ScalarReplAggregates.cpp revision 3c3af5d15569d708d2dc8f13351bc77056b4d70d
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 /// needsCleanup - This is set to true if there is some use of the alloca 89 /// that requires cleanup. 90 bool needsCleanup : 1; 91 92 /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 93 bool isMemCpySrc : 1; 94 95 /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 96 bool isMemCpyDst : 1; 97 98 AllocaInfo() 99 : isUnsafe(false), needsCleanup(false), 100 isMemCpySrc(false), isMemCpyDst(false) {} 101 }; 102 103 unsigned SRThreshold; 104 105 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; } 106 107 int isSafeAllocaToScalarRepl(AllocaInst *AI); 108 109 void isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 110 AllocaInfo &Info); 111 void isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t &Offset, 112 AllocaInfo &Info); 113 void isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize, 114 const Type *MemOpType, bool isStore, AllocaInfo &Info); 115 bool TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size); 116 uint64_t FindElementAndOffset(const Type *&T, uint64_t &Offset, 117 const Type *&IdxTy); 118 119 void DoScalarReplacement(AllocaInst *AI, 120 std::vector<AllocaInst*> &WorkList); 121 void DeleteDeadInstructions(); 122 void CleanupAllocaUsers(Value *V); 123 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocaInst *Base); 124 125 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 126 SmallVector<AllocaInst*, 32> &NewElts); 127 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 128 SmallVector<AllocaInst*, 32> &NewElts); 129 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 130 SmallVector<AllocaInst*, 32> &NewElts); 131 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 132 AllocaInst *AI, 133 SmallVector<AllocaInst*, 32> &NewElts); 134 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 135 SmallVector<AllocaInst*, 32> &NewElts); 136 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 137 SmallVector<AllocaInst*, 32> &NewElts); 138 139 bool CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy, 140 bool &SawVec, uint64_t Offset, unsigned AllocaSize); 141 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); 142 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType, 143 uint64_t Offset, IRBuilder<> &Builder); 144 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 145 uint64_t Offset, IRBuilder<> &Builder); 146 static Instruction *isOnlyCopiedFromConstantGlobal(AllocaInst *AI); 147 }; 148} 149 150char SROA::ID = 0; 151static RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates"); 152 153// Public interface to the ScalarReplAggregates pass 154FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) { 155 return new SROA(Threshold); 156} 157 158 159bool SROA::runOnFunction(Function &F) { 160 TD = getAnalysisIfAvailable<TargetData>(); 161 162 bool Changed = performPromotion(F); 163 164 // FIXME: ScalarRepl currently depends on TargetData more than it 165 // theoretically needs to. It should be refactored in order to support 166 // target-independent IR. Until this is done, just skip the actual 167 // scalar-replacement portion of this pass. 168 if (!TD) return Changed; 169 170 while (1) { 171 bool LocalChange = performScalarRepl(F); 172 if (!LocalChange) break; // No need to repromote if no scalarrepl 173 Changed = true; 174 LocalChange = performPromotion(F); 175 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 176 } 177 178 return Changed; 179} 180 181 182bool SROA::performPromotion(Function &F) { 183 std::vector<AllocaInst*> Allocas; 184 DominatorTree &DT = getAnalysis<DominatorTree>(); 185 DominanceFrontier &DF = getAnalysis<DominanceFrontier>(); 186 187 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 188 189 bool Changed = false; 190 191 while (1) { 192 Allocas.clear(); 193 194 // Find allocas that are safe to promote, by looking at all instructions in 195 // the entry node 196 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 197 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 198 if (isAllocaPromotable(AI)) 199 Allocas.push_back(AI); 200 201 if (Allocas.empty()) break; 202 203 PromoteMemToReg(Allocas, DT, DF); 204 NumPromoted += Allocas.size(); 205 Changed = true; 206 } 207 208 return Changed; 209} 210 211/// getNumSAElements - Return the number of elements in the specific struct or 212/// array. 213static uint64_t getNumSAElements(const Type *T) { 214 if (const StructType *ST = dyn_cast<StructType>(T)) 215 return ST->getNumElements(); 216 return cast<ArrayType>(T)->getNumElements(); 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(errs() << "Found alloca equal to global: " << *AI << '\n'); 256 DEBUG(errs() << " 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 // Do not promote any struct whose size is too big. 276 if (AllocaSize > SRThreshold) continue; 277 278 if ((isa<StructType>(AI->getAllocatedType()) || 279 isa<ArrayType>(AI->getAllocatedType())) && 280 // Do not promote any struct into more than "32" separate vars. 281 getNumSAElements(AI->getAllocatedType()) <= SRThreshold/4) { 282 // Check that all of the users of the allocation are capable of being 283 // transformed. 284 switch (isSafeAllocaToScalarRepl(AI)) { 285 default: llvm_unreachable("Unexpected value!"); 286 case 0: // Not safe to scalar replace. 287 break; 288 case 1: // Safe, but requires cleanup/canonicalizations first 289 CleanupAllocaUsers(AI); 290 // FALL THROUGH. 291 case 3: // Safe to scalar replace. 292 DoScalarReplacement(AI, WorkList); 293 Changed = true; 294 continue; 295 } 296 } 297 298 // If we can turn this aggregate value (potentially with casts) into a 299 // simple scalar value that can be mem2reg'd into a register value. 300 // IsNotTrivial tracks whether this is something that mem2reg could have 301 // promoted itself. If so, we don't want to transform it needlessly. Note 302 // that we can't just check based on the type: the alloca may be of an i32 303 // but that has pointer arithmetic to set byte 3 of it or something. 304 bool IsNotTrivial = false; 305 const Type *VectorTy = 0; 306 bool HadAVector = false; 307 if (CanConvertToScalar(AI, IsNotTrivial, VectorTy, HadAVector, 308 0, unsigned(AllocaSize)) && IsNotTrivial) { 309 AllocaInst *NewAI; 310 // If we were able to find a vector type that can handle this with 311 // insert/extract elements, and if there was at least one use that had 312 // a vector type, promote this to a vector. We don't want to promote 313 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 314 // we just get a lot of insert/extracts. If at least one vector is 315 // involved, then we probably really do have a union of vector/array. 316 if (VectorTy && isa<VectorType>(VectorTy) && HadAVector) { 317 DEBUG(errs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = " 318 << *VectorTy << '\n'); 319 320 // Create and insert the vector alloca. 321 NewAI = new AllocaInst(VectorTy, 0, "", AI->getParent()->begin()); 322 ConvertUsesToScalar(AI, NewAI, 0); 323 } else { 324 DEBUG(errs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"); 325 326 // Create and insert the integer alloca. 327 const Type *NewTy = IntegerType::get(AI->getContext(), AllocaSize*8); 328 NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 329 ConvertUsesToScalar(AI, NewAI, 0); 330 } 331 NewAI->takeName(AI); 332 AI->eraseFromParent(); 333 ++NumConverted; 334 Changed = true; 335 continue; 336 } 337 338 // Otherwise, couldn't process this alloca. 339 } 340 341 return Changed; 342} 343 344/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 345/// predicate, do SROA now. 346void SROA::DoScalarReplacement(AllocaInst *AI, 347 std::vector<AllocaInst*> &WorkList) { 348 DEBUG(errs() << "Found inst to SROA: " << *AI << '\n'); 349 SmallVector<AllocaInst*, 32> ElementAllocas; 350 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 351 ElementAllocas.reserve(ST->getNumContainedTypes()); 352 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 353 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 354 AI->getAlignment(), 355 AI->getName() + "." + Twine(i), AI); 356 ElementAllocas.push_back(NA); 357 WorkList.push_back(NA); // Add to worklist for recursive processing 358 } 359 } else { 360 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 361 ElementAllocas.reserve(AT->getNumElements()); 362 const Type *ElTy = AT->getElementType(); 363 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 364 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 365 AI->getName() + "." + Twine(i), AI); 366 ElementAllocas.push_back(NA); 367 WorkList.push_back(NA); // Add to worklist for recursive processing 368 } 369 } 370 371 // Now that we have created the new alloca instructions, rewrite all the 372 // uses of the old alloca. 373 RewriteForScalarRepl(AI, AI, 0, ElementAllocas); 374 375 // Now erase any instructions that were made dead while rewriting the alloca. 376 DeleteDeadInstructions(); 377 AI->eraseFromParent(); 378 379 NumReplaced++; 380} 381 382/// DeleteDeadInstructions - Erase instructions on the DeadInstrs list, 383/// recursively including all their operands that become trivially dead. 384void SROA::DeleteDeadInstructions() { 385 while (!DeadInsts.empty()) { 386 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val()); 387 388 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 389 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 390 // Zero out the operand and see if it becomes trivially dead. 391 // (But, don't add allocas to the dead instruction list -- they are 392 // already on the worklist and will be deleted separately.) 393 *OI = 0; 394 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U)) 395 DeadInsts.push_back(U); 396 } 397 398 I->eraseFromParent(); 399 } 400} 401 402/// isSafeForScalarRepl - Check if instruction I is a safe use with regard to 403/// performing scalar replacement of alloca AI. The results are flagged in 404/// the Info parameter. Offset indicates the position within AI that is 405/// referenced by this instruction. 406void SROA::isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 407 AllocaInfo &Info) { 408 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 409 Instruction *User = cast<Instruction>(*UI); 410 411 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 412 isSafeForScalarRepl(BC, AI, Offset, Info); 413 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 414 uint64_t GEPOffset = Offset; 415 isSafeGEP(GEPI, AI, GEPOffset, Info); 416 if (!Info.isUnsafe) 417 isSafeForScalarRepl(GEPI, AI, GEPOffset, Info); 418 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) { 419 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 420 if (Length) 421 isSafeMemAccess(AI, Offset, Length->getZExtValue(), 0, 422 UI.getOperandNo() == 1, Info); 423 else 424 MarkUnsafe(Info); 425 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 426 if (!LI->isVolatile()) { 427 const Type *LIType = LI->getType(); 428 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(LIType), 429 LIType, false, Info); 430 } else 431 MarkUnsafe(Info); 432 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 433 // Store is ok if storing INTO the pointer, not storing the pointer 434 if (!SI->isVolatile() && SI->getOperand(0) != I) { 435 const Type *SIType = SI->getOperand(0)->getType(); 436 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(SIType), 437 SIType, true, Info); 438 } else 439 MarkUnsafe(Info); 440 } else if (isa<DbgInfoIntrinsic>(UI)) { 441 // If one user is DbgInfoIntrinsic then check if all users are 442 // DbgInfoIntrinsics. 443 if (OnlyUsedByDbgInfoIntrinsics(I)) { 444 Info.needsCleanup = true; 445 return; 446 } 447 MarkUnsafe(Info); 448 } else { 449 DEBUG(errs() << " Transformation preventing inst: " << *User << '\n'); 450 MarkUnsafe(Info); 451 } 452 if (Info.isUnsafe) return; 453 } 454} 455 456/// isSafeGEP - Check if a GEP instruction can be handled for scalar 457/// replacement. It is safe when all the indices are constant, in-bounds 458/// references, and when the resulting offset corresponds to an element within 459/// the alloca type. The results are flagged in the Info parameter. Upon 460/// return, Offset is adjusted as specified by the GEP indices. 461void SROA::isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, 462 uint64_t &Offset, AllocaInfo &Info) { 463 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI); 464 if (GEPIt == E) 465 return; 466 467 // The first GEP index must be zero. 468 if (!isa<ConstantInt>(GEPIt.getOperand()) || 469 !cast<ConstantInt>(GEPIt.getOperand())->isZero()) 470 return MarkUnsafe(Info); 471 if (++GEPIt == E) 472 return; 473 474 // Walk through the GEP type indices, checking the types that this indexes 475 // into. 476 for (; GEPIt != E; ++GEPIt) { 477 // Ignore struct elements, no extra checking needed for these. 478 if (isa<StructType>(*GEPIt)) 479 continue; 480 481 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 482 if (!IdxVal) 483 return MarkUnsafe(Info); 484 485 if (const ArrayType *AT = dyn_cast<ArrayType>(*GEPIt)) { 486 // This GEP indexes an array. Verify that this is an in-range constant 487 // integer. Specifically, consider A[0][i]. We cannot know that the user 488 // isn't doing invalid things like allowing i to index an out-of-range 489 // subscript that accesses A[1]. Because of this, we have to reject SROA 490 // of any accesses into structs where any of the components are variables. 491 if (IdxVal->getZExtValue() >= AT->getNumElements()) 492 return MarkUnsafe(Info); 493 } else { 494 const VectorType *VT = cast<VectorType>(*GEPIt); 495 if (IdxVal->getZExtValue() >= VT->getNumElements()) 496 return MarkUnsafe(Info); 497 } 498 } 499 500 // All the indices are safe. Now compute the offset due to this GEP and 501 // check if the alloca has a component element at that offset. 502 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 503 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 504 &Indices[0], Indices.size()); 505 if (!TypeHasComponent(AI->getAllocatedType(), Offset, 0)) 506 MarkUnsafe(Info); 507} 508 509/// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI 510/// alloca or has an offset and size that corresponds to a component element 511/// within it. The offset checked here may have been formed from a GEP with a 512/// pointer bitcasted to a different type. 513void SROA::isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize, 514 const Type *MemOpType, bool isStore, 515 AllocaInfo &Info) { 516 // Check if this is a load/store of the entire alloca. 517 if (Offset == 0 && MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) { 518 bool UsesAggregateType = (MemOpType == AI->getAllocatedType()); 519 // This is safe for MemIntrinsics (where MemOpType is 0), integer types 520 // (which are essentially the same as the MemIntrinsics, especially with 521 // regard to copying padding between elements), or references using the 522 // aggregate type of the alloca. 523 if (!MemOpType || isa<IntegerType>(MemOpType) || UsesAggregateType) { 524 if (!UsesAggregateType) { 525 if (isStore) 526 Info.isMemCpyDst = true; 527 else 528 Info.isMemCpySrc = true; 529 } 530 return; 531 } 532 } 533 // Check if the offset/size correspond to a component within the alloca type. 534 const Type *T = AI->getAllocatedType(); 535 if (TypeHasComponent(T, Offset, MemSize)) 536 return; 537 538 return MarkUnsafe(Info); 539} 540 541/// TypeHasComponent - Return true if T has a component type with the 542/// specified offset and size. If Size is zero, do not check the size. 543bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) { 544 const Type *EltTy; 545 uint64_t EltSize; 546 if (const StructType *ST = dyn_cast<StructType>(T)) { 547 const StructLayout *Layout = TD->getStructLayout(ST); 548 unsigned EltIdx = Layout->getElementContainingOffset(Offset); 549 EltTy = ST->getContainedType(EltIdx); 550 EltSize = TD->getTypeAllocSize(EltTy); 551 Offset -= Layout->getElementOffset(EltIdx); 552 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) { 553 EltTy = AT->getElementType(); 554 EltSize = TD->getTypeAllocSize(EltTy); 555 Offset %= EltSize; 556 } else { 557 return false; 558 } 559 if (Offset == 0 && (Size == 0 || EltSize == Size)) 560 return true; 561 // Check if the component spans multiple elements. 562 if (Offset + Size > EltSize) 563 return false; 564 return TypeHasComponent(EltTy, Offset, Size); 565} 566 567/// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite 568/// the instruction I, which references it, to use the separate elements. 569/// Offset indicates the position within AI that is referenced by this 570/// instruction. 571void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset, 572 SmallVector<AllocaInst*, 32> &NewElts) { 573 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) { 574 Instruction *User = cast<Instruction>(*UI); 575 576 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) { 577 RewriteBitCast(BC, AI, Offset, NewElts); 578 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) { 579 RewriteGEP(GEPI, AI, Offset, NewElts); 580 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 581 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 582 uint64_t MemSize = Length->getZExtValue(); 583 if (Offset == 0 && 584 MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) 585 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts); 586 // Otherwise the intrinsic can only touch a single element and the 587 // address operand will be updated, so nothing else needs to be done. 588 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 589 const Type *LIType = LI->getType(); 590 if (LIType == AI->getAllocatedType()) { 591 // Replace: 592 // %res = load { i32, i32 }* %alloc 593 // with: 594 // %load.0 = load i32* %alloc.0 595 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 596 // %load.1 = load i32* %alloc.1 597 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 598 // (Also works for arrays instead of structs) 599 Value *Insert = UndefValue::get(LIType); 600 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 601 Value *Load = new LoadInst(NewElts[i], "load", LI); 602 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 603 } 604 LI->replaceAllUsesWith(Insert); 605 DeadInsts.push_back(LI); 606 } else if (isa<IntegerType>(LIType) && 607 TD->getTypeAllocSize(LIType) == 608 TD->getTypeAllocSize(AI->getAllocatedType())) { 609 // If this is a load of the entire alloca to an integer, rewrite it. 610 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 611 } 612 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 613 Value *Val = SI->getOperand(0); 614 const Type *SIType = Val->getType(); 615 if (SIType == AI->getAllocatedType()) { 616 // Replace: 617 // store { i32, i32 } %val, { i32, i32 }* %alloc 618 // with: 619 // %val.0 = extractvalue { i32, i32 } %val, 0 620 // store i32 %val.0, i32* %alloc.0 621 // %val.1 = extractvalue { i32, i32 } %val, 1 622 // store i32 %val.1, i32* %alloc.1 623 // (Also works for arrays instead of structs) 624 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 625 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 626 new StoreInst(Extract, NewElts[i], SI); 627 } 628 DeadInsts.push_back(SI); 629 } else if (isa<IntegerType>(SIType) && 630 TD->getTypeAllocSize(SIType) == 631 TD->getTypeAllocSize(AI->getAllocatedType())) { 632 // If this is a store of the entire alloca from an integer, rewrite it. 633 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 634 } 635 } 636 } 637} 638 639/// RewriteBitCast - Update a bitcast reference to the alloca being replaced 640/// and recursively continue updating all of its uses. 641void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset, 642 SmallVector<AllocaInst*, 32> &NewElts) { 643 RewriteForScalarRepl(BC, AI, Offset, NewElts); 644 if (BC->getOperand(0) != AI) 645 return; 646 647 // The bitcast references the original alloca. Replace its uses with 648 // references to the first new element alloca. 649 Instruction *Val = NewElts[0]; 650 if (Val->getType() != BC->getDestTy()) { 651 Val = new BitCastInst(Val, BC->getDestTy(), "", BC); 652 Val->takeName(BC); 653 } 654 BC->replaceAllUsesWith(Val); 655 DeadInsts.push_back(BC); 656} 657 658/// FindElementAndOffset - Return the index of the element containing Offset 659/// within the specified type, which must be either a struct or an array. 660/// Sets T to the type of the element and Offset to the offset within that 661/// element. IdxTy is set to the type of the index result to be used in a 662/// GEP instruction. 663uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset, 664 const Type *&IdxTy) { 665 uint64_t Idx = 0; 666 if (const StructType *ST = dyn_cast<StructType>(T)) { 667 const StructLayout *Layout = TD->getStructLayout(ST); 668 Idx = Layout->getElementContainingOffset(Offset); 669 T = ST->getContainedType(Idx); 670 Offset -= Layout->getElementOffset(Idx); 671 IdxTy = Type::getInt32Ty(T->getContext()); 672 return Idx; 673 } 674 const ArrayType *AT = cast<ArrayType>(T); 675 T = AT->getElementType(); 676 uint64_t EltSize = TD->getTypeAllocSize(T); 677 Idx = Offset / EltSize; 678 Offset -= Idx * EltSize; 679 IdxTy = Type::getInt64Ty(T->getContext()); 680 return Idx; 681} 682 683/// RewriteGEP - Check if this GEP instruction moves the pointer across 684/// elements of the alloca that are being split apart, and if so, rewrite 685/// the GEP to be relative to the new element. 686void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset, 687 SmallVector<AllocaInst*, 32> &NewElts) { 688 uint64_t OldOffset = Offset; 689 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end()); 690 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), 691 &Indices[0], Indices.size()); 692 693 RewriteForScalarRepl(GEPI, AI, Offset, NewElts); 694 695 const Type *T = AI->getAllocatedType(); 696 const Type *IdxTy; 697 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy); 698 if (GEPI->getOperand(0) == AI) 699 OldIdx = ~0ULL; // Force the GEP to be rewritten. 700 701 T = AI->getAllocatedType(); 702 uint64_t EltOffset = Offset; 703 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy); 704 705 // If this GEP does not move the pointer across elements of the alloca 706 // being split, then it does not needs to be rewritten. 707 if (Idx == OldIdx) 708 return; 709 710 const Type *i32Ty = Type::getInt32Ty(AI->getContext()); 711 SmallVector<Value*, 8> NewArgs; 712 NewArgs.push_back(Constant::getNullValue(i32Ty)); 713 while (EltOffset != 0) { 714 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy); 715 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx)); 716 } 717 Instruction *Val = NewElts[Idx]; 718 if (NewArgs.size() > 1) { 719 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(), 720 NewArgs.end(), "", GEPI); 721 Val->takeName(GEPI); 722 } 723 if (Val->getType() != GEPI->getType()) 724 Val = new BitCastInst(Val, GEPI->getType(), Val->getNameStr(), GEPI); 725 GEPI->replaceAllUsesWith(Val); 726 DeadInsts.push_back(GEPI); 727} 728 729/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 730/// Rewrite it to copy or set the elements of the scalarized memory. 731void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst, 732 AllocaInst *AI, 733 SmallVector<AllocaInst*, 32> &NewElts) { 734 // If this is a memcpy/memmove, construct the other pointer as the 735 // appropriate type. The "Other" pointer is the pointer that goes to memory 736 // that doesn't have anything to do with the alloca that we are promoting. For 737 // memset, this Value* stays null. 738 Value *OtherPtr = 0; 739 LLVMContext &Context = MI->getContext(); 740 unsigned MemAlignment = MI->getAlignment(); 741 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 742 if (Inst == MTI->getRawDest()) 743 OtherPtr = MTI->getRawSource(); 744 else { 745 assert(Inst == MTI->getRawSource()); 746 OtherPtr = MTI->getRawDest(); 747 } 748 } 749 750 // If there is an other pointer, we want to convert it to the same pointer 751 // type as AI has, so we can GEP through it safely. 752 if (OtherPtr) { 753 754 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an 755 // optimization, but it's also required to detect the corner case where 756 // both pointer operands are referencing the same memory, and where 757 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This 758 // function is only called for mem intrinsics that access the whole 759 // aggregate, so non-zero GEPs are not an issue here.) 760 while (1) { 761 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr)) { 762 OtherPtr = BC->getOperand(0); 763 continue; 764 } 765 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr)) { 766 // All zero GEPs are effectively bitcasts. 767 if (GEP->hasAllZeroIndices()) { 768 OtherPtr = GEP->getOperand(0); 769 continue; 770 } 771 } 772 break; 773 } 774 // If OtherPtr has already been rewritten, this intrinsic will be dead. 775 if (OtherPtr == NewElts[0]) 776 return; 777 778 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr)) 779 if (BCE->getOpcode() == Instruction::BitCast) 780 OtherPtr = BCE->getOperand(0); 781 782 // If the pointer is not the right type, insert a bitcast to the right 783 // type. 784 if (OtherPtr->getType() != AI->getType()) 785 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(), 786 MI); 787 } 788 789 // Process each element of the aggregate. 790 Value *TheFn = MI->getOperand(0); 791 const Type *BytePtrTy = MI->getRawDest()->getType(); 792 bool SROADest = MI->getRawDest() == Inst; 793 794 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 795 796 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 797 // If this is a memcpy/memmove, emit a GEP of the other element address. 798 Value *OtherElt = 0; 799 unsigned OtherEltAlign = MemAlignment; 800 801 if (OtherPtr == AI) { 802 OtherElt = NewElts[i]; 803 OtherEltAlign = 0; 804 } else if (OtherPtr) { 805 Value *Idx[2] = { Zero, 806 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 807 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2, 808 OtherPtr->getNameStr()+"."+Twine(i), 809 MI); 810 uint64_t EltOffset; 811 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 812 if (const StructType *ST = 813 dyn_cast<StructType>(OtherPtrTy->getElementType())) { 814 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 815 } else { 816 const Type *EltTy = 817 cast<SequentialType>(OtherPtr->getType())->getElementType(); 818 EltOffset = TD->getTypeAllocSize(EltTy)*i; 819 } 820 821 // The alignment of the other pointer is the guaranteed alignment of the 822 // element, which is affected by both the known alignment of the whole 823 // mem intrinsic and the alignment of the element. If the alignment of 824 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 825 // known alignment is just 4 bytes. 826 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 827 } 828 829 Value *EltPtr = NewElts[i]; 830 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 831 832 // If we got down to a scalar, insert a load or store as appropriate. 833 if (EltTy->isSingleValueType()) { 834 if (isa<MemTransferInst>(MI)) { 835 if (SROADest) { 836 // From Other to Alloca. 837 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 838 new StoreInst(Elt, EltPtr, MI); 839 } else { 840 // From Alloca to Other. 841 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 842 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 843 } 844 continue; 845 } 846 assert(isa<MemSetInst>(MI)); 847 848 // If the stored element is zero (common case), just store a null 849 // constant. 850 Constant *StoreVal; 851 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) { 852 if (CI->isZero()) { 853 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 854 } else { 855 // If EltTy is a vector type, get the element type. 856 const Type *ValTy = EltTy->getScalarType(); 857 858 // Construct an integer with the right value. 859 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 860 APInt OneVal(EltSize, CI->getZExtValue()); 861 APInt TotalVal(OneVal); 862 // Set each byte. 863 for (unsigned i = 0; 8*i < EltSize; ++i) { 864 TotalVal = TotalVal.shl(8); 865 TotalVal |= OneVal; 866 } 867 868 // Convert the integer value to the appropriate type. 869 StoreVal = ConstantInt::get(Context, TotalVal); 870 if (isa<PointerType>(ValTy)) 871 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 872 else if (ValTy->isFloatingPoint()) 873 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 874 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 875 876 // If the requested value was a vector constant, create it. 877 if (EltTy != ValTy) { 878 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 879 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 880 StoreVal = ConstantVector::get(&Elts[0], NumElts); 881 } 882 } 883 new StoreInst(StoreVal, EltPtr, MI); 884 continue; 885 } 886 // Otherwise, if we're storing a byte variable, use a memset call for 887 // this element. 888 } 889 890 // Cast the element pointer to BytePtrTy. 891 if (EltPtr->getType() != BytePtrTy) 892 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI); 893 894 // Cast the other pointer (if we have one) to BytePtrTy. 895 if (OtherElt && OtherElt->getType() != BytePtrTy) 896 OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(), 897 MI); 898 899 unsigned EltSize = TD->getTypeAllocSize(EltTy); 900 901 // Finally, insert the meminst for this element. 902 if (isa<MemTransferInst>(MI)) { 903 Value *Ops[] = { 904 SROADest ? EltPtr : OtherElt, // Dest ptr 905 SROADest ? OtherElt : EltPtr, // Src ptr 906 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 907 // Align 908 ConstantInt::get(Type::getInt32Ty(MI->getContext()), OtherEltAlign) 909 }; 910 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 911 } else { 912 assert(isa<MemSetInst>(MI)); 913 Value *Ops[] = { 914 EltPtr, MI->getOperand(2), // Dest, Value, 915 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 916 Zero // Align 917 }; 918 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 919 } 920 } 921 DeadInsts.push_back(MI); 922} 923 924/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 925/// overwrites the entire allocation. Extract out the pieces of the stored 926/// integer and store them individually. 927void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 928 SmallVector<AllocaInst*, 32> &NewElts){ 929 // Extract each element out of the integer according to its structure offset 930 // and store the element value to the individual alloca. 931 Value *SrcVal = SI->getOperand(0); 932 const Type *AllocaEltTy = AI->getAllocatedType(); 933 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 934 935 // Handle tail padding by extending the operand 936 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 937 SrcVal = new ZExtInst(SrcVal, 938 IntegerType::get(SI->getContext(), AllocaSizeBits), 939 "", SI); 940 941 DEBUG(errs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 942 << '\n'); 943 944 // There are two forms here: AI could be an array or struct. Both cases 945 // have different ways to compute the element offset. 946 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 947 const StructLayout *Layout = TD->getStructLayout(EltSTy); 948 949 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 950 // Get the number of bits to shift SrcVal to get the value. 951 const Type *FieldTy = EltSTy->getElementType(i); 952 uint64_t Shift = Layout->getElementOffsetInBits(i); 953 954 if (TD->isBigEndian()) 955 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 956 957 Value *EltVal = SrcVal; 958 if (Shift) { 959 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 960 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 961 "sroa.store.elt", SI); 962 } 963 964 // Truncate down to an integer of the right size. 965 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 966 967 // Ignore zero sized fields like {}, they obviously contain no data. 968 if (FieldSizeBits == 0) continue; 969 970 if (FieldSizeBits != AllocaSizeBits) 971 EltVal = new TruncInst(EltVal, 972 IntegerType::get(SI->getContext(), FieldSizeBits), 973 "", SI); 974 Value *DestField = NewElts[i]; 975 if (EltVal->getType() == FieldTy) { 976 // Storing to an integer field of this size, just do it. 977 } else if (FieldTy->isFloatingPoint() || isa<VectorType>(FieldTy)) { 978 // Bitcast to the right element type (for fp/vector values). 979 EltVal = new BitCastInst(EltVal, FieldTy, "", SI); 980 } else { 981 // Otherwise, bitcast the dest pointer (for aggregates). 982 DestField = new BitCastInst(DestField, 983 PointerType::getUnqual(EltVal->getType()), 984 "", SI); 985 } 986 new StoreInst(EltVal, DestField, SI); 987 } 988 989 } else { 990 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 991 const Type *ArrayEltTy = ATy->getElementType(); 992 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 993 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 994 995 uint64_t Shift; 996 997 if (TD->isBigEndian()) 998 Shift = AllocaSizeBits-ElementOffset; 999 else 1000 Shift = 0; 1001 1002 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1003 // Ignore zero sized fields like {}, they obviously contain no data. 1004 if (ElementSizeBits == 0) continue; 1005 1006 Value *EltVal = SrcVal; 1007 if (Shift) { 1008 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 1009 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 1010 "sroa.store.elt", SI); 1011 } 1012 1013 // Truncate down to an integer of the right size. 1014 if (ElementSizeBits != AllocaSizeBits) 1015 EltVal = new TruncInst(EltVal, 1016 IntegerType::get(SI->getContext(), 1017 ElementSizeBits),"",SI); 1018 Value *DestField = NewElts[i]; 1019 if (EltVal->getType() == ArrayEltTy) { 1020 // Storing to an integer field of this size, just do it. 1021 } else if (ArrayEltTy->isFloatingPoint() || isa<VectorType>(ArrayEltTy)) { 1022 // Bitcast to the right element type (for fp/vector values). 1023 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI); 1024 } else { 1025 // Otherwise, bitcast the dest pointer (for aggregates). 1026 DestField = new BitCastInst(DestField, 1027 PointerType::getUnqual(EltVal->getType()), 1028 "", SI); 1029 } 1030 new StoreInst(EltVal, DestField, SI); 1031 1032 if (TD->isBigEndian()) 1033 Shift -= ElementOffset; 1034 else 1035 Shift += ElementOffset; 1036 } 1037 } 1038 1039 DeadInsts.push_back(SI); 1040} 1041 1042/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 1043/// an integer. Load the individual pieces to form the aggregate value. 1044void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 1045 SmallVector<AllocaInst*, 32> &NewElts) { 1046 // Extract each element out of the NewElts according to its structure offset 1047 // and form the result value. 1048 const Type *AllocaEltTy = AI->getAllocatedType(); 1049 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1050 1051 DEBUG(errs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 1052 << '\n'); 1053 1054 // There are two forms here: AI could be an array or struct. Both cases 1055 // have different ways to compute the element offset. 1056 const StructLayout *Layout = 0; 1057 uint64_t ArrayEltBitOffset = 0; 1058 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1059 Layout = TD->getStructLayout(EltSTy); 1060 } else { 1061 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 1062 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1063 } 1064 1065 Value *ResultVal = 1066 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 1067 1068 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1069 // Load the value from the alloca. If the NewElt is an aggregate, cast 1070 // the pointer to an integer of the same size before doing the load. 1071 Value *SrcField = NewElts[i]; 1072 const Type *FieldTy = 1073 cast<PointerType>(SrcField->getType())->getElementType(); 1074 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1075 1076 // Ignore zero sized fields like {}, they obviously contain no data. 1077 if (FieldSizeBits == 0) continue; 1078 1079 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 1080 FieldSizeBits); 1081 if (!isa<IntegerType>(FieldTy) && !FieldTy->isFloatingPoint() && 1082 !isa<VectorType>(FieldTy)) 1083 SrcField = new BitCastInst(SrcField, 1084 PointerType::getUnqual(FieldIntTy), 1085 "", LI); 1086 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 1087 1088 // If SrcField is a fp or vector of the right size but that isn't an 1089 // integer type, bitcast to an integer so we can shift it. 1090 if (SrcField->getType() != FieldIntTy) 1091 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 1092 1093 // Zero extend the field to be the same size as the final alloca so that 1094 // we can shift and insert it. 1095 if (SrcField->getType() != ResultVal->getType()) 1096 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 1097 1098 // Determine the number of bits to shift SrcField. 1099 uint64_t Shift; 1100 if (Layout) // Struct case. 1101 Shift = Layout->getElementOffsetInBits(i); 1102 else // Array case. 1103 Shift = i*ArrayEltBitOffset; 1104 1105 if (TD->isBigEndian()) 1106 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 1107 1108 if (Shift) { 1109 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 1110 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 1111 } 1112 1113 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 1114 } 1115 1116 // Handle tail padding by truncating the result 1117 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 1118 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 1119 1120 LI->replaceAllUsesWith(ResultVal); 1121 DeadInsts.push_back(LI); 1122} 1123 1124/// HasPadding - Return true if the specified type has any structure or 1125/// alignment padding, false otherwise. 1126static bool HasPadding(const Type *Ty, const TargetData &TD) { 1127 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 1128 const StructLayout *SL = TD.getStructLayout(STy); 1129 unsigned PrevFieldBitOffset = 0; 1130 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1131 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 1132 1133 // Padding in sub-elements? 1134 if (HasPadding(STy->getElementType(i), TD)) 1135 return true; 1136 1137 // Check to see if there is any padding between this element and the 1138 // previous one. 1139 if (i) { 1140 unsigned PrevFieldEnd = 1141 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 1142 if (PrevFieldEnd < FieldBitOffset) 1143 return true; 1144 } 1145 1146 PrevFieldBitOffset = FieldBitOffset; 1147 } 1148 1149 // Check for tail padding. 1150 if (unsigned EltCount = STy->getNumElements()) { 1151 unsigned PrevFieldEnd = PrevFieldBitOffset + 1152 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 1153 if (PrevFieldEnd < SL->getSizeInBits()) 1154 return true; 1155 } 1156 1157 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1158 return HasPadding(ATy->getElementType(), TD); 1159 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) { 1160 return HasPadding(VTy->getElementType(), TD); 1161 } 1162 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 1163} 1164 1165/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 1166/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 1167/// or 1 if safe after canonicalization has been performed. 1168int SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 1169 // Loop over the use list of the alloca. We can only transform it if all of 1170 // the users are safe to transform. 1171 AllocaInfo Info; 1172 1173 isSafeForScalarRepl(AI, AI, 0, Info); 1174 if (Info.isUnsafe) { 1175 DEBUG(errs() << "Cannot transform: " << *AI << '\n'); 1176 return 0; 1177 } 1178 1179 // Okay, we know all the users are promotable. If the aggregate is a memcpy 1180 // source and destination, we have to be careful. In particular, the memcpy 1181 // could be moving around elements that live in structure padding of the LLVM 1182 // types, but may actually be used. In these cases, we refuse to promote the 1183 // struct. 1184 if (Info.isMemCpySrc && Info.isMemCpyDst && 1185 HasPadding(AI->getAllocatedType(), *TD)) 1186 return 0; 1187 1188 // If we require cleanup, return 1, otherwise return 3. 1189 return Info.needsCleanup ? 1 : 3; 1190} 1191 1192/// CleanupAllocaUsers - If SROA reported that it can promote the specified 1193/// allocation, but only if cleaned up, perform the cleanups required. 1194void SROA::CleanupAllocaUsers(Value *V) { 1195 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); 1196 UI != E; ) { 1197 User *U = *UI++; 1198 Instruction *I = cast<Instruction>(U); 1199 SmallVector<DbgInfoIntrinsic *, 2> DbgInUses; 1200 if (!isa<StoreInst>(I) && OnlyUsedByDbgInfoIntrinsics(I, &DbgInUses)) { 1201 // Safe to remove debug info uses. 1202 while (!DbgInUses.empty()) { 1203 DbgInfoIntrinsic *DI = DbgInUses.back(); DbgInUses.pop_back(); 1204 DI->eraseFromParent(); 1205 } 1206 I->eraseFromParent(); 1207 } 1208 } 1209} 1210 1211/// MergeInType - Add the 'In' type to the accumulated type (Accum) so far at 1212/// the offset specified by Offset (which is specified in bytes). 1213/// 1214/// There are two cases we handle here: 1215/// 1) A union of vector types of the same size and potentially its elements. 1216/// Here we turn element accesses into insert/extract element operations. 1217/// This promotes a <4 x float> with a store of float to the third element 1218/// into a <4 x float> that uses insert element. 1219/// 2) A fully general blob of memory, which we turn into some (potentially 1220/// large) integer type with extract and insert operations where the loads 1221/// and stores would mutate the memory. 1222static void MergeInType(const Type *In, uint64_t Offset, const Type *&VecTy, 1223 unsigned AllocaSize, const TargetData &TD, 1224 LLVMContext &Context) { 1225 // If this could be contributing to a vector, analyze it. 1226 if (VecTy != Type::getVoidTy(Context)) { // either null or a vector type. 1227 1228 // If the In type is a vector that is the same size as the alloca, see if it 1229 // matches the existing VecTy. 1230 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 1231 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 1232 // If we're storing/loading a vector of the right size, allow it as a 1233 // vector. If this the first vector we see, remember the type so that 1234 // we know the element size. 1235 if (VecTy == 0) 1236 VecTy = VInTy; 1237 return; 1238 } 1239 } else if (In->isFloatTy() || In->isDoubleTy() || 1240 (isa<IntegerType>(In) && In->getPrimitiveSizeInBits() >= 8 && 1241 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 1242 // If we're accessing something that could be an element of a vector, see 1243 // if the implied vector agrees with what we already have and if Offset is 1244 // compatible with it. 1245 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 1246 if (Offset % EltSize == 0 && 1247 AllocaSize % EltSize == 0 && 1248 (VecTy == 0 || 1249 cast<VectorType>(VecTy)->getElementType() 1250 ->getPrimitiveSizeInBits()/8 == EltSize)) { 1251 if (VecTy == 0) 1252 VecTy = VectorType::get(In, AllocaSize/EltSize); 1253 return; 1254 } 1255 } 1256 } 1257 1258 // Otherwise, we have a case that we can't handle with an optimized vector 1259 // form. We can still turn this into a large integer. 1260 VecTy = Type::getVoidTy(Context); 1261} 1262 1263/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 1264/// its accesses to a single vector type, return true and set VecTy to 1265/// the new type. If we could convert the alloca into a single promotable 1266/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 1267/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 1268/// is the current offset from the base of the alloca being analyzed. 1269/// 1270/// If we see at least one access to the value that is as a vector type, set the 1271/// SawVec flag. 1272bool SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy, 1273 bool &SawVec, uint64_t Offset, 1274 unsigned AllocaSize) { 1275 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1276 Instruction *User = cast<Instruction>(*UI); 1277 1278 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1279 // Don't break volatile loads. 1280 if (LI->isVolatile()) 1281 return false; 1282 MergeInType(LI->getType(), Offset, VecTy, 1283 AllocaSize, *TD, V->getContext()); 1284 SawVec |= isa<VectorType>(LI->getType()); 1285 continue; 1286 } 1287 1288 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1289 // Storing the pointer, not into the value? 1290 if (SI->getOperand(0) == V || SI->isVolatile()) return 0; 1291 MergeInType(SI->getOperand(0)->getType(), Offset, 1292 VecTy, AllocaSize, *TD, V->getContext()); 1293 SawVec |= isa<VectorType>(SI->getOperand(0)->getType()); 1294 continue; 1295 } 1296 1297 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 1298 if (!CanConvertToScalar(BCI, IsNotTrivial, VecTy, SawVec, Offset, 1299 AllocaSize)) 1300 return false; 1301 IsNotTrivial = true; 1302 continue; 1303 } 1304 1305 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1306 // If this is a GEP with a variable indices, we can't handle it. 1307 if (!GEP->hasAllConstantIndices()) 1308 return false; 1309 1310 // Compute the offset that this GEP adds to the pointer. 1311 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1312 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getPointerOperandType(), 1313 &Indices[0], Indices.size()); 1314 // See if all uses can be converted. 1315 if (!CanConvertToScalar(GEP, IsNotTrivial, VecTy, SawVec,Offset+GEPOffset, 1316 AllocaSize)) 1317 return false; 1318 IsNotTrivial = true; 1319 continue; 1320 } 1321 1322 // If this is a constant sized memset of a constant value (e.g. 0) we can 1323 // handle it. 1324 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 1325 // Store of constant value and constant size. 1326 if (isa<ConstantInt>(MSI->getValue()) && 1327 isa<ConstantInt>(MSI->getLength())) { 1328 IsNotTrivial = true; 1329 continue; 1330 } 1331 } 1332 1333 // If this is a memcpy or memmove into or out of the whole allocation, we 1334 // can handle it like a load or store of the scalar type. 1335 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 1336 if (ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength())) 1337 if (Len->getZExtValue() == AllocaSize && Offset == 0) { 1338 IsNotTrivial = true; 1339 continue; 1340 } 1341 } 1342 1343 // Ignore dbg intrinsic. 1344 if (isa<DbgInfoIntrinsic>(User)) 1345 continue; 1346 1347 // Otherwise, we cannot handle this! 1348 return false; 1349 } 1350 1351 return true; 1352} 1353 1354/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 1355/// directly. This happens when we are converting an "integer union" to a 1356/// single integer scalar, or when we are converting a "vector union" to a 1357/// vector with insert/extractelement instructions. 1358/// 1359/// Offset is an offset from the original alloca, in bits that need to be 1360/// shifted to the right. By the end of this, there should be no uses of Ptr. 1361void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset) { 1362 while (!Ptr->use_empty()) { 1363 Instruction *User = cast<Instruction>(Ptr->use_back()); 1364 1365 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 1366 ConvertUsesToScalar(CI, NewAI, Offset); 1367 CI->eraseFromParent(); 1368 continue; 1369 } 1370 1371 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1372 // Compute the offset that this GEP adds to the pointer. 1373 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1374 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getPointerOperandType(), 1375 &Indices[0], Indices.size()); 1376 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 1377 GEP->eraseFromParent(); 1378 continue; 1379 } 1380 1381 IRBuilder<> Builder(User->getParent(), User); 1382 1383 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1384 // The load is a bit extract from NewAI shifted right by Offset bits. 1385 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 1386 Value *NewLoadVal 1387 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 1388 LI->replaceAllUsesWith(NewLoadVal); 1389 LI->eraseFromParent(); 1390 continue; 1391 } 1392 1393 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1394 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 1395 // FIXME: Remove once builder has Twine API. 1396 Value *Old = Builder.CreateLoad(NewAI, 1397 (NewAI->getName()+".in").str().c_str()); 1398 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 1399 Builder); 1400 Builder.CreateStore(New, NewAI); 1401 SI->eraseFromParent(); 1402 continue; 1403 } 1404 1405 // If this is a constant sized memset of a constant value (e.g. 0) we can 1406 // transform it into a store of the expanded constant value. 1407 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 1408 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 1409 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 1410 if (NumBytes != 0) { 1411 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 1412 1413 // Compute the value replicated the right number of times. 1414 APInt APVal(NumBytes*8, Val); 1415 1416 // Splat the value if non-zero. 1417 if (Val) 1418 for (unsigned i = 1; i != NumBytes; ++i) 1419 APVal |= APVal << 8; 1420 1421 // FIXME: Remove once builder has Twine API. 1422 Value *Old = Builder.CreateLoad(NewAI, 1423 (NewAI->getName()+".in").str().c_str()); 1424 Value *New = ConvertScalar_InsertValue( 1425 ConstantInt::get(User->getContext(), APVal), 1426 Old, Offset, Builder); 1427 Builder.CreateStore(New, NewAI); 1428 } 1429 MSI->eraseFromParent(); 1430 continue; 1431 } 1432 1433 // If this is a memcpy or memmove into or out of the whole allocation, we 1434 // can handle it like a load or store of the scalar type. 1435 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 1436 assert(Offset == 0 && "must be store to start of alloca"); 1437 1438 // If the source and destination are both to the same alloca, then this is 1439 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 1440 // as appropriate. 1441 AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject()); 1442 1443 if (MTI->getSource()->getUnderlyingObject() != OrigAI) { 1444 // Dest must be OrigAI, change this to be a load from the original 1445 // pointer (bitcasted), then a store to our new alloca. 1446 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 1447 Value *SrcPtr = MTI->getSource(); 1448 SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType()); 1449 1450 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 1451 SrcVal->setAlignment(MTI->getAlignment()); 1452 Builder.CreateStore(SrcVal, NewAI); 1453 } else if (MTI->getDest()->getUnderlyingObject() != OrigAI) { 1454 // Src must be OrigAI, change this to be a load from NewAI then a store 1455 // through the original dest pointer (bitcasted). 1456 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 1457 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 1458 1459 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType()); 1460 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 1461 NewStore->setAlignment(MTI->getAlignment()); 1462 } else { 1463 // Noop transfer. Src == Dst 1464 } 1465 1466 1467 MTI->eraseFromParent(); 1468 continue; 1469 } 1470 1471 // If user is a dbg info intrinsic then it is safe to remove it. 1472 if (isa<DbgInfoIntrinsic>(User)) { 1473 User->eraseFromParent(); 1474 continue; 1475 } 1476 1477 llvm_unreachable("Unsupported operation!"); 1478 } 1479} 1480 1481/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 1482/// or vector value FromVal, extracting the bits from the offset specified by 1483/// Offset. This returns the value, which is of type ToType. 1484/// 1485/// This happens when we are converting an "integer union" to a single 1486/// integer scalar, or when we are converting a "vector union" to a vector with 1487/// insert/extractelement instructions. 1488/// 1489/// Offset is an offset from the original alloca, in bits that need to be 1490/// shifted to the right. 1491Value *SROA::ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 1492 uint64_t Offset, IRBuilder<> &Builder) { 1493 // If the load is of the whole new alloca, no conversion is needed. 1494 if (FromVal->getType() == ToType && Offset == 0) 1495 return FromVal; 1496 1497 // If the result alloca is a vector type, this is either an element 1498 // access or a bitcast to another vector type of the same size. 1499 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) { 1500 if (isa<VectorType>(ToType)) 1501 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 1502 1503 // Otherwise it must be an element access. 1504 unsigned Elt = 0; 1505 if (Offset) { 1506 unsigned EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 1507 Elt = Offset/EltSize; 1508 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 1509 } 1510 // Return the element extracted out of it. 1511 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get( 1512 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp"); 1513 if (V->getType() != ToType) 1514 V = Builder.CreateBitCast(V, ToType, "tmp"); 1515 return V; 1516 } 1517 1518 // If ToType is a first class aggregate, extract out each of the pieces and 1519 // use insertvalue's to form the FCA. 1520 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 1521 const StructLayout &Layout = *TD->getStructLayout(ST); 1522 Value *Res = UndefValue::get(ST); 1523 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1524 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 1525 Offset+Layout.getElementOffsetInBits(i), 1526 Builder); 1527 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1528 } 1529 return Res; 1530 } 1531 1532 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 1533 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 1534 Value *Res = UndefValue::get(AT); 1535 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1536 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 1537 Offset+i*EltSize, Builder); 1538 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1539 } 1540 return Res; 1541 } 1542 1543 // Otherwise, this must be a union that was converted to an integer value. 1544 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 1545 1546 // If this is a big-endian system and the load is narrower than the 1547 // full alloca type, we need to do a shift to get the right bits. 1548 int ShAmt = 0; 1549 if (TD->isBigEndian()) { 1550 // On big-endian machines, the lowest bit is stored at the bit offset 1551 // from the pointer given by getTypeStoreSizeInBits. This matters for 1552 // integers with a bitwidth that is not a multiple of 8. 1553 ShAmt = TD->getTypeStoreSizeInBits(NTy) - 1554 TD->getTypeStoreSizeInBits(ToType) - Offset; 1555 } else { 1556 ShAmt = Offset; 1557 } 1558 1559 // Note: we support negative bitwidths (with shl) which are not defined. 1560 // We do this to support (f.e.) loads off the end of a structure where 1561 // only some bits are used. 1562 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 1563 FromVal = Builder.CreateLShr(FromVal, 1564 ConstantInt::get(FromVal->getType(), 1565 ShAmt), "tmp"); 1566 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 1567 FromVal = Builder.CreateShl(FromVal, 1568 ConstantInt::get(FromVal->getType(), 1569 -ShAmt), "tmp"); 1570 1571 // Finally, unconditionally truncate the integer to the right width. 1572 unsigned LIBitWidth = TD->getTypeSizeInBits(ToType); 1573 if (LIBitWidth < NTy->getBitWidth()) 1574 FromVal = 1575 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 1576 LIBitWidth), "tmp"); 1577 else if (LIBitWidth > NTy->getBitWidth()) 1578 FromVal = 1579 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 1580 LIBitWidth), "tmp"); 1581 1582 // If the result is an integer, this is a trunc or bitcast. 1583 if (isa<IntegerType>(ToType)) { 1584 // Should be done. 1585 } else if (ToType->isFloatingPoint() || isa<VectorType>(ToType)) { 1586 // Just do a bitcast, we know the sizes match up. 1587 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 1588 } else { 1589 // Otherwise must be a pointer. 1590 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 1591 } 1592 assert(FromVal->getType() == ToType && "Didn't convert right?"); 1593 return FromVal; 1594} 1595 1596/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 1597/// or vector value "Old" at the offset specified by Offset. 1598/// 1599/// This happens when we are converting an "integer union" to a 1600/// single integer scalar, or when we are converting a "vector union" to a 1601/// vector with insert/extractelement instructions. 1602/// 1603/// Offset is an offset from the original alloca, in bits that need to be 1604/// shifted to the right. 1605Value *SROA::ConvertScalar_InsertValue(Value *SV, Value *Old, 1606 uint64_t Offset, IRBuilder<> &Builder) { 1607 1608 // Convert the stored type to the actual type, shift it left to insert 1609 // then 'or' into place. 1610 const Type *AllocaType = Old->getType(); 1611 LLVMContext &Context = Old->getContext(); 1612 1613 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 1614 uint64_t VecSize = TD->getTypeAllocSizeInBits(VTy); 1615 uint64_t ValSize = TD->getTypeAllocSizeInBits(SV->getType()); 1616 1617 // Changing the whole vector with memset or with an access of a different 1618 // vector type? 1619 if (ValSize == VecSize) 1620 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 1621 1622 uint64_t EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 1623 1624 // Must be an element insertion. 1625 unsigned Elt = Offset/EltSize; 1626 1627 if (SV->getType() != VTy->getElementType()) 1628 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 1629 1630 SV = Builder.CreateInsertElement(Old, SV, 1631 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt), 1632 "tmp"); 1633 return SV; 1634 } 1635 1636 // If SV is a first-class aggregate value, insert each value recursively. 1637 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 1638 const StructLayout &Layout = *TD->getStructLayout(ST); 1639 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1640 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1641 Old = ConvertScalar_InsertValue(Elt, Old, 1642 Offset+Layout.getElementOffsetInBits(i), 1643 Builder); 1644 } 1645 return Old; 1646 } 1647 1648 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 1649 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 1650 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1651 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1652 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 1653 } 1654 return Old; 1655 } 1656 1657 // If SV is a float, convert it to the appropriate integer type. 1658 // If it is a pointer, do the same. 1659 unsigned SrcWidth = TD->getTypeSizeInBits(SV->getType()); 1660 unsigned DestWidth = TD->getTypeSizeInBits(AllocaType); 1661 unsigned SrcStoreWidth = TD->getTypeStoreSizeInBits(SV->getType()); 1662 unsigned DestStoreWidth = TD->getTypeStoreSizeInBits(AllocaType); 1663 if (SV->getType()->isFloatingPoint() || isa<VectorType>(SV->getType())) 1664 SV = Builder.CreateBitCast(SV, 1665 IntegerType::get(SV->getContext(),SrcWidth), "tmp"); 1666 else if (isa<PointerType>(SV->getType())) 1667 SV = Builder.CreatePtrToInt(SV, TD->getIntPtrType(SV->getContext()), "tmp"); 1668 1669 // Zero extend or truncate the value if needed. 1670 if (SV->getType() != AllocaType) { 1671 if (SV->getType()->getPrimitiveSizeInBits() < 1672 AllocaType->getPrimitiveSizeInBits()) 1673 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 1674 else { 1675 // Truncation may be needed if storing more than the alloca can hold 1676 // (undefined behavior). 1677 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 1678 SrcWidth = DestWidth; 1679 SrcStoreWidth = DestStoreWidth; 1680 } 1681 } 1682 1683 // If this is a big-endian system and the store is narrower than the 1684 // full alloca type, we need to do a shift to get the right bits. 1685 int ShAmt = 0; 1686 if (TD->isBigEndian()) { 1687 // On big-endian machines, the lowest bit is stored at the bit offset 1688 // from the pointer given by getTypeStoreSizeInBits. This matters for 1689 // integers with a bitwidth that is not a multiple of 8. 1690 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 1691 } else { 1692 ShAmt = Offset; 1693 } 1694 1695 // Note: we support negative bitwidths (with shr) which are not defined. 1696 // We do this to support (f.e.) stores off the end of a structure where 1697 // only some bits in the structure are set. 1698 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 1699 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 1700 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 1701 ShAmt), "tmp"); 1702 Mask <<= ShAmt; 1703 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1704 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 1705 -ShAmt), "tmp"); 1706 Mask = Mask.lshr(-ShAmt); 1707 } 1708 1709 // Mask out the bits we are about to insert from the old value, and or 1710 // in the new bits. 1711 if (SrcWidth != DestWidth) { 1712 assert(DestWidth > SrcWidth); 1713 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 1714 SV = Builder.CreateOr(Old, SV, "ins"); 1715 } 1716 return SV; 1717} 1718 1719 1720 1721/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1722/// some part of a constant global variable. This intentionally only accepts 1723/// constant expressions because we don't can't rewrite arbitrary instructions. 1724static bool PointsToConstantGlobal(Value *V) { 1725 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1726 return GV->isConstant(); 1727 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1728 if (CE->getOpcode() == Instruction::BitCast || 1729 CE->getOpcode() == Instruction::GetElementPtr) 1730 return PointsToConstantGlobal(CE->getOperand(0)); 1731 return false; 1732} 1733 1734/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1735/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1736/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1737/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1738/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1739/// the alloca, and if the source pointer is a pointer to a constant global, we 1740/// can optimize this. 1741static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy, 1742 bool isOffset) { 1743 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1744 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) 1745 // Ignore non-volatile loads, they are always ok. 1746 if (!LI->isVolatile()) 1747 continue; 1748 1749 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) { 1750 // If uses of the bitcast are ok, we are ok. 1751 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 1752 return false; 1753 continue; 1754 } 1755 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) { 1756 // If the GEP has all zero indices, it doesn't offset the pointer. If it 1757 // doesn't, it does. 1758 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 1759 isOffset || !GEP->hasAllZeroIndices())) 1760 return false; 1761 continue; 1762 } 1763 1764 // If this is isn't our memcpy/memmove, reject it as something we can't 1765 // handle. 1766 if (!isa<MemTransferInst>(*UI)) 1767 return false; 1768 1769 // If we already have seen a copy, reject the second one. 1770 if (TheCopy) return false; 1771 1772 // If the pointer has been offset from the start of the alloca, we can't 1773 // safely handle this. 1774 if (isOffset) return false; 1775 1776 // If the memintrinsic isn't using the alloca as the dest, reject it. 1777 if (UI.getOperandNo() != 1) return false; 1778 1779 MemIntrinsic *MI = cast<MemIntrinsic>(*UI); 1780 1781 // If the source of the memcpy/move is not a constant global, reject it. 1782 if (!PointsToConstantGlobal(MI->getOperand(2))) 1783 return false; 1784 1785 // Otherwise, the transform is safe. Remember the copy instruction. 1786 TheCopy = MI; 1787 } 1788 return true; 1789} 1790 1791/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 1792/// modified by a copy from a constant global. If we can prove this, we can 1793/// replace any uses of the alloca with uses of the global directly. 1794Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) { 1795 Instruction *TheCopy = 0; 1796 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 1797 return TheCopy; 1798 return 0; 1799} 1800