ScalarReplAggregates.cpp revision 6be41eb7f00319f5ffa1a5435dcd1e81b3ce932d
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, 235 236 // Accesses via GEPs that are consistent with element access of a vector 237 // type. This will not be converted into a vector unless there is a later 238 // access using an actual vector type. 239 ImplicitVector, 240 241 // Accesses via vector operations and GEPs that are consistent with the 242 // layout of a vector type. 243 Vector, 244 245 // An integer bag-of-bits with bitwise operations for insertion and 246 // extraction. Any combination of types can be converted into this kind 247 // of scalar. 248 Integer 249 } ScalarKind; 250 251 /// VectorTy - This tracks the type that we should promote the vector to if 252 /// it is possible to turn it into a vector. This starts out null, and if it 253 /// isn't possible to turn into a vector type, it gets set to VoidTy. 254 const VectorType *VectorTy; 255 256 /// HadNonMemTransferAccess - True if there is at least one access to the 257 /// alloca that is not a MemTransferInst. We don't want to turn structs into 258 /// large integers unless there is some potential for optimization. 259 bool HadNonMemTransferAccess; 260 261public: 262 explicit ConvertToScalarInfo(unsigned Size, const TargetData &td) 263 : AllocaSize(Size), TD(td), IsNotTrivial(false), ScalarKind(Unknown), 264 VectorTy(0), HadNonMemTransferAccess(false) { } 265 266 AllocaInst *TryConvert(AllocaInst *AI); 267 268private: 269 bool CanConvertToScalar(Value *V, uint64_t Offset); 270 void MergeInTypeForLoadOrStore(const Type *In, uint64_t Offset); 271 bool MergeInVectorType(const VectorType *VInTy, uint64_t Offset); 272 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); 273 274 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType, 275 uint64_t Offset, IRBuilder<> &Builder); 276 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 277 uint64_t Offset, IRBuilder<> &Builder); 278}; 279} // end anonymous namespace. 280 281 282/// TryConvert - Analyze the specified alloca, and if it is safe to do so, 283/// rewrite it to be a new alloca which is mem2reg'able. This returns the new 284/// alloca if possible or null if not. 285AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) { 286 // If we can't convert this scalar, or if mem2reg can trivially do it, bail 287 // out. 288 if (!CanConvertToScalar(AI, 0) || !IsNotTrivial) 289 return 0; 290 291 // If an alloca has only memset / memcpy uses, it may still have an Unknown 292 // ScalarKind. Treat it as an Integer below. 293 if (ScalarKind == Unknown) 294 ScalarKind = Integer; 295 296 // If we were able to find a vector type that can handle this with 297 // insert/extract elements, and if there was at least one use that had 298 // a vector type, promote this to a vector. We don't want to promote 299 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 300 // we just get a lot of insert/extracts. If at least one vector is 301 // involved, then we probably really do have a union of vector/array. 302 const Type *NewTy; 303 if (ScalarKind == Vector) { 304 assert(VectorTy && "Missing type for vector scalar."); 305 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = " 306 << *VectorTy << '\n'); 307 NewTy = VectorTy; // Use the vector type. 308 } else { 309 unsigned BitWidth = AllocaSize * 8; 310 if ((ScalarKind == ImplicitVector || ScalarKind == Integer) && 311 !HadNonMemTransferAccess && !TD.fitsInLegalInteger(BitWidth)) 312 return 0; 313 314 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"); 315 // Create and insert the integer alloca. 316 NewTy = IntegerType::get(AI->getContext(), BitWidth); 317 } 318 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 319 ConvertUsesToScalar(AI, NewAI, 0); 320 return NewAI; 321} 322 323/// MergeInTypeForLoadOrStore - Add the 'In' type to the accumulated vector type 324/// (VectorTy) so far at the offset specified by Offset (which is specified in 325/// bytes). 326/// 327/// There are three cases we handle here: 328/// 1) A union of vector types of the same size and potentially its elements. 329/// Here we turn element accesses into insert/extract element operations. 330/// This promotes a <4 x float> with a store of float to the third element 331/// into a <4 x float> that uses insert element. 332/// 2) A union of vector types with power-of-2 size differences, e.g. a float, 333/// <2 x float> and <4 x float>. Here we turn element accesses into insert 334/// and extract element operations, and <2 x float> accesses into a cast to 335/// <2 x double>, an extract, and a cast back to <2 x float>. 336/// 3) A fully general blob of memory, which we turn into some (potentially 337/// large) integer type with extract and insert operations where the loads 338/// and stores would mutate the memory. We mark this by setting VectorTy 339/// to VoidTy. 340void ConvertToScalarInfo::MergeInTypeForLoadOrStore(const Type *In, 341 uint64_t Offset) { 342 // If we already decided to turn this into a blob of integer memory, there is 343 // nothing to be done. 344 if (ScalarKind == Integer) 345 return; 346 347 // If this could be contributing to a vector, analyze it. 348 349 // If the In type is a vector that is the same size as the alloca, see if it 350 // matches the existing VecTy. 351 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 352 if (MergeInVectorType(VInTy, Offset)) 353 return; 354 } else if (In->isFloatTy() || In->isDoubleTy() || 355 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 && 356 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 357 // Full width accesses can be ignored, because they can always be turned 358 // into bitcasts. 359 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 360 if (EltSize == AllocaSize) 361 return; 362 363 // If we're accessing something that could be an element of a vector, see 364 // if the implied vector agrees with what we already have and if Offset is 365 // compatible with it. 366 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 && 367 (!VectorTy || Offset * 8 < VectorTy->getPrimitiveSizeInBits())) { 368 if (!VectorTy) { 369 ScalarKind = ImplicitVector; 370 VectorTy = VectorType::get(In, AllocaSize/EltSize); 371 return; 372 } 373 374 unsigned CurrentEltSize = VectorTy->getElementType() 375 ->getPrimitiveSizeInBits()/8; 376 if (EltSize == CurrentEltSize) 377 return; 378 379 if (In->isIntegerTy() && isPowerOf2_32(AllocaSize / EltSize)) 380 return; 381 } 382 } 383 384 // Otherwise, we have a case that we can't handle with an optimized vector 385 // form. We can still turn this into a large integer. 386 ScalarKind = Integer; 387} 388 389/// MergeInVectorType - Handles the vector case of MergeInTypeForLoadOrStore, 390/// returning true if the type was successfully merged and false otherwise. 391bool ConvertToScalarInfo::MergeInVectorType(const VectorType *VInTy, 392 uint64_t Offset) { 393 // TODO: Support nonzero offsets? 394 if (Offset != 0) 395 return false; 396 397 // Only allow vectors that are a power-of-2 away from the size of the alloca. 398 if (!isPowerOf2_64(AllocaSize / (VInTy->getBitWidth() / 8))) 399 return false; 400 401 // If this the first vector we see, remember the type so that we know the 402 // element size. 403 if (!VectorTy) { 404 ScalarKind = Vector; 405 VectorTy = VInTy; 406 return true; 407 } 408 409 unsigned BitWidth = VectorTy->getBitWidth(); 410 unsigned InBitWidth = VInTy->getBitWidth(); 411 412 // Vectors of the same size can be converted using a simple bitcast. 413 if (InBitWidth == BitWidth && AllocaSize == (InBitWidth / 8)) { 414 ScalarKind = Vector; 415 return true; 416 } 417 418 const Type *ElementTy = VectorTy->getElementType(); 419 const Type *InElementTy = VInTy->getElementType(); 420 421 // Do not allow mixed integer and floating-point accesses from vectors of 422 // different sizes. 423 if (ElementTy->isFloatingPointTy() != InElementTy->isFloatingPointTy()) 424 return false; 425 426 if (ElementTy->isFloatingPointTy()) { 427 // Only allow floating-point vectors of different sizes if they have the 428 // same element type. 429 // TODO: This could be loosened a bit, but would anything benefit? 430 if (ElementTy != InElementTy) 431 return false; 432 433 // There are no arbitrary-precision floating-point types, which limits the 434 // number of legal vector types with larger element types that we can form 435 // to bitcast and extract a subvector. 436 // TODO: We could support some more cases with mixed fp128 and double here. 437 if (!(BitWidth == 64 || BitWidth == 128) || 438 !(InBitWidth == 64 || InBitWidth == 128)) 439 return false; 440 } else { 441 assert(ElementTy->isIntegerTy() && "Vector elements must be either integer " 442 "or floating-point."); 443 unsigned BitWidth = ElementTy->getPrimitiveSizeInBits(); 444 unsigned InBitWidth = InElementTy->getPrimitiveSizeInBits(); 445 446 // Do not allow integer types smaller than a byte or types whose widths are 447 // not a multiple of a byte. 448 if (BitWidth < 8 || InBitWidth < 8 || 449 BitWidth % 8 != 0 || InBitWidth % 8 != 0) 450 return false; 451 } 452 453 // Pick the largest of the two vector types. 454 ScalarKind = Vector; 455 if (InBitWidth > BitWidth) 456 VectorTy = VInTy; 457 458 return true; 459} 460 461/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 462/// its accesses to a single vector type, return true and set VecTy to 463/// the new type. If we could convert the alloca into a single promotable 464/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 465/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 466/// is the current offset from the base of the alloca being analyzed. 467/// 468/// If we see at least one access to the value that is as a vector type, set the 469/// SawVec flag. 470bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) { 471 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 472 Instruction *User = cast<Instruction>(*UI); 473 474 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 475 // Don't break volatile loads. 476 if (LI->isVolatile()) 477 return false; 478 // Don't touch MMX operations. 479 if (LI->getType()->isX86_MMXTy()) 480 return false; 481 HadNonMemTransferAccess = true; 482 MergeInTypeForLoadOrStore(LI->getType(), Offset); 483 continue; 484 } 485 486 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 487 // Storing the pointer, not into the value? 488 if (SI->getOperand(0) == V || SI->isVolatile()) return false; 489 // Don't touch MMX operations. 490 if (SI->getOperand(0)->getType()->isX86_MMXTy()) 491 return false; 492 HadNonMemTransferAccess = true; 493 MergeInTypeForLoadOrStore(SI->getOperand(0)->getType(), Offset); 494 continue; 495 } 496 497 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 498 IsNotTrivial = true; // Can't be mem2reg'd. 499 if (!CanConvertToScalar(BCI, Offset)) 500 return false; 501 continue; 502 } 503 504 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 505 // If this is a GEP with a variable indices, we can't handle it. 506 if (!GEP->hasAllConstantIndices()) 507 return false; 508 509 // Compute the offset that this GEP adds to the pointer. 510 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 511 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 512 &Indices[0], Indices.size()); 513 // See if all uses can be converted. 514 if (!CanConvertToScalar(GEP, Offset+GEPOffset)) 515 return false; 516 IsNotTrivial = true; // Can't be mem2reg'd. 517 HadNonMemTransferAccess = true; 518 continue; 519 } 520 521 // If this is a constant sized memset of a constant value (e.g. 0) we can 522 // handle it. 523 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 524 // Store of constant value. 525 if (!isa<ConstantInt>(MSI->getValue())) 526 return false; 527 528 // Store of constant size. 529 ConstantInt *Len = dyn_cast<ConstantInt>(MSI->getLength()); 530 if (!Len) 531 return false; 532 533 // If the size differs from the alloca, we can only convert the alloca to 534 // an integer bag-of-bits. 535 // FIXME: This should handle all of the cases that are currently accepted 536 // as vector element insertions. 537 if (Len->getZExtValue() != AllocaSize || Offset != 0) 538 ScalarKind = Integer; 539 540 IsNotTrivial = true; // Can't be mem2reg'd. 541 HadNonMemTransferAccess = true; 542 continue; 543 } 544 545 // If this is a memcpy or memmove into or out of the whole allocation, we 546 // can handle it like a load or store of the scalar type. 547 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 548 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()); 549 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0) 550 return false; 551 552 IsNotTrivial = true; // Can't be mem2reg'd. 553 continue; 554 } 555 556 // Otherwise, we cannot handle this! 557 return false; 558 } 559 560 return true; 561} 562 563/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 564/// directly. This happens when we are converting an "integer union" to a 565/// single integer scalar, or when we are converting a "vector union" to a 566/// vector with insert/extractelement instructions. 567/// 568/// Offset is an offset from the original alloca, in bits that need to be 569/// shifted to the right. By the end of this, there should be no uses of Ptr. 570void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, 571 uint64_t Offset) { 572 while (!Ptr->use_empty()) { 573 Instruction *User = cast<Instruction>(Ptr->use_back()); 574 575 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 576 ConvertUsesToScalar(CI, NewAI, Offset); 577 CI->eraseFromParent(); 578 continue; 579 } 580 581 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 582 // Compute the offset that this GEP adds to the pointer. 583 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 584 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 585 &Indices[0], Indices.size()); 586 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 587 GEP->eraseFromParent(); 588 continue; 589 } 590 591 IRBuilder<> Builder(User); 592 593 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 594 // The load is a bit extract from NewAI shifted right by Offset bits. 595 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 596 Value *NewLoadVal 597 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 598 LI->replaceAllUsesWith(NewLoadVal); 599 LI->eraseFromParent(); 600 continue; 601 } 602 603 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 604 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 605 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 606 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 607 Builder); 608 Builder.CreateStore(New, NewAI); 609 SI->eraseFromParent(); 610 611 // If the load we just inserted is now dead, then the inserted store 612 // overwrote the entire thing. 613 if (Old->use_empty()) 614 Old->eraseFromParent(); 615 continue; 616 } 617 618 // If this is a constant sized memset of a constant value (e.g. 0) we can 619 // transform it into a store of the expanded constant value. 620 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 621 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 622 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 623 if (NumBytes != 0) { 624 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 625 626 // Compute the value replicated the right number of times. 627 APInt APVal(NumBytes*8, Val); 628 629 // Splat the value if non-zero. 630 if (Val) 631 for (unsigned i = 1; i != NumBytes; ++i) 632 APVal |= APVal << 8; 633 634 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 635 Value *New = ConvertScalar_InsertValue( 636 ConstantInt::get(User->getContext(), APVal), 637 Old, Offset, Builder); 638 Builder.CreateStore(New, NewAI); 639 640 // If the load we just inserted is now dead, then the memset overwrote 641 // the entire thing. 642 if (Old->use_empty()) 643 Old->eraseFromParent(); 644 } 645 MSI->eraseFromParent(); 646 continue; 647 } 648 649 // If this is a memcpy or memmove into or out of the whole allocation, we 650 // can handle it like a load or store of the scalar type. 651 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 652 assert(Offset == 0 && "must be store to start of alloca"); 653 654 // If the source and destination are both to the same alloca, then this is 655 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 656 // as appropriate. 657 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, &TD, 0)); 658 659 if (GetUnderlyingObject(MTI->getSource(), &TD, 0) != OrigAI) { 660 // Dest must be OrigAI, change this to be a load from the original 661 // pointer (bitcasted), then a store to our new alloca. 662 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 663 Value *SrcPtr = MTI->getSource(); 664 const PointerType* SPTy = cast<PointerType>(SrcPtr->getType()); 665 const PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 666 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 667 AIPTy = PointerType::get(AIPTy->getElementType(), 668 SPTy->getAddressSpace()); 669 } 670 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy); 671 672 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 673 SrcVal->setAlignment(MTI->getAlignment()); 674 Builder.CreateStore(SrcVal, NewAI); 675 } else if (GetUnderlyingObject(MTI->getDest(), &TD, 0) != OrigAI) { 676 // Src must be OrigAI, change this to be a load from NewAI then a store 677 // through the original dest pointer (bitcasted). 678 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 679 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 680 681 const PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType()); 682 const PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 683 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 684 AIPTy = PointerType::get(AIPTy->getElementType(), 685 DPTy->getAddressSpace()); 686 } 687 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy); 688 689 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 690 NewStore->setAlignment(MTI->getAlignment()); 691 } else { 692 // Noop transfer. Src == Dst 693 } 694 695 MTI->eraseFromParent(); 696 continue; 697 } 698 699 llvm_unreachable("Unsupported operation!"); 700 } 701} 702 703/// getScaledElementType - Gets a scaled element type for a partial vector 704/// access of an alloca. The input types must be integer or floating-point 705/// scalar or vector types, and the resulting type is an integer, float or 706/// double. 707static const Type *getScaledElementType(const Type *Ty1, const Type *Ty2, 708 unsigned NewBitWidth) { 709 bool IsFP1 = Ty1->isFloatingPointTy() || 710 (Ty1->isVectorTy() && 711 cast<VectorType>(Ty1)->getElementType()->isFloatingPointTy()); 712 bool IsFP2 = Ty2->isFloatingPointTy() || 713 (Ty2->isVectorTy() && 714 cast<VectorType>(Ty2)->getElementType()->isFloatingPointTy()); 715 716 LLVMContext &Context = Ty1->getContext(); 717 718 // Prefer floating-point types over integer types, as integer types may have 719 // been created by earlier scalar replacement. 720 if (IsFP1 || IsFP2) { 721 if (NewBitWidth == 32) 722 return Type::getFloatTy(Context); 723 if (NewBitWidth == 64) 724 return Type::getDoubleTy(Context); 725 } 726 727 return Type::getIntNTy(Context, NewBitWidth); 728} 729 730/// CreateShuffleVectorCast - Creates a shuffle vector to convert one vector 731/// to another vector of the same element type which has the same allocation 732/// size but different primitive sizes (e.g. <3 x i32> and <4 x i32>). 733static Value *CreateShuffleVectorCast(Value *FromVal, const Type *ToType, 734 IRBuilder<> &Builder) { 735 const Type *FromType = FromVal->getType(); 736 const VectorType *FromVTy = cast<VectorType>(FromType); 737 const VectorType *ToVTy = cast<VectorType>(ToType); 738 assert((ToVTy->getElementType() == FromVTy->getElementType()) && 739 "Vectors must have the same element type"); 740 Value *UnV = UndefValue::get(FromType); 741 unsigned numEltsFrom = FromVTy->getNumElements(); 742 unsigned numEltsTo = ToVTy->getNumElements(); 743 744 SmallVector<Constant*, 3> Args; 745 const Type* Int32Ty = Builder.getInt32Ty(); 746 unsigned minNumElts = std::min(numEltsFrom, numEltsTo); 747 unsigned i; 748 for (i=0; i != minNumElts; ++i) 749 Args.push_back(ConstantInt::get(Int32Ty, i)); 750 751 if (i < numEltsTo) { 752 Constant* UnC = UndefValue::get(Int32Ty); 753 for (; i != numEltsTo; ++i) 754 Args.push_back(UnC); 755 } 756 Constant *Mask = ConstantVector::get(Args); 757 return Builder.CreateShuffleVector(FromVal, UnV, Mask, "tmpV"); 758} 759 760/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 761/// or vector value FromVal, extracting the bits from the offset specified by 762/// Offset. This returns the value, which is of type ToType. 763/// 764/// This happens when we are converting an "integer union" to a single 765/// integer scalar, or when we are converting a "vector union" to a vector with 766/// insert/extractelement instructions. 767/// 768/// Offset is an offset from the original alloca, in bits that need to be 769/// shifted to the right. 770Value *ConvertToScalarInfo:: 771ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 772 uint64_t Offset, IRBuilder<> &Builder) { 773 // If the load is of the whole new alloca, no conversion is needed. 774 const Type *FromType = FromVal->getType(); 775 if (FromType == ToType && Offset == 0) 776 return FromVal; 777 778 // If the result alloca is a vector type, this is either an element 779 // access or a bitcast to another vector type of the same size. 780 if (const VectorType *VTy = dyn_cast<VectorType>(FromType)) { 781 unsigned FromTypeSize = TD.getTypeAllocSize(FromType); 782 unsigned ToTypeSize = TD.getTypeAllocSize(ToType); 783 if (FromTypeSize == ToTypeSize) { 784 // If the two types have the same primitive size, use a bit cast. 785 // Otherwise, it is two vectors with the same element type that has 786 // the same allocation size but different number of elements so use 787 // a shuffle vector. 788 if (FromType->getPrimitiveSizeInBits() == 789 ToType->getPrimitiveSizeInBits()) 790 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 791 else 792 return CreateShuffleVectorCast(FromVal, ToType, Builder); 793 } 794 795 if (isPowerOf2_64(FromTypeSize / ToTypeSize)) { 796 assert(!(ToType->isVectorTy() && Offset != 0) && "Can't extract a value " 797 "of a smaller vector type at a nonzero offset."); 798 799 const Type *CastElementTy = getScaledElementType(FromType, ToType, 800 ToTypeSize * 8); 801 unsigned NumCastVectorElements = FromTypeSize / ToTypeSize; 802 803 LLVMContext &Context = FromVal->getContext(); 804 const Type *CastTy = VectorType::get(CastElementTy, 805 NumCastVectorElements); 806 Value *Cast = Builder.CreateBitCast(FromVal, CastTy, "tmp"); 807 808 unsigned EltSize = TD.getTypeAllocSizeInBits(CastElementTy); 809 unsigned Elt = Offset/EltSize; 810 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 811 Value *Extract = Builder.CreateExtractElement(Cast, ConstantInt::get( 812 Type::getInt32Ty(Context), Elt), "tmp"); 813 return Builder.CreateBitCast(Extract, ToType, "tmp"); 814 } 815 816 // Otherwise it must be an element access. 817 unsigned Elt = 0; 818 if (Offset) { 819 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 820 Elt = Offset/EltSize; 821 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 822 } 823 // Return the element extracted out of it. 824 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get( 825 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp"); 826 if (V->getType() != ToType) 827 V = Builder.CreateBitCast(V, ToType, "tmp"); 828 return V; 829 } 830 831 // If ToType is a first class aggregate, extract out each of the pieces and 832 // use insertvalue's to form the FCA. 833 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 834 const StructLayout &Layout = *TD.getStructLayout(ST); 835 Value *Res = UndefValue::get(ST); 836 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 837 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 838 Offset+Layout.getElementOffsetInBits(i), 839 Builder); 840 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 841 } 842 return Res; 843 } 844 845 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 846 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 847 Value *Res = UndefValue::get(AT); 848 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 849 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 850 Offset+i*EltSize, Builder); 851 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 852 } 853 return Res; 854 } 855 856 // Otherwise, this must be a union that was converted to an integer value. 857 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 858 859 // If this is a big-endian system and the load is narrower than the 860 // full alloca type, we need to do a shift to get the right bits. 861 int ShAmt = 0; 862 if (TD.isBigEndian()) { 863 // On big-endian machines, the lowest bit is stored at the bit offset 864 // from the pointer given by getTypeStoreSizeInBits. This matters for 865 // integers with a bitwidth that is not a multiple of 8. 866 ShAmt = TD.getTypeStoreSizeInBits(NTy) - 867 TD.getTypeStoreSizeInBits(ToType) - Offset; 868 } else { 869 ShAmt = Offset; 870 } 871 872 // Note: we support negative bitwidths (with shl) which are not defined. 873 // We do this to support (f.e.) loads off the end of a structure where 874 // only some bits are used. 875 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 876 FromVal = Builder.CreateLShr(FromVal, 877 ConstantInt::get(FromVal->getType(), 878 ShAmt), "tmp"); 879 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 880 FromVal = Builder.CreateShl(FromVal, 881 ConstantInt::get(FromVal->getType(), 882 -ShAmt), "tmp"); 883 884 // Finally, unconditionally truncate the integer to the right width. 885 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType); 886 if (LIBitWidth < NTy->getBitWidth()) 887 FromVal = 888 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 889 LIBitWidth), "tmp"); 890 else if (LIBitWidth > NTy->getBitWidth()) 891 FromVal = 892 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 893 LIBitWidth), "tmp"); 894 895 // If the result is an integer, this is a trunc or bitcast. 896 if (ToType->isIntegerTy()) { 897 // Should be done. 898 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 899 // Just do a bitcast, we know the sizes match up. 900 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 901 } else { 902 // Otherwise must be a pointer. 903 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 904 } 905 assert(FromVal->getType() == ToType && "Didn't convert right?"); 906 return FromVal; 907} 908 909/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 910/// or vector value "Old" at the offset specified by Offset. 911/// 912/// This happens when we are converting an "integer union" to a 913/// single integer scalar, or when we are converting a "vector union" to a 914/// vector with insert/extractelement instructions. 915/// 916/// Offset is an offset from the original alloca, in bits that need to be 917/// shifted to the right. 918Value *ConvertToScalarInfo:: 919ConvertScalar_InsertValue(Value *SV, Value *Old, 920 uint64_t Offset, IRBuilder<> &Builder) { 921 // Convert the stored type to the actual type, shift it left to insert 922 // then 'or' into place. 923 const Type *AllocaType = Old->getType(); 924 LLVMContext &Context = Old->getContext(); 925 926 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 927 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy); 928 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType()); 929 930 // Changing the whole vector with memset or with an access of a different 931 // vector type? 932 if (ValSize == VecSize) { 933 // If the two types have the same primitive size, use a bit cast. 934 // Otherwise, it is two vectors with the same element type that has 935 // the same allocation size but different number of elements so use 936 // a shuffle vector. 937 if (VTy->getPrimitiveSizeInBits() == 938 SV->getType()->getPrimitiveSizeInBits()) 939 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 940 else 941 return CreateShuffleVectorCast(SV, VTy, Builder); 942 } 943 944 if (isPowerOf2_64(VecSize / ValSize)) { 945 assert(!(SV->getType()->isVectorTy() && Offset != 0) && "Can't insert a " 946 "value of a smaller vector type at a nonzero offset."); 947 948 const Type *CastElementTy = getScaledElementType(VTy, SV->getType(), 949 ValSize); 950 unsigned NumCastVectorElements = VecSize / ValSize; 951 952 LLVMContext &Context = SV->getContext(); 953 const Type *OldCastTy = VectorType::get(CastElementTy, 954 NumCastVectorElements); 955 Value *OldCast = Builder.CreateBitCast(Old, OldCastTy, "tmp"); 956 957 Value *SVCast = Builder.CreateBitCast(SV, CastElementTy, "tmp"); 958 959 unsigned EltSize = TD.getTypeAllocSizeInBits(CastElementTy); 960 unsigned Elt = Offset/EltSize; 961 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 962 Value *Insert = 963 Builder.CreateInsertElement(OldCast, SVCast, ConstantInt::get( 964 Type::getInt32Ty(Context), Elt), "tmp"); 965 return Builder.CreateBitCast(Insert, AllocaType, "tmp"); 966 } 967 968 // Must be an element insertion. 969 assert(SV->getType() == VTy->getElementType()); 970 uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 971 unsigned Elt = Offset/EltSize; 972 return Builder.CreateInsertElement(Old, SV, 973 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt), 974 "tmp"); 975 } 976 977 // If SV is a first-class aggregate value, insert each value recursively. 978 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 979 const StructLayout &Layout = *TD.getStructLayout(ST); 980 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 981 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 982 Old = ConvertScalar_InsertValue(Elt, Old, 983 Offset+Layout.getElementOffsetInBits(i), 984 Builder); 985 } 986 return Old; 987 } 988 989 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 990 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 991 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 992 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 993 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 994 } 995 return Old; 996 } 997 998 // If SV is a float, convert it to the appropriate integer type. 999 // If it is a pointer, do the same. 1000 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 1001 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); 1002 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); 1003 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); 1004 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 1005 SV = Builder.CreateBitCast(SV, 1006 IntegerType::get(SV->getContext(),SrcWidth), "tmp"); 1007 else if (SV->getType()->isPointerTy()) 1008 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()), "tmp"); 1009 1010 // Zero extend or truncate the value if needed. 1011 if (SV->getType() != AllocaType) { 1012 if (SV->getType()->getPrimitiveSizeInBits() < 1013 AllocaType->getPrimitiveSizeInBits()) 1014 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 1015 else { 1016 // Truncation may be needed if storing more than the alloca can hold 1017 // (undefined behavior). 1018 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 1019 SrcWidth = DestWidth; 1020 SrcStoreWidth = DestStoreWidth; 1021 } 1022 } 1023 1024 // If this is a big-endian system and the store is narrower than the 1025 // full alloca type, we need to do a shift to get the right bits. 1026 int ShAmt = 0; 1027 if (TD.isBigEndian()) { 1028 // On big-endian machines, the lowest bit is stored at the bit offset 1029 // from the pointer given by getTypeStoreSizeInBits. This matters for 1030 // integers with a bitwidth that is not a multiple of 8. 1031 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 1032 } else { 1033 ShAmt = Offset; 1034 } 1035 1036 // Note: we support negative bitwidths (with shr) which are not defined. 1037 // We do this to support (f.e.) stores off the end of a structure where 1038 // only some bits in the structure are set. 1039 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 1040 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 1041 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 1042 ShAmt), "tmp"); 1043 Mask <<= ShAmt; 1044 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1045 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 1046 -ShAmt), "tmp"); 1047 Mask = Mask.lshr(-ShAmt); 1048 } 1049 1050 // Mask out the bits we are about to insert from the old value, and or 1051 // in the new bits. 1052 if (SrcWidth != DestWidth) { 1053 assert(DestWidth > SrcWidth); 1054 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 1055 SV = Builder.CreateOr(Old, SV, "ins"); 1056 } 1057 return SV; 1058} 1059 1060 1061//===----------------------------------------------------------------------===// 1062// SRoA Driver 1063//===----------------------------------------------------------------------===// 1064 1065 1066bool SROA::runOnFunction(Function &F) { 1067 TD = getAnalysisIfAvailable<TargetData>(); 1068 1069 bool Changed = performPromotion(F); 1070 1071 // FIXME: ScalarRepl currently depends on TargetData more than it 1072 // theoretically needs to. It should be refactored in order to support 1073 // target-independent IR. Until this is done, just skip the actual 1074 // scalar-replacement portion of this pass. 1075 if (!TD) return Changed; 1076 1077 while (1) { 1078 bool LocalChange = performScalarRepl(F); 1079 if (!LocalChange) break; // No need to repromote if no scalarrepl 1080 Changed = true; 1081 LocalChange = performPromotion(F); 1082 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 1083 } 1084 1085 return Changed; 1086} 1087 1088namespace { 1089class AllocaPromoter : public LoadAndStorePromoter { 1090 AllocaInst *AI; 1091public: 1092 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S, 1093 DbgDeclareInst *DD, DIBuilder *&DB) 1094 : LoadAndStorePromoter(Insts, S, DD, DB), AI(0) {} 1095 1096 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) { 1097 // Remember which alloca we're promoting (for isInstInList). 1098 this->AI = AI; 1099 LoadAndStorePromoter::run(Insts); 1100 AI->eraseFromParent(); 1101 } 1102 1103 virtual bool isInstInList(Instruction *I, 1104 const SmallVectorImpl<Instruction*> &Insts) const { 1105 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 1106 return LI->getOperand(0) == AI; 1107 return cast<StoreInst>(I)->getPointerOperand() == AI; 1108 } 1109}; 1110} // end anon namespace 1111 1112/// isSafeSelectToSpeculate - Select instructions that use an alloca and are 1113/// subsequently loaded can be rewritten to load both input pointers and then 1114/// select between the result, allowing the load of the alloca to be promoted. 1115/// From this: 1116/// %P2 = select i1 %cond, i32* %Alloca, i32* %Other 1117/// %V = load i32* %P2 1118/// to: 1119/// %V1 = load i32* %Alloca -> will be mem2reg'd 1120/// %V2 = load i32* %Other 1121/// %V = select i1 %cond, i32 %V1, i32 %V2 1122/// 1123/// We can do this to a select if its only uses are loads and if the operand to 1124/// the select can be loaded unconditionally. 1125static bool isSafeSelectToSpeculate(SelectInst *SI, const TargetData *TD) { 1126 bool TDerefable = SI->getTrueValue()->isDereferenceablePointer(); 1127 bool FDerefable = SI->getFalseValue()->isDereferenceablePointer(); 1128 1129 for (Value::use_iterator UI = SI->use_begin(), UE = SI->use_end(); 1130 UI != UE; ++UI) { 1131 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1132 if (LI == 0 || LI->isVolatile()) return false; 1133 1134 // Both operands to the select need to be dereferencable, either absolutely 1135 // (e.g. allocas) or at this point because we can see other accesses to it. 1136 if (!TDerefable && !isSafeToLoadUnconditionally(SI->getTrueValue(), LI, 1137 LI->getAlignment(), TD)) 1138 return false; 1139 if (!FDerefable && !isSafeToLoadUnconditionally(SI->getFalseValue(), LI, 1140 LI->getAlignment(), TD)) 1141 return false; 1142 } 1143 1144 return true; 1145} 1146 1147/// isSafePHIToSpeculate - PHI instructions that use an alloca and are 1148/// subsequently loaded can be rewritten to load both input pointers in the pred 1149/// blocks and then PHI the results, allowing the load of the alloca to be 1150/// promoted. 1151/// From this: 1152/// %P2 = phi [i32* %Alloca, i32* %Other] 1153/// %V = load i32* %P2 1154/// to: 1155/// %V1 = load i32* %Alloca -> will be mem2reg'd 1156/// ... 1157/// %V2 = load i32* %Other 1158/// ... 1159/// %V = phi [i32 %V1, i32 %V2] 1160/// 1161/// We can do this to a select if its only uses are loads and if the operand to 1162/// the select can be loaded unconditionally. 1163static bool isSafePHIToSpeculate(PHINode *PN, const TargetData *TD) { 1164 // For now, we can only do this promotion if the load is in the same block as 1165 // the PHI, and if there are no stores between the phi and load. 1166 // TODO: Allow recursive phi users. 1167 // TODO: Allow stores. 1168 BasicBlock *BB = PN->getParent(); 1169 unsigned MaxAlign = 0; 1170 for (Value::use_iterator UI = PN->use_begin(), UE = PN->use_end(); 1171 UI != UE; ++UI) { 1172 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1173 if (LI == 0 || LI->isVolatile()) return false; 1174 1175 // For now we only allow loads in the same block as the PHI. This is a 1176 // common case that happens when instcombine merges two loads through a PHI. 1177 if (LI->getParent() != BB) return false; 1178 1179 // Ensure that there are no instructions between the PHI and the load that 1180 // could store. 1181 for (BasicBlock::iterator BBI = PN; &*BBI != LI; ++BBI) 1182 if (BBI->mayWriteToMemory()) 1183 return false; 1184 1185 MaxAlign = std::max(MaxAlign, LI->getAlignment()); 1186 } 1187 1188 // Okay, we know that we have one or more loads in the same block as the PHI. 1189 // We can transform this if it is safe to push the loads into the predecessor 1190 // blocks. The only thing to watch out for is that we can't put a possibly 1191 // trapping load in the predecessor if it is a critical edge. 1192 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1193 BasicBlock *Pred = PN->getIncomingBlock(i); 1194 1195 // If the predecessor has a single successor, then the edge isn't critical. 1196 if (Pred->getTerminator()->getNumSuccessors() == 1) 1197 continue; 1198 1199 Value *InVal = PN->getIncomingValue(i); 1200 1201 // If the InVal is an invoke in the pred, we can't put a load on the edge. 1202 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal)) 1203 if (II->getParent() == Pred) 1204 return false; 1205 1206 // If this pointer is always safe to load, or if we can prove that there is 1207 // already a load in the block, then we can move the load to the pred block. 1208 if (InVal->isDereferenceablePointer() || 1209 isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign, TD)) 1210 continue; 1211 1212 return false; 1213 } 1214 1215 return true; 1216} 1217 1218 1219/// tryToMakeAllocaBePromotable - This returns true if the alloca only has 1220/// direct (non-volatile) loads and stores to it. If the alloca is close but 1221/// not quite there, this will transform the code to allow promotion. As such, 1222/// it is a non-pure predicate. 1223static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const TargetData *TD) { 1224 SetVector<Instruction*, SmallVector<Instruction*, 4>, 1225 SmallPtrSet<Instruction*, 4> > InstsToRewrite; 1226 1227 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end(); 1228 UI != UE; ++UI) { 1229 User *U = *UI; 1230 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 1231 if (LI->isVolatile()) 1232 return false; 1233 continue; 1234 } 1235 1236 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 1237 if (SI->getOperand(0) == AI || SI->isVolatile()) 1238 return false; // Don't allow a store OF the AI, only INTO the AI. 1239 continue; 1240 } 1241 1242 if (SelectInst *SI = dyn_cast<SelectInst>(U)) { 1243 // If the condition being selected on is a constant, fold the select, yes 1244 // this does (rarely) happen early on. 1245 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition())) { 1246 Value *Result = SI->getOperand(1+CI->isZero()); 1247 SI->replaceAllUsesWith(Result); 1248 SI->eraseFromParent(); 1249 1250 // This is very rare and we just scrambled the use list of AI, start 1251 // over completely. 1252 return tryToMakeAllocaBePromotable(AI, TD); 1253 } 1254 1255 // If it is safe to turn "load (select c, AI, ptr)" into a select of two 1256 // loads, then we can transform this by rewriting the select. 1257 if (!isSafeSelectToSpeculate(SI, TD)) 1258 return false; 1259 1260 InstsToRewrite.insert(SI); 1261 continue; 1262 } 1263 1264 if (PHINode *PN = dyn_cast<PHINode>(U)) { 1265 if (PN->use_empty()) { // Dead PHIs can be stripped. 1266 InstsToRewrite.insert(PN); 1267 continue; 1268 } 1269 1270 // If it is safe to turn "load (phi [AI, ptr, ...])" into a PHI of loads 1271 // in the pred blocks, then we can transform this by rewriting the PHI. 1272 if (!isSafePHIToSpeculate(PN, TD)) 1273 return false; 1274 1275 InstsToRewrite.insert(PN); 1276 continue; 1277 } 1278 1279 return false; 1280 } 1281 1282 // If there are no instructions to rewrite, then all uses are load/stores and 1283 // we're done! 1284 if (InstsToRewrite.empty()) 1285 return true; 1286 1287 // If we have instructions that need to be rewritten for this to be promotable 1288 // take care of it now. 1289 for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) { 1290 if (SelectInst *SI = dyn_cast<SelectInst>(InstsToRewrite[i])) { 1291 // Selects in InstsToRewrite only have load uses. Rewrite each as two 1292 // loads with a new select. 1293 while (!SI->use_empty()) { 1294 LoadInst *LI = cast<LoadInst>(SI->use_back()); 1295 1296 IRBuilder<> Builder(LI); 1297 LoadInst *TrueLoad = 1298 Builder.CreateLoad(SI->getTrueValue(), LI->getName()+".t"); 1299 LoadInst *FalseLoad = 1300 Builder.CreateLoad(SI->getFalseValue(), LI->getName()+".t"); 1301 1302 // Transfer alignment and TBAA info if present. 1303 TrueLoad->setAlignment(LI->getAlignment()); 1304 FalseLoad->setAlignment(LI->getAlignment()); 1305 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) { 1306 TrueLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1307 FalseLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1308 } 1309 1310 Value *V = Builder.CreateSelect(SI->getCondition(), TrueLoad, FalseLoad); 1311 V->takeName(LI); 1312 LI->replaceAllUsesWith(V); 1313 LI->eraseFromParent(); 1314 } 1315 1316 // Now that all the loads are gone, the select is gone too. 1317 SI->eraseFromParent(); 1318 continue; 1319 } 1320 1321 // Otherwise, we have a PHI node which allows us to push the loads into the 1322 // predecessors. 1323 PHINode *PN = cast<PHINode>(InstsToRewrite[i]); 1324 if (PN->use_empty()) { 1325 PN->eraseFromParent(); 1326 continue; 1327 } 1328 1329 const Type *LoadTy = cast<PointerType>(PN->getType())->getElementType(); 1330 PHINode *NewPN = PHINode::Create(LoadTy, PN->getNumIncomingValues(), 1331 PN->getName()+".ld", PN); 1332 1333 // Get the TBAA tag and alignment to use from one of the loads. It doesn't 1334 // matter which one we get and if any differ, it doesn't matter. 1335 LoadInst *SomeLoad = cast<LoadInst>(PN->use_back()); 1336 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa); 1337 unsigned Align = SomeLoad->getAlignment(); 1338 1339 // Rewrite all loads of the PN to use the new PHI. 1340 while (!PN->use_empty()) { 1341 LoadInst *LI = cast<LoadInst>(PN->use_back()); 1342 LI->replaceAllUsesWith(NewPN); 1343 LI->eraseFromParent(); 1344 } 1345 1346 // Inject loads into all of the pred blocks. Keep track of which blocks we 1347 // insert them into in case we have multiple edges from the same block. 1348 DenseMap<BasicBlock*, LoadInst*> InsertedLoads; 1349 1350 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1351 BasicBlock *Pred = PN->getIncomingBlock(i); 1352 LoadInst *&Load = InsertedLoads[Pred]; 1353 if (Load == 0) { 1354 Load = new LoadInst(PN->getIncomingValue(i), 1355 PN->getName() + "." + Pred->getName(), 1356 Pred->getTerminator()); 1357 Load->setAlignment(Align); 1358 if (TBAATag) Load->setMetadata(LLVMContext::MD_tbaa, TBAATag); 1359 } 1360 1361 NewPN->addIncoming(Load, Pred); 1362 } 1363 1364 PN->eraseFromParent(); 1365 } 1366 1367 ++NumAdjusted; 1368 return true; 1369} 1370 1371bool SROA::performPromotion(Function &F) { 1372 std::vector<AllocaInst*> Allocas; 1373 DominatorTree *DT = 0; 1374 if (HasDomTree) 1375 DT = &getAnalysis<DominatorTree>(); 1376 1377 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 1378 1379 bool Changed = false; 1380 SmallVector<Instruction*, 64> Insts; 1381 DIBuilder *DIB = 0; 1382 while (1) { 1383 Allocas.clear(); 1384 1385 // Find allocas that are safe to promote, by looking at all instructions in 1386 // the entry node 1387 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 1388 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 1389 if (tryToMakeAllocaBePromotable(AI, TD)) 1390 Allocas.push_back(AI); 1391 1392 if (Allocas.empty()) break; 1393 1394 if (HasDomTree) 1395 PromoteMemToReg(Allocas, *DT); 1396 else { 1397 SSAUpdater SSA; 1398 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) { 1399 AllocaInst *AI = Allocas[i]; 1400 1401 // Build list of instructions to promote. 1402 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 1403 UI != E; ++UI) 1404 Insts.push_back(cast<Instruction>(*UI)); 1405 1406 DbgDeclareInst *DDI = FindAllocaDbgDeclare(AI); 1407 if (DDI && !DIB) 1408 DIB = new DIBuilder(*AI->getParent()->getParent()->getParent()); 1409 AllocaPromoter(Insts, SSA, DDI, DIB).run(AI, Insts); 1410 Insts.clear(); 1411 } 1412 } 1413 NumPromoted += Allocas.size(); 1414 Changed = true; 1415 } 1416 1417 // FIXME: Is there a better way to handle the lazy initialization of DIB 1418 // so that there doesn't need to be an explicit delete? 1419 delete DIB; 1420 1421 return Changed; 1422} 1423 1424 1425/// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 1426/// SROA. It must be a struct or array type with a small number of elements. 1427static bool ShouldAttemptScalarRepl(AllocaInst *AI) { 1428 const Type *T = AI->getAllocatedType(); 1429 // Do not promote any struct into more than 32 separate vars. 1430 if (const StructType *ST = dyn_cast<StructType>(T)) 1431 return ST->getNumElements() <= 32; 1432 // Arrays are much less likely to be safe for SROA; only consider 1433 // them if they are very small. 1434 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) 1435 return AT->getNumElements() <= 8; 1436 return false; 1437} 1438 1439 1440// performScalarRepl - This algorithm is a simple worklist driven algorithm, 1441// which runs on all of the malloc/alloca instructions in the function, removing 1442// them if they are only used by getelementptr instructions. 1443// 1444bool SROA::performScalarRepl(Function &F) { 1445 std::vector<AllocaInst*> WorkList; 1446 1447 // Scan the entry basic block, adding allocas to the worklist. 1448 BasicBlock &BB = F.getEntryBlock(); 1449 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 1450 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 1451 WorkList.push_back(A); 1452 1453 // Process the worklist 1454 bool Changed = false; 1455 while (!WorkList.empty()) { 1456 AllocaInst *AI = WorkList.back(); 1457 WorkList.pop_back(); 1458 1459 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 1460 // with unused elements. 1461 if (AI->use_empty()) { 1462 AI->eraseFromParent(); 1463 Changed = true; 1464 continue; 1465 } 1466 1467 // If this alloca is impossible for us to promote, reject it early. 1468 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 1469 continue; 1470 1471 // Check to see if this allocation is only modified by a memcpy/memmove from 1472 // a constant global. If this is the case, we can change all users to use 1473 // the constant global instead. This is commonly produced by the CFE by 1474 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 1475 // is only subsequently read. 1476 if (MemTransferInst *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 1477 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); 1478 DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n'); 1479 Constant *TheSrc = cast<Constant>(TheCopy->getSource()); 1480 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 1481 TheCopy->eraseFromParent(); // Don't mutate the global. 1482 AI->eraseFromParent(); 1483 ++NumGlobals; 1484 Changed = true; 1485 continue; 1486 } 1487 1488 // Check to see if we can perform the core SROA transformation. We cannot 1489 // transform the allocation instruction if it is an array allocation 1490 // (allocations OF arrays are ok though), and an allocation of a scalar 1491 // value cannot be decomposed at all. 1492 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 1493 1494 // Do not promote [0 x %struct]. 1495 if (AllocaSize == 0) continue; 1496 1497 // Do not promote any struct whose size is too big. 1498 if (AllocaSize > SRThreshold) continue; 1499 1500 // If the alloca looks like a good candidate for scalar replacement, and if 1501 // all its users can be transformed, then split up the aggregate into its 1502 // separate elements. 1503 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 1504 DoScalarReplacement(AI, WorkList); 1505 Changed = true; 1506 continue; 1507 } 1508 1509 // If we can turn this aggregate value (potentially with casts) into a 1510 // simple scalar value that can be mem2reg'd into a register value. 1511 // IsNotTrivial tracks whether this is something that mem2reg could have 1512 // promoted itself. If so, we don't want to transform it needlessly. Note 1513 // that we can't just check based on the type: the alloca may be of an i32 1514 // but that has pointer arithmetic to set byte 3 of it or something. 1515 if (AllocaInst *NewAI = 1516 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) { 1517 NewAI->takeName(AI); 1518 AI->eraseFromParent(); 1519 ++NumConverted; 1520 Changed = true; 1521 continue; 1522 } 1523 1524 // Otherwise, couldn't process this alloca. 1525 } 1526 1527 return Changed; 1528} 1529 1530/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 1531/// predicate, do SROA now. 1532void SROA::DoScalarReplacement(AllocaInst *AI, 1533 std::vector<AllocaInst*> &WorkList) { 1534 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 1535 SmallVector<AllocaInst*, 32> ElementAllocas; 1536 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 1537 ElementAllocas.reserve(ST->getNumContainedTypes()); 1538 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 1539 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 1540 AI->getAlignment(), 1541 AI->getName() + "." + Twine(i), AI); 1542 ElementAllocas.push_back(NA); 1543 WorkList.push_back(NA); // Add to worklist for recursive processing 1544 } 1545 } else { 1546 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 1547 ElementAllocas.reserve(AT->getNumElements()); 1548 const Type *ElTy = AT->getElementType(); 1549 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1550 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 1551 AI->getName() + "." + Twine(i), AI); 1552 ElementAllocas.push_back(NA); 1553 WorkList.push_back(NA); // Add to worklist for recursive processing 1554 } 1555 } 1556 1557 // Now that we have created the new alloca instructions, rewrite all the 1558 // uses of the old alloca. 1559 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 1560 1561 // Now erase any instructions that were made dead while rewriting the alloca. 1562 DeleteDeadInstructions(); 1563 AI->eraseFromParent(); 1564 1565 ++NumReplaced; 1566} 1567 1568/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 1569/// recursively including all their operands that become trivially dead. 1570void SROA::DeleteDeadInstructions() { 1571 while (!DeadInsts.empty()) { 1572 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 1573 1574 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 1575 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 1576 // Zero out the operand and see if it becomes trivially dead. 1577 // (But, don't add allocas to the dead instruction list -- they are 1578 // already on the worklist and will be deleted separately.) 1579 *OI = 0; 1580 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 1581 DeadInsts.push_back(U); 1582 } 1583 1584 I->eraseFromParent(); 1585 } 1586} 1587 1588/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 1589/// performing scalar replacement of alloca AI. The results are flagged in 1590/// the Info parameter. Offset indicates the position within AI that is 1591/// referenced by this instruction. 1592void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset, 1593 AllocaInfo &Info) { 1594 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1595 Instruction *User = cast<Instruction>(*UI); 1596 1597 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1598 isSafeForScalarRepl(BC, Offset, Info); 1599 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1600 uint64_t GEPOffset = Offset; 1601 isSafeGEP(GEPI, GEPOffset, Info); 1602 if (!Info.isUnsafe) 1603 isSafeForScalarRepl(GEPI, GEPOffset, Info); 1604 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1605 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1606 if (Length == 0) 1607 return MarkUnsafe(Info, User); 1608 isSafeMemAccess(Offset, Length->getZExtValue(), 0, 1609 UI.getOperandNo() == 0, Info, MI, 1610 true /*AllowWholeAccess*/); 1611 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1612 if (LI->isVolatile()) 1613 return MarkUnsafe(Info, User); 1614 const Type *LIType = LI->getType(); 1615 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1616 LIType, false, Info, LI, true /*AllowWholeAccess*/); 1617 Info.hasALoadOrStore = true; 1618 1619 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1620 // Store is ok if storing INTO the pointer, not storing the pointer 1621 if (SI->isVolatile() || SI->getOperand(0) == I) 1622 return MarkUnsafe(Info, User); 1623 1624 const Type *SIType = SI->getOperand(0)->getType(); 1625 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1626 SIType, true, Info, SI, true /*AllowWholeAccess*/); 1627 Info.hasALoadOrStore = true; 1628 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1629 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1630 } else { 1631 return MarkUnsafe(Info, User); 1632 } 1633 if (Info.isUnsafe) return; 1634 } 1635} 1636 1637 1638/// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer 1639/// derived from the alloca, we can often still split the alloca into elements. 1640/// This is useful if we have a large alloca where one element is phi'd 1641/// together somewhere: we can SRoA and promote all the other elements even if 1642/// we end up not being able to promote this one. 1643/// 1644/// All we require is that the uses of the PHI do not index into other parts of 1645/// the alloca. The most important use case for this is single load and stores 1646/// that are PHI'd together, which can happen due to code sinking. 1647void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset, 1648 AllocaInfo &Info) { 1649 // If we've already checked this PHI, don't do it again. 1650 if (PHINode *PN = dyn_cast<PHINode>(I)) 1651 if (!Info.CheckedPHIs.insert(PN)) 1652 return; 1653 1654 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1655 Instruction *User = cast<Instruction>(*UI); 1656 1657 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1658 isSafePHISelectUseForScalarRepl(BC, Offset, Info); 1659 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1660 // Only allow "bitcast" GEPs for simplicity. We could generalize this, 1661 // but would have to prove that we're staying inside of an element being 1662 // promoted. 1663 if (!GEPI->hasAllZeroIndices()) 1664 return MarkUnsafe(Info, User); 1665 isSafePHISelectUseForScalarRepl(GEPI, Offset, Info); 1666 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1667 if (LI->isVolatile()) 1668 return MarkUnsafe(Info, User); 1669 const Type *LIType = LI->getType(); 1670 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1671 LIType, false, Info, LI, false /*AllowWholeAccess*/); 1672 Info.hasALoadOrStore = true; 1673 1674 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1675 // Store is ok if storing INTO the pointer, not storing the pointer 1676 if (SI->isVolatile() || SI->getOperand(0) == I) 1677 return MarkUnsafe(Info, User); 1678 1679 const Type *SIType = SI->getOperand(0)->getType(); 1680 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1681 SIType, true, Info, SI, false /*AllowWholeAccess*/); 1682 Info.hasALoadOrStore = true; 1683 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1684 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1685 } else { 1686 return MarkUnsafe(Info, User); 1687 } 1688 if (Info.isUnsafe) return; 1689 } 1690} 1691 1692/// isSafeGEP - Check if a GEP instruction can be handled for scalar 1693/// replacement. It is safe when all the indices are constant, in-bounds 1694/// references, and when the resulting offset corresponds to an element within 1695/// the alloca type. The results are flagged in the Info parameter. Upon 1696/// return, Offset is adjusted as specified by the GEP indices. 1697void SROA::isSafeGEP(GetElementPtrInst *GEPI, 1698 uint64_t &Offset, AllocaInfo &Info) { 1699 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 1700 if (GEPIt == E) 1701 return; 1702 1703 // Walk through the GEP type indices, checking the types that this indexes 1704 // into. 1705 for (; GEPIt != E; ++GEPIt) { 1706 // Ignore struct elements, no extra checking needed for these. 1707 if ((*GEPIt)->isStructTy()) 1708 continue; 1709 1710 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 1711 if (!IdxVal) 1712 return MarkUnsafe(Info, GEPI); 1713 } 1714 1715 // Compute the offset due to this GEP and check if the alloca has a 1716 // component element at that offset. 1717 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1718 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1719 &Indices[0], Indices.size()); 1720 if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 0)) 1721 MarkUnsafe(Info, GEPI); 1722} 1723 1724/// isHomogeneousAggregate - Check if type T is a struct or array containing 1725/// elements of the same type (which is always true for arrays). If so, 1726/// return true with NumElts and EltTy set to the number of elements and the 1727/// element type, respectively. 1728static bool isHomogeneousAggregate(const Type *T, unsigned &NumElts, 1729 const Type *&EltTy) { 1730 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1731 NumElts = AT->getNumElements(); 1732 EltTy = (NumElts == 0 ? 0 : AT->getElementType()); 1733 return true; 1734 } 1735 if (const StructType *ST = dyn_cast<StructType>(T)) { 1736 NumElts = ST->getNumContainedTypes(); 1737 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0)); 1738 for (unsigned n = 1; n < NumElts; ++n) { 1739 if (ST->getContainedType(n) != EltTy) 1740 return false; 1741 } 1742 return true; 1743 } 1744 return false; 1745} 1746 1747/// isCompatibleAggregate - Check if T1 and T2 are either the same type or are 1748/// "homogeneous" aggregates with the same element type and number of elements. 1749static bool isCompatibleAggregate(const Type *T1, const Type *T2) { 1750 if (T1 == T2) 1751 return true; 1752 1753 unsigned NumElts1, NumElts2; 1754 const Type *EltTy1, *EltTy2; 1755 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) && 1756 isHomogeneousAggregate(T2, NumElts2, EltTy2) && 1757 NumElts1 == NumElts2 && 1758 EltTy1 == EltTy2) 1759 return true; 1760 1761 return false; 1762} 1763 1764/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1765/// alloca or has an offset and size that corresponds to a component element 1766/// within it. The offset checked here may have been formed from a GEP with a 1767/// pointer bitcasted to a different type. 1768/// 1769/// If AllowWholeAccess is true, then this allows uses of the entire alloca as a 1770/// unit. If false, it only allows accesses known to be in a single element. 1771void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 1772 const Type *MemOpType, bool isStore, 1773 AllocaInfo &Info, Instruction *TheAccess, 1774 bool AllowWholeAccess) { 1775 // Check if this is a load/store of the entire alloca. 1776 if (Offset == 0 && AllowWholeAccess && 1777 MemSize == TD->getTypeAllocSize(Info.AI->getAllocatedType())) { 1778 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer 1779 // loads/stores (which are essentially the same as the MemIntrinsics with 1780 // regard to copying padding between elements). But, if an alloca is 1781 // flagged as both a source and destination of such operations, we'll need 1782 // to check later for padding between elements. 1783 if (!MemOpType || MemOpType->isIntegerTy()) { 1784 if (isStore) 1785 Info.isMemCpyDst = true; 1786 else 1787 Info.isMemCpySrc = true; 1788 return; 1789 } 1790 // This is also safe for references using a type that is compatible with 1791 // the type of the alloca, so that loads/stores can be rewritten using 1792 // insertvalue/extractvalue. 1793 if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) { 1794 Info.hasSubelementAccess = true; 1795 return; 1796 } 1797 } 1798 // Check if the offset/size correspond to a component within the alloca type. 1799 const Type *T = Info.AI->getAllocatedType(); 1800 if (TypeHasComponent(T, Offset, MemSize)) { 1801 Info.hasSubelementAccess = true; 1802 return; 1803 } 1804 1805 return MarkUnsafe(Info, TheAccess); 1806} 1807 1808/// TypeHasComponent - Return true if T has a component type with the 1809/// specified offset and size. If Size is zero, do not check the size. 1810bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) { 1811 const Type *EltTy; 1812 uint64_t EltSize; 1813 if (const StructType *ST = dyn_cast<StructType>(T)) { 1814 const StructLayout *Layout = TD->getStructLayout(ST); 1815 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1816 EltTy = ST->getContainedType(EltIdx); 1817 EltSize = TD->getTypeAllocSize(EltTy); 1818 Offset -= Layout->getElementOffset(EltIdx); 1819 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 1820 EltTy = AT->getElementType(); 1821 EltSize = TD->getTypeAllocSize(EltTy); 1822 if (Offset >= AT->getNumElements() * EltSize) 1823 return false; 1824 Offset %= EltSize; 1825 } else { 1826 return false; 1827 } 1828 if (Offset == 0 && (Size == 0 || EltSize == Size)) 1829 return true; 1830 // Check if the component spans multiple elements. 1831 if (Offset + Size > EltSize) 1832 return false; 1833 return TypeHasComponent(EltTy, Offset, Size); 1834} 1835 1836/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1837/// the instruction I, which references it, to use the separate elements. 1838/// Offset indicates the position within AI that is referenced by this 1839/// instruction. 1840void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1841 SmallVector<AllocaInst*, 32> &NewElts) { 1842 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) { 1843 Use &TheUse = UI.getUse(); 1844 Instruction *User = cast<Instruction>(*UI++); 1845 1846 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1847 RewriteBitCast(BC, AI, Offset, NewElts); 1848 continue; 1849 } 1850 1851 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1852 RewriteGEP(GEPI, AI, Offset, NewElts); 1853 continue; 1854 } 1855 1856 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1857 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1858 uint64_t MemSize = Length->getZExtValue(); 1859 if (Offset == 0 && 1860 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 1861 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1862 // Otherwise the intrinsic can only touch a single element and the 1863 // address operand will be updated, so nothing else needs to be done. 1864 continue; 1865 } 1866 1867 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1868 const Type *LIType = LI->getType(); 1869 1870 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) { 1871 // Replace: 1872 // %res = load { i32, i32 }* %alloc 1873 // with: 1874 // %load.0 = load i32* %alloc.0 1875 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1876 // %load.1 = load i32* %alloc.1 1877 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1878 // (Also works for arrays instead of structs) 1879 Value *Insert = UndefValue::get(LIType); 1880 IRBuilder<> Builder(LI); 1881 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1882 Value *Load = Builder.CreateLoad(NewElts[i], "load"); 1883 Insert = Builder.CreateInsertValue(Insert, Load, i, "insert"); 1884 } 1885 LI->replaceAllUsesWith(Insert); 1886 DeadInsts.push_back(LI); 1887 } else if (LIType->isIntegerTy() && 1888 TD->getTypeAllocSize(LIType) == 1889 TD->getTypeAllocSize(AI->getAllocatedType())) { 1890 // If this is a load of the entire alloca to an integer, rewrite it. 1891 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1892 } 1893 continue; 1894 } 1895 1896 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1897 Value *Val = SI->getOperand(0); 1898 const Type *SIType = Val->getType(); 1899 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) { 1900 // Replace: 1901 // store { i32, i32 } %val, { i32, i32 }* %alloc 1902 // with: 1903 // %val.0 = extractvalue { i32, i32 } %val, 0 1904 // store i32 %val.0, i32* %alloc.0 1905 // %val.1 = extractvalue { i32, i32 } %val, 1 1906 // store i32 %val.1, i32* %alloc.1 1907 // (Also works for arrays instead of structs) 1908 IRBuilder<> Builder(SI); 1909 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1910 Value *Extract = Builder.CreateExtractValue(Val, i, Val->getName()); 1911 Builder.CreateStore(Extract, NewElts[i]); 1912 } 1913 DeadInsts.push_back(SI); 1914 } else if (SIType->isIntegerTy() && 1915 TD->getTypeAllocSize(SIType) == 1916 TD->getTypeAllocSize(AI->getAllocatedType())) { 1917 // If this is a store of the entire alloca from an integer, rewrite it. 1918 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 1919 } 1920 continue; 1921 } 1922 1923 if (isa<SelectInst>(User) || isa<PHINode>(User)) { 1924 // If we have a PHI user of the alloca itself (as opposed to a GEP or 1925 // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to 1926 // the new pointer. 1927 if (!isa<AllocaInst>(I)) continue; 1928 1929 assert(Offset == 0 && NewElts[0] && 1930 "Direct alloca use should have a zero offset"); 1931 1932 // If we have a use of the alloca, we know the derived uses will be 1933 // utilizing just the first element of the scalarized result. Insert a 1934 // bitcast of the first alloca before the user as required. 1935 AllocaInst *NewAI = NewElts[0]; 1936 BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI); 1937 NewAI->moveBefore(BCI); 1938 TheUse = BCI; 1939 continue; 1940 } 1941 } 1942} 1943 1944/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 1945/// and recursively continue updating all of its uses. 1946void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 1947 SmallVector<AllocaInst*, 32> &NewElts) { 1948 RewriteForScalarRepl(BC, AI, Offset, NewElts); 1949 if (BC->getOperand(0) != AI) 1950 return; 1951 1952 // The bitcast references the original alloca. Replace its uses with 1953 // references to the first new element alloca. 1954 Instruction *Val = NewElts[0]; 1955 if (Val->getType() != BC->getDestTy()) { 1956 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 1957 Val->takeName(BC); 1958 } 1959 BC->replaceAllUsesWith(Val); 1960 DeadInsts.push_back(BC); 1961} 1962 1963/// FindElementAndOffset - Return the index of the element containing Offset 1964/// within the specified type, which must be either a struct or an array. 1965/// Sets T to the type of the element and Offset to the offset within that 1966/// element. IdxTy is set to the type of the index result to be used in a 1967/// GEP instruction. 1968uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset, 1969 const Type *&IdxTy) { 1970 uint64_t Idx = 0; 1971 if (const StructType *ST = dyn_cast<StructType>(T)) { 1972 const StructLayout *Layout = TD->getStructLayout(ST); 1973 Idx = Layout->getElementContainingOffset(Offset); 1974 T = ST->getContainedType(Idx); 1975 Offset -= Layout->getElementOffset(Idx); 1976 IdxTy = Type::getInt32Ty(T->getContext()); 1977 return Idx; 1978 } 1979 const ArrayType *AT = cast<ArrayType>(T); 1980 T = AT->getElementType(); 1981 uint64_t EltSize = TD->getTypeAllocSize(T); 1982 Idx = Offset / EltSize; 1983 Offset -= Idx * EltSize; 1984 IdxTy = Type::getInt64Ty(T->getContext()); 1985 return Idx; 1986} 1987 1988/// RewriteGEP - Check if this GEP instruction moves the pointer across 1989/// elements of the alloca that are being split apart, and if so, rewrite 1990/// the GEP to be relative to the new element. 1991void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 1992 SmallVector<AllocaInst*, 32> &NewElts) { 1993 uint64_t OldOffset = Offset; 1994 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1995 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 1996 &Indices[0], Indices.size()); 1997 1998 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 1999 2000 const Type *T = AI->getAllocatedType(); 2001 const Type *IdxTy; 2002 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 2003 if (GEPI->getOperand(0) == AI) 2004 OldIdx = ~0ULL; // Force the GEP to be rewritten. 2005 2006 T = AI->getAllocatedType(); 2007 uint64_t EltOffset = Offset; 2008 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 2009 2010 // If this GEP does not move the pointer across elements of the alloca 2011 // being split, then it does not needs to be rewritten. 2012 if (Idx == OldIdx) 2013 return; 2014 2015 const Type *i32Ty = Type::getInt32Ty(AI->getContext()); 2016 SmallVector<Value*, 8> NewArgs; 2017 NewArgs.push_back(Constant::getNullValue(i32Ty)); 2018 while (EltOffset != 0) { 2019 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 2020 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 2021 } 2022 Instruction *Val = NewElts[Idx]; 2023 if (NewArgs.size() > 1) { 2024 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(), 2025 NewArgs.end(), "", GEPI); 2026 Val->takeName(GEPI); 2027 } 2028 if (Val->getType() != GEPI->getType()) 2029 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 2030 GEPI->replaceAllUsesWith(Val); 2031 DeadInsts.push_back(GEPI); 2032} 2033 2034/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 2035/// Rewrite it to copy or set the elements of the scalarized memory. 2036void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 2037 AllocaInst *AI, 2038 SmallVector<AllocaInst*, 32> &NewElts) { 2039 // If this is a memcpy/memmove, construct the other pointer as the 2040 // appropriate type. The "Other" pointer is the pointer that goes to memory 2041 // that doesn't have anything to do with the alloca that we are promoting. For 2042 // memset, this Value* stays null. 2043 Value *OtherPtr = 0; 2044 unsigned MemAlignment = MI->getAlignment(); 2045 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 2046 if (Inst == MTI->getRawDest()) 2047 OtherPtr = MTI->getRawSource(); 2048 else { 2049 assert(Inst == MTI->getRawSource()); 2050 OtherPtr = MTI->getRawDest(); 2051 } 2052 } 2053 2054 // If there is an other pointer, we want to convert it to the same pointer 2055 // type as AI has, so we can GEP through it safely. 2056 if (OtherPtr) { 2057 unsigned AddrSpace = 2058 cast<PointerType>(OtherPtr->getType())->getAddressSpace(); 2059 2060 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 2061 // optimization, but it's also required to detect the corner case where 2062 // both pointer operands are referencing the same memory, and where 2063 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 2064 // function is only called for mem intrinsics that access the whole 2065 // aggregate, so non-zero GEPs are not an issue here.) 2066 OtherPtr = OtherPtr->stripPointerCasts(); 2067 2068 // Copying the alloca to itself is a no-op: just delete it. 2069 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 2070 // This code will run twice for a no-op memcpy -- once for each operand. 2071 // Put only one reference to MI on the DeadInsts list. 2072 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), 2073 E = DeadInsts.end(); I != E; ++I) 2074 if (*I == MI) return; 2075 DeadInsts.push_back(MI); 2076 return; 2077 } 2078 2079 // If the pointer is not the right type, insert a bitcast to the right 2080 // type. 2081 const Type *NewTy = 2082 PointerType::get(AI->getType()->getElementType(), AddrSpace); 2083 2084 if (OtherPtr->getType() != NewTy) 2085 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI); 2086 } 2087 2088 // Process each element of the aggregate. 2089 bool SROADest = MI->getRawDest() == Inst; 2090 2091 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 2092 2093 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2094 // If this is a memcpy/memmove, emit a GEP of the other element address. 2095 Value *OtherElt = 0; 2096 unsigned OtherEltAlign = MemAlignment; 2097 2098 if (OtherPtr) { 2099 Value *Idx[2] = { Zero, 2100 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 2101 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2, 2102 OtherPtr->getName()+"."+Twine(i), 2103 MI); 2104 uint64_t EltOffset; 2105 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 2106 const Type *OtherTy = OtherPtrTy->getElementType(); 2107 if (const StructType *ST = dyn_cast<StructType>(OtherTy)) { 2108 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 2109 } else { 2110 const Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); 2111 EltOffset = TD->getTypeAllocSize(EltTy)*i; 2112 } 2113 2114 // The alignment of the other pointer is the guaranteed alignment of the 2115 // element, which is affected by both the known alignment of the whole 2116 // mem intrinsic and the alignment of the element. If the alignment of 2117 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 2118 // known alignment is just 4 bytes. 2119 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 2120 } 2121 2122 Value *EltPtr = NewElts[i]; 2123 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 2124 2125 // If we got down to a scalar, insert a load or store as appropriate. 2126 if (EltTy->isSingleValueType()) { 2127 if (isa<MemTransferInst>(MI)) { 2128 if (SROADest) { 2129 // From Other to Alloca. 2130 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 2131 new StoreInst(Elt, EltPtr, MI); 2132 } else { 2133 // From Alloca to Other. 2134 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 2135 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 2136 } 2137 continue; 2138 } 2139 assert(isa<MemSetInst>(MI)); 2140 2141 // If the stored element is zero (common case), just store a null 2142 // constant. 2143 Constant *StoreVal; 2144 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) { 2145 if (CI->isZero()) { 2146 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 2147 } else { 2148 // If EltTy is a vector type, get the element type. 2149 const Type *ValTy = EltTy->getScalarType(); 2150 2151 // Construct an integer with the right value. 2152 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 2153 APInt OneVal(EltSize, CI->getZExtValue()); 2154 APInt TotalVal(OneVal); 2155 // Set each byte. 2156 for (unsigned i = 0; 8*i < EltSize; ++i) { 2157 TotalVal = TotalVal.shl(8); 2158 TotalVal |= OneVal; 2159 } 2160 2161 // Convert the integer value to the appropriate type. 2162 StoreVal = ConstantInt::get(CI->getContext(), TotalVal); 2163 if (ValTy->isPointerTy()) 2164 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 2165 else if (ValTy->isFloatingPointTy()) 2166 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 2167 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 2168 2169 // If the requested value was a vector constant, create it. 2170 if (EltTy != ValTy) { 2171 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 2172 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 2173 StoreVal = ConstantVector::get(Elts); 2174 } 2175 } 2176 new StoreInst(StoreVal, EltPtr, MI); 2177 continue; 2178 } 2179 // Otherwise, if we're storing a byte variable, use a memset call for 2180 // this element. 2181 } 2182 2183 unsigned EltSize = TD->getTypeAllocSize(EltTy); 2184 2185 IRBuilder<> Builder(MI); 2186 2187 // Finally, insert the meminst for this element. 2188 if (isa<MemSetInst>(MI)) { 2189 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize, 2190 MI->isVolatile()); 2191 } else { 2192 assert(isa<MemTransferInst>(MI)); 2193 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr 2194 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr 2195 2196 if (isa<MemCpyInst>(MI)) 2197 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile()); 2198 else 2199 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile()); 2200 } 2201 } 2202 DeadInsts.push_back(MI); 2203} 2204 2205/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 2206/// overwrites the entire allocation. Extract out the pieces of the stored 2207/// integer and store them individually. 2208void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 2209 SmallVector<AllocaInst*, 32> &NewElts){ 2210 // Extract each element out of the integer according to its structure offset 2211 // and store the element value to the individual alloca. 2212 Value *SrcVal = SI->getOperand(0); 2213 const Type *AllocaEltTy = AI->getAllocatedType(); 2214 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2215 2216 IRBuilder<> Builder(SI); 2217 2218 // Handle tail padding by extending the operand 2219 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 2220 SrcVal = Builder.CreateZExt(SrcVal, 2221 IntegerType::get(SI->getContext(), AllocaSizeBits)); 2222 2223 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 2224 << '\n'); 2225 2226 // There are two forms here: AI could be an array or struct. Both cases 2227 // have different ways to compute the element offset. 2228 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2229 const StructLayout *Layout = TD->getStructLayout(EltSTy); 2230 2231 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2232 // Get the number of bits to shift SrcVal to get the value. 2233 const Type *FieldTy = EltSTy->getElementType(i); 2234 uint64_t Shift = Layout->getElementOffsetInBits(i); 2235 2236 if (TD->isBigEndian()) 2237 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 2238 2239 Value *EltVal = SrcVal; 2240 if (Shift) { 2241 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2242 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2243 } 2244 2245 // Truncate down to an integer of the right size. 2246 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2247 2248 // Ignore zero sized fields like {}, they obviously contain no data. 2249 if (FieldSizeBits == 0) continue; 2250 2251 if (FieldSizeBits != AllocaSizeBits) 2252 EltVal = Builder.CreateTrunc(EltVal, 2253 IntegerType::get(SI->getContext(), FieldSizeBits)); 2254 Value *DestField = NewElts[i]; 2255 if (EltVal->getType() == FieldTy) { 2256 // Storing to an integer field of this size, just do it. 2257 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 2258 // Bitcast to the right element type (for fp/vector values). 2259 EltVal = Builder.CreateBitCast(EltVal, FieldTy); 2260 } else { 2261 // Otherwise, bitcast the dest pointer (for aggregates). 2262 DestField = Builder.CreateBitCast(DestField, 2263 PointerType::getUnqual(EltVal->getType())); 2264 } 2265 new StoreInst(EltVal, DestField, SI); 2266 } 2267 2268 } else { 2269 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 2270 const Type *ArrayEltTy = ATy->getElementType(); 2271 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2272 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 2273 2274 uint64_t Shift; 2275 2276 if (TD->isBigEndian()) 2277 Shift = AllocaSizeBits-ElementOffset; 2278 else 2279 Shift = 0; 2280 2281 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2282 // Ignore zero sized fields like {}, they obviously contain no data. 2283 if (ElementSizeBits == 0) continue; 2284 2285 Value *EltVal = SrcVal; 2286 if (Shift) { 2287 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2288 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2289 } 2290 2291 // Truncate down to an integer of the right size. 2292 if (ElementSizeBits != AllocaSizeBits) 2293 EltVal = Builder.CreateTrunc(EltVal, 2294 IntegerType::get(SI->getContext(), 2295 ElementSizeBits)); 2296 Value *DestField = NewElts[i]; 2297 if (EltVal->getType() == ArrayEltTy) { 2298 // Storing to an integer field of this size, just do it. 2299 } else if (ArrayEltTy->isFloatingPointTy() || 2300 ArrayEltTy->isVectorTy()) { 2301 // Bitcast to the right element type (for fp/vector values). 2302 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy); 2303 } else { 2304 // Otherwise, bitcast the dest pointer (for aggregates). 2305 DestField = Builder.CreateBitCast(DestField, 2306 PointerType::getUnqual(EltVal->getType())); 2307 } 2308 new StoreInst(EltVal, DestField, SI); 2309 2310 if (TD->isBigEndian()) 2311 Shift -= ElementOffset; 2312 else 2313 Shift += ElementOffset; 2314 } 2315 } 2316 2317 DeadInsts.push_back(SI); 2318} 2319 2320/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 2321/// an integer. Load the individual pieces to form the aggregate value. 2322void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 2323 SmallVector<AllocaInst*, 32> &NewElts) { 2324 // Extract each element out of the NewElts according to its structure offset 2325 // and form the result value. 2326 const Type *AllocaEltTy = AI->getAllocatedType(); 2327 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2328 2329 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 2330 << '\n'); 2331 2332 // There are two forms here: AI could be an array or struct. Both cases 2333 // have different ways to compute the element offset. 2334 const StructLayout *Layout = 0; 2335 uint64_t ArrayEltBitOffset = 0; 2336 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2337 Layout = TD->getStructLayout(EltSTy); 2338 } else { 2339 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 2340 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2341 } 2342 2343 Value *ResultVal = 2344 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 2345 2346 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2347 // Load the value from the alloca. If the NewElt is an aggregate, cast 2348 // the pointer to an integer of the same size before doing the load. 2349 Value *SrcField = NewElts[i]; 2350 const Type *FieldTy = 2351 cast<PointerType>(SrcField->getType())->getElementType(); 2352 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2353 2354 // Ignore zero sized fields like {}, they obviously contain no data. 2355 if (FieldSizeBits == 0) continue; 2356 2357 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 2358 FieldSizeBits); 2359 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 2360 !FieldTy->isVectorTy()) 2361 SrcField = new BitCastInst(SrcField, 2362 PointerType::getUnqual(FieldIntTy), 2363 "", LI); 2364 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 2365 2366 // If SrcField is a fp or vector of the right size but that isn't an 2367 // integer type, bitcast to an integer so we can shift it. 2368 if (SrcField->getType() != FieldIntTy) 2369 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 2370 2371 // Zero extend the field to be the same size as the final alloca so that 2372 // we can shift and insert it. 2373 if (SrcField->getType() != ResultVal->getType()) 2374 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 2375 2376 // Determine the number of bits to shift SrcField. 2377 uint64_t Shift; 2378 if (Layout) // Struct case. 2379 Shift = Layout->getElementOffsetInBits(i); 2380 else // Array case. 2381 Shift = i*ArrayEltBitOffset; 2382 2383 if (TD->isBigEndian()) 2384 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 2385 2386 if (Shift) { 2387 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 2388 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 2389 } 2390 2391 // Don't create an 'or x, 0' on the first iteration. 2392 if (!isa<Constant>(ResultVal) || 2393 !cast<Constant>(ResultVal)->isNullValue()) 2394 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 2395 else 2396 ResultVal = SrcField; 2397 } 2398 2399 // Handle tail padding by truncating the result 2400 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 2401 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 2402 2403 LI->replaceAllUsesWith(ResultVal); 2404 DeadInsts.push_back(LI); 2405} 2406 2407/// HasPadding - Return true if the specified type has any structure or 2408/// alignment padding in between the elements that would be split apart 2409/// by SROA; return false otherwise. 2410static bool HasPadding(const Type *Ty, const TargetData &TD) { 2411 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 2412 Ty = ATy->getElementType(); 2413 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 2414 } 2415 2416 // SROA currently handles only Arrays and Structs. 2417 const StructType *STy = cast<StructType>(Ty); 2418 const StructLayout *SL = TD.getStructLayout(STy); 2419 unsigned PrevFieldBitOffset = 0; 2420 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 2421 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 2422 2423 // Check to see if there is any padding between this element and the 2424 // previous one. 2425 if (i) { 2426 unsigned PrevFieldEnd = 2427 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 2428 if (PrevFieldEnd < FieldBitOffset) 2429 return true; 2430 } 2431 PrevFieldBitOffset = FieldBitOffset; 2432 } 2433 // Check for tail padding. 2434 if (unsigned EltCount = STy->getNumElements()) { 2435 unsigned PrevFieldEnd = PrevFieldBitOffset + 2436 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 2437 if (PrevFieldEnd < SL->getSizeInBits()) 2438 return true; 2439 } 2440 return false; 2441} 2442 2443/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 2444/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 2445/// or 1 if safe after canonicalization has been performed. 2446bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 2447 // Loop over the use list of the alloca. We can only transform it if all of 2448 // the users are safe to transform. 2449 AllocaInfo Info(AI); 2450 2451 isSafeForScalarRepl(AI, 0, Info); 2452 if (Info.isUnsafe) { 2453 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 2454 return false; 2455 } 2456 2457 // Okay, we know all the users are promotable. If the aggregate is a memcpy 2458 // source and destination, we have to be careful. In particular, the memcpy 2459 // could be moving around elements that live in structure padding of the LLVM 2460 // types, but may actually be used. In these cases, we refuse to promote the 2461 // struct. 2462 if (Info.isMemCpySrc && Info.isMemCpyDst && 2463 HasPadding(AI->getAllocatedType(), *TD)) 2464 return false; 2465 2466 // If the alloca never has an access to just *part* of it, but is accessed 2467 // via loads and stores, then we should use ConvertToScalarInfo to promote 2468 // the alloca instead of promoting each piece at a time and inserting fission 2469 // and fusion code. 2470 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) { 2471 // If the struct/array just has one element, use basic SRoA. 2472 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 2473 if (ST->getNumElements() > 1) return false; 2474 } else { 2475 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1) 2476 return false; 2477 } 2478 } 2479 2480 return true; 2481} 2482 2483 2484 2485/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 2486/// some part of a constant global variable. This intentionally only accepts 2487/// constant expressions because we don't can't rewrite arbitrary instructions. 2488static bool PointsToConstantGlobal(Value *V) { 2489 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 2490 return GV->isConstant(); 2491 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 2492 if (CE->getOpcode() == Instruction::BitCast || 2493 CE->getOpcode() == Instruction::GetElementPtr) 2494 return PointsToConstantGlobal(CE->getOperand(0)); 2495 return false; 2496} 2497 2498/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 2499/// pointer to an alloca. Ignore any reads of the pointer, return false if we 2500/// see any stores or other unknown uses. If we see pointer arithmetic, keep 2501/// track of whether it moves the pointer (with isOffset) but otherwise traverse 2502/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 2503/// the alloca, and if the source pointer is a pointer to a constant global, we 2504/// can optimize this. 2505static bool isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, 2506 bool isOffset) { 2507 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 2508 User *U = cast<Instruction>(*UI); 2509 2510 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 2511 // Ignore non-volatile loads, they are always ok. 2512 if (LI->isVolatile()) return false; 2513 continue; 2514 } 2515 2516 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 2517 // If uses of the bitcast are ok, we are ok. 2518 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 2519 return false; 2520 continue; 2521 } 2522 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 2523 // If the GEP has all zero indices, it doesn't offset the pointer. If it 2524 // doesn't, it does. 2525 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 2526 isOffset || !GEP->hasAllZeroIndices())) 2527 return false; 2528 continue; 2529 } 2530 2531 if (CallSite CS = U) { 2532 // If this is the function being called then we treat it like a load and 2533 // ignore it. 2534 if (CS.isCallee(UI)) 2535 continue; 2536 2537 // If this is a readonly/readnone call site, then we know it is just a 2538 // load (but one that potentially returns the value itself), so we can 2539 // ignore it if we know that the value isn't captured. 2540 unsigned ArgNo = CS.getArgumentNo(UI); 2541 if (CS.onlyReadsMemory() && 2542 (CS.getInstruction()->use_empty() || 2543 CS.paramHasAttr(ArgNo+1, Attribute::NoCapture))) 2544 continue; 2545 2546 // If this is being passed as a byval argument, the caller is making a 2547 // copy, so it is only a read of the alloca. 2548 if (CS.paramHasAttr(ArgNo+1, Attribute::ByVal)) 2549 continue; 2550 } 2551 2552 // If this is isn't our memcpy/memmove, reject it as something we can't 2553 // handle. 2554 MemTransferInst *MI = dyn_cast<MemTransferInst>(U); 2555 if (MI == 0) 2556 return false; 2557 2558 // If the transfer is using the alloca as a source of the transfer, then 2559 // ignore it since it is a load (unless the transfer is volatile). 2560 if (UI.getOperandNo() == 1) { 2561 if (MI->isVolatile()) return false; 2562 continue; 2563 } 2564 2565 // If we already have seen a copy, reject the second one. 2566 if (TheCopy) return false; 2567 2568 // If the pointer has been offset from the start of the alloca, we can't 2569 // safely handle this. 2570 if (isOffset) return false; 2571 2572 // If the memintrinsic isn't using the alloca as the dest, reject it. 2573 if (UI.getOperandNo() != 0) return false; 2574 2575 // If the source of the memcpy/move is not a constant global, reject it. 2576 if (!PointsToConstantGlobal(MI->getSource())) 2577 return false; 2578 2579 // Otherwise, the transform is safe. Remember the copy instruction. 2580 TheCopy = MI; 2581 } 2582 return true; 2583} 2584 2585/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 2586/// modified by a copy from a constant global. If we can prove this, we can 2587/// replace any uses of the alloca with uses of the global directly. 2588MemTransferInst *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) { 2589 MemTransferInst *TheCopy = 0; 2590 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 2591 return TheCopy; 2592 return 0; 2593} 2594