ScalarReplAggregates.cpp revision e6f329476c6598db1b45c3d573eac9cf65179831
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 } 505 506 507 isSafeElementUse(GEP, AreAllZeroIndices, AI, Info); 508 if (Info.isUnsafe) return; 509 break; 510 } 511 case Instruction::BitCast: 512 if (isFirstElt) { 513 isSafeUseOfBitCastedAllocation(cast<BitCastInst>(User), AI, Info); 514 if (Info.isUnsafe) return; 515 break; 516 } 517 DEBUG(errs() << " Transformation preventing inst: " << *User << '\n'); 518 return MarkUnsafe(Info); 519 case Instruction::Call: 520 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 521 if (isFirstElt) { 522 isSafeMemIntrinsicOnAllocation(MI, AI, I.getOperandNo(), Info); 523 if (Info.isUnsafe) return; 524 break; 525 } 526 } 527 DEBUG(errs() << " Transformation preventing inst: " << *User << '\n'); 528 return MarkUnsafe(Info); 529 default: 530 DEBUG(errs() << " Transformation preventing inst: " << *User << '\n'); 531 return MarkUnsafe(Info); 532 } 533 } 534 return; // All users look ok :) 535} 536 537/// AllUsersAreLoads - Return true if all users of this value are loads. 538static bool AllUsersAreLoads(Value *Ptr) { 539 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 540 I != E; ++I) 541 if (cast<Instruction>(*I)->getOpcode() != Instruction::Load) 542 return false; 543 return true; 544} 545 546/// isSafeUseOfAllocation - Check to see if this user is an allowed use for an 547/// aggregate allocation. 548/// 549void SROA::isSafeUseOfAllocation(Instruction *User, AllocaInst *AI, 550 AllocaInfo &Info) { 551 if (BitCastInst *C = dyn_cast<BitCastInst>(User)) 552 return isSafeUseOfBitCastedAllocation(C, AI, Info); 553 554 if (LoadInst *LI = dyn_cast<LoadInst>(User)) 555 if (!LI->isVolatile()) 556 return;// Loads (returning a first class aggregrate) are always rewritable 557 558 if (StoreInst *SI = dyn_cast<StoreInst>(User)) 559 if (!SI->isVolatile() && SI->getOperand(0) != AI) 560 return;// Store is ok if storing INTO the pointer, not storing the pointer 561 562 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User); 563 if (GEPI == 0) 564 return MarkUnsafe(Info); 565 566 gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI); 567 568 // The GEP is not safe to transform if not of the form "GEP <ptr>, 0, <cst>". 569 if (I == E || 570 I.getOperand() != Constant::getNullValue(I.getOperand()->getType())) { 571 return MarkUnsafe(Info); 572 } 573 574 ++I; 575 if (I == E) return MarkUnsafe(Info); // ran out of GEP indices?? 576 577 bool IsAllZeroIndices = true; 578 579 // If the first index is a non-constant index into an array, see if we can 580 // handle it as a special case. 581 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 582 if (!isa<ConstantInt>(I.getOperand())) { 583 IsAllZeroIndices = 0; 584 uint64_t NumElements = AT->getNumElements(); 585 586 // If this is an array index and the index is not constant, we cannot 587 // promote... that is unless the array has exactly one or two elements in 588 // it, in which case we CAN promote it, but we have to canonicalize this 589 // out if this is the only problem. 590 if ((NumElements == 1 || NumElements == 2) && 591 AllUsersAreLoads(GEPI)) { 592 Info.needsCleanup = true; 593 return; // Canonicalization required! 594 } 595 return MarkUnsafe(Info); 596 } 597 } 598 599 // Walk through the GEP type indices, checking the types that this indexes 600 // into. 601 for (; I != E; ++I) { 602 // Ignore struct elements, no extra checking needed for these. 603 if (isa<StructType>(*I)) 604 continue; 605 606 ConstantInt *IdxVal = dyn_cast<ConstantInt>(I.getOperand()); 607 if (!IdxVal) return MarkUnsafe(Info); 608 609 // Are all indices still zero? 610 IsAllZeroIndices &= IdxVal->isZero(); 611 612 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 613 // This GEP indexes an array. Verify that this is an in-range constant 614 // integer. Specifically, consider A[0][i]. We cannot know that the user 615 // isn't doing invalid things like allowing i to index an out-of-range 616 // subscript that accesses A[1]. Because of this, we have to reject SROA 617 // of any accesses into structs where any of the components are variables. 618 if (IdxVal->getZExtValue() >= AT->getNumElements()) 619 return MarkUnsafe(Info); 620 } else if (const VectorType *VT = dyn_cast<VectorType>(*I)) { 621 if (IdxVal->getZExtValue() >= VT->getNumElements()) 622 return MarkUnsafe(Info); 623 } 624 } 625 626 // If there are any non-simple uses of this getelementptr, make sure to reject 627 // them. 628 return isSafeElementUse(GEPI, IsAllZeroIndices, AI, Info); 629} 630 631/// isSafeMemIntrinsicOnAllocation - Return true if the specified memory 632/// intrinsic can be promoted by SROA. At this point, we know that the operand 633/// of the memintrinsic is a pointer to the beginning of the allocation. 634void SROA::isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocaInst *AI, 635 unsigned OpNo, AllocaInfo &Info) { 636 // If not constant length, give up. 637 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 638 if (!Length) return MarkUnsafe(Info); 639 640 // If not the whole aggregate, give up. 641 if (Length->getZExtValue() != 642 TD->getTypeAllocSize(AI->getType()->getElementType())) 643 return MarkUnsafe(Info); 644 645 // We only know about memcpy/memset/memmove. 646 if (!isa<MemIntrinsic>(MI)) 647 return MarkUnsafe(Info); 648 649 // Otherwise, we can transform it. Determine whether this is a memcpy/set 650 // into or out of the aggregate. 651 if (OpNo == 1) 652 Info.isMemCpyDst = true; 653 else { 654 assert(OpNo == 2); 655 Info.isMemCpySrc = true; 656 } 657} 658 659/// isSafeUseOfBitCastedAllocation - Return true if all users of this bitcast 660/// are 661void SROA::isSafeUseOfBitCastedAllocation(BitCastInst *BC, AllocaInst *AI, 662 AllocaInfo &Info) { 663 for (Value::use_iterator UI = BC->use_begin(), E = BC->use_end(); 664 UI != E; ++UI) { 665 if (BitCastInst *BCU = dyn_cast<BitCastInst>(UI)) { 666 isSafeUseOfBitCastedAllocation(BCU, AI, Info); 667 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) { 668 isSafeMemIntrinsicOnAllocation(MI, AI, UI.getOperandNo(), Info); 669 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI)) { 670 if (SI->isVolatile()) 671 return MarkUnsafe(Info); 672 673 // If storing the entire alloca in one chunk through a bitcasted pointer 674 // to integer, we can transform it. This happens (for example) when you 675 // cast a {i32,i32}* to i64* and store through it. This is similar to the 676 // memcpy case and occurs in various "byval" cases and emulated memcpys. 677 if (isa<IntegerType>(SI->getOperand(0)->getType()) && 678 TD->getTypeAllocSize(SI->getOperand(0)->getType()) == 679 TD->getTypeAllocSize(AI->getType()->getElementType())) { 680 Info.isMemCpyDst = true; 681 continue; 682 } 683 return MarkUnsafe(Info); 684 } else if (LoadInst *LI = dyn_cast<LoadInst>(UI)) { 685 if (LI->isVolatile()) 686 return MarkUnsafe(Info); 687 688 // If loading the entire alloca in one chunk through a bitcasted pointer 689 // to integer, we can transform it. This happens (for example) when you 690 // cast a {i32,i32}* to i64* and load through it. This is similar to the 691 // memcpy case and occurs in various "byval" cases and emulated memcpys. 692 if (isa<IntegerType>(LI->getType()) && 693 TD->getTypeAllocSize(LI->getType()) == 694 TD->getTypeAllocSize(AI->getType()->getElementType())) { 695 Info.isMemCpySrc = true; 696 continue; 697 } 698 return MarkUnsafe(Info); 699 } else if (isa<DbgInfoIntrinsic>(UI)) { 700 // If one user is DbgInfoIntrinsic then check if all users are 701 // DbgInfoIntrinsics. 702 if (OnlyUsedByDbgInfoIntrinsics(BC)) { 703 Info.needsCleanup = true; 704 return; 705 } 706 else 707 MarkUnsafe(Info); 708 } 709 else { 710 return MarkUnsafe(Info); 711 } 712 if (Info.isUnsafe) return; 713 } 714} 715 716/// RewriteBitCastUserOfAlloca - BCInst (transitively) bitcasts AI, or indexes 717/// to its first element. Transform users of the cast to use the new values 718/// instead. 719void SROA::RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocaInst *AI, 720 SmallVector<AllocaInst*, 32> &NewElts) { 721 Value::use_iterator UI = BCInst->use_begin(), UE = BCInst->use_end(); 722 while (UI != UE) { 723 Instruction *User = cast<Instruction>(*UI++); 724 if (BitCastInst *BCU = dyn_cast<BitCastInst>(User)) { 725 RewriteBitCastUserOfAlloca(BCU, AI, NewElts); 726 if (BCU->use_empty()) BCU->eraseFromParent(); 727 continue; 728 } 729 730 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 731 // This must be memcpy/memmove/memset of the entire aggregate. 732 // Split into one per element. 733 RewriteMemIntrinUserOfAlloca(MI, BCInst, AI, NewElts); 734 continue; 735 } 736 737 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 738 // If this is a store of the entire alloca from an integer, rewrite it. 739 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 740 continue; 741 } 742 743 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 744 // If this is a load of the entire alloca to an integer, rewrite it. 745 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 746 continue; 747 } 748 749 // Otherwise it must be some other user of a gep of the first pointer. Just 750 // leave these alone. 751 continue; 752 } 753} 754 755/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 756/// Rewrite it to copy or set the elements of the scalarized memory. 757void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst, 758 AllocaInst *AI, 759 SmallVector<AllocaInst*, 32> &NewElts) { 760 761 // If this is a memcpy/memmove, construct the other pointer as the 762 // appropriate type. The "Other" pointer is the pointer that goes to memory 763 // that doesn't have anything to do with the alloca that we are promoting. For 764 // memset, this Value* stays null. 765 Value *OtherPtr = 0; 766 LLVMContext &Context = MI->getContext(); 767 unsigned MemAlignment = MI->getAlignment(); 768 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 769 if (BCInst == MTI->getRawDest()) 770 OtherPtr = MTI->getRawSource(); 771 else { 772 assert(BCInst == MTI->getRawSource()); 773 OtherPtr = MTI->getRawDest(); 774 } 775 } 776 777 // If there is an other pointer, we want to convert it to the same pointer 778 // type as AI has, so we can GEP through it safely. 779 if (OtherPtr) { 780 // It is likely that OtherPtr is a bitcast, if so, remove it. 781 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr)) 782 OtherPtr = BC->getOperand(0); 783 // All zero GEPs are effectively bitcasts. 784 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr)) 785 if (GEP->hasAllZeroIndices()) 786 OtherPtr = GEP->getOperand(0); 787 788 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr)) 789 if (BCE->getOpcode() == Instruction::BitCast) 790 OtherPtr = BCE->getOperand(0); 791 792 // If the pointer is not the right type, insert a bitcast to the right 793 // type. 794 if (OtherPtr->getType() != AI->getType()) 795 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(), 796 MI); 797 } 798 799 // Process each element of the aggregate. 800 Value *TheFn = MI->getOperand(0); 801 const Type *BytePtrTy = MI->getRawDest()->getType(); 802 bool SROADest = MI->getRawDest() == BCInst; 803 804 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 805 806 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 807 // If this is a memcpy/memmove, emit a GEP of the other element address. 808 Value *OtherElt = 0; 809 unsigned OtherEltAlign = MemAlignment; 810 811 if (OtherPtr) { 812 Value *Idx[2] = { Zero, 813 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 814 OtherElt = GetElementPtrInst::Create(OtherPtr, Idx, Idx + 2, 815 OtherPtr->getNameStr()+"."+Twine(i), 816 MI); 817 uint64_t EltOffset; 818 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 819 if (const StructType *ST = 820 dyn_cast<StructType>(OtherPtrTy->getElementType())) { 821 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 822 } else { 823 const Type *EltTy = 824 cast<SequentialType>(OtherPtr->getType())->getElementType(); 825 EltOffset = TD->getTypeAllocSize(EltTy)*i; 826 } 827 828 // The alignment of the other pointer is the guaranteed alignment of the 829 // element, which is affected by both the known alignment of the whole 830 // mem intrinsic and the alignment of the element. If the alignment of 831 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 832 // known alignment is just 4 bytes. 833 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 834 } 835 836 Value *EltPtr = NewElts[i]; 837 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 838 839 // If we got down to a scalar, insert a load or store as appropriate. 840 if (EltTy->isSingleValueType()) { 841 if (isa<MemTransferInst>(MI)) { 842 if (SROADest) { 843 // From Other to Alloca. 844 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 845 new StoreInst(Elt, EltPtr, MI); 846 } else { 847 // From Alloca to Other. 848 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 849 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 850 } 851 continue; 852 } 853 assert(isa<MemSetInst>(MI)); 854 855 // If the stored element is zero (common case), just store a null 856 // constant. 857 Constant *StoreVal; 858 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) { 859 if (CI->isZero()) { 860 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 861 } else { 862 // If EltTy is a vector type, get the element type. 863 const Type *ValTy = EltTy->getScalarType(); 864 865 // Construct an integer with the right value. 866 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 867 APInt OneVal(EltSize, CI->getZExtValue()); 868 APInt TotalVal(OneVal); 869 // Set each byte. 870 for (unsigned i = 0; 8*i < EltSize; ++i) { 871 TotalVal = TotalVal.shl(8); 872 TotalVal |= OneVal; 873 } 874 875 // Convert the integer value to the appropriate type. 876 StoreVal = ConstantInt::get(Context, TotalVal); 877 if (isa<PointerType>(ValTy)) 878 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 879 else if (ValTy->isFloatingPoint()) 880 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 881 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 882 883 // If the requested value was a vector constant, create it. 884 if (EltTy != ValTy) { 885 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 886 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 887 StoreVal = ConstantVector::get(&Elts[0], NumElts); 888 } 889 } 890 new StoreInst(StoreVal, EltPtr, MI); 891 continue; 892 } 893 // Otherwise, if we're storing a byte variable, use a memset call for 894 // this element. 895 } 896 897 // Cast the element pointer to BytePtrTy. 898 if (EltPtr->getType() != BytePtrTy) 899 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI); 900 901 // Cast the other pointer (if we have one) to BytePtrTy. 902 if (OtherElt && OtherElt->getType() != BytePtrTy) 903 OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(), 904 MI); 905 906 unsigned EltSize = TD->getTypeAllocSize(EltTy); 907 908 // Finally, insert the meminst for this element. 909 if (isa<MemTransferInst>(MI)) { 910 Value *Ops[] = { 911 SROADest ? EltPtr : OtherElt, // Dest ptr 912 SROADest ? OtherElt : EltPtr, // Src ptr 913 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 914 // Align 915 ConstantInt::get(Type::getInt32Ty(MI->getContext()), OtherEltAlign) 916 }; 917 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 918 } else { 919 assert(isa<MemSetInst>(MI)); 920 Value *Ops[] = { 921 EltPtr, MI->getOperand(2), // Dest, Value, 922 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 923 Zero // Align 924 }; 925 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 926 } 927 } 928 MI->eraseFromParent(); 929} 930 931/// RewriteStoreUserOfWholeAlloca - We found an store of an integer that 932/// overwrites the entire allocation. Extract out the pieces of the stored 933/// integer and store them individually. 934void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI, 935 SmallVector<AllocaInst*, 32> &NewElts){ 936 // Extract each element out of the integer according to its structure offset 937 // and store the element value to the individual alloca. 938 Value *SrcVal = SI->getOperand(0); 939 const Type *AllocaEltTy = AI->getType()->getElementType(); 940 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 941 942 // If this isn't a store of an integer to the whole alloca, it may be a store 943 // to the first element. Just ignore the store in this case and normal SROA 944 // will handle it. 945 if (!isa<IntegerType>(SrcVal->getType()) || 946 TD->getTypeAllocSizeInBits(SrcVal->getType()) != AllocaSizeBits) 947 return; 948 // Handle tail padding by extending the operand 949 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 950 SrcVal = new ZExtInst(SrcVal, 951 IntegerType::get(SI->getContext(), AllocaSizeBits), 952 "", SI); 953 954 DEBUG(errs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI 955 << '\n'); 956 957 // There are two forms here: AI could be an array or struct. Both cases 958 // have different ways to compute the element offset. 959 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 960 const StructLayout *Layout = TD->getStructLayout(EltSTy); 961 962 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 963 // Get the number of bits to shift SrcVal to get the value. 964 const Type *FieldTy = EltSTy->getElementType(i); 965 uint64_t Shift = Layout->getElementOffsetInBits(i); 966 967 if (TD->isBigEndian()) 968 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 969 970 Value *EltVal = SrcVal; 971 if (Shift) { 972 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 973 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 974 "sroa.store.elt", SI); 975 } 976 977 // Truncate down to an integer of the right size. 978 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 979 980 // Ignore zero sized fields like {}, they obviously contain no data. 981 if (FieldSizeBits == 0) continue; 982 983 if (FieldSizeBits != AllocaSizeBits) 984 EltVal = new TruncInst(EltVal, 985 IntegerType::get(SI->getContext(), FieldSizeBits), 986 "", SI); 987 Value *DestField = NewElts[i]; 988 if (EltVal->getType() == FieldTy) { 989 // Storing to an integer field of this size, just do it. 990 } else if (FieldTy->isFloatingPoint() || isa<VectorType>(FieldTy)) { 991 // Bitcast to the right element type (for fp/vector values). 992 EltVal = new BitCastInst(EltVal, FieldTy, "", SI); 993 } else { 994 // Otherwise, bitcast the dest pointer (for aggregates). 995 DestField = new BitCastInst(DestField, 996 PointerType::getUnqual(EltVal->getType()), 997 "", SI); 998 } 999 new StoreInst(EltVal, DestField, SI); 1000 } 1001 1002 } else { 1003 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 1004 const Type *ArrayEltTy = ATy->getElementType(); 1005 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1006 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 1007 1008 uint64_t Shift; 1009 1010 if (TD->isBigEndian()) 1011 Shift = AllocaSizeBits-ElementOffset; 1012 else 1013 Shift = 0; 1014 1015 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1016 // Ignore zero sized fields like {}, they obviously contain no data. 1017 if (ElementSizeBits == 0) continue; 1018 1019 Value *EltVal = SrcVal; 1020 if (Shift) { 1021 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 1022 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 1023 "sroa.store.elt", SI); 1024 } 1025 1026 // Truncate down to an integer of the right size. 1027 if (ElementSizeBits != AllocaSizeBits) 1028 EltVal = new TruncInst(EltVal, 1029 IntegerType::get(SI->getContext(), 1030 ElementSizeBits),"",SI); 1031 Value *DestField = NewElts[i]; 1032 if (EltVal->getType() == ArrayEltTy) { 1033 // Storing to an integer field of this size, just do it. 1034 } else if (ArrayEltTy->isFloatingPoint() || isa<VectorType>(ArrayEltTy)) { 1035 // Bitcast to the right element type (for fp/vector values). 1036 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI); 1037 } else { 1038 // Otherwise, bitcast the dest pointer (for aggregates). 1039 DestField = new BitCastInst(DestField, 1040 PointerType::getUnqual(EltVal->getType()), 1041 "", SI); 1042 } 1043 new StoreInst(EltVal, DestField, SI); 1044 1045 if (TD->isBigEndian()) 1046 Shift -= ElementOffset; 1047 else 1048 Shift += ElementOffset; 1049 } 1050 } 1051 1052 SI->eraseFromParent(); 1053} 1054 1055/// RewriteLoadUserOfWholeAlloca - We found an load of the entire allocation to 1056/// an integer. Load the individual pieces to form the aggregate value. 1057void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI, 1058 SmallVector<AllocaInst*, 32> &NewElts) { 1059 // Extract each element out of the NewElts according to its structure offset 1060 // and form the result value. 1061 const Type *AllocaEltTy = AI->getType()->getElementType(); 1062 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1063 1064 // If this isn't a load of the whole alloca to an integer, it may be a load 1065 // of the first element. Just ignore the load in this case and normal SROA 1066 // will handle it. 1067 if (!isa<IntegerType>(LI->getType()) || 1068 TD->getTypeAllocSizeInBits(LI->getType()) != AllocaSizeBits) 1069 return; 1070 1071 DEBUG(errs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI 1072 << '\n'); 1073 1074 // There are two forms here: AI could be an array or struct. Both cases 1075 // have different ways to compute the element offset. 1076 const StructLayout *Layout = 0; 1077 uint64_t ArrayEltBitOffset = 0; 1078 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1079 Layout = TD->getStructLayout(EltSTy); 1080 } else { 1081 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 1082 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1083 } 1084 1085 Value *ResultVal = 1086 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 1087 1088 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1089 // Load the value from the alloca. If the NewElt is an aggregate, cast 1090 // the pointer to an integer of the same size before doing the load. 1091 Value *SrcField = NewElts[i]; 1092 const Type *FieldTy = 1093 cast<PointerType>(SrcField->getType())->getElementType(); 1094 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1095 1096 // Ignore zero sized fields like {}, they obviously contain no data. 1097 if (FieldSizeBits == 0) continue; 1098 1099 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 1100 FieldSizeBits); 1101 if (!isa<IntegerType>(FieldTy) && !FieldTy->isFloatingPoint() && 1102 !isa<VectorType>(FieldTy)) 1103 SrcField = new BitCastInst(SrcField, 1104 PointerType::getUnqual(FieldIntTy), 1105 "", LI); 1106 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 1107 1108 // If SrcField is a fp or vector of the right size but that isn't an 1109 // integer type, bitcast to an integer so we can shift it. 1110 if (SrcField->getType() != FieldIntTy) 1111 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 1112 1113 // Zero extend the field to be the same size as the final alloca so that 1114 // we can shift and insert it. 1115 if (SrcField->getType() != ResultVal->getType()) 1116 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 1117 1118 // Determine the number of bits to shift SrcField. 1119 uint64_t Shift; 1120 if (Layout) // Struct case. 1121 Shift = Layout->getElementOffsetInBits(i); 1122 else // Array case. 1123 Shift = i*ArrayEltBitOffset; 1124 1125 if (TD->isBigEndian()) 1126 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 1127 1128 if (Shift) { 1129 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 1130 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 1131 } 1132 1133 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 1134 } 1135 1136 // Handle tail padding by truncating the result 1137 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 1138 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 1139 1140 LI->replaceAllUsesWith(ResultVal); 1141 LI->eraseFromParent(); 1142} 1143 1144 1145/// HasPadding - Return true if the specified type has any structure or 1146/// alignment padding, false otherwise. 1147static bool HasPadding(const Type *Ty, const TargetData &TD) { 1148 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 1149 const StructLayout *SL = TD.getStructLayout(STy); 1150 unsigned PrevFieldBitOffset = 0; 1151 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1152 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 1153 1154 // Padding in sub-elements? 1155 if (HasPadding(STy->getElementType(i), TD)) 1156 return true; 1157 1158 // Check to see if there is any padding between this element and the 1159 // previous one. 1160 if (i) { 1161 unsigned PrevFieldEnd = 1162 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 1163 if (PrevFieldEnd < FieldBitOffset) 1164 return true; 1165 } 1166 1167 PrevFieldBitOffset = FieldBitOffset; 1168 } 1169 1170 // Check for tail padding. 1171 if (unsigned EltCount = STy->getNumElements()) { 1172 unsigned PrevFieldEnd = PrevFieldBitOffset + 1173 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 1174 if (PrevFieldEnd < SL->getSizeInBits()) 1175 return true; 1176 } 1177 1178 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1179 return HasPadding(ATy->getElementType(), TD); 1180 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) { 1181 return HasPadding(VTy->getElementType(), TD); 1182 } 1183 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 1184} 1185 1186/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 1187/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 1188/// or 1 if safe after canonicalization has been performed. 1189/// 1190int SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) { 1191 // Loop over the use list of the alloca. We can only transform it if all of 1192 // the users are safe to transform. 1193 AllocaInfo Info; 1194 1195 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); 1196 I != E; ++I) { 1197 isSafeUseOfAllocation(cast<Instruction>(*I), AI, Info); 1198 if (Info.isUnsafe) { 1199 DEBUG(errs() << "Cannot transform: " << *AI << "\n due to user: " 1200 << **I << '\n'); 1201 return 0; 1202 } 1203 } 1204 1205 // Okay, we know all the users are promotable. If the aggregate is a memcpy 1206 // source and destination, we have to be careful. In particular, the memcpy 1207 // could be moving around elements that live in structure padding of the LLVM 1208 // types, but may actually be used. In these cases, we refuse to promote the 1209 // struct. 1210 if (Info.isMemCpySrc && Info.isMemCpyDst && 1211 HasPadding(AI->getType()->getElementType(), *TD)) 1212 return 0; 1213 1214 // If we require cleanup, return 1, otherwise return 3. 1215 return Info.needsCleanup ? 1 : 3; 1216} 1217 1218/// CleanupGEP - GEP is used by an Alloca, which can be prompted after the GEP 1219/// is canonicalized here. 1220void SROA::CleanupGEP(GetElementPtrInst *GEPI) { 1221 gep_type_iterator I = gep_type_begin(GEPI); 1222 ++I; 1223 1224 const ArrayType *AT = dyn_cast<ArrayType>(*I); 1225 if (!AT) 1226 return; 1227 1228 uint64_t NumElements = AT->getNumElements(); 1229 1230 if (isa<ConstantInt>(I.getOperand())) 1231 return; 1232 1233 if (NumElements == 1) { 1234 GEPI->setOperand(2, 1235 Constant::getNullValue(Type::getInt32Ty(GEPI->getContext()))); 1236 return; 1237 } 1238 1239 assert(NumElements == 2 && "Unhandled case!"); 1240 // All users of the GEP must be loads. At each use of the GEP, insert 1241 // two loads of the appropriate indexed GEP and select between them. 1242 Value *IsOne = new ICmpInst(GEPI, ICmpInst::ICMP_NE, I.getOperand(), 1243 Constant::getNullValue(I.getOperand()->getType()), 1244 "isone"); 1245 // Insert the new GEP instructions, which are properly indexed. 1246 SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end()); 1247 Indices[1] = Constant::getNullValue(Type::getInt32Ty(GEPI->getContext())); 1248 Value *ZeroIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 1249 Indices.begin(), 1250 Indices.end(), 1251 GEPI->getName()+".0", GEPI); 1252 Indices[1] = ConstantInt::get(Type::getInt32Ty(GEPI->getContext()), 1); 1253 Value *OneIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 1254 Indices.begin(), 1255 Indices.end(), 1256 GEPI->getName()+".1", GEPI); 1257 // Replace all loads of the variable index GEP with loads from both 1258 // indexes and a select. 1259 while (!GEPI->use_empty()) { 1260 LoadInst *LI = cast<LoadInst>(GEPI->use_back()); 1261 Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI); 1262 Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI); 1263 Value *R = SelectInst::Create(IsOne, One, Zero, LI->getName(), LI); 1264 LI->replaceAllUsesWith(R); 1265 LI->eraseFromParent(); 1266 } 1267 GEPI->eraseFromParent(); 1268} 1269 1270 1271/// CleanupAllocaUsers - If SROA reported that it can promote the specified 1272/// allocation, but only if cleaned up, perform the cleanups required. 1273void SROA::CleanupAllocaUsers(AllocaInst *AI) { 1274 // At this point, we know that the end result will be SROA'd and promoted, so 1275 // we can insert ugly code if required so long as sroa+mem2reg will clean it 1276 // up. 1277 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 1278 UI != E; ) { 1279 User *U = *UI++; 1280 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) 1281 CleanupGEP(GEPI); 1282 else { 1283 Instruction *I = cast<Instruction>(U); 1284 SmallVector<DbgInfoIntrinsic *, 2> DbgInUses; 1285 if (!isa<StoreInst>(I) && OnlyUsedByDbgInfoIntrinsics(I, &DbgInUses)) { 1286 // Safe to remove debug info uses. 1287 while (!DbgInUses.empty()) { 1288 DbgInfoIntrinsic *DI = DbgInUses.back(); DbgInUses.pop_back(); 1289 DI->eraseFromParent(); 1290 } 1291 I->eraseFromParent(); 1292 } 1293 } 1294 } 1295} 1296 1297/// MergeInType - Add the 'In' type to the accumulated type (Accum) so far at 1298/// the offset specified by Offset (which is specified in bytes). 1299/// 1300/// There are two cases we handle here: 1301/// 1) A union of vector types of the same size and potentially its elements. 1302/// Here we turn element accesses into insert/extract element operations. 1303/// This promotes a <4 x float> with a store of float to the third element 1304/// into a <4 x float> that uses insert element. 1305/// 2) A fully general blob of memory, which we turn into some (potentially 1306/// large) integer type with extract and insert operations where the loads 1307/// and stores would mutate the memory. 1308static void MergeInType(const Type *In, uint64_t Offset, const Type *&VecTy, 1309 unsigned AllocaSize, const TargetData &TD, 1310 LLVMContext &Context) { 1311 // If this could be contributing to a vector, analyze it. 1312 if (VecTy != Type::getVoidTy(Context)) { // either null or a vector type. 1313 1314 // If the In type is a vector that is the same size as the alloca, see if it 1315 // matches the existing VecTy. 1316 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 1317 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 1318 // If we're storing/loading a vector of the right size, allow it as a 1319 // vector. If this the first vector we see, remember the type so that 1320 // we know the element size. 1321 if (VecTy == 0) 1322 VecTy = VInTy; 1323 return; 1324 } 1325 } else if (In->isFloatTy() || In->isDoubleTy() || 1326 (isa<IntegerType>(In) && In->getPrimitiveSizeInBits() >= 8 && 1327 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 1328 // If we're accessing something that could be an element of a vector, see 1329 // if the implied vector agrees with what we already have and if Offset is 1330 // compatible with it. 1331 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 1332 if (Offset % EltSize == 0 && 1333 AllocaSize % EltSize == 0 && 1334 (VecTy == 0 || 1335 cast<VectorType>(VecTy)->getElementType() 1336 ->getPrimitiveSizeInBits()/8 == EltSize)) { 1337 if (VecTy == 0) 1338 VecTy = VectorType::get(In, AllocaSize/EltSize); 1339 return; 1340 } 1341 } 1342 } 1343 1344 // Otherwise, we have a case that we can't handle with an optimized vector 1345 // form. We can still turn this into a large integer. 1346 VecTy = Type::getVoidTy(Context); 1347} 1348 1349/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 1350/// its accesses to use a to single vector type, return true, and set VecTy to 1351/// the new type. If we could convert the alloca into a single promotable 1352/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 1353/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 1354/// is the current offset from the base of the alloca being analyzed. 1355/// 1356/// If we see at least one access to the value that is as a vector type, set the 1357/// SawVec flag. 1358/// 1359bool SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy, 1360 bool &SawVec, uint64_t Offset, 1361 unsigned AllocaSize) { 1362 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1363 Instruction *User = cast<Instruction>(*UI); 1364 1365 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1366 // Don't break volatile loads. 1367 if (LI->isVolatile()) 1368 return false; 1369 MergeInType(LI->getType(), Offset, VecTy, 1370 AllocaSize, *TD, V->getContext()); 1371 SawVec |= isa<VectorType>(LI->getType()); 1372 continue; 1373 } 1374 1375 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1376 // Storing the pointer, not into the value? 1377 if (SI->getOperand(0) == V || SI->isVolatile()) return 0; 1378 MergeInType(SI->getOperand(0)->getType(), Offset, 1379 VecTy, AllocaSize, *TD, V->getContext()); 1380 SawVec |= isa<VectorType>(SI->getOperand(0)->getType()); 1381 continue; 1382 } 1383 1384 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 1385 if (!CanConvertToScalar(BCI, IsNotTrivial, VecTy, SawVec, Offset, 1386 AllocaSize)) 1387 return false; 1388 IsNotTrivial = true; 1389 continue; 1390 } 1391 1392 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1393 // If this is a GEP with a variable indices, we can't handle it. 1394 if (!GEP->hasAllConstantIndices()) 1395 return false; 1396 1397 // Compute the offset that this GEP adds to the pointer. 1398 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1399 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(), 1400 &Indices[0], Indices.size()); 1401 // See if all uses can be converted. 1402 if (!CanConvertToScalar(GEP, IsNotTrivial, VecTy, SawVec,Offset+GEPOffset, 1403 AllocaSize)) 1404 return false; 1405 IsNotTrivial = true; 1406 continue; 1407 } 1408 1409 // If this is a constant sized memset of a constant value (e.g. 0) we can 1410 // handle it. 1411 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 1412 // Store of constant value and constant size. 1413 if (isa<ConstantInt>(MSI->getValue()) && 1414 isa<ConstantInt>(MSI->getLength())) { 1415 IsNotTrivial = true; 1416 continue; 1417 } 1418 } 1419 1420 // If this is a memcpy or memmove into or out of the whole allocation, we 1421 // can handle it like a load or store of the scalar type. 1422 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 1423 if (ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength())) 1424 if (Len->getZExtValue() == AllocaSize && Offset == 0) { 1425 IsNotTrivial = true; 1426 continue; 1427 } 1428 } 1429 1430 // Ignore dbg intrinsic. 1431 if (isa<DbgInfoIntrinsic>(User)) 1432 continue; 1433 1434 // Otherwise, we cannot handle this! 1435 return false; 1436 } 1437 1438 return true; 1439} 1440 1441 1442/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 1443/// directly. This happens when we are converting an "integer union" to a 1444/// single integer scalar, or when we are converting a "vector union" to a 1445/// vector with insert/extractelement instructions. 1446/// 1447/// Offset is an offset from the original alloca, in bits that need to be 1448/// shifted to the right. By the end of this, there should be no uses of Ptr. 1449void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset) { 1450 while (!Ptr->use_empty()) { 1451 Instruction *User = cast<Instruction>(Ptr->use_back()); 1452 1453 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 1454 ConvertUsesToScalar(CI, NewAI, Offset); 1455 CI->eraseFromParent(); 1456 continue; 1457 } 1458 1459 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1460 // Compute the offset that this GEP adds to the pointer. 1461 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1462 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(), 1463 &Indices[0], Indices.size()); 1464 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 1465 GEP->eraseFromParent(); 1466 continue; 1467 } 1468 1469 IRBuilder<> Builder(User->getParent(), User); 1470 1471 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1472 // The load is a bit extract from NewAI shifted right by Offset bits. 1473 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 1474 Value *NewLoadVal 1475 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 1476 LI->replaceAllUsesWith(NewLoadVal); 1477 LI->eraseFromParent(); 1478 continue; 1479 } 1480 1481 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1482 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 1483 // FIXME: Remove once builder has Twine API. 1484 Value *Old = Builder.CreateLoad(NewAI, (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, (NewAI->getName()+".in").str().c_str()); 1510 Value *New = ConvertScalar_InsertValue( 1511 ConstantInt::get(User->getContext(), APVal), 1512 Old, Offset, Builder); 1513 Builder.CreateStore(New, NewAI); 1514 } 1515 MSI->eraseFromParent(); 1516 continue; 1517 } 1518 1519 // If this is a memcpy or memmove into or out of the whole allocation, we 1520 // can handle it like a load or store of the scalar type. 1521 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 1522 assert(Offset == 0 && "must be store to start of alloca"); 1523 1524 // If the source and destination are both to the same alloca, then this is 1525 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 1526 // as appropriate. 1527 AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject()); 1528 1529 if (MTI->getSource()->getUnderlyingObject() != OrigAI) { 1530 // Dest must be OrigAI, change this to be a load from the original 1531 // pointer (bitcasted), then a store to our new alloca. 1532 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 1533 Value *SrcPtr = MTI->getSource(); 1534 SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType()); 1535 1536 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 1537 SrcVal->setAlignment(MTI->getAlignment()); 1538 Builder.CreateStore(SrcVal, NewAI); 1539 } else if (MTI->getDest()->getUnderlyingObject() != OrigAI) { 1540 // Src must be OrigAI, change this to be a load from NewAI then a store 1541 // through the original dest pointer (bitcasted). 1542 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 1543 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 1544 1545 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType()); 1546 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 1547 NewStore->setAlignment(MTI->getAlignment()); 1548 } else { 1549 // Noop transfer. Src == Dst 1550 } 1551 1552 1553 MTI->eraseFromParent(); 1554 continue; 1555 } 1556 1557 // If user is a dbg info intrinsic then it is safe to remove it. 1558 if (isa<DbgInfoIntrinsic>(User)) { 1559 User->eraseFromParent(); 1560 continue; 1561 } 1562 1563 llvm_unreachable("Unsupported operation!"); 1564 } 1565} 1566 1567/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 1568/// or vector value FromVal, extracting the bits from the offset specified by 1569/// Offset. This returns the value, which is of type ToType. 1570/// 1571/// This happens when we are converting an "integer union" to a single 1572/// integer scalar, or when we are converting a "vector union" to a vector with 1573/// insert/extractelement instructions. 1574/// 1575/// Offset is an offset from the original alloca, in bits that need to be 1576/// shifted to the right. 1577Value *SROA::ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 1578 uint64_t Offset, IRBuilder<> &Builder) { 1579 // If the load is of the whole new alloca, no conversion is needed. 1580 if (FromVal->getType() == ToType && Offset == 0) 1581 return FromVal; 1582 1583 // If the result alloca is a vector type, this is either an element 1584 // access or a bitcast to another vector type of the same size. 1585 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) { 1586 if (isa<VectorType>(ToType)) 1587 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 1588 1589 // Otherwise it must be an element access. 1590 unsigned Elt = 0; 1591 if (Offset) { 1592 unsigned EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 1593 Elt = Offset/EltSize; 1594 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 1595 } 1596 // Return the element extracted out of it. 1597 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get( 1598 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp"); 1599 if (V->getType() != ToType) 1600 V = Builder.CreateBitCast(V, ToType, "tmp"); 1601 return V; 1602 } 1603 1604 // If ToType is a first class aggregate, extract out each of the pieces and 1605 // use insertvalue's to form the FCA. 1606 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 1607 const StructLayout &Layout = *TD->getStructLayout(ST); 1608 Value *Res = UndefValue::get(ST); 1609 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1610 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 1611 Offset+Layout.getElementOffsetInBits(i), 1612 Builder); 1613 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1614 } 1615 return Res; 1616 } 1617 1618 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 1619 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 1620 Value *Res = UndefValue::get(AT); 1621 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1622 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 1623 Offset+i*EltSize, Builder); 1624 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1625 } 1626 return Res; 1627 } 1628 1629 // Otherwise, this must be a union that was converted to an integer value. 1630 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 1631 1632 // If this is a big-endian system and the load is narrower than the 1633 // full alloca type, we need to do a shift to get the right bits. 1634 int ShAmt = 0; 1635 if (TD->isBigEndian()) { 1636 // On big-endian machines, the lowest bit is stored at the bit offset 1637 // from the pointer given by getTypeStoreSizeInBits. This matters for 1638 // integers with a bitwidth that is not a multiple of 8. 1639 ShAmt = TD->getTypeStoreSizeInBits(NTy) - 1640 TD->getTypeStoreSizeInBits(ToType) - Offset; 1641 } else { 1642 ShAmt = Offset; 1643 } 1644 1645 // Note: we support negative bitwidths (with shl) which are not defined. 1646 // We do this to support (f.e.) loads off the end of a structure where 1647 // only some bits are used. 1648 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 1649 FromVal = Builder.CreateLShr(FromVal, 1650 ConstantInt::get(FromVal->getType(), 1651 ShAmt), "tmp"); 1652 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 1653 FromVal = Builder.CreateShl(FromVal, 1654 ConstantInt::get(FromVal->getType(), 1655 -ShAmt), "tmp"); 1656 1657 // Finally, unconditionally truncate the integer to the right width. 1658 unsigned LIBitWidth = TD->getTypeSizeInBits(ToType); 1659 if (LIBitWidth < NTy->getBitWidth()) 1660 FromVal = 1661 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 1662 LIBitWidth), "tmp"); 1663 else if (LIBitWidth > NTy->getBitWidth()) 1664 FromVal = 1665 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 1666 LIBitWidth), "tmp"); 1667 1668 // If the result is an integer, this is a trunc or bitcast. 1669 if (isa<IntegerType>(ToType)) { 1670 // Should be done. 1671 } else if (ToType->isFloatingPoint() || isa<VectorType>(ToType)) { 1672 // Just do a bitcast, we know the sizes match up. 1673 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 1674 } else { 1675 // Otherwise must be a pointer. 1676 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 1677 } 1678 assert(FromVal->getType() == ToType && "Didn't convert right?"); 1679 return FromVal; 1680} 1681 1682 1683/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 1684/// or vector value "Old" at the offset specified by Offset. 1685/// 1686/// This happens when we are converting an "integer union" to a 1687/// single integer scalar, or when we are converting a "vector union" to a 1688/// vector with insert/extractelement instructions. 1689/// 1690/// Offset is an offset from the original alloca, in bits that need to be 1691/// shifted to the right. 1692Value *SROA::ConvertScalar_InsertValue(Value *SV, Value *Old, 1693 uint64_t Offset, IRBuilder<> &Builder) { 1694 1695 // Convert the stored type to the actual type, shift it left to insert 1696 // then 'or' into place. 1697 const Type *AllocaType = Old->getType(); 1698 LLVMContext &Context = Old->getContext(); 1699 1700 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 1701 uint64_t VecSize = TD->getTypeAllocSizeInBits(VTy); 1702 uint64_t ValSize = TD->getTypeAllocSizeInBits(SV->getType()); 1703 1704 // Changing the whole vector with memset or with an access of a different 1705 // vector type? 1706 if (ValSize == VecSize) 1707 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 1708 1709 uint64_t EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 1710 1711 // Must be an element insertion. 1712 unsigned Elt = Offset/EltSize; 1713 1714 if (SV->getType() != VTy->getElementType()) 1715 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 1716 1717 SV = Builder.CreateInsertElement(Old, SV, 1718 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt), 1719 "tmp"); 1720 return SV; 1721 } 1722 1723 // If SV is a first-class aggregate value, insert each value recursively. 1724 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 1725 const StructLayout &Layout = *TD->getStructLayout(ST); 1726 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1727 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1728 Old = ConvertScalar_InsertValue(Elt, Old, 1729 Offset+Layout.getElementOffsetInBits(i), 1730 Builder); 1731 } 1732 return Old; 1733 } 1734 1735 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 1736 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 1737 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1738 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1739 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 1740 } 1741 return Old; 1742 } 1743 1744 // If SV is a float, convert it to the appropriate integer type. 1745 // If it is a pointer, do the same. 1746 unsigned SrcWidth = TD->getTypeSizeInBits(SV->getType()); 1747 unsigned DestWidth = TD->getTypeSizeInBits(AllocaType); 1748 unsigned SrcStoreWidth = TD->getTypeStoreSizeInBits(SV->getType()); 1749 unsigned DestStoreWidth = TD->getTypeStoreSizeInBits(AllocaType); 1750 if (SV->getType()->isFloatingPoint() || isa<VectorType>(SV->getType())) 1751 SV = Builder.CreateBitCast(SV, 1752 IntegerType::get(SV->getContext(),SrcWidth), "tmp"); 1753 else if (isa<PointerType>(SV->getType())) 1754 SV = Builder.CreatePtrToInt(SV, TD->getIntPtrType(SV->getContext()), "tmp"); 1755 1756 // Zero extend or truncate the value if needed. 1757 if (SV->getType() != AllocaType) { 1758 if (SV->getType()->getPrimitiveSizeInBits() < 1759 AllocaType->getPrimitiveSizeInBits()) 1760 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 1761 else { 1762 // Truncation may be needed if storing more than the alloca can hold 1763 // (undefined behavior). 1764 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 1765 SrcWidth = DestWidth; 1766 SrcStoreWidth = DestStoreWidth; 1767 } 1768 } 1769 1770 // If this is a big-endian system and the store is narrower than the 1771 // full alloca type, we need to do a shift to get the right bits. 1772 int ShAmt = 0; 1773 if (TD->isBigEndian()) { 1774 // On big-endian machines, the lowest bit is stored at the bit offset 1775 // from the pointer given by getTypeStoreSizeInBits. This matters for 1776 // integers with a bitwidth that is not a multiple of 8. 1777 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 1778 } else { 1779 ShAmt = Offset; 1780 } 1781 1782 // Note: we support negative bitwidths (with shr) which are not defined. 1783 // We do this to support (f.e.) stores off the end of a structure where 1784 // only some bits in the structure are set. 1785 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 1786 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 1787 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 1788 ShAmt), "tmp"); 1789 Mask <<= ShAmt; 1790 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1791 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 1792 -ShAmt), "tmp"); 1793 Mask = Mask.lshr(-ShAmt); 1794 } 1795 1796 // Mask out the bits we are about to insert from the old value, and or 1797 // in the new bits. 1798 if (SrcWidth != DestWidth) { 1799 assert(DestWidth > SrcWidth); 1800 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 1801 SV = Builder.CreateOr(Old, SV, "ins"); 1802 } 1803 return SV; 1804} 1805 1806 1807 1808/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1809/// some part of a constant global variable. This intentionally only accepts 1810/// constant expressions because we don't can't rewrite arbitrary instructions. 1811static bool PointsToConstantGlobal(Value *V) { 1812 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1813 return GV->isConstant(); 1814 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1815 if (CE->getOpcode() == Instruction::BitCast || 1816 CE->getOpcode() == Instruction::GetElementPtr) 1817 return PointsToConstantGlobal(CE->getOperand(0)); 1818 return false; 1819} 1820 1821/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1822/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1823/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1824/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1825/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1826/// the alloca, and if the source pointer is a pointer to a constant global, we 1827/// can optimize this. 1828static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy, 1829 bool isOffset) { 1830 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1831 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) 1832 // Ignore non-volatile loads, they are always ok. 1833 if (!LI->isVolatile()) 1834 continue; 1835 1836 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) { 1837 // If uses of the bitcast are ok, we are ok. 1838 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 1839 return false; 1840 continue; 1841 } 1842 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) { 1843 // If the GEP has all zero indices, it doesn't offset the pointer. If it 1844 // doesn't, it does. 1845 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 1846 isOffset || !GEP->hasAllZeroIndices())) 1847 return false; 1848 continue; 1849 } 1850 1851 // If this is isn't our memcpy/memmove, reject it as something we can't 1852 // handle. 1853 if (!isa<MemTransferInst>(*UI)) 1854 return false; 1855 1856 // If we already have seen a copy, reject the second one. 1857 if (TheCopy) return false; 1858 1859 // If the pointer has been offset from the start of the alloca, we can't 1860 // safely handle this. 1861 if (isOffset) return false; 1862 1863 // If the memintrinsic isn't using the alloca as the dest, reject it. 1864 if (UI.getOperandNo() != 1) return false; 1865 1866 MemIntrinsic *MI = cast<MemIntrinsic>(*UI); 1867 1868 // If the source of the memcpy/move is not a constant global, reject it. 1869 if (!PointsToConstantGlobal(MI->getOperand(2))) 1870 return false; 1871 1872 // Otherwise, the transform is safe. Remember the copy instruction. 1873 TheCopy = MI; 1874 } 1875 return true; 1876} 1877 1878/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 1879/// modified by a copy from a constant global. If we can prove this, we can 1880/// replace any uses of the alloca with uses of the global directly. 1881Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) { 1882 Instruction *TheCopy = 0; 1883 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 1884 return TheCopy; 1885 return 0; 1886} 1887