ScalarReplAggregates.cpp revision 39fdd6947cb800a0db7f7012fe843c98b414602f
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/// isSafeElementUse - Check to see if this use is an allowed use for a 454/// getelementptr instruction of an array aggregate allocation. isFirstElt 455/// indicates whether Ptr is known to the start of the aggregate. 456void SROA::isSafeElementUse(Value *Ptr, bool isFirstElt, AllocaInst *AI, 457 AllocaInfo &Info) { 458 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 459 I != E; ++I) { 460 Instruction *User = cast<Instruction>(*I); 461 switch (User->getOpcode()) { 462 case Instruction::Load: break; 463 case Instruction::Store: 464 // Store is ok if storing INTO the pointer, not storing the pointer 465 if (User->getOperand(0) == Ptr) return MarkUnsafe(Info); 466 break; 467 case Instruction::GetElementPtr: { 468 GetElementPtrInst *GEP = cast<GetElementPtrInst>(User); 469 bool AreAllZeroIndices = isFirstElt; 470 if (GEP->getNumOperands() > 1 && 471 (!isa<ConstantInt>(GEP->getOperand(1)) || 472 !cast<ConstantInt>(GEP->getOperand(1))->isZero())) 473 // Using pointer arithmetic to navigate the array. 474 return MarkUnsafe(Info); 475 476 // Verify that any array subscripts are in range. 477 for (gep_type_iterator GEPIt = gep_type_begin(GEP), 478 E = gep_type_end(GEP); GEPIt != E; ++GEPIt) { 479 // Ignore struct elements, no extra checking needed for these. 480 if (isa<StructType>(*GEPIt)) 481 continue; 482 483 // This GEP indexes an array. Verify that this is an in-range 484 // constant integer. Specifically, consider A[0][i]. We cannot know that 485 // the user isn't doing invalid things like allowing i to index an 486 // out-of-range subscript that accesses A[1]. Because of this, we have 487 // to reject SROA of any accesses into structs where any of the 488 // components are variables. 489 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand()); 490 if (!IdxVal) return MarkUnsafe(Info); 491 492 // Are all indices still zero? 493 AreAllZeroIndices &= IdxVal->isZero(); 494 495 if (const ArrayType *AT = dyn_cast<ArrayType>(*GEPIt)) { 496 if (IdxVal->getZExtValue() >= AT->getNumElements()) 497 return MarkUnsafe(Info); 498 } else if (const VectorType *VT = dyn_cast<VectorType>(*GEPIt)) { 499 if (IdxVal->getZExtValue() >= VT->getNumElements()) 500 return MarkUnsafe(Info); 501 } 502 } 503 504 isSafeElementUse(GEP, AreAllZeroIndices, AI, Info); 505 if (Info.isUnsafe) return; 506 break; 507 } 508 case Instruction::BitCast: 509 if (isFirstElt) { 510 isSafeUseOfBitCastedAllocation(cast<BitCastInst>(User), AI, Info); 511 if (Info.isUnsafe) return; 512 break; 513 } 514 DEBUG(errs() << " Transformation preventing inst: " << *User << '\n'); 515 return MarkUnsafe(Info); 516 case Instruction::Call: 517 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 518 if (isFirstElt) { 519 isSafeMemIntrinsicOnAllocation(MI, AI, I.getOperandNo(), Info); 520 if (Info.isUnsafe) return; 521 break; 522 } 523 } 524 DEBUG(errs() << " Transformation preventing inst: " << *User << '\n'); 525 return MarkUnsafe(Info); 526 default: 527 DEBUG(errs() << " Transformation preventing inst: " << *User << '\n'); 528 return MarkUnsafe(Info); 529 } 530 } 531 return; // All users look ok :) 532} 533 534/// AllUsersAreLoads - Return true if all users of this value are loads. 535static bool AllUsersAreLoads(Value *Ptr) { 536 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 537 I != E; ++I) 538 if (cast<Instruction>(*I)->getOpcode() != Instruction::Load) 539 return false; 540 return true; 541} 542 543/// isSafeUseOfAllocation - Check to see if this user is an allowed use for an 544/// aggregate allocation. 545void SROA::isSafeUseOfAllocation(Instruction *User, AllocaInst *AI, 546 AllocaInfo &Info) { 547 if (BitCastInst *C = dyn_cast<BitCastInst>(User)) 548 return isSafeUseOfBitCastedAllocation(C, AI, Info); 549 550 if (LoadInst *LI = dyn_cast<LoadInst>(User)) 551 if (!LI->isVolatile()) 552 return;// Loads (returning a first class aggregrate) are always rewritable 553 554 if (StoreInst *SI = dyn_cast<StoreInst>(User)) 555 if (!SI->isVolatile() && SI->getOperand(0) != AI) 556 return;// Store is ok if storing INTO the pointer, not storing the pointer 557 558 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User); 559 if (GEPI == 0) 560 return MarkUnsafe(Info); 561 562 gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI); 563 564 // The GEP is not safe to transform if not of the form "GEP <ptr>, 0, <cst>". 565 if (I == E || 566 I.getOperand() != Constant::getNullValue(I.getOperand()->getType())) { 567 return MarkUnsafe(Info); 568 } 569 570 ++I; 571 if (I == E) return MarkUnsafe(Info); // ran out of GEP indices?? 572 573 bool IsAllZeroIndices = true; 574 575 // If the first index is a non-constant index into an array, see if we can 576 // handle it as a special case. 577 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 578 if (!isa<ConstantInt>(I.getOperand())) { 579 IsAllZeroIndices = 0; 580 uint64_t NumElements = AT->getNumElements(); 581 582 // If this is an array index and the index is not constant, we cannot 583 // promote... that is unless the array has exactly one or two elements in 584 // it, in which case we CAN promote it, but we have to canonicalize this 585 // out if this is the only problem. 586 if ((NumElements == 1 || NumElements == 2) && 587 AllUsersAreLoads(GEPI)) { 588 Info.needsCleanup = true; 589 return; // Canonicalization required! 590 } 591 return MarkUnsafe(Info); 592 } 593 } 594 595 // Walk through the GEP type indices, checking the types that this indexes 596 // into. 597 for (; I != E; ++I) { 598 // Ignore struct elements, no extra checking needed for these. 599 if (isa<StructType>(*I)) 600 continue; 601 602 ConstantInt *IdxVal = dyn_cast<ConstantInt>(I.getOperand()); 603 if (!IdxVal) return MarkUnsafe(Info); 604 605 // Are all indices still zero? 606 IsAllZeroIndices &= IdxVal->isZero(); 607 608 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 609 // This GEP indexes an array. Verify that this is an in-range constant 610 // integer. Specifically, consider A[0][i]. We cannot know that the user 611 // isn't doing invalid things like allowing i to index an out-of-range 612 // subscript that accesses A[1]. Because of this, we have to reject SROA 613 // of any accesses into structs where any of the components are variables. 614 if (IdxVal->getZExtValue() >= AT->getNumElements()) 615 return MarkUnsafe(Info); 616 } else if (const VectorType *VT = dyn_cast<VectorType>(*I)) { 617 if (IdxVal->getZExtValue() >= VT->getNumElements()) 618 return MarkUnsafe(Info); 619 } 620 } 621 622 // If there are any non-simple uses of this getelementptr, make sure to reject 623 // them. 624 return isSafeElementUse(GEPI, IsAllZeroIndices, AI, Info); 625} 626 627/// isSafeMemIntrinsicOnAllocation - Check if the specified memory 628/// intrinsic can be promoted by SROA. At this point, we know that the operand 629/// of the memintrinsic is a pointer to the beginning of the allocation. 630void SROA::isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocaInst *AI, 631 unsigned OpNo, AllocaInfo &Info) { 632 // If not constant length, give up. 633 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 634 if (!Length) return MarkUnsafe(Info); 635 636 // If not the whole aggregate, give up. 637 if (Length->getZExtValue() != 638 TD->getTypeAllocSize(AI->getType()->getElementType())) 639 return MarkUnsafe(Info); 640 641 // We only know about memcpy/memset/memmove. 642 if (!isa<MemIntrinsic>(MI)) 643 return MarkUnsafe(Info); 644 645 // Otherwise, we can transform it. Determine whether this is a memcpy/set 646 // into or out of the aggregate. 647 if (OpNo == 1) 648 Info.isMemCpyDst = true; 649 else { 650 assert(OpNo == 2); 651 Info.isMemCpySrc = true; 652 } 653} 654 655/// isSafeUseOfBitCastedAllocation - Check if all users of this bitcast 656/// from an alloca are safe for SROA of that alloca. 657void SROA::isSafeUseOfBitCastedAllocation(BitCastInst *BC, AllocaInst *AI, 658 AllocaInfo &Info) { 659 for (Value::use_iterator UI = BC->use_begin(), E = BC->use_end(); 660 UI != E; ++UI) { 661 if (BitCastInst *BCU = dyn_cast<BitCastInst>(UI)) { 662 isSafeUseOfBitCastedAllocation(BCU, AI, Info); 663 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) { 664 isSafeMemIntrinsicOnAllocation(MI, AI, UI.getOperandNo(), Info); 665 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI)) { 666 if (SI->isVolatile()) 667 return MarkUnsafe(Info); 668 669 // If storing the entire alloca in one chunk through a bitcasted pointer 670 // to integer, we can transform it. This happens (for example) when you 671 // cast a {i32,i32}* to i64* and store through it. This is similar to the 672 // memcpy case and occurs in various "byval" cases and emulated memcpys. 673 if (isa<IntegerType>(SI->getOperand(0)->getType()) && 674 TD->getTypeAllocSize(SI->getOperand(0)->getType()) == 675 TD->getTypeAllocSize(AI->getType()->getElementType())) { 676 Info.isMemCpyDst = true; 677 continue; 678 } 679 return MarkUnsafe(Info); 680 } else if (LoadInst *LI = dyn_cast<LoadInst>(UI)) { 681 if (LI->isVolatile()) 682 return MarkUnsafe(Info); 683 684 // If loading the entire alloca in one chunk through a bitcasted pointer 685 // to integer, we can transform it. This happens (for example) when you 686 // cast a {i32,i32}* to i64* and load through it. This is similar to the 687 // memcpy case and occurs in various "byval" cases and emulated memcpys. 688 if (isa<IntegerType>(LI->getType()) && 689 TD->getTypeAllocSize(LI->getType()) == 690 TD->getTypeAllocSize(AI->getType()->getElementType())) { 691 Info.isMemCpySrc = true; 692 continue; 693 } 694 return MarkUnsafe(Info); 695 } else if (isa<DbgInfoIntrinsic>(UI)) { 696 // If one user is DbgInfoIntrinsic then check if all users are 697 // DbgInfoIntrinsics. 698 if (OnlyUsedByDbgInfoIntrinsics(BC)) { 699 Info.needsCleanup = true; 700 return; 701 } 702 else 703 MarkUnsafe(Info); 704 } 705 else { 706 return MarkUnsafe(Info); 707 } 708 if (Info.isUnsafe) return; 709 } 710} 711 712/// RewriteBitCastUserOfAlloca - BCInst (transitively) bitcasts AI, or indexes 713/// to its first element. Transform users of the cast to use the new values 714/// instead. 715void SROA::RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocaInst *AI, 716 SmallVector<AllocaInst*, 32> &NewElts) { 717 Value::use_iterator UI = BCInst->use_begin(), UE = BCInst->use_end(); 718 while (UI != UE) { 719 Instruction *User = cast<Instruction>(*UI++); 720 if (BitCastInst *BCU = dyn_cast<BitCastInst>(User)) { 721 RewriteBitCastUserOfAlloca(BCU, AI, NewElts); 722 if (BCU->use_empty()) BCU->eraseFromParent(); 723 continue; 724 } 725 726 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 727 // This must be memcpy/memmove/memset of the entire aggregate. 728 // Split into one per element. 729 RewriteMemIntrinUserOfAlloca(MI, BCInst, AI, NewElts); 730 continue; 731 } 732 733 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 734 // If this is a store of the entire alloca from an integer, rewrite it. 735 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 736 continue; 737 } 738 739 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 740 // If this is a load of the entire alloca to an integer, rewrite it. 741 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 742 continue; 743 } 744 745 // Otherwise it must be some other user of a gep of the first pointer. Just 746 // leave these alone. 747 continue; 748 } 749} 750 751/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 752/// Rewrite it to copy or set the elements of the scalarized memory. 753void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst, 754 AllocaInst *AI, 755 SmallVector<AllocaInst*, 32> &NewElts) { 756 757 // If this is a memcpy/memmove, construct the other pointer as the 758 // appropriate type. The "Other" pointer is the pointer that goes to memory 759 // that doesn't have anything to do with the alloca that we are promoting. For 760 // memset, this Value* stays null. 761 Value *OtherPtr = 0; 762 LLVMContext &Context = MI->getContext(); 763 unsigned MemAlignment = MI->getAlignment(); 764 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 765 if (BCInst == MTI->getRawDest()) 766 OtherPtr = MTI->getRawSource(); 767 else { 768 assert(BCInst == MTI->getRawSource()); 769 OtherPtr = MTI->getRawDest(); 770 } 771 } 772 773 // If there is an other pointer, we want to convert it to the same pointer 774 // type as AI has, so we can GEP through it safely. 775 if (OtherPtr) { 776 // It is likely that OtherPtr is a bitcast, if so, remove it. 777 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr)) 778 OtherPtr = BC->getOperand(0); 779 // All zero GEPs are effectively bitcasts. 780 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr)) 781 if (GEP->hasAllZeroIndices()) 782 OtherPtr = GEP->getOperand(0); 783 784 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr)) 785 if (BCE->getOpcode() == Instruction::BitCast) 786 OtherPtr = BCE->getOperand(0); 787 788 // If the pointer is not the right type, insert a bitcast to the right 789 // type. 790 if (OtherPtr->getType() != AI->getType()) 791 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(), 792 MI); 793 } 794 795 // Process each element of the aggregate. 796 Value *TheFn = MI->getOperand(0); 797 const Type *BytePtrTy = MI->getRawDest()->getType(); 798 bool SROADest = MI->getRawDest() == BCInst; 799 800 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 801 802 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 803 // If this is a memcpy/memmove, emit a GEP of the other element address. 804 Value *OtherElt = 0; 805 unsigned OtherEltAlign = MemAlignment; 806 807 if (OtherPtr) { 808 Value *Idx[2] = { Zero, 809 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 810 OtherElt = GetElementPtrInst::Create(OtherPtr, Idx, Idx + 2, 811 OtherPtr->getNameStr()+"."+Twine(i), 812 MI); 813 uint64_t EltOffset; 814 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 815 if (const StructType *ST = 816 dyn_cast<StructType>(OtherPtrTy->getElementType())) { 817 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 818 } else { 819 const Type *EltTy = 820 cast<SequentialType>(OtherPtr->getType())->getElementType(); 821 EltOffset = TD->getTypeAllocSize(EltTy)*i; 822 } 823 824 // The alignment of the other pointer is the guaranteed alignment of the 825 // element, which is affected by both the known alignment of the whole 826 // mem intrinsic and the alignment of the element. If the alignment of 827 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 828 // known alignment is just 4 bytes. 829 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 830 } 831 832 Value *EltPtr = NewElts[i]; 833 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 834 835 // If we got down to a scalar, insert a load or store as appropriate. 836 if (EltTy->isSingleValueType()) { 837 if (isa<MemTransferInst>(MI)) { 838 if (SROADest) { 839 // From Other to Alloca. 840 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 841 new StoreInst(Elt, EltPtr, MI); 842 } else { 843 // From Alloca to Other. 844 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 845 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 846 } 847 continue; 848 } 849 assert(isa<MemSetInst>(MI)); 850 851 // If the stored element is zero (common case), just store a null 852 // constant. 853 Constant *StoreVal; 854 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) { 855 if (CI->isZero()) { 856 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 857 } else { 858 // If EltTy is a vector type, get the element type. 859 const Type *ValTy = EltTy->getScalarType(); 860 861 // Construct an integer with the right value. 862 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 863 APInt OneVal(EltSize, CI->getZExtValue()); 864 APInt TotalVal(OneVal); 865 // Set each byte. 866 for (unsigned i = 0; 8*i < EltSize; ++i) { 867 TotalVal = TotalVal.shl(8); 868 TotalVal |= OneVal; 869 } 870 871 // Convert the integer value to the appropriate type. 872 StoreVal = ConstantInt::get(Context, TotalVal); 873 if (isa<PointerType>(ValTy)) 874 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 875 else if (ValTy->isFloatingPoint()) 876 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 877 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 878 879 // If the requested value was a vector constant, create it. 880 if (EltTy != ValTy) { 881 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 882 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 883 StoreVal = ConstantVector::get(&Elts[0], NumElts); 884 } 885 } 886 new StoreInst(StoreVal, EltPtr, MI); 887 continue; 888 } 889 // Otherwise, if we're storing a byte variable, use a memset call for 890 // this element. 891 } 892 893 // Cast the element pointer to BytePtrTy. 894 if (EltPtr->getType() != BytePtrTy) 895 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI); 896 897 // Cast the other pointer (if we have one) to BytePtrTy. 898 if (OtherElt && OtherElt->getType() != BytePtrTy) 899 OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(), 900 MI); 901 902 unsigned EltSize = TD->getTypeAllocSize(EltTy); 903 904 // Finally, insert the meminst for this element. 905 if (isa<MemTransferInst>(MI)) { 906 Value *Ops[] = { 907 SROADest ? EltPtr : OtherElt, // Dest ptr 908 SROADest ? OtherElt : EltPtr, // Src ptr 909 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 910 // Align 911 ConstantInt::get(Type::getInt32Ty(MI->getContext()), OtherEltAlign) 912 }; 913 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 914 } else { 915 assert(isa<MemSetInst>(MI)); 916 Value *Ops[] = { 917 EltPtr, MI->getOperand(2), // Dest, Value, 918 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 919 Zero // Align 920 }; 921 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 922 } 923 } 924 MI->eraseFromParent(); 925} 926 927/// RewriteStoreUserOfWholeAlloca - We found a store of an integer that 928/// overwrites the entire allocation. Extract out the pieces of the stored 929/// integer and store them individually. 930void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 931 SmallVector<AllocaInst*, 32> &NewElts){ 932 // Extract each element out of the integer according to its structure offset 933 // and store the element value to the individual alloca. 934 Value *SrcVal = SI->getOperand(0); 935 const Type *AllocaEltTy = AI->getType()->getElementType(); 936 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 937 938 // If this isn't a store of an integer to the whole alloca, it may be a store 939 // to the first element. Just ignore the store in this case and normal SROA 940 // will handle it. 941 if (!isa<IntegerType>(SrcVal->getType()) || 942 TD->getTypeAllocSizeInBits(SrcVal->getType()) != AllocaSizeBits) 943 return; 944 // Handle tail padding by extending the operand 945 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 946 SrcVal = new ZExtInst(SrcVal, 947 IntegerType::get(SI->getContext(), AllocaSizeBits), 948 "", SI); 949 950 DEBUG(errs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 951 << '\n'); 952 953 // There are two forms here: AI could be an array or struct. Both cases 954 // have different ways to compute the element offset. 955 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 956 const StructLayout *Layout = TD->getStructLayout(EltSTy); 957 958 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 959 // Get the number of bits to shift SrcVal to get the value. 960 const Type *FieldTy = EltSTy->getElementType(i); 961 uint64_t Shift = Layout->getElementOffsetInBits(i); 962 963 if (TD->isBigEndian()) 964 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 965 966 Value *EltVal = SrcVal; 967 if (Shift) { 968 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 969 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 970 "sroa.store.elt", SI); 971 } 972 973 // Truncate down to an integer of the right size. 974 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 975 976 // Ignore zero sized fields like {}, they obviously contain no data. 977 if (FieldSizeBits == 0) continue; 978 979 if (FieldSizeBits != AllocaSizeBits) 980 EltVal = new TruncInst(EltVal, 981 IntegerType::get(SI->getContext(), FieldSizeBits), 982 "", SI); 983 Value *DestField = NewElts[i]; 984 if (EltVal->getType() == FieldTy) { 985 // Storing to an integer field of this size, just do it. 986 } else if (FieldTy->isFloatingPoint() || isa<VectorType>(FieldTy)) { 987 // Bitcast to the right element type (for fp/vector values). 988 EltVal = new BitCastInst(EltVal, FieldTy, "", SI); 989 } else { 990 // Otherwise, bitcast the dest pointer (for aggregates). 991 DestField = new BitCastInst(DestField, 992 PointerType::getUnqual(EltVal->getType()), 993 "", SI); 994 } 995 new StoreInst(EltVal, DestField, SI); 996 } 997 998 } else { 999 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 1000 const Type *ArrayEltTy = ATy->getElementType(); 1001 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1002 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 1003 1004 uint64_t Shift; 1005 1006 if (TD->isBigEndian()) 1007 Shift = AllocaSizeBits-ElementOffset; 1008 else 1009 Shift = 0; 1010 1011 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1012 // Ignore zero sized fields like {}, they obviously contain no data. 1013 if (ElementSizeBits == 0) continue; 1014 1015 Value *EltVal = SrcVal; 1016 if (Shift) { 1017 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 1018 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 1019 "sroa.store.elt", SI); 1020 } 1021 1022 // Truncate down to an integer of the right size. 1023 if (ElementSizeBits != AllocaSizeBits) 1024 EltVal = new TruncInst(EltVal, 1025 IntegerType::get(SI->getContext(), 1026 ElementSizeBits),"",SI); 1027 Value *DestField = NewElts[i]; 1028 if (EltVal->getType() == ArrayEltTy) { 1029 // Storing to an integer field of this size, just do it. 1030 } else if (ArrayEltTy->isFloatingPoint() || isa<VectorType>(ArrayEltTy)) { 1031 // Bitcast to the right element type (for fp/vector values). 1032 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI); 1033 } else { 1034 // Otherwise, bitcast the dest pointer (for aggregates). 1035 DestField = new BitCastInst(DestField, 1036 PointerType::getUnqual(EltVal->getType()), 1037 "", SI); 1038 } 1039 new StoreInst(EltVal, DestField, SI); 1040 1041 if (TD->isBigEndian()) 1042 Shift -= ElementOffset; 1043 else 1044 Shift += ElementOffset; 1045 } 1046 } 1047 1048 SI->eraseFromParent(); 1049} 1050 1051/// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to 1052/// an integer. Load the individual pieces to form the aggregate value. 1053void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 1054 SmallVector<AllocaInst*, 32> &NewElts) { 1055 // Extract each element out of the NewElts according to its structure offset 1056 // and form the result value. 1057 const Type *AllocaEltTy = AI->getType()->getElementType(); 1058 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1059 1060 // If this isn't a load of the whole alloca to an integer, it may be a load 1061 // of the first element. Just ignore the load in this case and normal SROA 1062 // will handle it. 1063 if (!isa<IntegerType>(LI->getType()) || 1064 TD->getTypeAllocSizeInBits(LI->getType()) != AllocaSizeBits) 1065 return; 1066 1067 DEBUG(errs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 1068 << '\n'); 1069 1070 // There are two forms here: AI could be an array or struct. Both cases 1071 // have different ways to compute the element offset. 1072 const StructLayout *Layout = 0; 1073 uint64_t ArrayEltBitOffset = 0; 1074 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1075 Layout = TD->getStructLayout(EltSTy); 1076 } else { 1077 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 1078 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1079 } 1080 1081 Value *ResultVal = 1082 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 1083 1084 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1085 // Load the value from the alloca. If the NewElt is an aggregate, cast 1086 // the pointer to an integer of the same size before doing the load. 1087 Value *SrcField = NewElts[i]; 1088 const Type *FieldTy = 1089 cast<PointerType>(SrcField->getType())->getElementType(); 1090 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1091 1092 // Ignore zero sized fields like {}, they obviously contain no data. 1093 if (FieldSizeBits == 0) continue; 1094 1095 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 1096 FieldSizeBits); 1097 if (!isa<IntegerType>(FieldTy) && !FieldTy->isFloatingPoint() && 1098 !isa<VectorType>(FieldTy)) 1099 SrcField = new BitCastInst(SrcField, 1100 PointerType::getUnqual(FieldIntTy), 1101 "", LI); 1102 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 1103 1104 // If SrcField is a fp or vector of the right size but that isn't an 1105 // integer type, bitcast to an integer so we can shift it. 1106 if (SrcField->getType() != FieldIntTy) 1107 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 1108 1109 // Zero extend the field to be the same size as the final alloca so that 1110 // we can shift and insert it. 1111 if (SrcField->getType() != ResultVal->getType()) 1112 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 1113 1114 // Determine the number of bits to shift SrcField. 1115 uint64_t Shift; 1116 if (Layout) // Struct case. 1117 Shift = Layout->getElementOffsetInBits(i); 1118 else // Array case. 1119 Shift = i*ArrayEltBitOffset; 1120 1121 if (TD->isBigEndian()) 1122 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 1123 1124 if (Shift) { 1125 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 1126 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 1127 } 1128 1129 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 1130 } 1131 1132 // Handle tail padding by truncating the result 1133 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 1134 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 1135 1136 LI->replaceAllUsesWith(ResultVal); 1137 LI->eraseFromParent(); 1138} 1139 1140 1141/// HasPadding - Return true if the specified type has any structure or 1142/// alignment padding, false otherwise. 1143static bool HasPadding(const Type *Ty, const TargetData &TD) { 1144 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 1145 const StructLayout *SL = TD.getStructLayout(STy); 1146 unsigned PrevFieldBitOffset = 0; 1147 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1148 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 1149 1150 // Padding in sub-elements? 1151 if (HasPadding(STy->getElementType(i), TD)) 1152 return true; 1153 1154 // Check to see if there is any padding between this element and the 1155 // previous one. 1156 if (i) { 1157 unsigned PrevFieldEnd = 1158 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 1159 if (PrevFieldEnd < FieldBitOffset) 1160 return true; 1161 } 1162 1163 PrevFieldBitOffset = FieldBitOffset; 1164 } 1165 1166 // Check for tail padding. 1167 if (unsigned EltCount = STy->getNumElements()) { 1168 unsigned PrevFieldEnd = PrevFieldBitOffset + 1169 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 1170 if (PrevFieldEnd < SL->getSizeInBits()) 1171 return true; 1172 } 1173 1174 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1175 return HasPadding(ATy->getElementType(), TD); 1176 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) { 1177 return HasPadding(VTy->getElementType(), TD); 1178 } 1179 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 1180} 1181 1182/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 1183/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 1184/// or 1 if safe after canonicalization has been performed. 1185int SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 1186 // Loop over the use list of the alloca. We can only transform it if all of 1187 // the users are safe to transform. 1188 AllocaInfo Info; 1189 1190 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); 1191 I != E; ++I) { 1192 isSafeUseOfAllocation(cast<Instruction>(*I), AI, Info); 1193 if (Info.isUnsafe) { 1194 DEBUG(errs() << "Cannot transform: " << *AI << "\n due to user: " 1195 << **I << '\n'); 1196 return 0; 1197 } 1198 } 1199 1200 // Okay, we know all the users are promotable. If the aggregate is a memcpy 1201 // source and destination, we have to be careful. In particular, the memcpy 1202 // could be moving around elements that live in structure padding of the LLVM 1203 // types, but may actually be used. In these cases, we refuse to promote the 1204 // struct. 1205 if (Info.isMemCpySrc && Info.isMemCpyDst && 1206 HasPadding(AI->getType()->getElementType(), *TD)) 1207 return 0; 1208 1209 // If we require cleanup, return 1, otherwise return 3. 1210 return Info.needsCleanup ? 1 : 3; 1211} 1212 1213/// CleanupGEP - GEP is used by an Alloca, which can be prompted after the GEP 1214/// is canonicalized here. 1215void SROA::CleanupGEP(GetElementPtrInst *GEPI) { 1216 gep_type_iterator I = gep_type_begin(GEPI); 1217 ++I; 1218 1219 const ArrayType *AT = dyn_cast<ArrayType>(*I); 1220 if (!AT) 1221 return; 1222 1223 uint64_t NumElements = AT->getNumElements(); 1224 1225 if (isa<ConstantInt>(I.getOperand())) 1226 return; 1227 1228 if (NumElements == 1) { 1229 GEPI->setOperand(2, 1230 Constant::getNullValue(Type::getInt32Ty(GEPI->getContext()))); 1231 return; 1232 } 1233 1234 assert(NumElements == 2 && "Unhandled case!"); 1235 // All users of the GEP must be loads. At each use of the GEP, insert 1236 // two loads of the appropriate indexed GEP and select between them. 1237 Value *IsOne = new ICmpInst(GEPI, ICmpInst::ICMP_NE, I.getOperand(), 1238 Constant::getNullValue(I.getOperand()->getType()), 1239 "isone"); 1240 // Insert the new GEP instructions, which are properly indexed. 1241 SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end()); 1242 Indices[1] = Constant::getNullValue(Type::getInt32Ty(GEPI->getContext())); 1243 Value *ZeroIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 1244 Indices.begin(), 1245 Indices.end(), 1246 GEPI->getName()+".0", GEPI); 1247 Indices[1] = ConstantInt::get(Type::getInt32Ty(GEPI->getContext()), 1); 1248 Value *OneIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 1249 Indices.begin(), 1250 Indices.end(), 1251 GEPI->getName()+".1", GEPI); 1252 // Replace all loads of the variable index GEP with loads from both 1253 // indexes and a select. 1254 while (!GEPI->use_empty()) { 1255 LoadInst *LI = cast<LoadInst>(GEPI->use_back()); 1256 Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI); 1257 Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI); 1258 Value *R = SelectInst::Create(IsOne, One, Zero, LI->getName(), LI); 1259 LI->replaceAllUsesWith(R); 1260 LI->eraseFromParent(); 1261 } 1262 GEPI->eraseFromParent(); 1263} 1264 1265 1266/// CleanupAllocaUsers - If SROA reported that it can promote the specified 1267/// allocation, but only if cleaned up, perform the cleanups required. 1268void SROA::CleanupAllocaUsers(AllocaInst *AI) { 1269 // At this point, we know that the end result will be SROA'd and promoted, so 1270 // we can insert ugly code if required so long as sroa+mem2reg will clean it 1271 // up. 1272 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 1273 UI != E; ) { 1274 User *U = *UI++; 1275 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) 1276 CleanupGEP(GEPI); 1277 else { 1278 Instruction *I = cast<Instruction>(U); 1279 SmallVector<DbgInfoIntrinsic *, 2> DbgInUses; 1280 if (!isa<StoreInst>(I) && OnlyUsedByDbgInfoIntrinsics(I, &DbgInUses)) { 1281 // Safe to remove debug info uses. 1282 while (!DbgInUses.empty()) { 1283 DbgInfoIntrinsic *DI = DbgInUses.back(); DbgInUses.pop_back(); 1284 DI->eraseFromParent(); 1285 } 1286 I->eraseFromParent(); 1287 } 1288 } 1289 } 1290} 1291 1292/// MergeInType - Add the 'In' type to the accumulated type (Accum) so far at 1293/// the offset specified by Offset (which is specified in bytes). 1294/// 1295/// There are two cases we handle here: 1296/// 1) A union of vector types of the same size and potentially its elements. 1297/// Here we turn element accesses into insert/extract element operations. 1298/// This promotes a <4 x float> with a store of float to the third element 1299/// into a <4 x float> that uses insert element. 1300/// 2) A fully general blob of memory, which we turn into some (potentially 1301/// large) integer type with extract and insert operations where the loads 1302/// and stores would mutate the memory. 1303static void MergeInType(const Type *In, uint64_t Offset, const Type *&VecTy, 1304 unsigned AllocaSize, const TargetData &TD, 1305 LLVMContext &Context) { 1306 // If this could be contributing to a vector, analyze it. 1307 if (VecTy != Type::getVoidTy(Context)) { // either null or a vector type. 1308 1309 // If the In type is a vector that is the same size as the alloca, see if it 1310 // matches the existing VecTy. 1311 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 1312 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 1313 // If we're storing/loading a vector of the right size, allow it as a 1314 // vector. If this the first vector we see, remember the type so that 1315 // we know the element size. 1316 if (VecTy == 0) 1317 VecTy = VInTy; 1318 return; 1319 } 1320 } else if (In->isFloatTy() || In->isDoubleTy() || 1321 (isa<IntegerType>(In) && In->getPrimitiveSizeInBits() >= 8 && 1322 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 1323 // If we're accessing something that could be an element of a vector, see 1324 // if the implied vector agrees with what we already have and if Offset is 1325 // compatible with it. 1326 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 1327 if (Offset % EltSize == 0 && 1328 AllocaSize % EltSize == 0 && 1329 (VecTy == 0 || 1330 cast<VectorType>(VecTy)->getElementType() 1331 ->getPrimitiveSizeInBits()/8 == EltSize)) { 1332 if (VecTy == 0) 1333 VecTy = VectorType::get(In, AllocaSize/EltSize); 1334 return; 1335 } 1336 } 1337 } 1338 1339 // Otherwise, we have a case that we can't handle with an optimized vector 1340 // form. We can still turn this into a large integer. 1341 VecTy = Type::getVoidTy(Context); 1342} 1343 1344/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 1345/// its accesses to use a to single vector type, return true, and set VecTy to 1346/// the new type. If we could convert the alloca into a single promotable 1347/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 1348/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 1349/// is the current offset from the base of the alloca being analyzed. 1350/// 1351/// If we see at least one access to the value that is as a vector type, set the 1352/// SawVec flag. 1353bool SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy, 1354 bool &SawVec, uint64_t Offset, 1355 unsigned AllocaSize) { 1356 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1357 Instruction *User = cast<Instruction>(*UI); 1358 1359 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1360 // Don't break volatile loads. 1361 if (LI->isVolatile()) 1362 return false; 1363 MergeInType(LI->getType(), Offset, VecTy, 1364 AllocaSize, *TD, V->getContext()); 1365 SawVec |= isa<VectorType>(LI->getType()); 1366 continue; 1367 } 1368 1369 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1370 // Storing the pointer, not into the value? 1371 if (SI->getOperand(0) == V || SI->isVolatile()) return 0; 1372 MergeInType(SI->getOperand(0)->getType(), Offset, 1373 VecTy, AllocaSize, *TD, V->getContext()); 1374 SawVec |= isa<VectorType>(SI->getOperand(0)->getType()); 1375 continue; 1376 } 1377 1378 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 1379 if (!CanConvertToScalar(BCI, IsNotTrivial, VecTy, SawVec, Offset, 1380 AllocaSize)) 1381 return false; 1382 IsNotTrivial = true; 1383 continue; 1384 } 1385 1386 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1387 // If this is a GEP with a variable indices, we can't handle it. 1388 if (!GEP->hasAllConstantIndices()) 1389 return false; 1390 1391 // Compute the offset that this GEP adds to the pointer. 1392 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1393 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(), 1394 &Indices[0], Indices.size()); 1395 // See if all uses can be converted. 1396 if (!CanConvertToScalar(GEP, IsNotTrivial, VecTy, SawVec,Offset+GEPOffset, 1397 AllocaSize)) 1398 return false; 1399 IsNotTrivial = true; 1400 continue; 1401 } 1402 1403 // If this is a constant sized memset of a constant value (e.g. 0) we can 1404 // handle it. 1405 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 1406 // Store of constant value and constant size. 1407 if (isa<ConstantInt>(MSI->getValue()) && 1408 isa<ConstantInt>(MSI->getLength())) { 1409 IsNotTrivial = true; 1410 continue; 1411 } 1412 } 1413 1414 // If this is a memcpy or memmove into or out of the whole allocation, we 1415 // can handle it like a load or store of the scalar type. 1416 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 1417 if (ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength())) 1418 if (Len->getZExtValue() == AllocaSize && Offset == 0) { 1419 IsNotTrivial = true; 1420 continue; 1421 } 1422 } 1423 1424 // Ignore dbg intrinsic. 1425 if (isa<DbgInfoIntrinsic>(User)) 1426 continue; 1427 1428 // Otherwise, we cannot handle this! 1429 return false; 1430 } 1431 1432 return true; 1433} 1434 1435/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 1436/// directly. This happens when we are converting an "integer union" to a 1437/// single integer scalar, or when we are converting a "vector union" to a 1438/// vector with insert/extractelement instructions. 1439/// 1440/// Offset is an offset from the original alloca, in bits that need to be 1441/// shifted to the right. By the end of this, there should be no uses of Ptr. 1442void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset) { 1443 while (!Ptr->use_empty()) { 1444 Instruction *User = cast<Instruction>(Ptr->use_back()); 1445 1446 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 1447 ConvertUsesToScalar(CI, NewAI, Offset); 1448 CI->eraseFromParent(); 1449 continue; 1450 } 1451 1452 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1453 // Compute the offset that this GEP adds to the pointer. 1454 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1455 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(), 1456 &Indices[0], Indices.size()); 1457 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 1458 GEP->eraseFromParent(); 1459 continue; 1460 } 1461 1462 IRBuilder<> Builder(User->getParent(), User); 1463 1464 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1465 // The load is a bit extract from NewAI shifted right by Offset bits. 1466 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 1467 Value *NewLoadVal 1468 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 1469 LI->replaceAllUsesWith(NewLoadVal); 1470 LI->eraseFromParent(); 1471 continue; 1472 } 1473 1474 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1475 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 1476 // FIXME: Remove once builder has Twine API. 1477 Value *Old = Builder.CreateLoad(NewAI, 1478 (NewAI->getName()+".in").str().c_str()); 1479 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 1480 Builder); 1481 Builder.CreateStore(New, NewAI); 1482 SI->eraseFromParent(); 1483 continue; 1484 } 1485 1486 // If this is a constant sized memset of a constant value (e.g. 0) we can 1487 // transform it into a store of the expanded constant value. 1488 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 1489 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 1490 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 1491 if (NumBytes != 0) { 1492 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 1493 1494 // Compute the value replicated the right number of times. 1495 APInt APVal(NumBytes*8, Val); 1496 1497 // Splat the value if non-zero. 1498 if (Val) 1499 for (unsigned i = 1; i != NumBytes; ++i) 1500 APVal |= APVal << 8; 1501 1502 // FIXME: Remove once builder has Twine API. 1503 Value *Old = Builder.CreateLoad(NewAI, 1504 (NewAI->getName()+".in").str().c_str()); 1505 Value *New = ConvertScalar_InsertValue( 1506 ConstantInt::get(User->getContext(), APVal), 1507 Old, Offset, Builder); 1508 Builder.CreateStore(New, NewAI); 1509 } 1510 MSI->eraseFromParent(); 1511 continue; 1512 } 1513 1514 // If this is a memcpy or memmove into or out of the whole allocation, we 1515 // can handle it like a load or store of the scalar type. 1516 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 1517 assert(Offset == 0 && "must be store to start of alloca"); 1518 1519 // If the source and destination are both to the same alloca, then this is 1520 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 1521 // as appropriate. 1522 AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject()); 1523 1524 if (MTI->getSource()->getUnderlyingObject() != OrigAI) { 1525 // Dest must be OrigAI, change this to be a load from the original 1526 // pointer (bitcasted), then a store to our new alloca. 1527 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 1528 Value *SrcPtr = MTI->getSource(); 1529 SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType()); 1530 1531 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 1532 SrcVal->setAlignment(MTI->getAlignment()); 1533 Builder.CreateStore(SrcVal, NewAI); 1534 } else if (MTI->getDest()->getUnderlyingObject() != OrigAI) { 1535 // Src must be OrigAI, change this to be a load from NewAI then a store 1536 // through the original dest pointer (bitcasted). 1537 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 1538 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 1539 1540 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType()); 1541 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 1542 NewStore->setAlignment(MTI->getAlignment()); 1543 } else { 1544 // Noop transfer. Src == Dst 1545 } 1546 1547 1548 MTI->eraseFromParent(); 1549 continue; 1550 } 1551 1552 // If user is a dbg info intrinsic then it is safe to remove it. 1553 if (isa<DbgInfoIntrinsic>(User)) { 1554 User->eraseFromParent(); 1555 continue; 1556 } 1557 1558 llvm_unreachable("Unsupported operation!"); 1559 } 1560} 1561 1562/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 1563/// or vector value FromVal, extracting the bits from the offset specified by 1564/// Offset. This returns the value, which is of type ToType. 1565/// 1566/// This happens when we are converting an "integer union" to a single 1567/// integer scalar, or when we are converting a "vector union" to a vector with 1568/// insert/extractelement instructions. 1569/// 1570/// Offset is an offset from the original alloca, in bits that need to be 1571/// shifted to the right. 1572Value *SROA::ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 1573 uint64_t Offset, IRBuilder<> &Builder) { 1574 // If the load is of the whole new alloca, no conversion is needed. 1575 if (FromVal->getType() == ToType && Offset == 0) 1576 return FromVal; 1577 1578 // If the result alloca is a vector type, this is either an element 1579 // access or a bitcast to another vector type of the same size. 1580 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) { 1581 if (isa<VectorType>(ToType)) 1582 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 1583 1584 // Otherwise it must be an element access. 1585 unsigned Elt = 0; 1586 if (Offset) { 1587 unsigned EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 1588 Elt = Offset/EltSize; 1589 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 1590 } 1591 // Return the element extracted out of it. 1592 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get( 1593 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp"); 1594 if (V->getType() != ToType) 1595 V = Builder.CreateBitCast(V, ToType, "tmp"); 1596 return V; 1597 } 1598 1599 // If ToType is a first class aggregate, extract out each of the pieces and 1600 // use insertvalue's to form the FCA. 1601 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 1602 const StructLayout &Layout = *TD->getStructLayout(ST); 1603 Value *Res = UndefValue::get(ST); 1604 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1605 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 1606 Offset+Layout.getElementOffsetInBits(i), 1607 Builder); 1608 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1609 } 1610 return Res; 1611 } 1612 1613 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 1614 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 1615 Value *Res = UndefValue::get(AT); 1616 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1617 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 1618 Offset+i*EltSize, Builder); 1619 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1620 } 1621 return Res; 1622 } 1623 1624 // Otherwise, this must be a union that was converted to an integer value. 1625 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 1626 1627 // If this is a big-endian system and the load is narrower than the 1628 // full alloca type, we need to do a shift to get the right bits. 1629 int ShAmt = 0; 1630 if (TD->isBigEndian()) { 1631 // On big-endian machines, the lowest bit is stored at the bit offset 1632 // from the pointer given by getTypeStoreSizeInBits. This matters for 1633 // integers with a bitwidth that is not a multiple of 8. 1634 ShAmt = TD->getTypeStoreSizeInBits(NTy) - 1635 TD->getTypeStoreSizeInBits(ToType) - Offset; 1636 } else { 1637 ShAmt = Offset; 1638 } 1639 1640 // Note: we support negative bitwidths (with shl) which are not defined. 1641 // We do this to support (f.e.) loads off the end of a structure where 1642 // only some bits are used. 1643 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 1644 FromVal = Builder.CreateLShr(FromVal, 1645 ConstantInt::get(FromVal->getType(), 1646 ShAmt), "tmp"); 1647 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 1648 FromVal = Builder.CreateShl(FromVal, 1649 ConstantInt::get(FromVal->getType(), 1650 -ShAmt), "tmp"); 1651 1652 // Finally, unconditionally truncate the integer to the right width. 1653 unsigned LIBitWidth = TD->getTypeSizeInBits(ToType); 1654 if (LIBitWidth < NTy->getBitWidth()) 1655 FromVal = 1656 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 1657 LIBitWidth), "tmp"); 1658 else if (LIBitWidth > NTy->getBitWidth()) 1659 FromVal = 1660 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 1661 LIBitWidth), "tmp"); 1662 1663 // If the result is an integer, this is a trunc or bitcast. 1664 if (isa<IntegerType>(ToType)) { 1665 // Should be done. 1666 } else if (ToType->isFloatingPoint() || isa<VectorType>(ToType)) { 1667 // Just do a bitcast, we know the sizes match up. 1668 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 1669 } else { 1670 // Otherwise must be a pointer. 1671 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 1672 } 1673 assert(FromVal->getType() == ToType && "Didn't convert right?"); 1674 return FromVal; 1675} 1676 1677/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 1678/// or vector value "Old" at the offset specified by Offset. 1679/// 1680/// This happens when we are converting an "integer union" to a 1681/// single integer scalar, or when we are converting a "vector union" to a 1682/// vector with insert/extractelement instructions. 1683/// 1684/// Offset is an offset from the original alloca, in bits that need to be 1685/// shifted to the right. 1686Value *SROA::ConvertScalar_InsertValue(Value *SV, Value *Old, 1687 uint64_t Offset, IRBuilder<> &Builder) { 1688 1689 // Convert the stored type to the actual type, shift it left to insert 1690 // then 'or' into place. 1691 const Type *AllocaType = Old->getType(); 1692 LLVMContext &Context = Old->getContext(); 1693 1694 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 1695 uint64_t VecSize = TD->getTypeAllocSizeInBits(VTy); 1696 uint64_t ValSize = TD->getTypeAllocSizeInBits(SV->getType()); 1697 1698 // Changing the whole vector with memset or with an access of a different 1699 // vector type? 1700 if (ValSize == VecSize) 1701 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 1702 1703 uint64_t EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 1704 1705 // Must be an element insertion. 1706 unsigned Elt = Offset/EltSize; 1707 1708 if (SV->getType() != VTy->getElementType()) 1709 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 1710 1711 SV = Builder.CreateInsertElement(Old, SV, 1712 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt), 1713 "tmp"); 1714 return SV; 1715 } 1716 1717 // If SV is a first-class aggregate value, insert each value recursively. 1718 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 1719 const StructLayout &Layout = *TD->getStructLayout(ST); 1720 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1721 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1722 Old = ConvertScalar_InsertValue(Elt, Old, 1723 Offset+Layout.getElementOffsetInBits(i), 1724 Builder); 1725 } 1726 return Old; 1727 } 1728 1729 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 1730 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 1731 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1732 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1733 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 1734 } 1735 return Old; 1736 } 1737 1738 // If SV is a float, convert it to the appropriate integer type. 1739 // If it is a pointer, do the same. 1740 unsigned SrcWidth = TD->getTypeSizeInBits(SV->getType()); 1741 unsigned DestWidth = TD->getTypeSizeInBits(AllocaType); 1742 unsigned SrcStoreWidth = TD->getTypeStoreSizeInBits(SV->getType()); 1743 unsigned DestStoreWidth = TD->getTypeStoreSizeInBits(AllocaType); 1744 if (SV->getType()->isFloatingPoint() || isa<VectorType>(SV->getType())) 1745 SV = Builder.CreateBitCast(SV, 1746 IntegerType::get(SV->getContext(),SrcWidth), "tmp"); 1747 else if (isa<PointerType>(SV->getType())) 1748 SV = Builder.CreatePtrToInt(SV, TD->getIntPtrType(SV->getContext()), "tmp"); 1749 1750 // Zero extend or truncate the value if needed. 1751 if (SV->getType() != AllocaType) { 1752 if (SV->getType()->getPrimitiveSizeInBits() < 1753 AllocaType->getPrimitiveSizeInBits()) 1754 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 1755 else { 1756 // Truncation may be needed if storing more than the alloca can hold 1757 // (undefined behavior). 1758 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 1759 SrcWidth = DestWidth; 1760 SrcStoreWidth = DestStoreWidth; 1761 } 1762 } 1763 1764 // If this is a big-endian system and the store is narrower than the 1765 // full alloca type, we need to do a shift to get the right bits. 1766 int ShAmt = 0; 1767 if (TD->isBigEndian()) { 1768 // On big-endian machines, the lowest bit is stored at the bit offset 1769 // from the pointer given by getTypeStoreSizeInBits. This matters for 1770 // integers with a bitwidth that is not a multiple of 8. 1771 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 1772 } else { 1773 ShAmt = Offset; 1774 } 1775 1776 // Note: we support negative bitwidths (with shr) which are not defined. 1777 // We do this to support (f.e.) stores off the end of a structure where 1778 // only some bits in the structure are set. 1779 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 1780 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 1781 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 1782 ShAmt), "tmp"); 1783 Mask <<= ShAmt; 1784 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1785 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 1786 -ShAmt), "tmp"); 1787 Mask = Mask.lshr(-ShAmt); 1788 } 1789 1790 // Mask out the bits we are about to insert from the old value, and or 1791 // in the new bits. 1792 if (SrcWidth != DestWidth) { 1793 assert(DestWidth > SrcWidth); 1794 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 1795 SV = Builder.CreateOr(Old, SV, "ins"); 1796 } 1797 return SV; 1798} 1799 1800 1801 1802/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1803/// some part of a constant global variable. This intentionally only accepts 1804/// constant expressions because we don't can't rewrite arbitrary instructions. 1805static bool PointsToConstantGlobal(Value *V) { 1806 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1807 return GV->isConstant(); 1808 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1809 if (CE->getOpcode() == Instruction::BitCast || 1810 CE->getOpcode() == Instruction::GetElementPtr) 1811 return PointsToConstantGlobal(CE->getOperand(0)); 1812 return false; 1813} 1814 1815/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1816/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1817/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1818/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1819/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1820/// the alloca, and if the source pointer is a pointer to a constant global, we 1821/// can optimize this. 1822static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy, 1823 bool isOffset) { 1824 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1825 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) 1826 // Ignore non-volatile loads, they are always ok. 1827 if (!LI->isVolatile()) 1828 continue; 1829 1830 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) { 1831 // If uses of the bitcast are ok, we are ok. 1832 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 1833 return false; 1834 continue; 1835 } 1836 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) { 1837 // If the GEP has all zero indices, it doesn't offset the pointer. If it 1838 // doesn't, it does. 1839 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 1840 isOffset || !GEP->hasAllZeroIndices())) 1841 return false; 1842 continue; 1843 } 1844 1845 // If this is isn't our memcpy/memmove, reject it as something we can't 1846 // handle. 1847 if (!isa<MemTransferInst>(*UI)) 1848 return false; 1849 1850 // If we already have seen a copy, reject the second one. 1851 if (TheCopy) return false; 1852 1853 // If the pointer has been offset from the start of the alloca, we can't 1854 // safely handle this. 1855 if (isOffset) return false; 1856 1857 // If the memintrinsic isn't using the alloca as the dest, reject it. 1858 if (UI.getOperandNo() != 1) return false; 1859 1860 MemIntrinsic *MI = cast<MemIntrinsic>(*UI); 1861 1862 // If the source of the memcpy/move is not a constant global, reject it. 1863 if (!PointsToConstantGlobal(MI->getOperand(2))) 1864 return false; 1865 1866 // Otherwise, the transform is safe. Remember the copy instruction. 1867 TheCopy = MI; 1868 } 1869 return true; 1870} 1871 1872/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 1873/// modified by a copy from a constant global. If we can prove this, we can 1874/// replace any uses of the alloca with uses of the global directly. 1875Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) { 1876 Instruction *TheCopy = 0; 1877 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 1878 return TheCopy; 1879 return 0; 1880} 1881