LLVM API Documentation
00001 //===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===// 00002 // 00003 // The LLVM Compiler Infrastructure 00004 // 00005 // This file is distributed under the University of Illinois Open Source 00006 // License. See LICENSE.TXT for details. 00007 // 00008 //===----------------------------------------------------------------------===// 00009 // 00010 // This transformation implements the well known scalar replacement of 00011 // aggregates transformation. This xform breaks up alloca instructions of 00012 // aggregate type (structure or array) into individual alloca instructions for 00013 // each member (if possible). Then, if possible, it transforms the individual 00014 // alloca instructions into nice clean scalar SSA form. 00015 // 00016 // This combines a simple SRoA algorithm with the Mem2Reg algorithm because 00017 // often interact, especially for C++ programs. As such, iterating between 00018 // SRoA, then Mem2Reg until we run out of things to promote works well. 00019 // 00020 //===----------------------------------------------------------------------===// 00021 00022 #define DEBUG_TYPE "scalarrepl" 00023 #include "llvm/Transforms/Scalar.h" 00024 #include "llvm/Constants.h" 00025 #include "llvm/DerivedTypes.h" 00026 #include "llvm/Function.h" 00027 #include "llvm/GlobalVariable.h" 00028 #include "llvm/Instructions.h" 00029 #include "llvm/IntrinsicInst.h" 00030 #include "llvm/LLVMContext.h" 00031 #include "llvm/Pass.h" 00032 #include "llvm/Analysis/Dominators.h" 00033 #include "llvm/Target/TargetData.h" 00034 #include "llvm/Transforms/Utils/PromoteMemToReg.h" 00035 #include "llvm/Transforms/Utils/Local.h" 00036 #include "llvm/Support/Debug.h" 00037 #include "llvm/Support/ErrorHandling.h" 00038 #include "llvm/Support/GetElementPtrTypeIterator.h" 00039 #include "llvm/Support/IRBuilder.h" 00040 #include "llvm/Support/MathExtras.h" 00041 #include "llvm/Support/Compiler.h" 00042 #include "llvm/ADT/SmallVector.h" 00043 #include "llvm/ADT/Statistic.h" 00044 using namespace llvm; 00045 00046 STATISTIC(NumReplaced, "Number of allocas broken up"); 00047 STATISTIC(NumPromoted, "Number of allocas promoted"); 00048 STATISTIC(NumConverted, "Number of aggregates converted to scalar"); 00049 STATISTIC(NumGlobals, "Number of allocas copied from constant global"); 00050 00051 namespace { 00052 struct VISIBILITY_HIDDEN SROA : public FunctionPass { 00053 static char ID; // Pass identification, replacement for typeid 00054 explicit SROA(signed T = -1) : FunctionPass(&ID) { 00055 if (T == -1) 00056 SRThreshold = 128; 00057 else 00058 SRThreshold = T; 00059 } 00060 00061 bool runOnFunction(Function &F); 00062 00063 bool performScalarRepl(Function &F); 00064 bool performPromotion(Function &F); 00065 00066 // getAnalysisUsage - This pass does not require any passes, but we know it 00067 // will not alter the CFG, so say so. 00068 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 00069 AU.addRequired<DominatorTree>(); 00070 AU.addRequired<DominanceFrontier>(); 00071 AU.setPreservesCFG(); 00072 } 00073 00074 private: 00075 TargetData *TD; 00076 00077 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures 00078 /// information about the uses. All these fields are initialized to false 00079 /// and set to true when something is learned. 00080 struct AllocaInfo { 00081 /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 00082 bool isUnsafe : 1; 00083 00084 /// needsCleanup - This is set to true if there is some use of the alloca 00085 /// that requires cleanup. 00086 bool needsCleanup : 1; 00087 00088 /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 00089 bool isMemCpySrc : 1; 00090 00091 /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 00092 bool isMemCpyDst : 1; 00093 00094 AllocaInfo() 00095 : isUnsafe(false), needsCleanup(false), 00096 isMemCpySrc(false), isMemCpyDst(false) {} 00097 }; 00098 00099 unsigned SRThreshold; 00100 00101 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; } 00102 00103 int isSafeAllocaToScalarRepl(AllocationInst *AI); 00104 00105 void isSafeUseOfAllocation(Instruction *User, AllocationInst *AI, 00106 AllocaInfo &Info); 00107 void isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI, 00108 AllocaInfo &Info); 00109 void isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI, 00110 unsigned OpNo, AllocaInfo &Info); 00111 void isSafeUseOfBitCastedAllocation(BitCastInst *User, AllocationInst *AI, 00112 AllocaInfo &Info); 00113 00114 void DoScalarReplacement(AllocationInst *AI, 00115 std::vector<AllocationInst*> &WorkList); 00116 void CleanupGEP(GetElementPtrInst *GEP); 00117 void CleanupAllocaUsers(AllocationInst *AI); 00118 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocationInst *Base); 00119 00120 void RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI, 00121 SmallVector<AllocaInst*, 32> &NewElts); 00122 00123 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst, 00124 AllocationInst *AI, 00125 SmallVector<AllocaInst*, 32> &NewElts); 00126 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocationInst *AI, 00127 SmallVector<AllocaInst*, 32> &NewElts); 00128 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI, 00129 SmallVector<AllocaInst*, 32> &NewElts); 00130 00131 bool CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy, 00132 bool &SawVec, uint64_t Offset, unsigned AllocaSize); 00133 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); 00134 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType, 00135 uint64_t Offset, IRBuilder<> &Builder); 00136 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 00137 uint64_t Offset, IRBuilder<> &Builder); 00138 static Instruction *isOnlyCopiedFromConstantGlobal(AllocationInst *AI); 00139 }; 00140 } 00141 00142 char SROA::ID = 0; 00143 static RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates"); 00144 00145 // Public interface to the ScalarReplAggregates pass 00146 FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) { 00147 return new SROA(Threshold); 00148 } 00149 00150 00151 bool SROA::runOnFunction(Function &F) { 00152 TD = getAnalysisIfAvailable<TargetData>(); 00153 00154 bool Changed = performPromotion(F); 00155 00156 // FIXME: ScalarRepl currently depends on TargetData more than it 00157 // theoretically needs to. It should be refactored in order to support 00158 // target-independent IR. Until this is done, just skip the actual 00159 // scalar-replacement portion of this pass. 00160 if (!TD) return Changed; 00161 00162 while (1) { 00163 bool LocalChange = performScalarRepl(F); 00164 if (!LocalChange) break; // No need to repromote if no scalarrepl 00165 Changed = true; 00166 LocalChange = performPromotion(F); 00167 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 00168 } 00169 00170 return Changed; 00171 } 00172 00173 00174 bool SROA::performPromotion(Function &F) { 00175 std::vector<AllocaInst*> Allocas; 00176 DominatorTree &DT = getAnalysis<DominatorTree>(); 00177 DominanceFrontier &DF = getAnalysis<DominanceFrontier>(); 00178 00179 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 00180 00181 bool Changed = false; 00182 00183 while (1) { 00184 Allocas.clear(); 00185 00186 // Find allocas that are safe to promote, by looking at all instructions in 00187 // the entry node 00188 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 00189 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 00190 if (isAllocaPromotable(AI)) 00191 Allocas.push_back(AI); 00192 00193 if (Allocas.empty()) break; 00194 00195 PromoteMemToReg(Allocas, DT, DF, F.getContext()); 00196 NumPromoted += Allocas.size(); 00197 Changed = true; 00198 } 00199 00200 return Changed; 00201 } 00202 00203 /// getNumSAElements - Return the number of elements in the specific struct or 00204 /// array. 00205 static uint64_t getNumSAElements(const Type *T) { 00206 if (const StructType *ST = dyn_cast<StructType>(T)) 00207 return ST->getNumElements(); 00208 return cast<ArrayType>(T)->getNumElements(); 00209 } 00210 00211 // performScalarRepl - This algorithm is a simple worklist driven algorithm, 00212 // which runs on all of the malloc/alloca instructions in the function, removing 00213 // them if they are only used by getelementptr instructions. 00214 // 00215 bool SROA::performScalarRepl(Function &F) { 00216 std::vector<AllocationInst*> WorkList; 00217 00218 // Scan the entry basic block, adding any alloca's and mallocs to the worklist 00219 BasicBlock &BB = F.getEntryBlock(); 00220 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 00221 if (AllocationInst *A = dyn_cast<AllocationInst>(I)) 00222 WorkList.push_back(A); 00223 00224 // Process the worklist 00225 bool Changed = false; 00226 while (!WorkList.empty()) { 00227 AllocationInst *AI = WorkList.back(); 00228 WorkList.pop_back(); 00229 00230 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 00231 // with unused elements. 00232 if (AI->use_empty()) { 00233 AI->eraseFromParent(); 00234 continue; 00235 } 00236 00237 // If this alloca is impossible for us to promote, reject it early. 00238 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 00239 continue; 00240 00241 // Check to see if this allocation is only modified by a memcpy/memmove from 00242 // a constant global. If this is the case, we can change all users to use 00243 // the constant global instead. This is commonly produced by the CFE by 00244 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 00245 // is only subsequently read. 00246 if (Instruction *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 00247 DOUT << "Found alloca equal to global: " << *AI; 00248 DOUT << " memcpy = " << *TheCopy; 00249 Constant *TheSrc = cast<Constant>(TheCopy->getOperand(2)); 00250 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 00251 TheCopy->eraseFromParent(); // Don't mutate the global. 00252 AI->eraseFromParent(); 00253 ++NumGlobals; 00254 Changed = true; 00255 continue; 00256 } 00257 00258 // Check to see if we can perform the core SROA transformation. We cannot 00259 // transform the allocation instruction if it is an array allocation 00260 // (allocations OF arrays are ok though), and an allocation of a scalar 00261 // value cannot be decomposed at all. 00262 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 00263 00264 // Do not promote [0 x %struct]. 00265 if (AllocaSize == 0) continue; 00266 00267 // Do not promote any struct whose size is too big. 00268 if (AllocaSize > SRThreshold) continue; 00269 00270 if ((isa<StructType>(AI->getAllocatedType()) || 00271 isa<ArrayType>(AI->getAllocatedType())) && 00272 // Do not promote any struct into more than "32" separate vars. 00273 getNumSAElements(AI->getAllocatedType()) <= SRThreshold/4) { 00274 // Check that all of the users of the allocation are capable of being 00275 // transformed. 00276 switch (isSafeAllocaToScalarRepl(AI)) { 00277 default: llvm_unreachable("Unexpected value!"); 00278 case 0: // Not safe to scalar replace. 00279 break; 00280 case 1: // Safe, but requires cleanup/canonicalizations first 00281 CleanupAllocaUsers(AI); 00282 // FALL THROUGH. 00283 case 3: // Safe to scalar replace. 00284 DoScalarReplacement(AI, WorkList); 00285 Changed = true; 00286 continue; 00287 } 00288 } 00289 00290 // If we can turn this aggregate value (potentially with casts) into a 00291 // simple scalar value that can be mem2reg'd into a register value. 00292 // IsNotTrivial tracks whether this is something that mem2reg could have 00293 // promoted itself. If so, we don't want to transform it needlessly. Note 00294 // that we can't just check based on the type: the alloca may be of an i32 00295 // but that has pointer arithmetic to set byte 3 of it or something. 00296 bool IsNotTrivial = false; 00297 const Type *VectorTy = 0; 00298 bool HadAVector = false; 00299 if (CanConvertToScalar(AI, IsNotTrivial, VectorTy, HadAVector, 00300 0, unsigned(AllocaSize)) && IsNotTrivial) { 00301 AllocaInst *NewAI; 00302 // If we were able to find a vector type that can handle this with 00303 // insert/extract elements, and if there was at least one use that had 00304 // a vector type, promote this to a vector. We don't want to promote 00305 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 00306 // we just get a lot of insert/extracts. If at least one vector is 00307 // involved, then we probably really do have a union of vector/array. 00308 if (VectorTy && isa<VectorType>(VectorTy) && HadAVector) { 00309 DOUT << "CONVERT TO VECTOR: " << *AI << " TYPE = " << *VectorTy <<"\n"; 00310 00311 // Create and insert the vector alloca. 00312 NewAI = new AllocaInst(VectorTy, 0, "", AI->getParent()->begin()); 00313 ConvertUsesToScalar(AI, NewAI, 0); 00314 } else { 00315 DOUT << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"; 00316 00317 // Create and insert the integer alloca. 00318 const Type *NewTy = IntegerType::get(AI->getContext(), AllocaSize*8); 00319 NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 00320 ConvertUsesToScalar(AI, NewAI, 0); 00321 } 00322 NewAI->takeName(AI); 00323 AI->eraseFromParent(); 00324 ++NumConverted; 00325 Changed = true; 00326 continue; 00327 } 00328 00329 // Otherwise, couldn't process this alloca. 00330 } 00331 00332 return Changed; 00333 } 00334 00335 /// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 00336 /// predicate, do SROA now. 00337 void SROA::DoScalarReplacement(AllocationInst *AI, 00338 std::vector<AllocationInst*> &WorkList) { 00339 DOUT << "Found inst to SROA: " << *AI; 00340 SmallVector<AllocaInst*, 32> ElementAllocas; 00341 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 00342 ElementAllocas.reserve(ST->getNumContainedTypes()); 00343 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 00344 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 00345 AI->getAlignment(), 00346 AI->getName() + "." + Twine(i), AI); 00347 ElementAllocas.push_back(NA); 00348 WorkList.push_back(NA); // Add to worklist for recursive processing 00349 } 00350 } else { 00351 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 00352 ElementAllocas.reserve(AT->getNumElements()); 00353 const Type *ElTy = AT->getElementType(); 00354 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 00355 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 00356 AI->getName() + "." + Twine(i), AI); 00357 ElementAllocas.push_back(NA); 00358 WorkList.push_back(NA); // Add to worklist for recursive processing 00359 } 00360 } 00361 00362 // Now that we have created the alloca instructions that we want to use, 00363 // expand the getelementptr instructions to use them. 00364 // 00365 while (!AI->use_empty()) { 00366 Instruction *User = cast<Instruction>(AI->use_back()); 00367 if (BitCastInst *BCInst = dyn_cast<BitCastInst>(User)) { 00368 RewriteBitCastUserOfAlloca(BCInst, AI, ElementAllocas); 00369 BCInst->eraseFromParent(); 00370 continue; 00371 } 00372 00373 // Replace: 00374 // %res = load { i32, i32 }* %alloc 00375 // with: 00376 // %load.0 = load i32* %alloc.0 00377 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 00378 // %load.1 = load i32* %alloc.1 00379 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 00380 // (Also works for arrays instead of structs) 00381 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 00382 Value *Insert = UndefValue::get(LI->getType()); 00383 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) { 00384 Value *Load = new LoadInst(ElementAllocas[i], "load", LI); 00385 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 00386 } 00387 LI->replaceAllUsesWith(Insert); 00388 LI->eraseFromParent(); 00389 continue; 00390 } 00391 00392 // Replace: 00393 // store { i32, i32 } %val, { i32, i32 }* %alloc 00394 // with: 00395 // %val.0 = extractvalue { i32, i32 } %val, 0 00396 // store i32 %val.0, i32* %alloc.0 00397 // %val.1 = extractvalue { i32, i32 } %val, 1 00398 // store i32 %val.1, i32* %alloc.1 00399 // (Also works for arrays instead of structs) 00400 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 00401 Value *Val = SI->getOperand(0); 00402 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) { 00403 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 00404 new StoreInst(Extract, ElementAllocas[i], SI); 00405 } 00406 SI->eraseFromParent(); 00407 continue; 00408 } 00409 00410 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User); 00411 // We now know that the GEP is of the form: GEP <ptr>, 0, <cst> 00412 unsigned Idx = 00413 (unsigned)cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue(); 00414 00415 assert(Idx < ElementAllocas.size() && "Index out of range?"); 00416 AllocaInst *AllocaToUse = ElementAllocas[Idx]; 00417 00418 Value *RepValue; 00419 if (GEPI->getNumOperands() == 3) { 00420 // Do not insert a new getelementptr instruction with zero indices, only 00421 // to have it optimized out later. 00422 RepValue = AllocaToUse; 00423 } else { 00424 // We are indexing deeply into the structure, so we still need a 00425 // getelement ptr instruction to finish the indexing. This may be 00426 // expanded itself once the worklist is rerun. 00427 // 00428 SmallVector<Value*, 8> NewArgs; 00429 NewArgs.push_back(Constant::getNullValue( 00430 Type::getInt32Ty(AI->getContext()))); 00431 NewArgs.append(GEPI->op_begin()+3, GEPI->op_end()); 00432 RepValue = GetElementPtrInst::Create(AllocaToUse, NewArgs.begin(), 00433 NewArgs.end(), "", GEPI); 00434 RepValue->takeName(GEPI); 00435 } 00436 00437 // If this GEP is to the start of the aggregate, check for memcpys. 00438 if (Idx == 0 && GEPI->hasAllZeroIndices()) 00439 RewriteBitCastUserOfAlloca(GEPI, AI, ElementAllocas); 00440 00441 // Move all of the users over to the new GEP. 00442 GEPI->replaceAllUsesWith(RepValue); 00443 // Delete the old GEP 00444 GEPI->eraseFromParent(); 00445 } 00446 00447 // Finally, delete the Alloca instruction 00448 AI->eraseFromParent(); 00449 NumReplaced++; 00450 } 00451 00452 00453 /// isSafeElementUse - Check to see if this use is an allowed use for a 00454 /// getelementptr instruction of an array aggregate allocation. isFirstElt 00455 /// indicates whether Ptr is known to the start of the aggregate. 00456 /// 00457 void SROA::isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI, 00458 AllocaInfo &Info) { 00459 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 00460 I != E; ++I) { 00461 Instruction *User = cast<Instruction>(*I); 00462 switch (User->getOpcode()) { 00463 case Instruction::Load: break; 00464 case Instruction::Store: 00465 // Store is ok if storing INTO the pointer, not storing the pointer 00466 if (User->getOperand(0) == Ptr) return MarkUnsafe(Info); 00467 break; 00468 case Instruction::GetElementPtr: { 00469 GetElementPtrInst *GEP = cast<GetElementPtrInst>(User); 00470 bool AreAllZeroIndices = isFirstElt; 00471 if (GEP->getNumOperands() > 1) { 00472 if (!isa<ConstantInt>(GEP->getOperand(1)) || 00473 !cast<ConstantInt>(GEP->getOperand(1))->isZero()) 00474 // Using pointer arithmetic to navigate the array. 00475 return MarkUnsafe(Info); 00476 00477 if (AreAllZeroIndices) 00478 AreAllZeroIndices = GEP->hasAllZeroIndices(); 00479 } 00480 isSafeElementUse(GEP, AreAllZeroIndices, AI, Info); 00481 if (Info.isUnsafe) return; 00482 break; 00483 } 00484 case Instruction::BitCast: 00485 if (isFirstElt) { 00486 isSafeUseOfBitCastedAllocation(cast<BitCastInst>(User), AI, Info); 00487 if (Info.isUnsafe) return; 00488 break; 00489 } 00490 DOUT << " Transformation preventing inst: " << *User; 00491 return MarkUnsafe(Info); 00492 case Instruction::Call: 00493 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 00494 if (isFirstElt) { 00495 isSafeMemIntrinsicOnAllocation(MI, AI, I.getOperandNo(), Info); 00496 if (Info.isUnsafe) return; 00497 break; 00498 } 00499 } 00500 DOUT << " Transformation preventing inst: " << *User; 00501 return MarkUnsafe(Info); 00502 default: 00503 DOUT << " Transformation preventing inst: " << *User; 00504 return MarkUnsafe(Info); 00505 } 00506 } 00507 return; // All users look ok :) 00508 } 00509 00510 /// AllUsersAreLoads - Return true if all users of this value are loads. 00511 static bool AllUsersAreLoads(Value *Ptr) { 00512 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 00513 I != E; ++I) 00514 if (cast<Instruction>(*I)->getOpcode() != Instruction::Load) 00515 return false; 00516 return true; 00517 } 00518 00519 /// isSafeUseOfAllocation - Check to see if this user is an allowed use for an 00520 /// aggregate allocation. 00521 /// 00522 void SROA::isSafeUseOfAllocation(Instruction *User, AllocationInst *AI, 00523 AllocaInfo &Info) { 00524 if (BitCastInst *C = dyn_cast<BitCastInst>(User)) 00525 return isSafeUseOfBitCastedAllocation(C, AI, Info); 00526 00527 if (LoadInst *LI = dyn_cast<LoadInst>(User)) 00528 if (!LI->isVolatile()) 00529 return;// Loads (returning a first class aggregrate) are always rewritable 00530 00531 if (StoreInst *SI = dyn_cast<StoreInst>(User)) 00532 if (!SI->isVolatile() && SI->getOperand(0) != AI) 00533 return;// Store is ok if storing INTO the pointer, not storing the pointer 00534 00535 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User); 00536 if (GEPI == 0) 00537 return MarkUnsafe(Info); 00538 00539 gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI); 00540 00541 // The GEP is not safe to transform if not of the form "GEP <ptr>, 0, <cst>". 00542 if (I == E || 00543 I.getOperand() != Constant::getNullValue(I.getOperand()->getType())) { 00544 return MarkUnsafe(Info); 00545 } 00546 00547 ++I; 00548 if (I == E) return MarkUnsafe(Info); // ran out of GEP indices?? 00549 00550 bool IsAllZeroIndices = true; 00551 00552 // If the first index is a non-constant index into an array, see if we can 00553 // handle it as a special case. 00554 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 00555 if (!isa<ConstantInt>(I.getOperand())) { 00556 IsAllZeroIndices = 0; 00557 uint64_t NumElements = AT->getNumElements(); 00558 00559 // If this is an array index and the index is not constant, we cannot 00560 // promote... that is unless the array has exactly one or two elements in 00561 // it, in which case we CAN promote it, but we have to canonicalize this 00562 // out if this is the only problem. 00563 if ((NumElements == 1 || NumElements == 2) && 00564 AllUsersAreLoads(GEPI)) { 00565 Info.needsCleanup = true; 00566 return; // Canonicalization required! 00567 } 00568 return MarkUnsafe(Info); 00569 } 00570 } 00571 00572 // Walk through the GEP type indices, checking the types that this indexes 00573 // into. 00574 for (; I != E; ++I) { 00575 // Ignore struct elements, no extra checking needed for these. 00576 if (isa<StructType>(*I)) 00577 continue; 00578 00579 ConstantInt *IdxVal = dyn_cast<ConstantInt>(I.getOperand()); 00580 if (!IdxVal) return MarkUnsafe(Info); 00581 00582 // Are all indices still zero? 00583 IsAllZeroIndices &= IdxVal->isZero(); 00584 00585 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 00586 // This GEP indexes an array. Verify that this is an in-range constant 00587 // integer. Specifically, consider A[0][i]. We cannot know that the user 00588 // isn't doing invalid things like allowing i to index an out-of-range 00589 // subscript that accesses A[1]. Because of this, we have to reject SROA 00590 // of any accesses into structs where any of the components are variables. 00591 if (IdxVal->getZExtValue() >= AT->getNumElements()) 00592 return MarkUnsafe(Info); 00593 } else if (const VectorType *VT = dyn_cast<VectorType>(*I)) { 00594 if (IdxVal->getZExtValue() >= VT->getNumElements()) 00595 return MarkUnsafe(Info); 00596 } 00597 } 00598 00599 // If there are any non-simple uses of this getelementptr, make sure to reject 00600 // them. 00601 return isSafeElementUse(GEPI, IsAllZeroIndices, AI, Info); 00602 } 00603 00604 /// isSafeMemIntrinsicOnAllocation - Return true if the specified memory 00605 /// intrinsic can be promoted by SROA. At this point, we know that the operand 00606 /// of the memintrinsic is a pointer to the beginning of the allocation. 00607 void SROA::isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI, 00608 unsigned OpNo, AllocaInfo &Info) { 00609 // If not constant length, give up. 00610 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 00611 if (!Length) return MarkUnsafe(Info); 00612 00613 // If not the whole aggregate, give up. 00614 if (Length->getZExtValue() != 00615 TD->getTypeAllocSize(AI->getType()->getElementType())) 00616 return MarkUnsafe(Info); 00617 00618 // We only know about memcpy/memset/memmove. 00619 if (!isa<MemIntrinsic>(MI)) 00620 return MarkUnsafe(Info); 00621 00622 // Otherwise, we can transform it. Determine whether this is a memcpy/set 00623 // into or out of the aggregate. 00624 if (OpNo == 1) 00625 Info.isMemCpyDst = true; 00626 else { 00627 assert(OpNo == 2); 00628 Info.isMemCpySrc = true; 00629 } 00630 } 00631 00632 /// isSafeUseOfBitCastedAllocation - Return true if all users of this bitcast 00633 /// are 00634 void SROA::isSafeUseOfBitCastedAllocation(BitCastInst *BC, AllocationInst *AI, 00635 AllocaInfo &Info) { 00636 for (Value::use_iterator UI = BC->use_begin(), E = BC->use_end(); 00637 UI != E; ++UI) { 00638 if (BitCastInst *BCU = dyn_cast<BitCastInst>(UI)) { 00639 isSafeUseOfBitCastedAllocation(BCU, AI, Info); 00640 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) { 00641 isSafeMemIntrinsicOnAllocation(MI, AI, UI.getOperandNo(), Info); 00642 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI)) { 00643 if (SI->isVolatile()) 00644 return MarkUnsafe(Info); 00645 00646 // If storing the entire alloca in one chunk through a bitcasted pointer 00647 // to integer, we can transform it. This happens (for example) when you 00648 // cast a {i32,i32}* to i64* and store through it. This is similar to the 00649 // memcpy case and occurs in various "byval" cases and emulated memcpys. 00650 if (isa<IntegerType>(SI->getOperand(0)->getType()) && 00651 TD->getTypeAllocSize(SI->getOperand(0)->getType()) == 00652 TD->getTypeAllocSize(AI->getType()->getElementType())) { 00653 Info.isMemCpyDst = true; 00654 continue; 00655 } 00656 return MarkUnsafe(Info); 00657 } else if (LoadInst *LI = dyn_cast<LoadInst>(UI)) { 00658 if (LI->isVolatile()) 00659 return MarkUnsafe(Info); 00660 00661 // If loading the entire alloca in one chunk through a bitcasted pointer 00662 // to integer, we can transform it. This happens (for example) when you 00663 // cast a {i32,i32}* to i64* and load through it. This is similar to the 00664 // memcpy case and occurs in various "byval" cases and emulated memcpys. 00665 if (isa<IntegerType>(LI->getType()) && 00666 TD->getTypeAllocSize(LI->getType()) == 00667 TD->getTypeAllocSize(AI->getType()->getElementType())) { 00668 Info.isMemCpySrc = true; 00669 continue; 00670 } 00671 return MarkUnsafe(Info); 00672 } else if (isa<DbgInfoIntrinsic>(UI)) { 00673 // If one user is DbgInfoIntrinsic then check if all users are 00674 // DbgInfoIntrinsics. 00675 if (OnlyUsedByDbgInfoIntrinsics(BC)) { 00676 Info.needsCleanup = true; 00677 return; 00678 } 00679 else 00680 MarkUnsafe(Info); 00681 } 00682 else { 00683 return MarkUnsafe(Info); 00684 } 00685 if (Info.isUnsafe) return; 00686 } 00687 } 00688 00689 /// RewriteBitCastUserOfAlloca - BCInst (transitively) bitcasts AI, or indexes 00690 /// to its first element. Transform users of the cast to use the new values 00691 /// instead. 00692 void SROA::RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI, 00693 SmallVector<AllocaInst*, 32> &NewElts) { 00694 Value::use_iterator UI = BCInst->use_begin(), UE = BCInst->use_end(); 00695 while (UI != UE) { 00696 Instruction *User = cast<Instruction>(*UI++); 00697 if (BitCastInst *BCU = dyn_cast<BitCastInst>(User)) { 00698 RewriteBitCastUserOfAlloca(BCU, AI, NewElts); 00699 if (BCU->use_empty()) BCU->eraseFromParent(); 00700 continue; 00701 } 00702 00703 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 00704 // This must be memcpy/memmove/memset of the entire aggregate. 00705 // Split into one per element. 00706 RewriteMemIntrinUserOfAlloca(MI, BCInst, AI, NewElts); 00707 continue; 00708 } 00709 00710 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 00711 // If this is a store of the entire alloca from an integer, rewrite it. 00712 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 00713 continue; 00714 } 00715 00716 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 00717 // If this is a load of the entire alloca to an integer, rewrite it. 00718 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 00719 continue; 00720 } 00721 00722 // Otherwise it must be some other user of a gep of the first pointer. Just 00723 // leave these alone. 00724 continue; 00725 } 00726 } 00727 00728 /// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 00729 /// Rewrite it to copy or set the elements of the scalarized memory. 00730 void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst, 00731 AllocationInst *AI, 00732 SmallVector<AllocaInst*, 32> &NewElts) { 00733 00734 // If this is a memcpy/memmove, construct the other pointer as the 00735 // appropriate type. The "Other" pointer is the pointer that goes to memory 00736 // that doesn't have anything to do with the alloca that we are promoting. For 00737 // memset, this Value* stays null. 00738 Value *OtherPtr = 0; 00739 LLVMContext &Context = MI->getContext(); 00740 unsigned MemAlignment = MI->getAlignment(); 00741 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 00742 if (BCInst == MTI->getRawDest()) 00743 OtherPtr = MTI->getRawSource(); 00744 else { 00745 assert(BCInst == MTI->getRawSource()); 00746 OtherPtr = MTI->getRawDest(); 00747 } 00748 } 00749 00750 // If there is an other pointer, we want to convert it to the same pointer 00751 // type as AI has, so we can GEP through it safely. 00752 if (OtherPtr) { 00753 // It is likely that OtherPtr is a bitcast, if so, remove it. 00754 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr)) 00755 OtherPtr = BC->getOperand(0); 00756 // All zero GEPs are effectively bitcasts. 00757 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr)) 00758 if (GEP->hasAllZeroIndices()) 00759 OtherPtr = GEP->getOperand(0); 00760 00761 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr)) 00762 if (BCE->getOpcode() == Instruction::BitCast) 00763 OtherPtr = BCE->getOperand(0); 00764 00765 // If the pointer is not the right type, insert a bitcast to the right 00766 // type. 00767 if (OtherPtr->getType() != AI->getType()) 00768 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(), 00769 MI); 00770 } 00771 00772 // Process each element of the aggregate. 00773 Value *TheFn = MI->getOperand(0); 00774 const Type *BytePtrTy = MI->getRawDest()->getType(); 00775 bool SROADest = MI->getRawDest() == BCInst; 00776 00777 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 00778 00779 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 00780 // If this is a memcpy/memmove, emit a GEP of the other element address. 00781 Value *OtherElt = 0; 00782 unsigned OtherEltAlign = MemAlignment; 00783 00784 if (OtherPtr) { 00785 Value *Idx[2] = { Zero, 00786 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 00787 OtherElt = GetElementPtrInst::Create(OtherPtr, Idx, Idx + 2, 00788 OtherPtr->getNameStr()+"."+Twine(i), 00789 MI); 00790 uint64_t EltOffset; 00791 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 00792 if (const StructType *ST = 00793 dyn_cast<StructType>(OtherPtrTy->getElementType())) { 00794 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 00795 } else { 00796 const Type *EltTy = 00797 cast<SequentialType>(OtherPtr->getType())->getElementType(); 00798 EltOffset = TD->getTypeAllocSize(EltTy)*i; 00799 } 00800 00801 // The alignment of the other pointer is the guaranteed alignment of the 00802 // element, which is affected by both the known alignment of the whole 00803 // mem intrinsic and the alignment of the element. If the alignment of 00804 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 00805 // known alignment is just 4 bytes. 00806 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 00807 } 00808 00809 Value *EltPtr = NewElts[i]; 00810 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 00811 00812 // If we got down to a scalar, insert a load or store as appropriate. 00813 if (EltTy->isSingleValueType()) { 00814 if (isa<MemTransferInst>(MI)) { 00815 if (SROADest) { 00816 // From Other to Alloca. 00817 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 00818 new StoreInst(Elt, EltPtr, MI); 00819 } else { 00820 // From Alloca to Other. 00821 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 00822 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 00823 } 00824 continue; 00825 } 00826 assert(isa<MemSetInst>(MI)); 00827 00828 // If the stored element is zero (common case), just store a null 00829 // constant. 00830 Constant *StoreVal; 00831 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) { 00832 if (CI->isZero()) { 00833 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 00834 } else { 00835 // If EltTy is a vector type, get the element type. 00836 const Type *ValTy = EltTy->getScalarType(); 00837 00838 // Construct an integer with the right value. 00839 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 00840 APInt OneVal(EltSize, CI->getZExtValue()); 00841 APInt TotalVal(OneVal); 00842 // Set each byte. 00843 for (unsigned i = 0; 8*i < EltSize; ++i) { 00844 TotalVal = TotalVal.shl(8); 00845 TotalVal |= OneVal; 00846 } 00847 00848 // Convert the integer value to the appropriate type. 00849 StoreVal = ConstantInt::get(Context, TotalVal); 00850 if (isa<PointerType>(ValTy)) 00851 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 00852 else if (ValTy->isFloatingPoint()) 00853 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 00854 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 00855 00856 // If the requested value was a vector constant, create it. 00857 if (EltTy != ValTy) { 00858 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 00859 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 00860 StoreVal = ConstantVector::get(&Elts[0], NumElts); 00861 } 00862 } 00863 new StoreInst(StoreVal, EltPtr, MI); 00864 continue; 00865 } 00866 // Otherwise, if we're storing a byte variable, use a memset call for 00867 // this element. 00868 } 00869 00870 // Cast the element pointer to BytePtrTy. 00871 if (EltPtr->getType() != BytePtrTy) 00872 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI); 00873 00874 // Cast the other pointer (if we have one) to BytePtrTy. 00875 if (OtherElt && OtherElt->getType() != BytePtrTy) 00876 OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(), 00877 MI); 00878 00879 unsigned EltSize = TD->getTypeAllocSize(EltTy); 00880 00881 // Finally, insert the meminst for this element. 00882 if (isa<MemTransferInst>(MI)) { 00883 Value *Ops[] = { 00884 SROADest ? EltPtr : OtherElt, // Dest ptr 00885 SROADest ? OtherElt : EltPtr, // Src ptr 00886 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 00887 // Align 00888 ConstantInt::get(Type::getInt32Ty(MI->getContext()), OtherEltAlign) 00889 }; 00890 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 00891 } else { 00892 assert(isa<MemSetInst>(MI)); 00893 Value *Ops[] = { 00894 EltPtr, MI->getOperand(2), // Dest, Value, 00895 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 00896 Zero // Align 00897 }; 00898 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 00899 } 00900 } 00901 MI->eraseFromParent(); 00902 } 00903 00904 /// RewriteStoreUserOfWholeAlloca - We found an store of an integer that 00905 /// overwrites the entire allocation. Extract out the pieces of the stored 00906 /// integer and store them individually. 00907 void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, 00908 AllocationInst *AI, 00909 SmallVector<AllocaInst*, 32> &NewElts){ 00910 // Extract each element out of the integer according to its structure offset 00911 // and store the element value to the individual alloca. 00912 Value *SrcVal = SI->getOperand(0); 00913 const Type *AllocaEltTy = AI->getType()->getElementType(); 00914 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 00915 00916 // If this isn't a store of an integer to the whole alloca, it may be a store 00917 // to the first element. Just ignore the store in this case and normal SROA 00918 // will handle it. 00919 if (!isa<IntegerType>(SrcVal->getType()) || 00920 TD->getTypeAllocSizeInBits(SrcVal->getType()) != AllocaSizeBits) 00921 return; 00922 // Handle tail padding by extending the operand 00923 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 00924 SrcVal = new ZExtInst(SrcVal, 00925 IntegerType::get(SI->getContext(), AllocaSizeBits), 00926 "", SI); 00927 00928 DOUT << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << *SI; 00929 00930 // There are two forms here: AI could be an array or struct. Both cases 00931 // have different ways to compute the element offset. 00932 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 00933 const StructLayout *Layout = TD->getStructLayout(EltSTy); 00934 00935 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 00936 // Get the number of bits to shift SrcVal to get the value. 00937 const Type *FieldTy = EltSTy->getElementType(i); 00938 uint64_t Shift = Layout->getElementOffsetInBits(i); 00939 00940 if (TD->isBigEndian()) 00941 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 00942 00943 Value *EltVal = SrcVal; 00944 if (Shift) { 00945 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 00946 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 00947 "sroa.store.elt", SI); 00948 } 00949 00950 // Truncate down to an integer of the right size. 00951 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 00952 00953 // Ignore zero sized fields like {}, they obviously contain no data. 00954 if (FieldSizeBits == 0) continue; 00955 00956 if (FieldSizeBits != AllocaSizeBits) 00957 EltVal = new TruncInst(EltVal, 00958 IntegerType::get(SI->getContext(), FieldSizeBits), 00959 "", SI); 00960 Value *DestField = NewElts[i]; 00961 if (EltVal->getType() == FieldTy) { 00962 // Storing to an integer field of this size, just do it. 00963 } else if (FieldTy->isFloatingPoint() || isa<VectorType>(FieldTy)) { 00964 // Bitcast to the right element type (for fp/vector values). 00965 EltVal = new BitCastInst(EltVal, FieldTy, "", SI); 00966 } else { 00967 // Otherwise, bitcast the dest pointer (for aggregates). 00968 DestField = new BitCastInst(DestField, 00969 PointerType::getUnqual(EltVal->getType()), 00970 "", SI); 00971 } 00972 new StoreInst(EltVal, DestField, SI); 00973 } 00974 00975 } else { 00976 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 00977 const Type *ArrayEltTy = ATy->getElementType(); 00978 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 00979 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 00980 00981 uint64_t Shift; 00982 00983 if (TD->isBigEndian()) 00984 Shift = AllocaSizeBits-ElementOffset; 00985 else 00986 Shift = 0; 00987 00988 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 00989 // Ignore zero sized fields like {}, they obviously contain no data. 00990 if (ElementSizeBits == 0) continue; 00991 00992 Value *EltVal = SrcVal; 00993 if (Shift) { 00994 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 00995 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 00996 "sroa.store.elt", SI); 00997 } 00998 00999 // Truncate down to an integer of the right size. 01000 if (ElementSizeBits != AllocaSizeBits) 01001 EltVal = new TruncInst(EltVal, 01002 IntegerType::get(SI->getContext(), 01003 ElementSizeBits),"",SI); 01004 Value *DestField = NewElts[i]; 01005 if (EltVal->getType() == ArrayEltTy) { 01006 // Storing to an integer field of this size, just do it. 01007 } else if (ArrayEltTy->isFloatingPoint() || isa<VectorType>(ArrayEltTy)) { 01008 // Bitcast to the right element type (for fp/vector values). 01009 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI); 01010 } else { 01011 // Otherwise, bitcast the dest pointer (for aggregates). 01012 DestField = new BitCastInst(DestField, 01013 PointerType::getUnqual(EltVal->getType()), 01014 "", SI); 01015 } 01016 new StoreInst(EltVal, DestField, SI); 01017 01018 if (TD->isBigEndian()) 01019 Shift -= ElementOffset; 01020 else 01021 Shift += ElementOffset; 01022 } 01023 } 01024 01025 SI->eraseFromParent(); 01026 } 01027 01028 /// RewriteLoadUserOfWholeAlloca - We found an load of the entire allocation to 01029 /// an integer. Load the individual pieces to form the aggregate value. 01030 void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI, 01031 SmallVector<AllocaInst*, 32> &NewElts) { 01032 // Extract each element out of the NewElts according to its structure offset 01033 // and form the result value. 01034 const Type *AllocaEltTy = AI->getType()->getElementType(); 01035 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 01036 01037 // If this isn't a load of the whole alloca to an integer, it may be a load 01038 // of the first element. Just ignore the load in this case and normal SROA 01039 // will handle it. 01040 if (!isa<IntegerType>(LI->getType()) || 01041 TD->getTypeAllocSizeInBits(LI->getType()) != AllocaSizeBits) 01042 return; 01043 01044 DOUT << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << *LI; 01045 01046 // There are two forms here: AI could be an array or struct. Both cases 01047 // have different ways to compute the element offset. 01048 const StructLayout *Layout = 0; 01049 uint64_t ArrayEltBitOffset = 0; 01050 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 01051 Layout = TD->getStructLayout(EltSTy); 01052 } else { 01053 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 01054 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 01055 } 01056 01057 Value *ResultVal = 01058 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 01059 01060 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 01061 // Load the value from the alloca. If the NewElt is an aggregate, cast 01062 // the pointer to an integer of the same size before doing the load. 01063 Value *SrcField = NewElts[i]; 01064 const Type *FieldTy = 01065 cast<PointerType>(SrcField->getType())->getElementType(); 01066 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 01067 01068 // Ignore zero sized fields like {}, they obviously contain no data. 01069 if (FieldSizeBits == 0) continue; 01070 01071 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 01072 FieldSizeBits); 01073 if (!isa<IntegerType>(FieldTy) && !FieldTy->isFloatingPoint() && 01074 !isa<VectorType>(FieldTy)) 01075 SrcField = new BitCastInst(SrcField, 01076 PointerType::getUnqual(FieldIntTy), 01077 "", LI); 01078 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 01079 01080 // If SrcField is a fp or vector of the right size but that isn't an 01081 // integer type, bitcast to an integer so we can shift it. 01082 if (SrcField->getType() != FieldIntTy) 01083 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 01084 01085 // Zero extend the field to be the same size as the final alloca so that 01086 // we can shift and insert it. 01087 if (SrcField->getType() != ResultVal->getType()) 01088 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 01089 01090 // Determine the number of bits to shift SrcField. 01091 uint64_t Shift; 01092 if (Layout) // Struct case. 01093 Shift = Layout->getElementOffsetInBits(i); 01094 else // Array case. 01095 Shift = i*ArrayEltBitOffset; 01096 01097 if (TD->isBigEndian()) 01098 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 01099 01100 if (Shift) { 01101 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 01102 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 01103 } 01104 01105 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 01106 } 01107 01108 // Handle tail padding by truncating the result 01109 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 01110 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 01111 01112 LI->replaceAllUsesWith(ResultVal); 01113 LI->eraseFromParent(); 01114 } 01115 01116 01117 /// HasPadding - Return true if the specified type has any structure or 01118 /// alignment padding, false otherwise. 01119 static bool HasPadding(const Type *Ty, const TargetData &TD) { 01120 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 01121 const StructLayout *SL = TD.getStructLayout(STy); 01122 unsigned PrevFieldBitOffset = 0; 01123 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 01124 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 01125 01126 // Padding in sub-elements? 01127 if (HasPadding(STy->getElementType(i), TD)) 01128 return true; 01129 01130 // Check to see if there is any padding between this element and the 01131 // previous one. 01132 if (i) { 01133 unsigned PrevFieldEnd = 01134 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 01135 if (PrevFieldEnd < FieldBitOffset) 01136 return true; 01137 } 01138 01139 PrevFieldBitOffset = FieldBitOffset; 01140 } 01141 01142 // Check for tail padding. 01143 if (unsigned EltCount = STy->getNumElements()) { 01144 unsigned PrevFieldEnd = PrevFieldBitOffset + 01145 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 01146 if (PrevFieldEnd < SL->getSizeInBits()) 01147 return true; 01148 } 01149 01150 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 01151 return HasPadding(ATy->getElementType(), TD); 01152 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) { 01153 return HasPadding(VTy->getElementType(), TD); 01154 } 01155 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 01156 } 01157 01158 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 01159 /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 01160 /// or 1 if safe after canonicalization has been performed. 01161 /// 01162 int SROA::isSafeAllocaToScalarRepl(AllocationInst *AI) { 01163 // Loop over the use list of the alloca. We can only transform it if all of 01164 // the users are safe to transform. 01165 AllocaInfo Info; 01166 01167 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); 01168 I != E; ++I) { 01169 isSafeUseOfAllocation(cast<Instruction>(*I), AI, Info); 01170 if (Info.isUnsafe) { 01171 DOUT << "Cannot transform: " << *AI << " due to user: " << **I; 01172 return 0; 01173 } 01174 } 01175 01176 // Okay, we know all the users are promotable. If the aggregate is a memcpy 01177 // source and destination, we have to be careful. In particular, the memcpy 01178 // could be moving around elements that live in structure padding of the LLVM 01179 // types, but may actually be used. In these cases, we refuse to promote the 01180 // struct. 01181 if (Info.isMemCpySrc && Info.isMemCpyDst && 01182 HasPadding(AI->getType()->getElementType(), *TD)) 01183 return 0; 01184 01185 // If we require cleanup, return 1, otherwise return 3. 01186 return Info.needsCleanup ? 1 : 3; 01187 } 01188 01189 /// CleanupGEP - GEP is used by an Alloca, which can be prompted after the GEP 01190 /// is canonicalized here. 01191 void SROA::CleanupGEP(GetElementPtrInst *GEPI) { 01192 gep_type_iterator I = gep_type_begin(GEPI); 01193 ++I; 01194 01195 const ArrayType *AT = dyn_cast<ArrayType>(*I); 01196 if (!AT) 01197 return; 01198 01199 uint64_t NumElements = AT->getNumElements(); 01200 01201 if (isa<ConstantInt>(I.getOperand())) 01202 return; 01203 01204 if (NumElements == 1) { 01205 GEPI->setOperand(2, 01206 Constant::getNullValue(Type::getInt32Ty(GEPI->getContext()))); 01207 return; 01208 } 01209 01210 assert(NumElements == 2 && "Unhandled case!"); 01211 // All users of the GEP must be loads. At each use of the GEP, insert 01212 // two loads of the appropriate indexed GEP and select between them. 01213 Value *IsOne = new ICmpInst(GEPI, ICmpInst::ICMP_NE, I.getOperand(), 01214 Constant::getNullValue(I.getOperand()->getType()), 01215 "isone"); 01216 // Insert the new GEP instructions, which are properly indexed. 01217 SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end()); 01218 Indices[1] = Constant::getNullValue(Type::getInt32Ty(GEPI->getContext())); 01219 Value *ZeroIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 01220 Indices.begin(), 01221 Indices.end(), 01222 GEPI->getName()+".0", GEPI); 01223 Indices[1] = ConstantInt::get(Type::getInt32Ty(GEPI->getContext()), 1); 01224 Value *OneIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 01225 Indices.begin(), 01226 Indices.end(), 01227 GEPI->getName()+".1", GEPI); 01228 // Replace all loads of the variable index GEP with loads from both 01229 // indexes and a select. 01230 while (!GEPI->use_empty()) { 01231 LoadInst *LI = cast<LoadInst>(GEPI->use_back()); 01232 Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI); 01233 Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI); 01234 Value *R = SelectInst::Create(IsOne, One, Zero, LI->getName(), LI); 01235 LI->replaceAllUsesWith(R); 01236 LI->eraseFromParent(); 01237 } 01238 GEPI->eraseFromParent(); 01239 } 01240 01241 01242 /// CleanupAllocaUsers - If SROA reported that it can promote the specified 01243 /// allocation, but only if cleaned up, perform the cleanups required. 01244 void SROA::CleanupAllocaUsers(AllocationInst *AI) { 01245 // At this point, we know that the end result will be SROA'd and promoted, so 01246 // we can insert ugly code if required so long as sroa+mem2reg will clean it 01247 // up. 01248 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 01249 UI != E; ) { 01250 User *U = *UI++; 01251 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) 01252 CleanupGEP(GEPI); 01253 else { 01254 Instruction *I = cast<Instruction>(U); 01255 SmallVector<DbgInfoIntrinsic *, 2> DbgInUses; 01256 if (!isa<StoreInst>(I) && OnlyUsedByDbgInfoIntrinsics(I, &DbgInUses)) { 01257 // Safe to remove debug info uses. 01258 while (!DbgInUses.empty()) { 01259 DbgInfoIntrinsic *DI = DbgInUses.back(); DbgInUses.pop_back(); 01260 DI->eraseFromParent(); 01261 } 01262 I->eraseFromParent(); 01263 } 01264 } 01265 } 01266 } 01267 01268 /// MergeInType - Add the 'In' type to the accumulated type (Accum) so far at 01269 /// the offset specified by Offset (which is specified in bytes). 01270 /// 01271 /// There are two cases we handle here: 01272 /// 1) A union of vector types of the same size and potentially its elements. 01273 /// Here we turn element accesses into insert/extract element operations. 01274 /// This promotes a <4 x float> with a store of float to the third element 01275 /// into a <4 x float> that uses insert element. 01276 /// 2) A fully general blob of memory, which we turn into some (potentially 01277 /// large) integer type with extract and insert operations where the loads 01278 /// and stores would mutate the memory. 01279 static void MergeInType(const Type *In, uint64_t Offset, const Type *&VecTy, 01280 unsigned AllocaSize, const TargetData &TD, 01281 LLVMContext &Context) { 01282 // If this could be contributing to a vector, analyze it. 01283 if (VecTy != Type::getVoidTy(Context)) { // either null or a vector type. 01284 01285 // If the In type is a vector that is the same size as the alloca, see if it 01286 // matches the existing VecTy. 01287 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 01288 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 01289 // If we're storing/loading a vector of the right size, allow it as a 01290 // vector. If this the first vector we see, remember the type so that 01291 // we know the element size. 01292 if (VecTy == 0) 01293 VecTy = VInTy; 01294 return; 01295 } 01296 } else if (In == Type::getFloatTy(Context) || 01297 In == Type::getDoubleTy(Context) || 01298 (isa<IntegerType>(In) && In->getPrimitiveSizeInBits() >= 8 && 01299 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 01300 // If we're accessing something that could be an element of a vector, see 01301 // if the implied vector agrees with what we already have and if Offset is 01302 // compatible with it. 01303 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 01304 if (Offset % EltSize == 0 && 01305 AllocaSize % EltSize == 0 && 01306 (VecTy == 0 || 01307 cast<VectorType>(VecTy)->getElementType() 01308 ->getPrimitiveSizeInBits()/8 == EltSize)) { 01309 if (VecTy == 0) 01310 VecTy = VectorType::get(In, AllocaSize/EltSize); 01311 return; 01312 } 01313 } 01314 } 01315 01316 // Otherwise, we have a case that we can't handle with an optimized vector 01317 // form. We can still turn this into a large integer. 01318 VecTy = Type::getVoidTy(Context); 01319 } 01320 01321 /// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 01322 /// its accesses to use a to single vector type, return true, and set VecTy to 01323 /// the new type. If we could convert the alloca into a single promotable 01324 /// integer, return true but set VecTy to VoidTy. Further, if the use is not a 01325 /// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 01326 /// is the current offset from the base of the alloca being analyzed. 01327 /// 01328 /// If we see at least one access to the value that is as a vector type, set the 01329 /// SawVec flag. 01330 /// 01331 bool SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy, 01332 bool &SawVec, uint64_t Offset, 01333 unsigned AllocaSize) { 01334 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 01335 Instruction *User = cast<Instruction>(*UI); 01336 01337 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 01338 // Don't break volatile loads. 01339 if (LI->isVolatile()) 01340 return false; 01341 MergeInType(LI->getType(), Offset, VecTy, 01342 AllocaSize, *TD, V->getContext()); 01343 SawVec |= isa<VectorType>(LI->getType()); 01344 continue; 01345 } 01346 01347 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 01348 // Storing the pointer, not into the value? 01349 if (SI->getOperand(0) == V || SI->isVolatile()) return 0; 01350 MergeInType(SI->getOperand(0)->getType(), Offset, 01351 VecTy, AllocaSize, *TD, V->getContext()); 01352 SawVec |= isa<VectorType>(SI->getOperand(0)->getType()); 01353 continue; 01354 } 01355 01356 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 01357 if (!CanConvertToScalar(BCI, IsNotTrivial, VecTy, SawVec, Offset, 01358 AllocaSize)) 01359 return false; 01360 IsNotTrivial = true; 01361 continue; 01362 } 01363 01364 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 01365 // If this is a GEP with a variable indices, we can't handle it. 01366 if (!GEP->hasAllConstantIndices()) 01367 return false; 01368 01369 // Compute the offset that this GEP adds to the pointer. 01370 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 01371 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(), 01372 &Indices[0], Indices.size()); 01373 // See if all uses can be converted. 01374 if (!CanConvertToScalar(GEP, IsNotTrivial, VecTy, SawVec,Offset+GEPOffset, 01375 AllocaSize)) 01376 return false; 01377 IsNotTrivial = true; 01378 continue; 01379 } 01380 01381 // If this is a constant sized memset of a constant value (e.g. 0) we can 01382 // handle it. 01383 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 01384 // Store of constant value and constant size. 01385 if (isa<ConstantInt>(MSI->getValue()) && 01386 isa<ConstantInt>(MSI->getLength())) { 01387 IsNotTrivial = true; 01388 continue; 01389 } 01390 } 01391 01392 // If this is a memcpy or memmove into or out of the whole allocation, we 01393 // can handle it like a load or store of the scalar type. 01394 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 01395 if (ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength())) 01396 if (Len->getZExtValue() == AllocaSize && Offset == 0) { 01397 IsNotTrivial = true; 01398 continue; 01399 } 01400 } 01401 01402 // Ignore dbg intrinsic. 01403 if (isa<DbgInfoIntrinsic>(User)) 01404 continue; 01405 01406 // Otherwise, we cannot handle this! 01407 return false; 01408 } 01409 01410 return true; 01411 } 01412 01413 01414 /// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 01415 /// directly. This happens when we are converting an "integer union" to a 01416 /// single integer scalar, or when we are converting a "vector union" to a 01417 /// vector with insert/extractelement instructions. 01418 /// 01419 /// Offset is an offset from the original alloca, in bits that need to be 01420 /// shifted to the right. By the end of this, there should be no uses of Ptr. 01421 void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset) { 01422 while (!Ptr->use_empty()) { 01423 Instruction *User = cast<Instruction>(Ptr->use_back()); 01424 01425 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 01426 ConvertUsesToScalar(CI, NewAI, Offset); 01427 CI->eraseFromParent(); 01428 continue; 01429 } 01430 01431 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 01432 // Compute the offset that this GEP adds to the pointer. 01433 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 01434 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(), 01435 &Indices[0], Indices.size()); 01436 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 01437 GEP->eraseFromParent(); 01438 continue; 01439 } 01440 01441 IRBuilder<> Builder(User->getParent(), User); 01442 01443 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 01444 // The load is a bit extract from NewAI shifted right by Offset bits. 01445 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 01446 Value *NewLoadVal 01447 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 01448 LI->replaceAllUsesWith(NewLoadVal); 01449 LI->eraseFromParent(); 01450 continue; 01451 } 01452 01453 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 01454 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 01455 // FIXME: Remove once builder has Twine API. 01456 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").str().c_str()); 01457 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 01458 Builder); 01459 Builder.CreateStore(New, NewAI); 01460 SI->eraseFromParent(); 01461 continue; 01462 } 01463 01464 // If this is a constant sized memset of a constant value (e.g. 0) we can 01465 // transform it into a store of the expanded constant value. 01466 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 01467 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 01468 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 01469 if (NumBytes != 0) { 01470 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 01471 01472 // Compute the value replicated the right number of times. 01473 APInt APVal(NumBytes*8, Val); 01474 01475 // Splat the value if non-zero. 01476 if (Val) 01477 for (unsigned i = 1; i != NumBytes; ++i) 01478 APVal |= APVal << 8; 01479 01480 // FIXME: Remove once builder has Twine API. 01481 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").str().c_str()); 01482 Value *New = ConvertScalar_InsertValue( 01483 ConstantInt::get(User->getContext(), APVal), 01484 Old, Offset, Builder); 01485 Builder.CreateStore(New, NewAI); 01486 } 01487 MSI->eraseFromParent(); 01488 continue; 01489 } 01490 01491 // If this is a memcpy or memmove into or out of the whole allocation, we 01492 // can handle it like a load or store of the scalar type. 01493 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 01494 assert(Offset == 0 && "must be store to start of alloca"); 01495 01496 // If the source and destination are both to the same alloca, then this is 01497 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 01498 // as appropriate. 01499 AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject()); 01500 01501 if (MTI->getSource()->getUnderlyingObject() != OrigAI) { 01502 // Dest must be OrigAI, change this to be a load from the original 01503 // pointer (bitcasted), then a store to our new alloca. 01504 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 01505 Value *SrcPtr = MTI->getSource(); 01506 SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType()); 01507 01508 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 01509 SrcVal->setAlignment(MTI->getAlignment()); 01510 Builder.CreateStore(SrcVal, NewAI); 01511 } else if (MTI->getDest()->getUnderlyingObject() != OrigAI) { 01512 // Src must be OrigAI, change this to be a load from NewAI then a store 01513 // through the original dest pointer (bitcasted). 01514 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 01515 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 01516 01517 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType()); 01518 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 01519 NewStore->setAlignment(MTI->getAlignment()); 01520 } else { 01521 // Noop transfer. Src == Dst 01522 } 01523 01524 01525 MTI->eraseFromParent(); 01526 continue; 01527 } 01528 01529 // If user is a dbg info intrinsic then it is safe to remove it. 01530 if (isa<DbgInfoIntrinsic>(User)) { 01531 User->eraseFromParent(); 01532 continue; 01533 } 01534 01535 llvm_unreachable("Unsupported operation!"); 01536 } 01537 } 01538 01539 /// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 01540 /// or vector value FromVal, extracting the bits from the offset specified by 01541 /// Offset. This returns the value, which is of type ToType. 01542 /// 01543 /// This happens when we are converting an "integer union" to a single 01544 /// integer scalar, or when we are converting a "vector union" to a vector with 01545 /// insert/extractelement instructions. 01546 /// 01547 /// Offset is an offset from the original alloca, in bits that need to be 01548 /// shifted to the right. 01549 Value *SROA::ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 01550 uint64_t Offset, IRBuilder<> &Builder) { 01551 // If the load is of the whole new alloca, no conversion is needed. 01552 if (FromVal->getType() == ToType && Offset == 0) 01553 return FromVal; 01554 01555 // If the result alloca is a vector type, this is either an element 01556 // access or a bitcast to another vector type of the same size. 01557 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) { 01558 if (isa<VectorType>(ToType)) 01559 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 01560 01561 // Otherwise it must be an element access. 01562 unsigned Elt = 0; 01563 if (Offset) { 01564 unsigned EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 01565 Elt = Offset/EltSize; 01566 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 01567 } 01568 // Return the element extracted out of it. 01569 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get( 01570 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp"); 01571 if (V->getType() != ToType) 01572 V = Builder.CreateBitCast(V, ToType, "tmp"); 01573 return V; 01574 } 01575 01576 // If ToType is a first class aggregate, extract out each of the pieces and 01577 // use insertvalue's to form the FCA. 01578 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 01579 const StructLayout &Layout = *TD->getStructLayout(ST); 01580 Value *Res = UndefValue::get(ST); 01581 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 01582 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 01583 Offset+Layout.getElementOffsetInBits(i), 01584 Builder); 01585 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 01586 } 01587 return Res; 01588 } 01589 01590 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 01591 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 01592 Value *Res = UndefValue::get(AT); 01593 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 01594 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 01595 Offset+i*EltSize, Builder); 01596 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 01597 } 01598 return Res; 01599 } 01600 01601 // Otherwise, this must be a union that was converted to an integer value. 01602 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 01603 01604 // If this is a big-endian system and the load is narrower than the 01605 // full alloca type, we need to do a shift to get the right bits. 01606 int ShAmt = 0; 01607 if (TD->isBigEndian()) { 01608 // On big-endian machines, the lowest bit is stored at the bit offset 01609 // from the pointer given by getTypeStoreSizeInBits. This matters for 01610 // integers with a bitwidth that is not a multiple of 8. 01611 ShAmt = TD->getTypeStoreSizeInBits(NTy) - 01612 TD->getTypeStoreSizeInBits(ToType) - Offset; 01613 } else { 01614 ShAmt = Offset; 01615 } 01616 01617 // Note: we support negative bitwidths (with shl) which are not defined. 01618 // We do this to support (f.e.) loads off the end of a structure where 01619 // only some bits are used. 01620 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 01621 FromVal = Builder.CreateLShr(FromVal, 01622 ConstantInt::get(FromVal->getType(), 01623 ShAmt), "tmp"); 01624 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 01625 FromVal = Builder.CreateShl(FromVal, 01626 ConstantInt::get(FromVal->getType(), 01627 -ShAmt), "tmp"); 01628 01629 // Finally, unconditionally truncate the integer to the right width. 01630 unsigned LIBitWidth = TD->getTypeSizeInBits(ToType); 01631 if (LIBitWidth < NTy->getBitWidth()) 01632 FromVal = 01633 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 01634 LIBitWidth), "tmp"); 01635 else if (LIBitWidth > NTy->getBitWidth()) 01636 FromVal = 01637 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 01638 LIBitWidth), "tmp"); 01639 01640 // If the result is an integer, this is a trunc or bitcast. 01641 if (isa<IntegerType>(ToType)) { 01642 // Should be done. 01643 } else if (ToType->isFloatingPoint() || isa<VectorType>(ToType)) { 01644 // Just do a bitcast, we know the sizes match up. 01645 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 01646 } else { 01647 // Otherwise must be a pointer. 01648 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 01649 } 01650 assert(FromVal->getType() == ToType && "Didn't convert right?"); 01651 return FromVal; 01652 } 01653 01654 01655 /// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 01656 /// or vector value "Old" at the offset specified by Offset. 01657 /// 01658 /// This happens when we are converting an "integer union" to a 01659 /// single integer scalar, or when we are converting a "vector union" to a 01660 /// vector with insert/extractelement instructions. 01661 /// 01662 /// Offset is an offset from the original alloca, in bits that need to be 01663 /// shifted to the right. 01664 Value *SROA::ConvertScalar_InsertValue(Value *SV, Value *Old, 01665 uint64_t Offset, IRBuilder<> &Builder) { 01666 01667 // Convert the stored type to the actual type, shift it left to insert 01668 // then 'or' into place. 01669 const Type *AllocaType = Old->getType(); 01670 LLVMContext &Context = Old->getContext(); 01671 01672 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 01673 uint64_t VecSize = TD->getTypeAllocSizeInBits(VTy); 01674 uint64_t ValSize = TD->getTypeAllocSizeInBits(SV->getType()); 01675 01676 // Changing the whole vector with memset or with an access of a different 01677 // vector type? 01678 if (ValSize == VecSize) 01679 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 01680 01681 uint64_t EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 01682 01683 // Must be an element insertion. 01684 unsigned Elt = Offset/EltSize; 01685 01686 if (SV->getType() != VTy->getElementType()) 01687 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 01688 01689 SV = Builder.CreateInsertElement(Old, SV, 01690 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt), 01691 "tmp"); 01692 return SV; 01693 } 01694 01695 // If SV is a first-class aggregate value, insert each value recursively. 01696 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 01697 const StructLayout &Layout = *TD->getStructLayout(ST); 01698 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 01699 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 01700 Old = ConvertScalar_InsertValue(Elt, Old, 01701 Offset+Layout.getElementOffsetInBits(i), 01702 Builder); 01703 } 01704 return Old; 01705 } 01706 01707 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 01708 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 01709 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 01710 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 01711 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 01712 } 01713 return Old; 01714 } 01715 01716 // If SV is a float, convert it to the appropriate integer type. 01717 // If it is a pointer, do the same. 01718 unsigned SrcWidth = TD->getTypeSizeInBits(SV->getType()); 01719 unsigned DestWidth = TD->getTypeSizeInBits(AllocaType); 01720 unsigned SrcStoreWidth = TD->getTypeStoreSizeInBits(SV->getType()); 01721 unsigned DestStoreWidth = TD->getTypeStoreSizeInBits(AllocaType); 01722 if (SV->getType()->isFloatingPoint() || isa<VectorType>(SV->getType())) 01723 SV = Builder.CreateBitCast(SV, 01724 IntegerType::get(SV->getContext(),SrcWidth), "tmp"); 01725 else if (isa<PointerType>(SV->getType())) 01726 SV = Builder.CreatePtrToInt(SV, TD->getIntPtrType(SV->getContext()), "tmp"); 01727 01728 // Zero extend or truncate the value if needed. 01729 if (SV->getType() != AllocaType) { 01730 if (SV->getType()->getPrimitiveSizeInBits() < 01731 AllocaType->getPrimitiveSizeInBits()) 01732 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 01733 else { 01734 // Truncation may be needed if storing more than the alloca can hold 01735 // (undefined behavior). 01736 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 01737 SrcWidth = DestWidth; 01738 SrcStoreWidth = DestStoreWidth; 01739 } 01740 } 01741 01742 // If this is a big-endian system and the store is narrower than the 01743 // full alloca type, we need to do a shift to get the right bits. 01744 int ShAmt = 0; 01745 if (TD->isBigEndian()) { 01746 // On big-endian machines, the lowest bit is stored at the bit offset 01747 // from the pointer given by getTypeStoreSizeInBits. This matters for 01748 // integers with a bitwidth that is not a multiple of 8. 01749 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 01750 } else { 01751 ShAmt = Offset; 01752 } 01753 01754 // Note: we support negative bitwidths (with shr) which are not defined. 01755 // We do this to support (f.e.) stores off the end of a structure where 01756 // only some bits in the structure are set. 01757 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 01758 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 01759 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 01760 ShAmt), "tmp"); 01761 Mask <<= ShAmt; 01762 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 01763 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 01764 -ShAmt), "tmp"); 01765 Mask = Mask.lshr(-ShAmt); 01766 } 01767 01768 // Mask out the bits we are about to insert from the old value, and or 01769 // in the new bits. 01770 if (SrcWidth != DestWidth) { 01771 assert(DestWidth > SrcWidth); 01772 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 01773 SV = Builder.CreateOr(Old, SV, "ins"); 01774 } 01775 return SV; 01776 } 01777 01778 01779 01780 /// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 01781 /// some part of a constant global variable. This intentionally only accepts 01782 /// constant expressions because we don't can't rewrite arbitrary instructions. 01783 static bool PointsToConstantGlobal(Value *V) { 01784 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 01785 return GV->isConstant(); 01786 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 01787 if (CE->getOpcode() == Instruction::BitCast || 01788 CE->getOpcode() == Instruction::GetElementPtr) 01789 return PointsToConstantGlobal(CE->getOperand(0)); 01790 return false; 01791 } 01792 01793 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 01794 /// pointer to an alloca. Ignore any reads of the pointer, return false if we 01795 /// see any stores or other unknown uses. If we see pointer arithmetic, keep 01796 /// track of whether it moves the pointer (with isOffset) but otherwise traverse 01797 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 01798 /// the alloca, and if the source pointer is a pointer to a constant global, we 01799 /// can optimize this. 01800 static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy, 01801 bool isOffset) { 01802 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 01803 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) 01804 // Ignore non-volatile loads, they are always ok. 01805 if (!LI->isVolatile()) 01806 continue; 01807 01808 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) { 01809 // If uses of the bitcast are ok, we are ok. 01810 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 01811 return false; 01812 continue; 01813 } 01814 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) { 01815 // If the GEP has all zero indices, it doesn't offset the pointer. If it 01816 // doesn't, it does. 01817 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 01818 isOffset || !GEP->hasAllZeroIndices())) 01819 return false; 01820 continue; 01821 } 01822 01823 // If this is isn't our memcpy/memmove, reject it as something we can't 01824 // handle. 01825 if (!isa<MemTransferInst>(*UI)) 01826 return false; 01827 01828 // If we already have seen a copy, reject the second one. 01829 if (TheCopy) return false; 01830 01831 // If the pointer has been offset from the start of the alloca, we can't 01832 // safely handle this. 01833 if (isOffset) return false; 01834 01835 // If the memintrinsic isn't using the alloca as the dest, reject it. 01836 if (UI.getOperandNo() != 1) return false; 01837 01838 MemIntrinsic *MI = cast<MemIntrinsic>(*UI); 01839 01840 // If the source of the memcpy/move is not a constant global, reject it. 01841 if (!PointsToConstantGlobal(MI->getOperand(2))) 01842 return false; 01843 01844 // Otherwise, the transform is safe. Remember the copy instruction. 01845 TheCopy = MI; 01846 } 01847 return true; 01848 } 01849 01850 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 01851 /// modified by a copy from a constant global. If we can prove this, we can 01852 /// replace any uses of the alloca with uses of the global directly. 01853 Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocationInst *AI) { 01854 Instruction *TheCopy = 0; 01855 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 01856 return TheCopy; 01857 return 0; 01858 }