IndVarSimplify.cpp revision ecb35ece5c42d89057da3c2c7bc2e95f08b1dbef
1//===- IndVarSimplify.cpp - Induction Variable Elimination ----------------===// 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 analyzes and transforms the induction variables (and 11// computations derived from them) into simpler forms suitable for subsequent 12// analysis and transformation. 13// 14// If the trip count of a loop is computable, this pass also makes the following 15// changes: 16// 1. The exit condition for the loop is canonicalized to compare the 17// induction value against the exit value. This turns loops like: 18// 'for (i = 7; i*i < 1000; ++i)' into 'for (i = 0; i != 25; ++i)' 19// 2. Any use outside of the loop of an expression derived from the indvar 20// is changed to compute the derived value outside of the loop, eliminating 21// the dependence on the exit value of the induction variable. If the only 22// purpose of the loop is to compute the exit value of some derived 23// expression, this transformation will make the loop dead. 24// 25//===----------------------------------------------------------------------===// 26 27#define DEBUG_TYPE "indvars" 28#include "llvm/Transforms/Scalar.h" 29#include "llvm/BasicBlock.h" 30#include "llvm/Constants.h" 31#include "llvm/Instructions.h" 32#include "llvm/IntrinsicInst.h" 33#include "llvm/LLVMContext.h" 34#include "llvm/Type.h" 35#include "llvm/Analysis/Dominators.h" 36#include "llvm/Analysis/IVUsers.h" 37#include "llvm/Analysis/ScalarEvolutionExpander.h" 38#include "llvm/Analysis/LoopInfo.h" 39#include "llvm/Analysis/LoopPass.h" 40#include "llvm/Support/CFG.h" 41#include "llvm/Support/CommandLine.h" 42#include "llvm/Support/Debug.h" 43#include "llvm/Support/raw_ostream.h" 44#include "llvm/Transforms/Utils/Local.h" 45#include "llvm/Transforms/Utils/BasicBlockUtils.h" 46#include "llvm/Transforms/Utils/SimplifyIndVar.h" 47#include "llvm/Target/TargetData.h" 48#include "llvm/ADT/DenseMap.h" 49#include "llvm/ADT/SmallVector.h" 50#include "llvm/ADT/Statistic.h" 51using namespace llvm; 52 53STATISTIC(NumRemoved , "Number of aux indvars removed"); 54STATISTIC(NumWidened , "Number of indvars widened"); 55STATISTIC(NumInserted , "Number of canonical indvars added"); 56STATISTIC(NumReplaced , "Number of exit values replaced"); 57STATISTIC(NumLFTR , "Number of loop exit tests replaced"); 58STATISTIC(NumElimExt , "Number of IV sign/zero extends eliminated"); 59STATISTIC(NumElimIV , "Number of congruent IVs eliminated"); 60 61static cl::opt<bool> EnableIVRewrite( 62 "enable-iv-rewrite", cl::Hidden, 63 cl::desc("Enable canonical induction variable rewriting")); 64 65// Trip count verification can be enabled by default under NDEBUG if we 66// implement a strong expression equivalence checker in SCEV. Until then, we 67// use the verify-indvars flag, which may assert in some cases. 68static cl::opt<bool> VerifyIndvars( 69 "verify-indvars", cl::Hidden, 70 cl::desc("Verify the ScalarEvolution result after running indvars")); 71 72namespace { 73 class IndVarSimplify : public LoopPass { 74 IVUsers *IU; 75 LoopInfo *LI; 76 ScalarEvolution *SE; 77 DominatorTree *DT; 78 TargetData *TD; 79 80 SmallVector<WeakVH, 16> DeadInsts; 81 bool Changed; 82 public: 83 84 static char ID; // Pass identification, replacement for typeid 85 IndVarSimplify() : LoopPass(ID), IU(0), LI(0), SE(0), DT(0), TD(0), 86 Changed(false) { 87 initializeIndVarSimplifyPass(*PassRegistry::getPassRegistry()); 88 } 89 90 virtual bool runOnLoop(Loop *L, LPPassManager &LPM); 91 92 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 93 AU.addRequired<DominatorTree>(); 94 AU.addRequired<LoopInfo>(); 95 AU.addRequired<ScalarEvolution>(); 96 AU.addRequiredID(LoopSimplifyID); 97 AU.addRequiredID(LCSSAID); 98 if (EnableIVRewrite) 99 AU.addRequired<IVUsers>(); 100 AU.addPreserved<ScalarEvolution>(); 101 AU.addPreservedID(LoopSimplifyID); 102 AU.addPreservedID(LCSSAID); 103 if (EnableIVRewrite) 104 AU.addPreserved<IVUsers>(); 105 AU.setPreservesCFG(); 106 } 107 108 private: 109 virtual void releaseMemory() { 110 DeadInsts.clear(); 111 } 112 113 bool isValidRewrite(Value *FromVal, Value *ToVal); 114 115 void HandleFloatingPointIV(Loop *L, PHINode *PH); 116 void RewriteNonIntegerIVs(Loop *L); 117 118 void SimplifyAndExtend(Loop *L, SCEVExpander &Rewriter, LPPassManager &LPM); 119 120 void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter); 121 122 void RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter); 123 124 Value *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount, 125 PHINode *IndVar, SCEVExpander &Rewriter); 126 127 void SinkUnusedInvariants(Loop *L); 128 }; 129} 130 131char IndVarSimplify::ID = 0; 132INITIALIZE_PASS_BEGIN(IndVarSimplify, "indvars", 133 "Induction Variable Simplification", false, false) 134INITIALIZE_PASS_DEPENDENCY(DominatorTree) 135INITIALIZE_PASS_DEPENDENCY(LoopInfo) 136INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 137INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 138INITIALIZE_PASS_DEPENDENCY(LCSSA) 139INITIALIZE_PASS_DEPENDENCY(IVUsers) 140INITIALIZE_PASS_END(IndVarSimplify, "indvars", 141 "Induction Variable Simplification", false, false) 142 143Pass *llvm::createIndVarSimplifyPass() { 144 return new IndVarSimplify(); 145} 146 147/// isValidRewrite - Return true if the SCEV expansion generated by the 148/// rewriter can replace the original value. SCEV guarantees that it 149/// produces the same value, but the way it is produced may be illegal IR. 150/// Ideally, this function will only be called for verification. 151bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) { 152 // If an SCEV expression subsumed multiple pointers, its expansion could 153 // reassociate the GEP changing the base pointer. This is illegal because the 154 // final address produced by a GEP chain must be inbounds relative to its 155 // underlying object. Otherwise basic alias analysis, among other things, 156 // could fail in a dangerous way. Ultimately, SCEV will be improved to avoid 157 // producing an expression involving multiple pointers. Until then, we must 158 // bail out here. 159 // 160 // Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject 161 // because it understands lcssa phis while SCEV does not. 162 Value *FromPtr = FromVal; 163 Value *ToPtr = ToVal; 164 if (GEPOperator *GEP = dyn_cast<GEPOperator>(FromVal)) { 165 FromPtr = GEP->getPointerOperand(); 166 } 167 if (GEPOperator *GEP = dyn_cast<GEPOperator>(ToVal)) { 168 ToPtr = GEP->getPointerOperand(); 169 } 170 if (FromPtr != FromVal || ToPtr != ToVal) { 171 // Quickly check the common case 172 if (FromPtr == ToPtr) 173 return true; 174 175 // SCEV may have rewritten an expression that produces the GEP's pointer 176 // operand. That's ok as long as the pointer operand has the same base 177 // pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the 178 // base of a recurrence. This handles the case in which SCEV expansion 179 // converts a pointer type recurrence into a nonrecurrent pointer base 180 // indexed by an integer recurrence. 181 const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr)); 182 const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr)); 183 if (FromBase == ToBase) 184 return true; 185 186 DEBUG(dbgs() << "INDVARS: GEP rewrite bail out " 187 << *FromBase << " != " << *ToBase << "\n"); 188 189 return false; 190 } 191 return true; 192} 193 194/// Determine the insertion point for this user. By default, insert immediately 195/// before the user. SCEVExpander or LICM will hoist loop invariants out of the 196/// loop. For PHI nodes, there may be multiple uses, so compute the nearest 197/// common dominator for the incoming blocks. 198static Instruction *getInsertPointForUses(Instruction *User, Value *Def, 199 DominatorTree *DT) { 200 PHINode *PHI = dyn_cast<PHINode>(User); 201 if (!PHI) 202 return User; 203 204 Instruction *InsertPt = 0; 205 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) { 206 if (PHI->getIncomingValue(i) != Def) 207 continue; 208 209 BasicBlock *InsertBB = PHI->getIncomingBlock(i); 210 if (!InsertPt) { 211 InsertPt = InsertBB->getTerminator(); 212 continue; 213 } 214 InsertBB = DT->findNearestCommonDominator(InsertPt->getParent(), InsertBB); 215 InsertPt = InsertBB->getTerminator(); 216 } 217 assert(InsertPt && "Missing phi operand"); 218 assert((!isa<Instruction>(Def) || 219 DT->dominates(cast<Instruction>(Def), InsertPt)) && 220 "def does not dominate all uses"); 221 return InsertPt; 222} 223 224//===----------------------------------------------------------------------===// 225// RewriteNonIntegerIVs and helpers. Prefer integer IVs. 226//===----------------------------------------------------------------------===// 227 228/// ConvertToSInt - Convert APF to an integer, if possible. 229static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) { 230 bool isExact = false; 231 if (&APF.getSemantics() == &APFloat::PPCDoubleDouble) 232 return false; 233 // See if we can convert this to an int64_t 234 uint64_t UIntVal; 235 if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero, 236 &isExact) != APFloat::opOK || !isExact) 237 return false; 238 IntVal = UIntVal; 239 return true; 240} 241 242/// HandleFloatingPointIV - If the loop has floating induction variable 243/// then insert corresponding integer induction variable if possible. 244/// For example, 245/// for(double i = 0; i < 10000; ++i) 246/// bar(i) 247/// is converted into 248/// for(int i = 0; i < 10000; ++i) 249/// bar((double)i); 250/// 251void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) { 252 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0)); 253 unsigned BackEdge = IncomingEdge^1; 254 255 // Check incoming value. 256 ConstantFP *InitValueVal = 257 dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge)); 258 259 int64_t InitValue; 260 if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue)) 261 return; 262 263 // Check IV increment. Reject this PN if increment operation is not 264 // an add or increment value can not be represented by an integer. 265 BinaryOperator *Incr = 266 dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge)); 267 if (Incr == 0 || Incr->getOpcode() != Instruction::FAdd) return; 268 269 // If this is not an add of the PHI with a constantfp, or if the constant fp 270 // is not an integer, bail out. 271 ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1)); 272 int64_t IncValue; 273 if (IncValueVal == 0 || Incr->getOperand(0) != PN || 274 !ConvertToSInt(IncValueVal->getValueAPF(), IncValue)) 275 return; 276 277 // Check Incr uses. One user is PN and the other user is an exit condition 278 // used by the conditional terminator. 279 Value::use_iterator IncrUse = Incr->use_begin(); 280 Instruction *U1 = cast<Instruction>(*IncrUse++); 281 if (IncrUse == Incr->use_end()) return; 282 Instruction *U2 = cast<Instruction>(*IncrUse++); 283 if (IncrUse != Incr->use_end()) return; 284 285 // Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't 286 // only used by a branch, we can't transform it. 287 FCmpInst *Compare = dyn_cast<FCmpInst>(U1); 288 if (!Compare) 289 Compare = dyn_cast<FCmpInst>(U2); 290 if (Compare == 0 || !Compare->hasOneUse() || 291 !isa<BranchInst>(Compare->use_back())) 292 return; 293 294 BranchInst *TheBr = cast<BranchInst>(Compare->use_back()); 295 296 // We need to verify that the branch actually controls the iteration count 297 // of the loop. If not, the new IV can overflow and no one will notice. 298 // The branch block must be in the loop and one of the successors must be out 299 // of the loop. 300 assert(TheBr->isConditional() && "Can't use fcmp if not conditional"); 301 if (!L->contains(TheBr->getParent()) || 302 (L->contains(TheBr->getSuccessor(0)) && 303 L->contains(TheBr->getSuccessor(1)))) 304 return; 305 306 307 // If it isn't a comparison with an integer-as-fp (the exit value), we can't 308 // transform it. 309 ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1)); 310 int64_t ExitValue; 311 if (ExitValueVal == 0 || 312 !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue)) 313 return; 314 315 // Find new predicate for integer comparison. 316 CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE; 317 switch (Compare->getPredicate()) { 318 default: return; // Unknown comparison. 319 case CmpInst::FCMP_OEQ: 320 case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break; 321 case CmpInst::FCMP_ONE: 322 case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break; 323 case CmpInst::FCMP_OGT: 324 case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break; 325 case CmpInst::FCMP_OGE: 326 case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break; 327 case CmpInst::FCMP_OLT: 328 case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break; 329 case CmpInst::FCMP_OLE: 330 case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break; 331 } 332 333 // We convert the floating point induction variable to a signed i32 value if 334 // we can. This is only safe if the comparison will not overflow in a way 335 // that won't be trapped by the integer equivalent operations. Check for this 336 // now. 337 // TODO: We could use i64 if it is native and the range requires it. 338 339 // The start/stride/exit values must all fit in signed i32. 340 if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue)) 341 return; 342 343 // If not actually striding (add x, 0.0), avoid touching the code. 344 if (IncValue == 0) 345 return; 346 347 // Positive and negative strides have different safety conditions. 348 if (IncValue > 0) { 349 // If we have a positive stride, we require the init to be less than the 350 // exit value. 351 if (InitValue >= ExitValue) 352 return; 353 354 uint32_t Range = uint32_t(ExitValue-InitValue); 355 // Check for infinite loop, either: 356 // while (i <= Exit) or until (i > Exit) 357 if (NewPred == CmpInst::ICMP_SLE || NewPred == CmpInst::ICMP_SGT) { 358 if (++Range == 0) return; // Range overflows. 359 } 360 361 unsigned Leftover = Range % uint32_t(IncValue); 362 363 // If this is an equality comparison, we require that the strided value 364 // exactly land on the exit value, otherwise the IV condition will wrap 365 // around and do things the fp IV wouldn't. 366 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) && 367 Leftover != 0) 368 return; 369 370 // If the stride would wrap around the i32 before exiting, we can't 371 // transform the IV. 372 if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue) 373 return; 374 375 } else { 376 // If we have a negative stride, we require the init to be greater than the 377 // exit value. 378 if (InitValue <= ExitValue) 379 return; 380 381 uint32_t Range = uint32_t(InitValue-ExitValue); 382 // Check for infinite loop, either: 383 // while (i >= Exit) or until (i < Exit) 384 if (NewPred == CmpInst::ICMP_SGE || NewPred == CmpInst::ICMP_SLT) { 385 if (++Range == 0) return; // Range overflows. 386 } 387 388 unsigned Leftover = Range % uint32_t(-IncValue); 389 390 // If this is an equality comparison, we require that the strided value 391 // exactly land on the exit value, otherwise the IV condition will wrap 392 // around and do things the fp IV wouldn't. 393 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) && 394 Leftover != 0) 395 return; 396 397 // If the stride would wrap around the i32 before exiting, we can't 398 // transform the IV. 399 if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue) 400 return; 401 } 402 403 IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext()); 404 405 // Insert new integer induction variable. 406 PHINode *NewPHI = PHINode::Create(Int32Ty, 2, PN->getName()+".int", PN); 407 NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue), 408 PN->getIncomingBlock(IncomingEdge)); 409 410 Value *NewAdd = 411 BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue), 412 Incr->getName()+".int", Incr); 413 NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge)); 414 415 ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd, 416 ConstantInt::get(Int32Ty, ExitValue), 417 Compare->getName()); 418 419 // In the following deletions, PN may become dead and may be deleted. 420 // Use a WeakVH to observe whether this happens. 421 WeakVH WeakPH = PN; 422 423 // Delete the old floating point exit comparison. The branch starts using the 424 // new comparison. 425 NewCompare->takeName(Compare); 426 Compare->replaceAllUsesWith(NewCompare); 427 RecursivelyDeleteTriviallyDeadInstructions(Compare); 428 429 // Delete the old floating point increment. 430 Incr->replaceAllUsesWith(UndefValue::get(Incr->getType())); 431 RecursivelyDeleteTriviallyDeadInstructions(Incr); 432 433 // If the FP induction variable still has uses, this is because something else 434 // in the loop uses its value. In order to canonicalize the induction 435 // variable, we chose to eliminate the IV and rewrite it in terms of an 436 // int->fp cast. 437 // 438 // We give preference to sitofp over uitofp because it is faster on most 439 // platforms. 440 if (WeakPH) { 441 Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv", 442 PN->getParent()->getFirstInsertionPt()); 443 PN->replaceAllUsesWith(Conv); 444 RecursivelyDeleteTriviallyDeadInstructions(PN); 445 } 446 447 // Add a new IVUsers entry for the newly-created integer PHI. 448 if (IU) 449 IU->AddUsersIfInteresting(NewPHI); 450 451 Changed = true; 452} 453 454void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) { 455 // First step. Check to see if there are any floating-point recurrences. 456 // If there are, change them into integer recurrences, permitting analysis by 457 // the SCEV routines. 458 // 459 BasicBlock *Header = L->getHeader(); 460 461 SmallVector<WeakVH, 8> PHIs; 462 for (BasicBlock::iterator I = Header->begin(); 463 PHINode *PN = dyn_cast<PHINode>(I); ++I) 464 PHIs.push_back(PN); 465 466 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) 467 if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i])) 468 HandleFloatingPointIV(L, PN); 469 470 // If the loop previously had floating-point IV, ScalarEvolution 471 // may not have been able to compute a trip count. Now that we've done some 472 // re-writing, the trip count may be computable. 473 if (Changed) 474 SE->forgetLoop(L); 475} 476 477//===----------------------------------------------------------------------===// 478// RewriteLoopExitValues - Optimize IV users outside the loop. 479// As a side effect, reduces the amount of IV processing within the loop. 480//===----------------------------------------------------------------------===// 481 482/// RewriteLoopExitValues - Check to see if this loop has a computable 483/// loop-invariant execution count. If so, this means that we can compute the 484/// final value of any expressions that are recurrent in the loop, and 485/// substitute the exit values from the loop into any instructions outside of 486/// the loop that use the final values of the current expressions. 487/// 488/// This is mostly redundant with the regular IndVarSimplify activities that 489/// happen later, except that it's more powerful in some cases, because it's 490/// able to brute-force evaluate arbitrary instructions as long as they have 491/// constant operands at the beginning of the loop. 492void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) { 493 // Verify the input to the pass in already in LCSSA form. 494 assert(L->isLCSSAForm(*DT)); 495 496 SmallVector<BasicBlock*, 8> ExitBlocks; 497 L->getUniqueExitBlocks(ExitBlocks); 498 499 // Find all values that are computed inside the loop, but used outside of it. 500 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan 501 // the exit blocks of the loop to find them. 502 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 503 BasicBlock *ExitBB = ExitBlocks[i]; 504 505 // If there are no PHI nodes in this exit block, then no values defined 506 // inside the loop are used on this path, skip it. 507 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin()); 508 if (!PN) continue; 509 510 unsigned NumPreds = PN->getNumIncomingValues(); 511 512 // Iterate over all of the PHI nodes. 513 BasicBlock::iterator BBI = ExitBB->begin(); 514 while ((PN = dyn_cast<PHINode>(BBI++))) { 515 if (PN->use_empty()) 516 continue; // dead use, don't replace it 517 518 // SCEV only supports integer expressions for now. 519 if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy()) 520 continue; 521 522 // It's necessary to tell ScalarEvolution about this explicitly so that 523 // it can walk the def-use list and forget all SCEVs, as it may not be 524 // watching the PHI itself. Once the new exit value is in place, there 525 // may not be a def-use connection between the loop and every instruction 526 // which got a SCEVAddRecExpr for that loop. 527 SE->forgetValue(PN); 528 529 // Iterate over all of the values in all the PHI nodes. 530 for (unsigned i = 0; i != NumPreds; ++i) { 531 // If the value being merged in is not integer or is not defined 532 // in the loop, skip it. 533 Value *InVal = PN->getIncomingValue(i); 534 if (!isa<Instruction>(InVal)) 535 continue; 536 537 // If this pred is for a subloop, not L itself, skip it. 538 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L) 539 continue; // The Block is in a subloop, skip it. 540 541 // Check that InVal is defined in the loop. 542 Instruction *Inst = cast<Instruction>(InVal); 543 if (!L->contains(Inst)) 544 continue; 545 546 // Okay, this instruction has a user outside of the current loop 547 // and varies predictably *inside* the loop. Evaluate the value it 548 // contains when the loop exits, if possible. 549 const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop()); 550 if (!SE->isLoopInvariant(ExitValue, L)) 551 continue; 552 553 Value *ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), Inst); 554 555 DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n' 556 << " LoopVal = " << *Inst << "\n"); 557 558 if (!isValidRewrite(Inst, ExitVal)) { 559 DeadInsts.push_back(ExitVal); 560 continue; 561 } 562 Changed = true; 563 ++NumReplaced; 564 565 PN->setIncomingValue(i, ExitVal); 566 567 // If this instruction is dead now, delete it. 568 RecursivelyDeleteTriviallyDeadInstructions(Inst); 569 570 if (NumPreds == 1) { 571 // Completely replace a single-pred PHI. This is safe, because the 572 // NewVal won't be variant in the loop, so we don't need an LCSSA phi 573 // node anymore. 574 PN->replaceAllUsesWith(ExitVal); 575 RecursivelyDeleteTriviallyDeadInstructions(PN); 576 } 577 } 578 if (NumPreds != 1) { 579 // Clone the PHI and delete the original one. This lets IVUsers and 580 // any other maps purge the original user from their records. 581 PHINode *NewPN = cast<PHINode>(PN->clone()); 582 NewPN->takeName(PN); 583 NewPN->insertBefore(PN); 584 PN->replaceAllUsesWith(NewPN); 585 PN->eraseFromParent(); 586 } 587 } 588 } 589 590 // The insertion point instruction may have been deleted; clear it out 591 // so that the rewriter doesn't trip over it later. 592 Rewriter.clearInsertPoint(); 593} 594 595//===----------------------------------------------------------------------===// 596// Rewrite IV users based on a canonical IV. 597// Only for use with -enable-iv-rewrite. 598//===----------------------------------------------------------------------===// 599 600/// FIXME: It is an extremely bad idea to indvar substitute anything more 601/// complex than affine induction variables. Doing so will put expensive 602/// polynomial evaluations inside of the loop, and the str reduction pass 603/// currently can only reduce affine polynomials. For now just disable 604/// indvar subst on anything more complex than an affine addrec, unless 605/// it can be expanded to a trivial value. 606static bool isSafe(const SCEV *S, const Loop *L, ScalarEvolution *SE) { 607 // Loop-invariant values are safe. 608 if (SE->isLoopInvariant(S, L)) return true; 609 610 // Affine addrecs are safe. Non-affine are not, because LSR doesn't know how 611 // to transform them into efficient code. 612 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) 613 return AR->isAffine(); 614 615 // An add is safe it all its operands are safe. 616 if (const SCEVCommutativeExpr *Commutative 617 = dyn_cast<SCEVCommutativeExpr>(S)) { 618 for (SCEVCommutativeExpr::op_iterator I = Commutative->op_begin(), 619 E = Commutative->op_end(); I != E; ++I) 620 if (!isSafe(*I, L, SE)) return false; 621 return true; 622 } 623 624 // A cast is safe if its operand is. 625 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) 626 return isSafe(C->getOperand(), L, SE); 627 628 // A udiv is safe if its operands are. 629 if (const SCEVUDivExpr *UD = dyn_cast<SCEVUDivExpr>(S)) 630 return isSafe(UD->getLHS(), L, SE) && 631 isSafe(UD->getRHS(), L, SE); 632 633 // SCEVUnknown is always safe. 634 if (isa<SCEVUnknown>(S)) 635 return true; 636 637 // Nothing else is safe. 638 return false; 639} 640 641void IndVarSimplify::RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter) { 642 // Rewrite all induction variable expressions in terms of the canonical 643 // induction variable. 644 // 645 // If there were induction variables of other sizes or offsets, manually 646 // add the offsets to the primary induction variable and cast, avoiding 647 // the need for the code evaluation methods to insert induction variables 648 // of different sizes. 649 for (IVUsers::iterator UI = IU->begin(), E = IU->end(); UI != E; ++UI) { 650 Value *Op = UI->getOperandValToReplace(); 651 Type *UseTy = Op->getType(); 652 Instruction *User = UI->getUser(); 653 654 // Compute the final addrec to expand into code. 655 const SCEV *AR = IU->getReplacementExpr(*UI); 656 657 // Evaluate the expression out of the loop, if possible. 658 if (!L->contains(UI->getUser())) { 659 const SCEV *ExitVal = SE->getSCEVAtScope(AR, L->getParentLoop()); 660 if (SE->isLoopInvariant(ExitVal, L)) 661 AR = ExitVal; 662 } 663 664 // FIXME: It is an extremely bad idea to indvar substitute anything more 665 // complex than affine induction variables. Doing so will put expensive 666 // polynomial evaluations inside of the loop, and the str reduction pass 667 // currently can only reduce affine polynomials. For now just disable 668 // indvar subst on anything more complex than an affine addrec, unless 669 // it can be expanded to a trivial value. 670 if (!isSafe(AR, L, SE)) 671 continue; 672 673 // Determine the insertion point for this user. By default, insert 674 // immediately before the user. The SCEVExpander class will automatically 675 // hoist loop invariants out of the loop. For PHI nodes, there may be 676 // multiple uses, so compute the nearest common dominator for the 677 // incoming blocks. 678 Instruction *InsertPt = getInsertPointForUses(User, Op, DT); 679 680 // Now expand it into actual Instructions and patch it into place. 681 Value *NewVal = Rewriter.expandCodeFor(AR, UseTy, InsertPt); 682 683 DEBUG(dbgs() << "INDVARS: Rewrote IV '" << *AR << "' " << *Op << '\n' 684 << " into = " << *NewVal << "\n"); 685 686 if (!isValidRewrite(Op, NewVal)) { 687 DeadInsts.push_back(NewVal); 688 continue; 689 } 690 // Inform ScalarEvolution that this value is changing. The change doesn't 691 // affect its value, but it does potentially affect which use lists the 692 // value will be on after the replacement, which affects ScalarEvolution's 693 // ability to walk use lists and drop dangling pointers when a value is 694 // deleted. 695 SE->forgetValue(User); 696 697 // Patch the new value into place. 698 if (Op->hasName()) 699 NewVal->takeName(Op); 700 if (Instruction *NewValI = dyn_cast<Instruction>(NewVal)) 701 NewValI->setDebugLoc(User->getDebugLoc()); 702 User->replaceUsesOfWith(Op, NewVal); 703 UI->setOperandValToReplace(NewVal); 704 705 ++NumRemoved; 706 Changed = true; 707 708 // The old value may be dead now. 709 DeadInsts.push_back(Op); 710 } 711} 712 713//===----------------------------------------------------------------------===// 714// IV Widening - Extend the width of an IV to cover its widest uses. 715//===----------------------------------------------------------------------===// 716 717namespace { 718 // Collect information about induction variables that are used by sign/zero 719 // extend operations. This information is recorded by CollectExtend and 720 // provides the input to WidenIV. 721 struct WideIVInfo { 722 PHINode *NarrowIV; 723 Type *WidestNativeType; // Widest integer type created [sz]ext 724 bool IsSigned; // Was an sext user seen before a zext? 725 726 WideIVInfo() : NarrowIV(0), WidestNativeType(0), IsSigned(false) {} 727 }; 728 729 class WideIVVisitor : public IVVisitor { 730 ScalarEvolution *SE; 731 const TargetData *TD; 732 733 public: 734 WideIVInfo WI; 735 736 WideIVVisitor(PHINode *NarrowIV, ScalarEvolution *SCEV, 737 const TargetData *TData) : 738 SE(SCEV), TD(TData) { WI.NarrowIV = NarrowIV; } 739 740 // Implement the interface used by simplifyUsersOfIV. 741 virtual void visitCast(CastInst *Cast); 742 }; 743} 744 745/// visitCast - Update information about the induction variable that is 746/// extended by this sign or zero extend operation. This is used to determine 747/// the final width of the IV before actually widening it. 748void WideIVVisitor::visitCast(CastInst *Cast) { 749 bool IsSigned = Cast->getOpcode() == Instruction::SExt; 750 if (!IsSigned && Cast->getOpcode() != Instruction::ZExt) 751 return; 752 753 Type *Ty = Cast->getType(); 754 uint64_t Width = SE->getTypeSizeInBits(Ty); 755 if (TD && !TD->isLegalInteger(Width)) 756 return; 757 758 if (!WI.WidestNativeType) { 759 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty); 760 WI.IsSigned = IsSigned; 761 return; 762 } 763 764 // We extend the IV to satisfy the sign of its first user, arbitrarily. 765 if (WI.IsSigned != IsSigned) 766 return; 767 768 if (Width > SE->getTypeSizeInBits(WI.WidestNativeType)) 769 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty); 770} 771 772namespace { 773 774/// NarrowIVDefUse - Record a link in the Narrow IV def-use chain along with the 775/// WideIV that computes the same value as the Narrow IV def. This avoids 776/// caching Use* pointers. 777struct NarrowIVDefUse { 778 Instruction *NarrowDef; 779 Instruction *NarrowUse; 780 Instruction *WideDef; 781 782 NarrowIVDefUse(): NarrowDef(0), NarrowUse(0), WideDef(0) {} 783 784 NarrowIVDefUse(Instruction *ND, Instruction *NU, Instruction *WD): 785 NarrowDef(ND), NarrowUse(NU), WideDef(WD) {} 786}; 787 788/// WidenIV - The goal of this transform is to remove sign and zero extends 789/// without creating any new induction variables. To do this, it creates a new 790/// phi of the wider type and redirects all users, either removing extends or 791/// inserting truncs whenever we stop propagating the type. 792/// 793class WidenIV { 794 // Parameters 795 PHINode *OrigPhi; 796 Type *WideType; 797 bool IsSigned; 798 799 // Context 800 LoopInfo *LI; 801 Loop *L; 802 ScalarEvolution *SE; 803 DominatorTree *DT; 804 805 // Result 806 PHINode *WidePhi; 807 Instruction *WideInc; 808 const SCEV *WideIncExpr; 809 SmallVectorImpl<WeakVH> &DeadInsts; 810 811 SmallPtrSet<Instruction*,16> Widened; 812 SmallVector<NarrowIVDefUse, 8> NarrowIVUsers; 813 814public: 815 WidenIV(const WideIVInfo &WI, LoopInfo *LInfo, 816 ScalarEvolution *SEv, DominatorTree *DTree, 817 SmallVectorImpl<WeakVH> &DI) : 818 OrigPhi(WI.NarrowIV), 819 WideType(WI.WidestNativeType), 820 IsSigned(WI.IsSigned), 821 LI(LInfo), 822 L(LI->getLoopFor(OrigPhi->getParent())), 823 SE(SEv), 824 DT(DTree), 825 WidePhi(0), 826 WideInc(0), 827 WideIncExpr(0), 828 DeadInsts(DI) { 829 assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV"); 830 } 831 832 PHINode *CreateWideIV(SCEVExpander &Rewriter); 833 834protected: 835 Value *getExtend(Value *NarrowOper, Type *WideType, bool IsSigned, 836 Instruction *Use); 837 838 Instruction *CloneIVUser(NarrowIVDefUse DU); 839 840 const SCEVAddRecExpr *GetWideRecurrence(Instruction *NarrowUse); 841 842 const SCEVAddRecExpr* GetExtendedOperandRecurrence(NarrowIVDefUse DU); 843 844 Instruction *WidenIVUse(NarrowIVDefUse DU); 845 846 void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef); 847}; 848} // anonymous namespace 849 850/// isLoopInvariant - Perform a quick domtree based check for loop invariance 851/// assuming that V is used within the loop. LoopInfo::isLoopInvariant() seems 852/// gratuitous for this purpose. 853static bool isLoopInvariant(Value *V, const Loop *L, const DominatorTree *DT) { 854 Instruction *Inst = dyn_cast<Instruction>(V); 855 if (!Inst) 856 return true; 857 858 return DT->properlyDominates(Inst->getParent(), L->getHeader()); 859} 860 861Value *WidenIV::getExtend(Value *NarrowOper, Type *WideType, bool IsSigned, 862 Instruction *Use) { 863 // Set the debug location and conservative insertion point. 864 IRBuilder<> Builder(Use); 865 // Hoist the insertion point into loop preheaders as far as possible. 866 for (const Loop *L = LI->getLoopFor(Use->getParent()); 867 L && L->getLoopPreheader() && isLoopInvariant(NarrowOper, L, DT); 868 L = L->getParentLoop()) 869 Builder.SetInsertPoint(L->getLoopPreheader()->getTerminator()); 870 871 return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) : 872 Builder.CreateZExt(NarrowOper, WideType); 873} 874 875/// CloneIVUser - Instantiate a wide operation to replace a narrow 876/// operation. This only needs to handle operations that can evaluation to 877/// SCEVAddRec. It can safely return 0 for any operation we decide not to clone. 878Instruction *WidenIV::CloneIVUser(NarrowIVDefUse DU) { 879 unsigned Opcode = DU.NarrowUse->getOpcode(); 880 switch (Opcode) { 881 default: 882 return 0; 883 case Instruction::Add: 884 case Instruction::Mul: 885 case Instruction::UDiv: 886 case Instruction::Sub: 887 case Instruction::And: 888 case Instruction::Or: 889 case Instruction::Xor: 890 case Instruction::Shl: 891 case Instruction::LShr: 892 case Instruction::AShr: 893 DEBUG(dbgs() << "Cloning IVUser: " << *DU.NarrowUse << "\n"); 894 895 // Replace NarrowDef operands with WideDef. Otherwise, we don't know 896 // anything about the narrow operand yet so must insert a [sz]ext. It is 897 // probably loop invariant and will be folded or hoisted. If it actually 898 // comes from a widened IV, it should be removed during a future call to 899 // WidenIVUse. 900 Value *LHS = (DU.NarrowUse->getOperand(0) == DU.NarrowDef) ? DU.WideDef : 901 getExtend(DU.NarrowUse->getOperand(0), WideType, IsSigned, DU.NarrowUse); 902 Value *RHS = (DU.NarrowUse->getOperand(1) == DU.NarrowDef) ? DU.WideDef : 903 getExtend(DU.NarrowUse->getOperand(1), WideType, IsSigned, DU.NarrowUse); 904 905 BinaryOperator *NarrowBO = cast<BinaryOperator>(DU.NarrowUse); 906 BinaryOperator *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(), 907 LHS, RHS, 908 NarrowBO->getName()); 909 IRBuilder<> Builder(DU.NarrowUse); 910 Builder.Insert(WideBO); 911 if (const OverflowingBinaryOperator *OBO = 912 dyn_cast<OverflowingBinaryOperator>(NarrowBO)) { 913 if (OBO->hasNoUnsignedWrap()) WideBO->setHasNoUnsignedWrap(); 914 if (OBO->hasNoSignedWrap()) WideBO->setHasNoSignedWrap(); 915 } 916 return WideBO; 917 } 918 llvm_unreachable(0); 919} 920 921/// No-wrap operations can transfer sign extension of their result to their 922/// operands. Generate the SCEV value for the widened operation without 923/// actually modifying the IR yet. If the expression after extending the 924/// operands is an AddRec for this loop, return it. 925const SCEVAddRecExpr* WidenIV::GetExtendedOperandRecurrence(NarrowIVDefUse DU) { 926 // Handle the common case of add<nsw/nuw> 927 if (DU.NarrowUse->getOpcode() != Instruction::Add) 928 return 0; 929 930 // One operand (NarrowDef) has already been extended to WideDef. Now determine 931 // if extending the other will lead to a recurrence. 932 unsigned ExtendOperIdx = DU.NarrowUse->getOperand(0) == DU.NarrowDef ? 1 : 0; 933 assert(DU.NarrowUse->getOperand(1-ExtendOperIdx) == DU.NarrowDef && "bad DU"); 934 935 const SCEV *ExtendOperExpr = 0; 936 const OverflowingBinaryOperator *OBO = 937 cast<OverflowingBinaryOperator>(DU.NarrowUse); 938 if (IsSigned && OBO->hasNoSignedWrap()) 939 ExtendOperExpr = SE->getSignExtendExpr( 940 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType); 941 else if(!IsSigned && OBO->hasNoUnsignedWrap()) 942 ExtendOperExpr = SE->getZeroExtendExpr( 943 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType); 944 else 945 return 0; 946 947 // When creating this AddExpr, don't apply the current operations NSW or NUW 948 // flags. This instruction may be guarded by control flow that the no-wrap 949 // behavior depends on. Non-control-equivalent instructions can be mapped to 950 // the same SCEV expression, and it would be incorrect to transfer NSW/NUW 951 // semantics to those operations. 952 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>( 953 SE->getAddExpr(SE->getSCEV(DU.WideDef), ExtendOperExpr)); 954 955 if (!AddRec || AddRec->getLoop() != L) 956 return 0; 957 return AddRec; 958} 959 960/// GetWideRecurrence - Is this instruction potentially interesting from 961/// IVUsers' perspective after widening it's type? In other words, can the 962/// extend be safely hoisted out of the loop with SCEV reducing the value to a 963/// recurrence on the same loop. If so, return the sign or zero extended 964/// recurrence. Otherwise return NULL. 965const SCEVAddRecExpr *WidenIV::GetWideRecurrence(Instruction *NarrowUse) { 966 if (!SE->isSCEVable(NarrowUse->getType())) 967 return 0; 968 969 const SCEV *NarrowExpr = SE->getSCEV(NarrowUse); 970 if (SE->getTypeSizeInBits(NarrowExpr->getType()) 971 >= SE->getTypeSizeInBits(WideType)) { 972 // NarrowUse implicitly widens its operand. e.g. a gep with a narrow 973 // index. So don't follow this use. 974 return 0; 975 } 976 977 const SCEV *WideExpr = IsSigned ? 978 SE->getSignExtendExpr(NarrowExpr, WideType) : 979 SE->getZeroExtendExpr(NarrowExpr, WideType); 980 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr); 981 if (!AddRec || AddRec->getLoop() != L) 982 return 0; 983 return AddRec; 984} 985 986/// WidenIVUse - Determine whether an individual user of the narrow IV can be 987/// widened. If so, return the wide clone of the user. 988Instruction *WidenIV::WidenIVUse(NarrowIVDefUse DU) { 989 990 // Stop traversing the def-use chain at inner-loop phis or post-loop phis. 991 if (isa<PHINode>(DU.NarrowUse) && 992 LI->getLoopFor(DU.NarrowUse->getParent()) != L) 993 return 0; 994 995 // Our raison d'etre! Eliminate sign and zero extension. 996 if (IsSigned ? isa<SExtInst>(DU.NarrowUse) : isa<ZExtInst>(DU.NarrowUse)) { 997 Value *NewDef = DU.WideDef; 998 if (DU.NarrowUse->getType() != WideType) { 999 unsigned CastWidth = SE->getTypeSizeInBits(DU.NarrowUse->getType()); 1000 unsigned IVWidth = SE->getTypeSizeInBits(WideType); 1001 if (CastWidth < IVWidth) { 1002 // The cast isn't as wide as the IV, so insert a Trunc. 1003 IRBuilder<> Builder(DU.NarrowUse); 1004 NewDef = Builder.CreateTrunc(DU.WideDef, DU.NarrowUse->getType()); 1005 } 1006 else { 1007 // A wider extend was hidden behind a narrower one. This may induce 1008 // another round of IV widening in which the intermediate IV becomes 1009 // dead. It should be very rare. 1010 DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi 1011 << " not wide enough to subsume " << *DU.NarrowUse << "\n"); 1012 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef); 1013 NewDef = DU.NarrowUse; 1014 } 1015 } 1016 if (NewDef != DU.NarrowUse) { 1017 DEBUG(dbgs() << "INDVARS: eliminating " << *DU.NarrowUse 1018 << " replaced by " << *DU.WideDef << "\n"); 1019 ++NumElimExt; 1020 DU.NarrowUse->replaceAllUsesWith(NewDef); 1021 DeadInsts.push_back(DU.NarrowUse); 1022 } 1023 // Now that the extend is gone, we want to expose it's uses for potential 1024 // further simplification. We don't need to directly inform SimplifyIVUsers 1025 // of the new users, because their parent IV will be processed later as a 1026 // new loop phi. If we preserved IVUsers analysis, we would also want to 1027 // push the uses of WideDef here. 1028 1029 // No further widening is needed. The deceased [sz]ext had done it for us. 1030 return 0; 1031 } 1032 1033 // Does this user itself evaluate to a recurrence after widening? 1034 const SCEVAddRecExpr *WideAddRec = GetWideRecurrence(DU.NarrowUse); 1035 if (!WideAddRec) { 1036 WideAddRec = GetExtendedOperandRecurrence(DU); 1037 } 1038 if (!WideAddRec) { 1039 // This user does not evaluate to a recurence after widening, so don't 1040 // follow it. Instead insert a Trunc to kill off the original use, 1041 // eventually isolating the original narrow IV so it can be removed. 1042 IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT)); 1043 Value *Trunc = Builder.CreateTrunc(DU.WideDef, DU.NarrowDef->getType()); 1044 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, Trunc); 1045 return 0; 1046 } 1047 // Assume block terminators cannot evaluate to a recurrence. We can't to 1048 // insert a Trunc after a terminator if there happens to be a critical edge. 1049 assert(DU.NarrowUse != DU.NarrowUse->getParent()->getTerminator() && 1050 "SCEV is not expected to evaluate a block terminator"); 1051 1052 // Reuse the IV increment that SCEVExpander created as long as it dominates 1053 // NarrowUse. 1054 Instruction *WideUse = 0; 1055 if (WideAddRec == WideIncExpr 1056 && SCEVExpander::hoistStep(WideInc, DU.NarrowUse, DT)) 1057 WideUse = WideInc; 1058 else { 1059 WideUse = CloneIVUser(DU); 1060 if (!WideUse) 1061 return 0; 1062 } 1063 // Evaluation of WideAddRec ensured that the narrow expression could be 1064 // extended outside the loop without overflow. This suggests that the wide use 1065 // evaluates to the same expression as the extended narrow use, but doesn't 1066 // absolutely guarantee it. Hence the following failsafe check. In rare cases 1067 // where it fails, we simply throw away the newly created wide use. 1068 if (WideAddRec != SE->getSCEV(WideUse)) { 1069 DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse 1070 << ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n"); 1071 DeadInsts.push_back(WideUse); 1072 return 0; 1073 } 1074 1075 // Returning WideUse pushes it on the worklist. 1076 return WideUse; 1077} 1078 1079/// pushNarrowIVUsers - Add eligible users of NarrowDef to NarrowIVUsers. 1080/// 1081void WidenIV::pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef) { 1082 for (Value::use_iterator UI = NarrowDef->use_begin(), 1083 UE = NarrowDef->use_end(); UI != UE; ++UI) { 1084 Instruction *NarrowUse = cast<Instruction>(*UI); 1085 1086 // Handle data flow merges and bizarre phi cycles. 1087 if (!Widened.insert(NarrowUse)) 1088 continue; 1089 1090 NarrowIVUsers.push_back(NarrowIVDefUse(NarrowDef, NarrowUse, WideDef)); 1091 } 1092} 1093 1094/// CreateWideIV - Process a single induction variable. First use the 1095/// SCEVExpander to create a wide induction variable that evaluates to the same 1096/// recurrence as the original narrow IV. Then use a worklist to forward 1097/// traverse the narrow IV's def-use chain. After WidenIVUse has processed all 1098/// interesting IV users, the narrow IV will be isolated for removal by 1099/// DeleteDeadPHIs. 1100/// 1101/// It would be simpler to delete uses as they are processed, but we must avoid 1102/// invalidating SCEV expressions. 1103/// 1104PHINode *WidenIV::CreateWideIV(SCEVExpander &Rewriter) { 1105 // Is this phi an induction variable? 1106 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi)); 1107 if (!AddRec) 1108 return NULL; 1109 1110 // Widen the induction variable expression. 1111 const SCEV *WideIVExpr = IsSigned ? 1112 SE->getSignExtendExpr(AddRec, WideType) : 1113 SE->getZeroExtendExpr(AddRec, WideType); 1114 1115 assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType && 1116 "Expect the new IV expression to preserve its type"); 1117 1118 // Can the IV be extended outside the loop without overflow? 1119 AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr); 1120 if (!AddRec || AddRec->getLoop() != L) 1121 return NULL; 1122 1123 // An AddRec must have loop-invariant operands. Since this AddRec is 1124 // materialized by a loop header phi, the expression cannot have any post-loop 1125 // operands, so they must dominate the loop header. 1126 assert(SE->properlyDominates(AddRec->getStart(), L->getHeader()) && 1127 SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader()) 1128 && "Loop header phi recurrence inputs do not dominate the loop"); 1129 1130 // The rewriter provides a value for the desired IV expression. This may 1131 // either find an existing phi or materialize a new one. Either way, we 1132 // expect a well-formed cyclic phi-with-increments. i.e. any operand not part 1133 // of the phi-SCC dominates the loop entry. 1134 Instruction *InsertPt = L->getHeader()->begin(); 1135 WidePhi = cast<PHINode>(Rewriter.expandCodeFor(AddRec, WideType, InsertPt)); 1136 1137 // Remembering the WideIV increment generated by SCEVExpander allows 1138 // WidenIVUse to reuse it when widening the narrow IV's increment. We don't 1139 // employ a general reuse mechanism because the call above is the only call to 1140 // SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses. 1141 if (BasicBlock *LatchBlock = L->getLoopLatch()) { 1142 WideInc = 1143 cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock)); 1144 WideIncExpr = SE->getSCEV(WideInc); 1145 } 1146 1147 DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n"); 1148 ++NumWidened; 1149 1150 // Traverse the def-use chain using a worklist starting at the original IV. 1151 assert(Widened.empty() && NarrowIVUsers.empty() && "expect initial state" ); 1152 1153 Widened.insert(OrigPhi); 1154 pushNarrowIVUsers(OrigPhi, WidePhi); 1155 1156 while (!NarrowIVUsers.empty()) { 1157 NarrowIVDefUse DU = NarrowIVUsers.pop_back_val(); 1158 1159 // Process a def-use edge. This may replace the use, so don't hold a 1160 // use_iterator across it. 1161 Instruction *WideUse = WidenIVUse(DU); 1162 1163 // Follow all def-use edges from the previous narrow use. 1164 if (WideUse) 1165 pushNarrowIVUsers(DU.NarrowUse, WideUse); 1166 1167 // WidenIVUse may have removed the def-use edge. 1168 if (DU.NarrowDef->use_empty()) 1169 DeadInsts.push_back(DU.NarrowDef); 1170 } 1171 return WidePhi; 1172} 1173 1174//===----------------------------------------------------------------------===// 1175// Simplification of IV users based on SCEV evaluation. 1176//===----------------------------------------------------------------------===// 1177 1178 1179/// SimplifyAndExtend - Iteratively perform simplification on a worklist of IV 1180/// users. Each successive simplification may push more users which may 1181/// themselves be candidates for simplification. 1182/// 1183/// Sign/Zero extend elimination is interleaved with IV simplification. 1184/// 1185void IndVarSimplify::SimplifyAndExtend(Loop *L, 1186 SCEVExpander &Rewriter, 1187 LPPassManager &LPM) { 1188 SmallVector<WideIVInfo, 8> WideIVs; 1189 1190 SmallVector<PHINode*, 8> LoopPhis; 1191 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) { 1192 LoopPhis.push_back(cast<PHINode>(I)); 1193 } 1194 // Each round of simplification iterates through the SimplifyIVUsers worklist 1195 // for all current phis, then determines whether any IVs can be 1196 // widened. Widening adds new phis to LoopPhis, inducing another round of 1197 // simplification on the wide IVs. 1198 while (!LoopPhis.empty()) { 1199 // Evaluate as many IV expressions as possible before widening any IVs. This 1200 // forces SCEV to set no-wrap flags before evaluating sign/zero 1201 // extension. The first time SCEV attempts to normalize sign/zero extension, 1202 // the result becomes final. So for the most predictable results, we delay 1203 // evaluation of sign/zero extend evaluation until needed, and avoid running 1204 // other SCEV based analysis prior to SimplifyAndExtend. 1205 do { 1206 PHINode *CurrIV = LoopPhis.pop_back_val(); 1207 1208 // Information about sign/zero extensions of CurrIV. 1209 WideIVVisitor WIV(CurrIV, SE, TD); 1210 1211 Changed |= simplifyUsersOfIV(CurrIV, SE, &LPM, DeadInsts, &WIV); 1212 1213 if (WIV.WI.WidestNativeType) { 1214 WideIVs.push_back(WIV.WI); 1215 } 1216 } while(!LoopPhis.empty()); 1217 1218 for (; !WideIVs.empty(); WideIVs.pop_back()) { 1219 WidenIV Widener(WideIVs.back(), LI, SE, DT, DeadInsts); 1220 if (PHINode *WidePhi = Widener.CreateWideIV(Rewriter)) { 1221 Changed = true; 1222 LoopPhis.push_back(WidePhi); 1223 } 1224 } 1225 } 1226} 1227 1228//===----------------------------------------------------------------------===// 1229// LinearFunctionTestReplace and its kin. Rewrite the loop exit condition. 1230//===----------------------------------------------------------------------===// 1231 1232/// Check for expressions that ScalarEvolution generates to compute 1233/// BackedgeTakenInfo. If these expressions have not been reduced, then 1234/// expanding them may incur additional cost (albeit in the loop preheader). 1235static bool isHighCostExpansion(const SCEV *S, BranchInst *BI, 1236 ScalarEvolution *SE) { 1237 // If the backedge-taken count is a UDiv, it's very likely a UDiv that 1238 // ScalarEvolution's HowFarToZero or HowManyLessThans produced to compute a 1239 // precise expression, rather than a UDiv from the user's code. If we can't 1240 // find a UDiv in the code with some simple searching, assume the former and 1241 // forego rewriting the loop. 1242 if (isa<SCEVUDivExpr>(S)) { 1243 ICmpInst *OrigCond = dyn_cast<ICmpInst>(BI->getCondition()); 1244 if (!OrigCond) return true; 1245 const SCEV *R = SE->getSCEV(OrigCond->getOperand(1)); 1246 R = SE->getMinusSCEV(R, SE->getConstant(R->getType(), 1)); 1247 if (R != S) { 1248 const SCEV *L = SE->getSCEV(OrigCond->getOperand(0)); 1249 L = SE->getMinusSCEV(L, SE->getConstant(L->getType(), 1)); 1250 if (L != S) 1251 return true; 1252 } 1253 } 1254 1255 if (EnableIVRewrite) 1256 return false; 1257 1258 // Recurse past add expressions, which commonly occur in the 1259 // BackedgeTakenCount. They may already exist in program code, and if not, 1260 // they are not too expensive rematerialize. 1261 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 1262 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end(); 1263 I != E; ++I) { 1264 if (isHighCostExpansion(*I, BI, SE)) 1265 return true; 1266 } 1267 return false; 1268 } 1269 1270 // HowManyLessThans uses a Max expression whenever the loop is not guarded by 1271 // the exit condition. 1272 if (isa<SCEVSMaxExpr>(S) || isa<SCEVUMaxExpr>(S)) 1273 return true; 1274 1275 // If we haven't recognized an expensive SCEV patter, assume its an expression 1276 // produced by program code. 1277 return false; 1278} 1279 1280/// canExpandBackedgeTakenCount - Return true if this loop's backedge taken 1281/// count expression can be safely and cheaply expanded into an instruction 1282/// sequence that can be used by LinearFunctionTestReplace. 1283/// 1284/// TODO: This fails for pointer-type loop counters with greater than one byte 1285/// strides, consequently preventing LFTR from running. For the purpose of LFTR 1286/// we could skip this check in the case that the LFTR loop counter (chosen by 1287/// FindLoopCounter) is also pointer type. Instead, we could directly convert 1288/// the loop test to an inequality test by checking the target data's alignment 1289/// of element types (given that the initial pointer value originates from or is 1290/// used by ABI constrained operation, as opposed to inttoptr/ptrtoint). 1291/// However, we don't yet have a strong motivation for converting loop tests 1292/// into inequality tests. 1293static bool canExpandBackedgeTakenCount(Loop *L, ScalarEvolution *SE) { 1294 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L); 1295 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) || 1296 BackedgeTakenCount->isZero()) 1297 return false; 1298 1299 if (!L->getExitingBlock()) 1300 return false; 1301 1302 // Can't rewrite non-branch yet. 1303 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1304 if (!BI) 1305 return false; 1306 1307 if (isHighCostExpansion(BackedgeTakenCount, BI, SE)) 1308 return false; 1309 1310 return true; 1311} 1312 1313/// getBackedgeIVType - Get the widest type used by the loop test after peeking 1314/// through Truncs. 1315/// 1316/// TODO: Unnecessary when ForceLFTR is removed. 1317static Type *getBackedgeIVType(Loop *L) { 1318 if (!L->getExitingBlock()) 1319 return 0; 1320 1321 // Can't rewrite non-branch yet. 1322 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1323 if (!BI) 1324 return 0; 1325 1326 ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition()); 1327 if (!Cond) 1328 return 0; 1329 1330 Type *Ty = 0; 1331 for(User::op_iterator OI = Cond->op_begin(), OE = Cond->op_end(); 1332 OI != OE; ++OI) { 1333 assert((!Ty || Ty == (*OI)->getType()) && "bad icmp operand types"); 1334 TruncInst *Trunc = dyn_cast<TruncInst>(*OI); 1335 if (!Trunc) 1336 continue; 1337 1338 return Trunc->getSrcTy(); 1339 } 1340 return Ty; 1341} 1342 1343/// getLoopPhiForCounter - Return the loop header phi IFF IncV adds a loop 1344/// invariant value to the phi. 1345static PHINode *getLoopPhiForCounter(Value *IncV, Loop *L, DominatorTree *DT) { 1346 Instruction *IncI = dyn_cast<Instruction>(IncV); 1347 if (!IncI) 1348 return 0; 1349 1350 switch (IncI->getOpcode()) { 1351 case Instruction::Add: 1352 case Instruction::Sub: 1353 break; 1354 case Instruction::GetElementPtr: 1355 // An IV counter must preserve its type. 1356 if (IncI->getNumOperands() == 2) 1357 break; 1358 default: 1359 return 0; 1360 } 1361 1362 PHINode *Phi = dyn_cast<PHINode>(IncI->getOperand(0)); 1363 if (Phi && Phi->getParent() == L->getHeader()) { 1364 if (isLoopInvariant(IncI->getOperand(1), L, DT)) 1365 return Phi; 1366 return 0; 1367 } 1368 if (IncI->getOpcode() == Instruction::GetElementPtr) 1369 return 0; 1370 1371 // Allow add/sub to be commuted. 1372 Phi = dyn_cast<PHINode>(IncI->getOperand(1)); 1373 if (Phi && Phi->getParent() == L->getHeader()) { 1374 if (isLoopInvariant(IncI->getOperand(0), L, DT)) 1375 return Phi; 1376 } 1377 return 0; 1378} 1379 1380/// needsLFTR - LinearFunctionTestReplace policy. Return true unless we can show 1381/// that the current exit test is already sufficiently canonical. 1382static bool needsLFTR(Loop *L, DominatorTree *DT) { 1383 assert(L->getExitingBlock() && "expected loop exit"); 1384 1385 BasicBlock *LatchBlock = L->getLoopLatch(); 1386 // Don't bother with LFTR if the loop is not properly simplified. 1387 if (!LatchBlock) 1388 return false; 1389 1390 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1391 assert(BI && "expected exit branch"); 1392 1393 // Do LFTR to simplify the exit condition to an ICMP. 1394 ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition()); 1395 if (!Cond) 1396 return true; 1397 1398 // Do LFTR to simplify the exit ICMP to EQ/NE 1399 ICmpInst::Predicate Pred = Cond->getPredicate(); 1400 if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ) 1401 return true; 1402 1403 // Look for a loop invariant RHS 1404 Value *LHS = Cond->getOperand(0); 1405 Value *RHS = Cond->getOperand(1); 1406 if (!isLoopInvariant(RHS, L, DT)) { 1407 if (!isLoopInvariant(LHS, L, DT)) 1408 return true; 1409 std::swap(LHS, RHS); 1410 } 1411 // Look for a simple IV counter LHS 1412 PHINode *Phi = dyn_cast<PHINode>(LHS); 1413 if (!Phi) 1414 Phi = getLoopPhiForCounter(LHS, L, DT); 1415 1416 if (!Phi) 1417 return true; 1418 1419 // Do LFTR if the exit condition's IV is *not* a simple counter. 1420 Value *IncV = Phi->getIncomingValueForBlock(L->getLoopLatch()); 1421 return Phi != getLoopPhiForCounter(IncV, L, DT); 1422} 1423 1424/// AlmostDeadIV - Return true if this IV has any uses other than the (soon to 1425/// be rewritten) loop exit test. 1426static bool AlmostDeadIV(PHINode *Phi, BasicBlock *LatchBlock, Value *Cond) { 1427 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock); 1428 Value *IncV = Phi->getIncomingValue(LatchIdx); 1429 1430 for (Value::use_iterator UI = Phi->use_begin(), UE = Phi->use_end(); 1431 UI != UE; ++UI) { 1432 if (*UI != Cond && *UI != IncV) return false; 1433 } 1434 1435 for (Value::use_iterator UI = IncV->use_begin(), UE = IncV->use_end(); 1436 UI != UE; ++UI) { 1437 if (*UI != Cond && *UI != Phi) return false; 1438 } 1439 return true; 1440} 1441 1442/// FindLoopCounter - Find an affine IV in canonical form. 1443/// 1444/// BECount may be an i8* pointer type. The pointer difference is already 1445/// valid count without scaling the address stride, so it remains a pointer 1446/// expression as far as SCEV is concerned. 1447/// 1448/// FIXME: Accept -1 stride and set IVLimit = IVInit - BECount 1449/// 1450/// FIXME: Accept non-unit stride as long as SCEV can reduce BECount * Stride. 1451/// This is difficult in general for SCEV because of potential overflow. But we 1452/// could at least handle constant BECounts. 1453static PHINode * 1454FindLoopCounter(Loop *L, const SCEV *BECount, 1455 ScalarEvolution *SE, DominatorTree *DT, const TargetData *TD) { 1456 uint64_t BCWidth = SE->getTypeSizeInBits(BECount->getType()); 1457 1458 Value *Cond = 1459 cast<BranchInst>(L->getExitingBlock()->getTerminator())->getCondition(); 1460 1461 // Loop over all of the PHI nodes, looking for a simple counter. 1462 PHINode *BestPhi = 0; 1463 const SCEV *BestInit = 0; 1464 BasicBlock *LatchBlock = L->getLoopLatch(); 1465 assert(LatchBlock && "needsLFTR should guarantee a loop latch"); 1466 1467 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) { 1468 PHINode *Phi = cast<PHINode>(I); 1469 if (!SE->isSCEVable(Phi->getType())) 1470 continue; 1471 1472 // Avoid comparing an integer IV against a pointer Limit. 1473 if (BECount->getType()->isPointerTy() && !Phi->getType()->isPointerTy()) 1474 continue; 1475 1476 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Phi)); 1477 if (!AR || AR->getLoop() != L || !AR->isAffine()) 1478 continue; 1479 1480 // AR may be a pointer type, while BECount is an integer type. 1481 // AR may be wider than BECount. With eq/ne tests overflow is immaterial. 1482 // AR may not be a narrower type, or we may never exit. 1483 uint64_t PhiWidth = SE->getTypeSizeInBits(AR->getType()); 1484 if (PhiWidth < BCWidth || (TD && !TD->isLegalInteger(PhiWidth))) 1485 continue; 1486 1487 const SCEV *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE)); 1488 if (!Step || !Step->isOne()) 1489 continue; 1490 1491 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock); 1492 Value *IncV = Phi->getIncomingValue(LatchIdx); 1493 if (getLoopPhiForCounter(IncV, L, DT) != Phi) 1494 continue; 1495 1496 const SCEV *Init = AR->getStart(); 1497 1498 if (BestPhi && !AlmostDeadIV(BestPhi, LatchBlock, Cond)) { 1499 // Don't force a live loop counter if another IV can be used. 1500 if (AlmostDeadIV(Phi, LatchBlock, Cond)) 1501 continue; 1502 1503 // Prefer to count-from-zero. This is a more "canonical" counter form. It 1504 // also prefers integer to pointer IVs. 1505 if (BestInit->isZero() != Init->isZero()) { 1506 if (BestInit->isZero()) 1507 continue; 1508 } 1509 // If two IVs both count from zero or both count from nonzero then the 1510 // narrower is likely a dead phi that has been widened. Use the wider phi 1511 // to allow the other to be eliminated. 1512 if (PhiWidth <= SE->getTypeSizeInBits(BestPhi->getType())) 1513 continue; 1514 } 1515 BestPhi = Phi; 1516 BestInit = Init; 1517 } 1518 return BestPhi; 1519} 1520 1521/// genLoopLimit - Help LinearFunctionTestReplace by generating a value that 1522/// holds the RHS of the new loop test. 1523static Value *genLoopLimit(PHINode *IndVar, const SCEV *IVCount, Loop *L, 1524 SCEVExpander &Rewriter, ScalarEvolution *SE) { 1525 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IndVar)); 1526 assert(AR && AR->getLoop() == L && AR->isAffine() && "bad loop counter"); 1527 const SCEV *IVInit = AR->getStart(); 1528 1529 // IVInit may be a pointer while IVCount is an integer when FindLoopCounter 1530 // finds a valid pointer IV. Sign extend BECount in order to materialize a 1531 // GEP. Avoid running SCEVExpander on a new pointer value, instead reusing 1532 // the existing GEPs whenever possible. 1533 if (IndVar->getType()->isPointerTy() 1534 && !IVCount->getType()->isPointerTy()) { 1535 1536 Type *OfsTy = SE->getEffectiveSCEVType(IVInit->getType()); 1537 const SCEV *IVOffset = SE->getTruncateOrSignExtend(IVCount, OfsTy); 1538 1539 // Expand the code for the iteration count. 1540 assert(SE->isLoopInvariant(IVOffset, L) && 1541 "Computed iteration count is not loop invariant!"); 1542 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1543 Value *GEPOffset = Rewriter.expandCodeFor(IVOffset, OfsTy, BI); 1544 1545 Value *GEPBase = IndVar->getIncomingValueForBlock(L->getLoopPreheader()); 1546 assert(AR->getStart() == SE->getSCEV(GEPBase) && "bad loop counter"); 1547 // We could handle pointer IVs other than i8*, but we need to compensate for 1548 // gep index scaling. See canExpandBackedgeTakenCount comments. 1549 assert(SE->getSizeOfExpr( 1550 cast<PointerType>(GEPBase->getType())->getElementType())->isOne() 1551 && "unit stride pointer IV must be i8*"); 1552 1553 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator()); 1554 return Builder.CreateGEP(GEPBase, GEPOffset, "lftr.limit"); 1555 } 1556 else { 1557 // In any other case, convert both IVInit and IVCount to integers before 1558 // comparing. This may result in SCEV expension of pointers, but in practice 1559 // SCEV will fold the pointer arithmetic away as such: 1560 // BECount = (IVEnd - IVInit - 1) => IVLimit = IVInit (postinc). 1561 // 1562 // Valid Cases: (1) both integers is most common; (2) both may be pointers 1563 // for simple memset-style loops; (3) IVInit is an integer and IVCount is a 1564 // pointer may occur when enable-iv-rewrite generates a canonical IV on top 1565 // of case #2. 1566 1567 const SCEV *IVLimit = 0; 1568 // For unit stride, IVCount = Start + BECount with 2's complement overflow. 1569 // For non-zero Start, compute IVCount here. 1570 if (AR->getStart()->isZero()) 1571 IVLimit = IVCount; 1572 else { 1573 assert(AR->getStepRecurrence(*SE)->isOne() && "only handles unit stride"); 1574 const SCEV *IVInit = AR->getStart(); 1575 1576 // For integer IVs, truncate the IV before computing IVInit + BECount. 1577 if (SE->getTypeSizeInBits(IVInit->getType()) 1578 > SE->getTypeSizeInBits(IVCount->getType())) 1579 IVInit = SE->getTruncateExpr(IVInit, IVCount->getType()); 1580 1581 IVLimit = SE->getAddExpr(IVInit, IVCount); 1582 } 1583 // Expand the code for the iteration count. 1584 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1585 IRBuilder<> Builder(BI); 1586 assert(SE->isLoopInvariant(IVLimit, L) && 1587 "Computed iteration count is not loop invariant!"); 1588 // Ensure that we generate the same type as IndVar, or a smaller integer 1589 // type. In the presence of null pointer values, we have an integer type 1590 // SCEV expression (IVInit) for a pointer type IV value (IndVar). 1591 Type *LimitTy = IVCount->getType()->isPointerTy() ? 1592 IndVar->getType() : IVCount->getType(); 1593 return Rewriter.expandCodeFor(IVLimit, LimitTy, BI); 1594 } 1595} 1596 1597/// LinearFunctionTestReplace - This method rewrites the exit condition of the 1598/// loop to be a canonical != comparison against the incremented loop induction 1599/// variable. This pass is able to rewrite the exit tests of any loop where the 1600/// SCEV analysis can determine a loop-invariant trip count of the loop, which 1601/// is actually a much broader range than just linear tests. 1602Value *IndVarSimplify:: 1603LinearFunctionTestReplace(Loop *L, 1604 const SCEV *BackedgeTakenCount, 1605 PHINode *IndVar, 1606 SCEVExpander &Rewriter) { 1607 assert(canExpandBackedgeTakenCount(L, SE) && "precondition"); 1608 1609 // LFTR can ignore IV overflow and truncate to the width of 1610 // BECount. This avoids materializing the add(zext(add)) expression. 1611 Type *CntTy = !EnableIVRewrite ? 1612 BackedgeTakenCount->getType() : IndVar->getType(); 1613 1614 const SCEV *IVCount = BackedgeTakenCount; 1615 1616 // If the exiting block is the same as the backedge block, we prefer to 1617 // compare against the post-incremented value, otherwise we must compare 1618 // against the preincremented value. 1619 Value *CmpIndVar; 1620 if (L->getExitingBlock() == L->getLoopLatch()) { 1621 // Add one to the "backedge-taken" count to get the trip count. 1622 // If this addition may overflow, we have to be more pessimistic and 1623 // cast the induction variable before doing the add. 1624 const SCEV *N = 1625 SE->getAddExpr(IVCount, SE->getConstant(IVCount->getType(), 1)); 1626 if (CntTy == IVCount->getType()) 1627 IVCount = N; 1628 else { 1629 const SCEV *Zero = SE->getConstant(IVCount->getType(), 0); 1630 if ((isa<SCEVConstant>(N) && !N->isZero()) || 1631 SE->isLoopEntryGuardedByCond(L, ICmpInst::ICMP_NE, N, Zero)) { 1632 // No overflow. Cast the sum. 1633 IVCount = SE->getTruncateOrZeroExtend(N, CntTy); 1634 } else { 1635 // Potential overflow. Cast before doing the add. 1636 IVCount = SE->getTruncateOrZeroExtend(IVCount, CntTy); 1637 IVCount = SE->getAddExpr(IVCount, SE->getConstant(CntTy, 1)); 1638 } 1639 } 1640 // The BackedgeTaken expression contains the number of times that the 1641 // backedge branches to the loop header. This is one less than the 1642 // number of times the loop executes, so use the incremented indvar. 1643 CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock()); 1644 } else { 1645 // We must use the preincremented value... 1646 IVCount = SE->getTruncateOrZeroExtend(IVCount, CntTy); 1647 CmpIndVar = IndVar; 1648 } 1649 1650 Value *ExitCnt = genLoopLimit(IndVar, IVCount, L, Rewriter, SE); 1651 assert(ExitCnt->getType()->isPointerTy() == IndVar->getType()->isPointerTy() 1652 && "genLoopLimit missed a cast"); 1653 1654 // Insert a new icmp_ne or icmp_eq instruction before the branch. 1655 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1656 ICmpInst::Predicate P; 1657 if (L->contains(BI->getSuccessor(0))) 1658 P = ICmpInst::ICMP_NE; 1659 else 1660 P = ICmpInst::ICMP_EQ; 1661 1662 DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n" 1663 << " LHS:" << *CmpIndVar << '\n' 1664 << " op:\t" 1665 << (P == ICmpInst::ICMP_NE ? "!=" : "==") << "\n" 1666 << " RHS:\t" << *ExitCnt << "\n" 1667 << " IVCount:\t" << *IVCount << "\n"); 1668 1669 IRBuilder<> Builder(BI); 1670 if (SE->getTypeSizeInBits(CmpIndVar->getType()) 1671 > SE->getTypeSizeInBits(ExitCnt->getType())) { 1672 CmpIndVar = Builder.CreateTrunc(CmpIndVar, ExitCnt->getType(), 1673 "lftr.wideiv"); 1674 } 1675 1676 Value *Cond = Builder.CreateICmp(P, CmpIndVar, ExitCnt, "exitcond"); 1677 Value *OrigCond = BI->getCondition(); 1678 // It's tempting to use replaceAllUsesWith here to fully replace the old 1679 // comparison, but that's not immediately safe, since users of the old 1680 // comparison may not be dominated by the new comparison. Instead, just 1681 // update the branch to use the new comparison; in the common case this 1682 // will make old comparison dead. 1683 BI->setCondition(Cond); 1684 DeadInsts.push_back(OrigCond); 1685 1686 ++NumLFTR; 1687 Changed = true; 1688 return Cond; 1689} 1690 1691//===----------------------------------------------------------------------===// 1692// SinkUnusedInvariants. A late subpass to cleanup loop preheaders. 1693//===----------------------------------------------------------------------===// 1694 1695/// If there's a single exit block, sink any loop-invariant values that 1696/// were defined in the preheader but not used inside the loop into the 1697/// exit block to reduce register pressure in the loop. 1698void IndVarSimplify::SinkUnusedInvariants(Loop *L) { 1699 BasicBlock *ExitBlock = L->getExitBlock(); 1700 if (!ExitBlock) return; 1701 1702 BasicBlock *Preheader = L->getLoopPreheader(); 1703 if (!Preheader) return; 1704 1705 Instruction *InsertPt = ExitBlock->getFirstInsertionPt(); 1706 BasicBlock::iterator I = Preheader->getTerminator(); 1707 while (I != Preheader->begin()) { 1708 --I; 1709 // New instructions were inserted at the end of the preheader. 1710 if (isa<PHINode>(I)) 1711 break; 1712 1713 // Don't move instructions which might have side effects, since the side 1714 // effects need to complete before instructions inside the loop. Also don't 1715 // move instructions which might read memory, since the loop may modify 1716 // memory. Note that it's okay if the instruction might have undefined 1717 // behavior: LoopSimplify guarantees that the preheader dominates the exit 1718 // block. 1719 if (I->mayHaveSideEffects() || I->mayReadFromMemory()) 1720 continue; 1721 1722 // Skip debug info intrinsics. 1723 if (isa<DbgInfoIntrinsic>(I)) 1724 continue; 1725 1726 // Skip landingpad instructions. 1727 if (isa<LandingPadInst>(I)) 1728 continue; 1729 1730 // Don't sink alloca: we never want to sink static alloca's out of the 1731 // entry block, and correctly sinking dynamic alloca's requires 1732 // checks for stacksave/stackrestore intrinsics. 1733 // FIXME: Refactor this check somehow? 1734 if (isa<AllocaInst>(I)) 1735 continue; 1736 1737 // Determine if there is a use in or before the loop (direct or 1738 // otherwise). 1739 bool UsedInLoop = false; 1740 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); 1741 UI != UE; ++UI) { 1742 User *U = *UI; 1743 BasicBlock *UseBB = cast<Instruction>(U)->getParent(); 1744 if (PHINode *P = dyn_cast<PHINode>(U)) { 1745 unsigned i = 1746 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()); 1747 UseBB = P->getIncomingBlock(i); 1748 } 1749 if (UseBB == Preheader || L->contains(UseBB)) { 1750 UsedInLoop = true; 1751 break; 1752 } 1753 } 1754 1755 // If there is, the def must remain in the preheader. 1756 if (UsedInLoop) 1757 continue; 1758 1759 // Otherwise, sink it to the exit block. 1760 Instruction *ToMove = I; 1761 bool Done = false; 1762 1763 if (I != Preheader->begin()) { 1764 // Skip debug info intrinsics. 1765 do { 1766 --I; 1767 } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin()); 1768 1769 if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin()) 1770 Done = true; 1771 } else { 1772 Done = true; 1773 } 1774 1775 ToMove->moveBefore(InsertPt); 1776 if (Done) break; 1777 InsertPt = ToMove; 1778 } 1779} 1780 1781//===----------------------------------------------------------------------===// 1782// IndVarSimplify driver. Manage several subpasses of IV simplification. 1783//===----------------------------------------------------------------------===// 1784 1785bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) { 1786 // If LoopSimplify form is not available, stay out of trouble. Some notes: 1787 // - LSR currently only supports LoopSimplify-form loops. Indvars' 1788 // canonicalization can be a pessimization without LSR to "clean up" 1789 // afterwards. 1790 // - We depend on having a preheader; in particular, 1791 // Loop::getCanonicalInductionVariable only supports loops with preheaders, 1792 // and we're in trouble if we can't find the induction variable even when 1793 // we've manually inserted one. 1794 if (!L->isLoopSimplifyForm()) 1795 return false; 1796 1797 if (EnableIVRewrite) 1798 IU = &getAnalysis<IVUsers>(); 1799 LI = &getAnalysis<LoopInfo>(); 1800 SE = &getAnalysis<ScalarEvolution>(); 1801 DT = &getAnalysis<DominatorTree>(); 1802 TD = getAnalysisIfAvailable<TargetData>(); 1803 1804 DeadInsts.clear(); 1805 Changed = false; 1806 1807 // If there are any floating-point recurrences, attempt to 1808 // transform them to use integer recurrences. 1809 RewriteNonIntegerIVs(L); 1810 1811 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L); 1812 1813 // Create a rewriter object which we'll use to transform the code with. 1814 SCEVExpander Rewriter(*SE, "indvars"); 1815#ifndef NDEBUG 1816 Rewriter.setDebugType(DEBUG_TYPE); 1817#endif 1818 1819 // Eliminate redundant IV users. 1820 // 1821 // Simplification works best when run before other consumers of SCEV. We 1822 // attempt to avoid evaluating SCEVs for sign/zero extend operations until 1823 // other expressions involving loop IVs have been evaluated. This helps SCEV 1824 // set no-wrap flags before normalizing sign/zero extension. 1825 if (!EnableIVRewrite) { 1826 Rewriter.disableCanonicalMode(); 1827 SimplifyAndExtend(L, Rewriter, LPM); 1828 } 1829 1830 // Check to see if this loop has a computable loop-invariant execution count. 1831 // If so, this means that we can compute the final value of any expressions 1832 // that are recurrent in the loop, and substitute the exit values from the 1833 // loop into any instructions outside of the loop that use the final values of 1834 // the current expressions. 1835 // 1836 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount)) 1837 RewriteLoopExitValues(L, Rewriter); 1838 1839 // Eliminate redundant IV users. 1840 if (EnableIVRewrite) 1841 Changed |= simplifyIVUsers(IU, SE, &LPM, DeadInsts); 1842 1843 // Eliminate redundant IV cycles. 1844 if (!EnableIVRewrite) 1845 NumElimIV += Rewriter.replaceCongruentIVs(L, DT, DeadInsts); 1846 1847 // Compute the type of the largest recurrence expression, and decide whether 1848 // a canonical induction variable should be inserted. 1849 Type *LargestType = 0; 1850 bool NeedCannIV = false; 1851 bool ExpandBECount = canExpandBackedgeTakenCount(L, SE); 1852 if (EnableIVRewrite && ExpandBECount) { 1853 // If we have a known trip count and a single exit block, we'll be 1854 // rewriting the loop exit test condition below, which requires a 1855 // canonical induction variable. 1856 NeedCannIV = true; 1857 Type *Ty = BackedgeTakenCount->getType(); 1858 if (!EnableIVRewrite) { 1859 // In this mode, SimplifyIVUsers may have already widened the IV used by 1860 // the backedge test and inserted a Trunc on the compare's operand. Get 1861 // the wider type to avoid creating a redundant narrow IV only used by the 1862 // loop test. 1863 LargestType = getBackedgeIVType(L); 1864 } 1865 if (!LargestType || 1866 SE->getTypeSizeInBits(Ty) > 1867 SE->getTypeSizeInBits(LargestType)) 1868 LargestType = SE->getEffectiveSCEVType(Ty); 1869 } 1870 if (EnableIVRewrite) { 1871 for (IVUsers::const_iterator I = IU->begin(), E = IU->end(); I != E; ++I) { 1872 NeedCannIV = true; 1873 Type *Ty = 1874 SE->getEffectiveSCEVType(I->getOperandValToReplace()->getType()); 1875 if (!LargestType || 1876 SE->getTypeSizeInBits(Ty) > 1877 SE->getTypeSizeInBits(LargestType)) 1878 LargestType = Ty; 1879 } 1880 } 1881 1882 // Now that we know the largest of the induction variable expressions 1883 // in this loop, insert a canonical induction variable of the largest size. 1884 PHINode *IndVar = 0; 1885 if (NeedCannIV) { 1886 // Check to see if the loop already has any canonical-looking induction 1887 // variables. If any are present and wider than the planned canonical 1888 // induction variable, temporarily remove them, so that the Rewriter 1889 // doesn't attempt to reuse them. 1890 SmallVector<PHINode *, 2> OldCannIVs; 1891 while (PHINode *OldCannIV = L->getCanonicalInductionVariable()) { 1892 if (SE->getTypeSizeInBits(OldCannIV->getType()) > 1893 SE->getTypeSizeInBits(LargestType)) 1894 OldCannIV->removeFromParent(); 1895 else 1896 break; 1897 OldCannIVs.push_back(OldCannIV); 1898 } 1899 1900 IndVar = Rewriter.getOrInsertCanonicalInductionVariable(L, LargestType); 1901 1902 ++NumInserted; 1903 Changed = true; 1904 DEBUG(dbgs() << "INDVARS: New CanIV: " << *IndVar << '\n'); 1905 1906 // Now that the official induction variable is established, reinsert 1907 // any old canonical-looking variables after it so that the IR remains 1908 // consistent. They will be deleted as part of the dead-PHI deletion at 1909 // the end of the pass. 1910 while (!OldCannIVs.empty()) { 1911 PHINode *OldCannIV = OldCannIVs.pop_back_val(); 1912 OldCannIV->insertBefore(L->getHeader()->getFirstInsertionPt()); 1913 } 1914 } 1915 else if (!EnableIVRewrite && ExpandBECount && needsLFTR(L, DT)) { 1916 IndVar = FindLoopCounter(L, BackedgeTakenCount, SE, DT, TD); 1917 } 1918 // If we have a trip count expression, rewrite the loop's exit condition 1919 // using it. We can currently only handle loops with a single exit. 1920 Value *NewICmp = 0; 1921 if (ExpandBECount && IndVar) { 1922 // Check preconditions for proper SCEVExpander operation. SCEV does not 1923 // express SCEVExpander's dependencies, such as LoopSimplify. Instead any 1924 // pass that uses the SCEVExpander must do it. This does not work well for 1925 // loop passes because SCEVExpander makes assumptions about all loops, while 1926 // LoopPassManager only forces the current loop to be simplified. 1927 // 1928 // FIXME: SCEV expansion has no way to bail out, so the caller must 1929 // explicitly check any assumptions made by SCEV. Brittle. 1930 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(BackedgeTakenCount); 1931 if (!AR || AR->getLoop()->getLoopPreheader()) 1932 NewICmp = 1933 LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar, Rewriter); 1934 } 1935 // Rewrite IV-derived expressions. 1936 if (EnableIVRewrite) 1937 RewriteIVExpressions(L, Rewriter); 1938 1939 // Clear the rewriter cache, because values that are in the rewriter's cache 1940 // can be deleted in the loop below, causing the AssertingVH in the cache to 1941 // trigger. 1942 Rewriter.clear(); 1943 1944 // Now that we're done iterating through lists, clean up any instructions 1945 // which are now dead. 1946 while (!DeadInsts.empty()) 1947 if (Instruction *Inst = 1948 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val())) 1949 RecursivelyDeleteTriviallyDeadInstructions(Inst); 1950 1951 // The Rewriter may not be used from this point on. 1952 1953 // Loop-invariant instructions in the preheader that aren't used in the 1954 // loop may be sunk below the loop to reduce register pressure. 1955 SinkUnusedInvariants(L); 1956 1957 // For completeness, inform IVUsers of the IV use in the newly-created 1958 // loop exit test instruction. 1959 if (IU && NewICmp) { 1960 ICmpInst *NewICmpInst = dyn_cast<ICmpInst>(NewICmp); 1961 if (NewICmpInst) 1962 IU->AddUsersIfInteresting(cast<Instruction>(NewICmpInst->getOperand(0))); 1963 } 1964 // Clean up dead instructions. 1965 Changed |= DeleteDeadPHIs(L->getHeader()); 1966 // Check a post-condition. 1967 assert(L->isLCSSAForm(*DT) && 1968 "Indvars did not leave the loop in lcssa form!"); 1969 1970 // Verify that LFTR, and any other change have not interfered with SCEV's 1971 // ability to compute trip count. 1972#ifndef NDEBUG 1973 if (!EnableIVRewrite && VerifyIndvars && 1974 !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) { 1975 SE->forgetLoop(L); 1976 const SCEV *NewBECount = SE->getBackedgeTakenCount(L); 1977 if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) < 1978 SE->getTypeSizeInBits(NewBECount->getType())) 1979 NewBECount = SE->getTruncateOrNoop(NewBECount, 1980 BackedgeTakenCount->getType()); 1981 else 1982 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount, 1983 NewBECount->getType()); 1984 assert(BackedgeTakenCount == NewBECount && "indvars must preserve SCEV"); 1985 } 1986#endif 1987 1988 return Changed; 1989} 1990