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 #include "llvm/Transforms/Scalar.h" 28 #include "llvm/ADT/DenseMap.h" 29 #include "llvm/ADT/SmallVector.h" 30 #include "llvm/ADT/Statistic.h" 31 #include "llvm/Analysis/GlobalsModRef.h" 32 #include "llvm/Analysis/LoopInfo.h" 33 #include "llvm/Analysis/LoopPass.h" 34 #include "llvm/Analysis/ScalarEvolutionExpander.h" 35 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h" 36 #include "llvm/Analysis/TargetLibraryInfo.h" 37 #include "llvm/Analysis/TargetTransformInfo.h" 38 #include "llvm/IR/BasicBlock.h" 39 #include "llvm/IR/CFG.h" 40 #include "llvm/IR/Constants.h" 41 #include "llvm/IR/DataLayout.h" 42 #include "llvm/IR/Dominators.h" 43 #include "llvm/IR/Instructions.h" 44 #include "llvm/IR/IntrinsicInst.h" 45 #include "llvm/IR/LLVMContext.h" 46 #include "llvm/IR/PatternMatch.h" 47 #include "llvm/IR/Type.h" 48 #include "llvm/Support/CommandLine.h" 49 #include "llvm/Support/Debug.h" 50 #include "llvm/Support/raw_ostream.h" 51 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 52 #include "llvm/Transforms/Utils/Local.h" 53 #include "llvm/Transforms/Utils/LoopUtils.h" 54 #include "llvm/Transforms/Utils/SimplifyIndVar.h" 55 using namespace llvm; 56 57 #define DEBUG_TYPE "indvars" 58 59 STATISTIC(NumWidened , "Number of indvars widened"); 60 STATISTIC(NumReplaced , "Number of exit values replaced"); 61 STATISTIC(NumLFTR , "Number of loop exit tests replaced"); 62 STATISTIC(NumElimExt , "Number of IV sign/zero extends eliminated"); 63 STATISTIC(NumElimIV , "Number of congruent IVs eliminated"); 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. 68 static cl::opt<bool> VerifyIndvars( 69 "verify-indvars", cl::Hidden, 70 cl::desc("Verify the ScalarEvolution result after running indvars")); 71 72 static cl::opt<bool> ReduceLiveIVs("liv-reduce", cl::Hidden, 73 cl::desc("Reduce live induction variables.")); 74 75 enum ReplaceExitVal { NeverRepl, OnlyCheapRepl, AlwaysRepl }; 76 77 static cl::opt<ReplaceExitVal> ReplaceExitValue( 78 "replexitval", cl::Hidden, cl::init(OnlyCheapRepl), 79 cl::desc("Choose the strategy to replace exit value in IndVarSimplify"), 80 cl::values(clEnumValN(NeverRepl, "never", "never replace exit value"), 81 clEnumValN(OnlyCheapRepl, "cheap", 82 "only replace exit value when the cost is cheap"), 83 clEnumValN(AlwaysRepl, "always", 84 "always replace exit value whenever possible"), 85 clEnumValEnd)); 86 87 namespace { 88 struct RewritePhi; 89 90 class IndVarSimplify : public LoopPass { 91 LoopInfo *LI; 92 ScalarEvolution *SE; 93 DominatorTree *DT; 94 TargetLibraryInfo *TLI; 95 const TargetTransformInfo *TTI; 96 97 SmallVector<WeakVH, 16> DeadInsts; 98 bool Changed; 99 public: 100 101 static char ID; // Pass identification, replacement for typeid 102 IndVarSimplify() 103 : LoopPass(ID), LI(nullptr), SE(nullptr), DT(nullptr), Changed(false) { 104 initializeIndVarSimplifyPass(*PassRegistry::getPassRegistry()); 105 } 106 107 bool runOnLoop(Loop *L, LPPassManager &LPM) override; 108 109 void getAnalysisUsage(AnalysisUsage &AU) const override { 110 AU.addRequired<DominatorTreeWrapperPass>(); 111 AU.addRequired<LoopInfoWrapperPass>(); 112 AU.addRequired<ScalarEvolutionWrapperPass>(); 113 AU.addRequiredID(LoopSimplifyID); 114 AU.addRequiredID(LCSSAID); 115 AU.addPreserved<GlobalsAAWrapperPass>(); 116 AU.addPreserved<ScalarEvolutionWrapperPass>(); 117 AU.addPreservedID(LoopSimplifyID); 118 AU.addPreservedID(LCSSAID); 119 AU.setPreservesCFG(); 120 } 121 122 private: 123 void releaseMemory() override { 124 DeadInsts.clear(); 125 } 126 127 bool isValidRewrite(Value *FromVal, Value *ToVal); 128 129 void handleFloatingPointIV(Loop *L, PHINode *PH); 130 void rewriteNonIntegerIVs(Loop *L); 131 132 void simplifyAndExtend(Loop *L, SCEVExpander &Rewriter, LoopInfo *LI); 133 134 bool canLoopBeDeleted(Loop *L, SmallVector<RewritePhi, 8> &RewritePhiSet); 135 void rewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter); 136 137 Value *linearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount, 138 PHINode *IndVar, SCEVExpander &Rewriter); 139 140 void sinkUnusedInvariants(Loop *L); 141 142 Value *expandSCEVIfNeeded(SCEVExpander &Rewriter, const SCEV *S, Loop *L, 143 Instruction *InsertPt, Type *Ty); 144 }; 145 } 146 147 char IndVarSimplify::ID = 0; 148 INITIALIZE_PASS_BEGIN(IndVarSimplify, "indvars", 149 "Induction Variable Simplification", false, false) 150 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 151 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) 152 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass) 153 INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 154 INITIALIZE_PASS_DEPENDENCY(LCSSA) 155 INITIALIZE_PASS_END(IndVarSimplify, "indvars", 156 "Induction Variable Simplification", false, false) 157 158 Pass *llvm::createIndVarSimplifyPass() { 159 return new IndVarSimplify(); 160 } 161 162 /// Return true if the SCEV expansion generated by the rewriter can replace the 163 /// original value. SCEV guarantees that it produces the same value, but the way 164 /// it is produced may be illegal IR. Ideally, this function will only be 165 /// called for verification. 166 bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) { 167 // If an SCEV expression subsumed multiple pointers, its expansion could 168 // reassociate the GEP changing the base pointer. This is illegal because the 169 // final address produced by a GEP chain must be inbounds relative to its 170 // underlying object. Otherwise basic alias analysis, among other things, 171 // could fail in a dangerous way. Ultimately, SCEV will be improved to avoid 172 // producing an expression involving multiple pointers. Until then, we must 173 // bail out here. 174 // 175 // Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject 176 // because it understands lcssa phis while SCEV does not. 177 Value *FromPtr = FromVal; 178 Value *ToPtr = ToVal; 179 if (auto *GEP = dyn_cast<GEPOperator>(FromVal)) { 180 FromPtr = GEP->getPointerOperand(); 181 } 182 if (auto *GEP = dyn_cast<GEPOperator>(ToVal)) { 183 ToPtr = GEP->getPointerOperand(); 184 } 185 if (FromPtr != FromVal || ToPtr != ToVal) { 186 // Quickly check the common case 187 if (FromPtr == ToPtr) 188 return true; 189 190 // SCEV may have rewritten an expression that produces the GEP's pointer 191 // operand. That's ok as long as the pointer operand has the same base 192 // pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the 193 // base of a recurrence. This handles the case in which SCEV expansion 194 // converts a pointer type recurrence into a nonrecurrent pointer base 195 // indexed by an integer recurrence. 196 197 // If the GEP base pointer is a vector of pointers, abort. 198 if (!FromPtr->getType()->isPointerTy() || !ToPtr->getType()->isPointerTy()) 199 return false; 200 201 const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr)); 202 const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr)); 203 if (FromBase == ToBase) 204 return true; 205 206 DEBUG(dbgs() << "INDVARS: GEP rewrite bail out " 207 << *FromBase << " != " << *ToBase << "\n"); 208 209 return false; 210 } 211 return true; 212 } 213 214 /// Determine the insertion point for this user. By default, insert immediately 215 /// before the user. SCEVExpander or LICM will hoist loop invariants out of the 216 /// loop. For PHI nodes, there may be multiple uses, so compute the nearest 217 /// common dominator for the incoming blocks. 218 static Instruction *getInsertPointForUses(Instruction *User, Value *Def, 219 DominatorTree *DT, LoopInfo *LI) { 220 PHINode *PHI = dyn_cast<PHINode>(User); 221 if (!PHI) 222 return User; 223 224 Instruction *InsertPt = nullptr; 225 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) { 226 if (PHI->getIncomingValue(i) != Def) 227 continue; 228 229 BasicBlock *InsertBB = PHI->getIncomingBlock(i); 230 if (!InsertPt) { 231 InsertPt = InsertBB->getTerminator(); 232 continue; 233 } 234 InsertBB = DT->findNearestCommonDominator(InsertPt->getParent(), InsertBB); 235 InsertPt = InsertBB->getTerminator(); 236 } 237 assert(InsertPt && "Missing phi operand"); 238 239 auto *DefI = dyn_cast<Instruction>(Def); 240 if (!DefI) 241 return InsertPt; 242 243 assert(DT->dominates(DefI, InsertPt) && "def does not dominate all uses"); 244 245 auto *L = LI->getLoopFor(DefI->getParent()); 246 assert(!L || L->contains(LI->getLoopFor(InsertPt->getParent()))); 247 248 for (auto *DTN = (*DT)[InsertPt->getParent()]; DTN; DTN = DTN->getIDom()) 249 if (LI->getLoopFor(DTN->getBlock()) == L) 250 return DTN->getBlock()->getTerminator(); 251 252 llvm_unreachable("DefI dominates InsertPt!"); 253 } 254 255 //===----------------------------------------------------------------------===// 256 // rewriteNonIntegerIVs and helpers. Prefer integer IVs. 257 //===----------------------------------------------------------------------===// 258 259 /// Convert APF to an integer, if possible. 260 static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) { 261 bool isExact = false; 262 // See if we can convert this to an int64_t 263 uint64_t UIntVal; 264 if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero, 265 &isExact) != APFloat::opOK || !isExact) 266 return false; 267 IntVal = UIntVal; 268 return true; 269 } 270 271 /// If the loop has floating induction variable then insert corresponding 272 /// integer induction variable if possible. 273 /// For example, 274 /// for(double i = 0; i < 10000; ++i) 275 /// bar(i) 276 /// is converted into 277 /// for(int i = 0; i < 10000; ++i) 278 /// bar((double)i); 279 /// 280 void IndVarSimplify::handleFloatingPointIV(Loop *L, PHINode *PN) { 281 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0)); 282 unsigned BackEdge = IncomingEdge^1; 283 284 // Check incoming value. 285 auto *InitValueVal = dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge)); 286 287 int64_t InitValue; 288 if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue)) 289 return; 290 291 // Check IV increment. Reject this PN if increment operation is not 292 // an add or increment value can not be represented by an integer. 293 auto *Incr = dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge)); 294 if (Incr == nullptr || Incr->getOpcode() != Instruction::FAdd) return; 295 296 // If this is not an add of the PHI with a constantfp, or if the constant fp 297 // is not an integer, bail out. 298 ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1)); 299 int64_t IncValue; 300 if (IncValueVal == nullptr || Incr->getOperand(0) != PN || 301 !ConvertToSInt(IncValueVal->getValueAPF(), IncValue)) 302 return; 303 304 // Check Incr uses. One user is PN and the other user is an exit condition 305 // used by the conditional terminator. 306 Value::user_iterator IncrUse = Incr->user_begin(); 307 Instruction *U1 = cast<Instruction>(*IncrUse++); 308 if (IncrUse == Incr->user_end()) return; 309 Instruction *U2 = cast<Instruction>(*IncrUse++); 310 if (IncrUse != Incr->user_end()) return; 311 312 // Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't 313 // only used by a branch, we can't transform it. 314 FCmpInst *Compare = dyn_cast<FCmpInst>(U1); 315 if (!Compare) 316 Compare = dyn_cast<FCmpInst>(U2); 317 if (!Compare || !Compare->hasOneUse() || 318 !isa<BranchInst>(Compare->user_back())) 319 return; 320 321 BranchInst *TheBr = cast<BranchInst>(Compare->user_back()); 322 323 // We need to verify that the branch actually controls the iteration count 324 // of the loop. If not, the new IV can overflow and no one will notice. 325 // The branch block must be in the loop and one of the successors must be out 326 // of the loop. 327 assert(TheBr->isConditional() && "Can't use fcmp if not conditional"); 328 if (!L->contains(TheBr->getParent()) || 329 (L->contains(TheBr->getSuccessor(0)) && 330 L->contains(TheBr->getSuccessor(1)))) 331 return; 332 333 334 // If it isn't a comparison with an integer-as-fp (the exit value), we can't 335 // transform it. 336 ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1)); 337 int64_t ExitValue; 338 if (ExitValueVal == nullptr || 339 !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue)) 340 return; 341 342 // Find new predicate for integer comparison. 343 CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE; 344 switch (Compare->getPredicate()) { 345 default: return; // Unknown comparison. 346 case CmpInst::FCMP_OEQ: 347 case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break; 348 case CmpInst::FCMP_ONE: 349 case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break; 350 case CmpInst::FCMP_OGT: 351 case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break; 352 case CmpInst::FCMP_OGE: 353 case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break; 354 case CmpInst::FCMP_OLT: 355 case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break; 356 case CmpInst::FCMP_OLE: 357 case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break; 358 } 359 360 // We convert the floating point induction variable to a signed i32 value if 361 // we can. This is only safe if the comparison will not overflow in a way 362 // that won't be trapped by the integer equivalent operations. Check for this 363 // now. 364 // TODO: We could use i64 if it is native and the range requires it. 365 366 // The start/stride/exit values must all fit in signed i32. 367 if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue)) 368 return; 369 370 // If not actually striding (add x, 0.0), avoid touching the code. 371 if (IncValue == 0) 372 return; 373 374 // Positive and negative strides have different safety conditions. 375 if (IncValue > 0) { 376 // If we have a positive stride, we require the init to be less than the 377 // exit value. 378 if (InitValue >= ExitValue) 379 return; 380 381 uint32_t Range = uint32_t(ExitValue-InitValue); 382 // Check for infinite loop, either: 383 // while (i <= Exit) or until (i > Exit) 384 if (NewPred == CmpInst::ICMP_SLE || NewPred == CmpInst::ICMP_SGT) { 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 } else { 403 // If we have a negative stride, we require the init to be greater than the 404 // exit value. 405 if (InitValue <= ExitValue) 406 return; 407 408 uint32_t Range = uint32_t(InitValue-ExitValue); 409 // Check for infinite loop, either: 410 // while (i >= Exit) or until (i < Exit) 411 if (NewPred == CmpInst::ICMP_SGE || NewPred == CmpInst::ICMP_SLT) { 412 if (++Range == 0) return; // Range overflows. 413 } 414 415 unsigned Leftover = Range % uint32_t(-IncValue); 416 417 // If this is an equality comparison, we require that the strided value 418 // exactly land on the exit value, otherwise the IV condition will wrap 419 // around and do things the fp IV wouldn't. 420 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) && 421 Leftover != 0) 422 return; 423 424 // If the stride would wrap around the i32 before exiting, we can't 425 // transform the IV. 426 if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue) 427 return; 428 } 429 430 IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext()); 431 432 // Insert new integer induction variable. 433 PHINode *NewPHI = PHINode::Create(Int32Ty, 2, PN->getName()+".int", PN); 434 NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue), 435 PN->getIncomingBlock(IncomingEdge)); 436 437 Value *NewAdd = 438 BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue), 439 Incr->getName()+".int", Incr); 440 NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge)); 441 442 ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd, 443 ConstantInt::get(Int32Ty, ExitValue), 444 Compare->getName()); 445 446 // In the following deletions, PN may become dead and may be deleted. 447 // Use a WeakVH to observe whether this happens. 448 WeakVH WeakPH = PN; 449 450 // Delete the old floating point exit comparison. The branch starts using the 451 // new comparison. 452 NewCompare->takeName(Compare); 453 Compare->replaceAllUsesWith(NewCompare); 454 RecursivelyDeleteTriviallyDeadInstructions(Compare, TLI); 455 456 // Delete the old floating point increment. 457 Incr->replaceAllUsesWith(UndefValue::get(Incr->getType())); 458 RecursivelyDeleteTriviallyDeadInstructions(Incr, TLI); 459 460 // If the FP induction variable still has uses, this is because something else 461 // in the loop uses its value. In order to canonicalize the induction 462 // variable, we chose to eliminate the IV and rewrite it in terms of an 463 // int->fp cast. 464 // 465 // We give preference to sitofp over uitofp because it is faster on most 466 // platforms. 467 if (WeakPH) { 468 Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv", 469 &*PN->getParent()->getFirstInsertionPt()); 470 PN->replaceAllUsesWith(Conv); 471 RecursivelyDeleteTriviallyDeadInstructions(PN, TLI); 472 } 473 Changed = true; 474 } 475 476 void IndVarSimplify::rewriteNonIntegerIVs(Loop *L) { 477 // First step. Check to see if there are any floating-point recurrences. 478 // If there are, change them into integer recurrences, permitting analysis by 479 // the SCEV routines. 480 // 481 BasicBlock *Header = L->getHeader(); 482 483 SmallVector<WeakVH, 8> PHIs; 484 for (BasicBlock::iterator I = Header->begin(); 485 PHINode *PN = dyn_cast<PHINode>(I); ++I) 486 PHIs.push_back(PN); 487 488 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) 489 if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i])) 490 handleFloatingPointIV(L, PN); 491 492 // If the loop previously had floating-point IV, ScalarEvolution 493 // may not have been able to compute a trip count. Now that we've done some 494 // re-writing, the trip count may be computable. 495 if (Changed) 496 SE->forgetLoop(L); 497 } 498 499 namespace { 500 // Collect information about PHI nodes which can be transformed in 501 // rewriteLoopExitValues. 502 struct RewritePhi { 503 PHINode *PN; 504 unsigned Ith; // Ith incoming value. 505 Value *Val; // Exit value after expansion. 506 bool HighCost; // High Cost when expansion. 507 bool SafePhi; // LCSSASafePhiForRAUW. 508 509 RewritePhi(PHINode *P, unsigned I, Value *V, bool H, bool S) 510 : PN(P), Ith(I), Val(V), HighCost(H), SafePhi(S) {} 511 }; 512 } 513 514 Value *IndVarSimplify::expandSCEVIfNeeded(SCEVExpander &Rewriter, const SCEV *S, 515 Loop *L, Instruction *InsertPt, 516 Type *ResultTy) { 517 // Before expanding S into an expensive LLVM expression, see if we can use an 518 // already existing value as the expansion for S. 519 if (Value *ExistingValue = Rewriter.findExistingExpansion(S, InsertPt, L)) 520 if (ExistingValue->getType() == ResultTy) 521 return ExistingValue; 522 523 // We didn't find anything, fall back to using SCEVExpander. 524 return Rewriter.expandCodeFor(S, ResultTy, InsertPt); 525 } 526 527 //===----------------------------------------------------------------------===// 528 // rewriteLoopExitValues - Optimize IV users outside the loop. 529 // As a side effect, reduces the amount of IV processing within the loop. 530 //===----------------------------------------------------------------------===// 531 532 /// Check to see if this loop has a computable loop-invariant execution count. 533 /// If so, this means that we can compute the final value of any expressions 534 /// that are recurrent in the loop, and substitute the exit values from the loop 535 /// into any instructions outside of the loop that use the final values of the 536 /// current expressions. 537 /// 538 /// This is mostly redundant with the regular IndVarSimplify activities that 539 /// happen later, except that it's more powerful in some cases, because it's 540 /// able to brute-force evaluate arbitrary instructions as long as they have 541 /// constant operands at the beginning of the loop. 542 void IndVarSimplify::rewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) { 543 // Check a pre-condition. 544 assert(L->isRecursivelyLCSSAForm(*DT) && "Indvars did not preserve LCSSA!"); 545 546 SmallVector<BasicBlock*, 8> ExitBlocks; 547 L->getUniqueExitBlocks(ExitBlocks); 548 549 SmallVector<RewritePhi, 8> RewritePhiSet; 550 // Find all values that are computed inside the loop, but used outside of it. 551 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan 552 // the exit blocks of the loop to find them. 553 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 554 BasicBlock *ExitBB = ExitBlocks[i]; 555 556 // If there are no PHI nodes in this exit block, then no values defined 557 // inside the loop are used on this path, skip it. 558 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin()); 559 if (!PN) continue; 560 561 unsigned NumPreds = PN->getNumIncomingValues(); 562 563 // We would like to be able to RAUW single-incoming value PHI nodes. We 564 // have to be certain this is safe even when this is an LCSSA PHI node. 565 // While the computed exit value is no longer varying in *this* loop, the 566 // exit block may be an exit block for an outer containing loop as well, 567 // the exit value may be varying in the outer loop, and thus it may still 568 // require an LCSSA PHI node. The safe case is when this is 569 // single-predecessor PHI node (LCSSA) and the exit block containing it is 570 // part of the enclosing loop, or this is the outer most loop of the nest. 571 // In either case the exit value could (at most) be varying in the same 572 // loop body as the phi node itself. Thus if it is in turn used outside of 573 // an enclosing loop it will only be via a separate LCSSA node. 574 bool LCSSASafePhiForRAUW = 575 NumPreds == 1 && 576 (!L->getParentLoop() || L->getParentLoop() == LI->getLoopFor(ExitBB)); 577 578 // Iterate over all of the PHI nodes. 579 BasicBlock::iterator BBI = ExitBB->begin(); 580 while ((PN = dyn_cast<PHINode>(BBI++))) { 581 if (PN->use_empty()) 582 continue; // dead use, don't replace it 583 584 // SCEV only supports integer expressions for now. 585 if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy()) 586 continue; 587 588 // It's necessary to tell ScalarEvolution about this explicitly so that 589 // it can walk the def-use list and forget all SCEVs, as it may not be 590 // watching the PHI itself. Once the new exit value is in place, there 591 // may not be a def-use connection between the loop and every instruction 592 // which got a SCEVAddRecExpr for that loop. 593 SE->forgetValue(PN); 594 595 // Iterate over all of the values in all the PHI nodes. 596 for (unsigned i = 0; i != NumPreds; ++i) { 597 // If the value being merged in is not integer or is not defined 598 // in the loop, skip it. 599 Value *InVal = PN->getIncomingValue(i); 600 if (!isa<Instruction>(InVal)) 601 continue; 602 603 // If this pred is for a subloop, not L itself, skip it. 604 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L) 605 continue; // The Block is in a subloop, skip it. 606 607 // Check that InVal is defined in the loop. 608 Instruction *Inst = cast<Instruction>(InVal); 609 if (!L->contains(Inst)) 610 continue; 611 612 // Okay, this instruction has a user outside of the current loop 613 // and varies predictably *inside* the loop. Evaluate the value it 614 // contains when the loop exits, if possible. 615 const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop()); 616 if (!SE->isLoopInvariant(ExitValue, L) || 617 !isSafeToExpand(ExitValue, *SE)) 618 continue; 619 620 // Computing the value outside of the loop brings no benefit if : 621 // - it is definitely used inside the loop in a way which can not be 622 // optimized away. 623 // - no use outside of the loop can take advantage of hoisting the 624 // computation out of the loop 625 if (ExitValue->getSCEVType()>=scMulExpr) { 626 unsigned NumHardInternalUses = 0; 627 unsigned NumSoftExternalUses = 0; 628 unsigned NumUses = 0; 629 for (auto IB = Inst->user_begin(), IE = Inst->user_end(); 630 IB != IE && NumUses <= 6; ++IB) { 631 Instruction *UseInstr = cast<Instruction>(*IB); 632 unsigned Opc = UseInstr->getOpcode(); 633 NumUses++; 634 if (L->contains(UseInstr)) { 635 if (Opc == Instruction::Call || Opc == Instruction::Ret) 636 NumHardInternalUses++; 637 } else { 638 if (Opc == Instruction::PHI) { 639 // Do not count the Phi as a use. LCSSA may have inserted 640 // plenty of trivial ones. 641 NumUses--; 642 for (auto PB = UseInstr->user_begin(), 643 PE = UseInstr->user_end(); 644 PB != PE && NumUses <= 6; ++PB, ++NumUses) { 645 unsigned PhiOpc = cast<Instruction>(*PB)->getOpcode(); 646 if (PhiOpc != Instruction::Call && PhiOpc != Instruction::Ret) 647 NumSoftExternalUses++; 648 } 649 continue; 650 } 651 if (Opc != Instruction::Call && Opc != Instruction::Ret) 652 NumSoftExternalUses++; 653 } 654 } 655 if (NumUses <= 6 && NumHardInternalUses && !NumSoftExternalUses) 656 continue; 657 } 658 659 bool HighCost = Rewriter.isHighCostExpansion(ExitValue, L, Inst); 660 Value *ExitVal = 661 expandSCEVIfNeeded(Rewriter, ExitValue, L, Inst, PN->getType()); 662 663 DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n' 664 << " LoopVal = " << *Inst << "\n"); 665 666 if (!isValidRewrite(Inst, ExitVal)) { 667 DeadInsts.push_back(ExitVal); 668 continue; 669 } 670 671 // Collect all the candidate PHINodes to be rewritten. 672 RewritePhiSet.push_back( 673 RewritePhi(PN, i, ExitVal, HighCost, LCSSASafePhiForRAUW)); 674 } 675 } 676 } 677 678 bool LoopCanBeDel = canLoopBeDeleted(L, RewritePhiSet); 679 680 // Transformation. 681 for (const RewritePhi &Phi : RewritePhiSet) { 682 PHINode *PN = Phi.PN; 683 Value *ExitVal = Phi.Val; 684 685 // Only do the rewrite when the ExitValue can be expanded cheaply. 686 // If LoopCanBeDel is true, rewrite exit value aggressively. 687 if (ReplaceExitValue == OnlyCheapRepl && !LoopCanBeDel && Phi.HighCost) { 688 DeadInsts.push_back(ExitVal); 689 continue; 690 } 691 692 Changed = true; 693 ++NumReplaced; 694 Instruction *Inst = cast<Instruction>(PN->getIncomingValue(Phi.Ith)); 695 PN->setIncomingValue(Phi.Ith, ExitVal); 696 697 // If this instruction is dead now, delete it. Don't do it now to avoid 698 // invalidating iterators. 699 if (isInstructionTriviallyDead(Inst, TLI)) 700 DeadInsts.push_back(Inst); 701 702 // If we determined that this PHI is safe to replace even if an LCSSA 703 // PHI, do so. 704 if (Phi.SafePhi) { 705 PN->replaceAllUsesWith(ExitVal); 706 PN->eraseFromParent(); 707 } 708 } 709 710 // The insertion point instruction may have been deleted; clear it out 711 // so that the rewriter doesn't trip over it later. 712 Rewriter.clearInsertPoint(); 713 } 714 715 /// Check whether it is possible to delete the loop after rewriting exit 716 /// value. If it is possible, ignore ReplaceExitValue and do rewriting 717 /// aggressively. 718 bool IndVarSimplify::canLoopBeDeleted( 719 Loop *L, SmallVector<RewritePhi, 8> &RewritePhiSet) { 720 721 BasicBlock *Preheader = L->getLoopPreheader(); 722 // If there is no preheader, the loop will not be deleted. 723 if (!Preheader) 724 return false; 725 726 // In LoopDeletion pass Loop can be deleted when ExitingBlocks.size() > 1. 727 // We obviate multiple ExitingBlocks case for simplicity. 728 // TODO: If we see testcase with multiple ExitingBlocks can be deleted 729 // after exit value rewriting, we can enhance the logic here. 730 SmallVector<BasicBlock *, 4> ExitingBlocks; 731 L->getExitingBlocks(ExitingBlocks); 732 SmallVector<BasicBlock *, 8> ExitBlocks; 733 L->getUniqueExitBlocks(ExitBlocks); 734 if (ExitBlocks.size() > 1 || ExitingBlocks.size() > 1) 735 return false; 736 737 BasicBlock *ExitBlock = ExitBlocks[0]; 738 BasicBlock::iterator BI = ExitBlock->begin(); 739 while (PHINode *P = dyn_cast<PHINode>(BI)) { 740 Value *Incoming = P->getIncomingValueForBlock(ExitingBlocks[0]); 741 742 // If the Incoming value of P is found in RewritePhiSet, we know it 743 // could be rewritten to use a loop invariant value in transformation 744 // phase later. Skip it in the loop invariant check below. 745 bool found = false; 746 for (const RewritePhi &Phi : RewritePhiSet) { 747 unsigned i = Phi.Ith; 748 if (Phi.PN == P && (Phi.PN)->getIncomingValue(i) == Incoming) { 749 found = true; 750 break; 751 } 752 } 753 754 Instruction *I; 755 if (!found && (I = dyn_cast<Instruction>(Incoming))) 756 if (!L->hasLoopInvariantOperands(I)) 757 return false; 758 759 ++BI; 760 } 761 762 for (auto *BB : L->blocks()) 763 if (any_of(*BB, [](Instruction &I) { return I.mayHaveSideEffects(); })) 764 return false; 765 766 return true; 767 } 768 769 //===----------------------------------------------------------------------===// 770 // IV Widening - Extend the width of an IV to cover its widest uses. 771 //===----------------------------------------------------------------------===// 772 773 namespace { 774 // Collect information about induction variables that are used by sign/zero 775 // extend operations. This information is recorded by CollectExtend and provides 776 // the input to WidenIV. 777 struct WideIVInfo { 778 PHINode *NarrowIV = nullptr; 779 Type *WidestNativeType = nullptr; // Widest integer type created [sz]ext 780 bool IsSigned = false; // Was a sext user seen before a zext? 781 }; 782 } 783 784 /// Update information about the induction variable that is extended by this 785 /// sign or zero extend operation. This is used to determine the final width of 786 /// the IV before actually widening it. 787 static void visitIVCast(CastInst *Cast, WideIVInfo &WI, ScalarEvolution *SE, 788 const TargetTransformInfo *TTI) { 789 bool IsSigned = Cast->getOpcode() == Instruction::SExt; 790 if (!IsSigned && Cast->getOpcode() != Instruction::ZExt) 791 return; 792 793 Type *Ty = Cast->getType(); 794 uint64_t Width = SE->getTypeSizeInBits(Ty); 795 if (!Cast->getModule()->getDataLayout().isLegalInteger(Width)) 796 return; 797 798 // Cast is either an sext or zext up to this point. 799 // We should not widen an indvar if arithmetics on the wider indvar are more 800 // expensive than those on the narrower indvar. We check only the cost of ADD 801 // because at least an ADD is required to increment the induction variable. We 802 // could compute more comprehensively the cost of all instructions on the 803 // induction variable when necessary. 804 if (TTI && 805 TTI->getArithmeticInstrCost(Instruction::Add, Ty) > 806 TTI->getArithmeticInstrCost(Instruction::Add, 807 Cast->getOperand(0)->getType())) { 808 return; 809 } 810 811 if (!WI.WidestNativeType) { 812 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty); 813 WI.IsSigned = IsSigned; 814 return; 815 } 816 817 // We extend the IV to satisfy the sign of its first user, arbitrarily. 818 if (WI.IsSigned != IsSigned) 819 return; 820 821 if (Width > SE->getTypeSizeInBits(WI.WidestNativeType)) 822 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty); 823 } 824 825 namespace { 826 827 /// Record a link in the Narrow IV def-use chain along with the WideIV that 828 /// computes the same value as the Narrow IV def. This avoids caching Use* 829 /// pointers. 830 struct NarrowIVDefUse { 831 Instruction *NarrowDef = nullptr; 832 Instruction *NarrowUse = nullptr; 833 Instruction *WideDef = nullptr; 834 835 // True if the narrow def is never negative. Tracking this information lets 836 // us use a sign extension instead of a zero extension or vice versa, when 837 // profitable and legal. 838 bool NeverNegative = false; 839 840 NarrowIVDefUse(Instruction *ND, Instruction *NU, Instruction *WD, 841 bool NeverNegative) 842 : NarrowDef(ND), NarrowUse(NU), WideDef(WD), 843 NeverNegative(NeverNegative) {} 844 }; 845 846 /// The goal of this transform is to remove sign and zero extends without 847 /// creating any new induction variables. To do this, it creates a new phi of 848 /// the wider type and redirects all users, either removing extends or inserting 849 /// truncs whenever we stop propagating the type. 850 /// 851 class WidenIV { 852 // Parameters 853 PHINode *OrigPhi; 854 Type *WideType; 855 bool IsSigned; 856 857 // Context 858 LoopInfo *LI; 859 Loop *L; 860 ScalarEvolution *SE; 861 DominatorTree *DT; 862 863 // Result 864 PHINode *WidePhi; 865 Instruction *WideInc; 866 const SCEV *WideIncExpr; 867 SmallVectorImpl<WeakVH> &DeadInsts; 868 869 SmallPtrSet<Instruction*,16> Widened; 870 SmallVector<NarrowIVDefUse, 8> NarrowIVUsers; 871 872 public: 873 WidenIV(const WideIVInfo &WI, LoopInfo *LInfo, 874 ScalarEvolution *SEv, DominatorTree *DTree, 875 SmallVectorImpl<WeakVH> &DI) : 876 OrigPhi(WI.NarrowIV), 877 WideType(WI.WidestNativeType), 878 IsSigned(WI.IsSigned), 879 LI(LInfo), 880 L(LI->getLoopFor(OrigPhi->getParent())), 881 SE(SEv), 882 DT(DTree), 883 WidePhi(nullptr), 884 WideInc(nullptr), 885 WideIncExpr(nullptr), 886 DeadInsts(DI) { 887 assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV"); 888 } 889 890 PHINode *createWideIV(SCEVExpander &Rewriter); 891 892 protected: 893 Value *createExtendInst(Value *NarrowOper, Type *WideType, bool IsSigned, 894 Instruction *Use); 895 896 Instruction *cloneIVUser(NarrowIVDefUse DU, const SCEVAddRecExpr *WideAR); 897 Instruction *cloneArithmeticIVUser(NarrowIVDefUse DU, 898 const SCEVAddRecExpr *WideAR); 899 Instruction *cloneBitwiseIVUser(NarrowIVDefUse DU); 900 901 const SCEVAddRecExpr *getWideRecurrence(Instruction *NarrowUse); 902 903 const SCEVAddRecExpr* getExtendedOperandRecurrence(NarrowIVDefUse DU); 904 905 const SCEV *getSCEVByOpCode(const SCEV *LHS, const SCEV *RHS, 906 unsigned OpCode) const; 907 908 Instruction *widenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter); 909 910 bool widenLoopCompare(NarrowIVDefUse DU); 911 912 void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef); 913 }; 914 } // anonymous namespace 915 916 /// Perform a quick domtree based check for loop invariance assuming that V is 917 /// used within the loop. LoopInfo::isLoopInvariant() seems gratuitous for this 918 /// purpose. 919 static bool isLoopInvariant(Value *V, const Loop *L, const DominatorTree *DT) { 920 Instruction *Inst = dyn_cast<Instruction>(V); 921 if (!Inst) 922 return true; 923 924 return DT->properlyDominates(Inst->getParent(), L->getHeader()); 925 } 926 927 Value *WidenIV::createExtendInst(Value *NarrowOper, Type *WideType, 928 bool IsSigned, Instruction *Use) { 929 // Set the debug location and conservative insertion point. 930 IRBuilder<> Builder(Use); 931 // Hoist the insertion point into loop preheaders as far as possible. 932 for (const Loop *L = LI->getLoopFor(Use->getParent()); 933 L && L->getLoopPreheader() && isLoopInvariant(NarrowOper, L, DT); 934 L = L->getParentLoop()) 935 Builder.SetInsertPoint(L->getLoopPreheader()->getTerminator()); 936 937 return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) : 938 Builder.CreateZExt(NarrowOper, WideType); 939 } 940 941 /// Instantiate a wide operation to replace a narrow operation. This only needs 942 /// to handle operations that can evaluation to SCEVAddRec. It can safely return 943 /// 0 for any operation we decide not to clone. 944 Instruction *WidenIV::cloneIVUser(NarrowIVDefUse DU, 945 const SCEVAddRecExpr *WideAR) { 946 unsigned Opcode = DU.NarrowUse->getOpcode(); 947 switch (Opcode) { 948 default: 949 return nullptr; 950 case Instruction::Add: 951 case Instruction::Mul: 952 case Instruction::UDiv: 953 case Instruction::Sub: 954 return cloneArithmeticIVUser(DU, WideAR); 955 956 case Instruction::And: 957 case Instruction::Or: 958 case Instruction::Xor: 959 case Instruction::Shl: 960 case Instruction::LShr: 961 case Instruction::AShr: 962 return cloneBitwiseIVUser(DU); 963 } 964 } 965 966 Instruction *WidenIV::cloneBitwiseIVUser(NarrowIVDefUse DU) { 967 Instruction *NarrowUse = DU.NarrowUse; 968 Instruction *NarrowDef = DU.NarrowDef; 969 Instruction *WideDef = DU.WideDef; 970 971 DEBUG(dbgs() << "Cloning bitwise IVUser: " << *NarrowUse << "\n"); 972 973 // Replace NarrowDef operands with WideDef. Otherwise, we don't know anything 974 // about the narrow operand yet so must insert a [sz]ext. It is probably loop 975 // invariant and will be folded or hoisted. If it actually comes from a 976 // widened IV, it should be removed during a future call to widenIVUse. 977 Value *LHS = (NarrowUse->getOperand(0) == NarrowDef) 978 ? WideDef 979 : createExtendInst(NarrowUse->getOperand(0), WideType, 980 IsSigned, NarrowUse); 981 Value *RHS = (NarrowUse->getOperand(1) == NarrowDef) 982 ? WideDef 983 : createExtendInst(NarrowUse->getOperand(1), WideType, 984 IsSigned, NarrowUse); 985 986 auto *NarrowBO = cast<BinaryOperator>(NarrowUse); 987 auto *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(), LHS, RHS, 988 NarrowBO->getName()); 989 IRBuilder<> Builder(NarrowUse); 990 Builder.Insert(WideBO); 991 WideBO->copyIRFlags(NarrowBO); 992 return WideBO; 993 } 994 995 Instruction *WidenIV::cloneArithmeticIVUser(NarrowIVDefUse DU, 996 const SCEVAddRecExpr *WideAR) { 997 Instruction *NarrowUse = DU.NarrowUse; 998 Instruction *NarrowDef = DU.NarrowDef; 999 Instruction *WideDef = DU.WideDef; 1000 1001 DEBUG(dbgs() << "Cloning arithmetic IVUser: " << *NarrowUse << "\n"); 1002 1003 unsigned IVOpIdx = (NarrowUse->getOperand(0) == NarrowDef) ? 0 : 1; 1004 1005 // We're trying to find X such that 1006 // 1007 // Widen(NarrowDef `op` NonIVNarrowDef) == WideAR == WideDef `op.wide` X 1008 // 1009 // We guess two solutions to X, sext(NonIVNarrowDef) and zext(NonIVNarrowDef), 1010 // and check using SCEV if any of them are correct. 1011 1012 // Returns true if extending NonIVNarrowDef according to `SignExt` is a 1013 // correct solution to X. 1014 auto GuessNonIVOperand = [&](bool SignExt) { 1015 const SCEV *WideLHS; 1016 const SCEV *WideRHS; 1017 1018 auto GetExtend = [this, SignExt](const SCEV *S, Type *Ty) { 1019 if (SignExt) 1020 return SE->getSignExtendExpr(S, Ty); 1021 return SE->getZeroExtendExpr(S, Ty); 1022 }; 1023 1024 if (IVOpIdx == 0) { 1025 WideLHS = SE->getSCEV(WideDef); 1026 const SCEV *NarrowRHS = SE->getSCEV(NarrowUse->getOperand(1)); 1027 WideRHS = GetExtend(NarrowRHS, WideType); 1028 } else { 1029 const SCEV *NarrowLHS = SE->getSCEV(NarrowUse->getOperand(0)); 1030 WideLHS = GetExtend(NarrowLHS, WideType); 1031 WideRHS = SE->getSCEV(WideDef); 1032 } 1033 1034 // WideUse is "WideDef `op.wide` X" as described in the comment. 1035 const SCEV *WideUse = nullptr; 1036 1037 switch (NarrowUse->getOpcode()) { 1038 default: 1039 llvm_unreachable("No other possibility!"); 1040 1041 case Instruction::Add: 1042 WideUse = SE->getAddExpr(WideLHS, WideRHS); 1043 break; 1044 1045 case Instruction::Mul: 1046 WideUse = SE->getMulExpr(WideLHS, WideRHS); 1047 break; 1048 1049 case Instruction::UDiv: 1050 WideUse = SE->getUDivExpr(WideLHS, WideRHS); 1051 break; 1052 1053 case Instruction::Sub: 1054 WideUse = SE->getMinusSCEV(WideLHS, WideRHS); 1055 break; 1056 } 1057 1058 return WideUse == WideAR; 1059 }; 1060 1061 bool SignExtend = IsSigned; 1062 if (!GuessNonIVOperand(SignExtend)) { 1063 SignExtend = !SignExtend; 1064 if (!GuessNonIVOperand(SignExtend)) 1065 return nullptr; 1066 } 1067 1068 Value *LHS = (NarrowUse->getOperand(0) == NarrowDef) 1069 ? WideDef 1070 : createExtendInst(NarrowUse->getOperand(0), WideType, 1071 SignExtend, NarrowUse); 1072 Value *RHS = (NarrowUse->getOperand(1) == NarrowDef) 1073 ? WideDef 1074 : createExtendInst(NarrowUse->getOperand(1), WideType, 1075 SignExtend, NarrowUse); 1076 1077 auto *NarrowBO = cast<BinaryOperator>(NarrowUse); 1078 auto *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(), LHS, RHS, 1079 NarrowBO->getName()); 1080 1081 IRBuilder<> Builder(NarrowUse); 1082 Builder.Insert(WideBO); 1083 WideBO->copyIRFlags(NarrowBO); 1084 return WideBO; 1085 } 1086 1087 const SCEV *WidenIV::getSCEVByOpCode(const SCEV *LHS, const SCEV *RHS, 1088 unsigned OpCode) const { 1089 if (OpCode == Instruction::Add) 1090 return SE->getAddExpr(LHS, RHS); 1091 if (OpCode == Instruction::Sub) 1092 return SE->getMinusSCEV(LHS, RHS); 1093 if (OpCode == Instruction::Mul) 1094 return SE->getMulExpr(LHS, RHS); 1095 1096 llvm_unreachable("Unsupported opcode."); 1097 } 1098 1099 /// No-wrap operations can transfer sign extension of their result to their 1100 /// operands. Generate the SCEV value for the widened operation without 1101 /// actually modifying the IR yet. If the expression after extending the 1102 /// operands is an AddRec for this loop, return it. 1103 const SCEVAddRecExpr* WidenIV::getExtendedOperandRecurrence(NarrowIVDefUse DU) { 1104 1105 // Handle the common case of add<nsw/nuw> 1106 const unsigned OpCode = DU.NarrowUse->getOpcode(); 1107 // Only Add/Sub/Mul instructions supported yet. 1108 if (OpCode != Instruction::Add && OpCode != Instruction::Sub && 1109 OpCode != Instruction::Mul) 1110 return nullptr; 1111 1112 // One operand (NarrowDef) has already been extended to WideDef. Now determine 1113 // if extending the other will lead to a recurrence. 1114 const unsigned ExtendOperIdx = 1115 DU.NarrowUse->getOperand(0) == DU.NarrowDef ? 1 : 0; 1116 assert(DU.NarrowUse->getOperand(1-ExtendOperIdx) == DU.NarrowDef && "bad DU"); 1117 1118 const SCEV *ExtendOperExpr = nullptr; 1119 const OverflowingBinaryOperator *OBO = 1120 cast<OverflowingBinaryOperator>(DU.NarrowUse); 1121 if (IsSigned && OBO->hasNoSignedWrap()) 1122 ExtendOperExpr = SE->getSignExtendExpr( 1123 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType); 1124 else if(!IsSigned && OBO->hasNoUnsignedWrap()) 1125 ExtendOperExpr = SE->getZeroExtendExpr( 1126 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType); 1127 else 1128 return nullptr; 1129 1130 // When creating this SCEV expr, don't apply the current operations NSW or NUW 1131 // flags. This instruction may be guarded by control flow that the no-wrap 1132 // behavior depends on. Non-control-equivalent instructions can be mapped to 1133 // the same SCEV expression, and it would be incorrect to transfer NSW/NUW 1134 // semantics to those operations. 1135 const SCEV *lhs = SE->getSCEV(DU.WideDef); 1136 const SCEV *rhs = ExtendOperExpr; 1137 1138 // Let's swap operands to the initial order for the case of non-commutative 1139 // operations, like SUB. See PR21014. 1140 if (ExtendOperIdx == 0) 1141 std::swap(lhs, rhs); 1142 const SCEVAddRecExpr *AddRec = 1143 dyn_cast<SCEVAddRecExpr>(getSCEVByOpCode(lhs, rhs, OpCode)); 1144 1145 if (!AddRec || AddRec->getLoop() != L) 1146 return nullptr; 1147 return AddRec; 1148 } 1149 1150 /// Is this instruction potentially interesting for further simplification after 1151 /// widening it's type? In other words, can the extend be safely hoisted out of 1152 /// the loop with SCEV reducing the value to a recurrence on the same loop. If 1153 /// so, return the sign or zero extended recurrence. Otherwise return NULL. 1154 const SCEVAddRecExpr *WidenIV::getWideRecurrence(Instruction *NarrowUse) { 1155 if (!SE->isSCEVable(NarrowUse->getType())) 1156 return nullptr; 1157 1158 const SCEV *NarrowExpr = SE->getSCEV(NarrowUse); 1159 if (SE->getTypeSizeInBits(NarrowExpr->getType()) 1160 >= SE->getTypeSizeInBits(WideType)) { 1161 // NarrowUse implicitly widens its operand. e.g. a gep with a narrow 1162 // index. So don't follow this use. 1163 return nullptr; 1164 } 1165 1166 const SCEV *WideExpr = IsSigned ? 1167 SE->getSignExtendExpr(NarrowExpr, WideType) : 1168 SE->getZeroExtendExpr(NarrowExpr, WideType); 1169 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr); 1170 if (!AddRec || AddRec->getLoop() != L) 1171 return nullptr; 1172 return AddRec; 1173 } 1174 1175 /// This IV user cannot be widen. Replace this use of the original narrow IV 1176 /// with a truncation of the new wide IV to isolate and eliminate the narrow IV. 1177 static void truncateIVUse(NarrowIVDefUse DU, DominatorTree *DT, LoopInfo *LI) { 1178 DEBUG(dbgs() << "INDVARS: Truncate IV " << *DU.WideDef 1179 << " for user " << *DU.NarrowUse << "\n"); 1180 IRBuilder<> Builder( 1181 getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT, LI)); 1182 Value *Trunc = Builder.CreateTrunc(DU.WideDef, DU.NarrowDef->getType()); 1183 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, Trunc); 1184 } 1185 1186 /// If the narrow use is a compare instruction, then widen the compare 1187 // (and possibly the other operand). The extend operation is hoisted into the 1188 // loop preheader as far as possible. 1189 bool WidenIV::widenLoopCompare(NarrowIVDefUse DU) { 1190 ICmpInst *Cmp = dyn_cast<ICmpInst>(DU.NarrowUse); 1191 if (!Cmp) 1192 return false; 1193 1194 // We can legally widen the comparison in the following two cases: 1195 // 1196 // - The signedness of the IV extension and comparison match 1197 // 1198 // - The narrow IV is always positive (and thus its sign extension is equal 1199 // to its zero extension). For instance, let's say we're zero extending 1200 // %narrow for the following use 1201 // 1202 // icmp slt i32 %narrow, %val ... (A) 1203 // 1204 // and %narrow is always positive. Then 1205 // 1206 // (A) == icmp slt i32 sext(%narrow), sext(%val) 1207 // == icmp slt i32 zext(%narrow), sext(%val) 1208 1209 if (!(DU.NeverNegative || IsSigned == Cmp->isSigned())) 1210 return false; 1211 1212 Value *Op = Cmp->getOperand(Cmp->getOperand(0) == DU.NarrowDef ? 1 : 0); 1213 unsigned CastWidth = SE->getTypeSizeInBits(Op->getType()); 1214 unsigned IVWidth = SE->getTypeSizeInBits(WideType); 1215 assert (CastWidth <= IVWidth && "Unexpected width while widening compare."); 1216 1217 // Widen the compare instruction. 1218 IRBuilder<> Builder( 1219 getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT, LI)); 1220 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef); 1221 1222 // Widen the other operand of the compare, if necessary. 1223 if (CastWidth < IVWidth) { 1224 Value *ExtOp = createExtendInst(Op, WideType, Cmp->isSigned(), Cmp); 1225 DU.NarrowUse->replaceUsesOfWith(Op, ExtOp); 1226 } 1227 return true; 1228 } 1229 1230 /// Determine whether an individual user of the narrow IV can be widened. If so, 1231 /// return the wide clone of the user. 1232 Instruction *WidenIV::widenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter) { 1233 1234 // Stop traversing the def-use chain at inner-loop phis or post-loop phis. 1235 if (PHINode *UsePhi = dyn_cast<PHINode>(DU.NarrowUse)) { 1236 if (LI->getLoopFor(UsePhi->getParent()) != L) { 1237 // For LCSSA phis, sink the truncate outside the loop. 1238 // After SimplifyCFG most loop exit targets have a single predecessor. 1239 // Otherwise fall back to a truncate within the loop. 1240 if (UsePhi->getNumOperands() != 1) 1241 truncateIVUse(DU, DT, LI); 1242 else { 1243 PHINode *WidePhi = 1244 PHINode::Create(DU.WideDef->getType(), 1, UsePhi->getName() + ".wide", 1245 UsePhi); 1246 WidePhi->addIncoming(DU.WideDef, UsePhi->getIncomingBlock(0)); 1247 IRBuilder<> Builder(&*WidePhi->getParent()->getFirstInsertionPt()); 1248 Value *Trunc = Builder.CreateTrunc(WidePhi, DU.NarrowDef->getType()); 1249 UsePhi->replaceAllUsesWith(Trunc); 1250 DeadInsts.emplace_back(UsePhi); 1251 DEBUG(dbgs() << "INDVARS: Widen lcssa phi " << *UsePhi 1252 << " to " << *WidePhi << "\n"); 1253 } 1254 return nullptr; 1255 } 1256 } 1257 // Our raison d'etre! Eliminate sign and zero extension. 1258 if (IsSigned ? isa<SExtInst>(DU.NarrowUse) : isa<ZExtInst>(DU.NarrowUse)) { 1259 Value *NewDef = DU.WideDef; 1260 if (DU.NarrowUse->getType() != WideType) { 1261 unsigned CastWidth = SE->getTypeSizeInBits(DU.NarrowUse->getType()); 1262 unsigned IVWidth = SE->getTypeSizeInBits(WideType); 1263 if (CastWidth < IVWidth) { 1264 // The cast isn't as wide as the IV, so insert a Trunc. 1265 IRBuilder<> Builder(DU.NarrowUse); 1266 NewDef = Builder.CreateTrunc(DU.WideDef, DU.NarrowUse->getType()); 1267 } 1268 else { 1269 // A wider extend was hidden behind a narrower one. This may induce 1270 // another round of IV widening in which the intermediate IV becomes 1271 // dead. It should be very rare. 1272 DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi 1273 << " not wide enough to subsume " << *DU.NarrowUse << "\n"); 1274 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef); 1275 NewDef = DU.NarrowUse; 1276 } 1277 } 1278 if (NewDef != DU.NarrowUse) { 1279 DEBUG(dbgs() << "INDVARS: eliminating " << *DU.NarrowUse 1280 << " replaced by " << *DU.WideDef << "\n"); 1281 ++NumElimExt; 1282 DU.NarrowUse->replaceAllUsesWith(NewDef); 1283 DeadInsts.emplace_back(DU.NarrowUse); 1284 } 1285 // Now that the extend is gone, we want to expose it's uses for potential 1286 // further simplification. We don't need to directly inform SimplifyIVUsers 1287 // of the new users, because their parent IV will be processed later as a 1288 // new loop phi. If we preserved IVUsers analysis, we would also want to 1289 // push the uses of WideDef here. 1290 1291 // No further widening is needed. The deceased [sz]ext had done it for us. 1292 return nullptr; 1293 } 1294 1295 // Does this user itself evaluate to a recurrence after widening? 1296 const SCEVAddRecExpr *WideAddRec = getWideRecurrence(DU.NarrowUse); 1297 if (!WideAddRec) 1298 WideAddRec = getExtendedOperandRecurrence(DU); 1299 1300 if (!WideAddRec) { 1301 // If use is a loop condition, try to promote the condition instead of 1302 // truncating the IV first. 1303 if (widenLoopCompare(DU)) 1304 return nullptr; 1305 1306 // This user does not evaluate to a recurence after widening, so don't 1307 // follow it. Instead insert a Trunc to kill off the original use, 1308 // eventually isolating the original narrow IV so it can be removed. 1309 truncateIVUse(DU, DT, LI); 1310 return nullptr; 1311 } 1312 // Assume block terminators cannot evaluate to a recurrence. We can't to 1313 // insert a Trunc after a terminator if there happens to be a critical edge. 1314 assert(DU.NarrowUse != DU.NarrowUse->getParent()->getTerminator() && 1315 "SCEV is not expected to evaluate a block terminator"); 1316 1317 // Reuse the IV increment that SCEVExpander created as long as it dominates 1318 // NarrowUse. 1319 Instruction *WideUse = nullptr; 1320 if (WideAddRec == WideIncExpr 1321 && Rewriter.hoistIVInc(WideInc, DU.NarrowUse)) 1322 WideUse = WideInc; 1323 else { 1324 WideUse = cloneIVUser(DU, WideAddRec); 1325 if (!WideUse) 1326 return nullptr; 1327 } 1328 // Evaluation of WideAddRec ensured that the narrow expression could be 1329 // extended outside the loop without overflow. This suggests that the wide use 1330 // evaluates to the same expression as the extended narrow use, but doesn't 1331 // absolutely guarantee it. Hence the following failsafe check. In rare cases 1332 // where it fails, we simply throw away the newly created wide use. 1333 if (WideAddRec != SE->getSCEV(WideUse)) { 1334 DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse 1335 << ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n"); 1336 DeadInsts.emplace_back(WideUse); 1337 return nullptr; 1338 } 1339 1340 // Returning WideUse pushes it on the worklist. 1341 return WideUse; 1342 } 1343 1344 /// Add eligible users of NarrowDef to NarrowIVUsers. 1345 /// 1346 void WidenIV::pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef) { 1347 const SCEV *NarrowSCEV = SE->getSCEV(NarrowDef); 1348 bool NeverNegative = 1349 SE->isKnownPredicate(ICmpInst::ICMP_SGE, NarrowSCEV, 1350 SE->getConstant(NarrowSCEV->getType(), 0)); 1351 for (User *U : NarrowDef->users()) { 1352 Instruction *NarrowUser = cast<Instruction>(U); 1353 1354 // Handle data flow merges and bizarre phi cycles. 1355 if (!Widened.insert(NarrowUser).second) 1356 continue; 1357 1358 NarrowIVUsers.push_back( 1359 NarrowIVDefUse(NarrowDef, NarrowUser, WideDef, NeverNegative)); 1360 } 1361 } 1362 1363 /// Process a single induction variable. First use the SCEVExpander to create a 1364 /// wide induction variable that evaluates to the same recurrence as the 1365 /// original narrow IV. Then use a worklist to forward traverse the narrow IV's 1366 /// def-use chain. After widenIVUse has processed all interesting IV users, the 1367 /// narrow IV will be isolated for removal by DeleteDeadPHIs. 1368 /// 1369 /// It would be simpler to delete uses as they are processed, but we must avoid 1370 /// invalidating SCEV expressions. 1371 /// 1372 PHINode *WidenIV::createWideIV(SCEVExpander &Rewriter) { 1373 // Is this phi an induction variable? 1374 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi)); 1375 if (!AddRec) 1376 return nullptr; 1377 1378 // Widen the induction variable expression. 1379 const SCEV *WideIVExpr = IsSigned ? 1380 SE->getSignExtendExpr(AddRec, WideType) : 1381 SE->getZeroExtendExpr(AddRec, WideType); 1382 1383 assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType && 1384 "Expect the new IV expression to preserve its type"); 1385 1386 // Can the IV be extended outside the loop without overflow? 1387 AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr); 1388 if (!AddRec || AddRec->getLoop() != L) 1389 return nullptr; 1390 1391 // An AddRec must have loop-invariant operands. Since this AddRec is 1392 // materialized by a loop header phi, the expression cannot have any post-loop 1393 // operands, so they must dominate the loop header. 1394 assert(SE->properlyDominates(AddRec->getStart(), L->getHeader()) && 1395 SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader()) 1396 && "Loop header phi recurrence inputs do not dominate the loop"); 1397 1398 // The rewriter provides a value for the desired IV expression. This may 1399 // either find an existing phi or materialize a new one. Either way, we 1400 // expect a well-formed cyclic phi-with-increments. i.e. any operand not part 1401 // of the phi-SCC dominates the loop entry. 1402 Instruction *InsertPt = &L->getHeader()->front(); 1403 WidePhi = cast<PHINode>(Rewriter.expandCodeFor(AddRec, WideType, InsertPt)); 1404 1405 // Remembering the WideIV increment generated by SCEVExpander allows 1406 // widenIVUse to reuse it when widening the narrow IV's increment. We don't 1407 // employ a general reuse mechanism because the call above is the only call to 1408 // SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses. 1409 if (BasicBlock *LatchBlock = L->getLoopLatch()) { 1410 WideInc = 1411 cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock)); 1412 WideIncExpr = SE->getSCEV(WideInc); 1413 } 1414 1415 DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n"); 1416 ++NumWidened; 1417 1418 // Traverse the def-use chain using a worklist starting at the original IV. 1419 assert(Widened.empty() && NarrowIVUsers.empty() && "expect initial state" ); 1420 1421 Widened.insert(OrigPhi); 1422 pushNarrowIVUsers(OrigPhi, WidePhi); 1423 1424 while (!NarrowIVUsers.empty()) { 1425 NarrowIVDefUse DU = NarrowIVUsers.pop_back_val(); 1426 1427 // Process a def-use edge. This may replace the use, so don't hold a 1428 // use_iterator across it. 1429 Instruction *WideUse = widenIVUse(DU, Rewriter); 1430 1431 // Follow all def-use edges from the previous narrow use. 1432 if (WideUse) 1433 pushNarrowIVUsers(DU.NarrowUse, WideUse); 1434 1435 // widenIVUse may have removed the def-use edge. 1436 if (DU.NarrowDef->use_empty()) 1437 DeadInsts.emplace_back(DU.NarrowDef); 1438 } 1439 return WidePhi; 1440 } 1441 1442 //===----------------------------------------------------------------------===// 1443 // Live IV Reduction - Minimize IVs live across the loop. 1444 //===----------------------------------------------------------------------===// 1445 1446 1447 //===----------------------------------------------------------------------===// 1448 // Simplification of IV users based on SCEV evaluation. 1449 //===----------------------------------------------------------------------===// 1450 1451 namespace { 1452 class IndVarSimplifyVisitor : public IVVisitor { 1453 ScalarEvolution *SE; 1454 const TargetTransformInfo *TTI; 1455 PHINode *IVPhi; 1456 1457 public: 1458 WideIVInfo WI; 1459 1460 IndVarSimplifyVisitor(PHINode *IV, ScalarEvolution *SCEV, 1461 const TargetTransformInfo *TTI, 1462 const DominatorTree *DTree) 1463 : SE(SCEV), TTI(TTI), IVPhi(IV) { 1464 DT = DTree; 1465 WI.NarrowIV = IVPhi; 1466 if (ReduceLiveIVs) 1467 setSplitOverflowIntrinsics(); 1468 } 1469 1470 // Implement the interface used by simplifyUsersOfIV. 1471 void visitCast(CastInst *Cast) override { visitIVCast(Cast, WI, SE, TTI); } 1472 }; 1473 } 1474 1475 /// Iteratively perform simplification on a worklist of IV users. Each 1476 /// successive simplification may push more users which may themselves be 1477 /// candidates for simplification. 1478 /// 1479 /// Sign/Zero extend elimination is interleaved with IV simplification. 1480 /// 1481 void IndVarSimplify::simplifyAndExtend(Loop *L, 1482 SCEVExpander &Rewriter, 1483 LoopInfo *LI) { 1484 SmallVector<WideIVInfo, 8> WideIVs; 1485 1486 SmallVector<PHINode*, 8> LoopPhis; 1487 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) { 1488 LoopPhis.push_back(cast<PHINode>(I)); 1489 } 1490 // Each round of simplification iterates through the SimplifyIVUsers worklist 1491 // for all current phis, then determines whether any IVs can be 1492 // widened. Widening adds new phis to LoopPhis, inducing another round of 1493 // simplification on the wide IVs. 1494 while (!LoopPhis.empty()) { 1495 // Evaluate as many IV expressions as possible before widening any IVs. This 1496 // forces SCEV to set no-wrap flags before evaluating sign/zero 1497 // extension. The first time SCEV attempts to normalize sign/zero extension, 1498 // the result becomes final. So for the most predictable results, we delay 1499 // evaluation of sign/zero extend evaluation until needed, and avoid running 1500 // other SCEV based analysis prior to simplifyAndExtend. 1501 do { 1502 PHINode *CurrIV = LoopPhis.pop_back_val(); 1503 1504 // Information about sign/zero extensions of CurrIV. 1505 IndVarSimplifyVisitor Visitor(CurrIV, SE, TTI, DT); 1506 1507 Changed |= simplifyUsersOfIV(CurrIV, SE, DT, LI, DeadInsts, &Visitor); 1508 1509 if (Visitor.WI.WidestNativeType) { 1510 WideIVs.push_back(Visitor.WI); 1511 } 1512 } while(!LoopPhis.empty()); 1513 1514 for (; !WideIVs.empty(); WideIVs.pop_back()) { 1515 WidenIV Widener(WideIVs.back(), LI, SE, DT, DeadInsts); 1516 if (PHINode *WidePhi = Widener.createWideIV(Rewriter)) { 1517 Changed = true; 1518 LoopPhis.push_back(WidePhi); 1519 } 1520 } 1521 } 1522 } 1523 1524 //===----------------------------------------------------------------------===// 1525 // linearFunctionTestReplace and its kin. Rewrite the loop exit condition. 1526 //===----------------------------------------------------------------------===// 1527 1528 /// Return true if this loop's backedge taken count expression can be safely and 1529 /// cheaply expanded into an instruction sequence that can be used by 1530 /// linearFunctionTestReplace. 1531 /// 1532 /// TODO: This fails for pointer-type loop counters with greater than one byte 1533 /// strides, consequently preventing LFTR from running. For the purpose of LFTR 1534 /// we could skip this check in the case that the LFTR loop counter (chosen by 1535 /// FindLoopCounter) is also pointer type. Instead, we could directly convert 1536 /// the loop test to an inequality test by checking the target data's alignment 1537 /// of element types (given that the initial pointer value originates from or is 1538 /// used by ABI constrained operation, as opposed to inttoptr/ptrtoint). 1539 /// However, we don't yet have a strong motivation for converting loop tests 1540 /// into inequality tests. 1541 static bool canExpandBackedgeTakenCount(Loop *L, ScalarEvolution *SE, 1542 SCEVExpander &Rewriter) { 1543 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L); 1544 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) || 1545 BackedgeTakenCount->isZero()) 1546 return false; 1547 1548 if (!L->getExitingBlock()) 1549 return false; 1550 1551 // Can't rewrite non-branch yet. 1552 if (!isa<BranchInst>(L->getExitingBlock()->getTerminator())) 1553 return false; 1554 1555 if (Rewriter.isHighCostExpansion(BackedgeTakenCount, L)) 1556 return false; 1557 1558 return true; 1559 } 1560 1561 /// Return the loop header phi IFF IncV adds a loop invariant value to the phi. 1562 static PHINode *getLoopPhiForCounter(Value *IncV, Loop *L, DominatorTree *DT) { 1563 Instruction *IncI = dyn_cast<Instruction>(IncV); 1564 if (!IncI) 1565 return nullptr; 1566 1567 switch (IncI->getOpcode()) { 1568 case Instruction::Add: 1569 case Instruction::Sub: 1570 break; 1571 case Instruction::GetElementPtr: 1572 // An IV counter must preserve its type. 1573 if (IncI->getNumOperands() == 2) 1574 break; 1575 default: 1576 return nullptr; 1577 } 1578 1579 PHINode *Phi = dyn_cast<PHINode>(IncI->getOperand(0)); 1580 if (Phi && Phi->getParent() == L->getHeader()) { 1581 if (isLoopInvariant(IncI->getOperand(1), L, DT)) 1582 return Phi; 1583 return nullptr; 1584 } 1585 if (IncI->getOpcode() == Instruction::GetElementPtr) 1586 return nullptr; 1587 1588 // Allow add/sub to be commuted. 1589 Phi = dyn_cast<PHINode>(IncI->getOperand(1)); 1590 if (Phi && Phi->getParent() == L->getHeader()) { 1591 if (isLoopInvariant(IncI->getOperand(0), L, DT)) 1592 return Phi; 1593 } 1594 return nullptr; 1595 } 1596 1597 /// Return the compare guarding the loop latch, or NULL for unrecognized tests. 1598 static ICmpInst *getLoopTest(Loop *L) { 1599 assert(L->getExitingBlock() && "expected loop exit"); 1600 1601 BasicBlock *LatchBlock = L->getLoopLatch(); 1602 // Don't bother with LFTR if the loop is not properly simplified. 1603 if (!LatchBlock) 1604 return nullptr; 1605 1606 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1607 assert(BI && "expected exit branch"); 1608 1609 return dyn_cast<ICmpInst>(BI->getCondition()); 1610 } 1611 1612 /// linearFunctionTestReplace policy. Return true unless we can show that the 1613 /// current exit test is already sufficiently canonical. 1614 static bool needsLFTR(Loop *L, DominatorTree *DT) { 1615 // Do LFTR to simplify the exit condition to an ICMP. 1616 ICmpInst *Cond = getLoopTest(L); 1617 if (!Cond) 1618 return true; 1619 1620 // Do LFTR to simplify the exit ICMP to EQ/NE 1621 ICmpInst::Predicate Pred = Cond->getPredicate(); 1622 if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ) 1623 return true; 1624 1625 // Look for a loop invariant RHS 1626 Value *LHS = Cond->getOperand(0); 1627 Value *RHS = Cond->getOperand(1); 1628 if (!isLoopInvariant(RHS, L, DT)) { 1629 if (!isLoopInvariant(LHS, L, DT)) 1630 return true; 1631 std::swap(LHS, RHS); 1632 } 1633 // Look for a simple IV counter LHS 1634 PHINode *Phi = dyn_cast<PHINode>(LHS); 1635 if (!Phi) 1636 Phi = getLoopPhiForCounter(LHS, L, DT); 1637 1638 if (!Phi) 1639 return true; 1640 1641 // Do LFTR if PHI node is defined in the loop, but is *not* a counter. 1642 int Idx = Phi->getBasicBlockIndex(L->getLoopLatch()); 1643 if (Idx < 0) 1644 return true; 1645 1646 // Do LFTR if the exit condition's IV is *not* a simple counter. 1647 Value *IncV = Phi->getIncomingValue(Idx); 1648 return Phi != getLoopPhiForCounter(IncV, L, DT); 1649 } 1650 1651 /// Recursive helper for hasConcreteDef(). Unfortunately, this currently boils 1652 /// down to checking that all operands are constant and listing instructions 1653 /// that may hide undef. 1654 static bool hasConcreteDefImpl(Value *V, SmallPtrSetImpl<Value*> &Visited, 1655 unsigned Depth) { 1656 if (isa<Constant>(V)) 1657 return !isa<UndefValue>(V); 1658 1659 if (Depth >= 6) 1660 return false; 1661 1662 // Conservatively handle non-constant non-instructions. For example, Arguments 1663 // may be undef. 1664 Instruction *I = dyn_cast<Instruction>(V); 1665 if (!I) 1666 return false; 1667 1668 // Load and return values may be undef. 1669 if(I->mayReadFromMemory() || isa<CallInst>(I) || isa<InvokeInst>(I)) 1670 return false; 1671 1672 // Optimistically handle other instructions. 1673 for (Value *Op : I->operands()) { 1674 if (!Visited.insert(Op).second) 1675 continue; 1676 if (!hasConcreteDefImpl(Op, Visited, Depth+1)) 1677 return false; 1678 } 1679 return true; 1680 } 1681 1682 /// Return true if the given value is concrete. We must prove that undef can 1683 /// never reach it. 1684 /// 1685 /// TODO: If we decide that this is a good approach to checking for undef, we 1686 /// may factor it into a common location. 1687 static bool hasConcreteDef(Value *V) { 1688 SmallPtrSet<Value*, 8> Visited; 1689 Visited.insert(V); 1690 return hasConcreteDefImpl(V, Visited, 0); 1691 } 1692 1693 /// Return true if this IV has any uses other than the (soon to be rewritten) 1694 /// loop exit test. 1695 static bool AlmostDeadIV(PHINode *Phi, BasicBlock *LatchBlock, Value *Cond) { 1696 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock); 1697 Value *IncV = Phi->getIncomingValue(LatchIdx); 1698 1699 for (User *U : Phi->users()) 1700 if (U != Cond && U != IncV) return false; 1701 1702 for (User *U : IncV->users()) 1703 if (U != Cond && U != Phi) return false; 1704 return true; 1705 } 1706 1707 /// Find an affine IV in canonical form. 1708 /// 1709 /// BECount may be an i8* pointer type. The pointer difference is already 1710 /// valid count without scaling the address stride, so it remains a pointer 1711 /// expression as far as SCEV is concerned. 1712 /// 1713 /// Currently only valid for LFTR. See the comments on hasConcreteDef below. 1714 /// 1715 /// FIXME: Accept -1 stride and set IVLimit = IVInit - BECount 1716 /// 1717 /// FIXME: Accept non-unit stride as long as SCEV can reduce BECount * Stride. 1718 /// This is difficult in general for SCEV because of potential overflow. But we 1719 /// could at least handle constant BECounts. 1720 static PHINode *FindLoopCounter(Loop *L, const SCEV *BECount, 1721 ScalarEvolution *SE, DominatorTree *DT) { 1722 uint64_t BCWidth = SE->getTypeSizeInBits(BECount->getType()); 1723 1724 Value *Cond = 1725 cast<BranchInst>(L->getExitingBlock()->getTerminator())->getCondition(); 1726 1727 // Loop over all of the PHI nodes, looking for a simple counter. 1728 PHINode *BestPhi = nullptr; 1729 const SCEV *BestInit = nullptr; 1730 BasicBlock *LatchBlock = L->getLoopLatch(); 1731 assert(LatchBlock && "needsLFTR should guarantee a loop latch"); 1732 1733 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) { 1734 PHINode *Phi = cast<PHINode>(I); 1735 if (!SE->isSCEVable(Phi->getType())) 1736 continue; 1737 1738 // Avoid comparing an integer IV against a pointer Limit. 1739 if (BECount->getType()->isPointerTy() && !Phi->getType()->isPointerTy()) 1740 continue; 1741 1742 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Phi)); 1743 if (!AR || AR->getLoop() != L || !AR->isAffine()) 1744 continue; 1745 1746 // AR may be a pointer type, while BECount is an integer type. 1747 // AR may be wider than BECount. With eq/ne tests overflow is immaterial. 1748 // AR may not be a narrower type, or we may never exit. 1749 uint64_t PhiWidth = SE->getTypeSizeInBits(AR->getType()); 1750 if (PhiWidth < BCWidth || 1751 !L->getHeader()->getModule()->getDataLayout().isLegalInteger(PhiWidth)) 1752 continue; 1753 1754 const SCEV *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE)); 1755 if (!Step || !Step->isOne()) 1756 continue; 1757 1758 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock); 1759 Value *IncV = Phi->getIncomingValue(LatchIdx); 1760 if (getLoopPhiForCounter(IncV, L, DT) != Phi) 1761 continue; 1762 1763 // Avoid reusing a potentially undef value to compute other values that may 1764 // have originally had a concrete definition. 1765 if (!hasConcreteDef(Phi)) { 1766 // We explicitly allow unknown phis as long as they are already used by 1767 // the loop test. In this case we assume that performing LFTR could not 1768 // increase the number of undef users. 1769 if (ICmpInst *Cond = getLoopTest(L)) { 1770 if (Phi != getLoopPhiForCounter(Cond->getOperand(0), L, DT) 1771 && Phi != getLoopPhiForCounter(Cond->getOperand(1), L, DT)) { 1772 continue; 1773 } 1774 } 1775 } 1776 const SCEV *Init = AR->getStart(); 1777 1778 if (BestPhi && !AlmostDeadIV(BestPhi, LatchBlock, Cond)) { 1779 // Don't force a live loop counter if another IV can be used. 1780 if (AlmostDeadIV(Phi, LatchBlock, Cond)) 1781 continue; 1782 1783 // Prefer to count-from-zero. This is a more "canonical" counter form. It 1784 // also prefers integer to pointer IVs. 1785 if (BestInit->isZero() != Init->isZero()) { 1786 if (BestInit->isZero()) 1787 continue; 1788 } 1789 // If two IVs both count from zero or both count from nonzero then the 1790 // narrower is likely a dead phi that has been widened. Use the wider phi 1791 // to allow the other to be eliminated. 1792 else if (PhiWidth <= SE->getTypeSizeInBits(BestPhi->getType())) 1793 continue; 1794 } 1795 BestPhi = Phi; 1796 BestInit = Init; 1797 } 1798 return BestPhi; 1799 } 1800 1801 /// Help linearFunctionTestReplace by generating a value that holds the RHS of 1802 /// the new loop test. 1803 static Value *genLoopLimit(PHINode *IndVar, const SCEV *IVCount, Loop *L, 1804 SCEVExpander &Rewriter, ScalarEvolution *SE) { 1805 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IndVar)); 1806 assert(AR && AR->getLoop() == L && AR->isAffine() && "bad loop counter"); 1807 const SCEV *IVInit = AR->getStart(); 1808 1809 // IVInit may be a pointer while IVCount is an integer when FindLoopCounter 1810 // finds a valid pointer IV. Sign extend BECount in order to materialize a 1811 // GEP. Avoid running SCEVExpander on a new pointer value, instead reusing 1812 // the existing GEPs whenever possible. 1813 if (IndVar->getType()->isPointerTy() 1814 && !IVCount->getType()->isPointerTy()) { 1815 1816 // IVOffset will be the new GEP offset that is interpreted by GEP as a 1817 // signed value. IVCount on the other hand represents the loop trip count, 1818 // which is an unsigned value. FindLoopCounter only allows induction 1819 // variables that have a positive unit stride of one. This means we don't 1820 // have to handle the case of negative offsets (yet) and just need to zero 1821 // extend IVCount. 1822 Type *OfsTy = SE->getEffectiveSCEVType(IVInit->getType()); 1823 const SCEV *IVOffset = SE->getTruncateOrZeroExtend(IVCount, OfsTy); 1824 1825 // Expand the code for the iteration count. 1826 assert(SE->isLoopInvariant(IVOffset, L) && 1827 "Computed iteration count is not loop invariant!"); 1828 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1829 Value *GEPOffset = Rewriter.expandCodeFor(IVOffset, OfsTy, BI); 1830 1831 Value *GEPBase = IndVar->getIncomingValueForBlock(L->getLoopPreheader()); 1832 assert(AR->getStart() == SE->getSCEV(GEPBase) && "bad loop counter"); 1833 // We could handle pointer IVs other than i8*, but we need to compensate for 1834 // gep index scaling. See canExpandBackedgeTakenCount comments. 1835 assert(SE->getSizeOfExpr(IntegerType::getInt64Ty(IndVar->getContext()), 1836 cast<PointerType>(GEPBase->getType())->getElementType())->isOne() 1837 && "unit stride pointer IV must be i8*"); 1838 1839 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator()); 1840 return Builder.CreateGEP(nullptr, GEPBase, GEPOffset, "lftr.limit"); 1841 } 1842 else { 1843 // In any other case, convert both IVInit and IVCount to integers before 1844 // comparing. This may result in SCEV expension of pointers, but in practice 1845 // SCEV will fold the pointer arithmetic away as such: 1846 // BECount = (IVEnd - IVInit - 1) => IVLimit = IVInit (postinc). 1847 // 1848 // Valid Cases: (1) both integers is most common; (2) both may be pointers 1849 // for simple memset-style loops. 1850 // 1851 // IVInit integer and IVCount pointer would only occur if a canonical IV 1852 // were generated on top of case #2, which is not expected. 1853 1854 const SCEV *IVLimit = nullptr; 1855 // For unit stride, IVCount = Start + BECount with 2's complement overflow. 1856 // For non-zero Start, compute IVCount here. 1857 if (AR->getStart()->isZero()) 1858 IVLimit = IVCount; 1859 else { 1860 assert(AR->getStepRecurrence(*SE)->isOne() && "only handles unit stride"); 1861 const SCEV *IVInit = AR->getStart(); 1862 1863 // For integer IVs, truncate the IV before computing IVInit + BECount. 1864 if (SE->getTypeSizeInBits(IVInit->getType()) 1865 > SE->getTypeSizeInBits(IVCount->getType())) 1866 IVInit = SE->getTruncateExpr(IVInit, IVCount->getType()); 1867 1868 IVLimit = SE->getAddExpr(IVInit, IVCount); 1869 } 1870 // Expand the code for the iteration count. 1871 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1872 IRBuilder<> Builder(BI); 1873 assert(SE->isLoopInvariant(IVLimit, L) && 1874 "Computed iteration count is not loop invariant!"); 1875 // Ensure that we generate the same type as IndVar, or a smaller integer 1876 // type. In the presence of null pointer values, we have an integer type 1877 // SCEV expression (IVInit) for a pointer type IV value (IndVar). 1878 Type *LimitTy = IVCount->getType()->isPointerTy() ? 1879 IndVar->getType() : IVCount->getType(); 1880 return Rewriter.expandCodeFor(IVLimit, LimitTy, BI); 1881 } 1882 } 1883 1884 /// This method rewrites the exit condition of the loop to be a canonical != 1885 /// comparison against the incremented loop induction variable. This pass is 1886 /// able to rewrite the exit tests of any loop where the SCEV analysis can 1887 /// determine a loop-invariant trip count of the loop, which is actually a much 1888 /// broader range than just linear tests. 1889 Value *IndVarSimplify:: 1890 linearFunctionTestReplace(Loop *L, 1891 const SCEV *BackedgeTakenCount, 1892 PHINode *IndVar, 1893 SCEVExpander &Rewriter) { 1894 assert(canExpandBackedgeTakenCount(L, SE, Rewriter) && "precondition"); 1895 1896 // Initialize CmpIndVar and IVCount to their preincremented values. 1897 Value *CmpIndVar = IndVar; 1898 const SCEV *IVCount = BackedgeTakenCount; 1899 1900 // If the exiting block is the same as the backedge block, we prefer to 1901 // compare against the post-incremented value, otherwise we must compare 1902 // against the preincremented value. 1903 if (L->getExitingBlock() == L->getLoopLatch()) { 1904 // Add one to the "backedge-taken" count to get the trip count. 1905 // This addition may overflow, which is valid as long as the comparison is 1906 // truncated to BackedgeTakenCount->getType(). 1907 IVCount = SE->getAddExpr(BackedgeTakenCount, 1908 SE->getOne(BackedgeTakenCount->getType())); 1909 // The BackedgeTaken expression contains the number of times that the 1910 // backedge branches to the loop header. This is one less than the 1911 // number of times the loop executes, so use the incremented indvar. 1912 CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock()); 1913 } 1914 1915 Value *ExitCnt = genLoopLimit(IndVar, IVCount, L, Rewriter, SE); 1916 assert(ExitCnt->getType()->isPointerTy() == IndVar->getType()->isPointerTy() 1917 && "genLoopLimit missed a cast"); 1918 1919 // Insert a new icmp_ne or icmp_eq instruction before the branch. 1920 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1921 ICmpInst::Predicate P; 1922 if (L->contains(BI->getSuccessor(0))) 1923 P = ICmpInst::ICMP_NE; 1924 else 1925 P = ICmpInst::ICMP_EQ; 1926 1927 DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n" 1928 << " LHS:" << *CmpIndVar << '\n' 1929 << " op:\t" 1930 << (P == ICmpInst::ICMP_NE ? "!=" : "==") << "\n" 1931 << " RHS:\t" << *ExitCnt << "\n" 1932 << " IVCount:\t" << *IVCount << "\n"); 1933 1934 IRBuilder<> Builder(BI); 1935 1936 // LFTR can ignore IV overflow and truncate to the width of 1937 // BECount. This avoids materializing the add(zext(add)) expression. 1938 unsigned CmpIndVarSize = SE->getTypeSizeInBits(CmpIndVar->getType()); 1939 unsigned ExitCntSize = SE->getTypeSizeInBits(ExitCnt->getType()); 1940 if (CmpIndVarSize > ExitCntSize) { 1941 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(SE->getSCEV(IndVar)); 1942 const SCEV *ARStart = AR->getStart(); 1943 const SCEV *ARStep = AR->getStepRecurrence(*SE); 1944 // For constant IVCount, avoid truncation. 1945 if (isa<SCEVConstant>(ARStart) && isa<SCEVConstant>(IVCount)) { 1946 const APInt &Start = cast<SCEVConstant>(ARStart)->getAPInt(); 1947 APInt Count = cast<SCEVConstant>(IVCount)->getAPInt(); 1948 // Note that the post-inc value of BackedgeTakenCount may have overflowed 1949 // above such that IVCount is now zero. 1950 if (IVCount != BackedgeTakenCount && Count == 0) { 1951 Count = APInt::getMaxValue(Count.getBitWidth()).zext(CmpIndVarSize); 1952 ++Count; 1953 } 1954 else 1955 Count = Count.zext(CmpIndVarSize); 1956 APInt NewLimit; 1957 if (cast<SCEVConstant>(ARStep)->getValue()->isNegative()) 1958 NewLimit = Start - Count; 1959 else 1960 NewLimit = Start + Count; 1961 ExitCnt = ConstantInt::get(CmpIndVar->getType(), NewLimit); 1962 1963 DEBUG(dbgs() << " Widen RHS:\t" << *ExitCnt << "\n"); 1964 } else { 1965 CmpIndVar = Builder.CreateTrunc(CmpIndVar, ExitCnt->getType(), 1966 "lftr.wideiv"); 1967 } 1968 } 1969 Value *Cond = Builder.CreateICmp(P, CmpIndVar, ExitCnt, "exitcond"); 1970 Value *OrigCond = BI->getCondition(); 1971 // It's tempting to use replaceAllUsesWith here to fully replace the old 1972 // comparison, but that's not immediately safe, since users of the old 1973 // comparison may not be dominated by the new comparison. Instead, just 1974 // update the branch to use the new comparison; in the common case this 1975 // will make old comparison dead. 1976 BI->setCondition(Cond); 1977 DeadInsts.push_back(OrigCond); 1978 1979 ++NumLFTR; 1980 Changed = true; 1981 return Cond; 1982 } 1983 1984 //===----------------------------------------------------------------------===// 1985 // sinkUnusedInvariants. A late subpass to cleanup loop preheaders. 1986 //===----------------------------------------------------------------------===// 1987 1988 /// If there's a single exit block, sink any loop-invariant values that 1989 /// were defined in the preheader but not used inside the loop into the 1990 /// exit block to reduce register pressure in the loop. 1991 void IndVarSimplify::sinkUnusedInvariants(Loop *L) { 1992 BasicBlock *ExitBlock = L->getExitBlock(); 1993 if (!ExitBlock) return; 1994 1995 BasicBlock *Preheader = L->getLoopPreheader(); 1996 if (!Preheader) return; 1997 1998 Instruction *InsertPt = &*ExitBlock->getFirstInsertionPt(); 1999 BasicBlock::iterator I(Preheader->getTerminator()); 2000 while (I != Preheader->begin()) { 2001 --I; 2002 // New instructions were inserted at the end of the preheader. 2003 if (isa<PHINode>(I)) 2004 break; 2005 2006 // Don't move instructions which might have side effects, since the side 2007 // effects need to complete before instructions inside the loop. Also don't 2008 // move instructions which might read memory, since the loop may modify 2009 // memory. Note that it's okay if the instruction might have undefined 2010 // behavior: LoopSimplify guarantees that the preheader dominates the exit 2011 // block. 2012 if (I->mayHaveSideEffects() || I->mayReadFromMemory()) 2013 continue; 2014 2015 // Skip debug info intrinsics. 2016 if (isa<DbgInfoIntrinsic>(I)) 2017 continue; 2018 2019 // Skip eh pad instructions. 2020 if (I->isEHPad()) 2021 continue; 2022 2023 // Don't sink alloca: we never want to sink static alloca's out of the 2024 // entry block, and correctly sinking dynamic alloca's requires 2025 // checks for stacksave/stackrestore intrinsics. 2026 // FIXME: Refactor this check somehow? 2027 if (isa<AllocaInst>(I)) 2028 continue; 2029 2030 // Determine if there is a use in or before the loop (direct or 2031 // otherwise). 2032 bool UsedInLoop = false; 2033 for (Use &U : I->uses()) { 2034 Instruction *User = cast<Instruction>(U.getUser()); 2035 BasicBlock *UseBB = User->getParent(); 2036 if (PHINode *P = dyn_cast<PHINode>(User)) { 2037 unsigned i = 2038 PHINode::getIncomingValueNumForOperand(U.getOperandNo()); 2039 UseBB = P->getIncomingBlock(i); 2040 } 2041 if (UseBB == Preheader || L->contains(UseBB)) { 2042 UsedInLoop = true; 2043 break; 2044 } 2045 } 2046 2047 // If there is, the def must remain in the preheader. 2048 if (UsedInLoop) 2049 continue; 2050 2051 // Otherwise, sink it to the exit block. 2052 Instruction *ToMove = &*I; 2053 bool Done = false; 2054 2055 if (I != Preheader->begin()) { 2056 // Skip debug info intrinsics. 2057 do { 2058 --I; 2059 } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin()); 2060 2061 if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin()) 2062 Done = true; 2063 } else { 2064 Done = true; 2065 } 2066 2067 ToMove->moveBefore(InsertPt); 2068 if (Done) break; 2069 InsertPt = ToMove; 2070 } 2071 } 2072 2073 //===----------------------------------------------------------------------===// 2074 // IndVarSimplify driver. Manage several subpasses of IV simplification. 2075 //===----------------------------------------------------------------------===// 2076 2077 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) { 2078 if (skipOptnoneFunction(L)) 2079 return false; 2080 2081 // If LoopSimplify form is not available, stay out of trouble. Some notes: 2082 // - LSR currently only supports LoopSimplify-form loops. Indvars' 2083 // canonicalization can be a pessimization without LSR to "clean up" 2084 // afterwards. 2085 // - We depend on having a preheader; in particular, 2086 // Loop::getCanonicalInductionVariable only supports loops with preheaders, 2087 // and we're in trouble if we can't find the induction variable even when 2088 // we've manually inserted one. 2089 if (!L->isLoopSimplifyForm()) 2090 return false; 2091 2092 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 2093 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); 2094 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 2095 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>(); 2096 TLI = TLIP ? &TLIP->getTLI() : nullptr; 2097 auto *TTIP = getAnalysisIfAvailable<TargetTransformInfoWrapperPass>(); 2098 TTI = TTIP ? &TTIP->getTTI(*L->getHeader()->getParent()) : nullptr; 2099 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout(); 2100 2101 DeadInsts.clear(); 2102 Changed = false; 2103 2104 // If there are any floating-point recurrences, attempt to 2105 // transform them to use integer recurrences. 2106 rewriteNonIntegerIVs(L); 2107 2108 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L); 2109 2110 // Create a rewriter object which we'll use to transform the code with. 2111 SCEVExpander Rewriter(*SE, DL, "indvars"); 2112 #ifndef NDEBUG 2113 Rewriter.setDebugType(DEBUG_TYPE); 2114 #endif 2115 2116 // Eliminate redundant IV users. 2117 // 2118 // Simplification works best when run before other consumers of SCEV. We 2119 // attempt to avoid evaluating SCEVs for sign/zero extend operations until 2120 // other expressions involving loop IVs have been evaluated. This helps SCEV 2121 // set no-wrap flags before normalizing sign/zero extension. 2122 Rewriter.disableCanonicalMode(); 2123 simplifyAndExtend(L, Rewriter, LI); 2124 2125 // Check to see if this loop has a computable loop-invariant execution count. 2126 // If so, this means that we can compute the final value of any expressions 2127 // that are recurrent in the loop, and substitute the exit values from the 2128 // loop into any instructions outside of the loop that use the final values of 2129 // the current expressions. 2130 // 2131 if (ReplaceExitValue != NeverRepl && 2132 !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) 2133 rewriteLoopExitValues(L, Rewriter); 2134 2135 // Eliminate redundant IV cycles. 2136 NumElimIV += Rewriter.replaceCongruentIVs(L, DT, DeadInsts); 2137 2138 // If we have a trip count expression, rewrite the loop's exit condition 2139 // using it. We can currently only handle loops with a single exit. 2140 if (canExpandBackedgeTakenCount(L, SE, Rewriter) && needsLFTR(L, DT)) { 2141 PHINode *IndVar = FindLoopCounter(L, BackedgeTakenCount, SE, DT); 2142 if (IndVar) { 2143 // Check preconditions for proper SCEVExpander operation. SCEV does not 2144 // express SCEVExpander's dependencies, such as LoopSimplify. Instead any 2145 // pass that uses the SCEVExpander must do it. This does not work well for 2146 // loop passes because SCEVExpander makes assumptions about all loops, 2147 // while LoopPassManager only forces the current loop to be simplified. 2148 // 2149 // FIXME: SCEV expansion has no way to bail out, so the caller must 2150 // explicitly check any assumptions made by SCEV. Brittle. 2151 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(BackedgeTakenCount); 2152 if (!AR || AR->getLoop()->getLoopPreheader()) 2153 (void)linearFunctionTestReplace(L, BackedgeTakenCount, IndVar, 2154 Rewriter); 2155 } 2156 } 2157 // Clear the rewriter cache, because values that are in the rewriter's cache 2158 // can be deleted in the loop below, causing the AssertingVH in the cache to 2159 // trigger. 2160 Rewriter.clear(); 2161 2162 // Now that we're done iterating through lists, clean up any instructions 2163 // which are now dead. 2164 while (!DeadInsts.empty()) 2165 if (Instruction *Inst = 2166 dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val())) 2167 RecursivelyDeleteTriviallyDeadInstructions(Inst, TLI); 2168 2169 // The Rewriter may not be used from this point on. 2170 2171 // Loop-invariant instructions in the preheader that aren't used in the 2172 // loop may be sunk below the loop to reduce register pressure. 2173 sinkUnusedInvariants(L); 2174 2175 // Clean up dead instructions. 2176 Changed |= DeleteDeadPHIs(L->getHeader(), TLI); 2177 2178 // Check a post-condition. 2179 assert(L->isRecursivelyLCSSAForm(*DT) && "Indvars did not preserve LCSSA!"); 2180 2181 // Verify that LFTR, and any other change have not interfered with SCEV's 2182 // ability to compute trip count. 2183 #ifndef NDEBUG 2184 if (VerifyIndvars && !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) { 2185 SE->forgetLoop(L); 2186 const SCEV *NewBECount = SE->getBackedgeTakenCount(L); 2187 if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) < 2188 SE->getTypeSizeInBits(NewBECount->getType())) 2189 NewBECount = SE->getTruncateOrNoop(NewBECount, 2190 BackedgeTakenCount->getType()); 2191 else 2192 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount, 2193 NewBECount->getType()); 2194 assert(BackedgeTakenCount == NewBECount && "indvars must preserve SCEV"); 2195 } 2196 #endif 2197 2198 return Changed; 2199 } 2200