IndVarSimplify.cpp revision 312832
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/IndVarSimplify.h" 28#include "llvm/Transforms/Scalar.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/LoopPassManager.h" 35#include "llvm/Analysis/ScalarEvolutionExpander.h" 36#include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h" 37#include "llvm/Analysis/TargetLibraryInfo.h" 38#include "llvm/Analysis/TargetTransformInfo.h" 39#include "llvm/IR/BasicBlock.h" 40#include "llvm/IR/CFG.h" 41#include "llvm/IR/Constants.h" 42#include "llvm/IR/DataLayout.h" 43#include "llvm/IR/Dominators.h" 44#include "llvm/IR/Instructions.h" 45#include "llvm/IR/IntrinsicInst.h" 46#include "llvm/IR/LLVMContext.h" 47#include "llvm/IR/PatternMatch.h" 48#include "llvm/IR/Type.h" 49#include "llvm/Support/CommandLine.h" 50#include "llvm/Support/Debug.h" 51#include "llvm/Support/raw_ostream.h" 52#include "llvm/Transforms/Utils/BasicBlockUtils.h" 53#include "llvm/Transforms/Utils/Local.h" 54#include "llvm/Transforms/Utils/LoopUtils.h" 55#include "llvm/Transforms/Utils/SimplifyIndVar.h" 56using namespace llvm; 57 58#define DEBUG_TYPE "indvars" 59 60STATISTIC(NumWidened , "Number of indvars widened"); 61STATISTIC(NumReplaced , "Number of exit values replaced"); 62STATISTIC(NumLFTR , "Number of loop exit tests replaced"); 63STATISTIC(NumElimExt , "Number of IV sign/zero extends eliminated"); 64STATISTIC(NumElimIV , "Number of congruent IVs eliminated"); 65 66// Trip count verification can be enabled by default under NDEBUG if we 67// implement a strong expression equivalence checker in SCEV. Until then, we 68// use the verify-indvars flag, which may assert in some cases. 69static cl::opt<bool> VerifyIndvars( 70 "verify-indvars", cl::Hidden, 71 cl::desc("Verify the ScalarEvolution result after running indvars")); 72 73enum ReplaceExitVal { NeverRepl, OnlyCheapRepl, AlwaysRepl }; 74 75static cl::opt<ReplaceExitVal> ReplaceExitValue( 76 "replexitval", cl::Hidden, cl::init(OnlyCheapRepl), 77 cl::desc("Choose the strategy to replace exit value in IndVarSimplify"), 78 cl::values(clEnumValN(NeverRepl, "never", "never replace exit value"), 79 clEnumValN(OnlyCheapRepl, "cheap", 80 "only replace exit value when the cost is cheap"), 81 clEnumValN(AlwaysRepl, "always", 82 "always replace exit value whenever possible"), 83 clEnumValEnd)); 84 85namespace { 86struct RewritePhi; 87 88class IndVarSimplify { 89 LoopInfo *LI; 90 ScalarEvolution *SE; 91 DominatorTree *DT; 92 const DataLayout &DL; 93 TargetLibraryInfo *TLI; 94 const TargetTransformInfo *TTI; 95 96 SmallVector<WeakVH, 16> DeadInsts; 97 bool Changed = false; 98 99 bool isValidRewrite(Value *FromVal, Value *ToVal); 100 101 void handleFloatingPointIV(Loop *L, PHINode *PH); 102 void rewriteNonIntegerIVs(Loop *L); 103 104 void simplifyAndExtend(Loop *L, SCEVExpander &Rewriter, LoopInfo *LI); 105 106 bool canLoopBeDeleted(Loop *L, SmallVector<RewritePhi, 8> &RewritePhiSet); 107 void rewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter); 108 void rewriteFirstIterationLoopExitValues(Loop *L); 109 110 Value *linearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount, 111 PHINode *IndVar, SCEVExpander &Rewriter); 112 113 void sinkUnusedInvariants(Loop *L); 114 115 Value *expandSCEVIfNeeded(SCEVExpander &Rewriter, const SCEV *S, Loop *L, 116 Instruction *InsertPt, Type *Ty); 117 118public: 119 IndVarSimplify(LoopInfo *LI, ScalarEvolution *SE, DominatorTree *DT, 120 const DataLayout &DL, TargetLibraryInfo *TLI, 121 TargetTransformInfo *TTI) 122 : LI(LI), SE(SE), DT(DT), DL(DL), TLI(TLI), TTI(TTI) {} 123 124 bool run(Loop *L); 125}; 126} 127 128/// Return true if the SCEV expansion generated by the rewriter can replace the 129/// original value. SCEV guarantees that it produces the same value, but the way 130/// it is produced may be illegal IR. Ideally, this function will only be 131/// called for verification. 132bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) { 133 // If an SCEV expression subsumed multiple pointers, its expansion could 134 // reassociate the GEP changing the base pointer. This is illegal because the 135 // final address produced by a GEP chain must be inbounds relative to its 136 // underlying object. Otherwise basic alias analysis, among other things, 137 // could fail in a dangerous way. Ultimately, SCEV will be improved to avoid 138 // producing an expression involving multiple pointers. Until then, we must 139 // bail out here. 140 // 141 // Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject 142 // because it understands lcssa phis while SCEV does not. 143 Value *FromPtr = FromVal; 144 Value *ToPtr = ToVal; 145 if (auto *GEP = dyn_cast<GEPOperator>(FromVal)) { 146 FromPtr = GEP->getPointerOperand(); 147 } 148 if (auto *GEP = dyn_cast<GEPOperator>(ToVal)) { 149 ToPtr = GEP->getPointerOperand(); 150 } 151 if (FromPtr != FromVal || ToPtr != ToVal) { 152 // Quickly check the common case 153 if (FromPtr == ToPtr) 154 return true; 155 156 // SCEV may have rewritten an expression that produces the GEP's pointer 157 // operand. That's ok as long as the pointer operand has the same base 158 // pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the 159 // base of a recurrence. This handles the case in which SCEV expansion 160 // converts a pointer type recurrence into a nonrecurrent pointer base 161 // indexed by an integer recurrence. 162 163 // If the GEP base pointer is a vector of pointers, abort. 164 if (!FromPtr->getType()->isPointerTy() || !ToPtr->getType()->isPointerTy()) 165 return false; 166 167 const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr)); 168 const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr)); 169 if (FromBase == ToBase) 170 return true; 171 172 DEBUG(dbgs() << "INDVARS: GEP rewrite bail out " 173 << *FromBase << " != " << *ToBase << "\n"); 174 175 return false; 176 } 177 return true; 178} 179 180/// Determine the insertion point for this user. By default, insert immediately 181/// before the user. SCEVExpander or LICM will hoist loop invariants out of the 182/// loop. For PHI nodes, there may be multiple uses, so compute the nearest 183/// common dominator for the incoming blocks. 184static Instruction *getInsertPointForUses(Instruction *User, Value *Def, 185 DominatorTree *DT, LoopInfo *LI) { 186 PHINode *PHI = dyn_cast<PHINode>(User); 187 if (!PHI) 188 return User; 189 190 Instruction *InsertPt = nullptr; 191 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) { 192 if (PHI->getIncomingValue(i) != Def) 193 continue; 194 195 BasicBlock *InsertBB = PHI->getIncomingBlock(i); 196 if (!InsertPt) { 197 InsertPt = InsertBB->getTerminator(); 198 continue; 199 } 200 InsertBB = DT->findNearestCommonDominator(InsertPt->getParent(), InsertBB); 201 InsertPt = InsertBB->getTerminator(); 202 } 203 assert(InsertPt && "Missing phi operand"); 204 205 auto *DefI = dyn_cast<Instruction>(Def); 206 if (!DefI) 207 return InsertPt; 208 209 assert(DT->dominates(DefI, InsertPt) && "def does not dominate all uses"); 210 211 auto *L = LI->getLoopFor(DefI->getParent()); 212 assert(!L || L->contains(LI->getLoopFor(InsertPt->getParent()))); 213 214 for (auto *DTN = (*DT)[InsertPt->getParent()]; DTN; DTN = DTN->getIDom()) 215 if (LI->getLoopFor(DTN->getBlock()) == L) 216 return DTN->getBlock()->getTerminator(); 217 218 llvm_unreachable("DefI dominates InsertPt!"); 219} 220 221//===----------------------------------------------------------------------===// 222// rewriteNonIntegerIVs and helpers. Prefer integer IVs. 223//===----------------------------------------------------------------------===// 224 225/// Convert APF to an integer, if possible. 226static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) { 227 bool isExact = false; 228 // See if we can convert this to an int64_t 229 uint64_t UIntVal; 230 if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero, 231 &isExact) != APFloat::opOK || !isExact) 232 return false; 233 IntVal = UIntVal; 234 return true; 235} 236 237/// If the loop has floating induction variable then insert corresponding 238/// integer induction variable if possible. 239/// For example, 240/// for(double i = 0; i < 10000; ++i) 241/// bar(i) 242/// is converted into 243/// for(int i = 0; i < 10000; ++i) 244/// bar((double)i); 245/// 246void IndVarSimplify::handleFloatingPointIV(Loop *L, PHINode *PN) { 247 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0)); 248 unsigned BackEdge = IncomingEdge^1; 249 250 // Check incoming value. 251 auto *InitValueVal = dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge)); 252 253 int64_t InitValue; 254 if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue)) 255 return; 256 257 // Check IV increment. Reject this PN if increment operation is not 258 // an add or increment value can not be represented by an integer. 259 auto *Incr = dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge)); 260 if (Incr == nullptr || Incr->getOpcode() != Instruction::FAdd) return; 261 262 // If this is not an add of the PHI with a constantfp, or if the constant fp 263 // is not an integer, bail out. 264 ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1)); 265 int64_t IncValue; 266 if (IncValueVal == nullptr || Incr->getOperand(0) != PN || 267 !ConvertToSInt(IncValueVal->getValueAPF(), IncValue)) 268 return; 269 270 // Check Incr uses. One user is PN and the other user is an exit condition 271 // used by the conditional terminator. 272 Value::user_iterator IncrUse = Incr->user_begin(); 273 Instruction *U1 = cast<Instruction>(*IncrUse++); 274 if (IncrUse == Incr->user_end()) return; 275 Instruction *U2 = cast<Instruction>(*IncrUse++); 276 if (IncrUse != Incr->user_end()) return; 277 278 // Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't 279 // only used by a branch, we can't transform it. 280 FCmpInst *Compare = dyn_cast<FCmpInst>(U1); 281 if (!Compare) 282 Compare = dyn_cast<FCmpInst>(U2); 283 if (!Compare || !Compare->hasOneUse() || 284 !isa<BranchInst>(Compare->user_back())) 285 return; 286 287 BranchInst *TheBr = cast<BranchInst>(Compare->user_back()); 288 289 // We need to verify that the branch actually controls the iteration count 290 // of the loop. If not, the new IV can overflow and no one will notice. 291 // The branch block must be in the loop and one of the successors must be out 292 // of the loop. 293 assert(TheBr->isConditional() && "Can't use fcmp if not conditional"); 294 if (!L->contains(TheBr->getParent()) || 295 (L->contains(TheBr->getSuccessor(0)) && 296 L->contains(TheBr->getSuccessor(1)))) 297 return; 298 299 300 // If it isn't a comparison with an integer-as-fp (the exit value), we can't 301 // transform it. 302 ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1)); 303 int64_t ExitValue; 304 if (ExitValueVal == nullptr || 305 !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue)) 306 return; 307 308 // Find new predicate for integer comparison. 309 CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE; 310 switch (Compare->getPredicate()) { 311 default: return; // Unknown comparison. 312 case CmpInst::FCMP_OEQ: 313 case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break; 314 case CmpInst::FCMP_ONE: 315 case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break; 316 case CmpInst::FCMP_OGT: 317 case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break; 318 case CmpInst::FCMP_OGE: 319 case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break; 320 case CmpInst::FCMP_OLT: 321 case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break; 322 case CmpInst::FCMP_OLE: 323 case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break; 324 } 325 326 // We convert the floating point induction variable to a signed i32 value if 327 // we can. This is only safe if the comparison will not overflow in a way 328 // that won't be trapped by the integer equivalent operations. Check for this 329 // now. 330 // TODO: We could use i64 if it is native and the range requires it. 331 332 // The start/stride/exit values must all fit in signed i32. 333 if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue)) 334 return; 335 336 // If not actually striding (add x, 0.0), avoid touching the code. 337 if (IncValue == 0) 338 return; 339 340 // Positive and negative strides have different safety conditions. 341 if (IncValue > 0) { 342 // If we have a positive stride, we require the init to be less than the 343 // exit value. 344 if (InitValue >= ExitValue) 345 return; 346 347 uint32_t Range = uint32_t(ExitValue-InitValue); 348 // Check for infinite loop, either: 349 // while (i <= Exit) or until (i > Exit) 350 if (NewPred == CmpInst::ICMP_SLE || NewPred == CmpInst::ICMP_SGT) { 351 if (++Range == 0) return; // Range overflows. 352 } 353 354 unsigned Leftover = Range % uint32_t(IncValue); 355 356 // If this is an equality comparison, we require that the strided value 357 // exactly land on the exit value, otherwise the IV condition will wrap 358 // around and do things the fp IV wouldn't. 359 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) && 360 Leftover != 0) 361 return; 362 363 // If the stride would wrap around the i32 before exiting, we can't 364 // transform the IV. 365 if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue) 366 return; 367 368 } else { 369 // If we have a negative stride, we require the init to be greater than the 370 // exit value. 371 if (InitValue <= ExitValue) 372 return; 373 374 uint32_t Range = uint32_t(InitValue-ExitValue); 375 // Check for infinite loop, either: 376 // while (i >= Exit) or until (i < Exit) 377 if (NewPred == CmpInst::ICMP_SGE || NewPred == CmpInst::ICMP_SLT) { 378 if (++Range == 0) return; // Range overflows. 379 } 380 381 unsigned Leftover = Range % uint32_t(-IncValue); 382 383 // If this is an equality comparison, we require that the strided value 384 // exactly land on the exit value, otherwise the IV condition will wrap 385 // around and do things the fp IV wouldn't. 386 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) && 387 Leftover != 0) 388 return; 389 390 // If the stride would wrap around the i32 before exiting, we can't 391 // transform the IV. 392 if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue) 393 return; 394 } 395 396 IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext()); 397 398 // Insert new integer induction variable. 399 PHINode *NewPHI = PHINode::Create(Int32Ty, 2, PN->getName()+".int", PN); 400 NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue), 401 PN->getIncomingBlock(IncomingEdge)); 402 403 Value *NewAdd = 404 BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue), 405 Incr->getName()+".int", Incr); 406 NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge)); 407 408 ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd, 409 ConstantInt::get(Int32Ty, ExitValue), 410 Compare->getName()); 411 412 // In the following deletions, PN may become dead and may be deleted. 413 // Use a WeakVH to observe whether this happens. 414 WeakVH WeakPH = PN; 415 416 // Delete the old floating point exit comparison. The branch starts using the 417 // new comparison. 418 NewCompare->takeName(Compare); 419 Compare->replaceAllUsesWith(NewCompare); 420 RecursivelyDeleteTriviallyDeadInstructions(Compare, TLI); 421 422 // Delete the old floating point increment. 423 Incr->replaceAllUsesWith(UndefValue::get(Incr->getType())); 424 RecursivelyDeleteTriviallyDeadInstructions(Incr, TLI); 425 426 // If the FP induction variable still has uses, this is because something else 427 // in the loop uses its value. In order to canonicalize the induction 428 // variable, we chose to eliminate the IV and rewrite it in terms of an 429 // int->fp cast. 430 // 431 // We give preference to sitofp over uitofp because it is faster on most 432 // platforms. 433 if (WeakPH) { 434 Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv", 435 &*PN->getParent()->getFirstInsertionPt()); 436 PN->replaceAllUsesWith(Conv); 437 RecursivelyDeleteTriviallyDeadInstructions(PN, TLI); 438 } 439 Changed = true; 440} 441 442void IndVarSimplify::rewriteNonIntegerIVs(Loop *L) { 443 // First step. Check to see if there are any floating-point recurrences. 444 // If there are, change them into integer recurrences, permitting analysis by 445 // the SCEV routines. 446 // 447 BasicBlock *Header = L->getHeader(); 448 449 SmallVector<WeakVH, 8> PHIs; 450 for (BasicBlock::iterator I = Header->begin(); 451 PHINode *PN = dyn_cast<PHINode>(I); ++I) 452 PHIs.push_back(PN); 453 454 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) 455 if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i])) 456 handleFloatingPointIV(L, PN); 457 458 // If the loop previously had floating-point IV, ScalarEvolution 459 // may not have been able to compute a trip count. Now that we've done some 460 // re-writing, the trip count may be computable. 461 if (Changed) 462 SE->forgetLoop(L); 463} 464 465namespace { 466// Collect information about PHI nodes which can be transformed in 467// rewriteLoopExitValues. 468struct RewritePhi { 469 PHINode *PN; 470 unsigned Ith; // Ith incoming value. 471 Value *Val; // Exit value after expansion. 472 bool HighCost; // High Cost when expansion. 473 474 RewritePhi(PHINode *P, unsigned I, Value *V, bool H) 475 : PN(P), Ith(I), Val(V), HighCost(H) {} 476}; 477} 478 479Value *IndVarSimplify::expandSCEVIfNeeded(SCEVExpander &Rewriter, const SCEV *S, 480 Loop *L, Instruction *InsertPt, 481 Type *ResultTy) { 482 // Before expanding S into an expensive LLVM expression, see if we can use an 483 // already existing value as the expansion for S. 484 if (Value *ExistingValue = Rewriter.getExactExistingExpansion(S, InsertPt, L)) 485 if (ExistingValue->getType() == ResultTy) 486 return ExistingValue; 487 488 // We didn't find anything, fall back to using SCEVExpander. 489 return Rewriter.expandCodeFor(S, ResultTy, InsertPt); 490} 491 492//===----------------------------------------------------------------------===// 493// rewriteLoopExitValues - Optimize IV users outside the loop. 494// As a side effect, reduces the amount of IV processing within the loop. 495//===----------------------------------------------------------------------===// 496 497/// Check to see if this loop has a computable loop-invariant execution count. 498/// If so, this means that we can compute the final value of any expressions 499/// that are recurrent in the loop, and substitute the exit values from the loop 500/// into any instructions outside of the loop that use the final values of the 501/// current expressions. 502/// 503/// This is mostly redundant with the regular IndVarSimplify activities that 504/// happen later, except that it's more powerful in some cases, because it's 505/// able to brute-force evaluate arbitrary instructions as long as they have 506/// constant operands at the beginning of the loop. 507void IndVarSimplify::rewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) { 508 // Check a pre-condition. 509 assert(L->isRecursivelyLCSSAForm(*DT) && "Indvars did not preserve LCSSA!"); 510 511 SmallVector<BasicBlock*, 8> ExitBlocks; 512 L->getUniqueExitBlocks(ExitBlocks); 513 514 SmallVector<RewritePhi, 8> RewritePhiSet; 515 // Find all values that are computed inside the loop, but used outside of it. 516 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan 517 // the exit blocks of the loop to find them. 518 for (BasicBlock *ExitBB : ExitBlocks) { 519 // If there are no PHI nodes in this exit block, then no values defined 520 // inside the loop are used on this path, skip it. 521 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin()); 522 if (!PN) continue; 523 524 unsigned NumPreds = PN->getNumIncomingValues(); 525 526 // Iterate over all of the PHI nodes. 527 BasicBlock::iterator BBI = ExitBB->begin(); 528 while ((PN = dyn_cast<PHINode>(BBI++))) { 529 if (PN->use_empty()) 530 continue; // dead use, don't replace it 531 532 if (!SE->isSCEVable(PN->getType())) 533 continue; 534 535 // It's necessary to tell ScalarEvolution about this explicitly so that 536 // it can walk the def-use list and forget all SCEVs, as it may not be 537 // watching the PHI itself. Once the new exit value is in place, there 538 // may not be a def-use connection between the loop and every instruction 539 // which got a SCEVAddRecExpr for that loop. 540 SE->forgetValue(PN); 541 542 // Iterate over all of the values in all the PHI nodes. 543 for (unsigned i = 0; i != NumPreds; ++i) { 544 // If the value being merged in is not integer or is not defined 545 // in the loop, skip it. 546 Value *InVal = PN->getIncomingValue(i); 547 if (!isa<Instruction>(InVal)) 548 continue; 549 550 // If this pred is for a subloop, not L itself, skip it. 551 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L) 552 continue; // The Block is in a subloop, skip it. 553 554 // Check that InVal is defined in the loop. 555 Instruction *Inst = cast<Instruction>(InVal); 556 if (!L->contains(Inst)) 557 continue; 558 559 // Okay, this instruction has a user outside of the current loop 560 // and varies predictably *inside* the loop. Evaluate the value it 561 // contains when the loop exits, if possible. 562 const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop()); 563 if (!SE->isLoopInvariant(ExitValue, L) || 564 !isSafeToExpand(ExitValue, *SE)) 565 continue; 566 567 // Computing the value outside of the loop brings no benefit if : 568 // - it is definitely used inside the loop in a way which can not be 569 // optimized away. 570 // - no use outside of the loop can take advantage of hoisting the 571 // computation out of the loop 572 if (ExitValue->getSCEVType()>=scMulExpr) { 573 unsigned NumHardInternalUses = 0; 574 unsigned NumSoftExternalUses = 0; 575 unsigned NumUses = 0; 576 for (auto IB = Inst->user_begin(), IE = Inst->user_end(); 577 IB != IE && NumUses <= 6; ++IB) { 578 Instruction *UseInstr = cast<Instruction>(*IB); 579 unsigned Opc = UseInstr->getOpcode(); 580 NumUses++; 581 if (L->contains(UseInstr)) { 582 if (Opc == Instruction::Call || Opc == Instruction::Ret) 583 NumHardInternalUses++; 584 } else { 585 if (Opc == Instruction::PHI) { 586 // Do not count the Phi as a use. LCSSA may have inserted 587 // plenty of trivial ones. 588 NumUses--; 589 for (auto PB = UseInstr->user_begin(), 590 PE = UseInstr->user_end(); 591 PB != PE && NumUses <= 6; ++PB, ++NumUses) { 592 unsigned PhiOpc = cast<Instruction>(*PB)->getOpcode(); 593 if (PhiOpc != Instruction::Call && PhiOpc != Instruction::Ret) 594 NumSoftExternalUses++; 595 } 596 continue; 597 } 598 if (Opc != Instruction::Call && Opc != Instruction::Ret) 599 NumSoftExternalUses++; 600 } 601 } 602 if (NumUses <= 6 && NumHardInternalUses && !NumSoftExternalUses) 603 continue; 604 } 605 606 bool HighCost = Rewriter.isHighCostExpansion(ExitValue, L, Inst); 607 Value *ExitVal = 608 expandSCEVIfNeeded(Rewriter, ExitValue, L, Inst, PN->getType()); 609 610 DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n' 611 << " LoopVal = " << *Inst << "\n"); 612 613 if (!isValidRewrite(Inst, ExitVal)) { 614 DeadInsts.push_back(ExitVal); 615 continue; 616 } 617 618 // Collect all the candidate PHINodes to be rewritten. 619 RewritePhiSet.emplace_back(PN, i, ExitVal, HighCost); 620 } 621 } 622 } 623 624 bool LoopCanBeDel = canLoopBeDeleted(L, RewritePhiSet); 625 626 // Transformation. 627 for (const RewritePhi &Phi : RewritePhiSet) { 628 PHINode *PN = Phi.PN; 629 Value *ExitVal = Phi.Val; 630 631 // Only do the rewrite when the ExitValue can be expanded cheaply. 632 // If LoopCanBeDel is true, rewrite exit value aggressively. 633 if (ReplaceExitValue == OnlyCheapRepl && !LoopCanBeDel && Phi.HighCost) { 634 DeadInsts.push_back(ExitVal); 635 continue; 636 } 637 638 Changed = true; 639 ++NumReplaced; 640 Instruction *Inst = cast<Instruction>(PN->getIncomingValue(Phi.Ith)); 641 PN->setIncomingValue(Phi.Ith, ExitVal); 642 643 // If this instruction is dead now, delete it. Don't do it now to avoid 644 // invalidating iterators. 645 if (isInstructionTriviallyDead(Inst, TLI)) 646 DeadInsts.push_back(Inst); 647 648 // Replace PN with ExitVal if that is legal and does not break LCSSA. 649 if (PN->getNumIncomingValues() == 1 && 650 LI->replacementPreservesLCSSAForm(PN, ExitVal)) { 651 PN->replaceAllUsesWith(ExitVal); 652 PN->eraseFromParent(); 653 } 654 } 655 656 // The insertion point instruction may have been deleted; clear it out 657 // so that the rewriter doesn't trip over it later. 658 Rewriter.clearInsertPoint(); 659} 660 661//===---------------------------------------------------------------------===// 662// rewriteFirstIterationLoopExitValues: Rewrite loop exit values if we know 663// they will exit at the first iteration. 664//===---------------------------------------------------------------------===// 665 666/// Check to see if this loop has loop invariant conditions which lead to loop 667/// exits. If so, we know that if the exit path is taken, it is at the first 668/// loop iteration. This lets us predict exit values of PHI nodes that live in 669/// loop header. 670void IndVarSimplify::rewriteFirstIterationLoopExitValues(Loop *L) { 671 // Verify the input to the pass is already in LCSSA form. 672 assert(L->isLCSSAForm(*DT)); 673 674 SmallVector<BasicBlock *, 8> ExitBlocks; 675 L->getUniqueExitBlocks(ExitBlocks); 676 auto *LoopHeader = L->getHeader(); 677 assert(LoopHeader && "Invalid loop"); 678 679 for (auto *ExitBB : ExitBlocks) { 680 BasicBlock::iterator BBI = ExitBB->begin(); 681 // If there are no more PHI nodes in this exit block, then no more 682 // values defined inside the loop are used on this path. 683 while (auto *PN = dyn_cast<PHINode>(BBI++)) { 684 for (unsigned IncomingValIdx = 0, E = PN->getNumIncomingValues(); 685 IncomingValIdx != E; ++IncomingValIdx) { 686 auto *IncomingBB = PN->getIncomingBlock(IncomingValIdx); 687 688 // We currently only support loop exits from loop header. If the 689 // incoming block is not loop header, we need to recursively check 690 // all conditions starting from loop header are loop invariants. 691 // Additional support might be added in the future. 692 if (IncomingBB != LoopHeader) 693 continue; 694 695 // Get condition that leads to the exit path. 696 auto *TermInst = IncomingBB->getTerminator(); 697 698 Value *Cond = nullptr; 699 if (auto *BI = dyn_cast<BranchInst>(TermInst)) { 700 // Must be a conditional branch, otherwise the block 701 // should not be in the loop. 702 Cond = BI->getCondition(); 703 } else if (auto *SI = dyn_cast<SwitchInst>(TermInst)) 704 Cond = SI->getCondition(); 705 else 706 continue; 707 708 if (!L->isLoopInvariant(Cond)) 709 continue; 710 711 auto *ExitVal = 712 dyn_cast<PHINode>(PN->getIncomingValue(IncomingValIdx)); 713 714 // Only deal with PHIs. 715 if (!ExitVal) 716 continue; 717 718 // If ExitVal is a PHI on the loop header, then we know its 719 // value along this exit because the exit can only be taken 720 // on the first iteration. 721 auto *LoopPreheader = L->getLoopPreheader(); 722 assert(LoopPreheader && "Invalid loop"); 723 int PreheaderIdx = ExitVal->getBasicBlockIndex(LoopPreheader); 724 if (PreheaderIdx != -1) { 725 assert(ExitVal->getParent() == LoopHeader && 726 "ExitVal must be in loop header"); 727 PN->setIncomingValue(IncomingValIdx, 728 ExitVal->getIncomingValue(PreheaderIdx)); 729 } 730 } 731 } 732 } 733} 734 735/// Check whether it is possible to delete the loop after rewriting exit 736/// value. If it is possible, ignore ReplaceExitValue and do rewriting 737/// aggressively. 738bool IndVarSimplify::canLoopBeDeleted( 739 Loop *L, SmallVector<RewritePhi, 8> &RewritePhiSet) { 740 741 BasicBlock *Preheader = L->getLoopPreheader(); 742 // If there is no preheader, the loop will not be deleted. 743 if (!Preheader) 744 return false; 745 746 // In LoopDeletion pass Loop can be deleted when ExitingBlocks.size() > 1. 747 // We obviate multiple ExitingBlocks case for simplicity. 748 // TODO: If we see testcase with multiple ExitingBlocks can be deleted 749 // after exit value rewriting, we can enhance the logic here. 750 SmallVector<BasicBlock *, 4> ExitingBlocks; 751 L->getExitingBlocks(ExitingBlocks); 752 SmallVector<BasicBlock *, 8> ExitBlocks; 753 L->getUniqueExitBlocks(ExitBlocks); 754 if (ExitBlocks.size() > 1 || ExitingBlocks.size() > 1) 755 return false; 756 757 BasicBlock *ExitBlock = ExitBlocks[0]; 758 BasicBlock::iterator BI = ExitBlock->begin(); 759 while (PHINode *P = dyn_cast<PHINode>(BI)) { 760 Value *Incoming = P->getIncomingValueForBlock(ExitingBlocks[0]); 761 762 // If the Incoming value of P is found in RewritePhiSet, we know it 763 // could be rewritten to use a loop invariant value in transformation 764 // phase later. Skip it in the loop invariant check below. 765 bool found = false; 766 for (const RewritePhi &Phi : RewritePhiSet) { 767 unsigned i = Phi.Ith; 768 if (Phi.PN == P && (Phi.PN)->getIncomingValue(i) == Incoming) { 769 found = true; 770 break; 771 } 772 } 773 774 Instruction *I; 775 if (!found && (I = dyn_cast<Instruction>(Incoming))) 776 if (!L->hasLoopInvariantOperands(I)) 777 return false; 778 779 ++BI; 780 } 781 782 for (auto *BB : L->blocks()) 783 if (any_of(*BB, [](Instruction &I) { return I.mayHaveSideEffects(); })) 784 return false; 785 786 return true; 787} 788 789//===----------------------------------------------------------------------===// 790// IV Widening - Extend the width of an IV to cover its widest uses. 791//===----------------------------------------------------------------------===// 792 793namespace { 794// Collect information about induction variables that are used by sign/zero 795// extend operations. This information is recorded by CollectExtend and provides 796// the input to WidenIV. 797struct WideIVInfo { 798 PHINode *NarrowIV = nullptr; 799 Type *WidestNativeType = nullptr; // Widest integer type created [sz]ext 800 bool IsSigned = false; // Was a sext user seen before a zext? 801}; 802} 803 804/// Update information about the induction variable that is extended by this 805/// sign or zero extend operation. This is used to determine the final width of 806/// the IV before actually widening it. 807static void visitIVCast(CastInst *Cast, WideIVInfo &WI, ScalarEvolution *SE, 808 const TargetTransformInfo *TTI) { 809 bool IsSigned = Cast->getOpcode() == Instruction::SExt; 810 if (!IsSigned && Cast->getOpcode() != Instruction::ZExt) 811 return; 812 813 Type *Ty = Cast->getType(); 814 uint64_t Width = SE->getTypeSizeInBits(Ty); 815 if (!Cast->getModule()->getDataLayout().isLegalInteger(Width)) 816 return; 817 818 // Check that `Cast` actually extends the induction variable (we rely on this 819 // later). This takes care of cases where `Cast` is extending a truncation of 820 // the narrow induction variable, and thus can end up being narrower than the 821 // "narrow" induction variable. 822 uint64_t NarrowIVWidth = SE->getTypeSizeInBits(WI.NarrowIV->getType()); 823 if (NarrowIVWidth >= Width) 824 return; 825 826 // Cast is either an sext or zext up to this point. 827 // We should not widen an indvar if arithmetics on the wider indvar are more 828 // expensive than those on the narrower indvar. We check only the cost of ADD 829 // because at least an ADD is required to increment the induction variable. We 830 // could compute more comprehensively the cost of all instructions on the 831 // induction variable when necessary. 832 if (TTI && 833 TTI->getArithmeticInstrCost(Instruction::Add, Ty) > 834 TTI->getArithmeticInstrCost(Instruction::Add, 835 Cast->getOperand(0)->getType())) { 836 return; 837 } 838 839 if (!WI.WidestNativeType) { 840 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty); 841 WI.IsSigned = IsSigned; 842 return; 843 } 844 845 // We extend the IV to satisfy the sign of its first user, arbitrarily. 846 if (WI.IsSigned != IsSigned) 847 return; 848 849 if (Width > SE->getTypeSizeInBits(WI.WidestNativeType)) 850 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty); 851} 852 853namespace { 854 855/// Record a link in the Narrow IV def-use chain along with the WideIV that 856/// computes the same value as the Narrow IV def. This avoids caching Use* 857/// pointers. 858struct NarrowIVDefUse { 859 Instruction *NarrowDef = nullptr; 860 Instruction *NarrowUse = nullptr; 861 Instruction *WideDef = nullptr; 862 863 // True if the narrow def is never negative. Tracking this information lets 864 // us use a sign extension instead of a zero extension or vice versa, when 865 // profitable and legal. 866 bool NeverNegative = false; 867 868 NarrowIVDefUse(Instruction *ND, Instruction *NU, Instruction *WD, 869 bool NeverNegative) 870 : NarrowDef(ND), NarrowUse(NU), WideDef(WD), 871 NeverNegative(NeverNegative) {} 872}; 873 874/// The goal of this transform is to remove sign and zero extends without 875/// creating any new induction variables. To do this, it creates a new phi of 876/// the wider type and redirects all users, either removing extends or inserting 877/// truncs whenever we stop propagating the type. 878/// 879class WidenIV { 880 // Parameters 881 PHINode *OrigPhi; 882 Type *WideType; 883 bool IsSigned; 884 885 // Context 886 LoopInfo *LI; 887 Loop *L; 888 ScalarEvolution *SE; 889 DominatorTree *DT; 890 891 // Result 892 PHINode *WidePhi; 893 Instruction *WideInc; 894 const SCEV *WideIncExpr; 895 SmallVectorImpl<WeakVH> &DeadInsts; 896 897 SmallPtrSet<Instruction*,16> Widened; 898 SmallVector<NarrowIVDefUse, 8> NarrowIVUsers; 899 900public: 901 WidenIV(const WideIVInfo &WI, LoopInfo *LInfo, 902 ScalarEvolution *SEv, DominatorTree *DTree, 903 SmallVectorImpl<WeakVH> &DI) : 904 OrigPhi(WI.NarrowIV), 905 WideType(WI.WidestNativeType), 906 IsSigned(WI.IsSigned), 907 LI(LInfo), 908 L(LI->getLoopFor(OrigPhi->getParent())), 909 SE(SEv), 910 DT(DTree), 911 WidePhi(nullptr), 912 WideInc(nullptr), 913 WideIncExpr(nullptr), 914 DeadInsts(DI) { 915 assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV"); 916 } 917 918 PHINode *createWideIV(SCEVExpander &Rewriter); 919 920protected: 921 Value *createExtendInst(Value *NarrowOper, Type *WideType, bool IsSigned, 922 Instruction *Use); 923 924 Instruction *cloneIVUser(NarrowIVDefUse DU, const SCEVAddRecExpr *WideAR); 925 Instruction *cloneArithmeticIVUser(NarrowIVDefUse DU, 926 const SCEVAddRecExpr *WideAR); 927 Instruction *cloneBitwiseIVUser(NarrowIVDefUse DU); 928 929 const SCEVAddRecExpr *getWideRecurrence(Instruction *NarrowUse); 930 931 const SCEVAddRecExpr* getExtendedOperandRecurrence(NarrowIVDefUse DU); 932 933 const SCEV *getSCEVByOpCode(const SCEV *LHS, const SCEV *RHS, 934 unsigned OpCode) const; 935 936 Instruction *widenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter); 937 938 bool widenLoopCompare(NarrowIVDefUse DU); 939 940 void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef); 941}; 942} // anonymous namespace 943 944/// Perform a quick domtree based check for loop invariance assuming that V is 945/// used within the loop. LoopInfo::isLoopInvariant() seems gratuitous for this 946/// purpose. 947static bool isLoopInvariant(Value *V, const Loop *L, const DominatorTree *DT) { 948 Instruction *Inst = dyn_cast<Instruction>(V); 949 if (!Inst) 950 return true; 951 952 return DT->properlyDominates(Inst->getParent(), L->getHeader()); 953} 954 955Value *WidenIV::createExtendInst(Value *NarrowOper, Type *WideType, 956 bool IsSigned, Instruction *Use) { 957 // Set the debug location and conservative insertion point. 958 IRBuilder<> Builder(Use); 959 // Hoist the insertion point into loop preheaders as far as possible. 960 for (const Loop *L = LI->getLoopFor(Use->getParent()); 961 L && L->getLoopPreheader() && isLoopInvariant(NarrowOper, L, DT); 962 L = L->getParentLoop()) 963 Builder.SetInsertPoint(L->getLoopPreheader()->getTerminator()); 964 965 return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) : 966 Builder.CreateZExt(NarrowOper, WideType); 967} 968 969/// Instantiate a wide operation to replace a narrow operation. This only needs 970/// to handle operations that can evaluation to SCEVAddRec. It can safely return 971/// 0 for any operation we decide not to clone. 972Instruction *WidenIV::cloneIVUser(NarrowIVDefUse DU, 973 const SCEVAddRecExpr *WideAR) { 974 unsigned Opcode = DU.NarrowUse->getOpcode(); 975 switch (Opcode) { 976 default: 977 return nullptr; 978 case Instruction::Add: 979 case Instruction::Mul: 980 case Instruction::UDiv: 981 case Instruction::Sub: 982 return cloneArithmeticIVUser(DU, WideAR); 983 984 case Instruction::And: 985 case Instruction::Or: 986 case Instruction::Xor: 987 case Instruction::Shl: 988 case Instruction::LShr: 989 case Instruction::AShr: 990 return cloneBitwiseIVUser(DU); 991 } 992} 993 994Instruction *WidenIV::cloneBitwiseIVUser(NarrowIVDefUse DU) { 995 Instruction *NarrowUse = DU.NarrowUse; 996 Instruction *NarrowDef = DU.NarrowDef; 997 Instruction *WideDef = DU.WideDef; 998 999 DEBUG(dbgs() << "Cloning bitwise IVUser: " << *NarrowUse << "\n"); 1000 1001 // Replace NarrowDef operands with WideDef. Otherwise, we don't know anything 1002 // about the narrow operand yet so must insert a [sz]ext. It is probably loop 1003 // invariant and will be folded or hoisted. If it actually comes from a 1004 // widened IV, it should be removed during a future call to widenIVUse. 1005 Value *LHS = (NarrowUse->getOperand(0) == NarrowDef) 1006 ? WideDef 1007 : createExtendInst(NarrowUse->getOperand(0), WideType, 1008 IsSigned, NarrowUse); 1009 Value *RHS = (NarrowUse->getOperand(1) == NarrowDef) 1010 ? WideDef 1011 : createExtendInst(NarrowUse->getOperand(1), WideType, 1012 IsSigned, NarrowUse); 1013 1014 auto *NarrowBO = cast<BinaryOperator>(NarrowUse); 1015 auto *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(), LHS, RHS, 1016 NarrowBO->getName()); 1017 IRBuilder<> Builder(NarrowUse); 1018 Builder.Insert(WideBO); 1019 WideBO->copyIRFlags(NarrowBO); 1020 return WideBO; 1021} 1022 1023Instruction *WidenIV::cloneArithmeticIVUser(NarrowIVDefUse DU, 1024 const SCEVAddRecExpr *WideAR) { 1025 Instruction *NarrowUse = DU.NarrowUse; 1026 Instruction *NarrowDef = DU.NarrowDef; 1027 Instruction *WideDef = DU.WideDef; 1028 1029 DEBUG(dbgs() << "Cloning arithmetic IVUser: " << *NarrowUse << "\n"); 1030 1031 unsigned IVOpIdx = (NarrowUse->getOperand(0) == NarrowDef) ? 0 : 1; 1032 1033 // We're trying to find X such that 1034 // 1035 // Widen(NarrowDef `op` NonIVNarrowDef) == WideAR == WideDef `op.wide` X 1036 // 1037 // We guess two solutions to X, sext(NonIVNarrowDef) and zext(NonIVNarrowDef), 1038 // and check using SCEV if any of them are correct. 1039 1040 // Returns true if extending NonIVNarrowDef according to `SignExt` is a 1041 // correct solution to X. 1042 auto GuessNonIVOperand = [&](bool SignExt) { 1043 const SCEV *WideLHS; 1044 const SCEV *WideRHS; 1045 1046 auto GetExtend = [this, SignExt](const SCEV *S, Type *Ty) { 1047 if (SignExt) 1048 return SE->getSignExtendExpr(S, Ty); 1049 return SE->getZeroExtendExpr(S, Ty); 1050 }; 1051 1052 if (IVOpIdx == 0) { 1053 WideLHS = SE->getSCEV(WideDef); 1054 const SCEV *NarrowRHS = SE->getSCEV(NarrowUse->getOperand(1)); 1055 WideRHS = GetExtend(NarrowRHS, WideType); 1056 } else { 1057 const SCEV *NarrowLHS = SE->getSCEV(NarrowUse->getOperand(0)); 1058 WideLHS = GetExtend(NarrowLHS, WideType); 1059 WideRHS = SE->getSCEV(WideDef); 1060 } 1061 1062 // WideUse is "WideDef `op.wide` X" as described in the comment. 1063 const SCEV *WideUse = nullptr; 1064 1065 switch (NarrowUse->getOpcode()) { 1066 default: 1067 llvm_unreachable("No other possibility!"); 1068 1069 case Instruction::Add: 1070 WideUse = SE->getAddExpr(WideLHS, WideRHS); 1071 break; 1072 1073 case Instruction::Mul: 1074 WideUse = SE->getMulExpr(WideLHS, WideRHS); 1075 break; 1076 1077 case Instruction::UDiv: 1078 WideUse = SE->getUDivExpr(WideLHS, WideRHS); 1079 break; 1080 1081 case Instruction::Sub: 1082 WideUse = SE->getMinusSCEV(WideLHS, WideRHS); 1083 break; 1084 } 1085 1086 return WideUse == WideAR; 1087 }; 1088 1089 bool SignExtend = IsSigned; 1090 if (!GuessNonIVOperand(SignExtend)) { 1091 SignExtend = !SignExtend; 1092 if (!GuessNonIVOperand(SignExtend)) 1093 return nullptr; 1094 } 1095 1096 Value *LHS = (NarrowUse->getOperand(0) == NarrowDef) 1097 ? WideDef 1098 : createExtendInst(NarrowUse->getOperand(0), WideType, 1099 SignExtend, NarrowUse); 1100 Value *RHS = (NarrowUse->getOperand(1) == NarrowDef) 1101 ? WideDef 1102 : createExtendInst(NarrowUse->getOperand(1), WideType, 1103 SignExtend, NarrowUse); 1104 1105 auto *NarrowBO = cast<BinaryOperator>(NarrowUse); 1106 auto *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(), LHS, RHS, 1107 NarrowBO->getName()); 1108 1109 IRBuilder<> Builder(NarrowUse); 1110 Builder.Insert(WideBO); 1111 WideBO->copyIRFlags(NarrowBO); 1112 return WideBO; 1113} 1114 1115const SCEV *WidenIV::getSCEVByOpCode(const SCEV *LHS, const SCEV *RHS, 1116 unsigned OpCode) const { 1117 if (OpCode == Instruction::Add) 1118 return SE->getAddExpr(LHS, RHS); 1119 if (OpCode == Instruction::Sub) 1120 return SE->getMinusSCEV(LHS, RHS); 1121 if (OpCode == Instruction::Mul) 1122 return SE->getMulExpr(LHS, RHS); 1123 1124 llvm_unreachable("Unsupported opcode."); 1125} 1126 1127/// No-wrap operations can transfer sign extension of their result to their 1128/// operands. Generate the SCEV value for the widened operation without 1129/// actually modifying the IR yet. If the expression after extending the 1130/// operands is an AddRec for this loop, return it. 1131const SCEVAddRecExpr* WidenIV::getExtendedOperandRecurrence(NarrowIVDefUse DU) { 1132 1133 // Handle the common case of add<nsw/nuw> 1134 const unsigned OpCode = DU.NarrowUse->getOpcode(); 1135 // Only Add/Sub/Mul instructions supported yet. 1136 if (OpCode != Instruction::Add && OpCode != Instruction::Sub && 1137 OpCode != Instruction::Mul) 1138 return nullptr; 1139 1140 // One operand (NarrowDef) has already been extended to WideDef. Now determine 1141 // if extending the other will lead to a recurrence. 1142 const unsigned ExtendOperIdx = 1143 DU.NarrowUse->getOperand(0) == DU.NarrowDef ? 1 : 0; 1144 assert(DU.NarrowUse->getOperand(1-ExtendOperIdx) == DU.NarrowDef && "bad DU"); 1145 1146 const SCEV *ExtendOperExpr = nullptr; 1147 const OverflowingBinaryOperator *OBO = 1148 cast<OverflowingBinaryOperator>(DU.NarrowUse); 1149 if (IsSigned && OBO->hasNoSignedWrap()) 1150 ExtendOperExpr = SE->getSignExtendExpr( 1151 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType); 1152 else if(!IsSigned && OBO->hasNoUnsignedWrap()) 1153 ExtendOperExpr = SE->getZeroExtendExpr( 1154 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType); 1155 else 1156 return nullptr; 1157 1158 // When creating this SCEV expr, don't apply the current operations NSW or NUW 1159 // flags. This instruction may be guarded by control flow that the no-wrap 1160 // behavior depends on. Non-control-equivalent instructions can be mapped to 1161 // the same SCEV expression, and it would be incorrect to transfer NSW/NUW 1162 // semantics to those operations. 1163 const SCEV *lhs = SE->getSCEV(DU.WideDef); 1164 const SCEV *rhs = ExtendOperExpr; 1165 1166 // Let's swap operands to the initial order for the case of non-commutative 1167 // operations, like SUB. See PR21014. 1168 if (ExtendOperIdx == 0) 1169 std::swap(lhs, rhs); 1170 const SCEVAddRecExpr *AddRec = 1171 dyn_cast<SCEVAddRecExpr>(getSCEVByOpCode(lhs, rhs, OpCode)); 1172 1173 if (!AddRec || AddRec->getLoop() != L) 1174 return nullptr; 1175 return AddRec; 1176} 1177 1178/// Is this instruction potentially interesting for further simplification after 1179/// widening it's type? In other words, can the extend be safely hoisted out of 1180/// the loop with SCEV reducing the value to a recurrence on the same loop. If 1181/// so, return the sign or zero extended recurrence. Otherwise return NULL. 1182const SCEVAddRecExpr *WidenIV::getWideRecurrence(Instruction *NarrowUse) { 1183 if (!SE->isSCEVable(NarrowUse->getType())) 1184 return nullptr; 1185 1186 const SCEV *NarrowExpr = SE->getSCEV(NarrowUse); 1187 if (SE->getTypeSizeInBits(NarrowExpr->getType()) 1188 >= SE->getTypeSizeInBits(WideType)) { 1189 // NarrowUse implicitly widens its operand. e.g. a gep with a narrow 1190 // index. So don't follow this use. 1191 return nullptr; 1192 } 1193 1194 const SCEV *WideExpr = IsSigned ? 1195 SE->getSignExtendExpr(NarrowExpr, WideType) : 1196 SE->getZeroExtendExpr(NarrowExpr, WideType); 1197 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr); 1198 if (!AddRec || AddRec->getLoop() != L) 1199 return nullptr; 1200 return AddRec; 1201} 1202 1203/// This IV user cannot be widen. Replace this use of the original narrow IV 1204/// with a truncation of the new wide IV to isolate and eliminate the narrow IV. 1205static void truncateIVUse(NarrowIVDefUse DU, DominatorTree *DT, LoopInfo *LI) { 1206 DEBUG(dbgs() << "INDVARS: Truncate IV " << *DU.WideDef 1207 << " for user " << *DU.NarrowUse << "\n"); 1208 IRBuilder<> Builder( 1209 getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT, LI)); 1210 Value *Trunc = Builder.CreateTrunc(DU.WideDef, DU.NarrowDef->getType()); 1211 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, Trunc); 1212} 1213 1214/// If the narrow use is a compare instruction, then widen the compare 1215// (and possibly the other operand). The extend operation is hoisted into the 1216// loop preheader as far as possible. 1217bool WidenIV::widenLoopCompare(NarrowIVDefUse DU) { 1218 ICmpInst *Cmp = dyn_cast<ICmpInst>(DU.NarrowUse); 1219 if (!Cmp) 1220 return false; 1221 1222 // We can legally widen the comparison in the following two cases: 1223 // 1224 // - The signedness of the IV extension and comparison match 1225 // 1226 // - The narrow IV is always positive (and thus its sign extension is equal 1227 // to its zero extension). For instance, let's say we're zero extending 1228 // %narrow for the following use 1229 // 1230 // icmp slt i32 %narrow, %val ... (A) 1231 // 1232 // and %narrow is always positive. Then 1233 // 1234 // (A) == icmp slt i32 sext(%narrow), sext(%val) 1235 // == icmp slt i32 zext(%narrow), sext(%val) 1236 1237 if (!(DU.NeverNegative || IsSigned == Cmp->isSigned())) 1238 return false; 1239 1240 Value *Op = Cmp->getOperand(Cmp->getOperand(0) == DU.NarrowDef ? 1 : 0); 1241 unsigned CastWidth = SE->getTypeSizeInBits(Op->getType()); 1242 unsigned IVWidth = SE->getTypeSizeInBits(WideType); 1243 assert (CastWidth <= IVWidth && "Unexpected width while widening compare."); 1244 1245 // Widen the compare instruction. 1246 IRBuilder<> Builder( 1247 getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT, LI)); 1248 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef); 1249 1250 // Widen the other operand of the compare, if necessary. 1251 if (CastWidth < IVWidth) { 1252 Value *ExtOp = createExtendInst(Op, WideType, Cmp->isSigned(), Cmp); 1253 DU.NarrowUse->replaceUsesOfWith(Op, ExtOp); 1254 } 1255 return true; 1256} 1257 1258/// Determine whether an individual user of the narrow IV can be widened. If so, 1259/// return the wide clone of the user. 1260Instruction *WidenIV::widenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter) { 1261 1262 // Stop traversing the def-use chain at inner-loop phis or post-loop phis. 1263 if (PHINode *UsePhi = dyn_cast<PHINode>(DU.NarrowUse)) { 1264 if (LI->getLoopFor(UsePhi->getParent()) != L) { 1265 // For LCSSA phis, sink the truncate outside the loop. 1266 // After SimplifyCFG most loop exit targets have a single predecessor. 1267 // Otherwise fall back to a truncate within the loop. 1268 if (UsePhi->getNumOperands() != 1) 1269 truncateIVUse(DU, DT, LI); 1270 else { 1271 // Widening the PHI requires us to insert a trunc. The logical place 1272 // for this trunc is in the same BB as the PHI. This is not possible if 1273 // the BB is terminated by a catchswitch. 1274 if (isa<CatchSwitchInst>(UsePhi->getParent()->getTerminator())) 1275 return nullptr; 1276 1277 PHINode *WidePhi = 1278 PHINode::Create(DU.WideDef->getType(), 1, UsePhi->getName() + ".wide", 1279 UsePhi); 1280 WidePhi->addIncoming(DU.WideDef, UsePhi->getIncomingBlock(0)); 1281 IRBuilder<> Builder(&*WidePhi->getParent()->getFirstInsertionPt()); 1282 Value *Trunc = Builder.CreateTrunc(WidePhi, DU.NarrowDef->getType()); 1283 UsePhi->replaceAllUsesWith(Trunc); 1284 DeadInsts.emplace_back(UsePhi); 1285 DEBUG(dbgs() << "INDVARS: Widen lcssa phi " << *UsePhi 1286 << " to " << *WidePhi << "\n"); 1287 } 1288 return nullptr; 1289 } 1290 } 1291 // Our raison d'etre! Eliminate sign and zero extension. 1292 if (IsSigned ? isa<SExtInst>(DU.NarrowUse) : isa<ZExtInst>(DU.NarrowUse)) { 1293 Value *NewDef = DU.WideDef; 1294 if (DU.NarrowUse->getType() != WideType) { 1295 unsigned CastWidth = SE->getTypeSizeInBits(DU.NarrowUse->getType()); 1296 unsigned IVWidth = SE->getTypeSizeInBits(WideType); 1297 if (CastWidth < IVWidth) { 1298 // The cast isn't as wide as the IV, so insert a Trunc. 1299 IRBuilder<> Builder(DU.NarrowUse); 1300 NewDef = Builder.CreateTrunc(DU.WideDef, DU.NarrowUse->getType()); 1301 } 1302 else { 1303 // A wider extend was hidden behind a narrower one. This may induce 1304 // another round of IV widening in which the intermediate IV becomes 1305 // dead. It should be very rare. 1306 DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi 1307 << " not wide enough to subsume " << *DU.NarrowUse << "\n"); 1308 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef); 1309 NewDef = DU.NarrowUse; 1310 } 1311 } 1312 if (NewDef != DU.NarrowUse) { 1313 DEBUG(dbgs() << "INDVARS: eliminating " << *DU.NarrowUse 1314 << " replaced by " << *DU.WideDef << "\n"); 1315 ++NumElimExt; 1316 DU.NarrowUse->replaceAllUsesWith(NewDef); 1317 DeadInsts.emplace_back(DU.NarrowUse); 1318 } 1319 // Now that the extend is gone, we want to expose it's uses for potential 1320 // further simplification. We don't need to directly inform SimplifyIVUsers 1321 // of the new users, because their parent IV will be processed later as a 1322 // new loop phi. If we preserved IVUsers analysis, we would also want to 1323 // push the uses of WideDef here. 1324 1325 // No further widening is needed. The deceased [sz]ext had done it for us. 1326 return nullptr; 1327 } 1328 1329 // Does this user itself evaluate to a recurrence after widening? 1330 const SCEVAddRecExpr *WideAddRec = getWideRecurrence(DU.NarrowUse); 1331 if (!WideAddRec) 1332 WideAddRec = getExtendedOperandRecurrence(DU); 1333 1334 if (!WideAddRec) { 1335 // If use is a loop condition, try to promote the condition instead of 1336 // truncating the IV first. 1337 if (widenLoopCompare(DU)) 1338 return nullptr; 1339 1340 // This user does not evaluate to a recurence after widening, so don't 1341 // follow it. Instead insert a Trunc to kill off the original use, 1342 // eventually isolating the original narrow IV so it can be removed. 1343 truncateIVUse(DU, DT, LI); 1344 return nullptr; 1345 } 1346 // Assume block terminators cannot evaluate to a recurrence. We can't to 1347 // insert a Trunc after a terminator if there happens to be a critical edge. 1348 assert(DU.NarrowUse != DU.NarrowUse->getParent()->getTerminator() && 1349 "SCEV is not expected to evaluate a block terminator"); 1350 1351 // Reuse the IV increment that SCEVExpander created as long as it dominates 1352 // NarrowUse. 1353 Instruction *WideUse = nullptr; 1354 if (WideAddRec == WideIncExpr && Rewriter.hoistIVInc(WideInc, DU.NarrowUse)) 1355 WideUse = WideInc; 1356 else { 1357 WideUse = cloneIVUser(DU, WideAddRec); 1358 if (!WideUse) 1359 return nullptr; 1360 } 1361 // Evaluation of WideAddRec ensured that the narrow expression could be 1362 // extended outside the loop without overflow. This suggests that the wide use 1363 // evaluates to the same expression as the extended narrow use, but doesn't 1364 // absolutely guarantee it. Hence the following failsafe check. In rare cases 1365 // where it fails, we simply throw away the newly created wide use. 1366 if (WideAddRec != SE->getSCEV(WideUse)) { 1367 DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse 1368 << ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n"); 1369 DeadInsts.emplace_back(WideUse); 1370 return nullptr; 1371 } 1372 1373 // Returning WideUse pushes it on the worklist. 1374 return WideUse; 1375} 1376 1377/// Add eligible users of NarrowDef to NarrowIVUsers. 1378/// 1379void WidenIV::pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef) { 1380 const SCEV *NarrowSCEV = SE->getSCEV(NarrowDef); 1381 bool NeverNegative = 1382 SE->isKnownPredicate(ICmpInst::ICMP_SGE, NarrowSCEV, 1383 SE->getConstant(NarrowSCEV->getType(), 0)); 1384 for (User *U : NarrowDef->users()) { 1385 Instruction *NarrowUser = cast<Instruction>(U); 1386 1387 // Handle data flow merges and bizarre phi cycles. 1388 if (!Widened.insert(NarrowUser).second) 1389 continue; 1390 1391 NarrowIVUsers.emplace_back(NarrowDef, NarrowUser, WideDef, NeverNegative); 1392 } 1393} 1394 1395/// Process a single induction variable. First use the SCEVExpander to create a 1396/// wide induction variable that evaluates to the same recurrence as the 1397/// original narrow IV. Then use a worklist to forward traverse the narrow IV's 1398/// def-use chain. After widenIVUse has processed all interesting IV users, the 1399/// narrow IV will be isolated for removal by DeleteDeadPHIs. 1400/// 1401/// It would be simpler to delete uses as they are processed, but we must avoid 1402/// invalidating SCEV expressions. 1403/// 1404PHINode *WidenIV::createWideIV(SCEVExpander &Rewriter) { 1405 // Is this phi an induction variable? 1406 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi)); 1407 if (!AddRec) 1408 return nullptr; 1409 1410 // Widen the induction variable expression. 1411 const SCEV *WideIVExpr = IsSigned ? 1412 SE->getSignExtendExpr(AddRec, WideType) : 1413 SE->getZeroExtendExpr(AddRec, WideType); 1414 1415 assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType && 1416 "Expect the new IV expression to preserve its type"); 1417 1418 // Can the IV be extended outside the loop without overflow? 1419 AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr); 1420 if (!AddRec || AddRec->getLoop() != L) 1421 return nullptr; 1422 1423 // An AddRec must have loop-invariant operands. Since this AddRec is 1424 // materialized by a loop header phi, the expression cannot have any post-loop 1425 // operands, so they must dominate the loop header. 1426 assert( 1427 SE->properlyDominates(AddRec->getStart(), L->getHeader()) && 1428 SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader()) && 1429 "Loop header phi recurrence inputs do not dominate the loop"); 1430 1431 // The rewriter provides a value for the desired IV expression. This may 1432 // either find an existing phi or materialize a new one. Either way, we 1433 // expect a well-formed cyclic phi-with-increments. i.e. any operand not part 1434 // of the phi-SCC dominates the loop entry. 1435 Instruction *InsertPt = &L->getHeader()->front(); 1436 WidePhi = cast<PHINode>(Rewriter.expandCodeFor(AddRec, WideType, InsertPt)); 1437 1438 // Remembering the WideIV increment generated by SCEVExpander allows 1439 // widenIVUse to reuse it when widening the narrow IV's increment. We don't 1440 // employ a general reuse mechanism because the call above is the only call to 1441 // SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses. 1442 if (BasicBlock *LatchBlock = L->getLoopLatch()) { 1443 WideInc = 1444 cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock)); 1445 WideIncExpr = SE->getSCEV(WideInc); 1446 } 1447 1448 DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n"); 1449 ++NumWidened; 1450 1451 // Traverse the def-use chain using a worklist starting at the original IV. 1452 assert(Widened.empty() && NarrowIVUsers.empty() && "expect initial state" ); 1453 1454 Widened.insert(OrigPhi); 1455 pushNarrowIVUsers(OrigPhi, WidePhi); 1456 1457 while (!NarrowIVUsers.empty()) { 1458 NarrowIVDefUse DU = NarrowIVUsers.pop_back_val(); 1459 1460 // Process a def-use edge. This may replace the use, so don't hold a 1461 // use_iterator across it. 1462 Instruction *WideUse = widenIVUse(DU, Rewriter); 1463 1464 // Follow all def-use edges from the previous narrow use. 1465 if (WideUse) 1466 pushNarrowIVUsers(DU.NarrowUse, WideUse); 1467 1468 // widenIVUse may have removed the def-use edge. 1469 if (DU.NarrowDef->use_empty()) 1470 DeadInsts.emplace_back(DU.NarrowDef); 1471 } 1472 return WidePhi; 1473} 1474 1475//===----------------------------------------------------------------------===// 1476// Live IV Reduction - Minimize IVs live across the loop. 1477//===----------------------------------------------------------------------===// 1478 1479 1480//===----------------------------------------------------------------------===// 1481// Simplification of IV users based on SCEV evaluation. 1482//===----------------------------------------------------------------------===// 1483 1484namespace { 1485class IndVarSimplifyVisitor : public IVVisitor { 1486 ScalarEvolution *SE; 1487 const TargetTransformInfo *TTI; 1488 PHINode *IVPhi; 1489 1490public: 1491 WideIVInfo WI; 1492 1493 IndVarSimplifyVisitor(PHINode *IV, ScalarEvolution *SCEV, 1494 const TargetTransformInfo *TTI, 1495 const DominatorTree *DTree) 1496 : SE(SCEV), TTI(TTI), IVPhi(IV) { 1497 DT = DTree; 1498 WI.NarrowIV = IVPhi; 1499 } 1500 1501 // Implement the interface used by simplifyUsersOfIV. 1502 void visitCast(CastInst *Cast) override { visitIVCast(Cast, WI, SE, TTI); } 1503}; 1504} 1505 1506/// Iteratively perform simplification on a worklist of IV users. Each 1507/// successive simplification may push more users which may themselves be 1508/// candidates for simplification. 1509/// 1510/// Sign/Zero extend elimination is interleaved with IV simplification. 1511/// 1512void IndVarSimplify::simplifyAndExtend(Loop *L, 1513 SCEVExpander &Rewriter, 1514 LoopInfo *LI) { 1515 SmallVector<WideIVInfo, 8> WideIVs; 1516 1517 SmallVector<PHINode*, 8> LoopPhis; 1518 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) { 1519 LoopPhis.push_back(cast<PHINode>(I)); 1520 } 1521 // Each round of simplification iterates through the SimplifyIVUsers worklist 1522 // for all current phis, then determines whether any IVs can be 1523 // widened. Widening adds new phis to LoopPhis, inducing another round of 1524 // simplification on the wide IVs. 1525 while (!LoopPhis.empty()) { 1526 // Evaluate as many IV expressions as possible before widening any IVs. This 1527 // forces SCEV to set no-wrap flags before evaluating sign/zero 1528 // extension. The first time SCEV attempts to normalize sign/zero extension, 1529 // the result becomes final. So for the most predictable results, we delay 1530 // evaluation of sign/zero extend evaluation until needed, and avoid running 1531 // other SCEV based analysis prior to simplifyAndExtend. 1532 do { 1533 PHINode *CurrIV = LoopPhis.pop_back_val(); 1534 1535 // Information about sign/zero extensions of CurrIV. 1536 IndVarSimplifyVisitor Visitor(CurrIV, SE, TTI, DT); 1537 1538 Changed |= simplifyUsersOfIV(CurrIV, SE, DT, LI, DeadInsts, &Visitor); 1539 1540 if (Visitor.WI.WidestNativeType) { 1541 WideIVs.push_back(Visitor.WI); 1542 } 1543 } while(!LoopPhis.empty()); 1544 1545 for (; !WideIVs.empty(); WideIVs.pop_back()) { 1546 WidenIV Widener(WideIVs.back(), LI, SE, DT, DeadInsts); 1547 if (PHINode *WidePhi = Widener.createWideIV(Rewriter)) { 1548 Changed = true; 1549 LoopPhis.push_back(WidePhi); 1550 } 1551 } 1552 } 1553} 1554 1555//===----------------------------------------------------------------------===// 1556// linearFunctionTestReplace and its kin. Rewrite the loop exit condition. 1557//===----------------------------------------------------------------------===// 1558 1559/// Return true if this loop's backedge taken count expression can be safely and 1560/// cheaply expanded into an instruction sequence that can be used by 1561/// linearFunctionTestReplace. 1562/// 1563/// TODO: This fails for pointer-type loop counters with greater than one byte 1564/// strides, consequently preventing LFTR from running. For the purpose of LFTR 1565/// we could skip this check in the case that the LFTR loop counter (chosen by 1566/// FindLoopCounter) is also pointer type. Instead, we could directly convert 1567/// the loop test to an inequality test by checking the target data's alignment 1568/// of element types (given that the initial pointer value originates from or is 1569/// used by ABI constrained operation, as opposed to inttoptr/ptrtoint). 1570/// However, we don't yet have a strong motivation for converting loop tests 1571/// into inequality tests. 1572static bool canExpandBackedgeTakenCount(Loop *L, ScalarEvolution *SE, 1573 SCEVExpander &Rewriter) { 1574 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L); 1575 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) || 1576 BackedgeTakenCount->isZero()) 1577 return false; 1578 1579 if (!L->getExitingBlock()) 1580 return false; 1581 1582 // Can't rewrite non-branch yet. 1583 if (!isa<BranchInst>(L->getExitingBlock()->getTerminator())) 1584 return false; 1585 1586 if (Rewriter.isHighCostExpansion(BackedgeTakenCount, L)) 1587 return false; 1588 1589 return true; 1590} 1591 1592/// Return the loop header phi IFF IncV adds a loop invariant value to the phi. 1593static PHINode *getLoopPhiForCounter(Value *IncV, Loop *L, DominatorTree *DT) { 1594 Instruction *IncI = dyn_cast<Instruction>(IncV); 1595 if (!IncI) 1596 return nullptr; 1597 1598 switch (IncI->getOpcode()) { 1599 case Instruction::Add: 1600 case Instruction::Sub: 1601 break; 1602 case Instruction::GetElementPtr: 1603 // An IV counter must preserve its type. 1604 if (IncI->getNumOperands() == 2) 1605 break; 1606 default: 1607 return nullptr; 1608 } 1609 1610 PHINode *Phi = dyn_cast<PHINode>(IncI->getOperand(0)); 1611 if (Phi && Phi->getParent() == L->getHeader()) { 1612 if (isLoopInvariant(IncI->getOperand(1), L, DT)) 1613 return Phi; 1614 return nullptr; 1615 } 1616 if (IncI->getOpcode() == Instruction::GetElementPtr) 1617 return nullptr; 1618 1619 // Allow add/sub to be commuted. 1620 Phi = dyn_cast<PHINode>(IncI->getOperand(1)); 1621 if (Phi && Phi->getParent() == L->getHeader()) { 1622 if (isLoopInvariant(IncI->getOperand(0), L, DT)) 1623 return Phi; 1624 } 1625 return nullptr; 1626} 1627 1628/// Return the compare guarding the loop latch, or NULL for unrecognized tests. 1629static ICmpInst *getLoopTest(Loop *L) { 1630 assert(L->getExitingBlock() && "expected loop exit"); 1631 1632 BasicBlock *LatchBlock = L->getLoopLatch(); 1633 // Don't bother with LFTR if the loop is not properly simplified. 1634 if (!LatchBlock) 1635 return nullptr; 1636 1637 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1638 assert(BI && "expected exit branch"); 1639 1640 return dyn_cast<ICmpInst>(BI->getCondition()); 1641} 1642 1643/// linearFunctionTestReplace policy. Return true unless we can show that the 1644/// current exit test is already sufficiently canonical. 1645static bool needsLFTR(Loop *L, DominatorTree *DT) { 1646 // Do LFTR to simplify the exit condition to an ICMP. 1647 ICmpInst *Cond = getLoopTest(L); 1648 if (!Cond) 1649 return true; 1650 1651 // Do LFTR to simplify the exit ICMP to EQ/NE 1652 ICmpInst::Predicate Pred = Cond->getPredicate(); 1653 if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ) 1654 return true; 1655 1656 // Look for a loop invariant RHS 1657 Value *LHS = Cond->getOperand(0); 1658 Value *RHS = Cond->getOperand(1); 1659 if (!isLoopInvariant(RHS, L, DT)) { 1660 if (!isLoopInvariant(LHS, L, DT)) 1661 return true; 1662 std::swap(LHS, RHS); 1663 } 1664 // Look for a simple IV counter LHS 1665 PHINode *Phi = dyn_cast<PHINode>(LHS); 1666 if (!Phi) 1667 Phi = getLoopPhiForCounter(LHS, L, DT); 1668 1669 if (!Phi) 1670 return true; 1671 1672 // Do LFTR if PHI node is defined in the loop, but is *not* a counter. 1673 int Idx = Phi->getBasicBlockIndex(L->getLoopLatch()); 1674 if (Idx < 0) 1675 return true; 1676 1677 // Do LFTR if the exit condition's IV is *not* a simple counter. 1678 Value *IncV = Phi->getIncomingValue(Idx); 1679 return Phi != getLoopPhiForCounter(IncV, L, DT); 1680} 1681 1682/// Recursive helper for hasConcreteDef(). Unfortunately, this currently boils 1683/// down to checking that all operands are constant and listing instructions 1684/// that may hide undef. 1685static bool hasConcreteDefImpl(Value *V, SmallPtrSetImpl<Value*> &Visited, 1686 unsigned Depth) { 1687 if (isa<Constant>(V)) 1688 return !isa<UndefValue>(V); 1689 1690 if (Depth >= 6) 1691 return false; 1692 1693 // Conservatively handle non-constant non-instructions. For example, Arguments 1694 // may be undef. 1695 Instruction *I = dyn_cast<Instruction>(V); 1696 if (!I) 1697 return false; 1698 1699 // Load and return values may be undef. 1700 if(I->mayReadFromMemory() || isa<CallInst>(I) || isa<InvokeInst>(I)) 1701 return false; 1702 1703 // Optimistically handle other instructions. 1704 for (Value *Op : I->operands()) { 1705 if (!Visited.insert(Op).second) 1706 continue; 1707 if (!hasConcreteDefImpl(Op, Visited, Depth+1)) 1708 return false; 1709 } 1710 return true; 1711} 1712 1713/// Return true if the given value is concrete. We must prove that undef can 1714/// never reach it. 1715/// 1716/// TODO: If we decide that this is a good approach to checking for undef, we 1717/// may factor it into a common location. 1718static bool hasConcreteDef(Value *V) { 1719 SmallPtrSet<Value*, 8> Visited; 1720 Visited.insert(V); 1721 return hasConcreteDefImpl(V, Visited, 0); 1722} 1723 1724/// Return true if this IV has any uses other than the (soon to be rewritten) 1725/// loop exit test. 1726static bool AlmostDeadIV(PHINode *Phi, BasicBlock *LatchBlock, Value *Cond) { 1727 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock); 1728 Value *IncV = Phi->getIncomingValue(LatchIdx); 1729 1730 for (User *U : Phi->users()) 1731 if (U != Cond && U != IncV) return false; 1732 1733 for (User *U : IncV->users()) 1734 if (U != Cond && U != Phi) return false; 1735 return true; 1736} 1737 1738/// Find an affine IV in canonical form. 1739/// 1740/// BECount may be an i8* pointer type. The pointer difference is already 1741/// valid count without scaling the address stride, so it remains a pointer 1742/// expression as far as SCEV is concerned. 1743/// 1744/// Currently only valid for LFTR. See the comments on hasConcreteDef below. 1745/// 1746/// FIXME: Accept -1 stride and set IVLimit = IVInit - BECount 1747/// 1748/// FIXME: Accept non-unit stride as long as SCEV can reduce BECount * Stride. 1749/// This is difficult in general for SCEV because of potential overflow. But we 1750/// could at least handle constant BECounts. 1751static PHINode *FindLoopCounter(Loop *L, const SCEV *BECount, 1752 ScalarEvolution *SE, DominatorTree *DT) { 1753 uint64_t BCWidth = SE->getTypeSizeInBits(BECount->getType()); 1754 1755 Value *Cond = 1756 cast<BranchInst>(L->getExitingBlock()->getTerminator())->getCondition(); 1757 1758 // Loop over all of the PHI nodes, looking for a simple counter. 1759 PHINode *BestPhi = nullptr; 1760 const SCEV *BestInit = nullptr; 1761 BasicBlock *LatchBlock = L->getLoopLatch(); 1762 assert(LatchBlock && "needsLFTR should guarantee a loop latch"); 1763 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout(); 1764 1765 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) { 1766 PHINode *Phi = cast<PHINode>(I); 1767 if (!SE->isSCEVable(Phi->getType())) 1768 continue; 1769 1770 // Avoid comparing an integer IV against a pointer Limit. 1771 if (BECount->getType()->isPointerTy() && !Phi->getType()->isPointerTy()) 1772 continue; 1773 1774 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Phi)); 1775 if (!AR || AR->getLoop() != L || !AR->isAffine()) 1776 continue; 1777 1778 // AR may be a pointer type, while BECount is an integer type. 1779 // AR may be wider than BECount. With eq/ne tests overflow is immaterial. 1780 // AR may not be a narrower type, or we may never exit. 1781 uint64_t PhiWidth = SE->getTypeSizeInBits(AR->getType()); 1782 if (PhiWidth < BCWidth || !DL.isLegalInteger(PhiWidth)) 1783 continue; 1784 1785 const SCEV *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE)); 1786 if (!Step || !Step->isOne()) 1787 continue; 1788 1789 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock); 1790 Value *IncV = Phi->getIncomingValue(LatchIdx); 1791 if (getLoopPhiForCounter(IncV, L, DT) != Phi) 1792 continue; 1793 1794 // Avoid reusing a potentially undef value to compute other values that may 1795 // have originally had a concrete definition. 1796 if (!hasConcreteDef(Phi)) { 1797 // We explicitly allow unknown phis as long as they are already used by 1798 // the loop test. In this case we assume that performing LFTR could not 1799 // increase the number of undef users. 1800 if (ICmpInst *Cond = getLoopTest(L)) { 1801 if (Phi != getLoopPhiForCounter(Cond->getOperand(0), L, DT) && 1802 Phi != getLoopPhiForCounter(Cond->getOperand(1), L, DT)) { 1803 continue; 1804 } 1805 } 1806 } 1807 const SCEV *Init = AR->getStart(); 1808 1809 if (BestPhi && !AlmostDeadIV(BestPhi, LatchBlock, Cond)) { 1810 // Don't force a live loop counter if another IV can be used. 1811 if (AlmostDeadIV(Phi, LatchBlock, Cond)) 1812 continue; 1813 1814 // Prefer to count-from-zero. This is a more "canonical" counter form. It 1815 // also prefers integer to pointer IVs. 1816 if (BestInit->isZero() != Init->isZero()) { 1817 if (BestInit->isZero()) 1818 continue; 1819 } 1820 // If two IVs both count from zero or both count from nonzero then the 1821 // narrower is likely a dead phi that has been widened. Use the wider phi 1822 // to allow the other to be eliminated. 1823 else if (PhiWidth <= SE->getTypeSizeInBits(BestPhi->getType())) 1824 continue; 1825 } 1826 BestPhi = Phi; 1827 BestInit = Init; 1828 } 1829 return BestPhi; 1830} 1831 1832/// Help linearFunctionTestReplace by generating a value that holds the RHS of 1833/// the new loop test. 1834static Value *genLoopLimit(PHINode *IndVar, const SCEV *IVCount, Loop *L, 1835 SCEVExpander &Rewriter, ScalarEvolution *SE) { 1836 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IndVar)); 1837 assert(AR && AR->getLoop() == L && AR->isAffine() && "bad loop counter"); 1838 const SCEV *IVInit = AR->getStart(); 1839 1840 // IVInit may be a pointer while IVCount is an integer when FindLoopCounter 1841 // finds a valid pointer IV. Sign extend BECount in order to materialize a 1842 // GEP. Avoid running SCEVExpander on a new pointer value, instead reusing 1843 // the existing GEPs whenever possible. 1844 if (IndVar->getType()->isPointerTy() && !IVCount->getType()->isPointerTy()) { 1845 // IVOffset will be the new GEP offset that is interpreted by GEP as a 1846 // signed value. IVCount on the other hand represents the loop trip count, 1847 // which is an unsigned value. FindLoopCounter only allows induction 1848 // variables that have a positive unit stride of one. This means we don't 1849 // have to handle the case of negative offsets (yet) and just need to zero 1850 // extend IVCount. 1851 Type *OfsTy = SE->getEffectiveSCEVType(IVInit->getType()); 1852 const SCEV *IVOffset = SE->getTruncateOrZeroExtend(IVCount, OfsTy); 1853 1854 // Expand the code for the iteration count. 1855 assert(SE->isLoopInvariant(IVOffset, L) && 1856 "Computed iteration count is not loop invariant!"); 1857 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1858 Value *GEPOffset = Rewriter.expandCodeFor(IVOffset, OfsTy, BI); 1859 1860 Value *GEPBase = IndVar->getIncomingValueForBlock(L->getLoopPreheader()); 1861 assert(AR->getStart() == SE->getSCEV(GEPBase) && "bad loop counter"); 1862 // We could handle pointer IVs other than i8*, but we need to compensate for 1863 // gep index scaling. See canExpandBackedgeTakenCount comments. 1864 assert(SE->getSizeOfExpr(IntegerType::getInt64Ty(IndVar->getContext()), 1865 cast<PointerType>(GEPBase->getType()) 1866 ->getElementType())->isOne() && 1867 "unit stride pointer IV must be i8*"); 1868 1869 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator()); 1870 return Builder.CreateGEP(nullptr, GEPBase, GEPOffset, "lftr.limit"); 1871 } else { 1872 // In any other case, convert both IVInit and IVCount to integers before 1873 // comparing. This may result in SCEV expension of pointers, but in practice 1874 // SCEV will fold the pointer arithmetic away as such: 1875 // BECount = (IVEnd - IVInit - 1) => IVLimit = IVInit (postinc). 1876 // 1877 // Valid Cases: (1) both integers is most common; (2) both may be pointers 1878 // for simple memset-style loops. 1879 // 1880 // IVInit integer and IVCount pointer would only occur if a canonical IV 1881 // were generated on top of case #2, which is not expected. 1882 1883 const SCEV *IVLimit = nullptr; 1884 // For unit stride, IVCount = Start + BECount with 2's complement overflow. 1885 // For non-zero Start, compute IVCount here. 1886 if (AR->getStart()->isZero()) 1887 IVLimit = IVCount; 1888 else { 1889 assert(AR->getStepRecurrence(*SE)->isOne() && "only handles unit stride"); 1890 const SCEV *IVInit = AR->getStart(); 1891 1892 // For integer IVs, truncate the IV before computing IVInit + BECount. 1893 if (SE->getTypeSizeInBits(IVInit->getType()) 1894 > SE->getTypeSizeInBits(IVCount->getType())) 1895 IVInit = SE->getTruncateExpr(IVInit, IVCount->getType()); 1896 1897 IVLimit = SE->getAddExpr(IVInit, IVCount); 1898 } 1899 // Expand the code for the iteration count. 1900 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1901 IRBuilder<> Builder(BI); 1902 assert(SE->isLoopInvariant(IVLimit, L) && 1903 "Computed iteration count is not loop invariant!"); 1904 // Ensure that we generate the same type as IndVar, or a smaller integer 1905 // type. In the presence of null pointer values, we have an integer type 1906 // SCEV expression (IVInit) for a pointer type IV value (IndVar). 1907 Type *LimitTy = IVCount->getType()->isPointerTy() ? 1908 IndVar->getType() : IVCount->getType(); 1909 return Rewriter.expandCodeFor(IVLimit, LimitTy, BI); 1910 } 1911} 1912 1913/// This method rewrites the exit condition of the loop to be a canonical != 1914/// comparison against the incremented loop induction variable. This pass is 1915/// able to rewrite the exit tests of any loop where the SCEV analysis can 1916/// determine a loop-invariant trip count of the loop, which is actually a much 1917/// broader range than just linear tests. 1918Value *IndVarSimplify:: 1919linearFunctionTestReplace(Loop *L, 1920 const SCEV *BackedgeTakenCount, 1921 PHINode *IndVar, 1922 SCEVExpander &Rewriter) { 1923 assert(canExpandBackedgeTakenCount(L, SE, Rewriter) && "precondition"); 1924 1925 // Initialize CmpIndVar and IVCount to their preincremented values. 1926 Value *CmpIndVar = IndVar; 1927 const SCEV *IVCount = BackedgeTakenCount; 1928 1929 // If the exiting block is the same as the backedge block, we prefer to 1930 // compare against the post-incremented value, otherwise we must compare 1931 // against the preincremented value. 1932 if (L->getExitingBlock() == L->getLoopLatch()) { 1933 // Add one to the "backedge-taken" count to get the trip count. 1934 // This addition may overflow, which is valid as long as the comparison is 1935 // truncated to BackedgeTakenCount->getType(). 1936 IVCount = SE->getAddExpr(BackedgeTakenCount, 1937 SE->getOne(BackedgeTakenCount->getType())); 1938 // The BackedgeTaken expression contains the number of times that the 1939 // backedge branches to the loop header. This is one less than the 1940 // number of times the loop executes, so use the incremented indvar. 1941 CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock()); 1942 } 1943 1944 Value *ExitCnt = genLoopLimit(IndVar, IVCount, L, Rewriter, SE); 1945 assert(ExitCnt->getType()->isPointerTy() == 1946 IndVar->getType()->isPointerTy() && 1947 "genLoopLimit missed a cast"); 1948 1949 // Insert a new icmp_ne or icmp_eq instruction before the branch. 1950 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1951 ICmpInst::Predicate P; 1952 if (L->contains(BI->getSuccessor(0))) 1953 P = ICmpInst::ICMP_NE; 1954 else 1955 P = ICmpInst::ICMP_EQ; 1956 1957 DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n" 1958 << " LHS:" << *CmpIndVar << '\n' 1959 << " op:\t" 1960 << (P == ICmpInst::ICMP_NE ? "!=" : "==") << "\n" 1961 << " RHS:\t" << *ExitCnt << "\n" 1962 << " IVCount:\t" << *IVCount << "\n"); 1963 1964 IRBuilder<> Builder(BI); 1965 1966 // LFTR can ignore IV overflow and truncate to the width of 1967 // BECount. This avoids materializing the add(zext(add)) expression. 1968 unsigned CmpIndVarSize = SE->getTypeSizeInBits(CmpIndVar->getType()); 1969 unsigned ExitCntSize = SE->getTypeSizeInBits(ExitCnt->getType()); 1970 if (CmpIndVarSize > ExitCntSize) { 1971 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(SE->getSCEV(IndVar)); 1972 const SCEV *ARStart = AR->getStart(); 1973 const SCEV *ARStep = AR->getStepRecurrence(*SE); 1974 // For constant IVCount, avoid truncation. 1975 if (isa<SCEVConstant>(ARStart) && isa<SCEVConstant>(IVCount)) { 1976 const APInt &Start = cast<SCEVConstant>(ARStart)->getAPInt(); 1977 APInt Count = cast<SCEVConstant>(IVCount)->getAPInt(); 1978 // Note that the post-inc value of BackedgeTakenCount may have overflowed 1979 // above such that IVCount is now zero. 1980 if (IVCount != BackedgeTakenCount && Count == 0) { 1981 Count = APInt::getMaxValue(Count.getBitWidth()).zext(CmpIndVarSize); 1982 ++Count; 1983 } 1984 else 1985 Count = Count.zext(CmpIndVarSize); 1986 APInt NewLimit; 1987 if (cast<SCEVConstant>(ARStep)->getValue()->isNegative()) 1988 NewLimit = Start - Count; 1989 else 1990 NewLimit = Start + Count; 1991 ExitCnt = ConstantInt::get(CmpIndVar->getType(), NewLimit); 1992 1993 DEBUG(dbgs() << " Widen RHS:\t" << *ExitCnt << "\n"); 1994 } else { 1995 CmpIndVar = Builder.CreateTrunc(CmpIndVar, ExitCnt->getType(), 1996 "lftr.wideiv"); 1997 } 1998 } 1999 Value *Cond = Builder.CreateICmp(P, CmpIndVar, ExitCnt, "exitcond"); 2000 Value *OrigCond = BI->getCondition(); 2001 // It's tempting to use replaceAllUsesWith here to fully replace the old 2002 // comparison, but that's not immediately safe, since users of the old 2003 // comparison may not be dominated by the new comparison. Instead, just 2004 // update the branch to use the new comparison; in the common case this 2005 // will make old comparison dead. 2006 BI->setCondition(Cond); 2007 DeadInsts.push_back(OrigCond); 2008 2009 ++NumLFTR; 2010 Changed = true; 2011 return Cond; 2012} 2013 2014//===----------------------------------------------------------------------===// 2015// sinkUnusedInvariants. A late subpass to cleanup loop preheaders. 2016//===----------------------------------------------------------------------===// 2017 2018/// If there's a single exit block, sink any loop-invariant values that 2019/// were defined in the preheader but not used inside the loop into the 2020/// exit block to reduce register pressure in the loop. 2021void IndVarSimplify::sinkUnusedInvariants(Loop *L) { 2022 BasicBlock *ExitBlock = L->getExitBlock(); 2023 if (!ExitBlock) return; 2024 2025 BasicBlock *Preheader = L->getLoopPreheader(); 2026 if (!Preheader) return; 2027 2028 Instruction *InsertPt = &*ExitBlock->getFirstInsertionPt(); 2029 BasicBlock::iterator I(Preheader->getTerminator()); 2030 while (I != Preheader->begin()) { 2031 --I; 2032 // New instructions were inserted at the end of the preheader. 2033 if (isa<PHINode>(I)) 2034 break; 2035 2036 // Don't move instructions which might have side effects, since the side 2037 // effects need to complete before instructions inside the loop. Also don't 2038 // move instructions which might read memory, since the loop may modify 2039 // memory. Note that it's okay if the instruction might have undefined 2040 // behavior: LoopSimplify guarantees that the preheader dominates the exit 2041 // block. 2042 if (I->mayHaveSideEffects() || I->mayReadFromMemory()) 2043 continue; 2044 2045 // Skip debug info intrinsics. 2046 if (isa<DbgInfoIntrinsic>(I)) 2047 continue; 2048 2049 // Skip eh pad instructions. 2050 if (I->isEHPad()) 2051 continue; 2052 2053 // Don't sink alloca: we never want to sink static alloca's out of the 2054 // entry block, and correctly sinking dynamic alloca's requires 2055 // checks for stacksave/stackrestore intrinsics. 2056 // FIXME: Refactor this check somehow? 2057 if (isa<AllocaInst>(I)) 2058 continue; 2059 2060 // Determine if there is a use in or before the loop (direct or 2061 // otherwise). 2062 bool UsedInLoop = false; 2063 for (Use &U : I->uses()) { 2064 Instruction *User = cast<Instruction>(U.getUser()); 2065 BasicBlock *UseBB = User->getParent(); 2066 if (PHINode *P = dyn_cast<PHINode>(User)) { 2067 unsigned i = 2068 PHINode::getIncomingValueNumForOperand(U.getOperandNo()); 2069 UseBB = P->getIncomingBlock(i); 2070 } 2071 if (UseBB == Preheader || L->contains(UseBB)) { 2072 UsedInLoop = true; 2073 break; 2074 } 2075 } 2076 2077 // If there is, the def must remain in the preheader. 2078 if (UsedInLoop) 2079 continue; 2080 2081 // Otherwise, sink it to the exit block. 2082 Instruction *ToMove = &*I; 2083 bool Done = false; 2084 2085 if (I != Preheader->begin()) { 2086 // Skip debug info intrinsics. 2087 do { 2088 --I; 2089 } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin()); 2090 2091 if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin()) 2092 Done = true; 2093 } else { 2094 Done = true; 2095 } 2096 2097 ToMove->moveBefore(InsertPt); 2098 if (Done) break; 2099 InsertPt = ToMove; 2100 } 2101} 2102 2103//===----------------------------------------------------------------------===// 2104// IndVarSimplify driver. Manage several subpasses of IV simplification. 2105//===----------------------------------------------------------------------===// 2106 2107bool IndVarSimplify::run(Loop *L) { 2108 // We need (and expect!) the incoming loop to be in LCSSA. 2109 assert(L->isRecursivelyLCSSAForm(*DT) && "LCSSA required to run indvars!"); 2110 2111 // If LoopSimplify form is not available, stay out of trouble. Some notes: 2112 // - LSR currently only supports LoopSimplify-form loops. Indvars' 2113 // canonicalization can be a pessimization without LSR to "clean up" 2114 // afterwards. 2115 // - We depend on having a preheader; in particular, 2116 // Loop::getCanonicalInductionVariable only supports loops with preheaders, 2117 // and we're in trouble if we can't find the induction variable even when 2118 // we've manually inserted one. 2119 if (!L->isLoopSimplifyForm()) 2120 return false; 2121 2122 // If there are any floating-point recurrences, attempt to 2123 // transform them to use integer recurrences. 2124 rewriteNonIntegerIVs(L); 2125 2126 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L); 2127 2128 // Create a rewriter object which we'll use to transform the code with. 2129 SCEVExpander Rewriter(*SE, DL, "indvars"); 2130#ifndef NDEBUG 2131 Rewriter.setDebugType(DEBUG_TYPE); 2132#endif 2133 2134 // Eliminate redundant IV users. 2135 // 2136 // Simplification works best when run before other consumers of SCEV. We 2137 // attempt to avoid evaluating SCEVs for sign/zero extend operations until 2138 // other expressions involving loop IVs have been evaluated. This helps SCEV 2139 // set no-wrap flags before normalizing sign/zero extension. 2140 Rewriter.disableCanonicalMode(); 2141 simplifyAndExtend(L, Rewriter, LI); 2142 2143 // Check to see if this loop has a computable loop-invariant execution count. 2144 // If so, this means that we can compute the final value of any expressions 2145 // that are recurrent in the loop, and substitute the exit values from the 2146 // loop into any instructions outside of the loop that use the final values of 2147 // the current expressions. 2148 // 2149 if (ReplaceExitValue != NeverRepl && 2150 !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) 2151 rewriteLoopExitValues(L, Rewriter); 2152 2153 // Eliminate redundant IV cycles. 2154 NumElimIV += Rewriter.replaceCongruentIVs(L, DT, DeadInsts); 2155 2156 // If we have a trip count expression, rewrite the loop's exit condition 2157 // using it. We can currently only handle loops with a single exit. 2158 if (canExpandBackedgeTakenCount(L, SE, Rewriter) && needsLFTR(L, DT)) { 2159 PHINode *IndVar = FindLoopCounter(L, BackedgeTakenCount, SE, DT); 2160 if (IndVar) { 2161 // Check preconditions for proper SCEVExpander operation. SCEV does not 2162 // express SCEVExpander's dependencies, such as LoopSimplify. Instead any 2163 // pass that uses the SCEVExpander must do it. This does not work well for 2164 // loop passes because SCEVExpander makes assumptions about all loops, 2165 // while LoopPassManager only forces the current loop to be simplified. 2166 // 2167 // FIXME: SCEV expansion has no way to bail out, so the caller must 2168 // explicitly check any assumptions made by SCEV. Brittle. 2169 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(BackedgeTakenCount); 2170 if (!AR || AR->getLoop()->getLoopPreheader()) 2171 (void)linearFunctionTestReplace(L, BackedgeTakenCount, IndVar, 2172 Rewriter); 2173 } 2174 } 2175 // Clear the rewriter cache, because values that are in the rewriter's cache 2176 // can be deleted in the loop below, causing the AssertingVH in the cache to 2177 // trigger. 2178 Rewriter.clear(); 2179 2180 // Now that we're done iterating through lists, clean up any instructions 2181 // which are now dead. 2182 while (!DeadInsts.empty()) 2183 if (Instruction *Inst = 2184 dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val())) 2185 RecursivelyDeleteTriviallyDeadInstructions(Inst, TLI); 2186 2187 // The Rewriter may not be used from this point on. 2188 2189 // Loop-invariant instructions in the preheader that aren't used in the 2190 // loop may be sunk below the loop to reduce register pressure. 2191 sinkUnusedInvariants(L); 2192 2193 // rewriteFirstIterationLoopExitValues does not rely on the computation of 2194 // trip count and therefore can further simplify exit values in addition to 2195 // rewriteLoopExitValues. 2196 rewriteFirstIterationLoopExitValues(L); 2197 2198 // Clean up dead instructions. 2199 Changed |= DeleteDeadPHIs(L->getHeader(), TLI); 2200 2201 // Check a post-condition. 2202 assert(L->isRecursivelyLCSSAForm(*DT) && "Indvars did not preserve LCSSA!"); 2203 2204 // Verify that LFTR, and any other change have not interfered with SCEV's 2205 // ability to compute trip count. 2206#ifndef NDEBUG 2207 if (VerifyIndvars && !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) { 2208 SE->forgetLoop(L); 2209 const SCEV *NewBECount = SE->getBackedgeTakenCount(L); 2210 if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) < 2211 SE->getTypeSizeInBits(NewBECount->getType())) 2212 NewBECount = SE->getTruncateOrNoop(NewBECount, 2213 BackedgeTakenCount->getType()); 2214 else 2215 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount, 2216 NewBECount->getType()); 2217 assert(BackedgeTakenCount == NewBECount && "indvars must preserve SCEV"); 2218 } 2219#endif 2220 2221 return Changed; 2222} 2223 2224PreservedAnalyses IndVarSimplifyPass::run(Loop &L, AnalysisManager<Loop> &AM) { 2225 auto &FAM = AM.getResult<FunctionAnalysisManagerLoopProxy>(L).getManager(); 2226 Function *F = L.getHeader()->getParent(); 2227 const DataLayout &DL = F->getParent()->getDataLayout(); 2228 2229 auto *LI = FAM.getCachedResult<LoopAnalysis>(*F); 2230 auto *SE = FAM.getCachedResult<ScalarEvolutionAnalysis>(*F); 2231 auto *DT = FAM.getCachedResult<DominatorTreeAnalysis>(*F); 2232 2233 assert((LI && SE && DT) && 2234 "Analyses required for indvarsimplify not available!"); 2235 2236 // Optional analyses. 2237 auto *TTI = FAM.getCachedResult<TargetIRAnalysis>(*F); 2238 auto *TLI = FAM.getCachedResult<TargetLibraryAnalysis>(*F); 2239 2240 IndVarSimplify IVS(LI, SE, DT, DL, TLI, TTI); 2241 if (!IVS.run(&L)) 2242 return PreservedAnalyses::all(); 2243 2244 // FIXME: This should also 'preserve the CFG'. 2245 return getLoopPassPreservedAnalyses(); 2246} 2247 2248namespace { 2249struct IndVarSimplifyLegacyPass : public LoopPass { 2250 static char ID; // Pass identification, replacement for typeid 2251 IndVarSimplifyLegacyPass() : LoopPass(ID) { 2252 initializeIndVarSimplifyLegacyPassPass(*PassRegistry::getPassRegistry()); 2253 } 2254 2255 bool runOnLoop(Loop *L, LPPassManager &LPM) override { 2256 if (skipLoop(L)) 2257 return false; 2258 2259 auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 2260 auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); 2261 auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 2262 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>(); 2263 auto *TLI = TLIP ? &TLIP->getTLI() : nullptr; 2264 auto *TTIP = getAnalysisIfAvailable<TargetTransformInfoWrapperPass>(); 2265 auto *TTI = TTIP ? &TTIP->getTTI(*L->getHeader()->getParent()) : nullptr; 2266 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout(); 2267 2268 IndVarSimplify IVS(LI, SE, DT, DL, TLI, TTI); 2269 return IVS.run(L); 2270 } 2271 2272 void getAnalysisUsage(AnalysisUsage &AU) const override { 2273 AU.setPreservesCFG(); 2274 getLoopAnalysisUsage(AU); 2275 } 2276}; 2277} 2278 2279char IndVarSimplifyLegacyPass::ID = 0; 2280INITIALIZE_PASS_BEGIN(IndVarSimplifyLegacyPass, "indvars", 2281 "Induction Variable Simplification", false, false) 2282INITIALIZE_PASS_DEPENDENCY(LoopPass) 2283INITIALIZE_PASS_END(IndVarSimplifyLegacyPass, "indvars", 2284 "Induction Variable Simplification", false, false) 2285 2286Pass *llvm::createIndVarSimplifyPass() { 2287 return new IndVarSimplifyLegacyPass(); 2288} 2289