ScalarReplAggregates.cpp revision 36b56886974eae4f9c5ebc96befd3e7bfe5de338
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 they 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/ADT/SetVector.h" 25#include "llvm/ADT/SmallVector.h" 26#include "llvm/ADT/Statistic.h" 27#include "llvm/Analysis/Loads.h" 28#include "llvm/Analysis/ValueTracking.h" 29#include "llvm/IR/CallSite.h" 30#include "llvm/IR/Constants.h" 31#include "llvm/IR/DIBuilder.h" 32#include "llvm/IR/DataLayout.h" 33#include "llvm/IR/DebugInfo.h" 34#include "llvm/IR/DerivedTypes.h" 35#include "llvm/IR/Dominators.h" 36#include "llvm/IR/Function.h" 37#include "llvm/IR/GetElementPtrTypeIterator.h" 38#include "llvm/IR/GlobalVariable.h" 39#include "llvm/IR/IRBuilder.h" 40#include "llvm/IR/Instructions.h" 41#include "llvm/IR/IntrinsicInst.h" 42#include "llvm/IR/LLVMContext.h" 43#include "llvm/IR/Module.h" 44#include "llvm/IR/Operator.h" 45#include "llvm/Pass.h" 46#include "llvm/Support/Debug.h" 47#include "llvm/Support/ErrorHandling.h" 48#include "llvm/Support/MathExtras.h" 49#include "llvm/Support/raw_ostream.h" 50#include "llvm/Transforms/Utils/Local.h" 51#include "llvm/Transforms/Utils/PromoteMemToReg.h" 52#include "llvm/Transforms/Utils/SSAUpdater.h" 53using namespace llvm; 54 55STATISTIC(NumReplaced, "Number of allocas broken up"); 56STATISTIC(NumPromoted, "Number of allocas promoted"); 57STATISTIC(NumAdjusted, "Number of scalar allocas adjusted to allow promotion"); 58STATISTIC(NumConverted, "Number of aggregates converted to scalar"); 59 60namespace { 61 struct SROA : public FunctionPass { 62 SROA(int T, bool hasDT, char &ID, int ST, int AT, int SLT) 63 : FunctionPass(ID), HasDomTree(hasDT) { 64 if (T == -1) 65 SRThreshold = 128; 66 else 67 SRThreshold = T; 68 if (ST == -1) 69 StructMemberThreshold = 32; 70 else 71 StructMemberThreshold = ST; 72 if (AT == -1) 73 ArrayElementThreshold = 8; 74 else 75 ArrayElementThreshold = AT; 76 if (SLT == -1) 77 // Do not limit the scalar integer load size if no threshold is given. 78 ScalarLoadThreshold = -1; 79 else 80 ScalarLoadThreshold = SLT; 81 } 82 83 bool runOnFunction(Function &F) override; 84 85 bool performScalarRepl(Function &F); 86 bool performPromotion(Function &F); 87 88 private: 89 bool HasDomTree; 90 const DataLayout *DL; 91 92 /// DeadInsts - Keep track of instructions we have made dead, so that 93 /// we can remove them after we are done working. 94 SmallVector<Value*, 32> DeadInsts; 95 96 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures 97 /// information about the uses. All these fields are initialized to false 98 /// and set to true when something is learned. 99 struct AllocaInfo { 100 /// The alloca to promote. 101 AllocaInst *AI; 102 103 /// CheckedPHIs - This is a set of verified PHI nodes, to prevent infinite 104 /// looping and avoid redundant work. 105 SmallPtrSet<PHINode*, 8> CheckedPHIs; 106 107 /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 108 bool isUnsafe : 1; 109 110 /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 111 bool isMemCpySrc : 1; 112 113 /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 114 bool isMemCpyDst : 1; 115 116 /// hasSubelementAccess - This is true if a subelement of the alloca is 117 /// ever accessed, or false if the alloca is only accessed with mem 118 /// intrinsics or load/store that only access the entire alloca at once. 119 bool hasSubelementAccess : 1; 120 121 /// hasALoadOrStore - This is true if there are any loads or stores to it. 122 /// The alloca may just be accessed with memcpy, for example, which would 123 /// not set this. 124 bool hasALoadOrStore : 1; 125 126 explicit AllocaInfo(AllocaInst *ai) 127 : AI(ai), isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false), 128 hasSubelementAccess(false), hasALoadOrStore(false) {} 129 }; 130 131 /// SRThreshold - The maximum alloca size to considered for SROA. 132 unsigned SRThreshold; 133 134 /// StructMemberThreshold - The maximum number of members a struct can 135 /// contain to be considered for SROA. 136 unsigned StructMemberThreshold; 137 138 /// ArrayElementThreshold - The maximum number of elements an array can 139 /// have to be considered for SROA. 140 unsigned ArrayElementThreshold; 141 142 /// ScalarLoadThreshold - The maximum size in bits of scalars to load when 143 /// converting to scalar 144 unsigned ScalarLoadThreshold; 145 146 void MarkUnsafe(AllocaInfo &I, Instruction *User) { 147 I.isUnsafe = true; 148 DEBUG(dbgs() << " Transformation preventing inst: " << *User << '\n'); 149 } 150 151 bool isSafeAllocaToScalarRepl(AllocaInst *AI); 152 153 void isSafeForScalarRepl(Instruction *I, uint64_t Offset, AllocaInfo &Info); 154 void isSafePHISelectUseForScalarRepl(Instruction *User, uint64_t Offset, 155 AllocaInfo &Info); 156 void isSafeGEP(GetElementPtrInst *GEPI, uint64_t &Offset, AllocaInfo &Info); 157 void isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 158 Type *MemOpType, bool isStore, AllocaInfo &Info, 159 Instruction *TheAccess, bool AllowWholeAccess); 160 bool TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size); 161 uint64_t FindElementAndOffset(Type *&T, uint64_t &Offset, 162 Type *&IdxTy); 163 164 void DoScalarReplacement(AllocaInst *AI, 165 std::vector<AllocaInst*> &WorkList); 166 void DeleteDeadInstructions(); 167 168 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 169 SmallVectorImpl<AllocaInst *> &NewElts); 170 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 171 SmallVectorImpl<AllocaInst *> &NewElts); 172 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 173 SmallVectorImpl<AllocaInst *> &NewElts); 174 void RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI, 175 uint64_t Offset, 176 SmallVectorImpl<AllocaInst *> &NewElts); 177 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 178 AllocaInst *AI, 179 SmallVectorImpl<AllocaInst *> &NewElts); 180 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 181 SmallVectorImpl<AllocaInst *> &NewElts); 182 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 183 SmallVectorImpl<AllocaInst *> &NewElts); 184 bool ShouldAttemptScalarRepl(AllocaInst *AI); 185 }; 186 187 // SROA_DT - SROA that uses DominatorTree. 188 struct SROA_DT : public SROA { 189 static char ID; 190 public: 191 SROA_DT(int T = -1, int ST = -1, int AT = -1, int SLT = -1) : 192 SROA(T, true, ID, ST, AT, SLT) { 193 initializeSROA_DTPass(*PassRegistry::getPassRegistry()); 194 } 195 196 // getAnalysisUsage - This pass does not require any passes, but we know it 197 // will not alter the CFG, so say so. 198 void getAnalysisUsage(AnalysisUsage &AU) const override { 199 AU.addRequired<DominatorTreeWrapperPass>(); 200 AU.setPreservesCFG(); 201 } 202 }; 203 204 // SROA_SSAUp - SROA that uses SSAUpdater. 205 struct SROA_SSAUp : public SROA { 206 static char ID; 207 public: 208 SROA_SSAUp(int T = -1, int ST = -1, int AT = -1, int SLT = -1) : 209 SROA(T, false, ID, ST, AT, SLT) { 210 initializeSROA_SSAUpPass(*PassRegistry::getPassRegistry()); 211 } 212 213 // getAnalysisUsage - This pass does not require any passes, but we know it 214 // will not alter the CFG, so say so. 215 void getAnalysisUsage(AnalysisUsage &AU) const override { 216 AU.setPreservesCFG(); 217 } 218 }; 219 220} 221 222char SROA_DT::ID = 0; 223char SROA_SSAUp::ID = 0; 224 225INITIALIZE_PASS_BEGIN(SROA_DT, "scalarrepl", 226 "Scalar Replacement of Aggregates (DT)", false, false) 227INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 228INITIALIZE_PASS_END(SROA_DT, "scalarrepl", 229 "Scalar Replacement of Aggregates (DT)", false, false) 230 231INITIALIZE_PASS_BEGIN(SROA_SSAUp, "scalarrepl-ssa", 232 "Scalar Replacement of Aggregates (SSAUp)", false, false) 233INITIALIZE_PASS_END(SROA_SSAUp, "scalarrepl-ssa", 234 "Scalar Replacement of Aggregates (SSAUp)", false, false) 235 236// Public interface to the ScalarReplAggregates pass 237FunctionPass *llvm::createScalarReplAggregatesPass(int Threshold, 238 bool UseDomTree, 239 int StructMemberThreshold, 240 int ArrayElementThreshold, 241 int ScalarLoadThreshold) { 242 if (UseDomTree) 243 return new SROA_DT(Threshold, StructMemberThreshold, ArrayElementThreshold, 244 ScalarLoadThreshold); 245 return new SROA_SSAUp(Threshold, StructMemberThreshold, 246 ArrayElementThreshold, ScalarLoadThreshold); 247} 248 249 250//===----------------------------------------------------------------------===// 251// Convert To Scalar Optimization. 252//===----------------------------------------------------------------------===// 253 254namespace { 255/// ConvertToScalarInfo - This class implements the "Convert To Scalar" 256/// optimization, which scans the uses of an alloca and determines if it can 257/// rewrite it in terms of a single new alloca that can be mem2reg'd. 258class ConvertToScalarInfo { 259 /// AllocaSize - The size of the alloca being considered in bytes. 260 unsigned AllocaSize; 261 const DataLayout &DL; 262 unsigned ScalarLoadThreshold; 263 264 /// IsNotTrivial - This is set to true if there is some access to the object 265 /// which means that mem2reg can't promote it. 266 bool IsNotTrivial; 267 268 /// ScalarKind - Tracks the kind of alloca being considered for promotion, 269 /// computed based on the uses of the alloca rather than the LLVM type system. 270 enum { 271 Unknown, 272 273 // Accesses via GEPs that are consistent with element access of a vector 274 // type. This will not be converted into a vector unless there is a later 275 // access using an actual vector type. 276 ImplicitVector, 277 278 // Accesses via vector operations and GEPs that are consistent with the 279 // layout of a vector type. 280 Vector, 281 282 // An integer bag-of-bits with bitwise operations for insertion and 283 // extraction. Any combination of types can be converted into this kind 284 // of scalar. 285 Integer 286 } ScalarKind; 287 288 /// VectorTy - This tracks the type that we should promote the vector to if 289 /// it is possible to turn it into a vector. This starts out null, and if it 290 /// isn't possible to turn into a vector type, it gets set to VoidTy. 291 VectorType *VectorTy; 292 293 /// HadNonMemTransferAccess - True if there is at least one access to the 294 /// alloca that is not a MemTransferInst. We don't want to turn structs into 295 /// large integers unless there is some potential for optimization. 296 bool HadNonMemTransferAccess; 297 298 /// HadDynamicAccess - True if some element of this alloca was dynamic. 299 /// We don't yet have support for turning a dynamic access into a large 300 /// integer. 301 bool HadDynamicAccess; 302 303public: 304 explicit ConvertToScalarInfo(unsigned Size, const DataLayout &DL, 305 unsigned SLT) 306 : AllocaSize(Size), DL(DL), ScalarLoadThreshold(SLT), IsNotTrivial(false), 307 ScalarKind(Unknown), VectorTy(0), HadNonMemTransferAccess(false), 308 HadDynamicAccess(false) { } 309 310 AllocaInst *TryConvert(AllocaInst *AI); 311 312private: 313 bool CanConvertToScalar(Value *V, uint64_t Offset, Value* NonConstantIdx); 314 void MergeInTypeForLoadOrStore(Type *In, uint64_t Offset); 315 bool MergeInVectorType(VectorType *VInTy, uint64_t Offset); 316 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset, 317 Value *NonConstantIdx); 318 319 Value *ConvertScalar_ExtractValue(Value *NV, Type *ToType, 320 uint64_t Offset, Value* NonConstantIdx, 321 IRBuilder<> &Builder); 322 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 323 uint64_t Offset, Value* NonConstantIdx, 324 IRBuilder<> &Builder); 325}; 326} // end anonymous namespace. 327 328 329/// TryConvert - Analyze the specified alloca, and if it is safe to do so, 330/// rewrite it to be a new alloca which is mem2reg'able. This returns the new 331/// alloca if possible or null if not. 332AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) { 333 // If we can't convert this scalar, or if mem2reg can trivially do it, bail 334 // out. 335 if (!CanConvertToScalar(AI, 0, 0) || !IsNotTrivial) 336 return 0; 337 338 // If an alloca has only memset / memcpy uses, it may still have an Unknown 339 // ScalarKind. Treat it as an Integer below. 340 if (ScalarKind == Unknown) 341 ScalarKind = Integer; 342 343 if (ScalarKind == Vector && VectorTy->getBitWidth() != AllocaSize * 8) 344 ScalarKind = Integer; 345 346 // If we were able to find a vector type that can handle this with 347 // insert/extract elements, and if there was at least one use that had 348 // a vector type, promote this to a vector. We don't want to promote 349 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 350 // we just get a lot of insert/extracts. If at least one vector is 351 // involved, then we probably really do have a union of vector/array. 352 Type *NewTy; 353 if (ScalarKind == Vector) { 354 assert(VectorTy && "Missing type for vector scalar."); 355 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = " 356 << *VectorTy << '\n'); 357 NewTy = VectorTy; // Use the vector type. 358 } else { 359 unsigned BitWidth = AllocaSize * 8; 360 361 // Do not convert to scalar integer if the alloca size exceeds the 362 // scalar load threshold. 363 if (BitWidth > ScalarLoadThreshold) 364 return 0; 365 366 if ((ScalarKind == ImplicitVector || ScalarKind == Integer) && 367 !HadNonMemTransferAccess && !DL.fitsInLegalInteger(BitWidth)) 368 return 0; 369 // Dynamic accesses on integers aren't yet supported. They need us to shift 370 // by a dynamic amount which could be difficult to work out as we might not 371 // know whether to use a left or right shift. 372 if (ScalarKind == Integer && HadDynamicAccess) 373 return 0; 374 375 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"); 376 // Create and insert the integer alloca. 377 NewTy = IntegerType::get(AI->getContext(), BitWidth); 378 } 379 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 380 ConvertUsesToScalar(AI, NewAI, 0, 0); 381 return NewAI; 382} 383 384/// MergeInTypeForLoadOrStore - Add the 'In' type to the accumulated vector type 385/// (VectorTy) so far at the offset specified by Offset (which is specified in 386/// bytes). 387/// 388/// There are two cases we handle here: 389/// 1) A union of vector types of the same size and potentially its elements. 390/// Here we turn element accesses into insert/extract element operations. 391/// This promotes a <4 x float> with a store of float to the third element 392/// into a <4 x float> that uses insert element. 393/// 2) A fully general blob of memory, which we turn into some (potentially 394/// large) integer type with extract and insert operations where the loads 395/// and stores would mutate the memory. We mark this by setting VectorTy 396/// to VoidTy. 397void ConvertToScalarInfo::MergeInTypeForLoadOrStore(Type *In, 398 uint64_t Offset) { 399 // If we already decided to turn this into a blob of integer memory, there is 400 // nothing to be done. 401 if (ScalarKind == Integer) 402 return; 403 404 // If this could be contributing to a vector, analyze it. 405 406 // If the In type is a vector that is the same size as the alloca, see if it 407 // matches the existing VecTy. 408 if (VectorType *VInTy = dyn_cast<VectorType>(In)) { 409 if (MergeInVectorType(VInTy, Offset)) 410 return; 411 } else if (In->isFloatTy() || In->isDoubleTy() || 412 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 && 413 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 414 // Full width accesses can be ignored, because they can always be turned 415 // into bitcasts. 416 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 417 if (EltSize == AllocaSize) 418 return; 419 420 // If we're accessing something that could be an element of a vector, see 421 // if the implied vector agrees with what we already have and if Offset is 422 // compatible with it. 423 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 && 424 (!VectorTy || EltSize == VectorTy->getElementType() 425 ->getPrimitiveSizeInBits()/8)) { 426 if (!VectorTy) { 427 ScalarKind = ImplicitVector; 428 VectorTy = VectorType::get(In, AllocaSize/EltSize); 429 } 430 return; 431 } 432 } 433 434 // Otherwise, we have a case that we can't handle with an optimized vector 435 // form. We can still turn this into a large integer. 436 ScalarKind = Integer; 437} 438 439/// MergeInVectorType - Handles the vector case of MergeInTypeForLoadOrStore, 440/// returning true if the type was successfully merged and false otherwise. 441bool ConvertToScalarInfo::MergeInVectorType(VectorType *VInTy, 442 uint64_t Offset) { 443 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 444 // If we're storing/loading a vector of the right size, allow it as a 445 // vector. If this the first vector we see, remember the type so that 446 // we know the element size. If this is a subsequent access, ignore it 447 // even if it is a differing type but the same size. Worst case we can 448 // bitcast the resultant vectors. 449 if (!VectorTy) 450 VectorTy = VInTy; 451 ScalarKind = Vector; 452 return true; 453 } 454 455 return false; 456} 457 458/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 459/// its accesses to a single vector type, return true and set VecTy to 460/// the new type. If we could convert the alloca into a single promotable 461/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 462/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 463/// is the current offset from the base of the alloca being analyzed. 464/// 465/// If we see at least one access to the value that is as a vector type, set the 466/// SawVec flag. 467bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset, 468 Value* NonConstantIdx) { 469 for (User *U : V->users()) { 470 Instruction *UI = cast<Instruction>(U); 471 472 if (LoadInst *LI = dyn_cast<LoadInst>(UI)) { 473 // Don't break volatile loads. 474 if (!LI->isSimple()) 475 return false; 476 // Don't touch MMX operations. 477 if (LI->getType()->isX86_MMXTy()) 478 return false; 479 HadNonMemTransferAccess = true; 480 MergeInTypeForLoadOrStore(LI->getType(), Offset); 481 continue; 482 } 483 484 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) { 485 // Storing the pointer, not into the value? 486 if (SI->getOperand(0) == V || !SI->isSimple()) return false; 487 // Don't touch MMX operations. 488 if (SI->getOperand(0)->getType()->isX86_MMXTy()) 489 return false; 490 HadNonMemTransferAccess = true; 491 MergeInTypeForLoadOrStore(SI->getOperand(0)->getType(), Offset); 492 continue; 493 } 494 495 if (BitCastInst *BCI = dyn_cast<BitCastInst>(UI)) { 496 if (!onlyUsedByLifetimeMarkers(BCI)) 497 IsNotTrivial = true; // Can't be mem2reg'd. 498 if (!CanConvertToScalar(BCI, Offset, NonConstantIdx)) 499 return false; 500 continue; 501 } 502 503 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(UI)) { 504 // If this is a GEP with a variable indices, we can't handle it. 505 PointerType* PtrTy = dyn_cast<PointerType>(GEP->getPointerOperandType()); 506 if (!PtrTy) 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 Value *GEPNonConstantIdx = 0; 512 if (!GEP->hasAllConstantIndices()) { 513 if (!isa<VectorType>(PtrTy->getElementType())) 514 return false; 515 if (NonConstantIdx) 516 return false; 517 GEPNonConstantIdx = Indices.pop_back_val(); 518 if (!GEPNonConstantIdx->getType()->isIntegerTy(32)) 519 return false; 520 HadDynamicAccess = true; 521 } else 522 GEPNonConstantIdx = NonConstantIdx; 523 uint64_t GEPOffset = DL.getIndexedOffset(PtrTy, 524 Indices); 525 // See if all uses can be converted. 526 if (!CanConvertToScalar(GEP, Offset+GEPOffset, GEPNonConstantIdx)) 527 return false; 528 IsNotTrivial = true; // Can't be mem2reg'd. 529 HadNonMemTransferAccess = true; 530 continue; 531 } 532 533 // If this is a constant sized memset of a constant value (e.g. 0) we can 534 // handle it. 535 if (MemSetInst *MSI = dyn_cast<MemSetInst>(UI)) { 536 // Store to dynamic index. 537 if (NonConstantIdx) 538 return false; 539 // Store of constant value. 540 if (!isa<ConstantInt>(MSI->getValue())) 541 return false; 542 543 // Store of constant size. 544 ConstantInt *Len = dyn_cast<ConstantInt>(MSI->getLength()); 545 if (!Len) 546 return false; 547 548 // If the size differs from the alloca, we can only convert the alloca to 549 // an integer bag-of-bits. 550 // FIXME: This should handle all of the cases that are currently accepted 551 // as vector element insertions. 552 if (Len->getZExtValue() != AllocaSize || Offset != 0) 553 ScalarKind = Integer; 554 555 IsNotTrivial = true; // Can't be mem2reg'd. 556 HadNonMemTransferAccess = true; 557 continue; 558 } 559 560 // If this is a memcpy or memmove into or out of the whole allocation, we 561 // can handle it like a load or store of the scalar type. 562 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(UI)) { 563 // Store to dynamic index. 564 if (NonConstantIdx) 565 return false; 566 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()); 567 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0) 568 return false; 569 570 IsNotTrivial = true; // Can't be mem2reg'd. 571 continue; 572 } 573 574 // If this is a lifetime intrinsic, we can handle it. 575 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(UI)) { 576 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 577 II->getIntrinsicID() == Intrinsic::lifetime_end) { 578 continue; 579 } 580 } 581 582 // Otherwise, we cannot handle this! 583 return false; 584 } 585 586 return true; 587} 588 589/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 590/// directly. This happens when we are converting an "integer union" to a 591/// single integer scalar, or when we are converting a "vector union" to a 592/// vector with insert/extractelement instructions. 593/// 594/// Offset is an offset from the original alloca, in bits that need to be 595/// shifted to the right. By the end of this, there should be no uses of Ptr. 596void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, 597 uint64_t Offset, 598 Value* NonConstantIdx) { 599 while (!Ptr->use_empty()) { 600 Instruction *User = cast<Instruction>(Ptr->user_back()); 601 602 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 603 ConvertUsesToScalar(CI, NewAI, Offset, NonConstantIdx); 604 CI->eraseFromParent(); 605 continue; 606 } 607 608 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 609 // Compute the offset that this GEP adds to the pointer. 610 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 611 Value* GEPNonConstantIdx = 0; 612 if (!GEP->hasAllConstantIndices()) { 613 assert(!NonConstantIdx && 614 "Dynamic GEP reading from dynamic GEP unsupported"); 615 GEPNonConstantIdx = Indices.pop_back_val(); 616 } else 617 GEPNonConstantIdx = NonConstantIdx; 618 uint64_t GEPOffset = DL.getIndexedOffset(GEP->getPointerOperandType(), 619 Indices); 620 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8, GEPNonConstantIdx); 621 GEP->eraseFromParent(); 622 continue; 623 } 624 625 IRBuilder<> Builder(User); 626 627 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 628 // The load is a bit extract from NewAI shifted right by Offset bits. 629 Value *LoadedVal = Builder.CreateLoad(NewAI); 630 Value *NewLoadVal 631 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, 632 NonConstantIdx, Builder); 633 LI->replaceAllUsesWith(NewLoadVal); 634 LI->eraseFromParent(); 635 continue; 636 } 637 638 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 639 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 640 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 641 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 642 NonConstantIdx, Builder); 643 Builder.CreateStore(New, NewAI); 644 SI->eraseFromParent(); 645 646 // If the load we just inserted is now dead, then the inserted store 647 // overwrote the entire thing. 648 if (Old->use_empty()) 649 Old->eraseFromParent(); 650 continue; 651 } 652 653 // If this is a constant sized memset of a constant value (e.g. 0) we can 654 // transform it into a store of the expanded constant value. 655 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 656 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 657 assert(!NonConstantIdx && "Cannot replace dynamic memset with insert"); 658 int64_t SNumBytes = cast<ConstantInt>(MSI->getLength())->getSExtValue(); 659 if (SNumBytes > 0 && (SNumBytes >> 32) == 0) { 660 unsigned NumBytes = static_cast<unsigned>(SNumBytes); 661 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 662 663 // Compute the value replicated the right number of times. 664 APInt APVal(NumBytes*8, Val); 665 666 // Splat the value if non-zero. 667 if (Val) 668 for (unsigned i = 1; i != NumBytes; ++i) 669 APVal |= APVal << 8; 670 671 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 672 Value *New = ConvertScalar_InsertValue( 673 ConstantInt::get(User->getContext(), APVal), 674 Old, Offset, 0, Builder); 675 Builder.CreateStore(New, NewAI); 676 677 // If the load we just inserted is now dead, then the memset overwrote 678 // the entire thing. 679 if (Old->use_empty()) 680 Old->eraseFromParent(); 681 } 682 MSI->eraseFromParent(); 683 continue; 684 } 685 686 // If this is a memcpy or memmove into or out of the whole allocation, we 687 // can handle it like a load or store of the scalar type. 688 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 689 assert(Offset == 0 && "must be store to start of alloca"); 690 assert(!NonConstantIdx && "Cannot replace dynamic transfer with insert"); 691 692 // If the source and destination are both to the same alloca, then this is 693 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 694 // as appropriate. 695 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, &DL, 0)); 696 697 if (GetUnderlyingObject(MTI->getSource(), &DL, 0) != OrigAI) { 698 // Dest must be OrigAI, change this to be a load from the original 699 // pointer (bitcasted), then a store to our new alloca. 700 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 701 Value *SrcPtr = MTI->getSource(); 702 PointerType* SPTy = cast<PointerType>(SrcPtr->getType()); 703 PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 704 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 705 AIPTy = PointerType::get(AIPTy->getElementType(), 706 SPTy->getAddressSpace()); 707 } 708 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy); 709 710 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 711 SrcVal->setAlignment(MTI->getAlignment()); 712 Builder.CreateStore(SrcVal, NewAI); 713 } else if (GetUnderlyingObject(MTI->getDest(), &DL, 0) != OrigAI) { 714 // Src must be OrigAI, change this to be a load from NewAI then a store 715 // through the original dest pointer (bitcasted). 716 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 717 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 718 719 PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType()); 720 PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 721 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 722 AIPTy = PointerType::get(AIPTy->getElementType(), 723 DPTy->getAddressSpace()); 724 } 725 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy); 726 727 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 728 NewStore->setAlignment(MTI->getAlignment()); 729 } else { 730 // Noop transfer. Src == Dst 731 } 732 733 MTI->eraseFromParent(); 734 continue; 735 } 736 737 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 738 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 739 II->getIntrinsicID() == Intrinsic::lifetime_end) { 740 // There's no need to preserve these, as the resulting alloca will be 741 // converted to a register anyways. 742 II->eraseFromParent(); 743 continue; 744 } 745 } 746 747 llvm_unreachable("Unsupported operation!"); 748 } 749} 750 751/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 752/// or vector value FromVal, extracting the bits from the offset specified by 753/// Offset. This returns the value, which is of type ToType. 754/// 755/// This happens when we are converting an "integer union" to a single 756/// integer scalar, or when we are converting a "vector union" to a vector with 757/// insert/extractelement instructions. 758/// 759/// Offset is an offset from the original alloca, in bits that need to be 760/// shifted to the right. 761Value *ConvertToScalarInfo:: 762ConvertScalar_ExtractValue(Value *FromVal, Type *ToType, 763 uint64_t Offset, Value* NonConstantIdx, 764 IRBuilder<> &Builder) { 765 // If the load is of the whole new alloca, no conversion is needed. 766 Type *FromType = FromVal->getType(); 767 if (FromType == ToType && Offset == 0) 768 return FromVal; 769 770 // If the result alloca is a vector type, this is either an element 771 // access or a bitcast to another vector type of the same size. 772 if (VectorType *VTy = dyn_cast<VectorType>(FromType)) { 773 unsigned FromTypeSize = DL.getTypeAllocSize(FromType); 774 unsigned ToTypeSize = DL.getTypeAllocSize(ToType); 775 if (FromTypeSize == ToTypeSize) 776 return Builder.CreateBitCast(FromVal, ToType); 777 778 // Otherwise it must be an element access. 779 unsigned Elt = 0; 780 if (Offset) { 781 unsigned EltSize = DL.getTypeAllocSizeInBits(VTy->getElementType()); 782 Elt = Offset/EltSize; 783 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 784 } 785 // Return the element extracted out of it. 786 Value *Idx; 787 if (NonConstantIdx) { 788 if (Elt) 789 Idx = Builder.CreateAdd(NonConstantIdx, 790 Builder.getInt32(Elt), 791 "dyn.offset"); 792 else 793 Idx = NonConstantIdx; 794 } else 795 Idx = Builder.getInt32(Elt); 796 Value *V = Builder.CreateExtractElement(FromVal, Idx); 797 if (V->getType() != ToType) 798 V = Builder.CreateBitCast(V, ToType); 799 return V; 800 } 801 802 // If ToType is a first class aggregate, extract out each of the pieces and 803 // use insertvalue's to form the FCA. 804 if (StructType *ST = dyn_cast<StructType>(ToType)) { 805 assert(!NonConstantIdx && 806 "Dynamic indexing into struct types not supported"); 807 const StructLayout &Layout = *DL.getStructLayout(ST); 808 Value *Res = UndefValue::get(ST); 809 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 810 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 811 Offset+Layout.getElementOffsetInBits(i), 812 0, Builder); 813 Res = Builder.CreateInsertValue(Res, Elt, i); 814 } 815 return Res; 816 } 817 818 if (ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 819 assert(!NonConstantIdx && 820 "Dynamic indexing into array types not supported"); 821 uint64_t EltSize = DL.getTypeAllocSizeInBits(AT->getElementType()); 822 Value *Res = UndefValue::get(AT); 823 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 824 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 825 Offset+i*EltSize, 0, Builder); 826 Res = Builder.CreateInsertValue(Res, Elt, i); 827 } 828 return Res; 829 } 830 831 // Otherwise, this must be a union that was converted to an integer value. 832 IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 833 834 // If this is a big-endian system and the load is narrower than the 835 // full alloca type, we need to do a shift to get the right bits. 836 int ShAmt = 0; 837 if (DL.isBigEndian()) { 838 // On big-endian machines, the lowest bit is stored at the bit offset 839 // from the pointer given by getTypeStoreSizeInBits. This matters for 840 // integers with a bitwidth that is not a multiple of 8. 841 ShAmt = DL.getTypeStoreSizeInBits(NTy) - 842 DL.getTypeStoreSizeInBits(ToType) - Offset; 843 } else { 844 ShAmt = Offset; 845 } 846 847 // Note: we support negative bitwidths (with shl) which are not defined. 848 // We do this to support (f.e.) loads off the end of a structure where 849 // only some bits are used. 850 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 851 FromVal = Builder.CreateLShr(FromVal, 852 ConstantInt::get(FromVal->getType(), ShAmt)); 853 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 854 FromVal = Builder.CreateShl(FromVal, 855 ConstantInt::get(FromVal->getType(), -ShAmt)); 856 857 // Finally, unconditionally truncate the integer to the right width. 858 unsigned LIBitWidth = DL.getTypeSizeInBits(ToType); 859 if (LIBitWidth < NTy->getBitWidth()) 860 FromVal = 861 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 862 LIBitWidth)); 863 else if (LIBitWidth > NTy->getBitWidth()) 864 FromVal = 865 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 866 LIBitWidth)); 867 868 // If the result is an integer, this is a trunc or bitcast. 869 if (ToType->isIntegerTy()) { 870 // Should be done. 871 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 872 // Just do a bitcast, we know the sizes match up. 873 FromVal = Builder.CreateBitCast(FromVal, ToType); 874 } else { 875 // Otherwise must be a pointer. 876 FromVal = Builder.CreateIntToPtr(FromVal, ToType); 877 } 878 assert(FromVal->getType() == ToType && "Didn't convert right?"); 879 return FromVal; 880} 881 882/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 883/// or vector value "Old" at the offset specified by Offset. 884/// 885/// This happens when we are converting an "integer union" to a 886/// single integer scalar, or when we are converting a "vector union" to a 887/// vector with insert/extractelement instructions. 888/// 889/// Offset is an offset from the original alloca, in bits that need to be 890/// shifted to the right. 891/// 892/// NonConstantIdx is an index value if there was a GEP with a non-constant 893/// index value. If this is 0 then all GEPs used to find this insert address 894/// are constant. 895Value *ConvertToScalarInfo:: 896ConvertScalar_InsertValue(Value *SV, Value *Old, 897 uint64_t Offset, Value* NonConstantIdx, 898 IRBuilder<> &Builder) { 899 // Convert the stored type to the actual type, shift it left to insert 900 // then 'or' into place. 901 Type *AllocaType = Old->getType(); 902 LLVMContext &Context = Old->getContext(); 903 904 if (VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 905 uint64_t VecSize = DL.getTypeAllocSizeInBits(VTy); 906 uint64_t ValSize = DL.getTypeAllocSizeInBits(SV->getType()); 907 908 // Changing the whole vector with memset or with an access of a different 909 // vector type? 910 if (ValSize == VecSize) 911 return Builder.CreateBitCast(SV, AllocaType); 912 913 // Must be an element insertion. 914 Type *EltTy = VTy->getElementType(); 915 if (SV->getType() != EltTy) 916 SV = Builder.CreateBitCast(SV, EltTy); 917 uint64_t EltSize = DL.getTypeAllocSizeInBits(EltTy); 918 unsigned Elt = Offset/EltSize; 919 Value *Idx; 920 if (NonConstantIdx) { 921 if (Elt) 922 Idx = Builder.CreateAdd(NonConstantIdx, 923 Builder.getInt32(Elt), 924 "dyn.offset"); 925 else 926 Idx = NonConstantIdx; 927 } else 928 Idx = Builder.getInt32(Elt); 929 return Builder.CreateInsertElement(Old, SV, Idx); 930 } 931 932 // If SV is a first-class aggregate value, insert each value recursively. 933 if (StructType *ST = dyn_cast<StructType>(SV->getType())) { 934 assert(!NonConstantIdx && 935 "Dynamic indexing into struct types not supported"); 936 const StructLayout &Layout = *DL.getStructLayout(ST); 937 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 938 Value *Elt = Builder.CreateExtractValue(SV, i); 939 Old = ConvertScalar_InsertValue(Elt, Old, 940 Offset+Layout.getElementOffsetInBits(i), 941 0, Builder); 942 } 943 return Old; 944 } 945 946 if (ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 947 assert(!NonConstantIdx && 948 "Dynamic indexing into array types not supported"); 949 uint64_t EltSize = DL.getTypeAllocSizeInBits(AT->getElementType()); 950 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 951 Value *Elt = Builder.CreateExtractValue(SV, i); 952 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, 0, Builder); 953 } 954 return Old; 955 } 956 957 // If SV is a float, convert it to the appropriate integer type. 958 // If it is a pointer, do the same. 959 unsigned SrcWidth = DL.getTypeSizeInBits(SV->getType()); 960 unsigned DestWidth = DL.getTypeSizeInBits(AllocaType); 961 unsigned SrcStoreWidth = DL.getTypeStoreSizeInBits(SV->getType()); 962 unsigned DestStoreWidth = DL.getTypeStoreSizeInBits(AllocaType); 963 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 964 SV = Builder.CreateBitCast(SV, IntegerType::get(SV->getContext(),SrcWidth)); 965 else if (SV->getType()->isPointerTy()) 966 SV = Builder.CreatePtrToInt(SV, DL.getIntPtrType(SV->getType())); 967 968 // Zero extend or truncate the value if needed. 969 if (SV->getType() != AllocaType) { 970 if (SV->getType()->getPrimitiveSizeInBits() < 971 AllocaType->getPrimitiveSizeInBits()) 972 SV = Builder.CreateZExt(SV, AllocaType); 973 else { 974 // Truncation may be needed if storing more than the alloca can hold 975 // (undefined behavior). 976 SV = Builder.CreateTrunc(SV, AllocaType); 977 SrcWidth = DestWidth; 978 SrcStoreWidth = DestStoreWidth; 979 } 980 } 981 982 // If this is a big-endian system and the store is narrower than the 983 // full alloca type, we need to do a shift to get the right bits. 984 int ShAmt = 0; 985 if (DL.isBigEndian()) { 986 // On big-endian machines, the lowest bit is stored at the bit offset 987 // from the pointer given by getTypeStoreSizeInBits. This matters for 988 // integers with a bitwidth that is not a multiple of 8. 989 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 990 } else { 991 ShAmt = Offset; 992 } 993 994 // Note: we support negative bitwidths (with shr) which are not defined. 995 // We do this to support (f.e.) stores off the end of a structure where 996 // only some bits in the structure are set. 997 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 998 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 999 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), ShAmt)); 1000 Mask <<= ShAmt; 1001 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1002 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), -ShAmt)); 1003 Mask = Mask.lshr(-ShAmt); 1004 } 1005 1006 // Mask out the bits we are about to insert from the old value, and or 1007 // in the new bits. 1008 if (SrcWidth != DestWidth) { 1009 assert(DestWidth > SrcWidth); 1010 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 1011 SV = Builder.CreateOr(Old, SV, "ins"); 1012 } 1013 return SV; 1014} 1015 1016 1017//===----------------------------------------------------------------------===// 1018// SRoA Driver 1019//===----------------------------------------------------------------------===// 1020 1021 1022bool SROA::runOnFunction(Function &F) { 1023 if (skipOptnoneFunction(F)) 1024 return false; 1025 1026 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>(); 1027 DL = DLP ? &DLP->getDataLayout() : 0; 1028 1029 bool Changed = performPromotion(F); 1030 1031 // FIXME: ScalarRepl currently depends on DataLayout more than it 1032 // theoretically needs to. It should be refactored in order to support 1033 // target-independent IR. Until this is done, just skip the actual 1034 // scalar-replacement portion of this pass. 1035 if (!DL) return Changed; 1036 1037 while (1) { 1038 bool LocalChange = performScalarRepl(F); 1039 if (!LocalChange) break; // No need to repromote if no scalarrepl 1040 Changed = true; 1041 LocalChange = performPromotion(F); 1042 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 1043 } 1044 1045 return Changed; 1046} 1047 1048namespace { 1049class AllocaPromoter : public LoadAndStorePromoter { 1050 AllocaInst *AI; 1051 DIBuilder *DIB; 1052 SmallVector<DbgDeclareInst *, 4> DDIs; 1053 SmallVector<DbgValueInst *, 4> DVIs; 1054public: 1055 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S, 1056 DIBuilder *DB) 1057 : LoadAndStorePromoter(Insts, S), AI(0), DIB(DB) {} 1058 1059 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) { 1060 // Remember which alloca we're promoting (for isInstInList). 1061 this->AI = AI; 1062 if (MDNode *DebugNode = MDNode::getIfExists(AI->getContext(), AI)) { 1063 for (User *U : DebugNode->users()) 1064 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(U)) 1065 DDIs.push_back(DDI); 1066 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(U)) 1067 DVIs.push_back(DVI); 1068 } 1069 1070 LoadAndStorePromoter::run(Insts); 1071 AI->eraseFromParent(); 1072 for (SmallVectorImpl<DbgDeclareInst *>::iterator I = DDIs.begin(), 1073 E = DDIs.end(); I != E; ++I) { 1074 DbgDeclareInst *DDI = *I; 1075 DDI->eraseFromParent(); 1076 } 1077 for (SmallVectorImpl<DbgValueInst *>::iterator I = DVIs.begin(), 1078 E = DVIs.end(); I != E; ++I) { 1079 DbgValueInst *DVI = *I; 1080 DVI->eraseFromParent(); 1081 } 1082 } 1083 1084 bool isInstInList(Instruction *I, 1085 const SmallVectorImpl<Instruction*> &Insts) const override { 1086 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 1087 return LI->getOperand(0) == AI; 1088 return cast<StoreInst>(I)->getPointerOperand() == AI; 1089 } 1090 1091 void updateDebugInfo(Instruction *Inst) const override { 1092 for (SmallVectorImpl<DbgDeclareInst *>::const_iterator I = DDIs.begin(), 1093 E = DDIs.end(); I != E; ++I) { 1094 DbgDeclareInst *DDI = *I; 1095 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) 1096 ConvertDebugDeclareToDebugValue(DDI, SI, *DIB); 1097 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) 1098 ConvertDebugDeclareToDebugValue(DDI, LI, *DIB); 1099 } 1100 for (SmallVectorImpl<DbgValueInst *>::const_iterator I = DVIs.begin(), 1101 E = DVIs.end(); I != E; ++I) { 1102 DbgValueInst *DVI = *I; 1103 Value *Arg = NULL; 1104 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 1105 // If an argument is zero extended then use argument directly. The ZExt 1106 // may be zapped by an optimization pass in future. 1107 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0))) 1108 Arg = dyn_cast<Argument>(ZExt->getOperand(0)); 1109 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0))) 1110 Arg = dyn_cast<Argument>(SExt->getOperand(0)); 1111 if (!Arg) 1112 Arg = SI->getOperand(0); 1113 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 1114 Arg = LI->getOperand(0); 1115 } else { 1116 continue; 1117 } 1118 Instruction *DbgVal = 1119 DIB->insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()), 1120 Inst); 1121 DbgVal->setDebugLoc(DVI->getDebugLoc()); 1122 } 1123 } 1124}; 1125} // end anon namespace 1126 1127/// isSafeSelectToSpeculate - Select instructions that use an alloca and are 1128/// subsequently loaded can be rewritten to load both input pointers and then 1129/// select between the result, allowing the load of the alloca to be promoted. 1130/// From this: 1131/// %P2 = select i1 %cond, i32* %Alloca, i32* %Other 1132/// %V = load i32* %P2 1133/// to: 1134/// %V1 = load i32* %Alloca -> will be mem2reg'd 1135/// %V2 = load i32* %Other 1136/// %V = select i1 %cond, i32 %V1, i32 %V2 1137/// 1138/// We can do this to a select if its only uses are loads and if the operand to 1139/// the select can be loaded unconditionally. 1140static bool isSafeSelectToSpeculate(SelectInst *SI, const DataLayout *DL) { 1141 bool TDerefable = SI->getTrueValue()->isDereferenceablePointer(); 1142 bool FDerefable = SI->getFalseValue()->isDereferenceablePointer(); 1143 1144 for (User *U : SI->users()) { 1145 LoadInst *LI = dyn_cast<LoadInst>(U); 1146 if (LI == 0 || !LI->isSimple()) return false; 1147 1148 // Both operands to the select need to be dereferencable, either absolutely 1149 // (e.g. allocas) or at this point because we can see other accesses to it. 1150 if (!TDerefable && !isSafeToLoadUnconditionally(SI->getTrueValue(), LI, 1151 LI->getAlignment(), DL)) 1152 return false; 1153 if (!FDerefable && !isSafeToLoadUnconditionally(SI->getFalseValue(), LI, 1154 LI->getAlignment(), DL)) 1155 return false; 1156 } 1157 1158 return true; 1159} 1160 1161/// isSafePHIToSpeculate - PHI instructions that use an alloca and are 1162/// subsequently loaded can be rewritten to load both input pointers in the pred 1163/// blocks and then PHI the results, allowing the load of the alloca to be 1164/// promoted. 1165/// From this: 1166/// %P2 = phi [i32* %Alloca, i32* %Other] 1167/// %V = load i32* %P2 1168/// to: 1169/// %V1 = load i32* %Alloca -> will be mem2reg'd 1170/// ... 1171/// %V2 = load i32* %Other 1172/// ... 1173/// %V = phi [i32 %V1, i32 %V2] 1174/// 1175/// We can do this to a select if its only uses are loads and if the operand to 1176/// the select can be loaded unconditionally. 1177static bool isSafePHIToSpeculate(PHINode *PN, const DataLayout *DL) { 1178 // For now, we can only do this promotion if the load is in the same block as 1179 // the PHI, and if there are no stores between the phi and load. 1180 // TODO: Allow recursive phi users. 1181 // TODO: Allow stores. 1182 BasicBlock *BB = PN->getParent(); 1183 unsigned MaxAlign = 0; 1184 for (User *U : PN->users()) { 1185 LoadInst *LI = dyn_cast<LoadInst>(U); 1186 if (LI == 0 || !LI->isSimple()) return false; 1187 1188 // For now we only allow loads in the same block as the PHI. This is a 1189 // common case that happens when instcombine merges two loads through a PHI. 1190 if (LI->getParent() != BB) return false; 1191 1192 // Ensure that there are no instructions between the PHI and the load that 1193 // could store. 1194 for (BasicBlock::iterator BBI = PN; &*BBI != LI; ++BBI) 1195 if (BBI->mayWriteToMemory()) 1196 return false; 1197 1198 MaxAlign = std::max(MaxAlign, LI->getAlignment()); 1199 } 1200 1201 // Okay, we know that we have one or more loads in the same block as the PHI. 1202 // We can transform this if it is safe to push the loads into the predecessor 1203 // blocks. The only thing to watch out for is that we can't put a possibly 1204 // trapping load in the predecessor if it is a critical edge. 1205 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1206 BasicBlock *Pred = PN->getIncomingBlock(i); 1207 Value *InVal = PN->getIncomingValue(i); 1208 1209 // If the terminator of the predecessor has side-effects (an invoke), 1210 // there is no safe place to put a load in the predecessor. 1211 if (Pred->getTerminator()->mayHaveSideEffects()) 1212 return false; 1213 1214 // If the value is produced by the terminator of the predecessor 1215 // (an invoke), there is no valid place to put a load in the predecessor. 1216 if (Pred->getTerminator() == InVal) 1217 return false; 1218 1219 // If the predecessor has a single successor, then the edge isn't critical. 1220 if (Pred->getTerminator()->getNumSuccessors() == 1) 1221 continue; 1222 1223 // If this pointer is always safe to load, or if we can prove that there is 1224 // already a load in the block, then we can move the load to the pred block. 1225 if (InVal->isDereferenceablePointer() || 1226 isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign, DL)) 1227 continue; 1228 1229 return false; 1230 } 1231 1232 return true; 1233} 1234 1235 1236/// tryToMakeAllocaBePromotable - This returns true if the alloca only has 1237/// direct (non-volatile) loads and stores to it. If the alloca is close but 1238/// not quite there, this will transform the code to allow promotion. As such, 1239/// it is a non-pure predicate. 1240static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const DataLayout *DL) { 1241 SetVector<Instruction*, SmallVector<Instruction*, 4>, 1242 SmallPtrSet<Instruction*, 4> > InstsToRewrite; 1243 for (User *U : AI->users()) { 1244 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 1245 if (!LI->isSimple()) 1246 return false; 1247 continue; 1248 } 1249 1250 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 1251 if (SI->getOperand(0) == AI || !SI->isSimple()) 1252 return false; // Don't allow a store OF the AI, only INTO the AI. 1253 continue; 1254 } 1255 1256 if (SelectInst *SI = dyn_cast<SelectInst>(U)) { 1257 // If the condition being selected on is a constant, fold the select, yes 1258 // this does (rarely) happen early on. 1259 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition())) { 1260 Value *Result = SI->getOperand(1+CI->isZero()); 1261 SI->replaceAllUsesWith(Result); 1262 SI->eraseFromParent(); 1263 1264 // This is very rare and we just scrambled the use list of AI, start 1265 // over completely. 1266 return tryToMakeAllocaBePromotable(AI, DL); 1267 } 1268 1269 // If it is safe to turn "load (select c, AI, ptr)" into a select of two 1270 // loads, then we can transform this by rewriting the select. 1271 if (!isSafeSelectToSpeculate(SI, DL)) 1272 return false; 1273 1274 InstsToRewrite.insert(SI); 1275 continue; 1276 } 1277 1278 if (PHINode *PN = dyn_cast<PHINode>(U)) { 1279 if (PN->use_empty()) { // Dead PHIs can be stripped. 1280 InstsToRewrite.insert(PN); 1281 continue; 1282 } 1283 1284 // If it is safe to turn "load (phi [AI, ptr, ...])" into a PHI of loads 1285 // in the pred blocks, then we can transform this by rewriting the PHI. 1286 if (!isSafePHIToSpeculate(PN, DL)) 1287 return false; 1288 1289 InstsToRewrite.insert(PN); 1290 continue; 1291 } 1292 1293 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 1294 if (onlyUsedByLifetimeMarkers(BCI)) { 1295 InstsToRewrite.insert(BCI); 1296 continue; 1297 } 1298 } 1299 1300 return false; 1301 } 1302 1303 // If there are no instructions to rewrite, then all uses are load/stores and 1304 // we're done! 1305 if (InstsToRewrite.empty()) 1306 return true; 1307 1308 // If we have instructions that need to be rewritten for this to be promotable 1309 // take care of it now. 1310 for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) { 1311 if (BitCastInst *BCI = dyn_cast<BitCastInst>(InstsToRewrite[i])) { 1312 // This could only be a bitcast used by nothing but lifetime intrinsics. 1313 for (BitCastInst::user_iterator I = BCI->user_begin(), E = BCI->user_end(); 1314 I != E;) 1315 cast<Instruction>(*I++)->eraseFromParent(); 1316 BCI->eraseFromParent(); 1317 continue; 1318 } 1319 1320 if (SelectInst *SI = dyn_cast<SelectInst>(InstsToRewrite[i])) { 1321 // Selects in InstsToRewrite only have load uses. Rewrite each as two 1322 // loads with a new select. 1323 while (!SI->use_empty()) { 1324 LoadInst *LI = cast<LoadInst>(SI->user_back()); 1325 1326 IRBuilder<> Builder(LI); 1327 LoadInst *TrueLoad = 1328 Builder.CreateLoad(SI->getTrueValue(), LI->getName()+".t"); 1329 LoadInst *FalseLoad = 1330 Builder.CreateLoad(SI->getFalseValue(), LI->getName()+".f"); 1331 1332 // Transfer alignment and TBAA info if present. 1333 TrueLoad->setAlignment(LI->getAlignment()); 1334 FalseLoad->setAlignment(LI->getAlignment()); 1335 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) { 1336 TrueLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1337 FalseLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1338 } 1339 1340 Value *V = Builder.CreateSelect(SI->getCondition(), TrueLoad, FalseLoad); 1341 V->takeName(LI); 1342 LI->replaceAllUsesWith(V); 1343 LI->eraseFromParent(); 1344 } 1345 1346 // Now that all the loads are gone, the select is gone too. 1347 SI->eraseFromParent(); 1348 continue; 1349 } 1350 1351 // Otherwise, we have a PHI node which allows us to push the loads into the 1352 // predecessors. 1353 PHINode *PN = cast<PHINode>(InstsToRewrite[i]); 1354 if (PN->use_empty()) { 1355 PN->eraseFromParent(); 1356 continue; 1357 } 1358 1359 Type *LoadTy = cast<PointerType>(PN->getType())->getElementType(); 1360 PHINode *NewPN = PHINode::Create(LoadTy, PN->getNumIncomingValues(), 1361 PN->getName()+".ld", PN); 1362 1363 // Get the TBAA tag and alignment to use from one of the loads. It doesn't 1364 // matter which one we get and if any differ, it doesn't matter. 1365 LoadInst *SomeLoad = cast<LoadInst>(PN->user_back()); 1366 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa); 1367 unsigned Align = SomeLoad->getAlignment(); 1368 1369 // Rewrite all loads of the PN to use the new PHI. 1370 while (!PN->use_empty()) { 1371 LoadInst *LI = cast<LoadInst>(PN->user_back()); 1372 LI->replaceAllUsesWith(NewPN); 1373 LI->eraseFromParent(); 1374 } 1375 1376 // Inject loads into all of the pred blocks. Keep track of which blocks we 1377 // insert them into in case we have multiple edges from the same block. 1378 DenseMap<BasicBlock*, LoadInst*> InsertedLoads; 1379 1380 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1381 BasicBlock *Pred = PN->getIncomingBlock(i); 1382 LoadInst *&Load = InsertedLoads[Pred]; 1383 if (Load == 0) { 1384 Load = new LoadInst(PN->getIncomingValue(i), 1385 PN->getName() + "." + Pred->getName(), 1386 Pred->getTerminator()); 1387 Load->setAlignment(Align); 1388 if (TBAATag) Load->setMetadata(LLVMContext::MD_tbaa, TBAATag); 1389 } 1390 1391 NewPN->addIncoming(Load, Pred); 1392 } 1393 1394 PN->eraseFromParent(); 1395 } 1396 1397 ++NumAdjusted; 1398 return true; 1399} 1400 1401bool SROA::performPromotion(Function &F) { 1402 std::vector<AllocaInst*> Allocas; 1403 DominatorTree *DT = 0; 1404 if (HasDomTree) 1405 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 1406 1407 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 1408 DIBuilder DIB(*F.getParent()); 1409 bool Changed = false; 1410 SmallVector<Instruction*, 64> Insts; 1411 while (1) { 1412 Allocas.clear(); 1413 1414 // Find allocas that are safe to promote, by looking at all instructions in 1415 // the entry node 1416 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 1417 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 1418 if (tryToMakeAllocaBePromotable(AI, DL)) 1419 Allocas.push_back(AI); 1420 1421 if (Allocas.empty()) break; 1422 1423 if (HasDomTree) 1424 PromoteMemToReg(Allocas, *DT); 1425 else { 1426 SSAUpdater SSA; 1427 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) { 1428 AllocaInst *AI = Allocas[i]; 1429 1430 // Build list of instructions to promote. 1431 for (User *U : AI->users()) 1432 Insts.push_back(cast<Instruction>(U)); 1433 AllocaPromoter(Insts, SSA, &DIB).run(AI, Insts); 1434 Insts.clear(); 1435 } 1436 } 1437 NumPromoted += Allocas.size(); 1438 Changed = true; 1439 } 1440 1441 return Changed; 1442} 1443 1444 1445/// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 1446/// SROA. It must be a struct or array type with a small number of elements. 1447bool SROA::ShouldAttemptScalarRepl(AllocaInst *AI) { 1448 Type *T = AI->getAllocatedType(); 1449 // Do not promote any struct that has too many members. 1450 if (StructType *ST = dyn_cast<StructType>(T)) 1451 return ST->getNumElements() <= StructMemberThreshold; 1452 // Do not promote any array that has too many elements. 1453 if (ArrayType *AT = dyn_cast<ArrayType>(T)) 1454 return AT->getNumElements() <= ArrayElementThreshold; 1455 return false; 1456} 1457 1458// performScalarRepl - This algorithm is a simple worklist driven algorithm, 1459// which runs on all of the alloca instructions in the entry block, removing 1460// them if they are only used by getelementptr instructions. 1461// 1462bool SROA::performScalarRepl(Function &F) { 1463 std::vector<AllocaInst*> WorkList; 1464 1465 // Scan the entry basic block, adding allocas to the worklist. 1466 BasicBlock &BB = F.getEntryBlock(); 1467 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 1468 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 1469 WorkList.push_back(A); 1470 1471 // Process the worklist 1472 bool Changed = false; 1473 while (!WorkList.empty()) { 1474 AllocaInst *AI = WorkList.back(); 1475 WorkList.pop_back(); 1476 1477 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 1478 // with unused elements. 1479 if (AI->use_empty()) { 1480 AI->eraseFromParent(); 1481 Changed = true; 1482 continue; 1483 } 1484 1485 // If this alloca is impossible for us to promote, reject it early. 1486 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 1487 continue; 1488 1489 // Check to see if we can perform the core SROA transformation. We cannot 1490 // transform the allocation instruction if it is an array allocation 1491 // (allocations OF arrays are ok though), and an allocation of a scalar 1492 // value cannot be decomposed at all. 1493 uint64_t AllocaSize = DL->getTypeAllocSize(AI->getAllocatedType()); 1494 1495 // Do not promote [0 x %struct]. 1496 if (AllocaSize == 0) continue; 1497 1498 // Do not promote any struct whose size is too big. 1499 if (AllocaSize > SRThreshold) continue; 1500 1501 // If the alloca looks like a good candidate for scalar replacement, and if 1502 // all its users can be transformed, then split up the aggregate into its 1503 // separate elements. 1504 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 1505 DoScalarReplacement(AI, WorkList); 1506 Changed = true; 1507 continue; 1508 } 1509 1510 // If we can turn this aggregate value (potentially with casts) into a 1511 // simple scalar value that can be mem2reg'd into a register value. 1512 // IsNotTrivial tracks whether this is something that mem2reg could have 1513 // promoted itself. If so, we don't want to transform it needlessly. Note 1514 // that we can't just check based on the type: the alloca may be of an i32 1515 // but that has pointer arithmetic to set byte 3 of it or something. 1516 if (AllocaInst *NewAI = ConvertToScalarInfo( 1517 (unsigned)AllocaSize, *DL, ScalarLoadThreshold).TryConvert(AI)) { 1518 NewAI->takeName(AI); 1519 AI->eraseFromParent(); 1520 ++NumConverted; 1521 Changed = true; 1522 continue; 1523 } 1524 1525 // Otherwise, couldn't process this alloca. 1526 } 1527 1528 return Changed; 1529} 1530 1531/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 1532/// predicate, do SROA now. 1533void SROA::DoScalarReplacement(AllocaInst *AI, 1534 std::vector<AllocaInst*> &WorkList) { 1535 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 1536 SmallVector<AllocaInst*, 32> ElementAllocas; 1537 if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 1538 ElementAllocas.reserve(ST->getNumContainedTypes()); 1539 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 1540 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 1541 AI->getAlignment(), 1542 AI->getName() + "." + Twine(i), AI); 1543 ElementAllocas.push_back(NA); 1544 WorkList.push_back(NA); // Add to worklist for recursive processing 1545 } 1546 } else { 1547 ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 1548 ElementAllocas.reserve(AT->getNumElements()); 1549 Type *ElTy = AT->getElementType(); 1550 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1551 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 1552 AI->getName() + "." + Twine(i), AI); 1553 ElementAllocas.push_back(NA); 1554 WorkList.push_back(NA); // Add to worklist for recursive processing 1555 } 1556 } 1557 1558 // Now that we have created the new alloca instructions, rewrite all the 1559 // uses of the old alloca. 1560 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 1561 1562 // Now erase any instructions that were made dead while rewriting the alloca. 1563 DeleteDeadInstructions(); 1564 AI->eraseFromParent(); 1565 1566 ++NumReplaced; 1567} 1568 1569/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 1570/// recursively including all their operands that become trivially dead. 1571void SROA::DeleteDeadInstructions() { 1572 while (!DeadInsts.empty()) { 1573 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 1574 1575 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 1576 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 1577 // Zero out the operand and see if it becomes trivially dead. 1578 // (But, don't add allocas to the dead instruction list -- they are 1579 // already on the worklist and will be deleted separately.) 1580 *OI = 0; 1581 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 1582 DeadInsts.push_back(U); 1583 } 1584 1585 I->eraseFromParent(); 1586 } 1587} 1588 1589/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 1590/// performing scalar replacement of alloca AI. The results are flagged in 1591/// the Info parameter. Offset indicates the position within AI that is 1592/// referenced by this instruction. 1593void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset, 1594 AllocaInfo &Info) { 1595 for (Use &U : I->uses()) { 1596 Instruction *User = cast<Instruction>(U.getUser()); 1597 1598 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1599 isSafeForScalarRepl(BC, Offset, Info); 1600 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1601 uint64_t GEPOffset = Offset; 1602 isSafeGEP(GEPI, GEPOffset, Info); 1603 if (!Info.isUnsafe) 1604 isSafeForScalarRepl(GEPI, GEPOffset, Info); 1605 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1606 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1607 if (Length == 0) 1608 return MarkUnsafe(Info, User); 1609 if (Length->isNegative()) 1610 return MarkUnsafe(Info, User); 1611 1612 isSafeMemAccess(Offset, Length->getZExtValue(), 0, 1613 U.getOperandNo() == 0, Info, MI, 1614 true /*AllowWholeAccess*/); 1615 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1616 if (!LI->isSimple()) 1617 return MarkUnsafe(Info, User); 1618 Type *LIType = LI->getType(); 1619 isSafeMemAccess(Offset, DL->getTypeAllocSize(LIType), 1620 LIType, false, Info, LI, true /*AllowWholeAccess*/); 1621 Info.hasALoadOrStore = true; 1622 1623 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1624 // Store is ok if storing INTO the pointer, not storing the pointer 1625 if (!SI->isSimple() || SI->getOperand(0) == I) 1626 return MarkUnsafe(Info, User); 1627 1628 Type *SIType = SI->getOperand(0)->getType(); 1629 isSafeMemAccess(Offset, DL->getTypeAllocSize(SIType), 1630 SIType, true, Info, SI, true /*AllowWholeAccess*/); 1631 Info.hasALoadOrStore = true; 1632 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 1633 if (II->getIntrinsicID() != Intrinsic::lifetime_start && 1634 II->getIntrinsicID() != Intrinsic::lifetime_end) 1635 return MarkUnsafe(Info, User); 1636 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1637 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1638 } else { 1639 return MarkUnsafe(Info, User); 1640 } 1641 if (Info.isUnsafe) return; 1642 } 1643} 1644 1645 1646/// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer 1647/// derived from the alloca, we can often still split the alloca into elements. 1648/// This is useful if we have a large alloca where one element is phi'd 1649/// together somewhere: we can SRoA and promote all the other elements even if 1650/// we end up not being able to promote this one. 1651/// 1652/// All we require is that the uses of the PHI do not index into other parts of 1653/// the alloca. The most important use case for this is single load and stores 1654/// that are PHI'd together, which can happen due to code sinking. 1655void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset, 1656 AllocaInfo &Info) { 1657 // If we've already checked this PHI, don't do it again. 1658 if (PHINode *PN = dyn_cast<PHINode>(I)) 1659 if (!Info.CheckedPHIs.insert(PN)) 1660 return; 1661 1662 for (User *U : I->users()) { 1663 Instruction *UI = cast<Instruction>(U); 1664 1665 if (BitCastInst *BC = dyn_cast<BitCastInst>(UI)) { 1666 isSafePHISelectUseForScalarRepl(BC, Offset, Info); 1667 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(UI)) { 1668 // Only allow "bitcast" GEPs for simplicity. We could generalize this, 1669 // but would have to prove that we're staying inside of an element being 1670 // promoted. 1671 if (!GEPI->hasAllZeroIndices()) 1672 return MarkUnsafe(Info, UI); 1673 isSafePHISelectUseForScalarRepl(GEPI, Offset, Info); 1674 } else if (LoadInst *LI = dyn_cast<LoadInst>(UI)) { 1675 if (!LI->isSimple()) 1676 return MarkUnsafe(Info, UI); 1677 Type *LIType = LI->getType(); 1678 isSafeMemAccess(Offset, DL->getTypeAllocSize(LIType), 1679 LIType, false, Info, LI, false /*AllowWholeAccess*/); 1680 Info.hasALoadOrStore = true; 1681 1682 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI)) { 1683 // Store is ok if storing INTO the pointer, not storing the pointer 1684 if (!SI->isSimple() || SI->getOperand(0) == I) 1685 return MarkUnsafe(Info, UI); 1686 1687 Type *SIType = SI->getOperand(0)->getType(); 1688 isSafeMemAccess(Offset, DL->getTypeAllocSize(SIType), 1689 SIType, true, Info, SI, false /*AllowWholeAccess*/); 1690 Info.hasALoadOrStore = true; 1691 } else if (isa<PHINode>(UI) || isa<SelectInst>(UI)) { 1692 isSafePHISelectUseForScalarRepl(UI, Offset, Info); 1693 } else { 1694 return MarkUnsafe(Info, UI); 1695 } 1696 if (Info.isUnsafe) return; 1697 } 1698} 1699 1700/// isSafeGEP - Check if a GEP instruction can be handled for scalar 1701/// replacement. It is safe when all the indices are constant, in-bounds 1702/// references, and when the resulting offset corresponds to an element within 1703/// the alloca type. The results are flagged in the Info parameter. Upon 1704/// return, Offset is adjusted as specified by the GEP indices. 1705void SROA::isSafeGEP(GetElementPtrInst *GEPI, 1706 uint64_t &Offset, AllocaInfo &Info) { 1707 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 1708 if (GEPIt == E) 1709 return; 1710 bool NonConstant = false; 1711 unsigned NonConstantIdxSize = 0; 1712 1713 // Walk through the GEP type indices, checking the types that this indexes 1714 // into. 1715 for (; GEPIt != E; ++GEPIt) { 1716 // Ignore struct elements, no extra checking needed for these. 1717 if ((*GEPIt)->isStructTy()) 1718 continue; 1719 1720 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 1721 if (!IdxVal) 1722 return MarkUnsafe(Info, GEPI); 1723 } 1724 1725 // Compute the offset due to this GEP and check if the alloca has a 1726 // component element at that offset. 1727 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1728 // If this GEP is non-constant then the last operand must have been a 1729 // dynamic index into a vector. Pop this now as it has no impact on the 1730 // constant part of the offset. 1731 if (NonConstant) 1732 Indices.pop_back(); 1733 Offset += DL->getIndexedOffset(GEPI->getPointerOperandType(), Indices); 1734 if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 1735 NonConstantIdxSize)) 1736 MarkUnsafe(Info, GEPI); 1737} 1738 1739/// isHomogeneousAggregate - Check if type T is a struct or array containing 1740/// elements of the same type (which is always true for arrays). If so, 1741/// return true with NumElts and EltTy set to the number of elements and the 1742/// element type, respectively. 1743static bool isHomogeneousAggregate(Type *T, unsigned &NumElts, 1744 Type *&EltTy) { 1745 if (ArrayType *AT = dyn_cast<ArrayType>(T)) { 1746 NumElts = AT->getNumElements(); 1747 EltTy = (NumElts == 0 ? 0 : AT->getElementType()); 1748 return true; 1749 } 1750 if (StructType *ST = dyn_cast<StructType>(T)) { 1751 NumElts = ST->getNumContainedTypes(); 1752 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0)); 1753 for (unsigned n = 1; n < NumElts; ++n) { 1754 if (ST->getContainedType(n) != EltTy) 1755 return false; 1756 } 1757 return true; 1758 } 1759 return false; 1760} 1761 1762/// isCompatibleAggregate - Check if T1 and T2 are either the same type or are 1763/// "homogeneous" aggregates with the same element type and number of elements. 1764static bool isCompatibleAggregate(Type *T1, Type *T2) { 1765 if (T1 == T2) 1766 return true; 1767 1768 unsigned NumElts1, NumElts2; 1769 Type *EltTy1, *EltTy2; 1770 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) && 1771 isHomogeneousAggregate(T2, NumElts2, EltTy2) && 1772 NumElts1 == NumElts2 && 1773 EltTy1 == EltTy2) 1774 return true; 1775 1776 return false; 1777} 1778 1779/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1780/// alloca or has an offset and size that corresponds to a component element 1781/// within it. The offset checked here may have been formed from a GEP with a 1782/// pointer bitcasted to a different type. 1783/// 1784/// If AllowWholeAccess is true, then this allows uses of the entire alloca as a 1785/// unit. If false, it only allows accesses known to be in a single element. 1786void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 1787 Type *MemOpType, bool isStore, 1788 AllocaInfo &Info, Instruction *TheAccess, 1789 bool AllowWholeAccess) { 1790 // Check if this is a load/store of the entire alloca. 1791 if (Offset == 0 && AllowWholeAccess && 1792 MemSize == DL->getTypeAllocSize(Info.AI->getAllocatedType())) { 1793 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer 1794 // loads/stores (which are essentially the same as the MemIntrinsics with 1795 // regard to copying padding between elements). But, if an alloca is 1796 // flagged as both a source and destination of such operations, we'll need 1797 // to check later for padding between elements. 1798 if (!MemOpType || MemOpType->isIntegerTy()) { 1799 if (isStore) 1800 Info.isMemCpyDst = true; 1801 else 1802 Info.isMemCpySrc = true; 1803 return; 1804 } 1805 // This is also safe for references using a type that is compatible with 1806 // the type of the alloca, so that loads/stores can be rewritten using 1807 // insertvalue/extractvalue. 1808 if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) { 1809 Info.hasSubelementAccess = true; 1810 return; 1811 } 1812 } 1813 // Check if the offset/size correspond to a component within the alloca type. 1814 Type *T = Info.AI->getAllocatedType(); 1815 if (TypeHasComponent(T, Offset, MemSize)) { 1816 Info.hasSubelementAccess = true; 1817 return; 1818 } 1819 1820 return MarkUnsafe(Info, TheAccess); 1821} 1822 1823/// TypeHasComponent - Return true if T has a component type with the 1824/// specified offset and size. If Size is zero, do not check the size. 1825bool SROA::TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size) { 1826 Type *EltTy; 1827 uint64_t EltSize; 1828 if (StructType *ST = dyn_cast<StructType>(T)) { 1829 const StructLayout *Layout = DL->getStructLayout(ST); 1830 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1831 EltTy = ST->getContainedType(EltIdx); 1832 EltSize = DL->getTypeAllocSize(EltTy); 1833 Offset -= Layout->getElementOffset(EltIdx); 1834 } else if (ArrayType *AT = dyn_cast<ArrayType>(T)) { 1835 EltTy = AT->getElementType(); 1836 EltSize = DL->getTypeAllocSize(EltTy); 1837 if (Offset >= AT->getNumElements() * EltSize) 1838 return false; 1839 Offset %= EltSize; 1840 } else if (VectorType *VT = dyn_cast<VectorType>(T)) { 1841 EltTy = VT->getElementType(); 1842 EltSize = DL->getTypeAllocSize(EltTy); 1843 if (Offset >= VT->getNumElements() * EltSize) 1844 return false; 1845 Offset %= EltSize; 1846 } else { 1847 return false; 1848 } 1849 if (Offset == 0 && (Size == 0 || EltSize == Size)) 1850 return true; 1851 // Check if the component spans multiple elements. 1852 if (Offset + Size > EltSize) 1853 return false; 1854 return TypeHasComponent(EltTy, Offset, Size); 1855} 1856 1857/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1858/// the instruction I, which references it, to use the separate elements. 1859/// Offset indicates the position within AI that is referenced by this 1860/// instruction. 1861void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1862 SmallVectorImpl<AllocaInst *> &NewElts) { 1863 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) { 1864 Use &TheUse = *UI++; 1865 Instruction *User = cast<Instruction>(TheUse.getUser()); 1866 1867 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1868 RewriteBitCast(BC, AI, Offset, NewElts); 1869 continue; 1870 } 1871 1872 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1873 RewriteGEP(GEPI, AI, Offset, NewElts); 1874 continue; 1875 } 1876 1877 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1878 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1879 uint64_t MemSize = Length->getZExtValue(); 1880 if (Offset == 0 && 1881 MemSize == DL->getTypeAllocSize(AI->getAllocatedType())) 1882 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1883 // Otherwise the intrinsic can only touch a single element and the 1884 // address operand will be updated, so nothing else needs to be done. 1885 continue; 1886 } 1887 1888 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 1889 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 1890 II->getIntrinsicID() == Intrinsic::lifetime_end) { 1891 RewriteLifetimeIntrinsic(II, AI, Offset, NewElts); 1892 } 1893 continue; 1894 } 1895 1896 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1897 Type *LIType = LI->getType(); 1898 1899 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) { 1900 // Replace: 1901 // %res = load { i32, i32 }* %alloc 1902 // with: 1903 // %load.0 = load i32* %alloc.0 1904 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1905 // %load.1 = load i32* %alloc.1 1906 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1907 // (Also works for arrays instead of structs) 1908 Value *Insert = UndefValue::get(LIType); 1909 IRBuilder<> Builder(LI); 1910 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1911 Value *Load = Builder.CreateLoad(NewElts[i], "load"); 1912 Insert = Builder.CreateInsertValue(Insert, Load, i, "insert"); 1913 } 1914 LI->replaceAllUsesWith(Insert); 1915 DeadInsts.push_back(LI); 1916 } else if (LIType->isIntegerTy() && 1917 DL->getTypeAllocSize(LIType) == 1918 DL->getTypeAllocSize(AI->getAllocatedType())) { 1919 // If this is a load of the entire alloca to an integer, rewrite it. 1920 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1921 } 1922 continue; 1923 } 1924 1925 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1926 Value *Val = SI->getOperand(0); 1927 Type *SIType = Val->getType(); 1928 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) { 1929 // Replace: 1930 // store { i32, i32 } %val, { i32, i32 }* %alloc 1931 // with: 1932 // %val.0 = extractvalue { i32, i32 } %val, 0 1933 // store i32 %val.0, i32* %alloc.0 1934 // %val.1 = extractvalue { i32, i32 } %val, 1 1935 // store i32 %val.1, i32* %alloc.1 1936 // (Also works for arrays instead of structs) 1937 IRBuilder<> Builder(SI); 1938 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1939 Value *Extract = Builder.CreateExtractValue(Val, i, Val->getName()); 1940 Builder.CreateStore(Extract, NewElts[i]); 1941 } 1942 DeadInsts.push_back(SI); 1943 } else if (SIType->isIntegerTy() && 1944 DL->getTypeAllocSize(SIType) == 1945 DL->getTypeAllocSize(AI->getAllocatedType())) { 1946 // If this is a store of the entire alloca from an integer, rewrite it. 1947 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 1948 } 1949 continue; 1950 } 1951 1952 if (isa<SelectInst>(User) || isa<PHINode>(User)) { 1953 // If we have a PHI user of the alloca itself (as opposed to a GEP or 1954 // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to 1955 // the new pointer. 1956 if (!isa<AllocaInst>(I)) continue; 1957 1958 assert(Offset == 0 && NewElts[0] && 1959 "Direct alloca use should have a zero offset"); 1960 1961 // If we have a use of the alloca, we know the derived uses will be 1962 // utilizing just the first element of the scalarized result. Insert a 1963 // bitcast of the first alloca before the user as required. 1964 AllocaInst *NewAI = NewElts[0]; 1965 BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI); 1966 NewAI->moveBefore(BCI); 1967 TheUse = BCI; 1968 continue; 1969 } 1970 } 1971} 1972 1973/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 1974/// and recursively continue updating all of its uses. 1975void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 1976 SmallVectorImpl<AllocaInst *> &NewElts) { 1977 RewriteForScalarRepl(BC, AI, Offset, NewElts); 1978 if (BC->getOperand(0) != AI) 1979 return; 1980 1981 // The bitcast references the original alloca. Replace its uses with 1982 // references to the alloca containing offset zero (which is normally at 1983 // index zero, but might not be in cases involving structs with elements 1984 // of size zero). 1985 Type *T = AI->getAllocatedType(); 1986 uint64_t EltOffset = 0; 1987 Type *IdxTy; 1988 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1989 Instruction *Val = NewElts[Idx]; 1990 if (Val->getType() != BC->getDestTy()) { 1991 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 1992 Val->takeName(BC); 1993 } 1994 BC->replaceAllUsesWith(Val); 1995 DeadInsts.push_back(BC); 1996} 1997 1998/// FindElementAndOffset - Return the index of the element containing Offset 1999/// within the specified type, which must be either a struct or an array. 2000/// Sets T to the type of the element and Offset to the offset within that 2001/// element. IdxTy is set to the type of the index result to be used in a 2002/// GEP instruction. 2003uint64_t SROA::FindElementAndOffset(Type *&T, uint64_t &Offset, 2004 Type *&IdxTy) { 2005 uint64_t Idx = 0; 2006 if (StructType *ST = dyn_cast<StructType>(T)) { 2007 const StructLayout *Layout = DL->getStructLayout(ST); 2008 Idx = Layout->getElementContainingOffset(Offset); 2009 T = ST->getContainedType(Idx); 2010 Offset -= Layout->getElementOffset(Idx); 2011 IdxTy = Type::getInt32Ty(T->getContext()); 2012 return Idx; 2013 } else if (ArrayType *AT = dyn_cast<ArrayType>(T)) { 2014 T = AT->getElementType(); 2015 uint64_t EltSize = DL->getTypeAllocSize(T); 2016 Idx = Offset / EltSize; 2017 Offset -= Idx * EltSize; 2018 IdxTy = Type::getInt64Ty(T->getContext()); 2019 return Idx; 2020 } 2021 VectorType *VT = cast<VectorType>(T); 2022 T = VT->getElementType(); 2023 uint64_t EltSize = DL->getTypeAllocSize(T); 2024 Idx = Offset / EltSize; 2025 Offset -= Idx * EltSize; 2026 IdxTy = Type::getInt64Ty(T->getContext()); 2027 return Idx; 2028} 2029 2030/// RewriteGEP - Check if this GEP instruction moves the pointer across 2031/// elements of the alloca that are being split apart, and if so, rewrite 2032/// the GEP to be relative to the new element. 2033void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 2034 SmallVectorImpl<AllocaInst *> &NewElts) { 2035 uint64_t OldOffset = Offset; 2036 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 2037 // If the GEP was dynamic then it must have been a dynamic vector lookup. 2038 // In this case, it must be the last GEP operand which is dynamic so keep that 2039 // aside until we've found the constant GEP offset then add it back in at the 2040 // end. 2041 Value* NonConstantIdx = 0; 2042 if (!GEPI->hasAllConstantIndices()) 2043 NonConstantIdx = Indices.pop_back_val(); 2044 Offset += DL->getIndexedOffset(GEPI->getPointerOperandType(), Indices); 2045 2046 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 2047 2048 Type *T = AI->getAllocatedType(); 2049 Type *IdxTy; 2050 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 2051 if (GEPI->getOperand(0) == AI) 2052 OldIdx = ~0ULL; // Force the GEP to be rewritten. 2053 2054 T = AI->getAllocatedType(); 2055 uint64_t EltOffset = Offset; 2056 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 2057 2058 // If this GEP does not move the pointer across elements of the alloca 2059 // being split, then it does not needs to be rewritten. 2060 if (Idx == OldIdx) 2061 return; 2062 2063 Type *i32Ty = Type::getInt32Ty(AI->getContext()); 2064 SmallVector<Value*, 8> NewArgs; 2065 NewArgs.push_back(Constant::getNullValue(i32Ty)); 2066 while (EltOffset != 0) { 2067 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 2068 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 2069 } 2070 if (NonConstantIdx) { 2071 Type* GepTy = T; 2072 // This GEP has a dynamic index. We need to add "i32 0" to index through 2073 // any structs or arrays in the original type until we get to the vector 2074 // to index. 2075 while (!isa<VectorType>(GepTy)) { 2076 NewArgs.push_back(Constant::getNullValue(i32Ty)); 2077 GepTy = cast<CompositeType>(GepTy)->getTypeAtIndex(0U); 2078 } 2079 NewArgs.push_back(NonConstantIdx); 2080 } 2081 Instruction *Val = NewElts[Idx]; 2082 if (NewArgs.size() > 1) { 2083 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs, "", GEPI); 2084 Val->takeName(GEPI); 2085 } 2086 if (Val->getType() != GEPI->getType()) 2087 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 2088 GEPI->replaceAllUsesWith(Val); 2089 DeadInsts.push_back(GEPI); 2090} 2091 2092/// RewriteLifetimeIntrinsic - II is a lifetime.start/lifetime.end. Rewrite it 2093/// to mark the lifetime of the scalarized memory. 2094void SROA::RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI, 2095 uint64_t Offset, 2096 SmallVectorImpl<AllocaInst *> &NewElts) { 2097 ConstantInt *OldSize = cast<ConstantInt>(II->getArgOperand(0)); 2098 // Put matching lifetime markers on everything from Offset up to 2099 // Offset+OldSize. 2100 Type *AIType = AI->getAllocatedType(); 2101 uint64_t NewOffset = Offset; 2102 Type *IdxTy; 2103 uint64_t Idx = FindElementAndOffset(AIType, NewOffset, IdxTy); 2104 2105 IRBuilder<> Builder(II); 2106 uint64_t Size = OldSize->getLimitedValue(); 2107 2108 if (NewOffset) { 2109 // Splice the first element and index 'NewOffset' bytes in. SROA will 2110 // split the alloca again later. 2111 Value *V = Builder.CreateBitCast(NewElts[Idx], Builder.getInt8PtrTy()); 2112 V = Builder.CreateGEP(V, Builder.getInt64(NewOffset)); 2113 2114 IdxTy = NewElts[Idx]->getAllocatedType(); 2115 uint64_t EltSize = DL->getTypeAllocSize(IdxTy) - NewOffset; 2116 if (EltSize > Size) { 2117 EltSize = Size; 2118 Size = 0; 2119 } else { 2120 Size -= EltSize; 2121 } 2122 if (II->getIntrinsicID() == Intrinsic::lifetime_start) 2123 Builder.CreateLifetimeStart(V, Builder.getInt64(EltSize)); 2124 else 2125 Builder.CreateLifetimeEnd(V, Builder.getInt64(EltSize)); 2126 ++Idx; 2127 } 2128 2129 for (; Idx != NewElts.size() && Size; ++Idx) { 2130 IdxTy = NewElts[Idx]->getAllocatedType(); 2131 uint64_t EltSize = DL->getTypeAllocSize(IdxTy); 2132 if (EltSize > Size) { 2133 EltSize = Size; 2134 Size = 0; 2135 } else { 2136 Size -= EltSize; 2137 } 2138 if (II->getIntrinsicID() == Intrinsic::lifetime_start) 2139 Builder.CreateLifetimeStart(NewElts[Idx], 2140 Builder.getInt64(EltSize)); 2141 else 2142 Builder.CreateLifetimeEnd(NewElts[Idx], 2143 Builder.getInt64(EltSize)); 2144 } 2145 DeadInsts.push_back(II); 2146} 2147 2148/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 2149/// Rewrite it to copy or set the elements of the scalarized memory. 2150void 2151SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 2152 AllocaInst *AI, 2153 SmallVectorImpl<AllocaInst *> &NewElts) { 2154 // If this is a memcpy/memmove, construct the other pointer as the 2155 // appropriate type. The "Other" pointer is the pointer that goes to memory 2156 // that doesn't have anything to do with the alloca that we are promoting. For 2157 // memset, this Value* stays null. 2158 Value *OtherPtr = 0; 2159 unsigned MemAlignment = MI->getAlignment(); 2160 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 2161 if (Inst == MTI->getRawDest()) 2162 OtherPtr = MTI->getRawSource(); 2163 else { 2164 assert(Inst == MTI->getRawSource()); 2165 OtherPtr = MTI->getRawDest(); 2166 } 2167 } 2168 2169 // If there is an other pointer, we want to convert it to the same pointer 2170 // type as AI has, so we can GEP through it safely. 2171 if (OtherPtr) { 2172 unsigned AddrSpace = 2173 cast<PointerType>(OtherPtr->getType())->getAddressSpace(); 2174 2175 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 2176 // optimization, but it's also required to detect the corner case where 2177 // both pointer operands are referencing the same memory, and where 2178 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 2179 // function is only called for mem intrinsics that access the whole 2180 // aggregate, so non-zero GEPs are not an issue here.) 2181 OtherPtr = OtherPtr->stripPointerCasts(); 2182 2183 // Copying the alloca to itself is a no-op: just delete it. 2184 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 2185 // This code will run twice for a no-op memcpy -- once for each operand. 2186 // Put only one reference to MI on the DeadInsts list. 2187 for (SmallVectorImpl<Value *>::const_iterator I = DeadInsts.begin(), 2188 E = DeadInsts.end(); I != E; ++I) 2189 if (*I == MI) return; 2190 DeadInsts.push_back(MI); 2191 return; 2192 } 2193 2194 // If the pointer is not the right type, insert a bitcast to the right 2195 // type. 2196 Type *NewTy = 2197 PointerType::get(AI->getType()->getElementType(), AddrSpace); 2198 2199 if (OtherPtr->getType() != NewTy) 2200 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI); 2201 } 2202 2203 // Process each element of the aggregate. 2204 bool SROADest = MI->getRawDest() == Inst; 2205 2206 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 2207 2208 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2209 // If this is a memcpy/memmove, emit a GEP of the other element address. 2210 Value *OtherElt = 0; 2211 unsigned OtherEltAlign = MemAlignment; 2212 2213 if (OtherPtr) { 2214 Value *Idx[2] = { Zero, 2215 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 2216 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, 2217 OtherPtr->getName()+"."+Twine(i), 2218 MI); 2219 uint64_t EltOffset; 2220 PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 2221 Type *OtherTy = OtherPtrTy->getElementType(); 2222 if (StructType *ST = dyn_cast<StructType>(OtherTy)) { 2223 EltOffset = DL->getStructLayout(ST)->getElementOffset(i); 2224 } else { 2225 Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); 2226 EltOffset = DL->getTypeAllocSize(EltTy)*i; 2227 } 2228 2229 // The alignment of the other pointer is the guaranteed alignment of the 2230 // element, which is affected by both the known alignment of the whole 2231 // mem intrinsic and the alignment of the element. If the alignment of 2232 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 2233 // known alignment is just 4 bytes. 2234 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 2235 } 2236 2237 Value *EltPtr = NewElts[i]; 2238 Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 2239 2240 // If we got down to a scalar, insert a load or store as appropriate. 2241 if (EltTy->isSingleValueType()) { 2242 if (isa<MemTransferInst>(MI)) { 2243 if (SROADest) { 2244 // From Other to Alloca. 2245 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 2246 new StoreInst(Elt, EltPtr, MI); 2247 } else { 2248 // From Alloca to Other. 2249 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 2250 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 2251 } 2252 continue; 2253 } 2254 assert(isa<MemSetInst>(MI)); 2255 2256 // If the stored element is zero (common case), just store a null 2257 // constant. 2258 Constant *StoreVal; 2259 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) { 2260 if (CI->isZero()) { 2261 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 2262 } else { 2263 // If EltTy is a vector type, get the element type. 2264 Type *ValTy = EltTy->getScalarType(); 2265 2266 // Construct an integer with the right value. 2267 unsigned EltSize = DL->getTypeSizeInBits(ValTy); 2268 APInt OneVal(EltSize, CI->getZExtValue()); 2269 APInt TotalVal(OneVal); 2270 // Set each byte. 2271 for (unsigned i = 0; 8*i < EltSize; ++i) { 2272 TotalVal = TotalVal.shl(8); 2273 TotalVal |= OneVal; 2274 } 2275 2276 // Convert the integer value to the appropriate type. 2277 StoreVal = ConstantInt::get(CI->getContext(), TotalVal); 2278 if (ValTy->isPointerTy()) 2279 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 2280 else if (ValTy->isFloatingPointTy()) 2281 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 2282 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 2283 2284 // If the requested value was a vector constant, create it. 2285 if (EltTy->isVectorTy()) { 2286 unsigned NumElts = cast<VectorType>(EltTy)->getNumElements(); 2287 StoreVal = ConstantVector::getSplat(NumElts, StoreVal); 2288 } 2289 } 2290 new StoreInst(StoreVal, EltPtr, MI); 2291 continue; 2292 } 2293 // Otherwise, if we're storing a byte variable, use a memset call for 2294 // this element. 2295 } 2296 2297 unsigned EltSize = DL->getTypeAllocSize(EltTy); 2298 if (!EltSize) 2299 continue; 2300 2301 IRBuilder<> Builder(MI); 2302 2303 // Finally, insert the meminst for this element. 2304 if (isa<MemSetInst>(MI)) { 2305 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize, 2306 MI->isVolatile()); 2307 } else { 2308 assert(isa<MemTransferInst>(MI)); 2309 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr 2310 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr 2311 2312 if (isa<MemCpyInst>(MI)) 2313 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile()); 2314 else 2315 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile()); 2316 } 2317 } 2318 DeadInsts.push_back(MI); 2319} 2320 2321/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 2322/// overwrites the entire allocation. Extract out the pieces of the stored 2323/// integer and store them individually. 2324void 2325SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 2326 SmallVectorImpl<AllocaInst *> &NewElts) { 2327 // Extract each element out of the integer according to its structure offset 2328 // and store the element value to the individual alloca. 2329 Value *SrcVal = SI->getOperand(0); 2330 Type *AllocaEltTy = AI->getAllocatedType(); 2331 uint64_t AllocaSizeBits = DL->getTypeAllocSizeInBits(AllocaEltTy); 2332 2333 IRBuilder<> Builder(SI); 2334 2335 // Handle tail padding by extending the operand 2336 if (DL->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 2337 SrcVal = Builder.CreateZExt(SrcVal, 2338 IntegerType::get(SI->getContext(), AllocaSizeBits)); 2339 2340 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 2341 << '\n'); 2342 2343 // There are two forms here: AI could be an array or struct. Both cases 2344 // have different ways to compute the element offset. 2345 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2346 const StructLayout *Layout = DL->getStructLayout(EltSTy); 2347 2348 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2349 // Get the number of bits to shift SrcVal to get the value. 2350 Type *FieldTy = EltSTy->getElementType(i); 2351 uint64_t Shift = Layout->getElementOffsetInBits(i); 2352 2353 if (DL->isBigEndian()) 2354 Shift = AllocaSizeBits-Shift-DL->getTypeAllocSizeInBits(FieldTy); 2355 2356 Value *EltVal = SrcVal; 2357 if (Shift) { 2358 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2359 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2360 } 2361 2362 // Truncate down to an integer of the right size. 2363 uint64_t FieldSizeBits = DL->getTypeSizeInBits(FieldTy); 2364 2365 // Ignore zero sized fields like {}, they obviously contain no data. 2366 if (FieldSizeBits == 0) continue; 2367 2368 if (FieldSizeBits != AllocaSizeBits) 2369 EltVal = Builder.CreateTrunc(EltVal, 2370 IntegerType::get(SI->getContext(), FieldSizeBits)); 2371 Value *DestField = NewElts[i]; 2372 if (EltVal->getType() == FieldTy) { 2373 // Storing to an integer field of this size, just do it. 2374 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 2375 // Bitcast to the right element type (for fp/vector values). 2376 EltVal = Builder.CreateBitCast(EltVal, FieldTy); 2377 } else { 2378 // Otherwise, bitcast the dest pointer (for aggregates). 2379 DestField = Builder.CreateBitCast(DestField, 2380 PointerType::getUnqual(EltVal->getType())); 2381 } 2382 new StoreInst(EltVal, DestField, SI); 2383 } 2384 2385 } else { 2386 ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 2387 Type *ArrayEltTy = ATy->getElementType(); 2388 uint64_t ElementOffset = DL->getTypeAllocSizeInBits(ArrayEltTy); 2389 uint64_t ElementSizeBits = DL->getTypeSizeInBits(ArrayEltTy); 2390 2391 uint64_t Shift; 2392 2393 if (DL->isBigEndian()) 2394 Shift = AllocaSizeBits-ElementOffset; 2395 else 2396 Shift = 0; 2397 2398 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2399 // Ignore zero sized fields like {}, they obviously contain no data. 2400 if (ElementSizeBits == 0) continue; 2401 2402 Value *EltVal = SrcVal; 2403 if (Shift) { 2404 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2405 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2406 } 2407 2408 // Truncate down to an integer of the right size. 2409 if (ElementSizeBits != AllocaSizeBits) 2410 EltVal = Builder.CreateTrunc(EltVal, 2411 IntegerType::get(SI->getContext(), 2412 ElementSizeBits)); 2413 Value *DestField = NewElts[i]; 2414 if (EltVal->getType() == ArrayEltTy) { 2415 // Storing to an integer field of this size, just do it. 2416 } else if (ArrayEltTy->isFloatingPointTy() || 2417 ArrayEltTy->isVectorTy()) { 2418 // Bitcast to the right element type (for fp/vector values). 2419 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy); 2420 } else { 2421 // Otherwise, bitcast the dest pointer (for aggregates). 2422 DestField = Builder.CreateBitCast(DestField, 2423 PointerType::getUnqual(EltVal->getType())); 2424 } 2425 new StoreInst(EltVal, DestField, SI); 2426 2427 if (DL->isBigEndian()) 2428 Shift -= ElementOffset; 2429 else 2430 Shift += ElementOffset; 2431 } 2432 } 2433 2434 DeadInsts.push_back(SI); 2435} 2436 2437/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 2438/// an integer. Load the individual pieces to form the aggregate value. 2439void 2440SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 2441 SmallVectorImpl<AllocaInst *> &NewElts) { 2442 // Extract each element out of the NewElts according to its structure offset 2443 // and form the result value. 2444 Type *AllocaEltTy = AI->getAllocatedType(); 2445 uint64_t AllocaSizeBits = DL->getTypeAllocSizeInBits(AllocaEltTy); 2446 2447 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 2448 << '\n'); 2449 2450 // There are two forms here: AI could be an array or struct. Both cases 2451 // have different ways to compute the element offset. 2452 const StructLayout *Layout = 0; 2453 uint64_t ArrayEltBitOffset = 0; 2454 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2455 Layout = DL->getStructLayout(EltSTy); 2456 } else { 2457 Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 2458 ArrayEltBitOffset = DL->getTypeAllocSizeInBits(ArrayEltTy); 2459 } 2460 2461 Value *ResultVal = 2462 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 2463 2464 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2465 // Load the value from the alloca. If the NewElt is an aggregate, cast 2466 // the pointer to an integer of the same size before doing the load. 2467 Value *SrcField = NewElts[i]; 2468 Type *FieldTy = 2469 cast<PointerType>(SrcField->getType())->getElementType(); 2470 uint64_t FieldSizeBits = DL->getTypeSizeInBits(FieldTy); 2471 2472 // Ignore zero sized fields like {}, they obviously contain no data. 2473 if (FieldSizeBits == 0) continue; 2474 2475 IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 2476 FieldSizeBits); 2477 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 2478 !FieldTy->isVectorTy()) 2479 SrcField = new BitCastInst(SrcField, 2480 PointerType::getUnqual(FieldIntTy), 2481 "", LI); 2482 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 2483 2484 // If SrcField is a fp or vector of the right size but that isn't an 2485 // integer type, bitcast to an integer so we can shift it. 2486 if (SrcField->getType() != FieldIntTy) 2487 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 2488 2489 // Zero extend the field to be the same size as the final alloca so that 2490 // we can shift and insert it. 2491 if (SrcField->getType() != ResultVal->getType()) 2492 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 2493 2494 // Determine the number of bits to shift SrcField. 2495 uint64_t Shift; 2496 if (Layout) // Struct case. 2497 Shift = Layout->getElementOffsetInBits(i); 2498 else // Array case. 2499 Shift = i*ArrayEltBitOffset; 2500 2501 if (DL->isBigEndian()) 2502 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 2503 2504 if (Shift) { 2505 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 2506 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 2507 } 2508 2509 // Don't create an 'or x, 0' on the first iteration. 2510 if (!isa<Constant>(ResultVal) || 2511 !cast<Constant>(ResultVal)->isNullValue()) 2512 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 2513 else 2514 ResultVal = SrcField; 2515 } 2516 2517 // Handle tail padding by truncating the result 2518 if (DL->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 2519 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 2520 2521 LI->replaceAllUsesWith(ResultVal); 2522 DeadInsts.push_back(LI); 2523} 2524 2525/// HasPadding - Return true if the specified type has any structure or 2526/// alignment padding in between the elements that would be split apart 2527/// by SROA; return false otherwise. 2528static bool HasPadding(Type *Ty, const DataLayout &DL) { 2529 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 2530 Ty = ATy->getElementType(); 2531 return DL.getTypeSizeInBits(Ty) != DL.getTypeAllocSizeInBits(Ty); 2532 } 2533 2534 // SROA currently handles only Arrays and Structs. 2535 StructType *STy = cast<StructType>(Ty); 2536 const StructLayout *SL = DL.getStructLayout(STy); 2537 unsigned PrevFieldBitOffset = 0; 2538 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 2539 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 2540 2541 // Check to see if there is any padding between this element and the 2542 // previous one. 2543 if (i) { 2544 unsigned PrevFieldEnd = 2545 PrevFieldBitOffset+DL.getTypeSizeInBits(STy->getElementType(i-1)); 2546 if (PrevFieldEnd < FieldBitOffset) 2547 return true; 2548 } 2549 PrevFieldBitOffset = FieldBitOffset; 2550 } 2551 // Check for tail padding. 2552 if (unsigned EltCount = STy->getNumElements()) { 2553 unsigned PrevFieldEnd = PrevFieldBitOffset + 2554 DL.getTypeSizeInBits(STy->getElementType(EltCount-1)); 2555 if (PrevFieldEnd < SL->getSizeInBits()) 2556 return true; 2557 } 2558 return false; 2559} 2560 2561/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 2562/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 2563/// or 1 if safe after canonicalization has been performed. 2564bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 2565 // Loop over the use list of the alloca. We can only transform it if all of 2566 // the users are safe to transform. 2567 AllocaInfo Info(AI); 2568 2569 isSafeForScalarRepl(AI, 0, Info); 2570 if (Info.isUnsafe) { 2571 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 2572 return false; 2573 } 2574 2575 // Okay, we know all the users are promotable. If the aggregate is a memcpy 2576 // source and destination, we have to be careful. In particular, the memcpy 2577 // could be moving around elements that live in structure padding of the LLVM 2578 // types, but may actually be used. In these cases, we refuse to promote the 2579 // struct. 2580 if (Info.isMemCpySrc && Info.isMemCpyDst && 2581 HasPadding(AI->getAllocatedType(), *DL)) 2582 return false; 2583 2584 // If the alloca never has an access to just *part* of it, but is accessed 2585 // via loads and stores, then we should use ConvertToScalarInfo to promote 2586 // the alloca instead of promoting each piece at a time and inserting fission 2587 // and fusion code. 2588 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) { 2589 // If the struct/array just has one element, use basic SRoA. 2590 if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 2591 if (ST->getNumElements() > 1) return false; 2592 } else { 2593 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1) 2594 return false; 2595 } 2596 } 2597 2598 return true; 2599} 2600