ScalarReplAggregates.cpp revision a9203109f4ac95aa7e9624f2838e3d89623ec902
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" 52using namespace llvm; 53 54STATISTIC(NumReplaced, "Number of allocas broken up"); 55STATISTIC(NumPromoted, "Number of allocas promoted"); 56STATISTIC(NumAdjusted, "Number of scalar allocas adjusted to allow promotion"); 57STATISTIC(NumConverted, "Number of aggregates converted to scalar"); 58STATISTIC(NumGlobals, "Number of allocas copied from constant global"); 59 60namespace { 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 RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 149 AllocaInst *AI, 150 SmallVector<AllocaInst*, 32> &NewElts); 151 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 152 SmallVector<AllocaInst*, 32> &NewElts); 153 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 154 SmallVector<AllocaInst*, 32> &NewElts); 155 156 static MemTransferInst *isOnlyCopiedFromConstantGlobal( 157 AllocaInst *AI, SmallVector<Instruction*, 4> &ToDelete); 158 }; 159 160 // SROA_DT - SROA that uses DominatorTree. 161 struct SROA_DT : public SROA { 162 static char ID; 163 public: 164 SROA_DT(int T = -1) : SROA(T, true, ID) { 165 initializeSROA_DTPass(*PassRegistry::getPassRegistry()); 166 } 167 168 // getAnalysisUsage - This pass does not require any passes, but we know it 169 // will not alter the CFG, so say so. 170 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 171 AU.addRequired<DominatorTree>(); 172 AU.setPreservesCFG(); 173 } 174 }; 175 176 // SROA_SSAUp - SROA that uses SSAUpdater. 177 struct SROA_SSAUp : public SROA { 178 static char ID; 179 public: 180 SROA_SSAUp(int T = -1) : SROA(T, false, ID) { 181 initializeSROA_SSAUpPass(*PassRegistry::getPassRegistry()); 182 } 183 184 // getAnalysisUsage - This pass does not require any passes, but we know it 185 // will not alter the CFG, so say so. 186 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 187 AU.setPreservesCFG(); 188 } 189 }; 190 191} 192 193char SROA_DT::ID = 0; 194char SROA_SSAUp::ID = 0; 195 196INITIALIZE_PASS_BEGIN(SROA_DT, "scalarrepl", 197 "Scalar Replacement of Aggregates (DT)", false, false) 198INITIALIZE_PASS_DEPENDENCY(DominatorTree) 199INITIALIZE_PASS_END(SROA_DT, "scalarrepl", 200 "Scalar Replacement of Aggregates (DT)", false, false) 201 202INITIALIZE_PASS_BEGIN(SROA_SSAUp, "scalarrepl-ssa", 203 "Scalar Replacement of Aggregates (SSAUp)", false, false) 204INITIALIZE_PASS_END(SROA_SSAUp, "scalarrepl-ssa", 205 "Scalar Replacement of Aggregates (SSAUp)", false, false) 206 207// Public interface to the ScalarReplAggregates pass 208FunctionPass *llvm::createScalarReplAggregatesPass(int Threshold, 209 bool UseDomTree) { 210 if (UseDomTree) 211 return new SROA_DT(Threshold); 212 return new SROA_SSAUp(Threshold); 213} 214 215 216//===----------------------------------------------------------------------===// 217// Convert To Scalar Optimization. 218//===----------------------------------------------------------------------===// 219 220namespace { 221/// ConvertToScalarInfo - This class implements the "Convert To Scalar" 222/// optimization, which scans the uses of an alloca and determines if it can 223/// rewrite it in terms of a single new alloca that can be mem2reg'd. 224class ConvertToScalarInfo { 225 /// AllocaSize - The size of the alloca being considered in bytes. 226 unsigned AllocaSize; 227 const TargetData &TD; 228 229 /// IsNotTrivial - This is set to true if there is some access to the object 230 /// which means that mem2reg can't promote it. 231 bool IsNotTrivial; 232 233 /// ScalarKind - Tracks the kind of alloca being considered for promotion, 234 /// computed based on the uses of the alloca rather than the LLVM type system. 235 enum { 236 Unknown, 237 238 // Accesses via GEPs that are consistent with element access of a vector 239 // type. This will not be converted into a vector unless there is a later 240 // access using an actual vector type. 241 ImplicitVector, 242 243 // Accesses via vector operations and GEPs that are consistent with the 244 // layout of a vector type. 245 Vector, 246 247 // An integer bag-of-bits with bitwise operations for insertion and 248 // extraction. Any combination of types can be converted into this kind 249 // of scalar. 250 Integer 251 } ScalarKind; 252 253 /// VectorTy - This tracks the type that we should promote the vector to if 254 /// it is possible to turn it into a vector. This starts out null, and if it 255 /// isn't possible to turn into a vector type, it gets set to VoidTy. 256 VectorType *VectorTy; 257 258 /// HadNonMemTransferAccess - True if there is at least one access to the 259 /// alloca that is not a MemTransferInst. We don't want to turn structs into 260 /// large integers unless there is some potential for optimization. 261 bool HadNonMemTransferAccess; 262 263public: 264 explicit ConvertToScalarInfo(unsigned Size, const TargetData &td) 265 : AllocaSize(Size), TD(td), IsNotTrivial(false), ScalarKind(Unknown), 266 VectorTy(0), HadNonMemTransferAccess(false) { } 267 268 AllocaInst *TryConvert(AllocaInst *AI); 269 270private: 271 bool CanConvertToScalar(Value *V, uint64_t Offset); 272 void MergeInTypeForLoadOrStore(Type *In, uint64_t Offset); 273 bool MergeInVectorType(VectorType *VInTy, uint64_t Offset); 274 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); 275 276 Value *ConvertScalar_ExtractValue(Value *NV, Type *ToType, 277 uint64_t Offset, IRBuilder<> &Builder); 278 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 279 uint64_t Offset, IRBuilder<> &Builder); 280}; 281} // end anonymous namespace. 282 283 284/// TryConvert - Analyze the specified alloca, and if it is safe to do so, 285/// rewrite it to be a new alloca which is mem2reg'able. This returns the new 286/// alloca if possible or null if not. 287AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) { 288 // If we can't convert this scalar, or if mem2reg can trivially do it, bail 289 // out. 290 if (!CanConvertToScalar(AI, 0) || !IsNotTrivial) 291 return 0; 292 293 // If an alloca has only memset / memcpy uses, it may still have an Unknown 294 // ScalarKind. Treat it as an Integer below. 295 if (ScalarKind == Unknown) 296 ScalarKind = Integer; 297 298 // FIXME: It should be possible to promote the vector type up to the alloca's 299 // size. 300 if (ScalarKind == Vector && VectorTy->getBitWidth() != AllocaSize * 8) 301 ScalarKind = Integer; 302 303 // If we were able to find a vector type that can handle this with 304 // insert/extract elements, and if there was at least one use that had 305 // a vector type, promote this to a vector. We don't want to promote 306 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 307 // we just get a lot of insert/extracts. If at least one vector is 308 // involved, then we probably really do have a union of vector/array. 309 Type *NewTy; 310 if (ScalarKind == Vector) { 311 assert(VectorTy && "Missing type for vector scalar."); 312 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = " 313 << *VectorTy << '\n'); 314 NewTy = VectorTy; // Use the vector type. 315 } else { 316 unsigned BitWidth = AllocaSize * 8; 317 if ((ScalarKind == ImplicitVector || ScalarKind == Integer) && 318 !HadNonMemTransferAccess && !TD.fitsInLegalInteger(BitWidth)) 319 return 0; 320 321 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"); 322 // Create and insert the integer alloca. 323 NewTy = IntegerType::get(AI->getContext(), BitWidth); 324 } 325 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 326 ConvertUsesToScalar(AI, NewAI, 0); 327 return NewAI; 328} 329 330/// MergeInTypeForLoadOrStore - Add the 'In' type to the accumulated vector type 331/// (VectorTy) so far at the offset specified by Offset (which is specified in 332/// bytes). 333/// 334/// There are three cases we handle here: 335/// 1) A union of vector types of the same size and potentially its elements. 336/// Here we turn element accesses into insert/extract element operations. 337/// This promotes a <4 x float> with a store of float to the third element 338/// into a <4 x float> that uses insert element. 339/// 2) A union of vector types with power-of-2 size differences, e.g. a float, 340/// <2 x float> and <4 x float>. Here we turn element accesses into insert 341/// and extract element operations, and <2 x float> accesses into a cast to 342/// <2 x double>, an extract, and a cast back to <2 x float>. 343/// 3) A fully general blob of memory, which we turn into some (potentially 344/// large) integer type with extract and insert operations where the loads 345/// and stores would mutate the memory. We mark this by setting VectorTy 346/// to VoidTy. 347void ConvertToScalarInfo::MergeInTypeForLoadOrStore(Type *In, 348 uint64_t Offset) { 349 // If we already decided to turn this into a blob of integer memory, there is 350 // nothing to be done. 351 if (ScalarKind == Integer) 352 return; 353 354 // If this could be contributing to a vector, analyze it. 355 356 // If the In type is a vector that is the same size as the alloca, see if it 357 // matches the existing VecTy. 358 if (VectorType *VInTy = dyn_cast<VectorType>(In)) { 359 if (MergeInVectorType(VInTy, Offset)) 360 return; 361 } else if (In->isFloatTy() || In->isDoubleTy() || 362 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 && 363 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 364 // Full width accesses can be ignored, because they can always be turned 365 // into bitcasts. 366 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 367 if (EltSize == AllocaSize) 368 return; 369 370 // If we're accessing something that could be an element of a vector, see 371 // if the implied vector agrees with what we already have and if Offset is 372 // compatible with it. 373 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 && 374 (!VectorTy || Offset * 8 < VectorTy->getPrimitiveSizeInBits())) { 375 if (!VectorTy) { 376 ScalarKind = ImplicitVector; 377 VectorTy = VectorType::get(In, AllocaSize/EltSize); 378 return; 379 } 380 381 unsigned CurrentEltSize = VectorTy->getElementType() 382 ->getPrimitiveSizeInBits()/8; 383 if (EltSize == CurrentEltSize) 384 return; 385 386 if (In->isIntegerTy() && isPowerOf2_32(AllocaSize / EltSize)) 387 return; 388 } 389 } 390 391 // Otherwise, we have a case that we can't handle with an optimized vector 392 // form. We can still turn this into a large integer. 393 ScalarKind = Integer; 394} 395 396/// MergeInVectorType - Handles the vector case of MergeInTypeForLoadOrStore, 397/// returning true if the type was successfully merged and false otherwise. 398bool ConvertToScalarInfo::MergeInVectorType(VectorType *VInTy, 399 uint64_t Offset) { 400 // TODO: Support nonzero offsets? 401 if (Offset != 0) 402 return false; 403 404 // Only allow vectors that are a power-of-2 away from the size of the alloca. 405 if (!isPowerOf2_64(AllocaSize / (VInTy->getBitWidth() / 8))) 406 return false; 407 408 // If this the first vector we see, remember the type so that we know the 409 // element size. 410 if (!VectorTy) { 411 ScalarKind = Vector; 412 VectorTy = VInTy; 413 return true; 414 } 415 416 unsigned BitWidth = VectorTy->getBitWidth(); 417 unsigned InBitWidth = VInTy->getBitWidth(); 418 419 // Vectors of the same size can be converted using a simple bitcast. 420 if (InBitWidth == BitWidth && AllocaSize == (InBitWidth / 8)) { 421 ScalarKind = Vector; 422 return true; 423 } 424 425 Type *ElementTy = VectorTy->getElementType(); 426 Type *InElementTy = VInTy->getElementType(); 427 428 // If they're the same alloc size, we'll be attempting to convert between 429 // them with a vector shuffle, which requires the element types to match. 430 if (TD.getTypeAllocSize(VectorTy) == TD.getTypeAllocSize(VInTy) && 431 ElementTy != InElementTy) 432 return false; 433 434 // Do not allow mixed integer and floating-point accesses from vectors of 435 // different sizes. 436 if (ElementTy->isFloatingPointTy() != InElementTy->isFloatingPointTy()) 437 return false; 438 439 if (ElementTy->isFloatingPointTy()) { 440 // Only allow floating-point vectors of different sizes if they have the 441 // same element type. 442 // TODO: This could be loosened a bit, but would anything benefit? 443 if (ElementTy != InElementTy) 444 return false; 445 446 // There are no arbitrary-precision floating-point types, which limits the 447 // number of legal vector types with larger element types that we can form 448 // to bitcast and extract a subvector. 449 // TODO: We could support some more cases with mixed fp128 and double here. 450 if (!(BitWidth == 64 || BitWidth == 128) || 451 !(InBitWidth == 64 || InBitWidth == 128)) 452 return false; 453 } else { 454 assert(ElementTy->isIntegerTy() && "Vector elements must be either integer " 455 "or floating-point."); 456 unsigned BitWidth = ElementTy->getPrimitiveSizeInBits(); 457 unsigned InBitWidth = InElementTy->getPrimitiveSizeInBits(); 458 459 // Do not allow integer types smaller than a byte or types whose widths are 460 // not a multiple of a byte. 461 if (BitWidth < 8 || InBitWidth < 8 || 462 BitWidth % 8 != 0 || InBitWidth % 8 != 0) 463 return false; 464 } 465 466 // Pick the largest of the two vector types. 467 ScalarKind = Vector; 468 if (InBitWidth > BitWidth) 469 VectorTy = VInTy; 470 471 return true; 472} 473 474/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 475/// its accesses to a single vector type, return true and set VecTy to 476/// the new type. If we could convert the alloca into a single promotable 477/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 478/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 479/// is the current offset from the base of the alloca being analyzed. 480/// 481/// If we see at least one access to the value that is as a vector type, set the 482/// SawVec flag. 483bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) { 484 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 485 Instruction *User = cast<Instruction>(*UI); 486 487 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 488 // Don't break volatile loads. 489 if (LI->isVolatile()) 490 return false; 491 // Don't touch MMX operations. 492 if (LI->getType()->isX86_MMXTy()) 493 return false; 494 HadNonMemTransferAccess = true; 495 MergeInTypeForLoadOrStore(LI->getType(), Offset); 496 continue; 497 } 498 499 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 500 // Storing the pointer, not into the value? 501 if (SI->getOperand(0) == V || SI->isVolatile()) return false; 502 // Don't touch MMX operations. 503 if (SI->getOperand(0)->getType()->isX86_MMXTy()) 504 return false; 505 HadNonMemTransferAccess = true; 506 MergeInTypeForLoadOrStore(SI->getOperand(0)->getType(), Offset); 507 continue; 508 } 509 510 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 511 IsNotTrivial = true; // Can't be mem2reg'd. 512 if (!CanConvertToScalar(BCI, Offset)) 513 return false; 514 continue; 515 } 516 517 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 518 // If this is a GEP with a variable indices, we can't handle it. 519 if (!GEP->hasAllConstantIndices()) 520 return false; 521 522 // Compute the offset that this GEP adds to the pointer. 523 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 524 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 525 Indices); 526 // See if all uses can be converted. 527 if (!CanConvertToScalar(GEP, Offset+GEPOffset)) 528 return false; 529 IsNotTrivial = true; // Can't be mem2reg'd. 530 HadNonMemTransferAccess = true; 531 continue; 532 } 533 534 // If this is a constant sized memset of a constant value (e.g. 0) we can 535 // handle it. 536 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 537 // Store of constant value. 538 if (!isa<ConstantInt>(MSI->getValue())) 539 return false; 540 541 // Store of constant size. 542 ConstantInt *Len = dyn_cast<ConstantInt>(MSI->getLength()); 543 if (!Len) 544 return false; 545 546 // If the size differs from the alloca, we can only convert the alloca to 547 // an integer bag-of-bits. 548 // FIXME: This should handle all of the cases that are currently accepted 549 // as vector element insertions. 550 if (Len->getZExtValue() != AllocaSize || Offset != 0) 551 ScalarKind = Integer; 552 553 IsNotTrivial = true; // Can't be mem2reg'd. 554 HadNonMemTransferAccess = true; 555 continue; 556 } 557 558 // If this is a memcpy or memmove into or out of the whole allocation, we 559 // can handle it like a load or store of the scalar type. 560 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 561 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()); 562 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0) 563 return false; 564 565 IsNotTrivial = true; // Can't be mem2reg'd. 566 continue; 567 } 568 569 // Otherwise, we cannot handle this! 570 return false; 571 } 572 573 return true; 574} 575 576/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 577/// directly. This happens when we are converting an "integer union" to a 578/// single integer scalar, or when we are converting a "vector union" to a 579/// vector with insert/extractelement instructions. 580/// 581/// Offset is an offset from the original alloca, in bits that need to be 582/// shifted to the right. By the end of this, there should be no uses of Ptr. 583void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, 584 uint64_t Offset) { 585 while (!Ptr->use_empty()) { 586 Instruction *User = cast<Instruction>(Ptr->use_back()); 587 588 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 589 ConvertUsesToScalar(CI, NewAI, Offset); 590 CI->eraseFromParent(); 591 continue; 592 } 593 594 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 595 // Compute the offset that this GEP adds to the pointer. 596 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 597 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 598 Indices); 599 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 600 GEP->eraseFromParent(); 601 continue; 602 } 603 604 IRBuilder<> Builder(User); 605 606 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 607 // The load is a bit extract from NewAI shifted right by Offset bits. 608 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 609 Value *NewLoadVal 610 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 611 LI->replaceAllUsesWith(NewLoadVal); 612 LI->eraseFromParent(); 613 continue; 614 } 615 616 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 617 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 618 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 619 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 620 Builder); 621 Builder.CreateStore(New, NewAI); 622 SI->eraseFromParent(); 623 624 // If the load we just inserted is now dead, then the inserted store 625 // overwrote the entire thing. 626 if (Old->use_empty()) 627 Old->eraseFromParent(); 628 continue; 629 } 630 631 // If this is a constant sized memset of a constant value (e.g. 0) we can 632 // transform it into a store of the expanded constant value. 633 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 634 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 635 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 636 if (NumBytes != 0) { 637 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 638 639 // Compute the value replicated the right number of times. 640 APInt APVal(NumBytes*8, Val); 641 642 // Splat the value if non-zero. 643 if (Val) 644 for (unsigned i = 1; i != NumBytes; ++i) 645 APVal |= APVal << 8; 646 647 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 648 Value *New = ConvertScalar_InsertValue( 649 ConstantInt::get(User->getContext(), APVal), 650 Old, Offset, Builder); 651 Builder.CreateStore(New, NewAI); 652 653 // If the load we just inserted is now dead, then the memset overwrote 654 // the entire thing. 655 if (Old->use_empty()) 656 Old->eraseFromParent(); 657 } 658 MSI->eraseFromParent(); 659 continue; 660 } 661 662 // If this is a memcpy or memmove into or out of the whole allocation, we 663 // can handle it like a load or store of the scalar type. 664 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 665 assert(Offset == 0 && "must be store to start of alloca"); 666 667 // If the source and destination are both to the same alloca, then this is 668 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 669 // as appropriate. 670 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, &TD, 0)); 671 672 if (GetUnderlyingObject(MTI->getSource(), &TD, 0) != OrigAI) { 673 // Dest must be OrigAI, change this to be a load from the original 674 // pointer (bitcasted), then a store to our new alloca. 675 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 676 Value *SrcPtr = MTI->getSource(); 677 PointerType* SPTy = cast<PointerType>(SrcPtr->getType()); 678 PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 679 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 680 AIPTy = PointerType::get(AIPTy->getElementType(), 681 SPTy->getAddressSpace()); 682 } 683 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy); 684 685 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 686 SrcVal->setAlignment(MTI->getAlignment()); 687 Builder.CreateStore(SrcVal, NewAI); 688 } else if (GetUnderlyingObject(MTI->getDest(), &TD, 0) != OrigAI) { 689 // Src must be OrigAI, change this to be a load from NewAI then a store 690 // through the original dest pointer (bitcasted). 691 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 692 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 693 694 PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType()); 695 PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 696 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 697 AIPTy = PointerType::get(AIPTy->getElementType(), 698 DPTy->getAddressSpace()); 699 } 700 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy); 701 702 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 703 NewStore->setAlignment(MTI->getAlignment()); 704 } else { 705 // Noop transfer. Src == Dst 706 } 707 708 MTI->eraseFromParent(); 709 continue; 710 } 711 712 llvm_unreachable("Unsupported operation!"); 713 } 714} 715 716/// getScaledElementType - Gets a scaled element type for a partial vector 717/// access of an alloca. The input types must be integer or floating-point 718/// scalar or vector types, and the resulting type is an integer, float or 719/// double. 720static Type *getScaledElementType(Type *Ty1, Type *Ty2, 721 unsigned NewBitWidth) { 722 bool IsFP1 = Ty1->isFloatingPointTy() || 723 (Ty1->isVectorTy() && 724 cast<VectorType>(Ty1)->getElementType()->isFloatingPointTy()); 725 bool IsFP2 = Ty2->isFloatingPointTy() || 726 (Ty2->isVectorTy() && 727 cast<VectorType>(Ty2)->getElementType()->isFloatingPointTy()); 728 729 LLVMContext &Context = Ty1->getContext(); 730 731 // Prefer floating-point types over integer types, as integer types may have 732 // been created by earlier scalar replacement. 733 if (IsFP1 || IsFP2) { 734 if (NewBitWidth == 32) 735 return Type::getFloatTy(Context); 736 if (NewBitWidth == 64) 737 return Type::getDoubleTy(Context); 738 } 739 740 return Type::getIntNTy(Context, NewBitWidth); 741} 742 743/// CreateShuffleVectorCast - Creates a shuffle vector to convert one vector 744/// to another vector of the same element type which has the same allocation 745/// size but different primitive sizes (e.g. <3 x i32> and <4 x i32>). 746static Value *CreateShuffleVectorCast(Value *FromVal, Type *ToType, 747 IRBuilder<> &Builder) { 748 Type *FromType = FromVal->getType(); 749 VectorType *FromVTy = cast<VectorType>(FromType); 750 VectorType *ToVTy = cast<VectorType>(ToType); 751 assert((ToVTy->getElementType() == FromVTy->getElementType()) && 752 "Vectors must have the same element type"); 753 Value *UnV = UndefValue::get(FromType); 754 unsigned numEltsFrom = FromVTy->getNumElements(); 755 unsigned numEltsTo = ToVTy->getNumElements(); 756 757 SmallVector<Constant*, 3> Args; 758 Type* Int32Ty = Builder.getInt32Ty(); 759 unsigned minNumElts = std::min(numEltsFrom, numEltsTo); 760 unsigned i; 761 for (i=0; i != minNumElts; ++i) 762 Args.push_back(ConstantInt::get(Int32Ty, i)); 763 764 if (i < numEltsTo) { 765 Constant* UnC = UndefValue::get(Int32Ty); 766 for (; i != numEltsTo; ++i) 767 Args.push_back(UnC); 768 } 769 Constant *Mask = ConstantVector::get(Args); 770 return Builder.CreateShuffleVector(FromVal, UnV, Mask, "tmpV"); 771} 772 773/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 774/// or vector value FromVal, extracting the bits from the offset specified by 775/// Offset. This returns the value, which is of type ToType. 776/// 777/// This happens when we are converting an "integer union" to a single 778/// integer scalar, or when we are converting a "vector union" to a vector with 779/// insert/extractelement instructions. 780/// 781/// Offset is an offset from the original alloca, in bits that need to be 782/// shifted to the right. 783Value *ConvertToScalarInfo:: 784ConvertScalar_ExtractValue(Value *FromVal, Type *ToType, 785 uint64_t Offset, IRBuilder<> &Builder) { 786 // If the load is of the whole new alloca, no conversion is needed. 787 Type *FromType = FromVal->getType(); 788 if (FromType == ToType && Offset == 0) 789 return FromVal; 790 791 // If the result alloca is a vector type, this is either an element 792 // access or a bitcast to another vector type of the same size. 793 if (VectorType *VTy = dyn_cast<VectorType>(FromType)) { 794 unsigned FromTypeSize = TD.getTypeAllocSize(FromType); 795 unsigned ToTypeSize = TD.getTypeAllocSize(ToType); 796 if (FromTypeSize == ToTypeSize) { 797 // If the two types have the same primitive size, use a bit cast. 798 // Otherwise, it is two vectors with the same element type that has 799 // the same allocation size but different number of elements so use 800 // a shuffle vector. 801 if (FromType->getPrimitiveSizeInBits() == 802 ToType->getPrimitiveSizeInBits()) 803 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 804 else 805 return CreateShuffleVectorCast(FromVal, ToType, Builder); 806 } 807 808 if (isPowerOf2_64(FromTypeSize / ToTypeSize)) { 809 assert(!(ToType->isVectorTy() && Offset != 0) && "Can't extract a value " 810 "of a smaller vector type at a nonzero offset."); 811 812 Type *CastElementTy = getScaledElementType(FromType, ToType, 813 ToTypeSize * 8); 814 unsigned NumCastVectorElements = FromTypeSize / ToTypeSize; 815 816 LLVMContext &Context = FromVal->getContext(); 817 Type *CastTy = VectorType::get(CastElementTy, 818 NumCastVectorElements); 819 Value *Cast = Builder.CreateBitCast(FromVal, CastTy, "tmp"); 820 821 unsigned EltSize = TD.getTypeAllocSizeInBits(CastElementTy); 822 unsigned Elt = Offset/EltSize; 823 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 824 Value *Extract = Builder.CreateExtractElement(Cast, ConstantInt::get( 825 Type::getInt32Ty(Context), Elt), "tmp"); 826 return Builder.CreateBitCast(Extract, ToType, "tmp"); 827 } 828 829 // Otherwise it must be an element access. 830 unsigned Elt = 0; 831 if (Offset) { 832 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 833 Elt = Offset/EltSize; 834 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 835 } 836 // Return the element extracted out of it. 837 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get( 838 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp"); 839 if (V->getType() != ToType) 840 V = Builder.CreateBitCast(V, ToType, "tmp"); 841 return V; 842 } 843 844 // If ToType is a first class aggregate, extract out each of the pieces and 845 // use insertvalue's to form the FCA. 846 if (StructType *ST = dyn_cast<StructType>(ToType)) { 847 const StructLayout &Layout = *TD.getStructLayout(ST); 848 Value *Res = UndefValue::get(ST); 849 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 850 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 851 Offset+Layout.getElementOffsetInBits(i), 852 Builder); 853 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 854 } 855 return Res; 856 } 857 858 if (ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 859 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 860 Value *Res = UndefValue::get(AT); 861 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 862 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 863 Offset+i*EltSize, Builder); 864 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 865 } 866 return Res; 867 } 868 869 // Otherwise, this must be a union that was converted to an integer value. 870 IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 871 872 // If this is a big-endian system and the load is narrower than the 873 // full alloca type, we need to do a shift to get the right bits. 874 int ShAmt = 0; 875 if (TD.isBigEndian()) { 876 // On big-endian machines, the lowest bit is stored at the bit offset 877 // from the pointer given by getTypeStoreSizeInBits. This matters for 878 // integers with a bitwidth that is not a multiple of 8. 879 ShAmt = TD.getTypeStoreSizeInBits(NTy) - 880 TD.getTypeStoreSizeInBits(ToType) - Offset; 881 } else { 882 ShAmt = Offset; 883 } 884 885 // Note: we support negative bitwidths (with shl) which are not defined. 886 // We do this to support (f.e.) loads off the end of a structure where 887 // only some bits are used. 888 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 889 FromVal = Builder.CreateLShr(FromVal, 890 ConstantInt::get(FromVal->getType(), 891 ShAmt), "tmp"); 892 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 893 FromVal = Builder.CreateShl(FromVal, 894 ConstantInt::get(FromVal->getType(), 895 -ShAmt), "tmp"); 896 897 // Finally, unconditionally truncate the integer to the right width. 898 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType); 899 if (LIBitWidth < NTy->getBitWidth()) 900 FromVal = 901 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 902 LIBitWidth), "tmp"); 903 else if (LIBitWidth > NTy->getBitWidth()) 904 FromVal = 905 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 906 LIBitWidth), "tmp"); 907 908 // If the result is an integer, this is a trunc or bitcast. 909 if (ToType->isIntegerTy()) { 910 // Should be done. 911 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 912 // Just do a bitcast, we know the sizes match up. 913 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 914 } else { 915 // Otherwise must be a pointer. 916 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 917 } 918 assert(FromVal->getType() == ToType && "Didn't convert right?"); 919 return FromVal; 920} 921 922/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 923/// or vector value "Old" at the offset specified by Offset. 924/// 925/// This happens when we are converting an "integer union" to a 926/// single integer scalar, or when we are converting a "vector union" to a 927/// vector with insert/extractelement instructions. 928/// 929/// Offset is an offset from the original alloca, in bits that need to be 930/// shifted to the right. 931Value *ConvertToScalarInfo:: 932ConvertScalar_InsertValue(Value *SV, Value *Old, 933 uint64_t Offset, IRBuilder<> &Builder) { 934 // Convert the stored type to the actual type, shift it left to insert 935 // then 'or' into place. 936 Type *AllocaType = Old->getType(); 937 LLVMContext &Context = Old->getContext(); 938 939 if (VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 940 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy); 941 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType()); 942 943 // Changing the whole vector with memset or with an access of a different 944 // vector type? 945 if (ValSize == VecSize) { 946 // If the two types have the same primitive size, use a bit cast. 947 // Otherwise, it is two vectors with the same element type that has 948 // the same allocation size but different number of elements so use 949 // a shuffle vector. 950 if (VTy->getPrimitiveSizeInBits() == 951 SV->getType()->getPrimitiveSizeInBits()) 952 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 953 else 954 return CreateShuffleVectorCast(SV, VTy, Builder); 955 } 956 957 if (isPowerOf2_64(VecSize / ValSize)) { 958 assert(!(SV->getType()->isVectorTy() && Offset != 0) && "Can't insert a " 959 "value of a smaller vector type at a nonzero offset."); 960 961 Type *CastElementTy = getScaledElementType(VTy, SV->getType(), 962 ValSize); 963 unsigned NumCastVectorElements = VecSize / ValSize; 964 965 LLVMContext &Context = SV->getContext(); 966 Type *OldCastTy = VectorType::get(CastElementTy, 967 NumCastVectorElements); 968 Value *OldCast = Builder.CreateBitCast(Old, OldCastTy, "tmp"); 969 970 Value *SVCast = Builder.CreateBitCast(SV, CastElementTy, "tmp"); 971 972 unsigned EltSize = TD.getTypeAllocSizeInBits(CastElementTy); 973 unsigned Elt = Offset/EltSize; 974 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 975 Value *Insert = 976 Builder.CreateInsertElement(OldCast, SVCast, ConstantInt::get( 977 Type::getInt32Ty(Context), Elt), "tmp"); 978 return Builder.CreateBitCast(Insert, AllocaType, "tmp"); 979 } 980 981 // Must be an element insertion. 982 assert(SV->getType() == VTy->getElementType()); 983 uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 984 unsigned Elt = Offset/EltSize; 985 return Builder.CreateInsertElement(Old, SV, 986 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt), 987 "tmp"); 988 } 989 990 // If SV is a first-class aggregate value, insert each value recursively. 991 if (StructType *ST = dyn_cast<StructType>(SV->getType())) { 992 const StructLayout &Layout = *TD.getStructLayout(ST); 993 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 994 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 995 Old = ConvertScalar_InsertValue(Elt, Old, 996 Offset+Layout.getElementOffsetInBits(i), 997 Builder); 998 } 999 return Old; 1000 } 1001 1002 if (ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 1003 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 1004 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1005 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1006 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 1007 } 1008 return Old; 1009 } 1010 1011 // If SV is a float, convert it to the appropriate integer type. 1012 // If it is a pointer, do the same. 1013 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 1014 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); 1015 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); 1016 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); 1017 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 1018 SV = Builder.CreateBitCast(SV, 1019 IntegerType::get(SV->getContext(),SrcWidth), "tmp"); 1020 else if (SV->getType()->isPointerTy()) 1021 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()), "tmp"); 1022 1023 // Zero extend or truncate the value if needed. 1024 if (SV->getType() != AllocaType) { 1025 if (SV->getType()->getPrimitiveSizeInBits() < 1026 AllocaType->getPrimitiveSizeInBits()) 1027 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 1028 else { 1029 // Truncation may be needed if storing more than the alloca can hold 1030 // (undefined behavior). 1031 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 1032 SrcWidth = DestWidth; 1033 SrcStoreWidth = DestStoreWidth; 1034 } 1035 } 1036 1037 // If this is a big-endian system and the store is narrower than the 1038 // full alloca type, we need to do a shift to get the right bits. 1039 int ShAmt = 0; 1040 if (TD.isBigEndian()) { 1041 // On big-endian machines, the lowest bit is stored at the bit offset 1042 // from the pointer given by getTypeStoreSizeInBits. This matters for 1043 // integers with a bitwidth that is not a multiple of 8. 1044 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 1045 } else { 1046 ShAmt = Offset; 1047 } 1048 1049 // Note: we support negative bitwidths (with shr) which are not defined. 1050 // We do this to support (f.e.) stores off the end of a structure where 1051 // only some bits in the structure are set. 1052 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 1053 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 1054 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 1055 ShAmt), "tmp"); 1056 Mask <<= ShAmt; 1057 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1058 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 1059 -ShAmt), "tmp"); 1060 Mask = Mask.lshr(-ShAmt); 1061 } 1062 1063 // Mask out the bits we are about to insert from the old value, and or 1064 // in the new bits. 1065 if (SrcWidth != DestWidth) { 1066 assert(DestWidth > SrcWidth); 1067 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 1068 SV = Builder.CreateOr(Old, SV, "ins"); 1069 } 1070 return SV; 1071} 1072 1073 1074//===----------------------------------------------------------------------===// 1075// SRoA Driver 1076//===----------------------------------------------------------------------===// 1077 1078 1079bool SROA::runOnFunction(Function &F) { 1080 TD = getAnalysisIfAvailable<TargetData>(); 1081 1082 bool Changed = performPromotion(F); 1083 1084 // FIXME: ScalarRepl currently depends on TargetData more than it 1085 // theoretically needs to. It should be refactored in order to support 1086 // target-independent IR. Until this is done, just skip the actual 1087 // scalar-replacement portion of this pass. 1088 if (!TD) return Changed; 1089 1090 while (1) { 1091 bool LocalChange = performScalarRepl(F); 1092 if (!LocalChange) break; // No need to repromote if no scalarrepl 1093 Changed = true; 1094 LocalChange = performPromotion(F); 1095 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 1096 } 1097 1098 return Changed; 1099} 1100 1101namespace { 1102class AllocaPromoter : public LoadAndStorePromoter { 1103 AllocaInst *AI; 1104 DIBuilder *DIB; 1105 SmallVector<DbgDeclareInst *, 4> DDIs; 1106 SmallVector<DbgValueInst *, 4> DVIs; 1107public: 1108 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S, 1109 DIBuilder *DB) 1110 : LoadAndStorePromoter(Insts, S), AI(0), DIB(DB) {} 1111 1112 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) { 1113 // Remember which alloca we're promoting (for isInstInList). 1114 this->AI = AI; 1115 if (MDNode *DebugNode = MDNode::getIfExists(AI->getContext(), AI)) 1116 for (Value::use_iterator UI = DebugNode->use_begin(), 1117 E = DebugNode->use_end(); UI != E; ++UI) 1118 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI)) 1119 DDIs.push_back(DDI); 1120 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI)) 1121 DVIs.push_back(DVI); 1122 1123 LoadAndStorePromoter::run(Insts); 1124 AI->eraseFromParent(); 1125 for (SmallVector<DbgDeclareInst *, 4>::iterator I = DDIs.begin(), 1126 E = DDIs.end(); I != E; ++I) { 1127 DbgDeclareInst *DDI = *I; 1128 DDI->eraseFromParent(); 1129 } 1130 for (SmallVector<DbgValueInst *, 4>::iterator I = DVIs.begin(), 1131 E = DVIs.end(); I != E; ++I) { 1132 DbgValueInst *DVI = *I; 1133 DVI->eraseFromParent(); 1134 } 1135 } 1136 1137 virtual bool isInstInList(Instruction *I, 1138 const SmallVectorImpl<Instruction*> &Insts) const { 1139 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 1140 return LI->getOperand(0) == AI; 1141 return cast<StoreInst>(I)->getPointerOperand() == AI; 1142 } 1143 1144 virtual void updateDebugInfo(Instruction *Inst) const { 1145 for (SmallVector<DbgDeclareInst *, 4>::const_iterator I = DDIs.begin(), 1146 E = DDIs.end(); I != E; ++I) { 1147 DbgDeclareInst *DDI = *I; 1148 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) 1149 ConvertDebugDeclareToDebugValue(DDI, SI, *DIB); 1150 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) 1151 ConvertDebugDeclareToDebugValue(DDI, LI, *DIB); 1152 } 1153 for (SmallVector<DbgValueInst *, 4>::const_iterator I = DVIs.begin(), 1154 E = DVIs.end(); I != E; ++I) { 1155 DbgValueInst *DVI = *I; 1156 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 1157 Instruction *DbgVal = NULL; 1158 // If an argument is zero extended then use argument directly. The ZExt 1159 // may be zapped by an optimization pass in future. 1160 Argument *ExtendedArg = NULL; 1161 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0))) 1162 ExtendedArg = dyn_cast<Argument>(ZExt->getOperand(0)); 1163 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0))) 1164 ExtendedArg = dyn_cast<Argument>(SExt->getOperand(0)); 1165 if (ExtendedArg) 1166 DbgVal = DIB->insertDbgValueIntrinsic(ExtendedArg, 0, 1167 DIVariable(DVI->getVariable()), 1168 SI); 1169 else 1170 DbgVal = DIB->insertDbgValueIntrinsic(SI->getOperand(0), 0, 1171 DIVariable(DVI->getVariable()), 1172 SI); 1173 DbgVal->setDebugLoc(DVI->getDebugLoc()); 1174 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 1175 Instruction *DbgVal = 1176 DIB->insertDbgValueIntrinsic(LI->getOperand(0), 0, 1177 DIVariable(DVI->getVariable()), LI); 1178 DbgVal->setDebugLoc(DVI->getDebugLoc()); 1179 } 1180 } 1181 } 1182}; 1183} // end anon namespace 1184 1185/// isSafeSelectToSpeculate - Select instructions that use an alloca and are 1186/// subsequently loaded can be rewritten to load both input pointers and then 1187/// select between the result, allowing the load of the alloca to be promoted. 1188/// From this: 1189/// %P2 = select i1 %cond, i32* %Alloca, i32* %Other 1190/// %V = load i32* %P2 1191/// to: 1192/// %V1 = load i32* %Alloca -> will be mem2reg'd 1193/// %V2 = load i32* %Other 1194/// %V = select i1 %cond, i32 %V1, i32 %V2 1195/// 1196/// We can do this to a select if its only uses are loads and if the operand to 1197/// the select can be loaded unconditionally. 1198static bool isSafeSelectToSpeculate(SelectInst *SI, const TargetData *TD) { 1199 bool TDerefable = SI->getTrueValue()->isDereferenceablePointer(); 1200 bool FDerefable = SI->getFalseValue()->isDereferenceablePointer(); 1201 1202 for (Value::use_iterator UI = SI->use_begin(), UE = SI->use_end(); 1203 UI != UE; ++UI) { 1204 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1205 if (LI == 0 || LI->isVolatile()) return false; 1206 1207 // Both operands to the select need to be dereferencable, either absolutely 1208 // (e.g. allocas) or at this point because we can see other accesses to it. 1209 if (!TDerefable && !isSafeToLoadUnconditionally(SI->getTrueValue(), LI, 1210 LI->getAlignment(), TD)) 1211 return false; 1212 if (!FDerefable && !isSafeToLoadUnconditionally(SI->getFalseValue(), LI, 1213 LI->getAlignment(), TD)) 1214 return false; 1215 } 1216 1217 return true; 1218} 1219 1220/// isSafePHIToSpeculate - PHI instructions that use an alloca and are 1221/// subsequently loaded can be rewritten to load both input pointers in the pred 1222/// blocks and then PHI the results, allowing the load of the alloca to be 1223/// promoted. 1224/// From this: 1225/// %P2 = phi [i32* %Alloca, i32* %Other] 1226/// %V = load i32* %P2 1227/// to: 1228/// %V1 = load i32* %Alloca -> will be mem2reg'd 1229/// ... 1230/// %V2 = load i32* %Other 1231/// ... 1232/// %V = phi [i32 %V1, i32 %V2] 1233/// 1234/// We can do this to a select if its only uses are loads and if the operand to 1235/// the select can be loaded unconditionally. 1236static bool isSafePHIToSpeculate(PHINode *PN, const TargetData *TD) { 1237 // For now, we can only do this promotion if the load is in the same block as 1238 // the PHI, and if there are no stores between the phi and load. 1239 // TODO: Allow recursive phi users. 1240 // TODO: Allow stores. 1241 BasicBlock *BB = PN->getParent(); 1242 unsigned MaxAlign = 0; 1243 for (Value::use_iterator UI = PN->use_begin(), UE = PN->use_end(); 1244 UI != UE; ++UI) { 1245 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1246 if (LI == 0 || LI->isVolatile()) return false; 1247 1248 // For now we only allow loads in the same block as the PHI. This is a 1249 // common case that happens when instcombine merges two loads through a PHI. 1250 if (LI->getParent() != BB) return false; 1251 1252 // Ensure that there are no instructions between the PHI and the load that 1253 // could store. 1254 for (BasicBlock::iterator BBI = PN; &*BBI != LI; ++BBI) 1255 if (BBI->mayWriteToMemory()) 1256 return false; 1257 1258 MaxAlign = std::max(MaxAlign, LI->getAlignment()); 1259 } 1260 1261 // Okay, we know that we have one or more loads in the same block as the PHI. 1262 // We can transform this if it is safe to push the loads into the predecessor 1263 // blocks. The only thing to watch out for is that we can't put a possibly 1264 // trapping load in the predecessor if it is a critical edge. 1265 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1266 BasicBlock *Pred = PN->getIncomingBlock(i); 1267 1268 // If the predecessor has a single successor, then the edge isn't critical. 1269 if (Pred->getTerminator()->getNumSuccessors() == 1) 1270 continue; 1271 1272 Value *InVal = PN->getIncomingValue(i); 1273 1274 // If the InVal is an invoke in the pred, we can't put a load on the edge. 1275 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal)) 1276 if (II->getParent() == Pred) 1277 return false; 1278 1279 // If this pointer is always safe to load, or if we can prove that there is 1280 // already a load in the block, then we can move the load to the pred block. 1281 if (InVal->isDereferenceablePointer() || 1282 isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign, TD)) 1283 continue; 1284 1285 return false; 1286 } 1287 1288 return true; 1289} 1290 1291 1292/// tryToMakeAllocaBePromotable - This returns true if the alloca only has 1293/// direct (non-volatile) loads and stores to it. If the alloca is close but 1294/// not quite there, this will transform the code to allow promotion. As such, 1295/// it is a non-pure predicate. 1296static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const TargetData *TD) { 1297 SetVector<Instruction*, SmallVector<Instruction*, 4>, 1298 SmallPtrSet<Instruction*, 4> > InstsToRewrite; 1299 1300 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end(); 1301 UI != UE; ++UI) { 1302 User *U = *UI; 1303 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 1304 if (LI->isVolatile()) 1305 return false; 1306 continue; 1307 } 1308 1309 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 1310 if (SI->getOperand(0) == AI || SI->isVolatile()) 1311 return false; // Don't allow a store OF the AI, only INTO the AI. 1312 continue; 1313 } 1314 1315 if (SelectInst *SI = dyn_cast<SelectInst>(U)) { 1316 // If the condition being selected on is a constant, fold the select, yes 1317 // this does (rarely) happen early on. 1318 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition())) { 1319 Value *Result = SI->getOperand(1+CI->isZero()); 1320 SI->replaceAllUsesWith(Result); 1321 SI->eraseFromParent(); 1322 1323 // This is very rare and we just scrambled the use list of AI, start 1324 // over completely. 1325 return tryToMakeAllocaBePromotable(AI, TD); 1326 } 1327 1328 // If it is safe to turn "load (select c, AI, ptr)" into a select of two 1329 // loads, then we can transform this by rewriting the select. 1330 if (!isSafeSelectToSpeculate(SI, TD)) 1331 return false; 1332 1333 InstsToRewrite.insert(SI); 1334 continue; 1335 } 1336 1337 if (PHINode *PN = dyn_cast<PHINode>(U)) { 1338 if (PN->use_empty()) { // Dead PHIs can be stripped. 1339 InstsToRewrite.insert(PN); 1340 continue; 1341 } 1342 1343 // If it is safe to turn "load (phi [AI, ptr, ...])" into a PHI of loads 1344 // in the pred blocks, then we can transform this by rewriting the PHI. 1345 if (!isSafePHIToSpeculate(PN, TD)) 1346 return false; 1347 1348 InstsToRewrite.insert(PN); 1349 continue; 1350 } 1351 1352 return false; 1353 } 1354 1355 // If there are no instructions to rewrite, then all uses are load/stores and 1356 // we're done! 1357 if (InstsToRewrite.empty()) 1358 return true; 1359 1360 // If we have instructions that need to be rewritten for this to be promotable 1361 // take care of it now. 1362 for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) { 1363 if (SelectInst *SI = dyn_cast<SelectInst>(InstsToRewrite[i])) { 1364 // Selects in InstsToRewrite only have load uses. Rewrite each as two 1365 // loads with a new select. 1366 while (!SI->use_empty()) { 1367 LoadInst *LI = cast<LoadInst>(SI->use_back()); 1368 1369 IRBuilder<> Builder(LI); 1370 LoadInst *TrueLoad = 1371 Builder.CreateLoad(SI->getTrueValue(), LI->getName()+".t"); 1372 LoadInst *FalseLoad = 1373 Builder.CreateLoad(SI->getFalseValue(), LI->getName()+".f"); 1374 1375 // Transfer alignment and TBAA info if present. 1376 TrueLoad->setAlignment(LI->getAlignment()); 1377 FalseLoad->setAlignment(LI->getAlignment()); 1378 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) { 1379 TrueLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1380 FalseLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1381 } 1382 1383 Value *V = Builder.CreateSelect(SI->getCondition(), TrueLoad, FalseLoad); 1384 V->takeName(LI); 1385 LI->replaceAllUsesWith(V); 1386 LI->eraseFromParent(); 1387 } 1388 1389 // Now that all the loads are gone, the select is gone too. 1390 SI->eraseFromParent(); 1391 continue; 1392 } 1393 1394 // Otherwise, we have a PHI node which allows us to push the loads into the 1395 // predecessors. 1396 PHINode *PN = cast<PHINode>(InstsToRewrite[i]); 1397 if (PN->use_empty()) { 1398 PN->eraseFromParent(); 1399 continue; 1400 } 1401 1402 Type *LoadTy = cast<PointerType>(PN->getType())->getElementType(); 1403 PHINode *NewPN = PHINode::Create(LoadTy, PN->getNumIncomingValues(), 1404 PN->getName()+".ld", PN); 1405 1406 // Get the TBAA tag and alignment to use from one of the loads. It doesn't 1407 // matter which one we get and if any differ, it doesn't matter. 1408 LoadInst *SomeLoad = cast<LoadInst>(PN->use_back()); 1409 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa); 1410 unsigned Align = SomeLoad->getAlignment(); 1411 1412 // Rewrite all loads of the PN to use the new PHI. 1413 while (!PN->use_empty()) { 1414 LoadInst *LI = cast<LoadInst>(PN->use_back()); 1415 LI->replaceAllUsesWith(NewPN); 1416 LI->eraseFromParent(); 1417 } 1418 1419 // Inject loads into all of the pred blocks. Keep track of which blocks we 1420 // insert them into in case we have multiple edges from the same block. 1421 DenseMap<BasicBlock*, LoadInst*> InsertedLoads; 1422 1423 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1424 BasicBlock *Pred = PN->getIncomingBlock(i); 1425 LoadInst *&Load = InsertedLoads[Pred]; 1426 if (Load == 0) { 1427 Load = new LoadInst(PN->getIncomingValue(i), 1428 PN->getName() + "." + Pred->getName(), 1429 Pred->getTerminator()); 1430 Load->setAlignment(Align); 1431 if (TBAATag) Load->setMetadata(LLVMContext::MD_tbaa, TBAATag); 1432 } 1433 1434 NewPN->addIncoming(Load, Pred); 1435 } 1436 1437 PN->eraseFromParent(); 1438 } 1439 1440 ++NumAdjusted; 1441 return true; 1442} 1443 1444bool SROA::performPromotion(Function &F) { 1445 std::vector<AllocaInst*> Allocas; 1446 DominatorTree *DT = 0; 1447 if (HasDomTree) 1448 DT = &getAnalysis<DominatorTree>(); 1449 1450 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 1451 DIBuilder DIB(*F.getParent()); 1452 bool Changed = false; 1453 SmallVector<Instruction*, 64> Insts; 1454 while (1) { 1455 Allocas.clear(); 1456 1457 // Find allocas that are safe to promote, by looking at all instructions in 1458 // the entry node 1459 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 1460 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 1461 if (tryToMakeAllocaBePromotable(AI, TD)) 1462 Allocas.push_back(AI); 1463 1464 if (Allocas.empty()) break; 1465 1466 if (HasDomTree) 1467 PromoteMemToReg(Allocas, *DT); 1468 else { 1469 SSAUpdater SSA; 1470 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) { 1471 AllocaInst *AI = Allocas[i]; 1472 1473 // Build list of instructions to promote. 1474 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 1475 UI != E; ++UI) 1476 Insts.push_back(cast<Instruction>(*UI)); 1477 AllocaPromoter(Insts, SSA, &DIB).run(AI, Insts); 1478 Insts.clear(); 1479 } 1480 } 1481 NumPromoted += Allocas.size(); 1482 Changed = true; 1483 } 1484 1485 return Changed; 1486} 1487 1488 1489/// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 1490/// SROA. It must be a struct or array type with a small number of elements. 1491static bool ShouldAttemptScalarRepl(AllocaInst *AI) { 1492 Type *T = AI->getAllocatedType(); 1493 // Do not promote any struct into more than 32 separate vars. 1494 if (StructType *ST = dyn_cast<StructType>(T)) 1495 return ST->getNumElements() <= 32; 1496 // Arrays are much less likely to be safe for SROA; only consider 1497 // them if they are very small. 1498 if (ArrayType *AT = dyn_cast<ArrayType>(T)) 1499 return AT->getNumElements() <= 8; 1500 return false; 1501} 1502 1503 1504// performScalarRepl - This algorithm is a simple worklist driven algorithm, 1505// which runs on all of the alloca instructions in the function, removing them 1506// if they are only used by getelementptr instructions. 1507// 1508bool SROA::performScalarRepl(Function &F) { 1509 std::vector<AllocaInst*> WorkList; 1510 1511 // Scan the entry basic block, adding allocas to the worklist. 1512 BasicBlock &BB = F.getEntryBlock(); 1513 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 1514 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 1515 WorkList.push_back(A); 1516 1517 // Process the worklist 1518 bool Changed = false; 1519 while (!WorkList.empty()) { 1520 AllocaInst *AI = WorkList.back(); 1521 WorkList.pop_back(); 1522 1523 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 1524 // with unused elements. 1525 if (AI->use_empty()) { 1526 AI->eraseFromParent(); 1527 Changed = true; 1528 continue; 1529 } 1530 1531 // If this alloca is impossible for us to promote, reject it early. 1532 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 1533 continue; 1534 1535 // Check to see if this allocation is only modified by a memcpy/memmove from 1536 // a constant global. If this is the case, we can change all users to use 1537 // the constant global instead. This is commonly produced by the CFE by 1538 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 1539 // is only subsequently read. 1540 SmallVector<Instruction *, 4> ToDelete; 1541 if (MemTransferInst *Copy = isOnlyCopiedFromConstantGlobal(AI, ToDelete)) { 1542 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); 1543 DEBUG(dbgs() << " memcpy = " << *Copy << '\n'); 1544 for (unsigned i = 0, e = ToDelete.size(); i != e; ++i) 1545 ToDelete[i]->eraseFromParent(); 1546 Constant *TheSrc = cast<Constant>(Copy->getSource()); 1547 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 1548 Copy->eraseFromParent(); // Don't mutate the global. 1549 AI->eraseFromParent(); 1550 ++NumGlobals; 1551 Changed = true; 1552 continue; 1553 } 1554 1555 // Check to see if we can perform the core SROA transformation. We cannot 1556 // transform the allocation instruction if it is an array allocation 1557 // (allocations OF arrays are ok though), and an allocation of a scalar 1558 // value cannot be decomposed at all. 1559 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 1560 1561 // Do not promote [0 x %struct]. 1562 if (AllocaSize == 0) continue; 1563 1564 // Do not promote any struct whose size is too big. 1565 if (AllocaSize > SRThreshold) continue; 1566 1567 // If the alloca looks like a good candidate for scalar replacement, and if 1568 // all its users can be transformed, then split up the aggregate into its 1569 // separate elements. 1570 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 1571 DoScalarReplacement(AI, WorkList); 1572 Changed = true; 1573 continue; 1574 } 1575 1576 // If we can turn this aggregate value (potentially with casts) into a 1577 // simple scalar value that can be mem2reg'd into a register value. 1578 // IsNotTrivial tracks whether this is something that mem2reg could have 1579 // promoted itself. If so, we don't want to transform it needlessly. Note 1580 // that we can't just check based on the type: the alloca may be of an i32 1581 // but that has pointer arithmetic to set byte 3 of it or something. 1582 if (AllocaInst *NewAI = 1583 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) { 1584 NewAI->takeName(AI); 1585 AI->eraseFromParent(); 1586 ++NumConverted; 1587 Changed = true; 1588 continue; 1589 } 1590 1591 // Otherwise, couldn't process this alloca. 1592 } 1593 1594 return Changed; 1595} 1596 1597/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 1598/// predicate, do SROA now. 1599void SROA::DoScalarReplacement(AllocaInst *AI, 1600 std::vector<AllocaInst*> &WorkList) { 1601 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 1602 SmallVector<AllocaInst*, 32> ElementAllocas; 1603 if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 1604 ElementAllocas.reserve(ST->getNumContainedTypes()); 1605 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 1606 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 1607 AI->getAlignment(), 1608 AI->getName() + "." + Twine(i), AI); 1609 ElementAllocas.push_back(NA); 1610 WorkList.push_back(NA); // Add to worklist for recursive processing 1611 } 1612 } else { 1613 ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 1614 ElementAllocas.reserve(AT->getNumElements()); 1615 Type *ElTy = AT->getElementType(); 1616 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1617 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 1618 AI->getName() + "." + Twine(i), AI); 1619 ElementAllocas.push_back(NA); 1620 WorkList.push_back(NA); // Add to worklist for recursive processing 1621 } 1622 } 1623 1624 // Now that we have created the new alloca instructions, rewrite all the 1625 // uses of the old alloca. 1626 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 1627 1628 // Now erase any instructions that were made dead while rewriting the alloca. 1629 DeleteDeadInstructions(); 1630 AI->eraseFromParent(); 1631 1632 ++NumReplaced; 1633} 1634 1635/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 1636/// recursively including all their operands that become trivially dead. 1637void SROA::DeleteDeadInstructions() { 1638 while (!DeadInsts.empty()) { 1639 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 1640 1641 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 1642 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 1643 // Zero out the operand and see if it becomes trivially dead. 1644 // (But, don't add allocas to the dead instruction list -- they are 1645 // already on the worklist and will be deleted separately.) 1646 *OI = 0; 1647 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 1648 DeadInsts.push_back(U); 1649 } 1650 1651 I->eraseFromParent(); 1652 } 1653} 1654 1655/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 1656/// performing scalar replacement of alloca AI. The results are flagged in 1657/// the Info parameter. Offset indicates the position within AI that is 1658/// referenced by this instruction. 1659void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset, 1660 AllocaInfo &Info) { 1661 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1662 Instruction *User = cast<Instruction>(*UI); 1663 1664 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1665 isSafeForScalarRepl(BC, Offset, Info); 1666 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1667 uint64_t GEPOffset = Offset; 1668 isSafeGEP(GEPI, GEPOffset, Info); 1669 if (!Info.isUnsafe) 1670 isSafeForScalarRepl(GEPI, GEPOffset, Info); 1671 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1672 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1673 if (Length == 0) 1674 return MarkUnsafe(Info, User); 1675 isSafeMemAccess(Offset, Length->getZExtValue(), 0, 1676 UI.getOperandNo() == 0, Info, MI, 1677 true /*AllowWholeAccess*/); 1678 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1679 if (LI->isVolatile()) 1680 return MarkUnsafe(Info, User); 1681 Type *LIType = LI->getType(); 1682 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1683 LIType, false, Info, LI, true /*AllowWholeAccess*/); 1684 Info.hasALoadOrStore = true; 1685 1686 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1687 // Store is ok if storing INTO the pointer, not storing the pointer 1688 if (SI->isVolatile() || SI->getOperand(0) == I) 1689 return MarkUnsafe(Info, User); 1690 1691 Type *SIType = SI->getOperand(0)->getType(); 1692 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1693 SIType, true, Info, SI, true /*AllowWholeAccess*/); 1694 Info.hasALoadOrStore = true; 1695 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1696 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1697 } else { 1698 return MarkUnsafe(Info, User); 1699 } 1700 if (Info.isUnsafe) return; 1701 } 1702} 1703 1704 1705/// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer 1706/// derived from the alloca, we can often still split the alloca into elements. 1707/// This is useful if we have a large alloca where one element is phi'd 1708/// together somewhere: we can SRoA and promote all the other elements even if 1709/// we end up not being able to promote this one. 1710/// 1711/// All we require is that the uses of the PHI do not index into other parts of 1712/// the alloca. The most important use case for this is single load and stores 1713/// that are PHI'd together, which can happen due to code sinking. 1714void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset, 1715 AllocaInfo &Info) { 1716 // If we've already checked this PHI, don't do it again. 1717 if (PHINode *PN = dyn_cast<PHINode>(I)) 1718 if (!Info.CheckedPHIs.insert(PN)) 1719 return; 1720 1721 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1722 Instruction *User = cast<Instruction>(*UI); 1723 1724 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1725 isSafePHISelectUseForScalarRepl(BC, Offset, Info); 1726 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1727 // Only allow "bitcast" GEPs for simplicity. We could generalize this, 1728 // but would have to prove that we're staying inside of an element being 1729 // promoted. 1730 if (!GEPI->hasAllZeroIndices()) 1731 return MarkUnsafe(Info, User); 1732 isSafePHISelectUseForScalarRepl(GEPI, Offset, Info); 1733 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1734 if (LI->isVolatile()) 1735 return MarkUnsafe(Info, User); 1736 Type *LIType = LI->getType(); 1737 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1738 LIType, false, Info, LI, false /*AllowWholeAccess*/); 1739 Info.hasALoadOrStore = true; 1740 1741 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1742 // Store is ok if storing INTO the pointer, not storing the pointer 1743 if (SI->isVolatile() || SI->getOperand(0) == I) 1744 return MarkUnsafe(Info, User); 1745 1746 Type *SIType = SI->getOperand(0)->getType(); 1747 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1748 SIType, true, Info, SI, false /*AllowWholeAccess*/); 1749 Info.hasALoadOrStore = true; 1750 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1751 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1752 } else { 1753 return MarkUnsafe(Info, User); 1754 } 1755 if (Info.isUnsafe) return; 1756 } 1757} 1758 1759/// isSafeGEP - Check if a GEP instruction can be handled for scalar 1760/// replacement. It is safe when all the indices are constant, in-bounds 1761/// references, and when the resulting offset corresponds to an element within 1762/// the alloca type. The results are flagged in the Info parameter. Upon 1763/// return, Offset is adjusted as specified by the GEP indices. 1764void SROA::isSafeGEP(GetElementPtrInst *GEPI, 1765 uint64_t &Offset, AllocaInfo &Info) { 1766 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 1767 if (GEPIt == E) 1768 return; 1769 1770 // Walk through the GEP type indices, checking the types that this indexes 1771 // into. 1772 for (; GEPIt != E; ++GEPIt) { 1773 // Ignore struct elements, no extra checking needed for these. 1774 if ((*GEPIt)->isStructTy()) 1775 continue; 1776 1777 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 1778 if (!IdxVal) 1779 return MarkUnsafe(Info, GEPI); 1780 } 1781 1782 // Compute the offset due to this GEP and check if the alloca has a 1783 // component element at that offset. 1784 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1785 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices); 1786 if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 0)) 1787 MarkUnsafe(Info, GEPI); 1788} 1789 1790/// isHomogeneousAggregate - Check if type T is a struct or array containing 1791/// elements of the same type (which is always true for arrays). If so, 1792/// return true with NumElts and EltTy set to the number of elements and the 1793/// element type, respectively. 1794static bool isHomogeneousAggregate(Type *T, unsigned &NumElts, 1795 Type *&EltTy) { 1796 if (ArrayType *AT = dyn_cast<ArrayType>(T)) { 1797 NumElts = AT->getNumElements(); 1798 EltTy = (NumElts == 0 ? 0 : AT->getElementType()); 1799 return true; 1800 } 1801 if (StructType *ST = dyn_cast<StructType>(T)) { 1802 NumElts = ST->getNumContainedTypes(); 1803 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0)); 1804 for (unsigned n = 1; n < NumElts; ++n) { 1805 if (ST->getContainedType(n) != EltTy) 1806 return false; 1807 } 1808 return true; 1809 } 1810 return false; 1811} 1812 1813/// isCompatibleAggregate - Check if T1 and T2 are either the same type or are 1814/// "homogeneous" aggregates with the same element type and number of elements. 1815static bool isCompatibleAggregate(Type *T1, Type *T2) { 1816 if (T1 == T2) 1817 return true; 1818 1819 unsigned NumElts1, NumElts2; 1820 Type *EltTy1, *EltTy2; 1821 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) && 1822 isHomogeneousAggregate(T2, NumElts2, EltTy2) && 1823 NumElts1 == NumElts2 && 1824 EltTy1 == EltTy2) 1825 return true; 1826 1827 return false; 1828} 1829 1830/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1831/// alloca or has an offset and size that corresponds to a component element 1832/// within it. The offset checked here may have been formed from a GEP with a 1833/// pointer bitcasted to a different type. 1834/// 1835/// If AllowWholeAccess is true, then this allows uses of the entire alloca as a 1836/// unit. If false, it only allows accesses known to be in a single element. 1837void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 1838 Type *MemOpType, bool isStore, 1839 AllocaInfo &Info, Instruction *TheAccess, 1840 bool AllowWholeAccess) { 1841 // Check if this is a load/store of the entire alloca. 1842 if (Offset == 0 && AllowWholeAccess && 1843 MemSize == TD->getTypeAllocSize(Info.AI->getAllocatedType())) { 1844 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer 1845 // loads/stores (which are essentially the same as the MemIntrinsics with 1846 // regard to copying padding between elements). But, if an alloca is 1847 // flagged as both a source and destination of such operations, we'll need 1848 // to check later for padding between elements. 1849 if (!MemOpType || MemOpType->isIntegerTy()) { 1850 if (isStore) 1851 Info.isMemCpyDst = true; 1852 else 1853 Info.isMemCpySrc = true; 1854 return; 1855 } 1856 // This is also safe for references using a type that is compatible with 1857 // the type of the alloca, so that loads/stores can be rewritten using 1858 // insertvalue/extractvalue. 1859 if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) { 1860 Info.hasSubelementAccess = true; 1861 return; 1862 } 1863 } 1864 // Check if the offset/size correspond to a component within the alloca type. 1865 Type *T = Info.AI->getAllocatedType(); 1866 if (TypeHasComponent(T, Offset, MemSize)) { 1867 Info.hasSubelementAccess = true; 1868 return; 1869 } 1870 1871 return MarkUnsafe(Info, TheAccess); 1872} 1873 1874/// TypeHasComponent - Return true if T has a component type with the 1875/// specified offset and size. If Size is zero, do not check the size. 1876bool SROA::TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size) { 1877 Type *EltTy; 1878 uint64_t EltSize; 1879 if (StructType *ST = dyn_cast<StructType>(T)) { 1880 const StructLayout *Layout = TD->getStructLayout(ST); 1881 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1882 EltTy = ST->getContainedType(EltIdx); 1883 EltSize = TD->getTypeAllocSize(EltTy); 1884 Offset -= Layout->getElementOffset(EltIdx); 1885 } else if (ArrayType *AT = dyn_cast<ArrayType>(T)) { 1886 EltTy = AT->getElementType(); 1887 EltSize = TD->getTypeAllocSize(EltTy); 1888 if (Offset >= AT->getNumElements() * EltSize) 1889 return false; 1890 Offset %= EltSize; 1891 } else { 1892 return false; 1893 } 1894 if (Offset == 0 && (Size == 0 || EltSize == Size)) 1895 return true; 1896 // Check if the component spans multiple elements. 1897 if (Offset + Size > EltSize) 1898 return false; 1899 return TypeHasComponent(EltTy, Offset, Size); 1900} 1901 1902/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1903/// the instruction I, which references it, to use the separate elements. 1904/// Offset indicates the position within AI that is referenced by this 1905/// instruction. 1906void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1907 SmallVector<AllocaInst*, 32> &NewElts) { 1908 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) { 1909 Use &TheUse = UI.getUse(); 1910 Instruction *User = cast<Instruction>(*UI++); 1911 1912 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1913 RewriteBitCast(BC, AI, Offset, NewElts); 1914 continue; 1915 } 1916 1917 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1918 RewriteGEP(GEPI, AI, Offset, NewElts); 1919 continue; 1920 } 1921 1922 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1923 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1924 uint64_t MemSize = Length->getZExtValue(); 1925 if (Offset == 0 && 1926 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 1927 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1928 // Otherwise the intrinsic can only touch a single element and the 1929 // address operand will be updated, so nothing else needs to be done. 1930 continue; 1931 } 1932 1933 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1934 Type *LIType = LI->getType(); 1935 1936 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) { 1937 // Replace: 1938 // %res = load { i32, i32 }* %alloc 1939 // with: 1940 // %load.0 = load i32* %alloc.0 1941 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1942 // %load.1 = load i32* %alloc.1 1943 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1944 // (Also works for arrays instead of structs) 1945 Value *Insert = UndefValue::get(LIType); 1946 IRBuilder<> Builder(LI); 1947 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1948 Value *Load = Builder.CreateLoad(NewElts[i], "load"); 1949 Insert = Builder.CreateInsertValue(Insert, Load, i, "insert"); 1950 } 1951 LI->replaceAllUsesWith(Insert); 1952 DeadInsts.push_back(LI); 1953 } else if (LIType->isIntegerTy() && 1954 TD->getTypeAllocSize(LIType) == 1955 TD->getTypeAllocSize(AI->getAllocatedType())) { 1956 // If this is a load of the entire alloca to an integer, rewrite it. 1957 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1958 } 1959 continue; 1960 } 1961 1962 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1963 Value *Val = SI->getOperand(0); 1964 Type *SIType = Val->getType(); 1965 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) { 1966 // Replace: 1967 // store { i32, i32 } %val, { i32, i32 }* %alloc 1968 // with: 1969 // %val.0 = extractvalue { i32, i32 } %val, 0 1970 // store i32 %val.0, i32* %alloc.0 1971 // %val.1 = extractvalue { i32, i32 } %val, 1 1972 // store i32 %val.1, i32* %alloc.1 1973 // (Also works for arrays instead of structs) 1974 IRBuilder<> Builder(SI); 1975 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1976 Value *Extract = Builder.CreateExtractValue(Val, i, Val->getName()); 1977 Builder.CreateStore(Extract, NewElts[i]); 1978 } 1979 DeadInsts.push_back(SI); 1980 } else if (SIType->isIntegerTy() && 1981 TD->getTypeAllocSize(SIType) == 1982 TD->getTypeAllocSize(AI->getAllocatedType())) { 1983 // If this is a store of the entire alloca from an integer, rewrite it. 1984 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 1985 } 1986 continue; 1987 } 1988 1989 if (isa<SelectInst>(User) || isa<PHINode>(User)) { 1990 // If we have a PHI user of the alloca itself (as opposed to a GEP or 1991 // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to 1992 // the new pointer. 1993 if (!isa<AllocaInst>(I)) continue; 1994 1995 assert(Offset == 0 && NewElts[0] && 1996 "Direct alloca use should have a zero offset"); 1997 1998 // If we have a use of the alloca, we know the derived uses will be 1999 // utilizing just the first element of the scalarized result. Insert a 2000 // bitcast of the first alloca before the user as required. 2001 AllocaInst *NewAI = NewElts[0]; 2002 BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI); 2003 NewAI->moveBefore(BCI); 2004 TheUse = BCI; 2005 continue; 2006 } 2007 } 2008} 2009 2010/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 2011/// and recursively continue updating all of its uses. 2012void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 2013 SmallVector<AllocaInst*, 32> &NewElts) { 2014 RewriteForScalarRepl(BC, AI, Offset, NewElts); 2015 if (BC->getOperand(0) != AI) 2016 return; 2017 2018 // The bitcast references the original alloca. Replace its uses with 2019 // references to the first new element alloca. 2020 Instruction *Val = NewElts[0]; 2021 if (Val->getType() != BC->getDestTy()) { 2022 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 2023 Val->takeName(BC); 2024 } 2025 BC->replaceAllUsesWith(Val); 2026 DeadInsts.push_back(BC); 2027} 2028 2029/// FindElementAndOffset - Return the index of the element containing Offset 2030/// within the specified type, which must be either a struct or an array. 2031/// Sets T to the type of the element and Offset to the offset within that 2032/// element. IdxTy is set to the type of the index result to be used in a 2033/// GEP instruction. 2034uint64_t SROA::FindElementAndOffset(Type *&T, uint64_t &Offset, 2035 Type *&IdxTy) { 2036 uint64_t Idx = 0; 2037 if (StructType *ST = dyn_cast<StructType>(T)) { 2038 const StructLayout *Layout = TD->getStructLayout(ST); 2039 Idx = Layout->getElementContainingOffset(Offset); 2040 T = ST->getContainedType(Idx); 2041 Offset -= Layout->getElementOffset(Idx); 2042 IdxTy = Type::getInt32Ty(T->getContext()); 2043 return Idx; 2044 } 2045 ArrayType *AT = cast<ArrayType>(T); 2046 T = AT->getElementType(); 2047 uint64_t EltSize = TD->getTypeAllocSize(T); 2048 Idx = Offset / EltSize; 2049 Offset -= Idx * EltSize; 2050 IdxTy = Type::getInt64Ty(T->getContext()); 2051 return Idx; 2052} 2053 2054/// RewriteGEP - Check if this GEP instruction moves the pointer across 2055/// elements of the alloca that are being split apart, and if so, rewrite 2056/// the GEP to be relative to the new element. 2057void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 2058 SmallVector<AllocaInst*, 32> &NewElts) { 2059 uint64_t OldOffset = Offset; 2060 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 2061 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices); 2062 2063 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 2064 2065 Type *T = AI->getAllocatedType(); 2066 Type *IdxTy; 2067 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 2068 if (GEPI->getOperand(0) == AI) 2069 OldIdx = ~0ULL; // Force the GEP to be rewritten. 2070 2071 T = AI->getAllocatedType(); 2072 uint64_t EltOffset = Offset; 2073 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 2074 2075 // If this GEP does not move the pointer across elements of the alloca 2076 // being split, then it does not needs to be rewritten. 2077 if (Idx == OldIdx) 2078 return; 2079 2080 Type *i32Ty = Type::getInt32Ty(AI->getContext()); 2081 SmallVector<Value*, 8> NewArgs; 2082 NewArgs.push_back(Constant::getNullValue(i32Ty)); 2083 while (EltOffset != 0) { 2084 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 2085 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 2086 } 2087 Instruction *Val = NewElts[Idx]; 2088 if (NewArgs.size() > 1) { 2089 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs, "", GEPI); 2090 Val->takeName(GEPI); 2091 } 2092 if (Val->getType() != GEPI->getType()) 2093 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 2094 GEPI->replaceAllUsesWith(Val); 2095 DeadInsts.push_back(GEPI); 2096} 2097 2098/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 2099/// Rewrite it to copy or set the elements of the scalarized memory. 2100void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 2101 AllocaInst *AI, 2102 SmallVector<AllocaInst*, 32> &NewElts) { 2103 // If this is a memcpy/memmove, construct the other pointer as the 2104 // appropriate type. The "Other" pointer is the pointer that goes to memory 2105 // that doesn't have anything to do with the alloca that we are promoting. For 2106 // memset, this Value* stays null. 2107 Value *OtherPtr = 0; 2108 unsigned MemAlignment = MI->getAlignment(); 2109 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 2110 if (Inst == MTI->getRawDest()) 2111 OtherPtr = MTI->getRawSource(); 2112 else { 2113 assert(Inst == MTI->getRawSource()); 2114 OtherPtr = MTI->getRawDest(); 2115 } 2116 } 2117 2118 // If there is an other pointer, we want to convert it to the same pointer 2119 // type as AI has, so we can GEP through it safely. 2120 if (OtherPtr) { 2121 unsigned AddrSpace = 2122 cast<PointerType>(OtherPtr->getType())->getAddressSpace(); 2123 2124 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 2125 // optimization, but it's also required to detect the corner case where 2126 // both pointer operands are referencing the same memory, and where 2127 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 2128 // function is only called for mem intrinsics that access the whole 2129 // aggregate, so non-zero GEPs are not an issue here.) 2130 OtherPtr = OtherPtr->stripPointerCasts(); 2131 2132 // Copying the alloca to itself is a no-op: just delete it. 2133 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 2134 // This code will run twice for a no-op memcpy -- once for each operand. 2135 // Put only one reference to MI on the DeadInsts list. 2136 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), 2137 E = DeadInsts.end(); I != E; ++I) 2138 if (*I == MI) return; 2139 DeadInsts.push_back(MI); 2140 return; 2141 } 2142 2143 // If the pointer is not the right type, insert a bitcast to the right 2144 // type. 2145 Type *NewTy = 2146 PointerType::get(AI->getType()->getElementType(), AddrSpace); 2147 2148 if (OtherPtr->getType() != NewTy) 2149 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI); 2150 } 2151 2152 // Process each element of the aggregate. 2153 bool SROADest = MI->getRawDest() == Inst; 2154 2155 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 2156 2157 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2158 // If this is a memcpy/memmove, emit a GEP of the other element address. 2159 Value *OtherElt = 0; 2160 unsigned OtherEltAlign = MemAlignment; 2161 2162 if (OtherPtr) { 2163 Value *Idx[2] = { Zero, 2164 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 2165 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, 2166 OtherPtr->getName()+"."+Twine(i), 2167 MI); 2168 uint64_t EltOffset; 2169 PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 2170 Type *OtherTy = OtherPtrTy->getElementType(); 2171 if (StructType *ST = dyn_cast<StructType>(OtherTy)) { 2172 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 2173 } else { 2174 Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); 2175 EltOffset = TD->getTypeAllocSize(EltTy)*i; 2176 } 2177 2178 // The alignment of the other pointer is the guaranteed alignment of the 2179 // element, which is affected by both the known alignment of the whole 2180 // mem intrinsic and the alignment of the element. If the alignment of 2181 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 2182 // known alignment is just 4 bytes. 2183 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 2184 } 2185 2186 Value *EltPtr = NewElts[i]; 2187 Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 2188 2189 // If we got down to a scalar, insert a load or store as appropriate. 2190 if (EltTy->isSingleValueType()) { 2191 if (isa<MemTransferInst>(MI)) { 2192 if (SROADest) { 2193 // From Other to Alloca. 2194 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 2195 new StoreInst(Elt, EltPtr, MI); 2196 } else { 2197 // From Alloca to Other. 2198 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 2199 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 2200 } 2201 continue; 2202 } 2203 assert(isa<MemSetInst>(MI)); 2204 2205 // If the stored element is zero (common case), just store a null 2206 // constant. 2207 Constant *StoreVal; 2208 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) { 2209 if (CI->isZero()) { 2210 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 2211 } else { 2212 // If EltTy is a vector type, get the element type. 2213 Type *ValTy = EltTy->getScalarType(); 2214 2215 // Construct an integer with the right value. 2216 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 2217 APInt OneVal(EltSize, CI->getZExtValue()); 2218 APInt TotalVal(OneVal); 2219 // Set each byte. 2220 for (unsigned i = 0; 8*i < EltSize; ++i) { 2221 TotalVal = TotalVal.shl(8); 2222 TotalVal |= OneVal; 2223 } 2224 2225 // Convert the integer value to the appropriate type. 2226 StoreVal = ConstantInt::get(CI->getContext(), TotalVal); 2227 if (ValTy->isPointerTy()) 2228 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 2229 else if (ValTy->isFloatingPointTy()) 2230 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 2231 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 2232 2233 // If the requested value was a vector constant, create it. 2234 if (EltTy != ValTy) { 2235 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 2236 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 2237 StoreVal = ConstantVector::get(Elts); 2238 } 2239 } 2240 new StoreInst(StoreVal, EltPtr, MI); 2241 continue; 2242 } 2243 // Otherwise, if we're storing a byte variable, use a memset call for 2244 // this element. 2245 } 2246 2247 unsigned EltSize = TD->getTypeAllocSize(EltTy); 2248 2249 IRBuilder<> Builder(MI); 2250 2251 // Finally, insert the meminst for this element. 2252 if (isa<MemSetInst>(MI)) { 2253 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize, 2254 MI->isVolatile()); 2255 } else { 2256 assert(isa<MemTransferInst>(MI)); 2257 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr 2258 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr 2259 2260 if (isa<MemCpyInst>(MI)) 2261 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile()); 2262 else 2263 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile()); 2264 } 2265 } 2266 DeadInsts.push_back(MI); 2267} 2268 2269/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 2270/// overwrites the entire allocation. Extract out the pieces of the stored 2271/// integer and store them individually. 2272void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 2273 SmallVector<AllocaInst*, 32> &NewElts){ 2274 // Extract each element out of the integer according to its structure offset 2275 // and store the element value to the individual alloca. 2276 Value *SrcVal = SI->getOperand(0); 2277 Type *AllocaEltTy = AI->getAllocatedType(); 2278 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2279 2280 IRBuilder<> Builder(SI); 2281 2282 // Handle tail padding by extending the operand 2283 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 2284 SrcVal = Builder.CreateZExt(SrcVal, 2285 IntegerType::get(SI->getContext(), AllocaSizeBits)); 2286 2287 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 2288 << '\n'); 2289 2290 // There are two forms here: AI could be an array or struct. Both cases 2291 // have different ways to compute the element offset. 2292 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2293 const StructLayout *Layout = TD->getStructLayout(EltSTy); 2294 2295 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2296 // Get the number of bits to shift SrcVal to get the value. 2297 Type *FieldTy = EltSTy->getElementType(i); 2298 uint64_t Shift = Layout->getElementOffsetInBits(i); 2299 2300 if (TD->isBigEndian()) 2301 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 2302 2303 Value *EltVal = SrcVal; 2304 if (Shift) { 2305 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2306 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2307 } 2308 2309 // Truncate down to an integer of the right size. 2310 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2311 2312 // Ignore zero sized fields like {}, they obviously contain no data. 2313 if (FieldSizeBits == 0) continue; 2314 2315 if (FieldSizeBits != AllocaSizeBits) 2316 EltVal = Builder.CreateTrunc(EltVal, 2317 IntegerType::get(SI->getContext(), FieldSizeBits)); 2318 Value *DestField = NewElts[i]; 2319 if (EltVal->getType() == FieldTy) { 2320 // Storing to an integer field of this size, just do it. 2321 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 2322 // Bitcast to the right element type (for fp/vector values). 2323 EltVal = Builder.CreateBitCast(EltVal, FieldTy); 2324 } else { 2325 // Otherwise, bitcast the dest pointer (for aggregates). 2326 DestField = Builder.CreateBitCast(DestField, 2327 PointerType::getUnqual(EltVal->getType())); 2328 } 2329 new StoreInst(EltVal, DestField, SI); 2330 } 2331 2332 } else { 2333 ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 2334 Type *ArrayEltTy = ATy->getElementType(); 2335 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2336 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 2337 2338 uint64_t Shift; 2339 2340 if (TD->isBigEndian()) 2341 Shift = AllocaSizeBits-ElementOffset; 2342 else 2343 Shift = 0; 2344 2345 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2346 // Ignore zero sized fields like {}, they obviously contain no data. 2347 if (ElementSizeBits == 0) continue; 2348 2349 Value *EltVal = SrcVal; 2350 if (Shift) { 2351 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2352 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2353 } 2354 2355 // Truncate down to an integer of the right size. 2356 if (ElementSizeBits != AllocaSizeBits) 2357 EltVal = Builder.CreateTrunc(EltVal, 2358 IntegerType::get(SI->getContext(), 2359 ElementSizeBits)); 2360 Value *DestField = NewElts[i]; 2361 if (EltVal->getType() == ArrayEltTy) { 2362 // Storing to an integer field of this size, just do it. 2363 } else if (ArrayEltTy->isFloatingPointTy() || 2364 ArrayEltTy->isVectorTy()) { 2365 // Bitcast to the right element type (for fp/vector values). 2366 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy); 2367 } else { 2368 // Otherwise, bitcast the dest pointer (for aggregates). 2369 DestField = Builder.CreateBitCast(DestField, 2370 PointerType::getUnqual(EltVal->getType())); 2371 } 2372 new StoreInst(EltVal, DestField, SI); 2373 2374 if (TD->isBigEndian()) 2375 Shift -= ElementOffset; 2376 else 2377 Shift += ElementOffset; 2378 } 2379 } 2380 2381 DeadInsts.push_back(SI); 2382} 2383 2384/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 2385/// an integer. Load the individual pieces to form the aggregate value. 2386void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 2387 SmallVector<AllocaInst*, 32> &NewElts) { 2388 // Extract each element out of the NewElts according to its structure offset 2389 // and form the result value. 2390 Type *AllocaEltTy = AI->getAllocatedType(); 2391 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2392 2393 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 2394 << '\n'); 2395 2396 // There are two forms here: AI could be an array or struct. Both cases 2397 // have different ways to compute the element offset. 2398 const StructLayout *Layout = 0; 2399 uint64_t ArrayEltBitOffset = 0; 2400 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2401 Layout = TD->getStructLayout(EltSTy); 2402 } else { 2403 Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 2404 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2405 } 2406 2407 Value *ResultVal = 2408 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 2409 2410 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2411 // Load the value from the alloca. If the NewElt is an aggregate, cast 2412 // the pointer to an integer of the same size before doing the load. 2413 Value *SrcField = NewElts[i]; 2414 Type *FieldTy = 2415 cast<PointerType>(SrcField->getType())->getElementType(); 2416 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2417 2418 // Ignore zero sized fields like {}, they obviously contain no data. 2419 if (FieldSizeBits == 0) continue; 2420 2421 IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 2422 FieldSizeBits); 2423 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 2424 !FieldTy->isVectorTy()) 2425 SrcField = new BitCastInst(SrcField, 2426 PointerType::getUnqual(FieldIntTy), 2427 "", LI); 2428 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 2429 2430 // If SrcField is a fp or vector of the right size but that isn't an 2431 // integer type, bitcast to an integer so we can shift it. 2432 if (SrcField->getType() != FieldIntTy) 2433 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 2434 2435 // Zero extend the field to be the same size as the final alloca so that 2436 // we can shift and insert it. 2437 if (SrcField->getType() != ResultVal->getType()) 2438 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 2439 2440 // Determine the number of bits to shift SrcField. 2441 uint64_t Shift; 2442 if (Layout) // Struct case. 2443 Shift = Layout->getElementOffsetInBits(i); 2444 else // Array case. 2445 Shift = i*ArrayEltBitOffset; 2446 2447 if (TD->isBigEndian()) 2448 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 2449 2450 if (Shift) { 2451 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 2452 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 2453 } 2454 2455 // Don't create an 'or x, 0' on the first iteration. 2456 if (!isa<Constant>(ResultVal) || 2457 !cast<Constant>(ResultVal)->isNullValue()) 2458 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 2459 else 2460 ResultVal = SrcField; 2461 } 2462 2463 // Handle tail padding by truncating the result 2464 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 2465 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 2466 2467 LI->replaceAllUsesWith(ResultVal); 2468 DeadInsts.push_back(LI); 2469} 2470 2471/// HasPadding - Return true if the specified type has any structure or 2472/// alignment padding in between the elements that would be split apart 2473/// by SROA; return false otherwise. 2474static bool HasPadding(Type *Ty, const TargetData &TD) { 2475 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 2476 Ty = ATy->getElementType(); 2477 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 2478 } 2479 2480 // SROA currently handles only Arrays and Structs. 2481 StructType *STy = cast<StructType>(Ty); 2482 const StructLayout *SL = TD.getStructLayout(STy); 2483 unsigned PrevFieldBitOffset = 0; 2484 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 2485 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 2486 2487 // Check to see if there is any padding between this element and the 2488 // previous one. 2489 if (i) { 2490 unsigned PrevFieldEnd = 2491 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 2492 if (PrevFieldEnd < FieldBitOffset) 2493 return true; 2494 } 2495 PrevFieldBitOffset = FieldBitOffset; 2496 } 2497 // Check for tail padding. 2498 if (unsigned EltCount = STy->getNumElements()) { 2499 unsigned PrevFieldEnd = PrevFieldBitOffset + 2500 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 2501 if (PrevFieldEnd < SL->getSizeInBits()) 2502 return true; 2503 } 2504 return false; 2505} 2506 2507/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 2508/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 2509/// or 1 if safe after canonicalization has been performed. 2510bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 2511 // Loop over the use list of the alloca. We can only transform it if all of 2512 // the users are safe to transform. 2513 AllocaInfo Info(AI); 2514 2515 isSafeForScalarRepl(AI, 0, Info); 2516 if (Info.isUnsafe) { 2517 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 2518 return false; 2519 } 2520 2521 // Okay, we know all the users are promotable. If the aggregate is a memcpy 2522 // source and destination, we have to be careful. In particular, the memcpy 2523 // could be moving around elements that live in structure padding of the LLVM 2524 // types, but may actually be used. In these cases, we refuse to promote the 2525 // struct. 2526 if (Info.isMemCpySrc && Info.isMemCpyDst && 2527 HasPadding(AI->getAllocatedType(), *TD)) 2528 return false; 2529 2530 // If the alloca never has an access to just *part* of it, but is accessed 2531 // via loads and stores, then we should use ConvertToScalarInfo to promote 2532 // the alloca instead of promoting each piece at a time and inserting fission 2533 // and fusion code. 2534 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) { 2535 // If the struct/array just has one element, use basic SRoA. 2536 if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 2537 if (ST->getNumElements() > 1) return false; 2538 } else { 2539 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1) 2540 return false; 2541 } 2542 } 2543 2544 return true; 2545} 2546 2547 2548 2549/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 2550/// some part of a constant global variable. This intentionally only accepts 2551/// constant expressions because we don't can't rewrite arbitrary instructions. 2552static bool PointsToConstantGlobal(Value *V) { 2553 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 2554 return GV->isConstant(); 2555 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 2556 if (CE->getOpcode() == Instruction::BitCast || 2557 CE->getOpcode() == Instruction::GetElementPtr) 2558 return PointsToConstantGlobal(CE->getOperand(0)); 2559 return false; 2560} 2561 2562/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 2563/// pointer to an alloca. Ignore any reads of the pointer, return false if we 2564/// see any stores or other unknown uses. If we see pointer arithmetic, keep 2565/// track of whether it moves the pointer (with isOffset) but otherwise traverse 2566/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 2567/// the alloca, and if the source pointer is a pointer to a constant global, we 2568/// can optimize this. 2569static bool 2570isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, 2571 bool isOffset, 2572 SmallVector<Instruction *, 4> &LifetimeMarkers) { 2573 // We track lifetime intrinsics as we encounter them. If we decide to go 2574 // ahead and replace the value with the global, this lets the caller quickly 2575 // eliminate the markers. 2576 2577 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 2578 User *U = cast<Instruction>(*UI); 2579 2580 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 2581 // Ignore non-volatile loads, they are always ok. 2582 if (LI->isVolatile()) return false; 2583 continue; 2584 } 2585 2586 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 2587 // If uses of the bitcast are ok, we are ok. 2588 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset, 2589 LifetimeMarkers)) 2590 return false; 2591 continue; 2592 } 2593 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 2594 // If the GEP has all zero indices, it doesn't offset the pointer. If it 2595 // doesn't, it does. 2596 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 2597 isOffset || !GEP->hasAllZeroIndices(), 2598 LifetimeMarkers)) 2599 return false; 2600 continue; 2601 } 2602 2603 if (CallSite CS = U) { 2604 // If this is the function being called then we treat it like a load and 2605 // ignore it. 2606 if (CS.isCallee(UI)) 2607 continue; 2608 2609 // If this is a readonly/readnone call site, then we know it is just a 2610 // load (but one that potentially returns the value itself), so we can 2611 // ignore it if we know that the value isn't captured. 2612 unsigned ArgNo = CS.getArgumentNo(UI); 2613 if (CS.onlyReadsMemory() && 2614 (CS.getInstruction()->use_empty() || 2615 CS.paramHasAttr(ArgNo+1, Attribute::NoCapture))) 2616 continue; 2617 2618 // If this is being passed as a byval argument, the caller is making a 2619 // copy, so it is only a read of the alloca. 2620 if (CS.paramHasAttr(ArgNo+1, Attribute::ByVal)) 2621 continue; 2622 } 2623 2624 // Lifetime intrinsics can be handled by the caller. 2625 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) { 2626 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 2627 II->getIntrinsicID() == Intrinsic::lifetime_end) { 2628 assert(II->use_empty() && "Lifetime markers have no result to use!"); 2629 LifetimeMarkers.push_back(II); 2630 continue; 2631 } 2632 } 2633 2634 // If this is isn't our memcpy/memmove, reject it as something we can't 2635 // handle. 2636 MemTransferInst *MI = dyn_cast<MemTransferInst>(U); 2637 if (MI == 0) 2638 return false; 2639 2640 // If the transfer is using the alloca as a source of the transfer, then 2641 // ignore it since it is a load (unless the transfer is volatile). 2642 if (UI.getOperandNo() == 1) { 2643 if (MI->isVolatile()) return false; 2644 continue; 2645 } 2646 2647 // If we already have seen a copy, reject the second one. 2648 if (TheCopy) return false; 2649 2650 // If the pointer has been offset from the start of the alloca, we can't 2651 // safely handle this. 2652 if (isOffset) return false; 2653 2654 // If the memintrinsic isn't using the alloca as the dest, reject it. 2655 if (UI.getOperandNo() != 0) return false; 2656 2657 // If the source of the memcpy/move is not a constant global, reject it. 2658 if (!PointsToConstantGlobal(MI->getSource())) 2659 return false; 2660 2661 // Otherwise, the transform is safe. Remember the copy instruction. 2662 TheCopy = MI; 2663 } 2664 return true; 2665} 2666 2667/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 2668/// modified by a copy from a constant global. If we can prove this, we can 2669/// replace any uses of the alloca with uses of the global directly. 2670MemTransferInst * 2671SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI, 2672 SmallVector<Instruction*, 4> &ToDelete) { 2673 MemTransferInst *TheCopy = 0; 2674 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false, ToDelete)) 2675 return TheCopy; 2676 return 0; 2677} 2678