ScalarReplAggregates.cpp revision c4472072641f702dbd99ae12b7da089e75c44a99
1//===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This transformation implements the well known scalar replacement of 11// aggregates transformation. This xform breaks up alloca instructions of 12// aggregate type (structure or array) into individual alloca instructions for 13// each member (if possible). Then, if possible, it transforms the individual 14// alloca instructions into nice clean scalar SSA form. 15// 16// This combines a simple SRoA algorithm with the Mem2Reg algorithm because 17// often interact, especially for C++ programs. As such, iterating between 18// SRoA, then Mem2Reg until we run out of things to promote works well. 19// 20//===----------------------------------------------------------------------===// 21 22#define DEBUG_TYPE "scalarrepl" 23#include "llvm/Transforms/Scalar.h" 24#include "llvm/Constants.h" 25#include "llvm/DerivedTypes.h" 26#include "llvm/Function.h" 27#include "llvm/GlobalVariable.h" 28#include "llvm/Instructions.h" 29#include "llvm/IntrinsicInst.h" 30#include "llvm/LLVMContext.h" 31#include "llvm/Pass.h" 32#include "llvm/Analysis/Dominators.h" 33#include "llvm/Target/TargetData.h" 34#include "llvm/Transforms/Utils/PromoteMemToReg.h" 35#include "llvm/Transforms/Utils/Local.h" 36#include "llvm/Support/Debug.h" 37#include "llvm/Support/ErrorHandling.h" 38#include "llvm/Support/GetElementPtrTypeIterator.h" 39#include "llvm/Support/IRBuilder.h" 40#include "llvm/Support/MathExtras.h" 41#include "llvm/Support/raw_ostream.h" 42#include "llvm/ADT/SmallVector.h" 43#include "llvm/ADT/Statistic.h" 44using namespace llvm; 45 46STATISTIC(NumReplaced, "Number of allocas broken up"); 47STATISTIC(NumPromoted, "Number of allocas promoted"); 48STATISTIC(NumConverted, "Number of aggregates converted to scalar"); 49STATISTIC(NumGlobals, "Number of allocas copied from constant global"); 50 51namespace { 52 struct ConvertToScalarInfo; 53 54 struct SROA : public FunctionPass { 55 static char ID; // Pass identification, replacement for typeid 56 explicit SROA(signed T = -1) : FunctionPass(&ID) { 57 if (T == -1) 58 SRThreshold = 128; 59 else 60 SRThreshold = T; 61 } 62 63 bool runOnFunction(Function &F); 64 65 bool performScalarRepl(Function &F); 66 bool performPromotion(Function &F); 67 68 // getAnalysisUsage - This pass does not require any passes, but we know it 69 // will not alter the CFG, so say so. 70 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 71 AU.addRequired<DominatorTree>(); 72 AU.addRequired<DominanceFrontier>(); 73 AU.setPreservesCFG(); 74 } 75 76 private: 77 TargetData *TD; 78 79 /// DeadInsts - Keep track of instructions we have made dead, so that 80 /// we can remove them after we are done working. 81 SmallVector<Value*, 32> DeadInsts; 82 83 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures 84 /// information about the uses. All these fields are initialized to false 85 /// and set to true when something is learned. 86 struct AllocaInfo { 87 /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 88 bool isUnsafe : 1; 89 90 /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 91 bool isMemCpySrc : 1; 92 93 /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 94 bool isMemCpyDst : 1; 95 96 AllocaInfo() 97 : isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false) {} 98 }; 99 100 unsigned SRThreshold; 101 102 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; } 103 104 bool isSafeAllocaToScalarRepl(AllocaInst *AI); 105 106 void isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 107 AllocaInfo &Info); 108 void isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t &Offset, 109 AllocaInfo &Info); 110 void isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize, 111 const Type *MemOpType, bool isStore, AllocaInfo &Info); 112 bool TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size); 113 uint64_t FindElementAndOffset(const Type *&T, uint64_t &Offset, 114 const Type *&IdxTy); 115 116 void DoScalarReplacement(AllocaInst *AI, 117 std::vector<AllocaInst*> &WorkList); 118 void DeleteDeadInstructions(); 119 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocaInst *Base); 120 121 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 122 SmallVector<AllocaInst*, 32> &NewElts); 123 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 124 SmallVector<AllocaInst*, 32> &NewElts); 125 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 126 SmallVector<AllocaInst*, 32> &NewElts); 127 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 128 AllocaInst *AI, 129 SmallVector<AllocaInst*, 32> &NewElts); 130 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 131 SmallVector<AllocaInst*, 32> &NewElts); 132 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 133 SmallVector<AllocaInst*, 32> &NewElts); 134 135 bool CanConvertToScalar(Value *V, ConvertToScalarInfo &ConvertInfo, 136 uint64_t Offset); 137 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); 138 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType, 139 uint64_t Offset, IRBuilder<> &Builder); 140 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 141 uint64_t Offset, IRBuilder<> &Builder); 142 static MemTransferInst *isOnlyCopiedFromConstantGlobal(AllocaInst *AI); 143 }; 144} 145 146char SROA::ID = 0; 147static RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates"); 148 149// Public interface to the ScalarReplAggregates pass 150FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) { 151 return new SROA(Threshold); 152} 153 154 155bool SROA::runOnFunction(Function &F) { 156 TD = getAnalysisIfAvailable<TargetData>(); 157 158 bool Changed = performPromotion(F); 159 160 // FIXME: ScalarRepl currently depends on TargetData more than it 161 // theoretically needs to. It should be refactored in order to support 162 // target-independent IR. Until this is done, just skip the actual 163 // scalar-replacement portion of this pass. 164 if (!TD) return Changed; 165 166 while (1) { 167 bool LocalChange = performScalarRepl(F); 168 if (!LocalChange) break; // No need to repromote if no scalarrepl 169 Changed = true; 170 LocalChange = performPromotion(F); 171 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 172 } 173 174 return Changed; 175} 176 177 178bool SROA::performPromotion(Function &F) { 179 std::vector<AllocaInst*> Allocas; 180 DominatorTree &DT = getAnalysis<DominatorTree>(); 181 DominanceFrontier &DF = getAnalysis<DominanceFrontier>(); 182 183 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 184 185 bool Changed = false; 186 187 while (1) { 188 Allocas.clear(); 189 190 // Find allocas that are safe to promote, by looking at all instructions in 191 // the entry node 192 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 193 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 194 if (isAllocaPromotable(AI)) 195 Allocas.push_back(AI); 196 197 if (Allocas.empty()) break; 198 199 PromoteMemToReg(Allocas, DT, DF); 200 NumPromoted += Allocas.size(); 201 Changed = true; 202 } 203 204 return Changed; 205} 206 207/// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 208/// SROA. It must be a struct or array type with a small number of elements. 209static bool ShouldAttemptScalarRepl(AllocaInst *AI) { 210 const Type *T = AI->getAllocatedType(); 211 // Do not promote any struct into more than 32 separate vars. 212 if (const StructType *ST = dyn_cast<StructType>(T)) 213 return ST->getNumElements() <= 32; 214 // Arrays are much less likely to be safe for SROA; only consider 215 // them if they are very small. 216 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) 217 return AT->getNumElements() <= 8; 218 return false; 219} 220 221namespace { 222struct ConvertToScalarInfo { 223 /// AllocaSize - The size of the alloca being considered. 224 unsigned AllocaSize; 225 226 bool IsNotTrivial; 227 const Type *VectorTy; 228 bool HadAVector; 229 230 explicit ConvertToScalarInfo(unsigned Size) : AllocaSize(Size) { 231 IsNotTrivial = false; 232 VectorTy = 0; 233 HadAVector = false; 234 } 235 236 bool shouldConvertToVector() const { 237 return VectorTy && VectorTy->isVectorTy() && HadAVector; 238 } 239}; 240} // end anonymous namespace. 241 242 243 244// performScalarRepl - This algorithm is a simple worklist driven algorithm, 245// which runs on all of the malloc/alloca instructions in the function, removing 246// them if they are only used by getelementptr instructions. 247// 248bool SROA::performScalarRepl(Function &F) { 249 std::vector<AllocaInst*> WorkList; 250 251 // Scan the entry basic block, adding allocas to the worklist. 252 BasicBlock &BB = F.getEntryBlock(); 253 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 254 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 255 WorkList.push_back(A); 256 257 // Process the worklist 258 bool Changed = false; 259 while (!WorkList.empty()) { 260 AllocaInst *AI = WorkList.back(); 261 WorkList.pop_back(); 262 263 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 264 // with unused elements. 265 if (AI->use_empty()) { 266 AI->eraseFromParent(); 267 Changed = true; 268 continue; 269 } 270 271 // If this alloca is impossible for us to promote, reject it early. 272 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 273 continue; 274 275 // Check to see if this allocation is only modified by a memcpy/memmove from 276 // a constant global. If this is the case, we can change all users to use 277 // the constant global instead. This is commonly produced by the CFE by 278 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 279 // is only subsequently read. 280 if (MemTransferInst *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 281 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); 282 DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n'); 283 Constant *TheSrc = cast<Constant>(TheCopy->getSource()); 284 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 285 TheCopy->eraseFromParent(); // Don't mutate the global. 286 AI->eraseFromParent(); 287 ++NumGlobals; 288 Changed = true; 289 continue; 290 } 291 292 // Check to see if we can perform the core SROA transformation. We cannot 293 // transform the allocation instruction if it is an array allocation 294 // (allocations OF arrays are ok though), and an allocation of a scalar 295 // value cannot be decomposed at all. 296 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 297 298 // Do not promote [0 x %struct]. 299 if (AllocaSize == 0) continue; 300 301 // Do not promote any struct whose size is too big. 302 if (AllocaSize > SRThreshold) continue; 303 304 // If the alloca looks like a good candidate for scalar replacement, and if 305 // all its users can be transformed, then split up the aggregate into its 306 // separate elements. 307 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 308 DoScalarReplacement(AI, WorkList); 309 Changed = true; 310 continue; 311 } 312 313 // If we can turn this aggregate value (potentially with casts) into a 314 // simple scalar value that can be mem2reg'd into a register value. 315 // IsNotTrivial tracks whether this is something that mem2reg could have 316 // promoted itself. If so, we don't want to transform it needlessly. Note 317 // that we can't just check based on the type: the alloca may be of an i32 318 // but that has pointer arithmetic to set byte 3 of it or something. 319 ConvertToScalarInfo ConvertInfo((unsigned)AllocaSize); 320 if (CanConvertToScalar(AI, ConvertInfo, 0) && ConvertInfo.IsNotTrivial) { 321 AllocaInst *NewAI; 322 // If we were able to find a vector type that can handle this with 323 // insert/extract elements, and if there was at least one use that had 324 // a vector type, promote this to a vector. We don't want to promote 325 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 326 // we just get a lot of insert/extracts. If at least one vector is 327 // involved, then we probably really do have a union of vector/array. 328 if (ConvertInfo.shouldConvertToVector()) { 329 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = " 330 << *ConvertInfo.VectorTy << '\n'); 331 332 // Create and insert the vector alloca. 333 NewAI = new AllocaInst(ConvertInfo.VectorTy, 0, "", 334 AI->getParent()->begin()); 335 ConvertUsesToScalar(AI, NewAI, 0); 336 } else { 337 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"); 338 339 // Create and insert the integer alloca. 340 const Type *NewTy = IntegerType::get(AI->getContext(), AllocaSize*8); 341 NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 342 ConvertUsesToScalar(AI, NewAI, 0); 343 } 344 NewAI->takeName(AI); 345 AI->eraseFromParent(); 346 ++NumConverted; 347 Changed = true; 348 continue; 349 } 350 351 // Otherwise, couldn't process this alloca. 352 } 353 354 return Changed; 355} 356 357/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 358/// predicate, do SROA now. 359void SROA::DoScalarReplacement(AllocaInst *AI, 360 std::vector<AllocaInst*> &WorkList) { 361 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 362 SmallVector<AllocaInst*, 32> ElementAllocas; 363 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 364 ElementAllocas.reserve(ST->getNumContainedTypes()); 365 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 366 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 367 AI->getAlignment(), 368 AI->getName() + "." + Twine(i), AI); 369 ElementAllocas.push_back(NA); 370 WorkList.push_back(NA); // Add to worklist for recursive processing 371 } 372 } else { 373 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 374 ElementAllocas.reserve(AT->getNumElements()); 375 const Type *ElTy = AT->getElementType(); 376 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 377 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 378 AI->getName() + "." + Twine(i), AI); 379 ElementAllocas.push_back(NA); 380 WorkList.push_back(NA); // Add to worklist for recursive processing 381 } 382 } 383 384 // Now that we have created the new alloca instructions, rewrite all the 385 // uses of the old alloca. 386 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 387 388 // Now erase any instructions that were made dead while rewriting the alloca. 389 DeleteDeadInstructions(); 390 AI->eraseFromParent(); 391 392 NumReplaced++; 393} 394 395/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 396/// recursively including all their operands that become trivially dead. 397void SROA::DeleteDeadInstructions() { 398 while (!DeadInsts.empty()) { 399 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 400 401 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 402 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 403 // Zero out the operand and see if it becomes trivially dead. 404 // (But, don't add allocas to the dead instruction list -- they are 405 // already on the worklist and will be deleted separately.) 406 *OI = 0; 407 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 408 DeadInsts.push_back(U); 409 } 410 411 I->eraseFromParent(); 412 } 413} 414 415/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 416/// performing scalar replacement of alloca AI. The results are flagged in 417/// the Info parameter. Offset indicates the position within AI that is 418/// referenced by this instruction. 419void SROA::isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 420 AllocaInfo &Info) { 421 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 422 Instruction *User = cast<Instruction>(*UI); 423 424 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 425 isSafeForScalarRepl(BC, AI, Offset, Info); 426 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 427 uint64_t GEPOffset = Offset; 428 isSafeGEP(GEPI, AI, GEPOffset, Info); 429 if (!Info.isUnsafe) 430 isSafeForScalarRepl(GEPI, AI, GEPOffset, Info); 431 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 432 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 433 if (Length) 434 isSafeMemAccess(AI, Offset, Length->getZExtValue(), 0, 435 UI.getOperandNo() == 0, Info); 436 else 437 MarkUnsafe(Info); 438 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 439 if (!LI->isVolatile()) { 440 const Type *LIType = LI->getType(); 441 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(LIType), 442 LIType, false, Info); 443 } else 444 MarkUnsafe(Info); 445 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 446 // Store is ok if storing INTO the pointer, not storing the pointer 447 if (!SI->isVolatile() && SI->getOperand(0) != I) { 448 const Type *SIType = SI->getOperand(0)->getType(); 449 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(SIType), 450 SIType, true, Info); 451 } else 452 MarkUnsafe(Info); 453 } else { 454 DEBUG(errs() << " Transformation preventing inst: " << *User << '\n'); 455 MarkUnsafe(Info); 456 } 457 if (Info.isUnsafe) return; 458 } 459} 460 461/// isSafeGEP - Check if a GEP instruction can be handled for scalar 462/// replacement. It is safe when all the indices are constant, in-bounds 463/// references, and when the resulting offset corresponds to an element within 464/// the alloca type. The results are flagged in the Info parameter. Upon 465/// return, Offset is adjusted as specified by the GEP indices. 466void SROA::isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, 467 uint64_t &Offset, AllocaInfo &Info) { 468 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 469 if (GEPIt == E) 470 return; 471 472 // Walk through the GEP type indices, checking the types that this indexes 473 // into. 474 for (; GEPIt != E; ++GEPIt) { 475 // Ignore struct elements, no extra checking needed for these. 476 if ((*GEPIt)->isStructTy()) 477 continue; 478 479 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 480 if (!IdxVal) 481 return MarkUnsafe(Info); 482 } 483 484 // Compute the offset due to this GEP and check if the alloca has a 485 // component element at that offset. 486 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 487 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 488 &Indices[0], Indices.size()); 489 if (!TypeHasComponent(AI->getAllocatedType(), Offset, 0)) 490 MarkUnsafe(Info); 491} 492 493/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 494/// alloca or has an offset and size that corresponds to a component element 495/// within it. The offset checked here may have been formed from a GEP with a 496/// pointer bitcasted to a different type. 497void SROA::isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize, 498 const Type *MemOpType, bool isStore, 499 AllocaInfo &Info) { 500 // Check if this is a load/store of the entire alloca. 501 if (Offset == 0 && MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) { 502 bool UsesAggregateType = (MemOpType == AI->getAllocatedType()); 503 // This is safe for MemIntrinsics (where MemOpType is 0), integer types 504 // (which are essentially the same as the MemIntrinsics, especially with 505 // regard to copying padding between elements), or references using the 506 // aggregate type of the alloca. 507 if (!MemOpType || MemOpType->isIntegerTy() || UsesAggregateType) { 508 if (!UsesAggregateType) { 509 if (isStore) 510 Info.isMemCpyDst = true; 511 else 512 Info.isMemCpySrc = true; 513 } 514 return; 515 } 516 } 517 // Check if the offset/size correspond to a component within the alloca type. 518 const Type *T = AI->getAllocatedType(); 519 if (TypeHasComponent(T, Offset, MemSize)) 520 return; 521 522 return MarkUnsafe(Info); 523} 524 525/// TypeHasComponent - Return true if T has a component type with the 526/// specified offset and size. If Size is zero, do not check the size. 527bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) { 528 const Type *EltTy; 529 uint64_t EltSize; 530 if (const StructType *ST = dyn_cast<StructType>(T)) { 531 const StructLayout *Layout = TD->getStructLayout(ST); 532 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 533 EltTy = ST->getContainedType(EltIdx); 534 EltSize = TD->getTypeAllocSize(EltTy); 535 Offset -= Layout->getElementOffset(EltIdx); 536 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 537 EltTy = AT->getElementType(); 538 EltSize = TD->getTypeAllocSize(EltTy); 539 if (Offset >= AT->getNumElements() * EltSize) 540 return false; 541 Offset %= EltSize; 542 } else { 543 return false; 544 } 545 if (Offset == 0 && (Size == 0 || EltSize == Size)) 546 return true; 547 // Check if the component spans multiple elements. 548 if (Offset + Size > EltSize) 549 return false; 550 return TypeHasComponent(EltTy, Offset, Size); 551} 552 553/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 554/// the instruction I, which references it, to use the separate elements. 555/// Offset indicates the position within AI that is referenced by this 556/// instruction. 557void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 558 SmallVector<AllocaInst*, 32> &NewElts) { 559 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 560 Instruction *User = cast<Instruction>(*UI); 561 562 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 563 RewriteBitCast(BC, AI, Offset, NewElts); 564 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 565 RewriteGEP(GEPI, AI, Offset, NewElts); 566 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 567 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 568 uint64_t MemSize = Length->getZExtValue(); 569 if (Offset == 0 && 570 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 571 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 572 // Otherwise the intrinsic can only touch a single element and the 573 // address operand will be updated, so nothing else needs to be done. 574 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 575 const Type *LIType = LI->getType(); 576 if (LIType == AI->getAllocatedType()) { 577 // Replace: 578 // %res = load { i32, i32 }* %alloc 579 // with: 580 // %load.0 = load i32* %alloc.0 581 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 582 // %load.1 = load i32* %alloc.1 583 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 584 // (Also works for arrays instead of structs) 585 Value *Insert = UndefValue::get(LIType); 586 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 587 Value *Load = new LoadInst(NewElts[i], "load", LI); 588 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 589 } 590 LI->replaceAllUsesWith(Insert); 591 DeadInsts.push_back(LI); 592 } else if (LIType->isIntegerTy() && 593 TD->getTypeAllocSize(LIType) == 594 TD->getTypeAllocSize(AI->getAllocatedType())) { 595 // If this is a load of the entire alloca to an integer, rewrite it. 596 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 597 } 598 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 599 Value *Val = SI->getOperand(0); 600 const Type *SIType = Val->getType(); 601 if (SIType == AI->getAllocatedType()) { 602 // Replace: 603 // store { i32, i32 } %val, { i32, i32 }* %alloc 604 // with: 605 // %val.0 = extractvalue { i32, i32 } %val, 0 606 // store i32 %val.0, i32* %alloc.0 607 // %val.1 = extractvalue { i32, i32 } %val, 1 608 // store i32 %val.1, i32* %alloc.1 609 // (Also works for arrays instead of structs) 610 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 611 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 612 new StoreInst(Extract, NewElts[i], SI); 613 } 614 DeadInsts.push_back(SI); 615 } else if (SIType->isIntegerTy() && 616 TD->getTypeAllocSize(SIType) == 617 TD->getTypeAllocSize(AI->getAllocatedType())) { 618 // If this is a store of the entire alloca from an integer, rewrite it. 619 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 620 } 621 } 622 } 623} 624 625/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 626/// and recursively continue updating all of its uses. 627void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 628 SmallVector<AllocaInst*, 32> &NewElts) { 629 RewriteForScalarRepl(BC, AI, Offset, NewElts); 630 if (BC->getOperand(0) != AI) 631 return; 632 633 // The bitcast references the original alloca. Replace its uses with 634 // references to the first new element alloca. 635 Instruction *Val = NewElts[0]; 636 if (Val->getType() != BC->getDestTy()) { 637 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 638 Val->takeName(BC); 639 } 640 BC->replaceAllUsesWith(Val); 641 DeadInsts.push_back(BC); 642} 643 644/// FindElementAndOffset - Return the index of the element containing Offset 645/// within the specified type, which must be either a struct or an array. 646/// Sets T to the type of the element and Offset to the offset within that 647/// element. IdxTy is set to the type of the index result to be used in a 648/// GEP instruction. 649uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset, 650 const Type *&IdxTy) { 651 uint64_t Idx = 0; 652 if (const StructType *ST = dyn_cast<StructType>(T)) { 653 const StructLayout *Layout = TD->getStructLayout(ST); 654 Idx = Layout->getElementContainingOffset(Offset); 655 T = ST->getContainedType(Idx); 656 Offset -= Layout->getElementOffset(Idx); 657 IdxTy = Type::getInt32Ty(T->getContext()); 658 return Idx; 659 } 660 const ArrayType *AT = cast<ArrayType>(T); 661 T = AT->getElementType(); 662 uint64_t EltSize = TD->getTypeAllocSize(T); 663 Idx = Offset / EltSize; 664 Offset -= Idx * EltSize; 665 IdxTy = Type::getInt64Ty(T->getContext()); 666 return Idx; 667} 668 669/// RewriteGEP - Check if this GEP instruction moves the pointer across 670/// elements of the alloca that are being split apart, and if so, rewrite 671/// the GEP to be relative to the new element. 672void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 673 SmallVector<AllocaInst*, 32> &NewElts) { 674 uint64_t OldOffset = Offset; 675 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 676 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 677 &Indices[0], Indices.size()); 678 679 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 680 681 const Type *T = AI->getAllocatedType(); 682 const Type *IdxTy; 683 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 684 if (GEPI->getOperand(0) == AI) 685 OldIdx = ~0ULL; // Force the GEP to be rewritten. 686 687 T = AI->getAllocatedType(); 688 uint64_t EltOffset = Offset; 689 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 690 691 // If this GEP does not move the pointer across elements of the alloca 692 // being split, then it does not needs to be rewritten. 693 if (Idx == OldIdx) 694 return; 695 696 const Type *i32Ty = Type::getInt32Ty(AI->getContext()); 697 SmallVector<Value*, 8> NewArgs; 698 NewArgs.push_back(Constant::getNullValue(i32Ty)); 699 while (EltOffset != 0) { 700 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 701 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 702 } 703 Instruction *Val = NewElts[Idx]; 704 if (NewArgs.size() > 1) { 705 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(), 706 NewArgs.end(), "", GEPI); 707 Val->takeName(GEPI); 708 } 709 if (Val->getType() != GEPI->getType()) 710 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 711 GEPI->replaceAllUsesWith(Val); 712 DeadInsts.push_back(GEPI); 713} 714 715/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 716/// Rewrite it to copy or set the elements of the scalarized memory. 717void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 718 AllocaInst *AI, 719 SmallVector<AllocaInst*, 32> &NewElts) { 720 // If this is a memcpy/memmove, construct the other pointer as the 721 // appropriate type. The "Other" pointer is the pointer that goes to memory 722 // that doesn't have anything to do with the alloca that we are promoting. For 723 // memset, this Value* stays null. 724 Value *OtherPtr = 0; 725 LLVMContext &Context = MI->getContext(); 726 unsigned MemAlignment = MI->getAlignment(); 727 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 728 if (Inst == MTI->getRawDest()) 729 OtherPtr = MTI->getRawSource(); 730 else { 731 assert(Inst == MTI->getRawSource()); 732 OtherPtr = MTI->getRawDest(); 733 } 734 } 735 736 // If there is an other pointer, we want to convert it to the same pointer 737 // type as AI has, so we can GEP through it safely. 738 if (OtherPtr) { 739 740 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 741 // optimization, but it's also required to detect the corner case where 742 // both pointer operands are referencing the same memory, and where 743 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 744 // function is only called for mem intrinsics that access the whole 745 // aggregate, so non-zero GEPs are not an issue here.) 746 while (1) { 747 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr)) { 748 OtherPtr = BC->getOperand(0); 749 continue; 750 } 751 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr)) { 752 // All zero GEPs are effectively bitcasts. 753 if (GEP->hasAllZeroIndices()) { 754 OtherPtr = GEP->getOperand(0); 755 continue; 756 } 757 } 758 break; 759 } 760 // Copying the alloca to itself is a no-op: just delete it. 761 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 762 // This code will run twice for a no-op memcpy -- once for each operand. 763 // Put only one reference to MI on the DeadInsts list. 764 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), 765 E = DeadInsts.end(); I != E; ++I) 766 if (*I == MI) return; 767 DeadInsts.push_back(MI); 768 return; 769 } 770 771 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr)) 772 if (BCE->getOpcode() == Instruction::BitCast) 773 OtherPtr = BCE->getOperand(0); 774 775 // If the pointer is not the right type, insert a bitcast to the right 776 // type. 777 if (OtherPtr->getType() != AI->getType()) 778 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(), 779 MI); 780 } 781 782 // Process each element of the aggregate. 783 Value *TheFn = MI->getCalledValue(); 784 const Type *BytePtrTy = MI->getRawDest()->getType(); 785 bool SROADest = MI->getRawDest() == Inst; 786 787 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 788 789 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 790 // If this is a memcpy/memmove, emit a GEP of the other element address. 791 Value *OtherElt = 0; 792 unsigned OtherEltAlign = MemAlignment; 793 794 if (OtherPtr) { 795 Value *Idx[2] = { Zero, 796 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 797 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2, 798 OtherPtr->getName()+"."+Twine(i), 799 MI); 800 uint64_t EltOffset; 801 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 802 if (const StructType *ST = 803 dyn_cast<StructType>(OtherPtrTy->getElementType())) { 804 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 805 } else { 806 const Type *EltTy = 807 cast<SequentialType>(OtherPtr->getType())->getElementType(); 808 EltOffset = TD->getTypeAllocSize(EltTy)*i; 809 } 810 811 // The alignment of the other pointer is the guaranteed alignment of the 812 // element, which is affected by both the known alignment of the whole 813 // mem intrinsic and the alignment of the element. If the alignment of 814 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 815 // known alignment is just 4 bytes. 816 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 817 } 818 819 Value *EltPtr = NewElts[i]; 820 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 821 822 // If we got down to a scalar, insert a load or store as appropriate. 823 if (EltTy->isSingleValueType()) { 824 if (isa<MemTransferInst>(MI)) { 825 if (SROADest) { 826 // From Other to Alloca. 827 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 828 new StoreInst(Elt, EltPtr, MI); 829 } else { 830 // From Alloca to Other. 831 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 832 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 833 } 834 continue; 835 } 836 assert(isa<MemSetInst>(MI)); 837 838 // If the stored element is zero (common case), just store a null 839 // constant. 840 Constant *StoreVal; 841 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(1))) { 842 if (CI->isZero()) { 843 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 844 } else { 845 // If EltTy is a vector type, get the element type. 846 const Type *ValTy = EltTy->getScalarType(); 847 848 // Construct an integer with the right value. 849 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 850 APInt OneVal(EltSize, CI->getZExtValue()); 851 APInt TotalVal(OneVal); 852 // Set each byte. 853 for (unsigned i = 0; 8*i < EltSize; ++i) { 854 TotalVal = TotalVal.shl(8); 855 TotalVal |= OneVal; 856 } 857 858 // Convert the integer value to the appropriate type. 859 StoreVal = ConstantInt::get(Context, TotalVal); 860 if (ValTy->isPointerTy()) 861 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 862 else if (ValTy->isFloatingPointTy()) 863 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 864 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 865 866 // If the requested value was a vector constant, create it. 867 if (EltTy != ValTy) { 868 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 869 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 870 StoreVal = ConstantVector::get(&Elts[0], NumElts); 871 } 872 } 873 new StoreInst(StoreVal, EltPtr, MI); 874 continue; 875 } 876 // Otherwise, if we're storing a byte variable, use a memset call for 877 // this element. 878 } 879 880 // Cast the element pointer to BytePtrTy. 881 if (EltPtr->getType() != BytePtrTy) 882 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getName(), MI); 883 884 // Cast the other pointer (if we have one) to BytePtrTy. 885 if (OtherElt && OtherElt->getType() != BytePtrTy) { 886 // Preserve address space of OtherElt 887 const PointerType* OtherPTy = cast<PointerType>(OtherElt->getType()); 888 const PointerType* PTy = cast<PointerType>(BytePtrTy); 889 if (OtherPTy->getElementType() != PTy->getElementType()) { 890 Type *NewOtherPTy = PointerType::get(PTy->getElementType(), 891 OtherPTy->getAddressSpace()); 892 OtherElt = new BitCastInst(OtherElt, NewOtherPTy, 893 OtherElt->getNameStr(), MI); 894 } 895 } 896 897 unsigned EltSize = TD->getTypeAllocSize(EltTy); 898 899 // Finally, insert the meminst for this element. 900 if (isa<MemTransferInst>(MI)) { 901 Value *Ops[] = { 902 SROADest ? EltPtr : OtherElt, // Dest ptr 903 SROADest ? OtherElt : EltPtr, // Src ptr 904 ConstantInt::get(MI->getOperand(2)->getType(), EltSize), // Size 905 // Align 906 ConstantInt::get(Type::getInt32Ty(MI->getContext()), OtherEltAlign), 907 MI->getVolatileCst() 908 }; 909 // In case we fold the address space overloaded memcpy of A to B 910 // with memcpy of B to C, change the function to be a memcpy of A to C. 911 const Type *Tys[] = { Ops[0]->getType(), Ops[1]->getType(), 912 Ops[2]->getType() }; 913 Module *M = MI->getParent()->getParent()->getParent(); 914 TheFn = Intrinsic::getDeclaration(M, MI->getIntrinsicID(), Tys, 3); 915 CallInst::Create(TheFn, Ops, Ops + 5, "", MI); 916 } else { 917 assert(isa<MemSetInst>(MI)); 918 Value *Ops[] = { 919 EltPtr, MI->getOperand(1), // Dest, Value, 920 ConstantInt::get(MI->getOperand(2)->getType(), EltSize), // Size 921 Zero, // Align 922 ConstantInt::get(Type::getInt1Ty(MI->getContext()), 0) // isVolatile 923 }; 924 const Type *Tys[] = { Ops[0]->getType(), Ops[2]->getType() }; 925 Module *M = MI->getParent()->getParent()->getParent(); 926 TheFn = Intrinsic::getDeclaration(M, Intrinsic::memset, Tys, 2); 927 CallInst::Create(TheFn, Ops, Ops + 5, "", MI); 928 } 929 } 930 DeadInsts.push_back(MI); 931} 932 933/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 934/// overwrites the entire allocation. Extract out the pieces of the stored 935/// integer and store them individually. 936void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 937 SmallVector<AllocaInst*, 32> &NewElts){ 938 // Extract each element out of the integer according to its structure offset 939 // and store the element value to the individual alloca. 940 Value *SrcVal = SI->getOperand(0); 941 const Type *AllocaEltTy = AI->getAllocatedType(); 942 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 943 944 // Handle tail padding by extending the operand 945 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 946 SrcVal = new ZExtInst(SrcVal, 947 IntegerType::get(SI->getContext(), AllocaSizeBits), 948 "", SI); 949 950 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 951 << '\n'); 952 953 // There are two forms here: AI could be an array or struct. Both cases 954 // have different ways to compute the element offset. 955 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 956 const StructLayout *Layout = TD->getStructLayout(EltSTy); 957 958 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 959 // Get the number of bits to shift SrcVal to get the value. 960 const Type *FieldTy = EltSTy->getElementType(i); 961 uint64_t Shift = Layout->getElementOffsetInBits(i); 962 963 if (TD->isBigEndian()) 964 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 965 966 Value *EltVal = SrcVal; 967 if (Shift) { 968 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 969 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 970 "sroa.store.elt", SI); 971 } 972 973 // Truncate down to an integer of the right size. 974 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 975 976 // Ignore zero sized fields like {}, they obviously contain no data. 977 if (FieldSizeBits == 0) continue; 978 979 if (FieldSizeBits != AllocaSizeBits) 980 EltVal = new TruncInst(EltVal, 981 IntegerType::get(SI->getContext(), FieldSizeBits), 982 "", SI); 983 Value *DestField = NewElts[i]; 984 if (EltVal->getType() == FieldTy) { 985 // Storing to an integer field of this size, just do it. 986 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 987 // Bitcast to the right element type (for fp/vector values). 988 EltVal = new BitCastInst(EltVal, FieldTy, "", SI); 989 } else { 990 // Otherwise, bitcast the dest pointer (for aggregates). 991 DestField = new BitCastInst(DestField, 992 PointerType::getUnqual(EltVal->getType()), 993 "", SI); 994 } 995 new StoreInst(EltVal, DestField, SI); 996 } 997 998 } else { 999 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 1000 const Type *ArrayEltTy = ATy->getElementType(); 1001 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1002 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 1003 1004 uint64_t Shift; 1005 1006 if (TD->isBigEndian()) 1007 Shift = AllocaSizeBits-ElementOffset; 1008 else 1009 Shift = 0; 1010 1011 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1012 // Ignore zero sized fields like {}, they obviously contain no data. 1013 if (ElementSizeBits == 0) continue; 1014 1015 Value *EltVal = SrcVal; 1016 if (Shift) { 1017 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 1018 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 1019 "sroa.store.elt", SI); 1020 } 1021 1022 // Truncate down to an integer of the right size. 1023 if (ElementSizeBits != AllocaSizeBits) 1024 EltVal = new TruncInst(EltVal, 1025 IntegerType::get(SI->getContext(), 1026 ElementSizeBits),"",SI); 1027 Value *DestField = NewElts[i]; 1028 if (EltVal->getType() == ArrayEltTy) { 1029 // Storing to an integer field of this size, just do it. 1030 } else if (ArrayEltTy->isFloatingPointTy() || 1031 ArrayEltTy->isVectorTy()) { 1032 // Bitcast to the right element type (for fp/vector values). 1033 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI); 1034 } else { 1035 // Otherwise, bitcast the dest pointer (for aggregates). 1036 DestField = new BitCastInst(DestField, 1037 PointerType::getUnqual(EltVal->getType()), 1038 "", SI); 1039 } 1040 new StoreInst(EltVal, DestField, SI); 1041 1042 if (TD->isBigEndian()) 1043 Shift -= ElementOffset; 1044 else 1045 Shift += ElementOffset; 1046 } 1047 } 1048 1049 DeadInsts.push_back(SI); 1050} 1051 1052/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 1053/// an integer. Load the individual pieces to form the aggregate value. 1054void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 1055 SmallVector<AllocaInst*, 32> &NewElts) { 1056 // Extract each element out of the NewElts according to its structure offset 1057 // and form the result value. 1058 const Type *AllocaEltTy = AI->getAllocatedType(); 1059 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1060 1061 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 1062 << '\n'); 1063 1064 // There are two forms here: AI could be an array or struct. Both cases 1065 // have different ways to compute the element offset. 1066 const StructLayout *Layout = 0; 1067 uint64_t ArrayEltBitOffset = 0; 1068 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1069 Layout = TD->getStructLayout(EltSTy); 1070 } else { 1071 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 1072 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1073 } 1074 1075 Value *ResultVal = 1076 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 1077 1078 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1079 // Load the value from the alloca. If the NewElt is an aggregate, cast 1080 // the pointer to an integer of the same size before doing the load. 1081 Value *SrcField = NewElts[i]; 1082 const Type *FieldTy = 1083 cast<PointerType>(SrcField->getType())->getElementType(); 1084 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1085 1086 // Ignore zero sized fields like {}, they obviously contain no data. 1087 if (FieldSizeBits == 0) continue; 1088 1089 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 1090 FieldSizeBits); 1091 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 1092 !FieldTy->isVectorTy()) 1093 SrcField = new BitCastInst(SrcField, 1094 PointerType::getUnqual(FieldIntTy), 1095 "", LI); 1096 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 1097 1098 // If SrcField is a fp or vector of the right size but that isn't an 1099 // integer type, bitcast to an integer so we can shift it. 1100 if (SrcField->getType() != FieldIntTy) 1101 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 1102 1103 // Zero extend the field to be the same size as the final alloca so that 1104 // we can shift and insert it. 1105 if (SrcField->getType() != ResultVal->getType()) 1106 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 1107 1108 // Determine the number of bits to shift SrcField. 1109 uint64_t Shift; 1110 if (Layout) // Struct case. 1111 Shift = Layout->getElementOffsetInBits(i); 1112 else // Array case. 1113 Shift = i*ArrayEltBitOffset; 1114 1115 if (TD->isBigEndian()) 1116 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 1117 1118 if (Shift) { 1119 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 1120 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 1121 } 1122 1123 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 1124 } 1125 1126 // Handle tail padding by truncating the result 1127 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 1128 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 1129 1130 LI->replaceAllUsesWith(ResultVal); 1131 DeadInsts.push_back(LI); 1132} 1133 1134/// HasPadding - Return true if the specified type has any structure or 1135/// alignment padding, false otherwise. 1136static bool HasPadding(const Type *Ty, const TargetData &TD) { 1137 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 1138 const StructLayout *SL = TD.getStructLayout(STy); 1139 unsigned PrevFieldBitOffset = 0; 1140 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1141 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 1142 1143 // Padding in sub-elements? 1144 if (HasPadding(STy->getElementType(i), TD)) 1145 return true; 1146 1147 // Check to see if there is any padding between this element and the 1148 // previous one. 1149 if (i) { 1150 unsigned PrevFieldEnd = 1151 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 1152 if (PrevFieldEnd < FieldBitOffset) 1153 return true; 1154 } 1155 1156 PrevFieldBitOffset = FieldBitOffset; 1157 } 1158 1159 // Check for tail padding. 1160 if (unsigned EltCount = STy->getNumElements()) { 1161 unsigned PrevFieldEnd = PrevFieldBitOffset + 1162 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 1163 if (PrevFieldEnd < SL->getSizeInBits()) 1164 return true; 1165 } 1166 1167 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1168 return HasPadding(ATy->getElementType(), TD); 1169 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) { 1170 return HasPadding(VTy->getElementType(), TD); 1171 } 1172 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 1173} 1174 1175/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 1176/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 1177/// or 1 if safe after canonicalization has been performed. 1178bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 1179 // Loop over the use list of the alloca. We can only transform it if all of 1180 // the users are safe to transform. 1181 AllocaInfo Info; 1182 1183 isSafeForScalarRepl(AI, AI, 0, Info); 1184 if (Info.isUnsafe) { 1185 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 1186 return false; 1187 } 1188 1189 // Okay, we know all the users are promotable. If the aggregate is a memcpy 1190 // source and destination, we have to be careful. In particular, the memcpy 1191 // could be moving around elements that live in structure padding of the LLVM 1192 // types, but may actually be used. In these cases, we refuse to promote the 1193 // struct. 1194 if (Info.isMemCpySrc && Info.isMemCpyDst && 1195 HasPadding(AI->getAllocatedType(), *TD)) 1196 return false; 1197 1198 return true; 1199} 1200 1201/// MergeInType - Add the 'In' type to the accumulated type (Accum) so far at 1202/// the offset specified by Offset (which is specified in bytes). 1203/// 1204/// There are two cases we handle here: 1205/// 1) A union of vector types of the same size and potentially its elements. 1206/// Here we turn element accesses into insert/extract element operations. 1207/// This promotes a <4 x float> with a store of float to the third element 1208/// into a <4 x float> that uses insert element. 1209/// 2) A fully general blob of memory, which we turn into some (potentially 1210/// large) integer type with extract and insert operations where the loads 1211/// and stores would mutate the memory. 1212static void MergeInType(const Type *In, uint64_t Offset, 1213 ConvertToScalarInfo &ConvertInfo, const TargetData &TD){ 1214 // Remember if we saw a vector type. 1215 ConvertInfo.HadAVector |= In->isVectorTy(); 1216 1217 if (ConvertInfo.VectorTy && ConvertInfo.VectorTy->isVoidTy()) 1218 return; 1219 1220 // If this could be contributing to a vector, analyze it. 1221 1222 // If the In type is a vector that is the same size as the alloca, see if it 1223 // matches the existing VecTy. 1224 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 1225 if (VInTy->getBitWidth()/8 == ConvertInfo.AllocaSize && Offset == 0) { 1226 // If we're storing/loading a vector of the right size, allow it as a 1227 // vector. If this the first vector we see, remember the type so that 1228 // we know the element size. 1229 if (ConvertInfo.VectorTy == 0) 1230 ConvertInfo.VectorTy = VInTy; 1231 return; 1232 } 1233 } else if (In->isFloatTy() || In->isDoubleTy() || 1234 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 && 1235 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 1236 // If we're accessing something that could be an element of a vector, see 1237 // if the implied vector agrees with what we already have and if Offset is 1238 // compatible with it. 1239 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 1240 if (Offset % EltSize == 0 && 1241 ConvertInfo.AllocaSize % EltSize == 0 && 1242 (ConvertInfo.VectorTy == 0 || 1243 cast<VectorType>(ConvertInfo.VectorTy)->getElementType() 1244 ->getPrimitiveSizeInBits()/8 == EltSize)) { 1245 if (ConvertInfo.VectorTy == 0) 1246 ConvertInfo.VectorTy = VectorType::get(In, 1247 ConvertInfo.AllocaSize/EltSize); 1248 return; 1249 } 1250 } 1251 1252 // Otherwise, we have a case that we can't handle with an optimized vector 1253 // form. We can still turn this into a large integer. 1254 ConvertInfo.VectorTy = Type::getVoidTy(In->getContext()); 1255} 1256 1257/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 1258/// its accesses to a single vector type, return true and set VecTy to 1259/// the new type. If we could convert the alloca into a single promotable 1260/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 1261/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 1262/// is the current offset from the base of the alloca being analyzed. 1263/// 1264/// If we see at least one access to the value that is as a vector type, set the 1265/// SawVec flag. 1266bool SROA::CanConvertToScalar(Value *V, ConvertToScalarInfo &ConvertInfo, 1267 uint64_t Offset) { 1268 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1269 Instruction *User = cast<Instruction>(*UI); 1270 1271 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1272 // Don't break volatile loads. 1273 if (LI->isVolatile()) 1274 return false; 1275 MergeInType(LI->getType(), Offset, ConvertInfo, *TD); 1276 continue; 1277 } 1278 1279 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1280 // Storing the pointer, not into the value? 1281 if (SI->getOperand(0) == V || SI->isVolatile()) return false; 1282 MergeInType(SI->getOperand(0)->getType(), Offset, ConvertInfo, *TD); 1283 continue; 1284 } 1285 1286 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 1287 if (!CanConvertToScalar(BCI, ConvertInfo, Offset)) 1288 return false; 1289 ConvertInfo.IsNotTrivial = true; 1290 continue; 1291 } 1292 1293 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1294 // If this is a GEP with a variable indices, we can't handle it. 1295 if (!GEP->hasAllConstantIndices()) 1296 return false; 1297 1298 // Compute the offset that this GEP adds to the pointer. 1299 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1300 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getPointerOperandType(), 1301 &Indices[0], Indices.size()); 1302 // See if all uses can be converted. 1303 if (!CanConvertToScalar(GEP, ConvertInfo, Offset+GEPOffset)) 1304 return false; 1305 ConvertInfo.IsNotTrivial = true; 1306 continue; 1307 } 1308 1309 // If this is a constant sized memset of a constant value (e.g. 0) we can 1310 // handle it. 1311 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 1312 // Store of constant value and constant size. 1313 if (isa<ConstantInt>(MSI->getValue()) && 1314 isa<ConstantInt>(MSI->getLength())) { 1315 ConvertInfo.IsNotTrivial = true; 1316 continue; 1317 } 1318 } 1319 1320 // If this is a memcpy or memmove into or out of the whole allocation, we 1321 // can handle it like a load or store of the scalar type. 1322 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 1323 if (ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength())) 1324 if (Len->getZExtValue() == ConvertInfo.AllocaSize && Offset == 0) { 1325 ConvertInfo.IsNotTrivial = true; 1326 continue; 1327 } 1328 } 1329 1330 // Otherwise, we cannot handle this! 1331 return false; 1332 } 1333 1334 return true; 1335} 1336 1337/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 1338/// directly. This happens when we are converting an "integer union" to a 1339/// single integer scalar, or when we are converting a "vector union" to a 1340/// vector with insert/extractelement instructions. 1341/// 1342/// Offset is an offset from the original alloca, in bits that need to be 1343/// shifted to the right. By the end of this, there should be no uses of Ptr. 1344void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset) { 1345 while (!Ptr->use_empty()) { 1346 Instruction *User = cast<Instruction>(Ptr->use_back()); 1347 1348 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 1349 ConvertUsesToScalar(CI, NewAI, Offset); 1350 CI->eraseFromParent(); 1351 continue; 1352 } 1353 1354 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1355 // Compute the offset that this GEP adds to the pointer. 1356 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1357 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getPointerOperandType(), 1358 &Indices[0], Indices.size()); 1359 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 1360 GEP->eraseFromParent(); 1361 continue; 1362 } 1363 1364 IRBuilder<> Builder(User->getParent(), User); 1365 1366 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1367 // The load is a bit extract from NewAI shifted right by Offset bits. 1368 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 1369 Value *NewLoadVal 1370 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 1371 LI->replaceAllUsesWith(NewLoadVal); 1372 LI->eraseFromParent(); 1373 continue; 1374 } 1375 1376 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1377 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 1378 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 1379 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 1380 Builder); 1381 Builder.CreateStore(New, NewAI); 1382 SI->eraseFromParent(); 1383 1384 // If the load we just inserted is now dead, then the inserted store 1385 // overwrote the entire thing. 1386 if (Old->use_empty()) 1387 Old->eraseFromParent(); 1388 continue; 1389 } 1390 1391 // If this is a constant sized memset of a constant value (e.g. 0) we can 1392 // transform it into a store of the expanded constant value. 1393 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 1394 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 1395 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 1396 if (NumBytes != 0) { 1397 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 1398 1399 // Compute the value replicated the right number of times. 1400 APInt APVal(NumBytes*8, Val); 1401 1402 // Splat the value if non-zero. 1403 if (Val) 1404 for (unsigned i = 1; i != NumBytes; ++i) 1405 APVal |= APVal << 8; 1406 1407 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 1408 Value *New = ConvertScalar_InsertValue( 1409 ConstantInt::get(User->getContext(), APVal), 1410 Old, Offset, Builder); 1411 Builder.CreateStore(New, NewAI); 1412 1413 // If the load we just inserted is now dead, then the memset overwrote 1414 // the entire thing. 1415 if (Old->use_empty()) 1416 Old->eraseFromParent(); 1417 } 1418 MSI->eraseFromParent(); 1419 continue; 1420 } 1421 1422 // If this is a memcpy or memmove into or out of the whole allocation, we 1423 // can handle it like a load or store of the scalar type. 1424 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 1425 assert(Offset == 0 && "must be store to start of alloca"); 1426 1427 // If the source and destination are both to the same alloca, then this is 1428 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 1429 // as appropriate. 1430 AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject(0)); 1431 1432 if (MTI->getSource()->getUnderlyingObject(0) != OrigAI) { 1433 // Dest must be OrigAI, change this to be a load from the original 1434 // pointer (bitcasted), then a store to our new alloca. 1435 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 1436 Value *SrcPtr = MTI->getSource(); 1437 SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType()); 1438 1439 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 1440 SrcVal->setAlignment(MTI->getAlignment()); 1441 Builder.CreateStore(SrcVal, NewAI); 1442 } else if (MTI->getDest()->getUnderlyingObject(0) != OrigAI) { 1443 // Src must be OrigAI, change this to be a load from NewAI then a store 1444 // through the original dest pointer (bitcasted). 1445 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 1446 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 1447 1448 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType()); 1449 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 1450 NewStore->setAlignment(MTI->getAlignment()); 1451 } else { 1452 // Noop transfer. Src == Dst 1453 } 1454 1455 MTI->eraseFromParent(); 1456 continue; 1457 } 1458 1459 llvm_unreachable("Unsupported operation!"); 1460 } 1461} 1462 1463/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 1464/// or vector value FromVal, extracting the bits from the offset specified by 1465/// Offset. This returns the value, which is of type ToType. 1466/// 1467/// This happens when we are converting an "integer union" to a single 1468/// integer scalar, or when we are converting a "vector union" to a vector with 1469/// insert/extractelement instructions. 1470/// 1471/// Offset is an offset from the original alloca, in bits that need to be 1472/// shifted to the right. 1473Value *SROA::ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 1474 uint64_t Offset, IRBuilder<> &Builder) { 1475 // If the load is of the whole new alloca, no conversion is needed. 1476 if (FromVal->getType() == ToType && Offset == 0) 1477 return FromVal; 1478 1479 // If the result alloca is a vector type, this is either an element 1480 // access or a bitcast to another vector type of the same size. 1481 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) { 1482 if (ToType->isVectorTy()) 1483 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 1484 1485 // Otherwise it must be an element access. 1486 unsigned Elt = 0; 1487 if (Offset) { 1488 unsigned EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 1489 Elt = Offset/EltSize; 1490 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 1491 } 1492 // Return the element extracted out of it. 1493 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get( 1494 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp"); 1495 if (V->getType() != ToType) 1496 V = Builder.CreateBitCast(V, ToType, "tmp"); 1497 return V; 1498 } 1499 1500 // If ToType is a first class aggregate, extract out each of the pieces and 1501 // use insertvalue's to form the FCA. 1502 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 1503 const StructLayout &Layout = *TD->getStructLayout(ST); 1504 Value *Res = UndefValue::get(ST); 1505 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1506 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 1507 Offset+Layout.getElementOffsetInBits(i), 1508 Builder); 1509 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1510 } 1511 return Res; 1512 } 1513 1514 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 1515 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 1516 Value *Res = UndefValue::get(AT); 1517 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1518 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 1519 Offset+i*EltSize, Builder); 1520 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1521 } 1522 return Res; 1523 } 1524 1525 // Otherwise, this must be a union that was converted to an integer value. 1526 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 1527 1528 // If this is a big-endian system and the load is narrower than the 1529 // full alloca type, we need to do a shift to get the right bits. 1530 int ShAmt = 0; 1531 if (TD->isBigEndian()) { 1532 // On big-endian machines, the lowest bit is stored at the bit offset 1533 // from the pointer given by getTypeStoreSizeInBits. This matters for 1534 // integers with a bitwidth that is not a multiple of 8. 1535 ShAmt = TD->getTypeStoreSizeInBits(NTy) - 1536 TD->getTypeStoreSizeInBits(ToType) - Offset; 1537 } else { 1538 ShAmt = Offset; 1539 } 1540 1541 // Note: we support negative bitwidths (with shl) which are not defined. 1542 // We do this to support (f.e.) loads off the end of a structure where 1543 // only some bits are used. 1544 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 1545 FromVal = Builder.CreateLShr(FromVal, 1546 ConstantInt::get(FromVal->getType(), 1547 ShAmt), "tmp"); 1548 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 1549 FromVal = Builder.CreateShl(FromVal, 1550 ConstantInt::get(FromVal->getType(), 1551 -ShAmt), "tmp"); 1552 1553 // Finally, unconditionally truncate the integer to the right width. 1554 unsigned LIBitWidth = TD->getTypeSizeInBits(ToType); 1555 if (LIBitWidth < NTy->getBitWidth()) 1556 FromVal = 1557 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 1558 LIBitWidth), "tmp"); 1559 else if (LIBitWidth > NTy->getBitWidth()) 1560 FromVal = 1561 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 1562 LIBitWidth), "tmp"); 1563 1564 // If the result is an integer, this is a trunc or bitcast. 1565 if (ToType->isIntegerTy()) { 1566 // Should be done. 1567 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 1568 // Just do a bitcast, we know the sizes match up. 1569 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 1570 } else { 1571 // Otherwise must be a pointer. 1572 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 1573 } 1574 assert(FromVal->getType() == ToType && "Didn't convert right?"); 1575 return FromVal; 1576} 1577 1578/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 1579/// or vector value "Old" at the offset specified by Offset. 1580/// 1581/// This happens when we are converting an "integer union" to a 1582/// single integer scalar, or when we are converting a "vector union" to a 1583/// vector with insert/extractelement instructions. 1584/// 1585/// Offset is an offset from the original alloca, in bits that need to be 1586/// shifted to the right. 1587Value *SROA::ConvertScalar_InsertValue(Value *SV, Value *Old, 1588 uint64_t Offset, IRBuilder<> &Builder) { 1589 1590 // Convert the stored type to the actual type, shift it left to insert 1591 // then 'or' into place. 1592 const Type *AllocaType = Old->getType(); 1593 LLVMContext &Context = Old->getContext(); 1594 1595 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 1596 uint64_t VecSize = TD->getTypeAllocSizeInBits(VTy); 1597 uint64_t ValSize = TD->getTypeAllocSizeInBits(SV->getType()); 1598 1599 // Changing the whole vector with memset or with an access of a different 1600 // vector type? 1601 if (ValSize == VecSize) 1602 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 1603 1604 uint64_t EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 1605 1606 // Must be an element insertion. 1607 unsigned Elt = Offset/EltSize; 1608 1609 if (SV->getType() != VTy->getElementType()) 1610 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 1611 1612 SV = Builder.CreateInsertElement(Old, SV, 1613 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt), 1614 "tmp"); 1615 return SV; 1616 } 1617 1618 // If SV is a first-class aggregate value, insert each value recursively. 1619 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 1620 const StructLayout &Layout = *TD->getStructLayout(ST); 1621 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1622 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1623 Old = ConvertScalar_InsertValue(Elt, Old, 1624 Offset+Layout.getElementOffsetInBits(i), 1625 Builder); 1626 } 1627 return Old; 1628 } 1629 1630 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 1631 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 1632 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1633 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1634 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 1635 } 1636 return Old; 1637 } 1638 1639 // If SV is a float, convert it to the appropriate integer type. 1640 // If it is a pointer, do the same. 1641 unsigned SrcWidth = TD->getTypeSizeInBits(SV->getType()); 1642 unsigned DestWidth = TD->getTypeSizeInBits(AllocaType); 1643 unsigned SrcStoreWidth = TD->getTypeStoreSizeInBits(SV->getType()); 1644 unsigned DestStoreWidth = TD->getTypeStoreSizeInBits(AllocaType); 1645 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 1646 SV = Builder.CreateBitCast(SV, 1647 IntegerType::get(SV->getContext(),SrcWidth), "tmp"); 1648 else if (SV->getType()->isPointerTy()) 1649 SV = Builder.CreatePtrToInt(SV, TD->getIntPtrType(SV->getContext()), "tmp"); 1650 1651 // Zero extend or truncate the value if needed. 1652 if (SV->getType() != AllocaType) { 1653 if (SV->getType()->getPrimitiveSizeInBits() < 1654 AllocaType->getPrimitiveSizeInBits()) 1655 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 1656 else { 1657 // Truncation may be needed if storing more than the alloca can hold 1658 // (undefined behavior). 1659 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 1660 SrcWidth = DestWidth; 1661 SrcStoreWidth = DestStoreWidth; 1662 } 1663 } 1664 1665 // If this is a big-endian system and the store is narrower than the 1666 // full alloca type, we need to do a shift to get the right bits. 1667 int ShAmt = 0; 1668 if (TD->isBigEndian()) { 1669 // On big-endian machines, the lowest bit is stored at the bit offset 1670 // from the pointer given by getTypeStoreSizeInBits. This matters for 1671 // integers with a bitwidth that is not a multiple of 8. 1672 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 1673 } else { 1674 ShAmt = Offset; 1675 } 1676 1677 // Note: we support negative bitwidths (with shr) which are not defined. 1678 // We do this to support (f.e.) stores off the end of a structure where 1679 // only some bits in the structure are set. 1680 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 1681 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 1682 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 1683 ShAmt), "tmp"); 1684 Mask <<= ShAmt; 1685 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1686 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 1687 -ShAmt), "tmp"); 1688 Mask = Mask.lshr(-ShAmt); 1689 } 1690 1691 // Mask out the bits we are about to insert from the old value, and or 1692 // in the new bits. 1693 if (SrcWidth != DestWidth) { 1694 assert(DestWidth > SrcWidth); 1695 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 1696 SV = Builder.CreateOr(Old, SV, "ins"); 1697 } 1698 return SV; 1699} 1700 1701 1702 1703/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1704/// some part of a constant global variable. This intentionally only accepts 1705/// constant expressions because we don't can't rewrite arbitrary instructions. 1706static bool PointsToConstantGlobal(Value *V) { 1707 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1708 return GV->isConstant(); 1709 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1710 if (CE->getOpcode() == Instruction::BitCast || 1711 CE->getOpcode() == Instruction::GetElementPtr) 1712 return PointsToConstantGlobal(CE->getOperand(0)); 1713 return false; 1714} 1715 1716/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1717/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1718/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1719/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1720/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1721/// the alloca, and if the source pointer is a pointer to a constant global, we 1722/// can optimize this. 1723static bool isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy, 1724 bool isOffset) { 1725 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1726 User *U = cast<Instruction>(*UI); 1727 1728 if (LoadInst *LI = dyn_cast<LoadInst>(U)) 1729 // Ignore non-volatile loads, they are always ok. 1730 if (!LI->isVolatile()) 1731 continue; 1732 1733 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 1734 // If uses of the bitcast are ok, we are ok. 1735 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 1736 return false; 1737 continue; 1738 } 1739 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 1740 // If the GEP has all zero indices, it doesn't offset the pointer. If it 1741 // doesn't, it does. 1742 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 1743 isOffset || !GEP->hasAllZeroIndices())) 1744 return false; 1745 continue; 1746 } 1747 1748 // If this is isn't our memcpy/memmove, reject it as something we can't 1749 // handle. 1750 MemTransferInst *MI = dyn_cast<MemTransferInst>(U); 1751 if (MI == 0) 1752 return false; 1753 1754 // If we already have seen a copy, reject the second one. 1755 if (TheCopy) return false; 1756 1757 // If the pointer has been offset from the start of the alloca, we can't 1758 // safely handle this. 1759 if (isOffset) return false; 1760 1761 // If the memintrinsic isn't using the alloca as the dest, reject it. 1762 if (UI.getOperandNo() != 0) return false; 1763 1764 // If the source of the memcpy/move is not a constant global, reject it. 1765 if (!PointsToConstantGlobal(MI->getSource())) 1766 return false; 1767 1768 // Otherwise, the transform is safe. Remember the copy instruction. 1769 TheCopy = MI; 1770 } 1771 return true; 1772} 1773 1774/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 1775/// modified by a copy from a constant global. If we can prove this, we can 1776/// replace any uses of the alloca with uses of the global directly. 1777MemTransferInst *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) { 1778 MemTransferInst *TheCopy = 0; 1779 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 1780 return TheCopy; 1781 return 0; 1782} 1783