ScalarReplAggregates.cpp revision 9ee17208115482441953127615231c59a2f4d052
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 SROA : public FunctionPass { 53 static char ID; // Pass identification, replacement for typeid 54 explicit SROA(signed T = -1) : FunctionPass(&ID) { 55 if (T == -1) 56 SRThreshold = 128; 57 else 58 SRThreshold = T; 59 } 60 61 bool runOnFunction(Function &F); 62 63 bool performScalarRepl(Function &F); 64 bool performPromotion(Function &F); 65 66 // getAnalysisUsage - This pass does not require any passes, but we know it 67 // will not alter the CFG, so say so. 68 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 69 AU.addRequired<DominatorTree>(); 70 AU.addRequired<DominanceFrontier>(); 71 AU.setPreservesCFG(); 72 } 73 74 private: 75 TargetData *TD; 76 77 /// DeadInsts - Keep track of instructions we have made dead, so that 78 /// we can remove them after we are done working. 79 SmallVector<Value*, 32> DeadInsts; 80 81 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures 82 /// information about the uses. All these fields are initialized to false 83 /// and set to true when something is learned. 84 struct AllocaInfo { 85 /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 86 bool isUnsafe : 1; 87 88 /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 89 bool isMemCpySrc : 1; 90 91 /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 92 bool isMemCpyDst : 1; 93 94 AllocaInfo() 95 : isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false) {} 96 }; 97 98 unsigned SRThreshold; 99 100 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; } 101 102 bool isSafeAllocaToScalarRepl(AllocaInst *AI); 103 104 void isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 105 AllocaInfo &Info); 106 void isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t &Offset, 107 AllocaInfo &Info); 108 void isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize, 109 const Type *MemOpType, bool isStore, AllocaInfo &Info); 110 bool TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size); 111 uint64_t FindElementAndOffset(const Type *&T, uint64_t &Offset, 112 const Type *&IdxTy); 113 114 void DoScalarReplacement(AllocaInst *AI, 115 std::vector<AllocaInst*> &WorkList); 116 void DeleteDeadInstructions(); 117 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocaInst *Base); 118 119 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 120 SmallVector<AllocaInst*, 32> &NewElts); 121 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 122 SmallVector<AllocaInst*, 32> &NewElts); 123 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 124 SmallVector<AllocaInst*, 32> &NewElts); 125 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 126 AllocaInst *AI, 127 SmallVector<AllocaInst*, 32> &NewElts); 128 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 129 SmallVector<AllocaInst*, 32> &NewElts); 130 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 131 SmallVector<AllocaInst*, 32> &NewElts); 132 133 bool CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy, 134 bool &SawVec, uint64_t Offset, unsigned AllocaSize); 135 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); 136 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType, 137 uint64_t Offset, IRBuilder<> &Builder); 138 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 139 uint64_t Offset, IRBuilder<> &Builder); 140 static Instruction *isOnlyCopiedFromConstantGlobal(AllocaInst *AI); 141 }; 142} 143 144char SROA::ID = 0; 145static RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates"); 146 147// Public interface to the ScalarReplAggregates pass 148FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) { 149 return new SROA(Threshold); 150} 151 152 153bool SROA::runOnFunction(Function &F) { 154 TD = getAnalysisIfAvailable<TargetData>(); 155 156 bool Changed = performPromotion(F); 157 158 // FIXME: ScalarRepl currently depends on TargetData more than it 159 // theoretically needs to. It should be refactored in order to support 160 // target-independent IR. Until this is done, just skip the actual 161 // scalar-replacement portion of this pass. 162 if (!TD) return Changed; 163 164 while (1) { 165 bool LocalChange = performScalarRepl(F); 166 if (!LocalChange) break; // No need to repromote if no scalarrepl 167 Changed = true; 168 LocalChange = performPromotion(F); 169 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 170 } 171 172 return Changed; 173} 174 175 176bool SROA::performPromotion(Function &F) { 177 std::vector<AllocaInst*> Allocas; 178 DominatorTree &DT = getAnalysis<DominatorTree>(); 179 DominanceFrontier &DF = getAnalysis<DominanceFrontier>(); 180 181 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 182 183 bool Changed = false; 184 185 while (1) { 186 Allocas.clear(); 187 188 // Find allocas that are safe to promote, by looking at all instructions in 189 // the entry node 190 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 191 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 192 if (isAllocaPromotable(AI)) 193 Allocas.push_back(AI); 194 195 if (Allocas.empty()) break; 196 197 PromoteMemToReg(Allocas, DT, DF); 198 NumPromoted += Allocas.size(); 199 Changed = true; 200 } 201 202 return Changed; 203} 204 205/// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for 206/// SROA. It must be a struct or array type with a small number of elements. 207static bool ShouldAttemptScalarRepl(AllocaInst *AI) { 208 const Type *T = AI->getAllocatedType(); 209 // Do not promote any struct into more than 32 separate vars. 210 if (const StructType *ST = dyn_cast<StructType>(T)) 211 return ST->getNumElements() <= 32; 212 // Arrays are much less likely to be safe for SROA; only consider 213 // them if they are very small. 214 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) 215 return AT->getNumElements() <= 8; 216 return false; 217} 218 219// performScalarRepl - This algorithm is a simple worklist driven algorithm, 220// which runs on all of the malloc/alloca instructions in the function, removing 221// them if they are only used by getelementptr instructions. 222// 223bool SROA::performScalarRepl(Function &F) { 224 std::vector<AllocaInst*> WorkList; 225 226 // Scan the entry basic block, adding any alloca's and mallocs to the worklist 227 BasicBlock &BB = F.getEntryBlock(); 228 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 229 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 230 WorkList.push_back(A); 231 232 // Process the worklist 233 bool Changed = false; 234 while (!WorkList.empty()) { 235 AllocaInst *AI = WorkList.back(); 236 WorkList.pop_back(); 237 238 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 239 // with unused elements. 240 if (AI->use_empty()) { 241 AI->eraseFromParent(); 242 continue; 243 } 244 245 // If this alloca is impossible for us to promote, reject it early. 246 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 247 continue; 248 249 // Check to see if this allocation is only modified by a memcpy/memmove from 250 // a constant global. If this is the case, we can change all users to use 251 // the constant global instead. This is commonly produced by the CFE by 252 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 253 // is only subsequently read. 254 if (Instruction *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 255 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n'); 256 DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n'); 257 Constant *TheSrc = cast<Constant>(TheCopy->getOperand(2)); 258 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 259 TheCopy->eraseFromParent(); // Don't mutate the global. 260 AI->eraseFromParent(); 261 ++NumGlobals; 262 Changed = true; 263 continue; 264 } 265 266 // Check to see if we can perform the core SROA transformation. We cannot 267 // transform the allocation instruction if it is an array allocation 268 // (allocations OF arrays are ok though), and an allocation of a scalar 269 // value cannot be decomposed at all. 270 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 271 272 // Do not promote [0 x %struct]. 273 if (AllocaSize == 0) continue; 274 275 // If the alloca looks like a good candidate for scalar replacement, and if 276 // all its users can be transformed, then split up the aggregate into its 277 // separate elements. 278 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) { 279 DoScalarReplacement(AI, WorkList); 280 Changed = true; 281 continue; 282 } 283 284 // Do not promote any struct whose size is too big. 285 if (AllocaSize > SRThreshold) continue; 286 287 // If we can turn this aggregate value (potentially with casts) into a 288 // simple scalar value that can be mem2reg'd into a register value. 289 // IsNotTrivial tracks whether this is something that mem2reg could have 290 // promoted itself. If so, we don't want to transform it needlessly. Note 291 // that we can't just check based on the type: the alloca may be of an i32 292 // but that has pointer arithmetic to set byte 3 of it or something. 293 bool IsNotTrivial = false; 294 const Type *VectorTy = 0; 295 bool HadAVector = false; 296 if (CanConvertToScalar(AI, IsNotTrivial, VectorTy, HadAVector, 297 0, unsigned(AllocaSize)) && IsNotTrivial) { 298 AllocaInst *NewAI; 299 // If we were able to find a vector type that can handle this with 300 // insert/extract elements, and if there was at least one use that had 301 // a vector type, promote this to a vector. We don't want to promote 302 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 303 // we just get a lot of insert/extracts. If at least one vector is 304 // involved, then we probably really do have a union of vector/array. 305 if (VectorTy && VectorTy->isVectorTy() && HadAVector) { 306 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = " 307 << *VectorTy << '\n'); 308 309 // Create and insert the vector alloca. 310 NewAI = new AllocaInst(VectorTy, 0, "", AI->getParent()->begin()); 311 ConvertUsesToScalar(AI, NewAI, 0); 312 } else { 313 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"); 314 315 // Create and insert the integer alloca. 316 const Type *NewTy = IntegerType::get(AI->getContext(), AllocaSize*8); 317 NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 318 ConvertUsesToScalar(AI, NewAI, 0); 319 } 320 NewAI->takeName(AI); 321 AI->eraseFromParent(); 322 ++NumConverted; 323 Changed = true; 324 continue; 325 } 326 327 // Otherwise, couldn't process this alloca. 328 } 329 330 return Changed; 331} 332 333/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 334/// predicate, do SROA now. 335void SROA::DoScalarReplacement(AllocaInst *AI, 336 std::vector<AllocaInst*> &WorkList) { 337 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n'); 338 SmallVector<AllocaInst*, 32> ElementAllocas; 339 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 340 ElementAllocas.reserve(ST->getNumContainedTypes()); 341 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 342 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 343 AI->getAlignment(), 344 AI->getName() + "." + Twine(i), AI); 345 ElementAllocas.push_back(NA); 346 WorkList.push_back(NA); // Add to worklist for recursive processing 347 } 348 } else { 349 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 350 ElementAllocas.reserve(AT->getNumElements()); 351 const Type *ElTy = AT->getElementType(); 352 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 353 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 354 AI->getName() + "." + Twine(i), AI); 355 ElementAllocas.push_back(NA); 356 WorkList.push_back(NA); // Add to worklist for recursive processing 357 } 358 } 359 360 // Now that we have created the new alloca instructions, rewrite all the 361 // uses of the old alloca. 362 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 363 364 // Now erase any instructions that were made dead while rewriting the alloca. 365 DeleteDeadInstructions(); 366 AI->eraseFromParent(); 367 368 NumReplaced++; 369} 370 371/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 372/// recursively including all their operands that become trivially dead. 373void SROA::DeleteDeadInstructions() { 374 while (!DeadInsts.empty()) { 375 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 376 377 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 378 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 379 // Zero out the operand and see if it becomes trivially dead. 380 // (But, don't add allocas to the dead instruction list -- they are 381 // already on the worklist and will be deleted separately.) 382 *OI = 0; 383 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 384 DeadInsts.push_back(U); 385 } 386 387 I->eraseFromParent(); 388 } 389} 390 391/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 392/// performing scalar replacement of alloca AI. The results are flagged in 393/// the Info parameter. Offset indicates the position within AI that is 394/// referenced by this instruction. 395void SROA::isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 396 AllocaInfo &Info) { 397 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 398 Instruction *User = cast<Instruction>(*UI); 399 400 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 401 isSafeForScalarRepl(BC, AI, Offset, Info); 402 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 403 uint64_t GEPOffset = Offset; 404 isSafeGEP(GEPI, AI, GEPOffset, Info); 405 if (!Info.isUnsafe) 406 isSafeForScalarRepl(GEPI, AI, GEPOffset, Info); 407 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) { 408 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 409 if (Length) 410 isSafeMemAccess(AI, Offset, Length->getZExtValue(), 0, 411 UI.getOperandNo() == 1, Info); 412 else 413 MarkUnsafe(Info); 414 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 415 if (!LI->isVolatile()) { 416 const Type *LIType = LI->getType(); 417 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(LIType), 418 LIType, false, Info); 419 } else 420 MarkUnsafe(Info); 421 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 422 // Store is ok if storing INTO the pointer, not storing the pointer 423 if (!SI->isVolatile() && SI->getOperand(0) != I) { 424 const Type *SIType = SI->getOperand(0)->getType(); 425 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(SIType), 426 SIType, true, Info); 427 } else 428 MarkUnsafe(Info); 429 } else { 430 DEBUG(errs() << " Transformation preventing inst: " << *User << '\n'); 431 MarkUnsafe(Info); 432 } 433 if (Info.isUnsafe) return; 434 } 435} 436 437/// isSafeGEP - Check if a GEP instruction can be handled for scalar 438/// replacement. It is safe when all the indices are constant, in-bounds 439/// references, and when the resulting offset corresponds to an element within 440/// the alloca type. The results are flagged in the Info parameter. Upon 441/// return, Offset is adjusted as specified by the GEP indices. 442void SROA::isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, 443 uint64_t &Offset, AllocaInfo &Info) { 444 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 445 if (GEPIt == E) 446 return; 447 448 // Walk through the GEP type indices, checking the types that this indexes 449 // into. 450 for (; GEPIt != E; ++GEPIt) { 451 // Ignore struct elements, no extra checking needed for these. 452 if ((*GEPIt)->isStructTy()) 453 continue; 454 455 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 456 if (!IdxVal) 457 return MarkUnsafe(Info); 458 } 459 460 // Compute the offset due to this GEP and check if the alloca has a 461 // component element at that offset. 462 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 463 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 464 &Indices[0], Indices.size()); 465 if (!TypeHasComponent(AI->getAllocatedType(), Offset, 0)) 466 MarkUnsafe(Info); 467} 468 469/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 470/// alloca or has an offset and size that corresponds to a component element 471/// within it. The offset checked here may have been formed from a GEP with a 472/// pointer bitcasted to a different type. 473void SROA::isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize, 474 const Type *MemOpType, bool isStore, 475 AllocaInfo &Info) { 476 // Check if this is a load/store of the entire alloca. 477 if (Offset == 0 && MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) { 478 bool UsesAggregateType = (MemOpType == AI->getAllocatedType()); 479 // This is safe for MemIntrinsics (where MemOpType is 0), integer types 480 // (which are essentially the same as the MemIntrinsics, especially with 481 // regard to copying padding between elements), or references using the 482 // aggregate type of the alloca. 483 if (!MemOpType || MemOpType->isIntegerTy() || UsesAggregateType) { 484 if (!UsesAggregateType) { 485 if (isStore) 486 Info.isMemCpyDst = true; 487 else 488 Info.isMemCpySrc = true; 489 } 490 return; 491 } 492 } 493 // Check if the offset/size correspond to a component within the alloca type. 494 const Type *T = AI->getAllocatedType(); 495 if (TypeHasComponent(T, Offset, MemSize)) 496 return; 497 498 return MarkUnsafe(Info); 499} 500 501/// TypeHasComponent - Return true if T has a component type with the 502/// specified offset and size. If Size is zero, do not check the size. 503bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) { 504 const Type *EltTy; 505 uint64_t EltSize; 506 if (const StructType *ST = dyn_cast<StructType>(T)) { 507 const StructLayout *Layout = TD->getStructLayout(ST); 508 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 509 EltTy = ST->getContainedType(EltIdx); 510 EltSize = TD->getTypeAllocSize(EltTy); 511 Offset -= Layout->getElementOffset(EltIdx); 512 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 513 EltTy = AT->getElementType(); 514 EltSize = TD->getTypeAllocSize(EltTy); 515 if (Offset >= AT->getNumElements() * EltSize) 516 return false; 517 Offset %= EltSize; 518 } else { 519 return false; 520 } 521 if (Offset == 0 && (Size == 0 || EltSize == Size)) 522 return true; 523 // Check if the component spans multiple elements. 524 if (Offset + Size > EltSize) 525 return false; 526 return TypeHasComponent(EltTy, Offset, Size); 527} 528 529/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 530/// the instruction I, which references it, to use the separate elements. 531/// Offset indicates the position within AI that is referenced by this 532/// instruction. 533void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 534 SmallVector<AllocaInst*, 32> &NewElts) { 535 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 536 Instruction *User = cast<Instruction>(*UI); 537 538 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 539 RewriteBitCast(BC, AI, Offset, NewElts); 540 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 541 RewriteGEP(GEPI, AI, Offset, NewElts); 542 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 543 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 544 uint64_t MemSize = Length->getZExtValue(); 545 if (Offset == 0 && 546 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 547 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 548 // Otherwise the intrinsic can only touch a single element and the 549 // address operand will be updated, so nothing else needs to be done. 550 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 551 const Type *LIType = LI->getType(); 552 if (LIType == AI->getAllocatedType()) { 553 // Replace: 554 // %res = load { i32, i32 }* %alloc 555 // with: 556 // %load.0 = load i32* %alloc.0 557 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 558 // %load.1 = load i32* %alloc.1 559 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 560 // (Also works for arrays instead of structs) 561 Value *Insert = UndefValue::get(LIType); 562 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 563 Value *Load = new LoadInst(NewElts[i], "load", LI); 564 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 565 } 566 LI->replaceAllUsesWith(Insert); 567 DeadInsts.push_back(LI); 568 } else if (LIType->isIntegerTy() && 569 TD->getTypeAllocSize(LIType) == 570 TD->getTypeAllocSize(AI->getAllocatedType())) { 571 // If this is a load of the entire alloca to an integer, rewrite it. 572 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 573 } 574 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 575 Value *Val = SI->getOperand(0); 576 const Type *SIType = Val->getType(); 577 if (SIType == AI->getAllocatedType()) { 578 // Replace: 579 // store { i32, i32 } %val, { i32, i32 }* %alloc 580 // with: 581 // %val.0 = extractvalue { i32, i32 } %val, 0 582 // store i32 %val.0, i32* %alloc.0 583 // %val.1 = extractvalue { i32, i32 } %val, 1 584 // store i32 %val.1, i32* %alloc.1 585 // (Also works for arrays instead of structs) 586 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 587 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 588 new StoreInst(Extract, NewElts[i], SI); 589 } 590 DeadInsts.push_back(SI); 591 } else if (SIType->isIntegerTy() && 592 TD->getTypeAllocSize(SIType) == 593 TD->getTypeAllocSize(AI->getAllocatedType())) { 594 // If this is a store of the entire alloca from an integer, rewrite it. 595 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 596 } 597 } 598 } 599} 600 601/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 602/// and recursively continue updating all of its uses. 603void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 604 SmallVector<AllocaInst*, 32> &NewElts) { 605 RewriteForScalarRepl(BC, AI, Offset, NewElts); 606 if (BC->getOperand(0) != AI) 607 return; 608 609 // The bitcast references the original alloca. Replace its uses with 610 // references to the first new element alloca. 611 Instruction *Val = NewElts[0]; 612 if (Val->getType() != BC->getDestTy()) { 613 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 614 Val->takeName(BC); 615 } 616 BC->replaceAllUsesWith(Val); 617 DeadInsts.push_back(BC); 618} 619 620/// FindElementAndOffset - Return the index of the element containing Offset 621/// within the specified type, which must be either a struct or an array. 622/// Sets T to the type of the element and Offset to the offset within that 623/// element. IdxTy is set to the type of the index result to be used in a 624/// GEP instruction. 625uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset, 626 const Type *&IdxTy) { 627 uint64_t Idx = 0; 628 if (const StructType *ST = dyn_cast<StructType>(T)) { 629 const StructLayout *Layout = TD->getStructLayout(ST); 630 Idx = Layout->getElementContainingOffset(Offset); 631 T = ST->getContainedType(Idx); 632 Offset -= Layout->getElementOffset(Idx); 633 IdxTy = Type::getInt32Ty(T->getContext()); 634 return Idx; 635 } 636 const ArrayType *AT = cast<ArrayType>(T); 637 T = AT->getElementType(); 638 uint64_t EltSize = TD->getTypeAllocSize(T); 639 Idx = Offset / EltSize; 640 Offset -= Idx * EltSize; 641 IdxTy = Type::getInt64Ty(T->getContext()); 642 return Idx; 643} 644 645/// RewriteGEP - Check if this GEP instruction moves the pointer across 646/// elements of the alloca that are being split apart, and if so, rewrite 647/// the GEP to be relative to the new element. 648void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 649 SmallVector<AllocaInst*, 32> &NewElts) { 650 uint64_t OldOffset = Offset; 651 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 652 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 653 &Indices[0], Indices.size()); 654 655 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 656 657 const Type *T = AI->getAllocatedType(); 658 const Type *IdxTy; 659 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 660 if (GEPI->getOperand(0) == AI) 661 OldIdx = ~0ULL; // Force the GEP to be rewritten. 662 663 T = AI->getAllocatedType(); 664 uint64_t EltOffset = Offset; 665 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 666 667 // If this GEP does not move the pointer across elements of the alloca 668 // being split, then it does not needs to be rewritten. 669 if (Idx == OldIdx) 670 return; 671 672 const Type *i32Ty = Type::getInt32Ty(AI->getContext()); 673 SmallVector<Value*, 8> NewArgs; 674 NewArgs.push_back(Constant::getNullValue(i32Ty)); 675 while (EltOffset != 0) { 676 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 677 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 678 } 679 Instruction *Val = NewElts[Idx]; 680 if (NewArgs.size() > 1) { 681 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(), 682 NewArgs.end(), "", GEPI); 683 Val->takeName(GEPI); 684 } 685 if (Val->getType() != GEPI->getType()) 686 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI); 687 GEPI->replaceAllUsesWith(Val); 688 DeadInsts.push_back(GEPI); 689} 690 691/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 692/// Rewrite it to copy or set the elements of the scalarized memory. 693void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 694 AllocaInst *AI, 695 SmallVector<AllocaInst*, 32> &NewElts) { 696 // If this is a memcpy/memmove, construct the other pointer as the 697 // appropriate type. The "Other" pointer is the pointer that goes to memory 698 // that doesn't have anything to do with the alloca that we are promoting. For 699 // memset, this Value* stays null. 700 Value *OtherPtr = 0; 701 LLVMContext &Context = MI->getContext(); 702 unsigned MemAlignment = MI->getAlignment(); 703 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 704 if (Inst == MTI->getRawDest()) 705 OtherPtr = MTI->getRawSource(); 706 else { 707 assert(Inst == MTI->getRawSource()); 708 OtherPtr = MTI->getRawDest(); 709 } 710 } 711 712 // If there is an other pointer, we want to convert it to the same pointer 713 // type as AI has, so we can GEP through it safely. 714 if (OtherPtr) { 715 716 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 717 // optimization, but it's also required to detect the corner case where 718 // both pointer operands are referencing the same memory, and where 719 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 720 // function is only called for mem intrinsics that access the whole 721 // aggregate, so non-zero GEPs are not an issue here.) 722 while (1) { 723 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr)) { 724 OtherPtr = BC->getOperand(0); 725 continue; 726 } 727 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr)) { 728 // All zero GEPs are effectively bitcasts. 729 if (GEP->hasAllZeroIndices()) { 730 OtherPtr = GEP->getOperand(0); 731 continue; 732 } 733 } 734 break; 735 } 736 // Copying the alloca to itself is a no-op: just delete it. 737 if (OtherPtr == AI || OtherPtr == NewElts[0]) { 738 // This code will run twice for a no-op memcpy -- once for each operand. 739 // Put only one reference to MI on the DeadInsts list. 740 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(), 741 E = DeadInsts.end(); I != E; ++I) 742 if (*I == MI) return; 743 DeadInsts.push_back(MI); 744 return; 745 } 746 747 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr)) 748 if (BCE->getOpcode() == Instruction::BitCast) 749 OtherPtr = BCE->getOperand(0); 750 751 // If the pointer is not the right type, insert a bitcast to the right 752 // type. 753 if (OtherPtr->getType() != AI->getType()) 754 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(), 755 MI); 756 } 757 758 // Process each element of the aggregate. 759 Value *TheFn = MI->getOperand(0); 760 const Type *BytePtrTy = MI->getRawDest()->getType(); 761 bool SROADest = MI->getRawDest() == Inst; 762 763 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 764 765 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 766 // If this is a memcpy/memmove, emit a GEP of the other element address. 767 Value *OtherElt = 0; 768 unsigned OtherEltAlign = MemAlignment; 769 770 if (OtherPtr) { 771 Value *Idx[2] = { Zero, 772 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 773 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2, 774 OtherPtr->getName()+"."+Twine(i), 775 MI); 776 uint64_t EltOffset; 777 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 778 if (const StructType *ST = 779 dyn_cast<StructType>(OtherPtrTy->getElementType())) { 780 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 781 } else { 782 const Type *EltTy = 783 cast<SequentialType>(OtherPtr->getType())->getElementType(); 784 EltOffset = TD->getTypeAllocSize(EltTy)*i; 785 } 786 787 // The alignment of the other pointer is the guaranteed alignment of the 788 // element, which is affected by both the known alignment of the whole 789 // mem intrinsic and the alignment of the element. If the alignment of 790 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 791 // known alignment is just 4 bytes. 792 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 793 } 794 795 Value *EltPtr = NewElts[i]; 796 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 797 798 // If we got down to a scalar, insert a load or store as appropriate. 799 if (EltTy->isSingleValueType()) { 800 if (isa<MemTransferInst>(MI)) { 801 if (SROADest) { 802 // From Other to Alloca. 803 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 804 new StoreInst(Elt, EltPtr, MI); 805 } else { 806 // From Alloca to Other. 807 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 808 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 809 } 810 continue; 811 } 812 assert(isa<MemSetInst>(MI)); 813 814 // If the stored element is zero (common case), just store a null 815 // constant. 816 Constant *StoreVal; 817 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) { 818 if (CI->isZero()) { 819 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 820 } else { 821 // If EltTy is a vector type, get the element type. 822 const Type *ValTy = EltTy->getScalarType(); 823 824 // Construct an integer with the right value. 825 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 826 APInt OneVal(EltSize, CI->getZExtValue()); 827 APInt TotalVal(OneVal); 828 // Set each byte. 829 for (unsigned i = 0; 8*i < EltSize; ++i) { 830 TotalVal = TotalVal.shl(8); 831 TotalVal |= OneVal; 832 } 833 834 // Convert the integer value to the appropriate type. 835 StoreVal = ConstantInt::get(Context, TotalVal); 836 if (ValTy->isPointerTy()) 837 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 838 else if (ValTy->isFloatingPointTy()) 839 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 840 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 841 842 // If the requested value was a vector constant, create it. 843 if (EltTy != ValTy) { 844 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 845 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 846 StoreVal = ConstantVector::get(&Elts[0], NumElts); 847 } 848 } 849 new StoreInst(StoreVal, EltPtr, MI); 850 continue; 851 } 852 // Otherwise, if we're storing a byte variable, use a memset call for 853 // this element. 854 } 855 856 // Cast the element pointer to BytePtrTy. 857 if (EltPtr->getType() != BytePtrTy) 858 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getName(), MI); 859 860 // Cast the other pointer (if we have one) to BytePtrTy. 861 if (OtherElt && OtherElt->getType() != BytePtrTy) { 862 // Preserve address space of OtherElt 863 const PointerType* OtherPTy = cast<PointerType>(OtherElt->getType()); 864 const PointerType* PTy = cast<PointerType>(BytePtrTy); 865 if (OtherPTy->getElementType() != PTy->getElementType()) { 866 Type *NewOtherPTy = PointerType::get(PTy->getElementType(), 867 OtherPTy->getAddressSpace()); 868 OtherElt = new BitCastInst(OtherElt, NewOtherPTy, 869 OtherElt->getNameStr(), MI); 870 } 871 } 872 873 unsigned EltSize = TD->getTypeAllocSize(EltTy); 874 875 // Finally, insert the meminst for this element. 876 if (isa<MemTransferInst>(MI)) { 877 Value *Ops[] = { 878 SROADest ? EltPtr : OtherElt, // Dest ptr 879 SROADest ? OtherElt : EltPtr, // Src ptr 880 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 881 // Align 882 ConstantInt::get(Type::getInt32Ty(MI->getContext()), OtherEltAlign), 883 MI->getVolatileCst() 884 }; 885 // In case we fold the address space overloaded memcpy of A to B 886 // with memcpy of B to C, change the function to be a memcpy of A to C. 887 const Type *Tys[] = { Ops[0]->getType(), Ops[1]->getType(), 888 Ops[2]->getType() }; 889 Module *M = MI->getParent()->getParent()->getParent(); 890 TheFn = Intrinsic::getDeclaration(M, MI->getIntrinsicID(), Tys, 3); 891 CallInst::Create(TheFn, Ops, Ops + 5, "", MI); 892 } else { 893 assert(isa<MemSetInst>(MI)); 894 Value *Ops[] = { 895 EltPtr, MI->getOperand(2), // Dest, Value, 896 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 897 Zero, // Align 898 ConstantInt::get(Type::getInt1Ty(MI->getContext()), 0) // isVolatile 899 }; 900 const Type *Tys[] = { Ops[0]->getType(), Ops[2]->getType() }; 901 Module *M = MI->getParent()->getParent()->getParent(); 902 TheFn = Intrinsic::getDeclaration(M, Intrinsic::memset, Tys, 2); 903 CallInst::Create(TheFn, Ops, Ops + 5, "", MI); 904 } 905 } 906 DeadInsts.push_back(MI); 907} 908 909/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 910/// overwrites the entire allocation. Extract out the pieces of the stored 911/// integer and store them individually. 912void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 913 SmallVector<AllocaInst*, 32> &NewElts){ 914 // Extract each element out of the integer according to its structure offset 915 // and store the element value to the individual alloca. 916 Value *SrcVal = SI->getOperand(0); 917 const Type *AllocaEltTy = AI->getAllocatedType(); 918 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 919 920 // Handle tail padding by extending the operand 921 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 922 SrcVal = new ZExtInst(SrcVal, 923 IntegerType::get(SI->getContext(), AllocaSizeBits), 924 "", SI); 925 926 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 927 << '\n'); 928 929 // There are two forms here: AI could be an array or struct. Both cases 930 // have different ways to compute the element offset. 931 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 932 const StructLayout *Layout = TD->getStructLayout(EltSTy); 933 934 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 935 // Get the number of bits to shift SrcVal to get the value. 936 const Type *FieldTy = EltSTy->getElementType(i); 937 uint64_t Shift = Layout->getElementOffsetInBits(i); 938 939 if (TD->isBigEndian()) 940 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 941 942 Value *EltVal = SrcVal; 943 if (Shift) { 944 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 945 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 946 "sroa.store.elt", SI); 947 } 948 949 // Truncate down to an integer of the right size. 950 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 951 952 // Ignore zero sized fields like {}, they obviously contain no data. 953 if (FieldSizeBits == 0) continue; 954 955 if (FieldSizeBits != AllocaSizeBits) 956 EltVal = new TruncInst(EltVal, 957 IntegerType::get(SI->getContext(), FieldSizeBits), 958 "", SI); 959 Value *DestField = NewElts[i]; 960 if (EltVal->getType() == FieldTy) { 961 // Storing to an integer field of this size, just do it. 962 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) { 963 // Bitcast to the right element type (for fp/vector values). 964 EltVal = new BitCastInst(EltVal, FieldTy, "", SI); 965 } else { 966 // Otherwise, bitcast the dest pointer (for aggregates). 967 DestField = new BitCastInst(DestField, 968 PointerType::getUnqual(EltVal->getType()), 969 "", SI); 970 } 971 new StoreInst(EltVal, DestField, SI); 972 } 973 974 } else { 975 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 976 const Type *ArrayEltTy = ATy->getElementType(); 977 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 978 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 979 980 uint64_t Shift; 981 982 if (TD->isBigEndian()) 983 Shift = AllocaSizeBits-ElementOffset; 984 else 985 Shift = 0; 986 987 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 988 // Ignore zero sized fields like {}, they obviously contain no data. 989 if (ElementSizeBits == 0) continue; 990 991 Value *EltVal = SrcVal; 992 if (Shift) { 993 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 994 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 995 "sroa.store.elt", SI); 996 } 997 998 // Truncate down to an integer of the right size. 999 if (ElementSizeBits != AllocaSizeBits) 1000 EltVal = new TruncInst(EltVal, 1001 IntegerType::get(SI->getContext(), 1002 ElementSizeBits),"",SI); 1003 Value *DestField = NewElts[i]; 1004 if (EltVal->getType() == ArrayEltTy) { 1005 // Storing to an integer field of this size, just do it. 1006 } else if (ArrayEltTy->isFloatingPointTy() || 1007 ArrayEltTy->isVectorTy()) { 1008 // Bitcast to the right element type (for fp/vector values). 1009 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI); 1010 } else { 1011 // Otherwise, bitcast the dest pointer (for aggregates). 1012 DestField = new BitCastInst(DestField, 1013 PointerType::getUnqual(EltVal->getType()), 1014 "", SI); 1015 } 1016 new StoreInst(EltVal, DestField, SI); 1017 1018 if (TD->isBigEndian()) 1019 Shift -= ElementOffset; 1020 else 1021 Shift += ElementOffset; 1022 } 1023 } 1024 1025 DeadInsts.push_back(SI); 1026} 1027 1028/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 1029/// an integer. Load the individual pieces to form the aggregate value. 1030void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 1031 SmallVector<AllocaInst*, 32> &NewElts) { 1032 // Extract each element out of the NewElts according to its structure offset 1033 // and form the result value. 1034 const Type *AllocaEltTy = AI->getAllocatedType(); 1035 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1036 1037 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 1038 << '\n'); 1039 1040 // There are two forms here: AI could be an array or struct. Both cases 1041 // have different ways to compute the element offset. 1042 const StructLayout *Layout = 0; 1043 uint64_t ArrayEltBitOffset = 0; 1044 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1045 Layout = TD->getStructLayout(EltSTy); 1046 } else { 1047 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 1048 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1049 } 1050 1051 Value *ResultVal = 1052 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 1053 1054 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1055 // Load the value from the alloca. If the NewElt is an aggregate, cast 1056 // the pointer to an integer of the same size before doing the load. 1057 Value *SrcField = NewElts[i]; 1058 const Type *FieldTy = 1059 cast<PointerType>(SrcField->getType())->getElementType(); 1060 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1061 1062 // Ignore zero sized fields like {}, they obviously contain no data. 1063 if (FieldSizeBits == 0) continue; 1064 1065 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 1066 FieldSizeBits); 1067 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() && 1068 !FieldTy->isVectorTy()) 1069 SrcField = new BitCastInst(SrcField, 1070 PointerType::getUnqual(FieldIntTy), 1071 "", LI); 1072 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 1073 1074 // If SrcField is a fp or vector of the right size but that isn't an 1075 // integer type, bitcast to an integer so we can shift it. 1076 if (SrcField->getType() != FieldIntTy) 1077 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 1078 1079 // Zero extend the field to be the same size as the final alloca so that 1080 // we can shift and insert it. 1081 if (SrcField->getType() != ResultVal->getType()) 1082 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 1083 1084 // Determine the number of bits to shift SrcField. 1085 uint64_t Shift; 1086 if (Layout) // Struct case. 1087 Shift = Layout->getElementOffsetInBits(i); 1088 else // Array case. 1089 Shift = i*ArrayEltBitOffset; 1090 1091 if (TD->isBigEndian()) 1092 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 1093 1094 if (Shift) { 1095 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 1096 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 1097 } 1098 1099 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 1100 } 1101 1102 // Handle tail padding by truncating the result 1103 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 1104 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 1105 1106 LI->replaceAllUsesWith(ResultVal); 1107 DeadInsts.push_back(LI); 1108} 1109 1110/// HasPadding - Return true if the specified type has any structure or 1111/// alignment padding, false otherwise. 1112static bool HasPadding(const Type *Ty, const TargetData &TD) { 1113 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 1114 const StructLayout *SL = TD.getStructLayout(STy); 1115 unsigned PrevFieldBitOffset = 0; 1116 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1117 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 1118 1119 // Padding in sub-elements? 1120 if (HasPadding(STy->getElementType(i), TD)) 1121 return true; 1122 1123 // Check to see if there is any padding between this element and the 1124 // previous one. 1125 if (i) { 1126 unsigned PrevFieldEnd = 1127 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 1128 if (PrevFieldEnd < FieldBitOffset) 1129 return true; 1130 } 1131 1132 PrevFieldBitOffset = FieldBitOffset; 1133 } 1134 1135 // Check for tail padding. 1136 if (unsigned EltCount = STy->getNumElements()) { 1137 unsigned PrevFieldEnd = PrevFieldBitOffset + 1138 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 1139 if (PrevFieldEnd < SL->getSizeInBits()) 1140 return true; 1141 } 1142 1143 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1144 return HasPadding(ATy->getElementType(), TD); 1145 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) { 1146 return HasPadding(VTy->getElementType(), TD); 1147 } 1148 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 1149} 1150 1151/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 1152/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 1153/// or 1 if safe after canonicalization has been performed. 1154bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 1155 // Loop over the use list of the alloca. We can only transform it if all of 1156 // the users are safe to transform. 1157 AllocaInfo Info; 1158 1159 isSafeForScalarRepl(AI, AI, 0, Info); 1160 if (Info.isUnsafe) { 1161 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n'); 1162 return false; 1163 } 1164 1165 // Okay, we know all the users are promotable. If the aggregate is a memcpy 1166 // source and destination, we have to be careful. In particular, the memcpy 1167 // could be moving around elements that live in structure padding of the LLVM 1168 // types, but may actually be used. In these cases, we refuse to promote the 1169 // struct. 1170 if (Info.isMemCpySrc && Info.isMemCpyDst && 1171 HasPadding(AI->getAllocatedType(), *TD)) 1172 return false; 1173 1174 return true; 1175} 1176 1177/// MergeInType - Add the 'In' type to the accumulated type (Accum) so far at 1178/// the offset specified by Offset (which is specified in bytes). 1179/// 1180/// There are two cases we handle here: 1181/// 1) A union of vector types of the same size and potentially its elements. 1182/// Here we turn element accesses into insert/extract element operations. 1183/// This promotes a <4 x float> with a store of float to the third element 1184/// into a <4 x float> that uses insert element. 1185/// 2) A fully general blob of memory, which we turn into some (potentially 1186/// large) integer type with extract and insert operations where the loads 1187/// and stores would mutate the memory. 1188static void MergeInType(const Type *In, uint64_t Offset, const Type *&VecTy, 1189 unsigned AllocaSize, const TargetData &TD, 1190 LLVMContext &Context) { 1191 // If this could be contributing to a vector, analyze it. 1192 if (VecTy != Type::getVoidTy(Context)) { // either null or a vector type. 1193 1194 // If the In type is a vector that is the same size as the alloca, see if it 1195 // matches the existing VecTy. 1196 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 1197 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 1198 // If we're storing/loading a vector of the right size, allow it as a 1199 // vector. If this the first vector we see, remember the type so that 1200 // we know the element size. 1201 if (VecTy == 0) 1202 VecTy = VInTy; 1203 return; 1204 } 1205 } else if (In->isFloatTy() || In->isDoubleTy() || 1206 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 && 1207 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 1208 // If we're accessing something that could be an element of a vector, see 1209 // if the implied vector agrees with what we already have and if Offset is 1210 // compatible with it. 1211 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 1212 if (Offset % EltSize == 0 && 1213 AllocaSize % EltSize == 0 && 1214 (VecTy == 0 || 1215 cast<VectorType>(VecTy)->getElementType() 1216 ->getPrimitiveSizeInBits()/8 == EltSize)) { 1217 if (VecTy == 0) 1218 VecTy = VectorType::get(In, AllocaSize/EltSize); 1219 return; 1220 } 1221 } 1222 } 1223 1224 // Otherwise, we have a case that we can't handle with an optimized vector 1225 // form. We can still turn this into a large integer. 1226 VecTy = Type::getVoidTy(Context); 1227} 1228 1229/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 1230/// its accesses to a single vector type, return true and set VecTy to 1231/// the new type. If we could convert the alloca into a single promotable 1232/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 1233/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 1234/// is the current offset from the base of the alloca being analyzed. 1235/// 1236/// If we see at least one access to the value that is as a vector type, set the 1237/// SawVec flag. 1238bool SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy, 1239 bool &SawVec, uint64_t Offset, 1240 unsigned AllocaSize) { 1241 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1242 Instruction *User = cast<Instruction>(*UI); 1243 1244 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1245 // Don't break volatile loads. 1246 if (LI->isVolatile()) 1247 return false; 1248 MergeInType(LI->getType(), Offset, VecTy, 1249 AllocaSize, *TD, V->getContext()); 1250 SawVec |= LI->getType()->isVectorTy(); 1251 continue; 1252 } 1253 1254 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1255 // Storing the pointer, not into the value? 1256 if (SI->getOperand(0) == V || SI->isVolatile()) return 0; 1257 MergeInType(SI->getOperand(0)->getType(), Offset, 1258 VecTy, AllocaSize, *TD, V->getContext()); 1259 SawVec |= SI->getOperand(0)->getType()->isVectorTy(); 1260 continue; 1261 } 1262 1263 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 1264 if (!CanConvertToScalar(BCI, IsNotTrivial, VecTy, SawVec, Offset, 1265 AllocaSize)) 1266 return false; 1267 IsNotTrivial = true; 1268 continue; 1269 } 1270 1271 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1272 // If this is a GEP with a variable indices, we can't handle it. 1273 if (!GEP->hasAllConstantIndices()) 1274 return false; 1275 1276 // Compute the offset that this GEP adds to the pointer. 1277 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1278 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getPointerOperandType(), 1279 &Indices[0], Indices.size()); 1280 // See if all uses can be converted. 1281 if (!CanConvertToScalar(GEP, IsNotTrivial, VecTy, SawVec,Offset+GEPOffset, 1282 AllocaSize)) 1283 return false; 1284 IsNotTrivial = true; 1285 continue; 1286 } 1287 1288 // If this is a constant sized memset of a constant value (e.g. 0) we can 1289 // handle it. 1290 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 1291 // Store of constant value and constant size. 1292 if (isa<ConstantInt>(MSI->getValue()) && 1293 isa<ConstantInt>(MSI->getLength())) { 1294 IsNotTrivial = true; 1295 continue; 1296 } 1297 } 1298 1299 // If this is a memcpy or memmove into or out of the whole allocation, we 1300 // can handle it like a load or store of the scalar type. 1301 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 1302 if (ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength())) 1303 if (Len->getZExtValue() == AllocaSize && Offset == 0) { 1304 IsNotTrivial = true; 1305 continue; 1306 } 1307 } 1308 1309 // Otherwise, we cannot handle this! 1310 return false; 1311 } 1312 1313 return true; 1314} 1315 1316/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 1317/// directly. This happens when we are converting an "integer union" to a 1318/// single integer scalar, or when we are converting a "vector union" to a 1319/// vector with insert/extractelement instructions. 1320/// 1321/// Offset is an offset from the original alloca, in bits that need to be 1322/// shifted to the right. By the end of this, there should be no uses of Ptr. 1323void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset) { 1324 while (!Ptr->use_empty()) { 1325 Instruction *User = cast<Instruction>(Ptr->use_back()); 1326 1327 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 1328 ConvertUsesToScalar(CI, NewAI, Offset); 1329 CI->eraseFromParent(); 1330 continue; 1331 } 1332 1333 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1334 // Compute the offset that this GEP adds to the pointer. 1335 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1336 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getPointerOperandType(), 1337 &Indices[0], Indices.size()); 1338 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 1339 GEP->eraseFromParent(); 1340 continue; 1341 } 1342 1343 IRBuilder<> Builder(User->getParent(), User); 1344 1345 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1346 // The load is a bit extract from NewAI shifted right by Offset bits. 1347 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 1348 Value *NewLoadVal 1349 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 1350 LI->replaceAllUsesWith(NewLoadVal); 1351 LI->eraseFromParent(); 1352 continue; 1353 } 1354 1355 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1356 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 1357 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 1358 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 1359 Builder); 1360 Builder.CreateStore(New, NewAI); 1361 SI->eraseFromParent(); 1362 1363 // If the load we just inserted is now dead, then the inserted store 1364 // overwrote the entire thing. 1365 if (Old->use_empty()) 1366 Old->eraseFromParent(); 1367 continue; 1368 } 1369 1370 // If this is a constant sized memset of a constant value (e.g. 0) we can 1371 // transform it into a store of the expanded constant value. 1372 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 1373 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 1374 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 1375 if (NumBytes != 0) { 1376 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 1377 1378 // Compute the value replicated the right number of times. 1379 APInt APVal(NumBytes*8, Val); 1380 1381 // Splat the value if non-zero. 1382 if (Val) 1383 for (unsigned i = 1; i != NumBytes; ++i) 1384 APVal |= APVal << 8; 1385 1386 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in"); 1387 Value *New = ConvertScalar_InsertValue( 1388 ConstantInt::get(User->getContext(), APVal), 1389 Old, Offset, Builder); 1390 Builder.CreateStore(New, NewAI); 1391 1392 // If the load we just inserted is now dead, then the memset overwrote 1393 // the entire thing. 1394 if (Old->use_empty()) 1395 Old->eraseFromParent(); 1396 } 1397 MSI->eraseFromParent(); 1398 continue; 1399 } 1400 1401 // If this is a memcpy or memmove into or out of the whole allocation, we 1402 // can handle it like a load or store of the scalar type. 1403 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 1404 assert(Offset == 0 && "must be store to start of alloca"); 1405 1406 // If the source and destination are both to the same alloca, then this is 1407 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 1408 // as appropriate. 1409 AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject(0)); 1410 1411 if (MTI->getSource()->getUnderlyingObject(0) != OrigAI) { 1412 // Dest must be OrigAI, change this to be a load from the original 1413 // pointer (bitcasted), then a store to our new alloca. 1414 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 1415 Value *SrcPtr = MTI->getSource(); 1416 SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType()); 1417 1418 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 1419 SrcVal->setAlignment(MTI->getAlignment()); 1420 Builder.CreateStore(SrcVal, NewAI); 1421 } else if (MTI->getDest()->getUnderlyingObject(0) != OrigAI) { 1422 // Src must be OrigAI, change this to be a load from NewAI then a store 1423 // through the original dest pointer (bitcasted). 1424 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 1425 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 1426 1427 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType()); 1428 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 1429 NewStore->setAlignment(MTI->getAlignment()); 1430 } else { 1431 // Noop transfer. Src == Dst 1432 } 1433 1434 MTI->eraseFromParent(); 1435 continue; 1436 } 1437 1438 llvm_unreachable("Unsupported operation!"); 1439 } 1440} 1441 1442/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 1443/// or vector value FromVal, extracting the bits from the offset specified by 1444/// Offset. This returns the value, which is of type ToType. 1445/// 1446/// This happens when we are converting an "integer union" to a single 1447/// integer scalar, or when we are converting a "vector union" to a vector with 1448/// insert/extractelement instructions. 1449/// 1450/// Offset is an offset from the original alloca, in bits that need to be 1451/// shifted to the right. 1452Value *SROA::ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 1453 uint64_t Offset, IRBuilder<> &Builder) { 1454 // If the load is of the whole new alloca, no conversion is needed. 1455 if (FromVal->getType() == ToType && Offset == 0) 1456 return FromVal; 1457 1458 // If the result alloca is a vector type, this is either an element 1459 // access or a bitcast to another vector type of the same size. 1460 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) { 1461 if (ToType->isVectorTy()) 1462 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 1463 1464 // Otherwise it must be an element access. 1465 unsigned Elt = 0; 1466 if (Offset) { 1467 unsigned EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 1468 Elt = Offset/EltSize; 1469 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 1470 } 1471 // Return the element extracted out of it. 1472 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get( 1473 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp"); 1474 if (V->getType() != ToType) 1475 V = Builder.CreateBitCast(V, ToType, "tmp"); 1476 return V; 1477 } 1478 1479 // If ToType is a first class aggregate, extract out each of the pieces and 1480 // use insertvalue's to form the FCA. 1481 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 1482 const StructLayout &Layout = *TD->getStructLayout(ST); 1483 Value *Res = UndefValue::get(ST); 1484 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1485 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 1486 Offset+Layout.getElementOffsetInBits(i), 1487 Builder); 1488 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1489 } 1490 return Res; 1491 } 1492 1493 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 1494 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 1495 Value *Res = UndefValue::get(AT); 1496 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1497 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 1498 Offset+i*EltSize, Builder); 1499 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1500 } 1501 return Res; 1502 } 1503 1504 // Otherwise, this must be a union that was converted to an integer value. 1505 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 1506 1507 // If this is a big-endian system and the load is narrower than the 1508 // full alloca type, we need to do a shift to get the right bits. 1509 int ShAmt = 0; 1510 if (TD->isBigEndian()) { 1511 // On big-endian machines, the lowest bit is stored at the bit offset 1512 // from the pointer given by getTypeStoreSizeInBits. This matters for 1513 // integers with a bitwidth that is not a multiple of 8. 1514 ShAmt = TD->getTypeStoreSizeInBits(NTy) - 1515 TD->getTypeStoreSizeInBits(ToType) - Offset; 1516 } else { 1517 ShAmt = Offset; 1518 } 1519 1520 // Note: we support negative bitwidths (with shl) which are not defined. 1521 // We do this to support (f.e.) loads off the end of a structure where 1522 // only some bits are used. 1523 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 1524 FromVal = Builder.CreateLShr(FromVal, 1525 ConstantInt::get(FromVal->getType(), 1526 ShAmt), "tmp"); 1527 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 1528 FromVal = Builder.CreateShl(FromVal, 1529 ConstantInt::get(FromVal->getType(), 1530 -ShAmt), "tmp"); 1531 1532 // Finally, unconditionally truncate the integer to the right width. 1533 unsigned LIBitWidth = TD->getTypeSizeInBits(ToType); 1534 if (LIBitWidth < NTy->getBitWidth()) 1535 FromVal = 1536 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 1537 LIBitWidth), "tmp"); 1538 else if (LIBitWidth > NTy->getBitWidth()) 1539 FromVal = 1540 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 1541 LIBitWidth), "tmp"); 1542 1543 // If the result is an integer, this is a trunc or bitcast. 1544 if (ToType->isIntegerTy()) { 1545 // Should be done. 1546 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) { 1547 // Just do a bitcast, we know the sizes match up. 1548 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 1549 } else { 1550 // Otherwise must be a pointer. 1551 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 1552 } 1553 assert(FromVal->getType() == ToType && "Didn't convert right?"); 1554 return FromVal; 1555} 1556 1557/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 1558/// or vector value "Old" at the offset specified by Offset. 1559/// 1560/// This happens when we are converting an "integer union" to a 1561/// single integer scalar, or when we are converting a "vector union" to a 1562/// vector with insert/extractelement instructions. 1563/// 1564/// Offset is an offset from the original alloca, in bits that need to be 1565/// shifted to the right. 1566Value *SROA::ConvertScalar_InsertValue(Value *SV, Value *Old, 1567 uint64_t Offset, IRBuilder<> &Builder) { 1568 1569 // Convert the stored type to the actual type, shift it left to insert 1570 // then 'or' into place. 1571 const Type *AllocaType = Old->getType(); 1572 LLVMContext &Context = Old->getContext(); 1573 1574 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 1575 uint64_t VecSize = TD->getTypeAllocSizeInBits(VTy); 1576 uint64_t ValSize = TD->getTypeAllocSizeInBits(SV->getType()); 1577 1578 // Changing the whole vector with memset or with an access of a different 1579 // vector type? 1580 if (ValSize == VecSize) 1581 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 1582 1583 uint64_t EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 1584 1585 // Must be an element insertion. 1586 unsigned Elt = Offset/EltSize; 1587 1588 if (SV->getType() != VTy->getElementType()) 1589 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 1590 1591 SV = Builder.CreateInsertElement(Old, SV, 1592 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt), 1593 "tmp"); 1594 return SV; 1595 } 1596 1597 // If SV is a first-class aggregate value, insert each value recursively. 1598 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 1599 const StructLayout &Layout = *TD->getStructLayout(ST); 1600 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1601 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1602 Old = ConvertScalar_InsertValue(Elt, Old, 1603 Offset+Layout.getElementOffsetInBits(i), 1604 Builder); 1605 } 1606 return Old; 1607 } 1608 1609 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 1610 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 1611 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1612 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1613 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 1614 } 1615 return Old; 1616 } 1617 1618 // If SV is a float, convert it to the appropriate integer type. 1619 // If it is a pointer, do the same. 1620 unsigned SrcWidth = TD->getTypeSizeInBits(SV->getType()); 1621 unsigned DestWidth = TD->getTypeSizeInBits(AllocaType); 1622 unsigned SrcStoreWidth = TD->getTypeStoreSizeInBits(SV->getType()); 1623 unsigned DestStoreWidth = TD->getTypeStoreSizeInBits(AllocaType); 1624 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy()) 1625 SV = Builder.CreateBitCast(SV, 1626 IntegerType::get(SV->getContext(),SrcWidth), "tmp"); 1627 else if (SV->getType()->isPointerTy()) 1628 SV = Builder.CreatePtrToInt(SV, TD->getIntPtrType(SV->getContext()), "tmp"); 1629 1630 // Zero extend or truncate the value if needed. 1631 if (SV->getType() != AllocaType) { 1632 if (SV->getType()->getPrimitiveSizeInBits() < 1633 AllocaType->getPrimitiveSizeInBits()) 1634 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 1635 else { 1636 // Truncation may be needed if storing more than the alloca can hold 1637 // (undefined behavior). 1638 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 1639 SrcWidth = DestWidth; 1640 SrcStoreWidth = DestStoreWidth; 1641 } 1642 } 1643 1644 // If this is a big-endian system and the store is narrower than the 1645 // full alloca type, we need to do a shift to get the right bits. 1646 int ShAmt = 0; 1647 if (TD->isBigEndian()) { 1648 // On big-endian machines, the lowest bit is stored at the bit offset 1649 // from the pointer given by getTypeStoreSizeInBits. This matters for 1650 // integers with a bitwidth that is not a multiple of 8. 1651 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 1652 } else { 1653 ShAmt = Offset; 1654 } 1655 1656 // Note: we support negative bitwidths (with shr) which are not defined. 1657 // We do this to support (f.e.) stores off the end of a structure where 1658 // only some bits in the structure are set. 1659 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 1660 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 1661 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 1662 ShAmt), "tmp"); 1663 Mask <<= ShAmt; 1664 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1665 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 1666 -ShAmt), "tmp"); 1667 Mask = Mask.lshr(-ShAmt); 1668 } 1669 1670 // Mask out the bits we are about to insert from the old value, and or 1671 // in the new bits. 1672 if (SrcWidth != DestWidth) { 1673 assert(DestWidth > SrcWidth); 1674 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 1675 SV = Builder.CreateOr(Old, SV, "ins"); 1676 } 1677 return SV; 1678} 1679 1680 1681 1682/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1683/// some part of a constant global variable. This intentionally only accepts 1684/// constant expressions because we don't can't rewrite arbitrary instructions. 1685static bool PointsToConstantGlobal(Value *V) { 1686 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1687 return GV->isConstant(); 1688 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1689 if (CE->getOpcode() == Instruction::BitCast || 1690 CE->getOpcode() == Instruction::GetElementPtr) 1691 return PointsToConstantGlobal(CE->getOperand(0)); 1692 return false; 1693} 1694 1695/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1696/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1697/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1698/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1699/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1700/// the alloca, and if the source pointer is a pointer to a constant global, we 1701/// can optimize this. 1702static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy, 1703 bool isOffset) { 1704 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1705 User *U = cast<Instruction>(*UI); 1706 1707 if (LoadInst *LI = dyn_cast<LoadInst>(U)) 1708 // Ignore non-volatile loads, they are always ok. 1709 if (!LI->isVolatile()) 1710 continue; 1711 1712 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) { 1713 // If uses of the bitcast are ok, we are ok. 1714 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 1715 return false; 1716 continue; 1717 } 1718 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) { 1719 // If the GEP has all zero indices, it doesn't offset the pointer. If it 1720 // doesn't, it does. 1721 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 1722 isOffset || !GEP->hasAllZeroIndices())) 1723 return false; 1724 continue; 1725 } 1726 1727 // If this is isn't our memcpy/memmove, reject it as something we can't 1728 // handle. 1729 if (!isa<MemTransferInst>(U)) 1730 return false; 1731 1732 // If we already have seen a copy, reject the second one. 1733 if (TheCopy) return false; 1734 1735 // If the pointer has been offset from the start of the alloca, we can't 1736 // safely handle this. 1737 if (isOffset) return false; 1738 1739 // If the memintrinsic isn't using the alloca as the dest, reject it. 1740 if (UI.getOperandNo() != 1) return false; 1741 1742 MemIntrinsic *MI = cast<MemIntrinsic>(U); 1743 1744 // If the source of the memcpy/move is not a constant global, reject it. 1745 if (!PointsToConstantGlobal(MI->getOperand(2))) 1746 return false; 1747 1748 // Otherwise, the transform is safe. Remember the copy instruction. 1749 TheCopy = MI; 1750 } 1751 return true; 1752} 1753 1754/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 1755/// modified by a copy from a constant global. If we can prove this, we can 1756/// replace any uses of the alloca with uses of the global directly. 1757Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) { 1758 Instruction *TheCopy = 0; 1759 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 1760 return TheCopy; 1761 return 0; 1762} 1763