ScalarReplAggregates.cpp revision 96326f9d312585532c95dcc31626f45f16cd5dd8
1//===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file was developed by the LLVM research group and is distributed under 6// the University of Illinois Open Source 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#include "llvm/Transforms/Scalar.h" 23#include "llvm/Constants.h" 24#include "llvm/DerivedTypes.h" 25#include "llvm/Function.h" 26#include "llvm/Pass.h" 27#include "llvm/Instructions.h" 28#include "llvm/Analysis/Dominators.h" 29#include "llvm/Support/GetElementPtrTypeIterator.h" 30#include "llvm/Target/TargetData.h" 31#include "llvm/Transforms/Utils/PromoteMemToReg.h" 32#include "llvm/Support/Debug.h" 33#include "llvm/ADT/Statistic.h" 34#include "llvm/ADT/StringExtras.h" 35using namespace llvm; 36 37namespace { 38 Statistic<> NumReplaced("scalarrepl", "Number of allocas broken up"); 39 Statistic<> NumPromoted("scalarrepl", "Number of allocas promoted"); 40 41 struct SROA : public FunctionPass { 42 bool runOnFunction(Function &F); 43 44 bool performScalarRepl(Function &F); 45 bool performPromotion(Function &F); 46 47 // getAnalysisUsage - This pass does not require any passes, but we know it 48 // will not alter the CFG, so say so. 49 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 50 AU.addRequired<DominatorTree>(); 51 AU.addRequired<DominanceFrontier>(); 52 AU.addRequired<TargetData>(); 53 AU.setPreservesCFG(); 54 } 55 56 private: 57 int isSafeElementUse(Value *Ptr); 58 int isSafeUseOfAllocation(Instruction *User); 59 int isSafeAllocaToScalarRepl(AllocationInst *AI); 60 void CanonicalizeAllocaUsers(AllocationInst *AI); 61 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocationInst *Base); 62 }; 63 64 RegisterOpt<SROA> X("scalarrepl", "Scalar Replacement of Aggregates"); 65} 66 67// Public interface to the ScalarReplAggregates pass 68FunctionPass *llvm::createScalarReplAggregatesPass() { return new SROA(); } 69 70 71bool SROA::runOnFunction(Function &F) { 72 bool Changed = performPromotion(F); 73 while (1) { 74 bool LocalChange = performScalarRepl(F); 75 if (!LocalChange) break; // No need to repromote if no scalarrepl 76 Changed = true; 77 LocalChange = performPromotion(F); 78 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 79 } 80 81 return Changed; 82} 83 84 85bool SROA::performPromotion(Function &F) { 86 std::vector<AllocaInst*> Allocas; 87 const TargetData &TD = getAnalysis<TargetData>(); 88 DominatorTree &DT = getAnalysis<DominatorTree>(); 89 DominanceFrontier &DF = getAnalysis<DominanceFrontier>(); 90 91 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 92 93 bool Changed = false; 94 95 while (1) { 96 Allocas.clear(); 97 98 // Find allocas that are safe to promote, by looking at all instructions in 99 // the entry node 100 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 101 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 102 if (isAllocaPromotable(AI, TD)) 103 Allocas.push_back(AI); 104 105 if (Allocas.empty()) break; 106 107 PromoteMemToReg(Allocas, DT, DF, TD); 108 NumPromoted += Allocas.size(); 109 Changed = true; 110 } 111 112 return Changed; 113} 114 115 116// performScalarRepl - This algorithm is a simple worklist driven algorithm, 117// which runs on all of the malloc/alloca instructions in the function, removing 118// them if they are only used by getelementptr instructions. 119// 120bool SROA::performScalarRepl(Function &F) { 121 std::vector<AllocationInst*> WorkList; 122 123 // Scan the entry basic block, adding any alloca's and mallocs to the worklist 124 BasicBlock &BB = F.getEntryBlock(); 125 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 126 if (AllocationInst *A = dyn_cast<AllocationInst>(I)) 127 WorkList.push_back(A); 128 129 // Process the worklist 130 bool Changed = false; 131 while (!WorkList.empty()) { 132 AllocationInst *AI = WorkList.back(); 133 WorkList.pop_back(); 134 135 // We cannot transform the allocation instruction if it is an array 136 // allocation (allocations OF arrays are ok though), and an allocation of a 137 // scalar value cannot be decomposed at all. 138 // 139 if (AI->isArrayAllocation() || 140 (!isa<StructType>(AI->getAllocatedType()) && 141 !isa<ArrayType>(AI->getAllocatedType()))) continue; 142 143 // Check that all of the users of the allocation are capable of being 144 // transformed. 145 switch (isSafeAllocaToScalarRepl(AI)) { 146 default: assert(0 && "Unexpected value!"); 147 case 0: // Not safe to scalar replace. 148 continue; 149 case 1: // Safe, but requires cleanup/canonicalizations first 150 CanonicalizeAllocaUsers(AI); 151 case 3: // Safe to scalar replace. 152 break; 153 } 154 155 DEBUG(std::cerr << "Found inst to xform: " << *AI); 156 Changed = true; 157 158 std::vector<AllocaInst*> ElementAllocas; 159 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 160 ElementAllocas.reserve(ST->getNumContainedTypes()); 161 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 162 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 163 AI->getName() + "." + utostr(i), AI); 164 ElementAllocas.push_back(NA); 165 WorkList.push_back(NA); // Add to worklist for recursive processing 166 } 167 } else { 168 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 169 ElementAllocas.reserve(AT->getNumElements()); 170 const Type *ElTy = AT->getElementType(); 171 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 172 AllocaInst *NA = new AllocaInst(ElTy, 0, 173 AI->getName() + "." + utostr(i), AI); 174 ElementAllocas.push_back(NA); 175 WorkList.push_back(NA); // Add to worklist for recursive processing 176 } 177 } 178 179 // Now that we have created the alloca instructions that we want to use, 180 // expand the getelementptr instructions to use them. 181 // 182 while (!AI->use_empty()) { 183 Instruction *User = cast<Instruction>(AI->use_back()); 184 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User); 185 // We now know that the GEP is of the form: GEP <ptr>, 0, <cst> 186 uint64_t Idx = cast<ConstantInt>(GEPI->getOperand(2))->getRawValue(); 187 188 assert(Idx < ElementAllocas.size() && "Index out of range?"); 189 AllocaInst *AllocaToUse = ElementAllocas[Idx]; 190 191 Value *RepValue; 192 if (GEPI->getNumOperands() == 3) { 193 // Do not insert a new getelementptr instruction with zero indices, only 194 // to have it optimized out later. 195 RepValue = AllocaToUse; 196 } else { 197 // We are indexing deeply into the structure, so we still need a 198 // getelement ptr instruction to finish the indexing. This may be 199 // expanded itself once the worklist is rerun. 200 // 201 std::string OldName = GEPI->getName(); // Steal the old name. 202 std::vector<Value*> NewArgs; 203 NewArgs.push_back(Constant::getNullValue(Type::IntTy)); 204 NewArgs.insert(NewArgs.end(), GEPI->op_begin()+3, GEPI->op_end()); 205 GEPI->setName(""); 206 RepValue = new GetElementPtrInst(AllocaToUse, NewArgs, OldName, GEPI); 207 } 208 209 // Move all of the users over to the new GEP. 210 GEPI->replaceAllUsesWith(RepValue); 211 // Delete the old GEP 212 GEPI->eraseFromParent(); 213 } 214 215 // Finally, delete the Alloca instruction 216 AI->getParent()->getInstList().erase(AI); 217 NumReplaced++; 218 } 219 220 return Changed; 221} 222 223 224/// isSafeElementUse - Check to see if this use is an allowed use for a 225/// getelementptr instruction of an array aggregate allocation. 226/// 227int SROA::isSafeElementUse(Value *Ptr) { 228 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 229 I != E; ++I) { 230 Instruction *User = cast<Instruction>(*I); 231 switch (User->getOpcode()) { 232 case Instruction::Load: break; 233 case Instruction::Store: 234 // Store is ok if storing INTO the pointer, not storing the pointer 235 if (User->getOperand(0) == Ptr) return 0; 236 break; 237 case Instruction::GetElementPtr: { 238 GetElementPtrInst *GEP = cast<GetElementPtrInst>(User); 239 if (GEP->getNumOperands() > 1) { 240 if (!isa<Constant>(GEP->getOperand(1)) || 241 !cast<Constant>(GEP->getOperand(1))->isNullValue()) 242 return 0; // Using pointer arithmetic to navigate the array... 243 } 244 if (!isSafeElementUse(GEP)) return 0; 245 break; 246 } 247 default: 248 DEBUG(std::cerr << " Transformation preventing inst: " << *User); 249 return 0; 250 } 251 } 252 return 3; // All users look ok :) 253} 254 255/// AllUsersAreLoads - Return true if all users of this value are loads. 256static bool AllUsersAreLoads(Value *Ptr) { 257 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 258 I != E; ++I) 259 if (cast<Instruction>(*I)->getOpcode() != Instruction::Load) 260 return false; 261 return true; 262} 263 264/// isSafeUseOfAllocation - Check to see if this user is an allowed use for an 265/// aggregate allocation. 266/// 267int SROA::isSafeUseOfAllocation(Instruction *User) { 268 if (!isa<GetElementPtrInst>(User)) return 0; 269 270 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User); 271 gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI); 272 273 // The GEP is safe to transform if it is of the form GEP <ptr>, 0, <cst> 274 if (I == E || 275 I.getOperand() != Constant::getNullValue(I.getOperand()->getType())) 276 return 0; 277 278 ++I; 279 if (I == E) return 0; // ran out of GEP indices?? 280 281 // If this is a use of an array allocation, do a bit more checking for sanity. 282 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 283 uint64_t NumElements = AT->getNumElements(); 284 285 if (ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand())) { 286 // Check to make sure that index falls within the array. If not, 287 // something funny is going on, so we won't do the optimization. 288 // 289 if (cast<ConstantInt>(GEPI->getOperand(2))->getRawValue() >= NumElements) 290 return 0; 291 292 } else { 293 // If this is an array index and the index is not constant, we cannot 294 // promote... that is unless the array has exactly one or two elements in 295 // it, in which case we CAN promote it, but we have to canonicalize this 296 // out if this is the only problem. 297 if (NumElements == 1 || NumElements == 2) 298 return AllUsersAreLoads(GEPI) ? 1 : 0; // Canonicalization required! 299 return 0; 300 } 301 } 302 303 // If there are any non-simple uses of this getelementptr, make sure to reject 304 // them. 305 return isSafeElementUse(GEPI); 306} 307 308/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 309/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 310/// or 1 if safe after canonicalization has been performed. 311/// 312int SROA::isSafeAllocaToScalarRepl(AllocationInst *AI) { 313 // Loop over the use list of the alloca. We can only transform it if all of 314 // the users are safe to transform. 315 // 316 int isSafe = 3; 317 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); 318 I != E; ++I) { 319 isSafe &= isSafeUseOfAllocation(cast<Instruction>(*I)); 320 if (isSafe == 0) { 321 DEBUG(std::cerr << "Cannot transform: " << *AI << " due to user: " 322 << **I); 323 return 0; 324 } 325 } 326 // If we require cleanup, isSafe is now 1, otherwise it is 3. 327 return isSafe; 328} 329 330/// CanonicalizeAllocaUsers - If SROA reported that it can promote the specified 331/// allocation, but only if cleaned up, perform the cleanups required. 332void SROA::CanonicalizeAllocaUsers(AllocationInst *AI) { 333 // At this point, we know that the end result will be SROA'd and promoted, so 334 // we can insert ugly code if required so long as sroa+mem2reg will clean it 335 // up. 336 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 337 UI != E; ) { 338 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(*UI++); 339 gep_type_iterator I = gep_type_begin(GEPI); 340 ++I; 341 342 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 343 uint64_t NumElements = AT->getNumElements(); 344 345 if (!isa<ConstantInt>(I.getOperand())) { 346 if (NumElements == 1) { 347 GEPI->setOperand(2, Constant::getNullValue(Type::IntTy)); 348 } else { 349 assert(NumElements == 2 && "Unhandled case!"); 350 // All users of the GEP must be loads. At each use of the GEP, insert 351 // two loads of the appropriate indexed GEP and select between them. 352 Value *IsOne = BinaryOperator::createSetNE(I.getOperand(), 353 Constant::getNullValue(I.getOperand()->getType()), 354 "isone", GEPI); 355 // Insert the new GEP instructions, which are properly indexed. 356 std::vector<Value*> Indices(GEPI->op_begin()+1, GEPI->op_end()); 357 Indices[1] = Constant::getNullValue(Type::IntTy); 358 Value *ZeroIdx = new GetElementPtrInst(GEPI->getOperand(0), Indices, 359 GEPI->getName()+".0", GEPI); 360 Indices[1] = ConstantInt::get(Type::IntTy, 1); 361 Value *OneIdx = new GetElementPtrInst(GEPI->getOperand(0), Indices, 362 GEPI->getName()+".1", GEPI); 363 // Replace all loads of the variable index GEP with loads from both 364 // indexes and a select. 365 while (!GEPI->use_empty()) { 366 LoadInst *LI = cast<LoadInst>(GEPI->use_back()); 367 Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI); 368 Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI); 369 Value *R = new SelectInst(IsOne, One, Zero, LI->getName(), LI); 370 LI->replaceAllUsesWith(R); 371 LI->eraseFromParent(); 372 } 373 GEPI->eraseFromParent(); 374 } 375 } 376 } 377 } 378} 379