1 //===- IndVarSimplify.cpp - Induction Variable Elimination ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
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.
14 // This transformation makes the following changes to each loop with an
15 // identifiable induction variable:
16 // 1. All loops are transformed to have a SINGLE canonical induction variable
17 // which starts at zero and steps by one.
18 // 2. The canonical induction variable is guaranteed to be the first PHI node
19 // in the loop header block.
20 // 3. The canonical induction variable is guaranteed to be in a wide enough
21 // type so that IV expressions need not be (directly) zero-extended or
23 // 4. Any pointer arithmetic recurrences are raised to use array subscripts.
25 // If the trip count of a loop is computable, this pass also makes the following
27 // 1. The exit condition for the loop is canonicalized to compare the
28 // induction value against the exit value. This turns loops like:
29 // 'for (i = 7; i*i < 1000; ++i)' into 'for (i = 0; i != 25; ++i)'
30 // 2. Any use outside of the loop of an expression derived from the indvar
31 // is changed to compute the derived value outside of the loop, eliminating
32 // the dependence on the exit value of the induction variable. If the only
33 // purpose of the loop is to compute the exit value of some derived
34 // expression, this transformation will make the loop dead.
36 // This transformation should be followed by strength reduction after all of the
37 // desired loop transformations have been performed.
39 //===----------------------------------------------------------------------===//
41 #define DEBUG_TYPE "indvars"
42 #include "llvm/Transforms/Scalar.h"
43 #include "llvm/BasicBlock.h"
44 #include "llvm/Constants.h"
45 #include "llvm/Instructions.h"
46 #include "llvm/IntrinsicInst.h"
47 #include "llvm/LLVMContext.h"
48 #include "llvm/Type.h"
49 #include "llvm/Analysis/Dominators.h"
50 #include "llvm/Analysis/IVUsers.h"
51 #include "llvm/Analysis/ScalarEvolutionExpander.h"
52 #include "llvm/Analysis/LoopInfo.h"
53 #include "llvm/Analysis/LoopPass.h"
54 #include "llvm/Support/CFG.h"
55 #include "llvm/Support/CommandLine.h"
56 #include "llvm/Support/Debug.h"
57 #include "llvm/Support/raw_ostream.h"
58 #include "llvm/Transforms/Utils/Local.h"
59 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
60 #include "llvm/ADT/SmallVector.h"
61 #include "llvm/ADT/Statistic.h"
62 #include "llvm/ADT/STLExtras.h"
65 STATISTIC(NumRemoved , "Number of aux indvars removed");
66 STATISTIC(NumInserted, "Number of canonical indvars added");
67 STATISTIC(NumReplaced, "Number of exit values replaced");
68 STATISTIC(NumLFTR , "Number of loop exit tests replaced");
71 class IndVarSimplify : public LoopPass {
79 static char ID; // Pass identification, replacement for typeid
80 IndVarSimplify() : LoopPass(ID) {
81 initializeIndVarSimplifyPass(*PassRegistry::getPassRegistry());
84 virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
86 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
87 AU.addRequired<DominatorTree>();
88 AU.addRequired<LoopInfo>();
89 AU.addRequired<ScalarEvolution>();
90 AU.addRequiredID(LoopSimplifyID);
91 AU.addRequiredID(LCSSAID);
92 AU.addRequired<IVUsers>();
93 AU.addPreserved<ScalarEvolution>();
94 AU.addPreservedID(LoopSimplifyID);
95 AU.addPreservedID(LCSSAID);
96 AU.addPreserved<IVUsers>();
102 void EliminateIVComparisons();
103 void EliminateIVRemainders();
104 void RewriteNonIntegerIVs(Loop *L);
106 ICmpInst *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
108 BasicBlock *ExitingBlock,
110 SCEVExpander &Rewriter);
111 void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter);
113 void RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter);
115 void SinkUnusedInvariants(Loop *L);
117 void HandleFloatingPointIV(Loop *L, PHINode *PH);
121 char IndVarSimplify::ID = 0;
122 INITIALIZE_PASS_BEGIN(IndVarSimplify, "indvars",
123 "Canonicalize Induction Variables", false, false)
124 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
125 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
126 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
127 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
128 INITIALIZE_PASS_DEPENDENCY(LCSSA)
129 INITIALIZE_PASS_DEPENDENCY(IVUsers)
130 INITIALIZE_PASS_END(IndVarSimplify, "indvars",
131 "Canonicalize Induction Variables", false, false)
133 Pass *llvm::createIndVarSimplifyPass() {
134 return new IndVarSimplify();
137 /// LinearFunctionTestReplace - This method rewrites the exit condition of the
138 /// loop to be a canonical != comparison against the incremented loop induction
139 /// variable. This pass is able to rewrite the exit tests of any loop where the
140 /// SCEV analysis can determine a loop-invariant trip count of the loop, which
141 /// is actually a much broader range than just linear tests.
142 ICmpInst *IndVarSimplify::LinearFunctionTestReplace(Loop *L,
143 const SCEV *BackedgeTakenCount,
145 BasicBlock *ExitingBlock,
147 SCEVExpander &Rewriter) {
148 // Special case: If the backedge-taken count is a UDiv, it's very likely a
149 // UDiv that ScalarEvolution produced in order to compute a precise
150 // expression, rather than a UDiv from the user's code. If we can't find a
151 // UDiv in the code with some simple searching, assume the former and forego
152 // rewriting the loop.
153 if (isa<SCEVUDivExpr>(BackedgeTakenCount)) {
154 ICmpInst *OrigCond = dyn_cast<ICmpInst>(BI->getCondition());
155 if (!OrigCond) return 0;
156 const SCEV *R = SE->getSCEV(OrigCond->getOperand(1));
157 R = SE->getMinusSCEV(R, SE->getConstant(R->getType(), 1));
158 if (R != BackedgeTakenCount) {
159 const SCEV *L = SE->getSCEV(OrigCond->getOperand(0));
160 L = SE->getMinusSCEV(L, SE->getConstant(L->getType(), 1));
161 if (L != BackedgeTakenCount)
166 // If the exiting block is not the same as the backedge block, we must compare
167 // against the preincremented value, otherwise we prefer to compare against
168 // the post-incremented value.
170 const SCEV *RHS = BackedgeTakenCount;
171 if (ExitingBlock == L->getLoopLatch()) {
172 // Add one to the "backedge-taken" count to get the trip count.
173 // If this addition may overflow, we have to be more pessimistic and
174 // cast the induction variable before doing the add.
175 const SCEV *Zero = SE->getConstant(BackedgeTakenCount->getType(), 0);
177 SE->getAddExpr(BackedgeTakenCount,
178 SE->getConstant(BackedgeTakenCount->getType(), 1));
179 if ((isa<SCEVConstant>(N) && !N->isZero()) ||
180 SE->isLoopEntryGuardedByCond(L, ICmpInst::ICMP_NE, N, Zero)) {
181 // No overflow. Cast the sum.
182 RHS = SE->getTruncateOrZeroExtend(N, IndVar->getType());
184 // Potential overflow. Cast before doing the add.
185 RHS = SE->getTruncateOrZeroExtend(BackedgeTakenCount,
187 RHS = SE->getAddExpr(RHS,
188 SE->getConstant(IndVar->getType(), 1));
191 // The BackedgeTaken expression contains the number of times that the
192 // backedge branches to the loop header. This is one less than the
193 // number of times the loop executes, so use the incremented indvar.
194 CmpIndVar = IndVar->getIncomingValueForBlock(ExitingBlock);
196 // We have to use the preincremented value...
197 RHS = SE->getTruncateOrZeroExtend(BackedgeTakenCount,
202 // Expand the code for the iteration count.
203 assert(SE->isLoopInvariant(RHS, L) &&
204 "Computed iteration count is not loop invariant!");
205 Value *ExitCnt = Rewriter.expandCodeFor(RHS, IndVar->getType(), BI);
207 // Insert a new icmp_ne or icmp_eq instruction before the branch.
208 ICmpInst::Predicate Opcode;
209 if (L->contains(BI->getSuccessor(0)))
210 Opcode = ICmpInst::ICMP_NE;
212 Opcode = ICmpInst::ICMP_EQ;
214 DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
215 << " LHS:" << *CmpIndVar << '\n'
217 << (Opcode == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
218 << " RHS:\t" << *RHS << "\n");
220 ICmpInst *Cond = new ICmpInst(BI, Opcode, CmpIndVar, ExitCnt, "exitcond");
222 Value *OrigCond = BI->getCondition();
223 // It's tempting to use replaceAllUsesWith here to fully replace the old
224 // comparison, but that's not immediately safe, since users of the old
225 // comparison may not be dominated by the new comparison. Instead, just
226 // update the branch to use the new comparison; in the common case this
227 // will make old comparison dead.
228 BI->setCondition(Cond);
229 RecursivelyDeleteTriviallyDeadInstructions(OrigCond);
236 /// RewriteLoopExitValues - Check to see if this loop has a computable
237 /// loop-invariant execution count. If so, this means that we can compute the
238 /// final value of any expressions that are recurrent in the loop, and
239 /// substitute the exit values from the loop into any instructions outside of
240 /// the loop that use the final values of the current expressions.
242 /// This is mostly redundant with the regular IndVarSimplify activities that
243 /// happen later, except that it's more powerful in some cases, because it's
244 /// able to brute-force evaluate arbitrary instructions as long as they have
245 /// constant operands at the beginning of the loop.
246 void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) {
247 // Verify the input to the pass in already in LCSSA form.
248 assert(L->isLCSSAForm(*DT));
250 SmallVector<BasicBlock*, 8> ExitBlocks;
251 L->getUniqueExitBlocks(ExitBlocks);
253 // Find all values that are computed inside the loop, but used outside of it.
254 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan
255 // the exit blocks of the loop to find them.
256 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
257 BasicBlock *ExitBB = ExitBlocks[i];
259 // If there are no PHI nodes in this exit block, then no values defined
260 // inside the loop are used on this path, skip it.
261 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
264 unsigned NumPreds = PN->getNumIncomingValues();
266 // Iterate over all of the PHI nodes.
267 BasicBlock::iterator BBI = ExitBB->begin();
268 while ((PN = dyn_cast<PHINode>(BBI++))) {
270 continue; // dead use, don't replace it
272 // SCEV only supports integer expressions for now.
273 if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy())
276 // It's necessary to tell ScalarEvolution about this explicitly so that
277 // it can walk the def-use list and forget all SCEVs, as it may not be
278 // watching the PHI itself. Once the new exit value is in place, there
279 // may not be a def-use connection between the loop and every instruction
280 // which got a SCEVAddRecExpr for that loop.
283 // Iterate over all of the values in all the PHI nodes.
284 for (unsigned i = 0; i != NumPreds; ++i) {
285 // If the value being merged in is not integer or is not defined
286 // in the loop, skip it.
287 Value *InVal = PN->getIncomingValue(i);
288 if (!isa<Instruction>(InVal))
291 // If this pred is for a subloop, not L itself, skip it.
292 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
293 continue; // The Block is in a subloop, skip it.
295 // Check that InVal is defined in the loop.
296 Instruction *Inst = cast<Instruction>(InVal);
297 if (!L->contains(Inst))
300 // Okay, this instruction has a user outside of the current loop
301 // and varies predictably *inside* the loop. Evaluate the value it
302 // contains when the loop exits, if possible.
303 const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
304 if (!SE->isLoopInvariant(ExitValue, L))
310 Value *ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), Inst);
312 DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
313 << " LoopVal = " << *Inst << "\n");
315 PN->setIncomingValue(i, ExitVal);
317 // If this instruction is dead now, delete it.
318 RecursivelyDeleteTriviallyDeadInstructions(Inst);
321 // Completely replace a single-pred PHI. This is safe, because the
322 // NewVal won't be variant in the loop, so we don't need an LCSSA phi
324 PN->replaceAllUsesWith(ExitVal);
325 RecursivelyDeleteTriviallyDeadInstructions(PN);
329 // Clone the PHI and delete the original one. This lets IVUsers and
330 // any other maps purge the original user from their records.
331 PHINode *NewPN = cast<PHINode>(PN->clone());
333 NewPN->insertBefore(PN);
334 PN->replaceAllUsesWith(NewPN);
335 PN->eraseFromParent();
340 // The insertion point instruction may have been deleted; clear it out
341 // so that the rewriter doesn't trip over it later.
342 Rewriter.clearInsertPoint();
345 void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
346 // First step. Check to see if there are any floating-point recurrences.
347 // If there are, change them into integer recurrences, permitting analysis by
348 // the SCEV routines.
350 BasicBlock *Header = L->getHeader();
352 SmallVector<WeakVH, 8> PHIs;
353 for (BasicBlock::iterator I = Header->begin();
354 PHINode *PN = dyn_cast<PHINode>(I); ++I)
357 for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
358 if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i]))
359 HandleFloatingPointIV(L, PN);
361 // If the loop previously had floating-point IV, ScalarEvolution
362 // may not have been able to compute a trip count. Now that we've done some
363 // re-writing, the trip count may be computable.
368 void IndVarSimplify::EliminateIVComparisons() {
369 SmallVector<WeakVH, 16> DeadInsts;
371 // Look for ICmp users.
372 for (IVUsers::iterator I = IU->begin(), E = IU->end(); I != E; ++I) {
373 IVStrideUse &UI = *I;
374 ICmpInst *ICmp = dyn_cast<ICmpInst>(UI.getUser());
377 bool Swapped = UI.getOperandValToReplace() == ICmp->getOperand(1);
378 ICmpInst::Predicate Pred = ICmp->getPredicate();
379 if (Swapped) Pred = ICmpInst::getSwappedPredicate(Pred);
381 // Get the SCEVs for the ICmp operands.
382 const SCEV *S = IU->getReplacementExpr(UI);
383 const SCEV *X = SE->getSCEV(ICmp->getOperand(!Swapped));
385 // Simplify unnecessary loops away.
386 const Loop *ICmpLoop = LI->getLoopFor(ICmp->getParent());
387 S = SE->getSCEVAtScope(S, ICmpLoop);
388 X = SE->getSCEVAtScope(X, ICmpLoop);
390 // If the condition is always true or always false, replace it with
392 if (SE->isKnownPredicate(Pred, S, X))
393 ICmp->replaceAllUsesWith(ConstantInt::getTrue(ICmp->getContext()));
394 else if (SE->isKnownPredicate(ICmpInst::getInversePredicate(Pred), S, X))
395 ICmp->replaceAllUsesWith(ConstantInt::getFalse(ICmp->getContext()));
399 DEBUG(dbgs() << "INDVARS: Eliminated comparison: " << *ICmp << '\n');
400 DeadInsts.push_back(ICmp);
403 // Now that we're done iterating through lists, clean up any instructions
404 // which are now dead.
405 while (!DeadInsts.empty())
406 if (Instruction *Inst =
407 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
408 RecursivelyDeleteTriviallyDeadInstructions(Inst);
411 void IndVarSimplify::EliminateIVRemainders() {
412 SmallVector<WeakVH, 16> DeadInsts;
414 // Look for SRem and URem users.
415 for (IVUsers::iterator I = IU->begin(), E = IU->end(); I != E; ++I) {
416 IVStrideUse &UI = *I;
417 BinaryOperator *Rem = dyn_cast<BinaryOperator>(UI.getUser());
420 bool isSigned = Rem->getOpcode() == Instruction::SRem;
421 if (!isSigned && Rem->getOpcode() != Instruction::URem)
424 // We're only interested in the case where we know something about
426 if (UI.getOperandValToReplace() != Rem->getOperand(0))
429 // Get the SCEVs for the ICmp operands.
430 const SCEV *S = SE->getSCEV(Rem->getOperand(0));
431 const SCEV *X = SE->getSCEV(Rem->getOperand(1));
433 // Simplify unnecessary loops away.
434 const Loop *ICmpLoop = LI->getLoopFor(Rem->getParent());
435 S = SE->getSCEVAtScope(S, ICmpLoop);
436 X = SE->getSCEVAtScope(X, ICmpLoop);
438 // i % n --> i if i is in [0,n).
439 if ((!isSigned || SE->isKnownNonNegative(S)) &&
440 SE->isKnownPredicate(isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
442 Rem->replaceAllUsesWith(Rem->getOperand(0));
444 // (i+1) % n --> (i+1)==n?0:(i+1) if i is in [0,n).
445 const SCEV *LessOne =
446 SE->getMinusSCEV(S, SE->getConstant(S->getType(), 1));
447 if ((!isSigned || SE->isKnownNonNegative(LessOne)) &&
448 SE->isKnownPredicate(isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
450 ICmpInst *ICmp = new ICmpInst(Rem, ICmpInst::ICMP_EQ,
451 Rem->getOperand(0), Rem->getOperand(1),
454 SelectInst::Create(ICmp,
455 ConstantInt::get(Rem->getType(), 0),
456 Rem->getOperand(0), "tmp", Rem);
457 Rem->replaceAllUsesWith(Sel);
462 // Inform IVUsers about the new users.
463 if (Instruction *I = dyn_cast<Instruction>(Rem->getOperand(0)))
464 IU->AddUsersIfInteresting(I);
466 DEBUG(dbgs() << "INDVARS: Simplified rem: " << *Rem << '\n');
467 DeadInsts.push_back(Rem);
470 // Now that we're done iterating through lists, clean up any instructions
471 // which are now dead.
472 while (!DeadInsts.empty())
473 if (Instruction *Inst =
474 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
475 RecursivelyDeleteTriviallyDeadInstructions(Inst);
478 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
479 // If LoopSimplify form is not available, stay out of trouble. Some notes:
480 // - LSR currently only supports LoopSimplify-form loops. Indvars'
481 // canonicalization can be a pessimization without LSR to "clean up"
483 // - We depend on having a preheader; in particular,
484 // Loop::getCanonicalInductionVariable only supports loops with preheaders,
485 // and we're in trouble if we can't find the induction variable even when
486 // we've manually inserted one.
487 if (!L->isLoopSimplifyForm())
490 IU = &getAnalysis<IVUsers>();
491 LI = &getAnalysis<LoopInfo>();
492 SE = &getAnalysis<ScalarEvolution>();
493 DT = &getAnalysis<DominatorTree>();
496 // If there are any floating-point recurrences, attempt to
497 // transform them to use integer recurrences.
498 RewriteNonIntegerIVs(L);
500 BasicBlock *ExitingBlock = L->getExitingBlock(); // may be null
501 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
503 // Create a rewriter object which we'll use to transform the code with.
504 SCEVExpander Rewriter(*SE);
506 // Check to see if this loop has a computable loop-invariant execution count.
507 // If so, this means that we can compute the final value of any expressions
508 // that are recurrent in the loop, and substitute the exit values from the
509 // loop into any instructions outside of the loop that use the final values of
510 // the current expressions.
512 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount))
513 RewriteLoopExitValues(L, Rewriter);
515 // Simplify ICmp IV users.
516 EliminateIVComparisons();
518 // Simplify SRem and URem IV users.
519 EliminateIVRemainders();
521 // Compute the type of the largest recurrence expression, and decide whether
522 // a canonical induction variable should be inserted.
523 const Type *LargestType = 0;
524 bool NeedCannIV = false;
525 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount)) {
526 LargestType = BackedgeTakenCount->getType();
527 LargestType = SE->getEffectiveSCEVType(LargestType);
528 // If we have a known trip count and a single exit block, we'll be
529 // rewriting the loop exit test condition below, which requires a
530 // canonical induction variable.
534 for (IVUsers::const_iterator I = IU->begin(), E = IU->end(); I != E; ++I) {
536 SE->getEffectiveSCEVType(I->getOperandValToReplace()->getType());
538 SE->getTypeSizeInBits(Ty) >
539 SE->getTypeSizeInBits(LargestType))
544 // Now that we know the largest of the induction variable expressions
545 // in this loop, insert a canonical induction variable of the largest size.
548 // Check to see if the loop already has any canonical-looking induction
549 // variables. If any are present and wider than the planned canonical
550 // induction variable, temporarily remove them, so that the Rewriter
551 // doesn't attempt to reuse them.
552 SmallVector<PHINode *, 2> OldCannIVs;
553 while (PHINode *OldCannIV = L->getCanonicalInductionVariable()) {
554 if (SE->getTypeSizeInBits(OldCannIV->getType()) >
555 SE->getTypeSizeInBits(LargestType))
556 OldCannIV->removeFromParent();
559 OldCannIVs.push_back(OldCannIV);
562 IndVar = Rewriter.getOrInsertCanonicalInductionVariable(L, LargestType);
566 DEBUG(dbgs() << "INDVARS: New CanIV: " << *IndVar << '\n');
568 // Now that the official induction variable is established, reinsert
569 // any old canonical-looking variables after it so that the IR remains
570 // consistent. They will be deleted as part of the dead-PHI deletion at
571 // the end of the pass.
572 while (!OldCannIVs.empty()) {
573 PHINode *OldCannIV = OldCannIVs.pop_back_val();
574 OldCannIV->insertBefore(L->getHeader()->getFirstNonPHI());
578 // If we have a trip count expression, rewrite the loop's exit condition
579 // using it. We can currently only handle loops with a single exit.
580 ICmpInst *NewICmp = 0;
581 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount) &&
582 !BackedgeTakenCount->isZero() &&
585 "LinearFunctionTestReplace requires a canonical induction variable");
586 // Can't rewrite non-branch yet.
587 if (BranchInst *BI = dyn_cast<BranchInst>(ExitingBlock->getTerminator()))
588 NewICmp = LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar,
589 ExitingBlock, BI, Rewriter);
592 // Rewrite IV-derived expressions. Clears the rewriter cache.
593 RewriteIVExpressions(L, Rewriter);
595 // The Rewriter may not be used from this point on.
597 // Loop-invariant instructions in the preheader that aren't used in the
598 // loop may be sunk below the loop to reduce register pressure.
599 SinkUnusedInvariants(L);
601 // For completeness, inform IVUsers of the IV use in the newly-created
602 // loop exit test instruction.
604 IU->AddUsersIfInteresting(cast<Instruction>(NewICmp->getOperand(0)));
606 // Clean up dead instructions.
607 Changed |= DeleteDeadPHIs(L->getHeader());
608 // Check a post-condition.
609 assert(L->isLCSSAForm(*DT) && "Indvars did not leave the loop in lcssa form!");
613 // FIXME: It is an extremely bad idea to indvar substitute anything more
614 // complex than affine induction variables. Doing so will put expensive
615 // polynomial evaluations inside of the loop, and the str reduction pass
616 // currently can only reduce affine polynomials. For now just disable
617 // indvar subst on anything more complex than an affine addrec, unless
618 // it can be expanded to a trivial value.
619 static bool isSafe(const SCEV *S, const Loop *L, ScalarEvolution *SE) {
620 // Loop-invariant values are safe.
621 if (SE->isLoopInvariant(S, L)) return true;
623 // Affine addrecs are safe. Non-affine are not, because LSR doesn't know how
624 // to transform them into efficient code.
625 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
626 return AR->isAffine();
628 // An add is safe it all its operands are safe.
629 if (const SCEVCommutativeExpr *Commutative = dyn_cast<SCEVCommutativeExpr>(S)) {
630 for (SCEVCommutativeExpr::op_iterator I = Commutative->op_begin(),
631 E = Commutative->op_end(); I != E; ++I)
632 if (!isSafe(*I, L, SE)) return false;
636 // A cast is safe if its operand is.
637 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
638 return isSafe(C->getOperand(), L, SE);
640 // A udiv is safe if its operands are.
641 if (const SCEVUDivExpr *UD = dyn_cast<SCEVUDivExpr>(S))
642 return isSafe(UD->getLHS(), L, SE) &&
643 isSafe(UD->getRHS(), L, SE);
645 // SCEVUnknown is always safe.
646 if (isa<SCEVUnknown>(S))
649 // Nothing else is safe.
653 void IndVarSimplify::RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter) {
654 SmallVector<WeakVH, 16> DeadInsts;
656 // Rewrite all induction variable expressions in terms of the canonical
657 // induction variable.
659 // If there were induction variables of other sizes or offsets, manually
660 // add the offsets to the primary induction variable and cast, avoiding
661 // the need for the code evaluation methods to insert induction variables
662 // of different sizes.
663 for (IVUsers::iterator UI = IU->begin(), E = IU->end(); UI != E; ++UI) {
664 Value *Op = UI->getOperandValToReplace();
665 const Type *UseTy = Op->getType();
666 Instruction *User = UI->getUser();
668 // Compute the final addrec to expand into code.
669 const SCEV *AR = IU->getReplacementExpr(*UI);
671 // Evaluate the expression out of the loop, if possible.
672 if (!L->contains(UI->getUser())) {
673 const SCEV *ExitVal = SE->getSCEVAtScope(AR, L->getParentLoop());
674 if (SE->isLoopInvariant(ExitVal, L))
678 // FIXME: It is an extremely bad idea to indvar substitute anything more
679 // complex than affine induction variables. Doing so will put expensive
680 // polynomial evaluations inside of the loop, and the str reduction pass
681 // currently can only reduce affine polynomials. For now just disable
682 // indvar subst on anything more complex than an affine addrec, unless
683 // it can be expanded to a trivial value.
684 if (!isSafe(AR, L, SE))
687 // Determine the insertion point for this user. By default, insert
688 // immediately before the user. The SCEVExpander class will automatically
689 // hoist loop invariants out of the loop. For PHI nodes, there may be
690 // multiple uses, so compute the nearest common dominator for the
692 Instruction *InsertPt = User;
693 if (PHINode *PHI = dyn_cast<PHINode>(InsertPt))
694 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
695 if (PHI->getIncomingValue(i) == Op) {
696 if (InsertPt == User)
697 InsertPt = PHI->getIncomingBlock(i)->getTerminator();
700 DT->findNearestCommonDominator(InsertPt->getParent(),
701 PHI->getIncomingBlock(i))
705 // Now expand it into actual Instructions and patch it into place.
706 Value *NewVal = Rewriter.expandCodeFor(AR, UseTy, InsertPt);
708 // Inform ScalarEvolution that this value is changing. The change doesn't
709 // affect its value, but it does potentially affect which use lists the
710 // value will be on after the replacement, which affects ScalarEvolution's
711 // ability to walk use lists and drop dangling pointers when a value is
713 SE->forgetValue(User);
715 // Patch the new value into place.
717 NewVal->takeName(Op);
718 User->replaceUsesOfWith(Op, NewVal);
719 UI->setOperandValToReplace(NewVal);
720 DEBUG(dbgs() << "INDVARS: Rewrote IV '" << *AR << "' " << *Op << '\n'
721 << " into = " << *NewVal << "\n");
725 // The old value may be dead now.
726 DeadInsts.push_back(Op);
729 // Clear the rewriter cache, because values that are in the rewriter's cache
730 // can be deleted in the loop below, causing the AssertingVH in the cache to
733 // Now that we're done iterating through lists, clean up any instructions
734 // which are now dead.
735 while (!DeadInsts.empty())
736 if (Instruction *Inst =
737 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
738 RecursivelyDeleteTriviallyDeadInstructions(Inst);
741 /// If there's a single exit block, sink any loop-invariant values that
742 /// were defined in the preheader but not used inside the loop into the
743 /// exit block to reduce register pressure in the loop.
744 void IndVarSimplify::SinkUnusedInvariants(Loop *L) {
745 BasicBlock *ExitBlock = L->getExitBlock();
746 if (!ExitBlock) return;
748 BasicBlock *Preheader = L->getLoopPreheader();
749 if (!Preheader) return;
751 Instruction *InsertPt = ExitBlock->getFirstNonPHI();
752 BasicBlock::iterator I = Preheader->getTerminator();
753 while (I != Preheader->begin()) {
755 // New instructions were inserted at the end of the preheader.
759 // Don't move instructions which might have side effects, since the side
760 // effects need to complete before instructions inside the loop. Also don't
761 // move instructions which might read memory, since the loop may modify
762 // memory. Note that it's okay if the instruction might have undefined
763 // behavior: LoopSimplify guarantees that the preheader dominates the exit
765 if (I->mayHaveSideEffects() || I->mayReadFromMemory())
768 // Skip debug info intrinsics.
769 if (isa<DbgInfoIntrinsic>(I))
772 // Don't sink static AllocaInsts out of the entry block, which would
773 // turn them into dynamic allocas!
774 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
775 if (AI->isStaticAlloca())
778 // Determine if there is a use in or before the loop (direct or
780 bool UsedInLoop = false;
781 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
784 BasicBlock *UseBB = cast<Instruction>(U)->getParent();
785 if (PHINode *P = dyn_cast<PHINode>(U)) {
787 PHINode::getIncomingValueNumForOperand(UI.getOperandNo());
788 UseBB = P->getIncomingBlock(i);
790 if (UseBB == Preheader || L->contains(UseBB)) {
796 // If there is, the def must remain in the preheader.
800 // Otherwise, sink it to the exit block.
801 Instruction *ToMove = I;
804 if (I != Preheader->begin()) {
805 // Skip debug info intrinsics.
808 } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin());
810 if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin())
816 ToMove->moveBefore(InsertPt);
822 /// ConvertToSInt - Convert APF to an integer, if possible.
823 static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) {
824 bool isExact = false;
825 if (&APF.getSemantics() == &APFloat::PPCDoubleDouble)
827 // See if we can convert this to an int64_t
829 if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero,
830 &isExact) != APFloat::opOK || !isExact)
836 /// HandleFloatingPointIV - If the loop has floating induction variable
837 /// then insert corresponding integer induction variable if possible.
839 /// for(double i = 0; i < 10000; ++i)
841 /// is converted into
842 /// for(int i = 0; i < 10000; ++i)
845 void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) {
846 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
847 unsigned BackEdge = IncomingEdge^1;
849 // Check incoming value.
850 ConstantFP *InitValueVal =
851 dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));
854 if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
857 // Check IV increment. Reject this PN if increment operation is not
858 // an add or increment value can not be represented by an integer.
859 BinaryOperator *Incr =
860 dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
861 if (Incr == 0 || Incr->getOpcode() != Instruction::FAdd) return;
863 // If this is not an add of the PHI with a constantfp, or if the constant fp
864 // is not an integer, bail out.
865 ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
867 if (IncValueVal == 0 || Incr->getOperand(0) != PN ||
868 !ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
871 // Check Incr uses. One user is PN and the other user is an exit condition
872 // used by the conditional terminator.
873 Value::use_iterator IncrUse = Incr->use_begin();
874 Instruction *U1 = cast<Instruction>(*IncrUse++);
875 if (IncrUse == Incr->use_end()) return;
876 Instruction *U2 = cast<Instruction>(*IncrUse++);
877 if (IncrUse != Incr->use_end()) return;
879 // Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't
880 // only used by a branch, we can't transform it.
881 FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
883 Compare = dyn_cast<FCmpInst>(U2);
884 if (Compare == 0 || !Compare->hasOneUse() ||
885 !isa<BranchInst>(Compare->use_back()))
888 BranchInst *TheBr = cast<BranchInst>(Compare->use_back());
890 // We need to verify that the branch actually controls the iteration count
891 // of the loop. If not, the new IV can overflow and no one will notice.
892 // The branch block must be in the loop and one of the successors must be out
894 assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
895 if (!L->contains(TheBr->getParent()) ||
896 (L->contains(TheBr->getSuccessor(0)) &&
897 L->contains(TheBr->getSuccessor(1))))
901 // If it isn't a comparison with an integer-as-fp (the exit value), we can't
903 ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
905 if (ExitValueVal == 0 ||
906 !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
909 // Find new predicate for integer comparison.
910 CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
911 switch (Compare->getPredicate()) {
912 default: return; // Unknown comparison.
913 case CmpInst::FCMP_OEQ:
914 case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
915 case CmpInst::FCMP_ONE:
916 case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
917 case CmpInst::FCMP_OGT:
918 case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
919 case CmpInst::FCMP_OGE:
920 case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
921 case CmpInst::FCMP_OLT:
922 case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
923 case CmpInst::FCMP_OLE:
924 case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
927 // We convert the floating point induction variable to a signed i32 value if
928 // we can. This is only safe if the comparison will not overflow in a way
929 // that won't be trapped by the integer equivalent operations. Check for this
931 // TODO: We could use i64 if it is native and the range requires it.
933 // The start/stride/exit values must all fit in signed i32.
934 if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
937 // If not actually striding (add x, 0.0), avoid touching the code.
941 // Positive and negative strides have different safety conditions.
943 // If we have a positive stride, we require the init to be less than the
944 // exit value and an equality or less than comparison.
945 if (InitValue >= ExitValue ||
946 NewPred == CmpInst::ICMP_SGT || NewPred == CmpInst::ICMP_SGE)
949 uint32_t Range = uint32_t(ExitValue-InitValue);
950 if (NewPred == CmpInst::ICMP_SLE) {
951 // Normalize SLE -> SLT, check for infinite loop.
952 if (++Range == 0) return; // Range overflows.
955 unsigned Leftover = Range % uint32_t(IncValue);
957 // If this is an equality comparison, we require that the strided value
958 // exactly land on the exit value, otherwise the IV condition will wrap
959 // around and do things the fp IV wouldn't.
960 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
964 // If the stride would wrap around the i32 before exiting, we can't
966 if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
970 // If we have a negative stride, we require the init to be greater than the
971 // exit value and an equality or greater than comparison.
972 if (InitValue >= ExitValue ||
973 NewPred == CmpInst::ICMP_SLT || NewPred == CmpInst::ICMP_SLE)
976 uint32_t Range = uint32_t(InitValue-ExitValue);
977 if (NewPred == CmpInst::ICMP_SGE) {
978 // Normalize SGE -> SGT, check for infinite loop.
979 if (++Range == 0) return; // Range overflows.
982 unsigned Leftover = Range % uint32_t(-IncValue);
984 // If this is an equality comparison, we require that the strided value
985 // exactly land on the exit value, otherwise the IV condition will wrap
986 // around and do things the fp IV wouldn't.
987 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
991 // If the stride would wrap around the i32 before exiting, we can't
993 if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
997 const IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
999 // Insert new integer induction variable.
1000 PHINode *NewPHI = PHINode::Create(Int32Ty, PN->getName()+".int", PN);
1001 NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
1002 PN->getIncomingBlock(IncomingEdge));
1005 BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
1006 Incr->getName()+".int", Incr);
1007 NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
1009 ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
1010 ConstantInt::get(Int32Ty, ExitValue),
1011 Compare->getName());
1013 // In the following deletions, PN may become dead and may be deleted.
1014 // Use a WeakVH to observe whether this happens.
1017 // Delete the old floating point exit comparison. The branch starts using the
1019 NewCompare->takeName(Compare);
1020 Compare->replaceAllUsesWith(NewCompare);
1021 RecursivelyDeleteTriviallyDeadInstructions(Compare);
1023 // Delete the old floating point increment.
1024 Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
1025 RecursivelyDeleteTriviallyDeadInstructions(Incr);
1027 // If the FP induction variable still has uses, this is because something else
1028 // in the loop uses its value. In order to canonicalize the induction
1029 // variable, we chose to eliminate the IV and rewrite it in terms of an
1032 // We give preference to sitofp over uitofp because it is faster on most
1035 Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
1036 PN->getParent()->getFirstNonPHI());
1037 PN->replaceAllUsesWith(Conv);
1038 RecursivelyDeleteTriviallyDeadInstructions(PN);
1041 // Add a new IVUsers entry for the newly-created integer PHI.
1042 IU->AddUsersIfInteresting(NewPHI);