ScalarReplAggregates.cpp revision c23197a26f34f559ea9797de51e187087c039c42
1//===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This transformation implements the well known scalar replacement of 11// aggregates transformation. This xform breaks up alloca instructions of 12// aggregate type (structure or array) into individual alloca instructions for 13// each member (if possible). Then, if possible, it transforms the individual 14// alloca instructions into nice clean scalar SSA form. 15// 16// This combines a simple SRoA algorithm with the Mem2Reg algorithm because 17// often interact, especially for C++ programs. As such, iterating between 18// SRoA, then Mem2Reg until we run out of things to promote works well. 19// 20//===----------------------------------------------------------------------===// 21 22#define DEBUG_TYPE "scalarrepl" 23#include "llvm/Transforms/Scalar.h" 24#include "llvm/Constants.h" 25#include "llvm/DerivedTypes.h" 26#include "llvm/Function.h" 27#include "llvm/GlobalVariable.h" 28#include "llvm/Instructions.h" 29#include "llvm/IntrinsicInst.h" 30#include "llvm/LLVMContext.h" 31#include "llvm/Pass.h" 32#include "llvm/Analysis/Dominators.h" 33#include "llvm/Target/TargetData.h" 34#include "llvm/Transforms/Utils/PromoteMemToReg.h" 35#include "llvm/Transforms/Utils/Local.h" 36#include "llvm/Support/Debug.h" 37#include "llvm/Support/ErrorHandling.h" 38#include "llvm/Support/GetElementPtrTypeIterator.h" 39#include "llvm/Support/IRBuilder.h" 40#include "llvm/Support/MathExtras.h" 41#include "llvm/Support/Compiler.h" 42#include "llvm/ADT/SmallVector.h" 43#include "llvm/ADT/Statistic.h" 44#include "llvm/ADT/StringExtras.h" 45using namespace llvm; 46 47STATISTIC(NumReplaced, "Number of allocas broken up"); 48STATISTIC(NumPromoted, "Number of allocas promoted"); 49STATISTIC(NumConverted, "Number of aggregates converted to scalar"); 50STATISTIC(NumGlobals, "Number of allocas copied from constant global"); 51 52namespace { 53 struct VISIBILITY_HIDDEN SROA : public FunctionPass { 54 static char ID; // Pass identification, replacement for typeid 55 explicit SROA(signed T = -1) : FunctionPass(&ID) { 56 if (T == -1) 57 SRThreshold = 128; 58 else 59 SRThreshold = T; 60 } 61 62 bool runOnFunction(Function &F); 63 64 bool performScalarRepl(Function &F); 65 bool performPromotion(Function &F); 66 67 // getAnalysisUsage - This pass does not require any passes, but we know it 68 // will not alter the CFG, so say so. 69 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 70 AU.addRequired<DominatorTree>(); 71 AU.addRequired<DominanceFrontier>(); 72 AU.addRequired<TargetData>(); 73 AU.setPreservesCFG(); 74 } 75 76 private: 77 TargetData *TD; 78 79 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures 80 /// information about the uses. All these fields are initialized to false 81 /// and set to true when something is learned. 82 struct AllocaInfo { 83 /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 84 bool isUnsafe : 1; 85 86 /// needsCleanup - This is set to true if there is some use of the alloca 87 /// that requires cleanup. 88 bool needsCleanup : 1; 89 90 /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 91 bool isMemCpySrc : 1; 92 93 /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 94 bool isMemCpyDst : 1; 95 96 AllocaInfo() 97 : isUnsafe(false), needsCleanup(false), 98 isMemCpySrc(false), isMemCpyDst(false) {} 99 }; 100 101 unsigned SRThreshold; 102 103 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; } 104 105 int isSafeAllocaToScalarRepl(AllocationInst *AI); 106 107 void isSafeUseOfAllocation(Instruction *User, AllocationInst *AI, 108 AllocaInfo &Info); 109 void isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI, 110 AllocaInfo &Info); 111 void isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI, 112 unsigned OpNo, AllocaInfo &Info); 113 void isSafeUseOfBitCastedAllocation(BitCastInst *User, AllocationInst *AI, 114 AllocaInfo &Info); 115 116 void DoScalarReplacement(AllocationInst *AI, 117 std::vector<AllocationInst*> &WorkList); 118 void CleanupGEP(GetElementPtrInst *GEP); 119 void CleanupAllocaUsers(AllocationInst *AI); 120 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocationInst *Base); 121 122 void RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI, 123 SmallVector<AllocaInst*, 32> &NewElts); 124 125 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst, 126 AllocationInst *AI, 127 SmallVector<AllocaInst*, 32> &NewElts); 128 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocationInst *AI, 129 SmallVector<AllocaInst*, 32> &NewElts); 130 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI, 131 SmallVector<AllocaInst*, 32> &NewElts); 132 133 bool CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy, 134 bool &SawVec, uint64_t Offset, unsigned AllocaSize); 135 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); 136 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType, 137 uint64_t Offset, IRBuilder<> &Builder); 138 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 139 uint64_t Offset, IRBuilder<> &Builder); 140 static Instruction *isOnlyCopiedFromConstantGlobal(AllocationInst *AI); 141 }; 142} 143 144char SROA::ID = 0; 145static RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates"); 146 147// Public interface to the ScalarReplAggregates pass 148FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) { 149 return new SROA(Threshold); 150} 151 152 153bool SROA::runOnFunction(Function &F) { 154 TD = &getAnalysis<TargetData>(); 155 156 bool Changed = performPromotion(F); 157 while (1) { 158 bool LocalChange = performScalarRepl(F); 159 if (!LocalChange) break; // No need to repromote if no scalarrepl 160 Changed = true; 161 LocalChange = performPromotion(F); 162 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 163 } 164 165 return Changed; 166} 167 168 169bool SROA::performPromotion(Function &F) { 170 std::vector<AllocaInst*> Allocas; 171 DominatorTree &DT = getAnalysis<DominatorTree>(); 172 DominanceFrontier &DF = getAnalysis<DominanceFrontier>(); 173 174 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 175 176 bool Changed = false; 177 178 while (1) { 179 Allocas.clear(); 180 181 // Find allocas that are safe to promote, by looking at all instructions in 182 // the entry node 183 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 184 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 185 if (isAllocaPromotable(AI)) 186 Allocas.push_back(AI); 187 188 if (Allocas.empty()) break; 189 190 PromoteMemToReg(Allocas, DT, DF, Context); 191 NumPromoted += Allocas.size(); 192 Changed = true; 193 } 194 195 return Changed; 196} 197 198/// getNumSAElements - Return the number of elements in the specific struct or 199/// array. 200static uint64_t getNumSAElements(const Type *T) { 201 if (const StructType *ST = dyn_cast<StructType>(T)) 202 return ST->getNumElements(); 203 return cast<ArrayType>(T)->getNumElements(); 204} 205 206// performScalarRepl - This algorithm is a simple worklist driven algorithm, 207// which runs on all of the malloc/alloca instructions in the function, removing 208// them if they are only used by getelementptr instructions. 209// 210bool SROA::performScalarRepl(Function &F) { 211 std::vector<AllocationInst*> WorkList; 212 213 // Scan the entry basic block, adding any alloca's and mallocs to the worklist 214 BasicBlock &BB = F.getEntryBlock(); 215 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 216 if (AllocationInst *A = dyn_cast<AllocationInst>(I)) 217 WorkList.push_back(A); 218 219 // Process the worklist 220 bool Changed = false; 221 while (!WorkList.empty()) { 222 AllocationInst *AI = WorkList.back(); 223 WorkList.pop_back(); 224 225 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 226 // with unused elements. 227 if (AI->use_empty()) { 228 AI->eraseFromParent(); 229 continue; 230 } 231 232 // If this alloca is impossible for us to promote, reject it early. 233 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 234 continue; 235 236 // Check to see if this allocation is only modified by a memcpy/memmove from 237 // a constant global. If this is the case, we can change all users to use 238 // the constant global instead. This is commonly produced by the CFE by 239 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 240 // is only subsequently read. 241 if (Instruction *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 242 DOUT << "Found alloca equal to global: " << *AI; 243 DOUT << " memcpy = " << *TheCopy; 244 Constant *TheSrc = cast<Constant>(TheCopy->getOperand(2)); 245 AI->replaceAllUsesWith( 246 Context->getConstantExprBitCast(TheSrc, AI->getType())); 247 TheCopy->eraseFromParent(); // Don't mutate the global. 248 AI->eraseFromParent(); 249 ++NumGlobals; 250 Changed = true; 251 continue; 252 } 253 254 // Check to see if we can perform the core SROA transformation. We cannot 255 // transform the allocation instruction if it is an array allocation 256 // (allocations OF arrays are ok though), and an allocation of a scalar 257 // value cannot be decomposed at all. 258 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 259 260 // Do not promote any struct whose size is too big. 261 if (AllocaSize > SRThreshold) continue; 262 263 if ((isa<StructType>(AI->getAllocatedType()) || 264 isa<ArrayType>(AI->getAllocatedType())) && 265 // Do not promote any struct into more than "32" separate vars. 266 getNumSAElements(AI->getAllocatedType()) <= SRThreshold/4) { 267 // Check that all of the users of the allocation are capable of being 268 // transformed. 269 switch (isSafeAllocaToScalarRepl(AI)) { 270 default: llvm_unreachable("Unexpected value!"); 271 case 0: // Not safe to scalar replace. 272 break; 273 case 1: // Safe, but requires cleanup/canonicalizations first 274 CleanupAllocaUsers(AI); 275 // FALL THROUGH. 276 case 3: // Safe to scalar replace. 277 DoScalarReplacement(AI, WorkList); 278 Changed = true; 279 continue; 280 } 281 } 282 283 // If we can turn this aggregate value (potentially with casts) into a 284 // simple scalar value that can be mem2reg'd into a register value. 285 // IsNotTrivial tracks whether this is something that mem2reg could have 286 // promoted itself. If so, we don't want to transform it needlessly. Note 287 // that we can't just check based on the type: the alloca may be of an i32 288 // but that has pointer arithmetic to set byte 3 of it or something. 289 bool IsNotTrivial = false; 290 const Type *VectorTy = 0; 291 bool HadAVector = false; 292 if (CanConvertToScalar(AI, IsNotTrivial, VectorTy, HadAVector, 293 0, unsigned(AllocaSize)) && IsNotTrivial) { 294 AllocaInst *NewAI; 295 // If we were able to find a vector type that can handle this with 296 // insert/extract elements, and if there was at least one use that had 297 // a vector type, promote this to a vector. We don't want to promote 298 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 299 // we just get a lot of insert/extracts. If at least one vector is 300 // involved, then we probably really do have a union of vector/array. 301 if (VectorTy && isa<VectorType>(VectorTy) && HadAVector) { 302 DOUT << "CONVERT TO VECTOR: " << *AI << " TYPE = " << *VectorTy <<"\n"; 303 304 // Create and insert the vector alloca. 305 NewAI = new AllocaInst(VectorTy, 0, "", AI->getParent()->begin()); 306 ConvertUsesToScalar(AI, NewAI, 0); 307 } else { 308 DOUT << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"; 309 310 // Create and insert the integer alloca. 311 const Type *NewTy = Context->getIntegerType(AllocaSize*8); 312 NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 313 ConvertUsesToScalar(AI, NewAI, 0); 314 } 315 NewAI->takeName(AI); 316 AI->eraseFromParent(); 317 ++NumConverted; 318 Changed = true; 319 continue; 320 } 321 322 // Otherwise, couldn't process this alloca. 323 } 324 325 return Changed; 326} 327 328/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 329/// predicate, do SROA now. 330void SROA::DoScalarReplacement(AllocationInst *AI, 331 std::vector<AllocationInst*> &WorkList) { 332 DOUT << "Found inst to SROA: " << *AI; 333 SmallVector<AllocaInst*, 32> ElementAllocas; 334 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 335 ElementAllocas.reserve(ST->getNumContainedTypes()); 336 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 337 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 338 AI->getAlignment(), 339 AI->getName() + "." + utostr(i), AI); 340 ElementAllocas.push_back(NA); 341 WorkList.push_back(NA); // Add to worklist for recursive processing 342 } 343 } else { 344 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 345 ElementAllocas.reserve(AT->getNumElements()); 346 const Type *ElTy = AT->getElementType(); 347 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 348 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 349 AI->getName() + "." + utostr(i), AI); 350 ElementAllocas.push_back(NA); 351 WorkList.push_back(NA); // Add to worklist for recursive processing 352 } 353 } 354 355 // Now that we have created the alloca instructions that we want to use, 356 // expand the getelementptr instructions to use them. 357 // 358 while (!AI->use_empty()) { 359 Instruction *User = cast<Instruction>(AI->use_back()); 360 if (BitCastInst *BCInst = dyn_cast<BitCastInst>(User)) { 361 RewriteBitCastUserOfAlloca(BCInst, AI, ElementAllocas); 362 BCInst->eraseFromParent(); 363 continue; 364 } 365 366 // Replace: 367 // %res = load { i32, i32 }* %alloc 368 // with: 369 // %load.0 = load i32* %alloc.0 370 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 371 // %load.1 = load i32* %alloc.1 372 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 373 // (Also works for arrays instead of structs) 374 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 375 Value *Insert = Context->getUndef(LI->getType()); 376 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) { 377 Value *Load = new LoadInst(ElementAllocas[i], "load", LI); 378 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 379 } 380 LI->replaceAllUsesWith(Insert); 381 LI->eraseFromParent(); 382 continue; 383 } 384 385 // Replace: 386 // store { i32, i32 } %val, { i32, i32 }* %alloc 387 // with: 388 // %val.0 = extractvalue { i32, i32 } %val, 0 389 // store i32 %val.0, i32* %alloc.0 390 // %val.1 = extractvalue { i32, i32 } %val, 1 391 // store i32 %val.1, i32* %alloc.1 392 // (Also works for arrays instead of structs) 393 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 394 Value *Val = SI->getOperand(0); 395 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) { 396 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 397 new StoreInst(Extract, ElementAllocas[i], SI); 398 } 399 SI->eraseFromParent(); 400 continue; 401 } 402 403 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User); 404 // We now know that the GEP is of the form: GEP <ptr>, 0, <cst> 405 unsigned Idx = 406 (unsigned)cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue(); 407 408 assert(Idx < ElementAllocas.size() && "Index out of range?"); 409 AllocaInst *AllocaToUse = ElementAllocas[Idx]; 410 411 Value *RepValue; 412 if (GEPI->getNumOperands() == 3) { 413 // Do not insert a new getelementptr instruction with zero indices, only 414 // to have it optimized out later. 415 RepValue = AllocaToUse; 416 } else { 417 // We are indexing deeply into the structure, so we still need a 418 // getelement ptr instruction to finish the indexing. This may be 419 // expanded itself once the worklist is rerun. 420 // 421 SmallVector<Value*, 8> NewArgs; 422 NewArgs.push_back(Context->getNullValue(Type::Int32Ty)); 423 NewArgs.append(GEPI->op_begin()+3, GEPI->op_end()); 424 RepValue = GetElementPtrInst::Create(AllocaToUse, NewArgs.begin(), 425 NewArgs.end(), "", GEPI); 426 RepValue->takeName(GEPI); 427 } 428 429 // If this GEP is to the start of the aggregate, check for memcpys. 430 if (Idx == 0 && GEPI->hasAllZeroIndices()) 431 RewriteBitCastUserOfAlloca(GEPI, AI, ElementAllocas); 432 433 // Move all of the users over to the new GEP. 434 GEPI->replaceAllUsesWith(RepValue); 435 // Delete the old GEP 436 GEPI->eraseFromParent(); 437 } 438 439 // Finally, delete the Alloca instruction 440 AI->eraseFromParent(); 441 NumReplaced++; 442} 443 444 445/// isSafeElementUse - Check to see if this use is an allowed use for a 446/// getelementptr instruction of an array aggregate allocation. isFirstElt 447/// indicates whether Ptr is known to the start of the aggregate. 448/// 449void SROA::isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI, 450 AllocaInfo &Info) { 451 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 452 I != E; ++I) { 453 Instruction *User = cast<Instruction>(*I); 454 switch (User->getOpcode()) { 455 case Instruction::Load: break; 456 case Instruction::Store: 457 // Store is ok if storing INTO the pointer, not storing the pointer 458 if (User->getOperand(0) == Ptr) return MarkUnsafe(Info); 459 break; 460 case Instruction::GetElementPtr: { 461 GetElementPtrInst *GEP = cast<GetElementPtrInst>(User); 462 bool AreAllZeroIndices = isFirstElt; 463 if (GEP->getNumOperands() > 1) { 464 if (!isa<ConstantInt>(GEP->getOperand(1)) || 465 !cast<ConstantInt>(GEP->getOperand(1))->isZero()) 466 // Using pointer arithmetic to navigate the array. 467 return MarkUnsafe(Info); 468 469 if (AreAllZeroIndices) 470 AreAllZeroIndices = GEP->hasAllZeroIndices(); 471 } 472 isSafeElementUse(GEP, AreAllZeroIndices, AI, Info); 473 if (Info.isUnsafe) return; 474 break; 475 } 476 case Instruction::BitCast: 477 if (isFirstElt) { 478 isSafeUseOfBitCastedAllocation(cast<BitCastInst>(User), AI, Info); 479 if (Info.isUnsafe) return; 480 break; 481 } 482 DOUT << " Transformation preventing inst: " << *User; 483 return MarkUnsafe(Info); 484 case Instruction::Call: 485 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 486 if (isFirstElt) { 487 isSafeMemIntrinsicOnAllocation(MI, AI, I.getOperandNo(), Info); 488 if (Info.isUnsafe) return; 489 break; 490 } 491 } 492 DOUT << " Transformation preventing inst: " << *User; 493 return MarkUnsafe(Info); 494 default: 495 DOUT << " Transformation preventing inst: " << *User; 496 return MarkUnsafe(Info); 497 } 498 } 499 return; // All users look ok :) 500} 501 502/// AllUsersAreLoads - Return true if all users of this value are loads. 503static bool AllUsersAreLoads(Value *Ptr) { 504 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 505 I != E; ++I) 506 if (cast<Instruction>(*I)->getOpcode() != Instruction::Load) 507 return false; 508 return true; 509} 510 511/// isSafeUseOfAllocation - Check to see if this user is an allowed use for an 512/// aggregate allocation. 513/// 514void SROA::isSafeUseOfAllocation(Instruction *User, AllocationInst *AI, 515 AllocaInfo &Info) { 516 if (BitCastInst *C = dyn_cast<BitCastInst>(User)) 517 return isSafeUseOfBitCastedAllocation(C, AI, Info); 518 519 if (LoadInst *LI = dyn_cast<LoadInst>(User)) 520 if (!LI->isVolatile()) 521 return;// Loads (returning a first class aggregrate) are always rewritable 522 523 if (StoreInst *SI = dyn_cast<StoreInst>(User)) 524 if (!SI->isVolatile() && SI->getOperand(0) != AI) 525 return;// Store is ok if storing INTO the pointer, not storing the pointer 526 527 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User); 528 if (GEPI == 0) 529 return MarkUnsafe(Info); 530 531 gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI); 532 533 // The GEP is not safe to transform if not of the form "GEP <ptr>, 0, <cst>". 534 if (I == E || 535 I.getOperand() != Context->getNullValue(I.getOperand()->getType())) { 536 return MarkUnsafe(Info); 537 } 538 539 ++I; 540 if (I == E) return MarkUnsafe(Info); // ran out of GEP indices?? 541 542 bool IsAllZeroIndices = true; 543 544 // If the first index is a non-constant index into an array, see if we can 545 // handle it as a special case. 546 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 547 if (!isa<ConstantInt>(I.getOperand())) { 548 IsAllZeroIndices = 0; 549 uint64_t NumElements = AT->getNumElements(); 550 551 // If this is an array index and the index is not constant, we cannot 552 // promote... that is unless the array has exactly one or two elements in 553 // it, in which case we CAN promote it, but we have to canonicalize this 554 // out if this is the only problem. 555 if ((NumElements == 1 || NumElements == 2) && 556 AllUsersAreLoads(GEPI)) { 557 Info.needsCleanup = true; 558 return; // Canonicalization required! 559 } 560 return MarkUnsafe(Info); 561 } 562 } 563 564 // Walk through the GEP type indices, checking the types that this indexes 565 // into. 566 for (; I != E; ++I) { 567 // Ignore struct elements, no extra checking needed for these. 568 if (isa<StructType>(*I)) 569 continue; 570 571 ConstantInt *IdxVal = dyn_cast<ConstantInt>(I.getOperand()); 572 if (!IdxVal) return MarkUnsafe(Info); 573 574 // Are all indices still zero? 575 IsAllZeroIndices &= IdxVal->isZero(); 576 577 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 578 // This GEP indexes an array. Verify that this is an in-range constant 579 // integer. Specifically, consider A[0][i]. We cannot know that the user 580 // isn't doing invalid things like allowing i to index an out-of-range 581 // subscript that accesses A[1]. Because of this, we have to reject SROA 582 // of any accesses into structs where any of the components are variables. 583 if (IdxVal->getZExtValue() >= AT->getNumElements()) 584 return MarkUnsafe(Info); 585 } else if (const VectorType *VT = dyn_cast<VectorType>(*I)) { 586 if (IdxVal->getZExtValue() >= VT->getNumElements()) 587 return MarkUnsafe(Info); 588 } 589 } 590 591 // If there are any non-simple uses of this getelementptr, make sure to reject 592 // them. 593 return isSafeElementUse(GEPI, IsAllZeroIndices, AI, Info); 594} 595 596/// isSafeMemIntrinsicOnAllocation - Return true if the specified memory 597/// intrinsic can be promoted by SROA. At this point, we know that the operand 598/// of the memintrinsic is a pointer to the beginning of the allocation. 599void SROA::isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI, 600 unsigned OpNo, AllocaInfo &Info) { 601 // If not constant length, give up. 602 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 603 if (!Length) return MarkUnsafe(Info); 604 605 // If not the whole aggregate, give up. 606 if (Length->getZExtValue() != 607 TD->getTypeAllocSize(AI->getType()->getElementType())) 608 return MarkUnsafe(Info); 609 610 // We only know about memcpy/memset/memmove. 611 if (!isa<MemIntrinsic>(MI)) 612 return MarkUnsafe(Info); 613 614 // Otherwise, we can transform it. Determine whether this is a memcpy/set 615 // into or out of the aggregate. 616 if (OpNo == 1) 617 Info.isMemCpyDst = true; 618 else { 619 assert(OpNo == 2); 620 Info.isMemCpySrc = true; 621 } 622} 623 624/// isSafeUseOfBitCastedAllocation - Return true if all users of this bitcast 625/// are 626void SROA::isSafeUseOfBitCastedAllocation(BitCastInst *BC, AllocationInst *AI, 627 AllocaInfo &Info) { 628 for (Value::use_iterator UI = BC->use_begin(), E = BC->use_end(); 629 UI != E; ++UI) { 630 if (BitCastInst *BCU = dyn_cast<BitCastInst>(UI)) { 631 isSafeUseOfBitCastedAllocation(BCU, AI, Info); 632 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) { 633 isSafeMemIntrinsicOnAllocation(MI, AI, UI.getOperandNo(), Info); 634 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI)) { 635 if (SI->isVolatile()) 636 return MarkUnsafe(Info); 637 638 // If storing the entire alloca in one chunk through a bitcasted pointer 639 // to integer, we can transform it. This happens (for example) when you 640 // cast a {i32,i32}* to i64* and store through it. This is similar to the 641 // memcpy case and occurs in various "byval" cases and emulated memcpys. 642 if (isa<IntegerType>(SI->getOperand(0)->getType()) && 643 TD->getTypeAllocSize(SI->getOperand(0)->getType()) == 644 TD->getTypeAllocSize(AI->getType()->getElementType())) { 645 Info.isMemCpyDst = true; 646 continue; 647 } 648 return MarkUnsafe(Info); 649 } else if (LoadInst *LI = dyn_cast<LoadInst>(UI)) { 650 if (LI->isVolatile()) 651 return MarkUnsafe(Info); 652 653 // If loading the entire alloca in one chunk through a bitcasted pointer 654 // to integer, we can transform it. This happens (for example) when you 655 // cast a {i32,i32}* to i64* and load through it. This is similar to the 656 // memcpy case and occurs in various "byval" cases and emulated memcpys. 657 if (isa<IntegerType>(LI->getType()) && 658 TD->getTypeAllocSize(LI->getType()) == 659 TD->getTypeAllocSize(AI->getType()->getElementType())) { 660 Info.isMemCpySrc = true; 661 continue; 662 } 663 return MarkUnsafe(Info); 664 } else if (isa<DbgInfoIntrinsic>(UI)) { 665 // If one user is DbgInfoIntrinsic then check if all users are 666 // DbgInfoIntrinsics. 667 if (OnlyUsedByDbgInfoIntrinsics(BC)) { 668 Info.needsCleanup = true; 669 return; 670 } 671 else 672 MarkUnsafe(Info); 673 } 674 else { 675 return MarkUnsafe(Info); 676 } 677 if (Info.isUnsafe) return; 678 } 679} 680 681/// RewriteBitCastUserOfAlloca - BCInst (transitively) bitcasts AI, or indexes 682/// to its first element. Transform users of the cast to use the new values 683/// instead. 684void SROA::RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI, 685 SmallVector<AllocaInst*, 32> &NewElts) { 686 Value::use_iterator UI = BCInst->use_begin(), UE = BCInst->use_end(); 687 while (UI != UE) { 688 Instruction *User = cast<Instruction>(*UI++); 689 if (BitCastInst *BCU = dyn_cast<BitCastInst>(User)) { 690 RewriteBitCastUserOfAlloca(BCU, AI, NewElts); 691 if (BCU->use_empty()) BCU->eraseFromParent(); 692 continue; 693 } 694 695 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 696 // This must be memcpy/memmove/memset of the entire aggregate. 697 // Split into one per element. 698 RewriteMemIntrinUserOfAlloca(MI, BCInst, AI, NewElts); 699 continue; 700 } 701 702 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 703 // If this is a store of the entire alloca from an integer, rewrite it. 704 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 705 continue; 706 } 707 708 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 709 // If this is a load of the entire alloca to an integer, rewrite it. 710 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 711 continue; 712 } 713 714 // Otherwise it must be some other user of a gep of the first pointer. Just 715 // leave these alone. 716 continue; 717 } 718} 719 720/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 721/// Rewrite it to copy or set the elements of the scalarized memory. 722void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst, 723 AllocationInst *AI, 724 SmallVector<AllocaInst*, 32> &NewElts) { 725 726 // If this is a memcpy/memmove, construct the other pointer as the 727 // appropriate type. The "Other" pointer is the pointer that goes to memory 728 // that doesn't have anything to do with the alloca that we are promoting. For 729 // memset, this Value* stays null. 730 Value *OtherPtr = 0; 731 unsigned MemAlignment = MI->getAlignment(); 732 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 733 if (BCInst == MTI->getRawDest()) 734 OtherPtr = MTI->getRawSource(); 735 else { 736 assert(BCInst == MTI->getRawSource()); 737 OtherPtr = MTI->getRawDest(); 738 } 739 } 740 741 // If there is an other pointer, we want to convert it to the same pointer 742 // type as AI has, so we can GEP through it safely. 743 if (OtherPtr) { 744 // It is likely that OtherPtr is a bitcast, if so, remove it. 745 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr)) 746 OtherPtr = BC->getOperand(0); 747 // All zero GEPs are effectively bitcasts. 748 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr)) 749 if (GEP->hasAllZeroIndices()) 750 OtherPtr = GEP->getOperand(0); 751 752 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr)) 753 if (BCE->getOpcode() == Instruction::BitCast) 754 OtherPtr = BCE->getOperand(0); 755 756 // If the pointer is not the right type, insert a bitcast to the right 757 // type. 758 if (OtherPtr->getType() != AI->getType()) 759 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(), 760 MI); 761 } 762 763 // Process each element of the aggregate. 764 Value *TheFn = MI->getOperand(0); 765 const Type *BytePtrTy = MI->getRawDest()->getType(); 766 bool SROADest = MI->getRawDest() == BCInst; 767 768 Constant *Zero = Context->getNullValue(Type::Int32Ty); 769 770 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 771 // If this is a memcpy/memmove, emit a GEP of the other element address. 772 Value *OtherElt = 0; 773 unsigned OtherEltAlign = MemAlignment; 774 775 if (OtherPtr) { 776 Value *Idx[2] = { Zero, Context->getConstantInt(Type::Int32Ty, i) }; 777 OtherElt = GetElementPtrInst::Create(OtherPtr, Idx, Idx + 2, 778 OtherPtr->getNameStr()+"."+utostr(i), 779 MI); 780 uint64_t EltOffset; 781 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 782 if (const StructType *ST = 783 dyn_cast<StructType>(OtherPtrTy->getElementType())) { 784 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 785 } else { 786 const Type *EltTy = 787 cast<SequentialType>(OtherPtr->getType())->getElementType(); 788 EltOffset = TD->getTypeAllocSize(EltTy)*i; 789 } 790 791 // The alignment of the other pointer is the guaranteed alignment of the 792 // element, which is affected by both the known alignment of the whole 793 // mem intrinsic and the alignment of the element. If the alignment of 794 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 795 // known alignment is just 4 bytes. 796 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 797 } 798 799 Value *EltPtr = NewElts[i]; 800 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 801 802 // If we got down to a scalar, insert a load or store as appropriate. 803 if (EltTy->isSingleValueType()) { 804 if (isa<MemTransferInst>(MI)) { 805 if (SROADest) { 806 // From Other to Alloca. 807 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 808 new StoreInst(Elt, EltPtr, MI); 809 } else { 810 // From Alloca to Other. 811 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 812 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 813 } 814 continue; 815 } 816 assert(isa<MemSetInst>(MI)); 817 818 // If the stored element is zero (common case), just store a null 819 // constant. 820 Constant *StoreVal; 821 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) { 822 if (CI->isZero()) { 823 StoreVal = Context->getNullValue(EltTy); // 0.0, null, 0, <0,0> 824 } else { 825 // If EltTy is a vector type, get the element type. 826 const Type *ValTy = EltTy->getScalarType(); 827 828 // Construct an integer with the right value. 829 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 830 APInt OneVal(EltSize, CI->getZExtValue()); 831 APInt TotalVal(OneVal); 832 // Set each byte. 833 for (unsigned i = 0; 8*i < EltSize; ++i) { 834 TotalVal = TotalVal.shl(8); 835 TotalVal |= OneVal; 836 } 837 838 // Convert the integer value to the appropriate type. 839 StoreVal = Context->getConstantInt(TotalVal); 840 if (isa<PointerType>(ValTy)) 841 StoreVal = Context->getConstantExprIntToPtr(StoreVal, ValTy); 842 else if (ValTy->isFloatingPoint()) 843 StoreVal = Context->getConstantExprBitCast(StoreVal, ValTy); 844 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 845 846 // If the requested value was a vector constant, create it. 847 if (EltTy != ValTy) { 848 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 849 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 850 StoreVal = Context->getConstantVector(&Elts[0], NumElts); 851 } 852 } 853 new StoreInst(StoreVal, EltPtr, MI); 854 continue; 855 } 856 // Otherwise, if we're storing a byte variable, use a memset call for 857 // this element. 858 } 859 860 // Cast the element pointer to BytePtrTy. 861 if (EltPtr->getType() != BytePtrTy) 862 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI); 863 864 // Cast the other pointer (if we have one) to BytePtrTy. 865 if (OtherElt && OtherElt->getType() != BytePtrTy) 866 OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(), 867 MI); 868 869 unsigned EltSize = TD->getTypeAllocSize(EltTy); 870 871 // Finally, insert the meminst for this element. 872 if (isa<MemTransferInst>(MI)) { 873 Value *Ops[] = { 874 SROADest ? EltPtr : OtherElt, // Dest ptr 875 SROADest ? OtherElt : EltPtr, // Src ptr 876 Context->getConstantInt(MI->getOperand(3)->getType(), EltSize), // Size 877 Context->getConstantInt(Type::Int32Ty, OtherEltAlign) // Align 878 }; 879 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 880 } else { 881 assert(isa<MemSetInst>(MI)); 882 Value *Ops[] = { 883 EltPtr, MI->getOperand(2), // Dest, Value, 884 Context->getConstantInt(MI->getOperand(3)->getType(), EltSize), // Size 885 Zero // Align 886 }; 887 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 888 } 889 } 890 MI->eraseFromParent(); 891} 892 893/// RewriteStoreUserOfWholeAlloca - We found an store of an integer that 894/// overwrites the entire allocation. Extract out the pieces of the stored 895/// integer and store them individually. 896void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, 897 AllocationInst *AI, 898 SmallVector<AllocaInst*, 32> &NewElts){ 899 // Extract each element out of the integer according to its structure offset 900 // and store the element value to the individual alloca. 901 Value *SrcVal = SI->getOperand(0); 902 const Type *AllocaEltTy = AI->getType()->getElementType(); 903 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 904 905 // If this isn't a store of an integer to the whole alloca, it may be a store 906 // to the first element. Just ignore the store in this case and normal SROA 907 // will handle it. 908 if (!isa<IntegerType>(SrcVal->getType()) || 909 TD->getTypeAllocSizeInBits(SrcVal->getType()) != AllocaSizeBits) 910 return; 911 // Handle tail padding by extending the operand 912 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 913 SrcVal = new ZExtInst(SrcVal, 914 Context->getIntegerType(AllocaSizeBits), "", SI); 915 916 DOUT << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << *SI; 917 918 // There are two forms here: AI could be an array or struct. Both cases 919 // have different ways to compute the element offset. 920 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 921 const StructLayout *Layout = TD->getStructLayout(EltSTy); 922 923 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 924 // Get the number of bits to shift SrcVal to get the value. 925 const Type *FieldTy = EltSTy->getElementType(i); 926 uint64_t Shift = Layout->getElementOffsetInBits(i); 927 928 if (TD->isBigEndian()) 929 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 930 931 Value *EltVal = SrcVal; 932 if (Shift) { 933 Value *ShiftVal = Context->getConstantInt(EltVal->getType(), Shift); 934 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 935 "sroa.store.elt", SI); 936 } 937 938 // Truncate down to an integer of the right size. 939 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 940 941 // Ignore zero sized fields like {}, they obviously contain no data. 942 if (FieldSizeBits == 0) continue; 943 944 if (FieldSizeBits != AllocaSizeBits) 945 EltVal = new TruncInst(EltVal, 946 Context->getIntegerType(FieldSizeBits), "", SI); 947 Value *DestField = NewElts[i]; 948 if (EltVal->getType() == FieldTy) { 949 // Storing to an integer field of this size, just do it. 950 } else if (FieldTy->isFloatingPoint() || isa<VectorType>(FieldTy)) { 951 // Bitcast to the right element type (for fp/vector values). 952 EltVal = new BitCastInst(EltVal, FieldTy, "", SI); 953 } else { 954 // Otherwise, bitcast the dest pointer (for aggregates). 955 DestField = new BitCastInst(DestField, 956 Context->getPointerTypeUnqual(EltVal->getType()), 957 "", SI); 958 } 959 new StoreInst(EltVal, DestField, SI); 960 } 961 962 } else { 963 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 964 const Type *ArrayEltTy = ATy->getElementType(); 965 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 966 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 967 968 uint64_t Shift; 969 970 if (TD->isBigEndian()) 971 Shift = AllocaSizeBits-ElementOffset; 972 else 973 Shift = 0; 974 975 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 976 // Ignore zero sized fields like {}, they obviously contain no data. 977 if (ElementSizeBits == 0) continue; 978 979 Value *EltVal = SrcVal; 980 if (Shift) { 981 Value *ShiftVal = Context->getConstantInt(EltVal->getType(), Shift); 982 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 983 "sroa.store.elt", SI); 984 } 985 986 // Truncate down to an integer of the right size. 987 if (ElementSizeBits != AllocaSizeBits) 988 EltVal = new TruncInst(EltVal, 989 Context->getIntegerType(ElementSizeBits),"",SI); 990 Value *DestField = NewElts[i]; 991 if (EltVal->getType() == ArrayEltTy) { 992 // Storing to an integer field of this size, just do it. 993 } else if (ArrayEltTy->isFloatingPoint() || isa<VectorType>(ArrayEltTy)) { 994 // Bitcast to the right element type (for fp/vector values). 995 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI); 996 } else { 997 // Otherwise, bitcast the dest pointer (for aggregates). 998 DestField = new BitCastInst(DestField, 999 Context->getPointerTypeUnqual(EltVal->getType()), 1000 "", SI); 1001 } 1002 new StoreInst(EltVal, DestField, SI); 1003 1004 if (TD->isBigEndian()) 1005 Shift -= ElementOffset; 1006 else 1007 Shift += ElementOffset; 1008 } 1009 } 1010 1011 SI->eraseFromParent(); 1012} 1013 1014/// RewriteLoadUserOfWholeAlloca - We found an load of the entire allocation to 1015/// an integer. Load the individual pieces to form the aggregate value. 1016void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI, 1017 SmallVector<AllocaInst*, 32> &NewElts) { 1018 // Extract each element out of the NewElts according to its structure offset 1019 // and form the result value. 1020 const Type *AllocaEltTy = AI->getType()->getElementType(); 1021 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1022 1023 // If this isn't a load of the whole alloca to an integer, it may be a load 1024 // of the first element. Just ignore the load in this case and normal SROA 1025 // will handle it. 1026 if (!isa<IntegerType>(LI->getType()) || 1027 TD->getTypeAllocSizeInBits(LI->getType()) != AllocaSizeBits) 1028 return; 1029 1030 DOUT << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << *LI; 1031 1032 // There are two forms here: AI could be an array or struct. Both cases 1033 // have different ways to compute the element offset. 1034 const StructLayout *Layout = 0; 1035 uint64_t ArrayEltBitOffset = 0; 1036 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1037 Layout = TD->getStructLayout(EltSTy); 1038 } else { 1039 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 1040 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1041 } 1042 1043 Value *ResultVal = 1044 Context->getNullValue(Context->getIntegerType(AllocaSizeBits)); 1045 1046 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1047 // Load the value from the alloca. If the NewElt is an aggregate, cast 1048 // the pointer to an integer of the same size before doing the load. 1049 Value *SrcField = NewElts[i]; 1050 const Type *FieldTy = 1051 cast<PointerType>(SrcField->getType())->getElementType(); 1052 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1053 1054 // Ignore zero sized fields like {}, they obviously contain no data. 1055 if (FieldSizeBits == 0) continue; 1056 1057 const IntegerType *FieldIntTy = Context->getIntegerType(FieldSizeBits); 1058 if (!isa<IntegerType>(FieldTy) && !FieldTy->isFloatingPoint() && 1059 !isa<VectorType>(FieldTy)) 1060 SrcField = new BitCastInst(SrcField, 1061 Context->getPointerTypeUnqual(FieldIntTy), 1062 "", LI); 1063 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 1064 1065 // If SrcField is a fp or vector of the right size but that isn't an 1066 // integer type, bitcast to an integer so we can shift it. 1067 if (SrcField->getType() != FieldIntTy) 1068 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 1069 1070 // Zero extend the field to be the same size as the final alloca so that 1071 // we can shift and insert it. 1072 if (SrcField->getType() != ResultVal->getType()) 1073 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 1074 1075 // Determine the number of bits to shift SrcField. 1076 uint64_t Shift; 1077 if (Layout) // Struct case. 1078 Shift = Layout->getElementOffsetInBits(i); 1079 else // Array case. 1080 Shift = i*ArrayEltBitOffset; 1081 1082 if (TD->isBigEndian()) 1083 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 1084 1085 if (Shift) { 1086 Value *ShiftVal = Context->getConstantInt(SrcField->getType(), Shift); 1087 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 1088 } 1089 1090 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 1091 } 1092 1093 // Handle tail padding by truncating the result 1094 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 1095 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 1096 1097 LI->replaceAllUsesWith(ResultVal); 1098 LI->eraseFromParent(); 1099} 1100 1101 1102/// HasPadding - Return true if the specified type has any structure or 1103/// alignment padding, false otherwise. 1104static bool HasPadding(const Type *Ty, const TargetData &TD) { 1105 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 1106 const StructLayout *SL = TD.getStructLayout(STy); 1107 unsigned PrevFieldBitOffset = 0; 1108 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1109 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 1110 1111 // Padding in sub-elements? 1112 if (HasPadding(STy->getElementType(i), TD)) 1113 return true; 1114 1115 // Check to see if there is any padding between this element and the 1116 // previous one. 1117 if (i) { 1118 unsigned PrevFieldEnd = 1119 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 1120 if (PrevFieldEnd < FieldBitOffset) 1121 return true; 1122 } 1123 1124 PrevFieldBitOffset = FieldBitOffset; 1125 } 1126 1127 // Check for tail padding. 1128 if (unsigned EltCount = STy->getNumElements()) { 1129 unsigned PrevFieldEnd = PrevFieldBitOffset + 1130 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 1131 if (PrevFieldEnd < SL->getSizeInBits()) 1132 return true; 1133 } 1134 1135 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1136 return HasPadding(ATy->getElementType(), TD); 1137 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) { 1138 return HasPadding(VTy->getElementType(), TD); 1139 } 1140 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 1141} 1142 1143/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 1144/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 1145/// or 1 if safe after canonicalization has been performed. 1146/// 1147int SROA::isSafeAllocaToScalarRepl(AllocationInst *AI) { 1148 // Loop over the use list of the alloca. We can only transform it if all of 1149 // the users are safe to transform. 1150 AllocaInfo Info; 1151 1152 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); 1153 I != E; ++I) { 1154 isSafeUseOfAllocation(cast<Instruction>(*I), AI, Info); 1155 if (Info.isUnsafe) { 1156 DOUT << "Cannot transform: " << *AI << " due to user: " << **I; 1157 return 0; 1158 } 1159 } 1160 1161 // Okay, we know all the users are promotable. If the aggregate is a memcpy 1162 // source and destination, we have to be careful. In particular, the memcpy 1163 // could be moving around elements that live in structure padding of the LLVM 1164 // types, but may actually be used. In these cases, we refuse to promote the 1165 // struct. 1166 if (Info.isMemCpySrc && Info.isMemCpyDst && 1167 HasPadding(AI->getType()->getElementType(), *TD)) 1168 return 0; 1169 1170 // If we require cleanup, return 1, otherwise return 3. 1171 return Info.needsCleanup ? 1 : 3; 1172} 1173 1174/// CleanupGEP - GEP is used by an Alloca, which can be prompted after the GEP 1175/// is canonicalized here. 1176void SROA::CleanupGEP(GetElementPtrInst *GEPI) { 1177 gep_type_iterator I = gep_type_begin(GEPI); 1178 ++I; 1179 1180 const ArrayType *AT = dyn_cast<ArrayType>(*I); 1181 if (!AT) 1182 return; 1183 1184 uint64_t NumElements = AT->getNumElements(); 1185 1186 if (isa<ConstantInt>(I.getOperand())) 1187 return; 1188 1189 if (NumElements == 1) { 1190 GEPI->setOperand(2, Context->getNullValue(Type::Int32Ty)); 1191 return; 1192 } 1193 1194 assert(NumElements == 2 && "Unhandled case!"); 1195 // All users of the GEP must be loads. At each use of the GEP, insert 1196 // two loads of the appropriate indexed GEP and select between them. 1197 Value *IsOne = new ICmpInst(GEPI, ICmpInst::ICMP_NE, I.getOperand(), 1198 Context->getNullValue(I.getOperand()->getType()), 1199 "isone"); 1200 // Insert the new GEP instructions, which are properly indexed. 1201 SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end()); 1202 Indices[1] = Context->getNullValue(Type::Int32Ty); 1203 Value *ZeroIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 1204 Indices.begin(), 1205 Indices.end(), 1206 GEPI->getName()+".0", GEPI); 1207 Indices[1] = Context->getConstantInt(Type::Int32Ty, 1); 1208 Value *OneIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 1209 Indices.begin(), 1210 Indices.end(), 1211 GEPI->getName()+".1", GEPI); 1212 // Replace all loads of the variable index GEP with loads from both 1213 // indexes and a select. 1214 while (!GEPI->use_empty()) { 1215 LoadInst *LI = cast<LoadInst>(GEPI->use_back()); 1216 Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI); 1217 Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI); 1218 Value *R = SelectInst::Create(IsOne, One, Zero, LI->getName(), LI); 1219 LI->replaceAllUsesWith(R); 1220 LI->eraseFromParent(); 1221 } 1222 GEPI->eraseFromParent(); 1223} 1224 1225 1226/// CleanupAllocaUsers - If SROA reported that it can promote the specified 1227/// allocation, but only if cleaned up, perform the cleanups required. 1228void SROA::CleanupAllocaUsers(AllocationInst *AI) { 1229 // At this point, we know that the end result will be SROA'd and promoted, so 1230 // we can insert ugly code if required so long as sroa+mem2reg will clean it 1231 // up. 1232 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 1233 UI != E; ) { 1234 User *U = *UI++; 1235 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) 1236 CleanupGEP(GEPI); 1237 else { 1238 Instruction *I = cast<Instruction>(U); 1239 SmallVector<DbgInfoIntrinsic *, 2> DbgInUses; 1240 if (!isa<StoreInst>(I) && OnlyUsedByDbgInfoIntrinsics(I, &DbgInUses)) { 1241 // Safe to remove debug info uses. 1242 while (!DbgInUses.empty()) { 1243 DbgInfoIntrinsic *DI = DbgInUses.back(); DbgInUses.pop_back(); 1244 DI->eraseFromParent(); 1245 } 1246 I->eraseFromParent(); 1247 } 1248 } 1249 } 1250} 1251 1252/// MergeInType - Add the 'In' type to the accumulated type (Accum) so far at 1253/// the offset specified by Offset (which is specified in bytes). 1254/// 1255/// There are two cases we handle here: 1256/// 1) A union of vector types of the same size and potentially its elements. 1257/// Here we turn element accesses into insert/extract element operations. 1258/// This promotes a <4 x float> with a store of float to the third element 1259/// into a <4 x float> that uses insert element. 1260/// 2) A fully general blob of memory, which we turn into some (potentially 1261/// large) integer type with extract and insert operations where the loads 1262/// and stores would mutate the memory. 1263static void MergeInType(const Type *In, uint64_t Offset, const Type *&VecTy, 1264 unsigned AllocaSize, const TargetData &TD, 1265 LLVMContext *Context) { 1266 // If this could be contributing to a vector, analyze it. 1267 if (VecTy != Type::VoidTy) { // either null or a vector type. 1268 1269 // If the In type is a vector that is the same size as the alloca, see if it 1270 // matches the existing VecTy. 1271 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 1272 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 1273 // If we're storing/loading a vector of the right size, allow it as a 1274 // vector. If this the first vector we see, remember the type so that 1275 // we know the element size. 1276 if (VecTy == 0) 1277 VecTy = VInTy; 1278 return; 1279 } 1280 } else if (In == Type::FloatTy || In == Type::DoubleTy || 1281 (isa<IntegerType>(In) && In->getPrimitiveSizeInBits() >= 8 && 1282 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 1283 // If we're accessing something that could be an element of a vector, see 1284 // if the implied vector agrees with what we already have and if Offset is 1285 // compatible with it. 1286 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 1287 if (Offset % EltSize == 0 && 1288 AllocaSize % EltSize == 0 && 1289 (VecTy == 0 || 1290 cast<VectorType>(VecTy)->getElementType() 1291 ->getPrimitiveSizeInBits()/8 == EltSize)) { 1292 if (VecTy == 0) 1293 VecTy = Context->getVectorType(In, AllocaSize/EltSize); 1294 return; 1295 } 1296 } 1297 } 1298 1299 // Otherwise, we have a case that we can't handle with an optimized vector 1300 // form. We can still turn this into a large integer. 1301 VecTy = Type::VoidTy; 1302} 1303 1304/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 1305/// its accesses to use a to single vector type, return true, and set VecTy to 1306/// the new type. If we could convert the alloca into a single promotable 1307/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 1308/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 1309/// is the current offset from the base of the alloca being analyzed. 1310/// 1311/// If we see at least one access to the value that is as a vector type, set the 1312/// SawVec flag. 1313/// 1314bool SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy, 1315 bool &SawVec, uint64_t Offset, 1316 unsigned AllocaSize) { 1317 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1318 Instruction *User = cast<Instruction>(*UI); 1319 1320 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1321 // Don't break volatile loads. 1322 if (LI->isVolatile()) 1323 return false; 1324 MergeInType(LI->getType(), Offset, VecTy, AllocaSize, *TD, Context); 1325 SawVec |= isa<VectorType>(LI->getType()); 1326 continue; 1327 } 1328 1329 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1330 // Storing the pointer, not into the value? 1331 if (SI->getOperand(0) == V || SI->isVolatile()) return 0; 1332 MergeInType(SI->getOperand(0)->getType(), Offset, 1333 VecTy, AllocaSize, *TD, Context); 1334 SawVec |= isa<VectorType>(SI->getOperand(0)->getType()); 1335 continue; 1336 } 1337 1338 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 1339 if (!CanConvertToScalar(BCI, IsNotTrivial, VecTy, SawVec, Offset, 1340 AllocaSize)) 1341 return false; 1342 IsNotTrivial = true; 1343 continue; 1344 } 1345 1346 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1347 // If this is a GEP with a variable indices, we can't handle it. 1348 if (!GEP->hasAllConstantIndices()) 1349 return false; 1350 1351 // Compute the offset that this GEP adds to the pointer. 1352 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1353 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(), 1354 &Indices[0], Indices.size()); 1355 // See if all uses can be converted. 1356 if (!CanConvertToScalar(GEP, IsNotTrivial, VecTy, SawVec,Offset+GEPOffset, 1357 AllocaSize)) 1358 return false; 1359 IsNotTrivial = true; 1360 continue; 1361 } 1362 1363 // If this is a constant sized memset of a constant value (e.g. 0) we can 1364 // handle it. 1365 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 1366 // Store of constant value and constant size. 1367 if (isa<ConstantInt>(MSI->getValue()) && 1368 isa<ConstantInt>(MSI->getLength())) { 1369 IsNotTrivial = true; 1370 continue; 1371 } 1372 } 1373 1374 // If this is a memcpy or memmove into or out of the whole allocation, we 1375 // can handle it like a load or store of the scalar type. 1376 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 1377 if (ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength())) 1378 if (Len->getZExtValue() == AllocaSize && Offset == 0) { 1379 IsNotTrivial = true; 1380 continue; 1381 } 1382 } 1383 1384 // Ignore dbg intrinsic. 1385 if (isa<DbgInfoIntrinsic>(User)) 1386 continue; 1387 1388 // Otherwise, we cannot handle this! 1389 return false; 1390 } 1391 1392 return true; 1393} 1394 1395 1396/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 1397/// directly. This happens when we are converting an "integer union" to a 1398/// single integer scalar, or when we are converting a "vector union" to a 1399/// vector with insert/extractelement instructions. 1400/// 1401/// Offset is an offset from the original alloca, in bits that need to be 1402/// shifted to the right. By the end of this, there should be no uses of Ptr. 1403void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset) { 1404 while (!Ptr->use_empty()) { 1405 Instruction *User = cast<Instruction>(Ptr->use_back()); 1406 1407 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 1408 ConvertUsesToScalar(CI, NewAI, Offset); 1409 CI->eraseFromParent(); 1410 continue; 1411 } 1412 1413 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1414 // Compute the offset that this GEP adds to the pointer. 1415 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1416 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(), 1417 &Indices[0], Indices.size()); 1418 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 1419 GEP->eraseFromParent(); 1420 continue; 1421 } 1422 1423 IRBuilder<> Builder(User->getParent(), User); 1424 1425 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1426 // The load is a bit extract from NewAI shifted right by Offset bits. 1427 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 1428 Value *NewLoadVal 1429 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 1430 LI->replaceAllUsesWith(NewLoadVal); 1431 LI->eraseFromParent(); 1432 continue; 1433 } 1434 1435 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1436 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 1437 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").c_str()); 1438 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 1439 Builder); 1440 Builder.CreateStore(New, NewAI); 1441 SI->eraseFromParent(); 1442 continue; 1443 } 1444 1445 // If this is a constant sized memset of a constant value (e.g. 0) we can 1446 // transform it into a store of the expanded constant value. 1447 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 1448 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 1449 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 1450 if (NumBytes != 0) { 1451 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 1452 1453 // Compute the value replicated the right number of times. 1454 APInt APVal(NumBytes*8, Val); 1455 1456 // Splat the value if non-zero. 1457 if (Val) 1458 for (unsigned i = 1; i != NumBytes; ++i) 1459 APVal |= APVal << 8; 1460 1461 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").c_str()); 1462 Value *New = ConvertScalar_InsertValue(Context->getConstantInt(APVal), 1463 Old, Offset, Builder); 1464 Builder.CreateStore(New, NewAI); 1465 } 1466 MSI->eraseFromParent(); 1467 continue; 1468 } 1469 1470 // If this is a memcpy or memmove into or out of the whole allocation, we 1471 // can handle it like a load or store of the scalar type. 1472 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 1473 assert(Offset == 0 && "must be store to start of alloca"); 1474 1475 // If the source and destination are both to the same alloca, then this is 1476 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 1477 // as appropriate. 1478 AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject()); 1479 1480 if (MTI->getSource()->getUnderlyingObject() != OrigAI) { 1481 // Dest must be OrigAI, change this to be a load from the original 1482 // pointer (bitcasted), then a store to our new alloca. 1483 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 1484 Value *SrcPtr = MTI->getSource(); 1485 SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType()); 1486 1487 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 1488 SrcVal->setAlignment(MTI->getAlignment()); 1489 Builder.CreateStore(SrcVal, NewAI); 1490 } else if (MTI->getDest()->getUnderlyingObject() != OrigAI) { 1491 // Src must be OrigAI, change this to be a load from NewAI then a store 1492 // through the original dest pointer (bitcasted). 1493 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 1494 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 1495 1496 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType()); 1497 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 1498 NewStore->setAlignment(MTI->getAlignment()); 1499 } else { 1500 // Noop transfer. Src == Dst 1501 } 1502 1503 1504 MTI->eraseFromParent(); 1505 continue; 1506 } 1507 1508 // If user is a dbg info intrinsic then it is safe to remove it. 1509 if (isa<DbgInfoIntrinsic>(User)) { 1510 User->eraseFromParent(); 1511 continue; 1512 } 1513 1514 llvm_unreachable("Unsupported operation!"); 1515 } 1516} 1517 1518/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 1519/// or vector value FromVal, extracting the bits from the offset specified by 1520/// Offset. This returns the value, which is of type ToType. 1521/// 1522/// This happens when we are converting an "integer union" to a single 1523/// integer scalar, or when we are converting a "vector union" to a vector with 1524/// insert/extractelement instructions. 1525/// 1526/// Offset is an offset from the original alloca, in bits that need to be 1527/// shifted to the right. 1528Value *SROA::ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 1529 uint64_t Offset, IRBuilder<> &Builder) { 1530 // If the load is of the whole new alloca, no conversion is needed. 1531 if (FromVal->getType() == ToType && Offset == 0) 1532 return FromVal; 1533 1534 // If the result alloca is a vector type, this is either an element 1535 // access or a bitcast to another vector type of the same size. 1536 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) { 1537 if (isa<VectorType>(ToType)) 1538 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 1539 1540 // Otherwise it must be an element access. 1541 unsigned Elt = 0; 1542 if (Offset) { 1543 unsigned EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 1544 Elt = Offset/EltSize; 1545 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 1546 } 1547 // Return the element extracted out of it. 1548 Value *V = Builder.CreateExtractElement(FromVal, 1549 Context->getConstantInt(Type::Int32Ty,Elt), 1550 "tmp"); 1551 if (V->getType() != ToType) 1552 V = Builder.CreateBitCast(V, ToType, "tmp"); 1553 return V; 1554 } 1555 1556 // If ToType is a first class aggregate, extract out each of the pieces and 1557 // use insertvalue's to form the FCA. 1558 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 1559 const StructLayout &Layout = *TD->getStructLayout(ST); 1560 Value *Res = Context->getUndef(ST); 1561 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1562 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 1563 Offset+Layout.getElementOffsetInBits(i), 1564 Builder); 1565 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1566 } 1567 return Res; 1568 } 1569 1570 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 1571 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 1572 Value *Res = Context->getUndef(AT); 1573 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1574 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 1575 Offset+i*EltSize, Builder); 1576 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1577 } 1578 return Res; 1579 } 1580 1581 // Otherwise, this must be a union that was converted to an integer value. 1582 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 1583 1584 // If this is a big-endian system and the load is narrower than the 1585 // full alloca type, we need to do a shift to get the right bits. 1586 int ShAmt = 0; 1587 if (TD->isBigEndian()) { 1588 // On big-endian machines, the lowest bit is stored at the bit offset 1589 // from the pointer given by getTypeStoreSizeInBits. This matters for 1590 // integers with a bitwidth that is not a multiple of 8. 1591 ShAmt = TD->getTypeStoreSizeInBits(NTy) - 1592 TD->getTypeStoreSizeInBits(ToType) - Offset; 1593 } else { 1594 ShAmt = Offset; 1595 } 1596 1597 // Note: we support negative bitwidths (with shl) which are not defined. 1598 // We do this to support (f.e.) loads off the end of a structure where 1599 // only some bits are used. 1600 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 1601 FromVal = Builder.CreateLShr(FromVal, 1602 Context->getConstantInt(FromVal->getType(), 1603 ShAmt), "tmp"); 1604 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 1605 FromVal = Builder.CreateShl(FromVal, 1606 Context->getConstantInt(FromVal->getType(), 1607 -ShAmt), "tmp"); 1608 1609 // Finally, unconditionally truncate the integer to the right width. 1610 unsigned LIBitWidth = TD->getTypeSizeInBits(ToType); 1611 if (LIBitWidth < NTy->getBitWidth()) 1612 FromVal = 1613 Builder.CreateTrunc(FromVal, Context->getIntegerType(LIBitWidth), "tmp"); 1614 else if (LIBitWidth > NTy->getBitWidth()) 1615 FromVal = 1616 Builder.CreateZExt(FromVal, Context->getIntegerType(LIBitWidth), "tmp"); 1617 1618 // If the result is an integer, this is a trunc or bitcast. 1619 if (isa<IntegerType>(ToType)) { 1620 // Should be done. 1621 } else if (ToType->isFloatingPoint() || isa<VectorType>(ToType)) { 1622 // Just do a bitcast, we know the sizes match up. 1623 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 1624 } else { 1625 // Otherwise must be a pointer. 1626 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 1627 } 1628 assert(FromVal->getType() == ToType && "Didn't convert right?"); 1629 return FromVal; 1630} 1631 1632 1633/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 1634/// or vector value "Old" at the offset specified by Offset. 1635/// 1636/// This happens when we are converting an "integer union" to a 1637/// single integer scalar, or when we are converting a "vector union" to a 1638/// vector with insert/extractelement instructions. 1639/// 1640/// Offset is an offset from the original alloca, in bits that need to be 1641/// shifted to the right. 1642Value *SROA::ConvertScalar_InsertValue(Value *SV, Value *Old, 1643 uint64_t Offset, IRBuilder<> &Builder) { 1644 1645 // Convert the stored type to the actual type, shift it left to insert 1646 // then 'or' into place. 1647 const Type *AllocaType = Old->getType(); 1648 1649 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 1650 uint64_t VecSize = TD->getTypeAllocSizeInBits(VTy); 1651 uint64_t ValSize = TD->getTypeAllocSizeInBits(SV->getType()); 1652 1653 // Changing the whole vector with memset or with an access of a different 1654 // vector type? 1655 if (ValSize == VecSize) 1656 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 1657 1658 uint64_t EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 1659 1660 // Must be an element insertion. 1661 unsigned Elt = Offset/EltSize; 1662 1663 if (SV->getType() != VTy->getElementType()) 1664 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 1665 1666 SV = Builder.CreateInsertElement(Old, SV, 1667 Context->getConstantInt(Type::Int32Ty, Elt), 1668 "tmp"); 1669 return SV; 1670 } 1671 1672 // If SV is a first-class aggregate value, insert each value recursively. 1673 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 1674 const StructLayout &Layout = *TD->getStructLayout(ST); 1675 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1676 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1677 Old = ConvertScalar_InsertValue(Elt, Old, 1678 Offset+Layout.getElementOffsetInBits(i), 1679 Builder); 1680 } 1681 return Old; 1682 } 1683 1684 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 1685 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 1686 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1687 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1688 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 1689 } 1690 return Old; 1691 } 1692 1693 // If SV is a float, convert it to the appropriate integer type. 1694 // If it is a pointer, do the same. 1695 unsigned SrcWidth = TD->getTypeSizeInBits(SV->getType()); 1696 unsigned DestWidth = TD->getTypeSizeInBits(AllocaType); 1697 unsigned SrcStoreWidth = TD->getTypeStoreSizeInBits(SV->getType()); 1698 unsigned DestStoreWidth = TD->getTypeStoreSizeInBits(AllocaType); 1699 if (SV->getType()->isFloatingPoint() || isa<VectorType>(SV->getType())) 1700 SV = Builder.CreateBitCast(SV, Context->getIntegerType(SrcWidth), "tmp"); 1701 else if (isa<PointerType>(SV->getType())) 1702 SV = Builder.CreatePtrToInt(SV, TD->getIntPtrType(), "tmp"); 1703 1704 // Zero extend or truncate the value if needed. 1705 if (SV->getType() != AllocaType) { 1706 if (SV->getType()->getPrimitiveSizeInBits() < 1707 AllocaType->getPrimitiveSizeInBits()) 1708 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 1709 else { 1710 // Truncation may be needed if storing more than the alloca can hold 1711 // (undefined behavior). 1712 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 1713 SrcWidth = DestWidth; 1714 SrcStoreWidth = DestStoreWidth; 1715 } 1716 } 1717 1718 // If this is a big-endian system and the store is narrower than the 1719 // full alloca type, we need to do a shift to get the right bits. 1720 int ShAmt = 0; 1721 if (TD->isBigEndian()) { 1722 // On big-endian machines, the lowest bit is stored at the bit offset 1723 // from the pointer given by getTypeStoreSizeInBits. This matters for 1724 // integers with a bitwidth that is not a multiple of 8. 1725 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 1726 } else { 1727 ShAmt = Offset; 1728 } 1729 1730 // Note: we support negative bitwidths (with shr) which are not defined. 1731 // We do this to support (f.e.) stores off the end of a structure where 1732 // only some bits in the structure are set. 1733 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 1734 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 1735 SV = Builder.CreateShl(SV, Context->getConstantInt(SV->getType(), 1736 ShAmt), "tmp"); 1737 Mask <<= ShAmt; 1738 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1739 SV = Builder.CreateLShr(SV, Context->getConstantInt(SV->getType(), 1740 -ShAmt), "tmp"); 1741 Mask = Mask.lshr(-ShAmt); 1742 } 1743 1744 // Mask out the bits we are about to insert from the old value, and or 1745 // in the new bits. 1746 if (SrcWidth != DestWidth) { 1747 assert(DestWidth > SrcWidth); 1748 Old = Builder.CreateAnd(Old, Context->getConstantInt(~Mask), "mask"); 1749 SV = Builder.CreateOr(Old, SV, "ins"); 1750 } 1751 return SV; 1752} 1753 1754 1755 1756/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1757/// some part of a constant global variable. This intentionally only accepts 1758/// constant expressions because we don't can't rewrite arbitrary instructions. 1759static bool PointsToConstantGlobal(Value *V) { 1760 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1761 return GV->isConstant(); 1762 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1763 if (CE->getOpcode() == Instruction::BitCast || 1764 CE->getOpcode() == Instruction::GetElementPtr) 1765 return PointsToConstantGlobal(CE->getOperand(0)); 1766 return false; 1767} 1768 1769/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1770/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1771/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1772/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1773/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1774/// the alloca, and if the source pointer is a pointer to a constant global, we 1775/// can optimize this. 1776static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy, 1777 bool isOffset) { 1778 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1779 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) 1780 // Ignore non-volatile loads, they are always ok. 1781 if (!LI->isVolatile()) 1782 continue; 1783 1784 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) { 1785 // If uses of the bitcast are ok, we are ok. 1786 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 1787 return false; 1788 continue; 1789 } 1790 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) { 1791 // If the GEP has all zero indices, it doesn't offset the pointer. If it 1792 // doesn't, it does. 1793 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 1794 isOffset || !GEP->hasAllZeroIndices())) 1795 return false; 1796 continue; 1797 } 1798 1799 // If this is isn't our memcpy/memmove, reject it as something we can't 1800 // handle. 1801 if (!isa<MemTransferInst>(*UI)) 1802 return false; 1803 1804 // If we already have seen a copy, reject the second one. 1805 if (TheCopy) return false; 1806 1807 // If the pointer has been offset from the start of the alloca, we can't 1808 // safely handle this. 1809 if (isOffset) return false; 1810 1811 // If the memintrinsic isn't using the alloca as the dest, reject it. 1812 if (UI.getOperandNo() != 1) return false; 1813 1814 MemIntrinsic *MI = cast<MemIntrinsic>(*UI); 1815 1816 // If the source of the memcpy/move is not a constant global, reject it. 1817 if (!PointsToConstantGlobal(MI->getOperand(2))) 1818 return false; 1819 1820 // Otherwise, the transform is safe. Remember the copy instruction. 1821 TheCopy = MI; 1822 } 1823 return true; 1824} 1825 1826/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 1827/// modified by a copy from a constant global. If we can prove this, we can 1828/// replace any uses of the alloca with uses of the global directly. 1829Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocationInst *AI) { 1830 Instruction *TheCopy = 0; 1831 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 1832 return TheCopy; 1833 return 0; 1834} 1835