ScalarReplAggregates.cpp revision a94d6e87c4c49f2e81b01d66d8bfb591277f8f96
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/Constants.h" 25#include "llvm/DIBuilder.h" 26#include "llvm/DebugInfo.h" 27#include "llvm/DerivedTypes.h" 28#include "llvm/Function.h" 29#include "llvm/GlobalVariable.h" 30#include "llvm/IRBuilder.h" 31#include "llvm/Instructions.h" 32#include "llvm/IntrinsicInst.h" 33#include "llvm/LLVMContext.h" 34#include "llvm/Module.h" 35#include "llvm/Operator.h" 36#include "llvm/Pass.h" 37#include "llvm/ADT/SetVector.h" 38#include "llvm/ADT/SmallVector.h" 39#include "llvm/ADT/Statistic.h" 40#include "llvm/Analysis/Dominators.h" 41#include "llvm/Analysis/Loads.h" 42#include "llvm/Analysis/ValueTracking.h" 43#include "llvm/Support/CallSite.h" 44#include "llvm/Support/Debug.h" 45#include "llvm/Support/ErrorHandling.h" 46#include "llvm/Support/GetElementPtrTypeIterator.h" 47#include "llvm/Support/MathExtras.h" 48#include "llvm/Support/raw_ostream.h" 49#include "llvm/Target/TargetData.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"); 59STATISTIC(NumGlobals, "Number of allocas copied from constant global"); 60 61namespace { 62 struct SROA : public FunctionPass { 63 SROA(int T, bool hasDT, char &ID, int ST, int AT, int SLT) 64 : FunctionPass(ID), HasDomTree(hasDT) { 65 if (T == -1) 66 SRThreshold = 128; 67 else 68 SRThreshold = T; 69 if (ST == -1) 70 StructMemberThreshold = 32; 71 else 72 StructMemberThreshold = ST; 73 if (AT == -1) 74 ArrayElementThreshold = 8; 75 else 76 ArrayElementThreshold = AT; 77 if (SLT == -1) 78 // Do not limit the scalar integer load size if no threshold is given. 79 ScalarLoadThreshold = -1; 80 else 81 ScalarLoadThreshold = SLT; 82 } 83 84 bool runOnFunction(Function &F); 85 86 bool performScalarRepl(Function &F); 87 bool performPromotion(Function &F); 88 89 private: 90 bool HasDomTree; 91 TargetData *TD; 92 93 /// DeadInsts - Keep track of instructions we have made dead, so that 94 /// we can remove them after we are done working. 95 SmallVector<Value*, 32> DeadInsts; 96 97 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures 98 /// information about the uses. All these fields are initialized to false 99 /// and set to true when something is learned. 100 struct AllocaInfo { 101 /// The alloca to promote. 102 AllocaInst *AI; 103 104 /// CheckedPHIs - This is a set of verified PHI nodes, to prevent infinite 105 /// looping and avoid redundant work. 106 SmallPtrSet<PHINode*, 8> CheckedPHIs; 107 108 /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 109 bool isUnsafe : 1; 110 111 /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 112 bool isMemCpySrc : 1; 113 114 /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 115 bool isMemCpyDst : 1; 116 117 /// hasSubelementAccess - This is true if a subelement of the alloca is 118 /// ever accessed, or false if the alloca is only accessed with mem 119 /// intrinsics or load/store that only access the entire alloca at once. 120 bool hasSubelementAccess : 1; 121 122 /// hasALoadOrStore - This is true if there are any loads or stores to it. 123 /// The alloca may just be accessed with memcpy, for example, which would 124 /// not set this. 125 bool hasALoadOrStore : 1; 126 127 explicit AllocaInfo(AllocaInst *ai) 128 : AI(ai), isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false), 129 hasSubelementAccess(false), hasALoadOrStore(false) {} 130 }; 131 132 /// SRThreshold - The maximum alloca size to considered for SROA. 133 unsigned SRThreshold; 134 135 /// StructMemberThreshold - The maximum number of members a struct can 136 /// contain to be considered for SROA. 137 unsigned StructMemberThreshold; 138 139 /// ArrayElementThreshold - The maximum number of elements an array can 140 /// have to be considered for SROA. 141 unsigned ArrayElementThreshold; 142 143 /// ScalarLoadThreshold - The maximum size in bits of scalars to load when 144 /// converting to scalar 145 unsigned ScalarLoadThreshold; 146 147 void MarkUnsafe(AllocaInfo &I, Instruction *User) { 148 I.isUnsafe = true; 149 DEBUG(dbgs() << " Transformation preventing inst: " << *User << '\n'); 150 } 151 152 bool isSafeAllocaToScalarRepl(AllocaInst *AI); 153 154 void isSafeForScalarRepl(Instruction *I, uint64_t Offset, AllocaInfo &Info); 155 void isSafePHISelectUseForScalarRepl(Instruction *User, uint64_t Offset, 156 AllocaInfo &Info); 157 void isSafeGEP(GetElementPtrInst *GEPI, uint64_t &Offset, AllocaInfo &Info); 158 void isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 159 Type *MemOpType, bool isStore, AllocaInfo &Info, 160 Instruction *TheAccess, bool AllowWholeAccess); 161 bool TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size); 162 uint64_t FindElementAndOffset(Type *&T, uint64_t &Offset, 163 Type *&IdxTy); 164 165 void DoScalarReplacement(AllocaInst *AI, 166 std::vector<AllocaInst*> &WorkList); 167 void DeleteDeadInstructions(); 168 169 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 170 SmallVector<AllocaInst*, 32> &NewElts); 171 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 172 SmallVector<AllocaInst*, 32> &NewElts); 173 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 174 SmallVector<AllocaInst*, 32> &NewElts); 175 void RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI, 176 uint64_t Offset, 177 SmallVector<AllocaInst*, 32> &NewElts); 178 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 179 AllocaInst *AI, 180 SmallVector<AllocaInst*, 32> &NewElts); 181 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 182 SmallVector<AllocaInst*, 32> &NewElts); 183 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 184 SmallVector<AllocaInst*, 32> &NewElts); 185 bool ShouldAttemptScalarRepl(AllocaInst *AI); 186 187 static MemTransferInst *isOnlyCopiedFromConstantGlobal( 188 AllocaInst *AI, SmallVector<Instruction*, 4> &ToDelete); 189 }; 190 191 // SROA_DT - SROA that uses DominatorTree. 192 struct SROA_DT : public SROA { 193 static char ID; 194 public: 195 SROA_DT(int T = -1, int ST = -1, int AT = -1, int SLT = -1) : 196 SROA(T, true, ID, ST, AT, SLT) { 197 initializeSROA_DTPass(*PassRegistry::getPassRegistry()); 198 } 199 200 // getAnalysisUsage - This pass does not require any passes, but we know it 201 // will not alter the CFG, so say so. 202 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 203 AU.addRequired<DominatorTree>(); 204 AU.setPreservesCFG(); 205 } 206 }; 207 208 // SROA_SSAUp - SROA that uses SSAUpdater. 209 struct SROA_SSAUp : public SROA { 210 static char ID; 211 public: 212 SROA_SSAUp(int T = -1, int ST = -1, int AT = -1, int SLT = -1) : 213 SROA(T, false, ID, ST, AT, SLT) { 214 initializeSROA_SSAUpPass(*PassRegistry::getPassRegistry()); 215 } 216 217 // getAnalysisUsage - This pass does not require any passes, but we know it 218 // will not alter the CFG, so say so. 219 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 220 AU.setPreservesCFG(); 221 } 222 }; 223 224} 225 226char SROA_DT::ID = 0; 227char SROA_SSAUp::ID = 0; 228 229INITIALIZE_PASS_BEGIN(SROA_DT, "scalarrepl", 230 "Scalar Replacement of Aggregates (DT)", false, false) 231INITIALIZE_PASS_DEPENDENCY(DominatorTree) 232INITIALIZE_PASS_END(SROA_DT, "scalarrepl", 233 "Scalar Replacement of Aggregates (DT)", false, false) 234 235INITIALIZE_PASS_BEGIN(SROA_SSAUp, "scalarrepl-ssa", 236 "Scalar Replacement of Aggregates (SSAUp)", false, false) 237INITIALIZE_PASS_END(SROA_SSAUp, "scalarrepl-ssa", 238 "Scalar Replacement of Aggregates (SSAUp)", false, false) 239 240// Public interface to the ScalarReplAggregates pass 241FunctionPass *llvm::createScalarReplAggregatesPass(int Threshold, 242 bool UseDomTree, 243 int StructMemberThreshold, 244 int ArrayElementThreshold, 245 int ScalarLoadThreshold) { 246 if (UseDomTree) 247 return new SROA_DT(Threshold, StructMemberThreshold, ArrayElementThreshold, 248 ScalarLoadThreshold); 249 return new SROA_SSAUp(Threshold, StructMemberThreshold, 250 ArrayElementThreshold, ScalarLoadThreshold); 251} 252 253 254//===----------------------------------------------------------------------===// 255// Convert To Scalar Optimization. 256//===----------------------------------------------------------------------===// 257 258namespace { 259/// ConvertToScalarInfo - This class implements the "Convert To Scalar" 260/// optimization, which scans the uses of an alloca and determines if it can 261/// rewrite it in terms of a single new alloca that can be mem2reg'd. 262class ConvertToScalarInfo { 263 /// AllocaSize - The size of the alloca being considered in bytes. 264 unsigned AllocaSize; 265 const TargetData &TD; 266 unsigned ScalarLoadThreshold; 267 268 /// IsNotTrivial - This is set to true if there is some access to the object 269 /// which means that mem2reg can't promote it. 270 bool IsNotTrivial; 271 272 /// ScalarKind - Tracks the kind of alloca being considered for promotion, 273 /// computed based on the uses of the alloca rather than the LLVM type system. 274 enum { 275 Unknown, 276 277 // Accesses via GEPs that are consistent with element access of a vector 278 // type. This will not be converted into a vector unless there is a later 279 // access using an actual vector type. 280 ImplicitVector, 281 282 // Accesses via vector operations and GEPs that are consistent with the 283 // layout of a vector type. 284 Vector, 285 286 // An integer bag-of-bits with bitwise operations for insertion and 287 // extraction. Any combination of types can be converted into this kind 288 // of scalar. 289 Integer 290 } ScalarKind; 291 292 /// VectorTy - This tracks the type that we should promote the vector to if 293 /// it is possible to turn it into a vector. This starts out null, and if it 294 /// isn't possible to turn into a vector type, it gets set to VoidTy. 295 VectorType *VectorTy; 296 297 /// HadNonMemTransferAccess - True if there is at least one access to the 298 /// alloca that is not a MemTransferInst. We don't want to turn structs into 299 /// large integers unless there is some potential for optimization. 300 bool HadNonMemTransferAccess; 301 302 /// HadDynamicAccess - True if some element of this alloca was dynamic. 303 /// We don't yet have support for turning a dynamic access into a large 304 /// integer. 305 bool HadDynamicAccess; 306 307public: 308 explicit ConvertToScalarInfo(unsigned Size, const TargetData &td, 309 unsigned SLT) 310 : AllocaSize(Size), TD(td), ScalarLoadThreshold(SLT), IsNotTrivial(false), 311 ScalarKind(Unknown), VectorTy(0), HadNonMemTransferAccess(false), 312 HadDynamicAccess(false) { } 313 314 AllocaInst *TryConvert(AllocaInst *AI); 315 316private: 317 bool CanConvertToScalar(Value *V, uint64_t Offset, Value* NonConstantIdx); 318 void MergeInTypeForLoadOrStore(Type *In, uint64_t Offset); 319 bool MergeInVectorType(VectorType *VInTy, uint64_t Offset); 320 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset, 321 Value *NonConstantIdx); 322 323 Value *ConvertScalar_ExtractValue(Value *NV, Type *ToType, 324 uint64_t Offset, Value* NonConstantIdx, 325 IRBuilder<> &Builder); 326 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 327 uint64_t Offset, Value* NonConstantIdx, 328 IRBuilder<> &Builder); 329}; 330} // end anonymous namespace. 331 332 333/// TryConvert - Analyze the specified alloca, and if it is safe to do so, 334/// rewrite it to be a new alloca which is mem2reg'able. This returns the new 335/// alloca if possible or null if not. 336AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) { 337 // If we can't convert this scalar, or if mem2reg can trivially do it, bail 338 // out. 339 if (!CanConvertToScalar(AI, 0, 0) || !IsNotTrivial) 340 return 0; 341 342 // If an alloca has only memset / memcpy uses, it may still have an Unknown 343 // ScalarKind. Treat it as an Integer below. 344 if (ScalarKind == Unknown) 345 ScalarKind = Integer; 346 347 if (ScalarKind == Vector && VectorTy->getBitWidth() != AllocaSize * 8) 348 ScalarKind = Integer; 349 350 // If we were able to find a vector type that can handle this with 351 // insert/extract elements, and if there was at least one use that had 352 // a vector type, promote this to a vector. We don't want to promote 353 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 354 // we just get a lot of insert/extracts. If at least one vector is 355 // involved, then we probably really do have a union of vector/array. 356 Type *NewTy; 357 if (ScalarKind == Vector) { 358 assert(VectorTy && "Missing type for vector scalar."); 359 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = " 360 << *VectorTy << '\n'); 361 NewTy = VectorTy; // Use the vector type. 362 } else { 363 unsigned BitWidth = AllocaSize * 8; 364 365 // Do not convert to scalar integer if the alloca size exceeds the 366 // scalar load threshold. 367 if (BitWidth > ScalarLoadThreshold) 368 return 0; 369 370 if ((ScalarKind == ImplicitVector || ScalarKind == Integer) && 371 !HadNonMemTransferAccess && !TD.fitsInLegalInteger(BitWidth)) 372 return 0; 373 // Dynamic accesses on integers aren't yet supported. They need us to shift 374 // by a dynamic amount which could be difficult to work out as we might not 375 // know whether to use a left or right shift. 376 if (ScalarKind == Integer && HadDynamicAccess) 377 return 0; 378 379 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"); 380 // Create and insert the integer alloca. 381 NewTy = IntegerType::get(AI->getContext(), BitWidth); 382 } 383 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 384 ConvertUsesToScalar(AI, NewAI, 0, 0); 385 return NewAI; 386} 387 388/// MergeInTypeForLoadOrStore - Add the 'In' type to the accumulated vector type 389/// (VectorTy) so far at the offset specified by Offset (which is specified in 390/// bytes). 391/// 392/// There are two cases we handle here: 393/// 1) A union of vector types of the same size and potentially its elements. 394/// Here we turn element accesses into insert/extract element operations. 395/// This promotes a <4 x float> with a store of float to the third element 396/// into a <4 x float> that uses insert element. 397/// 2) A fully general blob of memory, which we turn into some (potentially 398/// large) integer type with extract and insert operations where the loads 399/// and stores would mutate the memory. We mark this by setting VectorTy 400/// to VoidTy. 401void ConvertToScalarInfo::MergeInTypeForLoadOrStore(Type *In, 402 uint64_t Offset) { 403 // If we already decided to turn this into a blob of integer memory, there is 404 // nothing to be done. 405 if (ScalarKind == Integer) 406 return; 407 408 // If this could be contributing to a vector, analyze it. 409 410 // If the In type is a vector that is the same size as the alloca, see if it 411 // matches the existing VecTy. 412 if (VectorType *VInTy = dyn_cast<VectorType>(In)) { 413 if (MergeInVectorType(VInTy, Offset)) 414 return; 415 } else if (In->isFloatTy() || In->isDoubleTy() || 416 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 && 417 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 418 // Full width accesses can be ignored, because they can always be turned 419 // into bitcasts. 420 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 421 if (EltSize == AllocaSize) 422 return; 423 424 // If we're accessing something that could be an element of a vector, see 425 // if the implied vector agrees with what we already have and if Offset is 426 // compatible with it. 427 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 && 428 (!VectorTy || EltSize == VectorTy->getElementType() 429 ->getPrimitiveSizeInBits()/8)) { 430 if (!VectorTy) { 431 ScalarKind = ImplicitVector; 432 VectorTy = VectorType::get(In, AllocaSize/EltSize); 433 } 434 return; 435 } 436 } 437 438 // Otherwise, we have a case that we can't handle with an optimized vector 439 // form. We can still turn this into a large integer. 440 ScalarKind = Integer; 441} 442 443/// MergeInVectorType - Handles the vector case of MergeInTypeForLoadOrStore, 444/// returning true if the type was successfully merged and false otherwise. 445bool ConvertToScalarInfo::MergeInVectorType(VectorType *VInTy, 446 uint64_t Offset) { 447 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 448 // If we're storing/loading a vector of the right size, allow it as a 449 // vector. If this the first vector we see, remember the type so that 450 // we know the element size. If this is a subsequent access, ignore it 451 // even if it is a differing type but the same size. Worst case we can 452 // bitcast the resultant vectors. 453 if (!VectorTy) 454 VectorTy = VInTy; 455 ScalarKind = Vector; 456 return true; 457 } 458 459 return false; 460} 461 462/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 463/// its accesses to a single vector type, return true and set VecTy to 464/// the new type. If we could convert the alloca into a single promotable 465/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 466/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 467/// is the current offset from the base of the alloca being analyzed. 468/// 469/// If we see at least one access to the value that is as a vector type, set the 470/// SawVec flag. 471bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset, 472 Value* NonConstantIdx) { 473 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 474 Instruction *User = cast<Instruction>(*UI); 475 476 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 477 // Don't break volatile loads. 478 if (!LI->isSimple()) 479 return false; 480 // Don't touch MMX operations. 481 if (LI->getType()->isX86_MMXTy()) 482 return false; 483 HadNonMemTransferAccess = true; 484 MergeInTypeForLoadOrStore(LI->getType(), Offset); 485 continue; 486 } 487 488 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 489 // Storing the pointer, not into the value? 490 if (SI->getOperand(0) == V || !SI->isSimple()) return false; 491 // Don't touch MMX operations. 492 if (SI->getOperand(0)->getType()->isX86_MMXTy()) 493 return false; 494 HadNonMemTransferAccess = true; 495 MergeInTypeForLoadOrStore(SI->getOperand(0)->getType(), Offset); 496 continue; 497 } 498 499 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 500 if (!onlyUsedByLifetimeMarkers(BCI)) 501 IsNotTrivial = true; // Can't be mem2reg'd. 502 if (!CanConvertToScalar(BCI, Offset, NonConstantIdx)) 503 return false; 504 continue; 505 } 506 507 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 508 // If this is a GEP with a variable indices, we can't handle it. 509 PointerType* PtrTy = dyn_cast<PointerType>(GEP->getPointerOperandType()); 510 if (!PtrTy) 511 return false; 512 513 // Compute the offset that this GEP adds to the pointer. 514 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 515 Value *GEPNonConstantIdx = 0; 516 if (!GEP->hasAllConstantIndices()) { 517 if (!isa<VectorType>(PtrTy->getElementType())) 518 return false; 519 if (NonConstantIdx) 520 return false; 521 GEPNonConstantIdx = Indices.pop_back_val(); 522 if (!GEPNonConstantIdx->getType()->isIntegerTy(32)) 523 return false; 524 HadDynamicAccess = true; 525 } else 526 GEPNonConstantIdx = NonConstantIdx; 527 uint64_t GEPOffset = TD.getIndexedOffset(PtrTy, 528 Indices); 529 // See if all uses can be converted. 530 if (!CanConvertToScalar(GEP, Offset+GEPOffset, GEPNonConstantIdx)) 531 return false; 532 IsNotTrivial = true; // Can't be mem2reg'd. 533 HadNonMemTransferAccess = true; 534 continue; 535 } 536 537 // If this is a constant sized memset of a constant value (e.g. 0) we can 538 // handle it. 539 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 540 // Store to dynamic index. 541 if (NonConstantIdx) 542 return false; 543 // Store of constant value. 544 if (!isa<ConstantInt>(MSI->getValue())) 545 return false; 546 547 // Store of constant size. 548 ConstantInt *Len = dyn_cast<ConstantInt>(MSI->getLength()); 549 if (!Len) 550 return false; 551 552 // If the size differs from the alloca, we can only convert the alloca to 553 // an integer bag-of-bits. 554 // FIXME: This should handle all of the cases that are currently accepted 555 // as vector element insertions. 556 if (Len->getZExtValue() != AllocaSize || Offset != 0) 557 ScalarKind = Integer; 558 559 IsNotTrivial = true; // Can't be mem2reg'd. 560 HadNonMemTransferAccess = true; 561 continue; 562 } 563 564 // If this is a memcpy or memmove into or out of the whole allocation, we 565 // can handle it like a load or store of the scalar type. 566 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 567 // Store to dynamic index. 568 if (NonConstantIdx) 569 return false; 570 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()); 571 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0) 572 return false; 573 574 IsNotTrivial = true; // Can't be mem2reg'd. 575 continue; 576 } 577 578 // If this is a lifetime intrinsic, we can handle it. 579 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 580 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 581 II->getIntrinsicID() == Intrinsic::lifetime_end) { 582 continue; 583 } 584 } 585 586 // Otherwise, we cannot handle this! 587 return false; 588 } 589 590 return true; 591} 592 593/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 594/// directly. This happens when we are converting an "integer union" to a 595/// single integer scalar, or when we are converting a "vector union" to a 596/// vector with insert/extractelement instructions. 597/// 598/// Offset is an offset from the original alloca, in bits that need to be 599/// shifted to the right. By the end of this, there should be no uses of Ptr. 600void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, 601 uint64_t Offset, 602 Value* NonConstantIdx) { 603 while (!Ptr->use_empty()) { 604 Instruction *User = cast<Instruction>(Ptr->use_back()); 605 606 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 607 ConvertUsesToScalar(CI, NewAI, Offset, NonConstantIdx); 608 CI->eraseFromParent(); 609 continue; 610 } 611 612 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 613 // Compute the offset that this GEP adds to the pointer. 614 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 615 if (!GEP->hasAllConstantIndices()) 616 NonConstantIdx = Indices.pop_back_val(); 617 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(), 618 Indices); 619 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8, NonConstantIdx); 620 GEP->eraseFromParent(); 621 continue; 622 } 623 624 IRBuilder<> Builder(User); 625 626 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 627 // The load is a bit extract from NewAI shifted right by Offset bits. 628 Value *LoadedVal = Builder.CreateLoad(NewAI); 629 Value *NewLoadVal 630 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, 631 NonConstantIdx, Builder); 632 LI->replaceAllUsesWith(NewLoadVal); 633 LI->eraseFromParent(); 634 continue; 635 } 636 637 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 638 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 639 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 640 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 641 NonConstantIdx, Builder); 642 Builder.CreateStore(New, NewAI); 643 SI->eraseFromParent(); 644 645 // If the load we just inserted is now dead, then the inserted store 646 // overwrote the entire thing. 647 if (Old->use_empty()) 648 Old->eraseFromParent(); 649 continue; 650 } 651 652 // If this is a constant sized memset of a constant value (e.g. 0) we can 653 // transform it into a store of the expanded constant value. 654 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 655 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 656 assert(!NonConstantIdx && "Cannot replace dynamic memset with insert"); 657 int64_t SNumBytes = cast<ConstantInt>(MSI->getLength())->getSExtValue(); 658 if (SNumBytes > 0 && (SNumBytes >> 32) == 0) { 659 unsigned NumBytes = static_cast<unsigned>(SNumBytes); 660 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 661 662 // Compute the value replicated the right number of times. 663 APInt APVal(NumBytes*8, Val); 664 665 // Splat the value if non-zero. 666 if (Val) 667 for (unsigned i = 1; i != NumBytes; ++i) 668 APVal |= APVal << 8; 669 670 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 671 Value *New = ConvertScalar_InsertValue( 672 ConstantInt::get(User->getContext(), APVal), 673 Old, Offset, 0, Builder); 674 Builder.CreateStore(New, NewAI); 675 676 // If the load we just inserted is now dead, then the memset overwrote 677 // the entire thing. 678 if (Old->use_empty()) 679 Old->eraseFromParent(); 680 } 681 MSI->eraseFromParent(); 682 continue; 683 } 684 685 // If this is a memcpy or memmove into or out of the whole allocation, we 686 // can handle it like a load or store of the scalar type. 687 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 688 assert(Offset == 0 && "must be store to start of alloca"); 689 assert(!NonConstantIdx && "Cannot replace dynamic transfer with insert"); 690 691 // If the source and destination are both to the same alloca, then this is 692 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 693 // as appropriate. 694 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, &TD, 0)); 695 696 if (GetUnderlyingObject(MTI->getSource(), &TD, 0) != OrigAI) { 697 // Dest must be OrigAI, change this to be a load from the original 698 // pointer (bitcasted), then a store to our new alloca. 699 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 700 Value *SrcPtr = MTI->getSource(); 701 PointerType* SPTy = cast<PointerType>(SrcPtr->getType()); 702 PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 703 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 704 AIPTy = PointerType::get(AIPTy->getElementType(), 705 SPTy->getAddressSpace()); 706 } 707 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy); 708 709 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 710 SrcVal->setAlignment(MTI->getAlignment()); 711 Builder.CreateStore(SrcVal, NewAI); 712 } else if (GetUnderlyingObject(MTI->getDest(), &TD, 0) != OrigAI) { 713 // Src must be OrigAI, change this to be a load from NewAI then a store 714 // through the original dest pointer (bitcasted). 715 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 716 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 717 718 PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType()); 719 PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 720 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 721 AIPTy = PointerType::get(AIPTy->getElementType(), 722 DPTy->getAddressSpace()); 723 } 724 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy); 725 726 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 727 NewStore->setAlignment(MTI->getAlignment()); 728 } else { 729 // Noop transfer. Src == Dst 730 } 731 732 MTI->eraseFromParent(); 733 continue; 734 } 735 736 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 737 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 738 II->getIntrinsicID() == Intrinsic::lifetime_end) { 739 // There's no need to preserve these, as the resulting alloca will be 740 // converted to a register anyways. 741 II->eraseFromParent(); 742 continue; 743 } 744 } 745 746 llvm_unreachable("Unsupported operation!"); 747 } 748} 749 750/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 751/// or vector value FromVal, extracting the bits from the offset specified by 752/// Offset. This returns the value, which is of type ToType. 753/// 754/// This happens when we are converting an "integer union" to a single 755/// integer scalar, or when we are converting a "vector union" to a vector with 756/// insert/extractelement instructions. 757/// 758/// Offset is an offset from the original alloca, in bits that need to be 759/// shifted to the right. 760Value *ConvertToScalarInfo:: 761ConvertScalar_ExtractValue(Value *FromVal, Type *ToType, 762 uint64_t Offset, Value* NonConstantIdx, 763 IRBuilder<> &Builder) { 764 // If the load is of the whole new alloca, no conversion is needed. 765 Type *FromType = FromVal->getType(); 766 if (FromType == ToType && Offset == 0) 767 return FromVal; 768 769 // If the result alloca is a vector type, this is either an element 770 // access or a bitcast to another vector type of the same size. 771 if (VectorType *VTy = dyn_cast<VectorType>(FromType)) { 772 unsigned FromTypeSize = TD.getTypeAllocSize(FromType); 773 unsigned ToTypeSize = TD.getTypeAllocSize(ToType); 774 if (FromTypeSize == ToTypeSize) 775 return Builder.CreateBitCast(FromVal, ToType); 776 777 // Otherwise it must be an element access. 778 unsigned Elt = 0; 779 if (Offset) { 780 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType()); 781 Elt = Offset/EltSize; 782 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 783 } 784 // Return the element extracted out of it. 785 Value *Idx; 786 if (NonConstantIdx) { 787 if (Elt) 788 Idx = Builder.CreateAdd(NonConstantIdx, 789 Builder.getInt32(Elt), 790 "dyn.offset"); 791 else 792 Idx = NonConstantIdx; 793 } else 794 Idx = Builder.getInt32(Elt); 795 Value *V = Builder.CreateExtractElement(FromVal, Idx); 796 if (V->getType() != ToType) 797 V = Builder.CreateBitCast(V, ToType); 798 return V; 799 } 800 801 // If ToType is a first class aggregate, extract out each of the pieces and 802 // use insertvalue's to form the FCA. 803 if (StructType *ST = dyn_cast<StructType>(ToType)) { 804 assert(!NonConstantIdx && 805 "Dynamic indexing into struct types not supported"); 806 const StructLayout &Layout = *TD.getStructLayout(ST); 807 Value *Res = UndefValue::get(ST); 808 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 809 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 810 Offset+Layout.getElementOffsetInBits(i), 811 0, Builder); 812 Res = Builder.CreateInsertValue(Res, Elt, i); 813 } 814 return Res; 815 } 816 817 if (ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 818 assert(!NonConstantIdx && 819 "Dynamic indexing into array types not supported"); 820 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 821 Value *Res = UndefValue::get(AT); 822 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 823 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 824 Offset+i*EltSize, 0, Builder); 825 Res = Builder.CreateInsertValue(Res, Elt, i); 826 } 827 return Res; 828 } 829 830 // Otherwise, this must be a union that was converted to an integer value. 831 IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 832 833 // If this is a big-endian system and the load is narrower than the 834 // full alloca type, we need to do a shift to get the right bits. 835 int ShAmt = 0; 836 if (TD.isBigEndian()) { 837 // On big-endian machines, the lowest bit is stored at the bit offset 838 // from the pointer given by getTypeStoreSizeInBits. This matters for 839 // integers with a bitwidth that is not a multiple of 8. 840 ShAmt = TD.getTypeStoreSizeInBits(NTy) - 841 TD.getTypeStoreSizeInBits(ToType) - Offset; 842 } else { 843 ShAmt = Offset; 844 } 845 846 // Note: we support negative bitwidths (with shl) which are not defined. 847 // We do this to support (f.e.) loads off the end of a structure where 848 // only some bits are used. 849 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 850 FromVal = Builder.CreateLShr(FromVal, 851 ConstantInt::get(FromVal->getType(), ShAmt)); 852 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 853 FromVal = Builder.CreateShl(FromVal, 854 ConstantInt::get(FromVal->getType(), -ShAmt)); 855 856 // Finally, unconditionally truncate the integer to the right width. 857 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType); 858 if (LIBitWidth < NTy->getBitWidth()) 859 FromVal = 860 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 861 LIBitWidth)); 862 else if (LIBitWidth > NTy->getBitWidth()) 863 FromVal = 864 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 865 LIBitWidth)); 866 867 // If the result is an integer, this is a trunc or bitcast. 868 if (ToType->isIntegerTy()) { 869 // Should be done. 870 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 871 // Just do a bitcast, we know the sizes match up. 872 FromVal = Builder.CreateBitCast(FromVal, ToType); 873 } else { 874 // Otherwise must be a pointer. 875 FromVal = Builder.CreateIntToPtr(FromVal, ToType); 876 } 877 assert(FromVal->getType() == ToType && "Didn't convert right?"); 878 return FromVal; 879} 880 881/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 882/// or vector value "Old" at the offset specified by Offset. 883/// 884/// This happens when we are converting an "integer union" to a 885/// single integer scalar, or when we are converting a "vector union" to a 886/// vector with insert/extractelement instructions. 887/// 888/// Offset is an offset from the original alloca, in bits that need to be 889/// shifted to the right. 890/// 891/// NonConstantIdx is an index value if there was a GEP with a non-constant 892/// index value. If this is 0 then all GEPs used to find this insert address 893/// are constant. 894Value *ConvertToScalarInfo:: 895ConvertScalar_InsertValue(Value *SV, Value *Old, 896 uint64_t Offset, Value* NonConstantIdx, 897 IRBuilder<> &Builder) { 898 // Convert the stored type to the actual type, shift it left to insert 899 // then 'or' into place. 900 Type *AllocaType = Old->getType(); 901 LLVMContext &Context = Old->getContext(); 902 903 if (VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 904 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy); 905 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType()); 906 907 // Changing the whole vector with memset or with an access of a different 908 // vector type? 909 if (ValSize == VecSize) 910 return Builder.CreateBitCast(SV, AllocaType); 911 912 // Must be an element insertion. 913 Type *EltTy = VTy->getElementType(); 914 if (SV->getType() != EltTy) 915 SV = Builder.CreateBitCast(SV, EltTy); 916 uint64_t EltSize = TD.getTypeAllocSizeInBits(EltTy); 917 unsigned Elt = Offset/EltSize; 918 Value *Idx; 919 if (NonConstantIdx) { 920 if (Elt) 921 Idx = Builder.CreateAdd(NonConstantIdx, 922 Builder.getInt32(Elt), 923 "dyn.offset"); 924 else 925 Idx = NonConstantIdx; 926 } else 927 Idx = Builder.getInt32(Elt); 928 return Builder.CreateInsertElement(Old, SV, Idx); 929 } 930 931 // If SV is a first-class aggregate value, insert each value recursively. 932 if (StructType *ST = dyn_cast<StructType>(SV->getType())) { 933 assert(!NonConstantIdx && 934 "Dynamic indexing into struct types not supported"); 935 const StructLayout &Layout = *TD.getStructLayout(ST); 936 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 937 Value *Elt = Builder.CreateExtractValue(SV, i); 938 Old = ConvertScalar_InsertValue(Elt, Old, 939 Offset+Layout.getElementOffsetInBits(i), 940 0, Builder); 941 } 942 return Old; 943 } 944 945 if (ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 946 assert(!NonConstantIdx && 947 "Dynamic indexing into array types not supported"); 948 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType()); 949 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 950 Value *Elt = Builder.CreateExtractValue(SV, i); 951 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, 0, Builder); 952 } 953 return Old; 954 } 955 956 // If SV is a float, convert it to the appropriate integer type. 957 // If it is a pointer, do the same. 958 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 959 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); 960 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); 961 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); 962 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 963 SV = Builder.CreateBitCast(SV, IntegerType::get(SV->getContext(),SrcWidth)); 964 else if (SV->getType()->isPointerTy()) 965 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext())); 966 967 // Zero extend or truncate the value if needed. 968 if (SV->getType() != AllocaType) { 969 if (SV->getType()->getPrimitiveSizeInBits() < 970 AllocaType->getPrimitiveSizeInBits()) 971 SV = Builder.CreateZExt(SV, AllocaType); 972 else { 973 // Truncation may be needed if storing more than the alloca can hold 974 // (undefined behavior). 975 SV = Builder.CreateTrunc(SV, AllocaType); 976 SrcWidth = DestWidth; 977 SrcStoreWidth = DestStoreWidth; 978 } 979 } 980 981 // If this is a big-endian system and the store is narrower than the 982 // full alloca type, we need to do a shift to get the right bits. 983 int ShAmt = 0; 984 if (TD.isBigEndian()) { 985 // On big-endian machines, the lowest bit is stored at the bit offset 986 // from the pointer given by getTypeStoreSizeInBits. This matters for 987 // integers with a bitwidth that is not a multiple of 8. 988 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 989 } else { 990 ShAmt = Offset; 991 } 992 993 // Note: we support negative bitwidths (with shr) which are not defined. 994 // We do this to support (f.e.) stores off the end of a structure where 995 // only some bits in the structure are set. 996 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 997 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 998 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), ShAmt)); 999 Mask <<= ShAmt; 1000 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1001 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), -ShAmt)); 1002 Mask = Mask.lshr(-ShAmt); 1003 } 1004 1005 // Mask out the bits we are about to insert from the old value, and or 1006 // in the new bits. 1007 if (SrcWidth != DestWidth) { 1008 assert(DestWidth > SrcWidth); 1009 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 1010 SV = Builder.CreateOr(Old, SV, "ins"); 1011 } 1012 return SV; 1013} 1014 1015 1016//===----------------------------------------------------------------------===// 1017// SRoA Driver 1018//===----------------------------------------------------------------------===// 1019 1020 1021bool SROA::runOnFunction(Function &F) { 1022 TD = getAnalysisIfAvailable<TargetData>(); 1023 1024 bool Changed = performPromotion(F); 1025 1026 // FIXME: ScalarRepl currently depends on TargetData more than it 1027 // theoretically needs to. It should be refactored in order to support 1028 // target-independent IR. Until this is done, just skip the actual 1029 // scalar-replacement portion of this pass. 1030 if (!TD) return Changed; 1031 1032 while (1) { 1033 bool LocalChange = performScalarRepl(F); 1034 if (!LocalChange) break; // No need to repromote if no scalarrepl 1035 Changed = true; 1036 LocalChange = performPromotion(F); 1037 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 1038 } 1039 1040 return Changed; 1041} 1042 1043namespace { 1044class AllocaPromoter : public LoadAndStorePromoter { 1045 AllocaInst *AI; 1046 DIBuilder *DIB; 1047 SmallVector<DbgDeclareInst *, 4> DDIs; 1048 SmallVector<DbgValueInst *, 4> DVIs; 1049public: 1050 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S, 1051 DIBuilder *DB) 1052 : LoadAndStorePromoter(Insts, S), AI(0), DIB(DB) {} 1053 1054 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) { 1055 // Remember which alloca we're promoting (for isInstInList). 1056 this->AI = AI; 1057 if (MDNode *DebugNode = MDNode::getIfExists(AI->getContext(), AI)) { 1058 for (Value::use_iterator UI = DebugNode->use_begin(), 1059 E = DebugNode->use_end(); UI != E; ++UI) 1060 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI)) 1061 DDIs.push_back(DDI); 1062 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI)) 1063 DVIs.push_back(DVI); 1064 } 1065 1066 LoadAndStorePromoter::run(Insts); 1067 AI->eraseFromParent(); 1068 for (SmallVector<DbgDeclareInst *, 4>::iterator I = DDIs.begin(), 1069 E = DDIs.end(); I != E; ++I) { 1070 DbgDeclareInst *DDI = *I; 1071 DDI->eraseFromParent(); 1072 } 1073 for (SmallVector<DbgValueInst *, 4>::iterator I = DVIs.begin(), 1074 E = DVIs.end(); I != E; ++I) { 1075 DbgValueInst *DVI = *I; 1076 DVI->eraseFromParent(); 1077 } 1078 } 1079 1080 virtual bool isInstInList(Instruction *I, 1081 const SmallVectorImpl<Instruction*> &Insts) const { 1082 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 1083 return LI->getOperand(0) == AI; 1084 return cast<StoreInst>(I)->getPointerOperand() == AI; 1085 } 1086 1087 virtual void updateDebugInfo(Instruction *Inst) const { 1088 for (SmallVector<DbgDeclareInst *, 4>::const_iterator I = DDIs.begin(), 1089 E = DDIs.end(); I != E; ++I) { 1090 DbgDeclareInst *DDI = *I; 1091 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) 1092 ConvertDebugDeclareToDebugValue(DDI, SI, *DIB); 1093 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) 1094 ConvertDebugDeclareToDebugValue(DDI, LI, *DIB); 1095 } 1096 for (SmallVector<DbgValueInst *, 4>::const_iterator I = DVIs.begin(), 1097 E = DVIs.end(); I != E; ++I) { 1098 DbgValueInst *DVI = *I; 1099 Value *Arg = NULL; 1100 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 1101 // If an argument is zero extended then use argument directly. The ZExt 1102 // may be zapped by an optimization pass in future. 1103 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0))) 1104 Arg = dyn_cast<Argument>(ZExt->getOperand(0)); 1105 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0))) 1106 Arg = dyn_cast<Argument>(SExt->getOperand(0)); 1107 if (!Arg) 1108 Arg = SI->getOperand(0); 1109 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 1110 Arg = LI->getOperand(0); 1111 } else { 1112 continue; 1113 } 1114 Instruction *DbgVal = 1115 DIB->insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()), 1116 Inst); 1117 DbgVal->setDebugLoc(DVI->getDebugLoc()); 1118 } 1119 } 1120}; 1121} // end anon namespace 1122 1123/// isSafeSelectToSpeculate - Select instructions that use an alloca and are 1124/// subsequently loaded can be rewritten to load both input pointers and then 1125/// select between the result, allowing the load of the alloca to be promoted. 1126/// From this: 1127/// %P2 = select i1 %cond, i32* %Alloca, i32* %Other 1128/// %V = load i32* %P2 1129/// to: 1130/// %V1 = load i32* %Alloca -> will be mem2reg'd 1131/// %V2 = load i32* %Other 1132/// %V = select i1 %cond, i32 %V1, i32 %V2 1133/// 1134/// We can do this to a select if its only uses are loads and if the operand to 1135/// the select can be loaded unconditionally. 1136static bool isSafeSelectToSpeculate(SelectInst *SI, const TargetData *TD) { 1137 bool TDerefable = SI->getTrueValue()->isDereferenceablePointer(); 1138 bool FDerefable = SI->getFalseValue()->isDereferenceablePointer(); 1139 1140 for (Value::use_iterator UI = SI->use_begin(), UE = SI->use_end(); 1141 UI != UE; ++UI) { 1142 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1143 if (LI == 0 || !LI->isSimple()) return false; 1144 1145 // Both operands to the select need to be dereferencable, either absolutely 1146 // (e.g. allocas) or at this point because we can see other accesses to it. 1147 if (!TDerefable && !isSafeToLoadUnconditionally(SI->getTrueValue(), LI, 1148 LI->getAlignment(), TD)) 1149 return false; 1150 if (!FDerefable && !isSafeToLoadUnconditionally(SI->getFalseValue(), LI, 1151 LI->getAlignment(), TD)) 1152 return false; 1153 } 1154 1155 return true; 1156} 1157 1158/// isSafePHIToSpeculate - PHI instructions that use an alloca and are 1159/// subsequently loaded can be rewritten to load both input pointers in the pred 1160/// blocks and then PHI the results, allowing the load of the alloca to be 1161/// promoted. 1162/// From this: 1163/// %P2 = phi [i32* %Alloca, i32* %Other] 1164/// %V = load i32* %P2 1165/// to: 1166/// %V1 = load i32* %Alloca -> will be mem2reg'd 1167/// ... 1168/// %V2 = load i32* %Other 1169/// ... 1170/// %V = phi [i32 %V1, i32 %V2] 1171/// 1172/// We can do this to a select if its only uses are loads and if the operand to 1173/// the select can be loaded unconditionally. 1174static bool isSafePHIToSpeculate(PHINode *PN, const TargetData *TD) { 1175 // For now, we can only do this promotion if the load is in the same block as 1176 // the PHI, and if there are no stores between the phi and load. 1177 // TODO: Allow recursive phi users. 1178 // TODO: Allow stores. 1179 BasicBlock *BB = PN->getParent(); 1180 unsigned MaxAlign = 0; 1181 for (Value::use_iterator UI = PN->use_begin(), UE = PN->use_end(); 1182 UI != UE; ++UI) { 1183 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1184 if (LI == 0 || !LI->isSimple()) return false; 1185 1186 // For now we only allow loads in the same block as the PHI. This is a 1187 // common case that happens when instcombine merges two loads through a PHI. 1188 if (LI->getParent() != BB) return false; 1189 1190 // Ensure that there are no instructions between the PHI and the load that 1191 // could store. 1192 for (BasicBlock::iterator BBI = PN; &*BBI != LI; ++BBI) 1193 if (BBI->mayWriteToMemory()) 1194 return false; 1195 1196 MaxAlign = std::max(MaxAlign, LI->getAlignment()); 1197 } 1198 1199 // Okay, we know that we have one or more loads in the same block as the PHI. 1200 // We can transform this if it is safe to push the loads into the predecessor 1201 // blocks. The only thing to watch out for is that we can't put a possibly 1202 // trapping load in the predecessor if it is a critical edge. 1203 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1204 BasicBlock *Pred = PN->getIncomingBlock(i); 1205 Value *InVal = PN->getIncomingValue(i); 1206 1207 // If the terminator of the predecessor has side-effects (an invoke), 1208 // there is no safe place to put a load in the predecessor. 1209 if (Pred->getTerminator()->mayHaveSideEffects()) 1210 return false; 1211 1212 // If the value is produced by the terminator of the predecessor 1213 // (an invoke), there is no valid place to put a load in the predecessor. 1214 if (Pred->getTerminator() == InVal) 1215 return false; 1216 1217 // If the predecessor has a single successor, then the edge isn't critical. 1218 if (Pred->getTerminator()->getNumSuccessors() == 1) 1219 continue; 1220 1221 // If this pointer is always safe to load, or if we can prove that there is 1222 // already a load in the block, then we can move the load to the pred block. 1223 if (InVal->isDereferenceablePointer() || 1224 isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign, TD)) 1225 continue; 1226 1227 return false; 1228 } 1229 1230 return true; 1231} 1232 1233 1234/// tryToMakeAllocaBePromotable - This returns true if the alloca only has 1235/// direct (non-volatile) loads and stores to it. If the alloca is close but 1236/// not quite there, this will transform the code to allow promotion. As such, 1237/// it is a non-pure predicate. 1238static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const TargetData *TD) { 1239 SetVector<Instruction*, SmallVector<Instruction*, 4>, 1240 SmallPtrSet<Instruction*, 4> > InstsToRewrite; 1241 1242 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end(); 1243 UI != UE; ++UI) { 1244 User *U = *UI; 1245 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 1246 if (!LI->isSimple()) 1247 return false; 1248 continue; 1249 } 1250 1251 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 1252 if (SI->getOperand(0) == AI || !SI->isSimple()) 1253 return false; // Don't allow a store OF the AI, only INTO the AI. 1254 continue; 1255 } 1256 1257 if (SelectInst *SI = dyn_cast<SelectInst>(U)) { 1258 // If the condition being selected on is a constant, fold the select, yes 1259 // this does (rarely) happen early on. 1260 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition())) { 1261 Value *Result = SI->getOperand(1+CI->isZero()); 1262 SI->replaceAllUsesWith(Result); 1263 SI->eraseFromParent(); 1264 1265 // This is very rare and we just scrambled the use list of AI, start 1266 // over completely. 1267 return tryToMakeAllocaBePromotable(AI, TD); 1268 } 1269 1270 // If it is safe to turn "load (select c, AI, ptr)" into a select of two 1271 // loads, then we can transform this by rewriting the select. 1272 if (!isSafeSelectToSpeculate(SI, TD)) 1273 return false; 1274 1275 InstsToRewrite.insert(SI); 1276 continue; 1277 } 1278 1279 if (PHINode *PN = dyn_cast<PHINode>(U)) { 1280 if (PN->use_empty()) { // Dead PHIs can be stripped. 1281 InstsToRewrite.insert(PN); 1282 continue; 1283 } 1284 1285 // If it is safe to turn "load (phi [AI, ptr, ...])" into a PHI of loads 1286 // in the pred blocks, then we can transform this by rewriting the PHI. 1287 if (!isSafePHIToSpeculate(PN, TD)) 1288 return false; 1289 1290 InstsToRewrite.insert(PN); 1291 continue; 1292 } 1293 1294 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 1295 if (onlyUsedByLifetimeMarkers(BCI)) { 1296 InstsToRewrite.insert(BCI); 1297 continue; 1298 } 1299 } 1300 1301 return false; 1302 } 1303 1304 // If there are no instructions to rewrite, then all uses are load/stores and 1305 // we're done! 1306 if (InstsToRewrite.empty()) 1307 return true; 1308 1309 // If we have instructions that need to be rewritten for this to be promotable 1310 // take care of it now. 1311 for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) { 1312 if (BitCastInst *BCI = dyn_cast<BitCastInst>(InstsToRewrite[i])) { 1313 // This could only be a bitcast used by nothing but lifetime intrinsics. 1314 for (BitCastInst::use_iterator I = BCI->use_begin(), E = BCI->use_end(); 1315 I != E;) { 1316 Use &U = I.getUse(); 1317 ++I; 1318 cast<Instruction>(U.getUser())->eraseFromParent(); 1319 } 1320 BCI->eraseFromParent(); 1321 continue; 1322 } 1323 1324 if (SelectInst *SI = dyn_cast<SelectInst>(InstsToRewrite[i])) { 1325 // Selects in InstsToRewrite only have load uses. Rewrite each as two 1326 // loads with a new select. 1327 while (!SI->use_empty()) { 1328 LoadInst *LI = cast<LoadInst>(SI->use_back()); 1329 1330 IRBuilder<> Builder(LI); 1331 LoadInst *TrueLoad = 1332 Builder.CreateLoad(SI->getTrueValue(), LI->getName()+".t"); 1333 LoadInst *FalseLoad = 1334 Builder.CreateLoad(SI->getFalseValue(), LI->getName()+".f"); 1335 1336 // Transfer alignment and TBAA info if present. 1337 TrueLoad->setAlignment(LI->getAlignment()); 1338 FalseLoad->setAlignment(LI->getAlignment()); 1339 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) { 1340 TrueLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1341 FalseLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1342 } 1343 1344 Value *V = Builder.CreateSelect(SI->getCondition(), TrueLoad, FalseLoad); 1345 V->takeName(LI); 1346 LI->replaceAllUsesWith(V); 1347 LI->eraseFromParent(); 1348 } 1349 1350 // Now that all the loads are gone, the select is gone too. 1351 SI->eraseFromParent(); 1352 continue; 1353 } 1354 1355 // Otherwise, we have a PHI node which allows us to push the loads into the 1356 // predecessors. 1357 PHINode *PN = cast<PHINode>(InstsToRewrite[i]); 1358 if (PN->use_empty()) { 1359 PN->eraseFromParent(); 1360 continue; 1361 } 1362 1363 Type *LoadTy = cast<PointerType>(PN->getType())->getElementType(); 1364 PHINode *NewPN = PHINode::Create(LoadTy, PN->getNumIncomingValues(), 1365 PN->getName()+".ld", PN); 1366 1367 // Get the TBAA tag and alignment to use from one of the loads. It doesn't 1368 // matter which one we get and if any differ, it doesn't matter. 1369 LoadInst *SomeLoad = cast<LoadInst>(PN->use_back()); 1370 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa); 1371 unsigned Align = SomeLoad->getAlignment(); 1372 1373 // Rewrite all loads of the PN to use the new PHI. 1374 while (!PN->use_empty()) { 1375 LoadInst *LI = cast<LoadInst>(PN->use_back()); 1376 LI->replaceAllUsesWith(NewPN); 1377 LI->eraseFromParent(); 1378 } 1379 1380 // Inject loads into all of the pred blocks. Keep track of which blocks we 1381 // insert them into in case we have multiple edges from the same block. 1382 DenseMap<BasicBlock*, LoadInst*> InsertedLoads; 1383 1384 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1385 BasicBlock *Pred = PN->getIncomingBlock(i); 1386 LoadInst *&Load = InsertedLoads[Pred]; 1387 if (Load == 0) { 1388 Load = new LoadInst(PN->getIncomingValue(i), 1389 PN->getName() + "." + Pred->getName(), 1390 Pred->getTerminator()); 1391 Load->setAlignment(Align); 1392 if (TBAATag) Load->setMetadata(LLVMContext::MD_tbaa, TBAATag); 1393 } 1394 1395 NewPN->addIncoming(Load, Pred); 1396 } 1397 1398 PN->eraseFromParent(); 1399 } 1400 1401 ++NumAdjusted; 1402 return true; 1403} 1404 1405bool SROA::performPromotion(Function &F) { 1406 std::vector<AllocaInst*> Allocas; 1407 DominatorTree *DT = 0; 1408 if (HasDomTree) 1409 DT = &getAnalysis<DominatorTree>(); 1410 1411 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 1412 DIBuilder DIB(*F.getParent()); 1413 bool Changed = false; 1414 SmallVector<Instruction*, 64> Insts; 1415 while (1) { 1416 Allocas.clear(); 1417 1418 // Find allocas that are safe to promote, by looking at all instructions in 1419 // the entry node 1420 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 1421 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 1422 if (tryToMakeAllocaBePromotable(AI, TD)) 1423 Allocas.push_back(AI); 1424 1425 if (Allocas.empty()) break; 1426 1427 if (HasDomTree) 1428 PromoteMemToReg(Allocas, *DT); 1429 else { 1430 SSAUpdater SSA; 1431 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) { 1432 AllocaInst *AI = Allocas[i]; 1433 1434 // Build list of instructions to promote. 1435 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 1436 UI != E; ++UI) 1437 Insts.push_back(cast<Instruction>(*UI)); 1438 AllocaPromoter(Insts, SSA, &DIB).run(AI, Insts); 1439 Insts.clear(); 1440 } 1441 } 1442 NumPromoted += Allocas.size(); 1443 Changed = true; 1444 } 1445 1446 return Changed; 1447} 1448 1449 1450/// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 1451/// SROA. It must be a struct or array type with a small number of elements. 1452bool SROA::ShouldAttemptScalarRepl(AllocaInst *AI) { 1453 Type *T = AI->getAllocatedType(); 1454 // Do not promote any struct that has too many members. 1455 if (StructType *ST = dyn_cast<StructType>(T)) 1456 return ST->getNumElements() <= StructMemberThreshold; 1457 // Do not promote any array that has too many elements. 1458 if (ArrayType *AT = dyn_cast<ArrayType>(T)) 1459 return AT->getNumElements() <= ArrayElementThreshold; 1460 return false; 1461} 1462 1463/// getPointeeAlignment - Compute the minimum alignment of the value pointed 1464/// to by the given pointer. 1465static unsigned getPointeeAlignment(Value *V, const TargetData &TD) { 1466 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1467 if (CE->getOpcode() == Instruction::BitCast || 1468 (CE->getOpcode() == Instruction::GetElementPtr && 1469 cast<GEPOperator>(CE)->hasAllZeroIndices())) 1470 return getPointeeAlignment(CE->getOperand(0), TD); 1471 1472 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1473 if (!GV->isDeclaration()) 1474 return TD.getPreferredAlignment(GV); 1475 1476 if (PointerType *PT = dyn_cast<PointerType>(V->getType())) 1477 return TD.getABITypeAlignment(PT->getElementType()); 1478 1479 return 0; 1480} 1481 1482 1483// performScalarRepl - This algorithm is a simple worklist driven algorithm, 1484// which runs on all of the alloca instructions in the function, removing them 1485// if they are only used by getelementptr instructions. 1486// 1487bool SROA::performScalarRepl(Function &F) { 1488 std::vector<AllocaInst*> WorkList; 1489 1490 // Scan the entry basic block, adding allocas to the worklist. 1491 BasicBlock &BB = F.getEntryBlock(); 1492 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 1493 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 1494 WorkList.push_back(A); 1495 1496 // Process the worklist 1497 bool Changed = false; 1498 while (!WorkList.empty()) { 1499 AllocaInst *AI = WorkList.back(); 1500 WorkList.pop_back(); 1501 1502 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 1503 // with unused elements. 1504 if (AI->use_empty()) { 1505 AI->eraseFromParent(); 1506 Changed = true; 1507 continue; 1508 } 1509 1510 // If this alloca is impossible for us to promote, reject it early. 1511 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 1512 continue; 1513 1514 // Check to see if this allocation is only modified by a memcpy/memmove from 1515 // a constant global whose alignment is equal to or exceeds that of the 1516 // allocation. If this is the case, we can change all users to use 1517 // the constant global instead. This is commonly produced by the CFE by 1518 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 1519 // is only subsequently read. 1520 SmallVector<Instruction *, 4> ToDelete; 1521 if (MemTransferInst *Copy = isOnlyCopiedFromConstantGlobal(AI, ToDelete)) { 1522 if (AI->getAlignment() <= getPointeeAlignment(Copy->getSource(), *TD)) { 1523 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); 1524 DEBUG(dbgs() << " memcpy = " << *Copy << '\n'); 1525 for (unsigned i = 0, e = ToDelete.size(); i != e; ++i) 1526 ToDelete[i]->eraseFromParent(); 1527 Constant *TheSrc = cast<Constant>(Copy->getSource()); 1528 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 1529 Copy->eraseFromParent(); // Don't mutate the global. 1530 AI->eraseFromParent(); 1531 ++NumGlobals; 1532 Changed = true; 1533 continue; 1534 } 1535 } 1536 1537 // Check to see if we can perform the core SROA transformation. We cannot 1538 // transform the allocation instruction if it is an array allocation 1539 // (allocations OF arrays are ok though), and an allocation of a scalar 1540 // value cannot be decomposed at all. 1541 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 1542 1543 // Do not promote [0 x %struct]. 1544 if (AllocaSize == 0) continue; 1545 1546 // Do not promote any struct whose size is too big. 1547 if (AllocaSize > SRThreshold) continue; 1548 1549 // If the alloca looks like a good candidate for scalar replacement, and if 1550 // all its users can be transformed, then split up the aggregate into its 1551 // separate elements. 1552 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 1553 DoScalarReplacement(AI, WorkList); 1554 Changed = true; 1555 continue; 1556 } 1557 1558 // If we can turn this aggregate value (potentially with casts) into a 1559 // simple scalar value that can be mem2reg'd into a register value. 1560 // IsNotTrivial tracks whether this is something that mem2reg could have 1561 // promoted itself. If so, we don't want to transform it needlessly. Note 1562 // that we can't just check based on the type: the alloca may be of an i32 1563 // but that has pointer arithmetic to set byte 3 of it or something. 1564 if (AllocaInst *NewAI = ConvertToScalarInfo( 1565 (unsigned)AllocaSize, *TD, ScalarLoadThreshold).TryConvert(AI)) { 1566 NewAI->takeName(AI); 1567 AI->eraseFromParent(); 1568 ++NumConverted; 1569 Changed = true; 1570 continue; 1571 } 1572 1573 // Otherwise, couldn't process this alloca. 1574 } 1575 1576 return Changed; 1577} 1578 1579/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 1580/// predicate, do SROA now. 1581void SROA::DoScalarReplacement(AllocaInst *AI, 1582 std::vector<AllocaInst*> &WorkList) { 1583 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 1584 SmallVector<AllocaInst*, 32> ElementAllocas; 1585 if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 1586 ElementAllocas.reserve(ST->getNumContainedTypes()); 1587 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 1588 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 1589 AI->getAlignment(), 1590 AI->getName() + "." + Twine(i), AI); 1591 ElementAllocas.push_back(NA); 1592 WorkList.push_back(NA); // Add to worklist for recursive processing 1593 } 1594 } else { 1595 ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 1596 ElementAllocas.reserve(AT->getNumElements()); 1597 Type *ElTy = AT->getElementType(); 1598 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1599 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 1600 AI->getName() + "." + Twine(i), AI); 1601 ElementAllocas.push_back(NA); 1602 WorkList.push_back(NA); // Add to worklist for recursive processing 1603 } 1604 } 1605 1606 // Now that we have created the new alloca instructions, rewrite all the 1607 // uses of the old alloca. 1608 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 1609 1610 // Now erase any instructions that were made dead while rewriting the alloca. 1611 DeleteDeadInstructions(); 1612 AI->eraseFromParent(); 1613 1614 ++NumReplaced; 1615} 1616 1617/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 1618/// recursively including all their operands that become trivially dead. 1619void SROA::DeleteDeadInstructions() { 1620 while (!DeadInsts.empty()) { 1621 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 1622 1623 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 1624 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 1625 // Zero out the operand and see if it becomes trivially dead. 1626 // (But, don't add allocas to the dead instruction list -- they are 1627 // already on the worklist and will be deleted separately.) 1628 *OI = 0; 1629 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 1630 DeadInsts.push_back(U); 1631 } 1632 1633 I->eraseFromParent(); 1634 } 1635} 1636 1637/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 1638/// performing scalar replacement of alloca AI. The results are flagged in 1639/// the Info parameter. Offset indicates the position within AI that is 1640/// referenced by this instruction. 1641void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset, 1642 AllocaInfo &Info) { 1643 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1644 Instruction *User = cast<Instruction>(*UI); 1645 1646 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1647 isSafeForScalarRepl(BC, Offset, Info); 1648 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1649 uint64_t GEPOffset = Offset; 1650 isSafeGEP(GEPI, GEPOffset, Info); 1651 if (!Info.isUnsafe) 1652 isSafeForScalarRepl(GEPI, GEPOffset, Info); 1653 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1654 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1655 if (Length == 0) 1656 return MarkUnsafe(Info, User); 1657 if (Length->isNegative()) 1658 return MarkUnsafe(Info, User); 1659 1660 isSafeMemAccess(Offset, Length->getZExtValue(), 0, 1661 UI.getOperandNo() == 0, Info, MI, 1662 true /*AllowWholeAccess*/); 1663 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1664 if (!LI->isSimple()) 1665 return MarkUnsafe(Info, User); 1666 Type *LIType = LI->getType(); 1667 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1668 LIType, false, Info, LI, true /*AllowWholeAccess*/); 1669 Info.hasALoadOrStore = true; 1670 1671 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1672 // Store is ok if storing INTO the pointer, not storing the pointer 1673 if (!SI->isSimple() || SI->getOperand(0) == I) 1674 return MarkUnsafe(Info, User); 1675 1676 Type *SIType = SI->getOperand(0)->getType(); 1677 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1678 SIType, true, Info, SI, true /*AllowWholeAccess*/); 1679 Info.hasALoadOrStore = true; 1680 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 1681 if (II->getIntrinsicID() != Intrinsic::lifetime_start && 1682 II->getIntrinsicID() != Intrinsic::lifetime_end) 1683 return MarkUnsafe(Info, User); 1684 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1685 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1686 } else { 1687 return MarkUnsafe(Info, User); 1688 } 1689 if (Info.isUnsafe) return; 1690 } 1691} 1692 1693 1694/// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer 1695/// derived from the alloca, we can often still split the alloca into elements. 1696/// This is useful if we have a large alloca where one element is phi'd 1697/// together somewhere: we can SRoA and promote all the other elements even if 1698/// we end up not being able to promote this one. 1699/// 1700/// All we require is that the uses of the PHI do not index into other parts of 1701/// the alloca. The most important use case for this is single load and stores 1702/// that are PHI'd together, which can happen due to code sinking. 1703void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset, 1704 AllocaInfo &Info) { 1705 // If we've already checked this PHI, don't do it again. 1706 if (PHINode *PN = dyn_cast<PHINode>(I)) 1707 if (!Info.CheckedPHIs.insert(PN)) 1708 return; 1709 1710 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 1711 Instruction *User = cast<Instruction>(*UI); 1712 1713 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1714 isSafePHISelectUseForScalarRepl(BC, Offset, Info); 1715 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1716 // Only allow "bitcast" GEPs for simplicity. We could generalize this, 1717 // but would have to prove that we're staying inside of an element being 1718 // promoted. 1719 if (!GEPI->hasAllZeroIndices()) 1720 return MarkUnsafe(Info, User); 1721 isSafePHISelectUseForScalarRepl(GEPI, Offset, Info); 1722 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1723 if (!LI->isSimple()) 1724 return MarkUnsafe(Info, User); 1725 Type *LIType = LI->getType(); 1726 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType), 1727 LIType, false, Info, LI, false /*AllowWholeAccess*/); 1728 Info.hasALoadOrStore = true; 1729 1730 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1731 // Store is ok if storing INTO the pointer, not storing the pointer 1732 if (!SI->isSimple() || SI->getOperand(0) == I) 1733 return MarkUnsafe(Info, User); 1734 1735 Type *SIType = SI->getOperand(0)->getType(); 1736 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType), 1737 SIType, true, Info, SI, false /*AllowWholeAccess*/); 1738 Info.hasALoadOrStore = true; 1739 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1740 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1741 } else { 1742 return MarkUnsafe(Info, User); 1743 } 1744 if (Info.isUnsafe) return; 1745 } 1746} 1747 1748/// isSafeGEP - Check if a GEP instruction can be handled for scalar 1749/// replacement. It is safe when all the indices are constant, in-bounds 1750/// references, and when the resulting offset corresponds to an element within 1751/// the alloca type. The results are flagged in the Info parameter. Upon 1752/// return, Offset is adjusted as specified by the GEP indices. 1753void SROA::isSafeGEP(GetElementPtrInst *GEPI, 1754 uint64_t &Offset, AllocaInfo &Info) { 1755 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 1756 if (GEPIt == E) 1757 return; 1758 bool NonConstant = false; 1759 unsigned NonConstantIdxSize = 0; 1760 1761 // Walk through the GEP type indices, checking the types that this indexes 1762 // into. 1763 for (; GEPIt != E; ++GEPIt) { 1764 // Ignore struct elements, no extra checking needed for these. 1765 if ((*GEPIt)->isStructTy()) 1766 continue; 1767 1768 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 1769 if (!IdxVal) { 1770 // Non constant GEPs are only a problem on arrays, structs, and pointers 1771 // Vectors can be dynamically indexed. 1772 // FIXME: Add support for dynamic indexing on arrays. This should be 1773 // ok on any subarrays of the alloca array, eg, a[0][i] is ok, but a[i][0] 1774 // isn't. 1775 if (!(*GEPIt)->isVectorTy()) 1776 return MarkUnsafe(Info, GEPI); 1777 NonConstant = true; 1778 NonConstantIdxSize = TD->getTypeAllocSize(*GEPIt); 1779 } 1780 } 1781 1782 // Compute the offset due to this GEP and check if the alloca has a 1783 // component element at that offset. 1784 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1785 // If this GEP is non constant then the last operand must have been a 1786 // dynamic index into a vector. Pop this now as it has no impact on the 1787 // constant part of the offset. 1788 if (NonConstant) 1789 Indices.pop_back(); 1790 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices); 1791 if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 1792 NonConstantIdxSize)) 1793 MarkUnsafe(Info, GEPI); 1794} 1795 1796/// isHomogeneousAggregate - Check if type T is a struct or array containing 1797/// elements of the same type (which is always true for arrays). If so, 1798/// return true with NumElts and EltTy set to the number of elements and the 1799/// element type, respectively. 1800static bool isHomogeneousAggregate(Type *T, unsigned &NumElts, 1801 Type *&EltTy) { 1802 if (ArrayType *AT = dyn_cast<ArrayType>(T)) { 1803 NumElts = AT->getNumElements(); 1804 EltTy = (NumElts == 0 ? 0 : AT->getElementType()); 1805 return true; 1806 } 1807 if (StructType *ST = dyn_cast<StructType>(T)) { 1808 NumElts = ST->getNumContainedTypes(); 1809 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0)); 1810 for (unsigned n = 1; n < NumElts; ++n) { 1811 if (ST->getContainedType(n) != EltTy) 1812 return false; 1813 } 1814 return true; 1815 } 1816 return false; 1817} 1818 1819/// isCompatibleAggregate - Check if T1 and T2 are either the same type or are 1820/// "homogeneous" aggregates with the same element type and number of elements. 1821static bool isCompatibleAggregate(Type *T1, Type *T2) { 1822 if (T1 == T2) 1823 return true; 1824 1825 unsigned NumElts1, NumElts2; 1826 Type *EltTy1, *EltTy2; 1827 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) && 1828 isHomogeneousAggregate(T2, NumElts2, EltTy2) && 1829 NumElts1 == NumElts2 && 1830 EltTy1 == EltTy2) 1831 return true; 1832 1833 return false; 1834} 1835 1836/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1837/// alloca or has an offset and size that corresponds to a component element 1838/// within it. The offset checked here may have been formed from a GEP with a 1839/// pointer bitcasted to a different type. 1840/// 1841/// If AllowWholeAccess is true, then this allows uses of the entire alloca as a 1842/// unit. If false, it only allows accesses known to be in a single element. 1843void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 1844 Type *MemOpType, bool isStore, 1845 AllocaInfo &Info, Instruction *TheAccess, 1846 bool AllowWholeAccess) { 1847 // Check if this is a load/store of the entire alloca. 1848 if (Offset == 0 && AllowWholeAccess && 1849 MemSize == TD->getTypeAllocSize(Info.AI->getAllocatedType())) { 1850 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer 1851 // loads/stores (which are essentially the same as the MemIntrinsics with 1852 // regard to copying padding between elements). But, if an alloca is 1853 // flagged as both a source and destination of such operations, we'll need 1854 // to check later for padding between elements. 1855 if (!MemOpType || MemOpType->isIntegerTy()) { 1856 if (isStore) 1857 Info.isMemCpyDst = true; 1858 else 1859 Info.isMemCpySrc = true; 1860 return; 1861 } 1862 // This is also safe for references using a type that is compatible with 1863 // the type of the alloca, so that loads/stores can be rewritten using 1864 // insertvalue/extractvalue. 1865 if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) { 1866 Info.hasSubelementAccess = true; 1867 return; 1868 } 1869 } 1870 // Check if the offset/size correspond to a component within the alloca type. 1871 Type *T = Info.AI->getAllocatedType(); 1872 if (TypeHasComponent(T, Offset, MemSize)) { 1873 Info.hasSubelementAccess = true; 1874 return; 1875 } 1876 1877 return MarkUnsafe(Info, TheAccess); 1878} 1879 1880/// TypeHasComponent - Return true if T has a component type with the 1881/// specified offset and size. If Size is zero, do not check the size. 1882bool SROA::TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size) { 1883 Type *EltTy; 1884 uint64_t EltSize; 1885 if (StructType *ST = dyn_cast<StructType>(T)) { 1886 const StructLayout *Layout = TD->getStructLayout(ST); 1887 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1888 EltTy = ST->getContainedType(EltIdx); 1889 EltSize = TD->getTypeAllocSize(EltTy); 1890 Offset -= Layout->getElementOffset(EltIdx); 1891 } else if (ArrayType *AT = dyn_cast<ArrayType>(T)) { 1892 EltTy = AT->getElementType(); 1893 EltSize = TD->getTypeAllocSize(EltTy); 1894 if (Offset >= AT->getNumElements() * EltSize) 1895 return false; 1896 Offset %= EltSize; 1897 } else if (VectorType *VT = dyn_cast<VectorType>(T)) { 1898 EltTy = VT->getElementType(); 1899 EltSize = TD->getTypeAllocSize(EltTy); 1900 if (Offset >= VT->getNumElements() * EltSize) 1901 return false; 1902 Offset %= EltSize; 1903 } else { 1904 return false; 1905 } 1906 if (Offset == 0 && (Size == 0 || EltSize == Size)) 1907 return true; 1908 // Check if the component spans multiple elements. 1909 if (Offset + Size > EltSize) 1910 return false; 1911 return TypeHasComponent(EltTy, Offset, Size); 1912} 1913 1914/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1915/// the instruction I, which references it, to use the separate elements. 1916/// Offset indicates the position within AI that is referenced by this 1917/// instruction. 1918void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1919 SmallVector<AllocaInst*, 32> &NewElts) { 1920 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) { 1921 Use &TheUse = UI.getUse(); 1922 Instruction *User = cast<Instruction>(*UI++); 1923 1924 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1925 RewriteBitCast(BC, AI, Offset, NewElts); 1926 continue; 1927 } 1928 1929 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1930 RewriteGEP(GEPI, AI, Offset, NewElts); 1931 continue; 1932 } 1933 1934 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1935 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1936 uint64_t MemSize = Length->getZExtValue(); 1937 if (Offset == 0 && 1938 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 1939 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1940 // Otherwise the intrinsic can only touch a single element and the 1941 // address operand will be updated, so nothing else needs to be done. 1942 continue; 1943 } 1944 1945 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 1946 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 1947 II->getIntrinsicID() == Intrinsic::lifetime_end) { 1948 RewriteLifetimeIntrinsic(II, AI, Offset, NewElts); 1949 } 1950 continue; 1951 } 1952 1953 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1954 Type *LIType = LI->getType(); 1955 1956 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) { 1957 // Replace: 1958 // %res = load { i32, i32 }* %alloc 1959 // with: 1960 // %load.0 = load i32* %alloc.0 1961 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1962 // %load.1 = load i32* %alloc.1 1963 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1964 // (Also works for arrays instead of structs) 1965 Value *Insert = UndefValue::get(LIType); 1966 IRBuilder<> Builder(LI); 1967 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1968 Value *Load = Builder.CreateLoad(NewElts[i], "load"); 1969 Insert = Builder.CreateInsertValue(Insert, Load, i, "insert"); 1970 } 1971 LI->replaceAllUsesWith(Insert); 1972 DeadInsts.push_back(LI); 1973 } else if (LIType->isIntegerTy() && 1974 TD->getTypeAllocSize(LIType) == 1975 TD->getTypeAllocSize(AI->getAllocatedType())) { 1976 // If this is a load of the entire alloca to an integer, rewrite it. 1977 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1978 } 1979 continue; 1980 } 1981 1982 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1983 Value *Val = SI->getOperand(0); 1984 Type *SIType = Val->getType(); 1985 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) { 1986 // Replace: 1987 // store { i32, i32 } %val, { i32, i32 }* %alloc 1988 // with: 1989 // %val.0 = extractvalue { i32, i32 } %val, 0 1990 // store i32 %val.0, i32* %alloc.0 1991 // %val.1 = extractvalue { i32, i32 } %val, 1 1992 // store i32 %val.1, i32* %alloc.1 1993 // (Also works for arrays instead of structs) 1994 IRBuilder<> Builder(SI); 1995 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1996 Value *Extract = Builder.CreateExtractValue(Val, i, Val->getName()); 1997 Builder.CreateStore(Extract, NewElts[i]); 1998 } 1999 DeadInsts.push_back(SI); 2000 } else if (SIType->isIntegerTy() && 2001 TD->getTypeAllocSize(SIType) == 2002 TD->getTypeAllocSize(AI->getAllocatedType())) { 2003 // If this is a store of the entire alloca from an integer, rewrite it. 2004 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 2005 } 2006 continue; 2007 } 2008 2009 if (isa<SelectInst>(User) || isa<PHINode>(User)) { 2010 // If we have a PHI user of the alloca itself (as opposed to a GEP or 2011 // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to 2012 // the new pointer. 2013 if (!isa<AllocaInst>(I)) continue; 2014 2015 assert(Offset == 0 && NewElts[0] && 2016 "Direct alloca use should have a zero offset"); 2017 2018 // If we have a use of the alloca, we know the derived uses will be 2019 // utilizing just the first element of the scalarized result. Insert a 2020 // bitcast of the first alloca before the user as required. 2021 AllocaInst *NewAI = NewElts[0]; 2022 BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI); 2023 NewAI->moveBefore(BCI); 2024 TheUse = BCI; 2025 continue; 2026 } 2027 } 2028} 2029 2030/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 2031/// and recursively continue updating all of its uses. 2032void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 2033 SmallVector<AllocaInst*, 32> &NewElts) { 2034 RewriteForScalarRepl(BC, AI, Offset, NewElts); 2035 if (BC->getOperand(0) != AI) 2036 return; 2037 2038 // The bitcast references the original alloca. Replace its uses with 2039 // references to the alloca containing offset zero (which is normally at 2040 // index zero, but might not be in cases involving structs with elements 2041 // of size zero). 2042 Type *T = AI->getAllocatedType(); 2043 uint64_t EltOffset = 0; 2044 Type *IdxTy; 2045 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 2046 Instruction *Val = NewElts[Idx]; 2047 if (Val->getType() != BC->getDestTy()) { 2048 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 2049 Val->takeName(BC); 2050 } 2051 BC->replaceAllUsesWith(Val); 2052 DeadInsts.push_back(BC); 2053} 2054 2055/// FindElementAndOffset - Return the index of the element containing Offset 2056/// within the specified type, which must be either a struct or an array. 2057/// Sets T to the type of the element and Offset to the offset within that 2058/// element. IdxTy is set to the type of the index result to be used in a 2059/// GEP instruction. 2060uint64_t SROA::FindElementAndOffset(Type *&T, uint64_t &Offset, 2061 Type *&IdxTy) { 2062 uint64_t Idx = 0; 2063 if (StructType *ST = dyn_cast<StructType>(T)) { 2064 const StructLayout *Layout = TD->getStructLayout(ST); 2065 Idx = Layout->getElementContainingOffset(Offset); 2066 T = ST->getContainedType(Idx); 2067 Offset -= Layout->getElementOffset(Idx); 2068 IdxTy = Type::getInt32Ty(T->getContext()); 2069 return Idx; 2070 } else if (ArrayType *AT = dyn_cast<ArrayType>(T)) { 2071 T = AT->getElementType(); 2072 uint64_t EltSize = TD->getTypeAllocSize(T); 2073 Idx = Offset / EltSize; 2074 Offset -= Idx * EltSize; 2075 IdxTy = Type::getInt64Ty(T->getContext()); 2076 return Idx; 2077 } 2078 VectorType *VT = cast<VectorType>(T); 2079 T = VT->getElementType(); 2080 uint64_t EltSize = TD->getTypeAllocSize(T); 2081 Idx = Offset / EltSize; 2082 Offset -= Idx * EltSize; 2083 IdxTy = Type::getInt64Ty(T->getContext()); 2084 return Idx; 2085} 2086 2087/// RewriteGEP - Check if this GEP instruction moves the pointer across 2088/// elements of the alloca that are being split apart, and if so, rewrite 2089/// the GEP to be relative to the new element. 2090void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 2091 SmallVector<AllocaInst*, 32> &NewElts) { 2092 uint64_t OldOffset = Offset; 2093 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 2094 // If the GEP was dynamic then it must have been a dynamic vector lookup. 2095 // In this case, it must be the last GEP operand which is dynamic so keep that 2096 // aside until we've found the constant GEP offset then add it back in at the 2097 // end. 2098 Value* NonConstantIdx = 0; 2099 if (!GEPI->hasAllConstantIndices()) 2100 NonConstantIdx = Indices.pop_back_val(); 2101 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices); 2102 2103 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 2104 2105 Type *T = AI->getAllocatedType(); 2106 Type *IdxTy; 2107 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 2108 if (GEPI->getOperand(0) == AI) 2109 OldIdx = ~0ULL; // Force the GEP to be rewritten. 2110 2111 T = AI->getAllocatedType(); 2112 uint64_t EltOffset = Offset; 2113 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 2114 2115 // If this GEP does not move the pointer across elements of the alloca 2116 // being split, then it does not needs to be rewritten. 2117 if (Idx == OldIdx) 2118 return; 2119 2120 Type *i32Ty = Type::getInt32Ty(AI->getContext()); 2121 SmallVector<Value*, 8> NewArgs; 2122 NewArgs.push_back(Constant::getNullValue(i32Ty)); 2123 while (EltOffset != 0) { 2124 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 2125 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 2126 } 2127 if (NonConstantIdx) { 2128 Type* GepTy = T; 2129 // This GEP has a dynamic index. We need to add "i32 0" to index through 2130 // any structs or arrays in the original type until we get to the vector 2131 // to index. 2132 while (!isa<VectorType>(GepTy)) { 2133 NewArgs.push_back(Constant::getNullValue(i32Ty)); 2134 GepTy = cast<CompositeType>(GepTy)->getTypeAtIndex(0U); 2135 } 2136 NewArgs.push_back(NonConstantIdx); 2137 } 2138 Instruction *Val = NewElts[Idx]; 2139 if (NewArgs.size() > 1) { 2140 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs, "", GEPI); 2141 Val->takeName(GEPI); 2142 } 2143 if (Val->getType() != GEPI->getType()) 2144 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 2145 GEPI->replaceAllUsesWith(Val); 2146 DeadInsts.push_back(GEPI); 2147} 2148 2149/// RewriteLifetimeIntrinsic - II is a lifetime.start/lifetime.end. Rewrite it 2150/// to mark the lifetime of the scalarized memory. 2151void SROA::RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI, 2152 uint64_t Offset, 2153 SmallVector<AllocaInst*, 32> &NewElts) { 2154 ConstantInt *OldSize = cast<ConstantInt>(II->getArgOperand(0)); 2155 // Put matching lifetime markers on everything from Offset up to 2156 // Offset+OldSize. 2157 Type *AIType = AI->getAllocatedType(); 2158 uint64_t NewOffset = Offset; 2159 Type *IdxTy; 2160 uint64_t Idx = FindElementAndOffset(AIType, NewOffset, IdxTy); 2161 2162 IRBuilder<> Builder(II); 2163 uint64_t Size = OldSize->getLimitedValue(); 2164 2165 if (NewOffset) { 2166 // Splice the first element and index 'NewOffset' bytes in. SROA will 2167 // split the alloca again later. 2168 Value *V = Builder.CreateBitCast(NewElts[Idx], Builder.getInt8PtrTy()); 2169 V = Builder.CreateGEP(V, Builder.getInt64(NewOffset)); 2170 2171 IdxTy = NewElts[Idx]->getAllocatedType(); 2172 uint64_t EltSize = TD->getTypeAllocSize(IdxTy) - NewOffset; 2173 if (EltSize > Size) { 2174 EltSize = Size; 2175 Size = 0; 2176 } else { 2177 Size -= EltSize; 2178 } 2179 if (II->getIntrinsicID() == Intrinsic::lifetime_start) 2180 Builder.CreateLifetimeStart(V, Builder.getInt64(EltSize)); 2181 else 2182 Builder.CreateLifetimeEnd(V, Builder.getInt64(EltSize)); 2183 ++Idx; 2184 } 2185 2186 for (; Idx != NewElts.size() && Size; ++Idx) { 2187 IdxTy = NewElts[Idx]->getAllocatedType(); 2188 uint64_t EltSize = TD->getTypeAllocSize(IdxTy); 2189 if (EltSize > Size) { 2190 EltSize = Size; 2191 Size = 0; 2192 } else { 2193 Size -= EltSize; 2194 } 2195 if (II->getIntrinsicID() == Intrinsic::lifetime_start) 2196 Builder.CreateLifetimeStart(NewElts[Idx], 2197 Builder.getInt64(EltSize)); 2198 else 2199 Builder.CreateLifetimeEnd(NewElts[Idx], 2200 Builder.getInt64(EltSize)); 2201 } 2202 DeadInsts.push_back(II); 2203} 2204 2205/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 2206/// Rewrite it to copy or set the elements of the scalarized memory. 2207void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 2208 AllocaInst *AI, 2209 SmallVector<AllocaInst*, 32> &NewElts) { 2210 // If this is a memcpy/memmove, construct the other pointer as the 2211 // appropriate type. The "Other" pointer is the pointer that goes to memory 2212 // that doesn't have anything to do with the alloca that we are promoting. For 2213 // memset, this Value* stays null. 2214 Value *OtherPtr = 0; 2215 unsigned MemAlignment = MI->getAlignment(); 2216 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 2217 if (Inst == MTI->getRawDest()) 2218 OtherPtr = MTI->getRawSource(); 2219 else { 2220 assert(Inst == MTI->getRawSource()); 2221 OtherPtr = MTI->getRawDest(); 2222 } 2223 } 2224 2225 // If there is an other pointer, we want to convert it to the same pointer 2226 // type as AI has, so we can GEP through it safely. 2227 if (OtherPtr) { 2228 unsigned AddrSpace = 2229 cast<PointerType>(OtherPtr->getType())->getAddressSpace(); 2230 2231 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 2232 // optimization, but it's also required to detect the corner case where 2233 // both pointer operands are referencing the same memory, and where 2234 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 2235 // function is only called for mem intrinsics that access the whole 2236 // aggregate, so non-zero GEPs are not an issue here.) 2237 OtherPtr = OtherPtr->stripPointerCasts(); 2238 2239 // Copying the alloca to itself is a no-op: just delete it. 2240 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 2241 // This code will run twice for a no-op memcpy -- once for each operand. 2242 // Put only one reference to MI on the DeadInsts list. 2243 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), 2244 E = DeadInsts.end(); I != E; ++I) 2245 if (*I == MI) return; 2246 DeadInsts.push_back(MI); 2247 return; 2248 } 2249 2250 // If the pointer is not the right type, insert a bitcast to the right 2251 // type. 2252 Type *NewTy = 2253 PointerType::get(AI->getType()->getElementType(), AddrSpace); 2254 2255 if (OtherPtr->getType() != NewTy) 2256 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI); 2257 } 2258 2259 // Process each element of the aggregate. 2260 bool SROADest = MI->getRawDest() == Inst; 2261 2262 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 2263 2264 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2265 // If this is a memcpy/memmove, emit a GEP of the other element address. 2266 Value *OtherElt = 0; 2267 unsigned OtherEltAlign = MemAlignment; 2268 2269 if (OtherPtr) { 2270 Value *Idx[2] = { Zero, 2271 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 2272 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, 2273 OtherPtr->getName()+"."+Twine(i), 2274 MI); 2275 uint64_t EltOffset; 2276 PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 2277 Type *OtherTy = OtherPtrTy->getElementType(); 2278 if (StructType *ST = dyn_cast<StructType>(OtherTy)) { 2279 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 2280 } else { 2281 Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); 2282 EltOffset = TD->getTypeAllocSize(EltTy)*i; 2283 } 2284 2285 // The alignment of the other pointer is the guaranteed alignment of the 2286 // element, which is affected by both the known alignment of the whole 2287 // mem intrinsic and the alignment of the element. If the alignment of 2288 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 2289 // known alignment is just 4 bytes. 2290 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 2291 } 2292 2293 Value *EltPtr = NewElts[i]; 2294 Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 2295 2296 // If we got down to a scalar, insert a load or store as appropriate. 2297 if (EltTy->isSingleValueType()) { 2298 if (isa<MemTransferInst>(MI)) { 2299 if (SROADest) { 2300 // From Other to Alloca. 2301 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 2302 new StoreInst(Elt, EltPtr, MI); 2303 } else { 2304 // From Alloca to Other. 2305 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 2306 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 2307 } 2308 continue; 2309 } 2310 assert(isa<MemSetInst>(MI)); 2311 2312 // If the stored element is zero (common case), just store a null 2313 // constant. 2314 Constant *StoreVal; 2315 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) { 2316 if (CI->isZero()) { 2317 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 2318 } else { 2319 // If EltTy is a vector type, get the element type. 2320 Type *ValTy = EltTy->getScalarType(); 2321 2322 // Construct an integer with the right value. 2323 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 2324 APInt OneVal(EltSize, CI->getZExtValue()); 2325 APInt TotalVal(OneVal); 2326 // Set each byte. 2327 for (unsigned i = 0; 8*i < EltSize; ++i) { 2328 TotalVal = TotalVal.shl(8); 2329 TotalVal |= OneVal; 2330 } 2331 2332 // Convert the integer value to the appropriate type. 2333 StoreVal = ConstantInt::get(CI->getContext(), TotalVal); 2334 if (ValTy->isPointerTy()) 2335 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 2336 else if (ValTy->isFloatingPointTy()) 2337 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 2338 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 2339 2340 // If the requested value was a vector constant, create it. 2341 if (EltTy->isVectorTy()) { 2342 unsigned NumElts = cast<VectorType>(EltTy)->getNumElements(); 2343 StoreVal = ConstantVector::getSplat(NumElts, StoreVal); 2344 } 2345 } 2346 new StoreInst(StoreVal, EltPtr, MI); 2347 continue; 2348 } 2349 // Otherwise, if we're storing a byte variable, use a memset call for 2350 // this element. 2351 } 2352 2353 unsigned EltSize = TD->getTypeAllocSize(EltTy); 2354 if (!EltSize) 2355 continue; 2356 2357 IRBuilder<> Builder(MI); 2358 2359 // Finally, insert the meminst for this element. 2360 if (isa<MemSetInst>(MI)) { 2361 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize, 2362 MI->isVolatile()); 2363 } else { 2364 assert(isa<MemTransferInst>(MI)); 2365 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr 2366 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr 2367 2368 if (isa<MemCpyInst>(MI)) 2369 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile()); 2370 else 2371 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile()); 2372 } 2373 } 2374 DeadInsts.push_back(MI); 2375} 2376 2377/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 2378/// overwrites the entire allocation. Extract out the pieces of the stored 2379/// integer and store them individually. 2380void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 2381 SmallVector<AllocaInst*, 32> &NewElts){ 2382 // Extract each element out of the integer according to its structure offset 2383 // and store the element value to the individual alloca. 2384 Value *SrcVal = SI->getOperand(0); 2385 Type *AllocaEltTy = AI->getAllocatedType(); 2386 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2387 2388 IRBuilder<> Builder(SI); 2389 2390 // Handle tail padding by extending the operand 2391 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 2392 SrcVal = Builder.CreateZExt(SrcVal, 2393 IntegerType::get(SI->getContext(), AllocaSizeBits)); 2394 2395 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 2396 << '\n'); 2397 2398 // There are two forms here: AI could be an array or struct. Both cases 2399 // have different ways to compute the element offset. 2400 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2401 const StructLayout *Layout = TD->getStructLayout(EltSTy); 2402 2403 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2404 // Get the number of bits to shift SrcVal to get the value. 2405 Type *FieldTy = EltSTy->getElementType(i); 2406 uint64_t Shift = Layout->getElementOffsetInBits(i); 2407 2408 if (TD->isBigEndian()) 2409 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 2410 2411 Value *EltVal = SrcVal; 2412 if (Shift) { 2413 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2414 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2415 } 2416 2417 // Truncate down to an integer of the right size. 2418 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2419 2420 // Ignore zero sized fields like {}, they obviously contain no data. 2421 if (FieldSizeBits == 0) continue; 2422 2423 if (FieldSizeBits != AllocaSizeBits) 2424 EltVal = Builder.CreateTrunc(EltVal, 2425 IntegerType::get(SI->getContext(), FieldSizeBits)); 2426 Value *DestField = NewElts[i]; 2427 if (EltVal->getType() == FieldTy) { 2428 // Storing to an integer field of this size, just do it. 2429 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 2430 // Bitcast to the right element type (for fp/vector values). 2431 EltVal = Builder.CreateBitCast(EltVal, FieldTy); 2432 } else { 2433 // Otherwise, bitcast the dest pointer (for aggregates). 2434 DestField = Builder.CreateBitCast(DestField, 2435 PointerType::getUnqual(EltVal->getType())); 2436 } 2437 new StoreInst(EltVal, DestField, SI); 2438 } 2439 2440 } else { 2441 ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 2442 Type *ArrayEltTy = ATy->getElementType(); 2443 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2444 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 2445 2446 uint64_t Shift; 2447 2448 if (TD->isBigEndian()) 2449 Shift = AllocaSizeBits-ElementOffset; 2450 else 2451 Shift = 0; 2452 2453 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2454 // Ignore zero sized fields like {}, they obviously contain no data. 2455 if (ElementSizeBits == 0) continue; 2456 2457 Value *EltVal = SrcVal; 2458 if (Shift) { 2459 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2460 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2461 } 2462 2463 // Truncate down to an integer of the right size. 2464 if (ElementSizeBits != AllocaSizeBits) 2465 EltVal = Builder.CreateTrunc(EltVal, 2466 IntegerType::get(SI->getContext(), 2467 ElementSizeBits)); 2468 Value *DestField = NewElts[i]; 2469 if (EltVal->getType() == ArrayEltTy) { 2470 // Storing to an integer field of this size, just do it. 2471 } else if (ArrayEltTy->isFloatingPointTy() || 2472 ArrayEltTy->isVectorTy()) { 2473 // Bitcast to the right element type (for fp/vector values). 2474 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy); 2475 } else { 2476 // Otherwise, bitcast the dest pointer (for aggregates). 2477 DestField = Builder.CreateBitCast(DestField, 2478 PointerType::getUnqual(EltVal->getType())); 2479 } 2480 new StoreInst(EltVal, DestField, SI); 2481 2482 if (TD->isBigEndian()) 2483 Shift -= ElementOffset; 2484 else 2485 Shift += ElementOffset; 2486 } 2487 } 2488 2489 DeadInsts.push_back(SI); 2490} 2491 2492/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 2493/// an integer. Load the individual pieces to form the aggregate value. 2494void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 2495 SmallVector<AllocaInst*, 32> &NewElts) { 2496 // Extract each element out of the NewElts according to its structure offset 2497 // and form the result value. 2498 Type *AllocaEltTy = AI->getAllocatedType(); 2499 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 2500 2501 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 2502 << '\n'); 2503 2504 // There are two forms here: AI could be an array or struct. Both cases 2505 // have different ways to compute the element offset. 2506 const StructLayout *Layout = 0; 2507 uint64_t ArrayEltBitOffset = 0; 2508 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2509 Layout = TD->getStructLayout(EltSTy); 2510 } else { 2511 Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 2512 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 2513 } 2514 2515 Value *ResultVal = 2516 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 2517 2518 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2519 // Load the value from the alloca. If the NewElt is an aggregate, cast 2520 // the pointer to an integer of the same size before doing the load. 2521 Value *SrcField = NewElts[i]; 2522 Type *FieldTy = 2523 cast<PointerType>(SrcField->getType())->getElementType(); 2524 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 2525 2526 // Ignore zero sized fields like {}, they obviously contain no data. 2527 if (FieldSizeBits == 0) continue; 2528 2529 IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 2530 FieldSizeBits); 2531 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 2532 !FieldTy->isVectorTy()) 2533 SrcField = new BitCastInst(SrcField, 2534 PointerType::getUnqual(FieldIntTy), 2535 "", LI); 2536 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 2537 2538 // If SrcField is a fp or vector of the right size but that isn't an 2539 // integer type, bitcast to an integer so we can shift it. 2540 if (SrcField->getType() != FieldIntTy) 2541 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 2542 2543 // Zero extend the field to be the same size as the final alloca so that 2544 // we can shift and insert it. 2545 if (SrcField->getType() != ResultVal->getType()) 2546 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 2547 2548 // Determine the number of bits to shift SrcField. 2549 uint64_t Shift; 2550 if (Layout) // Struct case. 2551 Shift = Layout->getElementOffsetInBits(i); 2552 else // Array case. 2553 Shift = i*ArrayEltBitOffset; 2554 2555 if (TD->isBigEndian()) 2556 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 2557 2558 if (Shift) { 2559 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 2560 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 2561 } 2562 2563 // Don't create an 'or x, 0' on the first iteration. 2564 if (!isa<Constant>(ResultVal) || 2565 !cast<Constant>(ResultVal)->isNullValue()) 2566 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 2567 else 2568 ResultVal = SrcField; 2569 } 2570 2571 // Handle tail padding by truncating the result 2572 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 2573 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 2574 2575 LI->replaceAllUsesWith(ResultVal); 2576 DeadInsts.push_back(LI); 2577} 2578 2579/// HasPadding - Return true if the specified type has any structure or 2580/// alignment padding in between the elements that would be split apart 2581/// by SROA; return false otherwise. 2582static bool HasPadding(Type *Ty, const TargetData &TD) { 2583 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 2584 Ty = ATy->getElementType(); 2585 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 2586 } 2587 2588 // SROA currently handles only Arrays and Structs. 2589 StructType *STy = cast<StructType>(Ty); 2590 const StructLayout *SL = TD.getStructLayout(STy); 2591 unsigned PrevFieldBitOffset = 0; 2592 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 2593 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 2594 2595 // Check to see if there is any padding between this element and the 2596 // previous one. 2597 if (i) { 2598 unsigned PrevFieldEnd = 2599 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 2600 if (PrevFieldEnd < FieldBitOffset) 2601 return true; 2602 } 2603 PrevFieldBitOffset = FieldBitOffset; 2604 } 2605 // Check for tail padding. 2606 if (unsigned EltCount = STy->getNumElements()) { 2607 unsigned PrevFieldEnd = PrevFieldBitOffset + 2608 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 2609 if (PrevFieldEnd < SL->getSizeInBits()) 2610 return true; 2611 } 2612 return false; 2613} 2614 2615/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 2616/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 2617/// or 1 if safe after canonicalization has been performed. 2618bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 2619 // Loop over the use list of the alloca. We can only transform it if all of 2620 // the users are safe to transform. 2621 AllocaInfo Info(AI); 2622 2623 isSafeForScalarRepl(AI, 0, Info); 2624 if (Info.isUnsafe) { 2625 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 2626 return false; 2627 } 2628 2629 // Okay, we know all the users are promotable. If the aggregate is a memcpy 2630 // source and destination, we have to be careful. In particular, the memcpy 2631 // could be moving around elements that live in structure padding of the LLVM 2632 // types, but may actually be used. In these cases, we refuse to promote the 2633 // struct. 2634 if (Info.isMemCpySrc && Info.isMemCpyDst && 2635 HasPadding(AI->getAllocatedType(), *TD)) 2636 return false; 2637 2638 // If the alloca never has an access to just *part* of it, but is accessed 2639 // via loads and stores, then we should use ConvertToScalarInfo to promote 2640 // the alloca instead of promoting each piece at a time and inserting fission 2641 // and fusion code. 2642 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) { 2643 // If the struct/array just has one element, use basic SRoA. 2644 if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 2645 if (ST->getNumElements() > 1) return false; 2646 } else { 2647 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1) 2648 return false; 2649 } 2650 } 2651 2652 return true; 2653} 2654 2655 2656 2657/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 2658/// some part of a constant global variable. This intentionally only accepts 2659/// constant expressions because we don't can't rewrite arbitrary instructions. 2660static bool PointsToConstantGlobal(Value *V) { 2661 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 2662 return GV->isConstant(); 2663 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 2664 if (CE->getOpcode() == Instruction::BitCast || 2665 CE->getOpcode() == Instruction::GetElementPtr) 2666 return PointsToConstantGlobal(CE->getOperand(0)); 2667 return false; 2668} 2669 2670/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 2671/// pointer to an alloca. Ignore any reads of the pointer, return false if we 2672/// see any stores or other unknown uses. If we see pointer arithmetic, keep 2673/// track of whether it moves the pointer (with isOffset) but otherwise traverse 2674/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 2675/// the alloca, and if the source pointer is a pointer to a constant global, we 2676/// can optimize this. 2677static bool 2678isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, 2679 bool isOffset, 2680 SmallVector<Instruction *, 4> &LifetimeMarkers) { 2681 // We track lifetime intrinsics as we encounter them. If we decide to go 2682 // ahead and replace the value with the global, this lets the caller quickly 2683 // eliminate the markers. 2684 2685 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 2686 User *U = cast<Instruction>(*UI); 2687 2688 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 2689 // Ignore non-volatile loads, they are always ok. 2690 if (!LI->isSimple()) return false; 2691 continue; 2692 } 2693 2694 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 2695 // If uses of the bitcast are ok, we are ok. 2696 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset, 2697 LifetimeMarkers)) 2698 return false; 2699 continue; 2700 } 2701 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 2702 // If the GEP has all zero indices, it doesn't offset the pointer. If it 2703 // doesn't, it does. 2704 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 2705 isOffset || !GEP->hasAllZeroIndices(), 2706 LifetimeMarkers)) 2707 return false; 2708 continue; 2709 } 2710 2711 if (CallSite CS = U) { 2712 // If this is the function being called then we treat it like a load and 2713 // ignore it. 2714 if (CS.isCallee(UI)) 2715 continue; 2716 2717 // If this is a readonly/readnone call site, then we know it is just a 2718 // load (but one that potentially returns the value itself), so we can 2719 // ignore it if we know that the value isn't captured. 2720 unsigned ArgNo = CS.getArgumentNo(UI); 2721 if (CS.onlyReadsMemory() && 2722 (CS.getInstruction()->use_empty() || CS.doesNotCapture(ArgNo))) 2723 continue; 2724 2725 // If this is being passed as a byval argument, the caller is making a 2726 // copy, so it is only a read of the alloca. 2727 if (CS.isByValArgument(ArgNo)) 2728 continue; 2729 } 2730 2731 // Lifetime intrinsics can be handled by the caller. 2732 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) { 2733 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 2734 II->getIntrinsicID() == Intrinsic::lifetime_end) { 2735 assert(II->use_empty() && "Lifetime markers have no result to use!"); 2736 LifetimeMarkers.push_back(II); 2737 continue; 2738 } 2739 } 2740 2741 // If this is isn't our memcpy/memmove, reject it as something we can't 2742 // handle. 2743 MemTransferInst *MI = dyn_cast<MemTransferInst>(U); 2744 if (MI == 0) 2745 return false; 2746 2747 // If the transfer is using the alloca as a source of the transfer, then 2748 // ignore it since it is a load (unless the transfer is volatile). 2749 if (UI.getOperandNo() == 1) { 2750 if (MI->isVolatile()) return false; 2751 continue; 2752 } 2753 2754 // If we already have seen a copy, reject the second one. 2755 if (TheCopy) return false; 2756 2757 // If the pointer has been offset from the start of the alloca, we can't 2758 // safely handle this. 2759 if (isOffset) return false; 2760 2761 // If the memintrinsic isn't using the alloca as the dest, reject it. 2762 if (UI.getOperandNo() != 0) return false; 2763 2764 // If the source of the memcpy/move is not a constant global, reject it. 2765 if (!PointsToConstantGlobal(MI->getSource())) 2766 return false; 2767 2768 // Otherwise, the transform is safe. Remember the copy instruction. 2769 TheCopy = MI; 2770 } 2771 return true; 2772} 2773 2774/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 2775/// modified by a copy from a constant global. If we can prove this, we can 2776/// replace any uses of the alloca with uses of the global directly. 2777MemTransferInst * 2778SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI, 2779 SmallVector<Instruction*, 4> &ToDelete) { 2780 MemTransferInst *TheCopy = 0; 2781 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false, ToDelete)) 2782 return TheCopy; 2783 return 0; 2784} 2785