ScalarReplAggregates.cpp revision 9e9a0d5fc26878e51a58a8b57900fcbf952c2691
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/Compiler.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 VISIBILITY_HIDDEN 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.addRequired<TargetData>(); 72 AU.setPreservesCFG(); 73 } 74 75 private: 76 TargetData *TD; 77 78 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures 79 /// information about the uses. All these fields are initialized to false 80 /// and set to true when something is learned. 81 struct AllocaInfo { 82 /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 83 bool isUnsafe : 1; 84 85 /// needsCleanup - This is set to true if there is some use of the alloca 86 /// that requires cleanup. 87 bool needsCleanup : 1; 88 89 /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 90 bool isMemCpySrc : 1; 91 92 /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 93 bool isMemCpyDst : 1; 94 95 AllocaInfo() 96 : isUnsafe(false), needsCleanup(false), 97 isMemCpySrc(false), isMemCpyDst(false) {} 98 }; 99 100 unsigned SRThreshold; 101 102 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; } 103 104 int isSafeAllocaToScalarRepl(AllocationInst *AI); 105 106 void isSafeUseOfAllocation(Instruction *User, AllocationInst *AI, 107 AllocaInfo &Info); 108 void isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI, 109 AllocaInfo &Info); 110 void isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI, 111 unsigned OpNo, AllocaInfo &Info); 112 void isSafeUseOfBitCastedAllocation(BitCastInst *User, AllocationInst *AI, 113 AllocaInfo &Info); 114 115 void DoScalarReplacement(AllocationInst *AI, 116 std::vector<AllocationInst*> &WorkList); 117 void CleanupGEP(GetElementPtrInst *GEP); 118 void CleanupAllocaUsers(AllocationInst *AI); 119 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocationInst *Base); 120 121 void RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI, 122 SmallVector<AllocaInst*, 32> &NewElts); 123 124 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst, 125 AllocationInst *AI, 126 SmallVector<AllocaInst*, 32> &NewElts); 127 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocationInst *AI, 128 SmallVector<AllocaInst*, 32> &NewElts); 129 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI, 130 SmallVector<AllocaInst*, 32> &NewElts); 131 132 bool CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy, 133 bool &SawVec, uint64_t Offset, unsigned AllocaSize); 134 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); 135 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType, 136 uint64_t Offset, IRBuilder<> &Builder); 137 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 138 uint64_t Offset, IRBuilder<> &Builder); 139 static Instruction *isOnlyCopiedFromConstantGlobal(AllocationInst *AI); 140 }; 141} 142 143char SROA::ID = 0; 144static RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates"); 145 146// Public interface to the ScalarReplAggregates pass 147FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) { 148 return new SROA(Threshold); 149} 150 151 152bool SROA::runOnFunction(Function &F) { 153 TD = &getAnalysis<TargetData>(); 154 155 bool Changed = performPromotion(F); 156 while (1) { 157 bool LocalChange = performScalarRepl(F); 158 if (!LocalChange) break; // No need to repromote if no scalarrepl 159 Changed = true; 160 LocalChange = performPromotion(F); 161 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 162 } 163 164 return Changed; 165} 166 167 168bool SROA::performPromotion(Function &F) { 169 std::vector<AllocaInst*> Allocas; 170 DominatorTree &DT = getAnalysis<DominatorTree>(); 171 DominanceFrontier &DF = getAnalysis<DominanceFrontier>(); 172 173 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 174 175 bool Changed = false; 176 177 while (1) { 178 Allocas.clear(); 179 180 // Find allocas that are safe to promote, by looking at all instructions in 181 // the entry node 182 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 183 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 184 if (isAllocaPromotable(AI)) 185 Allocas.push_back(AI); 186 187 if (Allocas.empty()) break; 188 189 PromoteMemToReg(Allocas, DT, DF, F.getContext()); 190 NumPromoted += Allocas.size(); 191 Changed = true; 192 } 193 194 return Changed; 195} 196 197/// getNumSAElements - Return the number of elements in the specific struct or 198/// array. 199static uint64_t getNumSAElements(const Type *T) { 200 if (const StructType *ST = dyn_cast<StructType>(T)) 201 return ST->getNumElements(); 202 return cast<ArrayType>(T)->getNumElements(); 203} 204 205// performScalarRepl - This algorithm is a simple worklist driven algorithm, 206// which runs on all of the malloc/alloca instructions in the function, removing 207// them if they are only used by getelementptr instructions. 208// 209bool SROA::performScalarRepl(Function &F) { 210 std::vector<AllocationInst*> WorkList; 211 212 // Scan the entry basic block, adding any alloca's and mallocs to the worklist 213 BasicBlock &BB = F.getEntryBlock(); 214 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 215 if (AllocationInst *A = dyn_cast<AllocationInst>(I)) 216 WorkList.push_back(A); 217 218 // Process the worklist 219 bool Changed = false; 220 while (!WorkList.empty()) { 221 AllocationInst *AI = WorkList.back(); 222 WorkList.pop_back(); 223 224 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 225 // with unused elements. 226 if (AI->use_empty()) { 227 AI->eraseFromParent(); 228 continue; 229 } 230 231 // If this alloca is impossible for us to promote, reject it early. 232 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 233 continue; 234 235 // Check to see if this allocation is only modified by a memcpy/memmove from 236 // a constant global. If this is the case, we can change all users to use 237 // the constant global instead. This is commonly produced by the CFE by 238 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 239 // is only subsequently read. 240 if (Instruction *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 241 DOUT << "Found alloca equal to global: " << *AI; 242 DOUT << " memcpy = " << *TheCopy; 243 Constant *TheSrc = cast<Constant>(TheCopy->getOperand(2)); 244 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 245 TheCopy->eraseFromParent(); // Don't mutate the global. 246 AI->eraseFromParent(); 247 ++NumGlobals; 248 Changed = true; 249 continue; 250 } 251 252 // Check to see if we can perform the core SROA transformation. We cannot 253 // transform the allocation instruction if it is an array allocation 254 // (allocations OF arrays are ok though), and an allocation of a scalar 255 // value cannot be decomposed at all. 256 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 257 258 // Do not promote any struct whose size is too big. 259 if (AllocaSize > SRThreshold) continue; 260 261 if ((isa<StructType>(AI->getAllocatedType()) || 262 isa<ArrayType>(AI->getAllocatedType())) && 263 // Do not promote any struct into more than "32" separate vars. 264 getNumSAElements(AI->getAllocatedType()) <= SRThreshold/4) { 265 // Check that all of the users of the allocation are capable of being 266 // transformed. 267 switch (isSafeAllocaToScalarRepl(AI)) { 268 default: llvm_unreachable("Unexpected value!"); 269 case 0: // Not safe to scalar replace. 270 break; 271 case 1: // Safe, but requires cleanup/canonicalizations first 272 CleanupAllocaUsers(AI); 273 // FALL THROUGH. 274 case 3: // Safe to scalar replace. 275 DoScalarReplacement(AI, WorkList); 276 Changed = true; 277 continue; 278 } 279 } 280 281 // If we can turn this aggregate value (potentially with casts) into a 282 // simple scalar value that can be mem2reg'd into a register value. 283 // IsNotTrivial tracks whether this is something that mem2reg could have 284 // promoted itself. If so, we don't want to transform it needlessly. Note 285 // that we can't just check based on the type: the alloca may be of an i32 286 // but that has pointer arithmetic to set byte 3 of it or something. 287 bool IsNotTrivial = false; 288 const Type *VectorTy = 0; 289 bool HadAVector = false; 290 if (CanConvertToScalar(AI, IsNotTrivial, VectorTy, HadAVector, 291 0, unsigned(AllocaSize)) && IsNotTrivial) { 292 AllocaInst *NewAI; 293 // If we were able to find a vector type that can handle this with 294 // insert/extract elements, and if there was at least one use that had 295 // a vector type, promote this to a vector. We don't want to promote 296 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 297 // we just get a lot of insert/extracts. If at least one vector is 298 // involved, then we probably really do have a union of vector/array. 299 if (VectorTy && isa<VectorType>(VectorTy) && HadAVector) { 300 DOUT << "CONVERT TO VECTOR: " << *AI << " TYPE = " << *VectorTy <<"\n"; 301 302 // Create and insert the vector alloca. 303 NewAI = new AllocaInst(VectorTy, 0, "", AI->getParent()->begin()); 304 ConvertUsesToScalar(AI, NewAI, 0); 305 } else { 306 DOUT << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"; 307 308 // Create and insert the integer alloca. 309 const Type *NewTy = IntegerType::get(AllocaSize*8); 310 NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 311 ConvertUsesToScalar(AI, NewAI, 0); 312 } 313 NewAI->takeName(AI); 314 AI->eraseFromParent(); 315 ++NumConverted; 316 Changed = true; 317 continue; 318 } 319 320 // Otherwise, couldn't process this alloca. 321 } 322 323 return Changed; 324} 325 326/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 327/// predicate, do SROA now. 328void SROA::DoScalarReplacement(AllocationInst *AI, 329 std::vector<AllocationInst*> &WorkList) { 330 DOUT << "Found inst to SROA: " << *AI; 331 SmallVector<AllocaInst*, 32> ElementAllocas; 332 LLVMContext &Context = AI->getContext(); 333 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 334 ElementAllocas.reserve(ST->getNumContainedTypes()); 335 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 336 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 337 AI->getAlignment(), 338 AI->getName() + "." + Twine(i), AI); 339 ElementAllocas.push_back(NA); 340 WorkList.push_back(NA); // Add to worklist for recursive processing 341 } 342 } else { 343 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 344 ElementAllocas.reserve(AT->getNumElements()); 345 const Type *ElTy = AT->getElementType(); 346 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 347 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 348 AI->getName() + "." + Twine(i), AI); 349 ElementAllocas.push_back(NA); 350 WorkList.push_back(NA); // Add to worklist for recursive processing 351 } 352 } 353 354 // Now that we have created the alloca instructions that we want to use, 355 // expand the getelementptr instructions to use them. 356 // 357 while (!AI->use_empty()) { 358 Instruction *User = cast<Instruction>(AI->use_back()); 359 if (BitCastInst *BCInst = dyn_cast<BitCastInst>(User)) { 360 RewriteBitCastUserOfAlloca(BCInst, AI, ElementAllocas); 361 BCInst->eraseFromParent(); 362 continue; 363 } 364 365 // Replace: 366 // %res = load { i32, i32 }* %alloc 367 // with: 368 // %load.0 = load i32* %alloc.0 369 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 370 // %load.1 = load i32* %alloc.1 371 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 372 // (Also works for arrays instead of structs) 373 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 374 Value *Insert = UndefValue::get(LI->getType()); 375 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) { 376 Value *Load = new LoadInst(ElementAllocas[i], "load", LI); 377 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 378 } 379 LI->replaceAllUsesWith(Insert); 380 LI->eraseFromParent(); 381 continue; 382 } 383 384 // Replace: 385 // store { i32, i32 } %val, { i32, i32 }* %alloc 386 // with: 387 // %val.0 = extractvalue { i32, i32 } %val, 0 388 // store i32 %val.0, i32* %alloc.0 389 // %val.1 = extractvalue { i32, i32 } %val, 1 390 // store i32 %val.1, i32* %alloc.1 391 // (Also works for arrays instead of structs) 392 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 393 Value *Val = SI->getOperand(0); 394 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) { 395 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 396 new StoreInst(Extract, ElementAllocas[i], SI); 397 } 398 SI->eraseFromParent(); 399 continue; 400 } 401 402 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User); 403 // We now know that the GEP is of the form: GEP <ptr>, 0, <cst> 404 unsigned Idx = 405 (unsigned)cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue(); 406 407 assert(Idx < ElementAllocas.size() && "Index out of range?"); 408 AllocaInst *AllocaToUse = ElementAllocas[Idx]; 409 410 Value *RepValue; 411 if (GEPI->getNumOperands() == 3) { 412 // Do not insert a new getelementptr instruction with zero indices, only 413 // to have it optimized out later. 414 RepValue = AllocaToUse; 415 } else { 416 // We are indexing deeply into the structure, so we still need a 417 // getelement ptr instruction to finish the indexing. This may be 418 // expanded itself once the worklist is rerun. 419 // 420 SmallVector<Value*, 8> NewArgs; 421 NewArgs.push_back(Context.getNullValue(Type::Int32Ty)); 422 NewArgs.append(GEPI->op_begin()+3, GEPI->op_end()); 423 RepValue = GetElementPtrInst::Create(AllocaToUse, NewArgs.begin(), 424 NewArgs.end(), "", GEPI); 425 RepValue->takeName(GEPI); 426 } 427 428 // If this GEP is to the start of the aggregate, check for memcpys. 429 if (Idx == 0 && GEPI->hasAllZeroIndices()) 430 RewriteBitCastUserOfAlloca(GEPI, AI, ElementAllocas); 431 432 // Move all of the users over to the new GEP. 433 GEPI->replaceAllUsesWith(RepValue); 434 // Delete the old GEP 435 GEPI->eraseFromParent(); 436 } 437 438 // Finally, delete the Alloca instruction 439 AI->eraseFromParent(); 440 NumReplaced++; 441} 442 443 444/// isSafeElementUse - Check to see if this use is an allowed use for a 445/// getelementptr instruction of an array aggregate allocation. isFirstElt 446/// indicates whether Ptr is known to the start of the aggregate. 447/// 448void SROA::isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI, 449 AllocaInfo &Info) { 450 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 451 I != E; ++I) { 452 Instruction *User = cast<Instruction>(*I); 453 switch (User->getOpcode()) { 454 case Instruction::Load: break; 455 case Instruction::Store: 456 // Store is ok if storing INTO the pointer, not storing the pointer 457 if (User->getOperand(0) == Ptr) return MarkUnsafe(Info); 458 break; 459 case Instruction::GetElementPtr: { 460 GetElementPtrInst *GEP = cast<GetElementPtrInst>(User); 461 bool AreAllZeroIndices = isFirstElt; 462 if (GEP->getNumOperands() > 1) { 463 if (!isa<ConstantInt>(GEP->getOperand(1)) || 464 !cast<ConstantInt>(GEP->getOperand(1))->isZero()) 465 // Using pointer arithmetic to navigate the array. 466 return MarkUnsafe(Info); 467 468 if (AreAllZeroIndices) 469 AreAllZeroIndices = GEP->hasAllZeroIndices(); 470 } 471 isSafeElementUse(GEP, AreAllZeroIndices, AI, Info); 472 if (Info.isUnsafe) return; 473 break; 474 } 475 case Instruction::BitCast: 476 if (isFirstElt) { 477 isSafeUseOfBitCastedAllocation(cast<BitCastInst>(User), AI, Info); 478 if (Info.isUnsafe) return; 479 break; 480 } 481 DOUT << " Transformation preventing inst: " << *User; 482 return MarkUnsafe(Info); 483 case Instruction::Call: 484 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 485 if (isFirstElt) { 486 isSafeMemIntrinsicOnAllocation(MI, AI, I.getOperandNo(), Info); 487 if (Info.isUnsafe) return; 488 break; 489 } 490 } 491 DOUT << " Transformation preventing inst: " << *User; 492 return MarkUnsafe(Info); 493 default: 494 DOUT << " Transformation preventing inst: " << *User; 495 return MarkUnsafe(Info); 496 } 497 } 498 return; // All users look ok :) 499} 500 501/// AllUsersAreLoads - Return true if all users of this value are loads. 502static bool AllUsersAreLoads(Value *Ptr) { 503 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 504 I != E; ++I) 505 if (cast<Instruction>(*I)->getOpcode() != Instruction::Load) 506 return false; 507 return true; 508} 509 510/// isSafeUseOfAllocation - Check to see if this user is an allowed use for an 511/// aggregate allocation. 512/// 513void SROA::isSafeUseOfAllocation(Instruction *User, AllocationInst *AI, 514 AllocaInfo &Info) { 515 LLVMContext &Context = User->getContext(); 516 if (BitCastInst *C = dyn_cast<BitCastInst>(User)) 517 return isSafeUseOfBitCastedAllocation(C, AI, Info); 518 519 if (LoadInst *LI = dyn_cast<LoadInst>(User)) 520 if (!LI->isVolatile()) 521 return;// Loads (returning a first class aggregrate) are always rewritable 522 523 if (StoreInst *SI = dyn_cast<StoreInst>(User)) 524 if (!SI->isVolatile() && SI->getOperand(0) != AI) 525 return;// Store is ok if storing INTO the pointer, not storing the pointer 526 527 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User); 528 if (GEPI == 0) 529 return MarkUnsafe(Info); 530 531 gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI); 532 533 // The GEP is not safe to transform if not of the form "GEP <ptr>, 0, <cst>". 534 if (I == E || 535 I.getOperand() != Context.getNullValue(I.getOperand()->getType())) { 536 return MarkUnsafe(Info); 537 } 538 539 ++I; 540 if (I == E) return MarkUnsafe(Info); // ran out of GEP indices?? 541 542 bool IsAllZeroIndices = true; 543 544 // If the first index is a non-constant index into an array, see if we can 545 // handle it as a special case. 546 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 547 if (!isa<ConstantInt>(I.getOperand())) { 548 IsAllZeroIndices = 0; 549 uint64_t NumElements = AT->getNumElements(); 550 551 // If this is an array index and the index is not constant, we cannot 552 // promote... that is unless the array has exactly one or two elements in 553 // it, in which case we CAN promote it, but we have to canonicalize this 554 // out if this is the only problem. 555 if ((NumElements == 1 || NumElements == 2) && 556 AllUsersAreLoads(GEPI)) { 557 Info.needsCleanup = true; 558 return; // Canonicalization required! 559 } 560 return MarkUnsafe(Info); 561 } 562 } 563 564 // Walk through the GEP type indices, checking the types that this indexes 565 // into. 566 for (; I != E; ++I) { 567 // Ignore struct elements, no extra checking needed for these. 568 if (isa<StructType>(*I)) 569 continue; 570 571 ConstantInt *IdxVal = dyn_cast<ConstantInt>(I.getOperand()); 572 if (!IdxVal) return MarkUnsafe(Info); 573 574 // Are all indices still zero? 575 IsAllZeroIndices &= IdxVal->isZero(); 576 577 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 578 // This GEP indexes an array. Verify that this is an in-range constant 579 // integer. Specifically, consider A[0][i]. We cannot know that the user 580 // isn't doing invalid things like allowing i to index an out-of-range 581 // subscript that accesses A[1]. Because of this, we have to reject SROA 582 // of any accesses into structs where any of the components are variables. 583 if (IdxVal->getZExtValue() >= AT->getNumElements()) 584 return MarkUnsafe(Info); 585 } else if (const VectorType *VT = dyn_cast<VectorType>(*I)) { 586 if (IdxVal->getZExtValue() >= VT->getNumElements()) 587 return MarkUnsafe(Info); 588 } 589 } 590 591 // If there are any non-simple uses of this getelementptr, make sure to reject 592 // them. 593 return isSafeElementUse(GEPI, IsAllZeroIndices, AI, Info); 594} 595 596/// isSafeMemIntrinsicOnAllocation - Return true if the specified memory 597/// intrinsic can be promoted by SROA. At this point, we know that the operand 598/// of the memintrinsic is a pointer to the beginning of the allocation. 599void SROA::isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI, 600 unsigned OpNo, AllocaInfo &Info) { 601 // If not constant length, give up. 602 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 603 if (!Length) return MarkUnsafe(Info); 604 605 // If not the whole aggregate, give up. 606 if (Length->getZExtValue() != 607 TD->getTypeAllocSize(AI->getType()->getElementType())) 608 return MarkUnsafe(Info); 609 610 // We only know about memcpy/memset/memmove. 611 if (!isa<MemIntrinsic>(MI)) 612 return MarkUnsafe(Info); 613 614 // Otherwise, we can transform it. Determine whether this is a memcpy/set 615 // into or out of the aggregate. 616 if (OpNo == 1) 617 Info.isMemCpyDst = true; 618 else { 619 assert(OpNo == 2); 620 Info.isMemCpySrc = true; 621 } 622} 623 624/// isSafeUseOfBitCastedAllocation - Return true if all users of this bitcast 625/// are 626void SROA::isSafeUseOfBitCastedAllocation(BitCastInst *BC, AllocationInst *AI, 627 AllocaInfo &Info) { 628 for (Value::use_iterator UI = BC->use_begin(), E = BC->use_end(); 629 UI != E; ++UI) { 630 if (BitCastInst *BCU = dyn_cast<BitCastInst>(UI)) { 631 isSafeUseOfBitCastedAllocation(BCU, AI, Info); 632 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) { 633 isSafeMemIntrinsicOnAllocation(MI, AI, UI.getOperandNo(), Info); 634 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI)) { 635 if (SI->isVolatile()) 636 return MarkUnsafe(Info); 637 638 // If storing the entire alloca in one chunk through a bitcasted pointer 639 // to integer, we can transform it. This happens (for example) when you 640 // cast a {i32,i32}* to i64* and store through it. This is similar to the 641 // memcpy case and occurs in various "byval" cases and emulated memcpys. 642 if (isa<IntegerType>(SI->getOperand(0)->getType()) && 643 TD->getTypeAllocSize(SI->getOperand(0)->getType()) == 644 TD->getTypeAllocSize(AI->getType()->getElementType())) { 645 Info.isMemCpyDst = true; 646 continue; 647 } 648 return MarkUnsafe(Info); 649 } else if (LoadInst *LI = dyn_cast<LoadInst>(UI)) { 650 if (LI->isVolatile()) 651 return MarkUnsafe(Info); 652 653 // If loading the entire alloca in one chunk through a bitcasted pointer 654 // to integer, we can transform it. This happens (for example) when you 655 // cast a {i32,i32}* to i64* and load through it. This is similar to the 656 // memcpy case and occurs in various "byval" cases and emulated memcpys. 657 if (isa<IntegerType>(LI->getType()) && 658 TD->getTypeAllocSize(LI->getType()) == 659 TD->getTypeAllocSize(AI->getType()->getElementType())) { 660 Info.isMemCpySrc = true; 661 continue; 662 } 663 return MarkUnsafe(Info); 664 } else if (isa<DbgInfoIntrinsic>(UI)) { 665 // If one user is DbgInfoIntrinsic then check if all users are 666 // DbgInfoIntrinsics. 667 if (OnlyUsedByDbgInfoIntrinsics(BC)) { 668 Info.needsCleanup = true; 669 return; 670 } 671 else 672 MarkUnsafe(Info); 673 } 674 else { 675 return MarkUnsafe(Info); 676 } 677 if (Info.isUnsafe) return; 678 } 679} 680 681/// RewriteBitCastUserOfAlloca - BCInst (transitively) bitcasts AI, or indexes 682/// to its first element. Transform users of the cast to use the new values 683/// instead. 684void SROA::RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI, 685 SmallVector<AllocaInst*, 32> &NewElts) { 686 Value::use_iterator UI = BCInst->use_begin(), UE = BCInst->use_end(); 687 while (UI != UE) { 688 Instruction *User = cast<Instruction>(*UI++); 689 if (BitCastInst *BCU = dyn_cast<BitCastInst>(User)) { 690 RewriteBitCastUserOfAlloca(BCU, AI, NewElts); 691 if (BCU->use_empty()) BCU->eraseFromParent(); 692 continue; 693 } 694 695 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 696 // This must be memcpy/memmove/memset of the entire aggregate. 697 // Split into one per element. 698 RewriteMemIntrinUserOfAlloca(MI, BCInst, AI, NewElts); 699 continue; 700 } 701 702 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 703 // If this is a store of the entire alloca from an integer, rewrite it. 704 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 705 continue; 706 } 707 708 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 709 // If this is a load of the entire alloca to an integer, rewrite it. 710 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 711 continue; 712 } 713 714 // Otherwise it must be some other user of a gep of the first pointer. Just 715 // leave these alone. 716 continue; 717 } 718} 719 720/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 721/// Rewrite it to copy or set the elements of the scalarized memory. 722void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst, 723 AllocationInst *AI, 724 SmallVector<AllocaInst*, 32> &NewElts) { 725 726 // If this is a memcpy/memmove, construct the other pointer as the 727 // appropriate type. The "Other" pointer is the pointer that goes to memory 728 // that doesn't have anything to do with the alloca that we are promoting. For 729 // memset, this Value* stays null. 730 Value *OtherPtr = 0; 731 LLVMContext &Context = MI->getContext(); 732 unsigned MemAlignment = MI->getAlignment(); 733 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 734 if (BCInst == MTI->getRawDest()) 735 OtherPtr = MTI->getRawSource(); 736 else { 737 assert(BCInst == MTI->getRawSource()); 738 OtherPtr = MTI->getRawDest(); 739 } 740 } 741 742 // If there is an other pointer, we want to convert it to the same pointer 743 // type as AI has, so we can GEP through it safely. 744 if (OtherPtr) { 745 // It is likely that OtherPtr is a bitcast, if so, remove it. 746 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr)) 747 OtherPtr = BC->getOperand(0); 748 // All zero GEPs are effectively bitcasts. 749 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr)) 750 if (GEP->hasAllZeroIndices()) 751 OtherPtr = GEP->getOperand(0); 752 753 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr)) 754 if (BCE->getOpcode() == Instruction::BitCast) 755 OtherPtr = BCE->getOperand(0); 756 757 // If the pointer is not the right type, insert a bitcast to the right 758 // type. 759 if (OtherPtr->getType() != AI->getType()) 760 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(), 761 MI); 762 } 763 764 // Process each element of the aggregate. 765 Value *TheFn = MI->getOperand(0); 766 const Type *BytePtrTy = MI->getRawDest()->getType(); 767 bool SROADest = MI->getRawDest() == BCInst; 768 769 Constant *Zero = Context.getNullValue(Type::Int32Ty); 770 771 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 772 // If this is a memcpy/memmove, emit a GEP of the other element address. 773 Value *OtherElt = 0; 774 unsigned OtherEltAlign = MemAlignment; 775 776 if (OtherPtr) { 777 Value *Idx[2] = { Zero, ConstantInt::get(Type::Int32Ty, i) }; 778 OtherElt = GetElementPtrInst::Create(OtherPtr, Idx, Idx + 2, 779 OtherPtr->getNameStr()+"."+Twine(i), 780 MI); 781 uint64_t EltOffset; 782 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 783 if (const StructType *ST = 784 dyn_cast<StructType>(OtherPtrTy->getElementType())) { 785 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 786 } else { 787 const Type *EltTy = 788 cast<SequentialType>(OtherPtr->getType())->getElementType(); 789 EltOffset = TD->getTypeAllocSize(EltTy)*i; 790 } 791 792 // The alignment of the other pointer is the guaranteed alignment of the 793 // element, which is affected by both the known alignment of the whole 794 // mem intrinsic and the alignment of the element. If the alignment of 795 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 796 // known alignment is just 4 bytes. 797 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 798 } 799 800 Value *EltPtr = NewElts[i]; 801 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 802 803 // If we got down to a scalar, insert a load or store as appropriate. 804 if (EltTy->isSingleValueType()) { 805 if (isa<MemTransferInst>(MI)) { 806 if (SROADest) { 807 // From Other to Alloca. 808 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 809 new StoreInst(Elt, EltPtr, MI); 810 } else { 811 // From Alloca to Other. 812 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 813 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 814 } 815 continue; 816 } 817 assert(isa<MemSetInst>(MI)); 818 819 // If the stored element is zero (common case), just store a null 820 // constant. 821 Constant *StoreVal; 822 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) { 823 if (CI->isZero()) { 824 StoreVal = Context.getNullValue(EltTy); // 0.0, null, 0, <0,0> 825 } else { 826 // If EltTy is a vector type, get the element type. 827 const Type *ValTy = EltTy->getScalarType(); 828 829 // Construct an integer with the right value. 830 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 831 APInt OneVal(EltSize, CI->getZExtValue()); 832 APInt TotalVal(OneVal); 833 // Set each byte. 834 for (unsigned i = 0; 8*i < EltSize; ++i) { 835 TotalVal = TotalVal.shl(8); 836 TotalVal |= OneVal; 837 } 838 839 // Convert the integer value to the appropriate type. 840 StoreVal = ConstantInt::get(Context, TotalVal); 841 if (isa<PointerType>(ValTy)) 842 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 843 else if (ValTy->isFloatingPoint()) 844 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 845 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 846 847 // If the requested value was a vector constant, create it. 848 if (EltTy != ValTy) { 849 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 850 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 851 StoreVal = ConstantVector::get(&Elts[0], NumElts); 852 } 853 } 854 new StoreInst(StoreVal, EltPtr, MI); 855 continue; 856 } 857 // Otherwise, if we're storing a byte variable, use a memset call for 858 // this element. 859 } 860 861 // Cast the element pointer to BytePtrTy. 862 if (EltPtr->getType() != BytePtrTy) 863 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI); 864 865 // Cast the other pointer (if we have one) to BytePtrTy. 866 if (OtherElt && OtherElt->getType() != BytePtrTy) 867 OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(), 868 MI); 869 870 unsigned EltSize = TD->getTypeAllocSize(EltTy); 871 872 // Finally, insert the meminst for this element. 873 if (isa<MemTransferInst>(MI)) { 874 Value *Ops[] = { 875 SROADest ? EltPtr : OtherElt, // Dest ptr 876 SROADest ? OtherElt : EltPtr, // Src ptr 877 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 878 ConstantInt::get(Type::Int32Ty, OtherEltAlign) // Align 879 }; 880 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 881 } else { 882 assert(isa<MemSetInst>(MI)); 883 Value *Ops[] = { 884 EltPtr, MI->getOperand(2), // Dest, Value, 885 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 886 Zero // Align 887 }; 888 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 889 } 890 } 891 MI->eraseFromParent(); 892} 893 894/// RewriteStoreUserOfWholeAlloca - We found an store of an integer that 895/// overwrites the entire allocation. Extract out the pieces of the stored 896/// integer and store them individually. 897void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, 898 AllocationInst *AI, 899 SmallVector<AllocaInst*, 32> &NewElts){ 900 // Extract each element out of the integer according to its structure offset 901 // and store the element value to the individual alloca. 902 Value *SrcVal = SI->getOperand(0); 903 const Type *AllocaEltTy = AI->getType()->getElementType(); 904 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 905 906 // If this isn't a store of an integer to the whole alloca, it may be a store 907 // to the first element. Just ignore the store in this case and normal SROA 908 // will handle it. 909 if (!isa<IntegerType>(SrcVal->getType()) || 910 TD->getTypeAllocSizeInBits(SrcVal->getType()) != AllocaSizeBits) 911 return; 912 // Handle tail padding by extending the operand 913 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 914 SrcVal = new ZExtInst(SrcVal, 915 IntegerType::get(AllocaSizeBits), "", SI); 916 917 DOUT << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << *SI; 918 919 // There are two forms here: AI could be an array or struct. Both cases 920 // have different ways to compute the element offset. 921 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 922 const StructLayout *Layout = TD->getStructLayout(EltSTy); 923 924 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 925 // Get the number of bits to shift SrcVal to get the value. 926 const Type *FieldTy = EltSTy->getElementType(i); 927 uint64_t Shift = Layout->getElementOffsetInBits(i); 928 929 if (TD->isBigEndian()) 930 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 931 932 Value *EltVal = SrcVal; 933 if (Shift) { 934 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 935 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 936 "sroa.store.elt", SI); 937 } 938 939 // Truncate down to an integer of the right size. 940 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 941 942 // Ignore zero sized fields like {}, they obviously contain no data. 943 if (FieldSizeBits == 0) continue; 944 945 if (FieldSizeBits != AllocaSizeBits) 946 EltVal = new TruncInst(EltVal, 947 IntegerType::get(FieldSizeBits), "", SI); 948 Value *DestField = NewElts[i]; 949 if (EltVal->getType() == FieldTy) { 950 // Storing to an integer field of this size, just do it. 951 } else if (FieldTy->isFloatingPoint() || isa<VectorType>(FieldTy)) { 952 // Bitcast to the right element type (for fp/vector values). 953 EltVal = new BitCastInst(EltVal, FieldTy, "", SI); 954 } else { 955 // Otherwise, bitcast the dest pointer (for aggregates). 956 DestField = new BitCastInst(DestField, 957 PointerType::getUnqual(EltVal->getType()), 958 "", SI); 959 } 960 new StoreInst(EltVal, DestField, SI); 961 } 962 963 } else { 964 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 965 const Type *ArrayEltTy = ATy->getElementType(); 966 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 967 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 968 969 uint64_t Shift; 970 971 if (TD->isBigEndian()) 972 Shift = AllocaSizeBits-ElementOffset; 973 else 974 Shift = 0; 975 976 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 977 // Ignore zero sized fields like {}, they obviously contain no data. 978 if (ElementSizeBits == 0) continue; 979 980 Value *EltVal = SrcVal; 981 if (Shift) { 982 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 983 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 984 "sroa.store.elt", SI); 985 } 986 987 // Truncate down to an integer of the right size. 988 if (ElementSizeBits != AllocaSizeBits) 989 EltVal = new TruncInst(EltVal, 990 IntegerType::get(ElementSizeBits),"",SI); 991 Value *DestField = NewElts[i]; 992 if (EltVal->getType() == ArrayEltTy) { 993 // Storing to an integer field of this size, just do it. 994 } else if (ArrayEltTy->isFloatingPoint() || isa<VectorType>(ArrayEltTy)) { 995 // Bitcast to the right element type (for fp/vector values). 996 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI); 997 } else { 998 // Otherwise, bitcast the dest pointer (for aggregates). 999 DestField = new BitCastInst(DestField, 1000 PointerType::getUnqual(EltVal->getType()), 1001 "", SI); 1002 } 1003 new StoreInst(EltVal, DestField, SI); 1004 1005 if (TD->isBigEndian()) 1006 Shift -= ElementOffset; 1007 else 1008 Shift += ElementOffset; 1009 } 1010 } 1011 1012 SI->eraseFromParent(); 1013} 1014 1015/// RewriteLoadUserOfWholeAlloca - We found an load of the entire allocation to 1016/// an integer. Load the individual pieces to form the aggregate value. 1017void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI, 1018 SmallVector<AllocaInst*, 32> &NewElts) { 1019 // Extract each element out of the NewElts according to its structure offset 1020 // and form the result value. 1021 const Type *AllocaEltTy = AI->getType()->getElementType(); 1022 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1023 1024 // If this isn't a load of the whole alloca to an integer, it may be a load 1025 // of the first element. Just ignore the load in this case and normal SROA 1026 // will handle it. 1027 if (!isa<IntegerType>(LI->getType()) || 1028 TD->getTypeAllocSizeInBits(LI->getType()) != AllocaSizeBits) 1029 return; 1030 1031 DOUT << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << *LI; 1032 1033 // There are two forms here: AI could be an array or struct. Both cases 1034 // have different ways to compute the element offset. 1035 const StructLayout *Layout = 0; 1036 uint64_t ArrayEltBitOffset = 0; 1037 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1038 Layout = TD->getStructLayout(EltSTy); 1039 } else { 1040 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 1041 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1042 } 1043 1044 LLVMContext &Context = LI->getContext(); 1045 1046 Value *ResultVal = 1047 Context.getNullValue(IntegerType::get(AllocaSizeBits)); 1048 1049 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1050 // Load the value from the alloca. If the NewElt is an aggregate, cast 1051 // the pointer to an integer of the same size before doing the load. 1052 Value *SrcField = NewElts[i]; 1053 const Type *FieldTy = 1054 cast<PointerType>(SrcField->getType())->getElementType(); 1055 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1056 1057 // Ignore zero sized fields like {}, they obviously contain no data. 1058 if (FieldSizeBits == 0) continue; 1059 1060 const IntegerType *FieldIntTy = IntegerType::get(FieldSizeBits); 1061 if (!isa<IntegerType>(FieldTy) && !FieldTy->isFloatingPoint() && 1062 !isa<VectorType>(FieldTy)) 1063 SrcField = new BitCastInst(SrcField, 1064 PointerType::getUnqual(FieldIntTy), 1065 "", LI); 1066 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 1067 1068 // If SrcField is a fp or vector of the right size but that isn't an 1069 // integer type, bitcast to an integer so we can shift it. 1070 if (SrcField->getType() != FieldIntTy) 1071 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 1072 1073 // Zero extend the field to be the same size as the final alloca so that 1074 // we can shift and insert it. 1075 if (SrcField->getType() != ResultVal->getType()) 1076 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 1077 1078 // Determine the number of bits to shift SrcField. 1079 uint64_t Shift; 1080 if (Layout) // Struct case. 1081 Shift = Layout->getElementOffsetInBits(i); 1082 else // Array case. 1083 Shift = i*ArrayEltBitOffset; 1084 1085 if (TD->isBigEndian()) 1086 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 1087 1088 if (Shift) { 1089 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 1090 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 1091 } 1092 1093 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 1094 } 1095 1096 // Handle tail padding by truncating the result 1097 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 1098 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 1099 1100 LI->replaceAllUsesWith(ResultVal); 1101 LI->eraseFromParent(); 1102} 1103 1104 1105/// HasPadding - Return true if the specified type has any structure or 1106/// alignment padding, false otherwise. 1107static bool HasPadding(const Type *Ty, const TargetData &TD) { 1108 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 1109 const StructLayout *SL = TD.getStructLayout(STy); 1110 unsigned PrevFieldBitOffset = 0; 1111 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1112 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 1113 1114 // Padding in sub-elements? 1115 if (HasPadding(STy->getElementType(i), TD)) 1116 return true; 1117 1118 // Check to see if there is any padding between this element and the 1119 // previous one. 1120 if (i) { 1121 unsigned PrevFieldEnd = 1122 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 1123 if (PrevFieldEnd < FieldBitOffset) 1124 return true; 1125 } 1126 1127 PrevFieldBitOffset = FieldBitOffset; 1128 } 1129 1130 // Check for tail padding. 1131 if (unsigned EltCount = STy->getNumElements()) { 1132 unsigned PrevFieldEnd = PrevFieldBitOffset + 1133 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 1134 if (PrevFieldEnd < SL->getSizeInBits()) 1135 return true; 1136 } 1137 1138 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1139 return HasPadding(ATy->getElementType(), TD); 1140 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) { 1141 return HasPadding(VTy->getElementType(), TD); 1142 } 1143 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 1144} 1145 1146/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 1147/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 1148/// or 1 if safe after canonicalization has been performed. 1149/// 1150int SROA::isSafeAllocaToScalarRepl(AllocationInst *AI) { 1151 // Loop over the use list of the alloca. We can only transform it if all of 1152 // the users are safe to transform. 1153 AllocaInfo Info; 1154 1155 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); 1156 I != E; ++I) { 1157 isSafeUseOfAllocation(cast<Instruction>(*I), AI, Info); 1158 if (Info.isUnsafe) { 1159 DOUT << "Cannot transform: " << *AI << " due to user: " << **I; 1160 return 0; 1161 } 1162 } 1163 1164 // Okay, we know all the users are promotable. If the aggregate is a memcpy 1165 // source and destination, we have to be careful. In particular, the memcpy 1166 // could be moving around elements that live in structure padding of the LLVM 1167 // types, but may actually be used. In these cases, we refuse to promote the 1168 // struct. 1169 if (Info.isMemCpySrc && Info.isMemCpyDst && 1170 HasPadding(AI->getType()->getElementType(), *TD)) 1171 return 0; 1172 1173 // If we require cleanup, return 1, otherwise return 3. 1174 return Info.needsCleanup ? 1 : 3; 1175} 1176 1177/// CleanupGEP - GEP is used by an Alloca, which can be prompted after the GEP 1178/// is canonicalized here. 1179void SROA::CleanupGEP(GetElementPtrInst *GEPI) { 1180 gep_type_iterator I = gep_type_begin(GEPI); 1181 ++I; 1182 1183 const ArrayType *AT = dyn_cast<ArrayType>(*I); 1184 if (!AT) 1185 return; 1186 1187 uint64_t NumElements = AT->getNumElements(); 1188 1189 if (isa<ConstantInt>(I.getOperand())) 1190 return; 1191 1192 LLVMContext &Context = GEPI->getContext(); 1193 1194 if (NumElements == 1) { 1195 GEPI->setOperand(2, Context.getNullValue(Type::Int32Ty)); 1196 return; 1197 } 1198 1199 assert(NumElements == 2 && "Unhandled case!"); 1200 // All users of the GEP must be loads. At each use of the GEP, insert 1201 // two loads of the appropriate indexed GEP and select between them. 1202 Value *IsOne = new ICmpInst(GEPI, ICmpInst::ICMP_NE, I.getOperand(), 1203 Context.getNullValue(I.getOperand()->getType()), 1204 "isone"); 1205 // Insert the new GEP instructions, which are properly indexed. 1206 SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end()); 1207 Indices[1] = Context.getNullValue(Type::Int32Ty); 1208 Value *ZeroIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 1209 Indices.begin(), 1210 Indices.end(), 1211 GEPI->getName()+".0", GEPI); 1212 Indices[1] = ConstantInt::get(Type::Int32Ty, 1); 1213 Value *OneIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 1214 Indices.begin(), 1215 Indices.end(), 1216 GEPI->getName()+".1", GEPI); 1217 // Replace all loads of the variable index GEP with loads from both 1218 // indexes and a select. 1219 while (!GEPI->use_empty()) { 1220 LoadInst *LI = cast<LoadInst>(GEPI->use_back()); 1221 Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI); 1222 Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI); 1223 Value *R = SelectInst::Create(IsOne, One, Zero, LI->getName(), LI); 1224 LI->replaceAllUsesWith(R); 1225 LI->eraseFromParent(); 1226 } 1227 GEPI->eraseFromParent(); 1228} 1229 1230 1231/// CleanupAllocaUsers - If SROA reported that it can promote the specified 1232/// allocation, but only if cleaned up, perform the cleanups required. 1233void SROA::CleanupAllocaUsers(AllocationInst *AI) { 1234 // At this point, we know that the end result will be SROA'd and promoted, so 1235 // we can insert ugly code if required so long as sroa+mem2reg will clean it 1236 // up. 1237 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 1238 UI != E; ) { 1239 User *U = *UI++; 1240 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) 1241 CleanupGEP(GEPI); 1242 else { 1243 Instruction *I = cast<Instruction>(U); 1244 SmallVector<DbgInfoIntrinsic *, 2> DbgInUses; 1245 if (!isa<StoreInst>(I) && OnlyUsedByDbgInfoIntrinsics(I, &DbgInUses)) { 1246 // Safe to remove debug info uses. 1247 while (!DbgInUses.empty()) { 1248 DbgInfoIntrinsic *DI = DbgInUses.back(); DbgInUses.pop_back(); 1249 DI->eraseFromParent(); 1250 } 1251 I->eraseFromParent(); 1252 } 1253 } 1254 } 1255} 1256 1257/// MergeInType - Add the 'In' type to the accumulated type (Accum) so far at 1258/// the offset specified by Offset (which is specified in bytes). 1259/// 1260/// There are two cases we handle here: 1261/// 1) A union of vector types of the same size and potentially its elements. 1262/// Here we turn element accesses into insert/extract element operations. 1263/// This promotes a <4 x float> with a store of float to the third element 1264/// into a <4 x float> that uses insert element. 1265/// 2) A fully general blob of memory, which we turn into some (potentially 1266/// large) integer type with extract and insert operations where the loads 1267/// and stores would mutate the memory. 1268static void MergeInType(const Type *In, uint64_t Offset, const Type *&VecTy, 1269 unsigned AllocaSize, const TargetData &TD, 1270 LLVMContext &Context) { 1271 // If this could be contributing to a vector, analyze it. 1272 if (VecTy != Type::VoidTy) { // either null or a vector type. 1273 1274 // If the In type is a vector that is the same size as the alloca, see if it 1275 // matches the existing VecTy. 1276 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 1277 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 1278 // If we're storing/loading a vector of the right size, allow it as a 1279 // vector. If this the first vector we see, remember the type so that 1280 // we know the element size. 1281 if (VecTy == 0) 1282 VecTy = VInTy; 1283 return; 1284 } 1285 } else if (In == Type::FloatTy || In == Type::DoubleTy || 1286 (isa<IntegerType>(In) && In->getPrimitiveSizeInBits() >= 8 && 1287 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 1288 // If we're accessing something that could be an element of a vector, see 1289 // if the implied vector agrees with what we already have and if Offset is 1290 // compatible with it. 1291 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 1292 if (Offset % EltSize == 0 && 1293 AllocaSize % EltSize == 0 && 1294 (VecTy == 0 || 1295 cast<VectorType>(VecTy)->getElementType() 1296 ->getPrimitiveSizeInBits()/8 == EltSize)) { 1297 if (VecTy == 0) 1298 VecTy = VectorType::get(In, AllocaSize/EltSize); 1299 return; 1300 } 1301 } 1302 } 1303 1304 // Otherwise, we have a case that we can't handle with an optimized vector 1305 // form. We can still turn this into a large integer. 1306 VecTy = Type::VoidTy; 1307} 1308 1309/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 1310/// its accesses to use a to single vector type, return true, and set VecTy to 1311/// the new type. If we could convert the alloca into a single promotable 1312/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 1313/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 1314/// is the current offset from the base of the alloca being analyzed. 1315/// 1316/// If we see at least one access to the value that is as a vector type, set the 1317/// SawVec flag. 1318/// 1319bool SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy, 1320 bool &SawVec, uint64_t Offset, 1321 unsigned AllocaSize) { 1322 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1323 Instruction *User = cast<Instruction>(*UI); 1324 1325 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1326 // Don't break volatile loads. 1327 if (LI->isVolatile()) 1328 return false; 1329 MergeInType(LI->getType(), Offset, VecTy, 1330 AllocaSize, *TD, V->getContext()); 1331 SawVec |= isa<VectorType>(LI->getType()); 1332 continue; 1333 } 1334 1335 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1336 // Storing the pointer, not into the value? 1337 if (SI->getOperand(0) == V || SI->isVolatile()) return 0; 1338 MergeInType(SI->getOperand(0)->getType(), Offset, 1339 VecTy, AllocaSize, *TD, V->getContext()); 1340 SawVec |= isa<VectorType>(SI->getOperand(0)->getType()); 1341 continue; 1342 } 1343 1344 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 1345 if (!CanConvertToScalar(BCI, IsNotTrivial, VecTy, SawVec, Offset, 1346 AllocaSize)) 1347 return false; 1348 IsNotTrivial = true; 1349 continue; 1350 } 1351 1352 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1353 // If this is a GEP with a variable indices, we can't handle it. 1354 if (!GEP->hasAllConstantIndices()) 1355 return false; 1356 1357 // Compute the offset that this GEP adds to the pointer. 1358 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1359 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(), 1360 &Indices[0], Indices.size()); 1361 // See if all uses can be converted. 1362 if (!CanConvertToScalar(GEP, IsNotTrivial, VecTy, SawVec,Offset+GEPOffset, 1363 AllocaSize)) 1364 return false; 1365 IsNotTrivial = true; 1366 continue; 1367 } 1368 1369 // If this is a constant sized memset of a constant value (e.g. 0) we can 1370 // handle it. 1371 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 1372 // Store of constant value and constant size. 1373 if (isa<ConstantInt>(MSI->getValue()) && 1374 isa<ConstantInt>(MSI->getLength())) { 1375 IsNotTrivial = true; 1376 continue; 1377 } 1378 } 1379 1380 // If this is a memcpy or memmove into or out of the whole allocation, we 1381 // can handle it like a load or store of the scalar type. 1382 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 1383 if (ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength())) 1384 if (Len->getZExtValue() == AllocaSize && Offset == 0) { 1385 IsNotTrivial = true; 1386 continue; 1387 } 1388 } 1389 1390 // Ignore dbg intrinsic. 1391 if (isa<DbgInfoIntrinsic>(User)) 1392 continue; 1393 1394 // Otherwise, we cannot handle this! 1395 return false; 1396 } 1397 1398 return true; 1399} 1400 1401 1402/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 1403/// directly. This happens when we are converting an "integer union" to a 1404/// single integer scalar, or when we are converting a "vector union" to a 1405/// vector with insert/extractelement instructions. 1406/// 1407/// Offset is an offset from the original alloca, in bits that need to be 1408/// shifted to the right. By the end of this, there should be no uses of Ptr. 1409void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset) { 1410 while (!Ptr->use_empty()) { 1411 Instruction *User = cast<Instruction>(Ptr->use_back()); 1412 1413 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 1414 ConvertUsesToScalar(CI, NewAI, Offset); 1415 CI->eraseFromParent(); 1416 continue; 1417 } 1418 1419 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1420 // Compute the offset that this GEP adds to the pointer. 1421 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1422 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(), 1423 &Indices[0], Indices.size()); 1424 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 1425 GEP->eraseFromParent(); 1426 continue; 1427 } 1428 1429 IRBuilder<> Builder(User->getParent(), User); 1430 1431 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1432 // The load is a bit extract from NewAI shifted right by Offset bits. 1433 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 1434 Value *NewLoadVal 1435 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 1436 LI->replaceAllUsesWith(NewLoadVal); 1437 LI->eraseFromParent(); 1438 continue; 1439 } 1440 1441 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1442 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 1443 // FIXME: Remove once builder has Twine API. 1444 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").str().c_str()); 1445 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 1446 Builder); 1447 Builder.CreateStore(New, NewAI); 1448 SI->eraseFromParent(); 1449 continue; 1450 } 1451 1452 // If this is a constant sized memset of a constant value (e.g. 0) we can 1453 // transform it into a store of the expanded constant value. 1454 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 1455 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 1456 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 1457 if (NumBytes != 0) { 1458 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 1459 1460 // Compute the value replicated the right number of times. 1461 APInt APVal(NumBytes*8, Val); 1462 1463 // Splat the value if non-zero. 1464 if (Val) 1465 for (unsigned i = 1; i != NumBytes; ++i) 1466 APVal |= APVal << 8; 1467 1468 // FIXME: Remove once builder has Twine API. 1469 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").str().c_str()); 1470 Value *New = ConvertScalar_InsertValue( 1471 ConstantInt::get(User->getContext(), APVal), 1472 Old, Offset, Builder); 1473 Builder.CreateStore(New, NewAI); 1474 } 1475 MSI->eraseFromParent(); 1476 continue; 1477 } 1478 1479 // If this is a memcpy or memmove into or out of the whole allocation, we 1480 // can handle it like a load or store of the scalar type. 1481 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 1482 assert(Offset == 0 && "must be store to start of alloca"); 1483 1484 // If the source and destination are both to the same alloca, then this is 1485 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 1486 // as appropriate. 1487 AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject()); 1488 1489 if (MTI->getSource()->getUnderlyingObject() != OrigAI) { 1490 // Dest must be OrigAI, change this to be a load from the original 1491 // pointer (bitcasted), then a store to our new alloca. 1492 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 1493 Value *SrcPtr = MTI->getSource(); 1494 SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType()); 1495 1496 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 1497 SrcVal->setAlignment(MTI->getAlignment()); 1498 Builder.CreateStore(SrcVal, NewAI); 1499 } else if (MTI->getDest()->getUnderlyingObject() != OrigAI) { 1500 // Src must be OrigAI, change this to be a load from NewAI then a store 1501 // through the original dest pointer (bitcasted). 1502 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 1503 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 1504 1505 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType()); 1506 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 1507 NewStore->setAlignment(MTI->getAlignment()); 1508 } else { 1509 // Noop transfer. Src == Dst 1510 } 1511 1512 1513 MTI->eraseFromParent(); 1514 continue; 1515 } 1516 1517 // If user is a dbg info intrinsic then it is safe to remove it. 1518 if (isa<DbgInfoIntrinsic>(User)) { 1519 User->eraseFromParent(); 1520 continue; 1521 } 1522 1523 llvm_unreachable("Unsupported operation!"); 1524 } 1525} 1526 1527/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 1528/// or vector value FromVal, extracting the bits from the offset specified by 1529/// Offset. This returns the value, which is of type ToType. 1530/// 1531/// This happens when we are converting an "integer union" to a single 1532/// integer scalar, or when we are converting a "vector union" to a vector with 1533/// insert/extractelement instructions. 1534/// 1535/// Offset is an offset from the original alloca, in bits that need to be 1536/// shifted to the right. 1537Value *SROA::ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 1538 uint64_t Offset, IRBuilder<> &Builder) { 1539 // If the load is of the whole new alloca, no conversion is needed. 1540 if (FromVal->getType() == ToType && Offset == 0) 1541 return FromVal; 1542 1543 // If the result alloca is a vector type, this is either an element 1544 // access or a bitcast to another vector type of the same size. 1545 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) { 1546 if (isa<VectorType>(ToType)) 1547 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 1548 1549 // Otherwise it must be an element access. 1550 unsigned Elt = 0; 1551 if (Offset) { 1552 unsigned EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 1553 Elt = Offset/EltSize; 1554 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 1555 } 1556 // Return the element extracted out of it. 1557 Value *V = Builder.CreateExtractElement(FromVal, 1558 ConstantInt::get(Type::Int32Ty,Elt), 1559 "tmp"); 1560 if (V->getType() != ToType) 1561 V = Builder.CreateBitCast(V, ToType, "tmp"); 1562 return V; 1563 } 1564 1565 // If ToType is a first class aggregate, extract out each of the pieces and 1566 // use insertvalue's to form the FCA. 1567 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 1568 const StructLayout &Layout = *TD->getStructLayout(ST); 1569 Value *Res = UndefValue::get(ST); 1570 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1571 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 1572 Offset+Layout.getElementOffsetInBits(i), 1573 Builder); 1574 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1575 } 1576 return Res; 1577 } 1578 1579 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 1580 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 1581 Value *Res = UndefValue::get(AT); 1582 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1583 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 1584 Offset+i*EltSize, Builder); 1585 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1586 } 1587 return Res; 1588 } 1589 1590 // Otherwise, this must be a union that was converted to an integer value. 1591 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 1592 1593 // If this is a big-endian system and the load is narrower than the 1594 // full alloca type, we need to do a shift to get the right bits. 1595 int ShAmt = 0; 1596 if (TD->isBigEndian()) { 1597 // On big-endian machines, the lowest bit is stored at the bit offset 1598 // from the pointer given by getTypeStoreSizeInBits. This matters for 1599 // integers with a bitwidth that is not a multiple of 8. 1600 ShAmt = TD->getTypeStoreSizeInBits(NTy) - 1601 TD->getTypeStoreSizeInBits(ToType) - Offset; 1602 } else { 1603 ShAmt = Offset; 1604 } 1605 1606 // Note: we support negative bitwidths (with shl) which are not defined. 1607 // We do this to support (f.e.) loads off the end of a structure where 1608 // only some bits are used. 1609 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 1610 FromVal = Builder.CreateLShr(FromVal, 1611 ConstantInt::get(FromVal->getType(), 1612 ShAmt), "tmp"); 1613 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 1614 FromVal = Builder.CreateShl(FromVal, 1615 ConstantInt::get(FromVal->getType(), 1616 -ShAmt), "tmp"); 1617 1618 // Finally, unconditionally truncate the integer to the right width. 1619 unsigned LIBitWidth = TD->getTypeSizeInBits(ToType); 1620 if (LIBitWidth < NTy->getBitWidth()) 1621 FromVal = 1622 Builder.CreateTrunc(FromVal, IntegerType::get(LIBitWidth), "tmp"); 1623 else if (LIBitWidth > NTy->getBitWidth()) 1624 FromVal = 1625 Builder.CreateZExt(FromVal, IntegerType::get(LIBitWidth), "tmp"); 1626 1627 // If the result is an integer, this is a trunc or bitcast. 1628 if (isa<IntegerType>(ToType)) { 1629 // Should be done. 1630 } else if (ToType->isFloatingPoint() || isa<VectorType>(ToType)) { 1631 // Just do a bitcast, we know the sizes match up. 1632 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 1633 } else { 1634 // Otherwise must be a pointer. 1635 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 1636 } 1637 assert(FromVal->getType() == ToType && "Didn't convert right?"); 1638 return FromVal; 1639} 1640 1641 1642/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 1643/// or vector value "Old" at the offset specified by Offset. 1644/// 1645/// This happens when we are converting an "integer union" to a 1646/// single integer scalar, or when we are converting a "vector union" to a 1647/// vector with insert/extractelement instructions. 1648/// 1649/// Offset is an offset from the original alloca, in bits that need to be 1650/// shifted to the right. 1651Value *SROA::ConvertScalar_InsertValue(Value *SV, Value *Old, 1652 uint64_t Offset, IRBuilder<> &Builder) { 1653 1654 // Convert the stored type to the actual type, shift it left to insert 1655 // then 'or' into place. 1656 const Type *AllocaType = Old->getType(); 1657 LLVMContext &Context = Old->getContext(); 1658 1659 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 1660 uint64_t VecSize = TD->getTypeAllocSizeInBits(VTy); 1661 uint64_t ValSize = TD->getTypeAllocSizeInBits(SV->getType()); 1662 1663 // Changing the whole vector with memset or with an access of a different 1664 // vector type? 1665 if (ValSize == VecSize) 1666 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 1667 1668 uint64_t EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 1669 1670 // Must be an element insertion. 1671 unsigned Elt = Offset/EltSize; 1672 1673 if (SV->getType() != VTy->getElementType()) 1674 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 1675 1676 SV = Builder.CreateInsertElement(Old, SV, 1677 ConstantInt::get(Type::Int32Ty, Elt), 1678 "tmp"); 1679 return SV; 1680 } 1681 1682 // If SV is a first-class aggregate value, insert each value recursively. 1683 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 1684 const StructLayout &Layout = *TD->getStructLayout(ST); 1685 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1686 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1687 Old = ConvertScalar_InsertValue(Elt, Old, 1688 Offset+Layout.getElementOffsetInBits(i), 1689 Builder); 1690 } 1691 return Old; 1692 } 1693 1694 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 1695 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 1696 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1697 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1698 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 1699 } 1700 return Old; 1701 } 1702 1703 // If SV is a float, convert it to the appropriate integer type. 1704 // If it is a pointer, do the same. 1705 unsigned SrcWidth = TD->getTypeSizeInBits(SV->getType()); 1706 unsigned DestWidth = TD->getTypeSizeInBits(AllocaType); 1707 unsigned SrcStoreWidth = TD->getTypeStoreSizeInBits(SV->getType()); 1708 unsigned DestStoreWidth = TD->getTypeStoreSizeInBits(AllocaType); 1709 if (SV->getType()->isFloatingPoint() || isa<VectorType>(SV->getType())) 1710 SV = Builder.CreateBitCast(SV, IntegerType::get(SrcWidth), "tmp"); 1711 else if (isa<PointerType>(SV->getType())) 1712 SV = Builder.CreatePtrToInt(SV, TD->getIntPtrType(), "tmp"); 1713 1714 // Zero extend or truncate the value if needed. 1715 if (SV->getType() != AllocaType) { 1716 if (SV->getType()->getPrimitiveSizeInBits() < 1717 AllocaType->getPrimitiveSizeInBits()) 1718 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 1719 else { 1720 // Truncation may be needed if storing more than the alloca can hold 1721 // (undefined behavior). 1722 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 1723 SrcWidth = DestWidth; 1724 SrcStoreWidth = DestStoreWidth; 1725 } 1726 } 1727 1728 // If this is a big-endian system and the store is narrower than the 1729 // full alloca type, we need to do a shift to get the right bits. 1730 int ShAmt = 0; 1731 if (TD->isBigEndian()) { 1732 // On big-endian machines, the lowest bit is stored at the bit offset 1733 // from the pointer given by getTypeStoreSizeInBits. This matters for 1734 // integers with a bitwidth that is not a multiple of 8. 1735 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 1736 } else { 1737 ShAmt = Offset; 1738 } 1739 1740 // Note: we support negative bitwidths (with shr) which are not defined. 1741 // We do this to support (f.e.) stores off the end of a structure where 1742 // only some bits in the structure are set. 1743 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 1744 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 1745 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 1746 ShAmt), "tmp"); 1747 Mask <<= ShAmt; 1748 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1749 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 1750 -ShAmt), "tmp"); 1751 Mask = Mask.lshr(-ShAmt); 1752 } 1753 1754 // Mask out the bits we are about to insert from the old value, and or 1755 // in the new bits. 1756 if (SrcWidth != DestWidth) { 1757 assert(DestWidth > SrcWidth); 1758 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 1759 SV = Builder.CreateOr(Old, SV, "ins"); 1760 } 1761 return SV; 1762} 1763 1764 1765 1766/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1767/// some part of a constant global variable. This intentionally only accepts 1768/// constant expressions because we don't can't rewrite arbitrary instructions. 1769static bool PointsToConstantGlobal(Value *V) { 1770 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1771 return GV->isConstant(); 1772 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1773 if (CE->getOpcode() == Instruction::BitCast || 1774 CE->getOpcode() == Instruction::GetElementPtr) 1775 return PointsToConstantGlobal(CE->getOperand(0)); 1776 return false; 1777} 1778 1779/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1780/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1781/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1782/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1783/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1784/// the alloca, and if the source pointer is a pointer to a constant global, we 1785/// can optimize this. 1786static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy, 1787 bool isOffset) { 1788 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1789 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) 1790 // Ignore non-volatile loads, they are always ok. 1791 if (!LI->isVolatile()) 1792 continue; 1793 1794 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) { 1795 // If uses of the bitcast are ok, we are ok. 1796 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 1797 return false; 1798 continue; 1799 } 1800 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) { 1801 // If the GEP has all zero indices, it doesn't offset the pointer. If it 1802 // doesn't, it does. 1803 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 1804 isOffset || !GEP->hasAllZeroIndices())) 1805 return false; 1806 continue; 1807 } 1808 1809 // If this is isn't our memcpy/memmove, reject it as something we can't 1810 // handle. 1811 if (!isa<MemTransferInst>(*UI)) 1812 return false; 1813 1814 // If we already have seen a copy, reject the second one. 1815 if (TheCopy) return false; 1816 1817 // If the pointer has been offset from the start of the alloca, we can't 1818 // safely handle this. 1819 if (isOffset) return false; 1820 1821 // If the memintrinsic isn't using the alloca as the dest, reject it. 1822 if (UI.getOperandNo() != 1) return false; 1823 1824 MemIntrinsic *MI = cast<MemIntrinsic>(*UI); 1825 1826 // If the source of the memcpy/move is not a constant global, reject it. 1827 if (!PointsToConstantGlobal(MI->getOperand(2))) 1828 return false; 1829 1830 // Otherwise, the transform is safe. Remember the copy instruction. 1831 TheCopy = MI; 1832 } 1833 return true; 1834} 1835 1836/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 1837/// modified by a copy from a constant global. If we can prove this, we can 1838/// replace any uses of the alloca with uses of the global directly. 1839Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocationInst *AI) { 1840 Instruction *TheCopy = 0; 1841 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 1842 return TheCopy; 1843 return 0; 1844} 1845