901 } 1902 ArrayType *AT = cast<ArrayType>(T); 1903 T = AT->getElementType(); 1904 uint64_t EltSize = TD->getTypeAllocSize(T); 1905 Idx = Offset / EltSize; 1906 Offset -= Idx * EltSize; 1907 IdxTy = Type::getInt64Ty(T->getContext()); 1908 return Idx; 1909 } 1910 1911 /// RewriteGEP - Check if this GEP instruction moves the pointer across 1912 /// elements of the alloca that are being split apart, and if so, rewrite 1913 /// the GEP to be relative to the new element. 1914 void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 1915 SmallVector<AllocaInst*, 32> &NewElts) { 1916 uint64_t OldOffset = Offset; 1917 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1918 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices); 1919 1920 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 1921 1922 Type *T = AI->getAllocatedType(); 1923 Type *IdxTy; 1924 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 1925 if (GEPI->getOperand(0) == AI) 1926 OldIdx = ~0ULL; // Force the GEP to be rewritten. 1927 1928 T = AI->getAllocatedType(); 1929 uint64_t EltOffset = Offset; 1930 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1931 1932 // If this GEP does not move the pointer across elements of the alloca 1933 // being split, then it does not needs to be rewritten. 1934 if (Idx == OldIdx) 1935 return; 1936 1937 Type *i32Ty = Type::getInt32Ty(AI->getContext()); 1938 SmallVector<Value*, 8> NewArgs; 1939 NewArgs.push_back(Constant::getNullValue(i32Ty)); 1940 while (EltOffset != 0) { 1941 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 1942 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 1943 } 1944 Instruction *Val = NewElts[Idx]; 1945 if (NewArgs.size() > 1) { 1946 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs, "", GEPI); 1947 Val->takeName(GEPI); 1948 } 1949 if (Val->getType() != GEPI->getType()) 1950 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 1951 GEPI->replaceAllUsesWith(Val); 1952 DeadInsts.push_back(GEPI); 1953 } 1954 1955 /// RewriteLifetimeIntrinsic - II is a lifetime.start/lifetime.end. Rewrite it 1956 /// to mark the lifetime of the scalarized memory. 1957 void SROA::RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI, 1958 uint64_t Offset, 1959 SmallVector<AllocaInst*, 32> &NewElts) { 1960 ConstantInt *OldSize = cast<ConstantInt>(II->getArgOperand(0)); 1961 // Put matching lifetime markers on everything from Offset up to 1962 // Offset+OldSize. 1963 Type *AIType = AI->getAllocatedType(); 1964 uint64_t NewOffset = Offset; 1965 Type *IdxTy; 1966 uint64_t Idx = FindElementAndOffset(AIType, NewOffset, IdxTy); 1967 1968 IRBuilder<> Builder(II); 1969 uint64_t Size = OldSize->getLimitedValue(); 1970 1971 if (NewOffset) { 1972 // Splice the first element and index 'NewOffset' bytes in. SROA will 1973 // split the alloca again later. 1974 Value *V = Builder.CreateBitCast(NewElts[Idx], Builder.getInt8PtrTy()); 1975 V = Builder.CreateGEP(V, Builder.getInt64(NewOffset)); 1976 1977 IdxTy = NewElts[Idx]->getAllocatedType(); 1978 uint64_t EltSize = TD->getTypeAllocSize(IdxTy) - NewOffset; 1979 if (EltSize > Size) { 1980 EltSize = Size; 1981 Size = 0; 1982 } else { 1983 Size -= EltSize; 1984 } 1985 if (II->getIntrinsicID() == Intrinsic::lifetime_start) 1986 Builder.CreateLifetimeStart(V, Builder.getInt64(EltSize)); 1987 else 1988 Builder.CreateLifetimeEnd(V, Builder.getInt64(EltSize)); 1989 ++Idx; 1990 } 1991 1992 for (; Idx != NewElts.size() && Size; ++Idx) { 1993 IdxTy = NewElts[Idx]->getAllocatedType(); 1994 uint64_t EltSize = TD->getTypeAllocSize(IdxTy); 1995 if (EltSize > Size) { 1996 EltSize = Size; 1997 Size = 0; 1998 } else { 1999 Size -= EltSize; 2000 } 2001 if (II->getIntrinsicID() == Intrinsic::lifetime_start) 2002 Builder.CreateLifetimeStart(NewElts[Idx], 2003 Builder.getInt64(EltSize)); 2004 else 2005 Builder.CreateLifetimeEnd(NewElts[Idx], 2006 Builder.getInt64(EltSize)); 2007 } 2008 DeadInsts.push_back(II); 2009 } 2010 2011 /// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 2012 /// Rewrite it to copy or set the elements of the scalarized memory. 2013 void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 2014 AllocaInst *AI, 2015 SmallVector<AllocaInst*, 32> &NewElts) { 2016 // If this is a memcpy/memmove, construct the other pointer as the 2017 // appropriate type. The "Other" pointer is the pointer that goes to memory 2018 // that doesn't have anything to do with the alloca that we are promoting. For 2019 // memset, this Value* stays null. 2020 Value *OtherPtr = 0; 2021 unsigned MemAlignment = MI->getAlignment(); 2022 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 2023 if (Inst == MTI->getRawDest()) 2024 OtherPtr = MTI->getRawSource(); 2025 else { 2026 assert(Inst == MTI->getRawSource()); 2027 OtherPtr = MTI->getRawDest(); 2028 } 2029 } 2030 2031 // If there is an other pointer, we want to convert it to the same pointer 2032 // type as AI has, so we can GEP through it safely. 2033 if (OtherPtr) { 2034 unsigned AddrSpace = 2035 cast<PointerType>(OtherPtr->getType())->getAddressSpace(); 2036 2037 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 2038 // optimization, but it's also required to detect the corner case where 2039 // both pointer operands are referencing the same memory, and where 2040 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 2041 // function is only called for mem intrinsics that access the whole 2042 // aggregate, so non-zero GEPs are not an issue here.) 2043 OtherPtr = OtherPtr->stripPointerCasts(); 2044 2045 // Copying the alloca to itself is a no-op: just delete it. 2046 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 2047 // This code will run twice for a no-op memcpy -- once for each operand. 2048 // Put only one reference to MI on the DeadInsts list. 2049 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), 2050 E = DeadInsts.end(); I != E; ++I) 2051 if (*I == MI) return; 2052 DeadInsts.push_back(MI); 2053 return; 2054 } 2055 2056 // If the pointer is not the right type, insert a bitcast to the right 2057 // type. 2058 Type *NewTy = 2059 PointerType::get(AI->getType()->getElementType(), AddrSpace); 2060 2061 if (OtherPtr->getType() != NewTy) 2062 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI); 2063 } 2064 2065 // Process each element of the aggregate. 2066 bool SROADest = MI->getRawDest() == Inst; 2067 2068 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 2069 2070 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2071 // If this is a memcpy/memmove, emit a GEP of the other element address. 2072 Value *OtherElt = 0; 2073 unsigned OtherEltAlign = MemAlignment; 2074 2075 if (OtherPtr) { 2076 Value *Idx[2] = { Zero, 2077 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 2078 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, 2079 OtherPtr->getName()+"."+Twine(i), 2080 MI); 2081 uint64_t EltOffset; 2082 PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 2083 Type *OtherTy = OtherPtrTy->getElementType(); 2084 if (StructType *ST = dyn_cast<StructType>(OtherTy)) { 2085 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 2086 } else { 2087 Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); 2088 EltOffset = TD->getTypeAllocSize(EltTy)*i; 2089 } 2090 2091 // The alignment of the other pointer is the guaranteed alignment of the 2092 // element, which is affected by both the known alignment of the whole 2093 // mem intrinsic and the alignment of the element. If the alignment of 2094 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 2095 // known alignment is just 4 bytes. 2096 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 2097 } 2098 2099 Value *EltPtr = NewElts[i]; 2100 Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 2101 2102 // If we got down to a scalar, insert a load or store as appropriate. 2103 if (EltTy->isSingleValueType()) { 2104 if (isa<MemTransferInst>(MI)) { 2105 if (SROADest) { 2106 // From Other to Alloca. 2107 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 2108 new StoreInst(Elt, EltPtr, MI); 2109 } else { 2110 // From Alloca to Other. 2111 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 2112 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 2113 } 2114 continue; 2115 } 2116 assert(isa<MemSetInst>(MI)); 2117 2118 // If the stored element is zero (common case), just store a null 2119 // constant. 2120 Constant *StoreVal; 2121 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) { 2122 if (CI->isZero()) { 2123 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 2124 } else { 2125 // If EltTy is a vector type, get the element type. 2126 Type *ValTy = EltTy->getScalarType(); 2127 2128 // Construct an integer with the right value. 2129 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 2130 APInt OneVal(EltSize, CI->getZExtValue()); 2131 APInt TotalVal(OneVal); 2132 // Set each byte. 2133 for (unsigned i = 0; 8*i < EltSize; ++i) { 2134 TotalVal = TotalVal.shl(8); 2135 TotalVal |= OneVal; 2136 } 2137 2138 // Convert the integer value to the appropriate type. 2139 StoreVal = ConstantInt::get(CI->getContext(), TotalVal); 2140 if (ValTy->isPointerTy()) 2141 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 2142 else if (ValTy->isFloatingPointTy()) 2143 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 2144 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 2145 2146 // If the requested value was a vector constant, create it. 2147 if (EltTy->isVectorTy()) { 2148 unsigned NumElts = cast<VectorType>(EltTy)->getNumElements(); 2149 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 2150 StoreVal = ConstantVector::get(Elts); 2151 } 2152 } 2153 new StoreInst(StoreVal, EltPtr, MI); 2154 continue; 2155 } 2156 // Otherwise, if we're storing a byte variable, use a memset call for 2157 // this element. 2158 } 2159 2160 unsigned EltSize = TD->getTypeAllocSize(EltTy); 2161 2162 IRBuilder<> Builder(MI); 2163 2164 // Finally, insert the meminst for this element. 2165 if (isa<MemSetInst>(MI)) { 2166 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize, 2167 MI->isVolatile()); 2168 } else { 2169 assert(isa<MemTransferInst>(MI)); 2170 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr 2171 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr 2172 2173 if (isa<MemCpyInst>(MI)) 2174 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile()); 2175 else 2176 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile()); 2177 } 2178 } 2179 DeadInsts.push_back(MI); 2180 } 2181 2182 /// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 2183 /// overwrites the entire allocation. Extract out the pieces of the stored 2184 /// integer and store them individually. 2185 void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 2186 SmallVector<AllocaInst*, 32> &NewElts){ 2187 // Extract each element out of the integer according to its structure offset 2188 // and store the element value to the individual alloca. 2189 Value *SrcVal = SI->getOperand(0); 2190 Type *AllocaEltTy = AI->getAllocatedType(); 2191 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2192 2193 IRBuilder<> Builder(SI); 2194 2195 // Handle tail padding by extending the operand 2196 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 2197 SrcVal = Builder.CreateZExt(SrcVal, 2198 IntegerType::get(SI->getContext(), AllocaSizeBits)); 2199 2200 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 2201 << '\n'); 2202 2203 // There are two forms here: AI could be an array or struct. Both cases 2204 // have different ways to compute the element offset. 2205 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2206 const StructLayout *Layout = TD->getStructLayout(EltSTy); 2207 2208 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2209 // Get the number of bits to shift SrcVal to get the value. 2210 Type *FieldTy = EltSTy->getElementType(i); 2211 uint64_t Shift = Layout->getElementOffsetInBits(i); 2212 2213 if (TD->isBigEndian()) 2214 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 2215 2216 Value *EltVal = SrcVal; 2217 if (Shift) { 2218 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2219 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2220 } 2221 2222 // Truncate down to an integer of the right size. 2223 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2224 2225 // Ignore zero sized fields like {}, they obviously contain no data. 2226 if (FieldSizeBits == 0) continue; 2227 2228 if (FieldSizeBits != AllocaSizeBits) 2229 EltVal = Builder.CreateTrunc(EltVal, 2230 IntegerType::get(SI->getContext(), FieldSizeBits)); 2231 Value *DestField = NewElts[i]; 2232 if (EltVal->getType() == FieldTy) { 2233 // Storing to an integer field of this size, just do it. 2234 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 2235 // Bitcast to the right element type (for fp/vector values). 2236 EltVal = Builder.CreateBitCast(EltVal, FieldTy); 2237 } else { 2238 // Otherwise, bitcast the dest pointer (for aggregates). 2239 DestField = Builder.CreateBitCast(DestField, 2240 PointerType::getUnqual(EltVal->getType())); 2241 } 2242 new StoreInst(EltVal, DestField, SI); 2243 } 2244 2245 } else { 2246 ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 2247 Type *ArrayEltTy = ATy->getElementType(); 2248 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2249 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 2250 2251 uint64_t Shift; 2252 2253 if (TD->isBigEndian()) 2254 Shift = AllocaSizeBits-ElementOffset; 2255 else 2256 Shift = 0; 2257 2258 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2259 // Ignore zero sized fields like {}, they obviously contain no data. 2260 if (ElementSizeBits == 0) continue; 2261 2262 Value *EltVal = SrcVal; 2263 if (Shift) { 2264 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2265 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2266 } 2267 2268 // Truncate down to an integer of the right size. 2269 if (ElementSizeBits != AllocaSizeBits) 2270 EltVal = Builder.CreateTrunc(EltVal, 2271 IntegerType::get(SI->getContext(), 2272 ElementSizeBits)); 2273 Value *DestField = NewElts[i]; 2274 if (EltVal->getType() == ArrayEltTy) { 2275 // Storing to an integer field of this size, just do it. 2276 } else if (ArrayEltTy->isFloatingPointTy() || 2277 ArrayEltTy->isVectorTy()) { 2278 // Bitcast to the right element type (for fp/vector values). 2279 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy); 2280 } else { 2281 // Otherwise, bitcast the dest pointer (for aggregates). 2282 DestField = Builder.CreateBitCast(DestField, 2283 PointerType::getUnqual(EltVal->getType())); 2284 } 2285 new StoreInst(EltVal, DestField, SI); 2286 2287 if (TD->isBigEndian()) 2288 Shift -= ElementOffset; 2289 else 2290 Shift += ElementOffset; 2291 } 2292 } 2293 2294 DeadInsts.push_back(SI); 2295 } 2296 2297 /// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 2298 /// an integer. Load the individual pieces to form the aggregate value. 2299 void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 2300 SmallVector<AllocaInst*, 32> &NewElts) { 2301 // Extract each element out of the NewElts according to its structure offset 2302 // and form the result value. 2303 Type *AllocaEltTy = AI->getAllocatedType(); 2304 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2305 2306 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 2307 << '\n'); 2308 2309 // There are two forms here: AI could be an array or struct. Both cases 2310 // have different ways to compute the element offset. 2311 const StructLayout *Layout = 0; 2312 uint64_t ArrayEltBitOffset = 0; 2313 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2314 Layout = TD->getStructLayout(EltSTy); 2315 } else { 2316 Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 2317 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2318 } 2319 2320 Value *ResultVal = 2321 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 2322 2323 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2324 // Load the value from the alloca. If the NewElt is an aggregate, cast 2325 // the pointer to an integer of the same size before doing the load. 2326 Value *SrcField = NewElts[i]; 2327 Type *FieldTy = 2328 cast<PointerType>(SrcField->getType())->getElementType(); 2329 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2330 2331 // Ignore zero sized fields like {}, they obviously contain no data. 2332 if (FieldSizeBits == 0) continue; 2333 2334 IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 2335 FieldSizeBits); 2336 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 2337 !FieldTy->isVectorTy()) 2338 SrcField = new BitCastInst(SrcField, 2339 PointerType::getUnqual(FieldIntTy), 2340 "", LI); 2341 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 2342 2343 // If SrcField is a fp or vector of the right size but that isn't an 2344 // integer type, bitcast to an integer so we can shift it. 2345 if (SrcField->getType() != FieldIntTy) 2346 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 2347 2348 // Zero extend the field to be the same size as the final alloca so that 2349 // we can shift and insert it. 2350 if (SrcField->getType() != ResultVal->getType()) 2351 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 2352 2353 // Determine the number of bits to shift SrcField. 2354 uint64_t Shift; 2355 if (Layout) // Struct case. 2356 Shift = Layout->getElementOffsetInBits(i); 2357 else // Array case. 2358 Shift = i*ArrayEltBitOffset; 2359 2360 if (TD->isBigEndian()) 2361 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 2362 2363 if (Shift) { 2364 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 2365 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 2366 } 2367 2368 // Don't create an 'or x, 0' on the first iteration. 2369 if (!isa<Constant>(ResultVal) || 2370 !cast<Constant>(ResultVal)->isNullValue()) 2371 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 2372 else 2373 ResultVal = SrcField; 2374 } 2375 2376 // Handle tail padding by truncating the result 2377 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 2378 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 2379 2380 LI->replaceAllUsesWith(ResultVal); 2381 DeadInsts.push_back(LI); 2382 } 2383 2384 /// HasPadding - Return true if the specified type has any structure or 2385 /// alignment padding in between the elements that would be split apart 2386 /// by SROA; return false otherwise. 2387 static bool HasPadding(Type *Ty, const TargetData &TD) { 2388 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 2389 Ty = ATy->getElementType(); 2390 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 2391 } 2392 2393 // SROA currently handles only Arrays and Structs. 2394 StructType *STy = cast<StructType>(Ty); 2395 const StructLayout *SL = TD.getStructLayout(STy); 2396 unsigned PrevFieldBitOffset = 0; 2397 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 2398 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 2399 2400 // Check to see if there is any padding between this element and the 2401 // previous one. 2402 if (i) { 2403 unsigned PrevFieldEnd = 2404 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 2405 if (PrevFieldEnd < FieldBitOffset) 2406 return true; 2407 } 2408 PrevFieldBitOffset = FieldBitOffset; 2409 } 2410 // Check for tail padding. 2411 if (unsigned EltCount = STy->getNumElements()) { 2412 unsigned PrevFieldEnd = PrevFieldBitOffset + 2413 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 2414 if (PrevFieldEnd < SL->getSizeInBits()) 2415 return true; 2416 } 2417 return false; 2418 } 2419 2420 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 2421 /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 2422 /// or 1 if safe after canonicalization has been performed. 2423 bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 2424 // Loop over the use list of the alloca. We can only transform it if all of 2425 // the users are safe to transform. 2426 AllocaInfo Info(AI); 2427 2428 isSafeForScalarRepl(AI, 0, Info); 2429 if (Info.isUnsafe) { 2430 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 2431 return false; 2432 } 2433 2434 // Okay, we know all the users are promotable. If the aggregate is a memcpy 2435 // source and destination, we have to be careful. In particular, the memcpy 2436 // could be moving around elements that live in structure padding of the LLVM 2437 // types, but may actually be used. In these cases, we refuse to promote the 2438 // struct. 2439 if (Info.isMemCpySrc && Info.isMemCpyDst && 2440 HasPadding(AI->getAllocatedType(), *TD)) 2441 return false; 2442 2443 // If the alloca never has an access to just *part* of it, but is accessed 2444 // via loads and stores, then we should use ConvertToScalarInfo to promote 2445 // the alloca instead of promoting each piece at a time and inserting fission 2446 // and fusion code. 2447 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) { 2448 // If the struct/array just has one element, use basic SRoA. 2449 if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 2450 if (ST->getNumElements() > 1) return false; 2451 } else { 2452 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1) 2453 return false; 2454 } 2455 } 2456 2457 return true; 2458 } 2459 2460 2461 2462 /// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 2463 /// some part of a constant global variable. This intentionally only accepts 2464 /// constant expressions because we don't can't rewrite arbitrary instructions. 2465 static bool PointsToConstantGlobal(Value *V) { 2466 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 2467 return GV->isConstant(); 2468 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 2469 if (CE->getOpcode() == Instruction::BitCast || 2470 CE->getOpcode() == Instruction::GetElementPtr) 2471 return PointsToConstantGlobal(CE->getOperand(0)); 2472 return false; 2473 } 2474 2475 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 2476 /// pointer to an alloca. Ignore any reads of the pointer, return false if we 2477 /// see any stores or other unknown uses. If we see pointer arithmetic, keep 2478 /// track of whether it moves the pointer (with isOffset) but otherwise traverse 2479 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 2480 /// the alloca, and if the source pointer is a pointer to a constant global, we 2481 /// can optimize this. 2482 static bool 2483 isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, 2484 bool isOffset, 2485 SmallVector<Instruction *, 4> &LifetimeMarkers) { 2486 // We track lifetime intrinsics as we encounter them. If we decide to go 2487 // ahead and replace the value with the global, this lets the caller quickly 2488 // eliminate the markers. 2489 2490 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 2491 User *U = cast<Instruction>(*UI); 2492 2493 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 2494 // Ignore non-volatile loads, they are always ok. 2495 if (!LI->isSimple()) return false; 2496 continue; 2497 } 2498 2499 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 2500 // If uses of the bitcast are ok, we are ok. 2501 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset, 2502 LifetimeMarkers)) 2503 return false; 2504 continue; 2505 } 2506 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 2507 // If the GEP has all zero indices, it doesn't offset the pointer. If it 2508 // doesn't, it does. 2509 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 2510 isOffset || !GEP->hasAllZeroIndices(), 2511 LifetimeMarkers)) 2512 return false; 2513 continue; 2514 } 2515 2516 if (CallSite CS = U) { 2517 // If this is the function being called then we treat it like a load and 2518 // ignore it. 2519 if (CS.isCallee(UI)) 2520 continue; 2521 2522 // If this is a readonly/readnone call site, then we know it is just a 2523 // load (but one that potentially returns the value itself), so we can 2524 // ignore it if we know that the value isn't captured. 2525 unsigned ArgNo = CS.getArgumentNo(UI); 2526 if (CS.onlyReadsMemory() && 2527 (CS.getInstruction()->use_empty() || 2528 CS.paramHasAttr(ArgNo+1, Attribute::NoCapture))) 2529 continue; 2530 2531 // If this is being passed as a byval argument, the caller is making a 2532 // copy, so it is only a read of the alloca. 2533 if (CS.paramHasAttr(ArgNo+1, Attribute::ByVal)) 2534 continue; 2535 } 2536 2537 // Lifetime intrinsics can be handled by the caller. 2538 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) { 2539 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 2540 II->getIntrinsicID() == Intrinsic::lifetime_end) { 2541 assert(II->use_empty() && "Lifetime markers have no result to use!"); 2542 LifetimeMarkers.push_back(II); 2543 continue; 2544 } 2545 } 2546 2547 // If this is isn't our memcpy/memmove, reject it as something we can't 2548 // handle. 2549 MemTransferInst *MI = dyn_cast<MemTransferInst>(U); 2550 if (MI == 0) 2551 return false; 2552 2553 // If the transfer is using the alloca as a source of the transfer, then 2554 // ignore it since it is a load (unless the transfer is volatile). 2555 if (UI.getOperandNo() == 1) { 2556 if (MI->isVolatile()) return false; 2557 continue; 2558 } 2559 2560 // If we already have seen a copy, reject the second one. 2561 if (TheCopy) return false; 2562 2563 // If the pointer has been offset from the start of the alloca, we can't 2564 // safely handle this. 2565 if (isOffset) return false; 2566 2567 // If the memintrinsic isn't using the alloca as the dest, reject it. 2568 if (UI.getOperandNo() != 0) return false; 2569 2570 // If the source of the memcpy/move is not a constant global, reject it. 2571 if (!PointsToConstantGlobal(MI->getSource())) 2572 return false; 2573 2574 // Otherwise, the transform is safe. Remember the copy instruction. 2575 TheCopy = MI; 2576 } 2577 return true; 2578 } 2579 2580 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 2581 /// modified by a copy from a constant global. If we can prove this, we can 2582 /// replace any uses of the alloca with uses of the global directly. 2583 MemTransferInst * 2584 SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI, 2585 SmallVector<Instruction*, 4> &ToDelete) { 2586 MemTransferInst *TheCopy = 0; 2587 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false, ToDelete)) 2588 return TheCopy; 2589 return 0; 2590 } 2591