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) {}
82 virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
84 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
85 AU.addRequired<DominatorTree>();
86 AU.addRequired<LoopInfo>();
87 AU.addRequired<ScalarEvolution>();
88 AU.addRequiredID(LoopSimplifyID);
89 AU.addRequiredID(LCSSAID);
90 AU.addRequired<IVUsers>();
91 AU.addPreserved<ScalarEvolution>();
92 AU.addPreservedID(LoopSimplifyID);
93 AU.addPreservedID(LCSSAID);
94 AU.addPreserved<IVUsers>();
100 void EliminateIVComparisons();
101 void EliminateIVRemainders();
102 void RewriteNonIntegerIVs(Loop *L);
104 ICmpInst *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
106 BasicBlock *ExitingBlock,
108 SCEVExpander &Rewriter);
109 void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter);
111 void RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter);
113 void SinkUnusedInvariants(Loop *L);
115 void HandleFloatingPointIV(Loop *L, PHINode *PH);
119 char IndVarSimplify::ID = 0;
120 INITIALIZE_PASS(IndVarSimplify, "indvars",
121 "Canonicalize Induction Variables", false, false);
123 Pass *llvm::createIndVarSimplifyPass() {
124 return new IndVarSimplify();
127 /// LinearFunctionTestReplace - This method rewrites the exit condition of the
128 /// loop to be a canonical != comparison against the incremented loop induction
129 /// variable. This pass is able to rewrite the exit tests of any loop where the
130 /// SCEV analysis can determine a loop-invariant trip count of the loop, which
131 /// is actually a much broader range than just linear tests.
132 ICmpInst *IndVarSimplify::LinearFunctionTestReplace(Loop *L,
133 const SCEV *BackedgeTakenCount,
135 BasicBlock *ExitingBlock,
137 SCEVExpander &Rewriter) {
138 // Special case: If the backedge-taken count is a UDiv, it's very likely a
139 // UDiv that ScalarEvolution produced in order to compute a precise
140 // expression, rather than a UDiv from the user's code. If we can't find a
141 // UDiv in the code with some simple searching, assume the former and forego
142 // rewriting the loop.
143 if (isa<SCEVUDivExpr>(BackedgeTakenCount)) {
144 ICmpInst *OrigCond = dyn_cast<ICmpInst>(BI->getCondition());
145 if (!OrigCond) return 0;
146 const SCEV *R = SE->getSCEV(OrigCond->getOperand(1));
147 R = SE->getMinusSCEV(R, SE->getConstant(R->getType(), 1));
148 if (R != BackedgeTakenCount) {
149 const SCEV *L = SE->getSCEV(OrigCond->getOperand(0));
150 L = SE->getMinusSCEV(L, SE->getConstant(L->getType(), 1));
151 if (L != BackedgeTakenCount)
156 // If the exiting block is not the same as the backedge block, we must compare
157 // against the preincremented value, otherwise we prefer to compare against
158 // the post-incremented value.
160 const SCEV *RHS = BackedgeTakenCount;
161 if (ExitingBlock == L->getLoopLatch()) {
162 // Add one to the "backedge-taken" count to get the trip count.
163 // If this addition may overflow, we have to be more pessimistic and
164 // cast the induction variable before doing the add.
165 const SCEV *Zero = SE->getConstant(BackedgeTakenCount->getType(), 0);
167 SE->getAddExpr(BackedgeTakenCount,
168 SE->getConstant(BackedgeTakenCount->getType(), 1));
169 if ((isa<SCEVConstant>(N) && !N->isZero()) ||
170 SE->isLoopEntryGuardedByCond(L, ICmpInst::ICMP_NE, N, Zero)) {
171 // No overflow. Cast the sum.
172 RHS = SE->getTruncateOrZeroExtend(N, IndVar->getType());
174 // Potential overflow. Cast before doing the add.
175 RHS = SE->getTruncateOrZeroExtend(BackedgeTakenCount,
177 RHS = SE->getAddExpr(RHS,
178 SE->getConstant(IndVar->getType(), 1));
181 // The BackedgeTaken expression contains the number of times that the
182 // backedge branches to the loop header. This is one less than the
183 // number of times the loop executes, so use the incremented indvar.
184 CmpIndVar = IndVar->getIncomingValueForBlock(ExitingBlock);
186 // We have to use the preincremented value...
187 RHS = SE->getTruncateOrZeroExtend(BackedgeTakenCount,
192 // Expand the code for the iteration count.
193 assert(RHS->isLoopInvariant(L) &&
194 "Computed iteration count is not loop invariant!");
195 Value *ExitCnt = Rewriter.expandCodeFor(RHS, IndVar->getType(), BI);
197 // Insert a new icmp_ne or icmp_eq instruction before the branch.
198 ICmpInst::Predicate Opcode;
199 if (L->contains(BI->getSuccessor(0)))
200 Opcode = ICmpInst::ICMP_NE;
202 Opcode = ICmpInst::ICMP_EQ;
204 DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
205 << " LHS:" << *CmpIndVar << '\n'
207 << (Opcode == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
208 << " RHS:\t" << *RHS << "\n");
210 ICmpInst *Cond = new ICmpInst(BI, Opcode, CmpIndVar, ExitCnt, "exitcond");
212 Value *OrigCond = BI->getCondition();
213 // It's tempting to use replaceAllUsesWith here to fully replace the old
214 // comparison, but that's not immediately safe, since users of the old
215 // comparison may not be dominated by the new comparison. Instead, just
216 // update the branch to use the new comparison; in the common case this
217 // will make old comparison dead.
218 BI->setCondition(Cond);
219 RecursivelyDeleteTriviallyDeadInstructions(OrigCond);
226 /// RewriteLoopExitValues - Check to see if this loop has a computable
227 /// loop-invariant execution count. If so, this means that we can compute the
228 /// final value of any expressions that are recurrent in the loop, and
229 /// substitute the exit values from the loop into any instructions outside of
230 /// the loop that use the final values of the current expressions.
232 /// This is mostly redundant with the regular IndVarSimplify activities that
233 /// happen later, except that it's more powerful in some cases, because it's
234 /// able to brute-force evaluate arbitrary instructions as long as they have
235 /// constant operands at the beginning of the loop.
236 void IndVarSimplify::RewriteLoopExitValues(Loop *L,
237 SCEVExpander &Rewriter) {
238 // Verify the input to the pass in already in LCSSA form.
239 assert(L->isLCSSAForm(*DT));
241 SmallVector<BasicBlock*, 8> ExitBlocks;
242 L->getUniqueExitBlocks(ExitBlocks);
244 // Find all values that are computed inside the loop, but used outside of it.
245 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan
246 // the exit blocks of the loop to find them.
247 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
248 BasicBlock *ExitBB = ExitBlocks[i];
250 // If there are no PHI nodes in this exit block, then no values defined
251 // inside the loop are used on this path, skip it.
252 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
255 unsigned NumPreds = PN->getNumIncomingValues();
257 // Iterate over all of the PHI nodes.
258 BasicBlock::iterator BBI = ExitBB->begin();
259 while ((PN = dyn_cast<PHINode>(BBI++))) {
261 continue; // dead use, don't replace it
263 // SCEV only supports integer expressions for now.
264 if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy())
267 // It's necessary to tell ScalarEvolution about this explicitly so that
268 // it can walk the def-use list and forget all SCEVs, as it may not be
269 // watching the PHI itself. Once the new exit value is in place, there
270 // may not be a def-use connection between the loop and every instruction
271 // which got a SCEVAddRecExpr for that loop.
274 // Iterate over all of the values in all the PHI nodes.
275 for (unsigned i = 0; i != NumPreds; ++i) {
276 // If the value being merged in is not integer or is not defined
277 // in the loop, skip it.
278 Value *InVal = PN->getIncomingValue(i);
279 if (!isa<Instruction>(InVal))
282 // If this pred is for a subloop, not L itself, skip it.
283 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
284 continue; // The Block is in a subloop, skip it.
286 // Check that InVal is defined in the loop.
287 Instruction *Inst = cast<Instruction>(InVal);
288 if (!L->contains(Inst))
291 // Okay, this instruction has a user outside of the current loop
292 // and varies predictably *inside* the loop. Evaluate the value it
293 // contains when the loop exits, if possible.
294 const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
295 if (!ExitValue->isLoopInvariant(L))
301 Value *ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), Inst);
303 DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
304 << " LoopVal = " << *Inst << "\n");
306 PN->setIncomingValue(i, ExitVal);
308 // If this instruction is dead now, delete it.
309 RecursivelyDeleteTriviallyDeadInstructions(Inst);
312 // Completely replace a single-pred PHI. This is safe, because the
313 // NewVal won't be variant in the loop, so we don't need an LCSSA phi
315 PN->replaceAllUsesWith(ExitVal);
316 RecursivelyDeleteTriviallyDeadInstructions(PN);
320 // Clone the PHI and delete the original one. This lets IVUsers and
321 // any other maps purge the original user from their records.
322 PHINode *NewPN = cast<PHINode>(PN->clone());
324 NewPN->insertBefore(PN);
325 PN->replaceAllUsesWith(NewPN);
326 PN->eraseFromParent();
331 // The insertion point instruction may have been deleted; clear it out
332 // so that the rewriter doesn't trip over it later.
333 Rewriter.clearInsertPoint();
336 void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
337 // First step. Check to see if there are any floating-point recurrences.
338 // If there are, change them into integer recurrences, permitting analysis by
339 // the SCEV routines.
341 BasicBlock *Header = L->getHeader();
343 SmallVector<WeakVH, 8> PHIs;
344 for (BasicBlock::iterator I = Header->begin();
345 PHINode *PN = dyn_cast<PHINode>(I); ++I)
348 for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
349 if (PHINode *PN = dyn_cast_or_null<PHINode>(PHIs[i]))
350 HandleFloatingPointIV(L, PN);
352 // If the loop previously had floating-point IV, ScalarEvolution
353 // may not have been able to compute a trip count. Now that we've done some
354 // re-writing, the trip count may be computable.
359 void IndVarSimplify::EliminateIVComparisons() {
360 SmallVector<WeakVH, 16> DeadInsts;
362 // Look for ICmp users.
363 for (IVUsers::iterator I = IU->begin(), E = IU->end(); I != E; ++I) {
364 IVStrideUse &UI = *I;
365 ICmpInst *ICmp = dyn_cast<ICmpInst>(UI.getUser());
368 bool Swapped = UI.getOperandValToReplace() == ICmp->getOperand(1);
369 ICmpInst::Predicate Pred = ICmp->getPredicate();
370 if (Swapped) Pred = ICmpInst::getSwappedPredicate(Pred);
372 // Get the SCEVs for the ICmp operands.
373 const SCEV *S = IU->getReplacementExpr(UI);
374 const SCEV *X = SE->getSCEV(ICmp->getOperand(!Swapped));
376 // Simplify unnecessary loops away.
377 const Loop *ICmpLoop = LI->getLoopFor(ICmp->getParent());
378 S = SE->getSCEVAtScope(S, ICmpLoop);
379 X = SE->getSCEVAtScope(X, ICmpLoop);
381 // If the condition is always true or always false, replace it with
383 if (SE->isKnownPredicate(Pred, S, X))
384 ICmp->replaceAllUsesWith(ConstantInt::getTrue(ICmp->getContext()));
385 else if (SE->isKnownPredicate(ICmpInst::getInversePredicate(Pred), S, X))
386 ICmp->replaceAllUsesWith(ConstantInt::getFalse(ICmp->getContext()));
390 DEBUG(dbgs() << "INDVARS: Eliminated comparison: " << *ICmp << '\n');
391 DeadInsts.push_back(ICmp);
394 // Now that we're done iterating through lists, clean up any instructions
395 // which are now dead.
396 while (!DeadInsts.empty())
397 if (Instruction *Inst =
398 dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val()))
399 RecursivelyDeleteTriviallyDeadInstructions(Inst);
402 void IndVarSimplify::EliminateIVRemainders() {
403 SmallVector<WeakVH, 16> DeadInsts;
405 // Look for SRem and URem users.
406 for (IVUsers::iterator I = IU->begin(), E = IU->end(); I != E; ++I) {
407 IVStrideUse &UI = *I;
408 BinaryOperator *Rem = dyn_cast<BinaryOperator>(UI.getUser());
411 bool isSigned = Rem->getOpcode() == Instruction::SRem;
412 if (!isSigned && Rem->getOpcode() != Instruction::URem)
415 // We're only interested in the case where we know something about
417 if (UI.getOperandValToReplace() != Rem->getOperand(0))
420 // Get the SCEVs for the ICmp operands.
421 const SCEV *S = SE->getSCEV(Rem->getOperand(0));
422 const SCEV *X = SE->getSCEV(Rem->getOperand(1));
424 // Simplify unnecessary loops away.
425 const Loop *ICmpLoop = LI->getLoopFor(Rem->getParent());
426 S = SE->getSCEVAtScope(S, ICmpLoop);
427 X = SE->getSCEVAtScope(X, ICmpLoop);
429 // i % n --> i if i is in [0,n).
430 if ((!isSigned || SE->isKnownNonNegative(S)) &&
431 SE->isKnownPredicate(isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
433 Rem->replaceAllUsesWith(Rem->getOperand(0));
435 // (i+1) % n --> (i+1)==n?0:(i+1) if i is in [0,n).
436 const SCEV *LessOne =
437 SE->getMinusSCEV(S, SE->getConstant(S->getType(), 1));
438 if ((!isSigned || SE->isKnownNonNegative(LessOne)) &&
439 SE->isKnownPredicate(isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
441 ICmpInst *ICmp = new ICmpInst(Rem, ICmpInst::ICMP_EQ,
442 Rem->getOperand(0), Rem->getOperand(1),
445 SelectInst::Create(ICmp,
446 ConstantInt::get(Rem->getType(), 0),
447 Rem->getOperand(0), "tmp", Rem);
448 Rem->replaceAllUsesWith(Sel);
453 // Inform IVUsers about the new users.
454 if (Instruction *I = dyn_cast<Instruction>(Rem->getOperand(0)))
455 IU->AddUsersIfInteresting(I);
457 DEBUG(dbgs() << "INDVARS: Simplified rem: " << *Rem << '\n');
458 DeadInsts.push_back(Rem);
461 // Now that we're done iterating through lists, clean up any instructions
462 // which are now dead.
463 while (!DeadInsts.empty())
464 if (Instruction *Inst =
465 dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val()))
466 RecursivelyDeleteTriviallyDeadInstructions(Inst);
469 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
470 // If LoopSimplify form is not available, stay out of trouble. Some notes:
471 // - LSR currently only supports LoopSimplify-form loops. Indvars'
472 // canonicalization can be a pessimization without LSR to "clean up"
474 // - We depend on having a preheader; in particular,
475 // Loop::getCanonicalInductionVariable only supports loops with preheaders,
476 // and we're in trouble if we can't find the induction variable even when
477 // we've manually inserted one.
478 if (!L->isLoopSimplifyForm())
481 IU = &getAnalysis<IVUsers>();
482 LI = &getAnalysis<LoopInfo>();
483 SE = &getAnalysis<ScalarEvolution>();
484 DT = &getAnalysis<DominatorTree>();
487 // If there are any floating-point recurrences, attempt to
488 // transform them to use integer recurrences.
489 RewriteNonIntegerIVs(L);
491 BasicBlock *ExitingBlock = L->getExitingBlock(); // may be null
492 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
494 // Create a rewriter object which we'll use to transform the code with.
495 SCEVExpander Rewriter(*SE);
497 // Check to see if this loop has a computable loop-invariant execution count.
498 // If so, this means that we can compute the final value of any expressions
499 // that are recurrent in the loop, and substitute the exit values from the
500 // loop into any instructions outside of the loop that use the final values of
501 // the current expressions.
503 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount))
504 RewriteLoopExitValues(L, Rewriter);
506 // Simplify ICmp IV users.
507 EliminateIVComparisons();
509 // Simplify SRem and URem IV users.
510 EliminateIVRemainders();
512 // Compute the type of the largest recurrence expression, and decide whether
513 // a canonical induction variable should be inserted.
514 const Type *LargestType = 0;
515 bool NeedCannIV = false;
516 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount)) {
517 LargestType = BackedgeTakenCount->getType();
518 LargestType = SE->getEffectiveSCEVType(LargestType);
519 // If we have a known trip count and a single exit block, we'll be
520 // rewriting the loop exit test condition below, which requires a
521 // canonical induction variable.
525 for (IVUsers::const_iterator I = IU->begin(), E = IU->end(); I != E; ++I) {
527 SE->getEffectiveSCEVType(I->getOperandValToReplace()->getType());
529 SE->getTypeSizeInBits(Ty) >
530 SE->getTypeSizeInBits(LargestType))
535 // Now that we know the largest of the induction variable expressions
536 // in this loop, insert a canonical induction variable of the largest size.
539 // Check to see if the loop already has any canonical-looking induction
540 // variables. If any are present and wider than the planned canonical
541 // induction variable, temporarily remove them, so that the Rewriter
542 // doesn't attempt to reuse them.
543 SmallVector<PHINode *, 2> OldCannIVs;
544 while (PHINode *OldCannIV = L->getCanonicalInductionVariable()) {
545 if (SE->getTypeSizeInBits(OldCannIV->getType()) >
546 SE->getTypeSizeInBits(LargestType))
547 OldCannIV->removeFromParent();
550 OldCannIVs.push_back(OldCannIV);
553 IndVar = Rewriter.getOrInsertCanonicalInductionVariable(L, LargestType);
557 DEBUG(dbgs() << "INDVARS: New CanIV: " << *IndVar << '\n');
559 // Now that the official induction variable is established, reinsert
560 // any old canonical-looking variables after it so that the IR remains
561 // consistent. They will be deleted as part of the dead-PHI deletion at
562 // the end of the pass.
563 while (!OldCannIVs.empty()) {
564 PHINode *OldCannIV = OldCannIVs.pop_back_val();
565 OldCannIV->insertBefore(L->getHeader()->getFirstNonPHI());
569 // If we have a trip count expression, rewrite the loop's exit condition
570 // using it. We can currently only handle loops with a single exit.
571 ICmpInst *NewICmp = 0;
572 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount) &&
573 !BackedgeTakenCount->isZero() &&
576 "LinearFunctionTestReplace requires a canonical induction variable");
577 // Can't rewrite non-branch yet.
578 if (BranchInst *BI = dyn_cast<BranchInst>(ExitingBlock->getTerminator()))
579 NewICmp = LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar,
580 ExitingBlock, BI, Rewriter);
583 // Rewrite IV-derived expressions. Clears the rewriter cache.
584 RewriteIVExpressions(L, Rewriter);
586 // The Rewriter may not be used from this point on.
588 // Loop-invariant instructions in the preheader that aren't used in the
589 // loop may be sunk below the loop to reduce register pressure.
590 SinkUnusedInvariants(L);
592 // For completeness, inform IVUsers of the IV use in the newly-created
593 // loop exit test instruction.
595 IU->AddUsersIfInteresting(cast<Instruction>(NewICmp->getOperand(0)));
597 // Clean up dead instructions.
598 Changed |= DeleteDeadPHIs(L->getHeader());
599 // Check a post-condition.
600 assert(L->isLCSSAForm(*DT) && "Indvars did not leave the loop in lcssa form!");
604 // FIXME: It is an extremely bad idea to indvar substitute anything more
605 // complex than affine induction variables. Doing so will put expensive
606 // polynomial evaluations inside of the loop, and the str reduction pass
607 // currently can only reduce affine polynomials. For now just disable
608 // indvar subst on anything more complex than an affine addrec, unless
609 // it can be expanded to a trivial value.
610 static bool isSafe(const SCEV *S, const Loop *L) {
611 // Loop-invariant values are safe.
612 if (S->isLoopInvariant(L)) return true;
614 // Affine addrecs are safe. Non-affine are not, because LSR doesn't know how
615 // to transform them into efficient code.
616 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
617 return AR->isAffine();
619 // An add is safe it all its operands are safe.
620 if (const SCEVCommutativeExpr *Commutative = dyn_cast<SCEVCommutativeExpr>(S)) {
621 for (SCEVCommutativeExpr::op_iterator I = Commutative->op_begin(),
622 E = Commutative->op_end(); I != E; ++I)
623 if (!isSafe(*I, L)) return false;
627 // A cast is safe if its operand is.
628 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
629 return isSafe(C->getOperand(), L);
631 // A udiv is safe if its operands are.
632 if (const SCEVUDivExpr *UD = dyn_cast<SCEVUDivExpr>(S))
633 return isSafe(UD->getLHS(), L) &&
634 isSafe(UD->getRHS(), L);
636 // SCEVUnknown is always safe.
637 if (isa<SCEVUnknown>(S))
640 // Nothing else is safe.
644 void IndVarSimplify::RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter) {
645 SmallVector<WeakVH, 16> DeadInsts;
647 // Rewrite all induction variable expressions in terms of the canonical
648 // induction variable.
650 // If there were induction variables of other sizes or offsets, manually
651 // add the offsets to the primary induction variable and cast, avoiding
652 // the need for the code evaluation methods to insert induction variables
653 // of different sizes.
654 for (IVUsers::iterator UI = IU->begin(), E = IU->end(); UI != E; ++UI) {
655 Value *Op = UI->getOperandValToReplace();
656 const Type *UseTy = Op->getType();
657 Instruction *User = UI->getUser();
659 // Compute the final addrec to expand into code.
660 const SCEV *AR = IU->getReplacementExpr(*UI);
662 // Evaluate the expression out of the loop, if possible.
663 if (!L->contains(UI->getUser())) {
664 const SCEV *ExitVal = SE->getSCEVAtScope(AR, L->getParentLoop());
665 if (ExitVal->isLoopInvariant(L))
669 // FIXME: It is an extremely bad idea to indvar substitute anything more
670 // complex than affine induction variables. Doing so will put expensive
671 // polynomial evaluations inside of the loop, and the str reduction pass
672 // currently can only reduce affine polynomials. For now just disable
673 // indvar subst on anything more complex than an affine addrec, unless
674 // it can be expanded to a trivial value.
678 // Determine the insertion point for this user. By default, insert
679 // immediately before the user. The SCEVExpander class will automatically
680 // hoist loop invariants out of the loop. For PHI nodes, there may be
681 // multiple uses, so compute the nearest common dominator for the
683 Instruction *InsertPt = User;
684 if (PHINode *PHI = dyn_cast<PHINode>(InsertPt))
685 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
686 if (PHI->getIncomingValue(i) == Op) {
687 if (InsertPt == User)
688 InsertPt = PHI->getIncomingBlock(i)->getTerminator();
691 DT->findNearestCommonDominator(InsertPt->getParent(),
692 PHI->getIncomingBlock(i))
696 // Now expand it into actual Instructions and patch it into place.
697 Value *NewVal = Rewriter.expandCodeFor(AR, UseTy, InsertPt);
699 // Inform ScalarEvolution that this value is changing. The change doesn't
700 // affect its value, but it does potentially affect which use lists the
701 // value will be on after the replacement, which affects ScalarEvolution's
702 // ability to walk use lists and drop dangling pointers when a value is
704 SE->forgetValue(User);
706 // Patch the new value into place.
708 NewVal->takeName(Op);
709 User->replaceUsesOfWith(Op, NewVal);
710 UI->setOperandValToReplace(NewVal);
711 DEBUG(dbgs() << "INDVARS: Rewrote IV '" << *AR << "' " << *Op << '\n'
712 << " into = " << *NewVal << "\n");
716 // The old value may be dead now.
717 DeadInsts.push_back(Op);
720 // Clear the rewriter cache, because values that are in the rewriter's cache
721 // can be deleted in the loop below, causing the AssertingVH in the cache to
724 // Now that we're done iterating through lists, clean up any instructions
725 // which are now dead.
726 while (!DeadInsts.empty())
727 if (Instruction *Inst =
728 dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val()))
729 RecursivelyDeleteTriviallyDeadInstructions(Inst);
732 /// If there's a single exit block, sink any loop-invariant values that
733 /// were defined in the preheader but not used inside the loop into the
734 /// exit block to reduce register pressure in the loop.
735 void IndVarSimplify::SinkUnusedInvariants(Loop *L) {
736 BasicBlock *ExitBlock = L->getExitBlock();
737 if (!ExitBlock) return;
739 BasicBlock *Preheader = L->getLoopPreheader();
740 if (!Preheader) return;
742 Instruction *InsertPt = ExitBlock->getFirstNonPHI();
743 BasicBlock::iterator I = Preheader->getTerminator();
744 while (I != Preheader->begin()) {
746 // New instructions were inserted at the end of the preheader.
750 // Don't move instructions which might have side effects, since the side
751 // effects need to complete before instructions inside the loop. Also don't
752 // move instructions which might read memory, since the loop may modify
753 // memory. Note that it's okay if the instruction might have undefined
754 // behavior: LoopSimplify guarantees that the preheader dominates the exit
756 if (I->mayHaveSideEffects() || I->mayReadFromMemory())
759 // Skip debug info intrinsics.
760 if (isa<DbgInfoIntrinsic>(I))
763 // Don't sink static AllocaInsts out of the entry block, which would
764 // turn them into dynamic allocas!
765 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
766 if (AI->isStaticAlloca())
769 // Determine if there is a use in or before the loop (direct or
771 bool UsedInLoop = false;
772 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
775 BasicBlock *UseBB = cast<Instruction>(U)->getParent();
776 if (PHINode *P = dyn_cast<PHINode>(U)) {
778 PHINode::getIncomingValueNumForOperand(UI.getOperandNo());
779 UseBB = P->getIncomingBlock(i);
781 if (UseBB == Preheader || L->contains(UseBB)) {
787 // If there is, the def must remain in the preheader.
791 // Otherwise, sink it to the exit block.
792 Instruction *ToMove = I;
795 if (I != Preheader->begin()) {
796 // Skip debug info intrinsics.
799 } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin());
801 if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin())
807 ToMove->moveBefore(InsertPt);
813 /// ConvertToSInt - Convert APF to an integer, if possible.
814 static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) {
815 bool isExact = false;
816 if (&APF.getSemantics() == &APFloat::PPCDoubleDouble)
818 // See if we can convert this to an int64_t
820 if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero,
821 &isExact) != APFloat::opOK || !isExact)
827 /// HandleFloatingPointIV - If the loop has floating induction variable
828 /// then insert corresponding integer induction variable if possible.
830 /// for(double i = 0; i < 10000; ++i)
832 /// is converted into
833 /// for(int i = 0; i < 10000; ++i)
836 void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) {
837 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
838 unsigned BackEdge = IncomingEdge^1;
840 // Check incoming value.
841 ConstantFP *InitValueVal =
842 dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));
845 if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
848 // Check IV increment. Reject this PN if increment operation is not
849 // an add or increment value can not be represented by an integer.
850 BinaryOperator *Incr =
851 dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
852 if (Incr == 0 || Incr->getOpcode() != Instruction::FAdd) return;
854 // If this is not an add of the PHI with a constantfp, or if the constant fp
855 // is not an integer, bail out.
856 ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
858 if (IncValueVal == 0 || Incr->getOperand(0) != PN ||
859 !ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
862 // Check Incr uses. One user is PN and the other user is an exit condition
863 // used by the conditional terminator.
864 Value::use_iterator IncrUse = Incr->use_begin();
865 Instruction *U1 = cast<Instruction>(*IncrUse++);
866 if (IncrUse == Incr->use_end()) return;
867 Instruction *U2 = cast<Instruction>(*IncrUse++);
868 if (IncrUse != Incr->use_end()) return;
870 // Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't
871 // only used by a branch, we can't transform it.
872 FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
874 Compare = dyn_cast<FCmpInst>(U2);
875 if (Compare == 0 || !Compare->hasOneUse() ||
876 !isa<BranchInst>(Compare->use_back()))
879 BranchInst *TheBr = cast<BranchInst>(Compare->use_back());
881 // We need to verify that the branch actually controls the iteration count
882 // of the loop. If not, the new IV can overflow and no one will notice.
883 // The branch block must be in the loop and one of the successors must be out
885 assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
886 if (!L->contains(TheBr->getParent()) ||
887 (L->contains(TheBr->getSuccessor(0)) &&
888 L->contains(TheBr->getSuccessor(1))))
892 // If it isn't a comparison with an integer-as-fp (the exit value), we can't
894 ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
896 if (ExitValueVal == 0 ||
897 !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
900 // Find new predicate for integer comparison.
901 CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
902 switch (Compare->getPredicate()) {
903 default: return; // Unknown comparison.
904 case CmpInst::FCMP_OEQ:
905 case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
906 case CmpInst::FCMP_ONE:
907 case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
908 case CmpInst::FCMP_OGT:
909 case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
910 case CmpInst::FCMP_OGE:
911 case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
912 case CmpInst::FCMP_OLT:
913 case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
914 case CmpInst::FCMP_OLE:
915 case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
918 // We convert the floating point induction variable to a signed i32 value if
919 // we can. This is only safe if the comparison will not overflow in a way
920 // that won't be trapped by the integer equivalent operations. Check for this
922 // TODO: We could use i64 if it is native and the range requires it.
924 // The start/stride/exit values must all fit in signed i32.
925 if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
928 // If not actually striding (add x, 0.0), avoid touching the code.
932 // Positive and negative strides have different safety conditions.
934 // If we have a positive stride, we require the init to be less than the
935 // exit value and an equality or less than comparison.
936 if (InitValue >= ExitValue ||
937 NewPred == CmpInst::ICMP_SGT || NewPred == CmpInst::ICMP_SGE)
940 uint32_t Range = uint32_t(ExitValue-InitValue);
941 if (NewPred == CmpInst::ICMP_SLE) {
942 // Normalize SLE -> SLT, check for infinite loop.
943 if (++Range == 0) return; // Range overflows.
946 unsigned Leftover = Range % uint32_t(IncValue);
948 // If this is an equality comparison, we require that the strided value
949 // exactly land on the exit value, otherwise the IV condition will wrap
950 // around and do things the fp IV wouldn't.
951 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
955 // If the stride would wrap around the i32 before exiting, we can't
957 if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
961 // If we have a negative stride, we require the init to be greater than the
962 // exit value and an equality or greater than comparison.
963 if (InitValue >= ExitValue ||
964 NewPred == CmpInst::ICMP_SLT || NewPred == CmpInst::ICMP_SLE)
967 uint32_t Range = uint32_t(InitValue-ExitValue);
968 if (NewPred == CmpInst::ICMP_SGE) {
969 // Normalize SGE -> SGT, check for infinite loop.
970 if (++Range == 0) return; // Range overflows.
973 unsigned Leftover = Range % uint32_t(-IncValue);
975 // If this is an equality comparison, we require that the strided value
976 // exactly land on the exit value, otherwise the IV condition will wrap
977 // around and do things the fp IV wouldn't.
978 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
982 // If the stride would wrap around the i32 before exiting, we can't
984 if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
988 const IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
990 // Insert new integer induction variable.
991 PHINode *NewPHI = PHINode::Create(Int32Ty, PN->getName()+".int", PN);
992 NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
993 PN->getIncomingBlock(IncomingEdge));
996 BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
997 Incr->getName()+".int", Incr);
998 NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
1000 ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
1001 ConstantInt::get(Int32Ty, ExitValue),
1002 Compare->getName());
1004 // In the following deletions, PN may become dead and may be deleted.
1005 // Use a WeakVH to observe whether this happens.
1008 // Delete the old floating point exit comparison. The branch starts using the
1010 NewCompare->takeName(Compare);
1011 Compare->replaceAllUsesWith(NewCompare);
1012 RecursivelyDeleteTriviallyDeadInstructions(Compare);
1014 // Delete the old floating point increment.
1015 Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
1016 RecursivelyDeleteTriviallyDeadInstructions(Incr);
1018 // If the FP induction variable still has uses, this is because something else
1019 // in the loop uses its value. In order to canonicalize the induction
1020 // variable, we chose to eliminate the IV and rewrite it in terms of an
1023 // We give preference to sitofp over uitofp because it is faster on most
1026 Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
1027 PN->getParent()->getFirstNonPHI());
1028 PN->replaceAllUsesWith(Conv);
1029 RecursivelyDeleteTriviallyDeadInstructions(PN);
1032 // Add a new IVUsers entry for the newly-created integer PHI.
1033 IU->AddUsersIfInteresting(NewPHI);