IndVarSimplify.cpp revision 36b56886974eae4f9c5ebc96befd3e7bfe5de338
1//===- IndVarSimplify.cpp - Induction Variable Elimination ----------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This transformation analyzes and transforms the induction variables (and 11// computations derived from them) into simpler forms suitable for subsequent 12// analysis and transformation. 13// 14// If the trip count of a loop is computable, this pass also makes the following 15// changes: 16// 1. The exit condition for the loop is canonicalized to compare the 17// induction value against the exit value. This turns loops like: 18// 'for (i = 7; i*i < 1000; ++i)' into 'for (i = 0; i != 25; ++i)' 19// 2. Any use outside of the loop of an expression derived from the indvar 20// is changed to compute the derived value outside of the loop, eliminating 21// the dependence on the exit value of the induction variable. If the only 22// purpose of the loop is to compute the exit value of some derived 23// expression, this transformation will make the loop dead. 24// 25//===----------------------------------------------------------------------===// 26 27#define DEBUG_TYPE "indvars" 28#include "llvm/Transforms/Scalar.h" 29#include "llvm/ADT/DenseMap.h" 30#include "llvm/ADT/SmallVector.h" 31#include "llvm/ADT/Statistic.h" 32#include "llvm/Analysis/LoopInfo.h" 33#include "llvm/Analysis/LoopPass.h" 34#include "llvm/Analysis/ScalarEvolutionExpander.h" 35#include "llvm/IR/BasicBlock.h" 36#include "llvm/IR/CFG.h" 37#include "llvm/IR/Constants.h" 38#include "llvm/IR/DataLayout.h" 39#include "llvm/IR/Dominators.h" 40#include "llvm/IR/Instructions.h" 41#include "llvm/IR/IntrinsicInst.h" 42#include "llvm/IR/LLVMContext.h" 43#include "llvm/IR/Type.h" 44#include "llvm/Support/CommandLine.h" 45#include "llvm/Support/Debug.h" 46#include "llvm/Support/raw_ostream.h" 47#include "llvm/Target/TargetLibraryInfo.h" 48#include "llvm/Transforms/Utils/BasicBlockUtils.h" 49#include "llvm/Transforms/Utils/Local.h" 50#include "llvm/Transforms/Utils/SimplifyIndVar.h" 51using namespace llvm; 52 53STATISTIC(NumWidened , "Number of indvars widened"); 54STATISTIC(NumReplaced , "Number of exit values replaced"); 55STATISTIC(NumLFTR , "Number of loop exit tests replaced"); 56STATISTIC(NumElimExt , "Number of IV sign/zero extends eliminated"); 57STATISTIC(NumElimIV , "Number of congruent IVs eliminated"); 58 59// Trip count verification can be enabled by default under NDEBUG if we 60// implement a strong expression equivalence checker in SCEV. Until then, we 61// use the verify-indvars flag, which may assert in some cases. 62static cl::opt<bool> VerifyIndvars( 63 "verify-indvars", cl::Hidden, 64 cl::desc("Verify the ScalarEvolution result after running indvars")); 65 66static cl::opt<bool> ReduceLiveIVs("liv-reduce", cl::Hidden, 67 cl::desc("Reduce live induction variables.")); 68 69namespace { 70 class IndVarSimplify : public LoopPass { 71 LoopInfo *LI; 72 ScalarEvolution *SE; 73 DominatorTree *DT; 74 const DataLayout *DL; 75 TargetLibraryInfo *TLI; 76 77 SmallVector<WeakVH, 16> DeadInsts; 78 bool Changed; 79 public: 80 81 static char ID; // Pass identification, replacement for typeid 82 IndVarSimplify() : LoopPass(ID), LI(0), SE(0), DT(0), DL(0), 83 Changed(false) { 84 initializeIndVarSimplifyPass(*PassRegistry::getPassRegistry()); 85 } 86 87 bool runOnLoop(Loop *L, LPPassManager &LPM) override; 88 89 void getAnalysisUsage(AnalysisUsage &AU) const override { 90 AU.addRequired<DominatorTreeWrapperPass>(); 91 AU.addRequired<LoopInfo>(); 92 AU.addRequired<ScalarEvolution>(); 93 AU.addRequiredID(LoopSimplifyID); 94 AU.addRequiredID(LCSSAID); 95 AU.addPreserved<ScalarEvolution>(); 96 AU.addPreservedID(LoopSimplifyID); 97 AU.addPreservedID(LCSSAID); 98 AU.setPreservesCFG(); 99 } 100 101 private: 102 void releaseMemory() override { 103 DeadInsts.clear(); 104 } 105 106 bool isValidRewrite(Value *FromVal, Value *ToVal); 107 108 void HandleFloatingPointIV(Loop *L, PHINode *PH); 109 void RewriteNonIntegerIVs(Loop *L); 110 111 void SimplifyAndExtend(Loop *L, SCEVExpander &Rewriter, LPPassManager &LPM); 112 113 void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter); 114 115 Value *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount, 116 PHINode *IndVar, SCEVExpander &Rewriter); 117 118 void SinkUnusedInvariants(Loop *L); 119 }; 120} 121 122char IndVarSimplify::ID = 0; 123INITIALIZE_PASS_BEGIN(IndVarSimplify, "indvars", 124 "Induction Variable Simplification", false, false) 125INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 126INITIALIZE_PASS_DEPENDENCY(LoopInfo) 127INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 128INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 129INITIALIZE_PASS_DEPENDENCY(LCSSA) 130INITIALIZE_PASS_END(IndVarSimplify, "indvars", 131 "Induction Variable Simplification", false, false) 132 133Pass *llvm::createIndVarSimplifyPass() { 134 return new IndVarSimplify(); 135} 136 137/// isValidRewrite - Return true if the SCEV expansion generated by the 138/// rewriter can replace the original value. SCEV guarantees that it 139/// produces the same value, but the way it is produced may be illegal IR. 140/// Ideally, this function will only be called for verification. 141bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) { 142 // If an SCEV expression subsumed multiple pointers, its expansion could 143 // reassociate the GEP changing the base pointer. This is illegal because the 144 // final address produced by a GEP chain must be inbounds relative to its 145 // underlying object. Otherwise basic alias analysis, among other things, 146 // could fail in a dangerous way. Ultimately, SCEV will be improved to avoid 147 // producing an expression involving multiple pointers. Until then, we must 148 // bail out here. 149 // 150 // Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject 151 // because it understands lcssa phis while SCEV does not. 152 Value *FromPtr = FromVal; 153 Value *ToPtr = ToVal; 154 if (GEPOperator *GEP = dyn_cast<GEPOperator>(FromVal)) { 155 FromPtr = GEP->getPointerOperand(); 156 } 157 if (GEPOperator *GEP = dyn_cast<GEPOperator>(ToVal)) { 158 ToPtr = GEP->getPointerOperand(); 159 } 160 if (FromPtr != FromVal || ToPtr != ToVal) { 161 // Quickly check the common case 162 if (FromPtr == ToPtr) 163 return true; 164 165 // SCEV may have rewritten an expression that produces the GEP's pointer 166 // operand. That's ok as long as the pointer operand has the same base 167 // pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the 168 // base of a recurrence. This handles the case in which SCEV expansion 169 // converts a pointer type recurrence into a nonrecurrent pointer base 170 // indexed by an integer recurrence. 171 172 // If the GEP base pointer is a vector of pointers, abort. 173 if (!FromPtr->getType()->isPointerTy() || !ToPtr->getType()->isPointerTy()) 174 return false; 175 176 const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr)); 177 const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr)); 178 if (FromBase == ToBase) 179 return true; 180 181 DEBUG(dbgs() << "INDVARS: GEP rewrite bail out " 182 << *FromBase << " != " << *ToBase << "\n"); 183 184 return false; 185 } 186 return true; 187} 188 189/// Determine the insertion point for this user. By default, insert immediately 190/// before the user. SCEVExpander or LICM will hoist loop invariants out of the 191/// loop. For PHI nodes, there may be multiple uses, so compute the nearest 192/// common dominator for the incoming blocks. 193static Instruction *getInsertPointForUses(Instruction *User, Value *Def, 194 DominatorTree *DT) { 195 PHINode *PHI = dyn_cast<PHINode>(User); 196 if (!PHI) 197 return User; 198 199 Instruction *InsertPt = 0; 200 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) { 201 if (PHI->getIncomingValue(i) != Def) 202 continue; 203 204 BasicBlock *InsertBB = PHI->getIncomingBlock(i); 205 if (!InsertPt) { 206 InsertPt = InsertBB->getTerminator(); 207 continue; 208 } 209 InsertBB = DT->findNearestCommonDominator(InsertPt->getParent(), InsertBB); 210 InsertPt = InsertBB->getTerminator(); 211 } 212 assert(InsertPt && "Missing phi operand"); 213 assert((!isa<Instruction>(Def) || 214 DT->dominates(cast<Instruction>(Def), InsertPt)) && 215 "def does not dominate all uses"); 216 return InsertPt; 217} 218 219//===----------------------------------------------------------------------===// 220// RewriteNonIntegerIVs and helpers. Prefer integer IVs. 221//===----------------------------------------------------------------------===// 222 223/// ConvertToSInt - Convert APF to an integer, if possible. 224static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) { 225 bool isExact = false; 226 // See if we can convert this to an int64_t 227 uint64_t UIntVal; 228 if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero, 229 &isExact) != APFloat::opOK || !isExact) 230 return false; 231 IntVal = UIntVal; 232 return true; 233} 234 235/// HandleFloatingPointIV - If the loop has floating induction variable 236/// then insert corresponding integer induction variable if possible. 237/// For example, 238/// for(double i = 0; i < 10000; ++i) 239/// bar(i) 240/// is converted into 241/// for(int i = 0; i < 10000; ++i) 242/// bar((double)i); 243/// 244void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) { 245 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0)); 246 unsigned BackEdge = IncomingEdge^1; 247 248 // Check incoming value. 249 ConstantFP *InitValueVal = 250 dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge)); 251 252 int64_t InitValue; 253 if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue)) 254 return; 255 256 // Check IV increment. Reject this PN if increment operation is not 257 // an add or increment value can not be represented by an integer. 258 BinaryOperator *Incr = 259 dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge)); 260 if (Incr == 0 || 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 == 0 || 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 == 0 || !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 == 0 || 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} 464 465//===----------------------------------------------------------------------===// 466// RewriteLoopExitValues - Optimize IV users outside the loop. 467// As a side effect, reduces the amount of IV processing within the loop. 468//===----------------------------------------------------------------------===// 469 470/// RewriteLoopExitValues - Check to see if this loop has a computable 471/// loop-invariant execution count. If so, this means that we can compute the 472/// final value of any expressions that are recurrent in the loop, and 473/// substitute the exit values from the loop into any instructions outside of 474/// the loop that use the final values of the current expressions. 475/// 476/// This is mostly redundant with the regular IndVarSimplify activities that 477/// happen later, except that it's more powerful in some cases, because it's 478/// able to brute-force evaluate arbitrary instructions as long as they have 479/// constant operands at the beginning of the loop. 480void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) { 481 // Verify the input to the pass in already in LCSSA form. 482 assert(L->isLCSSAForm(*DT)); 483 484 SmallVector<BasicBlock*, 8> ExitBlocks; 485 L->getUniqueExitBlocks(ExitBlocks); 486 487 // Find all values that are computed inside the loop, but used outside of it. 488 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan 489 // the exit blocks of the loop to find them. 490 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 491 BasicBlock *ExitBB = ExitBlocks[i]; 492 493 // If there are no PHI nodes in this exit block, then no values defined 494 // inside the loop are used on this path, skip it. 495 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin()); 496 if (!PN) continue; 497 498 unsigned NumPreds = PN->getNumIncomingValues(); 499 500 // We would like to be able to RAUW single-incoming value PHI nodes. We 501 // have to be certain this is safe even when this is an LCSSA PHI node. 502 // While the computed exit value is no longer varying in *this* loop, the 503 // exit block may be an exit block for an outer containing loop as well, 504 // the exit value may be varying in the outer loop, and thus it may still 505 // require an LCSSA PHI node. The safe case is when this is 506 // single-predecessor PHI node (LCSSA) and the exit block containing it is 507 // part of the enclosing loop, or this is the outer most loop of the nest. 508 // In either case the exit value could (at most) be varying in the same 509 // loop body as the phi node itself. Thus if it is in turn used outside of 510 // an enclosing loop it will only be via a separate LCSSA node. 511 bool LCSSASafePhiForRAUW = 512 NumPreds == 1 && 513 (!L->getParentLoop() || L->getParentLoop() == LI->getLoopFor(ExitBB)); 514 515 // Iterate over all of the PHI nodes. 516 BasicBlock::iterator BBI = ExitBB->begin(); 517 while ((PN = dyn_cast<PHINode>(BBI++))) { 518 if (PN->use_empty()) 519 continue; // dead use, don't replace it 520 521 // SCEV only supports integer expressions for now. 522 if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy()) 523 continue; 524 525 // It's necessary to tell ScalarEvolution about this explicitly so that 526 // it can walk the def-use list and forget all SCEVs, as it may not be 527 // watching the PHI itself. Once the new exit value is in place, there 528 // may not be a def-use connection between the loop and every instruction 529 // which got a SCEVAddRecExpr for that loop. 530 SE->forgetValue(PN); 531 532 // Iterate over all of the values in all the PHI nodes. 533 for (unsigned i = 0; i != NumPreds; ++i) { 534 // If the value being merged in is not integer or is not defined 535 // in the loop, skip it. 536 Value *InVal = PN->getIncomingValue(i); 537 if (!isa<Instruction>(InVal)) 538 continue; 539 540 // If this pred is for a subloop, not L itself, skip it. 541 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L) 542 continue; // The Block is in a subloop, skip it. 543 544 // Check that InVal is defined in the loop. 545 Instruction *Inst = cast<Instruction>(InVal); 546 if (!L->contains(Inst)) 547 continue; 548 549 // Okay, this instruction has a user outside of the current loop 550 // and varies predictably *inside* the loop. Evaluate the value it 551 // contains when the loop exits, if possible. 552 const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop()); 553 if (!SE->isLoopInvariant(ExitValue, L) || 554 !isSafeToExpand(ExitValue, *SE)) 555 continue; 556 557 // Computing the value outside of the loop brings no benefit if : 558 // - it is definitely used inside the loop in a way which can not be 559 // optimized away. 560 // - no use outside of the loop can take advantage of hoisting the 561 // computation out of the loop 562 if (ExitValue->getSCEVType()>=scMulExpr) { 563 unsigned NumHardInternalUses = 0; 564 unsigned NumSoftExternalUses = 0; 565 unsigned NumUses = 0; 566 for (auto IB = Inst->user_begin(), IE = Inst->user_end(); 567 IB != IE && NumUses <= 6; ++IB) { 568 Instruction *UseInstr = cast<Instruction>(*IB); 569 unsigned Opc = UseInstr->getOpcode(); 570 NumUses++; 571 if (L->contains(UseInstr)) { 572 if (Opc == Instruction::Call || Opc == Instruction::Ret) 573 NumHardInternalUses++; 574 } else { 575 if (Opc == Instruction::PHI) { 576 // Do not count the Phi as a use. LCSSA may have inserted 577 // plenty of trivial ones. 578 NumUses--; 579 for (auto PB = UseInstr->user_begin(), 580 PE = UseInstr->user_end(); 581 PB != PE && NumUses <= 6; ++PB, ++NumUses) { 582 unsigned PhiOpc = cast<Instruction>(*PB)->getOpcode(); 583 if (PhiOpc != Instruction::Call && PhiOpc != Instruction::Ret) 584 NumSoftExternalUses++; 585 } 586 continue; 587 } 588 if (Opc != Instruction::Call && Opc != Instruction::Ret) 589 NumSoftExternalUses++; 590 } 591 } 592 if (NumUses <= 6 && NumHardInternalUses && !NumSoftExternalUses) 593 continue; 594 } 595 596 Value *ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), Inst); 597 598 DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n' 599 << " LoopVal = " << *Inst << "\n"); 600 601 if (!isValidRewrite(Inst, ExitVal)) { 602 DeadInsts.push_back(ExitVal); 603 continue; 604 } 605 Changed = true; 606 ++NumReplaced; 607 608 PN->setIncomingValue(i, ExitVal); 609 610 // If this instruction is dead now, delete it. Don't do it now to avoid 611 // invalidating iterators. 612 if (isInstructionTriviallyDead(Inst, TLI)) 613 DeadInsts.push_back(Inst); 614 615 // If we determined that this PHI is safe to replace even if an LCSSA 616 // PHI, do so. 617 if (LCSSASafePhiForRAUW) { 618 PN->replaceAllUsesWith(ExitVal); 619 PN->eraseFromParent(); 620 } 621 } 622 623 // If we were unable to completely replace the PHI node, clone the PHI 624 // and delete the original one. This lets IVUsers and any other maps 625 // purge the original user from their records. 626 if (!LCSSASafePhiForRAUW) { 627 PHINode *NewPN = cast<PHINode>(PN->clone()); 628 NewPN->takeName(PN); 629 NewPN->insertBefore(PN); 630 PN->replaceAllUsesWith(NewPN); 631 PN->eraseFromParent(); 632 } 633 } 634 } 635 636 // The insertion point instruction may have been deleted; clear it out 637 // so that the rewriter doesn't trip over it later. 638 Rewriter.clearInsertPoint(); 639} 640 641//===----------------------------------------------------------------------===// 642// IV Widening - Extend the width of an IV to cover its widest uses. 643//===----------------------------------------------------------------------===// 644 645namespace { 646 // Collect information about induction variables that are used by sign/zero 647 // extend operations. This information is recorded by CollectExtend and 648 // provides the input to WidenIV. 649 struct WideIVInfo { 650 PHINode *NarrowIV; 651 Type *WidestNativeType; // Widest integer type created [sz]ext 652 bool IsSigned; // Was an sext user seen before a zext? 653 654 WideIVInfo() : NarrowIV(0), WidestNativeType(0), IsSigned(false) {} 655 }; 656} 657 658/// visitCast - Update information about the induction variable that is 659/// extended by this sign or zero extend operation. This is used to determine 660/// the final width of the IV before actually widening it. 661static void visitIVCast(CastInst *Cast, WideIVInfo &WI, ScalarEvolution *SE, 662 const DataLayout *DL) { 663 bool IsSigned = Cast->getOpcode() == Instruction::SExt; 664 if (!IsSigned && Cast->getOpcode() != Instruction::ZExt) 665 return; 666 667 Type *Ty = Cast->getType(); 668 uint64_t Width = SE->getTypeSizeInBits(Ty); 669 if (DL && !DL->isLegalInteger(Width)) 670 return; 671 672 if (!WI.WidestNativeType) { 673 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty); 674 WI.IsSigned = IsSigned; 675 return; 676 } 677 678 // We extend the IV to satisfy the sign of its first user, arbitrarily. 679 if (WI.IsSigned != IsSigned) 680 return; 681 682 if (Width > SE->getTypeSizeInBits(WI.WidestNativeType)) 683 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty); 684} 685 686namespace { 687 688/// NarrowIVDefUse - Record a link in the Narrow IV def-use chain along with the 689/// WideIV that computes the same value as the Narrow IV def. This avoids 690/// caching Use* pointers. 691struct NarrowIVDefUse { 692 Instruction *NarrowDef; 693 Instruction *NarrowUse; 694 Instruction *WideDef; 695 696 NarrowIVDefUse(): NarrowDef(0), NarrowUse(0), WideDef(0) {} 697 698 NarrowIVDefUse(Instruction *ND, Instruction *NU, Instruction *WD): 699 NarrowDef(ND), NarrowUse(NU), WideDef(WD) {} 700}; 701 702/// WidenIV - The goal of this transform is to remove sign and zero extends 703/// without creating any new induction variables. To do this, it creates a new 704/// phi of the wider type and redirects all users, either removing extends or 705/// inserting truncs whenever we stop propagating the type. 706/// 707class WidenIV { 708 // Parameters 709 PHINode *OrigPhi; 710 Type *WideType; 711 bool IsSigned; 712 713 // Context 714 LoopInfo *LI; 715 Loop *L; 716 ScalarEvolution *SE; 717 DominatorTree *DT; 718 719 // Result 720 PHINode *WidePhi; 721 Instruction *WideInc; 722 const SCEV *WideIncExpr; 723 SmallVectorImpl<WeakVH> &DeadInsts; 724 725 SmallPtrSet<Instruction*,16> Widened; 726 SmallVector<NarrowIVDefUse, 8> NarrowIVUsers; 727 728public: 729 WidenIV(const WideIVInfo &WI, LoopInfo *LInfo, 730 ScalarEvolution *SEv, DominatorTree *DTree, 731 SmallVectorImpl<WeakVH> &DI) : 732 OrigPhi(WI.NarrowIV), 733 WideType(WI.WidestNativeType), 734 IsSigned(WI.IsSigned), 735 LI(LInfo), 736 L(LI->getLoopFor(OrigPhi->getParent())), 737 SE(SEv), 738 DT(DTree), 739 WidePhi(0), 740 WideInc(0), 741 WideIncExpr(0), 742 DeadInsts(DI) { 743 assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV"); 744 } 745 746 PHINode *CreateWideIV(SCEVExpander &Rewriter); 747 748protected: 749 Value *getExtend(Value *NarrowOper, Type *WideType, bool IsSigned, 750 Instruction *Use); 751 752 Instruction *CloneIVUser(NarrowIVDefUse DU); 753 754 const SCEVAddRecExpr *GetWideRecurrence(Instruction *NarrowUse); 755 756 const SCEVAddRecExpr* GetExtendedOperandRecurrence(NarrowIVDefUse DU); 757 758 Instruction *WidenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter); 759 760 void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef); 761}; 762} // anonymous namespace 763 764/// isLoopInvariant - Perform a quick domtree based check for loop invariance 765/// assuming that V is used within the loop. LoopInfo::isLoopInvariant() seems 766/// gratuitous for this purpose. 767static bool isLoopInvariant(Value *V, const Loop *L, const DominatorTree *DT) { 768 Instruction *Inst = dyn_cast<Instruction>(V); 769 if (!Inst) 770 return true; 771 772 return DT->properlyDominates(Inst->getParent(), L->getHeader()); 773} 774 775Value *WidenIV::getExtend(Value *NarrowOper, Type *WideType, bool IsSigned, 776 Instruction *Use) { 777 // Set the debug location and conservative insertion point. 778 IRBuilder<> Builder(Use); 779 // Hoist the insertion point into loop preheaders as far as possible. 780 for (const Loop *L = LI->getLoopFor(Use->getParent()); 781 L && L->getLoopPreheader() && isLoopInvariant(NarrowOper, L, DT); 782 L = L->getParentLoop()) 783 Builder.SetInsertPoint(L->getLoopPreheader()->getTerminator()); 784 785 return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) : 786 Builder.CreateZExt(NarrowOper, WideType); 787} 788 789/// CloneIVUser - Instantiate a wide operation to replace a narrow 790/// operation. This only needs to handle operations that can evaluation to 791/// SCEVAddRec. It can safely return 0 for any operation we decide not to clone. 792Instruction *WidenIV::CloneIVUser(NarrowIVDefUse DU) { 793 unsigned Opcode = DU.NarrowUse->getOpcode(); 794 switch (Opcode) { 795 default: 796 return 0; 797 case Instruction::Add: 798 case Instruction::Mul: 799 case Instruction::UDiv: 800 case Instruction::Sub: 801 case Instruction::And: 802 case Instruction::Or: 803 case Instruction::Xor: 804 case Instruction::Shl: 805 case Instruction::LShr: 806 case Instruction::AShr: 807 DEBUG(dbgs() << "Cloning IVUser: " << *DU.NarrowUse << "\n"); 808 809 // Replace NarrowDef operands with WideDef. Otherwise, we don't know 810 // anything about the narrow operand yet so must insert a [sz]ext. It is 811 // probably loop invariant and will be folded or hoisted. If it actually 812 // comes from a widened IV, it should be removed during a future call to 813 // WidenIVUse. 814 Value *LHS = (DU.NarrowUse->getOperand(0) == DU.NarrowDef) ? DU.WideDef : 815 getExtend(DU.NarrowUse->getOperand(0), WideType, IsSigned, DU.NarrowUse); 816 Value *RHS = (DU.NarrowUse->getOperand(1) == DU.NarrowDef) ? DU.WideDef : 817 getExtend(DU.NarrowUse->getOperand(1), WideType, IsSigned, DU.NarrowUse); 818 819 BinaryOperator *NarrowBO = cast<BinaryOperator>(DU.NarrowUse); 820 BinaryOperator *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(), 821 LHS, RHS, 822 NarrowBO->getName()); 823 IRBuilder<> Builder(DU.NarrowUse); 824 Builder.Insert(WideBO); 825 if (const OverflowingBinaryOperator *OBO = 826 dyn_cast<OverflowingBinaryOperator>(NarrowBO)) { 827 if (OBO->hasNoUnsignedWrap()) WideBO->setHasNoUnsignedWrap(); 828 if (OBO->hasNoSignedWrap()) WideBO->setHasNoSignedWrap(); 829 } 830 return WideBO; 831 } 832} 833 834/// No-wrap operations can transfer sign extension of their result to their 835/// operands. Generate the SCEV value for the widened operation without 836/// actually modifying the IR yet. If the expression after extending the 837/// operands is an AddRec for this loop, return it. 838const SCEVAddRecExpr* WidenIV::GetExtendedOperandRecurrence(NarrowIVDefUse DU) { 839 // Handle the common case of add<nsw/nuw> 840 if (DU.NarrowUse->getOpcode() != Instruction::Add) 841 return 0; 842 843 // One operand (NarrowDef) has already been extended to WideDef. Now determine 844 // if extending the other will lead to a recurrence. 845 unsigned ExtendOperIdx = DU.NarrowUse->getOperand(0) == DU.NarrowDef ? 1 : 0; 846 assert(DU.NarrowUse->getOperand(1-ExtendOperIdx) == DU.NarrowDef && "bad DU"); 847 848 const SCEV *ExtendOperExpr = 0; 849 const OverflowingBinaryOperator *OBO = 850 cast<OverflowingBinaryOperator>(DU.NarrowUse); 851 if (IsSigned && OBO->hasNoSignedWrap()) 852 ExtendOperExpr = SE->getSignExtendExpr( 853 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType); 854 else if(!IsSigned && OBO->hasNoUnsignedWrap()) 855 ExtendOperExpr = SE->getZeroExtendExpr( 856 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType); 857 else 858 return 0; 859 860 // When creating this AddExpr, don't apply the current operations NSW or NUW 861 // flags. This instruction may be guarded by control flow that the no-wrap 862 // behavior depends on. Non-control-equivalent instructions can be mapped to 863 // the same SCEV expression, and it would be incorrect to transfer NSW/NUW 864 // semantics to those operations. 865 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>( 866 SE->getAddExpr(SE->getSCEV(DU.WideDef), ExtendOperExpr)); 867 868 if (!AddRec || AddRec->getLoop() != L) 869 return 0; 870 return AddRec; 871} 872 873/// GetWideRecurrence - Is this instruction potentially interesting from 874/// IVUsers' perspective after widening it's type? In other words, can the 875/// extend be safely hoisted out of the loop with SCEV reducing the value to a 876/// recurrence on the same loop. If so, return the sign or zero extended 877/// recurrence. Otherwise return NULL. 878const SCEVAddRecExpr *WidenIV::GetWideRecurrence(Instruction *NarrowUse) { 879 if (!SE->isSCEVable(NarrowUse->getType())) 880 return 0; 881 882 const SCEV *NarrowExpr = SE->getSCEV(NarrowUse); 883 if (SE->getTypeSizeInBits(NarrowExpr->getType()) 884 >= SE->getTypeSizeInBits(WideType)) { 885 // NarrowUse implicitly widens its operand. e.g. a gep with a narrow 886 // index. So don't follow this use. 887 return 0; 888 } 889 890 const SCEV *WideExpr = IsSigned ? 891 SE->getSignExtendExpr(NarrowExpr, WideType) : 892 SE->getZeroExtendExpr(NarrowExpr, WideType); 893 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr); 894 if (!AddRec || AddRec->getLoop() != L) 895 return 0; 896 return AddRec; 897} 898 899/// This IV user cannot be widen. Replace this use of the original narrow IV 900/// with a truncation of the new wide IV to isolate and eliminate the narrow IV. 901static void truncateIVUse(NarrowIVDefUse DU, DominatorTree *DT) { 902 DEBUG(dbgs() << "INDVARS: Truncate IV " << *DU.WideDef 903 << " for user " << *DU.NarrowUse << "\n"); 904 IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT)); 905 Value *Trunc = Builder.CreateTrunc(DU.WideDef, DU.NarrowDef->getType()); 906 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, Trunc); 907} 908 909/// WidenIVUse - Determine whether an individual user of the narrow IV can be 910/// widened. If so, return the wide clone of the user. 911Instruction *WidenIV::WidenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter) { 912 913 // Stop traversing the def-use chain at inner-loop phis or post-loop phis. 914 if (PHINode *UsePhi = dyn_cast<PHINode>(DU.NarrowUse)) { 915 if (LI->getLoopFor(UsePhi->getParent()) != L) { 916 // For LCSSA phis, sink the truncate outside the loop. 917 // After SimplifyCFG most loop exit targets have a single predecessor. 918 // Otherwise fall back to a truncate within the loop. 919 if (UsePhi->getNumOperands() != 1) 920 truncateIVUse(DU, DT); 921 else { 922 PHINode *WidePhi = 923 PHINode::Create(DU.WideDef->getType(), 1, UsePhi->getName() + ".wide", 924 UsePhi); 925 WidePhi->addIncoming(DU.WideDef, UsePhi->getIncomingBlock(0)); 926 IRBuilder<> Builder(WidePhi->getParent()->getFirstInsertionPt()); 927 Value *Trunc = Builder.CreateTrunc(WidePhi, DU.NarrowDef->getType()); 928 UsePhi->replaceAllUsesWith(Trunc); 929 DeadInsts.push_back(UsePhi); 930 DEBUG(dbgs() << "INDVARS: Widen lcssa phi " << *UsePhi 931 << " to " << *WidePhi << "\n"); 932 } 933 return 0; 934 } 935 } 936 // Our raison d'etre! Eliminate sign and zero extension. 937 if (IsSigned ? isa<SExtInst>(DU.NarrowUse) : isa<ZExtInst>(DU.NarrowUse)) { 938 Value *NewDef = DU.WideDef; 939 if (DU.NarrowUse->getType() != WideType) { 940 unsigned CastWidth = SE->getTypeSizeInBits(DU.NarrowUse->getType()); 941 unsigned IVWidth = SE->getTypeSizeInBits(WideType); 942 if (CastWidth < IVWidth) { 943 // The cast isn't as wide as the IV, so insert a Trunc. 944 IRBuilder<> Builder(DU.NarrowUse); 945 NewDef = Builder.CreateTrunc(DU.WideDef, DU.NarrowUse->getType()); 946 } 947 else { 948 // A wider extend was hidden behind a narrower one. This may induce 949 // another round of IV widening in which the intermediate IV becomes 950 // dead. It should be very rare. 951 DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi 952 << " not wide enough to subsume " << *DU.NarrowUse << "\n"); 953 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef); 954 NewDef = DU.NarrowUse; 955 } 956 } 957 if (NewDef != DU.NarrowUse) { 958 DEBUG(dbgs() << "INDVARS: eliminating " << *DU.NarrowUse 959 << " replaced by " << *DU.WideDef << "\n"); 960 ++NumElimExt; 961 DU.NarrowUse->replaceAllUsesWith(NewDef); 962 DeadInsts.push_back(DU.NarrowUse); 963 } 964 // Now that the extend is gone, we want to expose it's uses for potential 965 // further simplification. We don't need to directly inform SimplifyIVUsers 966 // of the new users, because their parent IV will be processed later as a 967 // new loop phi. If we preserved IVUsers analysis, we would also want to 968 // push the uses of WideDef here. 969 970 // No further widening is needed. The deceased [sz]ext had done it for us. 971 return 0; 972 } 973 974 // Does this user itself evaluate to a recurrence after widening? 975 const SCEVAddRecExpr *WideAddRec = GetWideRecurrence(DU.NarrowUse); 976 if (!WideAddRec) { 977 WideAddRec = GetExtendedOperandRecurrence(DU); 978 } 979 if (!WideAddRec) { 980 // This user does not evaluate to a recurence after widening, so don't 981 // follow it. Instead insert a Trunc to kill off the original use, 982 // eventually isolating the original narrow IV so it can be removed. 983 truncateIVUse(DU, DT); 984 return 0; 985 } 986 // Assume block terminators cannot evaluate to a recurrence. We can't to 987 // insert a Trunc after a terminator if there happens to be a critical edge. 988 assert(DU.NarrowUse != DU.NarrowUse->getParent()->getTerminator() && 989 "SCEV is not expected to evaluate a block terminator"); 990 991 // Reuse the IV increment that SCEVExpander created as long as it dominates 992 // NarrowUse. 993 Instruction *WideUse = 0; 994 if (WideAddRec == WideIncExpr 995 && Rewriter.hoistIVInc(WideInc, DU.NarrowUse)) 996 WideUse = WideInc; 997 else { 998 WideUse = CloneIVUser(DU); 999 if (!WideUse) 1000 return 0; 1001 } 1002 // Evaluation of WideAddRec ensured that the narrow expression could be 1003 // extended outside the loop without overflow. This suggests that the wide use 1004 // evaluates to the same expression as the extended narrow use, but doesn't 1005 // absolutely guarantee it. Hence the following failsafe check. In rare cases 1006 // where it fails, we simply throw away the newly created wide use. 1007 if (WideAddRec != SE->getSCEV(WideUse)) { 1008 DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse 1009 << ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n"); 1010 DeadInsts.push_back(WideUse); 1011 return 0; 1012 } 1013 1014 // Returning WideUse pushes it on the worklist. 1015 return WideUse; 1016} 1017 1018/// pushNarrowIVUsers - Add eligible users of NarrowDef to NarrowIVUsers. 1019/// 1020void WidenIV::pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef) { 1021 for (User *U : NarrowDef->users()) { 1022 Instruction *NarrowUser = cast<Instruction>(U); 1023 1024 // Handle data flow merges and bizarre phi cycles. 1025 if (!Widened.insert(NarrowUser)) 1026 continue; 1027 1028 NarrowIVUsers.push_back(NarrowIVDefUse(NarrowDef, NarrowUser, WideDef)); 1029 } 1030} 1031 1032/// CreateWideIV - Process a single induction variable. First use the 1033/// SCEVExpander to create a wide induction variable that evaluates to the same 1034/// recurrence as the original narrow IV. Then use a worklist to forward 1035/// traverse the narrow IV's def-use chain. After WidenIVUse has processed all 1036/// interesting IV users, the narrow IV will be isolated for removal by 1037/// DeleteDeadPHIs. 1038/// 1039/// It would be simpler to delete uses as they are processed, but we must avoid 1040/// invalidating SCEV expressions. 1041/// 1042PHINode *WidenIV::CreateWideIV(SCEVExpander &Rewriter) { 1043 // Is this phi an induction variable? 1044 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi)); 1045 if (!AddRec) 1046 return NULL; 1047 1048 // Widen the induction variable expression. 1049 const SCEV *WideIVExpr = IsSigned ? 1050 SE->getSignExtendExpr(AddRec, WideType) : 1051 SE->getZeroExtendExpr(AddRec, WideType); 1052 1053 assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType && 1054 "Expect the new IV expression to preserve its type"); 1055 1056 // Can the IV be extended outside the loop without overflow? 1057 AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr); 1058 if (!AddRec || AddRec->getLoop() != L) 1059 return NULL; 1060 1061 // An AddRec must have loop-invariant operands. Since this AddRec is 1062 // materialized by a loop header phi, the expression cannot have any post-loop 1063 // operands, so they must dominate the loop header. 1064 assert(SE->properlyDominates(AddRec->getStart(), L->getHeader()) && 1065 SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader()) 1066 && "Loop header phi recurrence inputs do not dominate the loop"); 1067 1068 // The rewriter provides a value for the desired IV expression. This may 1069 // either find an existing phi or materialize a new one. Either way, we 1070 // expect a well-formed cyclic phi-with-increments. i.e. any operand not part 1071 // of the phi-SCC dominates the loop entry. 1072 Instruction *InsertPt = L->getHeader()->begin(); 1073 WidePhi = cast<PHINode>(Rewriter.expandCodeFor(AddRec, WideType, InsertPt)); 1074 1075 // Remembering the WideIV increment generated by SCEVExpander allows 1076 // WidenIVUse to reuse it when widening the narrow IV's increment. We don't 1077 // employ a general reuse mechanism because the call above is the only call to 1078 // SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses. 1079 if (BasicBlock *LatchBlock = L->getLoopLatch()) { 1080 WideInc = 1081 cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock)); 1082 WideIncExpr = SE->getSCEV(WideInc); 1083 } 1084 1085 DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n"); 1086 ++NumWidened; 1087 1088 // Traverse the def-use chain using a worklist starting at the original IV. 1089 assert(Widened.empty() && NarrowIVUsers.empty() && "expect initial state" ); 1090 1091 Widened.insert(OrigPhi); 1092 pushNarrowIVUsers(OrigPhi, WidePhi); 1093 1094 while (!NarrowIVUsers.empty()) { 1095 NarrowIVDefUse DU = NarrowIVUsers.pop_back_val(); 1096 1097 // Process a def-use edge. This may replace the use, so don't hold a 1098 // use_iterator across it. 1099 Instruction *WideUse = WidenIVUse(DU, Rewriter); 1100 1101 // Follow all def-use edges from the previous narrow use. 1102 if (WideUse) 1103 pushNarrowIVUsers(DU.NarrowUse, WideUse); 1104 1105 // WidenIVUse may have removed the def-use edge. 1106 if (DU.NarrowDef->use_empty()) 1107 DeadInsts.push_back(DU.NarrowDef); 1108 } 1109 return WidePhi; 1110} 1111 1112//===----------------------------------------------------------------------===// 1113// Live IV Reduction - Minimize IVs live across the loop. 1114//===----------------------------------------------------------------------===// 1115 1116 1117//===----------------------------------------------------------------------===// 1118// Simplification of IV users based on SCEV evaluation. 1119//===----------------------------------------------------------------------===// 1120 1121namespace { 1122 class IndVarSimplifyVisitor : public IVVisitor { 1123 ScalarEvolution *SE; 1124 const DataLayout *DL; 1125 PHINode *IVPhi; 1126 1127 public: 1128 WideIVInfo WI; 1129 1130 IndVarSimplifyVisitor(PHINode *IV, ScalarEvolution *SCEV, 1131 const DataLayout *DL, const DominatorTree *DTree): 1132 SE(SCEV), DL(DL), IVPhi(IV) { 1133 DT = DTree; 1134 WI.NarrowIV = IVPhi; 1135 if (ReduceLiveIVs) 1136 setSplitOverflowIntrinsics(); 1137 } 1138 1139 // Implement the interface used by simplifyUsersOfIV. 1140 void visitCast(CastInst *Cast) override { visitIVCast(Cast, WI, SE, DL); } 1141 }; 1142} 1143 1144/// SimplifyAndExtend - Iteratively perform simplification on a worklist of IV 1145/// users. Each successive simplification may push more users which may 1146/// themselves be candidates for simplification. 1147/// 1148/// Sign/Zero extend elimination is interleaved with IV simplification. 1149/// 1150void IndVarSimplify::SimplifyAndExtend(Loop *L, 1151 SCEVExpander &Rewriter, 1152 LPPassManager &LPM) { 1153 SmallVector<WideIVInfo, 8> WideIVs; 1154 1155 SmallVector<PHINode*, 8> LoopPhis; 1156 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) { 1157 LoopPhis.push_back(cast<PHINode>(I)); 1158 } 1159 // Each round of simplification iterates through the SimplifyIVUsers worklist 1160 // for all current phis, then determines whether any IVs can be 1161 // widened. Widening adds new phis to LoopPhis, inducing another round of 1162 // simplification on the wide IVs. 1163 while (!LoopPhis.empty()) { 1164 // Evaluate as many IV expressions as possible before widening any IVs. This 1165 // forces SCEV to set no-wrap flags before evaluating sign/zero 1166 // extension. The first time SCEV attempts to normalize sign/zero extension, 1167 // the result becomes final. So for the most predictable results, we delay 1168 // evaluation of sign/zero extend evaluation until needed, and avoid running 1169 // other SCEV based analysis prior to SimplifyAndExtend. 1170 do { 1171 PHINode *CurrIV = LoopPhis.pop_back_val(); 1172 1173 // Information about sign/zero extensions of CurrIV. 1174 IndVarSimplifyVisitor Visitor(CurrIV, SE, DL, DT); 1175 1176 Changed |= simplifyUsersOfIV(CurrIV, SE, &LPM, DeadInsts, &Visitor); 1177 1178 if (Visitor.WI.WidestNativeType) { 1179 WideIVs.push_back(Visitor.WI); 1180 } 1181 } while(!LoopPhis.empty()); 1182 1183 for (; !WideIVs.empty(); WideIVs.pop_back()) { 1184 WidenIV Widener(WideIVs.back(), LI, SE, DT, DeadInsts); 1185 if (PHINode *WidePhi = Widener.CreateWideIV(Rewriter)) { 1186 Changed = true; 1187 LoopPhis.push_back(WidePhi); 1188 } 1189 } 1190 } 1191} 1192 1193//===----------------------------------------------------------------------===// 1194// LinearFunctionTestReplace and its kin. Rewrite the loop exit condition. 1195//===----------------------------------------------------------------------===// 1196 1197/// Check for expressions that ScalarEvolution generates to compute 1198/// BackedgeTakenInfo. If these expressions have not been reduced, then 1199/// expanding them may incur additional cost (albeit in the loop preheader). 1200static bool isHighCostExpansion(const SCEV *S, BranchInst *BI, 1201 SmallPtrSet<const SCEV*, 8> &Processed, 1202 ScalarEvolution *SE) { 1203 if (!Processed.insert(S)) 1204 return false; 1205 1206 // If the backedge-taken count is a UDiv, it's very likely a UDiv that 1207 // ScalarEvolution's HowFarToZero or HowManyLessThans produced to compute a 1208 // precise expression, rather than a UDiv from the user's code. If we can't 1209 // find a UDiv in the code with some simple searching, assume the former and 1210 // forego rewriting the loop. 1211 if (isa<SCEVUDivExpr>(S)) { 1212 ICmpInst *OrigCond = dyn_cast<ICmpInst>(BI->getCondition()); 1213 if (!OrigCond) return true; 1214 const SCEV *R = SE->getSCEV(OrigCond->getOperand(1)); 1215 R = SE->getMinusSCEV(R, SE->getConstant(R->getType(), 1)); 1216 if (R != S) { 1217 const SCEV *L = SE->getSCEV(OrigCond->getOperand(0)); 1218 L = SE->getMinusSCEV(L, SE->getConstant(L->getType(), 1)); 1219 if (L != S) 1220 return true; 1221 } 1222 } 1223 1224 // Recurse past add expressions, which commonly occur in the 1225 // BackedgeTakenCount. They may already exist in program code, and if not, 1226 // they are not too expensive rematerialize. 1227 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 1228 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end(); 1229 I != E; ++I) { 1230 if (isHighCostExpansion(*I, BI, Processed, SE)) 1231 return true; 1232 } 1233 return false; 1234 } 1235 1236 // HowManyLessThans uses a Max expression whenever the loop is not guarded by 1237 // the exit condition. 1238 if (isa<SCEVSMaxExpr>(S) || isa<SCEVUMaxExpr>(S)) 1239 return true; 1240 1241 // If we haven't recognized an expensive SCEV pattern, assume it's an 1242 // expression produced by program code. 1243 return false; 1244} 1245 1246/// canExpandBackedgeTakenCount - Return true if this loop's backedge taken 1247/// count expression can be safely and cheaply expanded into an instruction 1248/// sequence that can be used by LinearFunctionTestReplace. 1249/// 1250/// TODO: This fails for pointer-type loop counters with greater than one byte 1251/// strides, consequently preventing LFTR from running. For the purpose of LFTR 1252/// we could skip this check in the case that the LFTR loop counter (chosen by 1253/// FindLoopCounter) is also pointer type. Instead, we could directly convert 1254/// the loop test to an inequality test by checking the target data's alignment 1255/// of element types (given that the initial pointer value originates from or is 1256/// used by ABI constrained operation, as opposed to inttoptr/ptrtoint). 1257/// However, we don't yet have a strong motivation for converting loop tests 1258/// into inequality tests. 1259static bool canExpandBackedgeTakenCount(Loop *L, ScalarEvolution *SE) { 1260 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L); 1261 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) || 1262 BackedgeTakenCount->isZero()) 1263 return false; 1264 1265 if (!L->getExitingBlock()) 1266 return false; 1267 1268 // Can't rewrite non-branch yet. 1269 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1270 if (!BI) 1271 return false; 1272 1273 SmallPtrSet<const SCEV*, 8> Processed; 1274 if (isHighCostExpansion(BackedgeTakenCount, BI, Processed, SE)) 1275 return false; 1276 1277 return true; 1278} 1279 1280/// getLoopPhiForCounter - Return the loop header phi IFF IncV adds a loop 1281/// invariant value to the phi. 1282static PHINode *getLoopPhiForCounter(Value *IncV, Loop *L, DominatorTree *DT) { 1283 Instruction *IncI = dyn_cast<Instruction>(IncV); 1284 if (!IncI) 1285 return 0; 1286 1287 switch (IncI->getOpcode()) { 1288 case Instruction::Add: 1289 case Instruction::Sub: 1290 break; 1291 case Instruction::GetElementPtr: 1292 // An IV counter must preserve its type. 1293 if (IncI->getNumOperands() == 2) 1294 break; 1295 default: 1296 return 0; 1297 } 1298 1299 PHINode *Phi = dyn_cast<PHINode>(IncI->getOperand(0)); 1300 if (Phi && Phi->getParent() == L->getHeader()) { 1301 if (isLoopInvariant(IncI->getOperand(1), L, DT)) 1302 return Phi; 1303 return 0; 1304 } 1305 if (IncI->getOpcode() == Instruction::GetElementPtr) 1306 return 0; 1307 1308 // Allow add/sub to be commuted. 1309 Phi = dyn_cast<PHINode>(IncI->getOperand(1)); 1310 if (Phi && Phi->getParent() == L->getHeader()) { 1311 if (isLoopInvariant(IncI->getOperand(0), L, DT)) 1312 return Phi; 1313 } 1314 return 0; 1315} 1316 1317/// Return the compare guarding the loop latch, or NULL for unrecognized tests. 1318static ICmpInst *getLoopTest(Loop *L) { 1319 assert(L->getExitingBlock() && "expected loop exit"); 1320 1321 BasicBlock *LatchBlock = L->getLoopLatch(); 1322 // Don't bother with LFTR if the loop is not properly simplified. 1323 if (!LatchBlock) 1324 return 0; 1325 1326 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1327 assert(BI && "expected exit branch"); 1328 1329 return dyn_cast<ICmpInst>(BI->getCondition()); 1330} 1331 1332/// needsLFTR - LinearFunctionTestReplace policy. Return true unless we can show 1333/// that the current exit test is already sufficiently canonical. 1334static bool needsLFTR(Loop *L, DominatorTree *DT) { 1335 // Do LFTR to simplify the exit condition to an ICMP. 1336 ICmpInst *Cond = getLoopTest(L); 1337 if (!Cond) 1338 return true; 1339 1340 // Do LFTR to simplify the exit ICMP to EQ/NE 1341 ICmpInst::Predicate Pred = Cond->getPredicate(); 1342 if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ) 1343 return true; 1344 1345 // Look for a loop invariant RHS 1346 Value *LHS = Cond->getOperand(0); 1347 Value *RHS = Cond->getOperand(1); 1348 if (!isLoopInvariant(RHS, L, DT)) { 1349 if (!isLoopInvariant(LHS, L, DT)) 1350 return true; 1351 std::swap(LHS, RHS); 1352 } 1353 // Look for a simple IV counter LHS 1354 PHINode *Phi = dyn_cast<PHINode>(LHS); 1355 if (!Phi) 1356 Phi = getLoopPhiForCounter(LHS, L, DT); 1357 1358 if (!Phi) 1359 return true; 1360 1361 // Do LFTR if PHI node is defined in the loop, but is *not* a counter. 1362 int Idx = Phi->getBasicBlockIndex(L->getLoopLatch()); 1363 if (Idx < 0) 1364 return true; 1365 1366 // Do LFTR if the exit condition's IV is *not* a simple counter. 1367 Value *IncV = Phi->getIncomingValue(Idx); 1368 return Phi != getLoopPhiForCounter(IncV, L, DT); 1369} 1370 1371/// Recursive helper for hasConcreteDef(). Unfortunately, this currently boils 1372/// down to checking that all operands are constant and listing instructions 1373/// that may hide undef. 1374static bool hasConcreteDefImpl(Value *V, SmallPtrSet<Value*, 8> &Visited, 1375 unsigned Depth) { 1376 if (isa<Constant>(V)) 1377 return !isa<UndefValue>(V); 1378 1379 if (Depth >= 6) 1380 return false; 1381 1382 // Conservatively handle non-constant non-instructions. For example, Arguments 1383 // may be undef. 1384 Instruction *I = dyn_cast<Instruction>(V); 1385 if (!I) 1386 return false; 1387 1388 // Load and return values may be undef. 1389 if(I->mayReadFromMemory() || isa<CallInst>(I) || isa<InvokeInst>(I)) 1390 return false; 1391 1392 // Optimistically handle other instructions. 1393 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) { 1394 if (!Visited.insert(*OI)) 1395 continue; 1396 if (!hasConcreteDefImpl(*OI, Visited, Depth+1)) 1397 return false; 1398 } 1399 return true; 1400} 1401 1402/// Return true if the given value is concrete. We must prove that undef can 1403/// never reach it. 1404/// 1405/// TODO: If we decide that this is a good approach to checking for undef, we 1406/// may factor it into a common location. 1407static bool hasConcreteDef(Value *V) { 1408 SmallPtrSet<Value*, 8> Visited; 1409 Visited.insert(V); 1410 return hasConcreteDefImpl(V, Visited, 0); 1411} 1412 1413/// AlmostDeadIV - Return true if this IV has any uses other than the (soon to 1414/// be rewritten) loop exit test. 1415static bool AlmostDeadIV(PHINode *Phi, BasicBlock *LatchBlock, Value *Cond) { 1416 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock); 1417 Value *IncV = Phi->getIncomingValue(LatchIdx); 1418 1419 for (User *U : Phi->users()) 1420 if (U != Cond && U != IncV) return false; 1421 1422 for (User *U : IncV->users()) 1423 if (U != Cond && U != Phi) return false; 1424 return true; 1425} 1426 1427/// FindLoopCounter - Find an affine IV in canonical form. 1428/// 1429/// BECount may be an i8* pointer type. The pointer difference is already 1430/// valid count without scaling the address stride, so it remains a pointer 1431/// expression as far as SCEV is concerned. 1432/// 1433/// Currently only valid for LFTR. See the comments on hasConcreteDef below. 1434/// 1435/// FIXME: Accept -1 stride and set IVLimit = IVInit - BECount 1436/// 1437/// FIXME: Accept non-unit stride as long as SCEV can reduce BECount * Stride. 1438/// This is difficult in general for SCEV because of potential overflow. But we 1439/// could at least handle constant BECounts. 1440static PHINode * 1441FindLoopCounter(Loop *L, const SCEV *BECount, 1442 ScalarEvolution *SE, DominatorTree *DT, const DataLayout *DL) { 1443 uint64_t BCWidth = SE->getTypeSizeInBits(BECount->getType()); 1444 1445 Value *Cond = 1446 cast<BranchInst>(L->getExitingBlock()->getTerminator())->getCondition(); 1447 1448 // Loop over all of the PHI nodes, looking for a simple counter. 1449 PHINode *BestPhi = 0; 1450 const SCEV *BestInit = 0; 1451 BasicBlock *LatchBlock = L->getLoopLatch(); 1452 assert(LatchBlock && "needsLFTR should guarantee a loop latch"); 1453 1454 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) { 1455 PHINode *Phi = cast<PHINode>(I); 1456 if (!SE->isSCEVable(Phi->getType())) 1457 continue; 1458 1459 // Avoid comparing an integer IV against a pointer Limit. 1460 if (BECount->getType()->isPointerTy() && !Phi->getType()->isPointerTy()) 1461 continue; 1462 1463 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Phi)); 1464 if (!AR || AR->getLoop() != L || !AR->isAffine()) 1465 continue; 1466 1467 // AR may be a pointer type, while BECount is an integer type. 1468 // AR may be wider than BECount. With eq/ne tests overflow is immaterial. 1469 // AR may not be a narrower type, or we may never exit. 1470 uint64_t PhiWidth = SE->getTypeSizeInBits(AR->getType()); 1471 if (PhiWidth < BCWidth || (DL && !DL->isLegalInteger(PhiWidth))) 1472 continue; 1473 1474 const SCEV *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE)); 1475 if (!Step || !Step->isOne()) 1476 continue; 1477 1478 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock); 1479 Value *IncV = Phi->getIncomingValue(LatchIdx); 1480 if (getLoopPhiForCounter(IncV, L, DT) != Phi) 1481 continue; 1482 1483 // Avoid reusing a potentially undef value to compute other values that may 1484 // have originally had a concrete definition. 1485 if (!hasConcreteDef(Phi)) { 1486 // We explicitly allow unknown phis as long as they are already used by 1487 // the loop test. In this case we assume that performing LFTR could not 1488 // increase the number of undef users. 1489 if (ICmpInst *Cond = getLoopTest(L)) { 1490 if (Phi != getLoopPhiForCounter(Cond->getOperand(0), L, DT) 1491 && Phi != getLoopPhiForCounter(Cond->getOperand(1), L, DT)) { 1492 continue; 1493 } 1494 } 1495 } 1496 const SCEV *Init = AR->getStart(); 1497 1498 if (BestPhi && !AlmostDeadIV(BestPhi, LatchBlock, Cond)) { 1499 // Don't force a live loop counter if another IV can be used. 1500 if (AlmostDeadIV(Phi, LatchBlock, Cond)) 1501 continue; 1502 1503 // Prefer to count-from-zero. This is a more "canonical" counter form. It 1504 // also prefers integer to pointer IVs. 1505 if (BestInit->isZero() != Init->isZero()) { 1506 if (BestInit->isZero()) 1507 continue; 1508 } 1509 // If two IVs both count from zero or both count from nonzero then the 1510 // narrower is likely a dead phi that has been widened. Use the wider phi 1511 // to allow the other to be eliminated. 1512 else if (PhiWidth <= SE->getTypeSizeInBits(BestPhi->getType())) 1513 continue; 1514 } 1515 BestPhi = Phi; 1516 BestInit = Init; 1517 } 1518 return BestPhi; 1519} 1520 1521/// genLoopLimit - Help LinearFunctionTestReplace by generating a value that 1522/// holds the RHS of the new loop test. 1523static Value *genLoopLimit(PHINode *IndVar, const SCEV *IVCount, Loop *L, 1524 SCEVExpander &Rewriter, ScalarEvolution *SE) { 1525 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IndVar)); 1526 assert(AR && AR->getLoop() == L && AR->isAffine() && "bad loop counter"); 1527 const SCEV *IVInit = AR->getStart(); 1528 1529 // IVInit may be a pointer while IVCount is an integer when FindLoopCounter 1530 // finds a valid pointer IV. Sign extend BECount in order to materialize a 1531 // GEP. Avoid running SCEVExpander on a new pointer value, instead reusing 1532 // the existing GEPs whenever possible. 1533 if (IndVar->getType()->isPointerTy() 1534 && !IVCount->getType()->isPointerTy()) { 1535 1536 // IVOffset will be the new GEP offset that is interpreted by GEP as a 1537 // signed value. IVCount on the other hand represents the loop trip count, 1538 // which is an unsigned value. FindLoopCounter only allows induction 1539 // variables that have a positive unit stride of one. This means we don't 1540 // have to handle the case of negative offsets (yet) and just need to zero 1541 // extend IVCount. 1542 Type *OfsTy = SE->getEffectiveSCEVType(IVInit->getType()); 1543 const SCEV *IVOffset = SE->getTruncateOrZeroExtend(IVCount, OfsTy); 1544 1545 // Expand the code for the iteration count. 1546 assert(SE->isLoopInvariant(IVOffset, L) && 1547 "Computed iteration count is not loop invariant!"); 1548 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1549 Value *GEPOffset = Rewriter.expandCodeFor(IVOffset, OfsTy, BI); 1550 1551 Value *GEPBase = IndVar->getIncomingValueForBlock(L->getLoopPreheader()); 1552 assert(AR->getStart() == SE->getSCEV(GEPBase) && "bad loop counter"); 1553 // We could handle pointer IVs other than i8*, but we need to compensate for 1554 // gep index scaling. See canExpandBackedgeTakenCount comments. 1555 assert(SE->getSizeOfExpr(IntegerType::getInt64Ty(IndVar->getContext()), 1556 cast<PointerType>(GEPBase->getType())->getElementType())->isOne() 1557 && "unit stride pointer IV must be i8*"); 1558 1559 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator()); 1560 return Builder.CreateGEP(GEPBase, GEPOffset, "lftr.limit"); 1561 } 1562 else { 1563 // In any other case, convert both IVInit and IVCount to integers before 1564 // comparing. This may result in SCEV expension of pointers, but in practice 1565 // SCEV will fold the pointer arithmetic away as such: 1566 // BECount = (IVEnd - IVInit - 1) => IVLimit = IVInit (postinc). 1567 // 1568 // Valid Cases: (1) both integers is most common; (2) both may be pointers 1569 // for simple memset-style loops. 1570 // 1571 // IVInit integer and IVCount pointer would only occur if a canonical IV 1572 // were generated on top of case #2, which is not expected. 1573 1574 const SCEV *IVLimit = 0; 1575 // For unit stride, IVCount = Start + BECount with 2's complement overflow. 1576 // For non-zero Start, compute IVCount here. 1577 if (AR->getStart()->isZero()) 1578 IVLimit = IVCount; 1579 else { 1580 assert(AR->getStepRecurrence(*SE)->isOne() && "only handles unit stride"); 1581 const SCEV *IVInit = AR->getStart(); 1582 1583 // For integer IVs, truncate the IV before computing IVInit + BECount. 1584 if (SE->getTypeSizeInBits(IVInit->getType()) 1585 > SE->getTypeSizeInBits(IVCount->getType())) 1586 IVInit = SE->getTruncateExpr(IVInit, IVCount->getType()); 1587 1588 IVLimit = SE->getAddExpr(IVInit, IVCount); 1589 } 1590 // Expand the code for the iteration count. 1591 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1592 IRBuilder<> Builder(BI); 1593 assert(SE->isLoopInvariant(IVLimit, L) && 1594 "Computed iteration count is not loop invariant!"); 1595 // Ensure that we generate the same type as IndVar, or a smaller integer 1596 // type. In the presence of null pointer values, we have an integer type 1597 // SCEV expression (IVInit) for a pointer type IV value (IndVar). 1598 Type *LimitTy = IVCount->getType()->isPointerTy() ? 1599 IndVar->getType() : IVCount->getType(); 1600 return Rewriter.expandCodeFor(IVLimit, LimitTy, BI); 1601 } 1602} 1603 1604/// LinearFunctionTestReplace - This method rewrites the exit condition of the 1605/// loop to be a canonical != comparison against the incremented loop induction 1606/// variable. This pass is able to rewrite the exit tests of any loop where the 1607/// SCEV analysis can determine a loop-invariant trip count of the loop, which 1608/// is actually a much broader range than just linear tests. 1609Value *IndVarSimplify:: 1610LinearFunctionTestReplace(Loop *L, 1611 const SCEV *BackedgeTakenCount, 1612 PHINode *IndVar, 1613 SCEVExpander &Rewriter) { 1614 assert(canExpandBackedgeTakenCount(L, SE) && "precondition"); 1615 1616 // Initialize CmpIndVar and IVCount to their preincremented values. 1617 Value *CmpIndVar = IndVar; 1618 const SCEV *IVCount = BackedgeTakenCount; 1619 1620 // If the exiting block is the same as the backedge block, we prefer to 1621 // compare against the post-incremented value, otherwise we must compare 1622 // against the preincremented value. 1623 if (L->getExitingBlock() == L->getLoopLatch()) { 1624 // Add one to the "backedge-taken" count to get the trip count. 1625 // This addition may overflow, which is valid as long as the comparison is 1626 // truncated to BackedgeTakenCount->getType(). 1627 IVCount = SE->getAddExpr(BackedgeTakenCount, 1628 SE->getConstant(BackedgeTakenCount->getType(), 1)); 1629 // The BackedgeTaken expression contains the number of times that the 1630 // backedge branches to the loop header. This is one less than the 1631 // number of times the loop executes, so use the incremented indvar. 1632 CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock()); 1633 } 1634 1635 Value *ExitCnt = genLoopLimit(IndVar, IVCount, L, Rewriter, SE); 1636 assert(ExitCnt->getType()->isPointerTy() == IndVar->getType()->isPointerTy() 1637 && "genLoopLimit missed a cast"); 1638 1639 // Insert a new icmp_ne or icmp_eq instruction before the branch. 1640 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1641 ICmpInst::Predicate P; 1642 if (L->contains(BI->getSuccessor(0))) 1643 P = ICmpInst::ICMP_NE; 1644 else 1645 P = ICmpInst::ICMP_EQ; 1646 1647 DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n" 1648 << " LHS:" << *CmpIndVar << '\n' 1649 << " op:\t" 1650 << (P == ICmpInst::ICMP_NE ? "!=" : "==") << "\n" 1651 << " RHS:\t" << *ExitCnt << "\n" 1652 << " IVCount:\t" << *IVCount << "\n"); 1653 1654 IRBuilder<> Builder(BI); 1655 1656 // LFTR can ignore IV overflow and truncate to the width of 1657 // BECount. This avoids materializing the add(zext(add)) expression. 1658 unsigned CmpIndVarSize = SE->getTypeSizeInBits(CmpIndVar->getType()); 1659 unsigned ExitCntSize = SE->getTypeSizeInBits(ExitCnt->getType()); 1660 if (CmpIndVarSize > ExitCntSize) { 1661 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(SE->getSCEV(IndVar)); 1662 const SCEV *ARStart = AR->getStart(); 1663 const SCEV *ARStep = AR->getStepRecurrence(*SE); 1664 // For constant IVCount, avoid truncation. 1665 if (isa<SCEVConstant>(ARStart) && isa<SCEVConstant>(IVCount)) { 1666 const APInt &Start = cast<SCEVConstant>(ARStart)->getValue()->getValue(); 1667 APInt Count = cast<SCEVConstant>(IVCount)->getValue()->getValue(); 1668 // Note that the post-inc value of BackedgeTakenCount may have overflowed 1669 // above such that IVCount is now zero. 1670 if (IVCount != BackedgeTakenCount && Count == 0) { 1671 Count = APInt::getMaxValue(Count.getBitWidth()).zext(CmpIndVarSize); 1672 ++Count; 1673 } 1674 else 1675 Count = Count.zext(CmpIndVarSize); 1676 APInt NewLimit; 1677 if (cast<SCEVConstant>(ARStep)->getValue()->isNegative()) 1678 NewLimit = Start - Count; 1679 else 1680 NewLimit = Start + Count; 1681 ExitCnt = ConstantInt::get(CmpIndVar->getType(), NewLimit); 1682 1683 DEBUG(dbgs() << " Widen RHS:\t" << *ExitCnt << "\n"); 1684 } else { 1685 CmpIndVar = Builder.CreateTrunc(CmpIndVar, ExitCnt->getType(), 1686 "lftr.wideiv"); 1687 } 1688 } 1689 Value *Cond = Builder.CreateICmp(P, CmpIndVar, ExitCnt, "exitcond"); 1690 Value *OrigCond = BI->getCondition(); 1691 // It's tempting to use replaceAllUsesWith here to fully replace the old 1692 // comparison, but that's not immediately safe, since users of the old 1693 // comparison may not be dominated by the new comparison. Instead, just 1694 // update the branch to use the new comparison; in the common case this 1695 // will make old comparison dead. 1696 BI->setCondition(Cond); 1697 DeadInsts.push_back(OrigCond); 1698 1699 ++NumLFTR; 1700 Changed = true; 1701 return Cond; 1702} 1703 1704//===----------------------------------------------------------------------===// 1705// SinkUnusedInvariants. A late subpass to cleanup loop preheaders. 1706//===----------------------------------------------------------------------===// 1707 1708/// If there's a single exit block, sink any loop-invariant values that 1709/// were defined in the preheader but not used inside the loop into the 1710/// exit block to reduce register pressure in the loop. 1711void IndVarSimplify::SinkUnusedInvariants(Loop *L) { 1712 BasicBlock *ExitBlock = L->getExitBlock(); 1713 if (!ExitBlock) return; 1714 1715 BasicBlock *Preheader = L->getLoopPreheader(); 1716 if (!Preheader) return; 1717 1718 Instruction *InsertPt = ExitBlock->getFirstInsertionPt(); 1719 BasicBlock::iterator I = Preheader->getTerminator(); 1720 while (I != Preheader->begin()) { 1721 --I; 1722 // New instructions were inserted at the end of the preheader. 1723 if (isa<PHINode>(I)) 1724 break; 1725 1726 // Don't move instructions which might have side effects, since the side 1727 // effects need to complete before instructions inside the loop. Also don't 1728 // move instructions which might read memory, since the loop may modify 1729 // memory. Note that it's okay if the instruction might have undefined 1730 // behavior: LoopSimplify guarantees that the preheader dominates the exit 1731 // block. 1732 if (I->mayHaveSideEffects() || I->mayReadFromMemory()) 1733 continue; 1734 1735 // Skip debug info intrinsics. 1736 if (isa<DbgInfoIntrinsic>(I)) 1737 continue; 1738 1739 // Skip landingpad instructions. 1740 if (isa<LandingPadInst>(I)) 1741 continue; 1742 1743 // Don't sink alloca: we never want to sink static alloca's out of the 1744 // entry block, and correctly sinking dynamic alloca's requires 1745 // checks for stacksave/stackrestore intrinsics. 1746 // FIXME: Refactor this check somehow? 1747 if (isa<AllocaInst>(I)) 1748 continue; 1749 1750 // Determine if there is a use in or before the loop (direct or 1751 // otherwise). 1752 bool UsedInLoop = false; 1753 for (Use &U : I->uses()) { 1754 Instruction *User = cast<Instruction>(U.getUser()); 1755 BasicBlock *UseBB = User->getParent(); 1756 if (PHINode *P = dyn_cast<PHINode>(User)) { 1757 unsigned i = 1758 PHINode::getIncomingValueNumForOperand(U.getOperandNo()); 1759 UseBB = P->getIncomingBlock(i); 1760 } 1761 if (UseBB == Preheader || L->contains(UseBB)) { 1762 UsedInLoop = true; 1763 break; 1764 } 1765 } 1766 1767 // If there is, the def must remain in the preheader. 1768 if (UsedInLoop) 1769 continue; 1770 1771 // Otherwise, sink it to the exit block. 1772 Instruction *ToMove = I; 1773 bool Done = false; 1774 1775 if (I != Preheader->begin()) { 1776 // Skip debug info intrinsics. 1777 do { 1778 --I; 1779 } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin()); 1780 1781 if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin()) 1782 Done = true; 1783 } else { 1784 Done = true; 1785 } 1786 1787 ToMove->moveBefore(InsertPt); 1788 if (Done) break; 1789 InsertPt = ToMove; 1790 } 1791} 1792 1793//===----------------------------------------------------------------------===// 1794// IndVarSimplify driver. Manage several subpasses of IV simplification. 1795//===----------------------------------------------------------------------===// 1796 1797bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) { 1798 if (skipOptnoneFunction(L)) 1799 return false; 1800 1801 // If LoopSimplify form is not available, stay out of trouble. Some notes: 1802 // - LSR currently only supports LoopSimplify-form loops. Indvars' 1803 // canonicalization can be a pessimization without LSR to "clean up" 1804 // afterwards. 1805 // - We depend on having a preheader; in particular, 1806 // Loop::getCanonicalInductionVariable only supports loops with preheaders, 1807 // and we're in trouble if we can't find the induction variable even when 1808 // we've manually inserted one. 1809 if (!L->isLoopSimplifyForm()) 1810 return false; 1811 1812 LI = &getAnalysis<LoopInfo>(); 1813 SE = &getAnalysis<ScalarEvolution>(); 1814 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 1815 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>(); 1816 DL = DLP ? &DLP->getDataLayout() : 0; 1817 TLI = getAnalysisIfAvailable<TargetLibraryInfo>(); 1818 1819 DeadInsts.clear(); 1820 Changed = false; 1821 1822 // If there are any floating-point recurrences, attempt to 1823 // transform them to use integer recurrences. 1824 RewriteNonIntegerIVs(L); 1825 1826 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L); 1827 1828 // Create a rewriter object which we'll use to transform the code with. 1829 SCEVExpander Rewriter(*SE, "indvars"); 1830#ifndef NDEBUG 1831 Rewriter.setDebugType(DEBUG_TYPE); 1832#endif 1833 1834 // Eliminate redundant IV users. 1835 // 1836 // Simplification works best when run before other consumers of SCEV. We 1837 // attempt to avoid evaluating SCEVs for sign/zero extend operations until 1838 // other expressions involving loop IVs have been evaluated. This helps SCEV 1839 // set no-wrap flags before normalizing sign/zero extension. 1840 Rewriter.disableCanonicalMode(); 1841 SimplifyAndExtend(L, Rewriter, LPM); 1842 1843 // Check to see if this loop has a computable loop-invariant execution count. 1844 // If so, this means that we can compute the final value of any expressions 1845 // that are recurrent in the loop, and substitute the exit values from the 1846 // loop into any instructions outside of the loop that use the final values of 1847 // the current expressions. 1848 // 1849 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount)) 1850 RewriteLoopExitValues(L, Rewriter); 1851 1852 // Eliminate redundant IV cycles. 1853 NumElimIV += Rewriter.replaceCongruentIVs(L, DT, DeadInsts); 1854 1855 // If we have a trip count expression, rewrite the loop's exit condition 1856 // using it. We can currently only handle loops with a single exit. 1857 if (canExpandBackedgeTakenCount(L, SE) && needsLFTR(L, DT)) { 1858 PHINode *IndVar = FindLoopCounter(L, BackedgeTakenCount, SE, DT, DL); 1859 if (IndVar) { 1860 // Check preconditions for proper SCEVExpander operation. SCEV does not 1861 // express SCEVExpander's dependencies, such as LoopSimplify. Instead any 1862 // pass that uses the SCEVExpander must do it. This does not work well for 1863 // loop passes because SCEVExpander makes assumptions about all loops, 1864 // while LoopPassManager only forces the current loop to be simplified. 1865 // 1866 // FIXME: SCEV expansion has no way to bail out, so the caller must 1867 // explicitly check any assumptions made by SCEV. Brittle. 1868 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(BackedgeTakenCount); 1869 if (!AR || AR->getLoop()->getLoopPreheader()) 1870 (void)LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar, 1871 Rewriter); 1872 } 1873 } 1874 // Clear the rewriter cache, because values that are in the rewriter's cache 1875 // can be deleted in the loop below, causing the AssertingVH in the cache to 1876 // trigger. 1877 Rewriter.clear(); 1878 1879 // Now that we're done iterating through lists, clean up any instructions 1880 // which are now dead. 1881 while (!DeadInsts.empty()) 1882 if (Instruction *Inst = 1883 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val())) 1884 RecursivelyDeleteTriviallyDeadInstructions(Inst, TLI); 1885 1886 // The Rewriter may not be used from this point on. 1887 1888 // Loop-invariant instructions in the preheader that aren't used in the 1889 // loop may be sunk below the loop to reduce register pressure. 1890 SinkUnusedInvariants(L); 1891 1892 // Clean up dead instructions. 1893 Changed |= DeleteDeadPHIs(L->getHeader(), TLI); 1894 // Check a post-condition. 1895 assert(L->isLCSSAForm(*DT) && 1896 "Indvars did not leave the loop in lcssa form!"); 1897 1898 // Verify that LFTR, and any other change have not interfered with SCEV's 1899 // ability to compute trip count. 1900#ifndef NDEBUG 1901 if (VerifyIndvars && !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) { 1902 SE->forgetLoop(L); 1903 const SCEV *NewBECount = SE->getBackedgeTakenCount(L); 1904 if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) < 1905 SE->getTypeSizeInBits(NewBECount->getType())) 1906 NewBECount = SE->getTruncateOrNoop(NewBECount, 1907 BackedgeTakenCount->getType()); 1908 else 1909 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount, 1910 NewBECount->getType()); 1911 assert(BackedgeTakenCount == NewBECount && "indvars must preserve SCEV"); 1912 } 1913#endif 1914 1915 return Changed; 1916} 1917