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 #include "llvm/Transforms/Scalar.h" 23 #include "llvm/ADT/SetVector.h" 24 #include "llvm/ADT/SmallVector.h" 25 #include "llvm/ADT/Statistic.h" 26 #include "llvm/Analysis/Loads.h" 27 #include "llvm/Analysis/ValueTracking.h" 28 #include "llvm/IR/CallSite.h" 29 #include "llvm/IR/Constants.h" 30 #include "llvm/IR/DIBuilder.h" 31 #include "llvm/IR/DataLayout.h" 32 #include "llvm/IR/DebugInfo.h" 33 #include "llvm/IR/DerivedTypes.h" 34 #include "llvm/IR/Dominators.h" 35 #include "llvm/IR/Function.h" 36 #include "llvm/IR/GetElementPtrTypeIterator.h" 37 #include "llvm/IR/GlobalVariable.h" 38 #include "llvm/IR/IRBuilder.h" 39 #include "llvm/IR/Instructions.h" 40 #include "llvm/IR/IntrinsicInst.h" 41 #include "llvm/IR/LLVMContext.h" 42 #include "llvm/IR/Module.h" 43 #include "llvm/IR/Operator.h" 44 #include "llvm/Pass.h" 45 #include "llvm/Support/Debug.h" 46 #include "llvm/Support/ErrorHandling.h" 47 #include "llvm/Support/MathExtras.h" 48 #include "llvm/Support/raw_ostream.h" 49 #include "llvm/Transforms/Utils/Local.h" 50 #include "llvm/Transforms/Utils/PromoteMemToReg.h" 51 #include "llvm/Transforms/Utils/SSAUpdater.h" 52 using namespace llvm; 53 54 #define DEBUG_TYPE "scalarrepl" 55 56 STATISTIC(NumReplaced, "Number of allocas broken up"); 57 STATISTIC(NumPromoted, "Number of allocas promoted"); 58 STATISTIC(NumAdjusted, "Number of scalar allocas adjusted to allow promotion"); 59 STATISTIC(NumConverted, "Number of aggregates converted to scalar"); 60 61 namespace { 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) override; 85 86 bool performScalarRepl(Function &F); 87 bool performPromotion(Function &F); 88 89 private: 90 bool HasDomTree; 91 const DataLayout *DL; 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 SmallVectorImpl<AllocaInst *> &NewElts); 171 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 172 SmallVectorImpl<AllocaInst *> &NewElts); 173 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 174 SmallVectorImpl<AllocaInst *> &NewElts); 175 void RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI, 176 uint64_t Offset, 177 SmallVectorImpl<AllocaInst *> &NewElts); 178 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 179 AllocaInst *AI, 180 SmallVectorImpl<AllocaInst *> &NewElts); 181 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 182 SmallVectorImpl<AllocaInst *> &NewElts); 183 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 184 SmallVectorImpl<AllocaInst *> &NewElts); 185 bool ShouldAttemptScalarRepl(AllocaInst *AI); 186 }; 187 188 // SROA_DT - SROA that uses DominatorTree. 189 struct SROA_DT : public SROA { 190 static char ID; 191 public: 192 SROA_DT(int T = -1, int ST = -1, int AT = -1, int SLT = -1) : 193 SROA(T, true, ID, ST, AT, SLT) { 194 initializeSROA_DTPass(*PassRegistry::getPassRegistry()); 195 } 196 197 // getAnalysisUsage - This pass does not require any passes, but we know it 198 // will not alter the CFG, so say so. 199 void getAnalysisUsage(AnalysisUsage &AU) const override { 200 AU.addRequired<DominatorTreeWrapperPass>(); 201 AU.setPreservesCFG(); 202 } 203 }; 204 205 // SROA_SSAUp - SROA that uses SSAUpdater. 206 struct SROA_SSAUp : public SROA { 207 static char ID; 208 public: 209 SROA_SSAUp(int T = -1, int ST = -1, int AT = -1, int SLT = -1) : 210 SROA(T, false, ID, ST, AT, SLT) { 211 initializeSROA_SSAUpPass(*PassRegistry::getPassRegistry()); 212 } 213 214 // getAnalysisUsage - This pass does not require any passes, but we know it 215 // will not alter the CFG, so say so. 216 void getAnalysisUsage(AnalysisUsage &AU) const override { 217 AU.setPreservesCFG(); 218 } 219 }; 220 221 } 222 223 char SROA_DT::ID = 0; 224 char SROA_SSAUp::ID = 0; 225 226 INITIALIZE_PASS_BEGIN(SROA_DT, "scalarrepl", 227 "Scalar Replacement of Aggregates (DT)", false, false) 228 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 229 INITIALIZE_PASS_END(SROA_DT, "scalarrepl", 230 "Scalar Replacement of Aggregates (DT)", false, false) 231 232 INITIALIZE_PASS_BEGIN(SROA_SSAUp, "scalarrepl-ssa", 233 "Scalar Replacement of Aggregates (SSAUp)", false, false) 234 INITIALIZE_PASS_END(SROA_SSAUp, "scalarrepl-ssa", 235 "Scalar Replacement of Aggregates (SSAUp)", false, false) 236 237 // Public interface to the ScalarReplAggregates pass 238 FunctionPass *llvm::createScalarReplAggregatesPass(int Threshold, 239 bool UseDomTree, 240 int StructMemberThreshold, 241 int ArrayElementThreshold, 242 int ScalarLoadThreshold) { 243 if (UseDomTree) 244 return new SROA_DT(Threshold, StructMemberThreshold, ArrayElementThreshold, 245 ScalarLoadThreshold); 246 return new SROA_SSAUp(Threshold, StructMemberThreshold, 247 ArrayElementThreshold, ScalarLoadThreshold); 248 } 249 250 251 //===----------------------------------------------------------------------===// 252 // Convert To Scalar Optimization. 253 //===----------------------------------------------------------------------===// 254 255 namespace { 256 /// ConvertToScalarInfo - This class implements the "Convert To Scalar" 257 /// optimization, which scans the uses of an alloca and determines if it can 258 /// rewrite it in terms of a single new alloca that can be mem2reg'd. 259 class ConvertToScalarInfo { 260 /// AllocaSize - The size of the alloca being considered in bytes. 261 unsigned AllocaSize; 262 const DataLayout &DL; 263 unsigned ScalarLoadThreshold; 264 265 /// IsNotTrivial - This is set to true if there is some access to the object 266 /// which means that mem2reg can't promote it. 267 bool IsNotTrivial; 268 269 /// ScalarKind - Tracks the kind of alloca being considered for promotion, 270 /// computed based on the uses of the alloca rather than the LLVM type system. 271 enum { 272 Unknown, 273 274 // Accesses via GEPs that are consistent with element access of a vector 275 // type. This will not be converted into a vector unless there is a later 276 // access using an actual vector type. 277 ImplicitVector, 278 279 // Accesses via vector operations and GEPs that are consistent with the 280 // layout of a vector type. 281 Vector, 282 283 // An integer bag-of-bits with bitwise operations for insertion and 284 // extraction. Any combination of types can be converted into this kind 285 // of scalar. 286 Integer 287 } ScalarKind; 288 289 /// VectorTy - This tracks the type that we should promote the vector to if 290 /// it is possible to turn it into a vector. This starts out null, and if it 291 /// isn't possible to turn into a vector type, it gets set to VoidTy. 292 VectorType *VectorTy; 293 294 /// HadNonMemTransferAccess - True if there is at least one access to the 295 /// alloca that is not a MemTransferInst. We don't want to turn structs into 296 /// large integers unless there is some potential for optimization. 297 bool HadNonMemTransferAccess; 298 299 /// HadDynamicAccess - True if some element of this alloca was dynamic. 300 /// We don't yet have support for turning a dynamic access into a large 301 /// integer. 302 bool HadDynamicAccess; 303 304 public: 305 explicit ConvertToScalarInfo(unsigned Size, const DataLayout &DL, 306 unsigned SLT) 307 : AllocaSize(Size), DL(DL), ScalarLoadThreshold(SLT), IsNotTrivial(false), 308 ScalarKind(Unknown), VectorTy(nullptr), HadNonMemTransferAccess(false), 309 HadDynamicAccess(false) { } 310 311 AllocaInst *TryConvert(AllocaInst *AI); 312 313 private: 314 bool CanConvertToScalar(Value *V, uint64_t Offset, Value* NonConstantIdx); 315 void MergeInTypeForLoadOrStore(Type *In, uint64_t Offset); 316 bool MergeInVectorType(VectorType *VInTy, uint64_t Offset); 317 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset, 318 Value *NonConstantIdx); 319 320 Value *ConvertScalar_ExtractValue(Value *NV, Type *ToType, 321 uint64_t Offset, Value* NonConstantIdx, 322 IRBuilder<> &Builder); 323 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 324 uint64_t Offset, Value* NonConstantIdx, 325 IRBuilder<> &Builder); 326 }; 327 } // end anonymous namespace. 328 329 330 /// TryConvert - Analyze the specified alloca, and if it is safe to do so, 331 /// rewrite it to be a new alloca which is mem2reg'able. This returns the new 332 /// alloca if possible or null if not. 333 AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) { 334 // If we can't convert this scalar, or if mem2reg can trivially do it, bail 335 // out. 336 if (!CanConvertToScalar(AI, 0, nullptr) || !IsNotTrivial) 337 return nullptr; 338 339 // If an alloca has only memset / memcpy uses, it may still have an Unknown 340 // ScalarKind. Treat it as an Integer below. 341 if (ScalarKind == Unknown) 342 ScalarKind = Integer; 343 344 if (ScalarKind == Vector && VectorTy->getBitWidth() != AllocaSize * 8) 345 ScalarKind = Integer; 346 347 // If we were able to find a vector type that can handle this with 348 // insert/extract elements, and if there was at least one use that had 349 // a vector type, promote this to a vector. We don't want to promote 350 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 351 // we just get a lot of insert/extracts. If at least one vector is 352 // involved, then we probably really do have a union of vector/array. 353 Type *NewTy; 354 if (ScalarKind == Vector) { 355 assert(VectorTy && "Missing type for vector scalar."); 356 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = " 357 << *VectorTy << '\n'); 358 NewTy = VectorTy; // Use the vector type. 359 } else { 360 unsigned BitWidth = AllocaSize * 8; 361 362 // Do not convert to scalar integer if the alloca size exceeds the 363 // scalar load threshold. 364 if (BitWidth > ScalarLoadThreshold) 365 return nullptr; 366 367 if ((ScalarKind == ImplicitVector || ScalarKind == Integer) && 368 !HadNonMemTransferAccess && !DL.fitsInLegalInteger(BitWidth)) 369 return nullptr; 370 // Dynamic accesses on integers aren't yet supported. They need us to shift 371 // by a dynamic amount which could be difficult to work out as we might not 372 // know whether to use a left or right shift. 373 if (ScalarKind == Integer && HadDynamicAccess) 374 return nullptr; 375 376 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"); 377 // Create and insert the integer alloca. 378 NewTy = IntegerType::get(AI->getContext(), BitWidth); 379 } 380 AllocaInst *NewAI = new AllocaInst(NewTy, nullptr, "", 381 AI->getParent()->begin()); 382 ConvertUsesToScalar(AI, NewAI, 0, nullptr); 383 return NewAI; 384 } 385 386 /// MergeInTypeForLoadOrStore - Add the 'In' type to the accumulated vector type 387 /// (VectorTy) so far at the offset specified by Offset (which is specified in 388 /// bytes). 389 /// 390 /// There are two cases we handle here: 391 /// 1) A union of vector types of the same size and potentially its elements. 392 /// Here we turn element accesses into insert/extract element operations. 393 /// This promotes a <4 x float> with a store of float to the third element 394 /// into a <4 x float> that uses insert element. 395 /// 2) A fully general blob of memory, which we turn into some (potentially 396 /// large) integer type with extract and insert operations where the loads 397 /// and stores would mutate the memory. We mark this by setting VectorTy 398 /// to VoidTy. 399 void ConvertToScalarInfo::MergeInTypeForLoadOrStore(Type *In, 400 uint64_t Offset) { 401 // If we already decided to turn this into a blob of integer memory, there is 402 // nothing to be done. 403 if (ScalarKind == Integer) 404 return; 405 406 // If this could be contributing to a vector, analyze it. 407 408 // If the In type is a vector that is the same size as the alloca, see if it 409 // matches the existing VecTy. 410 if (VectorType *VInTy = dyn_cast<VectorType>(In)) { 411 if (MergeInVectorType(VInTy, Offset)) 412 return; 413 } else if (In->isFloatTy() || In->isDoubleTy() || 414 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 && 415 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 416 // Full width accesses can be ignored, because they can always be turned 417 // into bitcasts. 418 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 419 if (EltSize == AllocaSize) 420 return; 421 422 // If we're accessing something that could be an element of a vector, see 423 // if the implied vector agrees with what we already have and if Offset is 424 // compatible with it. 425 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 && 426 (!VectorTy || EltSize == VectorTy->getElementType() 427 ->getPrimitiveSizeInBits()/8)) { 428 if (!VectorTy) { 429 ScalarKind = ImplicitVector; 430 VectorTy = VectorType::get(In, AllocaSize/EltSize); 431 } 432 return; 433 } 434 } 435 436 // Otherwise, we have a case that we can't handle with an optimized vector 437 // form. We can still turn this into a large integer. 438 ScalarKind = Integer; 439 } 440 441 /// MergeInVectorType - Handles the vector case of MergeInTypeForLoadOrStore, 442 /// returning true if the type was successfully merged and false otherwise. 443 bool ConvertToScalarInfo::MergeInVectorType(VectorType *VInTy, 444 uint64_t Offset) { 445 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 446 // If we're storing/loading a vector of the right size, allow it as a 447 // vector. If this the first vector we see, remember the type so that 448 // we know the element size. If this is a subsequent access, ignore it 449 // even if it is a differing type but the same size. Worst case we can 450 // bitcast the resultant vectors. 451 if (!VectorTy) 452 VectorTy = VInTy; 453 ScalarKind = Vector; 454 return true; 455 } 456 457 return false; 458 } 459 460 /// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 461 /// its accesses to a single vector type, return true and set VecTy to 462 /// the new type. If we could convert the alloca into a single promotable 463 /// integer, return true but set VecTy to VoidTy. Further, if the use is not a 464 /// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 465 /// is the current offset from the base of the alloca being analyzed. 466 /// 467 /// If we see at least one access to the value that is as a vector type, set the 468 /// SawVec flag. 469 bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset, 470 Value* NonConstantIdx) { 471 for (User *U : V->users()) { 472 Instruction *UI = cast<Instruction>(U); 473 474 if (LoadInst *LI = dyn_cast<LoadInst>(UI)) { 475 // Don't break volatile loads. 476 if (!LI->isSimple()) 477 return false; 478 // Don't touch MMX operations. 479 if (LI->getType()->isX86_MMXTy()) 480 return false; 481 HadNonMemTransferAccess = true; 482 MergeInTypeForLoadOrStore(LI->getType(), Offset); 483 continue; 484 } 485 486 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) { 487 // Storing the pointer, not into the value? 488 if (SI->getOperand(0) == V || !SI->isSimple()) return false; 489 // Don't touch MMX operations. 490 if (SI->getOperand(0)->getType()->isX86_MMXTy()) 491 return false; 492 HadNonMemTransferAccess = true; 493 MergeInTypeForLoadOrStore(SI->getOperand(0)->getType(), Offset); 494 continue; 495 } 496 497 if (BitCastInst *BCI = dyn_cast<BitCastInst>(UI)) { 498 if (!onlyUsedByLifetimeMarkers(BCI)) 499 IsNotTrivial = true; // Can't be mem2reg'd. 500 if (!CanConvertToScalar(BCI, Offset, NonConstantIdx)) 501 return false; 502 continue; 503 } 504 505 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(UI)) { 506 // If this is a GEP with a variable indices, we can't handle it. 507 PointerType* PtrTy = dyn_cast<PointerType>(GEP->getPointerOperandType()); 508 if (!PtrTy) 509 return false; 510 511 // Compute the offset that this GEP adds to the pointer. 512 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 513 Value *GEPNonConstantIdx = nullptr; 514 if (!GEP->hasAllConstantIndices()) { 515 if (!isa<VectorType>(PtrTy->getElementType())) 516 return false; 517 if (NonConstantIdx) 518 return false; 519 GEPNonConstantIdx = Indices.pop_back_val(); 520 if (!GEPNonConstantIdx->getType()->isIntegerTy(32)) 521 return false; 522 HadDynamicAccess = true; 523 } else 524 GEPNonConstantIdx = NonConstantIdx; 525 uint64_t GEPOffset = DL.getIndexedOffset(PtrTy, 526 Indices); 527 // See if all uses can be converted. 528 if (!CanConvertToScalar(GEP, Offset+GEPOffset, GEPNonConstantIdx)) 529 return false; 530 IsNotTrivial = true; // Can't be mem2reg'd. 531 HadNonMemTransferAccess = true; 532 continue; 533 } 534 535 // If this is a constant sized memset of a constant value (e.g. 0) we can 536 // handle it. 537 if (MemSetInst *MSI = dyn_cast<MemSetInst>(UI)) { 538 // Store to dynamic index. 539 if (NonConstantIdx) 540 return false; 541 // Store of constant value. 542 if (!isa<ConstantInt>(MSI->getValue())) 543 return false; 544 545 // Store of constant size. 546 ConstantInt *Len = dyn_cast<ConstantInt>(MSI->getLength()); 547 if (!Len) 548 return false; 549 550 // If the size differs from the alloca, we can only convert the alloca to 551 // an integer bag-of-bits. 552 // FIXME: This should handle all of the cases that are currently accepted 553 // as vector element insertions. 554 if (Len->getZExtValue() != AllocaSize || Offset != 0) 555 ScalarKind = Integer; 556 557 IsNotTrivial = true; // Can't be mem2reg'd. 558 HadNonMemTransferAccess = true; 559 continue; 560 } 561 562 // If this is a memcpy or memmove into or out of the whole allocation, we 563 // can handle it like a load or store of the scalar type. 564 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(UI)) { 565 // Store to dynamic index. 566 if (NonConstantIdx) 567 return false; 568 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength()); 569 if (!Len || Len->getZExtValue() != AllocaSize || Offset != 0) 570 return false; 571 572 IsNotTrivial = true; // Can't be mem2reg'd. 573 continue; 574 } 575 576 // If this is a lifetime intrinsic, we can handle it. 577 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(UI)) { 578 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 579 II->getIntrinsicID() == Intrinsic::lifetime_end) { 580 continue; 581 } 582 } 583 584 // Otherwise, we cannot handle this! 585 return false; 586 } 587 588 return true; 589 } 590 591 /// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 592 /// directly. This happens when we are converting an "integer union" to a 593 /// single integer scalar, or when we are converting a "vector union" to a 594 /// vector with insert/extractelement instructions. 595 /// 596 /// Offset is an offset from the original alloca, in bits that need to be 597 /// shifted to the right. By the end of this, there should be no uses of Ptr. 598 void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, 599 uint64_t Offset, 600 Value* NonConstantIdx) { 601 while (!Ptr->use_empty()) { 602 Instruction *User = cast<Instruction>(Ptr->user_back()); 603 604 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 605 ConvertUsesToScalar(CI, NewAI, Offset, NonConstantIdx); 606 CI->eraseFromParent(); 607 continue; 608 } 609 610 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 611 // Compute the offset that this GEP adds to the pointer. 612 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 613 Value* GEPNonConstantIdx = nullptr; 614 if (!GEP->hasAllConstantIndices()) { 615 assert(!NonConstantIdx && 616 "Dynamic GEP reading from dynamic GEP unsupported"); 617 GEPNonConstantIdx = Indices.pop_back_val(); 618 } else 619 GEPNonConstantIdx = NonConstantIdx; 620 uint64_t GEPOffset = DL.getIndexedOffset(GEP->getPointerOperandType(), 621 Indices); 622 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8, GEPNonConstantIdx); 623 GEP->eraseFromParent(); 624 continue; 625 } 626 627 IRBuilder<> Builder(User); 628 629 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 630 // The load is a bit extract from NewAI shifted right by Offset bits. 631 Value *LoadedVal = Builder.CreateLoad(NewAI); 632 Value *NewLoadVal 633 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, 634 NonConstantIdx, Builder); 635 LI->replaceAllUsesWith(NewLoadVal); 636 LI->eraseFromParent(); 637 continue; 638 } 639 640 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 641 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 642 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 643 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 644 NonConstantIdx, Builder); 645 Builder.CreateStore(New, NewAI); 646 SI->eraseFromParent(); 647 648 // If the load we just inserted is now dead, then the inserted store 649 // overwrote the entire thing. 650 if (Old->use_empty()) 651 Old->eraseFromParent(); 652 continue; 653 } 654 655 // If this is a constant sized memset of a constant value (e.g. 0) we can 656 // transform it into a store of the expanded constant value. 657 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 658 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 659 assert(!NonConstantIdx && "Cannot replace dynamic memset with insert"); 660 int64_t SNumBytes = cast<ConstantInt>(MSI->getLength())->getSExtValue(); 661 if (SNumBytes > 0 && (SNumBytes >> 32) == 0) { 662 unsigned NumBytes = static_cast<unsigned>(SNumBytes); 663 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 664 665 // Compute the value replicated the right number of times. 666 APInt APVal(NumBytes*8, Val); 667 668 // Splat the value if non-zero. 669 if (Val) 670 for (unsigned i = 1; i != NumBytes; ++i) 671 APVal |= APVal << 8; 672 673 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 674 Value *New = ConvertScalar_InsertValue( 675 ConstantInt::get(User->getContext(), APVal), 676 Old, Offset, nullptr, Builder); 677 Builder.CreateStore(New, NewAI); 678 679 // If the load we just inserted is now dead, then the memset overwrote 680 // the entire thing. 681 if (Old->use_empty()) 682 Old->eraseFromParent(); 683 } 684 MSI->eraseFromParent(); 685 continue; 686 } 687 688 // If this is a memcpy or memmove into or out of the whole allocation, we 689 // can handle it like a load or store of the scalar type. 690 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 691 assert(Offset == 0 && "must be store to start of alloca"); 692 assert(!NonConstantIdx && "Cannot replace dynamic transfer with insert"); 693 694 // If the source and destination are both to the same alloca, then this is 695 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 696 // as appropriate. 697 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, &DL, 0)); 698 699 if (GetUnderlyingObject(MTI->getSource(), &DL, 0) != OrigAI) { 700 // Dest must be OrigAI, change this to be a load from the original 701 // pointer (bitcasted), then a store to our new alloca. 702 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 703 Value *SrcPtr = MTI->getSource(); 704 PointerType* SPTy = cast<PointerType>(SrcPtr->getType()); 705 PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 706 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 707 AIPTy = PointerType::get(AIPTy->getElementType(), 708 SPTy->getAddressSpace()); 709 } 710 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy); 711 712 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 713 SrcVal->setAlignment(MTI->getAlignment()); 714 Builder.CreateStore(SrcVal, NewAI); 715 } else if (GetUnderlyingObject(MTI->getDest(), &DL, 0) != OrigAI) { 716 // Src must be OrigAI, change this to be a load from NewAI then a store 717 // through the original dest pointer (bitcasted). 718 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 719 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 720 721 PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType()); 722 PointerType* AIPTy = cast<PointerType>(NewAI->getType()); 723 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) { 724 AIPTy = PointerType::get(AIPTy->getElementType(), 725 DPTy->getAddressSpace()); 726 } 727 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy); 728 729 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 730 NewStore->setAlignment(MTI->getAlignment()); 731 } else { 732 // Noop transfer. Src == Dst 733 } 734 735 MTI->eraseFromParent(); 736 continue; 737 } 738 739 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 740 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 741 II->getIntrinsicID() == Intrinsic::lifetime_end) { 742 // There's no need to preserve these, as the resulting alloca will be 743 // converted to a register anyways. 744 II->eraseFromParent(); 745 continue; 746 } 747 } 748 749 llvm_unreachable("Unsupported operation!"); 750 } 751 } 752 753 /// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 754 /// or vector value FromVal, extracting the bits from the offset specified by 755 /// Offset. This returns the value, which is of type ToType. 756 /// 757 /// This happens when we are converting an "integer union" to a single 758 /// integer scalar, or when we are converting a "vector union" to a vector with 759 /// insert/extractelement instructions. 760 /// 761 /// Offset is an offset from the original alloca, in bits that need to be 762 /// shifted to the right. 763 Value *ConvertToScalarInfo:: 764 ConvertScalar_ExtractValue(Value *FromVal, Type *ToType, 765 uint64_t Offset, Value* NonConstantIdx, 766 IRBuilder<> &Builder) { 767 // If the load is of the whole new alloca, no conversion is needed. 768 Type *FromType = FromVal->getType(); 769 if (FromType == ToType && Offset == 0) 770 return FromVal; 771 772 // If the result alloca is a vector type, this is either an element 773 // access or a bitcast to another vector type of the same size. 774 if (VectorType *VTy = dyn_cast<VectorType>(FromType)) { 775 unsigned FromTypeSize = DL.getTypeAllocSize(FromType); 776 unsigned ToTypeSize = DL.getTypeAllocSize(ToType); 777 if (FromTypeSize == ToTypeSize) 778 return Builder.CreateBitCast(FromVal, ToType); 779 780 // Otherwise it must be an element access. 781 unsigned Elt = 0; 782 if (Offset) { 783 unsigned EltSize = DL.getTypeAllocSizeInBits(VTy->getElementType()); 784 Elt = Offset/EltSize; 785 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 786 } 787 // Return the element extracted out of it. 788 Value *Idx; 789 if (NonConstantIdx) { 790 if (Elt) 791 Idx = Builder.CreateAdd(NonConstantIdx, 792 Builder.getInt32(Elt), 793 "dyn.offset"); 794 else 795 Idx = NonConstantIdx; 796 } else 797 Idx = Builder.getInt32(Elt); 798 Value *V = Builder.CreateExtractElement(FromVal, Idx); 799 if (V->getType() != ToType) 800 V = Builder.CreateBitCast(V, ToType); 801 return V; 802 } 803 804 // If ToType is a first class aggregate, extract out each of the pieces and 805 // use insertvalue's to form the FCA. 806 if (StructType *ST = dyn_cast<StructType>(ToType)) { 807 assert(!NonConstantIdx && 808 "Dynamic indexing into struct types not supported"); 809 const StructLayout &Layout = *DL.getStructLayout(ST); 810 Value *Res = UndefValue::get(ST); 811 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 812 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 813 Offset+Layout.getElementOffsetInBits(i), 814 nullptr, Builder); 815 Res = Builder.CreateInsertValue(Res, Elt, i); 816 } 817 return Res; 818 } 819 820 if (ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 821 assert(!NonConstantIdx && 822 "Dynamic indexing into array types not supported"); 823 uint64_t EltSize = DL.getTypeAllocSizeInBits(AT->getElementType()); 824 Value *Res = UndefValue::get(AT); 825 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 826 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 827 Offset+i*EltSize, nullptr, 828 Builder); 829 Res = Builder.CreateInsertValue(Res, Elt, i); 830 } 831 return Res; 832 } 833 834 // Otherwise, this must be a union that was converted to an integer value. 835 IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 836 837 // If this is a big-endian system and the load is narrower than the 838 // full alloca type, we need to do a shift to get the right bits. 839 int ShAmt = 0; 840 if (DL.isBigEndian()) { 841 // On big-endian machines, the lowest bit is stored at the bit offset 842 // from the pointer given by getTypeStoreSizeInBits. This matters for 843 // integers with a bitwidth that is not a multiple of 8. 844 ShAmt = DL.getTypeStoreSizeInBits(NTy) - 845 DL.getTypeStoreSizeInBits(ToType) - Offset; 846 } else { 847 ShAmt = Offset; 848 } 849 850 // Note: we support negative bitwidths (with shl) which are not defined. 851 // We do this to support (f.e.) loads off the end of a structure where 852 // only some bits are used. 853 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 854 FromVal = Builder.CreateLShr(FromVal, 855 ConstantInt::get(FromVal->getType(), ShAmt)); 856 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 857 FromVal = Builder.CreateShl(FromVal, 858 ConstantInt::get(FromVal->getType(), -ShAmt)); 859 860 // Finally, unconditionally truncate the integer to the right width. 861 unsigned LIBitWidth = DL.getTypeSizeInBits(ToType); 862 if (LIBitWidth < NTy->getBitWidth()) 863 FromVal = 864 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 865 LIBitWidth)); 866 else if (LIBitWidth > NTy->getBitWidth()) 867 FromVal = 868 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 869 LIBitWidth)); 870 871 // If the result is an integer, this is a trunc or bitcast. 872 if (ToType->isIntegerTy()) { 873 // Should be done. 874 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 875 // Just do a bitcast, we know the sizes match up. 876 FromVal = Builder.CreateBitCast(FromVal, ToType); 877 } else { 878 // Otherwise must be a pointer. 879 FromVal = Builder.CreateIntToPtr(FromVal, ToType); 880 } 881 assert(FromVal->getType() == ToType && "Didn't convert right?"); 882 return FromVal; 883 } 884 885 /// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 886 /// or vector value "Old" at the offset specified by Offset. 887 /// 888 /// This happens when we are converting an "integer union" to a 889 /// single integer scalar, or when we are converting a "vector union" to a 890 /// vector with insert/extractelement instructions. 891 /// 892 /// Offset is an offset from the original alloca, in bits that need to be 893 /// shifted to the right. 894 /// 895 /// NonConstantIdx is an index value if there was a GEP with a non-constant 896 /// index value. If this is 0 then all GEPs used to find this insert address 897 /// are constant. 898 Value *ConvertToScalarInfo:: 899 ConvertScalar_InsertValue(Value *SV, Value *Old, 900 uint64_t Offset, Value* NonConstantIdx, 901 IRBuilder<> &Builder) { 902 // Convert the stored type to the actual type, shift it left to insert 903 // then 'or' into place. 904 Type *AllocaType = Old->getType(); 905 LLVMContext &Context = Old->getContext(); 906 907 if (VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 908 uint64_t VecSize = DL.getTypeAllocSizeInBits(VTy); 909 uint64_t ValSize = DL.getTypeAllocSizeInBits(SV->getType()); 910 911 // Changing the whole vector with memset or with an access of a different 912 // vector type? 913 if (ValSize == VecSize) 914 return Builder.CreateBitCast(SV, AllocaType); 915 916 // Must be an element insertion. 917 Type *EltTy = VTy->getElementType(); 918 if (SV->getType() != EltTy) 919 SV = Builder.CreateBitCast(SV, EltTy); 920 uint64_t EltSize = DL.getTypeAllocSizeInBits(EltTy); 921 unsigned Elt = Offset/EltSize; 922 Value *Idx; 923 if (NonConstantIdx) { 924 if (Elt) 925 Idx = Builder.CreateAdd(NonConstantIdx, 926 Builder.getInt32(Elt), 927 "dyn.offset"); 928 else 929 Idx = NonConstantIdx; 930 } else 931 Idx = Builder.getInt32(Elt); 932 return Builder.CreateInsertElement(Old, SV, Idx); 933 } 934 935 // If SV is a first-class aggregate value, insert each value recursively. 936 if (StructType *ST = dyn_cast<StructType>(SV->getType())) { 937 assert(!NonConstantIdx && 938 "Dynamic indexing into struct types not supported"); 939 const StructLayout &Layout = *DL.getStructLayout(ST); 940 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 941 Value *Elt = Builder.CreateExtractValue(SV, i); 942 Old = ConvertScalar_InsertValue(Elt, Old, 943 Offset+Layout.getElementOffsetInBits(i), 944 nullptr, Builder); 945 } 946 return Old; 947 } 948 949 if (ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 950 assert(!NonConstantIdx && 951 "Dynamic indexing into array types not supported"); 952 uint64_t EltSize = DL.getTypeAllocSizeInBits(AT->getElementType()); 953 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 954 Value *Elt = Builder.CreateExtractValue(SV, i); 955 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, nullptr, 956 Builder); 957 } 958 return Old; 959 } 960 961 // If SV is a float, convert it to the appropriate integer type. 962 // If it is a pointer, do the same. 963 unsigned SrcWidth = DL.getTypeSizeInBits(SV->getType()); 964 unsigned DestWidth = DL.getTypeSizeInBits(AllocaType); 965 unsigned SrcStoreWidth = DL.getTypeStoreSizeInBits(SV->getType()); 966 unsigned DestStoreWidth = DL.getTypeStoreSizeInBits(AllocaType); 967 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 968 SV = Builder.CreateBitCast(SV, IntegerType::get(SV->getContext(),SrcWidth)); 969 else if (SV->getType()->isPointerTy()) 970 SV = Builder.CreatePtrToInt(SV, DL.getIntPtrType(SV->getType())); 971 972 // Zero extend or truncate the value if needed. 973 if (SV->getType() != AllocaType) { 974 if (SV->getType()->getPrimitiveSizeInBits() < 975 AllocaType->getPrimitiveSizeInBits()) 976 SV = Builder.CreateZExt(SV, AllocaType); 977 else { 978 // Truncation may be needed if storing more than the alloca can hold 979 // (undefined behavior). 980 SV = Builder.CreateTrunc(SV, AllocaType); 981 SrcWidth = DestWidth; 982 SrcStoreWidth = DestStoreWidth; 983 } 984 } 985 986 // If this is a big-endian system and the store is narrower than the 987 // full alloca type, we need to do a shift to get the right bits. 988 int ShAmt = 0; 989 if (DL.isBigEndian()) { 990 // On big-endian machines, the lowest bit is stored at the bit offset 991 // from the pointer given by getTypeStoreSizeInBits. This matters for 992 // integers with a bitwidth that is not a multiple of 8. 993 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 994 } else { 995 ShAmt = Offset; 996 } 997 998 // Note: we support negative bitwidths (with shr) which are not defined. 999 // We do this to support (f.e.) stores off the end of a structure where 1000 // only some bits in the structure are set. 1001 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 1002 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 1003 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), ShAmt)); 1004 Mask <<= ShAmt; 1005 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1006 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), -ShAmt)); 1007 Mask = Mask.lshr(-ShAmt); 1008 } 1009 1010 // Mask out the bits we are about to insert from the old value, and or 1011 // in the new bits. 1012 if (SrcWidth != DestWidth) { 1013 assert(DestWidth > SrcWidth); 1014 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 1015 SV = Builder.CreateOr(Old, SV, "ins"); 1016 } 1017 return SV; 1018 } 1019 1020 1021 //===----------------------------------------------------------------------===// 1022 // SRoA Driver 1023 //===----------------------------------------------------------------------===// 1024 1025 1026 bool SROA::runOnFunction(Function &F) { 1027 if (skipOptnoneFunction(F)) 1028 return false; 1029 1030 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>(); 1031 DL = DLP ? &DLP->getDataLayout() : nullptr; 1032 1033 bool Changed = performPromotion(F); 1034 1035 // FIXME: ScalarRepl currently depends on DataLayout more than it 1036 // theoretically needs to. It should be refactored in order to support 1037 // target-independent IR. Until this is done, just skip the actual 1038 // scalar-replacement portion of this pass. 1039 if (!DL) return Changed; 1040 1041 while (1) { 1042 bool LocalChange = performScalarRepl(F); 1043 if (!LocalChange) break; // No need to repromote if no scalarrepl 1044 Changed = true; 1045 LocalChange = performPromotion(F); 1046 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 1047 } 1048 1049 return Changed; 1050 } 1051 1052 namespace { 1053 class AllocaPromoter : public LoadAndStorePromoter { 1054 AllocaInst *AI; 1055 DIBuilder *DIB; 1056 SmallVector<DbgDeclareInst *, 4> DDIs; 1057 SmallVector<DbgValueInst *, 4> DVIs; 1058 public: 1059 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S, 1060 DIBuilder *DB) 1061 : LoadAndStorePromoter(Insts, S), AI(nullptr), DIB(DB) {} 1062 1063 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) { 1064 // Remember which alloca we're promoting (for isInstInList). 1065 this->AI = AI; 1066 if (MDNode *DebugNode = MDNode::getIfExists(AI->getContext(), AI)) { 1067 for (User *U : DebugNode->users()) 1068 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(U)) 1069 DDIs.push_back(DDI); 1070 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(U)) 1071 DVIs.push_back(DVI); 1072 } 1073 1074 LoadAndStorePromoter::run(Insts); 1075 AI->eraseFromParent(); 1076 for (SmallVectorImpl<DbgDeclareInst *>::iterator I = DDIs.begin(), 1077 E = DDIs.end(); I != E; ++I) { 1078 DbgDeclareInst *DDI = *I; 1079 DDI->eraseFromParent(); 1080 } 1081 for (SmallVectorImpl<DbgValueInst *>::iterator I = DVIs.begin(), 1082 E = DVIs.end(); I != E; ++I) { 1083 DbgValueInst *DVI = *I; 1084 DVI->eraseFromParent(); 1085 } 1086 } 1087 1088 bool isInstInList(Instruction *I, 1089 const SmallVectorImpl<Instruction*> &Insts) const override { 1090 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 1091 return LI->getOperand(0) == AI; 1092 return cast<StoreInst>(I)->getPointerOperand() == AI; 1093 } 1094 1095 void updateDebugInfo(Instruction *Inst) const override { 1096 for (SmallVectorImpl<DbgDeclareInst *>::const_iterator I = DDIs.begin(), 1097 E = DDIs.end(); I != E; ++I) { 1098 DbgDeclareInst *DDI = *I; 1099 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) 1100 ConvertDebugDeclareToDebugValue(DDI, SI, *DIB); 1101 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) 1102 ConvertDebugDeclareToDebugValue(DDI, LI, *DIB); 1103 } 1104 for (SmallVectorImpl<DbgValueInst *>::const_iterator I = DVIs.begin(), 1105 E = DVIs.end(); I != E; ++I) { 1106 DbgValueInst *DVI = *I; 1107 Value *Arg = nullptr; 1108 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 1109 // If an argument is zero extended then use argument directly. The ZExt 1110 // may be zapped by an optimization pass in future. 1111 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0))) 1112 Arg = dyn_cast<Argument>(ZExt->getOperand(0)); 1113 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0))) 1114 Arg = dyn_cast<Argument>(SExt->getOperand(0)); 1115 if (!Arg) 1116 Arg = SI->getOperand(0); 1117 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 1118 Arg = LI->getOperand(0); 1119 } else { 1120 continue; 1121 } 1122 Instruction *DbgVal = 1123 DIB->insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()), 1124 Inst); 1125 DbgVal->setDebugLoc(DVI->getDebugLoc()); 1126 } 1127 } 1128 }; 1129 } // end anon namespace 1130 1131 /// isSafeSelectToSpeculate - Select instructions that use an alloca and are 1132 /// subsequently loaded can be rewritten to load both input pointers and then 1133 /// select between the result, allowing the load of the alloca to be promoted. 1134 /// From this: 1135 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other 1136 /// %V = load i32* %P2 1137 /// to: 1138 /// %V1 = load i32* %Alloca -> will be mem2reg'd 1139 /// %V2 = load i32* %Other 1140 /// %V = select i1 %cond, i32 %V1, i32 %V2 1141 /// 1142 /// We can do this to a select if its only uses are loads and if the operand to 1143 /// the select can be loaded unconditionally. 1144 static bool isSafeSelectToSpeculate(SelectInst *SI, const DataLayout *DL) { 1145 bool TDerefable = SI->getTrueValue()->isDereferenceablePointer(DL); 1146 bool FDerefable = SI->getFalseValue()->isDereferenceablePointer(DL); 1147 1148 for (User *U : SI->users()) { 1149 LoadInst *LI = dyn_cast<LoadInst>(U); 1150 if (!LI || !LI->isSimple()) return false; 1151 1152 // Both operands to the select need to be dereferencable, either absolutely 1153 // (e.g. allocas) or at this point because we can see other accesses to it. 1154 if (!TDerefable && !isSafeToLoadUnconditionally(SI->getTrueValue(), LI, 1155 LI->getAlignment(), DL)) 1156 return false; 1157 if (!FDerefable && !isSafeToLoadUnconditionally(SI->getFalseValue(), LI, 1158 LI->getAlignment(), DL)) 1159 return false; 1160 } 1161 1162 return true; 1163 } 1164 1165 /// isSafePHIToSpeculate - PHI instructions that use an alloca and are 1166 /// subsequently loaded can be rewritten to load both input pointers in the pred 1167 /// blocks and then PHI the results, allowing the load of the alloca to be 1168 /// promoted. 1169 /// From this: 1170 /// %P2 = phi [i32* %Alloca, i32* %Other] 1171 /// %V = load i32* %P2 1172 /// to: 1173 /// %V1 = load i32* %Alloca -> will be mem2reg'd 1174 /// ... 1175 /// %V2 = load i32* %Other 1176 /// ... 1177 /// %V = phi [i32 %V1, i32 %V2] 1178 /// 1179 /// We can do this to a select if its only uses are loads and if the operand to 1180 /// the select can be loaded unconditionally. 1181 static bool isSafePHIToSpeculate(PHINode *PN, const DataLayout *DL) { 1182 // For now, we can only do this promotion if the load is in the same block as 1183 // the PHI, and if there are no stores between the phi and load. 1184 // TODO: Allow recursive phi users. 1185 // TODO: Allow stores. 1186 BasicBlock *BB = PN->getParent(); 1187 unsigned MaxAlign = 0; 1188 for (User *U : PN->users()) { 1189 LoadInst *LI = dyn_cast<LoadInst>(U); 1190 if (!LI || !LI->isSimple()) return false; 1191 1192 // For now we only allow loads in the same block as the PHI. This is a 1193 // common case that happens when instcombine merges two loads through a PHI. 1194 if (LI->getParent() != BB) return false; 1195 1196 // Ensure that there are no instructions between the PHI and the load that 1197 // could store. 1198 for (BasicBlock::iterator BBI = PN; &*BBI != LI; ++BBI) 1199 if (BBI->mayWriteToMemory()) 1200 return false; 1201 1202 MaxAlign = std::max(MaxAlign, LI->getAlignment()); 1203 } 1204 1205 // Okay, we know that we have one or more loads in the same block as the PHI. 1206 // We can transform this if it is safe to push the loads into the predecessor 1207 // blocks. The only thing to watch out for is that we can't put a possibly 1208 // trapping load in the predecessor if it is a critical edge. 1209 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1210 BasicBlock *Pred = PN->getIncomingBlock(i); 1211 Value *InVal = PN->getIncomingValue(i); 1212 1213 // If the terminator of the predecessor has side-effects (an invoke), 1214 // there is no safe place to put a load in the predecessor. 1215 if (Pred->getTerminator()->mayHaveSideEffects()) 1216 return false; 1217 1218 // If the value is produced by the terminator of the predecessor 1219 // (an invoke), there is no valid place to put a load in the predecessor. 1220 if (Pred->getTerminator() == InVal) 1221 return false; 1222 1223 // If the predecessor has a single successor, then the edge isn't critical. 1224 if (Pred->getTerminator()->getNumSuccessors() == 1) 1225 continue; 1226 1227 // If this pointer is always safe to load, or if we can prove that there is 1228 // already a load in the block, then we can move the load to the pred block. 1229 if (InVal->isDereferenceablePointer(DL) || 1230 isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign, DL)) 1231 continue; 1232 1233 return false; 1234 } 1235 1236 return true; 1237 } 1238 1239 1240 /// tryToMakeAllocaBePromotable - This returns true if the alloca only has 1241 /// direct (non-volatile) loads and stores to it. If the alloca is close but 1242 /// not quite there, this will transform the code to allow promotion. As such, 1243 /// it is a non-pure predicate. 1244 static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const DataLayout *DL) { 1245 SetVector<Instruction*, SmallVector<Instruction*, 4>, 1246 SmallPtrSet<Instruction*, 4> > InstsToRewrite; 1247 for (User *U : AI->users()) { 1248 if (LoadInst *LI = dyn_cast<LoadInst>(U)) { 1249 if (!LI->isSimple()) 1250 return false; 1251 continue; 1252 } 1253 1254 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 1255 if (SI->getOperand(0) == AI || !SI->isSimple()) 1256 return false; // Don't allow a store OF the AI, only INTO the AI. 1257 continue; 1258 } 1259 1260 if (SelectInst *SI = dyn_cast<SelectInst>(U)) { 1261 // If the condition being selected on is a constant, fold the select, yes 1262 // this does (rarely) happen early on. 1263 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition())) { 1264 Value *Result = SI->getOperand(1+CI->isZero()); 1265 SI->replaceAllUsesWith(Result); 1266 SI->eraseFromParent(); 1267 1268 // This is very rare and we just scrambled the use list of AI, start 1269 // over completely. 1270 return tryToMakeAllocaBePromotable(AI, DL); 1271 } 1272 1273 // If it is safe to turn "load (select c, AI, ptr)" into a select of two 1274 // loads, then we can transform this by rewriting the select. 1275 if (!isSafeSelectToSpeculate(SI, DL)) 1276 return false; 1277 1278 InstsToRewrite.insert(SI); 1279 continue; 1280 } 1281 1282 if (PHINode *PN = dyn_cast<PHINode>(U)) { 1283 if (PN->use_empty()) { // Dead PHIs can be stripped. 1284 InstsToRewrite.insert(PN); 1285 continue; 1286 } 1287 1288 // If it is safe to turn "load (phi [AI, ptr, ...])" into a PHI of loads 1289 // in the pred blocks, then we can transform this by rewriting the PHI. 1290 if (!isSafePHIToSpeculate(PN, DL)) 1291 return false; 1292 1293 InstsToRewrite.insert(PN); 1294 continue; 1295 } 1296 1297 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 1298 if (onlyUsedByLifetimeMarkers(BCI)) { 1299 InstsToRewrite.insert(BCI); 1300 continue; 1301 } 1302 } 1303 1304 return false; 1305 } 1306 1307 // If there are no instructions to rewrite, then all uses are load/stores and 1308 // we're done! 1309 if (InstsToRewrite.empty()) 1310 return true; 1311 1312 // If we have instructions that need to be rewritten for this to be promotable 1313 // take care of it now. 1314 for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) { 1315 if (BitCastInst *BCI = dyn_cast<BitCastInst>(InstsToRewrite[i])) { 1316 // This could only be a bitcast used by nothing but lifetime intrinsics. 1317 for (BitCastInst::user_iterator I = BCI->user_begin(), E = BCI->user_end(); 1318 I != E;) 1319 cast<Instruction>(*I++)->eraseFromParent(); 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->user_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->user_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->user_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) { 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 1405 bool SROA::performPromotion(Function &F) { 1406 std::vector<AllocaInst*> Allocas; 1407 DominatorTree *DT = nullptr; 1408 if (HasDomTree) 1409 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 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, DL)) 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 (User *U : AI->users()) 1436 Insts.push_back(cast<Instruction>(U)); 1437 AllocaPromoter(Insts, SSA, &DIB).run(AI, Insts); 1438 Insts.clear(); 1439 } 1440 } 1441 NumPromoted += Allocas.size(); 1442 Changed = true; 1443 } 1444 1445 return Changed; 1446 } 1447 1448 1449 /// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 1450 /// SROA. It must be a struct or array type with a small number of elements. 1451 bool SROA::ShouldAttemptScalarRepl(AllocaInst *AI) { 1452 Type *T = AI->getAllocatedType(); 1453 // Do not promote any struct that has too many members. 1454 if (StructType *ST = dyn_cast<StructType>(T)) 1455 return ST->getNumElements() <= StructMemberThreshold; 1456 // Do not promote any array that has too many elements. 1457 if (ArrayType *AT = dyn_cast<ArrayType>(T)) 1458 return AT->getNumElements() <= ArrayElementThreshold; 1459 return false; 1460 } 1461 1462 // performScalarRepl - This algorithm is a simple worklist driven algorithm, 1463 // which runs on all of the alloca instructions in the entry block, removing 1464 // them if they are only used by getelementptr instructions. 1465 // 1466 bool SROA::performScalarRepl(Function &F) { 1467 std::vector<AllocaInst*> WorkList; 1468 1469 // Scan the entry basic block, adding allocas to the worklist. 1470 BasicBlock &BB = F.getEntryBlock(); 1471 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 1472 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 1473 WorkList.push_back(A); 1474 1475 // Process the worklist 1476 bool Changed = false; 1477 while (!WorkList.empty()) { 1478 AllocaInst *AI = WorkList.back(); 1479 WorkList.pop_back(); 1480 1481 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 1482 // with unused elements. 1483 if (AI->use_empty()) { 1484 AI->eraseFromParent(); 1485 Changed = true; 1486 continue; 1487 } 1488 1489 // If this alloca is impossible for us to promote, reject it early. 1490 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 1491 continue; 1492 1493 // Check to see if we can perform the core SROA transformation. We cannot 1494 // transform the allocation instruction if it is an array allocation 1495 // (allocations OF arrays are ok though), and an allocation of a scalar 1496 // value cannot be decomposed at all. 1497 uint64_t AllocaSize = DL->getTypeAllocSize(AI->getAllocatedType()); 1498 1499 // Do not promote [0 x %struct]. 1500 if (AllocaSize == 0) continue; 1501 1502 // Do not promote any struct whose size is too big. 1503 if (AllocaSize > SRThreshold) continue; 1504 1505 // If the alloca looks like a good candidate for scalar replacement, and if 1506 // all its users can be transformed, then split up the aggregate into its 1507 // separate elements. 1508 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 1509 DoScalarReplacement(AI, WorkList); 1510 Changed = true; 1511 continue; 1512 } 1513 1514 // If we can turn this aggregate value (potentially with casts) into a 1515 // simple scalar value that can be mem2reg'd into a register value. 1516 // IsNotTrivial tracks whether this is something that mem2reg could have 1517 // promoted itself. If so, we don't want to transform it needlessly. Note 1518 // that we can't just check based on the type: the alloca may be of an i32 1519 // but that has pointer arithmetic to set byte 3 of it or something. 1520 if (AllocaInst *NewAI = ConvertToScalarInfo( 1521 (unsigned)AllocaSize, *DL, ScalarLoadThreshold).TryConvert(AI)) { 1522 NewAI->takeName(AI); 1523 AI->eraseFromParent(); 1524 ++NumConverted; 1525 Changed = true; 1526 continue; 1527 } 1528 1529 // Otherwise, couldn't process this alloca. 1530 } 1531 1532 return Changed; 1533 } 1534 1535 /// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 1536 /// predicate, do SROA now. 1537 void SROA::DoScalarReplacement(AllocaInst *AI, 1538 std::vector<AllocaInst*> &WorkList) { 1539 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 1540 SmallVector<AllocaInst*, 32> ElementAllocas; 1541 if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 1542 ElementAllocas.reserve(ST->getNumContainedTypes()); 1543 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 1544 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), nullptr, 1545 AI->getAlignment(), 1546 AI->getName() + "." + Twine(i), AI); 1547 ElementAllocas.push_back(NA); 1548 WorkList.push_back(NA); // Add to worklist for recursive processing 1549 } 1550 } else { 1551 ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 1552 ElementAllocas.reserve(AT->getNumElements()); 1553 Type *ElTy = AT->getElementType(); 1554 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1555 AllocaInst *NA = new AllocaInst(ElTy, nullptr, AI->getAlignment(), 1556 AI->getName() + "." + Twine(i), AI); 1557 ElementAllocas.push_back(NA); 1558 WorkList.push_back(NA); // Add to worklist for recursive processing 1559 } 1560 } 1561 1562 // Now that we have created the new alloca instructions, rewrite all the 1563 // uses of the old alloca. 1564 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 1565 1566 // Now erase any instructions that were made dead while rewriting the alloca. 1567 DeleteDeadInstructions(); 1568 AI->eraseFromParent(); 1569 1570 ++NumReplaced; 1571 } 1572 1573 /// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 1574 /// recursively including all their operands that become trivially dead. 1575 void SROA::DeleteDeadInstructions() { 1576 while (!DeadInsts.empty()) { 1577 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 1578 1579 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 1580 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 1581 // Zero out the operand and see if it becomes trivially dead. 1582 // (But, don't add allocas to the dead instruction list -- they are 1583 // already on the worklist and will be deleted separately.) 1584 *OI = nullptr; 1585 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 1586 DeadInsts.push_back(U); 1587 } 1588 1589 I->eraseFromParent(); 1590 } 1591 } 1592 1593 /// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 1594 /// performing scalar replacement of alloca AI. The results are flagged in 1595 /// the Info parameter. Offset indicates the position within AI that is 1596 /// referenced by this instruction. 1597 void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset, 1598 AllocaInfo &Info) { 1599 for (Use &U : I->uses()) { 1600 Instruction *User = cast<Instruction>(U.getUser()); 1601 1602 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1603 isSafeForScalarRepl(BC, Offset, Info); 1604 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1605 uint64_t GEPOffset = Offset; 1606 isSafeGEP(GEPI, GEPOffset, Info); 1607 if (!Info.isUnsafe) 1608 isSafeForScalarRepl(GEPI, GEPOffset, Info); 1609 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1610 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1611 if (!Length || Length->isNegative()) 1612 return MarkUnsafe(Info, User); 1613 1614 isSafeMemAccess(Offset, Length->getZExtValue(), nullptr, 1615 U.getOperandNo() == 0, Info, MI, 1616 true /*AllowWholeAccess*/); 1617 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1618 if (!LI->isSimple()) 1619 return MarkUnsafe(Info, User); 1620 Type *LIType = LI->getType(); 1621 isSafeMemAccess(Offset, DL->getTypeAllocSize(LIType), 1622 LIType, false, Info, LI, true /*AllowWholeAccess*/); 1623 Info.hasALoadOrStore = true; 1624 1625 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1626 // Store is ok if storing INTO the pointer, not storing the pointer 1627 if (!SI->isSimple() || SI->getOperand(0) == I) 1628 return MarkUnsafe(Info, User); 1629 1630 Type *SIType = SI->getOperand(0)->getType(); 1631 isSafeMemAccess(Offset, DL->getTypeAllocSize(SIType), 1632 SIType, true, Info, SI, true /*AllowWholeAccess*/); 1633 Info.hasALoadOrStore = true; 1634 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 1635 if (II->getIntrinsicID() != Intrinsic::lifetime_start && 1636 II->getIntrinsicID() != Intrinsic::lifetime_end) 1637 return MarkUnsafe(Info, User); 1638 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) { 1639 isSafePHISelectUseForScalarRepl(User, Offset, Info); 1640 } else { 1641 return MarkUnsafe(Info, User); 1642 } 1643 if (Info.isUnsafe) return; 1644 } 1645 } 1646 1647 1648 /// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer 1649 /// derived from the alloca, we can often still split the alloca into elements. 1650 /// This is useful if we have a large alloca where one element is phi'd 1651 /// together somewhere: we can SRoA and promote all the other elements even if 1652 /// we end up not being able to promote this one. 1653 /// 1654 /// All we require is that the uses of the PHI do not index into other parts of 1655 /// the alloca. The most important use case for this is single load and stores 1656 /// that are PHI'd together, which can happen due to code sinking. 1657 void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset, 1658 AllocaInfo &Info) { 1659 // If we've already checked this PHI, don't do it again. 1660 if (PHINode *PN = dyn_cast<PHINode>(I)) 1661 if (!Info.CheckedPHIs.insert(PN)) 1662 return; 1663 1664 for (User *U : I->users()) { 1665 Instruction *UI = cast<Instruction>(U); 1666 1667 if (BitCastInst *BC = dyn_cast<BitCastInst>(UI)) { 1668 isSafePHISelectUseForScalarRepl(BC, Offset, Info); 1669 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(UI)) { 1670 // Only allow "bitcast" GEPs for simplicity. We could generalize this, 1671 // but would have to prove that we're staying inside of an element being 1672 // promoted. 1673 if (!GEPI->hasAllZeroIndices()) 1674 return MarkUnsafe(Info, UI); 1675 isSafePHISelectUseForScalarRepl(GEPI, Offset, Info); 1676 } else if (LoadInst *LI = dyn_cast<LoadInst>(UI)) { 1677 if (!LI->isSimple()) 1678 return MarkUnsafe(Info, UI); 1679 Type *LIType = LI->getType(); 1680 isSafeMemAccess(Offset, DL->getTypeAllocSize(LIType), 1681 LIType, false, Info, LI, false /*AllowWholeAccess*/); 1682 Info.hasALoadOrStore = true; 1683 1684 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI)) { 1685 // Store is ok if storing INTO the pointer, not storing the pointer 1686 if (!SI->isSimple() || SI->getOperand(0) == I) 1687 return MarkUnsafe(Info, UI); 1688 1689 Type *SIType = SI->getOperand(0)->getType(); 1690 isSafeMemAccess(Offset, DL->getTypeAllocSize(SIType), 1691 SIType, true, Info, SI, false /*AllowWholeAccess*/); 1692 Info.hasALoadOrStore = true; 1693 } else if (isa<PHINode>(UI) || isa<SelectInst>(UI)) { 1694 isSafePHISelectUseForScalarRepl(UI, Offset, Info); 1695 } else { 1696 return MarkUnsafe(Info, UI); 1697 } 1698 if (Info.isUnsafe) return; 1699 } 1700 } 1701 1702 /// isSafeGEP - Check if a GEP instruction can be handled for scalar 1703 /// replacement. It is safe when all the indices are constant, in-bounds 1704 /// references, and when the resulting offset corresponds to an element within 1705 /// the alloca type. The results are flagged in the Info parameter. Upon 1706 /// return, Offset is adjusted as specified by the GEP indices. 1707 void SROA::isSafeGEP(GetElementPtrInst *GEPI, 1708 uint64_t &Offset, AllocaInfo &Info) { 1709 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 1710 if (GEPIt == E) 1711 return; 1712 bool NonConstant = false; 1713 unsigned NonConstantIdxSize = 0; 1714 1715 // Walk through the GEP type indices, checking the types that this indexes 1716 // into. 1717 for (; GEPIt != E; ++GEPIt) { 1718 // Ignore struct elements, no extra checking needed for these. 1719 if ((*GEPIt)->isStructTy()) 1720 continue; 1721 1722 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 1723 if (!IdxVal) 1724 return MarkUnsafe(Info, GEPI); 1725 } 1726 1727 // Compute the offset due to this GEP and check if the alloca has a 1728 // component element at that offset. 1729 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 1730 // If this GEP is non-constant then the last operand must have been a 1731 // dynamic index into a vector. Pop this now as it has no impact on the 1732 // constant part of the offset. 1733 if (NonConstant) 1734 Indices.pop_back(); 1735 Offset += DL->getIndexedOffset(GEPI->getPointerOperandType(), Indices); 1736 if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 1737 NonConstantIdxSize)) 1738 MarkUnsafe(Info, GEPI); 1739 } 1740 1741 /// isHomogeneousAggregate - Check if type T is a struct or array containing 1742 /// elements of the same type (which is always true for arrays). If so, 1743 /// return true with NumElts and EltTy set to the number of elements and the 1744 /// element type, respectively. 1745 static bool isHomogeneousAggregate(Type *T, unsigned &NumElts, 1746 Type *&EltTy) { 1747 if (ArrayType *AT = dyn_cast<ArrayType>(T)) { 1748 NumElts = AT->getNumElements(); 1749 EltTy = (NumElts == 0 ? nullptr : AT->getElementType()); 1750 return true; 1751 } 1752 if (StructType *ST = dyn_cast<StructType>(T)) { 1753 NumElts = ST->getNumContainedTypes(); 1754 EltTy = (NumElts == 0 ? nullptr : ST->getContainedType(0)); 1755 for (unsigned n = 1; n < NumElts; ++n) { 1756 if (ST->getContainedType(n) != EltTy) 1757 return false; 1758 } 1759 return true; 1760 } 1761 return false; 1762 } 1763 1764 /// isCompatibleAggregate - Check if T1 and T2 are either the same type or are 1765 /// "homogeneous" aggregates with the same element type and number of elements. 1766 static bool isCompatibleAggregate(Type *T1, Type *T2) { 1767 if (T1 == T2) 1768 return true; 1769 1770 unsigned NumElts1, NumElts2; 1771 Type *EltTy1, *EltTy2; 1772 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) && 1773 isHomogeneousAggregate(T2, NumElts2, EltTy2) && 1774 NumElts1 == NumElts2 && 1775 EltTy1 == EltTy2) 1776 return true; 1777 1778 return false; 1779 } 1780 1781 /// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 1782 /// alloca or has an offset and size that corresponds to a component element 1783 /// within it. The offset checked here may have been formed from a GEP with a 1784 /// pointer bitcasted to a different type. 1785 /// 1786 /// If AllowWholeAccess is true, then this allows uses of the entire alloca as a 1787 /// unit. If false, it only allows accesses known to be in a single element. 1788 void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize, 1789 Type *MemOpType, bool isStore, 1790 AllocaInfo &Info, Instruction *TheAccess, 1791 bool AllowWholeAccess) { 1792 // Check if this is a load/store of the entire alloca. 1793 if (Offset == 0 && AllowWholeAccess && 1794 MemSize == DL->getTypeAllocSize(Info.AI->getAllocatedType())) { 1795 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer 1796 // loads/stores (which are essentially the same as the MemIntrinsics with 1797 // regard to copying padding between elements). But, if an alloca is 1798 // flagged as both a source and destination of such operations, we'll need 1799 // to check later for padding between elements. 1800 if (!MemOpType || MemOpType->isIntegerTy()) { 1801 if (isStore) 1802 Info.isMemCpyDst = true; 1803 else 1804 Info.isMemCpySrc = true; 1805 return; 1806 } 1807 // This is also safe for references using a type that is compatible with 1808 // the type of the alloca, so that loads/stores can be rewritten using 1809 // insertvalue/extractvalue. 1810 if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) { 1811 Info.hasSubelementAccess = true; 1812 return; 1813 } 1814 } 1815 // Check if the offset/size correspond to a component within the alloca type. 1816 Type *T = Info.AI->getAllocatedType(); 1817 if (TypeHasComponent(T, Offset, MemSize)) { 1818 Info.hasSubelementAccess = true; 1819 return; 1820 } 1821 1822 return MarkUnsafe(Info, TheAccess); 1823 } 1824 1825 /// TypeHasComponent - Return true if T has a component type with the 1826 /// specified offset and size. If Size is zero, do not check the size. 1827 bool SROA::TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size) { 1828 Type *EltTy; 1829 uint64_t EltSize; 1830 if (StructType *ST = dyn_cast<StructType>(T)) { 1831 const StructLayout *Layout = DL->getStructLayout(ST); 1832 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 1833 EltTy = ST->getContainedType(EltIdx); 1834 EltSize = DL->getTypeAllocSize(EltTy); 1835 Offset -= Layout->getElementOffset(EltIdx); 1836 } else if (ArrayType *AT = dyn_cast<ArrayType>(T)) { 1837 EltTy = AT->getElementType(); 1838 EltSize = DL->getTypeAllocSize(EltTy); 1839 if (Offset >= AT->getNumElements() * EltSize) 1840 return false; 1841 Offset %= EltSize; 1842 } else if (VectorType *VT = dyn_cast<VectorType>(T)) { 1843 EltTy = VT->getElementType(); 1844 EltSize = DL->getTypeAllocSize(EltTy); 1845 if (Offset >= VT->getNumElements() * EltSize) 1846 return false; 1847 Offset %= EltSize; 1848 } else { 1849 return false; 1850 } 1851 if (Offset == 0 && (Size == 0 || EltSize == Size)) 1852 return true; 1853 // Check if the component spans multiple elements. 1854 if (Offset + Size > EltSize) 1855 return false; 1856 return TypeHasComponent(EltTy, Offset, Size); 1857 } 1858 1859 /// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 1860 /// the instruction I, which references it, to use the separate elements. 1861 /// Offset indicates the position within AI that is referenced by this 1862 /// instruction. 1863 void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 1864 SmallVectorImpl<AllocaInst *> &NewElts) { 1865 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) { 1866 Use &TheUse = *UI++; 1867 Instruction *User = cast<Instruction>(TheUse.getUser()); 1868 1869 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 1870 RewriteBitCast(BC, AI, Offset, NewElts); 1871 continue; 1872 } 1873 1874 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 1875 RewriteGEP(GEPI, AI, Offset, NewElts); 1876 continue; 1877 } 1878 1879 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 1880 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 1881 uint64_t MemSize = Length->getZExtValue(); 1882 if (Offset == 0 && 1883 MemSize == DL->getTypeAllocSize(AI->getAllocatedType())) 1884 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 1885 // Otherwise the intrinsic can only touch a single element and the 1886 // address operand will be updated, so nothing else needs to be done. 1887 continue; 1888 } 1889 1890 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) { 1891 if (II->getIntrinsicID() == Intrinsic::lifetime_start || 1892 II->getIntrinsicID() == Intrinsic::lifetime_end) { 1893 RewriteLifetimeIntrinsic(II, AI, Offset, NewElts); 1894 } 1895 continue; 1896 } 1897 1898 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1899 Type *LIType = LI->getType(); 1900 1901 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) { 1902 // Replace: 1903 // %res = load { i32, i32 }* %alloc 1904 // with: 1905 // %load.0 = load i32* %alloc.0 1906 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 1907 // %load.1 = load i32* %alloc.1 1908 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 1909 // (Also works for arrays instead of structs) 1910 Value *Insert = UndefValue::get(LIType); 1911 IRBuilder<> Builder(LI); 1912 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1913 Value *Load = Builder.CreateLoad(NewElts[i], "load"); 1914 Insert = Builder.CreateInsertValue(Insert, Load, i, "insert"); 1915 } 1916 LI->replaceAllUsesWith(Insert); 1917 DeadInsts.push_back(LI); 1918 } else if (LIType->isIntegerTy() && 1919 DL->getTypeAllocSize(LIType) == 1920 DL->getTypeAllocSize(AI->getAllocatedType())) { 1921 // If this is a load of the entire alloca to an integer, rewrite it. 1922 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 1923 } 1924 continue; 1925 } 1926 1927 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1928 Value *Val = SI->getOperand(0); 1929 Type *SIType = Val->getType(); 1930 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) { 1931 // Replace: 1932 // store { i32, i32 } %val, { i32, i32 }* %alloc 1933 // with: 1934 // %val.0 = extractvalue { i32, i32 } %val, 0 1935 // store i32 %val.0, i32* %alloc.0 1936 // %val.1 = extractvalue { i32, i32 } %val, 1 1937 // store i32 %val.1, i32* %alloc.1 1938 // (Also works for arrays instead of structs) 1939 IRBuilder<> Builder(SI); 1940 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1941 Value *Extract = Builder.CreateExtractValue(Val, i, Val->getName()); 1942 Builder.CreateStore(Extract, NewElts[i]); 1943 } 1944 DeadInsts.push_back(SI); 1945 } else if (SIType->isIntegerTy() && 1946 DL->getTypeAllocSize(SIType) == 1947 DL->getTypeAllocSize(AI->getAllocatedType())) { 1948 // If this is a store of the entire alloca from an integer, rewrite it. 1949 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 1950 } 1951 continue; 1952 } 1953 1954 if (isa<SelectInst>(User) || isa<PHINode>(User)) { 1955 // If we have a PHI user of the alloca itself (as opposed to a GEP or 1956 // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to 1957 // the new pointer. 1958 if (!isa<AllocaInst>(I)) continue; 1959 1960 assert(Offset == 0 && NewElts[0] && 1961 "Direct alloca use should have a zero offset"); 1962 1963 // If we have a use of the alloca, we know the derived uses will be 1964 // utilizing just the first element of the scalarized result. Insert a 1965 // bitcast of the first alloca before the user as required. 1966 AllocaInst *NewAI = NewElts[0]; 1967 BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI); 1968 NewAI->moveBefore(BCI); 1969 TheUse = BCI; 1970 continue; 1971 } 1972 } 1973 } 1974 1975 /// RewriteBitCast - Update a bitcast reference to the alloca being replaced 1976 /// and recursively continue updating all of its uses. 1977 void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 1978 SmallVectorImpl<AllocaInst *> &NewElts) { 1979 RewriteForScalarRepl(BC, AI, Offset, NewElts); 1980 if (BC->getOperand(0) != AI) 1981 return; 1982 1983 // The bitcast references the original alloca. Replace its uses with 1984 // references to the alloca containing offset zero (which is normally at 1985 // index zero, but might not be in cases involving structs with elements 1986 // of size zero). 1987 Type *T = AI->getAllocatedType(); 1988 uint64_t EltOffset = 0; 1989 Type *IdxTy; 1990 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 1991 Instruction *Val = NewElts[Idx]; 1992 if (Val->getType() != BC->getDestTy()) { 1993 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 1994 Val->takeName(BC); 1995 } 1996 BC->replaceAllUsesWith(Val); 1997 DeadInsts.push_back(BC); 1998 } 1999 2000 /// FindElementAndOffset - Return the index of the element containing Offset 2001 /// within the specified type, which must be either a struct or an array. 2002 /// Sets T to the type of the element and Offset to the offset within that 2003 /// element. IdxTy is set to the type of the index result to be used in a 2004 /// GEP instruction. 2005 uint64_t SROA::FindElementAndOffset(Type *&T, uint64_t &Offset, 2006 Type *&IdxTy) { 2007 uint64_t Idx = 0; 2008 if (StructType *ST = dyn_cast<StructType>(T)) { 2009 const StructLayout *Layout = DL->getStructLayout(ST); 2010 Idx = Layout->getElementContainingOffset(Offset); 2011 T = ST->getContainedType(Idx); 2012 Offset -= Layout->getElementOffset(Idx); 2013 IdxTy = Type::getInt32Ty(T->getContext()); 2014 return Idx; 2015 } else if (ArrayType *AT = dyn_cast<ArrayType>(T)) { 2016 T = AT->getElementType(); 2017 uint64_t EltSize = DL->getTypeAllocSize(T); 2018 Idx = Offset / EltSize; 2019 Offset -= Idx * EltSize; 2020 IdxTy = Type::getInt64Ty(T->getContext()); 2021 return Idx; 2022 } 2023 VectorType *VT = cast<VectorType>(T); 2024 T = VT->getElementType(); 2025 uint64_t EltSize = DL->getTypeAllocSize(T); 2026 Idx = Offset / EltSize; 2027 Offset -= Idx * EltSize; 2028 IdxTy = Type::getInt64Ty(T->getContext()); 2029 return Idx; 2030 } 2031 2032 /// RewriteGEP - Check if this GEP instruction moves the pointer across 2033 /// elements of the alloca that are being split apart, and if so, rewrite 2034 /// the GEP to be relative to the new element. 2035 void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 2036 SmallVectorImpl<AllocaInst *> &NewElts) { 2037 uint64_t OldOffset = Offset; 2038 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 2039 // If the GEP was dynamic then it must have been a dynamic vector lookup. 2040 // In this case, it must be the last GEP operand which is dynamic so keep that 2041 // aside until we've found the constant GEP offset then add it back in at the 2042 // end. 2043 Value* NonConstantIdx = nullptr; 2044 if (!GEPI->hasAllConstantIndices()) 2045 NonConstantIdx = Indices.pop_back_val(); 2046 Offset += DL->getIndexedOffset(GEPI->getPointerOperandType(), Indices); 2047 2048 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 2049 2050 Type *T = AI->getAllocatedType(); 2051 Type *IdxTy; 2052 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 2053 if (GEPI->getOperand(0) == AI) 2054 OldIdx = ~0ULL; // Force the GEP to be rewritten. 2055 2056 T = AI->getAllocatedType(); 2057 uint64_t EltOffset = Offset; 2058 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 2059 2060 // If this GEP does not move the pointer across elements of the alloca 2061 // being split, then it does not needs to be rewritten. 2062 if (Idx == OldIdx) 2063 return; 2064 2065 Type *i32Ty = Type::getInt32Ty(AI->getContext()); 2066 SmallVector<Value*, 8> NewArgs; 2067 NewArgs.push_back(Constant::getNullValue(i32Ty)); 2068 while (EltOffset != 0) { 2069 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 2070 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 2071 } 2072 if (NonConstantIdx) { 2073 Type* GepTy = T; 2074 // This GEP has a dynamic index. We need to add "i32 0" to index through 2075 // any structs or arrays in the original type until we get to the vector 2076 // to index. 2077 while (!isa<VectorType>(GepTy)) { 2078 NewArgs.push_back(Constant::getNullValue(i32Ty)); 2079 GepTy = cast<CompositeType>(GepTy)->getTypeAtIndex(0U); 2080 } 2081 NewArgs.push_back(NonConstantIdx); 2082 } 2083 Instruction *Val = NewElts[Idx]; 2084 if (NewArgs.size() > 1) { 2085 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs, "", GEPI); 2086 Val->takeName(GEPI); 2087 } 2088 if (Val->getType() != GEPI->getType()) 2089 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 2090 GEPI->replaceAllUsesWith(Val); 2091 DeadInsts.push_back(GEPI); 2092 } 2093 2094 /// RewriteLifetimeIntrinsic - II is a lifetime.start/lifetime.end. Rewrite it 2095 /// to mark the lifetime of the scalarized memory. 2096 void SROA::RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI, 2097 uint64_t Offset, 2098 SmallVectorImpl<AllocaInst *> &NewElts) { 2099 ConstantInt *OldSize = cast<ConstantInt>(II->getArgOperand(0)); 2100 // Put matching lifetime markers on everything from Offset up to 2101 // Offset+OldSize. 2102 Type *AIType = AI->getAllocatedType(); 2103 uint64_t NewOffset = Offset; 2104 Type *IdxTy; 2105 uint64_t Idx = FindElementAndOffset(AIType, NewOffset, IdxTy); 2106 2107 IRBuilder<> Builder(II); 2108 uint64_t Size = OldSize->getLimitedValue(); 2109 2110 if (NewOffset) { 2111 // Splice the first element and index 'NewOffset' bytes in. SROA will 2112 // split the alloca again later. 2113 unsigned AS = AI->getType()->getAddressSpace(); 2114 Value *V = Builder.CreateBitCast(NewElts[Idx], Builder.getInt8PtrTy(AS)); 2115 V = Builder.CreateGEP(V, Builder.getInt64(NewOffset)); 2116 2117 IdxTy = NewElts[Idx]->getAllocatedType(); 2118 uint64_t EltSize = DL->getTypeAllocSize(IdxTy) - NewOffset; 2119 if (EltSize > Size) { 2120 EltSize = Size; 2121 Size = 0; 2122 } else { 2123 Size -= EltSize; 2124 } 2125 if (II->getIntrinsicID() == Intrinsic::lifetime_start) 2126 Builder.CreateLifetimeStart(V, Builder.getInt64(EltSize)); 2127 else 2128 Builder.CreateLifetimeEnd(V, Builder.getInt64(EltSize)); 2129 ++Idx; 2130 } 2131 2132 for (; Idx != NewElts.size() && Size; ++Idx) { 2133 IdxTy = NewElts[Idx]->getAllocatedType(); 2134 uint64_t EltSize = DL->getTypeAllocSize(IdxTy); 2135 if (EltSize > Size) { 2136 EltSize = Size; 2137 Size = 0; 2138 } else { 2139 Size -= EltSize; 2140 } 2141 if (II->getIntrinsicID() == Intrinsic::lifetime_start) 2142 Builder.CreateLifetimeStart(NewElts[Idx], 2143 Builder.getInt64(EltSize)); 2144 else 2145 Builder.CreateLifetimeEnd(NewElts[Idx], 2146 Builder.getInt64(EltSize)); 2147 } 2148 DeadInsts.push_back(II); 2149 } 2150 2151 /// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 2152 /// Rewrite it to copy or set the elements of the scalarized memory. 2153 void 2154 SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 2155 AllocaInst *AI, 2156 SmallVectorImpl<AllocaInst *> &NewElts) { 2157 // If this is a memcpy/memmove, construct the other pointer as the 2158 // appropriate type. The "Other" pointer is the pointer that goes to memory 2159 // that doesn't have anything to do with the alloca that we are promoting. For 2160 // memset, this Value* stays null. 2161 Value *OtherPtr = nullptr; 2162 unsigned MemAlignment = MI->getAlignment(); 2163 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 2164 if (Inst == MTI->getRawDest()) 2165 OtherPtr = MTI->getRawSource(); 2166 else { 2167 assert(Inst == MTI->getRawSource()); 2168 OtherPtr = MTI->getRawDest(); 2169 } 2170 } 2171 2172 // If there is an other pointer, we want to convert it to the same pointer 2173 // type as AI has, so we can GEP through it safely. 2174 if (OtherPtr) { 2175 unsigned AddrSpace = 2176 cast<PointerType>(OtherPtr->getType())->getAddressSpace(); 2177 2178 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 2179 // optimization, but it's also required to detect the corner case where 2180 // both pointer operands are referencing the same memory, and where 2181 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 2182 // function is only called for mem intrinsics that access the whole 2183 // aggregate, so non-zero GEPs are not an issue here.) 2184 OtherPtr = OtherPtr->stripPointerCasts(); 2185 2186 // Copying the alloca to itself is a no-op: just delete it. 2187 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 2188 // This code will run twice for a no-op memcpy -- once for each operand. 2189 // Put only one reference to MI on the DeadInsts list. 2190 for (SmallVectorImpl<Value *>::const_iterator I = DeadInsts.begin(), 2191 E = DeadInsts.end(); I != E; ++I) 2192 if (*I == MI) return; 2193 DeadInsts.push_back(MI); 2194 return; 2195 } 2196 2197 // If the pointer is not the right type, insert a bitcast to the right 2198 // type. 2199 Type *NewTy = 2200 PointerType::get(AI->getType()->getElementType(), AddrSpace); 2201 2202 if (OtherPtr->getType() != NewTy) 2203 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI); 2204 } 2205 2206 // Process each element of the aggregate. 2207 bool SROADest = MI->getRawDest() == Inst; 2208 2209 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 2210 2211 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2212 // If this is a memcpy/memmove, emit a GEP of the other element address. 2213 Value *OtherElt = nullptr; 2214 unsigned OtherEltAlign = MemAlignment; 2215 2216 if (OtherPtr) { 2217 Value *Idx[2] = { Zero, 2218 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 2219 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, 2220 OtherPtr->getName()+"."+Twine(i), 2221 MI); 2222 uint64_t EltOffset; 2223 PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 2224 Type *OtherTy = OtherPtrTy->getElementType(); 2225 if (StructType *ST = dyn_cast<StructType>(OtherTy)) { 2226 EltOffset = DL->getStructLayout(ST)->getElementOffset(i); 2227 } else { 2228 Type *EltTy = cast<SequentialType>(OtherTy)->getElementType(); 2229 EltOffset = DL->getTypeAllocSize(EltTy)*i; 2230 } 2231 2232 // The alignment of the other pointer is the guaranteed alignment of the 2233 // element, which is affected by both the known alignment of the whole 2234 // mem intrinsic and the alignment of the element. If the alignment of 2235 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 2236 // known alignment is just 4 bytes. 2237 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 2238 } 2239 2240 Value *EltPtr = NewElts[i]; 2241 Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 2242 2243 // If we got down to a scalar, insert a load or store as appropriate. 2244 if (EltTy->isSingleValueType()) { 2245 if (isa<MemTransferInst>(MI)) { 2246 if (SROADest) { 2247 // From Other to Alloca. 2248 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 2249 new StoreInst(Elt, EltPtr, MI); 2250 } else { 2251 // From Alloca to Other. 2252 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 2253 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 2254 } 2255 continue; 2256 } 2257 assert(isa<MemSetInst>(MI)); 2258 2259 // If the stored element is zero (common case), just store a null 2260 // constant. 2261 Constant *StoreVal; 2262 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) { 2263 if (CI->isZero()) { 2264 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 2265 } else { 2266 // If EltTy is a vector type, get the element type. 2267 Type *ValTy = EltTy->getScalarType(); 2268 2269 // Construct an integer with the right value. 2270 unsigned EltSize = DL->getTypeSizeInBits(ValTy); 2271 APInt OneVal(EltSize, CI->getZExtValue()); 2272 APInt TotalVal(OneVal); 2273 // Set each byte. 2274 for (unsigned i = 0; 8*i < EltSize; ++i) { 2275 TotalVal = TotalVal.shl(8); 2276 TotalVal |= OneVal; 2277 } 2278 2279 // Convert the integer value to the appropriate type. 2280 StoreVal = ConstantInt::get(CI->getContext(), TotalVal); 2281 if (ValTy->isPointerTy()) 2282 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 2283 else if (ValTy->isFloatingPointTy()) 2284 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 2285 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 2286 2287 // If the requested value was a vector constant, create it. 2288 if (EltTy->isVectorTy()) { 2289 unsigned NumElts = cast<VectorType>(EltTy)->getNumElements(); 2290 StoreVal = ConstantVector::getSplat(NumElts, StoreVal); 2291 } 2292 } 2293 new StoreInst(StoreVal, EltPtr, MI); 2294 continue; 2295 } 2296 // Otherwise, if we're storing a byte variable, use a memset call for 2297 // this element. 2298 } 2299 2300 unsigned EltSize = DL->getTypeAllocSize(EltTy); 2301 if (!EltSize) 2302 continue; 2303 2304 IRBuilder<> Builder(MI); 2305 2306 // Finally, insert the meminst for this element. 2307 if (isa<MemSetInst>(MI)) { 2308 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize, 2309 MI->isVolatile()); 2310 } else { 2311 assert(isa<MemTransferInst>(MI)); 2312 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr 2313 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr 2314 2315 if (isa<MemCpyInst>(MI)) 2316 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile()); 2317 else 2318 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile()); 2319 } 2320 } 2321 DeadInsts.push_back(MI); 2322 } 2323 2324 /// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 2325 /// overwrites the entire allocation. Extract out the pieces of the stored 2326 /// integer and store them individually. 2327 void 2328 SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 2329 SmallVectorImpl<AllocaInst *> &NewElts) { 2330 // Extract each element out of the integer according to its structure offset 2331 // and store the element value to the individual alloca. 2332 Value *SrcVal = SI->getOperand(0); 2333 Type *AllocaEltTy = AI->getAllocatedType(); 2334 uint64_t AllocaSizeBits = DL->getTypeAllocSizeInBits(AllocaEltTy); 2335 2336 IRBuilder<> Builder(SI); 2337 2338 // Handle tail padding by extending the operand 2339 if (DL->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 2340 SrcVal = Builder.CreateZExt(SrcVal, 2341 IntegerType::get(SI->getContext(), AllocaSizeBits)); 2342 2343 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 2344 << '\n'); 2345 2346 // There are two forms here: AI could be an array or struct. Both cases 2347 // have different ways to compute the element offset. 2348 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2349 const StructLayout *Layout = DL->getStructLayout(EltSTy); 2350 2351 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2352 // Get the number of bits to shift SrcVal to get the value. 2353 Type *FieldTy = EltSTy->getElementType(i); 2354 uint64_t Shift = Layout->getElementOffsetInBits(i); 2355 2356 if (DL->isBigEndian()) 2357 Shift = AllocaSizeBits-Shift-DL->getTypeAllocSizeInBits(FieldTy); 2358 2359 Value *EltVal = SrcVal; 2360 if (Shift) { 2361 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2362 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2363 } 2364 2365 // Truncate down to an integer of the right size. 2366 uint64_t FieldSizeBits = DL->getTypeSizeInBits(FieldTy); 2367 2368 // Ignore zero sized fields like {}, they obviously contain no data. 2369 if (FieldSizeBits == 0) continue; 2370 2371 if (FieldSizeBits != AllocaSizeBits) 2372 EltVal = Builder.CreateTrunc(EltVal, 2373 IntegerType::get(SI->getContext(), FieldSizeBits)); 2374 Value *DestField = NewElts[i]; 2375 if (EltVal->getType() == FieldTy) { 2376 // Storing to an integer field of this size, just do it. 2377 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 2378 // Bitcast to the right element type (for fp/vector values). 2379 EltVal = Builder.CreateBitCast(EltVal, FieldTy); 2380 } else { 2381 // Otherwise, bitcast the dest pointer (for aggregates). 2382 DestField = Builder.CreateBitCast(DestField, 2383 PointerType::getUnqual(EltVal->getType())); 2384 } 2385 new StoreInst(EltVal, DestField, SI); 2386 } 2387 2388 } else { 2389 ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 2390 Type *ArrayEltTy = ATy->getElementType(); 2391 uint64_t ElementOffset = DL->getTypeAllocSizeInBits(ArrayEltTy); 2392 uint64_t ElementSizeBits = DL->getTypeSizeInBits(ArrayEltTy); 2393 2394 uint64_t Shift; 2395 2396 if (DL->isBigEndian()) 2397 Shift = AllocaSizeBits-ElementOffset; 2398 else 2399 Shift = 0; 2400 2401 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2402 // Ignore zero sized fields like {}, they obviously contain no data. 2403 if (ElementSizeBits == 0) continue; 2404 2405 Value *EltVal = SrcVal; 2406 if (Shift) { 2407 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 2408 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt"); 2409 } 2410 2411 // Truncate down to an integer of the right size. 2412 if (ElementSizeBits != AllocaSizeBits) 2413 EltVal = Builder.CreateTrunc(EltVal, 2414 IntegerType::get(SI->getContext(), 2415 ElementSizeBits)); 2416 Value *DestField = NewElts[i]; 2417 if (EltVal->getType() == ArrayEltTy) { 2418 // Storing to an integer field of this size, just do it. 2419 } else if (ArrayEltTy->isFloatingPointTy() || 2420 ArrayEltTy->isVectorTy()) { 2421 // Bitcast to the right element type (for fp/vector values). 2422 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy); 2423 } else { 2424 // Otherwise, bitcast the dest pointer (for aggregates). 2425 DestField = Builder.CreateBitCast(DestField, 2426 PointerType::getUnqual(EltVal->getType())); 2427 } 2428 new StoreInst(EltVal, DestField, SI); 2429 2430 if (DL->isBigEndian()) 2431 Shift -= ElementOffset; 2432 else 2433 Shift += ElementOffset; 2434 } 2435 } 2436 2437 DeadInsts.push_back(SI); 2438 } 2439 2440 /// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 2441 /// an integer. Load the individual pieces to form the aggregate value. 2442 void 2443 SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 2444 SmallVectorImpl<AllocaInst *> &NewElts) { 2445 // Extract each element out of the NewElts according to its structure offset 2446 // and form the result value. 2447 Type *AllocaEltTy = AI->getAllocatedType(); 2448 uint64_t AllocaSizeBits = DL->getTypeAllocSizeInBits(AllocaEltTy); 2449 2450 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 2451 << '\n'); 2452 2453 // There are two forms here: AI could be an array or struct. Both cases 2454 // have different ways to compute the element offset. 2455 const StructLayout *Layout = nullptr; 2456 uint64_t ArrayEltBitOffset = 0; 2457 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 2458 Layout = DL->getStructLayout(EltSTy); 2459 } else { 2460 Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 2461 ArrayEltBitOffset = DL->getTypeAllocSizeInBits(ArrayEltTy); 2462 } 2463 2464 Value *ResultVal = 2465 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 2466 2467 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 2468 // Load the value from the alloca. If the NewElt is an aggregate, cast 2469 // the pointer to an integer of the same size before doing the load. 2470 Value *SrcField = NewElts[i]; 2471 Type *FieldTy = 2472 cast<PointerType>(SrcField->getType())->getElementType(); 2473 uint64_t FieldSizeBits = DL->getTypeSizeInBits(FieldTy); 2474 2475 // Ignore zero sized fields like {}, they obviously contain no data. 2476 if (FieldSizeBits == 0) continue; 2477 2478 IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 2479 FieldSizeBits); 2480 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 2481 !FieldTy->isVectorTy()) 2482 SrcField = new BitCastInst(SrcField, 2483 PointerType::getUnqual(FieldIntTy), 2484 "", LI); 2485 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 2486 2487 // If SrcField is a fp or vector of the right size but that isn't an 2488 // integer type, bitcast to an integer so we can shift it. 2489 if (SrcField->getType() != FieldIntTy) 2490 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 2491 2492 // Zero extend the field to be the same size as the final alloca so that 2493 // we can shift and insert it. 2494 if (SrcField->getType() != ResultVal->getType()) 2495 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 2496 2497 // Determine the number of bits to shift SrcField. 2498 uint64_t Shift; 2499 if (Layout) // Struct case. 2500 Shift = Layout->getElementOffsetInBits(i); 2501 else // Array case. 2502 Shift = i*ArrayEltBitOffset; 2503 2504 if (DL->isBigEndian()) 2505 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 2506 2507 if (Shift) { 2508 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 2509 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 2510 } 2511 2512 // Don't create an 'or x, 0' on the first iteration. 2513 if (!isa<Constant>(ResultVal) || 2514 !cast<Constant>(ResultVal)->isNullValue()) 2515 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 2516 else 2517 ResultVal = SrcField; 2518 } 2519 2520 // Handle tail padding by truncating the result 2521 if (DL->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 2522 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 2523 2524 LI->replaceAllUsesWith(ResultVal); 2525 DeadInsts.push_back(LI); 2526 } 2527 2528 /// HasPadding - Return true if the specified type has any structure or 2529 /// alignment padding in between the elements that would be split apart 2530 /// by SROA; return false otherwise. 2531 static bool HasPadding(Type *Ty, const DataLayout &DL) { 2532 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 2533 Ty = ATy->getElementType(); 2534 return DL.getTypeSizeInBits(Ty) != DL.getTypeAllocSizeInBits(Ty); 2535 } 2536 2537 // SROA currently handles only Arrays and Structs. 2538 StructType *STy = cast<StructType>(Ty); 2539 const StructLayout *SL = DL.getStructLayout(STy); 2540 unsigned PrevFieldBitOffset = 0; 2541 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 2542 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 2543 2544 // Check to see if there is any padding between this element and the 2545 // previous one. 2546 if (i) { 2547 unsigned PrevFieldEnd = 2548 PrevFieldBitOffset+DL.getTypeSizeInBits(STy->getElementType(i-1)); 2549 if (PrevFieldEnd < FieldBitOffset) 2550 return true; 2551 } 2552 PrevFieldBitOffset = FieldBitOffset; 2553 } 2554 // Check for tail padding. 2555 if (unsigned EltCount = STy->getNumElements()) { 2556 unsigned PrevFieldEnd = PrevFieldBitOffset + 2557 DL.getTypeSizeInBits(STy->getElementType(EltCount-1)); 2558 if (PrevFieldEnd < SL->getSizeInBits()) 2559 return true; 2560 } 2561 return false; 2562 } 2563 2564 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 2565 /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 2566 /// or 1 if safe after canonicalization has been performed. 2567 bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 2568 // Loop over the use list of the alloca. We can only transform it if all of 2569 // the users are safe to transform. 2570 AllocaInfo Info(AI); 2571 2572 isSafeForScalarRepl(AI, 0, Info); 2573 if (Info.isUnsafe) { 2574 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 2575 return false; 2576 } 2577 2578 // Okay, we know all the users are promotable. If the aggregate is a memcpy 2579 // source and destination, we have to be careful. In particular, the memcpy 2580 // could be moving around elements that live in structure padding of the LLVM 2581 // types, but may actually be used. In these cases, we refuse to promote the 2582 // struct. 2583 if (Info.isMemCpySrc && Info.isMemCpyDst && 2584 HasPadding(AI->getAllocatedType(), *DL)) 2585 return false; 2586 2587 // If the alloca never has an access to just *part* of it, but is accessed 2588 // via loads and stores, then we should use ConvertToScalarInfo to promote 2589 // the alloca instead of promoting each piece at a time and inserting fission 2590 // and fusion code. 2591 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) { 2592 // If the struct/array just has one element, use basic SRoA. 2593 if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 2594 if (ST->getNumElements() > 1) return false; 2595 } else { 2596 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1) 2597 return false; 2598 } 2599 } 2600 2601 return true; 2602 } 2603