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 RewriteNonIntegerIVs(Loop *L);
103 ICmpInst *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
105 BasicBlock *ExitingBlock,
107 SCEVExpander &Rewriter);
108 void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter);
110 void RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter);
112 void SinkUnusedInvariants(Loop *L);
114 void HandleFloatingPointIV(Loop *L, PHINode *PH);
118 char IndVarSimplify::ID = 0;
119 static RegisterPass<IndVarSimplify>
120 X("indvars", "Canonicalize Induction Variables");
122 Pass *llvm::createIndVarSimplifyPass() {
123 return new IndVarSimplify();
126 /// LinearFunctionTestReplace - This method rewrites the exit condition of the
127 /// loop to be a canonical != comparison against the incremented loop induction
128 /// variable. This pass is able to rewrite the exit tests of any loop where the
129 /// SCEV analysis can determine a loop-invariant trip count of the loop, which
130 /// is actually a much broader range than just linear tests.
131 ICmpInst *IndVarSimplify::LinearFunctionTestReplace(Loop *L,
132 const SCEV *BackedgeTakenCount,
134 BasicBlock *ExitingBlock,
136 SCEVExpander &Rewriter) {
137 // If the exiting block is not the same as the backedge block, we must compare
138 // against the preincremented value, otherwise we prefer to compare against
139 // the post-incremented value.
141 const SCEV *RHS = BackedgeTakenCount;
142 if (ExitingBlock == L->getLoopLatch()) {
143 // Add one to the "backedge-taken" count to get the trip count.
144 // If this addition may overflow, we have to be more pessimistic and
145 // cast the induction variable before doing the add.
146 const SCEV *Zero = SE->getIntegerSCEV(0, BackedgeTakenCount->getType());
148 SE->getAddExpr(BackedgeTakenCount,
149 SE->getIntegerSCEV(1, BackedgeTakenCount->getType()));
150 if ((isa<SCEVConstant>(N) && !N->isZero()) ||
151 SE->isLoopEntryGuardedByCond(L, ICmpInst::ICMP_NE, N, Zero)) {
152 // No overflow. Cast the sum.
153 RHS = SE->getTruncateOrZeroExtend(N, IndVar->getType());
155 // Potential overflow. Cast before doing the add.
156 RHS = SE->getTruncateOrZeroExtend(BackedgeTakenCount,
158 RHS = SE->getAddExpr(RHS,
159 SE->getIntegerSCEV(1, IndVar->getType()));
162 // The BackedgeTaken expression contains the number of times that the
163 // backedge branches to the loop header. This is one less than the
164 // number of times the loop executes, so use the incremented indvar.
165 CmpIndVar = L->getCanonicalInductionVariableIncrement();
167 // We have to use the preincremented value...
168 RHS = SE->getTruncateOrZeroExtend(BackedgeTakenCount,
173 // Expand the code for the iteration count.
174 assert(RHS->isLoopInvariant(L) &&
175 "Computed iteration count is not loop invariant!");
176 Value *ExitCnt = Rewriter.expandCodeFor(RHS, IndVar->getType(), BI);
178 // Insert a new icmp_ne or icmp_eq instruction before the branch.
179 ICmpInst::Predicate Opcode;
180 if (L->contains(BI->getSuccessor(0)))
181 Opcode = ICmpInst::ICMP_NE;
183 Opcode = ICmpInst::ICMP_EQ;
185 DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
186 << " LHS:" << *CmpIndVar << '\n'
188 << (Opcode == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
189 << " RHS:\t" << *RHS << "\n");
191 ICmpInst *Cond = new ICmpInst(BI, Opcode, CmpIndVar, ExitCnt, "exitcond");
193 Value *OrigCond = BI->getCondition();
194 // It's tempting to use replaceAllUsesWith here to fully replace the old
195 // comparison, but that's not immediately safe, since users of the old
196 // comparison may not be dominated by the new comparison. Instead, just
197 // update the branch to use the new comparison; in the common case this
198 // will make old comparison dead.
199 BI->setCondition(Cond);
200 RecursivelyDeleteTriviallyDeadInstructions(OrigCond);
207 /// RewriteLoopExitValues - Check to see if this loop has a computable
208 /// loop-invariant execution count. If so, this means that we can compute the
209 /// final value of any expressions that are recurrent in the loop, and
210 /// substitute the exit values from the loop into any instructions outside of
211 /// the loop that use the final values of the current expressions.
213 /// This is mostly redundant with the regular IndVarSimplify activities that
214 /// happen later, except that it's more powerful in some cases, because it's
215 /// able to brute-force evaluate arbitrary instructions as long as they have
216 /// constant operands at the beginning of the loop.
217 void IndVarSimplify::RewriteLoopExitValues(Loop *L,
218 SCEVExpander &Rewriter) {
219 // Verify the input to the pass in already in LCSSA form.
220 assert(L->isLCSSAForm(*DT));
222 SmallVector<BasicBlock*, 8> ExitBlocks;
223 L->getUniqueExitBlocks(ExitBlocks);
225 // Find all values that are computed inside the loop, but used outside of it.
226 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan
227 // the exit blocks of the loop to find them.
228 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
229 BasicBlock *ExitBB = ExitBlocks[i];
231 // If there are no PHI nodes in this exit block, then no values defined
232 // inside the loop are used on this path, skip it.
233 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
236 unsigned NumPreds = PN->getNumIncomingValues();
238 // Iterate over all of the PHI nodes.
239 BasicBlock::iterator BBI = ExitBB->begin();
240 while ((PN = dyn_cast<PHINode>(BBI++))) {
242 continue; // dead use, don't replace it
244 // SCEV only supports integer expressions for now.
245 if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy())
248 // It's necessary to tell ScalarEvolution about this explicitly so that
249 // it can walk the def-use list and forget all SCEVs, as it may not be
250 // watching the PHI itself. Once the new exit value is in place, there
251 // may not be a def-use connection between the loop and every instruction
252 // which got a SCEVAddRecExpr for that loop.
255 // Iterate over all of the values in all the PHI nodes.
256 for (unsigned i = 0; i != NumPreds; ++i) {
257 // If the value being merged in is not integer or is not defined
258 // in the loop, skip it.
259 Value *InVal = PN->getIncomingValue(i);
260 if (!isa<Instruction>(InVal))
263 // If this pred is for a subloop, not L itself, skip it.
264 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
265 continue; // The Block is in a subloop, skip it.
267 // Check that InVal is defined in the loop.
268 Instruction *Inst = cast<Instruction>(InVal);
269 if (!L->contains(Inst))
272 // Okay, this instruction has a user outside of the current loop
273 // and varies predictably *inside* the loop. Evaluate the value it
274 // contains when the loop exits, if possible.
275 const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
276 if (!ExitValue->isLoopInvariant(L))
282 Value *ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), Inst);
284 DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
285 << " LoopVal = " << *Inst << "\n");
287 PN->setIncomingValue(i, ExitVal);
289 // If this instruction is dead now, delete it.
290 RecursivelyDeleteTriviallyDeadInstructions(Inst);
293 // Completely replace a single-pred PHI. This is safe, because the
294 // NewVal won't be variant in the loop, so we don't need an LCSSA phi
296 PN->replaceAllUsesWith(ExitVal);
297 RecursivelyDeleteTriviallyDeadInstructions(PN);
301 // Clone the PHI and delete the original one. This lets IVUsers and
302 // any other maps purge the original user from their records.
303 PHINode *NewPN = cast<PHINode>(PN->clone());
305 NewPN->insertBefore(PN);
306 PN->replaceAllUsesWith(NewPN);
307 PN->eraseFromParent();
312 // The insertion point instruction may have been deleted; clear it out
313 // so that the rewriter doesn't trip over it later.
314 Rewriter.clearInsertPoint();
317 void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
318 // First step. Check to see if there are any floating-point recurrences.
319 // If there are, change them into integer recurrences, permitting analysis by
320 // the SCEV routines.
322 BasicBlock *Header = L->getHeader();
324 SmallVector<WeakVH, 8> PHIs;
325 for (BasicBlock::iterator I = Header->begin();
326 PHINode *PN = dyn_cast<PHINode>(I); ++I)
329 for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
330 if (PHINode *PN = dyn_cast_or_null<PHINode>(PHIs[i]))
331 HandleFloatingPointIV(L, PN);
333 // If the loop previously had floating-point IV, ScalarEvolution
334 // may not have been able to compute a trip count. Now that we've done some
335 // re-writing, the trip count may be computable.
340 void IndVarSimplify::EliminateIVComparisons() {
341 // Look for ICmp users.
342 for (IVUsers::iterator I = IU->begin(), E = IU->end(); I != E;) {
343 IVStrideUse &UI = *I++;
344 ICmpInst *ICmp = dyn_cast<ICmpInst>(UI.getUser());
347 bool Swapped = UI.getOperandValToReplace() == ICmp->getOperand(1);
348 ICmpInst::Predicate Pred = ICmp->getPredicate();
349 if (Swapped) Pred = ICmpInst::getSwappedPredicate(Pred);
351 // Get the SCEVs for the ICmp operands.
352 const SCEV *S = IU->getReplacementExpr(UI);
353 const SCEV *X = SE->getSCEV(ICmp->getOperand(!Swapped));
355 // Simplify unnecessary loops away.
356 const Loop *ICmpLoop = LI->getLoopFor(ICmp->getParent());
357 S = SE->getSCEVAtScope(S, ICmpLoop);
358 X = SE->getSCEVAtScope(X, ICmpLoop);
360 // If the condition is always true or always false, replace it with
362 if (SE->isKnownPredicate(Pred, S, X))
363 ICmp->replaceAllUsesWith(ConstantInt::getTrue(ICmp->getContext()));
364 else if (SE->isKnownPredicate(ICmpInst::getInversePredicate(Pred), S, X))
365 ICmp->replaceAllUsesWith(ConstantInt::getFalse(ICmp->getContext()));
369 DEBUG(dbgs() << "INDVARS: Eliminated comparison: " << *ICmp << '\n');
370 ICmp->eraseFromParent();
374 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
375 IU = &getAnalysis<IVUsers>();
376 LI = &getAnalysis<LoopInfo>();
377 SE = &getAnalysis<ScalarEvolution>();
378 DT = &getAnalysis<DominatorTree>();
381 // If there are any floating-point recurrences, attempt to
382 // transform them to use integer recurrences.
383 RewriteNonIntegerIVs(L);
385 BasicBlock *ExitingBlock = L->getExitingBlock(); // may be null
386 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
388 // Create a rewriter object which we'll use to transform the code with.
389 SCEVExpander Rewriter(*SE);
391 // Check to see if this loop has a computable loop-invariant execution count.
392 // If so, this means that we can compute the final value of any expressions
393 // that are recurrent in the loop, and substitute the exit values from the
394 // loop into any instructions outside of the loop that use the final values of
395 // the current expressions.
397 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount))
398 RewriteLoopExitValues(L, Rewriter);
400 // Simplify ICmp IV users.
401 EliminateIVComparisons();
403 // Compute the type of the largest recurrence expression, and decide whether
404 // a canonical induction variable should be inserted.
405 const Type *LargestType = 0;
406 bool NeedCannIV = false;
407 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount)) {
408 LargestType = BackedgeTakenCount->getType();
409 LargestType = SE->getEffectiveSCEVType(LargestType);
410 // If we have a known trip count and a single exit block, we'll be
411 // rewriting the loop exit test condition below, which requires a
412 // canonical induction variable.
416 for (IVUsers::const_iterator I = IU->begin(), E = IU->end(); I != E; ++I) {
418 SE->getEffectiveSCEVType(I->getOperandValToReplace()->getType());
420 SE->getTypeSizeInBits(Ty) >
421 SE->getTypeSizeInBits(LargestType))
426 // Now that we know the largest of the induction variable expressions
427 // in this loop, insert a canonical induction variable of the largest size.
430 // Check to see if the loop already has any canonical-looking induction
431 // variables. If any are present and wider than the planned canonical
432 // induction variable, temporarily remove them, so that the Rewriter
433 // doesn't attempt to reuse them.
434 SmallVector<PHINode *, 2> OldCannIVs;
435 while (PHINode *OldCannIV = L->getCanonicalInductionVariable()) {
436 if (SE->getTypeSizeInBits(OldCannIV->getType()) >
437 SE->getTypeSizeInBits(LargestType))
438 OldCannIV->removeFromParent();
441 OldCannIVs.push_back(OldCannIV);
444 IndVar = Rewriter.getOrInsertCanonicalInductionVariable(L, LargestType);
448 DEBUG(dbgs() << "INDVARS: New CanIV: " << *IndVar << '\n');
450 // Now that the official induction variable is established, reinsert
451 // any old canonical-looking variables after it so that the IR remains
452 // consistent. They will be deleted as part of the dead-PHI deletion at
453 // the end of the pass.
454 while (!OldCannIVs.empty()) {
455 PHINode *OldCannIV = OldCannIVs.pop_back_val();
456 OldCannIV->insertBefore(L->getHeader()->getFirstNonPHI());
460 // If we have a trip count expression, rewrite the loop's exit condition
461 // using it. We can currently only handle loops with a single exit.
462 ICmpInst *NewICmp = 0;
463 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount) &&
464 !BackedgeTakenCount->isZero() &&
467 "LinearFunctionTestReplace requires a canonical induction variable");
468 // Can't rewrite non-branch yet.
469 if (BranchInst *BI = dyn_cast<BranchInst>(ExitingBlock->getTerminator()))
470 NewICmp = LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar,
471 ExitingBlock, BI, Rewriter);
474 // Rewrite IV-derived expressions. Clears the rewriter cache.
475 RewriteIVExpressions(L, Rewriter);
477 // The Rewriter may not be used from this point on.
479 // Loop-invariant instructions in the preheader that aren't used in the
480 // loop may be sunk below the loop to reduce register pressure.
481 SinkUnusedInvariants(L);
483 // For completeness, inform IVUsers of the IV use in the newly-created
484 // loop exit test instruction.
486 IU->AddUsersIfInteresting(cast<Instruction>(NewICmp->getOperand(0)));
488 // Clean up dead instructions.
489 Changed |= DeleteDeadPHIs(L->getHeader());
490 // Check a post-condition.
491 assert(L->isLCSSAForm(*DT) && "Indvars did not leave the loop in lcssa form!");
495 // FIXME: It is an extremely bad idea to indvar substitute anything more
496 // complex than affine induction variables. Doing so will put expensive
497 // polynomial evaluations inside of the loop, and the str reduction pass
498 // currently can only reduce affine polynomials. For now just disable
499 // indvar subst on anything more complex than an affine addrec, unless
500 // it can be expanded to a trivial value.
501 static bool isSafe(const SCEV *S, const Loop *L) {
502 // Loop-invariant values are safe.
503 if (S->isLoopInvariant(L)) return true;
505 // Affine addrecs are safe. Non-affine are not, because LSR doesn't know how
506 // to transform them into efficient code.
507 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
508 return AR->isAffine();
510 // An add is safe it all its operands are safe.
511 if (const SCEVCommutativeExpr *Commutative = dyn_cast<SCEVCommutativeExpr>(S)) {
512 for (SCEVCommutativeExpr::op_iterator I = Commutative->op_begin(),
513 E = Commutative->op_end(); I != E; ++I)
514 if (!isSafe(*I, L)) return false;
518 // A cast is safe if its operand is.
519 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
520 return isSafe(C->getOperand(), L);
522 // A udiv is safe if its operands are.
523 if (const SCEVUDivExpr *UD = dyn_cast<SCEVUDivExpr>(S))
524 return isSafe(UD->getLHS(), L) &&
525 isSafe(UD->getRHS(), L);
527 // SCEVUnknown is always safe.
528 if (isa<SCEVUnknown>(S))
531 // Nothing else is safe.
535 void IndVarSimplify::RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter) {
536 SmallVector<WeakVH, 16> DeadInsts;
538 // Rewrite all induction variable expressions in terms of the canonical
539 // induction variable.
541 // If there were induction variables of other sizes or offsets, manually
542 // add the offsets to the primary induction variable and cast, avoiding
543 // the need for the code evaluation methods to insert induction variables
544 // of different sizes.
545 for (IVUsers::iterator UI = IU->begin(), E = IU->end(); UI != E; ++UI) {
546 Value *Op = UI->getOperandValToReplace();
547 const Type *UseTy = Op->getType();
548 Instruction *User = UI->getUser();
550 // Compute the final addrec to expand into code.
551 const SCEV *AR = IU->getReplacementExpr(*UI);
553 // Evaluate the expression out of the loop, if possible.
554 if (!L->contains(UI->getUser())) {
555 const SCEV *ExitVal = SE->getSCEVAtScope(AR, L->getParentLoop());
556 if (ExitVal->isLoopInvariant(L))
560 // FIXME: It is an extremely bad idea to indvar substitute anything more
561 // complex than affine induction variables. Doing so will put expensive
562 // polynomial evaluations inside of the loop, and the str reduction pass
563 // currently can only reduce affine polynomials. For now just disable
564 // indvar subst on anything more complex than an affine addrec, unless
565 // it can be expanded to a trivial value.
569 // Determine the insertion point for this user. By default, insert
570 // immediately before the user. The SCEVExpander class will automatically
571 // hoist loop invariants out of the loop. For PHI nodes, there may be
572 // multiple uses, so compute the nearest common dominator for the
574 Instruction *InsertPt = User;
575 if (PHINode *PHI = dyn_cast<PHINode>(InsertPt))
576 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
577 if (PHI->getIncomingValue(i) == Op) {
578 if (InsertPt == User)
579 InsertPt = PHI->getIncomingBlock(i)->getTerminator();
582 DT->findNearestCommonDominator(InsertPt->getParent(),
583 PHI->getIncomingBlock(i))
587 // Now expand it into actual Instructions and patch it into place.
588 Value *NewVal = Rewriter.expandCodeFor(AR, UseTy, InsertPt);
590 // Inform ScalarEvolution that this value is changing. The change doesn't
591 // affect its value, but it does potentially affect which use lists the
592 // value will be on after the replacement, which affects ScalarEvolution's
593 // ability to walk use lists and drop dangling pointers when a value is
595 SE->forgetValue(User);
597 // Patch the new value into place.
599 NewVal->takeName(Op);
600 User->replaceUsesOfWith(Op, NewVal);
601 UI->setOperandValToReplace(NewVal);
602 DEBUG(dbgs() << "INDVARS: Rewrote IV '" << *AR << "' " << *Op << '\n'
603 << " into = " << *NewVal << "\n");
607 // The old value may be dead now.
608 DeadInsts.push_back(Op);
611 // Clear the rewriter cache, because values that are in the rewriter's cache
612 // can be deleted in the loop below, causing the AssertingVH in the cache to
615 // Now that we're done iterating through lists, clean up any instructions
616 // which are now dead.
617 while (!DeadInsts.empty())
618 if (Instruction *Inst =
619 dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val()))
620 RecursivelyDeleteTriviallyDeadInstructions(Inst);
623 /// If there's a single exit block, sink any loop-invariant values that
624 /// were defined in the preheader but not used inside the loop into the
625 /// exit block to reduce register pressure in the loop.
626 void IndVarSimplify::SinkUnusedInvariants(Loop *L) {
627 BasicBlock *ExitBlock = L->getExitBlock();
628 if (!ExitBlock) return;
630 BasicBlock *Preheader = L->getLoopPreheader();
631 if (!Preheader) return;
633 Instruction *InsertPt = ExitBlock->getFirstNonPHI();
634 BasicBlock::iterator I = Preheader->getTerminator();
635 while (I != Preheader->begin()) {
637 // New instructions were inserted at the end of the preheader.
641 // Don't move instructions which might have side effects, since the side
642 // effects need to complete before instructions inside the loop. Also don't
643 // move instructions which might read memory, since the loop may modify
644 // memory. Note that it's okay if the instruction might have undefined
645 // behavior: LoopSimplify guarantees that the preheader dominates the exit
647 if (I->mayHaveSideEffects() || I->mayReadFromMemory())
650 // Skip debug info intrinsics.
651 if (isa<DbgInfoIntrinsic>(I))
654 // Don't sink static AllocaInsts out of the entry block, which would
655 // turn them into dynamic allocas!
656 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
657 if (AI->isStaticAlloca())
660 // Determine if there is a use in or before the loop (direct or
662 bool UsedInLoop = false;
663 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
665 BasicBlock *UseBB = cast<Instruction>(UI)->getParent();
666 if (PHINode *P = dyn_cast<PHINode>(UI)) {
668 PHINode::getIncomingValueNumForOperand(UI.getOperandNo());
669 UseBB = P->getIncomingBlock(i);
671 if (UseBB == Preheader || L->contains(UseBB)) {
677 // If there is, the def must remain in the preheader.
681 // Otherwise, sink it to the exit block.
682 Instruction *ToMove = I;
685 if (I != Preheader->begin()) {
686 // Skip debug info intrinsics.
689 } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin());
691 if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin())
697 ToMove->moveBefore(InsertPt);
703 /// ConvertToSInt - Convert APF to an integer, if possible.
704 static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) {
705 bool isExact = false;
706 if (&APF.getSemantics() == &APFloat::PPCDoubleDouble)
708 // See if we can convert this to an int64_t
710 if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero,
711 &isExact) != APFloat::opOK || !isExact)
717 /// HandleFloatingPointIV - If the loop has floating induction variable
718 /// then insert corresponding integer induction variable if possible.
720 /// for(double i = 0; i < 10000; ++i)
722 /// is converted into
723 /// for(int i = 0; i < 10000; ++i)
726 void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) {
727 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
728 unsigned BackEdge = IncomingEdge^1;
730 // Check incoming value.
731 ConstantFP *InitValueVal =
732 dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));
735 if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
738 // Check IV increment. Reject this PN if increment operation is not
739 // an add or increment value can not be represented by an integer.
740 BinaryOperator *Incr =
741 dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
742 if (Incr == 0 || Incr->getOpcode() != Instruction::FAdd) return;
744 // If this is not an add of the PHI with a constantfp, or if the constant fp
745 // is not an integer, bail out.
746 ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
748 if (IncValueVal == 0 || Incr->getOperand(0) != PN ||
749 !ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
752 // Check Incr uses. One user is PN and the other user is an exit condition
753 // used by the conditional terminator.
754 Value::use_iterator IncrUse = Incr->use_begin();
755 Instruction *U1 = cast<Instruction>(IncrUse++);
756 if (IncrUse == Incr->use_end()) return;
757 Instruction *U2 = cast<Instruction>(IncrUse++);
758 if (IncrUse != Incr->use_end()) return;
760 // Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't
761 // only used by a branch, we can't transform it.
762 FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
764 Compare = dyn_cast<FCmpInst>(U2);
765 if (Compare == 0 || !Compare->hasOneUse() ||
766 !isa<BranchInst>(Compare->use_back()))
769 BranchInst *TheBr = cast<BranchInst>(Compare->use_back());
771 // We need to verify that the branch actually controls the iteration count
772 // of the loop. If not, the new IV can overflow and no one will notice.
773 // The branch block must be in the loop and one of the successors must be out
775 assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
776 if (!L->contains(TheBr->getParent()) ||
777 (L->contains(TheBr->getSuccessor(0)) &&
778 L->contains(TheBr->getSuccessor(1))))
782 // If it isn't a comparison with an integer-as-fp (the exit value), we can't
784 ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
786 if (ExitValueVal == 0 ||
787 !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
790 // Find new predicate for integer comparison.
791 CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
792 switch (Compare->getPredicate()) {
793 default: return; // Unknown comparison.
794 case CmpInst::FCMP_OEQ:
795 case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
796 case CmpInst::FCMP_ONE:
797 case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
798 case CmpInst::FCMP_OGT:
799 case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
800 case CmpInst::FCMP_OGE:
801 case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
802 case CmpInst::FCMP_OLT:
803 case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
804 case CmpInst::FCMP_OLE:
805 case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
808 // We convert the floating point induction variable to a signed i32 value if
809 // we can. This is only safe if the comparison will not overflow in a way
810 // that won't be trapped by the integer equivalent operations. Check for this
812 // TODO: We could use i64 if it is native and the range requires it.
814 // The start/stride/exit values must all fit in signed i32.
815 if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
818 // If not actually striding (add x, 0.0), avoid touching the code.
822 // Positive and negative strides have different safety conditions.
824 // If we have a positive stride, we require the init to be less than the
825 // exit value and an equality or less than comparison.
826 if (InitValue >= ExitValue ||
827 NewPred == CmpInst::ICMP_SGT || NewPred == CmpInst::ICMP_SGE)
830 uint32_t Range = uint32_t(ExitValue-InitValue);
831 if (NewPred == CmpInst::ICMP_SLE) {
832 // Normalize SLE -> SLT, check for infinite loop.
833 if (++Range == 0) return; // Range overflows.
836 unsigned Leftover = Range % uint32_t(IncValue);
838 // If this is an equality comparison, we require that the strided value
839 // exactly land on the exit value, otherwise the IV condition will wrap
840 // around and do things the fp IV wouldn't.
841 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
845 // If the stride would wrap around the i32 before exiting, we can't
847 if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
851 // If we have a negative stride, we require the init to be greater than the
852 // exit value and an equality or greater than comparison.
853 if (InitValue >= ExitValue ||
854 NewPred == CmpInst::ICMP_SLT || NewPred == CmpInst::ICMP_SLE)
857 uint32_t Range = uint32_t(InitValue-ExitValue);
858 if (NewPred == CmpInst::ICMP_SGE) {
859 // Normalize SGE -> SGT, check for infinite loop.
860 if (++Range == 0) return; // Range overflows.
863 unsigned Leftover = Range % uint32_t(-IncValue);
865 // If this is an equality comparison, we require that the strided value
866 // exactly land on the exit value, otherwise the IV condition will wrap
867 // around and do things the fp IV wouldn't.
868 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
872 // If the stride would wrap around the i32 before exiting, we can't
874 if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
878 const IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
880 // Insert new integer induction variable.
881 PHINode *NewPHI = PHINode::Create(Int32Ty, PN->getName()+".int", PN);
882 NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
883 PN->getIncomingBlock(IncomingEdge));
886 BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
887 Incr->getName()+".int", Incr);
888 NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
890 ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
891 ConstantInt::get(Int32Ty, ExitValue),
894 // In the following deletions, PN may become dead and may be deleted.
895 // Use a WeakVH to observe whether this happens.
898 // Delete the old floating point exit comparison. The branch starts using the
900 NewCompare->takeName(Compare);
901 Compare->replaceAllUsesWith(NewCompare);
902 RecursivelyDeleteTriviallyDeadInstructions(Compare);
904 // Delete the old floating point increment.
905 Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
906 RecursivelyDeleteTriviallyDeadInstructions(Incr);
908 // If the FP induction variable still has uses, this is because something else
909 // in the loop uses its value. In order to canonicalize the induction
910 // variable, we chose to eliminate the IV and rewrite it in terms of an
913 // We give preference to sitofp over uitofp because it is faster on most
916 Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
917 PN->getParent()->getFirstNonPHI());
918 PN->replaceAllUsesWith(Conv);
919 RecursivelyDeleteTriviallyDeadInstructions(PN);
922 // Add a new IVUsers entry for the newly-created integer PHI.
923 IU->AddUsersIfInteresting(NewPHI);