ScalarReplAggregates.cpp revision 194612
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->getScalarType(); 824 825 // Construct an integer with the right value. 826 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 827 APInt OneVal(EltSize, CI->getZExtValue()); 828 APInt TotalVal(OneVal); 829 // Set each byte. 830 for (unsigned i = 0; 8*i < EltSize; ++i) { 831 TotalVal = TotalVal.shl(8); 832 TotalVal |= OneVal; 833 } 834 835 // Convert the integer value to the appropriate type. 836 StoreVal = ConstantInt::get(TotalVal); 837 if (isa<PointerType>(ValTy)) 838 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 839 else if (ValTy->isFloatingPoint()) 840 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 841 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 842 843 // If the requested value was a vector constant, create it. 844 if (EltTy != ValTy) { 845 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 846 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 847 StoreVal = ConstantVector::get(&Elts[0], NumElts); 848 } 849 } 850 new StoreInst(StoreVal, EltPtr, MI); 851 continue; 852 } 853 // Otherwise, if we're storing a byte variable, use a memset call for 854 // this element. 855 } 856 857 // Cast the element pointer to BytePtrTy. 858 if (EltPtr->getType() != BytePtrTy) 859 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI); 860 861 // Cast the other pointer (if we have one) to BytePtrTy. 862 if (OtherElt && OtherElt->getType() != BytePtrTy) 863 OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(), 864 MI); 865 866 unsigned EltSize = TD->getTypeAllocSize(EltTy); 867 868 // Finally, insert the meminst for this element. 869 if (isa<MemTransferInst>(MI)) { 870 Value *Ops[] = { 871 SROADest ? EltPtr : OtherElt, // Dest ptr 872 SROADest ? OtherElt : EltPtr, // Src ptr 873 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 874 ConstantInt::get(Type::Int32Ty, OtherEltAlign) // Align 875 }; 876 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 877 } else { 878 assert(isa<MemSetInst>(MI)); 879 Value *Ops[] = { 880 EltPtr, MI->getOperand(2), // Dest, Value, 881 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 882 Zero // Align 883 }; 884 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 885 } 886 } 887 MI->eraseFromParent(); 888} 889 890/// RewriteStoreUserOfWholeAlloca - We found an store of an integer that 891/// overwrites the entire allocation. Extract out the pieces of the stored 892/// integer and store them individually. 893void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, 894 AllocationInst *AI, 895 SmallVector<AllocaInst*, 32> &NewElts){ 896 // Extract each element out of the integer according to its structure offset 897 // and store the element value to the individual alloca. 898 Value *SrcVal = SI->getOperand(0); 899 const Type *AllocaEltTy = AI->getType()->getElementType(); 900 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 901 902 // If this isn't a store of an integer to the whole alloca, it may be a store 903 // to the first element. Just ignore the store in this case and normal SROA 904 // will handle it. 905 if (!isa<IntegerType>(SrcVal->getType()) || 906 TD->getTypeAllocSizeInBits(SrcVal->getType()) != AllocaSizeBits) 907 return; 908 // Handle tail padding by extending the operand 909 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 910 SrcVal = new ZExtInst(SrcVal, IntegerType::get(AllocaSizeBits), "", SI); 911 912 DOUT << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << *SI; 913 914 // There are two forms here: AI could be an array or struct. Both cases 915 // have different ways to compute the element offset. 916 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 917 const StructLayout *Layout = TD->getStructLayout(EltSTy); 918 919 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 920 // Get the number of bits to shift SrcVal to get the value. 921 const Type *FieldTy = EltSTy->getElementType(i); 922 uint64_t Shift = Layout->getElementOffsetInBits(i); 923 924 if (TD->isBigEndian()) 925 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 926 927 Value *EltVal = SrcVal; 928 if (Shift) { 929 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 930 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 931 "sroa.store.elt", SI); 932 } 933 934 // Truncate down to an integer of the right size. 935 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 936 937 // Ignore zero sized fields like {}, they obviously contain no data. 938 if (FieldSizeBits == 0) continue; 939 940 if (FieldSizeBits != AllocaSizeBits) 941 EltVal = new TruncInst(EltVal, IntegerType::get(FieldSizeBits), "", SI); 942 Value *DestField = NewElts[i]; 943 if (EltVal->getType() == FieldTy) { 944 // Storing to an integer field of this size, just do it. 945 } else if (FieldTy->isFloatingPoint() || isa<VectorType>(FieldTy)) { 946 // Bitcast to the right element type (for fp/vector values). 947 EltVal = new BitCastInst(EltVal, FieldTy, "", SI); 948 } else { 949 // Otherwise, bitcast the dest pointer (for aggregates). 950 DestField = new BitCastInst(DestField, 951 PointerType::getUnqual(EltVal->getType()), 952 "", SI); 953 } 954 new StoreInst(EltVal, DestField, SI); 955 } 956 957 } else { 958 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 959 const Type *ArrayEltTy = ATy->getElementType(); 960 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 961 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 962 963 uint64_t Shift; 964 965 if (TD->isBigEndian()) 966 Shift = AllocaSizeBits-ElementOffset; 967 else 968 Shift = 0; 969 970 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 971 // Ignore zero sized fields like {}, they obviously contain no data. 972 if (ElementSizeBits == 0) continue; 973 974 Value *EltVal = SrcVal; 975 if (Shift) { 976 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 977 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 978 "sroa.store.elt", SI); 979 } 980 981 // Truncate down to an integer of the right size. 982 if (ElementSizeBits != AllocaSizeBits) 983 EltVal = new TruncInst(EltVal, IntegerType::get(ElementSizeBits),"",SI); 984 Value *DestField = NewElts[i]; 985 if (EltVal->getType() == ArrayEltTy) { 986 // Storing to an integer field of this size, just do it. 987 } else if (ArrayEltTy->isFloatingPoint() || isa<VectorType>(ArrayEltTy)) { 988 // Bitcast to the right element type (for fp/vector values). 989 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI); 990 } else { 991 // Otherwise, bitcast the dest pointer (for aggregates). 992 DestField = new BitCastInst(DestField, 993 PointerType::getUnqual(EltVal->getType()), 994 "", SI); 995 } 996 new StoreInst(EltVal, DestField, SI); 997 998 if (TD->isBigEndian()) 999 Shift -= ElementOffset; 1000 else 1001 Shift += ElementOffset; 1002 } 1003 } 1004 1005 SI->eraseFromParent(); 1006} 1007 1008/// RewriteLoadUserOfWholeAlloca - We found an load of the entire allocation to 1009/// an integer. Load the individual pieces to form the aggregate value. 1010void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI, 1011 SmallVector<AllocaInst*, 32> &NewElts) { 1012 // Extract each element out of the NewElts according to its structure offset 1013 // and form the result value. 1014 const Type *AllocaEltTy = AI->getType()->getElementType(); 1015 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1016 1017 // If this isn't a load of the whole alloca to an integer, it may be a load 1018 // of the first element. Just ignore the load in this case and normal SROA 1019 // will handle it. 1020 if (!isa<IntegerType>(LI->getType()) || 1021 TD->getTypeAllocSizeInBits(LI->getType()) != AllocaSizeBits) 1022 return; 1023 1024 DOUT << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << *LI; 1025 1026 // There are two forms here: AI could be an array or struct. Both cases 1027 // have different ways to compute the element offset. 1028 const StructLayout *Layout = 0; 1029 uint64_t ArrayEltBitOffset = 0; 1030 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1031 Layout = TD->getStructLayout(EltSTy); 1032 } else { 1033 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 1034 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1035 } 1036 1037 Value *ResultVal = Constant::getNullValue(IntegerType::get(AllocaSizeBits)); 1038 1039 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1040 // Load the value from the alloca. If the NewElt is an aggregate, cast 1041 // the pointer to an integer of the same size before doing the load. 1042 Value *SrcField = NewElts[i]; 1043 const Type *FieldTy = 1044 cast<PointerType>(SrcField->getType())->getElementType(); 1045 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1046 1047 // Ignore zero sized fields like {}, they obviously contain no data. 1048 if (FieldSizeBits == 0) continue; 1049 1050 const IntegerType *FieldIntTy = IntegerType::get(FieldSizeBits); 1051 if (!isa<IntegerType>(FieldTy) && !FieldTy->isFloatingPoint() && 1052 !isa<VectorType>(FieldTy)) 1053 SrcField = new BitCastInst(SrcField, PointerType::getUnqual(FieldIntTy), 1054 "", LI); 1055 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 1056 1057 // If SrcField is a fp or vector of the right size but that isn't an 1058 // integer type, bitcast to an integer so we can shift it. 1059 if (SrcField->getType() != FieldIntTy) 1060 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 1061 1062 // Zero extend the field to be the same size as the final alloca so that 1063 // we can shift and insert it. 1064 if (SrcField->getType() != ResultVal->getType()) 1065 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 1066 1067 // Determine the number of bits to shift SrcField. 1068 uint64_t Shift; 1069 if (Layout) // Struct case. 1070 Shift = Layout->getElementOffsetInBits(i); 1071 else // Array case. 1072 Shift = i*ArrayEltBitOffset; 1073 1074 if (TD->isBigEndian()) 1075 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 1076 1077 if (Shift) { 1078 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 1079 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 1080 } 1081 1082 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 1083 } 1084 1085 // Handle tail padding by truncating the result 1086 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 1087 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 1088 1089 LI->replaceAllUsesWith(ResultVal); 1090 LI->eraseFromParent(); 1091} 1092 1093 1094/// HasPadding - Return true if the specified type has any structure or 1095/// alignment padding, false otherwise. 1096static bool HasPadding(const Type *Ty, const TargetData &TD) { 1097 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 1098 const StructLayout *SL = TD.getStructLayout(STy); 1099 unsigned PrevFieldBitOffset = 0; 1100 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1101 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 1102 1103 // Padding in sub-elements? 1104 if (HasPadding(STy->getElementType(i), TD)) 1105 return true; 1106 1107 // Check to see if there is any padding between this element and the 1108 // previous one. 1109 if (i) { 1110 unsigned PrevFieldEnd = 1111 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 1112 if (PrevFieldEnd < FieldBitOffset) 1113 return true; 1114 } 1115 1116 PrevFieldBitOffset = FieldBitOffset; 1117 } 1118 1119 // Check for tail padding. 1120 if (unsigned EltCount = STy->getNumElements()) { 1121 unsigned PrevFieldEnd = PrevFieldBitOffset + 1122 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 1123 if (PrevFieldEnd < SL->getSizeInBits()) 1124 return true; 1125 } 1126 1127 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1128 return HasPadding(ATy->getElementType(), TD); 1129 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) { 1130 return HasPadding(VTy->getElementType(), TD); 1131 } 1132 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 1133} 1134 1135/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 1136/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 1137/// or 1 if safe after canonicalization has been performed. 1138/// 1139int SROA::isSafeAllocaToScalarRepl(AllocationInst *AI) { 1140 // Loop over the use list of the alloca. We can only transform it if all of 1141 // the users are safe to transform. 1142 AllocaInfo Info; 1143 1144 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); 1145 I != E; ++I) { 1146 isSafeUseOfAllocation(cast<Instruction>(*I), AI, Info); 1147 if (Info.isUnsafe) { 1148 DOUT << "Cannot transform: " << *AI << " due to user: " << **I; 1149 return 0; 1150 } 1151 } 1152 1153 // Okay, we know all the users are promotable. If the aggregate is a memcpy 1154 // source and destination, we have to be careful. In particular, the memcpy 1155 // could be moving around elements that live in structure padding of the LLVM 1156 // types, but may actually be used. In these cases, we refuse to promote the 1157 // struct. 1158 if (Info.isMemCpySrc && Info.isMemCpyDst && 1159 HasPadding(AI->getType()->getElementType(), *TD)) 1160 return 0; 1161 1162 // If we require cleanup, return 1, otherwise return 3. 1163 return Info.needsCleanup ? 1 : 3; 1164} 1165 1166/// CleanupGEP - GEP is used by an Alloca, which can be prompted after the GEP 1167/// is canonicalized here. 1168void SROA::CleanupGEP(GetElementPtrInst *GEPI) { 1169 gep_type_iterator I = gep_type_begin(GEPI); 1170 ++I; 1171 1172 const ArrayType *AT = dyn_cast<ArrayType>(*I); 1173 if (!AT) 1174 return; 1175 1176 uint64_t NumElements = AT->getNumElements(); 1177 1178 if (isa<ConstantInt>(I.getOperand())) 1179 return; 1180 1181 if (NumElements == 1) { 1182 GEPI->setOperand(2, Constant::getNullValue(Type::Int32Ty)); 1183 return; 1184 } 1185 1186 assert(NumElements == 2 && "Unhandled case!"); 1187 // All users of the GEP must be loads. At each use of the GEP, insert 1188 // two loads of the appropriate indexed GEP and select between them. 1189 Value *IsOne = new ICmpInst(ICmpInst::ICMP_NE, I.getOperand(), 1190 Constant::getNullValue(I.getOperand()->getType()), 1191 "isone", GEPI); 1192 // Insert the new GEP instructions, which are properly indexed. 1193 SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end()); 1194 Indices[1] = Constant::getNullValue(Type::Int32Ty); 1195 Value *ZeroIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 1196 Indices.begin(), 1197 Indices.end(), 1198 GEPI->getName()+".0", GEPI); 1199 Indices[1] = ConstantInt::get(Type::Int32Ty, 1); 1200 Value *OneIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 1201 Indices.begin(), 1202 Indices.end(), 1203 GEPI->getName()+".1", GEPI); 1204 // Replace all loads of the variable index GEP with loads from both 1205 // indexes and a select. 1206 while (!GEPI->use_empty()) { 1207 LoadInst *LI = cast<LoadInst>(GEPI->use_back()); 1208 Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI); 1209 Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI); 1210 Value *R = SelectInst::Create(IsOne, One, Zero, LI->getName(), LI); 1211 LI->replaceAllUsesWith(R); 1212 LI->eraseFromParent(); 1213 } 1214 GEPI->eraseFromParent(); 1215} 1216 1217 1218/// CleanupAllocaUsers - If SROA reported that it can promote the specified 1219/// allocation, but only if cleaned up, perform the cleanups required. 1220void SROA::CleanupAllocaUsers(AllocationInst *AI) { 1221 // At this point, we know that the end result will be SROA'd and promoted, so 1222 // we can insert ugly code if required so long as sroa+mem2reg will clean it 1223 // up. 1224 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 1225 UI != E; ) { 1226 User *U = *UI++; 1227 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) 1228 CleanupGEP(GEPI); 1229 else { 1230 Instruction *I = cast<Instruction>(U); 1231 SmallVector<DbgInfoIntrinsic *, 2> DbgInUses; 1232 if (!isa<StoreInst>(I) && OnlyUsedByDbgInfoIntrinsics(I, &DbgInUses)) { 1233 // Safe to remove debug info uses. 1234 while (!DbgInUses.empty()) { 1235 DbgInfoIntrinsic *DI = DbgInUses.back(); DbgInUses.pop_back(); 1236 DI->eraseFromParent(); 1237 } 1238 I->eraseFromParent(); 1239 } 1240 } 1241 } 1242} 1243 1244/// MergeInType - Add the 'In' type to the accumulated type (Accum) so far at 1245/// the offset specified by Offset (which is specified in bytes). 1246/// 1247/// There are two cases we handle here: 1248/// 1) A union of vector types of the same size and potentially its elements. 1249/// Here we turn element accesses into insert/extract element operations. 1250/// This promotes a <4 x float> with a store of float to the third element 1251/// into a <4 x float> that uses insert element. 1252/// 2) A fully general blob of memory, which we turn into some (potentially 1253/// large) integer type with extract and insert operations where the loads 1254/// and stores would mutate the memory. 1255static void MergeInType(const Type *In, uint64_t Offset, const Type *&VecTy, 1256 unsigned AllocaSize, const TargetData &TD) { 1257 // If this could be contributing to a vector, analyze it. 1258 if (VecTy != Type::VoidTy) { // either null or a vector type. 1259 1260 // If the In type is a vector that is the same size as the alloca, see if it 1261 // matches the existing VecTy. 1262 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 1263 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 1264 // If we're storing/loading a vector of the right size, allow it as a 1265 // vector. If this the first vector we see, remember the type so that 1266 // we know the element size. 1267 if (VecTy == 0) 1268 VecTy = VInTy; 1269 return; 1270 } 1271 } else if (In == Type::FloatTy || In == Type::DoubleTy || 1272 (isa<IntegerType>(In) && In->getPrimitiveSizeInBits() >= 8 && 1273 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 1274 // If we're accessing something that could be an element of a vector, see 1275 // if the implied vector agrees with what we already have and if Offset is 1276 // compatible with it. 1277 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 1278 if (Offset % EltSize == 0 && 1279 AllocaSize % EltSize == 0 && 1280 (VecTy == 0 || 1281 cast<VectorType>(VecTy)->getElementType() 1282 ->getPrimitiveSizeInBits()/8 == EltSize)) { 1283 if (VecTy == 0) 1284 VecTy = VectorType::get(In, AllocaSize/EltSize); 1285 return; 1286 } 1287 } 1288 } 1289 1290 // Otherwise, we have a case that we can't handle with an optimized vector 1291 // form. We can still turn this into a large integer. 1292 VecTy = Type::VoidTy; 1293} 1294 1295/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 1296/// its accesses to use a to single vector type, return true, and set VecTy to 1297/// the new type. If we could convert the alloca into a single promotable 1298/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 1299/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 1300/// is the current offset from the base of the alloca being analyzed. 1301/// 1302/// If we see at least one access to the value that is as a vector type, set the 1303/// SawVec flag. 1304/// 1305bool SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy, 1306 bool &SawVec, uint64_t Offset, 1307 unsigned AllocaSize) { 1308 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1309 Instruction *User = cast<Instruction>(*UI); 1310 1311 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1312 // Don't break volatile loads. 1313 if (LI->isVolatile()) 1314 return false; 1315 MergeInType(LI->getType(), Offset, VecTy, AllocaSize, *TD); 1316 SawVec |= isa<VectorType>(LI->getType()); 1317 continue; 1318 } 1319 1320 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1321 // Storing the pointer, not into the value? 1322 if (SI->getOperand(0) == V || SI->isVolatile()) return 0; 1323 MergeInType(SI->getOperand(0)->getType(), Offset, VecTy, AllocaSize, *TD); 1324 SawVec |= isa<VectorType>(SI->getOperand(0)->getType()); 1325 continue; 1326 } 1327 1328 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 1329 if (!CanConvertToScalar(BCI, IsNotTrivial, VecTy, SawVec, Offset, 1330 AllocaSize)) 1331 return false; 1332 IsNotTrivial = true; 1333 continue; 1334 } 1335 1336 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1337 // If this is a GEP with a variable indices, we can't handle it. 1338 if (!GEP->hasAllConstantIndices()) 1339 return false; 1340 1341 // Compute the offset that this GEP adds to the pointer. 1342 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1343 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(), 1344 &Indices[0], Indices.size()); 1345 // See if all uses can be converted. 1346 if (!CanConvertToScalar(GEP, IsNotTrivial, VecTy, SawVec,Offset+GEPOffset, 1347 AllocaSize)) 1348 return false; 1349 IsNotTrivial = true; 1350 continue; 1351 } 1352 1353 // If this is a constant sized memset of a constant value (e.g. 0) we can 1354 // handle it. 1355 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 1356 // Store of constant value and constant size. 1357 if (isa<ConstantInt>(MSI->getValue()) && 1358 isa<ConstantInt>(MSI->getLength())) { 1359 IsNotTrivial = true; 1360 continue; 1361 } 1362 } 1363 1364 // If this is a memcpy or memmove into or out of the whole allocation, we 1365 // can handle it like a load or store of the scalar type. 1366 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 1367 if (ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength())) 1368 if (Len->getZExtValue() == AllocaSize && Offset == 0) { 1369 IsNotTrivial = true; 1370 continue; 1371 } 1372 } 1373 1374 // Ignore dbg intrinsic. 1375 if (isa<DbgInfoIntrinsic>(User)) 1376 continue; 1377 1378 // Otherwise, we cannot handle this! 1379 return false; 1380 } 1381 1382 return true; 1383} 1384 1385 1386/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 1387/// directly. This happens when we are converting an "integer union" to a 1388/// single integer scalar, or when we are converting a "vector union" to a 1389/// vector with insert/extractelement instructions. 1390/// 1391/// Offset is an offset from the original alloca, in bits that need to be 1392/// shifted to the right. By the end of this, there should be no uses of Ptr. 1393void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset) { 1394 while (!Ptr->use_empty()) { 1395 Instruction *User = cast<Instruction>(Ptr->use_back()); 1396 1397 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 1398 ConvertUsesToScalar(CI, NewAI, Offset); 1399 CI->eraseFromParent(); 1400 continue; 1401 } 1402 1403 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1404 // Compute the offset that this GEP adds to the pointer. 1405 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1406 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(), 1407 &Indices[0], Indices.size()); 1408 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 1409 GEP->eraseFromParent(); 1410 continue; 1411 } 1412 1413 IRBuilder<> Builder(User->getParent(), User); 1414 1415 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1416 // The load is a bit extract from NewAI shifted right by Offset bits. 1417 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 1418 Value *NewLoadVal 1419 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 1420 LI->replaceAllUsesWith(NewLoadVal); 1421 LI->eraseFromParent(); 1422 continue; 1423 } 1424 1425 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1426 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 1427 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").c_str()); 1428 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 1429 Builder); 1430 Builder.CreateStore(New, NewAI); 1431 SI->eraseFromParent(); 1432 continue; 1433 } 1434 1435 // If this is a constant sized memset of a constant value (e.g. 0) we can 1436 // transform it into a store of the expanded constant value. 1437 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 1438 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 1439 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 1440 if (NumBytes != 0) { 1441 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 1442 1443 // Compute the value replicated the right number of times. 1444 APInt APVal(NumBytes*8, Val); 1445 1446 // Splat the value if non-zero. 1447 if (Val) 1448 for (unsigned i = 1; i != NumBytes; ++i) 1449 APVal |= APVal << 8; 1450 1451 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").c_str()); 1452 Value *New = ConvertScalar_InsertValue(ConstantInt::get(APVal), Old, 1453 Offset, Builder); 1454 Builder.CreateStore(New, NewAI); 1455 } 1456 MSI->eraseFromParent(); 1457 continue; 1458 } 1459 1460 // If this is a memcpy or memmove into or out of the whole allocation, we 1461 // can handle it like a load or store of the scalar type. 1462 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 1463 assert(Offset == 0 && "must be store to start of alloca"); 1464 1465 // If the source and destination are both to the same alloca, then this is 1466 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 1467 // as appropriate. 1468 AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject()); 1469 1470 if (MTI->getSource()->getUnderlyingObject() != OrigAI) { 1471 // Dest must be OrigAI, change this to be a load from the original 1472 // pointer (bitcasted), then a store to our new alloca. 1473 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 1474 Value *SrcPtr = MTI->getSource(); 1475 SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType()); 1476 1477 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 1478 SrcVal->setAlignment(MTI->getAlignment()); 1479 Builder.CreateStore(SrcVal, NewAI); 1480 } else if (MTI->getDest()->getUnderlyingObject() != OrigAI) { 1481 // Src must be OrigAI, change this to be a load from NewAI then a store 1482 // through the original dest pointer (bitcasted). 1483 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 1484 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 1485 1486 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType()); 1487 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 1488 NewStore->setAlignment(MTI->getAlignment()); 1489 } else { 1490 // Noop transfer. Src == Dst 1491 } 1492 1493 1494 MTI->eraseFromParent(); 1495 continue; 1496 } 1497 1498 // If user is a dbg info intrinsic then it is safe to remove it. 1499 if (isa<DbgInfoIntrinsic>(User)) { 1500 User->eraseFromParent(); 1501 continue; 1502 } 1503 1504 assert(0 && "Unsupported operation!"); 1505 abort(); 1506 } 1507} 1508 1509/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 1510/// or vector value FromVal, extracting the bits from the offset specified by 1511/// Offset. This returns the value, which is of type ToType. 1512/// 1513/// This happens when we are converting an "integer union" to a single 1514/// integer scalar, or when we are converting a "vector union" to a vector with 1515/// insert/extractelement instructions. 1516/// 1517/// Offset is an offset from the original alloca, in bits that need to be 1518/// shifted to the right. 1519Value *SROA::ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 1520 uint64_t Offset, IRBuilder<> &Builder) { 1521 // If the load is of the whole new alloca, no conversion is needed. 1522 if (FromVal->getType() == ToType && Offset == 0) 1523 return FromVal; 1524 1525 // If the result alloca is a vector type, this is either an element 1526 // access or a bitcast to another vector type of the same size. 1527 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) { 1528 if (isa<VectorType>(ToType)) 1529 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 1530 1531 // Otherwise it must be an element access. 1532 unsigned Elt = 0; 1533 if (Offset) { 1534 unsigned EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 1535 Elt = Offset/EltSize; 1536 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 1537 } 1538 // Return the element extracted out of it. 1539 Value *V = Builder.CreateExtractElement(FromVal, 1540 ConstantInt::get(Type::Int32Ty,Elt), 1541 "tmp"); 1542 if (V->getType() != ToType) 1543 V = Builder.CreateBitCast(V, ToType, "tmp"); 1544 return V; 1545 } 1546 1547 // If ToType is a first class aggregate, extract out each of the pieces and 1548 // use insertvalue's to form the FCA. 1549 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 1550 const StructLayout &Layout = *TD->getStructLayout(ST); 1551 Value *Res = UndefValue::get(ST); 1552 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1553 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 1554 Offset+Layout.getElementOffsetInBits(i), 1555 Builder); 1556 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1557 } 1558 return Res; 1559 } 1560 1561 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 1562 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 1563 Value *Res = UndefValue::get(AT); 1564 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1565 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 1566 Offset+i*EltSize, Builder); 1567 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1568 } 1569 return Res; 1570 } 1571 1572 // Otherwise, this must be a union that was converted to an integer value. 1573 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 1574 1575 // If this is a big-endian system and the load is narrower than the 1576 // full alloca type, we need to do a shift to get the right bits. 1577 int ShAmt = 0; 1578 if (TD->isBigEndian()) { 1579 // On big-endian machines, the lowest bit is stored at the bit offset 1580 // from the pointer given by getTypeStoreSizeInBits. This matters for 1581 // integers with a bitwidth that is not a multiple of 8. 1582 ShAmt = TD->getTypeStoreSizeInBits(NTy) - 1583 TD->getTypeStoreSizeInBits(ToType) - Offset; 1584 } else { 1585 ShAmt = Offset; 1586 } 1587 1588 // Note: we support negative bitwidths (with shl) which are not defined. 1589 // We do this to support (f.e.) loads off the end of a structure where 1590 // only some bits are used. 1591 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 1592 FromVal = Builder.CreateLShr(FromVal, ConstantInt::get(FromVal->getType(), 1593 ShAmt), "tmp"); 1594 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 1595 FromVal = Builder.CreateShl(FromVal, ConstantInt::get(FromVal->getType(), 1596 -ShAmt), "tmp"); 1597 1598 // Finally, unconditionally truncate the integer to the right width. 1599 unsigned LIBitWidth = TD->getTypeSizeInBits(ToType); 1600 if (LIBitWidth < NTy->getBitWidth()) 1601 FromVal = Builder.CreateTrunc(FromVal, IntegerType::get(LIBitWidth), "tmp"); 1602 else if (LIBitWidth > NTy->getBitWidth()) 1603 FromVal = Builder.CreateZExt(FromVal, IntegerType::get(LIBitWidth), "tmp"); 1604 1605 // If the result is an integer, this is a trunc or bitcast. 1606 if (isa<IntegerType>(ToType)) { 1607 // Should be done. 1608 } else if (ToType->isFloatingPoint() || isa<VectorType>(ToType)) { 1609 // Just do a bitcast, we know the sizes match up. 1610 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 1611 } else { 1612 // Otherwise must be a pointer. 1613 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 1614 } 1615 assert(FromVal->getType() == ToType && "Didn't convert right?"); 1616 return FromVal; 1617} 1618 1619 1620/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 1621/// or vector value "Old" at the offset specified by Offset. 1622/// 1623/// This happens when we are converting an "integer union" to a 1624/// single integer scalar, or when we are converting a "vector union" to a 1625/// vector with insert/extractelement instructions. 1626/// 1627/// Offset is an offset from the original alloca, in bits that need to be 1628/// shifted to the right. 1629Value *SROA::ConvertScalar_InsertValue(Value *SV, Value *Old, 1630 uint64_t Offset, IRBuilder<> &Builder) { 1631 1632 // Convert the stored type to the actual type, shift it left to insert 1633 // then 'or' into place. 1634 const Type *AllocaType = Old->getType(); 1635 1636 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 1637 uint64_t VecSize = TD->getTypeAllocSizeInBits(VTy); 1638 uint64_t ValSize = TD->getTypeAllocSizeInBits(SV->getType()); 1639 1640 // Changing the whole vector with memset or with an access of a different 1641 // vector type? 1642 if (ValSize == VecSize) 1643 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 1644 1645 uint64_t EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 1646 1647 // Must be an element insertion. 1648 unsigned Elt = Offset/EltSize; 1649 1650 if (SV->getType() != VTy->getElementType()) 1651 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 1652 1653 SV = Builder.CreateInsertElement(Old, SV, 1654 ConstantInt::get(Type::Int32Ty, Elt), 1655 "tmp"); 1656 return SV; 1657 } 1658 1659 // If SV is a first-class aggregate value, insert each value recursively. 1660 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 1661 const StructLayout &Layout = *TD->getStructLayout(ST); 1662 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1663 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1664 Old = ConvertScalar_InsertValue(Elt, Old, 1665 Offset+Layout.getElementOffsetInBits(i), 1666 Builder); 1667 } 1668 return Old; 1669 } 1670 1671 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 1672 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 1673 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1674 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1675 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 1676 } 1677 return Old; 1678 } 1679 1680 // If SV is a float, convert it to the appropriate integer type. 1681 // If it is a pointer, do the same. 1682 unsigned SrcWidth = TD->getTypeSizeInBits(SV->getType()); 1683 unsigned DestWidth = TD->getTypeSizeInBits(AllocaType); 1684 unsigned SrcStoreWidth = TD->getTypeStoreSizeInBits(SV->getType()); 1685 unsigned DestStoreWidth = TD->getTypeStoreSizeInBits(AllocaType); 1686 if (SV->getType()->isFloatingPoint() || isa<VectorType>(SV->getType())) 1687 SV = Builder.CreateBitCast(SV, IntegerType::get(SrcWidth), "tmp"); 1688 else if (isa<PointerType>(SV->getType())) 1689 SV = Builder.CreatePtrToInt(SV, TD->getIntPtrType(), "tmp"); 1690 1691 // Zero extend or truncate the value if needed. 1692 if (SV->getType() != AllocaType) { 1693 if (SV->getType()->getPrimitiveSizeInBits() < 1694 AllocaType->getPrimitiveSizeInBits()) 1695 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 1696 else { 1697 // Truncation may be needed if storing more than the alloca can hold 1698 // (undefined behavior). 1699 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 1700 SrcWidth = DestWidth; 1701 SrcStoreWidth = DestStoreWidth; 1702 } 1703 } 1704 1705 // If this is a big-endian system and the store is narrower than the 1706 // full alloca type, we need to do a shift to get the right bits. 1707 int ShAmt = 0; 1708 if (TD->isBigEndian()) { 1709 // On big-endian machines, the lowest bit is stored at the bit offset 1710 // from the pointer given by getTypeStoreSizeInBits. This matters for 1711 // integers with a bitwidth that is not a multiple of 8. 1712 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 1713 } else { 1714 ShAmt = Offset; 1715 } 1716 1717 // Note: we support negative bitwidths (with shr) which are not defined. 1718 // We do this to support (f.e.) stores off the end of a structure where 1719 // only some bits in the structure are set. 1720 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 1721 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 1722 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), ShAmt), "tmp"); 1723 Mask <<= ShAmt; 1724 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1725 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), -ShAmt), "tmp"); 1726 Mask = Mask.lshr(-ShAmt); 1727 } 1728 1729 // Mask out the bits we are about to insert from the old value, and or 1730 // in the new bits. 1731 if (SrcWidth != DestWidth) { 1732 assert(DestWidth > SrcWidth); 1733 Old = Builder.CreateAnd(Old, ConstantInt::get(~Mask), "mask"); 1734 SV = Builder.CreateOr(Old, SV, "ins"); 1735 } 1736 return SV; 1737} 1738 1739 1740 1741/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1742/// some part of a constant global variable. This intentionally only accepts 1743/// constant expressions because we don't can't rewrite arbitrary instructions. 1744static bool PointsToConstantGlobal(Value *V) { 1745 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1746 return GV->isConstant(); 1747 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1748 if (CE->getOpcode() == Instruction::BitCast || 1749 CE->getOpcode() == Instruction::GetElementPtr) 1750 return PointsToConstantGlobal(CE->getOperand(0)); 1751 return false; 1752} 1753 1754/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1755/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1756/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1757/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1758/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1759/// the alloca, and if the source pointer is a pointer to a constant global, we 1760/// can optimize this. 1761static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy, 1762 bool isOffset) { 1763 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1764 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) 1765 // Ignore non-volatile loads, they are always ok. 1766 if (!LI->isVolatile()) 1767 continue; 1768 1769 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) { 1770 // If uses of the bitcast are ok, we are ok. 1771 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 1772 return false; 1773 continue; 1774 } 1775 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) { 1776 // If the GEP has all zero indices, it doesn't offset the pointer. If it 1777 // doesn't, it does. 1778 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 1779 isOffset || !GEP->hasAllZeroIndices())) 1780 return false; 1781 continue; 1782 } 1783 1784 // If this is isn't our memcpy/memmove, reject it as something we can't 1785 // handle. 1786 if (!isa<MemTransferInst>(*UI)) 1787 return false; 1788 1789 // If we already have seen a copy, reject the second one. 1790 if (TheCopy) return false; 1791 1792 // If the pointer has been offset from the start of the alloca, we can't 1793 // safely handle this. 1794 if (isOffset) return false; 1795 1796 // If the memintrinsic isn't using the alloca as the dest, reject it. 1797 if (UI.getOperandNo() != 1) return false; 1798 1799 MemIntrinsic *MI = cast<MemIntrinsic>(*UI); 1800 1801 // If the source of the memcpy/move is not a constant global, reject it. 1802 if (!PointsToConstantGlobal(MI->getOperand(2))) 1803 return false; 1804 1805 // Otherwise, the transform is safe. Remember the copy instruction. 1806 TheCopy = MI; 1807 } 1808 return true; 1809} 1810 1811/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 1812/// modified by a copy from a constant global. If we can prove this, we can 1813/// replace any uses of the alloca with uses of the global directly. 1814Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocationInst *AI) { 1815 Instruction *TheCopy = 0; 1816 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 1817 return TheCopy; 1818 return 0; 1819} 1820