LoopStrengthReduce.cpp revision 7e608bbb5dfe4f827e64e91b0bb68a1d95d737ae
1//===- LoopStrengthReduce.cpp - Strength Reduce GEPs in Loops -------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file was developed by Nate Begeman and is distributed under the 6// University of Illinois Open Source License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This pass performs a strength reduction on array references inside loops that 11// have as one or more of their components the loop induction variable. This is 12// accomplished by creating a new Value to hold the initial value of the array 13// access for the first iteration, and then creating a new GEP instruction in 14// the loop to increment the value by the appropriate amount. 15// 16//===----------------------------------------------------------------------===// 17 18#include "llvm/Transforms/Scalar.h" 19#include "llvm/Constants.h" 20#include "llvm/Instructions.h" 21#include "llvm/Type.h" 22#include "llvm/DerivedTypes.h" 23#include "llvm/Analysis/Dominators.h" 24#include "llvm/Analysis/LoopInfo.h" 25#include "llvm/Analysis/ScalarEvolutionExpander.h" 26#include "llvm/Support/CFG.h" 27#include "llvm/Support/GetElementPtrTypeIterator.h" 28#include "llvm/Transforms/Utils/Local.h" 29#include "llvm/Target/TargetData.h" 30#include "llvm/ADT/Statistic.h" 31#include "llvm/Support/Debug.h" 32#include <algorithm> 33#include <set> 34using namespace llvm; 35 36namespace { 37 Statistic<> NumReduced ("loop-reduce", "Number of GEPs strength reduced"); 38 39 class GEPCache { 40 public: 41 GEPCache() : CachedPHINode(0), Map() {} 42 43 GEPCache *get(Value *v) { 44 std::map<Value *, GEPCache>::iterator I = Map.find(v); 45 if (I == Map.end()) 46 I = Map.insert(std::pair<Value *, GEPCache>(v, GEPCache())).first; 47 return &I->second; 48 } 49 50 PHINode *CachedPHINode; 51 std::map<Value *, GEPCache> Map; 52 }; 53 54 struct IVUse { 55 /// Users - Keep track of all of the users of this stride as well as the 56 /// initial value. 57 std::vector<std::pair<SCEVHandle, Instruction*> > Users; 58 std::vector<Instruction *> UserOperands; 59 60 void addUser(SCEVHandle &SH, Instruction *U, Instruction *V) { 61 Users.push_back(std::make_pair(SH, U)); 62 UserOperands.push_back(V); 63 } 64 }; 65 66 67 class LoopStrengthReduce : public FunctionPass { 68 LoopInfo *LI; 69 DominatorSet *DS; 70 ScalarEvolution *SE; 71 const TargetData *TD; 72 const Type *UIntPtrTy; 73 bool Changed; 74 75 /// MaxTargetAMSize - This is the maximum power-of-two scale value that the 76 /// target can handle for free with its addressing modes. 77 unsigned MaxTargetAMSize; 78 79 /// IVUsesByStride - Keep track of all uses of induction variables that we 80 /// are interested in. The key of the map is the stride of the access. 81 std::map<Value*, IVUse> IVUsesByStride; 82 83 /// CastedBasePointers - As we need to lower getelementptr instructions, we 84 /// cast the pointer input to uintptr_t. This keeps track of the casted 85 /// values for the pointers we have processed so far. 86 std::map<Value*, Value*> CastedBasePointers; 87 88 /// DeadInsts - Keep track of instructions we may have made dead, so that 89 /// we can remove them after we are done working. 90 std::set<Instruction*> DeadInsts; 91 public: 92 LoopStrengthReduce(unsigned MTAMS = 1) 93 : MaxTargetAMSize(MTAMS) { 94 } 95 96 virtual bool runOnFunction(Function &) { 97 LI = &getAnalysis<LoopInfo>(); 98 DS = &getAnalysis<DominatorSet>(); 99 SE = &getAnalysis<ScalarEvolution>(); 100 TD = &getAnalysis<TargetData>(); 101 UIntPtrTy = TD->getIntPtrType(); 102 Changed = false; 103 104 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I) 105 runOnLoop(*I); 106 return Changed; 107 } 108 109 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 110 AU.setPreservesCFG(); 111 AU.addRequiredID(LoopSimplifyID); 112 AU.addRequired<LoopInfo>(); 113 AU.addRequired<DominatorSet>(); 114 AU.addRequired<TargetData>(); 115 AU.addRequired<ScalarEvolution>(); 116 } 117 private: 118 void runOnLoop(Loop *L); 119 bool AddUsersIfInteresting(Instruction *I, Loop *L); 120 void AnalyzeGetElementPtrUsers(GetElementPtrInst *GEP, Instruction *I, 121 Loop *L); 122 123 void StrengthReduceStridedIVUsers(Value *Stride, IVUse &Uses, Loop *L, 124 bool isOnlyStride); 125 126 void strengthReduceGEP(GetElementPtrInst *GEPI, Loop *L, 127 GEPCache* GEPCache, 128 Instruction *InsertBefore, 129 std::set<Instruction*> &DeadInsts); 130 void DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts); 131 }; 132 RegisterOpt<LoopStrengthReduce> X("loop-reduce", 133 "Strength Reduce GEP Uses of Ind. Vars"); 134} 135 136FunctionPass *llvm::createLoopStrengthReducePass(unsigned MaxTargetAMSize) { 137 return new LoopStrengthReduce(MaxTargetAMSize); 138} 139 140/// DeleteTriviallyDeadInstructions - If any of the instructions is the 141/// specified set are trivially dead, delete them and see if this makes any of 142/// their operands subsequently dead. 143void LoopStrengthReduce:: 144DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts) { 145 while (!Insts.empty()) { 146 Instruction *I = *Insts.begin(); 147 Insts.erase(Insts.begin()); 148 if (isInstructionTriviallyDead(I)) { 149 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 150 if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i))) 151 Insts.insert(U); 152 I->getParent()->getInstList().erase(I); 153 Changed = true; 154 } 155 } 156} 157 158 159/// CanReduceSCEV - Return true if we can strength reduce this scalar evolution 160/// in the specified loop. 161static bool CanReduceSCEV(const SCEVHandle &SH, Loop *L) { 162 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SH); 163 if (!AddRec || AddRec->getLoop() != L) return false; 164 165 // FIXME: Generalize to non-affine IV's. 166 if (!AddRec->isAffine()) return false; 167 168 // FIXME: generalize to IV's with more complex strides (must emit stride 169 // expression outside of loop!) 170 if (isa<SCEVConstant>(AddRec->getOperand(1))) 171 return true; 172 173 // We handle steps by unsigned values, because we know we won't have to insert 174 // a cast for them. 175 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(AddRec->getOperand(1))) 176 if (SU->getValue()->getType()->isUnsigned()) 177 return true; 178 179 // Otherwise, no, we can't handle it yet. 180 return false; 181} 182 183 184/// GetAdjustedIndex - Adjust the specified GEP sequential type index to match 185/// the size of the pointer type, and scale it by the type size. 186static SCEVHandle GetAdjustedIndex(const SCEVHandle &Idx, uint64_t TySize, 187 const Type *UIntPtrTy) { 188 SCEVHandle Result = Idx; 189 if (Result->getType()->getUnsignedVersion() != UIntPtrTy) { 190 if (UIntPtrTy->getPrimitiveSize() < Result->getType()->getPrimitiveSize()) 191 Result = SCEVTruncateExpr::get(Result, UIntPtrTy); 192 else 193 Result = SCEVZeroExtendExpr::get(Result, UIntPtrTy); 194 } 195 196 // This index is scaled by the type size being indexed. 197 if (TySize != 1) 198 Result = SCEVMulExpr::get(Result, 199 SCEVConstant::get(ConstantUInt::get(UIntPtrTy, 200 TySize))); 201 return Result; 202} 203 204/// AnalyzeGetElementPtrUsers - Analyze all of the users of the specified 205/// getelementptr instruction, adding them to the IVUsesByStride table. Note 206/// that we only want to analyze a getelementptr instruction once, and it can 207/// have multiple operands that are uses of the indvar (e.g. A[i][i]). Because 208/// of this, we only process a GEP instruction if its first recurrent operand is 209/// "op", otherwise we will either have already processed it or we will sometime 210/// later. 211void LoopStrengthReduce::AnalyzeGetElementPtrUsers(GetElementPtrInst *GEP, 212 Instruction *Op, Loop *L) { 213 // Analyze all of the subscripts of this getelementptr instruction, looking 214 // for uses that are determined by the trip count of L. First, skip all 215 // operands the are not dependent on the IV. 216 217 // Build up the base expression. Insert an LLVM cast of the pointer to 218 // uintptr_t first. 219 Value *BasePtr; 220 if (Constant *CB = dyn_cast<Constant>(GEP->getOperand(0))) 221 BasePtr = ConstantExpr::getCast(CB, UIntPtrTy); 222 else { 223 Value *&BP = CastedBasePointers[GEP->getOperand(0)]; 224 if (BP == 0) { 225 BasicBlock::iterator InsertPt; 226 if (isa<Argument>(GEP->getOperand(0))) { 227 InsertPt = GEP->getParent()->getParent()->begin()->begin(); 228 } else { 229 InsertPt = cast<Instruction>(GEP->getOperand(0)); 230 if (InvokeInst *II = dyn_cast<InvokeInst>(GEP->getOperand(0))) 231 InsertPt = II->getNormalDest()->begin(); 232 else 233 ++InsertPt; 234 } 235 BP = new CastInst(GEP->getOperand(0), UIntPtrTy, 236 GEP->getOperand(0)->getName(), InsertPt); 237 } 238 BasePtr = BP; 239 } 240 241 SCEVHandle Base = SCEVUnknown::get(BasePtr); 242 243 gep_type_iterator GTI = gep_type_begin(GEP); 244 unsigned i = 1; 245 for (; GEP->getOperand(i) != Op; ++i, ++GTI) { 246 // If this is a use of a recurrence that we can analyze, and it comes before 247 // Op does in the GEP operand list, we will handle this when we process this 248 // operand. 249 if (const StructType *STy = dyn_cast<StructType>(*GTI)) { 250 const StructLayout *SL = TD->getStructLayout(STy); 251 unsigned Idx = cast<ConstantUInt>(GEP->getOperand(i))->getValue(); 252 uint64_t Offset = SL->MemberOffsets[Idx]; 253 Base = SCEVAddExpr::get(Base, SCEVUnknown::getIntegerSCEV(Offset, 254 UIntPtrTy)); 255 } else { 256 SCEVHandle Idx = SE->getSCEV(GEP->getOperand(i)); 257 258 // If this operand is reducible, and it's not the one we are looking at 259 // currently, do not process the GEP at this time. 260 if (CanReduceSCEV(Idx, L)) 261 return; 262 Base = SCEVAddExpr::get(Base, GetAdjustedIndex(Idx, 263 TD->getTypeSize(GTI.getIndexedType()), UIntPtrTy)); 264 } 265 } 266 267 // Get the index, convert it to intptr_t. 268 SCEVHandle GEPIndexExpr = 269 GetAdjustedIndex(SE->getSCEV(Op), TD->getTypeSize(GTI.getIndexedType()), 270 UIntPtrTy); 271 272 // Process all remaining subscripts in the GEP instruction. 273 for (++i, ++GTI; i != GEP->getNumOperands(); ++i, ++GTI) 274 if (const StructType *STy = dyn_cast<StructType>(*GTI)) { 275 const StructLayout *SL = TD->getStructLayout(STy); 276 unsigned Idx = cast<ConstantUInt>(GEP->getOperand(i))->getValue(); 277 uint64_t Offset = SL->MemberOffsets[Idx]; 278 Base = SCEVAddExpr::get(Base, SCEVUnknown::getIntegerSCEV(Offset, 279 UIntPtrTy)); 280 } else { 281 SCEVHandle Idx = SE->getSCEV(GEP->getOperand(i)); 282 if (CanReduceSCEV(Idx, L)) { // Another IV subscript 283 GEPIndexExpr = SCEVAddExpr::get(GEPIndexExpr, 284 GetAdjustedIndex(Idx, TD->getTypeSize(GTI.getIndexedType()), 285 UIntPtrTy)); 286 assert(CanReduceSCEV(GEPIndexExpr, L) && 287 "Cannot reduce the sum of two reducible SCEV's??"); 288 } else { 289 Base = SCEVAddExpr::get(Base, GetAdjustedIndex(Idx, 290 TD->getTypeSize(GTI.getIndexedType()), UIntPtrTy)); 291 } 292 } 293 294 assert(CanReduceSCEV(GEPIndexExpr, L) && "Non reducible idx??"); 295 296 // FIXME: If the base is not loop invariant, we currently cannot emit this. 297 if (!Base->isLoopInvariant(L)) { 298 DEBUG(std::cerr << "IGNORING GEP due to non-invaiant base: " 299 << *Base << "\n"); 300 return; 301 } 302 303 Base = SCEVAddExpr::get(Base, cast<SCEVAddRecExpr>(GEPIndexExpr)->getStart()); 304 SCEVHandle Stride = cast<SCEVAddRecExpr>(GEPIndexExpr)->getOperand(1); 305 306 DEBUG(std::cerr << "GEP BASE : " << *Base << "\n"); 307 DEBUG(std::cerr << "GEP STRIDE: " << *Stride << "\n"); 308 309 Value *Step = 0; // Step of ISE. 310 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) 311 /// Always get the step value as an unsigned value. 312 Step = ConstantExpr::getCast(SC->getValue(), 313 SC->getValue()->getType()->getUnsignedVersion()); 314 else 315 Step = cast<SCEVUnknown>(Stride)->getValue(); 316 assert(Step->getType()->isUnsigned() && "Bad step value!"); 317 318 319 // Now that we know the base and stride contributed by the GEP instruction, 320 // process all users. 321 for (Value::use_iterator UI = GEP->use_begin(), E = GEP->use_end(); 322 UI != E; ++UI) { 323 Instruction *User = cast<Instruction>(*UI); 324 325 // Do not infinitely recurse on PHI nodes. 326 if (isa<PHINode>(User) && User->getParent() == L->getHeader()) 327 continue; 328 329 // If this is an instruction defined in a nested loop, or outside this loop, 330 // don't mess with it. 331 if (LI->getLoopFor(User->getParent()) != L) 332 continue; 333 334 DEBUG(std::cerr << "FOUND USER: " << *User 335 << " OF STRIDE: " << *Step << " BASE = " << *Base << "\n"); 336 337 338 // Okay, we found a user that we cannot reduce. Analyze the instruction 339 // and decide what to do with it. 340 IVUsesByStride[Step].addUser(Base, User, GEP); 341 } 342} 343 344/// AddUsersIfInteresting - Inspect the specified instruction. If it is a 345/// reducible SCEV, recursively add its users to the IVUsesByStride set and 346/// return true. Otherwise, return false. 347bool LoopStrengthReduce::AddUsersIfInteresting(Instruction *I, Loop *L) { 348 if (I->getType() == Type::VoidTy) return false; 349 SCEVHandle ISE = SE->getSCEV(I); 350 if (!CanReduceSCEV(ISE, L)) return false; 351 352 SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(ISE); 353 SCEVHandle Start = AR->getStart(); 354 355 // Get the step value, canonicalizing to an unsigned integer type so that 356 // lookups in the map will match. 357 Value *Step = 0; // Step of ISE. 358 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(AR->getOperand(1))) 359 /// Always get the step value as an unsigned value. 360 Step = ConstantExpr::getCast(SC->getValue(), 361 SC->getValue()->getType()->getUnsignedVersion()); 362 else 363 Step = cast<SCEVUnknown>(AR->getOperand(1))->getValue(); 364 assert(Step->getType()->isUnsigned() && "Bad step value!"); 365 366 std::set<GetElementPtrInst*> AnalyzedGEPs; 367 368 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;++UI){ 369 Instruction *User = cast<Instruction>(*UI); 370 371 // Do not infinitely recurse on PHI nodes. 372 if (isa<PHINode>(User) && User->getParent() == L->getHeader()) 373 continue; 374 375 // If this is an instruction defined in a nested loop, or outside this loop, 376 // don't mess with it. 377 if (LI->getLoopFor(User->getParent()) != L) 378 continue; 379 380 // Next, see if this user is analyzable itself! 381 if (!AddUsersIfInteresting(User, L)) { 382 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 383 // If this is a getelementptr instruction, figure out what linear 384 // expression of induction variable is actually being used. 385 // 386 if (AnalyzedGEPs.insert(GEP).second) // Not already analyzed? 387 AnalyzeGetElementPtrUsers(GEP, I, L); 388 } else { 389 DEBUG(std::cerr << "FOUND USER: " << *User 390 << " OF SCEV: " << *ISE << "\n"); 391 392 // Okay, we found a user that we cannot reduce. Analyze the instruction 393 // and decide what to do with it. 394 IVUsesByStride[Step].addUser(Start, User, I); 395 } 396 } 397 } 398 return true; 399} 400 401namespace { 402 /// BasedUser - For a particular base value, keep information about how we've 403 /// partitioned the expression so far. 404 struct BasedUser { 405 /// Inst - The instruction using the induction variable. 406 Instruction *Inst; 407 408 /// Op - The value to replace with the EmittedBase. 409 Value *Op; 410 411 /// Imm - The immediate value that should be added to the base immediately 412 /// before Inst, because it will be folded into the imm field of the 413 /// instruction. 414 SCEVHandle Imm; 415 416 /// EmittedBase - The actual value* to use for the base value of this 417 /// operation. This is null if we should just use zero so far. 418 Value *EmittedBase; 419 420 BasedUser(Instruction *I, Value *V, const SCEVHandle &IMM) 421 : Inst(I), Op(V), Imm(IMM), EmittedBase(0) {} 422 423 424 // No need to compare these. 425 bool operator<(const BasedUser &BU) const { return 0; } 426 427 void dump() const; 428 }; 429} 430 431void BasedUser::dump() const { 432 std::cerr << " Imm=" << *Imm; 433 if (EmittedBase) 434 std::cerr << " EB=" << *EmittedBase; 435 436 std::cerr << " Inst: " << *Inst; 437} 438 439/// isTargetConstant - Return true if the following can be referenced by the 440/// immediate field of a target instruction. 441static bool isTargetConstant(const SCEVHandle &V) { 442 443 // FIXME: Look at the target to decide if &GV is a legal constant immediate. 444 if (isa<SCEVConstant>(V)) return true; 445 446 return false; // ENABLE this for x86 447 448 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V)) 449 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(SU->getValue())) 450 if (CE->getOpcode() == Instruction::Cast) 451 if (isa<GlobalValue>(CE->getOperand(0))) 452 // FIXME: should check to see that the dest is uintptr_t! 453 return true; 454 return false; 455} 456 457/// GetImmediateValues - Look at Val, and pull out any additions of constants 458/// that can fit into the immediate field of instructions in the target. 459static SCEVHandle GetImmediateValues(SCEVHandle Val, bool isAddress) { 460 if (!isAddress) 461 return SCEVUnknown::getIntegerSCEV(0, Val->getType()); 462 if (isTargetConstant(Val)) 463 return Val; 464 465 SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val); 466 if (SAE) { 467 unsigned i = 0; 468 for (; i != SAE->getNumOperands(); ++i) 469 if (isTargetConstant(SAE->getOperand(i))) { 470 SCEVHandle ImmVal = SAE->getOperand(i); 471 472 // If there are any other immediates that we can handle here, pull them 473 // out too. 474 for (++i; i != SAE->getNumOperands(); ++i) 475 if (isTargetConstant(SAE->getOperand(i))) 476 ImmVal = SCEVAddExpr::get(ImmVal, SAE->getOperand(i)); 477 return ImmVal; 478 } 479 } 480 481 return SCEVUnknown::getIntegerSCEV(0, Val->getType()); 482} 483 484/// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single 485/// stride of IV. All of the users may have different starting values, and this 486/// may not be the only stride (we know it is if isOnlyStride is true). 487void LoopStrengthReduce::StrengthReduceStridedIVUsers(Value *Stride, 488 IVUse &Uses, Loop *L, 489 bool isOnlyStride) { 490 // Transform our list of users and offsets to a bit more complex table. In 491 // this new vector, the first entry for each element is the base of the 492 // strided access, and the second is the BasedUser object for the use. We 493 // progressively move information from the first to the second entry, until we 494 // eventually emit the object. 495 std::vector<std::pair<SCEVHandle, BasedUser> > UsersToProcess; 496 UsersToProcess.reserve(Uses.Users.size()); 497 498 SCEVHandle ZeroBase = SCEVUnknown::getIntegerSCEV(0, 499 Uses.Users[0].first->getType()); 500 501 for (unsigned i = 0, e = Uses.Users.size(); i != e; ++i) 502 UsersToProcess.push_back(std::make_pair(Uses.Users[i].first, 503 BasedUser(Uses.Users[i].second, 504 Uses.UserOperands[i], 505 ZeroBase))); 506 507 // First pass, figure out what we can represent in the immediate fields of 508 // instructions. If we can represent anything there, move it to the imm 509 // fields of the BasedUsers. 510 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) { 511 bool isAddress = isa<LoadInst>(UsersToProcess[i].second.Inst) || 512 isa<StoreInst>(UsersToProcess[i].second.Inst); 513 UsersToProcess[i].second.Imm = GetImmediateValues(UsersToProcess[i].first, 514 isAddress); 515 UsersToProcess[i].first = SCEV::getMinusSCEV(UsersToProcess[i].first, 516 UsersToProcess[i].second.Imm); 517 518 DEBUG(std::cerr << "BASE: " << *UsersToProcess[i].first); 519 DEBUG(UsersToProcess[i].second.dump()); 520 } 521 522 SCEVExpander Rewriter(*SE, *LI); 523 BasicBlock *Preheader = L->getLoopPreheader(); 524 Instruction *PreInsertPt = Preheader->getTerminator(); 525 Instruction *PhiInsertBefore = L->getHeader()->begin(); 526 527 assert(isa<PHINode>(PhiInsertBefore) && 528 "How could this loop have IV's without any phis?"); 529 PHINode *SomeLoopPHI = cast<PHINode>(PhiInsertBefore); 530 assert(SomeLoopPHI->getNumIncomingValues() == 2 && 531 "This loop isn't canonicalized right"); 532 BasicBlock *LatchBlock = 533 SomeLoopPHI->getIncomingBlock(SomeLoopPHI->getIncomingBlock(0) == Preheader); 534 535 // FIXME: This loop needs increasing levels of intelligence. 536 // STAGE 0: just emit everything as its own base. <-- We are here 537 // STAGE 1: factor out common vars from bases, and try and push resulting 538 // constants into Imm field. 539 // STAGE 2: factor out large constants to try and make more constants 540 // acceptable for target loads and stores. 541 std::sort(UsersToProcess.begin(), UsersToProcess.end()); 542 543 while (!UsersToProcess.empty()) { 544 // Create a new Phi for this base, and stick it in the loop header. 545 Value *Replaced = UsersToProcess.front().second.Op; 546 const Type *ReplacedTy = Replaced->getType(); 547 PHINode *NewPHI = new PHINode(ReplacedTy, Replaced->getName()+".str", 548 PhiInsertBefore); 549 550 // Emit the initial base value into the loop preheader, and add it to the 551 // Phi node. 552 Value *BaseV = Rewriter.expandCodeFor(UsersToProcess.front().first, 553 PreInsertPt, ReplacedTy); 554 NewPHI->addIncoming(BaseV, Preheader); 555 556 // Emit the increment of the base value before the terminator of the loop 557 // latch block, and add it to the Phi node. 558 SCEVHandle Inc = SCEVAddExpr::get(SCEVUnknown::get(NewPHI), 559 SCEVUnknown::get(Stride)); 560 561 Value *IncV = Rewriter.expandCodeFor(Inc, LatchBlock->getTerminator(), 562 ReplacedTy); 563 IncV->setName(NewPHI->getName()+".inc"); 564 NewPHI->addIncoming(IncV, LatchBlock); 565 566 // Emit the code to add the immediate offset to the Phi value, just before 567 // the instruction that we identified as using this stride and base. 568 // First, empty the SCEVExpander's expression map so that we are guaranteed 569 // to have the code emitted where we expect it. 570 Rewriter.clear(); 571 SCEVHandle NewValSCEV = SCEVAddExpr::get(SCEVUnknown::get(NewPHI), 572 UsersToProcess.front().second.Imm); 573 Value *newVal = Rewriter.expandCodeFor(NewValSCEV, 574 UsersToProcess.front().second.Inst, 575 ReplacedTy); 576 577 // Replace the use of the operand Value with the new Phi we just created. 578 DEBUG(std::cerr << "REPLACING: " << *Replaced << "IN: " << 579 *UsersToProcess.front().second.Inst << "WITH: "<< *newVal << '\n'); 580 UsersToProcess.front().second.Inst->replaceUsesOfWith(Replaced, newVal); 581 582 // Mark old value we replaced as possibly dead, so that it is elminated 583 // if we just replaced the last use of that value. 584 DeadInsts.insert(cast<Instruction>(Replaced)); 585 586 UsersToProcess.erase(UsersToProcess.begin()); 587 ++NumReduced; 588 589 // TODO: Next, find out which base index is the most common, pull it out. 590 } 591 592 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but 593 // different starting values, into different PHIs. 594 595 // BEFORE writing this, it's probably useful to handle GEP's. 596 597 // NOTE: pull all constants together, for REG+IMM addressing, include &GV in 598 // 'IMM' if the target supports it. 599} 600 601 602void LoopStrengthReduce::runOnLoop(Loop *L) { 603 // First step, transform all loops nesting inside of this loop. 604 for (LoopInfo::iterator I = L->begin(), E = L->end(); I != E; ++I) 605 runOnLoop(*I); 606 607 // Next, find all uses of induction variables in this loop, and catagorize 608 // them by stride. Start by finding all of the PHI nodes in the header for 609 // this loop. If they are induction variables, inspect their uses. 610 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) 611 AddUsersIfInteresting(I, L); 612 613 // If we have nothing to do, return. 614 //if (IVUsesByStride.empty()) return; 615 616 // FIXME: We can widen subreg IV's here for RISC targets. e.g. instead of 617 // doing computation in byte values, promote to 32-bit values if safe. 618 619 // FIXME: Attempt to reuse values across multiple IV's. In particular, we 620 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should be 621 // codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC. Need 622 // to be careful that IV's are all the same type. Only works for intptr_t 623 // indvars. 624 625 // If we only have one stride, we can more aggressively eliminate some things. 626 bool HasOneStride = IVUsesByStride.size() == 1; 627 628 for (std::map<Value*, IVUse>::iterator SI = IVUsesByStride.begin(), 629 E = IVUsesByStride.end(); SI != E; ++SI) 630 StrengthReduceStridedIVUsers(SI->first, SI->second, L, HasOneStride); 631 632 // Clean up after ourselves 633 if (!DeadInsts.empty()) { 634 DeleteTriviallyDeadInstructions(DeadInsts); 635 636 BasicBlock::iterator I = L->getHeader()->begin(); 637 PHINode *PN; 638 while ((PN = dyn_cast<PHINode>(I))) { 639 ++I; // Preincrement iterator to avoid invalidating it when deleting PN. 640 641 // At this point, we know that we have killed one or more GEP instructions. 642 // It is worth checking to see if the cann indvar is also dead, so that we 643 // can remove it as well. The requirements for the cann indvar to be 644 // considered dead are: 645 // 1. the cann indvar has one use 646 // 2. the use is an add instruction 647 // 3. the add has one use 648 // 4. the add is used by the cann indvar 649 // If all four cases above are true, then we can remove both the add and 650 // the cann indvar. 651 // FIXME: this needs to eliminate an induction variable even if it's being 652 // compared against some value to decide loop termination. 653 if (PN->hasOneUse()) { 654 BinaryOperator *BO = dyn_cast<BinaryOperator>(*(PN->use_begin())); 655 if (BO && BO->hasOneUse()) { 656 if (PN == *(BO->use_begin())) { 657 DeadInsts.insert(BO); 658 // Break the cycle, then delete the PHI. 659 PN->replaceAllUsesWith(UndefValue::get(PN->getType())); 660 PN->eraseFromParent(); 661 } 662 } 663 } 664 } 665 DeleteTriviallyDeadInstructions(DeadInsts); 666 } 667 668 IVUsesByStride.clear(); 669 CastedBasePointers.clear(); 670 return; 671} 672