ScalarReplAggregates.cpp revision 9837a5c95981c239b34c1d0dfb51d4096409cbe1
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/DIBuilder.h" 34#include "llvm/Analysis/Dominators.h" 35#include "llvm/Analysis/Loads.h" 36#include "llvm/Analysis/ValueTracking.h" 37#include "llvm/Target/TargetData.h" 38#include "llvm/Transforms/Utils/PromoteMemToReg.h" 39#include "llvm/Transforms/Utils/Local.h" 40#include "llvm/Transforms/Utils/SSAUpdater.h" 41#include "llvm/Support/CallSite.h" 42#include "llvm/Support/Debug.h" 43#include "llvm/Support/ErrorHandling.h" 44#include "llvm/Support/GetElementPtrTypeIterator.h" 45#include "llvm/Support/IRBuilder.h" 46#include "llvm/Support/MathExtras.h" 47#include "llvm/Support/raw_ostream.h" 48#include "llvm/ADT/SetVector.h" 49#include "llvm/ADT/SmallVector.h" 50#include "llvm/ADT/Statistic.h" 51using namespace llvm; 52 53STATISTIC(NumReplaced, "Number of allocas broken up"); 54STATISTIC(NumPromoted, "Number of allocas promoted"); 55STATISTIC(NumAdjusted, "Number of scalar allocas adjusted to allow promotion"); 56STATISTIC(NumConverted, "Number of aggregates converted to scalar"); 57STATISTIC(NumGlobals, "Number of allocas copied from constant global"); 58 59namespace { 60 struct SROA : public FunctionPass { 61 SROA(int T, bool hasDT, char &ID) 62 : FunctionPass(ID), HasDomTree(hasDT) { 63 if (T == -1) 64 SRThreshold = 128; 65 else 66 SRThreshold = T; 67 } 68 69 bool runOnFunction(Function &F); 70 71 bool performScalarRepl(Function &F); 72 bool performPromotion(Function &F); 73 74 private: 75 bool HasDomTree; 76 TargetData *TD; 77 78 /// DeadInsts - Keep track of instructions we have made dead, so that 79 /// we can remove them after we are done working. 80 SmallVector<Value*, 32> DeadInsts; 81 82 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures 83 /// information about the uses. All these fields are initialized to false 84 /// and set to true when something is learned. 85 struct AllocaInfo { 86 /// The alloca to promote. 87 AllocaInst *AI; 88 89 /// CheckedPHIs - This is a set of verified PHI nodes, to prevent infinite 90 /// looping and avoid redundant work. 91 SmallPtrSet<PHINode*, 8> CheckedPHIs; 92 93 /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 94 bool isUnsafe : 1; 95 96 /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 97 bool isMemCpySrc : 1; 98 99 /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 100 bool isMemCpyDst : 1; 101 102 /// hasSubelementAccess - This is true if a subelement of the alloca is 103 /// ever accessed, or false if the alloca is only accessed with mem 104 /// intrinsics or load/store that only access the entire alloca at once. 105 bool hasSubelementAccess : 1; 106 107 /// hasALoadOrStore - This is true if there are any loads or stores to it. 108 /// The alloca may just be accessed with memcpy, for example, which would 109 /// not set this. 110 bool hasALoadOrStore : 1; 111 112 explicit AllocaInfo(AllocaInst *ai) 113 : AI(ai), isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false), 114 hasSubelementAccess(false), hasALoadOrStore(false) {} 115 }; 116 117 unsigned SRThreshold; 118 119 void MarkUnsafe(AllocaInfo &I, Instruction *User) { 120 I.isUnsafe = true; 121 DEBUG(dbgs() << " Transformation preventing inst: " << *User << '\n'); 122 } 123 124 bool isSafeAllocaToScalarRepl(AllocaInst *AI); 125 126 void isSafeForScalarRepl(Instruction *I, uint64_t Offset, AllocaInfo &Info); 127 void isSafePHISelectUseForScalarRepl(Instruction *User, uint64_t Offset, 128 AllocaInfo &Info); 129 void isSafeGEP(GetElementPtrInst *GEPI, uint64_t &Offset, AllocaInfo &Info); 130 void isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 131 const Type *MemOpType, bool isStore, AllocaInfo &Info, 132 Instruction *TheAccess, bool AllowWholeAccess); 133 bool TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size); 134 uint64_t FindElementAndOffset(const Type *&T, uint64_t &Offset, 135 const Type *&IdxTy); 136 137 void DoScalarReplacement(AllocaInst *AI, 138 std::vector<AllocaInst*> &WorkList); 139 void DeleteDeadInstructions(); 140 141 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 142 SmallVector<AllocaInst*, 32> &NewElts); 143 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 144 SmallVector<AllocaInst*, 32> &NewElts); 145 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 146 SmallVector<AllocaInst*, 32> &NewElts); 147 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 148 AllocaInst *AI, 149 SmallVector<AllocaInst*, 32> &NewElts); 150 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 151 SmallVector<AllocaInst*, 32> &NewElts); 152 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 153 SmallVector<AllocaInst*, 32> &NewElts); 154 155 static MemTransferInst *isOnlyCopiedFromConstantGlobal(AllocaInst *AI); 156 }; 157 158 // SROA_DT - SROA that uses DominatorTree. 159 struct SROA_DT : public SROA { 160 static char ID; 161 public: 162 SROA_DT(int T = -1) : SROA(T, true, ID) { 163 initializeSROA_DTPass(*PassRegistry::getPassRegistry()); 164 } 165 166 // getAnalysisUsage - This pass does not require any passes, but we know it 167 // will not alter the CFG, so say so. 168 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 169 AU.addRequired<DominatorTree>(); 170 AU.setPreservesCFG(); 171 } 172 }; 173 174 // SROA_SSAUp - SROA that uses SSAUpdater. 175 struct SROA_SSAUp : public SROA { 176 static char ID; 177 public: 178 SROA_SSAUp(int T = -1) : SROA(T, false, ID) { 179 initializeSROA_SSAUpPass(*PassRegistry::getPassRegistry()); 180 } 181 182 // getAnalysisUsage - This pass does not require any passes, but we know it 183 // will not alter the CFG, so say so. 184 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 185 AU.setPreservesCFG(); 186 } 187 }; 188 189} 190 191char SROA_DT::ID = 0; 192char SROA_SSAUp::ID = 0; 193 194INITIALIZE_PASS_BEGIN(SROA_DT, "scalarrepl", 195 "Scalar Replacement of Aggregates (DT)", false, false) 196INITIALIZE_PASS_DEPENDENCY(DominatorTree) 197INITIALIZE_PASS_END(SROA_DT, "scalarrepl", 198 "Scalar Replacement of Aggregates (DT)", false, false) 199 200INITIALIZE_PASS_BEGIN(SROA_SSAUp, "scalarrepl-ssa", 201 "Scalar Replacement of Aggregates (SSAUp)", false, false) 202INITIALIZE_PASS_END(SROA_SSAUp, "scalarrepl-ssa", 203 "Scalar Replacement of Aggregates (SSAUp)", false, false) 204 205// Public interface to the ScalarReplAggregates pass 206FunctionPass *llvm::createScalarReplAggregatesPass(int Threshold, 207 bool UseDomTree) { 208 if (UseDomTree) 209 return new SROA_DT(Threshold); 210 return new SROA_SSAUp(Threshold); 211} 212 213 214//===----------------------------------------------------------------------===// 215// Convert To Scalar Optimization. 216//===----------------------------------------------------------------------===// 217 218namespace { 219/// ConvertToScalarInfo - This class implements the "Convert To Scalar" 220/// optimization, which scans the uses of an alloca and determines if it can 221/// rewrite it in terms of a single new alloca that can be mem2reg'd. 222class ConvertToScalarInfo { 223 /// AllocaSize - The size of the alloca being considered in bytes. 224 unsigned AllocaSize; 225 const TargetData &TD; 226 227 /// IsNotTrivial - This is set to true if there is some access to the object 228 /// which means that mem2reg can't promote it. 229 bool IsNotTrivial; 230 231 /// ScalarKind - Tracks the kind of alloca being considered for promotion, 232 /// computed based on the uses of the alloca rather than the LLVM type system. 233 enum { 234 Unknown, 235 Vector, 236 Integer 237 } ScalarKind; 238 239 /// VectorTy - This tracks the type that we should promote the vector to if 240 /// it is possible to turn it into a vector. This starts out null, and if it 241 /// isn't possible to turn into a vector type, it gets set to VoidTy. 242 const VectorType *VectorTy; 243 244 /// HadAVector - True if there is at least one vector access to the alloca. 245 /// We don't want to turn random arrays into vectors and use vector element 246 /// insert/extract, but if there are element accesses to something that is 247 /// also declared as a vector, we do want to promote to a vector. 248 bool HadAVector; 249 250 /// HadNonMemTransferAccess - True if there is at least one access to the 251 /// alloca that is not a MemTransferInst. We don't want to turn structs into 252 /// large integers unless there is some potential for optimization. 253 bool HadNonMemTransferAccess; 254 255public: 256 explicit ConvertToScalarInfo(unsigned Size, const TargetData &td) 257 : AllocaSize(Size), TD(td), IsNotTrivial(false), ScalarKind(Unknown), 258 VectorTy(0), HadAVector(false), HadNonMemTransferAccess(false) { } 259 260 AllocaInst *TryConvert(AllocaInst *AI); 261 262private: 263 bool CanConvertToScalar(Value *V, uint64_t Offset); 264 void MergeInType(const Type *In, uint64_t Offset); 265 bool MergeInVectorType(const VectorType *VInTy, uint64_t Offset); 266 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); 267 268 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType, 269 uint64_t Offset, IRBuilder<> &Builder); 270 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 271 uint64_t Offset, IRBuilder<> &Builder); 272}; 273} // end anonymous namespace. 274 275 276/// TryConvert - Analyze the specified alloca, and if it is safe to do so, 277/// rewrite it to be a new alloca which is mem2reg'able. This returns the new 278/// alloca if possible or null if not. 279AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) { 280 // If we can't convert this scalar, or if mem2reg can trivially do it, bail 281 // out. 282 if (!CanConvertToScalar(AI, 0) || !IsNotTrivial) 283 return 0; 284 285 // If we were able to find a vector type that can handle this with 286 // insert/extract elements, and if there was at least one use that had 287 // a vector type, promote this to a vector. We don't want to promote 288 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 289 // we just get a lot of insert/extracts. If at least one vector is 290 // involved, then we probably really do have a union of vector/array. 291 const Type *NewTy; 292 if (VectorTy && HadAVector) { 293 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = " 294 << *VectorTy << '\n'); 295 NewTy = VectorTy; // Use the vector type. 296 } else { 297 unsigned BitWidth = AllocaSize * 8; 298 if (!HadAVector && !HadNonMemTransferAccess && 299 !TD.fitsInLegalInteger(BitWidth)) 300 return 0; 301 302 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"); 303 // Create and insert the integer alloca. 304 NewTy = IntegerType::get(AI->getContext(), BitWidth); 305 } 306 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 307 ConvertUsesToScalar(AI, NewAI, 0); 308 return NewAI; 309} 310 311/// MergeInType - Add the 'In' type to the accumulated vector type (VectorTy) 312/// so far at the offset specified by Offset (which is specified in bytes). 313/// 314/// There are three cases we handle here: 315/// 1) A union of vector types of the same size and potentially its elements. 316/// Here we turn element accesses into insert/extract element operations. 317/// This promotes a <4 x float> with a store of float to the third element 318/// into a <4 x float> that uses insert element. 319/// 2) A union of vector types with power-of-2 size differences, e.g. a float, 320/// <2 x float> and <4 x float>. Here we turn element accesses into insert 321/// and extract element operations, and <2 x float> accesses into a cast to 322/// <2 x double>, an extract, and a cast back to <2 x float>. 323/// 3) A fully general blob of memory, which we turn into some (potentially 324/// large) integer type with extract and insert operations where the loads 325/// and stores would mutate the memory. We mark this by setting VectorTy 326/// to VoidTy. 327void ConvertToScalarInfo::MergeInType(const Type *In, uint64_t Offset) { 328 // If we already decided to turn this into a blob of integer memory, there is 329 // nothing to be done. 330 if (ScalarKind == Integer) 331 return; 332 333 // If this could be contributing to a vector, analyze it. 334 335 // If the In type is a vector that is the same size as the alloca, see if it 336 // matches the existing VecTy. 337 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 338 if (MergeInVectorType(VInTy, Offset)) 339 return; 340 } else if (In->isFloatTy() || In->isDoubleTy() || 341 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 && 342 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 343 // Full width accesses can be ignored, because they can always be turned 344 // into bitcasts. 345 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 346 if (EltSize == AllocaSize) 347 return; 348 349 // If we're accessing something that could be an element of a vector, see 350 // if the implied vector agrees with what we already have and if Offset is 351 // compatible with it. 352 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 && 353 (!VectorTy || Offset * 8 < VectorTy->getPrimitiveSizeInBits())) { 354 if (!VectorTy) { 355 ScalarKind = Vector; 356 VectorTy = VectorType::get(In, AllocaSize/EltSize); 357 return; 358 } 359 360 unsigned CurrentEltSize = VectorTy->getElementType() 361 ->getPrimitiveSizeInBits()/8; 362 if (EltSize == CurrentEltSize) 363 return; 364 365 if (In->isIntegerTy() && isPowerOf2_32(AllocaSize / EltSize)) 366 return; 367 } 368 } 369 370 // Otherwise, we have a case that we can't handle with an optimized vector 371 // form. We can still turn this into a large integer. 372 ScalarKind = Integer; 373 VectorTy = 0; 374} 375 376/// MergeInVectorType - Handles the vector case of MergeInType, returning true 377/// if the type was successfully merged and false otherwise. 378bool ConvertToScalarInfo::MergeInVectorType(const VectorType *VInTy, 379 uint64_t Offset) { 380 // Remember if we saw a vector type. 381 HadAVector = true; 382 383 // TODO: Support nonzero offsets? 384 if (Offset != 0) 385 return false; 386 387 // Only allow vectors that are a power-of-2 away from the size of the alloca. 388 if (!isPowerOf2_64(AllocaSize / (VInTy->getBitWidth() / 8))) 389 return false; 390 391 // If this the first vector we see, remember the type so that we know the 392 // element size. 393 if (!VectorTy) { 394 ScalarKind = Vector; 395 VectorTy = VInTy; 396 return true; 397 } 398 399 unsigned BitWidth = VectorTy->getBitWidth(); 400 unsigned InBitWidth = VInTy->getBitWidth(); 401 402 // Vectors of the same size can be converted using a simple bitcast. 403 if (InBitWidth == BitWidth && AllocaSize == (InBitWidth / 8)) 404 return true; 405 406 const Type *ElementTy = VectorTy->getElementType(); 407 const Type *InElementTy = VInTy->getElementType(); 408 409 // Do not allow mixed integer and floating-point accesses from vectors of 410 // different sizes. 411 if (ElementTy->isFloatingPointTy() != InElementTy->isFloatingPointTy()) 412 return false; 413 414 if (ElementTy->isFloatingPointTy()) { 415 // Only allow floating-point vectors of different sizes if they have the 416 // same element type. 417 // TODO: This could be loosened a bit, but would anything benefit? 418 if (ElementTy != InElementTy) 419 return false; 420 421 // There are no arbitrary-precision floating-point types, which limits the 422 // number of legal vector types with larger element types that we can form 423 // to bitcast and extract a subvector. 424 // TODO: We could support some more cases with mixed fp128 and double here. 425 if (!(BitWidth == 64 || BitWidth == 128) || 426 !(InBitWidth == 64 || InBitWidth == 128)) 427 return false; 428 } else { 429 assert(ElementTy->isIntegerTy() && "Vector elements must be either integer " 430 "or floating-point."); 431 unsigned BitWidth = ElementTy->getPrimitiveSizeInBits(); 432 unsigned InBitWidth = InElementTy->getPrimitiveSizeInBits(); 433 434 // Do not allow integer types smaller than a byte or types whose widths are 435 // not a multiple of a byte. 436 if (BitWidth < 8 || InBitWidth < 8 || 437 BitWidth % 8 != 0 || InBitWidth % 8 != 0) 438 return false; 439 } 440 441 // Pick the largest of the two vector types. 442 if (InBitWidth > BitWidth) 443 VectorTy = VInTy; 444 445 return true; 446} 447 448/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 449/// its accesses to a single vector type, return true and set VecTy to 450/// the new type. If we could convert the alloca into a single promotable 451/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 452/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 453/// is the current offset from the base of the alloca being analyzed. 454/// 455/// If we see at least one access to the value that is as a vector type, set the 456/// SawVec flag. 457bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) { 458 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 459 Instruction *User = cast<Instruction>(*UI); 460 461 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 462 // Don't break volatile loads. 463 if (LI->isVolatile()) 464 return false; 465 // Don't touch MMX operations. 466 if (LI->getType()->isX86_MMXTy()) 467 return false; 468 HadNonMemTransferAccess = true; 469 MergeInType(LI->getType(), Offset); 470 continue; 471 } 472 473 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 474 // Storing the pointer, not into the value? 475 if (SI->getOperand(0) == V || SI->isVolatile()) return false; 476 // Don't touch MMX operations. 477 if (SI->getOperand(0)->getType()->isX86_MMXTy()) 478 return false; 479 HadNonMemTransferAccess = true; 480 MergeInType(SI->getOperand(0)->getType(), Offset); 481 continue; 482 } 483 484 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 485 IsNotTrivial = true; // Can't be mem2reg'd. 486 if (!CanConvertToScalar(BCI, Offset)) 487 return false; 488 continue; 489 } 490 491 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 492 // If this is a GEP with a variable indices, we can't handle it. 493 if (!GEP->hasAllConstantIndices()) 494 return false; 495 496 // Compute the offset that this GEP adds to the pointer. 497 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 498 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 499 &Indices[0], Indices.size()); 500 // See if all uses can be converted. 501 if (!CanConvertToScalar(GEP, Offset+GEPOffset)) 502 return false; 503 IsNotTrivial = true; // Can't be mem2reg'd. 504 HadNonMemTransferAccess = true; 505 continue; 506 } 507 508 // If this is a constant sized memset of a constant value (e.g. 0) we can 509 // handle it. 510 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 511 // Store of constant value and constant size. 512 if (!isa<ConstantInt>(MSI->getValue()) || 513 !isa<ConstantInt>(MSI->getLength())) 514 return false; 515 IsNotTrivial = true; // Can't be mem2reg'd. 516 HadNonMemTransferAccess = true; 517 continue; 518 } 519 520 // If this is a memcpy or memmove into or out of the whole allocation, we 521 // can handle it like a load or store of the scalar type. 522 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 523 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()); 524 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0) 525 return false; 526 527 IsNotTrivial = true; // Can't be mem2reg'd. 528 continue; 529 } 530 531 // Otherwise, we cannot handle this! 532 return false; 533 } 534 535 return true; 536} 537 538/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 539/// directly. This happens when we are converting an "integer union" to a 540/// single integer scalar, or when we are converting a "vector union" to a 541/// vector with insert/extractelement instructions. 542/// 543/// Offset is an offset from the original alloca, in bits that need to be 544/// shifted to the right. By the end of this, there should be no uses of Ptr. 545void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, 546 uint64_t Offset) { 547 while (!Ptr->use_empty()) { 548 Instruction *User = cast<Instruction>(Ptr->use_back()); 549 550 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 551 ConvertUsesToScalar(CI, NewAI, Offset); 552 CI->eraseFromParent(); 553 continue; 554 } 555 556 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 557 // Compute the offset that this GEP adds to the pointer. 558 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 559 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 560 &Indices[0], Indices.size()); 561 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 562 GEP->eraseFromParent(); 563 continue; 564 } 565 566 IRBuilder<> Builder(User); 567 568 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 569 // The load is a bit extract from NewAI shifted right by Offset bits. 570 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 571 Value *NewLoadVal 572 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 573 LI->replaceAllUsesWith(NewLoadVal); 574 LI->eraseFromParent(); 575 continue; 576 } 577 578 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 579 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 580 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 581 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 582 Builder); 583 Builder.CreateStore(New, NewAI); 584 SI->eraseFromParent(); 585 586 // If the load we just inserted is now dead, then the inserted store 587 // overwrote the entire thing. 588 if (Old->use_empty()) 589 Old->eraseFromParent(); 590 continue; 591 } 592 593 // If this is a constant sized memset of a constant value (e.g. 0) we can 594 // transform it into a store of the expanded constant value. 595 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 596 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 597 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 598 if (NumBytes != 0) { 599 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 600 601 // Compute the value replicated the right number of times. 602 APInt APVal(NumBytes*8, Val); 603 604 // Splat the value if non-zero. 605 if (Val) 606 for (unsigned i = 1; i != NumBytes; ++i) 607 APVal |= APVal << 8; 608 609 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 610 Value *New = ConvertScalar_InsertValue( 611 ConstantInt::get(User->getContext(), APVal), 612 Old, Offset, Builder); 613 Builder.CreateStore(New, NewAI); 614 615 // If the load we just inserted is now dead, then the memset overwrote 616 // the entire thing. 617 if (Old->use_empty()) 618 Old->eraseFromParent(); 619 } 620 MSI->eraseFromParent(); 621 continue; 622 } 623 624 // If this is a memcpy or memmove into or out of the whole allocation, we 625 // can handle it like a load or store of the scalar type. 626 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 627 assert(Offset == 0 && "must be store to start of alloca"); 628 629 // If the source and destination are both to the same alloca, then this is 630 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 631 // as appropriate. 632 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, &TD, 0)); 633 634 if (GetUnderlyingObject(MTI->getSource(), &TD, 0) != OrigAI) { 635 // Dest must be OrigAI, change this to be a load from the original 636 // pointer (bitcasted), then a store to our new alloca. 637 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 638 Value *SrcPtr = MTI->getSource(); 639 const PointerType* SPTy = cast<PointerType>(SrcPtr->getType()); 640 const PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 641 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 642 AIPTy = PointerType::get(AIPTy->getElementType(), 643 SPTy->getAddressSpace()); 644 } 645 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy); 646 647 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 648 SrcVal->setAlignment(MTI->getAlignment()); 649 Builder.CreateStore(SrcVal, NewAI); 650 } else if (GetUnderlyingObject(MTI->getDest(), &TD, 0) != OrigAI) { 651 // Src must be OrigAI, change this to be a load from NewAI then a store 652 // through the original dest pointer (bitcasted). 653 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 654 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 655 656 const PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType()); 657 const PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 658 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 659 AIPTy = PointerType::get(AIPTy->getElementType(), 660 DPTy->getAddressSpace()); 661 } 662 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy); 663 664 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 665 NewStore->setAlignment(MTI->getAlignment()); 666 } else { 667 // Noop transfer. Src == Dst 668 } 669 670 MTI->eraseFromParent(); 671 continue; 672 } 673 674 llvm_unreachable("Unsupported operation!"); 675 } 676} 677 678/// getScaledElementType - Gets a scaled element type for a partial vector 679/// access of an alloca. The input types must be integer or floating-point 680/// scalar or vector types, and the resulting type is an integer, float or 681/// double. 682static const Type *getScaledElementType(const Type *Ty1, const Type *Ty2, 683 unsigned NewBitWidth) { 684 bool IsFP1 = Ty1->isFloatingPointTy() || 685 (Ty1->isVectorTy() && 686 cast<VectorType>(Ty1)->getElementType()->isFloatingPointTy()); 687 bool IsFP2 = Ty2->isFloatingPointTy() || 688 (Ty2->isVectorTy() && 689 cast<VectorType>(Ty2)->getElementType()->isFloatingPointTy()); 690 691 LLVMContext &Context = Ty1->getContext(); 692 693 // Prefer floating-point types over integer types, as integer types may have 694 // been created by earlier scalar replacement. 695 if (IsFP1 || IsFP2) { 696 if (NewBitWidth == 32) 697 return Type::getFloatTy(Context); 698 if (NewBitWidth == 64) 699 return Type::getDoubleTy(Context); 700 } 701 702 return Type::getIntNTy(Context, NewBitWidth); 703} 704 705/// CreateShuffleVectorCast - Creates a shuffle vector to convert one vector 706/// to another vector of the same element type which has the same allocation 707/// size but different primitive sizes (e.g. <3 x i32> and <4 x i32>). 708static Value *CreateShuffleVectorCast(Value *FromVal, const Type *ToType, 709 IRBuilder<> &Builder) { 710 const Type *FromType = FromVal->getType(); 711 const VectorType *FromVTy = cast<VectorType>(FromType); 712 const VectorType *ToVTy = cast<VectorType>(ToType); 713 assert((ToVTy->getElementType() == FromVTy->getElementType()) && 714 "Vectors must have the same element type"); 715 Value *UnV = UndefValue::get(FromType); 716 unsigned numEltsFrom = FromVTy->getNumElements(); 717 unsigned numEltsTo = ToVTy->getNumElements(); 718 719 SmallVector<Constant*, 3> Args; 720 const Type* Int32Ty = Builder.getInt32Ty(); 721 unsigned minNumElts = std::min(numEltsFrom, numEltsTo); 722 unsigned i; 723 for (i=0; i != minNumElts; ++i) 724 Args.push_back(ConstantInt::get(Int32Ty, i)); 725 726 if (i < numEltsTo) { 727 Constant* UnC = UndefValue::get(Int32Ty); 728 for (; i != numEltsTo; ++i) 729 Args.push_back(UnC); 730 } 731 Constant *Mask = ConstantVector::get(Args); 732 return Builder.CreateShuffleVector(FromVal, UnV, Mask, "tmpV"); 733} 734 735/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 736/// or vector value FromVal, extracting the bits from the offset specified by 737/// Offset. This returns the value, which is of type ToType. 738/// 739/// This happens when we are converting an "integer union" to a single 740/// integer scalar, or when we are converting a "vector union" to a vector with 741/// insert/extractelement instructions. 742/// 743/// Offset is an offset from the original alloca, in bits that need to be 744/// shifted to the right. 745Value *ConvertToScalarInfo:: 746ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 747 uint64_t Offset, IRBuilder<> &Builder) { 748 // If the load is of the whole new alloca, no conversion is needed. 749 const Type *FromType = FromVal->getType(); 750 if (FromType == ToType && Offset == 0) 751 return FromVal; 752 753 // If the result alloca is a vector type, this is either an element 754 // access or a bitcast to another vector type of the same size. 755 if (const VectorType *VTy = dyn_cast<VectorType>(FromType)) { 756 unsigned FromTypeSize = TD.getTypeAllocSize(FromType); 757 unsigned ToTypeSize = TD.getTypeAllocSize(ToType); 758 if (FromTypeSize == ToTypeSize) { 759 // If the two types have the same primitive size, use a bit cast. 760 // Otherwise, it is two vectors with the same element type that has 761 // the same allocation size but different number of elements so use 762 // a shuffle vector. 763 if (FromType->getPrimitiveSizeInBits() == 764 ToType->getPrimitiveSizeInBits()) 765 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 766 else 767 return CreateShuffleVectorCast(FromVal, ToType, Builder); 768 } 769 770 if (isPowerOf2_64(FromTypeSize / ToTypeSize)) { 771 assert(!(ToType->isVectorTy() && Offset != 0) && "Can't extract a value " 772 "of a smaller vector type at a nonzero offset."); 773 774 const Type *CastElementTy = getScaledElementType(FromType, ToType, 775 ToTypeSize * 8); 776 unsigned NumCastVectorElements = FromTypeSize / ToTypeSize; 777 778 LLVMContext &Context = FromVal->getContext(); 779 const Type *CastTy = VectorType::get(CastElementTy, 780 NumCastVectorElements); 781 Value *Cast = Builder.CreateBitCast(FromVal, CastTy, "tmp"); 782 783 unsigned EltSize = TD.getTypeAllocSizeInBits(CastElementTy); 784 unsigned Elt = Offset/EltSize; 785 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 786 Value *Extract = Builder.CreateExtractElement(Cast, ConstantInt::get( 787 Type::getInt32Ty(Context), Elt), "tmp"); 788 return Builder.CreateBitCast(Extract, ToType, "tmp"); 789 } 790 791 // Otherwise it must be an element access. 792 unsigned Elt = 0; 793 if (Offset) { 794 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 795 Elt = Offset/EltSize; 796 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 797 } 798 // Return the element extracted out of it. 799 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get( 800 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp"); 801 if (V->getType() != ToType) 802 V = Builder.CreateBitCast(V, ToType, "tmp"); 803 return V; 804 } 805 806 // If ToType is a first class aggregate, extract out each of the pieces and 807 // use insertvalue's to form the FCA. 808 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 809 const StructLayout &Layout = *TD.getStructLayout(ST); 810 Value *Res = UndefValue::get(ST); 811 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 812 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 813 Offset+Layout.getElementOffsetInBits(i), 814 Builder); 815 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 816 } 817 return Res; 818 } 819 820 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 821 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 822 Value *Res = UndefValue::get(AT); 823 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 824 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 825 Offset+i*EltSize, Builder); 826 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 827 } 828 return Res; 829 } 830 831 // Otherwise, this must be a union that was converted to an integer value. 832 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 833 834 // If this is a big-endian system and the load is narrower than the 835 // full alloca type, we need to do a shift to get the right bits. 836 int ShAmt = 0; 837 if (TD.isBigEndian()) { 838 // On big-endian machines, the lowest bit is stored at the bit offset 839 // from the pointer given by getTypeStoreSizeInBits. This matters for 840 // integers with a bitwidth that is not a multiple of 8. 841 ShAmt = TD.getTypeStoreSizeInBits(NTy) - 842 TD.getTypeStoreSizeInBits(ToType) - Offset; 843 } else { 844 ShAmt = Offset; 845 } 846 847 // Note: we support negative bitwidths (with shl) which are not defined. 848 // We do this to support (f.e.) loads off the end of a structure where 849 // only some bits are used. 850 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 851 FromVal = Builder.CreateLShr(FromVal, 852 ConstantInt::get(FromVal->getType(), 853 ShAmt), "tmp"); 854 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 855 FromVal = Builder.CreateShl(FromVal, 856 ConstantInt::get(FromVal->getType(), 857 -ShAmt), "tmp"); 858 859 // Finally, unconditionally truncate the integer to the right width. 860 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType); 861 if (LIBitWidth < NTy->getBitWidth()) 862 FromVal = 863 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 864 LIBitWidth), "tmp"); 865 else if (LIBitWidth > NTy->getBitWidth()) 866 FromVal = 867 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 868 LIBitWidth), "tmp"); 869 870 // If the result is an integer, this is a trunc or bitcast. 871 if (ToType->isIntegerTy()) { 872 // Should be done. 873 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 874 // Just do a bitcast, we know the sizes match up. 875 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 876 } else { 877 // Otherwise must be a pointer. 878 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 879 } 880 assert(FromVal->getType() == ToType && "Didn't convert right?"); 881 return FromVal; 882} 883 884/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 885/// or vector value "Old" at the offset specified by Offset. 886/// 887/// This happens when we are converting an "integer union" to a 888/// single integer scalar, or when we are converting a "vector union" to a 889/// vector with insert/extractelement instructions. 890/// 891/// Offset is an offset from the original alloca, in bits that need to be 892/// shifted to the right. 893Value *ConvertToScalarInfo:: 894ConvertScalar_InsertValue(Value *SV, Value *Old, 895 uint64_t Offset, IRBuilder<> &Builder) { 896 // Convert the stored type to the actual type, shift it left to insert 897 // then 'or' into place. 898 const Type *AllocaType = Old->getType(); 899 LLVMContext &Context = Old->getContext(); 900 901 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 902 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy); 903 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType()); 904 905 // Changing the whole vector with memset or with an access of a different 906 // vector type? 907 if (ValSize == VecSize) { 908 // If the two types have the same primitive size, use a bit cast. 909 // Otherwise, it is two vectors with the same element type that has 910 // the same allocation size but different number of elements so use 911 // a shuffle vector. 912 if (VTy->getPrimitiveSizeInBits() == 913 SV->getType()->getPrimitiveSizeInBits()) 914 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 915 else 916 return CreateShuffleVectorCast(SV, VTy, Builder); 917 } 918 919 if (isPowerOf2_64(VecSize / ValSize)) { 920 assert(!(SV->getType()->isVectorTy() && Offset != 0) && "Can't insert a " 921 "value of a smaller vector type at a nonzero offset."); 922 923 const Type *CastElementTy = getScaledElementType(VTy, SV->getType(), 924 ValSize); 925 unsigned NumCastVectorElements = VecSize / ValSize; 926 927 LLVMContext &Context = SV->getContext(); 928 const Type *OldCastTy = VectorType::get(CastElementTy, 929 NumCastVectorElements); 930 Value *OldCast = Builder.CreateBitCast(Old, OldCastTy, "tmp"); 931 932 Value *SVCast = Builder.CreateBitCast(SV, CastElementTy, "tmp"); 933 934 unsigned EltSize = TD.getTypeAllocSizeInBits(CastElementTy); 935 unsigned Elt = Offset/EltSize; 936 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 937 Value *Insert = 938 Builder.CreateInsertElement(OldCast, SVCast, ConstantInt::get( 939 Type::getInt32Ty(Context), Elt), "tmp"); 940 return Builder.CreateBitCast(Insert, AllocaType, "tmp"); 941 } 942 943 // Must be an element insertion. 944 assert(SV->getType() == VTy->getElementType()); 945 uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 946 unsigned Elt = Offset/EltSize; 947 return Builder.CreateInsertElement(Old, SV, 948 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt), 949 "tmp"); 950 } 951 952 // If SV is a first-class aggregate value, insert each value recursively. 953 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 954 const StructLayout &Layout = *TD.getStructLayout(ST); 955 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 956 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 957 Old = ConvertScalar_InsertValue(Elt, Old, 958 Offset+Layout.getElementOffsetInBits(i), 959 Builder); 960 } 961 return Old; 962 } 963 964 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 965 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 966 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 967 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 968 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 969 } 970 return Old; 971 } 972 973 // If SV is a float, convert it to the appropriate integer type. 974 // If it is a pointer, do the same. 975 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 976 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); 977 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); 978 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); 979 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 980 SV = Builder.CreateBitCast(SV, 981 IntegerType::get(SV->getContext(),SrcWidth), "tmp"); 982 else if (SV->getType()->isPointerTy()) 983 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()), "tmp"); 984 985 // Zero extend or truncate the value if needed. 986 if (SV->getType() != AllocaType) { 987 if (SV->getType()->getPrimitiveSizeInBits() < 988 AllocaType->getPrimitiveSizeInBits()) 989 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 990 else { 991 // Truncation may be needed if storing more than the alloca can hold 992 // (undefined behavior). 993 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 994 SrcWidth = DestWidth; 995 SrcStoreWidth = DestStoreWidth; 996 } 997 } 998 999 // If this is a big-endian system and the store is narrower than the 1000 // full alloca type, we need to do a shift to get the right bits. 1001 int ShAmt = 0; 1002 if (TD.isBigEndian()) { 1003 // On big-endian machines, the lowest bit is stored at the bit offset 1004 // from the pointer given by getTypeStoreSizeInBits. This matters for 1005 // integers with a bitwidth that is not a multiple of 8. 1006 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 1007 } else { 1008 ShAmt = Offset; 1009 } 1010 1011 // Note: we support negative bitwidths (with shr) which are not defined. 1012 // We do this to support (f.e.) stores off the end of a structure where 1013 // only some bits in the structure are set. 1014 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 1015 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 1016 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 1017 ShAmt), "tmp"); 1018 Mask <<= ShAmt; 1019 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1020 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 1021 -ShAmt), "tmp"); 1022 Mask = Mask.lshr(-ShAmt); 1023 } 1024 1025 // Mask out the bits we are about to insert from the old value, and or 1026 // in the new bits. 1027 if (SrcWidth != DestWidth) { 1028 assert(DestWidth > SrcWidth); 1029 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 1030 SV = Builder.CreateOr(Old, SV, "ins"); 1031 } 1032 return SV; 1033} 1034 1035 1036//===----------------------------------------------------------------------===// 1037// SRoA Driver 1038//===----------------------------------------------------------------------===// 1039 1040 1041bool SROA::runOnFunction(Function &F) { 1042 TD = getAnalysisIfAvailable<TargetData>(); 1043 1044 bool Changed = performPromotion(F); 1045 1046 // FIXME: ScalarRepl currently depends on TargetData more than it 1047 // theoretically needs to. It should be refactored in order to support 1048 // target-independent IR. Until this is done, just skip the actual 1049 // scalar-replacement portion of this pass. 1050 if (!TD) return Changed; 1051 1052 while (1) { 1053 bool LocalChange = performScalarRepl(F); 1054 if (!LocalChange) break; // No need to repromote if no scalarrepl 1055 Changed = true; 1056 LocalChange = performPromotion(F); 1057 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 1058 } 1059 1060 return Changed; 1061} 1062 1063namespace { 1064class AllocaPromoter : public LoadAndStorePromoter { 1065 AllocaInst *AI; 1066public: 1067 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S, 1068 DbgDeclareInst *DD, DIBuilder *&DB) 1069 : LoadAndStorePromoter(Insts, S, DD, DB), AI(0) {} 1070 1071 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) { 1072 // Remember which alloca we're promoting (for isInstInList). 1073 this->AI = AI; 1074 LoadAndStorePromoter::run(Insts); 1075 AI->eraseFromParent(); 1076 } 1077 1078 virtual bool isInstInList(Instruction *I, 1079 const SmallVectorImpl<Instruction*> &Insts) const { 1080 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 1081 return LI->getOperand(0) == AI; 1082 return cast<StoreInst>(I)->getPointerOperand() == AI; 1083 } 1084}; 1085} // end anon namespace 1086 1087/// isSafeSelectToSpeculate - Select instructions that use an alloca and are 1088/// subsequently loaded can be rewritten to load both input pointers and then 1089/// select between the result, allowing the load of the alloca to be promoted. 1090/// From this: 1091/// %P2 = select i1 %cond, i32* %Alloca, i32* %Other 1092/// %V = load i32* %P2 1093/// to: 1094/// %V1 = load i32* %Alloca -> will be mem2reg'd 1095/// %V2 = load i32* %Other 1096/// %V = select i1 %cond, i32 %V1, i32 %V2 1097/// 1098/// We can do this to a select if its only uses are loads and if the operand to 1099/// the select can be loaded unconditionally. 1100static bool isSafeSelectToSpeculate(SelectInst *SI, const TargetData *TD) { 1101 bool TDerefable = SI->getTrueValue()->isDereferenceablePointer(); 1102 bool FDerefable = SI->getFalseValue()->isDereferenceablePointer(); 1103 1104 for (Value::use_iterator UI = SI->use_begin(), UE = SI->use_end(); 1105 UI != UE; ++UI) { 1106 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1107 if (LI == 0 || LI->isVolatile()) return false; 1108 1109 // Both operands to the select need to be dereferencable, either absolutely 1110 // (e.g. allocas) or at this point because we can see other accesses to it. 1111 if (!TDerefable && !isSafeToLoadUnconditionally(SI->getTrueValue(), LI, 1112 LI->getAlignment(), TD)) 1113 return false; 1114 if (!FDerefable && !isSafeToLoadUnconditionally(SI->getFalseValue(), LI, 1115 LI->getAlignment(), TD)) 1116 return false; 1117 } 1118 1119 return true; 1120} 1121 1122/// isSafePHIToSpeculate - PHI instructions that use an alloca and are 1123/// subsequently loaded can be rewritten to load both input pointers in the pred 1124/// blocks and then PHI the results, allowing the load of the alloca to be 1125/// promoted. 1126/// From this: 1127/// %P2 = phi [i32* %Alloca, i32* %Other] 1128/// %V = load i32* %P2 1129/// to: 1130/// %V1 = load i32* %Alloca -> will be mem2reg'd 1131/// ... 1132/// %V2 = load i32* %Other 1133/// ... 1134/// %V = phi [i32 %V1, i32 %V2] 1135/// 1136/// We can do this to a select if its only uses are loads and if the operand to 1137/// the select can be loaded unconditionally. 1138static bool isSafePHIToSpeculate(PHINode *PN, const TargetData *TD) { 1139 // For now, we can only do this promotion if the load is in the same block as 1140 // the PHI, and if there are no stores between the phi and load. 1141 // TODO: Allow recursive phi users. 1142 // TODO: Allow stores. 1143 BasicBlock *BB = PN->getParent(); 1144 unsigned MaxAlign = 0; 1145 for (Value::use_iterator UI = PN->use_begin(), UE = PN->use_end(); 1146 UI != UE; ++UI) { 1147 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1148 if (LI == 0 || LI->isVolatile()) return false; 1149 1150 // For now we only allow loads in the same block as the PHI. This is a 1151 // common case that happens when instcombine merges two loads through a PHI. 1152 if (LI->getParent() != BB) return false; 1153 1154 // Ensure that there are no instructions between the PHI and the load that 1155 // could store. 1156 for (BasicBlock::iterator BBI = PN; &*BBI != LI; ++BBI) 1157 if (BBI->mayWriteToMemory()) 1158 return false; 1159 1160 MaxAlign = std::max(MaxAlign, LI->getAlignment()); 1161 } 1162 1163 // Okay, we know that we have one or more loads in the same block as the PHI. 1164 // We can transform this if it is safe to push the loads into the predecessor 1165 // blocks. The only thing to watch out for is that we can't put a possibly 1166 // trapping load in the predecessor if it is a critical edge. 1167 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1168 BasicBlock *Pred = PN->getIncomingBlock(i); 1169 1170 // If the predecessor has a single successor, then the edge isn't critical. 1171 if (Pred->getTerminator()->getNumSuccessors() == 1) 1172 continue; 1173 1174 Value *InVal = PN->getIncomingValue(i); 1175 1176 // If the InVal is an invoke in the pred, we can't put a load on the edge. 1177 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal)) 1178 if (II->getParent() == Pred) 1179 return false; 1180 1181 // If this pointer is always safe to load, or if we can prove that there is 1182 // already a load in the block, then we can move the load to the pred block. 1183 if (InVal->isDereferenceablePointer() || 1184 isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign, TD)) 1185 continue; 1186 1187 return false; 1188 } 1189 1190 return true; 1191} 1192 1193 1194/// tryToMakeAllocaBePromotable - This returns true if the alloca only has 1195/// direct (non-volatile) loads and stores to it. If the alloca is close but 1196/// not quite there, this will transform the code to allow promotion. As such, 1197/// it is a non-pure predicate. 1198static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const TargetData *TD) { 1199 SetVector<Instruction*, SmallVector<Instruction*, 4>, 1200 SmallPtrSet<Instruction*, 4> > InstsToRewrite; 1201 1202 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end(); 1203 UI != UE; ++UI) { 1204 User *U = *UI; 1205 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 1206 if (LI->isVolatile()) 1207 return false; 1208 continue; 1209 } 1210 1211 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 1212 if (SI->getOperand(0) == AI || SI->isVolatile()) 1213 return false; // Don't allow a store OF the AI, only INTO the AI. 1214 continue; 1215 } 1216 1217 if (SelectInst *SI = dyn_cast<SelectInst>(U)) { 1218 // If the condition being selected on is a constant, fold the select, yes 1219 // this does (rarely) happen early on. 1220 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition())) { 1221 Value *Result = SI->getOperand(1+CI->isZero()); 1222 SI->replaceAllUsesWith(Result); 1223 SI->eraseFromParent(); 1224 1225 // This is very rare and we just scrambled the use list of AI, start 1226 // over completely. 1227 return tryToMakeAllocaBePromotable(AI, TD); 1228 } 1229 1230 // If it is safe to turn "load (select c, AI, ptr)" into a select of two 1231 // loads, then we can transform this by rewriting the select. 1232 if (!isSafeSelectToSpeculate(SI, TD)) 1233 return false; 1234 1235 InstsToRewrite.insert(SI); 1236 continue; 1237 } 1238 1239 if (PHINode *PN = dyn_cast<PHINode>(U)) { 1240 if (PN->use_empty()) { // Dead PHIs can be stripped. 1241 InstsToRewrite.insert(PN); 1242 continue; 1243 } 1244 1245 // If it is safe to turn "load (phi [AI, ptr, ...])" into a PHI of loads 1246 // in the pred blocks, then we can transform this by rewriting the PHI. 1247 if (!isSafePHIToSpeculate(PN, TD)) 1248 return false; 1249 1250 InstsToRewrite.insert(PN); 1251 continue; 1252 } 1253 1254 return false; 1255 } 1256 1257 // If there are no instructions to rewrite, then all uses are load/stores and 1258 // we're done! 1259 if (InstsToRewrite.empty()) 1260 return true; 1261 1262 // If we have instructions that need to be rewritten for this to be promotable 1263 // take care of it now. 1264 for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) { 1265 if (SelectInst *SI = dyn_cast<SelectInst>(InstsToRewrite[i])) { 1266 // Selects in InstsToRewrite only have load uses. Rewrite each as two 1267 // loads with a new select. 1268 while (!SI->use_empty()) { 1269 LoadInst *LI = cast<LoadInst>(SI->use_back()); 1270 1271 IRBuilder<> Builder(LI); 1272 LoadInst *TrueLoad = 1273 Builder.CreateLoad(SI->getTrueValue(), LI->getName()+".t"); 1274 LoadInst *FalseLoad = 1275 Builder.CreateLoad(SI->getFalseValue(), LI->getName()+".t"); 1276 1277 // Transfer alignment and TBAA info if present. 1278 TrueLoad->setAlignment(LI->getAlignment()); 1279 FalseLoad->setAlignment(LI->getAlignment()); 1280 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) { 1281 TrueLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1282 FalseLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1283 } 1284 1285 Value *V = Builder.CreateSelect(SI->getCondition(), TrueLoad, FalseLoad); 1286 V->takeName(LI); 1287 LI->replaceAllUsesWith(V); 1288 LI->eraseFromParent(); 1289 } 1290 1291 // Now that all the loads are gone, the select is gone too. 1292 SI->eraseFromParent(); 1293 continue; 1294 } 1295 1296 // Otherwise, we have a PHI node which allows us to push the loads into the 1297 // predecessors. 1298 PHINode *PN = cast<PHINode>(InstsToRewrite[i]); 1299 if (PN->use_empty()) { 1300 PN->eraseFromParent(); 1301 continue; 1302 } 1303 1304 const Type *LoadTy = cast<PointerType>(PN->getType())->getElementType(); 1305 PHINode *NewPN = PHINode::Create(LoadTy, PN->getNumIncomingValues(), 1306 PN->getName()+".ld", PN); 1307 1308 // Get the TBAA tag and alignment to use from one of the loads. It doesn't 1309 // matter which one we get and if any differ, it doesn't matter. 1310 LoadInst *SomeLoad = cast<LoadInst>(PN->use_back()); 1311 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa); 1312 unsigned Align = SomeLoad->getAlignment(); 1313 1314 // Rewrite all loads of the PN to use the new PHI. 1315 while (!PN->use_empty()) { 1316 LoadInst *LI = cast<LoadInst>(PN->use_back()); 1317 LI->replaceAllUsesWith(NewPN); 1318 LI->eraseFromParent(); 1319 } 1320 1321 // Inject loads into all of the pred blocks. Keep track of which blocks we 1322 // insert them into in case we have multiple edges from the same block. 1323 DenseMap<BasicBlock*, LoadInst*> InsertedLoads; 1324 1325 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1326 BasicBlock *Pred = PN->getIncomingBlock(i); 1327 LoadInst *&Load = InsertedLoads[Pred]; 1328 if (Load == 0) { 1329 Load = new LoadInst(PN->getIncomingValue(i), 1330 PN->getName() + "." + Pred->getName(), 1331 Pred->getTerminator()); 1332 Load->setAlignment(Align); 1333 if (TBAATag) Load->setMetadata(LLVMContext::MD_tbaa, TBAATag); 1334 } 1335 1336 NewPN->addIncoming(Load, Pred); 1337 } 1338 1339 PN->eraseFromParent(); 1340 } 1341 1342 ++NumAdjusted; 1343 return true; 1344} 1345 1346bool SROA::performPromotion(Function &F) { 1347 std::vector<AllocaInst*> Allocas; 1348 DominatorTree *DT = 0; 1349 if (HasDomTree) 1350 DT = &getAnalysis<DominatorTree>(); 1351 1352 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 1353 1354 bool Changed = false; 1355 SmallVector<Instruction*, 64> Insts; 1356 DIBuilder *DIB = 0; 1357 while (1) { 1358 Allocas.clear(); 1359 1360 // Find allocas that are safe to promote, by looking at all instructions in 1361 // the entry node 1362 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 1363 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 1364 if (tryToMakeAllocaBePromotable(AI, TD)) 1365 Allocas.push_back(AI); 1366 1367 if (Allocas.empty()) break; 1368 1369 if (HasDomTree) 1370 PromoteMemToReg(Allocas, *DT); 1371 else { 1372 SSAUpdater SSA; 1373 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) { 1374 AllocaInst *AI = Allocas[i]; 1375 1376 // Build list of instructions to promote. 1377 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 1378 UI != E; ++UI) 1379 Insts.push_back(cast<Instruction>(*UI)); 1380 1381 DbgDeclareInst *DDI = FindAllocaDbgDeclare(AI); 1382 if (DDI && !DIB) 1383 DIB = new DIBuilder(*AI->getParent()->getParent()->getParent()); 1384 AllocaPromoter(Insts, SSA, DDI, DIB).run(AI, Insts); 1385 Insts.clear(); 1386 } 1387 } 1388 NumPromoted += Allocas.size(); 1389 Changed = true; 1390 } 1391 1392 // FIXME: Is there a better way to handle the lazy initialization of DIB 1393 // so that there doesn't need to be an explicit delete? 1394 delete DIB; 1395 1396 return Changed; 1397} 1398 1399 1400/// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 1401/// SROA. It must be a struct or array type with a small number of elements. 1402static bool ShouldAttemptScalarRepl(AllocaInst *AI) { 1403 const Type *T = AI->getAllocatedType(); 1404 // Do not promote any struct into more than 32 separate vars. 1405 if (const StructType *ST = dyn_cast<StructType>(T)) 1406 return ST->getNumElements() <= 32; 1407 // Arrays are much less likely to be safe for SROA; only consider 1408 // them if they are very small. 1409 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) 1410 return AT->getNumElements() <= 8; 1411 return false; 1412} 1413 1414 1415// performScalarRepl - This algorithm is a simple worklist driven algorithm, 1416// which runs on all of the malloc/alloca instructions in the function, removing 1417// them if they are only used by getelementptr instructions. 1418// 1419bool SROA::performScalarRepl(Function &F) { 1420 std::vector<AllocaInst*> WorkList; 1421 1422 // Scan the entry basic block, adding allocas to the worklist. 1423 BasicBlock &BB = F.getEntryBlock(); 1424 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 1425 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 1426 WorkList.push_back(A); 1427 1428 // Process the worklist 1429 bool Changed = false; 1430 while (!WorkList.empty()) { 1431 AllocaInst *AI = WorkList.back(); 1432 WorkList.pop_back(); 1433 1434 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 1435 // with unused elements. 1436 if (AI->use_empty()) { 1437 AI->eraseFromParent(); 1438 Changed = true; 1439 continue; 1440 } 1441 1442 // If this alloca is impossible for us to promote, reject it early. 1443 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 1444 continue; 1445 1446 // Check to see if this allocation is only modified by a memcpy/memmove from 1447 // a constant global. If this is the case, we can change all users to use 1448 // the constant global instead. This is commonly produced by the CFE by 1449 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 1450 // is only subsequently read. 1451 if (MemTransferInst *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 1452 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); 1453 DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n'); 1454 Constant *TheSrc = cast<Constant>(TheCopy->getSource()); 1455 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 1456 TheCopy->eraseFromParent(); // Don't mutate the global. 1457 AI->eraseFromParent(); 1458 ++NumGlobals; 1459 Changed = true; 1460 continue; 1461 } 1462 1463 // Check to see if we can perform the core SROA transformation. We cannot 1464 // transform the allocation instruction if it is an array allocation 1465 // (allocations OF arrays are ok though), and an allocation of a scalar 1466 // value cannot be decomposed at all. 1467 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 1468 1469 // Do not promote [0 x %struct]. 1470 if (AllocaSize == 0) continue; 1471 1472 // Do not promote any struct whose size is too big. 1473 if (AllocaSize > SRThreshold) continue; 1474 1475 // If the alloca looks like a good candidate for scalar replacement, and if 1476 // all its users can be transformed, then split up the aggregate into its 1477 // separate elements. 1478 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 1479 DoScalarReplacement(AI, WorkList); 1480 Changed = true; 1481 continue; 1482 } 1483 1484 // If we can turn this aggregate value (potentially with casts) into a 1485 // simple scalar value that can be mem2reg'd into a register value. 1486 // IsNotTrivial tracks whether this is something that mem2reg could have 1487 // promoted itself. If so, we don't want to transform it needlessly. Note 1488 // that we can't just check based on the type: the alloca may be of an i32 1489 // but that has pointer arithmetic to set byte 3 of it or something. 1490 if (AllocaInst *NewAI = 1491 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) { 1492 NewAI->takeName(AI); 1493 AI->eraseFromParent(); 1494 ++NumConverted; 1495 Changed = true; 1496 continue; 1497 } 1498 1499 // Otherwise, couldn't process this alloca. 1500 } 1501 1502 return Changed; 1503} 1504 1505/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 1506/// predicate, do SROA now. 1507void SROA::DoScalarReplacement(AllocaInst *AI, 1508 std::vector<AllocaInst*> &WorkList) { 1509 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 1510 SmallVector<AllocaInst*, 32> ElementAllocas; 1511 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 1512 ElementAllocas.reserve(ST->getNumContainedTypes()); 1513 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 1514 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 1515 AI->getAlignment(), 1516 AI->getName() + "." + Twine(i), AI); 1517 ElementAllocas.push_back(NA); 1518 WorkList.push_back(NA); // Add to worklist for recursive processing 1519 } 1520 } else { 1521 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 1522 ElementAllocas.reserve(AT->getNumElements()); 1523 const Type *ElTy = AT->getElementType(); 1524 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1525 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 1526 AI->getName() + "." + Twine(i), AI); 1527 ElementAllocas.push_back(NA); 1528 WorkList.push_back(NA); // Add to worklist for recursive processing 1529 } 1530 } 1531 1532 // Now that we have created the new alloca instructions, rewrite all the 1533 // uses of the old alloca. 1534 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 1535 1536 // Now erase any instructions that were made dead while rewriting the alloca. 1537 DeleteDeadInstructions(); 1538 AI->eraseFromParent(); 1539 1540 ++NumReplaced; 1541} 1542 1543/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 1544/// recursively including all their operands that become trivially dead. 1545void SROA::DeleteDeadInstructions() { 1546 while (!DeadInsts.empty()) { 1547 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 1548 1549 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 1550 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 1551 // Zero out the operand and see if it becomes trivially dead. 1552 // (But, don't add allocas to the dead instruction list -- they are 1553 // already on the worklist and will be deleted separately.) 1554 *OI = 0; 1555 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 1556 DeadInsts.push_back(U); 1557 } 1558 1559 I->eraseFromParent(); 1560 } 1561} 1562 1563/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 1564/// performing scalar replacement of alloca AI. The results are flagged in 1565/// the Info parameter. Offset indicates the position within AI that is 1566/// referenced by this instruction. 1567void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset, 1568 AllocaInfo &Info) { 1569 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1570 Instruction *User = cast<Instruction>(*UI); 1571 1572 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1573 isSafeForScalarRepl(BC, Offset, Info); 1574 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1575 uint64_t GEPOffset = Offset; 1576 isSafeGEP(GEPI, GEPOffset, Info); 1577 if (!Info.isUnsafe) 1578 isSafeForScalarRepl(GEPI, GEPOffset, Info); 1579 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1580 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1581 if (Length == 0) 1582 return MarkUnsafe(Info, User); 1583 isSafeMemAccess(Offset, Length->getZExtValue(), 0, 1584 UI.getOperandNo() == 0, Info, MI, 1585 true /*AllowWholeAccess*/); 1586 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1587 if (LI->isVolatile()) 1588 return MarkUnsafe(Info, User); 1589 const Type *LIType = LI->getType(); 1590 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1591 LIType, false, Info, LI, true /*AllowWholeAccess*/); 1592 Info.hasALoadOrStore = true; 1593 1594 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1595 // Store is ok if storing INTO the pointer, not storing the pointer 1596 if (SI->isVolatile() || SI->getOperand(0) == I) 1597 return MarkUnsafe(Info, User); 1598 1599 const Type *SIType = SI->getOperand(0)->getType(); 1600 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1601 SIType, true, Info, SI, true /*AllowWholeAccess*/); 1602 Info.hasALoadOrStore = true; 1603 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1604 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1605 } else { 1606 return MarkUnsafe(Info, User); 1607 } 1608 if (Info.isUnsafe) return; 1609 } 1610} 1611 1612 1613/// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer 1614/// derived from the alloca, we can often still split the alloca into elements. 1615/// This is useful if we have a large alloca where one element is phi'd 1616/// together somewhere: we can SRoA and promote all the other elements even if 1617/// we end up not being able to promote this one. 1618/// 1619/// All we require is that the uses of the PHI do not index into other parts of 1620/// the alloca. The most important use case for this is single load and stores 1621/// that are PHI'd together, which can happen due to code sinking. 1622void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset, 1623 AllocaInfo &Info) { 1624 // If we've already checked this PHI, don't do it again. 1625 if (PHINode *PN = dyn_cast<PHINode>(I)) 1626 if (!Info.CheckedPHIs.insert(PN)) 1627 return; 1628 1629 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1630 Instruction *User = cast<Instruction>(*UI); 1631 1632 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1633 isSafePHISelectUseForScalarRepl(BC, Offset, Info); 1634 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1635 // Only allow "bitcast" GEPs for simplicity. We could generalize this, 1636 // but would have to prove that we're staying inside of an element being 1637 // promoted. 1638 if (!GEPI->hasAllZeroIndices()) 1639 return MarkUnsafe(Info, User); 1640 isSafePHISelectUseForScalarRepl(GEPI, Offset, Info); 1641 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1642 if (LI->isVolatile()) 1643 return MarkUnsafe(Info, User); 1644 const Type *LIType = LI->getType(); 1645 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1646 LIType, false, Info, LI, false /*AllowWholeAccess*/); 1647 Info.hasALoadOrStore = true; 1648 1649 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1650 // Store is ok if storing INTO the pointer, not storing the pointer 1651 if (SI->isVolatile() || SI->getOperand(0) == I) 1652 return MarkUnsafe(Info, User); 1653 1654 const Type *SIType = SI->getOperand(0)->getType(); 1655 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1656 SIType, true, Info, SI, false /*AllowWholeAccess*/); 1657 Info.hasALoadOrStore = true; 1658 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1659 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1660 } else { 1661 return MarkUnsafe(Info, User); 1662 } 1663 if (Info.isUnsafe) return; 1664 } 1665} 1666 1667/// isSafeGEP - Check if a GEP instruction can be handled for scalar 1668/// replacement. It is safe when all the indices are constant, in-bounds 1669/// references, and when the resulting offset corresponds to an element within 1670/// the alloca type. The results are flagged in the Info parameter. Upon 1671/// return, Offset is adjusted as specified by the GEP indices. 1672void SROA::isSafeGEP(GetElementPtrInst *GEPI, 1673 uint64_t &Offset, AllocaInfo &Info) { 1674 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 1675 if (GEPIt == E) 1676 return; 1677 1678 // Walk through the GEP type indices, checking the types that this indexes 1679 // into. 1680 for (; GEPIt != E; ++GEPIt) { 1681 // Ignore struct elements, no extra checking needed for these. 1682 if ((*GEPIt)->isStructTy()) 1683 continue; 1684 1685 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 1686 if (!IdxVal) 1687 return MarkUnsafe(Info, GEPI); 1688 } 1689 1690 // Compute the offset due to this GEP and check if the alloca has a 1691 // component element at that offset. 1692 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1693 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1694 &Indices[0], Indices.size()); 1695 if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 0)) 1696 MarkUnsafe(Info, GEPI); 1697} 1698 1699/// isHomogeneousAggregate - Check if type T is a struct or array containing 1700/// elements of the same type (which is always true for arrays). If so, 1701/// return true with NumElts and EltTy set to the number of elements and the 1702/// element type, respectively. 1703static bool isHomogeneousAggregate(const Type *T, unsigned &NumElts, 1704 const Type *&EltTy) { 1705 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1706 NumElts = AT->getNumElements(); 1707 EltTy = (NumElts == 0 ? 0 : AT->getElementType()); 1708 return true; 1709 } 1710 if (const StructType *ST = dyn_cast<StructType>(T)) { 1711 NumElts = ST->getNumContainedTypes(); 1712 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0)); 1713 for (unsigned n = 1; n < NumElts; ++n) { 1714 if (ST->getContainedType(n) != EltTy) 1715 return false; 1716 } 1717 return true; 1718 } 1719 return false; 1720} 1721 1722/// isCompatibleAggregate - Check if T1 and T2 are either the same type or are 1723/// "homogeneous" aggregates with the same element type and number of elements. 1724static bool isCompatibleAggregate(const Type *T1, const Type *T2) { 1725 if (T1 == T2) 1726 return true; 1727 1728 unsigned NumElts1, NumElts2; 1729 const Type *EltTy1, *EltTy2; 1730 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) && 1731 isHomogeneousAggregate(T2, NumElts2, EltTy2) && 1732 NumElts1 == NumElts2 && 1733 EltTy1 == EltTy2) 1734 return true; 1735 1736 return false; 1737} 1738 1739/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1740/// alloca or has an offset and size that corresponds to a component element 1741/// within it. The offset checked here may have been formed from a GEP with a 1742/// pointer bitcasted to a different type. 1743/// 1744/// If AllowWholeAccess is true, then this allows uses of the entire alloca as a 1745/// unit. If false, it only allows accesses known to be in a single element. 1746void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 1747 const Type *MemOpType, bool isStore, 1748 AllocaInfo &Info, Instruction *TheAccess, 1749 bool AllowWholeAccess) { 1750 // Check if this is a load/store of the entire alloca. 1751 if (Offset == 0 && AllowWholeAccess && 1752 MemSize == TD->getTypeAllocSize(Info.AI->getAllocatedType())) { 1753 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer 1754 // loads/stores (which are essentially the same as the MemIntrinsics with 1755 // regard to copying padding between elements). But, if an alloca is 1756 // flagged as both a source and destination of such operations, we'll need 1757 // to check later for padding between elements. 1758 if (!MemOpType || MemOpType->isIntegerTy()) { 1759 if (isStore) 1760 Info.isMemCpyDst = true; 1761 else 1762 Info.isMemCpySrc = true; 1763 return; 1764 } 1765 // This is also safe for references using a type that is compatible with 1766 // the type of the alloca, so that loads/stores can be rewritten using 1767 // insertvalue/extractvalue. 1768 if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) { 1769 Info.hasSubelementAccess = true; 1770 return; 1771 } 1772 } 1773 // Check if the offset/size correspond to a component within the alloca type. 1774 const Type *T = Info.AI->getAllocatedType(); 1775 if (TypeHasComponent(T, Offset, MemSize)) { 1776 Info.hasSubelementAccess = true; 1777 return; 1778 } 1779 1780 return MarkUnsafe(Info, TheAccess); 1781} 1782 1783/// TypeHasComponent - Return true if T has a component type with the 1784/// specified offset and size. If Size is zero, do not check the size. 1785bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) { 1786 const Type *EltTy; 1787 uint64_t EltSize; 1788 if (const StructType *ST = dyn_cast<StructType>(T)) { 1789 const StructLayout *Layout = TD->getStructLayout(ST); 1790 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1791 EltTy = ST->getContainedType(EltIdx); 1792 EltSize = TD->getTypeAllocSize(EltTy); 1793 Offset -= Layout->getElementOffset(EltIdx); 1794 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1795 EltTy = AT->getElementType(); 1796 EltSize = TD->getTypeAllocSize(EltTy); 1797 if (Offset >= AT->getNumElements() * EltSize) 1798 return false; 1799 Offset %= EltSize; 1800 } else { 1801 return false; 1802 } 1803 if (Offset == 0 && (Size == 0 || EltSize == Size)) 1804 return true; 1805 // Check if the component spans multiple elements. 1806 if (Offset + Size > EltSize) 1807 return false; 1808 return TypeHasComponent(EltTy, Offset, Size); 1809} 1810 1811/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1812/// the instruction I, which references it, to use the separate elements. 1813/// Offset indicates the position within AI that is referenced by this 1814/// instruction. 1815void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1816 SmallVector<AllocaInst*, 32> &NewElts) { 1817 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) { 1818 Use &TheUse = UI.getUse(); 1819 Instruction *User = cast<Instruction>(*UI++); 1820 1821 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1822 RewriteBitCast(BC, AI, Offset, NewElts); 1823 continue; 1824 } 1825 1826 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1827 RewriteGEP(GEPI, AI, Offset, NewElts); 1828 continue; 1829 } 1830 1831 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1832 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1833 uint64_t MemSize = Length->getZExtValue(); 1834 if (Offset == 0 && 1835 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 1836 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1837 // Otherwise the intrinsic can only touch a single element and the 1838 // address operand will be updated, so nothing else needs to be done. 1839 continue; 1840 } 1841 1842 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1843 const Type *LIType = LI->getType(); 1844 1845 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) { 1846 // Replace: 1847 // %res = load { i32, i32 }* %alloc 1848 // with: 1849 // %load.0 = load i32* %alloc.0 1850 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1851 // %load.1 = load i32* %alloc.1 1852 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1853 // (Also works for arrays instead of structs) 1854 Value *Insert = UndefValue::get(LIType); 1855 IRBuilder<> Builder(LI); 1856 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1857 Value *Load = Builder.CreateLoad(NewElts[i], "load"); 1858 Insert = Builder.CreateInsertValue(Insert, Load, i, "insert"); 1859 } 1860 LI->replaceAllUsesWith(Insert); 1861 DeadInsts.push_back(LI); 1862 } else if (LIType->isIntegerTy() && 1863 TD->getTypeAllocSize(LIType) == 1864 TD->getTypeAllocSize(AI->getAllocatedType())) { 1865 // If this is a load of the entire alloca to an integer, rewrite it. 1866 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1867 } 1868 continue; 1869 } 1870 1871 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1872 Value *Val = SI->getOperand(0); 1873 const Type *SIType = Val->getType(); 1874 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) { 1875 // Replace: 1876 // store { i32, i32 } %val, { i32, i32 }* %alloc 1877 // with: 1878 // %val.0 = extractvalue { i32, i32 } %val, 0 1879 // store i32 %val.0, i32* %alloc.0 1880 // %val.1 = extractvalue { i32, i32 } %val, 1 1881 // store i32 %val.1, i32* %alloc.1 1882 // (Also works for arrays instead of structs) 1883 IRBuilder<> Builder(SI); 1884 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1885 Value *Extract = Builder.CreateExtractValue(Val, i, Val->getName()); 1886 Builder.CreateStore(Extract, NewElts[i]); 1887 } 1888 DeadInsts.push_back(SI); 1889 } else if (SIType->isIntegerTy() && 1890 TD->getTypeAllocSize(SIType) == 1891 TD->getTypeAllocSize(AI->getAllocatedType())) { 1892 // If this is a store of the entire alloca from an integer, rewrite it. 1893 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 1894 } 1895 continue; 1896 } 1897 1898 if (isa<SelectInst>(User) || isa<PHINode>(User)) { 1899 // If we have a PHI user of the alloca itself (as opposed to a GEP or 1900 // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to 1901 // the new pointer. 1902 if (!isa<AllocaInst>(I)) continue; 1903 1904 assert(Offset == 0 && NewElts[0] && 1905 "Direct alloca use should have a zero offset"); 1906 1907 // If we have a use of the alloca, we know the derived uses will be 1908 // utilizing just the first element of the scalarized result. Insert a 1909 // bitcast of the first alloca before the user as required. 1910 AllocaInst *NewAI = NewElts[0]; 1911 BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI); 1912 NewAI->moveBefore(BCI); 1913 TheUse = BCI; 1914 continue; 1915 } 1916 } 1917} 1918 1919/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 1920/// and recursively continue updating all of its uses. 1921void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 1922 SmallVector<AllocaInst*, 32> &NewElts) { 1923 RewriteForScalarRepl(BC, AI, Offset, NewElts); 1924 if (BC->getOperand(0) != AI) 1925 return; 1926 1927 // The bitcast references the original alloca. Replace its uses with 1928 // references to the first new element alloca. 1929 Instruction *Val = NewElts[0]; 1930 if (Val->getType() != BC->getDestTy()) { 1931 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 1932 Val->takeName(BC); 1933 } 1934 BC->replaceAllUsesWith(Val); 1935 DeadInsts.push_back(BC); 1936} 1937 1938/// FindElementAndOffset - Return the index of the element containing Offset 1939/// within the specified type, which must be either a struct or an array. 1940/// Sets T to the type of the element and Offset to the offset within that 1941/// element. IdxTy is set to the type of the index result to be used in a 1942/// GEP instruction. 1943uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset, 1944 const Type *&IdxTy) { 1945 uint64_t Idx = 0; 1946 if (const StructType *ST = dyn_cast<StructType>(T)) { 1947 const StructLayout *Layout = TD->getStructLayout(ST); 1948 Idx = Layout->getElementContainingOffset(Offset); 1949 T = ST->getContainedType(Idx); 1950 Offset -= Layout->getElementOffset(Idx); 1951 IdxTy = Type::getInt32Ty(T->getContext()); 1952 return Idx; 1953 } 1954 const ArrayType *AT = cast<ArrayType>(T); 1955 T = AT->getElementType(); 1956 uint64_t EltSize = TD->getTypeAllocSize(T); 1957 Idx = Offset / EltSize; 1958 Offset -= Idx * EltSize; 1959 IdxTy = Type::getInt64Ty(T->getContext()); 1960 return Idx; 1961} 1962 1963/// RewriteGEP - Check if this GEP instruction moves the pointer across 1964/// elements of the alloca that are being split apart, and if so, rewrite 1965/// the GEP to be relative to the new element. 1966void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 1967 SmallVector<AllocaInst*, 32> &NewElts) { 1968 uint64_t OldOffset = Offset; 1969 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1970 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1971 &Indices[0], Indices.size()); 1972 1973 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 1974 1975 const Type *T = AI->getAllocatedType(); 1976 const Type *IdxTy; 1977 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 1978 if (GEPI->getOperand(0) == AI) 1979 OldIdx = ~0ULL; // Force the GEP to be rewritten. 1980 1981 T = AI->getAllocatedType(); 1982 uint64_t EltOffset = Offset; 1983 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1984 1985 // If this GEP does not move the pointer across elements of the alloca 1986 // being split, then it does not needs to be rewritten. 1987 if (Idx == OldIdx) 1988 return; 1989 1990 const Type *i32Ty = Type::getInt32Ty(AI->getContext()); 1991 SmallVector<Value*, 8> NewArgs; 1992 NewArgs.push_back(Constant::getNullValue(i32Ty)); 1993 while (EltOffset != 0) { 1994 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 1995 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 1996 } 1997 Instruction *Val = NewElts[Idx]; 1998 if (NewArgs.size() > 1) { 1999 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(), 2000 NewArgs.end(), "", GEPI); 2001 Val->takeName(GEPI); 2002 } 2003 if (Val->getType() != GEPI->getType()) 2004 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 2005 GEPI->replaceAllUsesWith(Val); 2006 DeadInsts.push_back(GEPI); 2007} 2008 2009/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 2010/// Rewrite it to copy or set the elements of the scalarized memory. 2011void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 2012 AllocaInst *AI, 2013 SmallVector<AllocaInst*, 32> &NewElts) { 2014 // If this is a memcpy/memmove, construct the other pointer as the 2015 // appropriate type. The "Other" pointer is the pointer that goes to memory 2016 // that doesn't have anything to do with the alloca that we are promoting. For 2017 // memset, this Value* stays null. 2018 Value *OtherPtr = 0; 2019 unsigned MemAlignment = MI->getAlignment(); 2020 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 2021 if (Inst == MTI->getRawDest()) 2022 OtherPtr = MTI->getRawSource(); 2023 else { 2024 assert(Inst == MTI->getRawSource()); 2025 OtherPtr = MTI->getRawDest(); 2026 } 2027 } 2028 2029 // If there is an other pointer, we want to convert it to the same pointer 2030 // type as AI has, so we can GEP through it safely. 2031 if (OtherPtr) { 2032 unsigned AddrSpace = 2033 cast<PointerType>(OtherPtr->getType())->getAddressSpace(); 2034 2035 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 2036 // optimization, but it's also required to detect the corner case where 2037 // both pointer operands are referencing the same memory, and where 2038 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 2039 // function is only called for mem intrinsics that access the whole 2040 // aggregate, so non-zero GEPs are not an issue here.) 2041 OtherPtr = OtherPtr->stripPointerCasts(); 2042 2043 // Copying the alloca to itself is a no-op: just delete it. 2044 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 2045 // This code will run twice for a no-op memcpy -- once for each operand. 2046 // Put only one reference to MI on the DeadInsts list. 2047 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), 2048 E = DeadInsts.end(); I != E; ++I) 2049 if (*I == MI) return; 2050 DeadInsts.push_back(MI); 2051 return; 2052 } 2053 2054 // If the pointer is not the right type, insert a bitcast to the right 2055 // type. 2056 const Type *NewTy = 2057 PointerType::get(AI->getType()->getElementType(), AddrSpace); 2058 2059 if (OtherPtr->getType() != NewTy) 2060 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI); 2061 } 2062 2063 // Process each element of the aggregate. 2064 bool SROADest = MI->getRawDest() == Inst; 2065 2066 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 2067 2068 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2069 // If this is a memcpy/memmove, emit a GEP of the other element address. 2070 Value *OtherElt = 0; 2071 unsigned OtherEltAlign = MemAlignment; 2072 2073 if (OtherPtr) { 2074 Value *Idx[2] = { Zero, 2075 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 2076 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2, 2077 OtherPtr->getName()+"."+Twine(i), 2078 MI); 2079 uint64_t EltOffset; 2080 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 2081 const Type *OtherTy = OtherPtrTy->getElementType(); 2082 if (const StructType *ST = dyn_cast<StructType>(OtherTy)) { 2083 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 2084 } else { 2085 const Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); 2086 EltOffset = TD->getTypeAllocSize(EltTy)*i; 2087 } 2088 2089 // The alignment of the other pointer is the guaranteed alignment of the 2090 // element, which is affected by both the known alignment of the whole 2091 // mem intrinsic and the alignment of the element. If the alignment of 2092 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 2093 // known alignment is just 4 bytes. 2094 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 2095 } 2096 2097 Value *EltPtr = NewElts[i]; 2098 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 2099 2100 // If we got down to a scalar, insert a load or store as appropriate. 2101 if (EltTy->isSingleValueType()) { 2102 if (isa<MemTransferInst>(MI)) { 2103 if (SROADest) { 2104 // From Other to Alloca. 2105 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 2106 new StoreInst(Elt, EltPtr, MI); 2107 } else { 2108 // From Alloca to Other. 2109 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 2110 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 2111 } 2112 continue; 2113 } 2114 assert(isa<MemSetInst>(MI)); 2115 2116 // If the stored element is zero (common case), just store a null 2117 // constant. 2118 Constant *StoreVal; 2119 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) { 2120 if (CI->isZero()) { 2121 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 2122 } else { 2123 // If EltTy is a vector type, get the element type. 2124 const Type *ValTy = EltTy->getScalarType(); 2125 2126 // Construct an integer with the right value. 2127 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 2128 APInt OneVal(EltSize, CI->getZExtValue()); 2129 APInt TotalVal(OneVal); 2130 // Set each byte. 2131 for (unsigned i = 0; 8*i < EltSize; ++i) { 2132 TotalVal = TotalVal.shl(8); 2133 TotalVal |= OneVal; 2134 } 2135 2136 // Convert the integer value to the appropriate type. 2137 StoreVal = ConstantInt::get(CI->getContext(), TotalVal); 2138 if (ValTy->isPointerTy()) 2139 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 2140 else if (ValTy->isFloatingPointTy()) 2141 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 2142 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 2143 2144 // If the requested value was a vector constant, create it. 2145 if (EltTy != ValTy) { 2146 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 2147 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 2148 StoreVal = ConstantVector::get(Elts); 2149 } 2150 } 2151 new StoreInst(StoreVal, EltPtr, MI); 2152 continue; 2153 } 2154 // Otherwise, if we're storing a byte variable, use a memset call for 2155 // this element. 2156 } 2157 2158 unsigned EltSize = TD->getTypeAllocSize(EltTy); 2159 2160 IRBuilder<> Builder(MI); 2161 2162 // Finally, insert the meminst for this element. 2163 if (isa<MemSetInst>(MI)) { 2164 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize, 2165 MI->isVolatile()); 2166 } else { 2167 assert(isa<MemTransferInst>(MI)); 2168 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr 2169 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr 2170 2171 if (isa<MemCpyInst>(MI)) 2172 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile()); 2173 else 2174 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile()); 2175 } 2176 } 2177 DeadInsts.push_back(MI); 2178} 2179 2180/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 2181/// overwrites the entire allocation. Extract out the pieces of the stored 2182/// integer and store them individually. 2183void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 2184 SmallVector<AllocaInst*, 32> &NewElts){ 2185 // Extract each element out of the integer according to its structure offset 2186 // and store the element value to the individual alloca. 2187 Value *SrcVal = SI->getOperand(0); 2188 const Type *AllocaEltTy = AI->getAllocatedType(); 2189 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2190 2191 IRBuilder<> Builder(SI); 2192 2193 // Handle tail padding by extending the operand 2194 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 2195 SrcVal = Builder.CreateZExt(SrcVal, 2196 IntegerType::get(SI->getContext(), AllocaSizeBits)); 2197 2198 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 2199 << '\n'); 2200 2201 // There are two forms here: AI could be an array or struct. Both cases 2202 // have different ways to compute the element offset. 2203 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2204 const StructLayout *Layout = TD->getStructLayout(EltSTy); 2205 2206 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2207 // Get the number of bits to shift SrcVal to get the value. 2208 const Type *FieldTy = EltSTy->getElementType(i); 2209 uint64_t Shift = Layout->getElementOffsetInBits(i); 2210 2211 if (TD->isBigEndian()) 2212 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 2213 2214 Value *EltVal = SrcVal; 2215 if (Shift) { 2216 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2217 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2218 } 2219 2220 // Truncate down to an integer of the right size. 2221 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2222 2223 // Ignore zero sized fields like {}, they obviously contain no data. 2224 if (FieldSizeBits == 0) continue; 2225 2226 if (FieldSizeBits != AllocaSizeBits) 2227 EltVal = Builder.CreateTrunc(EltVal, 2228 IntegerType::get(SI->getContext(), FieldSizeBits)); 2229 Value *DestField = NewElts[i]; 2230 if (EltVal->getType() == FieldTy) { 2231 // Storing to an integer field of this size, just do it. 2232 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 2233 // Bitcast to the right element type (for fp/vector values). 2234 EltVal = Builder.CreateBitCast(EltVal, FieldTy); 2235 } else { 2236 // Otherwise, bitcast the dest pointer (for aggregates). 2237 DestField = Builder.CreateBitCast(DestField, 2238 PointerType::getUnqual(EltVal->getType())); 2239 } 2240 new StoreInst(EltVal, DestField, SI); 2241 } 2242 2243 } else { 2244 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 2245 const Type *ArrayEltTy = ATy->getElementType(); 2246 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2247 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 2248 2249 uint64_t Shift; 2250 2251 if (TD->isBigEndian()) 2252 Shift = AllocaSizeBits-ElementOffset; 2253 else 2254 Shift = 0; 2255 2256 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2257 // Ignore zero sized fields like {}, they obviously contain no data. 2258 if (ElementSizeBits == 0) continue; 2259 2260 Value *EltVal = SrcVal; 2261 if (Shift) { 2262 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2263 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2264 } 2265 2266 // Truncate down to an integer of the right size. 2267 if (ElementSizeBits != AllocaSizeBits) 2268 EltVal = Builder.CreateTrunc(EltVal, 2269 IntegerType::get(SI->getContext(), 2270 ElementSizeBits)); 2271 Value *DestField = NewElts[i]; 2272 if (EltVal->getType() == ArrayEltTy) { 2273 // Storing to an integer field of this size, just do it. 2274 } else if (ArrayEltTy->isFloatingPointTy() || 2275 ArrayEltTy->isVectorTy()) { 2276 // Bitcast to the right element type (for fp/vector values). 2277 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy); 2278 } else { 2279 // Otherwise, bitcast the dest pointer (for aggregates). 2280 DestField = Builder.CreateBitCast(DestField, 2281 PointerType::getUnqual(EltVal->getType())); 2282 } 2283 new StoreInst(EltVal, DestField, SI); 2284 2285 if (TD->isBigEndian()) 2286 Shift -= ElementOffset; 2287 else 2288 Shift += ElementOffset; 2289 } 2290 } 2291 2292 DeadInsts.push_back(SI); 2293} 2294 2295/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 2296/// an integer. Load the individual pieces to form the aggregate value. 2297void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 2298 SmallVector<AllocaInst*, 32> &NewElts) { 2299 // Extract each element out of the NewElts according to its structure offset 2300 // and form the result value. 2301 const Type *AllocaEltTy = AI->getAllocatedType(); 2302 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2303 2304 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 2305 << '\n'); 2306 2307 // There are two forms here: AI could be an array or struct. Both cases 2308 // have different ways to compute the element offset. 2309 const StructLayout *Layout = 0; 2310 uint64_t ArrayEltBitOffset = 0; 2311 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2312 Layout = TD->getStructLayout(EltSTy); 2313 } else { 2314 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 2315 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2316 } 2317 2318 Value *ResultVal = 2319 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 2320 2321 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2322 // Load the value from the alloca. If the NewElt is an aggregate, cast 2323 // the pointer to an integer of the same size before doing the load. 2324 Value *SrcField = NewElts[i]; 2325 const Type *FieldTy = 2326 cast<PointerType>(SrcField->getType())->getElementType(); 2327 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2328 2329 // Ignore zero sized fields like {}, they obviously contain no data. 2330 if (FieldSizeBits == 0) continue; 2331 2332 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 2333 FieldSizeBits); 2334 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 2335 !FieldTy->isVectorTy()) 2336 SrcField = new BitCastInst(SrcField, 2337 PointerType::getUnqual(FieldIntTy), 2338 "", LI); 2339 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 2340 2341 // If SrcField is a fp or vector of the right size but that isn't an 2342 // integer type, bitcast to an integer so we can shift it. 2343 if (SrcField->getType() != FieldIntTy) 2344 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 2345 2346 // Zero extend the field to be the same size as the final alloca so that 2347 // we can shift and insert it. 2348 if (SrcField->getType() != ResultVal->getType()) 2349 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 2350 2351 // Determine the number of bits to shift SrcField. 2352 uint64_t Shift; 2353 if (Layout) // Struct case. 2354 Shift = Layout->getElementOffsetInBits(i); 2355 else // Array case. 2356 Shift = i*ArrayEltBitOffset; 2357 2358 if (TD->isBigEndian()) 2359 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 2360 2361 if (Shift) { 2362 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 2363 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 2364 } 2365 2366 // Don't create an 'or x, 0' on the first iteration. 2367 if (!isa<Constant>(ResultVal) || 2368 !cast<Constant>(ResultVal)->isNullValue()) 2369 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 2370 else 2371 ResultVal = SrcField; 2372 } 2373 2374 // Handle tail padding by truncating the result 2375 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 2376 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 2377 2378 LI->replaceAllUsesWith(ResultVal); 2379 DeadInsts.push_back(LI); 2380} 2381 2382/// HasPadding - Return true if the specified type has any structure or 2383/// alignment padding in between the elements that would be split apart 2384/// by SROA; return false otherwise. 2385static bool HasPadding(const Type *Ty, const TargetData &TD) { 2386 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 2387 Ty = ATy->getElementType(); 2388 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 2389 } 2390 2391 // SROA currently handles only Arrays and Structs. 2392 const StructType *STy = cast<StructType>(Ty); 2393 const StructLayout *SL = TD.getStructLayout(STy); 2394 unsigned PrevFieldBitOffset = 0; 2395 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 2396 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 2397 2398 // Check to see if there is any padding between this element and the 2399 // previous one. 2400 if (i) { 2401 unsigned PrevFieldEnd = 2402 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 2403 if (PrevFieldEnd < FieldBitOffset) 2404 return true; 2405 } 2406 PrevFieldBitOffset = FieldBitOffset; 2407 } 2408 // Check for tail padding. 2409 if (unsigned EltCount = STy->getNumElements()) { 2410 unsigned PrevFieldEnd = PrevFieldBitOffset + 2411 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 2412 if (PrevFieldEnd < SL->getSizeInBits()) 2413 return true; 2414 } 2415 return false; 2416} 2417 2418/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 2419/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 2420/// or 1 if safe after canonicalization has been performed. 2421bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 2422 // Loop over the use list of the alloca. We can only transform it if all of 2423 // the users are safe to transform. 2424 AllocaInfo Info(AI); 2425 2426 isSafeForScalarRepl(AI, 0, Info); 2427 if (Info.isUnsafe) { 2428 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 2429 return false; 2430 } 2431 2432 // Okay, we know all the users are promotable. If the aggregate is a memcpy 2433 // source and destination, we have to be careful. In particular, the memcpy 2434 // could be moving around elements that live in structure padding of the LLVM 2435 // types, but may actually be used. In these cases, we refuse to promote the 2436 // struct. 2437 if (Info.isMemCpySrc && Info.isMemCpyDst && 2438 HasPadding(AI->getAllocatedType(), *TD)) 2439 return false; 2440 2441 // If the alloca never has an access to just *part* of it, but is accessed 2442 // via loads and stores, then we should use ConvertToScalarInfo to promote 2443 // the alloca instead of promoting each piece at a time and inserting fission 2444 // and fusion code. 2445 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) { 2446 // If the struct/array just has one element, use basic SRoA. 2447 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 2448 if (ST->getNumElements() > 1) return false; 2449 } else { 2450 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1) 2451 return false; 2452 } 2453 } 2454 2455 return true; 2456} 2457 2458 2459 2460/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 2461/// some part of a constant global variable. This intentionally only accepts 2462/// constant expressions because we don't can't rewrite arbitrary instructions. 2463static bool PointsToConstantGlobal(Value *V) { 2464 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 2465 return GV->isConstant(); 2466 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 2467 if (CE->getOpcode() == Instruction::BitCast || 2468 CE->getOpcode() == Instruction::GetElementPtr) 2469 return PointsToConstantGlobal(CE->getOperand(0)); 2470 return false; 2471} 2472 2473/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 2474/// pointer to an alloca. Ignore any reads of the pointer, return false if we 2475/// see any stores or other unknown uses. If we see pointer arithmetic, keep 2476/// track of whether it moves the pointer (with isOffset) but otherwise traverse 2477/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 2478/// the alloca, and if the source pointer is a pointer to a constant global, we 2479/// can optimize this. 2480static bool isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, 2481 bool isOffset) { 2482 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 2483 User *U = cast<Instruction>(*UI); 2484 2485 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 2486 // Ignore non-volatile loads, they are always ok. 2487 if (LI->isVolatile()) return false; 2488 continue; 2489 } 2490 2491 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 2492 // If uses of the bitcast are ok, we are ok. 2493 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 2494 return false; 2495 continue; 2496 } 2497 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 2498 // If the GEP has all zero indices, it doesn't offset the pointer. If it 2499 // doesn't, it does. 2500 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 2501 isOffset || !GEP->hasAllZeroIndices())) 2502 return false; 2503 continue; 2504 } 2505 2506 if (CallSite CS = U) { 2507 // If this is the function being called then we treat it like a load and 2508 // ignore it. 2509 if (CS.isCallee(UI)) 2510 continue; 2511 2512 // If this is a readonly/readnone call site, then we know it is just a 2513 // load (but one that potentially returns the value itself), so we can 2514 // ignore it if we know that the value isn't captured. 2515 unsigned ArgNo = CS.getArgumentNo(UI); 2516 if (CS.onlyReadsMemory() && 2517 (CS.getInstruction()->use_empty() || 2518 CS.paramHasAttr(ArgNo+1, Attribute::NoCapture))) 2519 continue; 2520 2521 // If this is being passed as a byval argument, the caller is making a 2522 // copy, so it is only a read of the alloca. 2523 if (CS.paramHasAttr(ArgNo+1, Attribute::ByVal)) 2524 continue; 2525 } 2526 2527 // If this is isn't our memcpy/memmove, reject it as something we can't 2528 // handle. 2529 MemTransferInst *MI = dyn_cast<MemTransferInst>(U); 2530 if (MI == 0) 2531 return false; 2532 2533 // If the transfer is using the alloca as a source of the transfer, then 2534 // ignore it since it is a load (unless the transfer is volatile). 2535 if (UI.getOperandNo() == 1) { 2536 if (MI->isVolatile()) return false; 2537 continue; 2538 } 2539 2540 // If we already have seen a copy, reject the second one. 2541 if (TheCopy) return false; 2542 2543 // If the pointer has been offset from the start of the alloca, we can't 2544 // safely handle this. 2545 if (isOffset) return false; 2546 2547 // If the memintrinsic isn't using the alloca as the dest, reject it. 2548 if (UI.getOperandNo() != 0) return false; 2549 2550 // If the source of the memcpy/move is not a constant global, reject it. 2551 if (!PointsToConstantGlobal(MI->getSource())) 2552 return false; 2553 2554 // Otherwise, the transform is safe. Remember the copy instruction. 2555 TheCopy = MI; 2556 } 2557 return true; 2558} 2559 2560/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 2561/// modified by a copy from a constant global. If we can prove this, we can 2562/// replace any uses of the alloca with uses of the global directly. 2563MemTransferInst *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) { 2564 MemTransferInst *TheCopy = 0; 2565 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 2566 return TheCopy; 2567 return 0; 2568} 2569