ScalarReplAggregates.cpp revision d614a1f0a89927639db37efd8840fe006dacd7a1
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 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures 78 /// information about the uses. All these fields are initialized to false 79 /// and set to true when something is learned. 80 struct AllocaInfo { 81 /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 82 bool isUnsafe : 1; 83 84 /// needsCleanup - This is set to true if there is some use of the alloca 85 /// that requires cleanup. 86 bool needsCleanup : 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), needsCleanup(false), 96 isMemCpySrc(false), isMemCpyDst(false) {} 97 }; 98 99 unsigned SRThreshold; 100 101 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; } 102 103 int isSafeAllocaToScalarRepl(AllocaInst *AI); 104 105 void isSafeUseOfAllocation(Instruction *User, AllocaInst *AI, 106 AllocaInfo &Info); 107 void isSafeElementUse(Value *Ptr, bool isFirstElt, AllocaInst *AI, 108 AllocaInfo &Info); 109 void isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocaInst *AI, 110 unsigned OpNo, AllocaInfo &Info); 111 void isSafeUseOfBitCastedAllocation(BitCastInst *User, AllocaInst *AI, 112 AllocaInfo &Info); 113 114 void DoScalarReplacement(AllocaInst *AI, 115 std::vector<AllocaInst*> &WorkList); 116 void CleanupGEP(GetElementPtrInst *GEP); 117 void CleanupAllocaUsers(AllocaInst *AI); 118 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocaInst *Base); 119 120 void RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocaInst *AI, 121 SmallVector<AllocaInst*, 32> &NewElts); 122 123 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst, 124 AllocaInst *AI, 125 SmallVector<AllocaInst*, 32> &NewElts); 126 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 127 SmallVector<AllocaInst*, 32> &NewElts); 128 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 129 SmallVector<AllocaInst*, 32> &NewElts); 130 131 bool CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy, 132 bool &SawVec, uint64_t Offset, unsigned AllocaSize); 133 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); 134 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType, 135 uint64_t Offset, IRBuilder<> &Builder); 136 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 137 uint64_t Offset, IRBuilder<> &Builder); 138 static Instruction *isOnlyCopiedFromConstantGlobal(AllocaInst *AI); 139 }; 140} 141 142char SROA::ID = 0; 143static RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates"); 144 145// Public interface to the ScalarReplAggregates pass 146FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) { 147 return new SROA(Threshold); 148} 149 150 151bool SROA::runOnFunction(Function &F) { 152 TD = getAnalysisIfAvailable<TargetData>(); 153 154 bool Changed = performPromotion(F); 155 156 // FIXME: ScalarRepl currently depends on TargetData more than it 157 // theoretically needs to. It should be refactored in order to support 158 // target-independent IR. Until this is done, just skip the actual 159 // scalar-replacement portion of this pass. 160 if (!TD) return Changed; 161 162 while (1) { 163 bool LocalChange = performScalarRepl(F); 164 if (!LocalChange) break; // No need to repromote if no scalarrepl 165 Changed = true; 166 LocalChange = performPromotion(F); 167 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 168 } 169 170 return Changed; 171} 172 173 174bool SROA::performPromotion(Function &F) { 175 std::vector<AllocaInst*> Allocas; 176 DominatorTree &DT = getAnalysis<DominatorTree>(); 177 DominanceFrontier &DF = getAnalysis<DominanceFrontier>(); 178 179 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 180 181 bool Changed = false; 182 183 while (1) { 184 Allocas.clear(); 185 186 // Find allocas that are safe to promote, by looking at all instructions in 187 // the entry node 188 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 189 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 190 if (isAllocaPromotable(AI)) 191 Allocas.push_back(AI); 192 193 if (Allocas.empty()) break; 194 195 PromoteMemToReg(Allocas, DT, DF); 196 NumPromoted += Allocas.size(); 197 Changed = true; 198 } 199 200 return Changed; 201} 202 203/// getNumSAElements - Return the number of elements in the specific struct or 204/// array. 205static uint64_t getNumSAElements(const Type *T) { 206 if (const StructType *ST = dyn_cast<StructType>(T)) 207 return ST->getNumElements(); 208 return cast<ArrayType>(T)->getNumElements(); 209} 210 211// performScalarRepl - This algorithm is a simple worklist driven algorithm, 212// which runs on all of the malloc/alloca instructions in the function, removing 213// them if they are only used by getelementptr instructions. 214// 215bool SROA::performScalarRepl(Function &F) { 216 std::vector<AllocaInst*> WorkList; 217 218 // Scan the entry basic block, adding any alloca's and mallocs to the worklist 219 BasicBlock &BB = F.getEntryBlock(); 220 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 221 if (AllocaInst *A = dyn_cast<AllocaInst>(I)) 222 WorkList.push_back(A); 223 224 // Process the worklist 225 bool Changed = false; 226 while (!WorkList.empty()) { 227 AllocaInst *AI = WorkList.back(); 228 WorkList.pop_back(); 229 230 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 231 // with unused elements. 232 if (AI->use_empty()) { 233 AI->eraseFromParent(); 234 continue; 235 } 236 237 // If this alloca is impossible for us to promote, reject it early. 238 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 239 continue; 240 241 // Check to see if this allocation is only modified by a memcpy/memmove from 242 // a constant global. If this is the case, we can change all users to use 243 // the constant global instead. This is commonly produced by the CFE by 244 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 245 // is only subsequently read. 246 if (Instruction *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 247 DEBUG(errs() << "Found alloca equal to global: " << *AI << '\n'); 248 DEBUG(errs() << " memcpy = " << *TheCopy << '\n'); 249 Constant *TheSrc = cast<Constant>(TheCopy->getOperand(2)); 250 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 251 TheCopy->eraseFromParent(); // Don't mutate the global. 252 AI->eraseFromParent(); 253 ++NumGlobals; 254 Changed = true; 255 continue; 256 } 257 258 // Check to see if we can perform the core SROA transformation. We cannot 259 // transform the allocation instruction if it is an array allocation 260 // (allocations OF arrays are ok though), and an allocation of a scalar 261 // value cannot be decomposed at all. 262 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 263 264 // Do not promote [0 x %struct]. 265 if (AllocaSize == 0) continue; 266 267 // Do not promote any struct whose size is too big. 268 if (AllocaSize > SRThreshold) continue; 269 270 if ((isa<StructType>(AI->getAllocatedType()) || 271 isa<ArrayType>(AI->getAllocatedType())) && 272 // Do not promote any struct into more than "32" separate vars. 273 getNumSAElements(AI->getAllocatedType()) <= SRThreshold/4) { 274 // Check that all of the users of the allocation are capable of being 275 // transformed. 276 switch (isSafeAllocaToScalarRepl(AI)) { 277 default: llvm_unreachable("Unexpected value!"); 278 case 0: // Not safe to scalar replace. 279 break; 280 case 1: // Safe, but requires cleanup/canonicalizations first 281 CleanupAllocaUsers(AI); 282 // FALL THROUGH. 283 case 3: // Safe to scalar replace. 284 DoScalarReplacement(AI, WorkList); 285 Changed = true; 286 continue; 287 } 288 } 289 290 // If we can turn this aggregate value (potentially with casts) into a 291 // simple scalar value that can be mem2reg'd into a register value. 292 // IsNotTrivial tracks whether this is something that mem2reg could have 293 // promoted itself. If so, we don't want to transform it needlessly. Note 294 // that we can't just check based on the type: the alloca may be of an i32 295 // but that has pointer arithmetic to set byte 3 of it or something. 296 bool IsNotTrivial = false; 297 const Type *VectorTy = 0; 298 bool HadAVector = false; 299 if (CanConvertToScalar(AI, IsNotTrivial, VectorTy, HadAVector, 300 0, unsigned(AllocaSize)) && IsNotTrivial) { 301 AllocaInst *NewAI; 302 // If we were able to find a vector type that can handle this with 303 // insert/extract elements, and if there was at least one use that had 304 // a vector type, promote this to a vector. We don't want to promote 305 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 306 // we just get a lot of insert/extracts. If at least one vector is 307 // involved, then we probably really do have a union of vector/array. 308 if (VectorTy && isa<VectorType>(VectorTy) && HadAVector) { 309 DEBUG(errs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = " 310 << *VectorTy << '\n'); 311 312 // Create and insert the vector alloca. 313 NewAI = new AllocaInst(VectorTy, 0, "", AI->getParent()->begin()); 314 ConvertUsesToScalar(AI, NewAI, 0); 315 } else { 316 DEBUG(errs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"); 317 318 // Create and insert the integer alloca. 319 const Type *NewTy = IntegerType::get(AI->getContext(), AllocaSize*8); 320 NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 321 ConvertUsesToScalar(AI, NewAI, 0); 322 } 323 NewAI->takeName(AI); 324 AI->eraseFromParent(); 325 ++NumConverted; 326 Changed = true; 327 continue; 328 } 329 330 // Otherwise, couldn't process this alloca. 331 } 332 333 return Changed; 334} 335 336/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 337/// predicate, do SROA now. 338void SROA::DoScalarReplacement(AllocaInst *AI, 339 std::vector<AllocaInst*> &WorkList) { 340 DEBUG(errs() << "Found inst to SROA: " << *AI << '\n'); 341 SmallVector<AllocaInst*, 32> ElementAllocas; 342 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 343 ElementAllocas.reserve(ST->getNumContainedTypes()); 344 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 345 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 346 AI->getAlignment(), 347 AI->getName() + "." + Twine(i), AI); 348 ElementAllocas.push_back(NA); 349 WorkList.push_back(NA); // Add to worklist for recursive processing 350 } 351 } else { 352 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 353 ElementAllocas.reserve(AT->getNumElements()); 354 const Type *ElTy = AT->getElementType(); 355 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 356 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 357 AI->getName() + "." + Twine(i), AI); 358 ElementAllocas.push_back(NA); 359 WorkList.push_back(NA); // Add to worklist for recursive processing 360 } 361 } 362 363 // Now that we have created the alloca instructions that we want to use, 364 // expand the getelementptr instructions to use them. 365 // 366 while (!AI->use_empty()) { 367 Instruction *User = cast<Instruction>(AI->use_back()); 368 if (BitCastInst *BCInst = dyn_cast<BitCastInst>(User)) { 369 RewriteBitCastUserOfAlloca(BCInst, AI, ElementAllocas); 370 BCInst->eraseFromParent(); 371 continue; 372 } 373 374 // Replace: 375 // %res = load { i32, i32 }* %alloc 376 // with: 377 // %load.0 = load i32* %alloc.0 378 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 379 // %load.1 = load i32* %alloc.1 380 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 381 // (Also works for arrays instead of structs) 382 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 383 Value *Insert = UndefValue::get(LI->getType()); 384 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) { 385 Value *Load = new LoadInst(ElementAllocas[i], "load", LI); 386 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 387 } 388 LI->replaceAllUsesWith(Insert); 389 LI->eraseFromParent(); 390 continue; 391 } 392 393 // Replace: 394 // store { i32, i32 } %val, { i32, i32 }* %alloc 395 // with: 396 // %val.0 = extractvalue { i32, i32 } %val, 0 397 // store i32 %val.0, i32* %alloc.0 398 // %val.1 = extractvalue { i32, i32 } %val, 1 399 // store i32 %val.1, i32* %alloc.1 400 // (Also works for arrays instead of structs) 401 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 402 Value *Val = SI->getOperand(0); 403 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) { 404 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 405 new StoreInst(Extract, ElementAllocas[i], SI); 406 } 407 SI->eraseFromParent(); 408 continue; 409 } 410 411 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User); 412 // We now know that the GEP is of the form: GEP <ptr>, 0, <cst> 413 unsigned Idx = 414 (unsigned)cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue(); 415 416 assert(Idx < ElementAllocas.size() && "Index out of range?"); 417 AllocaInst *AllocaToUse = ElementAllocas[Idx]; 418 419 Value *RepValue; 420 if (GEPI->getNumOperands() == 3) { 421 // Do not insert a new getelementptr instruction with zero indices, only 422 // to have it optimized out later. 423 RepValue = AllocaToUse; 424 } else { 425 // We are indexing deeply into the structure, so we still need a 426 // getelement ptr instruction to finish the indexing. This may be 427 // expanded itself once the worklist is rerun. 428 // 429 SmallVector<Value*, 8> NewArgs; 430 NewArgs.push_back(Constant::getNullValue( 431 Type::getInt32Ty(AI->getContext()))); 432 NewArgs.append(GEPI->op_begin()+3, GEPI->op_end()); 433 RepValue = GetElementPtrInst::Create(AllocaToUse, NewArgs.begin(), 434 NewArgs.end(), "", GEPI); 435 RepValue->takeName(GEPI); 436 } 437 438 // If this GEP is to the start of the aggregate, check for memcpys. 439 if (Idx == 0 && GEPI->hasAllZeroIndices()) 440 RewriteBitCastUserOfAlloca(GEPI, AI, ElementAllocas); 441 442 // Move all of the users over to the new GEP. 443 GEPI->replaceAllUsesWith(RepValue); 444 // Delete the old GEP 445 GEPI->eraseFromParent(); 446 } 447 448 // Finally, delete the Alloca instruction 449 AI->eraseFromParent(); 450 NumReplaced++; 451} 452 453 454/// isSafeElementUse - Check to see if this use is an allowed use for a 455/// getelementptr instruction of an array aggregate allocation. isFirstElt 456/// indicates whether Ptr is known to the start of the aggregate. 457/// 458void SROA::isSafeElementUse(Value *Ptr, bool isFirstElt, AllocaInst *AI, 459 AllocaInfo &Info) { 460 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 461 I != E; ++I) { 462 Instruction *User = cast<Instruction>(*I); 463 switch (User->getOpcode()) { 464 case Instruction::Load: break; 465 case Instruction::Store: 466 // Store is ok if storing INTO the pointer, not storing the pointer 467 if (User->getOperand(0) == Ptr) return MarkUnsafe(Info); 468 break; 469 case Instruction::GetElementPtr: { 470 GetElementPtrInst *GEP = cast<GetElementPtrInst>(User); 471 bool AreAllZeroIndices = isFirstElt; 472 if (GEP->getNumOperands() > 1 && 473 (!isa<ConstantInt>(GEP->getOperand(1)) || 474 !cast<ConstantInt>(GEP->getOperand(1))->isZero())) 475 // Using pointer arithmetic to navigate the array. 476 return MarkUnsafe(Info); 477 478 // Verify that any array subscripts are in range. 479 for (gep_type_iterator GEPIt = gep_type_begin(GEP), 480 E = gep_type_end(GEP); GEPIt != E; ++GEPIt) { 481 // Ignore struct elements, no extra checking needed for these. 482 if (isa<StructType>(*GEPIt)) 483 continue; 484 485 // This GEP indexes an array. Verify that this is an in-range 486 // constant integer. Specifically, consider A[0][i]. We cannot know that 487 // the user isn't doing invalid things like allowing i to index an 488 // out-of-range subscript that accesses A[1]. Because of this, we have 489 // to reject SROA of any accesses into structs where any of the 490 // components are variables. 491 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 492 if (!IdxVal) return MarkUnsafe(Info); 493 494 // Are all indices still zero? 495 AreAllZeroIndices &= IdxVal->isZero(); 496 497 if (const ArrayType *AT = dyn_cast<ArrayType>(*GEPIt)) { 498 if (IdxVal->getZExtValue() >= AT->getNumElements()) 499 return MarkUnsafe(Info); 500 } else if (const VectorType *VT = dyn_cast<VectorType>(*GEPIt)) { 501 if (IdxVal->getZExtValue() >= VT->getNumElements()) 502 return MarkUnsafe(Info); 503 } 504 } 505 506 isSafeElementUse(GEP, AreAllZeroIndices, AI, Info); 507 if (Info.isUnsafe) return; 508 break; 509 } 510 case Instruction::BitCast: 511 if (isFirstElt) { 512 isSafeUseOfBitCastedAllocation(cast<BitCastInst>(User), AI, Info); 513 if (Info.isUnsafe) return; 514 break; 515 } 516 DEBUG(errs() << " Transformation preventing inst: " << *User << '\n'); 517 return MarkUnsafe(Info); 518 case Instruction::Call: 519 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 520 if (isFirstElt) { 521 isSafeMemIntrinsicOnAllocation(MI, AI, I.getOperandNo(), Info); 522 if (Info.isUnsafe) return; 523 break; 524 } 525 } 526 DEBUG(errs() << " Transformation preventing inst: " << *User << '\n'); 527 return MarkUnsafe(Info); 528 default: 529 DEBUG(errs() << " Transformation preventing inst: " << *User << '\n'); 530 return MarkUnsafe(Info); 531 } 532 } 533 return; // All users look ok :) 534} 535 536/// AllUsersAreLoads - Return true if all users of this value are loads. 537static bool AllUsersAreLoads(Value *Ptr) { 538 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 539 I != E; ++I) 540 if (cast<Instruction>(*I)->getOpcode() != Instruction::Load) 541 return false; 542 return true; 543} 544 545/// isSafeUseOfAllocation - Check to see if this user is an allowed use for an 546/// aggregate allocation. 547/// 548void SROA::isSafeUseOfAllocation(Instruction *User, AllocaInst *AI, 549 AllocaInfo &Info) { 550 if (BitCastInst *C = dyn_cast<BitCastInst>(User)) 551 return isSafeUseOfBitCastedAllocation(C, AI, Info); 552 553 if (LoadInst *LI = dyn_cast<LoadInst>(User)) 554 if (!LI->isVolatile()) 555 return;// Loads (returning a first class aggregrate) are always rewritable 556 557 if (StoreInst *SI = dyn_cast<StoreInst>(User)) 558 if (!SI->isVolatile() && SI->getOperand(0) != AI) 559 return;// Store is ok if storing INTO the pointer, not storing the pointer 560 561 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User); 562 if (GEPI == 0) 563 return MarkUnsafe(Info); 564 565 gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI); 566 567 // The GEP is not safe to transform if not of the form "GEP <ptr>, 0, <cst>". 568 if (I == E || 569 I.getOperand() != Constant::getNullValue(I.getOperand()->getType())) { 570 return MarkUnsafe(Info); 571 } 572 573 ++I; 574 if (I == E) return MarkUnsafe(Info); // ran out of GEP indices?? 575 576 bool IsAllZeroIndices = true; 577 578 // If the first index is a non-constant index into an array, see if we can 579 // handle it as a special case. 580 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 581 if (!isa<ConstantInt>(I.getOperand())) { 582 IsAllZeroIndices = 0; 583 uint64_t NumElements = AT->getNumElements(); 584 585 // If this is an array index and the index is not constant, we cannot 586 // promote... that is unless the array has exactly one or two elements in 587 // it, in which case we CAN promote it, but we have to canonicalize this 588 // out if this is the only problem. 589 if ((NumElements == 1 || NumElements == 2) && 590 AllUsersAreLoads(GEPI)) { 591 Info.needsCleanup = true; 592 return; // Canonicalization required! 593 } 594 return MarkUnsafe(Info); 595 } 596 } 597 598 // Walk through the GEP type indices, checking the types that this indexes 599 // into. 600 for (; I != E; ++I) { 601 // Ignore struct elements, no extra checking needed for these. 602 if (isa<StructType>(*I)) 603 continue; 604 605 ConstantInt *IdxVal = dyn_cast<ConstantInt>(I.getOperand()); 606 if (!IdxVal) return MarkUnsafe(Info); 607 608 // Are all indices still zero? 609 IsAllZeroIndices &= IdxVal->isZero(); 610 611 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 612 // This GEP indexes an array. Verify that this is an in-range constant 613 // integer. Specifically, consider A[0][i]. We cannot know that the user 614 // isn't doing invalid things like allowing i to index an out-of-range 615 // subscript that accesses A[1]. Because of this, we have to reject SROA 616 // of any accesses into structs where any of the components are variables. 617 if (IdxVal->getZExtValue() >= AT->getNumElements()) 618 return MarkUnsafe(Info); 619 } else if (const VectorType *VT = dyn_cast<VectorType>(*I)) { 620 if (IdxVal->getZExtValue() >= VT->getNumElements()) 621 return MarkUnsafe(Info); 622 } 623 } 624 625 // If there are any non-simple uses of this getelementptr, make sure to reject 626 // them. 627 return isSafeElementUse(GEPI, IsAllZeroIndices, AI, Info); 628} 629 630/// isSafeMemIntrinsicOnAllocation - Return true if the specified memory 631/// intrinsic can be promoted by SROA. At this point, we know that the operand 632/// of the memintrinsic is a pointer to the beginning of the allocation. 633void SROA::isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocaInst *AI, 634 unsigned OpNo, AllocaInfo &Info) { 635 // If not constant length, give up. 636 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 637 if (!Length) return MarkUnsafe(Info); 638 639 // If not the whole aggregate, give up. 640 if (Length->getZExtValue() != 641 TD->getTypeAllocSize(AI->getType()->getElementType())) 642 return MarkUnsafe(Info); 643 644 // We only know about memcpy/memset/memmove. 645 if (!isa<MemIntrinsic>(MI)) 646 return MarkUnsafe(Info); 647 648 // Otherwise, we can transform it. Determine whether this is a memcpy/set 649 // into or out of the aggregate. 650 if (OpNo == 1) 651 Info.isMemCpyDst = true; 652 else { 653 assert(OpNo == 2); 654 Info.isMemCpySrc = true; 655 } 656} 657 658/// isSafeUseOfBitCastedAllocation - Return true if all users of this bitcast 659/// are 660void SROA::isSafeUseOfBitCastedAllocation(BitCastInst *BC, AllocaInst *AI, 661 AllocaInfo &Info) { 662 for (Value::use_iterator UI = BC->use_begin(), E = BC->use_end(); 663 UI != E; ++UI) { 664 if (BitCastInst *BCU = dyn_cast<BitCastInst>(UI)) { 665 isSafeUseOfBitCastedAllocation(BCU, AI, Info); 666 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) { 667 isSafeMemIntrinsicOnAllocation(MI, AI, UI.getOperandNo(), Info); 668 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI)) { 669 if (SI->isVolatile()) 670 return MarkUnsafe(Info); 671 672 // If storing the entire alloca in one chunk through a bitcasted pointer 673 // to integer, we can transform it. This happens (for example) when you 674 // cast a {i32,i32}* to i64* and store through it. This is similar to the 675 // memcpy case and occurs in various "byval" cases and emulated memcpys. 676 if (isa<IntegerType>(SI->getOperand(0)->getType()) && 677 TD->getTypeAllocSize(SI->getOperand(0)->getType()) == 678 TD->getTypeAllocSize(AI->getType()->getElementType())) { 679 Info.isMemCpyDst = true; 680 continue; 681 } 682 return MarkUnsafe(Info); 683 } else if (LoadInst *LI = dyn_cast<LoadInst>(UI)) { 684 if (LI->isVolatile()) 685 return MarkUnsafe(Info); 686 687 // If loading the entire alloca in one chunk through a bitcasted pointer 688 // to integer, we can transform it. This happens (for example) when you 689 // cast a {i32,i32}* to i64* and load through it. This is similar to the 690 // memcpy case and occurs in various "byval" cases and emulated memcpys. 691 if (isa<IntegerType>(LI->getType()) && 692 TD->getTypeAllocSize(LI->getType()) == 693 TD->getTypeAllocSize(AI->getType()->getElementType())) { 694 Info.isMemCpySrc = true; 695 continue; 696 } 697 return MarkUnsafe(Info); 698 } else if (isa<DbgInfoIntrinsic>(UI)) { 699 // If one user is DbgInfoIntrinsic then check if all users are 700 // DbgInfoIntrinsics. 701 if (OnlyUsedByDbgInfoIntrinsics(BC)) { 702 Info.needsCleanup = true; 703 return; 704 } 705 else 706 MarkUnsafe(Info); 707 } 708 else { 709 return MarkUnsafe(Info); 710 } 711 if (Info.isUnsafe) return; 712 } 713} 714 715/// RewriteBitCastUserOfAlloca - BCInst (transitively) bitcasts AI, or indexes 716/// to its first element. Transform users of the cast to use the new values 717/// instead. 718void SROA::RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocaInst *AI, 719 SmallVector<AllocaInst*, 32> &NewElts) { 720 Value::use_iterator UI = BCInst->use_begin(), UE = BCInst->use_end(); 721 while (UI != UE) { 722 Instruction *User = cast<Instruction>(*UI++); 723 if (BitCastInst *BCU = dyn_cast<BitCastInst>(User)) { 724 RewriteBitCastUserOfAlloca(BCU, AI, NewElts); 725 if (BCU->use_empty()) BCU->eraseFromParent(); 726 continue; 727 } 728 729 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 730 // This must be memcpy/memmove/memset of the entire aggregate. 731 // Split into one per element. 732 RewriteMemIntrinUserOfAlloca(MI, BCInst, AI, NewElts); 733 continue; 734 } 735 736 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 737 // If this is a store of the entire alloca from an integer, rewrite it. 738 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 739 continue; 740 } 741 742 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 743 // If this is a load of the entire alloca to an integer, rewrite it. 744 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 745 continue; 746 } 747 748 // Otherwise it must be some other user of a gep of the first pointer. Just 749 // leave these alone. 750 continue; 751 } 752} 753 754/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 755/// Rewrite it to copy or set the elements of the scalarized memory. 756void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst, 757 AllocaInst *AI, 758 SmallVector<AllocaInst*, 32> &NewElts) { 759 760 // If this is a memcpy/memmove, construct the other pointer as the 761 // appropriate type. The "Other" pointer is the pointer that goes to memory 762 // that doesn't have anything to do with the alloca that we are promoting. For 763 // memset, this Value* stays null. 764 Value *OtherPtr = 0; 765 LLVMContext &Context = MI->getContext(); 766 unsigned MemAlignment = MI->getAlignment(); 767 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 768 if (BCInst == MTI->getRawDest()) 769 OtherPtr = MTI->getRawSource(); 770 else { 771 assert(BCInst == MTI->getRawSource()); 772 OtherPtr = MTI->getRawDest(); 773 } 774 } 775 776 // If there is an other pointer, we want to convert it to the same pointer 777 // type as AI has, so we can GEP through it safely. 778 if (OtherPtr) { 779 // It is likely that OtherPtr is a bitcast, if so, remove it. 780 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr)) 781 OtherPtr = BC->getOperand(0); 782 // All zero GEPs are effectively bitcasts. 783 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr)) 784 if (GEP->hasAllZeroIndices()) 785 OtherPtr = GEP->getOperand(0); 786 787 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr)) 788 if (BCE->getOpcode() == Instruction::BitCast) 789 OtherPtr = BCE->getOperand(0); 790 791 // If the pointer is not the right type, insert a bitcast to the right 792 // type. 793 if (OtherPtr->getType() != AI->getType()) 794 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(), 795 MI); 796 } 797 798 // Process each element of the aggregate. 799 Value *TheFn = MI->getOperand(0); 800 const Type *BytePtrTy = MI->getRawDest()->getType(); 801 bool SROADest = MI->getRawDest() == BCInst; 802 803 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 804 805 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 806 // If this is a memcpy/memmove, emit a GEP of the other element address. 807 Value *OtherElt = 0; 808 unsigned OtherEltAlign = MemAlignment; 809 810 if (OtherPtr) { 811 Value *Idx[2] = { Zero, 812 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 813 OtherElt = GetElementPtrInst::Create(OtherPtr, Idx, Idx + 2, 814 OtherPtr->getNameStr()+"."+Twine(i), 815 MI); 816 uint64_t EltOffset; 817 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 818 if (const StructType *ST = 819 dyn_cast<StructType>(OtherPtrTy->getElementType())) { 820 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 821 } else { 822 const Type *EltTy = 823 cast<SequentialType>(OtherPtr->getType())->getElementType(); 824 EltOffset = TD->getTypeAllocSize(EltTy)*i; 825 } 826 827 // The alignment of the other pointer is the guaranteed alignment of the 828 // element, which is affected by both the known alignment of the whole 829 // mem intrinsic and the alignment of the element. If the alignment of 830 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 831 // known alignment is just 4 bytes. 832 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 833 } 834 835 Value *EltPtr = NewElts[i]; 836 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 837 838 // If we got down to a scalar, insert a load or store as appropriate. 839 if (EltTy->isSingleValueType()) { 840 if (isa<MemTransferInst>(MI)) { 841 if (SROADest) { 842 // From Other to Alloca. 843 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 844 new StoreInst(Elt, EltPtr, MI); 845 } else { 846 // From Alloca to Other. 847 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 848 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 849 } 850 continue; 851 } 852 assert(isa<MemSetInst>(MI)); 853 854 // If the stored element is zero (common case), just store a null 855 // constant. 856 Constant *StoreVal; 857 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) { 858 if (CI->isZero()) { 859 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 860 } else { 861 // If EltTy is a vector type, get the element type. 862 const Type *ValTy = EltTy->getScalarType(); 863 864 // Construct an integer with the right value. 865 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 866 APInt OneVal(EltSize, CI->getZExtValue()); 867 APInt TotalVal(OneVal); 868 // Set each byte. 869 for (unsigned i = 0; 8*i < EltSize; ++i) { 870 TotalVal = TotalVal.shl(8); 871 TotalVal |= OneVal; 872 } 873 874 // Convert the integer value to the appropriate type. 875 StoreVal = ConstantInt::get(Context, TotalVal); 876 if (isa<PointerType>(ValTy)) 877 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 878 else if (ValTy->isFloatingPoint()) 879 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 880 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 881 882 // If the requested value was a vector constant, create it. 883 if (EltTy != ValTy) { 884 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 885 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 886 StoreVal = ConstantVector::get(&Elts[0], NumElts); 887 } 888 } 889 new StoreInst(StoreVal, EltPtr, MI); 890 continue; 891 } 892 // Otherwise, if we're storing a byte variable, use a memset call for 893 // this element. 894 } 895 896 // Cast the element pointer to BytePtrTy. 897 if (EltPtr->getType() != BytePtrTy) 898 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI); 899 900 // Cast the other pointer (if we have one) to BytePtrTy. 901 if (OtherElt && OtherElt->getType() != BytePtrTy) 902 OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(), 903 MI); 904 905 unsigned EltSize = TD->getTypeAllocSize(EltTy); 906 907 // Finally, insert the meminst for this element. 908 if (isa<MemTransferInst>(MI)) { 909 Value *Ops[] = { 910 SROADest ? EltPtr : OtherElt, // Dest ptr 911 SROADest ? OtherElt : EltPtr, // Src ptr 912 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 913 // Align 914 ConstantInt::get(Type::getInt32Ty(MI->getContext()), OtherEltAlign) 915 }; 916 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 917 } else { 918 assert(isa<MemSetInst>(MI)); 919 Value *Ops[] = { 920 EltPtr, MI->getOperand(2), // Dest, Value, 921 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 922 Zero // Align 923 }; 924 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 925 } 926 } 927 MI->eraseFromParent(); 928} 929 930/// RewriteStoreUserOfWholeAlloca - We found an store of an integer that 931/// overwrites the entire allocation. Extract out the pieces of the stored 932/// integer and store them individually. 933void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 934 SmallVector<AllocaInst*, 32> &NewElts){ 935 // Extract each element out of the integer according to its structure offset 936 // and store the element value to the individual alloca. 937 Value *SrcVal = SI->getOperand(0); 938 const Type *AllocaEltTy = AI->getType()->getElementType(); 939 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 940 941 // If this isn't a store of an integer to the whole alloca, it may be a store 942 // to the first element. Just ignore the store in this case and normal SROA 943 // will handle it. 944 if (!isa<IntegerType>(SrcVal->getType()) || 945 TD->getTypeAllocSizeInBits(SrcVal->getType()) != AllocaSizeBits) 946 return; 947 // Handle tail padding by extending the operand 948 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 949 SrcVal = new ZExtInst(SrcVal, 950 IntegerType::get(SI->getContext(), AllocaSizeBits), 951 "", SI); 952 953 DEBUG(errs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 954 << '\n'); 955 956 // There are two forms here: AI could be an array or struct. Both cases 957 // have different ways to compute the element offset. 958 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 959 const StructLayout *Layout = TD->getStructLayout(EltSTy); 960 961 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 962 // Get the number of bits to shift SrcVal to get the value. 963 const Type *FieldTy = EltSTy->getElementType(i); 964 uint64_t Shift = Layout->getElementOffsetInBits(i); 965 966 if (TD->isBigEndian()) 967 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 968 969 Value *EltVal = SrcVal; 970 if (Shift) { 971 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 972 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 973 "sroa.store.elt", SI); 974 } 975 976 // Truncate down to an integer of the right size. 977 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 978 979 // Ignore zero sized fields like {}, they obviously contain no data. 980 if (FieldSizeBits == 0) continue; 981 982 if (FieldSizeBits != AllocaSizeBits) 983 EltVal = new TruncInst(EltVal, 984 IntegerType::get(SI->getContext(), FieldSizeBits), 985 "", SI); 986 Value *DestField = NewElts[i]; 987 if (EltVal->getType() == FieldTy) { 988 // Storing to an integer field of this size, just do it. 989 } else if (FieldTy->isFloatingPoint() || isa<VectorType>(FieldTy)) { 990 // Bitcast to the right element type (for fp/vector values). 991 EltVal = new BitCastInst(EltVal, FieldTy, "", SI); 992 } else { 993 // Otherwise, bitcast the dest pointer (for aggregates). 994 DestField = new BitCastInst(DestField, 995 PointerType::getUnqual(EltVal->getType()), 996 "", SI); 997 } 998 new StoreInst(EltVal, DestField, SI); 999 } 1000 1001 } else { 1002 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 1003 const Type *ArrayEltTy = ATy->getElementType(); 1004 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1005 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 1006 1007 uint64_t Shift; 1008 1009 if (TD->isBigEndian()) 1010 Shift = AllocaSizeBits-ElementOffset; 1011 else 1012 Shift = 0; 1013 1014 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1015 // Ignore zero sized fields like {}, they obviously contain no data. 1016 if (ElementSizeBits == 0) continue; 1017 1018 Value *EltVal = SrcVal; 1019 if (Shift) { 1020 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 1021 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 1022 "sroa.store.elt", SI); 1023 } 1024 1025 // Truncate down to an integer of the right size. 1026 if (ElementSizeBits != AllocaSizeBits) 1027 EltVal = new TruncInst(EltVal, 1028 IntegerType::get(SI->getContext(), 1029 ElementSizeBits),"",SI); 1030 Value *DestField = NewElts[i]; 1031 if (EltVal->getType() == ArrayEltTy) { 1032 // Storing to an integer field of this size, just do it. 1033 } else if (ArrayEltTy->isFloatingPoint() || isa<VectorType>(ArrayEltTy)) { 1034 // Bitcast to the right element type (for fp/vector values). 1035 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI); 1036 } else { 1037 // Otherwise, bitcast the dest pointer (for aggregates). 1038 DestField = new BitCastInst(DestField, 1039 PointerType::getUnqual(EltVal->getType()), 1040 "", SI); 1041 } 1042 new StoreInst(EltVal, DestField, SI); 1043 1044 if (TD->isBigEndian()) 1045 Shift -= ElementOffset; 1046 else 1047 Shift += ElementOffset; 1048 } 1049 } 1050 1051 SI->eraseFromParent(); 1052} 1053 1054/// RewriteLoadUserOfWholeAlloca - We found an load of the entire allocation to 1055/// an integer. Load the individual pieces to form the aggregate value. 1056void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 1057 SmallVector<AllocaInst*, 32> &NewElts) { 1058 // Extract each element out of the NewElts according to its structure offset 1059 // and form the result value. 1060 const Type *AllocaEltTy = AI->getType()->getElementType(); 1061 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1062 1063 // If this isn't a load of the whole alloca to an integer, it may be a load 1064 // of the first element. Just ignore the load in this case and normal SROA 1065 // will handle it. 1066 if (!isa<IntegerType>(LI->getType()) || 1067 TD->getTypeAllocSizeInBits(LI->getType()) != AllocaSizeBits) 1068 return; 1069 1070 DEBUG(errs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 1071 << '\n'); 1072 1073 // There are two forms here: AI could be an array or struct. Both cases 1074 // have different ways to compute the element offset. 1075 const StructLayout *Layout = 0; 1076 uint64_t ArrayEltBitOffset = 0; 1077 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1078 Layout = TD->getStructLayout(EltSTy); 1079 } else { 1080 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 1081 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1082 } 1083 1084 Value *ResultVal = 1085 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 1086 1087 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1088 // Load the value from the alloca. If the NewElt is an aggregate, cast 1089 // the pointer to an integer of the same size before doing the load. 1090 Value *SrcField = NewElts[i]; 1091 const Type *FieldTy = 1092 cast<PointerType>(SrcField->getType())->getElementType(); 1093 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1094 1095 // Ignore zero sized fields like {}, they obviously contain no data. 1096 if (FieldSizeBits == 0) continue; 1097 1098 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 1099 FieldSizeBits); 1100 if (!isa<IntegerType>(FieldTy) && !FieldTy->isFloatingPoint() && 1101 !isa<VectorType>(FieldTy)) 1102 SrcField = new BitCastInst(SrcField, 1103 PointerType::getUnqual(FieldIntTy), 1104 "", LI); 1105 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 1106 1107 // If SrcField is a fp or vector of the right size but that isn't an 1108 // integer type, bitcast to an integer so we can shift it. 1109 if (SrcField->getType() != FieldIntTy) 1110 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 1111 1112 // Zero extend the field to be the same size as the final alloca so that 1113 // we can shift and insert it. 1114 if (SrcField->getType() != ResultVal->getType()) 1115 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 1116 1117 // Determine the number of bits to shift SrcField. 1118 uint64_t Shift; 1119 if (Layout) // Struct case. 1120 Shift = Layout->getElementOffsetInBits(i); 1121 else // Array case. 1122 Shift = i*ArrayEltBitOffset; 1123 1124 if (TD->isBigEndian()) 1125 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 1126 1127 if (Shift) { 1128 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 1129 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 1130 } 1131 1132 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 1133 } 1134 1135 // Handle tail padding by truncating the result 1136 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 1137 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 1138 1139 LI->replaceAllUsesWith(ResultVal); 1140 LI->eraseFromParent(); 1141} 1142 1143 1144/// HasPadding - Return true if the specified type has any structure or 1145/// alignment padding, false otherwise. 1146static bool HasPadding(const Type *Ty, const TargetData &TD) { 1147 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 1148 const StructLayout *SL = TD.getStructLayout(STy); 1149 unsigned PrevFieldBitOffset = 0; 1150 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1151 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 1152 1153 // Padding in sub-elements? 1154 if (HasPadding(STy->getElementType(i), TD)) 1155 return true; 1156 1157 // Check to see if there is any padding between this element and the 1158 // previous one. 1159 if (i) { 1160 unsigned PrevFieldEnd = 1161 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 1162 if (PrevFieldEnd < FieldBitOffset) 1163 return true; 1164 } 1165 1166 PrevFieldBitOffset = FieldBitOffset; 1167 } 1168 1169 // Check for tail padding. 1170 if (unsigned EltCount = STy->getNumElements()) { 1171 unsigned PrevFieldEnd = PrevFieldBitOffset + 1172 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 1173 if (PrevFieldEnd < SL->getSizeInBits()) 1174 return true; 1175 } 1176 1177 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1178 return HasPadding(ATy->getElementType(), TD); 1179 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) { 1180 return HasPadding(VTy->getElementType(), TD); 1181 } 1182 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 1183} 1184 1185/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 1186/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 1187/// or 1 if safe after canonicalization has been performed. 1188/// 1189int SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 1190 // Loop over the use list of the alloca. We can only transform it if all of 1191 // the users are safe to transform. 1192 AllocaInfo Info; 1193 1194 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); 1195 I != E; ++I) { 1196 isSafeUseOfAllocation(cast<Instruction>(*I), AI, Info); 1197 if (Info.isUnsafe) { 1198 DEBUG(errs() << "Cannot transform: " << *AI << "\n due to user: " 1199 << **I << '\n'); 1200 return 0; 1201 } 1202 } 1203 1204 // Okay, we know all the users are promotable. If the aggregate is a memcpy 1205 // source and destination, we have to be careful. In particular, the memcpy 1206 // could be moving around elements that live in structure padding of the LLVM 1207 // types, but may actually be used. In these cases, we refuse to promote the 1208 // struct. 1209 if (Info.isMemCpySrc && Info.isMemCpyDst && 1210 HasPadding(AI->getType()->getElementType(), *TD)) 1211 return 0; 1212 1213 // If we require cleanup, return 1, otherwise return 3. 1214 return Info.needsCleanup ? 1 : 3; 1215} 1216 1217/// CleanupGEP - GEP is used by an Alloca, which can be prompted after the GEP 1218/// is canonicalized here. 1219void SROA::CleanupGEP(GetElementPtrInst *GEPI) { 1220 gep_type_iterator I = gep_type_begin(GEPI); 1221 ++I; 1222 1223 const ArrayType *AT = dyn_cast<ArrayType>(*I); 1224 if (!AT) 1225 return; 1226 1227 uint64_t NumElements = AT->getNumElements(); 1228 1229 if (isa<ConstantInt>(I.getOperand())) 1230 return; 1231 1232 if (NumElements == 1) { 1233 GEPI->setOperand(2, 1234 Constant::getNullValue(Type::getInt32Ty(GEPI->getContext()))); 1235 return; 1236 } 1237 1238 assert(NumElements == 2 && "Unhandled case!"); 1239 // All users of the GEP must be loads. At each use of the GEP, insert 1240 // two loads of the appropriate indexed GEP and select between them. 1241 Value *IsOne = new ICmpInst(GEPI, ICmpInst::ICMP_NE, I.getOperand(), 1242 Constant::getNullValue(I.getOperand()->getType()), 1243 "isone"); 1244 // Insert the new GEP instructions, which are properly indexed. 1245 SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end()); 1246 Indices[1] = Constant::getNullValue(Type::getInt32Ty(GEPI->getContext())); 1247 Value *ZeroIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 1248 Indices.begin(), 1249 Indices.end(), 1250 GEPI->getName()+".0", GEPI); 1251 Indices[1] = ConstantInt::get(Type::getInt32Ty(GEPI->getContext()), 1); 1252 Value *OneIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 1253 Indices.begin(), 1254 Indices.end(), 1255 GEPI->getName()+".1", GEPI); 1256 // Replace all loads of the variable index GEP with loads from both 1257 // indexes and a select. 1258 while (!GEPI->use_empty()) { 1259 LoadInst *LI = cast<LoadInst>(GEPI->use_back()); 1260 Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI); 1261 Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI); 1262 Value *R = SelectInst::Create(IsOne, One, Zero, LI->getName(), LI); 1263 LI->replaceAllUsesWith(R); 1264 LI->eraseFromParent(); 1265 } 1266 GEPI->eraseFromParent(); 1267} 1268 1269 1270/// CleanupAllocaUsers - If SROA reported that it can promote the specified 1271/// allocation, but only if cleaned up, perform the cleanups required. 1272void SROA::CleanupAllocaUsers(AllocaInst *AI) { 1273 // At this point, we know that the end result will be SROA'd and promoted, so 1274 // we can insert ugly code if required so long as sroa+mem2reg will clean it 1275 // up. 1276 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 1277 UI != E; ) { 1278 User *U = *UI++; 1279 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) 1280 CleanupGEP(GEPI); 1281 else { 1282 Instruction *I = cast<Instruction>(U); 1283 SmallVector<DbgInfoIntrinsic *, 2> DbgInUses; 1284 if (!isa<StoreInst>(I) && OnlyUsedByDbgInfoIntrinsics(I, &DbgInUses)) { 1285 // Safe to remove debug info uses. 1286 while (!DbgInUses.empty()) { 1287 DbgInfoIntrinsic *DI = DbgInUses.back(); DbgInUses.pop_back(); 1288 DI->eraseFromParent(); 1289 } 1290 I->eraseFromParent(); 1291 } 1292 } 1293 } 1294} 1295 1296/// MergeInType - Add the 'In' type to the accumulated type (Accum) so far at 1297/// the offset specified by Offset (which is specified in bytes). 1298/// 1299/// There are two cases we handle here: 1300/// 1) A union of vector types of the same size and potentially its elements. 1301/// Here we turn element accesses into insert/extract element operations. 1302/// This promotes a <4 x float> with a store of float to the third element 1303/// into a <4 x float> that uses insert element. 1304/// 2) A fully general blob of memory, which we turn into some (potentially 1305/// large) integer type with extract and insert operations where the loads 1306/// and stores would mutate the memory. 1307static void MergeInType(const Type *In, uint64_t Offset, const Type *&VecTy, 1308 unsigned AllocaSize, const TargetData &TD, 1309 LLVMContext &Context) { 1310 // If this could be contributing to a vector, analyze it. 1311 if (VecTy != Type::getVoidTy(Context)) { // either null or a vector type. 1312 1313 // If the In type is a vector that is the same size as the alloca, see if it 1314 // matches the existing VecTy. 1315 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 1316 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 1317 // If we're storing/loading a vector of the right size, allow it as a 1318 // vector. If this the first vector we see, remember the type so that 1319 // we know the element size. 1320 if (VecTy == 0) 1321 VecTy = VInTy; 1322 return; 1323 } 1324 } else if (In->isFloatTy() || In->isDoubleTy() || 1325 (isa<IntegerType>(In) && In->getPrimitiveSizeInBits() >= 8 && 1326 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 1327 // If we're accessing something that could be an element of a vector, see 1328 // if the implied vector agrees with what we already have and if Offset is 1329 // compatible with it. 1330 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 1331 if (Offset % EltSize == 0 && 1332 AllocaSize % EltSize == 0 && 1333 (VecTy == 0 || 1334 cast<VectorType>(VecTy)->getElementType() 1335 ->getPrimitiveSizeInBits()/8 == EltSize)) { 1336 if (VecTy == 0) 1337 VecTy = VectorType::get(In, AllocaSize/EltSize); 1338 return; 1339 } 1340 } 1341 } 1342 1343 // Otherwise, we have a case that we can't handle with an optimized vector 1344 // form. We can still turn this into a large integer. 1345 VecTy = Type::getVoidTy(Context); 1346} 1347 1348/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 1349/// its accesses to use a to single vector type, return true, and set VecTy to 1350/// the new type. If we could convert the alloca into a single promotable 1351/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 1352/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 1353/// is the current offset from the base of the alloca being analyzed. 1354/// 1355/// If we see at least one access to the value that is as a vector type, set the 1356/// SawVec flag. 1357/// 1358bool SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy, 1359 bool &SawVec, uint64_t Offset, 1360 unsigned AllocaSize) { 1361 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1362 Instruction *User = cast<Instruction>(*UI); 1363 1364 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1365 // Don't break volatile loads. 1366 if (LI->isVolatile()) 1367 return false; 1368 MergeInType(LI->getType(), Offset, VecTy, 1369 AllocaSize, *TD, V->getContext()); 1370 SawVec |= isa<VectorType>(LI->getType()); 1371 continue; 1372 } 1373 1374 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1375 // Storing the pointer, not into the value? 1376 if (SI->getOperand(0) == V || SI->isVolatile()) return 0; 1377 MergeInType(SI->getOperand(0)->getType(), Offset, 1378 VecTy, AllocaSize, *TD, V->getContext()); 1379 SawVec |= isa<VectorType>(SI->getOperand(0)->getType()); 1380 continue; 1381 } 1382 1383 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 1384 if (!CanConvertToScalar(BCI, IsNotTrivial, VecTy, SawVec, Offset, 1385 AllocaSize)) 1386 return false; 1387 IsNotTrivial = true; 1388 continue; 1389 } 1390 1391 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1392 // If this is a GEP with a variable indices, we can't handle it. 1393 if (!GEP->hasAllConstantIndices()) 1394 return false; 1395 1396 // Compute the offset that this GEP adds to the pointer. 1397 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1398 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(), 1399 &Indices[0], Indices.size()); 1400 // See if all uses can be converted. 1401 if (!CanConvertToScalar(GEP, IsNotTrivial, VecTy, SawVec,Offset+GEPOffset, 1402 AllocaSize)) 1403 return false; 1404 IsNotTrivial = true; 1405 continue; 1406 } 1407 1408 // If this is a constant sized memset of a constant value (e.g. 0) we can 1409 // handle it. 1410 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 1411 // Store of constant value and constant size. 1412 if (isa<ConstantInt>(MSI->getValue()) && 1413 isa<ConstantInt>(MSI->getLength())) { 1414 IsNotTrivial = true; 1415 continue; 1416 } 1417 } 1418 1419 // If this is a memcpy or memmove into or out of the whole allocation, we 1420 // can handle it like a load or store of the scalar type. 1421 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 1422 if (ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength())) 1423 if (Len->getZExtValue() == AllocaSize && Offset == 0) { 1424 IsNotTrivial = true; 1425 continue; 1426 } 1427 } 1428 1429 // Ignore dbg intrinsic. 1430 if (isa<DbgInfoIntrinsic>(User)) 1431 continue; 1432 1433 // Otherwise, we cannot handle this! 1434 return false; 1435 } 1436 1437 return true; 1438} 1439 1440 1441/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 1442/// directly. This happens when we are converting an "integer union" to a 1443/// single integer scalar, or when we are converting a "vector union" to a 1444/// vector with insert/extractelement instructions. 1445/// 1446/// Offset is an offset from the original alloca, in bits that need to be 1447/// shifted to the right. By the end of this, there should be no uses of Ptr. 1448void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset) { 1449 while (!Ptr->use_empty()) { 1450 Instruction *User = cast<Instruction>(Ptr->use_back()); 1451 1452 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 1453 ConvertUsesToScalar(CI, NewAI, Offset); 1454 CI->eraseFromParent(); 1455 continue; 1456 } 1457 1458 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1459 // Compute the offset that this GEP adds to the pointer. 1460 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1461 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(), 1462 &Indices[0], Indices.size()); 1463 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 1464 GEP->eraseFromParent(); 1465 continue; 1466 } 1467 1468 IRBuilder<> Builder(User->getParent(), User); 1469 1470 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1471 // The load is a bit extract from NewAI shifted right by Offset bits. 1472 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 1473 Value *NewLoadVal 1474 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 1475 LI->replaceAllUsesWith(NewLoadVal); 1476 LI->eraseFromParent(); 1477 continue; 1478 } 1479 1480 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1481 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 1482 // FIXME: Remove once builder has Twine API. 1483 Value *Old = Builder.CreateLoad(NewAI, 1484 (NewAI->getName()+".in").str().c_str()); 1485 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 1486 Builder); 1487 Builder.CreateStore(New, NewAI); 1488 SI->eraseFromParent(); 1489 continue; 1490 } 1491 1492 // If this is a constant sized memset of a constant value (e.g. 0) we can 1493 // transform it into a store of the expanded constant value. 1494 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 1495 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 1496 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 1497 if (NumBytes != 0) { 1498 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 1499 1500 // Compute the value replicated the right number of times. 1501 APInt APVal(NumBytes*8, Val); 1502 1503 // Splat the value if non-zero. 1504 if (Val) 1505 for (unsigned i = 1; i != NumBytes; ++i) 1506 APVal |= APVal << 8; 1507 1508 // FIXME: Remove once builder has Twine API. 1509 Value *Old = Builder.CreateLoad(NewAI, 1510 (NewAI->getName()+".in").str().c_str()); 1511 Value *New = ConvertScalar_InsertValue( 1512 ConstantInt::get(User->getContext(), APVal), 1513 Old, Offset, Builder); 1514 Builder.CreateStore(New, NewAI); 1515 } 1516 MSI->eraseFromParent(); 1517 continue; 1518 } 1519 1520 // If this is a memcpy or memmove into or out of the whole allocation, we 1521 // can handle it like a load or store of the scalar type. 1522 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 1523 assert(Offset == 0 && "must be store to start of alloca"); 1524 1525 // If the source and destination are both to the same alloca, then this is 1526 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 1527 // as appropriate. 1528 AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject()); 1529 1530 if (MTI->getSource()->getUnderlyingObject() != OrigAI) { 1531 // Dest must be OrigAI, change this to be a load from the original 1532 // pointer (bitcasted), then a store to our new alloca. 1533 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 1534 Value *SrcPtr = MTI->getSource(); 1535 SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType()); 1536 1537 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 1538 SrcVal->setAlignment(MTI->getAlignment()); 1539 Builder.CreateStore(SrcVal, NewAI); 1540 } else if (MTI->getDest()->getUnderlyingObject() != OrigAI) { 1541 // Src must be OrigAI, change this to be a load from NewAI then a store 1542 // through the original dest pointer (bitcasted). 1543 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 1544 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 1545 1546 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType()); 1547 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 1548 NewStore->setAlignment(MTI->getAlignment()); 1549 } else { 1550 // Noop transfer. Src == Dst 1551 } 1552 1553 1554 MTI->eraseFromParent(); 1555 continue; 1556 } 1557 1558 // If user is a dbg info intrinsic then it is safe to remove it. 1559 if (isa<DbgInfoIntrinsic>(User)) { 1560 User->eraseFromParent(); 1561 continue; 1562 } 1563 1564 llvm_unreachable("Unsupported operation!"); 1565 } 1566} 1567 1568/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 1569/// or vector value FromVal, extracting the bits from the offset specified by 1570/// Offset. This returns the value, which is of type ToType. 1571/// 1572/// This happens when we are converting an "integer union" to a single 1573/// integer scalar, or when we are converting a "vector union" to a vector with 1574/// insert/extractelement instructions. 1575/// 1576/// Offset is an offset from the original alloca, in bits that need to be 1577/// shifted to the right. 1578Value *SROA::ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 1579 uint64_t Offset, IRBuilder<> &Builder) { 1580 // If the load is of the whole new alloca, no conversion is needed. 1581 if (FromVal->getType() == ToType && Offset == 0) 1582 return FromVal; 1583 1584 // If the result alloca is a vector type, this is either an element 1585 // access or a bitcast to another vector type of the same size. 1586 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) { 1587 if (isa<VectorType>(ToType)) 1588 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 1589 1590 // Otherwise it must be an element access. 1591 unsigned Elt = 0; 1592 if (Offset) { 1593 unsigned EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 1594 Elt = Offset/EltSize; 1595 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 1596 } 1597 // Return the element extracted out of it. 1598 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get( 1599 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp"); 1600 if (V->getType() != ToType) 1601 V = Builder.CreateBitCast(V, ToType, "tmp"); 1602 return V; 1603 } 1604 1605 // If ToType is a first class aggregate, extract out each of the pieces and 1606 // use insertvalue's to form the FCA. 1607 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 1608 const StructLayout &Layout = *TD->getStructLayout(ST); 1609 Value *Res = UndefValue::get(ST); 1610 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1611 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 1612 Offset+Layout.getElementOffsetInBits(i), 1613 Builder); 1614 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1615 } 1616 return Res; 1617 } 1618 1619 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 1620 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 1621 Value *Res = UndefValue::get(AT); 1622 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1623 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 1624 Offset+i*EltSize, Builder); 1625 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1626 } 1627 return Res; 1628 } 1629 1630 // Otherwise, this must be a union that was converted to an integer value. 1631 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 1632 1633 // If this is a big-endian system and the load is narrower than the 1634 // full alloca type, we need to do a shift to get the right bits. 1635 int ShAmt = 0; 1636 if (TD->isBigEndian()) { 1637 // On big-endian machines, the lowest bit is stored at the bit offset 1638 // from the pointer given by getTypeStoreSizeInBits. This matters for 1639 // integers with a bitwidth that is not a multiple of 8. 1640 ShAmt = TD->getTypeStoreSizeInBits(NTy) - 1641 TD->getTypeStoreSizeInBits(ToType) - Offset; 1642 } else { 1643 ShAmt = Offset; 1644 } 1645 1646 // Note: we support negative bitwidths (with shl) which are not defined. 1647 // We do this to support (f.e.) loads off the end of a structure where 1648 // only some bits are used. 1649 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 1650 FromVal = Builder.CreateLShr(FromVal, 1651 ConstantInt::get(FromVal->getType(), 1652 ShAmt), "tmp"); 1653 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 1654 FromVal = Builder.CreateShl(FromVal, 1655 ConstantInt::get(FromVal->getType(), 1656 -ShAmt), "tmp"); 1657 1658 // Finally, unconditionally truncate the integer to the right width. 1659 unsigned LIBitWidth = TD->getTypeSizeInBits(ToType); 1660 if (LIBitWidth < NTy->getBitWidth()) 1661 FromVal = 1662 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 1663 LIBitWidth), "tmp"); 1664 else if (LIBitWidth > NTy->getBitWidth()) 1665 FromVal = 1666 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 1667 LIBitWidth), "tmp"); 1668 1669 // If the result is an integer, this is a trunc or bitcast. 1670 if (isa<IntegerType>(ToType)) { 1671 // Should be done. 1672 } else if (ToType->isFloatingPoint() || isa<VectorType>(ToType)) { 1673 // Just do a bitcast, we know the sizes match up. 1674 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 1675 } else { 1676 // Otherwise must be a pointer. 1677 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 1678 } 1679 assert(FromVal->getType() == ToType && "Didn't convert right?"); 1680 return FromVal; 1681} 1682 1683 1684/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 1685/// or vector value "Old" at the offset specified by Offset. 1686/// 1687/// This happens when we are converting an "integer union" to a 1688/// single integer scalar, or when we are converting a "vector union" to a 1689/// vector with insert/extractelement instructions. 1690/// 1691/// Offset is an offset from the original alloca, in bits that need to be 1692/// shifted to the right. 1693Value *SROA::ConvertScalar_InsertValue(Value *SV, Value *Old, 1694 uint64_t Offset, IRBuilder<> &Builder) { 1695 1696 // Convert the stored type to the actual type, shift it left to insert 1697 // then 'or' into place. 1698 const Type *AllocaType = Old->getType(); 1699 LLVMContext &Context = Old->getContext(); 1700 1701 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 1702 uint64_t VecSize = TD->getTypeAllocSizeInBits(VTy); 1703 uint64_t ValSize = TD->getTypeAllocSizeInBits(SV->getType()); 1704 1705 // Changing the whole vector with memset or with an access of a different 1706 // vector type? 1707 if (ValSize == VecSize) 1708 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 1709 1710 uint64_t EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 1711 1712 // Must be an element insertion. 1713 unsigned Elt = Offset/EltSize; 1714 1715 if (SV->getType() != VTy->getElementType()) 1716 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 1717 1718 SV = Builder.CreateInsertElement(Old, SV, 1719 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt), 1720 "tmp"); 1721 return SV; 1722 } 1723 1724 // If SV is a first-class aggregate value, insert each value recursively. 1725 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 1726 const StructLayout &Layout = *TD->getStructLayout(ST); 1727 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1728 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1729 Old = ConvertScalar_InsertValue(Elt, Old, 1730 Offset+Layout.getElementOffsetInBits(i), 1731 Builder); 1732 } 1733 return Old; 1734 } 1735 1736 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 1737 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 1738 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1739 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1740 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 1741 } 1742 return Old; 1743 } 1744 1745 // If SV is a float, convert it to the appropriate integer type. 1746 // If it is a pointer, do the same. 1747 unsigned SrcWidth = TD->getTypeSizeInBits(SV->getType()); 1748 unsigned DestWidth = TD->getTypeSizeInBits(AllocaType); 1749 unsigned SrcStoreWidth = TD->getTypeStoreSizeInBits(SV->getType()); 1750 unsigned DestStoreWidth = TD->getTypeStoreSizeInBits(AllocaType); 1751 if (SV->getType()->isFloatingPoint() || isa<VectorType>(SV->getType())) 1752 SV = Builder.CreateBitCast(SV, 1753 IntegerType::get(SV->getContext(),SrcWidth), "tmp"); 1754 else if (isa<PointerType>(SV->getType())) 1755 SV = Builder.CreatePtrToInt(SV, TD->getIntPtrType(SV->getContext()), "tmp"); 1756 1757 // Zero extend or truncate the value if needed. 1758 if (SV->getType() != AllocaType) { 1759 if (SV->getType()->getPrimitiveSizeInBits() < 1760 AllocaType->getPrimitiveSizeInBits()) 1761 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 1762 else { 1763 // Truncation may be needed if storing more than the alloca can hold 1764 // (undefined behavior). 1765 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 1766 SrcWidth = DestWidth; 1767 SrcStoreWidth = DestStoreWidth; 1768 } 1769 } 1770 1771 // If this is a big-endian system and the store is narrower than the 1772 // full alloca type, we need to do a shift to get the right bits. 1773 int ShAmt = 0; 1774 if (TD->isBigEndian()) { 1775 // On big-endian machines, the lowest bit is stored at the bit offset 1776 // from the pointer given by getTypeStoreSizeInBits. This matters for 1777 // integers with a bitwidth that is not a multiple of 8. 1778 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 1779 } else { 1780 ShAmt = Offset; 1781 } 1782 1783 // Note: we support negative bitwidths (with shr) which are not defined. 1784 // We do this to support (f.e.) stores off the end of a structure where 1785 // only some bits in the structure are set. 1786 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 1787 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 1788 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 1789 ShAmt), "tmp"); 1790 Mask <<= ShAmt; 1791 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1792 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 1793 -ShAmt), "tmp"); 1794 Mask = Mask.lshr(-ShAmt); 1795 } 1796 1797 // Mask out the bits we are about to insert from the old value, and or 1798 // in the new bits. 1799 if (SrcWidth != DestWidth) { 1800 assert(DestWidth > SrcWidth); 1801 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 1802 SV = Builder.CreateOr(Old, SV, "ins"); 1803 } 1804 return SV; 1805} 1806 1807 1808 1809/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1810/// some part of a constant global variable. This intentionally only accepts 1811/// constant expressions because we don't can't rewrite arbitrary instructions. 1812static bool PointsToConstantGlobal(Value *V) { 1813 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1814 return GV->isConstant(); 1815 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1816 if (CE->getOpcode() == Instruction::BitCast || 1817 CE->getOpcode() == Instruction::GetElementPtr) 1818 return PointsToConstantGlobal(CE->getOperand(0)); 1819 return false; 1820} 1821 1822/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1823/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1824/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1825/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1826/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1827/// the alloca, and if the source pointer is a pointer to a constant global, we 1828/// can optimize this. 1829static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy, 1830 bool isOffset) { 1831 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1832 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) 1833 // Ignore non-volatile loads, they are always ok. 1834 if (!LI->isVolatile()) 1835 continue; 1836 1837 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) { 1838 // If uses of the bitcast are ok, we are ok. 1839 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 1840 return false; 1841 continue; 1842 } 1843 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) { 1844 // If the GEP has all zero indices, it doesn't offset the pointer. If it 1845 // doesn't, it does. 1846 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 1847 isOffset || !GEP->hasAllZeroIndices())) 1848 return false; 1849 continue; 1850 } 1851 1852 // If this is isn't our memcpy/memmove, reject it as something we can't 1853 // handle. 1854 if (!isa<MemTransferInst>(*UI)) 1855 return false; 1856 1857 // If we already have seen a copy, reject the second one. 1858 if (TheCopy) return false; 1859 1860 // If the pointer has been offset from the start of the alloca, we can't 1861 // safely handle this. 1862 if (isOffset) return false; 1863 1864 // If the memintrinsic isn't using the alloca as the dest, reject it. 1865 if (UI.getOperandNo() != 1) return false; 1866 1867 MemIntrinsic *MI = cast<MemIntrinsic>(*UI); 1868 1869 // If the source of the memcpy/move is not a constant global, reject it. 1870 if (!PointsToConstantGlobal(MI->getOperand(2))) 1871 return false; 1872 1873 // Otherwise, the transform is safe. Remember the copy instruction. 1874 TheCopy = MI; 1875 } 1876 return true; 1877} 1878 1879/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 1880/// modified by a copy from a constant global. If we can prove this, we can 1881/// replace any uses of the alloca with uses of the global directly. 1882Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) { 1883 Instruction *TheCopy = 0; 1884 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 1885 return TheCopy; 1886 return 0; 1887} 1888