ScalarReplAggregates.cpp revision 02518140ac3310d0357c26a87b2372d85da9c2f4
1//===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This transformation implements the well known scalar replacement of 11// aggregates transformation. This xform breaks up alloca instructions of 12// aggregate type (structure or array) into individual alloca instructions for 13// each member (if possible). Then, if possible, it transforms the individual 14// alloca instructions into nice clean scalar SSA form. 15// 16// This combines a simple SRoA algorithm with the Mem2Reg algorithm because 17// often interact, especially for C++ programs. As such, iterating between 18// SRoA, then Mem2Reg until we run out of things to promote works well. 19// 20//===----------------------------------------------------------------------===// 21 22#define DEBUG_TYPE "scalarrepl" 23#include "llvm/Transforms/Scalar.h" 24#include "llvm/Constants.h" 25#include "llvm/DerivedTypes.h" 26#include "llvm/Function.h" 27#include "llvm/GlobalVariable.h" 28#include "llvm/Instructions.h" 29#include "llvm/IntrinsicInst.h" 30#include "llvm/Pass.h" 31#include "llvm/Analysis/Dominators.h" 32#include "llvm/Target/TargetData.h" 33#include "llvm/Transforms/Utils/PromoteMemToReg.h" 34#include "llvm/Support/Debug.h" 35#include "llvm/Support/GetElementPtrTypeIterator.h" 36#include "llvm/Support/MathExtras.h" 37#include "llvm/Support/Compiler.h" 38#include "llvm/ADT/SmallVector.h" 39#include "llvm/ADT/Statistic.h" 40#include "llvm/ADT/StringExtras.h" 41using namespace llvm; 42 43STATISTIC(NumReplaced, "Number of allocas broken up"); 44STATISTIC(NumPromoted, "Number of allocas promoted"); 45STATISTIC(NumConverted, "Number of aggregates converted to scalar"); 46STATISTIC(NumGlobals, "Number of allocas copied from constant global"); 47 48namespace { 49 struct VISIBILITY_HIDDEN SROA : public FunctionPass { 50 static char ID; // Pass identification, replacement for typeid 51 explicit SROA(signed T = -1) : FunctionPass((intptr_t)&ID) { 52 if (T == -1) 53 SRThreshold = 128; 54 else 55 SRThreshold = T; 56 } 57 58 bool runOnFunction(Function &F); 59 60 bool performScalarRepl(Function &F); 61 bool performPromotion(Function &F); 62 63 // getAnalysisUsage - This pass does not require any passes, but we know it 64 // will not alter the CFG, so say so. 65 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 66 AU.addRequired<DominatorTree>(); 67 AU.addRequired<DominanceFrontier>(); 68 AU.addRequired<TargetData>(); 69 AU.setPreservesCFG(); 70 } 71 72 private: 73 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures 74 /// information about the uses. All these fields are initialized to false 75 /// and set to true when something is learned. 76 struct AllocaInfo { 77 /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 78 bool isUnsafe : 1; 79 80 /// needsCanon - This is set to true if there is some use of the alloca 81 /// that requires canonicalization. 82 bool needsCanon : 1; 83 84 /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 85 bool isMemCpySrc : 1; 86 87 /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 88 bool isMemCpyDst : 1; 89 90 AllocaInfo() 91 : isUnsafe(false), needsCanon(false), 92 isMemCpySrc(false), isMemCpyDst(false) {} 93 }; 94 95 unsigned SRThreshold; 96 97 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; } 98 99 int isSafeAllocaToScalarRepl(AllocationInst *AI); 100 101 void isSafeUseOfAllocation(Instruction *User, AllocationInst *AI, 102 AllocaInfo &Info); 103 void isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI, 104 AllocaInfo &Info); 105 void isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI, 106 unsigned OpNo, AllocaInfo &Info); 107 void isSafeUseOfBitCastedAllocation(BitCastInst *User, AllocationInst *AI, 108 AllocaInfo &Info); 109 110 void DoScalarReplacement(AllocationInst *AI, 111 std::vector<AllocationInst*> &WorkList); 112 void CanonicalizeAllocaUsers(AllocationInst *AI); 113 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocationInst *Base); 114 115 void RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI, 116 SmallVector<AllocaInst*, 32> &NewElts); 117 118 const Type *CanConvertToScalar(Value *V, bool &IsNotTrivial); 119 void ConvertToScalar(AllocationInst *AI, const Type *Ty); 120 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, unsigned Offset); 121 Value *ConvertUsesOfLoadToScalar(LoadInst *LI, AllocaInst *NewAI, 122 unsigned Offset); 123 Value *ConvertUsesOfStoreToScalar(StoreInst *SI, AllocaInst *NewAI, 124 unsigned Offset); 125 static Instruction *isOnlyCopiedFromConstantGlobal(AllocationInst *AI); 126 }; 127} 128 129char SROA::ID = 0; 130static RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates"); 131 132// Public interface to the ScalarReplAggregates pass 133FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) { 134 return new SROA(Threshold); 135} 136 137 138bool SROA::runOnFunction(Function &F) { 139 bool Changed = performPromotion(F); 140 while (1) { 141 bool LocalChange = performScalarRepl(F); 142 if (!LocalChange) break; // No need to repromote if no scalarrepl 143 Changed = true; 144 LocalChange = performPromotion(F); 145 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 146 } 147 148 return Changed; 149} 150 151 152bool SROA::performPromotion(Function &F) { 153 std::vector<AllocaInst*> Allocas; 154 DominatorTree &DT = getAnalysis<DominatorTree>(); 155 DominanceFrontier &DF = getAnalysis<DominanceFrontier>(); 156 157 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 158 159 bool Changed = false; 160 161 while (1) { 162 Allocas.clear(); 163 164 // Find allocas that are safe to promote, by looking at all instructions in 165 // the entry node 166 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 167 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 168 if (isAllocaPromotable(AI)) 169 Allocas.push_back(AI); 170 171 if (Allocas.empty()) break; 172 173 PromoteMemToReg(Allocas, DT, DF); 174 NumPromoted += Allocas.size(); 175 Changed = true; 176 } 177 178 return Changed; 179} 180 181// performScalarRepl - This algorithm is a simple worklist driven algorithm, 182// which runs on all of the malloc/alloca instructions in the function, removing 183// them if they are only used by getelementptr instructions. 184// 185bool SROA::performScalarRepl(Function &F) { 186 std::vector<AllocationInst*> WorkList; 187 188 // Scan the entry basic block, adding any alloca's and mallocs to the worklist 189 BasicBlock &BB = F.getEntryBlock(); 190 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 191 if (AllocationInst *A = dyn_cast<AllocationInst>(I)) 192 WorkList.push_back(A); 193 194 const TargetData &TD = getAnalysis<TargetData>(); 195 196 // Process the worklist 197 bool Changed = false; 198 while (!WorkList.empty()) { 199 AllocationInst *AI = WorkList.back(); 200 WorkList.pop_back(); 201 202 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 203 // with unused elements. 204 if (AI->use_empty()) { 205 AI->eraseFromParent(); 206 continue; 207 } 208 209 // If we can turn this aggregate value (potentially with casts) into a 210 // simple scalar value that can be mem2reg'd into a register value. 211 bool IsNotTrivial = false; 212 if (const Type *ActualType = CanConvertToScalar(AI, IsNotTrivial)) 213 if (IsNotTrivial && ActualType != Type::VoidTy) { 214 ConvertToScalar(AI, ActualType); 215 Changed = true; 216 continue; 217 } 218 219 // Check to see if we can perform the core SROA transformation. We cannot 220 // transform the allocation instruction if it is an array allocation 221 // (allocations OF arrays are ok though), and an allocation of a scalar 222 // value cannot be decomposed at all. 223 if (!AI->isArrayAllocation() && 224 (isa<StructType>(AI->getAllocatedType()) || 225 isa<ArrayType>(AI->getAllocatedType())) && 226 AI->getAllocatedType()->isSized() && 227 TD.getABITypeSize(AI->getAllocatedType()) < SRThreshold) { 228 // Check that all of the users of the allocation are capable of being 229 // transformed. 230 switch (isSafeAllocaToScalarRepl(AI)) { 231 default: assert(0 && "Unexpected value!"); 232 case 0: // Not safe to scalar replace. 233 break; 234 case 1: // Safe, but requires cleanup/canonicalizations first 235 CanonicalizeAllocaUsers(AI); 236 // FALL THROUGH. 237 case 3: // Safe to scalar replace. 238 DoScalarReplacement(AI, WorkList); 239 Changed = true; 240 continue; 241 } 242 } 243 244 // Check to see if this allocation is only modified by a memcpy/memmove from 245 // a constant global. If this is the case, we can change all users to use 246 // the constant global instead. This is commonly produced by the CFE by 247 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 248 // is only subsequently read. 249 if (Instruction *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 250 DOUT << "Found alloca equal to global: " << *AI; 251 DOUT << " memcpy = " << *TheCopy; 252 Constant *TheSrc = cast<Constant>(TheCopy->getOperand(2)); 253 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 254 TheCopy->eraseFromParent(); // Don't mutate the global. 255 AI->eraseFromParent(); 256 ++NumGlobals; 257 Changed = true; 258 continue; 259 } 260 261 // Otherwise, couldn't process this. 262 } 263 264 return Changed; 265} 266 267/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 268/// predicate, do SROA now. 269void SROA::DoScalarReplacement(AllocationInst *AI, 270 std::vector<AllocationInst*> &WorkList) { 271 DOUT << "Found inst to SROA: " << *AI; 272 SmallVector<AllocaInst*, 32> ElementAllocas; 273 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 274 ElementAllocas.reserve(ST->getNumContainedTypes()); 275 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 276 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 277 AI->getAlignment(), 278 AI->getName() + "." + utostr(i), AI); 279 ElementAllocas.push_back(NA); 280 WorkList.push_back(NA); // Add to worklist for recursive processing 281 } 282 } else { 283 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 284 ElementAllocas.reserve(AT->getNumElements()); 285 const Type *ElTy = AT->getElementType(); 286 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 287 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 288 AI->getName() + "." + utostr(i), AI); 289 ElementAllocas.push_back(NA); 290 WorkList.push_back(NA); // Add to worklist for recursive processing 291 } 292 } 293 294 // Now that we have created the alloca instructions that we want to use, 295 // expand the getelementptr instructions to use them. 296 // 297 while (!AI->use_empty()) { 298 Instruction *User = cast<Instruction>(AI->use_back()); 299 if (BitCastInst *BCInst = dyn_cast<BitCastInst>(User)) { 300 RewriteBitCastUserOfAlloca(BCInst, AI, ElementAllocas); 301 BCInst->eraseFromParent(); 302 continue; 303 } 304 305 // Replace %res = load { i32, i32 }* %alloc 306 // by 307 // %load.0 = load i32* %alloc.0 308 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 309 // %load.1 = load i32* %alloc.1 310 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 311 // (Also works for arrays instead of structs) 312 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 313 Value *Insert = UndefValue::get(LI->getType()); 314 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) { 315 Value *Load = new LoadInst(ElementAllocas[i], "load", LI); 316 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 317 } 318 LI->replaceAllUsesWith(Insert); 319 LI->eraseFromParent(); 320 continue; 321 } 322 323 // Replace store { i32, i32 } %val, { i32, i32 }* %alloc 324 // by 325 // %val.0 = extractvalue { i32, i32 } %val, 0 326 // store i32 %val.0, i32* %alloc.0 327 // %val.1 = extractvalue { i32, i32 } %val, 1 328 // store i32 %val.1, i32* %alloc.1 329 // (Also works for arrays instead of structs) 330 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 331 Value *Val = SI->getOperand(0); 332 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) { 333 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 334 new StoreInst(Extract, ElementAllocas[i], SI); 335 } 336 SI->eraseFromParent(); 337 continue; 338 } 339 340 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User); 341 // We now know that the GEP is of the form: GEP <ptr>, 0, <cst> 342 unsigned Idx = 343 (unsigned)cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue(); 344 345 assert(Idx < ElementAllocas.size() && "Index out of range?"); 346 AllocaInst *AllocaToUse = ElementAllocas[Idx]; 347 348 Value *RepValue; 349 if (GEPI->getNumOperands() == 3) { 350 // Do not insert a new getelementptr instruction with zero indices, only 351 // to have it optimized out later. 352 RepValue = AllocaToUse; 353 } else { 354 // We are indexing deeply into the structure, so we still need a 355 // getelement ptr instruction to finish the indexing. This may be 356 // expanded itself once the worklist is rerun. 357 // 358 SmallVector<Value*, 8> NewArgs; 359 NewArgs.push_back(Constant::getNullValue(Type::Int32Ty)); 360 NewArgs.append(GEPI->op_begin()+3, GEPI->op_end()); 361 RepValue = GetElementPtrInst::Create(AllocaToUse, NewArgs.begin(), 362 NewArgs.end(), "", GEPI); 363 RepValue->takeName(GEPI); 364 } 365 366 // If this GEP is to the start of the aggregate, check for memcpys. 367 if (Idx == 0) { 368 bool IsStartOfAggregateGEP = true; 369 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i) { 370 if (!isa<ConstantInt>(GEPI->getOperand(i))) { 371 IsStartOfAggregateGEP = false; 372 break; 373 } 374 if (!cast<ConstantInt>(GEPI->getOperand(i))->isZero()) { 375 IsStartOfAggregateGEP = false; 376 break; 377 } 378 } 379 380 if (IsStartOfAggregateGEP) 381 RewriteBitCastUserOfAlloca(GEPI, AI, ElementAllocas); 382 } 383 384 385 // Move all of the users over to the new GEP. 386 GEPI->replaceAllUsesWith(RepValue); 387 // Delete the old GEP 388 GEPI->eraseFromParent(); 389 } 390 391 // Finally, delete the Alloca instruction 392 AI->eraseFromParent(); 393 NumReplaced++; 394} 395 396 397/// isSafeElementUse - Check to see if this use is an allowed use for a 398/// getelementptr instruction of an array aggregate allocation. isFirstElt 399/// indicates whether Ptr is known to the start of the aggregate. 400/// 401void SROA::isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI, 402 AllocaInfo &Info) { 403 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 404 I != E; ++I) { 405 Instruction *User = cast<Instruction>(*I); 406 switch (User->getOpcode()) { 407 case Instruction::Load: break; 408 case Instruction::Store: 409 // Store is ok if storing INTO the pointer, not storing the pointer 410 if (User->getOperand(0) == Ptr) return MarkUnsafe(Info); 411 break; 412 case Instruction::GetElementPtr: { 413 GetElementPtrInst *GEP = cast<GetElementPtrInst>(User); 414 bool AreAllZeroIndices = isFirstElt; 415 if (GEP->getNumOperands() > 1) { 416 if (!isa<ConstantInt>(GEP->getOperand(1)) || 417 !cast<ConstantInt>(GEP->getOperand(1))->isZero()) 418 // Using pointer arithmetic to navigate the array. 419 return MarkUnsafe(Info); 420 421 if (AreAllZeroIndices) { 422 for (unsigned i = 2, e = GEP->getNumOperands(); i != e; ++i) { 423 if (!isa<ConstantInt>(GEP->getOperand(i)) || 424 !cast<ConstantInt>(GEP->getOperand(i))->isZero()) { 425 AreAllZeroIndices = false; 426 break; 427 } 428 } 429 } 430 } 431 isSafeElementUse(GEP, AreAllZeroIndices, AI, Info); 432 if (Info.isUnsafe) return; 433 break; 434 } 435 case Instruction::BitCast: 436 if (isFirstElt) { 437 isSafeUseOfBitCastedAllocation(cast<BitCastInst>(User), AI, Info); 438 if (Info.isUnsafe) return; 439 break; 440 } 441 DOUT << " Transformation preventing inst: " << *User; 442 return MarkUnsafe(Info); 443 case Instruction::Call: 444 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 445 if (isFirstElt) { 446 isSafeMemIntrinsicOnAllocation(MI, AI, I.getOperandNo(), Info); 447 if (Info.isUnsafe) return; 448 break; 449 } 450 } 451 DOUT << " Transformation preventing inst: " << *User; 452 return MarkUnsafe(Info); 453 default: 454 DOUT << " Transformation preventing inst: " << *User; 455 return MarkUnsafe(Info); 456 } 457 } 458 return; // All users look ok :) 459} 460 461/// AllUsersAreLoads - Return true if all users of this value are loads. 462static bool AllUsersAreLoads(Value *Ptr) { 463 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 464 I != E; ++I) 465 if (cast<Instruction>(*I)->getOpcode() != Instruction::Load) 466 return false; 467 return true; 468} 469 470/// isSafeUseOfAllocation - Check to see if this user is an allowed use for an 471/// aggregate allocation. 472/// 473void SROA::isSafeUseOfAllocation(Instruction *User, AllocationInst *AI, 474 AllocaInfo &Info) { 475 if (BitCastInst *C = dyn_cast<BitCastInst>(User)) 476 return isSafeUseOfBitCastedAllocation(C, AI, Info); 477 478 if (isa<LoadInst>(User)) 479 return; // Loads (returning a first class aggregrate) are always rewritable 480 481 if (isa<StoreInst>(User) && User->getOperand(0) != AI) 482 return; // Store is ok if storing INTO the pointer, not storing the pointer 483 484 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User); 485 if (GEPI == 0) 486 return MarkUnsafe(Info); 487 488 gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI); 489 490 // The GEP is not safe to transform if not of the form "GEP <ptr>, 0, <cst>". 491 if (I == E || 492 I.getOperand() != Constant::getNullValue(I.getOperand()->getType())) { 493 return MarkUnsafe(Info); 494 } 495 496 ++I; 497 if (I == E) return MarkUnsafe(Info); // ran out of GEP indices?? 498 499 bool IsAllZeroIndices = true; 500 501 // If this is a use of an array allocation, do a bit more checking for sanity. 502 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 503 uint64_t NumElements = AT->getNumElements(); 504 505 if (ConstantInt *Idx = dyn_cast<ConstantInt>(I.getOperand())) { 506 IsAllZeroIndices &= Idx->isZero(); 507 508 // Check to make sure that index falls within the array. If not, 509 // something funny is going on, so we won't do the optimization. 510 // 511 if (Idx->getZExtValue() >= NumElements) 512 return MarkUnsafe(Info); 513 514 // We cannot scalar repl this level of the array unless any array 515 // sub-indices are in-range constants. In particular, consider: 516 // A[0][i]. We cannot know that the user isn't doing invalid things like 517 // allowing i to index an out-of-range subscript that accesses A[1]. 518 // 519 // Scalar replacing *just* the outer index of the array is probably not 520 // going to be a win anyway, so just give up. 521 for (++I; I != E && (isa<ArrayType>(*I) || isa<VectorType>(*I)); ++I) { 522 uint64_t NumElements; 523 if (const ArrayType *SubArrayTy = dyn_cast<ArrayType>(*I)) 524 NumElements = SubArrayTy->getNumElements(); 525 else 526 NumElements = cast<VectorType>(*I)->getNumElements(); 527 528 ConstantInt *IdxVal = dyn_cast<ConstantInt>(I.getOperand()); 529 if (!IdxVal) return MarkUnsafe(Info); 530 if (IdxVal->getZExtValue() >= NumElements) 531 return MarkUnsafe(Info); 532 IsAllZeroIndices &= IdxVal->isZero(); 533 } 534 535 } else { 536 IsAllZeroIndices = 0; 537 538 // If this is an array index and the index is not constant, we cannot 539 // promote... that is unless the array has exactly one or two elements in 540 // it, in which case we CAN promote it, but we have to canonicalize this 541 // out if this is the only problem. 542 if ((NumElements == 1 || NumElements == 2) && 543 AllUsersAreLoads(GEPI)) { 544 Info.needsCanon = true; 545 return; // Canonicalization required! 546 } 547 return MarkUnsafe(Info); 548 } 549 } 550 551 // If there are any non-simple uses of this getelementptr, make sure to reject 552 // them. 553 return isSafeElementUse(GEPI, IsAllZeroIndices, AI, Info); 554} 555 556/// isSafeMemIntrinsicOnAllocation - Return true if the specified memory 557/// intrinsic can be promoted by SROA. At this point, we know that the operand 558/// of the memintrinsic is a pointer to the beginning of the allocation. 559void SROA::isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI, 560 unsigned OpNo, AllocaInfo &Info) { 561 // If not constant length, give up. 562 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 563 if (!Length) return MarkUnsafe(Info); 564 565 // If not the whole aggregate, give up. 566 const TargetData &TD = getAnalysis<TargetData>(); 567 if (Length->getZExtValue() != 568 TD.getABITypeSize(AI->getType()->getElementType())) 569 return MarkUnsafe(Info); 570 571 // We only know about memcpy/memset/memmove. 572 if (!isa<MemCpyInst>(MI) && !isa<MemSetInst>(MI) && !isa<MemMoveInst>(MI)) 573 return MarkUnsafe(Info); 574 575 // Otherwise, we can transform it. Determine whether this is a memcpy/set 576 // into or out of the aggregate. 577 if (OpNo == 1) 578 Info.isMemCpyDst = true; 579 else { 580 assert(OpNo == 2); 581 Info.isMemCpySrc = true; 582 } 583} 584 585/// isSafeUseOfBitCastedAllocation - Return true if all users of this bitcast 586/// are 587void SROA::isSafeUseOfBitCastedAllocation(BitCastInst *BC, AllocationInst *AI, 588 AllocaInfo &Info) { 589 for (Value::use_iterator UI = BC->use_begin(), E = BC->use_end(); 590 UI != E; ++UI) { 591 if (BitCastInst *BCU = dyn_cast<BitCastInst>(UI)) { 592 isSafeUseOfBitCastedAllocation(BCU, AI, Info); 593 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) { 594 isSafeMemIntrinsicOnAllocation(MI, AI, UI.getOperandNo(), Info); 595 } else { 596 return MarkUnsafe(Info); 597 } 598 if (Info.isUnsafe) return; 599 } 600} 601 602/// RewriteBitCastUserOfAlloca - BCInst (transitively) bitcasts AI, or indexes 603/// to its first element. Transform users of the cast to use the new values 604/// instead. 605void SROA::RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI, 606 SmallVector<AllocaInst*, 32> &NewElts) { 607 Constant *Zero = Constant::getNullValue(Type::Int32Ty); 608 const TargetData &TD = getAnalysis<TargetData>(); 609 610 Value::use_iterator UI = BCInst->use_begin(), UE = BCInst->use_end(); 611 while (UI != UE) { 612 if (BitCastInst *BCU = dyn_cast<BitCastInst>(*UI)) { 613 RewriteBitCastUserOfAlloca(BCU, AI, NewElts); 614 ++UI; 615 BCU->eraseFromParent(); 616 continue; 617 } 618 619 // Otherwise, must be memcpy/memmove/memset of the entire aggregate. Split 620 // into one per element. 621 MemIntrinsic *MI = dyn_cast<MemIntrinsic>(*UI); 622 623 // If it's not a mem intrinsic, it must be some other user of a gep of the 624 // first pointer. Just leave these alone. 625 if (!MI) { 626 ++UI; 627 continue; 628 } 629 630 // If this is a memcpy/memmove, construct the other pointer as the 631 // appropriate type. 632 Value *OtherPtr = 0; 633 if (MemCpyInst *MCI = dyn_cast<MemCpyInst>(MI)) { 634 if (BCInst == MCI->getRawDest()) 635 OtherPtr = MCI->getRawSource(); 636 else { 637 assert(BCInst == MCI->getRawSource()); 638 OtherPtr = MCI->getRawDest(); 639 } 640 } else if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) { 641 if (BCInst == MMI->getRawDest()) 642 OtherPtr = MMI->getRawSource(); 643 else { 644 assert(BCInst == MMI->getRawSource()); 645 OtherPtr = MMI->getRawDest(); 646 } 647 } 648 649 // If there is an other pointer, we want to convert it to the same pointer 650 // type as AI has, so we can GEP through it. 651 if (OtherPtr) { 652 // It is likely that OtherPtr is a bitcast, if so, remove it. 653 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr)) 654 OtherPtr = BC->getOperand(0); 655 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr)) 656 if (BCE->getOpcode() == Instruction::BitCast) 657 OtherPtr = BCE->getOperand(0); 658 659 // If the pointer is not the right type, insert a bitcast to the right 660 // type. 661 if (OtherPtr->getType() != AI->getType()) 662 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(), 663 MI); 664 } 665 666 // Process each element of the aggregate. 667 Value *TheFn = MI->getOperand(0); 668 const Type *BytePtrTy = MI->getRawDest()->getType(); 669 bool SROADest = MI->getRawDest() == BCInst; 670 671 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 672 // If this is a memcpy/memmove, emit a GEP of the other element address. 673 Value *OtherElt = 0; 674 if (OtherPtr) { 675 Value *Idx[2]; 676 Idx[0] = Zero; 677 Idx[1] = ConstantInt::get(Type::Int32Ty, i); 678 OtherElt = GetElementPtrInst::Create(OtherPtr, Idx, Idx + 2, 679 OtherPtr->getNameStr()+"."+utostr(i), 680 MI); 681 } 682 683 Value *EltPtr = NewElts[i]; 684 const Type *EltTy =cast<PointerType>(EltPtr->getType())->getElementType(); 685 686 // If we got down to a scalar, insert a load or store as appropriate. 687 if (EltTy->isSingleValueType()) { 688 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) { 689 Value *Elt = new LoadInst(SROADest ? OtherElt : EltPtr, "tmp", 690 MI); 691 new StoreInst(Elt, SROADest ? EltPtr : OtherElt, MI); 692 continue; 693 } else { 694 assert(isa<MemSetInst>(MI)); 695 696 // If the stored element is zero (common case), just store a null 697 // constant. 698 Constant *StoreVal; 699 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) { 700 if (CI->isZero()) { 701 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 702 } else { 703 // If EltTy is a vector type, get the element type. 704 const Type *ValTy = EltTy; 705 if (const VectorType *VTy = dyn_cast<VectorType>(ValTy)) 706 ValTy = VTy->getElementType(); 707 708 // Construct an integer with the right value. 709 unsigned EltSize = TD.getTypeSizeInBits(ValTy); 710 APInt OneVal(EltSize, CI->getZExtValue()); 711 APInt TotalVal(OneVal); 712 // Set each byte. 713 for (unsigned i = 0; 8*i < EltSize; ++i) { 714 TotalVal = TotalVal.shl(8); 715 TotalVal |= OneVal; 716 } 717 718 // Convert the integer value to the appropriate type. 719 StoreVal = ConstantInt::get(TotalVal); 720 if (isa<PointerType>(ValTy)) 721 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 722 else if (ValTy->isFloatingPoint()) 723 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 724 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 725 726 // If the requested value was a vector constant, create it. 727 if (EltTy != ValTy) { 728 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 729 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 730 StoreVal = ConstantVector::get(&Elts[0], NumElts); 731 } 732 } 733 new StoreInst(StoreVal, EltPtr, MI); 734 continue; 735 } 736 // Otherwise, if we're storing a byte variable, use a memset call for 737 // this element. 738 } 739 } 740 741 // Cast the element pointer to BytePtrTy. 742 if (EltPtr->getType() != BytePtrTy) 743 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI); 744 745 // Cast the other pointer (if we have one) to BytePtrTy. 746 if (OtherElt && OtherElt->getType() != BytePtrTy) 747 OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(), 748 MI); 749 750 unsigned EltSize = TD.getABITypeSize(EltTy); 751 752 // Finally, insert the meminst for this element. 753 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) { 754 Value *Ops[] = { 755 SROADest ? EltPtr : OtherElt, // Dest ptr 756 SROADest ? OtherElt : EltPtr, // Src ptr 757 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 758 Zero // Align 759 }; 760 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 761 } else { 762 assert(isa<MemSetInst>(MI)); 763 Value *Ops[] = { 764 EltPtr, MI->getOperand(2), // Dest, Value, 765 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 766 Zero // Align 767 }; 768 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 769 } 770 } 771 772 // Finally, MI is now dead, as we've modified its actions to occur on all of 773 // the elements of the aggregate. 774 ++UI; 775 MI->eraseFromParent(); 776 } 777} 778 779/// HasPadding - Return true if the specified type has any structure or 780/// alignment padding, false otherwise. 781static bool HasPadding(const Type *Ty, const TargetData &TD) { 782 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 783 const StructLayout *SL = TD.getStructLayout(STy); 784 unsigned PrevFieldBitOffset = 0; 785 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 786 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 787 788 // Padding in sub-elements? 789 if (HasPadding(STy->getElementType(i), TD)) 790 return true; 791 792 // Check to see if there is any padding between this element and the 793 // previous one. 794 if (i) { 795 unsigned PrevFieldEnd = 796 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 797 if (PrevFieldEnd < FieldBitOffset) 798 return true; 799 } 800 801 PrevFieldBitOffset = FieldBitOffset; 802 } 803 804 // Check for tail padding. 805 if (unsigned EltCount = STy->getNumElements()) { 806 unsigned PrevFieldEnd = PrevFieldBitOffset + 807 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 808 if (PrevFieldEnd < SL->getSizeInBits()) 809 return true; 810 } 811 812 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 813 return HasPadding(ATy->getElementType(), TD); 814 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) { 815 return HasPadding(VTy->getElementType(), TD); 816 } 817 return TD.getTypeSizeInBits(Ty) != TD.getABITypeSizeInBits(Ty); 818} 819 820/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 821/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 822/// or 1 if safe after canonicalization has been performed. 823/// 824int SROA::isSafeAllocaToScalarRepl(AllocationInst *AI) { 825 // Loop over the use list of the alloca. We can only transform it if all of 826 // the users are safe to transform. 827 AllocaInfo Info; 828 829 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); 830 I != E; ++I) { 831 isSafeUseOfAllocation(cast<Instruction>(*I), AI, Info); 832 if (Info.isUnsafe) { 833 DOUT << "Cannot transform: " << *AI << " due to user: " << **I; 834 return 0; 835 } 836 } 837 838 // Okay, we know all the users are promotable. If the aggregate is a memcpy 839 // source and destination, we have to be careful. In particular, the memcpy 840 // could be moving around elements that live in structure padding of the LLVM 841 // types, but may actually be used. In these cases, we refuse to promote the 842 // struct. 843 if (Info.isMemCpySrc && Info.isMemCpyDst && 844 HasPadding(AI->getType()->getElementType(), getAnalysis<TargetData>())) 845 return 0; 846 847 // If we require cleanup, return 1, otherwise return 3. 848 return Info.needsCanon ? 1 : 3; 849} 850 851/// CanonicalizeAllocaUsers - If SROA reported that it can promote the specified 852/// allocation, but only if cleaned up, perform the cleanups required. 853void SROA::CanonicalizeAllocaUsers(AllocationInst *AI) { 854 // At this point, we know that the end result will be SROA'd and promoted, so 855 // we can insert ugly code if required so long as sroa+mem2reg will clean it 856 // up. 857 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 858 UI != E; ) { 859 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI++); 860 if (!GEPI) continue; 861 gep_type_iterator I = gep_type_begin(GEPI); 862 ++I; 863 864 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 865 uint64_t NumElements = AT->getNumElements(); 866 867 if (!isa<ConstantInt>(I.getOperand())) { 868 if (NumElements == 1) { 869 GEPI->setOperand(2, Constant::getNullValue(Type::Int32Ty)); 870 } else { 871 assert(NumElements == 2 && "Unhandled case!"); 872 // All users of the GEP must be loads. At each use of the GEP, insert 873 // two loads of the appropriate indexed GEP and select between them. 874 Value *IsOne = new ICmpInst(ICmpInst::ICMP_NE, I.getOperand(), 875 Constant::getNullValue(I.getOperand()->getType()), 876 "isone", GEPI); 877 // Insert the new GEP instructions, which are properly indexed. 878 SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end()); 879 Indices[1] = Constant::getNullValue(Type::Int32Ty); 880 Value *ZeroIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 881 Indices.begin(), 882 Indices.end(), 883 GEPI->getName()+".0", GEPI); 884 Indices[1] = ConstantInt::get(Type::Int32Ty, 1); 885 Value *OneIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 886 Indices.begin(), 887 Indices.end(), 888 GEPI->getName()+".1", GEPI); 889 // Replace all loads of the variable index GEP with loads from both 890 // indexes and a select. 891 while (!GEPI->use_empty()) { 892 LoadInst *LI = cast<LoadInst>(GEPI->use_back()); 893 Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI); 894 Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI); 895 Value *R = SelectInst::Create(IsOne, One, Zero, LI->getName(), LI); 896 LI->replaceAllUsesWith(R); 897 LI->eraseFromParent(); 898 } 899 GEPI->eraseFromParent(); 900 } 901 } 902 } 903 } 904} 905 906/// MergeInType - Add the 'In' type to the accumulated type so far. If the 907/// types are incompatible, return true, otherwise update Accum and return 908/// false. 909/// 910/// There are three cases we handle here: 911/// 1) An effectively-integer union, where the pieces are stored into as 912/// smaller integers (common with byte swap and other idioms). 913/// 2) A union of vector types of the same size and potentially its elements. 914/// Here we turn element accesses into insert/extract element operations. 915/// 3) A union of scalar types, such as int/float or int/pointer. Here we 916/// merge together into integers, allowing the xform to work with #1 as 917/// well. 918static bool MergeInType(const Type *In, const Type *&Accum, 919 const TargetData &TD) { 920 // If this is our first type, just use it. 921 const VectorType *PTy; 922 if (Accum == Type::VoidTy || In == Accum) { 923 Accum = In; 924 } else if (In == Type::VoidTy) { 925 // Noop. 926 } else if (In->isInteger() && Accum->isInteger()) { // integer union. 927 // Otherwise pick whichever type is larger. 928 if (cast<IntegerType>(In)->getBitWidth() > 929 cast<IntegerType>(Accum)->getBitWidth()) 930 Accum = In; 931 } else if (isa<PointerType>(In) && isa<PointerType>(Accum)) { 932 // Pointer unions just stay as one of the pointers. 933 } else if (isa<VectorType>(In) || isa<VectorType>(Accum)) { 934 if ((PTy = dyn_cast<VectorType>(Accum)) && 935 PTy->getElementType() == In) { 936 // Accum is a vector, and we are accessing an element: ok. 937 } else if ((PTy = dyn_cast<VectorType>(In)) && 938 PTy->getElementType() == Accum) { 939 // In is a vector, and accum is an element: ok, remember In. 940 Accum = In; 941 } else if ((PTy = dyn_cast<VectorType>(In)) && isa<VectorType>(Accum) && 942 PTy->getBitWidth() == cast<VectorType>(Accum)->getBitWidth()) { 943 // Two vectors of the same size: keep Accum. 944 } else { 945 // Cannot insert an short into a <4 x int> or handle 946 // <2 x int> -> <4 x int> 947 return true; 948 } 949 } else { 950 // Pointer/FP/Integer unions merge together as integers. 951 switch (Accum->getTypeID()) { 952 case Type::PointerTyID: Accum = TD.getIntPtrType(); break; 953 case Type::FloatTyID: Accum = Type::Int32Ty; break; 954 case Type::DoubleTyID: Accum = Type::Int64Ty; break; 955 case Type::X86_FP80TyID: return true; 956 case Type::FP128TyID: return true; 957 case Type::PPC_FP128TyID: return true; 958 default: 959 assert(Accum->isInteger() && "Unknown FP type!"); 960 break; 961 } 962 963 switch (In->getTypeID()) { 964 case Type::PointerTyID: In = TD.getIntPtrType(); break; 965 case Type::FloatTyID: In = Type::Int32Ty; break; 966 case Type::DoubleTyID: In = Type::Int64Ty; break; 967 case Type::X86_FP80TyID: return true; 968 case Type::FP128TyID: return true; 969 case Type::PPC_FP128TyID: return true; 970 default: 971 assert(In->isInteger() && "Unknown FP type!"); 972 break; 973 } 974 return MergeInType(In, Accum, TD); 975 } 976 return false; 977} 978 979/// getUIntAtLeastAsBigAs - Return an unsigned integer type that is at least 980/// as big as the specified type. If there is no suitable type, this returns 981/// null. 982const Type *getUIntAtLeastAsBigAs(unsigned NumBits) { 983 if (NumBits > 64) return 0; 984 if (NumBits > 32) return Type::Int64Ty; 985 if (NumBits > 16) return Type::Int32Ty; 986 if (NumBits > 8) return Type::Int16Ty; 987 return Type::Int8Ty; 988} 989 990/// CanConvertToScalar - V is a pointer. If we can convert the pointee to a 991/// single scalar integer type, return that type. Further, if the use is not 992/// a completely trivial use that mem2reg could promote, set IsNotTrivial. If 993/// there are no uses of this pointer, return Type::VoidTy to differentiate from 994/// failure. 995/// 996const Type *SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial) { 997 const Type *UsedType = Type::VoidTy; // No uses, no forced type. 998 const TargetData &TD = getAnalysis<TargetData>(); 999 const PointerType *PTy = cast<PointerType>(V->getType()); 1000 1001 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1002 Instruction *User = cast<Instruction>(*UI); 1003 1004 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1005 // FIXME: Loads of a first class aggregrate value could be converted to a 1006 // series of loads and insertvalues 1007 if (!LI->getType()->isSingleValueType()) 1008 return 0; 1009 1010 if (MergeInType(LI->getType(), UsedType, TD)) 1011 return 0; 1012 1013 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1014 // Storing the pointer, not into the value? 1015 if (SI->getOperand(0) == V) return 0; 1016 1017 // FIXME: Stores of a first class aggregrate value could be converted to a 1018 // series of extractvalues and stores 1019 if (!SI->getOperand(0)->getType()->isSingleValueType()) 1020 return 0; 1021 1022 // NOTE: We could handle storing of FP imms into integers here! 1023 1024 if (MergeInType(SI->getOperand(0)->getType(), UsedType, TD)) 1025 return 0; 1026 } else if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 1027 IsNotTrivial = true; 1028 const Type *SubTy = CanConvertToScalar(CI, IsNotTrivial); 1029 if (!SubTy || MergeInType(SubTy, UsedType, TD)) return 0; 1030 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1031 // Check to see if this is stepping over an element: GEP Ptr, int C 1032 if (GEP->getNumOperands() == 2 && isa<ConstantInt>(GEP->getOperand(1))) { 1033 unsigned Idx = cast<ConstantInt>(GEP->getOperand(1))->getZExtValue(); 1034 unsigned ElSize = TD.getABITypeSize(PTy->getElementType()); 1035 unsigned BitOffset = Idx*ElSize*8; 1036 if (BitOffset > 64 || !isPowerOf2_32(ElSize)) return 0; 1037 1038 IsNotTrivial = true; 1039 const Type *SubElt = CanConvertToScalar(GEP, IsNotTrivial); 1040 if (SubElt == 0) return 0; 1041 if (SubElt != Type::VoidTy && SubElt->isInteger()) { 1042 const Type *NewTy = 1043 getUIntAtLeastAsBigAs(TD.getABITypeSizeInBits(SubElt)+BitOffset); 1044 if (NewTy == 0 || MergeInType(NewTy, UsedType, TD)) return 0; 1045 continue; 1046 } 1047 } else if (GEP->getNumOperands() == 3 && 1048 isa<ConstantInt>(GEP->getOperand(1)) && 1049 isa<ConstantInt>(GEP->getOperand(2)) && 1050 cast<ConstantInt>(GEP->getOperand(1))->isZero()) { 1051 // We are stepping into an element, e.g. a structure or an array: 1052 // GEP Ptr, int 0, uint C 1053 const Type *AggTy = PTy->getElementType(); 1054 unsigned Idx = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue(); 1055 1056 if (const ArrayType *ATy = dyn_cast<ArrayType>(AggTy)) { 1057 if (Idx >= ATy->getNumElements()) return 0; // Out of range. 1058 } else if (const VectorType *VectorTy = dyn_cast<VectorType>(AggTy)) { 1059 // Getting an element of the vector. 1060 if (Idx >= VectorTy->getNumElements()) return 0; // Out of range. 1061 1062 // Merge in the vector type. 1063 if (MergeInType(VectorTy, UsedType, TD)) return 0; 1064 1065 const Type *SubTy = CanConvertToScalar(GEP, IsNotTrivial); 1066 if (SubTy == 0) return 0; 1067 1068 if (SubTy != Type::VoidTy && MergeInType(SubTy, UsedType, TD)) 1069 return 0; 1070 1071 // We'll need to change this to an insert/extract element operation. 1072 IsNotTrivial = true; 1073 continue; // Everything looks ok 1074 1075 } else if (isa<StructType>(AggTy)) { 1076 // Structs are always ok. 1077 } else { 1078 return 0; 1079 } 1080 const Type *NTy = getUIntAtLeastAsBigAs(TD.getABITypeSizeInBits(AggTy)); 1081 if (NTy == 0 || MergeInType(NTy, UsedType, TD)) return 0; 1082 const Type *SubTy = CanConvertToScalar(GEP, IsNotTrivial); 1083 if (SubTy == 0) return 0; 1084 if (SubTy != Type::VoidTy && MergeInType(SubTy, UsedType, TD)) 1085 return 0; 1086 continue; // Everything looks ok 1087 } 1088 return 0; 1089 } else { 1090 // Cannot handle this! 1091 return 0; 1092 } 1093 } 1094 1095 return UsedType; 1096} 1097 1098/// ConvertToScalar - The specified alloca passes the CanConvertToScalar 1099/// predicate and is non-trivial. Convert it to something that can be trivially 1100/// promoted into a register by mem2reg. 1101void SROA::ConvertToScalar(AllocationInst *AI, const Type *ActualTy) { 1102 DOUT << "CONVERT TO SCALAR: " << *AI << " TYPE = " 1103 << *ActualTy << "\n"; 1104 ++NumConverted; 1105 1106 BasicBlock *EntryBlock = AI->getParent(); 1107 assert(EntryBlock == &EntryBlock->getParent()->getEntryBlock() && 1108 "Not in the entry block!"); 1109 EntryBlock->getInstList().remove(AI); // Take the alloca out of the program. 1110 1111 // Create and insert the alloca. 1112 AllocaInst *NewAI = new AllocaInst(ActualTy, 0, AI->getName(), 1113 EntryBlock->begin()); 1114 ConvertUsesToScalar(AI, NewAI, 0); 1115 delete AI; 1116} 1117 1118 1119/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 1120/// directly. This happens when we are converting an "integer union" to a 1121/// single integer scalar, or when we are converting a "vector union" to a 1122/// vector with insert/extractelement instructions. 1123/// 1124/// Offset is an offset from the original alloca, in bits that need to be 1125/// shifted to the right. By the end of this, there should be no uses of Ptr. 1126void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, unsigned Offset) { 1127 while (!Ptr->use_empty()) { 1128 Instruction *User = cast<Instruction>(Ptr->use_back()); 1129 1130 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1131 Value *NV = ConvertUsesOfLoadToScalar(LI, NewAI, Offset); 1132 LI->replaceAllUsesWith(NV); 1133 LI->eraseFromParent(); 1134 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1135 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 1136 1137 Value *SV = ConvertUsesOfStoreToScalar(SI, NewAI, Offset); 1138 new StoreInst(SV, NewAI, SI); 1139 SI->eraseFromParent(); 1140 1141 } else if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 1142 ConvertUsesToScalar(CI, NewAI, Offset); 1143 CI->eraseFromParent(); 1144 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1145 const PointerType *AggPtrTy = 1146 cast<PointerType>(GEP->getOperand(0)->getType()); 1147 const TargetData &TD = getAnalysis<TargetData>(); 1148 unsigned AggSizeInBits = 1149 TD.getABITypeSizeInBits(AggPtrTy->getElementType()); 1150 1151 // Check to see if this is stepping over an element: GEP Ptr, int C 1152 unsigned NewOffset = Offset; 1153 if (GEP->getNumOperands() == 2) { 1154 unsigned Idx = cast<ConstantInt>(GEP->getOperand(1))->getZExtValue(); 1155 unsigned BitOffset = Idx*AggSizeInBits; 1156 1157 NewOffset += BitOffset; 1158 } else if (GEP->getNumOperands() == 3) { 1159 // We know that operand #2 is zero. 1160 unsigned Idx = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue(); 1161 const Type *AggTy = AggPtrTy->getElementType(); 1162 if (const SequentialType *SeqTy = dyn_cast<SequentialType>(AggTy)) { 1163 unsigned ElSizeBits = 1164 TD.getABITypeSizeInBits(SeqTy->getElementType()); 1165 1166 NewOffset += ElSizeBits*Idx; 1167 } else if (const StructType *STy = dyn_cast<StructType>(AggTy)) { 1168 unsigned EltBitOffset = 1169 TD.getStructLayout(STy)->getElementOffsetInBits(Idx); 1170 1171 NewOffset += EltBitOffset; 1172 } else { 1173 assert(0 && "Unsupported operation!"); 1174 abort(); 1175 } 1176 } else { 1177 assert(0 && "Unsupported operation!"); 1178 abort(); 1179 } 1180 ConvertUsesToScalar(GEP, NewAI, NewOffset); 1181 GEP->eraseFromParent(); 1182 } else { 1183 assert(0 && "Unsupported operation!"); 1184 abort(); 1185 } 1186 } 1187} 1188 1189/// ConvertUsesOfLoadToScalar - Convert all of the users the specified load to 1190/// use the new alloca directly, returning the value that should replace the 1191/// load. This happens when we are converting an "integer union" to a 1192/// single integer scalar, or when we are converting a "vector union" to a 1193/// vector with insert/extractelement instructions. 1194/// 1195/// Offset is an offset from the original alloca, in bits that need to be 1196/// shifted to the right. By the end of this, there should be no uses of Ptr. 1197Value *SROA::ConvertUsesOfLoadToScalar(LoadInst *LI, AllocaInst *NewAI, 1198 unsigned Offset) { 1199 // The load is a bit extract from NewAI shifted right by Offset bits. 1200 Value *NV = new LoadInst(NewAI, LI->getName(), LI); 1201 1202 if (NV->getType() == LI->getType() && Offset == 0) { 1203 // We win, no conversion needed. 1204 return NV; 1205 } 1206 1207 // If the result type of the 'union' is a pointer, then this must be ptr->ptr 1208 // cast. Anything else would result in NV being an integer. 1209 if (isa<PointerType>(NV->getType())) { 1210 assert(isa<PointerType>(LI->getType())); 1211 return new BitCastInst(NV, LI->getType(), LI->getName(), LI); 1212 } 1213 1214 if (const VectorType *VTy = dyn_cast<VectorType>(NV->getType())) { 1215 // If the result alloca is a vector type, this is either an element 1216 // access or a bitcast to another vector type. 1217 if (isa<VectorType>(LI->getType())) 1218 return new BitCastInst(NV, LI->getType(), LI->getName(), LI); 1219 1220 // Otherwise it must be an element access. 1221 const TargetData &TD = getAnalysis<TargetData>(); 1222 unsigned Elt = 0; 1223 if (Offset) { 1224 unsigned EltSize = TD.getABITypeSizeInBits(VTy->getElementType()); 1225 Elt = Offset/EltSize; 1226 Offset -= EltSize*Elt; 1227 } 1228 NV = new ExtractElementInst(NV, ConstantInt::get(Type::Int32Ty, Elt), 1229 "tmp", LI); 1230 1231 // If we're done, return this element. 1232 if (NV->getType() == LI->getType() && Offset == 0) 1233 return NV; 1234 } 1235 1236 const IntegerType *NTy = cast<IntegerType>(NV->getType()); 1237 1238 // If this is a big-endian system and the load is narrower than the 1239 // full alloca type, we need to do a shift to get the right bits. 1240 int ShAmt = 0; 1241 const TargetData &TD = getAnalysis<TargetData>(); 1242 if (TD.isBigEndian()) { 1243 // On big-endian machines, the lowest bit is stored at the bit offset 1244 // from the pointer given by getTypeStoreSizeInBits. This matters for 1245 // integers with a bitwidth that is not a multiple of 8. 1246 ShAmt = TD.getTypeStoreSizeInBits(NTy) - 1247 TD.getTypeStoreSizeInBits(LI->getType()) - Offset; 1248 } else { 1249 ShAmt = Offset; 1250 } 1251 1252 // Note: we support negative bitwidths (with shl) which are not defined. 1253 // We do this to support (f.e.) loads off the end of a structure where 1254 // only some bits are used. 1255 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 1256 NV = BinaryOperator::CreateLShr(NV, 1257 ConstantInt::get(NV->getType(),ShAmt), 1258 LI->getName(), LI); 1259 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 1260 NV = BinaryOperator::CreateShl(NV, 1261 ConstantInt::get(NV->getType(),-ShAmt), 1262 LI->getName(), LI); 1263 1264 // Finally, unconditionally truncate the integer to the right width. 1265 unsigned LIBitWidth = TD.getTypeSizeInBits(LI->getType()); 1266 if (LIBitWidth < NTy->getBitWidth()) 1267 NV = new TruncInst(NV, IntegerType::get(LIBitWidth), 1268 LI->getName(), LI); 1269 1270 // If the result is an integer, this is a trunc or bitcast. 1271 if (isa<IntegerType>(LI->getType())) { 1272 // Should be done. 1273 } else if (LI->getType()->isFloatingPoint()) { 1274 // Just do a bitcast, we know the sizes match up. 1275 NV = new BitCastInst(NV, LI->getType(), LI->getName(), LI); 1276 } else { 1277 // Otherwise must be a pointer. 1278 NV = new IntToPtrInst(NV, LI->getType(), LI->getName(), LI); 1279 } 1280 assert(NV->getType() == LI->getType() && "Didn't convert right?"); 1281 return NV; 1282} 1283 1284 1285/// ConvertUsesOfStoreToScalar - Convert the specified store to a load+store 1286/// pair of the new alloca directly, returning the value that should be stored 1287/// to the alloca. This happens when we are converting an "integer union" to a 1288/// single integer scalar, or when we are converting a "vector union" to a 1289/// vector with insert/extractelement instructions. 1290/// 1291/// Offset is an offset from the original alloca, in bits that need to be 1292/// shifted to the right. By the end of this, there should be no uses of Ptr. 1293Value *SROA::ConvertUsesOfStoreToScalar(StoreInst *SI, AllocaInst *NewAI, 1294 unsigned Offset) { 1295 1296 // Convert the stored type to the actual type, shift it left to insert 1297 // then 'or' into place. 1298 Value *SV = SI->getOperand(0); 1299 const Type *AllocaType = NewAI->getType()->getElementType(); 1300 if (SV->getType() == AllocaType && Offset == 0) { 1301 // All is well. 1302 } else if (const VectorType *PTy = dyn_cast<VectorType>(AllocaType)) { 1303 Value *Old = new LoadInst(NewAI, NewAI->getName()+".in", SI); 1304 1305 // If the result alloca is a vector type, this is either an element 1306 // access or a bitcast to another vector type. 1307 if (isa<VectorType>(SV->getType())) { 1308 SV = new BitCastInst(SV, AllocaType, SV->getName(), SI); 1309 } else { 1310 // Must be an element insertion. 1311 const TargetData &TD = getAnalysis<TargetData>(); 1312 unsigned Elt = Offset/TD.getABITypeSizeInBits(PTy->getElementType()); 1313 SV = InsertElementInst::Create(Old, SV, 1314 ConstantInt::get(Type::Int32Ty, Elt), 1315 "tmp", SI); 1316 } 1317 } else if (isa<PointerType>(AllocaType)) { 1318 // If the alloca type is a pointer, then all the elements must be 1319 // pointers. 1320 if (SV->getType() != AllocaType) 1321 SV = new BitCastInst(SV, AllocaType, SV->getName(), SI); 1322 } else { 1323 Value *Old = new LoadInst(NewAI, NewAI->getName()+".in", SI); 1324 1325 // If SV is a float, convert it to the appropriate integer type. 1326 // If it is a pointer, do the same, and also handle ptr->ptr casts 1327 // here. 1328 const TargetData &TD = getAnalysis<TargetData>(); 1329 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 1330 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); 1331 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); 1332 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); 1333 if (SV->getType()->isFloatingPoint()) 1334 SV = new BitCastInst(SV, IntegerType::get(SrcWidth), 1335 SV->getName(), SI); 1336 else if (isa<PointerType>(SV->getType())) 1337 SV = new PtrToIntInst(SV, TD.getIntPtrType(), SV->getName(), SI); 1338 1339 // Always zero extend the value if needed. 1340 if (SV->getType() != AllocaType) 1341 SV = new ZExtInst(SV, AllocaType, SV->getName(), SI); 1342 1343 // If this is a big-endian system and the store is narrower than the 1344 // full alloca type, we need to do a shift to get the right bits. 1345 int ShAmt = 0; 1346 if (TD.isBigEndian()) { 1347 // On big-endian machines, the lowest bit is stored at the bit offset 1348 // from the pointer given by getTypeStoreSizeInBits. This matters for 1349 // integers with a bitwidth that is not a multiple of 8. 1350 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 1351 } else { 1352 ShAmt = Offset; 1353 } 1354 1355 // Note: we support negative bitwidths (with shr) which are not defined. 1356 // We do this to support (f.e.) stores off the end of a structure where 1357 // only some bits in the structure are set. 1358 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 1359 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 1360 SV = BinaryOperator::CreateShl(SV, 1361 ConstantInt::get(SV->getType(), ShAmt), 1362 SV->getName(), SI); 1363 Mask <<= ShAmt; 1364 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1365 SV = BinaryOperator::CreateLShr(SV, 1366 ConstantInt::get(SV->getType(),-ShAmt), 1367 SV->getName(), SI); 1368 Mask = Mask.lshr(ShAmt); 1369 } 1370 1371 // Mask out the bits we are about to insert from the old value, and or 1372 // in the new bits. 1373 if (SrcWidth != DestWidth) { 1374 assert(DestWidth > SrcWidth); 1375 Old = BinaryOperator::CreateAnd(Old, ConstantInt::get(~Mask), 1376 Old->getName()+".mask", SI); 1377 SV = BinaryOperator::CreateOr(Old, SV, SV->getName()+".ins", SI); 1378 } 1379 } 1380 return SV; 1381} 1382 1383 1384 1385/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1386/// some part of a constant global variable. This intentionally only accepts 1387/// constant expressions because we don't can't rewrite arbitrary instructions. 1388static bool PointsToConstantGlobal(Value *V) { 1389 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1390 return GV->isConstant(); 1391 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1392 if (CE->getOpcode() == Instruction::BitCast || 1393 CE->getOpcode() == Instruction::GetElementPtr) 1394 return PointsToConstantGlobal(CE->getOperand(0)); 1395 return false; 1396} 1397 1398/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1399/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1400/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1401/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1402/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1403/// the alloca, and if the source pointer is a pointer to a constant global, we 1404/// can optimize this. 1405static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy, 1406 bool isOffset) { 1407 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1408 if (isa<LoadInst>(*UI)) { 1409 // Ignore loads, they are always ok. 1410 continue; 1411 } 1412 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) { 1413 // If uses of the bitcast are ok, we are ok. 1414 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 1415 return false; 1416 continue; 1417 } 1418 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) { 1419 // If the GEP has all zero indices, it doesn't offset the pointer. If it 1420 // doesn't, it does. 1421 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 1422 isOffset || !GEP->hasAllZeroIndices())) 1423 return false; 1424 continue; 1425 } 1426 1427 // If this is isn't our memcpy/memmove, reject it as something we can't 1428 // handle. 1429 if (!isa<MemCpyInst>(*UI) && !isa<MemMoveInst>(*UI)) 1430 return false; 1431 1432 // If we already have seen a copy, reject the second one. 1433 if (TheCopy) return false; 1434 1435 // If the pointer has been offset from the start of the alloca, we can't 1436 // safely handle this. 1437 if (isOffset) return false; 1438 1439 // If the memintrinsic isn't using the alloca as the dest, reject it. 1440 if (UI.getOperandNo() != 1) return false; 1441 1442 MemIntrinsic *MI = cast<MemIntrinsic>(*UI); 1443 1444 // If the source of the memcpy/move is not a constant global, reject it. 1445 if (!PointsToConstantGlobal(MI->getOperand(2))) 1446 return false; 1447 1448 // Otherwise, the transform is safe. Remember the copy instruction. 1449 TheCopy = MI; 1450 } 1451 return true; 1452} 1453 1454/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 1455/// modified by a copy from a constant global. If we can prove this, we can 1456/// replace any uses of the alloca with uses of the global directly. 1457Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocationInst *AI) { 1458 Instruction *TheCopy = 0; 1459 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 1460 return TheCopy; 1461 return 0; 1462} 1463