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/Pass.h" 31#include "llvm/Analysis/Dominators.h" 32#include "llvm/Target/TargetData.h" 33#include "llvm/Transforms/Utils/PromoteMemToReg.h" 34#include "llvm/Transforms/Utils/Local.h" 35#include "llvm/Support/Debug.h" 36#include "llvm/Support/GetElementPtrTypeIterator.h" 37#include "llvm/Support/IRBuilder.h" 38#include "llvm/Support/MathExtras.h" 39#include "llvm/Support/Compiler.h" 40#include "llvm/ADT/SmallVector.h" 41#include "llvm/ADT/Statistic.h" 42#include "llvm/ADT/StringExtras.h" 43using namespace llvm; 44 45STATISTIC(NumReplaced, "Number of allocas broken up"); 46STATISTIC(NumPromoted, "Number of allocas promoted"); 47STATISTIC(NumConverted, "Number of aggregates converted to scalar"); 48STATISTIC(NumGlobals, "Number of allocas copied from constant global"); 49 50namespace { 51 struct VISIBILITY_HIDDEN SROA : public FunctionPass { 52 static char ID; // Pass identification, replacement for typeid 53 explicit SROA(signed T = -1) : FunctionPass(&ID) { 54 if (T == -1) 55 SRThreshold = 128; 56 else 57 SRThreshold = T; 58 } 59 60 bool runOnFunction(Function &F); 61 62 bool performScalarRepl(Function &F); 63 bool performPromotion(Function &F); 64 65 // getAnalysisUsage - This pass does not require any passes, but we know it 66 // will not alter the CFG, so say so. 67 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 68 AU.addRequired<DominatorTree>(); 69 AU.addRequired<DominanceFrontier>(); 70 AU.addRequired<TargetData>(); 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(AllocationInst *AI); 104 105 void isSafeUseOfAllocation(Instruction *User, AllocationInst *AI, 106 AllocaInfo &Info); 107 void isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI, 108 AllocaInfo &Info); 109 void isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI, 110 unsigned OpNo, AllocaInfo &Info); 111 void isSafeUseOfBitCastedAllocation(BitCastInst *User, AllocationInst *AI, 112 AllocaInfo &Info); 113 114 void DoScalarReplacement(AllocationInst *AI, 115 std::vector<AllocationInst*> &WorkList); 116 void CleanupGEP(GetElementPtrInst *GEP); 117 void CleanupAllocaUsers(AllocationInst *AI); 118 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocationInst *Base); 119 120 void RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI, 121 SmallVector<AllocaInst*, 32> &NewElts); 122 123 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst, 124 AllocationInst *AI, 125 SmallVector<AllocaInst*, 32> &NewElts); 126 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocationInst *AI, 127 SmallVector<AllocaInst*, 32> &NewElts); 128 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *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(AllocationInst *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 = &getAnalysis<TargetData>(); 153 154 bool Changed = performPromotion(F); 155 while (1) { 156 bool LocalChange = performScalarRepl(F); 157 if (!LocalChange) break; // No need to repromote if no scalarrepl 158 Changed = true; 159 LocalChange = performPromotion(F); 160 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 161 } 162 163 return Changed; 164} 165 166 167bool SROA::performPromotion(Function &F) { 168 std::vector<AllocaInst*> Allocas; 169 DominatorTree &DT = getAnalysis<DominatorTree>(); 170 DominanceFrontier &DF = getAnalysis<DominanceFrontier>(); 171 172 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 173 174 bool Changed = false; 175 176 while (1) { 177 Allocas.clear(); 178 179 // Find allocas that are safe to promote, by looking at all instructions in 180 // the entry node 181 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 182 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 183 if (isAllocaPromotable(AI)) 184 Allocas.push_back(AI); 185 186 if (Allocas.empty()) break; 187 188 PromoteMemToReg(Allocas, DT, DF); 189 NumPromoted += Allocas.size(); 190 Changed = true; 191 } 192 193 return Changed; 194} 195 196/// getNumSAElements - Return the number of elements in the specific struct or 197/// array. 198static uint64_t getNumSAElements(const Type *T) { 199 if (const StructType *ST = dyn_cast<StructType>(T)) 200 return ST->getNumElements(); 201 return cast<ArrayType>(T)->getNumElements(); 202} 203 204// performScalarRepl - This algorithm is a simple worklist driven algorithm, 205// which runs on all of the malloc/alloca instructions in the function, removing 206// them if they are only used by getelementptr instructions. 207// 208bool SROA::performScalarRepl(Function &F) { 209 std::vector<AllocationInst*> WorkList; 210 211 // Scan the entry basic block, adding any alloca's and mallocs to the worklist 212 BasicBlock &BB = F.getEntryBlock(); 213 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 214 if (AllocationInst *A = dyn_cast<AllocationInst>(I)) 215 WorkList.push_back(A); 216 217 // Process the worklist 218 bool Changed = false; 219 while (!WorkList.empty()) { 220 AllocationInst *AI = WorkList.back(); 221 WorkList.pop_back(); 222 223 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 224 // with unused elements. 225 if (AI->use_empty()) { 226 AI->eraseFromParent(); 227 continue; 228 } 229 230 // If this alloca is impossible for us to promote, reject it early. 231 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 232 continue; 233 234 // Check to see if this allocation is only modified by a memcpy/memmove from 235 // a constant global. If this is the case, we can change all users to use 236 // the constant global instead. This is commonly produced by the CFE by 237 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 238 // is only subsequently read. 239 if (Instruction *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 240 DOUT << "Found alloca equal to global: " << *AI; 241 DOUT << " memcpy = " << *TheCopy; 242 Constant *TheSrc = cast<Constant>(TheCopy->getOperand(2)); 243 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 244 TheCopy->eraseFromParent(); // Don't mutate the global. 245 AI->eraseFromParent(); 246 ++NumGlobals; 247 Changed = true; 248 continue; 249 } 250 251 // Check to see if we can perform the core SROA transformation. We cannot 252 // transform the allocation instruction if it is an array allocation 253 // (allocations OF arrays are ok though), and an allocation of a scalar 254 // value cannot be decomposed at all. 255 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 256 257 // Do not promote any struct whose size is too big. 258 if (AllocaSize > SRThreshold) continue; 259 260 if ((isa<StructType>(AI->getAllocatedType()) || 261 isa<ArrayType>(AI->getAllocatedType())) && 262 // Do not promote any struct into more than "32" separate vars. 263 getNumSAElements(AI->getAllocatedType()) <= SRThreshold/4) { 264 // Check that all of the users of the allocation are capable of being 265 // transformed. 266 switch (isSafeAllocaToScalarRepl(AI)) { 267 default: assert(0 && "Unexpected value!"); 268 case 0: // Not safe to scalar replace. 269 break; 270 case 1: // Safe, but requires cleanup/canonicalizations first 271 CleanupAllocaUsers(AI); 272 // FALL THROUGH. 273 case 3: // Safe to scalar replace. 274 DoScalarReplacement(AI, WorkList); 275 Changed = true; 276 continue; 277 } 278 } 279 280 // If we can turn this aggregate value (potentially with casts) into a 281 // simple scalar value that can be mem2reg'd into a register value. 282 // IsNotTrivial tracks whether this is something that mem2reg could have 283 // promoted itself. If so, we don't want to transform it needlessly. Note 284 // that we can't just check based on the type: the alloca may be of an i32 285 // but that has pointer arithmetic to set byte 3 of it or something. 286 bool IsNotTrivial = false; 287 const Type *VectorTy = 0; 288 bool HadAVector = false; 289 if (CanConvertToScalar(AI, IsNotTrivial, VectorTy, HadAVector, 290 0, unsigned(AllocaSize)) && IsNotTrivial) { 291 AllocaInst *NewAI; 292 // If we were able to find a vector type that can handle this with 293 // insert/extract elements, and if there was at least one use that had 294 // a vector type, promote this to a vector. We don't want to promote 295 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 296 // we just get a lot of insert/extracts. If at least one vector is 297 // involved, then we probably really do have a union of vector/array. 298 if (VectorTy && isa<VectorType>(VectorTy) && HadAVector) { 299 DOUT << "CONVERT TO VECTOR: " << *AI << " TYPE = " << *VectorTy <<"\n"; 300 301 // Create and insert the vector alloca. 302 NewAI = new AllocaInst(VectorTy, 0, "", AI->getParent()->begin()); 303 ConvertUsesToScalar(AI, NewAI, 0); 304 } else { 305 DOUT << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"; 306 307 // Create and insert the integer alloca. 308 const Type *NewTy = IntegerType::get(AllocaSize*8); 309 NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 310 ConvertUsesToScalar(AI, NewAI, 0); 311 } 312 NewAI->takeName(AI); 313 AI->eraseFromParent(); 314 ++NumConverted; 315 Changed = true; 316 continue; 317 } 318 319 // Otherwise, couldn't process this alloca. 320 } 321 322 return Changed; 323} 324 325/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 326/// predicate, do SROA now. 327void SROA::DoScalarReplacement(AllocationInst *AI, 328 std::vector<AllocationInst*> &WorkList) { 329 DOUT << "Found inst to SROA: " << *AI; 330 SmallVector<AllocaInst*, 32> ElementAllocas; 331 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 332 ElementAllocas.reserve(ST->getNumContainedTypes()); 333 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 334 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 335 AI->getAlignment(), 336 AI->getName() + "." + utostr(i), AI); 337 ElementAllocas.push_back(NA); 338 WorkList.push_back(NA); // Add to worklist for recursive processing 339 } 340 } else { 341 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 342 ElementAllocas.reserve(AT->getNumElements()); 343 const Type *ElTy = AT->getElementType(); 344 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 345 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 346 AI->getName() + "." + utostr(i), AI); 347 ElementAllocas.push_back(NA); 348 WorkList.push_back(NA); // Add to worklist for recursive processing 349 } 350 } 351 352 // Now that we have created the alloca instructions that we want to use, 353 // expand the getelementptr instructions to use them. 354 // 355 while (!AI->use_empty()) { 356 Instruction *User = cast<Instruction>(AI->use_back()); 357 if (BitCastInst *BCInst = dyn_cast<BitCastInst>(User)) { 358 RewriteBitCastUserOfAlloca(BCInst, AI, ElementAllocas); 359 BCInst->eraseFromParent(); 360 continue; 361 } 362 363 // Replace: 364 // %res = load { i32, i32 }* %alloc 365 // with: 366 // %load.0 = load i32* %alloc.0 367 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 368 // %load.1 = load i32* %alloc.1 369 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 370 // (Also works for arrays instead of structs) 371 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 372 Value *Insert = UndefValue::get(LI->getType()); 373 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) { 374 Value *Load = new LoadInst(ElementAllocas[i], "load", LI); 375 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 376 } 377 LI->replaceAllUsesWith(Insert); 378 LI->eraseFromParent(); 379 continue; 380 } 381 382 // Replace: 383 // store { i32, i32 } %val, { i32, i32 }* %alloc 384 // with: 385 // %val.0 = extractvalue { i32, i32 } %val, 0 386 // store i32 %val.0, i32* %alloc.0 387 // %val.1 = extractvalue { i32, i32 } %val, 1 388 // store i32 %val.1, i32* %alloc.1 389 // (Also works for arrays instead of structs) 390 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 391 Value *Val = SI->getOperand(0); 392 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) { 393 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 394 new StoreInst(Extract, ElementAllocas[i], SI); 395 } 396 SI->eraseFromParent(); 397 continue; 398 } 399 400 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User); 401 // We now know that the GEP is of the form: GEP <ptr>, 0, <cst> 402 unsigned Idx = 403 (unsigned)cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue(); 404 405 assert(Idx < ElementAllocas.size() && "Index out of range?"); 406 AllocaInst *AllocaToUse = ElementAllocas[Idx]; 407 408 Value *RepValue; 409 if (GEPI->getNumOperands() == 3) { 410 // Do not insert a new getelementptr instruction with zero indices, only 411 // to have it optimized out later. 412 RepValue = AllocaToUse; 413 } else { 414 // We are indexing deeply into the structure, so we still need a 415 // getelement ptr instruction to finish the indexing. This may be 416 // expanded itself once the worklist is rerun. 417 // 418 SmallVector<Value*, 8> NewArgs; 419 NewArgs.push_back(Constant::getNullValue(Type::Int32Ty)); 420 NewArgs.append(GEPI->op_begin()+3, GEPI->op_end()); 421 RepValue = GetElementPtrInst::Create(AllocaToUse, NewArgs.begin(), 422 NewArgs.end(), "", GEPI); 423 RepValue->takeName(GEPI); 424 } 425 426 // If this GEP is to the start of the aggregate, check for memcpys. 427 if (Idx == 0 && GEPI->hasAllZeroIndices()) 428 RewriteBitCastUserOfAlloca(GEPI, AI, ElementAllocas); 429 430 // Move all of the users over to the new GEP. 431 GEPI->replaceAllUsesWith(RepValue); 432 // Delete the old GEP 433 GEPI->eraseFromParent(); 434 } 435 436 // Finally, delete the Alloca instruction 437 AI->eraseFromParent(); 438 NumReplaced++; 439} 440 441 442/// isSafeElementUse - Check to see if this use is an allowed use for a 443/// getelementptr instruction of an array aggregate allocation. isFirstElt 444/// indicates whether Ptr is known to the start of the aggregate. 445/// 446void SROA::isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI, 447 AllocaInfo &Info) { 448 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 449 I != E; ++I) { 450 Instruction *User = cast<Instruction>(*I); 451 switch (User->getOpcode()) { 452 case Instruction::Load: break; 453 case Instruction::Store: 454 // Store is ok if storing INTO the pointer, not storing the pointer 455 if (User->getOperand(0) == Ptr) return MarkUnsafe(Info); 456 break; 457 case Instruction::GetElementPtr: { 458 GetElementPtrInst *GEP = cast<GetElementPtrInst>(User); 459 bool AreAllZeroIndices = isFirstElt; 460 if (GEP->getNumOperands() > 1) { 461 if (!isa<ConstantInt>(GEP->getOperand(1)) || 462 !cast<ConstantInt>(GEP->getOperand(1))->isZero()) 463 // Using pointer arithmetic to navigate the array. 464 return MarkUnsafe(Info); 465 466 if (AreAllZeroIndices) 467 AreAllZeroIndices = GEP->hasAllZeroIndices(); 468 } 469 isSafeElementUse(GEP, AreAllZeroIndices, AI, Info); 470 if (Info.isUnsafe) return; 471 break; 472 } 473 case Instruction::BitCast: 474 if (isFirstElt) { 475 isSafeUseOfBitCastedAllocation(cast<BitCastInst>(User), AI, Info); 476 if (Info.isUnsafe) return; 477 break; 478 } 479 DOUT << " Transformation preventing inst: " << *User; 480 return MarkUnsafe(Info); 481 case Instruction::Call: 482 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 483 if (isFirstElt) { 484 isSafeMemIntrinsicOnAllocation(MI, AI, I.getOperandNo(), Info); 485 if (Info.isUnsafe) return; 486 break; 487 } 488 } 489 DOUT << " Transformation preventing inst: " << *User; 490 return MarkUnsafe(Info); 491 default: 492 DOUT << " Transformation preventing inst: " << *User; 493 return MarkUnsafe(Info); 494 } 495 } 496 return; // All users look ok :) 497} 498 499/// AllUsersAreLoads - Return true if all users of this value are loads. 500static bool AllUsersAreLoads(Value *Ptr) { 501 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 502 I != E; ++I) 503 if (cast<Instruction>(*I)->getOpcode() != Instruction::Load) 504 return false; 505 return true; 506} 507 508/// isSafeUseOfAllocation - Check to see if this user is an allowed use for an 509/// aggregate allocation. 510/// 511void SROA::isSafeUseOfAllocation(Instruction *User, AllocationInst *AI, 512 AllocaInfo &Info) { 513 if (BitCastInst *C = dyn_cast<BitCastInst>(User)) 514 return isSafeUseOfBitCastedAllocation(C, AI, Info); 515 516 if (LoadInst *LI = dyn_cast<LoadInst>(User)) 517 if (!LI->isVolatile()) 518 return;// Loads (returning a first class aggregrate) are always rewritable 519 520 if (StoreInst *SI = dyn_cast<StoreInst>(User)) 521 if (!SI->isVolatile() && SI->getOperand(0) != AI) 522 return;// Store is ok if storing INTO the pointer, not storing the pointer 523 524 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User); 525 if (GEPI == 0) 526 return MarkUnsafe(Info); 527 528 gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI); 529 530 // The GEP is not safe to transform if not of the form "GEP <ptr>, 0, <cst>". 531 if (I == E || 532 I.getOperand() != Constant::getNullValue(I.getOperand()->getType())) { 533 return MarkUnsafe(Info); 534 } 535 536 ++I; 537 if (I == E) return MarkUnsafe(Info); // ran out of GEP indices?? 538 539 bool IsAllZeroIndices = true; 540 541 // If the first index is a non-constant index into an array, see if we can 542 // handle it as a special case. 543 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 544 if (!isa<ConstantInt>(I.getOperand())) { 545 IsAllZeroIndices = 0; 546 uint64_t NumElements = AT->getNumElements(); 547 548 // If this is an array index and the index is not constant, we cannot 549 // promote... that is unless the array has exactly one or two elements in 550 // it, in which case we CAN promote it, but we have to canonicalize this 551 // out if this is the only problem. 552 if ((NumElements == 1 || NumElements == 2) && 553 AllUsersAreLoads(GEPI)) { 554 Info.needsCleanup = true; 555 return; // Canonicalization required! 556 } 557 return MarkUnsafe(Info); 558 } 559 } 560 561 // Walk through the GEP type indices, checking the types that this indexes 562 // into. 563 for (; I != E; ++I) { 564 // Ignore struct elements, no extra checking needed for these. 565 if (isa<StructType>(*I)) 566 continue; 567 568 ConstantInt *IdxVal = dyn_cast<ConstantInt>(I.getOperand()); 569 if (!IdxVal) return MarkUnsafe(Info); 570 571 // Are all indices still zero? 572 IsAllZeroIndices &= IdxVal->isZero(); 573 574 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 575 // This GEP indexes an array. Verify that this is an in-range constant 576 // integer. Specifically, consider A[0][i]. We cannot know that the user 577 // isn't doing invalid things like allowing i to index an out-of-range 578 // subscript that accesses A[1]. Because of this, we have to reject SROA 579 // of any accesses into structs where any of the components are variables. 580 if (IdxVal->getZExtValue() >= AT->getNumElements()) 581 return MarkUnsafe(Info); 582 } else if (const VectorType *VT = dyn_cast<VectorType>(*I)) { 583 if (IdxVal->getZExtValue() >= VT->getNumElements()) 584 return MarkUnsafe(Info); 585 } 586 } 587 588 // If there are any non-simple uses of this getelementptr, make sure to reject 589 // them. 590 return isSafeElementUse(GEPI, IsAllZeroIndices, AI, Info); 591} 592 593/// isSafeMemIntrinsicOnAllocation - Return true if the specified memory 594/// intrinsic can be promoted by SROA. At this point, we know that the operand 595/// of the memintrinsic is a pointer to the beginning of the allocation. 596void SROA::isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI, 597 unsigned OpNo, AllocaInfo &Info) { 598 // If not constant length, give up. 599 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 600 if (!Length) return MarkUnsafe(Info); 601 602 // If not the whole aggregate, give up. 603 if (Length->getZExtValue() != 604 TD->getTypeAllocSize(AI->getType()->getElementType())) 605 return MarkUnsafe(Info); 606 607 // We only know about memcpy/memset/memmove. 608 if (!isa<MemIntrinsic>(MI)) 609 return MarkUnsafe(Info); 610 611 // Otherwise, we can transform it. Determine whether this is a memcpy/set 612 // into or out of the aggregate. 613 if (OpNo == 1) 614 Info.isMemCpyDst = true; 615 else { 616 assert(OpNo == 2); 617 Info.isMemCpySrc = true; 618 } 619} 620 621/// isSafeUseOfBitCastedAllocation - Return true if all users of this bitcast 622/// are 623void SROA::isSafeUseOfBitCastedAllocation(BitCastInst *BC, AllocationInst *AI, 624 AllocaInfo &Info) { 625 for (Value::use_iterator UI = BC->use_begin(), E = BC->use_end(); 626 UI != E; ++UI) { 627 if (BitCastInst *BCU = dyn_cast<BitCastInst>(UI)) { 628 isSafeUseOfBitCastedAllocation(BCU, AI, Info); 629 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) { 630 isSafeMemIntrinsicOnAllocation(MI, AI, UI.getOperandNo(), Info); 631 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI)) { 632 if (SI->isVolatile()) 633 return MarkUnsafe(Info); 634 635 // If storing the entire alloca in one chunk through a bitcasted pointer 636 // to integer, we can transform it. This happens (for example) when you 637 // cast a {i32,i32}* to i64* and store through it. This is similar to the 638 // memcpy case and occurs in various "byval" cases and emulated memcpys. 639 if (isa<IntegerType>(SI->getOperand(0)->getType()) && 640 TD->getTypeAllocSize(SI->getOperand(0)->getType()) == 641 TD->getTypeAllocSize(AI->getType()->getElementType())) { 642 Info.isMemCpyDst = true; 643 continue; 644 } 645 return MarkUnsafe(Info); 646 } else if (LoadInst *LI = dyn_cast<LoadInst>(UI)) { 647 if (LI->isVolatile()) 648 return MarkUnsafe(Info); 649 650 // If loading the entire alloca in one chunk through a bitcasted pointer 651 // to integer, we can transform it. This happens (for example) when you 652 // cast a {i32,i32}* to i64* and load through it. This is similar to the 653 // memcpy case and occurs in various "byval" cases and emulated memcpys. 654 if (isa<IntegerType>(LI->getType()) && 655 TD->getTypeAllocSize(LI->getType()) == 656 TD->getTypeAllocSize(AI->getType()->getElementType())) { 657 Info.isMemCpySrc = true; 658 continue; 659 } 660 return MarkUnsafe(Info); 661 } else if (isa<DbgInfoIntrinsic>(UI)) { 662 // If one user is DbgInfoIntrinsic then check if all users are 663 // DbgInfoIntrinsics. 664 if (OnlyUsedByDbgInfoIntrinsics(BC)) { 665 Info.needsCleanup = true; 666 return; 667 } 668 else 669 MarkUnsafe(Info); 670 } 671 else { 672 return MarkUnsafe(Info); 673 } 674 if (Info.isUnsafe) return; 675 } 676} 677 678/// RewriteBitCastUserOfAlloca - BCInst (transitively) bitcasts AI, or indexes 679/// to its first element. Transform users of the cast to use the new values 680/// instead. 681void SROA::RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI, 682 SmallVector<AllocaInst*, 32> &NewElts) { 683 Value::use_iterator UI = BCInst->use_begin(), UE = BCInst->use_end(); 684 while (UI != UE) { 685 Instruction *User = cast<Instruction>(*UI++); 686 if (BitCastInst *BCU = dyn_cast<BitCastInst>(User)) { 687 RewriteBitCastUserOfAlloca(BCU, AI, NewElts); 688 if (BCU->use_empty()) BCU->eraseFromParent(); 689 continue; 690 } 691 692 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 693 // This must be memcpy/memmove/memset of the entire aggregate. 694 // Split into one per element. 695 RewriteMemIntrinUserOfAlloca(MI, BCInst, AI, NewElts); 696 continue; 697 } 698 699 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 700 // If this is a store of the entire alloca from an integer, rewrite it. 701 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 702 continue; 703 } 704 705 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 706 // If this is a load of the entire alloca to an integer, rewrite it. 707 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 708 continue; 709 } 710 711 // Otherwise it must be some other user of a gep of the first pointer. Just 712 // leave these alone. 713 continue; 714 } 715} 716 717/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 718/// Rewrite it to copy or set the elements of the scalarized memory. 719void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst, 720 AllocationInst *AI, 721 SmallVector<AllocaInst*, 32> &NewElts) { 722 723 // If this is a memcpy/memmove, construct the other pointer as the 724 // appropriate type. The "Other" pointer is the pointer that goes to memory 725 // that doesn't have anything to do with the alloca that we are promoting. For 726 // memset, this Value* stays null. 727 Value *OtherPtr = 0; 728 unsigned MemAlignment = MI->getAlignment(); 729 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 730 if (BCInst == MTI->getRawDest()) 731 OtherPtr = MTI->getRawSource(); 732 else { 733 assert(BCInst == MTI->getRawSource()); 734 OtherPtr = MTI->getRawDest(); 735 } 736 } 737 738 // If there is an other pointer, we want to convert it to the same pointer 739 // type as AI has, so we can GEP through it safely. 740 if (OtherPtr) { 741 // It is likely that OtherPtr is a bitcast, if so, remove it. 742 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr)) 743 OtherPtr = BC->getOperand(0); 744 // All zero GEPs are effectively bitcasts. 745 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr)) 746 if (GEP->hasAllZeroIndices()) 747 OtherPtr = GEP->getOperand(0); 748 749 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr)) 750 if (BCE->getOpcode() == Instruction::BitCast) 751 OtherPtr = BCE->getOperand(0); 752 753 // If the pointer is not the right type, insert a bitcast to the right 754 // type. 755 if (OtherPtr->getType() != AI->getType()) 756 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(), 757 MI); 758 } 759 760 // Process each element of the aggregate. 761 Value *TheFn = MI->getOperand(0); 762 const Type *BytePtrTy = MI->getRawDest()->getType(); 763 bool SROADest = MI->getRawDest() == BCInst; 764 765 Constant *Zero = Constant::getNullValue(Type::Int32Ty); 766 767 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 768 // If this is a memcpy/memmove, emit a GEP of the other element address. 769 Value *OtherElt = 0; 770 unsigned OtherEltAlign = MemAlignment; 771 772 if (OtherPtr) { 773 Value *Idx[2] = { Zero, ConstantInt::get(Type::Int32Ty, i) }; 774 OtherElt = GetElementPtrInst::Create(OtherPtr, Idx, Idx + 2, 775 OtherPtr->getNameStr()+"."+utostr(i), 776 MI); 777 uint64_t EltOffset; 778 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 779 if (const StructType *ST = 780 dyn_cast<StructType>(OtherPtrTy->getElementType())) { 781 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 782 } else { 783 const Type *EltTy = 784 cast<SequentialType>(OtherPtr->getType())->getElementType(); 785 EltOffset = TD->getTypeAllocSize(EltTy)*i; 786 } 787 788 // The alignment of the other pointer is the guaranteed alignment of the 789 // element, which is affected by both the known alignment of the whole 790 // mem intrinsic and the alignment of the element. If the alignment of 791 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 792 // known alignment is just 4 bytes. 793 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 794 } 795 796 Value *EltPtr = NewElts[i]; 797 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 798 799 // If we got down to a scalar, insert a load or store as appropriate. 800 if (EltTy->isSingleValueType()) { 801 if (isa<MemTransferInst>(MI)) { 802 if (SROADest) { 803 // From Other to Alloca. 804 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 805 new StoreInst(Elt, EltPtr, MI); 806 } else { 807 // From Alloca to Other. 808 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 809 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 810 } 811 continue; 812 } 813 assert(isa<MemSetInst>(MI)); 814 815 // If the stored element is zero (common case), just store a null 816 // constant. 817 Constant *StoreVal; 818 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) { 819 if (CI->isZero()) { 820 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 821 } else { 822 // If EltTy is a vector type, get the element type. 823 const Type *ValTy = EltTy; 824 if (const VectorType *VTy = dyn_cast<VectorType>(ValTy)) 825 ValTy = VTy->getElementType(); 826 827 // Construct an integer with the right value. 828 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 829 APInt OneVal(EltSize, CI->getZExtValue()); 830 APInt TotalVal(OneVal); 831 // Set each byte. 832 for (unsigned i = 0; 8*i < EltSize; ++i) { 833 TotalVal = TotalVal.shl(8); 834 TotalVal |= OneVal; 835 } 836 837 // Convert the integer value to the appropriate type. 838 StoreVal = ConstantInt::get(TotalVal); 839 if (isa<PointerType>(ValTy)) 840 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 841 else if (ValTy->isFloatingPoint()) 842 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 843 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 844 845 // If the requested value was a vector constant, create it. 846 if (EltTy != ValTy) { 847 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 848 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 849 StoreVal = ConstantVector::get(&Elts[0], NumElts); 850 } 851 } 852 new StoreInst(StoreVal, EltPtr, MI); 853 continue; 854 } 855 // Otherwise, if we're storing a byte variable, use a memset call for 856 // this element. 857 } 858 859 // Cast the element pointer to BytePtrTy. 860 if (EltPtr->getType() != BytePtrTy) 861 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI); 862 863 // Cast the other pointer (if we have one) to BytePtrTy. 864 if (OtherElt && OtherElt->getType() != BytePtrTy) 865 OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(), 866 MI); 867 868 unsigned EltSize = TD->getTypeAllocSize(EltTy); 869 870 // Finally, insert the meminst for this element. 871 if (isa<MemTransferInst>(MI)) { 872 Value *Ops[] = { 873 SROADest ? EltPtr : OtherElt, // Dest ptr 874 SROADest ? OtherElt : EltPtr, // Src ptr 875 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 876 ConstantInt::get(Type::Int32Ty, OtherEltAlign) // Align 877 }; 878 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 879 } else { 880 assert(isa<MemSetInst>(MI)); 881 Value *Ops[] = { 882 EltPtr, MI->getOperand(2), // Dest, Value, 883 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 884 Zero // Align 885 }; 886 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 887 } 888 } 889 MI->eraseFromParent(); 890} 891 892/// RewriteStoreUserOfWholeAlloca - We found an store of an integer that 893/// overwrites the entire allocation. Extract out the pieces of the stored 894/// integer and store them individually. 895void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, 896 AllocationInst *AI, 897 SmallVector<AllocaInst*, 32> &NewElts){ 898 // Extract each element out of the integer according to its structure offset 899 // and store the element value to the individual alloca. 900 Value *SrcVal = SI->getOperand(0); 901 const Type *AllocaEltTy = AI->getType()->getElementType(); 902 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 903 904 // If this isn't a store of an integer to the whole alloca, it may be a store 905 // to the first element. Just ignore the store in this case and normal SROA 906 // will handle it. 907 if (!isa<IntegerType>(SrcVal->getType()) || 908 TD->getTypeAllocSizeInBits(SrcVal->getType()) != AllocaSizeBits) 909 return; 910 // Handle tail padding by extending the operand 911 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 912 SrcVal = new ZExtInst(SrcVal, IntegerType::get(AllocaSizeBits), "", SI); 913 914 DOUT << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << *SI; 915 916 // There are two forms here: AI could be an array or struct. Both cases 917 // have different ways to compute the element offset. 918 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 919 const StructLayout *Layout = TD->getStructLayout(EltSTy); 920 921 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 922 // Get the number of bits to shift SrcVal to get the value. 923 const Type *FieldTy = EltSTy->getElementType(i); 924 uint64_t Shift = Layout->getElementOffsetInBits(i); 925 926 if (TD->isBigEndian()) 927 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 928 929 Value *EltVal = SrcVal; 930 if (Shift) { 931 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 932 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 933 "sroa.store.elt", SI); 934 } 935 936 // Truncate down to an integer of the right size. 937 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 938 939 // Ignore zero sized fields like {}, they obviously contain no data. 940 if (FieldSizeBits == 0) continue; 941 942 if (FieldSizeBits != AllocaSizeBits) 943 EltVal = new TruncInst(EltVal, IntegerType::get(FieldSizeBits), "", SI); 944 Value *DestField = NewElts[i]; 945 if (EltVal->getType() == FieldTy) { 946 // Storing to an integer field of this size, just do it. 947 } else if (FieldTy->isFloatingPoint() || isa<VectorType>(FieldTy)) { 948 // Bitcast to the right element type (for fp/vector values). 949 EltVal = new BitCastInst(EltVal, FieldTy, "", SI); 950 } else { 951 // Otherwise, bitcast the dest pointer (for aggregates). 952 DestField = new BitCastInst(DestField, 953 PointerType::getUnqual(EltVal->getType()), 954 "", SI); 955 } 956 new StoreInst(EltVal, DestField, SI); 957 } 958 959 } else { 960 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 961 const Type *ArrayEltTy = ATy->getElementType(); 962 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 963 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 964 965 uint64_t Shift; 966 967 if (TD->isBigEndian()) 968 Shift = AllocaSizeBits-ElementOffset; 969 else 970 Shift = 0; 971 972 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 973 // Ignore zero sized fields like {}, they obviously contain no data. 974 if (ElementSizeBits == 0) continue; 975 976 Value *EltVal = SrcVal; 977 if (Shift) { 978 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 979 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 980 "sroa.store.elt", SI); 981 } 982 983 // Truncate down to an integer of the right size. 984 if (ElementSizeBits != AllocaSizeBits) 985 EltVal = new TruncInst(EltVal, IntegerType::get(ElementSizeBits),"",SI); 986 Value *DestField = NewElts[i]; 987 if (EltVal->getType() == ArrayEltTy) { 988 // Storing to an integer field of this size, just do it. 989 } else if (ArrayEltTy->isFloatingPoint() || isa<VectorType>(ArrayEltTy)) { 990 // Bitcast to the right element type (for fp/vector values). 991 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI); 992 } else { 993 // Otherwise, bitcast the dest pointer (for aggregates). 994 DestField = new BitCastInst(DestField, 995 PointerType::getUnqual(EltVal->getType()), 996 "", SI); 997 } 998 new StoreInst(EltVal, DestField, SI); 999 1000 if (TD->isBigEndian()) 1001 Shift -= ElementOffset; 1002 else 1003 Shift += ElementOffset; 1004 } 1005 } 1006 1007 SI->eraseFromParent(); 1008} 1009 1010/// RewriteLoadUserOfWholeAlloca - We found an load of the entire allocation to 1011/// an integer. Load the individual pieces to form the aggregate value. 1012void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI, 1013 SmallVector<AllocaInst*, 32> &NewElts) { 1014 // Extract each element out of the NewElts according to its structure offset 1015 // and form the result value. 1016 const Type *AllocaEltTy = AI->getType()->getElementType(); 1017 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1018 1019 // If this isn't a load of the whole alloca to an integer, it may be a load 1020 // of the first element. Just ignore the load in this case and normal SROA 1021 // will handle it. 1022 if (!isa<IntegerType>(LI->getType()) || 1023 TD->getTypeAllocSizeInBits(LI->getType()) != AllocaSizeBits) 1024 return; 1025 1026 DOUT << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << *LI; 1027 1028 // There are two forms here: AI could be an array or struct. Both cases 1029 // have different ways to compute the element offset. 1030 const StructLayout *Layout = 0; 1031 uint64_t ArrayEltBitOffset = 0; 1032 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1033 Layout = TD->getStructLayout(EltSTy); 1034 } else { 1035 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 1036 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1037 } 1038 1039 Value *ResultVal = Constant::getNullValue(IntegerType::get(AllocaSizeBits)); 1040 1041 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1042 // Load the value from the alloca. If the NewElt is an aggregate, cast 1043 // the pointer to an integer of the same size before doing the load. 1044 Value *SrcField = NewElts[i]; 1045 const Type *FieldTy = 1046 cast<PointerType>(SrcField->getType())->getElementType(); 1047 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1048 1049 // Ignore zero sized fields like {}, they obviously contain no data. 1050 if (FieldSizeBits == 0) continue; 1051 1052 const IntegerType *FieldIntTy = IntegerType::get(FieldSizeBits); 1053 if (!isa<IntegerType>(FieldTy) && !FieldTy->isFloatingPoint() && 1054 !isa<VectorType>(FieldTy)) 1055 SrcField = new BitCastInst(SrcField, PointerType::getUnqual(FieldIntTy), 1056 "", LI); 1057 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 1058 1059 // If SrcField is a fp or vector of the right size but that isn't an 1060 // integer type, bitcast to an integer so we can shift it. 1061 if (SrcField->getType() != FieldIntTy) 1062 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 1063 1064 // Zero extend the field to be the same size as the final alloca so that 1065 // we can shift and insert it. 1066 if (SrcField->getType() != ResultVal->getType()) 1067 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 1068 1069 // Determine the number of bits to shift SrcField. 1070 uint64_t Shift; 1071 if (Layout) // Struct case. 1072 Shift = Layout->getElementOffsetInBits(i); 1073 else // Array case. 1074 Shift = i*ArrayEltBitOffset; 1075 1076 if (TD->isBigEndian()) 1077 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 1078 1079 if (Shift) { 1080 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 1081 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 1082 } 1083 1084 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 1085 } 1086 1087 // Handle tail padding by truncating the result 1088 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 1089 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 1090 1091 LI->replaceAllUsesWith(ResultVal); 1092 LI->eraseFromParent(); 1093} 1094 1095 1096/// HasPadding - Return true if the specified type has any structure or 1097/// alignment padding, false otherwise. 1098static bool HasPadding(const Type *Ty, const TargetData &TD) { 1099 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 1100 const StructLayout *SL = TD.getStructLayout(STy); 1101 unsigned PrevFieldBitOffset = 0; 1102 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1103 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 1104 1105 // Padding in sub-elements? 1106 if (HasPadding(STy->getElementType(i), TD)) 1107 return true; 1108 1109 // Check to see if there is any padding between this element and the 1110 // previous one. 1111 if (i) { 1112 unsigned PrevFieldEnd = 1113 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 1114 if (PrevFieldEnd < FieldBitOffset) 1115 return true; 1116 } 1117 1118 PrevFieldBitOffset = FieldBitOffset; 1119 } 1120 1121 // Check for tail padding. 1122 if (unsigned EltCount = STy->getNumElements()) { 1123 unsigned PrevFieldEnd = PrevFieldBitOffset + 1124 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 1125 if (PrevFieldEnd < SL->getSizeInBits()) 1126 return true; 1127 } 1128 1129 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1130 return HasPadding(ATy->getElementType(), TD); 1131 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) { 1132 return HasPadding(VTy->getElementType(), TD); 1133 } 1134 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 1135} 1136 1137/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 1138/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 1139/// or 1 if safe after canonicalization has been performed. 1140/// 1141int SROA::isSafeAllocaToScalarRepl(AllocationInst *AI) { 1142 // Loop over the use list of the alloca. We can only transform it if all of 1143 // the users are safe to transform. 1144 AllocaInfo Info; 1145 1146 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); 1147 I != E; ++I) { 1148 isSafeUseOfAllocation(cast<Instruction>(*I), AI, Info); 1149 if (Info.isUnsafe) { 1150 DOUT << "Cannot transform: " << *AI << " due to user: " << **I; 1151 return 0; 1152 } 1153 } 1154 1155 // Okay, we know all the users are promotable. If the aggregate is a memcpy 1156 // source and destination, we have to be careful. In particular, the memcpy 1157 // could be moving around elements that live in structure padding of the LLVM 1158 // types, but may actually be used. In these cases, we refuse to promote the 1159 // struct. 1160 if (Info.isMemCpySrc && Info.isMemCpyDst && 1161 HasPadding(AI->getType()->getElementType(), *TD)) 1162 return 0; 1163 1164 // If we require cleanup, return 1, otherwise return 3. 1165 return Info.needsCleanup ? 1 : 3; 1166} 1167 1168/// CleanupGEP - GEP is used by an Alloca, which can be prompted after the GEP 1169/// is canonicalized here. 1170void SROA::CleanupGEP(GetElementPtrInst *GEPI) { 1171 gep_type_iterator I = gep_type_begin(GEPI); 1172 ++I; 1173 1174 const ArrayType *AT = dyn_cast<ArrayType>(*I); 1175 if (!AT) 1176 return; 1177 1178 uint64_t NumElements = AT->getNumElements(); 1179 1180 if (isa<ConstantInt>(I.getOperand())) 1181 return; 1182 1183 if (NumElements == 1) { 1184 GEPI->setOperand(2, Constant::getNullValue(Type::Int32Ty)); 1185 return; 1186 } 1187 1188 assert(NumElements == 2 && "Unhandled case!"); 1189 // All users of the GEP must be loads. At each use of the GEP, insert 1190 // two loads of the appropriate indexed GEP and select between them. 1191 Value *IsOne = new ICmpInst(ICmpInst::ICMP_NE, I.getOperand(), 1192 Constant::getNullValue(I.getOperand()->getType()), 1193 "isone", GEPI); 1194 // Insert the new GEP instructions, which are properly indexed. 1195 SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end()); 1196 Indices[1] = Constant::getNullValue(Type::Int32Ty); 1197 Value *ZeroIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 1198 Indices.begin(), 1199 Indices.end(), 1200 GEPI->getName()+".0", GEPI); 1201 Indices[1] = ConstantInt::get(Type::Int32Ty, 1); 1202 Value *OneIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 1203 Indices.begin(), 1204 Indices.end(), 1205 GEPI->getName()+".1", GEPI); 1206 // Replace all loads of the variable index GEP with loads from both 1207 // indexes and a select. 1208 while (!GEPI->use_empty()) { 1209 LoadInst *LI = cast<LoadInst>(GEPI->use_back()); 1210 Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI); 1211 Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI); 1212 Value *R = SelectInst::Create(IsOne, One, Zero, LI->getName(), LI); 1213 LI->replaceAllUsesWith(R); 1214 LI->eraseFromParent(); 1215 } 1216 GEPI->eraseFromParent(); 1217} 1218 1219 1220/// CleanupAllocaUsers - If SROA reported that it can promote the specified 1221/// allocation, but only if cleaned up, perform the cleanups required. 1222void SROA::CleanupAllocaUsers(AllocationInst *AI) { 1223 // At this point, we know that the end result will be SROA'd and promoted, so 1224 // we can insert ugly code if required so long as sroa+mem2reg will clean it 1225 // up. 1226 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 1227 UI != E; ) { 1228 User *U = *UI++; 1229 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) 1230 CleanupGEP(GEPI);
| 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/Pass.h" 31#include "llvm/Analysis/Dominators.h" 32#include "llvm/Target/TargetData.h" 33#include "llvm/Transforms/Utils/PromoteMemToReg.h" 34#include "llvm/Transforms/Utils/Local.h" 35#include "llvm/Support/Debug.h" 36#include "llvm/Support/GetElementPtrTypeIterator.h" 37#include "llvm/Support/IRBuilder.h" 38#include "llvm/Support/MathExtras.h" 39#include "llvm/Support/Compiler.h" 40#include "llvm/ADT/SmallVector.h" 41#include "llvm/ADT/Statistic.h" 42#include "llvm/ADT/StringExtras.h" 43using namespace llvm; 44 45STATISTIC(NumReplaced, "Number of allocas broken up"); 46STATISTIC(NumPromoted, "Number of allocas promoted"); 47STATISTIC(NumConverted, "Number of aggregates converted to scalar"); 48STATISTIC(NumGlobals, "Number of allocas copied from constant global"); 49 50namespace { 51 struct VISIBILITY_HIDDEN SROA : public FunctionPass { 52 static char ID; // Pass identification, replacement for typeid 53 explicit SROA(signed T = -1) : FunctionPass(&ID) { 54 if (T == -1) 55 SRThreshold = 128; 56 else 57 SRThreshold = T; 58 } 59 60 bool runOnFunction(Function &F); 61 62 bool performScalarRepl(Function &F); 63 bool performPromotion(Function &F); 64 65 // getAnalysisUsage - This pass does not require any passes, but we know it 66 // will not alter the CFG, so say so. 67 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 68 AU.addRequired<DominatorTree>(); 69 AU.addRequired<DominanceFrontier>(); 70 AU.addRequired<TargetData>(); 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(AllocationInst *AI); 104 105 void isSafeUseOfAllocation(Instruction *User, AllocationInst *AI, 106 AllocaInfo &Info); 107 void isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI, 108 AllocaInfo &Info); 109 void isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI, 110 unsigned OpNo, AllocaInfo &Info); 111 void isSafeUseOfBitCastedAllocation(BitCastInst *User, AllocationInst *AI, 112 AllocaInfo &Info); 113 114 void DoScalarReplacement(AllocationInst *AI, 115 std::vector<AllocationInst*> &WorkList); 116 void CleanupGEP(GetElementPtrInst *GEP); 117 void CleanupAllocaUsers(AllocationInst *AI); 118 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocationInst *Base); 119 120 void RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI, 121 SmallVector<AllocaInst*, 32> &NewElts); 122 123 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst, 124 AllocationInst *AI, 125 SmallVector<AllocaInst*, 32> &NewElts); 126 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocationInst *AI, 127 SmallVector<AllocaInst*, 32> &NewElts); 128 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *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(AllocationInst *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 = &getAnalysis<TargetData>(); 153 154 bool Changed = performPromotion(F); 155 while (1) { 156 bool LocalChange = performScalarRepl(F); 157 if (!LocalChange) break; // No need to repromote if no scalarrepl 158 Changed = true; 159 LocalChange = performPromotion(F); 160 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 161 } 162 163 return Changed; 164} 165 166 167bool SROA::performPromotion(Function &F) { 168 std::vector<AllocaInst*> Allocas; 169 DominatorTree &DT = getAnalysis<DominatorTree>(); 170 DominanceFrontier &DF = getAnalysis<DominanceFrontier>(); 171 172 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 173 174 bool Changed = false; 175 176 while (1) { 177 Allocas.clear(); 178 179 // Find allocas that are safe to promote, by looking at all instructions in 180 // the entry node 181 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 182 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 183 if (isAllocaPromotable(AI)) 184 Allocas.push_back(AI); 185 186 if (Allocas.empty()) break; 187 188 PromoteMemToReg(Allocas, DT, DF); 189 NumPromoted += Allocas.size(); 190 Changed = true; 191 } 192 193 return Changed; 194} 195 196/// getNumSAElements - Return the number of elements in the specific struct or 197/// array. 198static uint64_t getNumSAElements(const Type *T) { 199 if (const StructType *ST = dyn_cast<StructType>(T)) 200 return ST->getNumElements(); 201 return cast<ArrayType>(T)->getNumElements(); 202} 203 204// performScalarRepl - This algorithm is a simple worklist driven algorithm, 205// which runs on all of the malloc/alloca instructions in the function, removing 206// them if they are only used by getelementptr instructions. 207// 208bool SROA::performScalarRepl(Function &F) { 209 std::vector<AllocationInst*> WorkList; 210 211 // Scan the entry basic block, adding any alloca's and mallocs to the worklist 212 BasicBlock &BB = F.getEntryBlock(); 213 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 214 if (AllocationInst *A = dyn_cast<AllocationInst>(I)) 215 WorkList.push_back(A); 216 217 // Process the worklist 218 bool Changed = false; 219 while (!WorkList.empty()) { 220 AllocationInst *AI = WorkList.back(); 221 WorkList.pop_back(); 222 223 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 224 // with unused elements. 225 if (AI->use_empty()) { 226 AI->eraseFromParent(); 227 continue; 228 } 229 230 // If this alloca is impossible for us to promote, reject it early. 231 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 232 continue; 233 234 // Check to see if this allocation is only modified by a memcpy/memmove from 235 // a constant global. If this is the case, we can change all users to use 236 // the constant global instead. This is commonly produced by the CFE by 237 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 238 // is only subsequently read. 239 if (Instruction *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 240 DOUT << "Found alloca equal to global: " << *AI; 241 DOUT << " memcpy = " << *TheCopy; 242 Constant *TheSrc = cast<Constant>(TheCopy->getOperand(2)); 243 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 244 TheCopy->eraseFromParent(); // Don't mutate the global. 245 AI->eraseFromParent(); 246 ++NumGlobals; 247 Changed = true; 248 continue; 249 } 250 251 // Check to see if we can perform the core SROA transformation. We cannot 252 // transform the allocation instruction if it is an array allocation 253 // (allocations OF arrays are ok though), and an allocation of a scalar 254 // value cannot be decomposed at all. 255 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 256 257 // Do not promote any struct whose size is too big. 258 if (AllocaSize > SRThreshold) continue; 259 260 if ((isa<StructType>(AI->getAllocatedType()) || 261 isa<ArrayType>(AI->getAllocatedType())) && 262 // Do not promote any struct into more than "32" separate vars. 263 getNumSAElements(AI->getAllocatedType()) <= SRThreshold/4) { 264 // Check that all of the users of the allocation are capable of being 265 // transformed. 266 switch (isSafeAllocaToScalarRepl(AI)) { 267 default: assert(0 && "Unexpected value!"); 268 case 0: // Not safe to scalar replace. 269 break; 270 case 1: // Safe, but requires cleanup/canonicalizations first 271 CleanupAllocaUsers(AI); 272 // FALL THROUGH. 273 case 3: // Safe to scalar replace. 274 DoScalarReplacement(AI, WorkList); 275 Changed = true; 276 continue; 277 } 278 } 279 280 // If we can turn this aggregate value (potentially with casts) into a 281 // simple scalar value that can be mem2reg'd into a register value. 282 // IsNotTrivial tracks whether this is something that mem2reg could have 283 // promoted itself. If so, we don't want to transform it needlessly. Note 284 // that we can't just check based on the type: the alloca may be of an i32 285 // but that has pointer arithmetic to set byte 3 of it or something. 286 bool IsNotTrivial = false; 287 const Type *VectorTy = 0; 288 bool HadAVector = false; 289 if (CanConvertToScalar(AI, IsNotTrivial, VectorTy, HadAVector, 290 0, unsigned(AllocaSize)) && IsNotTrivial) { 291 AllocaInst *NewAI; 292 // If we were able to find a vector type that can handle this with 293 // insert/extract elements, and if there was at least one use that had 294 // a vector type, promote this to a vector. We don't want to promote 295 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 296 // we just get a lot of insert/extracts. If at least one vector is 297 // involved, then we probably really do have a union of vector/array. 298 if (VectorTy && isa<VectorType>(VectorTy) && HadAVector) { 299 DOUT << "CONVERT TO VECTOR: " << *AI << " TYPE = " << *VectorTy <<"\n"; 300 301 // Create and insert the vector alloca. 302 NewAI = new AllocaInst(VectorTy, 0, "", AI->getParent()->begin()); 303 ConvertUsesToScalar(AI, NewAI, 0); 304 } else { 305 DOUT << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"; 306 307 // Create and insert the integer alloca. 308 const Type *NewTy = IntegerType::get(AllocaSize*8); 309 NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 310 ConvertUsesToScalar(AI, NewAI, 0); 311 } 312 NewAI->takeName(AI); 313 AI->eraseFromParent(); 314 ++NumConverted; 315 Changed = true; 316 continue; 317 } 318 319 // Otherwise, couldn't process this alloca. 320 } 321 322 return Changed; 323} 324 325/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 326/// predicate, do SROA now. 327void SROA::DoScalarReplacement(AllocationInst *AI, 328 std::vector<AllocationInst*> &WorkList) { 329 DOUT << "Found inst to SROA: " << *AI; 330 SmallVector<AllocaInst*, 32> ElementAllocas; 331 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 332 ElementAllocas.reserve(ST->getNumContainedTypes()); 333 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 334 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 335 AI->getAlignment(), 336 AI->getName() + "." + utostr(i), AI); 337 ElementAllocas.push_back(NA); 338 WorkList.push_back(NA); // Add to worklist for recursive processing 339 } 340 } else { 341 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 342 ElementAllocas.reserve(AT->getNumElements()); 343 const Type *ElTy = AT->getElementType(); 344 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 345 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 346 AI->getName() + "." + utostr(i), AI); 347 ElementAllocas.push_back(NA); 348 WorkList.push_back(NA); // Add to worklist for recursive processing 349 } 350 } 351 352 // Now that we have created the alloca instructions that we want to use, 353 // expand the getelementptr instructions to use them. 354 // 355 while (!AI->use_empty()) { 356 Instruction *User = cast<Instruction>(AI->use_back()); 357 if (BitCastInst *BCInst = dyn_cast<BitCastInst>(User)) { 358 RewriteBitCastUserOfAlloca(BCInst, AI, ElementAllocas); 359 BCInst->eraseFromParent(); 360 continue; 361 } 362 363 // Replace: 364 // %res = load { i32, i32 }* %alloc 365 // with: 366 // %load.0 = load i32* %alloc.0 367 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 368 // %load.1 = load i32* %alloc.1 369 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 370 // (Also works for arrays instead of structs) 371 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 372 Value *Insert = UndefValue::get(LI->getType()); 373 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) { 374 Value *Load = new LoadInst(ElementAllocas[i], "load", LI); 375 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 376 } 377 LI->replaceAllUsesWith(Insert); 378 LI->eraseFromParent(); 379 continue; 380 } 381 382 // Replace: 383 // store { i32, i32 } %val, { i32, i32 }* %alloc 384 // with: 385 // %val.0 = extractvalue { i32, i32 } %val, 0 386 // store i32 %val.0, i32* %alloc.0 387 // %val.1 = extractvalue { i32, i32 } %val, 1 388 // store i32 %val.1, i32* %alloc.1 389 // (Also works for arrays instead of structs) 390 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 391 Value *Val = SI->getOperand(0); 392 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) { 393 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 394 new StoreInst(Extract, ElementAllocas[i], SI); 395 } 396 SI->eraseFromParent(); 397 continue; 398 } 399 400 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User); 401 // We now know that the GEP is of the form: GEP <ptr>, 0, <cst> 402 unsigned Idx = 403 (unsigned)cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue(); 404 405 assert(Idx < ElementAllocas.size() && "Index out of range?"); 406 AllocaInst *AllocaToUse = ElementAllocas[Idx]; 407 408 Value *RepValue; 409 if (GEPI->getNumOperands() == 3) { 410 // Do not insert a new getelementptr instruction with zero indices, only 411 // to have it optimized out later. 412 RepValue = AllocaToUse; 413 } else { 414 // We are indexing deeply into the structure, so we still need a 415 // getelement ptr instruction to finish the indexing. This may be 416 // expanded itself once the worklist is rerun. 417 // 418 SmallVector<Value*, 8> NewArgs; 419 NewArgs.push_back(Constant::getNullValue(Type::Int32Ty)); 420 NewArgs.append(GEPI->op_begin()+3, GEPI->op_end()); 421 RepValue = GetElementPtrInst::Create(AllocaToUse, NewArgs.begin(), 422 NewArgs.end(), "", GEPI); 423 RepValue->takeName(GEPI); 424 } 425 426 // If this GEP is to the start of the aggregate, check for memcpys. 427 if (Idx == 0 && GEPI->hasAllZeroIndices()) 428 RewriteBitCastUserOfAlloca(GEPI, AI, ElementAllocas); 429 430 // Move all of the users over to the new GEP. 431 GEPI->replaceAllUsesWith(RepValue); 432 // Delete the old GEP 433 GEPI->eraseFromParent(); 434 } 435 436 // Finally, delete the Alloca instruction 437 AI->eraseFromParent(); 438 NumReplaced++; 439} 440 441 442/// isSafeElementUse - Check to see if this use is an allowed use for a 443/// getelementptr instruction of an array aggregate allocation. isFirstElt 444/// indicates whether Ptr is known to the start of the aggregate. 445/// 446void SROA::isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI, 447 AllocaInfo &Info) { 448 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 449 I != E; ++I) { 450 Instruction *User = cast<Instruction>(*I); 451 switch (User->getOpcode()) { 452 case Instruction::Load: break; 453 case Instruction::Store: 454 // Store is ok if storing INTO the pointer, not storing the pointer 455 if (User->getOperand(0) == Ptr) return MarkUnsafe(Info); 456 break; 457 case Instruction::GetElementPtr: { 458 GetElementPtrInst *GEP = cast<GetElementPtrInst>(User); 459 bool AreAllZeroIndices = isFirstElt; 460 if (GEP->getNumOperands() > 1) { 461 if (!isa<ConstantInt>(GEP->getOperand(1)) || 462 !cast<ConstantInt>(GEP->getOperand(1))->isZero()) 463 // Using pointer arithmetic to navigate the array. 464 return MarkUnsafe(Info); 465 466 if (AreAllZeroIndices) 467 AreAllZeroIndices = GEP->hasAllZeroIndices(); 468 } 469 isSafeElementUse(GEP, AreAllZeroIndices, AI, Info); 470 if (Info.isUnsafe) return; 471 break; 472 } 473 case Instruction::BitCast: 474 if (isFirstElt) { 475 isSafeUseOfBitCastedAllocation(cast<BitCastInst>(User), AI, Info); 476 if (Info.isUnsafe) return; 477 break; 478 } 479 DOUT << " Transformation preventing inst: " << *User; 480 return MarkUnsafe(Info); 481 case Instruction::Call: 482 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 483 if (isFirstElt) { 484 isSafeMemIntrinsicOnAllocation(MI, AI, I.getOperandNo(), Info); 485 if (Info.isUnsafe) return; 486 break; 487 } 488 } 489 DOUT << " Transformation preventing inst: " << *User; 490 return MarkUnsafe(Info); 491 default: 492 DOUT << " Transformation preventing inst: " << *User; 493 return MarkUnsafe(Info); 494 } 495 } 496 return; // All users look ok :) 497} 498 499/// AllUsersAreLoads - Return true if all users of this value are loads. 500static bool AllUsersAreLoads(Value *Ptr) { 501 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 502 I != E; ++I) 503 if (cast<Instruction>(*I)->getOpcode() != Instruction::Load) 504 return false; 505 return true; 506} 507 508/// isSafeUseOfAllocation - Check to see if this user is an allowed use for an 509/// aggregate allocation. 510/// 511void SROA::isSafeUseOfAllocation(Instruction *User, AllocationInst *AI, 512 AllocaInfo &Info) { 513 if (BitCastInst *C = dyn_cast<BitCastInst>(User)) 514 return isSafeUseOfBitCastedAllocation(C, AI, Info); 515 516 if (LoadInst *LI = dyn_cast<LoadInst>(User)) 517 if (!LI->isVolatile()) 518 return;// Loads (returning a first class aggregrate) are always rewritable 519 520 if (StoreInst *SI = dyn_cast<StoreInst>(User)) 521 if (!SI->isVolatile() && SI->getOperand(0) != AI) 522 return;// Store is ok if storing INTO the pointer, not storing the pointer 523 524 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User); 525 if (GEPI == 0) 526 return MarkUnsafe(Info); 527 528 gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI); 529 530 // The GEP is not safe to transform if not of the form "GEP <ptr>, 0, <cst>". 531 if (I == E || 532 I.getOperand() != Constant::getNullValue(I.getOperand()->getType())) { 533 return MarkUnsafe(Info); 534 } 535 536 ++I; 537 if (I == E) return MarkUnsafe(Info); // ran out of GEP indices?? 538 539 bool IsAllZeroIndices = true; 540 541 // If the first index is a non-constant index into an array, see if we can 542 // handle it as a special case. 543 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 544 if (!isa<ConstantInt>(I.getOperand())) { 545 IsAllZeroIndices = 0; 546 uint64_t NumElements = AT->getNumElements(); 547 548 // If this is an array index and the index is not constant, we cannot 549 // promote... that is unless the array has exactly one or two elements in 550 // it, in which case we CAN promote it, but we have to canonicalize this 551 // out if this is the only problem. 552 if ((NumElements == 1 || NumElements == 2) && 553 AllUsersAreLoads(GEPI)) { 554 Info.needsCleanup = true; 555 return; // Canonicalization required! 556 } 557 return MarkUnsafe(Info); 558 } 559 } 560 561 // Walk through the GEP type indices, checking the types that this indexes 562 // into. 563 for (; I != E; ++I) { 564 // Ignore struct elements, no extra checking needed for these. 565 if (isa<StructType>(*I)) 566 continue; 567 568 ConstantInt *IdxVal = dyn_cast<ConstantInt>(I.getOperand()); 569 if (!IdxVal) return MarkUnsafe(Info); 570 571 // Are all indices still zero? 572 IsAllZeroIndices &= IdxVal->isZero(); 573 574 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 575 // This GEP indexes an array. Verify that this is an in-range constant 576 // integer. Specifically, consider A[0][i]. We cannot know that the user 577 // isn't doing invalid things like allowing i to index an out-of-range 578 // subscript that accesses A[1]. Because of this, we have to reject SROA 579 // of any accesses into structs where any of the components are variables. 580 if (IdxVal->getZExtValue() >= AT->getNumElements()) 581 return MarkUnsafe(Info); 582 } else if (const VectorType *VT = dyn_cast<VectorType>(*I)) { 583 if (IdxVal->getZExtValue() >= VT->getNumElements()) 584 return MarkUnsafe(Info); 585 } 586 } 587 588 // If there are any non-simple uses of this getelementptr, make sure to reject 589 // them. 590 return isSafeElementUse(GEPI, IsAllZeroIndices, AI, Info); 591} 592 593/// isSafeMemIntrinsicOnAllocation - Return true if the specified memory 594/// intrinsic can be promoted by SROA. At this point, we know that the operand 595/// of the memintrinsic is a pointer to the beginning of the allocation. 596void SROA::isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI, 597 unsigned OpNo, AllocaInfo &Info) { 598 // If not constant length, give up. 599 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 600 if (!Length) return MarkUnsafe(Info); 601 602 // If not the whole aggregate, give up. 603 if (Length->getZExtValue() != 604 TD->getTypeAllocSize(AI->getType()->getElementType())) 605 return MarkUnsafe(Info); 606 607 // We only know about memcpy/memset/memmove. 608 if (!isa<MemIntrinsic>(MI)) 609 return MarkUnsafe(Info); 610 611 // Otherwise, we can transform it. Determine whether this is a memcpy/set 612 // into or out of the aggregate. 613 if (OpNo == 1) 614 Info.isMemCpyDst = true; 615 else { 616 assert(OpNo == 2); 617 Info.isMemCpySrc = true; 618 } 619} 620 621/// isSafeUseOfBitCastedAllocation - Return true if all users of this bitcast 622/// are 623void SROA::isSafeUseOfBitCastedAllocation(BitCastInst *BC, AllocationInst *AI, 624 AllocaInfo &Info) { 625 for (Value::use_iterator UI = BC->use_begin(), E = BC->use_end(); 626 UI != E; ++UI) { 627 if (BitCastInst *BCU = dyn_cast<BitCastInst>(UI)) { 628 isSafeUseOfBitCastedAllocation(BCU, AI, Info); 629 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) { 630 isSafeMemIntrinsicOnAllocation(MI, AI, UI.getOperandNo(), Info); 631 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI)) { 632 if (SI->isVolatile()) 633 return MarkUnsafe(Info); 634 635 // If storing the entire alloca in one chunk through a bitcasted pointer 636 // to integer, we can transform it. This happens (for example) when you 637 // cast a {i32,i32}* to i64* and store through it. This is similar to the 638 // memcpy case and occurs in various "byval" cases and emulated memcpys. 639 if (isa<IntegerType>(SI->getOperand(0)->getType()) && 640 TD->getTypeAllocSize(SI->getOperand(0)->getType()) == 641 TD->getTypeAllocSize(AI->getType()->getElementType())) { 642 Info.isMemCpyDst = true; 643 continue; 644 } 645 return MarkUnsafe(Info); 646 } else if (LoadInst *LI = dyn_cast<LoadInst>(UI)) { 647 if (LI->isVolatile()) 648 return MarkUnsafe(Info); 649 650 // If loading the entire alloca in one chunk through a bitcasted pointer 651 // to integer, we can transform it. This happens (for example) when you 652 // cast a {i32,i32}* to i64* and load through it. This is similar to the 653 // memcpy case and occurs in various "byval" cases and emulated memcpys. 654 if (isa<IntegerType>(LI->getType()) && 655 TD->getTypeAllocSize(LI->getType()) == 656 TD->getTypeAllocSize(AI->getType()->getElementType())) { 657 Info.isMemCpySrc = true; 658 continue; 659 } 660 return MarkUnsafe(Info); 661 } else if (isa<DbgInfoIntrinsic>(UI)) { 662 // If one user is DbgInfoIntrinsic then check if all users are 663 // DbgInfoIntrinsics. 664 if (OnlyUsedByDbgInfoIntrinsics(BC)) { 665 Info.needsCleanup = true; 666 return; 667 } 668 else 669 MarkUnsafe(Info); 670 } 671 else { 672 return MarkUnsafe(Info); 673 } 674 if (Info.isUnsafe) return; 675 } 676} 677 678/// RewriteBitCastUserOfAlloca - BCInst (transitively) bitcasts AI, or indexes 679/// to its first element. Transform users of the cast to use the new values 680/// instead. 681void SROA::RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI, 682 SmallVector<AllocaInst*, 32> &NewElts) { 683 Value::use_iterator UI = BCInst->use_begin(), UE = BCInst->use_end(); 684 while (UI != UE) { 685 Instruction *User = cast<Instruction>(*UI++); 686 if (BitCastInst *BCU = dyn_cast<BitCastInst>(User)) { 687 RewriteBitCastUserOfAlloca(BCU, AI, NewElts); 688 if (BCU->use_empty()) BCU->eraseFromParent(); 689 continue; 690 } 691 692 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 693 // This must be memcpy/memmove/memset of the entire aggregate. 694 // Split into one per element. 695 RewriteMemIntrinUserOfAlloca(MI, BCInst, AI, NewElts); 696 continue; 697 } 698 699 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 700 // If this is a store of the entire alloca from an integer, rewrite it. 701 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 702 continue; 703 } 704 705 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 706 // If this is a load of the entire alloca to an integer, rewrite it. 707 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 708 continue; 709 } 710 711 // Otherwise it must be some other user of a gep of the first pointer. Just 712 // leave these alone. 713 continue; 714 } 715} 716 717/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 718/// Rewrite it to copy or set the elements of the scalarized memory. 719void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst, 720 AllocationInst *AI, 721 SmallVector<AllocaInst*, 32> &NewElts) { 722 723 // If this is a memcpy/memmove, construct the other pointer as the 724 // appropriate type. The "Other" pointer is the pointer that goes to memory 725 // that doesn't have anything to do with the alloca that we are promoting. For 726 // memset, this Value* stays null. 727 Value *OtherPtr = 0; 728 unsigned MemAlignment = MI->getAlignment(); 729 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 730 if (BCInst == MTI->getRawDest()) 731 OtherPtr = MTI->getRawSource(); 732 else { 733 assert(BCInst == MTI->getRawSource()); 734 OtherPtr = MTI->getRawDest(); 735 } 736 } 737 738 // If there is an other pointer, we want to convert it to the same pointer 739 // type as AI has, so we can GEP through it safely. 740 if (OtherPtr) { 741 // It is likely that OtherPtr is a bitcast, if so, remove it. 742 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr)) 743 OtherPtr = BC->getOperand(0); 744 // All zero GEPs are effectively bitcasts. 745 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr)) 746 if (GEP->hasAllZeroIndices()) 747 OtherPtr = GEP->getOperand(0); 748 749 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr)) 750 if (BCE->getOpcode() == Instruction::BitCast) 751 OtherPtr = BCE->getOperand(0); 752 753 // If the pointer is not the right type, insert a bitcast to the right 754 // type. 755 if (OtherPtr->getType() != AI->getType()) 756 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(), 757 MI); 758 } 759 760 // Process each element of the aggregate. 761 Value *TheFn = MI->getOperand(0); 762 const Type *BytePtrTy = MI->getRawDest()->getType(); 763 bool SROADest = MI->getRawDest() == BCInst; 764 765 Constant *Zero = Constant::getNullValue(Type::Int32Ty); 766 767 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 768 // If this is a memcpy/memmove, emit a GEP of the other element address. 769 Value *OtherElt = 0; 770 unsigned OtherEltAlign = MemAlignment; 771 772 if (OtherPtr) { 773 Value *Idx[2] = { Zero, ConstantInt::get(Type::Int32Ty, i) }; 774 OtherElt = GetElementPtrInst::Create(OtherPtr, Idx, Idx + 2, 775 OtherPtr->getNameStr()+"."+utostr(i), 776 MI); 777 uint64_t EltOffset; 778 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 779 if (const StructType *ST = 780 dyn_cast<StructType>(OtherPtrTy->getElementType())) { 781 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 782 } else { 783 const Type *EltTy = 784 cast<SequentialType>(OtherPtr->getType())->getElementType(); 785 EltOffset = TD->getTypeAllocSize(EltTy)*i; 786 } 787 788 // The alignment of the other pointer is the guaranteed alignment of the 789 // element, which is affected by both the known alignment of the whole 790 // mem intrinsic and the alignment of the element. If the alignment of 791 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 792 // known alignment is just 4 bytes. 793 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 794 } 795 796 Value *EltPtr = NewElts[i]; 797 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 798 799 // If we got down to a scalar, insert a load or store as appropriate. 800 if (EltTy->isSingleValueType()) { 801 if (isa<MemTransferInst>(MI)) { 802 if (SROADest) { 803 // From Other to Alloca. 804 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 805 new StoreInst(Elt, EltPtr, MI); 806 } else { 807 // From Alloca to Other. 808 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 809 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 810 } 811 continue; 812 } 813 assert(isa<MemSetInst>(MI)); 814 815 // If the stored element is zero (common case), just store a null 816 // constant. 817 Constant *StoreVal; 818 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) { 819 if (CI->isZero()) { 820 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 821 } else { 822 // If EltTy is a vector type, get the element type. 823 const Type *ValTy = EltTy; 824 if (const VectorType *VTy = dyn_cast<VectorType>(ValTy)) 825 ValTy = VTy->getElementType(); 826 827 // Construct an integer with the right value. 828 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 829 APInt OneVal(EltSize, CI->getZExtValue()); 830 APInt TotalVal(OneVal); 831 // Set each byte. 832 for (unsigned i = 0; 8*i < EltSize; ++i) { 833 TotalVal = TotalVal.shl(8); 834 TotalVal |= OneVal; 835 } 836 837 // Convert the integer value to the appropriate type. 838 StoreVal = ConstantInt::get(TotalVal); 839 if (isa<PointerType>(ValTy)) 840 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 841 else if (ValTy->isFloatingPoint()) 842 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 843 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 844 845 // If the requested value was a vector constant, create it. 846 if (EltTy != ValTy) { 847 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 848 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 849 StoreVal = ConstantVector::get(&Elts[0], NumElts); 850 } 851 } 852 new StoreInst(StoreVal, EltPtr, MI); 853 continue; 854 } 855 // Otherwise, if we're storing a byte variable, use a memset call for 856 // this element. 857 } 858 859 // Cast the element pointer to BytePtrTy. 860 if (EltPtr->getType() != BytePtrTy) 861 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI); 862 863 // Cast the other pointer (if we have one) to BytePtrTy. 864 if (OtherElt && OtherElt->getType() != BytePtrTy) 865 OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(), 866 MI); 867 868 unsigned EltSize = TD->getTypeAllocSize(EltTy); 869 870 // Finally, insert the meminst for this element. 871 if (isa<MemTransferInst>(MI)) { 872 Value *Ops[] = { 873 SROADest ? EltPtr : OtherElt, // Dest ptr 874 SROADest ? OtherElt : EltPtr, // Src ptr 875 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 876 ConstantInt::get(Type::Int32Ty, OtherEltAlign) // Align 877 }; 878 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 879 } else { 880 assert(isa<MemSetInst>(MI)); 881 Value *Ops[] = { 882 EltPtr, MI->getOperand(2), // Dest, Value, 883 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 884 Zero // Align 885 }; 886 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 887 } 888 } 889 MI->eraseFromParent(); 890} 891 892/// RewriteStoreUserOfWholeAlloca - We found an store of an integer that 893/// overwrites the entire allocation. Extract out the pieces of the stored 894/// integer and store them individually. 895void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, 896 AllocationInst *AI, 897 SmallVector<AllocaInst*, 32> &NewElts){ 898 // Extract each element out of the integer according to its structure offset 899 // and store the element value to the individual alloca. 900 Value *SrcVal = SI->getOperand(0); 901 const Type *AllocaEltTy = AI->getType()->getElementType(); 902 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 903 904 // If this isn't a store of an integer to the whole alloca, it may be a store 905 // to the first element. Just ignore the store in this case and normal SROA 906 // will handle it. 907 if (!isa<IntegerType>(SrcVal->getType()) || 908 TD->getTypeAllocSizeInBits(SrcVal->getType()) != AllocaSizeBits) 909 return; 910 // Handle tail padding by extending the operand 911 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 912 SrcVal = new ZExtInst(SrcVal, IntegerType::get(AllocaSizeBits), "", SI); 913 914 DOUT << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << *SI; 915 916 // There are two forms here: AI could be an array or struct. Both cases 917 // have different ways to compute the element offset. 918 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 919 const StructLayout *Layout = TD->getStructLayout(EltSTy); 920 921 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 922 // Get the number of bits to shift SrcVal to get the value. 923 const Type *FieldTy = EltSTy->getElementType(i); 924 uint64_t Shift = Layout->getElementOffsetInBits(i); 925 926 if (TD->isBigEndian()) 927 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 928 929 Value *EltVal = SrcVal; 930 if (Shift) { 931 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 932 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 933 "sroa.store.elt", SI); 934 } 935 936 // Truncate down to an integer of the right size. 937 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 938 939 // Ignore zero sized fields like {}, they obviously contain no data. 940 if (FieldSizeBits == 0) continue; 941 942 if (FieldSizeBits != AllocaSizeBits) 943 EltVal = new TruncInst(EltVal, IntegerType::get(FieldSizeBits), "", SI); 944 Value *DestField = NewElts[i]; 945 if (EltVal->getType() == FieldTy) { 946 // Storing to an integer field of this size, just do it. 947 } else if (FieldTy->isFloatingPoint() || isa<VectorType>(FieldTy)) { 948 // Bitcast to the right element type (for fp/vector values). 949 EltVal = new BitCastInst(EltVal, FieldTy, "", SI); 950 } else { 951 // Otherwise, bitcast the dest pointer (for aggregates). 952 DestField = new BitCastInst(DestField, 953 PointerType::getUnqual(EltVal->getType()), 954 "", SI); 955 } 956 new StoreInst(EltVal, DestField, SI); 957 } 958 959 } else { 960 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 961 const Type *ArrayEltTy = ATy->getElementType(); 962 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 963 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 964 965 uint64_t Shift; 966 967 if (TD->isBigEndian()) 968 Shift = AllocaSizeBits-ElementOffset; 969 else 970 Shift = 0; 971 972 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 973 // Ignore zero sized fields like {}, they obviously contain no data. 974 if (ElementSizeBits == 0) continue; 975 976 Value *EltVal = SrcVal; 977 if (Shift) { 978 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 979 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 980 "sroa.store.elt", SI); 981 } 982 983 // Truncate down to an integer of the right size. 984 if (ElementSizeBits != AllocaSizeBits) 985 EltVal = new TruncInst(EltVal, IntegerType::get(ElementSizeBits),"",SI); 986 Value *DestField = NewElts[i]; 987 if (EltVal->getType() == ArrayEltTy) { 988 // Storing to an integer field of this size, just do it. 989 } else if (ArrayEltTy->isFloatingPoint() || isa<VectorType>(ArrayEltTy)) { 990 // Bitcast to the right element type (for fp/vector values). 991 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI); 992 } else { 993 // Otherwise, bitcast the dest pointer (for aggregates). 994 DestField = new BitCastInst(DestField, 995 PointerType::getUnqual(EltVal->getType()), 996 "", SI); 997 } 998 new StoreInst(EltVal, DestField, SI); 999 1000 if (TD->isBigEndian()) 1001 Shift -= ElementOffset; 1002 else 1003 Shift += ElementOffset; 1004 } 1005 } 1006 1007 SI->eraseFromParent(); 1008} 1009 1010/// RewriteLoadUserOfWholeAlloca - We found an load of the entire allocation to 1011/// an integer. Load the individual pieces to form the aggregate value. 1012void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI, 1013 SmallVector<AllocaInst*, 32> &NewElts) { 1014 // Extract each element out of the NewElts according to its structure offset 1015 // and form the result value. 1016 const Type *AllocaEltTy = AI->getType()->getElementType(); 1017 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1018 1019 // If this isn't a load of the whole alloca to an integer, it may be a load 1020 // of the first element. Just ignore the load in this case and normal SROA 1021 // will handle it. 1022 if (!isa<IntegerType>(LI->getType()) || 1023 TD->getTypeAllocSizeInBits(LI->getType()) != AllocaSizeBits) 1024 return; 1025 1026 DOUT << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << *LI; 1027 1028 // There are two forms here: AI could be an array or struct. Both cases 1029 // have different ways to compute the element offset. 1030 const StructLayout *Layout = 0; 1031 uint64_t ArrayEltBitOffset = 0; 1032 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1033 Layout = TD->getStructLayout(EltSTy); 1034 } else { 1035 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 1036 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1037 } 1038 1039 Value *ResultVal = Constant::getNullValue(IntegerType::get(AllocaSizeBits)); 1040 1041 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1042 // Load the value from the alloca. If the NewElt is an aggregate, cast 1043 // the pointer to an integer of the same size before doing the load. 1044 Value *SrcField = NewElts[i]; 1045 const Type *FieldTy = 1046 cast<PointerType>(SrcField->getType())->getElementType(); 1047 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1048 1049 // Ignore zero sized fields like {}, they obviously contain no data. 1050 if (FieldSizeBits == 0) continue; 1051 1052 const IntegerType *FieldIntTy = IntegerType::get(FieldSizeBits); 1053 if (!isa<IntegerType>(FieldTy) && !FieldTy->isFloatingPoint() && 1054 !isa<VectorType>(FieldTy)) 1055 SrcField = new BitCastInst(SrcField, PointerType::getUnqual(FieldIntTy), 1056 "", LI); 1057 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 1058 1059 // If SrcField is a fp or vector of the right size but that isn't an 1060 // integer type, bitcast to an integer so we can shift it. 1061 if (SrcField->getType() != FieldIntTy) 1062 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 1063 1064 // Zero extend the field to be the same size as the final alloca so that 1065 // we can shift and insert it. 1066 if (SrcField->getType() != ResultVal->getType()) 1067 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 1068 1069 // Determine the number of bits to shift SrcField. 1070 uint64_t Shift; 1071 if (Layout) // Struct case. 1072 Shift = Layout->getElementOffsetInBits(i); 1073 else // Array case. 1074 Shift = i*ArrayEltBitOffset; 1075 1076 if (TD->isBigEndian()) 1077 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 1078 1079 if (Shift) { 1080 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 1081 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 1082 } 1083 1084 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 1085 } 1086 1087 // Handle tail padding by truncating the result 1088 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 1089 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 1090 1091 LI->replaceAllUsesWith(ResultVal); 1092 LI->eraseFromParent(); 1093} 1094 1095 1096/// HasPadding - Return true if the specified type has any structure or 1097/// alignment padding, false otherwise. 1098static bool HasPadding(const Type *Ty, const TargetData &TD) { 1099 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 1100 const StructLayout *SL = TD.getStructLayout(STy); 1101 unsigned PrevFieldBitOffset = 0; 1102 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1103 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 1104 1105 // Padding in sub-elements? 1106 if (HasPadding(STy->getElementType(i), TD)) 1107 return true; 1108 1109 // Check to see if there is any padding between this element and the 1110 // previous one. 1111 if (i) { 1112 unsigned PrevFieldEnd = 1113 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 1114 if (PrevFieldEnd < FieldBitOffset) 1115 return true; 1116 } 1117 1118 PrevFieldBitOffset = FieldBitOffset; 1119 } 1120 1121 // Check for tail padding. 1122 if (unsigned EltCount = STy->getNumElements()) { 1123 unsigned PrevFieldEnd = PrevFieldBitOffset + 1124 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 1125 if (PrevFieldEnd < SL->getSizeInBits()) 1126 return true; 1127 } 1128 1129 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1130 return HasPadding(ATy->getElementType(), TD); 1131 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) { 1132 return HasPadding(VTy->getElementType(), TD); 1133 } 1134 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 1135} 1136 1137/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 1138/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 1139/// or 1 if safe after canonicalization has been performed. 1140/// 1141int SROA::isSafeAllocaToScalarRepl(AllocationInst *AI) { 1142 // Loop over the use list of the alloca. We can only transform it if all of 1143 // the users are safe to transform. 1144 AllocaInfo Info; 1145 1146 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); 1147 I != E; ++I) { 1148 isSafeUseOfAllocation(cast<Instruction>(*I), AI, Info); 1149 if (Info.isUnsafe) { 1150 DOUT << "Cannot transform: " << *AI << " due to user: " << **I; 1151 return 0; 1152 } 1153 } 1154 1155 // Okay, we know all the users are promotable. If the aggregate is a memcpy 1156 // source and destination, we have to be careful. In particular, the memcpy 1157 // could be moving around elements that live in structure padding of the LLVM 1158 // types, but may actually be used. In these cases, we refuse to promote the 1159 // struct. 1160 if (Info.isMemCpySrc && Info.isMemCpyDst && 1161 HasPadding(AI->getType()->getElementType(), *TD)) 1162 return 0; 1163 1164 // If we require cleanup, return 1, otherwise return 3. 1165 return Info.needsCleanup ? 1 : 3; 1166} 1167 1168/// CleanupGEP - GEP is used by an Alloca, which can be prompted after the GEP 1169/// is canonicalized here. 1170void SROA::CleanupGEP(GetElementPtrInst *GEPI) { 1171 gep_type_iterator I = gep_type_begin(GEPI); 1172 ++I; 1173 1174 const ArrayType *AT = dyn_cast<ArrayType>(*I); 1175 if (!AT) 1176 return; 1177 1178 uint64_t NumElements = AT->getNumElements(); 1179 1180 if (isa<ConstantInt>(I.getOperand())) 1181 return; 1182 1183 if (NumElements == 1) { 1184 GEPI->setOperand(2, Constant::getNullValue(Type::Int32Ty)); 1185 return; 1186 } 1187 1188 assert(NumElements == 2 && "Unhandled case!"); 1189 // All users of the GEP must be loads. At each use of the GEP, insert 1190 // two loads of the appropriate indexed GEP and select between them. 1191 Value *IsOne = new ICmpInst(ICmpInst::ICMP_NE, I.getOperand(), 1192 Constant::getNullValue(I.getOperand()->getType()), 1193 "isone", GEPI); 1194 // Insert the new GEP instructions, which are properly indexed. 1195 SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end()); 1196 Indices[1] = Constant::getNullValue(Type::Int32Ty); 1197 Value *ZeroIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 1198 Indices.begin(), 1199 Indices.end(), 1200 GEPI->getName()+".0", GEPI); 1201 Indices[1] = ConstantInt::get(Type::Int32Ty, 1); 1202 Value *OneIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 1203 Indices.begin(), 1204 Indices.end(), 1205 GEPI->getName()+".1", GEPI); 1206 // Replace all loads of the variable index GEP with loads from both 1207 // indexes and a select. 1208 while (!GEPI->use_empty()) { 1209 LoadInst *LI = cast<LoadInst>(GEPI->use_back()); 1210 Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI); 1211 Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI); 1212 Value *R = SelectInst::Create(IsOne, One, Zero, LI->getName(), LI); 1213 LI->replaceAllUsesWith(R); 1214 LI->eraseFromParent(); 1215 } 1216 GEPI->eraseFromParent(); 1217} 1218 1219 1220/// CleanupAllocaUsers - If SROA reported that it can promote the specified 1221/// allocation, but only if cleaned up, perform the cleanups required. 1222void SROA::CleanupAllocaUsers(AllocationInst *AI) { 1223 // At this point, we know that the end result will be SROA'd and promoted, so 1224 // we can insert ugly code if required so long as sroa+mem2reg will clean it 1225 // up. 1226 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 1227 UI != E; ) { 1228 User *U = *UI++; 1229 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) 1230 CleanupGEP(GEPI);
|
1232 SmallVector<DbgInfoIntrinsic *, 2> DbgInUses; 1233 if (!isa<StoreInst>(I) && OnlyUsedByDbgInfoIntrinsics(I, &DbgInUses)) { 1234 // Safe to remove debug info uses. 1235 while (!DbgInUses.empty()) { 1236 DbgInfoIntrinsic *DI = DbgInUses.back(); DbgInUses.pop_back(); 1237 DI->eraseFromParent(); 1238 } 1239 I->eraseFromParent(); 1240 } 1241 } 1242 } 1243} 1244 1245/// MergeInType - Add the 'In' type to the accumulated type (Accum) so far at 1246/// the offset specified by Offset (which is specified in bytes). 1247/// 1248/// There are two cases we handle here: 1249/// 1) A union of vector types of the same size and potentially its elements. 1250/// Here we turn element accesses into insert/extract element operations. 1251/// This promotes a <4 x float> with a store of float to the third element 1252/// into a <4 x float> that uses insert element. 1253/// 2) A fully general blob of memory, which we turn into some (potentially 1254/// large) integer type with extract and insert operations where the loads 1255/// and stores would mutate the memory. 1256static void MergeInType(const Type *In, uint64_t Offset, const Type *&VecTy, 1257 unsigned AllocaSize, const TargetData &TD) { 1258 // If this could be contributing to a vector, analyze it. 1259 if (VecTy != Type::VoidTy) { // either null or a vector type. 1260 1261 // If the In type is a vector that is the same size as the alloca, see if it 1262 // matches the existing VecTy. 1263 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 1264 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 1265 // If we're storing/loading a vector of the right size, allow it as a 1266 // vector. If this the first vector we see, remember the type so that 1267 // we know the element size. 1268 if (VecTy == 0) 1269 VecTy = VInTy; 1270 return; 1271 } 1272 } else if (In == Type::FloatTy || In == Type::DoubleTy || 1273 (isa<IntegerType>(In) && In->getPrimitiveSizeInBits() >= 8 && 1274 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 1275 // If we're accessing something that could be an element of a vector, see 1276 // if the implied vector agrees with what we already have and if Offset is 1277 // compatible with it. 1278 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 1279 if (Offset % EltSize == 0 && 1280 AllocaSize % EltSize == 0 && 1281 (VecTy == 0 || 1282 cast<VectorType>(VecTy)->getElementType() 1283 ->getPrimitiveSizeInBits()/8 == EltSize)) { 1284 if (VecTy == 0) 1285 VecTy = VectorType::get(In, AllocaSize/EltSize); 1286 return; 1287 } 1288 } 1289 } 1290 1291 // Otherwise, we have a case that we can't handle with an optimized vector 1292 // form. We can still turn this into a large integer. 1293 VecTy = Type::VoidTy; 1294} 1295 1296/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 1297/// its accesses to use a to single vector type, return true, and set VecTy to 1298/// the new type. If we could convert the alloca into a single promotable 1299/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 1300/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 1301/// is the current offset from the base of the alloca being analyzed. 1302/// 1303/// If we see at least one access to the value that is as a vector type, set the 1304/// SawVec flag. 1305/// 1306bool SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy, 1307 bool &SawVec, uint64_t Offset, 1308 unsigned AllocaSize) { 1309 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1310 Instruction *User = cast<Instruction>(*UI); 1311 1312 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1313 // Don't break volatile loads. 1314 if (LI->isVolatile()) 1315 return false; 1316 MergeInType(LI->getType(), Offset, VecTy, AllocaSize, *TD); 1317 SawVec |= isa<VectorType>(LI->getType()); 1318 continue; 1319 } 1320 1321 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1322 // Storing the pointer, not into the value? 1323 if (SI->getOperand(0) == V || SI->isVolatile()) return 0; 1324 MergeInType(SI->getOperand(0)->getType(), Offset, VecTy, AllocaSize, *TD); 1325 SawVec |= isa<VectorType>(SI->getOperand(0)->getType()); 1326 continue; 1327 } 1328 1329 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 1330 if (!CanConvertToScalar(BCI, IsNotTrivial, VecTy, SawVec, Offset, 1331 AllocaSize)) 1332 return false; 1333 IsNotTrivial = true; 1334 continue; 1335 } 1336 1337 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1338 // If this is a GEP with a variable indices, we can't handle it. 1339 if (!GEP->hasAllConstantIndices()) 1340 return false; 1341 1342 // Compute the offset that this GEP adds to the pointer. 1343 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1344 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(), 1345 &Indices[0], Indices.size()); 1346 // See if all uses can be converted. 1347 if (!CanConvertToScalar(GEP, IsNotTrivial, VecTy, SawVec,Offset+GEPOffset, 1348 AllocaSize)) 1349 return false; 1350 IsNotTrivial = true; 1351 continue; 1352 } 1353 1354 // If this is a constant sized memset of a constant value (e.g. 0) we can 1355 // handle it. 1356 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 1357 // Store of constant value and constant size. 1358 if (isa<ConstantInt>(MSI->getValue()) && 1359 isa<ConstantInt>(MSI->getLength())) { 1360 IsNotTrivial = true; 1361 continue; 1362 } 1363 } 1364 1365 // If this is a memcpy or memmove into or out of the whole allocation, we 1366 // can handle it like a load or store of the scalar type. 1367 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 1368 if (ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength())) 1369 if (Len->getZExtValue() == AllocaSize && Offset == 0) { 1370 IsNotTrivial = true; 1371 continue; 1372 } 1373 } 1374 1375 // Ignore dbg intrinsic. 1376 if (isa<DbgInfoIntrinsic>(User)) 1377 continue; 1378 1379 // Otherwise, we cannot handle this! 1380 return false; 1381 } 1382 1383 return true; 1384} 1385 1386 1387/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 1388/// directly. This happens when we are converting an "integer union" to a 1389/// single integer scalar, or when we are converting a "vector union" to a 1390/// vector with insert/extractelement instructions. 1391/// 1392/// Offset is an offset from the original alloca, in bits that need to be 1393/// shifted to the right. By the end of this, there should be no uses of Ptr. 1394void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset) { 1395 while (!Ptr->use_empty()) { 1396 Instruction *User = cast<Instruction>(Ptr->use_back()); 1397 1398 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 1399 ConvertUsesToScalar(CI, NewAI, Offset); 1400 CI->eraseFromParent(); 1401 continue; 1402 } 1403 1404 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1405 // Compute the offset that this GEP adds to the pointer. 1406 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1407 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(), 1408 &Indices[0], Indices.size()); 1409 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 1410 GEP->eraseFromParent(); 1411 continue; 1412 } 1413 1414 IRBuilder<> Builder(User->getParent(), User); 1415 1416 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1417 // The load is a bit extract from NewAI shifted right by Offset bits. 1418 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 1419 Value *NewLoadVal 1420 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 1421 LI->replaceAllUsesWith(NewLoadVal); 1422 LI->eraseFromParent(); 1423 continue; 1424 } 1425 1426 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1427 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 1428 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").c_str()); 1429 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 1430 Builder); 1431 Builder.CreateStore(New, NewAI); 1432 SI->eraseFromParent(); 1433 continue; 1434 } 1435 1436 // If this is a constant sized memset of a constant value (e.g. 0) we can 1437 // transform it into a store of the expanded constant value. 1438 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 1439 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 1440 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 1441 if (NumBytes != 0) { 1442 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 1443 1444 // Compute the value replicated the right number of times. 1445 APInt APVal(NumBytes*8, Val); 1446 1447 // Splat the value if non-zero. 1448 if (Val) 1449 for (unsigned i = 1; i != NumBytes; ++i) 1450 APVal |= APVal << 8; 1451 1452 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").c_str()); 1453 Value *New = ConvertScalar_InsertValue(ConstantInt::get(APVal), Old, 1454 Offset, Builder); 1455 Builder.CreateStore(New, NewAI); 1456 } 1457 MSI->eraseFromParent(); 1458 continue; 1459 } 1460 1461 // If this is a memcpy or memmove into or out of the whole allocation, we 1462 // can handle it like a load or store of the scalar type. 1463 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 1464 assert(Offset == 0 && "must be store to start of alloca"); 1465 1466 // If the source and destination are both to the same alloca, then this is 1467 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 1468 // as appropriate. 1469 AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject()); 1470 1471 if (MTI->getSource()->getUnderlyingObject() != OrigAI) { 1472 // Dest must be OrigAI, change this to be a load from the original 1473 // pointer (bitcasted), then a store to our new alloca. 1474 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 1475 Value *SrcPtr = MTI->getSource(); 1476 SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType()); 1477 1478 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 1479 SrcVal->setAlignment(MTI->getAlignment()); 1480 Builder.CreateStore(SrcVal, NewAI); 1481 } else if (MTI->getDest()->getUnderlyingObject() != OrigAI) { 1482 // Src must be OrigAI, change this to be a load from NewAI then a store 1483 // through the original dest pointer (bitcasted). 1484 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 1485 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 1486 1487 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType()); 1488 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 1489 NewStore->setAlignment(MTI->getAlignment()); 1490 } else { 1491 // Noop transfer. Src == Dst 1492 } 1493 1494 1495 MTI->eraseFromParent(); 1496 continue; 1497 } 1498 1499 // If user is a dbg info intrinsic then it is safe to remove it. 1500 if (isa<DbgInfoIntrinsic>(User)) { 1501 User->eraseFromParent(); 1502 continue; 1503 } 1504 1505 assert(0 && "Unsupported operation!"); 1506 abort(); 1507 } 1508} 1509 1510/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 1511/// or vector value FromVal, extracting the bits from the offset specified by 1512/// Offset. This returns the value, which is of type ToType. 1513/// 1514/// This happens when we are converting an "integer union" to a single 1515/// integer scalar, or when we are converting a "vector union" to a vector with 1516/// insert/extractelement instructions. 1517/// 1518/// Offset is an offset from the original alloca, in bits that need to be 1519/// shifted to the right. 1520Value *SROA::ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 1521 uint64_t Offset, IRBuilder<> &Builder) { 1522 // If the load is of the whole new alloca, no conversion is needed. 1523 if (FromVal->getType() == ToType && Offset == 0) 1524 return FromVal; 1525 1526 // If the result alloca is a vector type, this is either an element 1527 // access or a bitcast to another vector type of the same size. 1528 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) { 1529 if (isa<VectorType>(ToType)) 1530 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 1531 1532 // Otherwise it must be an element access. 1533 unsigned Elt = 0; 1534 if (Offset) { 1535 unsigned EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 1536 Elt = Offset/EltSize; 1537 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 1538 } 1539 // Return the element extracted out of it. 1540 Value *V = Builder.CreateExtractElement(FromVal, 1541 ConstantInt::get(Type::Int32Ty,Elt), 1542 "tmp"); 1543 if (V->getType() != ToType) 1544 V = Builder.CreateBitCast(V, ToType, "tmp"); 1545 return V; 1546 } 1547 1548 // If ToType is a first class aggregate, extract out each of the pieces and 1549 // use insertvalue's to form the FCA. 1550 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 1551 const StructLayout &Layout = *TD->getStructLayout(ST); 1552 Value *Res = UndefValue::get(ST); 1553 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1554 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 1555 Offset+Layout.getElementOffsetInBits(i), 1556 Builder); 1557 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1558 } 1559 return Res; 1560 } 1561 1562 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 1563 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 1564 Value *Res = UndefValue::get(AT); 1565 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1566 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 1567 Offset+i*EltSize, Builder); 1568 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1569 } 1570 return Res; 1571 } 1572 1573 // Otherwise, this must be a union that was converted to an integer value. 1574 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 1575 1576 // If this is a big-endian system and the load is narrower than the 1577 // full alloca type, we need to do a shift to get the right bits. 1578 int ShAmt = 0; 1579 if (TD->isBigEndian()) { 1580 // On big-endian machines, the lowest bit is stored at the bit offset 1581 // from the pointer given by getTypeStoreSizeInBits. This matters for 1582 // integers with a bitwidth that is not a multiple of 8. 1583 ShAmt = TD->getTypeStoreSizeInBits(NTy) - 1584 TD->getTypeStoreSizeInBits(ToType) - Offset; 1585 } else { 1586 ShAmt = Offset; 1587 } 1588 1589 // Note: we support negative bitwidths (with shl) which are not defined. 1590 // We do this to support (f.e.) loads off the end of a structure where 1591 // only some bits are used. 1592 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 1593 FromVal = Builder.CreateLShr(FromVal, ConstantInt::get(FromVal->getType(), 1594 ShAmt), "tmp"); 1595 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 1596 FromVal = Builder.CreateShl(FromVal, ConstantInt::get(FromVal->getType(), 1597 -ShAmt), "tmp"); 1598 1599 // Finally, unconditionally truncate the integer to the right width. 1600 unsigned LIBitWidth = TD->getTypeSizeInBits(ToType); 1601 if (LIBitWidth < NTy->getBitWidth()) 1602 FromVal = Builder.CreateTrunc(FromVal, IntegerType::get(LIBitWidth), "tmp"); 1603 else if (LIBitWidth > NTy->getBitWidth()) 1604 FromVal = Builder.CreateZExt(FromVal, IntegerType::get(LIBitWidth), "tmp"); 1605 1606 // If the result is an integer, this is a trunc or bitcast. 1607 if (isa<IntegerType>(ToType)) { 1608 // Should be done. 1609 } else if (ToType->isFloatingPoint() || isa<VectorType>(ToType)) { 1610 // Just do a bitcast, we know the sizes match up. 1611 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 1612 } else { 1613 // Otherwise must be a pointer. 1614 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 1615 } 1616 assert(FromVal->getType() == ToType && "Didn't convert right?"); 1617 return FromVal; 1618} 1619 1620 1621/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 1622/// or vector value "Old" at the offset specified by Offset. 1623/// 1624/// This happens when we are converting an "integer union" to a 1625/// single integer scalar, or when we are converting a "vector union" to a 1626/// vector with insert/extractelement instructions. 1627/// 1628/// Offset is an offset from the original alloca, in bits that need to be 1629/// shifted to the right. 1630Value *SROA::ConvertScalar_InsertValue(Value *SV, Value *Old, 1631 uint64_t Offset, IRBuilder<> &Builder) { 1632 1633 // Convert the stored type to the actual type, shift it left to insert 1634 // then 'or' into place. 1635 const Type *AllocaType = Old->getType(); 1636 1637 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 1638 uint64_t VecSize = TD->getTypeAllocSizeInBits(VTy); 1639 uint64_t ValSize = TD->getTypeAllocSizeInBits(SV->getType()); 1640 1641 // Changing the whole vector with memset or with an access of a different 1642 // vector type? 1643 if (ValSize == VecSize) 1644 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 1645 1646 uint64_t EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 1647 1648 // Must be an element insertion. 1649 unsigned Elt = Offset/EltSize; 1650 1651 if (SV->getType() != VTy->getElementType()) 1652 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 1653 1654 SV = Builder.CreateInsertElement(Old, SV, 1655 ConstantInt::get(Type::Int32Ty, Elt), 1656 "tmp"); 1657 return SV; 1658 } 1659 1660 // If SV is a first-class aggregate value, insert each value recursively. 1661 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 1662 const StructLayout &Layout = *TD->getStructLayout(ST); 1663 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1664 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1665 Old = ConvertScalar_InsertValue(Elt, Old, 1666 Offset+Layout.getElementOffsetInBits(i), 1667 Builder); 1668 } 1669 return Old; 1670 } 1671 1672 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 1673 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 1674 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1675 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1676 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 1677 } 1678 return Old; 1679 } 1680 1681 // If SV is a float, convert it to the appropriate integer type. 1682 // If it is a pointer, do the same. 1683 unsigned SrcWidth = TD->getTypeSizeInBits(SV->getType()); 1684 unsigned DestWidth = TD->getTypeSizeInBits(AllocaType); 1685 unsigned SrcStoreWidth = TD->getTypeStoreSizeInBits(SV->getType()); 1686 unsigned DestStoreWidth = TD->getTypeStoreSizeInBits(AllocaType); 1687 if (SV->getType()->isFloatingPoint() || isa<VectorType>(SV->getType())) 1688 SV = Builder.CreateBitCast(SV, IntegerType::get(SrcWidth), "tmp"); 1689 else if (isa<PointerType>(SV->getType())) 1690 SV = Builder.CreatePtrToInt(SV, TD->getIntPtrType(), "tmp"); 1691 1692 // Zero extend or truncate the value if needed. 1693 if (SV->getType() != AllocaType) { 1694 if (SV->getType()->getPrimitiveSizeInBits() < 1695 AllocaType->getPrimitiveSizeInBits()) 1696 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 1697 else { 1698 // Truncation may be needed if storing more than the alloca can hold 1699 // (undefined behavior). 1700 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 1701 SrcWidth = DestWidth; 1702 SrcStoreWidth = DestStoreWidth; 1703 } 1704 } 1705 1706 // If this is a big-endian system and the store is narrower than the 1707 // full alloca type, we need to do a shift to get the right bits. 1708 int ShAmt = 0; 1709 if (TD->isBigEndian()) { 1710 // On big-endian machines, the lowest bit is stored at the bit offset 1711 // from the pointer given by getTypeStoreSizeInBits. This matters for 1712 // integers with a bitwidth that is not a multiple of 8. 1713 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 1714 } else { 1715 ShAmt = Offset; 1716 } 1717 1718 // Note: we support negative bitwidths (with shr) which are not defined. 1719 // We do this to support (f.e.) stores off the end of a structure where 1720 // only some bits in the structure are set. 1721 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 1722 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 1723 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), ShAmt), "tmp"); 1724 Mask <<= ShAmt; 1725 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1726 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), -ShAmt), "tmp"); 1727 Mask = Mask.lshr(-ShAmt); 1728 } 1729 1730 // Mask out the bits we are about to insert from the old value, and or 1731 // in the new bits. 1732 if (SrcWidth != DestWidth) { 1733 assert(DestWidth > SrcWidth); 1734 Old = Builder.CreateAnd(Old, ConstantInt::get(~Mask), "mask"); 1735 SV = Builder.CreateOr(Old, SV, "ins"); 1736 } 1737 return SV; 1738} 1739 1740 1741 1742/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1743/// some part of a constant global variable. This intentionally only accepts 1744/// constant expressions because we don't can't rewrite arbitrary instructions. 1745static bool PointsToConstantGlobal(Value *V) { 1746 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1747 return GV->isConstant(); 1748 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1749 if (CE->getOpcode() == Instruction::BitCast || 1750 CE->getOpcode() == Instruction::GetElementPtr) 1751 return PointsToConstantGlobal(CE->getOperand(0)); 1752 return false; 1753} 1754 1755/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1756/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1757/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1758/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1759/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1760/// the alloca, and if the source pointer is a pointer to a constant global, we 1761/// can optimize this. 1762static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy, 1763 bool isOffset) { 1764 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1765 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) 1766 // Ignore non-volatile loads, they are always ok. 1767 if (!LI->isVolatile()) 1768 continue; 1769 1770 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) { 1771 // If uses of the bitcast are ok, we are ok. 1772 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 1773 return false; 1774 continue; 1775 } 1776 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) { 1777 // If the GEP has all zero indices, it doesn't offset the pointer. If it 1778 // doesn't, it does. 1779 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 1780 isOffset || !GEP->hasAllZeroIndices())) 1781 return false; 1782 continue; 1783 } 1784 1785 // If this is isn't our memcpy/memmove, reject it as something we can't 1786 // handle. 1787 if (!isa<MemTransferInst>(*UI)) 1788 return false; 1789 1790 // If we already have seen a copy, reject the second one. 1791 if (TheCopy) return false; 1792 1793 // If the pointer has been offset from the start of the alloca, we can't 1794 // safely handle this. 1795 if (isOffset) return false; 1796 1797 // If the memintrinsic isn't using the alloca as the dest, reject it. 1798 if (UI.getOperandNo() != 1) return false; 1799 1800 MemIntrinsic *MI = cast<MemIntrinsic>(*UI); 1801 1802 // If the source of the memcpy/move is not a constant global, reject it. 1803 if (!PointsToConstantGlobal(MI->getOperand(2))) 1804 return false; 1805 1806 // Otherwise, the transform is safe. Remember the copy instruction. 1807 TheCopy = MI; 1808 } 1809 return true; 1810} 1811 1812/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 1813/// modified by a copy from a constant global. If we can prove this, we can 1814/// replace any uses of the alloca with uses of the global directly. 1815Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocationInst *AI) { 1816 Instruction *TheCopy = 0; 1817 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 1818 return TheCopy; 1819 return 0; 1820}
| 1233 SmallVector<DbgInfoIntrinsic *, 2> DbgInUses; 1234 if (!isa<StoreInst>(I) && OnlyUsedByDbgInfoIntrinsics(I, &DbgInUses)) { 1235 // Safe to remove debug info uses. 1236 while (!DbgInUses.empty()) { 1237 DbgInfoIntrinsic *DI = DbgInUses.back(); DbgInUses.pop_back(); 1238 DI->eraseFromParent(); 1239 } 1240 I->eraseFromParent(); 1241 } 1242 } 1243 } 1244} 1245 1246/// MergeInType - Add the 'In' type to the accumulated type (Accum) so far at 1247/// the offset specified by Offset (which is specified in bytes). 1248/// 1249/// There are two cases we handle here: 1250/// 1) A union of vector types of the same size and potentially its elements. 1251/// Here we turn element accesses into insert/extract element operations. 1252/// This promotes a <4 x float> with a store of float to the third element 1253/// into a <4 x float> that uses insert element. 1254/// 2) A fully general blob of memory, which we turn into some (potentially 1255/// large) integer type with extract and insert operations where the loads 1256/// and stores would mutate the memory. 1257static void MergeInType(const Type *In, uint64_t Offset, const Type *&VecTy, 1258 unsigned AllocaSize, const TargetData &TD) { 1259 // If this could be contributing to a vector, analyze it. 1260 if (VecTy != Type::VoidTy) { // either null or a vector type. 1261 1262 // If the In type is a vector that is the same size as the alloca, see if it 1263 // matches the existing VecTy. 1264 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 1265 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 1266 // If we're storing/loading a vector of the right size, allow it as a 1267 // vector. If this the first vector we see, remember the type so that 1268 // we know the element size. 1269 if (VecTy == 0) 1270 VecTy = VInTy; 1271 return; 1272 } 1273 } else if (In == Type::FloatTy || In == Type::DoubleTy || 1274 (isa<IntegerType>(In) && In->getPrimitiveSizeInBits() >= 8 && 1275 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 1276 // If we're accessing something that could be an element of a vector, see 1277 // if the implied vector agrees with what we already have and if Offset is 1278 // compatible with it. 1279 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 1280 if (Offset % EltSize == 0 && 1281 AllocaSize % EltSize == 0 && 1282 (VecTy == 0 || 1283 cast<VectorType>(VecTy)->getElementType() 1284 ->getPrimitiveSizeInBits()/8 == EltSize)) { 1285 if (VecTy == 0) 1286 VecTy = VectorType::get(In, AllocaSize/EltSize); 1287 return; 1288 } 1289 } 1290 } 1291 1292 // Otherwise, we have a case that we can't handle with an optimized vector 1293 // form. We can still turn this into a large integer. 1294 VecTy = Type::VoidTy; 1295} 1296 1297/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 1298/// its accesses to use a to single vector type, return true, and set VecTy to 1299/// the new type. If we could convert the alloca into a single promotable 1300/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 1301/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 1302/// is the current offset from the base of the alloca being analyzed. 1303/// 1304/// If we see at least one access to the value that is as a vector type, set the 1305/// SawVec flag. 1306/// 1307bool SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy, 1308 bool &SawVec, uint64_t Offset, 1309 unsigned AllocaSize) { 1310 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1311 Instruction *User = cast<Instruction>(*UI); 1312 1313 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1314 // Don't break volatile loads. 1315 if (LI->isVolatile()) 1316 return false; 1317 MergeInType(LI->getType(), Offset, VecTy, AllocaSize, *TD); 1318 SawVec |= isa<VectorType>(LI->getType()); 1319 continue; 1320 } 1321 1322 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1323 // Storing the pointer, not into the value? 1324 if (SI->getOperand(0) == V || SI->isVolatile()) return 0; 1325 MergeInType(SI->getOperand(0)->getType(), Offset, VecTy, AllocaSize, *TD); 1326 SawVec |= isa<VectorType>(SI->getOperand(0)->getType()); 1327 continue; 1328 } 1329 1330 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 1331 if (!CanConvertToScalar(BCI, IsNotTrivial, VecTy, SawVec, Offset, 1332 AllocaSize)) 1333 return false; 1334 IsNotTrivial = true; 1335 continue; 1336 } 1337 1338 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1339 // If this is a GEP with a variable indices, we can't handle it. 1340 if (!GEP->hasAllConstantIndices()) 1341 return false; 1342 1343 // Compute the offset that this GEP adds to the pointer. 1344 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1345 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(), 1346 &Indices[0], Indices.size()); 1347 // See if all uses can be converted. 1348 if (!CanConvertToScalar(GEP, IsNotTrivial, VecTy, SawVec,Offset+GEPOffset, 1349 AllocaSize)) 1350 return false; 1351 IsNotTrivial = true; 1352 continue; 1353 } 1354 1355 // If this is a constant sized memset of a constant value (e.g. 0) we can 1356 // handle it. 1357 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 1358 // Store of constant value and constant size. 1359 if (isa<ConstantInt>(MSI->getValue()) && 1360 isa<ConstantInt>(MSI->getLength())) { 1361 IsNotTrivial = true; 1362 continue; 1363 } 1364 } 1365 1366 // If this is a memcpy or memmove into or out of the whole allocation, we 1367 // can handle it like a load or store of the scalar type. 1368 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 1369 if (ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength())) 1370 if (Len->getZExtValue() == AllocaSize && Offset == 0) { 1371 IsNotTrivial = true; 1372 continue; 1373 } 1374 } 1375 1376 // Ignore dbg intrinsic. 1377 if (isa<DbgInfoIntrinsic>(User)) 1378 continue; 1379 1380 // Otherwise, we cannot handle this! 1381 return false; 1382 } 1383 1384 return true; 1385} 1386 1387 1388/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 1389/// directly. This happens when we are converting an "integer union" to a 1390/// single integer scalar, or when we are converting a "vector union" to a 1391/// vector with insert/extractelement instructions. 1392/// 1393/// Offset is an offset from the original alloca, in bits that need to be 1394/// shifted to the right. By the end of this, there should be no uses of Ptr. 1395void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset) { 1396 while (!Ptr->use_empty()) { 1397 Instruction *User = cast<Instruction>(Ptr->use_back()); 1398 1399 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 1400 ConvertUsesToScalar(CI, NewAI, Offset); 1401 CI->eraseFromParent(); 1402 continue; 1403 } 1404 1405 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1406 // Compute the offset that this GEP adds to the pointer. 1407 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1408 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(), 1409 &Indices[0], Indices.size()); 1410 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 1411 GEP->eraseFromParent(); 1412 continue; 1413 } 1414 1415 IRBuilder<> Builder(User->getParent(), User); 1416 1417 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1418 // The load is a bit extract from NewAI shifted right by Offset bits. 1419 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 1420 Value *NewLoadVal 1421 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 1422 LI->replaceAllUsesWith(NewLoadVal); 1423 LI->eraseFromParent(); 1424 continue; 1425 } 1426 1427 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1428 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 1429 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").c_str()); 1430 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 1431 Builder); 1432 Builder.CreateStore(New, NewAI); 1433 SI->eraseFromParent(); 1434 continue; 1435 } 1436 1437 // If this is a constant sized memset of a constant value (e.g. 0) we can 1438 // transform it into a store of the expanded constant value. 1439 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 1440 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 1441 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 1442 if (NumBytes != 0) { 1443 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 1444 1445 // Compute the value replicated the right number of times. 1446 APInt APVal(NumBytes*8, Val); 1447 1448 // Splat the value if non-zero. 1449 if (Val) 1450 for (unsigned i = 1; i != NumBytes; ++i) 1451 APVal |= APVal << 8; 1452 1453 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").c_str()); 1454 Value *New = ConvertScalar_InsertValue(ConstantInt::get(APVal), Old, 1455 Offset, Builder); 1456 Builder.CreateStore(New, NewAI); 1457 } 1458 MSI->eraseFromParent(); 1459 continue; 1460 } 1461 1462 // If this is a memcpy or memmove into or out of the whole allocation, we 1463 // can handle it like a load or store of the scalar type. 1464 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 1465 assert(Offset == 0 && "must be store to start of alloca"); 1466 1467 // If the source and destination are both to the same alloca, then this is 1468 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 1469 // as appropriate. 1470 AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject()); 1471 1472 if (MTI->getSource()->getUnderlyingObject() != OrigAI) { 1473 // Dest must be OrigAI, change this to be a load from the original 1474 // pointer (bitcasted), then a store to our new alloca. 1475 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 1476 Value *SrcPtr = MTI->getSource(); 1477 SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType()); 1478 1479 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 1480 SrcVal->setAlignment(MTI->getAlignment()); 1481 Builder.CreateStore(SrcVal, NewAI); 1482 } else if (MTI->getDest()->getUnderlyingObject() != OrigAI) { 1483 // Src must be OrigAI, change this to be a load from NewAI then a store 1484 // through the original dest pointer (bitcasted). 1485 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 1486 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 1487 1488 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType()); 1489 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 1490 NewStore->setAlignment(MTI->getAlignment()); 1491 } else { 1492 // Noop transfer. Src == Dst 1493 } 1494 1495 1496 MTI->eraseFromParent(); 1497 continue; 1498 } 1499 1500 // If user is a dbg info intrinsic then it is safe to remove it. 1501 if (isa<DbgInfoIntrinsic>(User)) { 1502 User->eraseFromParent(); 1503 continue; 1504 } 1505 1506 assert(0 && "Unsupported operation!"); 1507 abort(); 1508 } 1509} 1510 1511/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 1512/// or vector value FromVal, extracting the bits from the offset specified by 1513/// Offset. This returns the value, which is of type ToType. 1514/// 1515/// This happens when we are converting an "integer union" to a single 1516/// integer scalar, or when we are converting a "vector union" to a vector with 1517/// insert/extractelement instructions. 1518/// 1519/// Offset is an offset from the original alloca, in bits that need to be 1520/// shifted to the right. 1521Value *SROA::ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 1522 uint64_t Offset, IRBuilder<> &Builder) { 1523 // If the load is of the whole new alloca, no conversion is needed. 1524 if (FromVal->getType() == ToType && Offset == 0) 1525 return FromVal; 1526 1527 // If the result alloca is a vector type, this is either an element 1528 // access or a bitcast to another vector type of the same size. 1529 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) { 1530 if (isa<VectorType>(ToType)) 1531 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 1532 1533 // Otherwise it must be an element access. 1534 unsigned Elt = 0; 1535 if (Offset) { 1536 unsigned EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 1537 Elt = Offset/EltSize; 1538 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 1539 } 1540 // Return the element extracted out of it. 1541 Value *V = Builder.CreateExtractElement(FromVal, 1542 ConstantInt::get(Type::Int32Ty,Elt), 1543 "tmp"); 1544 if (V->getType() != ToType) 1545 V = Builder.CreateBitCast(V, ToType, "tmp"); 1546 return V; 1547 } 1548 1549 // If ToType is a first class aggregate, extract out each of the pieces and 1550 // use insertvalue's to form the FCA. 1551 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 1552 const StructLayout &Layout = *TD->getStructLayout(ST); 1553 Value *Res = UndefValue::get(ST); 1554 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1555 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 1556 Offset+Layout.getElementOffsetInBits(i), 1557 Builder); 1558 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1559 } 1560 return Res; 1561 } 1562 1563 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 1564 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 1565 Value *Res = UndefValue::get(AT); 1566 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1567 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 1568 Offset+i*EltSize, Builder); 1569 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1570 } 1571 return Res; 1572 } 1573 1574 // Otherwise, this must be a union that was converted to an integer value. 1575 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 1576 1577 // If this is a big-endian system and the load is narrower than the 1578 // full alloca type, we need to do a shift to get the right bits. 1579 int ShAmt = 0; 1580 if (TD->isBigEndian()) { 1581 // On big-endian machines, the lowest bit is stored at the bit offset 1582 // from the pointer given by getTypeStoreSizeInBits. This matters for 1583 // integers with a bitwidth that is not a multiple of 8. 1584 ShAmt = TD->getTypeStoreSizeInBits(NTy) - 1585 TD->getTypeStoreSizeInBits(ToType) - Offset; 1586 } else { 1587 ShAmt = Offset; 1588 } 1589 1590 // Note: we support negative bitwidths (with shl) which are not defined. 1591 // We do this to support (f.e.) loads off the end of a structure where 1592 // only some bits are used. 1593 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 1594 FromVal = Builder.CreateLShr(FromVal, ConstantInt::get(FromVal->getType(), 1595 ShAmt), "tmp"); 1596 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 1597 FromVal = Builder.CreateShl(FromVal, ConstantInt::get(FromVal->getType(), 1598 -ShAmt), "tmp"); 1599 1600 // Finally, unconditionally truncate the integer to the right width. 1601 unsigned LIBitWidth = TD->getTypeSizeInBits(ToType); 1602 if (LIBitWidth < NTy->getBitWidth()) 1603 FromVal = Builder.CreateTrunc(FromVal, IntegerType::get(LIBitWidth), "tmp"); 1604 else if (LIBitWidth > NTy->getBitWidth()) 1605 FromVal = Builder.CreateZExt(FromVal, IntegerType::get(LIBitWidth), "tmp"); 1606 1607 // If the result is an integer, this is a trunc or bitcast. 1608 if (isa<IntegerType>(ToType)) { 1609 // Should be done. 1610 } else if (ToType->isFloatingPoint() || isa<VectorType>(ToType)) { 1611 // Just do a bitcast, we know the sizes match up. 1612 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 1613 } else { 1614 // Otherwise must be a pointer. 1615 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 1616 } 1617 assert(FromVal->getType() == ToType && "Didn't convert right?"); 1618 return FromVal; 1619} 1620 1621 1622/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 1623/// or vector value "Old" at the offset specified by Offset. 1624/// 1625/// This happens when we are converting an "integer union" to a 1626/// single integer scalar, or when we are converting a "vector union" to a 1627/// vector with insert/extractelement instructions. 1628/// 1629/// Offset is an offset from the original alloca, in bits that need to be 1630/// shifted to the right. 1631Value *SROA::ConvertScalar_InsertValue(Value *SV, Value *Old, 1632 uint64_t Offset, IRBuilder<> &Builder) { 1633 1634 // Convert the stored type to the actual type, shift it left to insert 1635 // then 'or' into place. 1636 const Type *AllocaType = Old->getType(); 1637 1638 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 1639 uint64_t VecSize = TD->getTypeAllocSizeInBits(VTy); 1640 uint64_t ValSize = TD->getTypeAllocSizeInBits(SV->getType()); 1641 1642 // Changing the whole vector with memset or with an access of a different 1643 // vector type? 1644 if (ValSize == VecSize) 1645 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 1646 1647 uint64_t EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 1648 1649 // Must be an element insertion. 1650 unsigned Elt = Offset/EltSize; 1651 1652 if (SV->getType() != VTy->getElementType()) 1653 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 1654 1655 SV = Builder.CreateInsertElement(Old, SV, 1656 ConstantInt::get(Type::Int32Ty, Elt), 1657 "tmp"); 1658 return SV; 1659 } 1660 1661 // If SV is a first-class aggregate value, insert each value recursively. 1662 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 1663 const StructLayout &Layout = *TD->getStructLayout(ST); 1664 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1665 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1666 Old = ConvertScalar_InsertValue(Elt, Old, 1667 Offset+Layout.getElementOffsetInBits(i), 1668 Builder); 1669 } 1670 return Old; 1671 } 1672 1673 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 1674 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 1675 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1676 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1677 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 1678 } 1679 return Old; 1680 } 1681 1682 // If SV is a float, convert it to the appropriate integer type. 1683 // If it is a pointer, do the same. 1684 unsigned SrcWidth = TD->getTypeSizeInBits(SV->getType()); 1685 unsigned DestWidth = TD->getTypeSizeInBits(AllocaType); 1686 unsigned SrcStoreWidth = TD->getTypeStoreSizeInBits(SV->getType()); 1687 unsigned DestStoreWidth = TD->getTypeStoreSizeInBits(AllocaType); 1688 if (SV->getType()->isFloatingPoint() || isa<VectorType>(SV->getType())) 1689 SV = Builder.CreateBitCast(SV, IntegerType::get(SrcWidth), "tmp"); 1690 else if (isa<PointerType>(SV->getType())) 1691 SV = Builder.CreatePtrToInt(SV, TD->getIntPtrType(), "tmp"); 1692 1693 // Zero extend or truncate the value if needed. 1694 if (SV->getType() != AllocaType) { 1695 if (SV->getType()->getPrimitiveSizeInBits() < 1696 AllocaType->getPrimitiveSizeInBits()) 1697 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 1698 else { 1699 // Truncation may be needed if storing more than the alloca can hold 1700 // (undefined behavior). 1701 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 1702 SrcWidth = DestWidth; 1703 SrcStoreWidth = DestStoreWidth; 1704 } 1705 } 1706 1707 // If this is a big-endian system and the store is narrower than the 1708 // full alloca type, we need to do a shift to get the right bits. 1709 int ShAmt = 0; 1710 if (TD->isBigEndian()) { 1711 // On big-endian machines, the lowest bit is stored at the bit offset 1712 // from the pointer given by getTypeStoreSizeInBits. This matters for 1713 // integers with a bitwidth that is not a multiple of 8. 1714 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 1715 } else { 1716 ShAmt = Offset; 1717 } 1718 1719 // Note: we support negative bitwidths (with shr) which are not defined. 1720 // We do this to support (f.e.) stores off the end of a structure where 1721 // only some bits in the structure are set. 1722 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 1723 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 1724 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), ShAmt), "tmp"); 1725 Mask <<= ShAmt; 1726 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1727 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), -ShAmt), "tmp"); 1728 Mask = Mask.lshr(-ShAmt); 1729 } 1730 1731 // Mask out the bits we are about to insert from the old value, and or 1732 // in the new bits. 1733 if (SrcWidth != DestWidth) { 1734 assert(DestWidth > SrcWidth); 1735 Old = Builder.CreateAnd(Old, ConstantInt::get(~Mask), "mask"); 1736 SV = Builder.CreateOr(Old, SV, "ins"); 1737 } 1738 return SV; 1739} 1740 1741 1742 1743/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1744/// some part of a constant global variable. This intentionally only accepts 1745/// constant expressions because we don't can't rewrite arbitrary instructions. 1746static bool PointsToConstantGlobal(Value *V) { 1747 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1748 return GV->isConstant(); 1749 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1750 if (CE->getOpcode() == Instruction::BitCast || 1751 CE->getOpcode() == Instruction::GetElementPtr) 1752 return PointsToConstantGlobal(CE->getOperand(0)); 1753 return false; 1754} 1755 1756/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1757/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1758/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1759/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1760/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1761/// the alloca, and if the source pointer is a pointer to a constant global, we 1762/// can optimize this. 1763static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy, 1764 bool isOffset) { 1765 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1766 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) 1767 // Ignore non-volatile loads, they are always ok. 1768 if (!LI->isVolatile()) 1769 continue; 1770 1771 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) { 1772 // If uses of the bitcast are ok, we are ok. 1773 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 1774 return false; 1775 continue; 1776 } 1777 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) { 1778 // If the GEP has all zero indices, it doesn't offset the pointer. If it 1779 // doesn't, it does. 1780 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 1781 isOffset || !GEP->hasAllZeroIndices())) 1782 return false; 1783 continue; 1784 } 1785 1786 // If this is isn't our memcpy/memmove, reject it as something we can't 1787 // handle. 1788 if (!isa<MemTransferInst>(*UI)) 1789 return false; 1790 1791 // If we already have seen a copy, reject the second one. 1792 if (TheCopy) return false; 1793 1794 // If the pointer has been offset from the start of the alloca, we can't 1795 // safely handle this. 1796 if (isOffset) return false; 1797 1798 // If the memintrinsic isn't using the alloca as the dest, reject it. 1799 if (UI.getOperandNo() != 1) return false; 1800 1801 MemIntrinsic *MI = cast<MemIntrinsic>(*UI); 1802 1803 // If the source of the memcpy/move is not a constant global, reject it. 1804 if (!PointsToConstantGlobal(MI->getOperand(2))) 1805 return false; 1806 1807 // Otherwise, the transform is safe. Remember the copy instruction. 1808 TheCopy = MI; 1809 } 1810 return true; 1811} 1812 1813/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 1814/// modified by a copy from a constant global. If we can prove this, we can 1815/// replace any uses of the alloca with uses of the global directly. 1816Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocationInst *AI) { 1817 Instruction *TheCopy = 0; 1818 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 1819 return TheCopy; 1820 return 0; 1821}
|