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 RewriteNonIntegerIVs(Loop *L);
102 ICmpInst *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
104 BasicBlock *ExitingBlock,
106 SCEVExpander &Rewriter);
107 void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter);
109 void RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter);
111 void SinkUnusedInvariants(Loop *L);
113 void HandleFloatingPointIV(Loop *L, PHINode *PH);
117 char IndVarSimplify::ID = 0;
118 static RegisterPass<IndVarSimplify>
119 X("indvars", "Canonicalize Induction Variables");
121 Pass *llvm::createIndVarSimplifyPass() {
122 return new IndVarSimplify();
125 /// LinearFunctionTestReplace - This method rewrites the exit condition of the
126 /// loop to be a canonical != comparison against the incremented loop induction
127 /// variable. This pass is able to rewrite the exit tests of any loop where the
128 /// SCEV analysis can determine a loop-invariant trip count of the loop, which
129 /// is actually a much broader range than just linear tests.
130 ICmpInst *IndVarSimplify::LinearFunctionTestReplace(Loop *L,
131 const SCEV *BackedgeTakenCount,
133 BasicBlock *ExitingBlock,
135 SCEVExpander &Rewriter) {
136 // If the exiting block is not the same as the backedge block, we must compare
137 // against the preincremented value, otherwise we prefer to compare against
138 // the post-incremented value.
140 const SCEV *RHS = BackedgeTakenCount;
141 if (ExitingBlock == L->getLoopLatch()) {
142 // Add one to the "backedge-taken" count to get the trip count.
143 // If this addition may overflow, we have to be more pessimistic and
144 // cast the induction variable before doing the add.
145 const SCEV *Zero = SE->getIntegerSCEV(0, BackedgeTakenCount->getType());
147 SE->getAddExpr(BackedgeTakenCount,
148 SE->getIntegerSCEV(1, BackedgeTakenCount->getType()));
149 if ((isa<SCEVConstant>(N) && !N->isZero()) ||
150 SE->isLoopGuardedByCond(L, ICmpInst::ICMP_NE, N, Zero)) {
151 // No overflow. Cast the sum.
152 RHS = SE->getTruncateOrZeroExtend(N, IndVar->getType());
154 // Potential overflow. Cast before doing the add.
155 RHS = SE->getTruncateOrZeroExtend(BackedgeTakenCount,
157 RHS = SE->getAddExpr(RHS,
158 SE->getIntegerSCEV(1, IndVar->getType()));
161 // The BackedgeTaken expression contains the number of times that the
162 // backedge branches to the loop header. This is one less than the
163 // number of times the loop executes, so use the incremented indvar.
164 CmpIndVar = L->getCanonicalInductionVariableIncrement();
166 // We have to use the preincremented value...
167 RHS = SE->getTruncateOrZeroExtend(BackedgeTakenCount,
172 // Expand the code for the iteration count.
173 assert(RHS->isLoopInvariant(L) &&
174 "Computed iteration count is not loop invariant!");
175 Value *ExitCnt = Rewriter.expandCodeFor(RHS, IndVar->getType(), BI);
177 // Insert a new icmp_ne or icmp_eq instruction before the branch.
178 ICmpInst::Predicate Opcode;
179 if (L->contains(BI->getSuccessor(0)))
180 Opcode = ICmpInst::ICMP_NE;
182 Opcode = ICmpInst::ICMP_EQ;
184 DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
185 << " LHS:" << *CmpIndVar << '\n'
187 << (Opcode == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
188 << " RHS:\t" << *RHS << "\n");
190 ICmpInst *Cond = new ICmpInst(BI, Opcode, CmpIndVar, ExitCnt, "exitcond");
192 Value *OrigCond = BI->getCondition();
193 // It's tempting to use replaceAllUsesWith here to fully replace the old
194 // comparison, but that's not immediately safe, since users of the old
195 // comparison may not be dominated by the new comparison. Instead, just
196 // update the branch to use the new comparison; in the common case this
197 // will make old comparison dead.
198 BI->setCondition(Cond);
199 RecursivelyDeleteTriviallyDeadInstructions(OrigCond);
206 /// RewriteLoopExitValues - Check to see if this loop has a computable
207 /// loop-invariant execution count. If so, this means that we can compute the
208 /// final value of any expressions that are recurrent in the loop, and
209 /// substitute the exit values from the loop into any instructions outside of
210 /// the loop that use the final values of the current expressions.
212 /// This is mostly redundant with the regular IndVarSimplify activities that
213 /// happen later, except that it's more powerful in some cases, because it's
214 /// able to brute-force evaluate arbitrary instructions as long as they have
215 /// constant operands at the beginning of the loop.
216 void IndVarSimplify::RewriteLoopExitValues(Loop *L,
217 SCEVExpander &Rewriter) {
218 // Verify the input to the pass in already in LCSSA form.
219 assert(L->isLCSSAForm(*DT));
221 SmallVector<BasicBlock*, 8> ExitBlocks;
222 L->getUniqueExitBlocks(ExitBlocks);
224 // Find all values that are computed inside the loop, but used outside of it.
225 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan
226 // the exit blocks of the loop to find them.
227 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
228 BasicBlock *ExitBB = ExitBlocks[i];
230 // If there are no PHI nodes in this exit block, then no values defined
231 // inside the loop are used on this path, skip it.
232 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
235 unsigned NumPreds = PN->getNumIncomingValues();
237 // Iterate over all of the PHI nodes.
238 BasicBlock::iterator BBI = ExitBB->begin();
239 while ((PN = dyn_cast<PHINode>(BBI++))) {
241 continue; // dead use, don't replace it
243 // SCEV only supports integer expressions for now.
244 if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy())
247 // It's necessary to tell ScalarEvolution about this explicitly so that
248 // it can walk the def-use list and forget all SCEVs, as it may not be
249 // watching the PHI itself. Once the new exit value is in place, there
250 // may not be a def-use connection between the loop and every instruction
251 // which got a SCEVAddRecExpr for that loop.
254 // Iterate over all of the values in all the PHI nodes.
255 for (unsigned i = 0; i != NumPreds; ++i) {
256 // If the value being merged in is not integer or is not defined
257 // in the loop, skip it.
258 Value *InVal = PN->getIncomingValue(i);
259 if (!isa<Instruction>(InVal))
262 // If this pred is for a subloop, not L itself, skip it.
263 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
264 continue; // The Block is in a subloop, skip it.
266 // Check that InVal is defined in the loop.
267 Instruction *Inst = cast<Instruction>(InVal);
268 if (!L->contains(Inst))
271 // Okay, this instruction has a user outside of the current loop
272 // and varies predictably *inside* the loop. Evaluate the value it
273 // contains when the loop exits, if possible.
274 const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
275 if (!ExitValue->isLoopInvariant(L))
281 Value *ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), Inst);
283 DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
284 << " LoopVal = " << *Inst << "\n");
286 PN->setIncomingValue(i, ExitVal);
288 // If this instruction is dead now, delete it.
289 RecursivelyDeleteTriviallyDeadInstructions(Inst);
292 // Completely replace a single-pred PHI. This is safe, because the
293 // NewVal won't be variant in the loop, so we don't need an LCSSA phi
295 PN->replaceAllUsesWith(ExitVal);
296 RecursivelyDeleteTriviallyDeadInstructions(PN);
300 // Clone the PHI and delete the original one. This lets IVUsers and
301 // any other maps purge the original user from their records.
302 PHINode *NewPN = cast<PHINode>(PN->clone());
304 NewPN->insertBefore(PN);
305 PN->replaceAllUsesWith(NewPN);
306 PN->eraseFromParent();
311 // The insertion point instruction may have been deleted; clear it out
312 // so that the rewriter doesn't trip over it later.
313 Rewriter.clearInsertPoint();
316 void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
317 // First step. Check to see if there are any floating-point recurrences.
318 // If there are, change them into integer recurrences, permitting analysis by
319 // the SCEV routines.
321 BasicBlock *Header = L->getHeader();
323 SmallVector<WeakVH, 8> PHIs;
324 for (BasicBlock::iterator I = Header->begin();
325 PHINode *PN = dyn_cast<PHINode>(I); ++I)
328 for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
329 if (PHINode *PN = dyn_cast_or_null<PHINode>(PHIs[i]))
330 HandleFloatingPointIV(L, PN);
332 // If the loop previously had floating-point IV, ScalarEvolution
333 // may not have been able to compute a trip count. Now that we've done some
334 // re-writing, the trip count may be computable.
339 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
340 IU = &getAnalysis<IVUsers>();
341 LI = &getAnalysis<LoopInfo>();
342 SE = &getAnalysis<ScalarEvolution>();
343 DT = &getAnalysis<DominatorTree>();
346 // If there are any floating-point recurrences, attempt to
347 // transform them to use integer recurrences.
348 RewriteNonIntegerIVs(L);
350 BasicBlock *ExitingBlock = L->getExitingBlock(); // may be null
351 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
353 // Create a rewriter object which we'll use to transform the code with.
354 SCEVExpander Rewriter(*SE);
356 // Check to see if this loop has a computable loop-invariant execution count.
357 // If so, this means that we can compute the final value of any expressions
358 // that are recurrent in the loop, and substitute the exit values from the
359 // loop into any instructions outside of the loop that use the final values of
360 // the current expressions.
362 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount))
363 RewriteLoopExitValues(L, Rewriter);
365 // Compute the type of the largest recurrence expression, and decide whether
366 // a canonical induction variable should be inserted.
367 const Type *LargestType = 0;
368 bool NeedCannIV = false;
369 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount)) {
370 LargestType = BackedgeTakenCount->getType();
371 LargestType = SE->getEffectiveSCEVType(LargestType);
372 // If we have a known trip count and a single exit block, we'll be
373 // rewriting the loop exit test condition below, which requires a
374 // canonical induction variable.
378 for (IVUsers::const_iterator I = IU->begin(), E = IU->end(); I != E; ++I) {
380 SE->getEffectiveSCEVType(I->getOperandValToReplace()->getType());
382 SE->getTypeSizeInBits(Ty) >
383 SE->getTypeSizeInBits(LargestType))
388 // Now that we know the largest of the induction variable expressions
389 // in this loop, insert a canonical induction variable of the largest size.
392 // Check to see if the loop already has any canonical-looking induction
393 // variables. If any are present and wider than the planned canonical
394 // induction variable, temporarily remove them, so that the Rewriter
395 // doesn't attempt to reuse them.
396 SmallVector<PHINode *, 2> OldCannIVs;
397 while (PHINode *OldCannIV = L->getCanonicalInductionVariable()) {
398 if (SE->getTypeSizeInBits(OldCannIV->getType()) >
399 SE->getTypeSizeInBits(LargestType))
400 OldCannIV->removeFromParent();
403 OldCannIVs.push_back(OldCannIV);
406 IndVar = Rewriter.getOrInsertCanonicalInductionVariable(L, LargestType);
410 DEBUG(dbgs() << "INDVARS: New CanIV: " << *IndVar << '\n');
412 // Now that the official induction variable is established, reinsert
413 // any old canonical-looking variables after it so that the IR remains
414 // consistent. They will be deleted as part of the dead-PHI deletion at
415 // the end of the pass.
416 while (!OldCannIVs.empty()) {
417 PHINode *OldCannIV = OldCannIVs.pop_back_val();
418 OldCannIV->insertBefore(L->getHeader()->getFirstNonPHI());
422 // If we have a trip count expression, rewrite the loop's exit condition
423 // using it. We can currently only handle loops with a single exit.
424 ICmpInst *NewICmp = 0;
425 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount) &&
426 !BackedgeTakenCount->isZero() &&
429 "LinearFunctionTestReplace requires a canonical induction variable");
430 // Can't rewrite non-branch yet.
431 if (BranchInst *BI = dyn_cast<BranchInst>(ExitingBlock->getTerminator()))
432 NewICmp = LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar,
433 ExitingBlock, BI, Rewriter);
436 // Rewrite IV-derived expressions. Clears the rewriter cache.
437 RewriteIVExpressions(L, Rewriter);
439 // The Rewriter may not be used from this point on.
441 // Loop-invariant instructions in the preheader that aren't used in the
442 // loop may be sunk below the loop to reduce register pressure.
443 SinkUnusedInvariants(L);
445 // For completeness, inform IVUsers of the IV use in the newly-created
446 // loop exit test instruction.
448 IU->AddUsersIfInteresting(cast<Instruction>(NewICmp->getOperand(0)));
450 // Clean up dead instructions.
451 Changed |= DeleteDeadPHIs(L->getHeader());
452 // Check a post-condition.
453 assert(L->isLCSSAForm(*DT) && "Indvars did not leave the loop in lcssa form!");
457 // FIXME: It is an extremely bad idea to indvar substitute anything more
458 // complex than affine induction variables. Doing so will put expensive
459 // polynomial evaluations inside of the loop, and the str reduction pass
460 // currently can only reduce affine polynomials. For now just disable
461 // indvar subst on anything more complex than an affine addrec, unless
462 // it can be expanded to a trivial value.
463 static bool isSafe(const SCEV *S, const Loop *L) {
464 // Loop-invariant values are safe.
465 if (S->isLoopInvariant(L)) return true;
467 // Affine addrecs are safe. Non-affine are not, because LSR doesn't know how
468 // to transform them into efficient code.
469 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
470 return AR->isAffine();
472 // An add is safe it all its operands are safe.
473 if (const SCEVCommutativeExpr *Commutative = dyn_cast<SCEVCommutativeExpr>(S)) {
474 for (SCEVCommutativeExpr::op_iterator I = Commutative->op_begin(),
475 E = Commutative->op_end(); I != E; ++I)
476 if (!isSafe(*I, L)) return false;
480 // A cast is safe if its operand is.
481 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
482 return isSafe(C->getOperand(), L);
484 // A udiv is safe if its operands are.
485 if (const SCEVUDivExpr *UD = dyn_cast<SCEVUDivExpr>(S))
486 return isSafe(UD->getLHS(), L) &&
487 isSafe(UD->getRHS(), L);
489 // SCEVUnknown is always safe.
490 if (isa<SCEVUnknown>(S))
493 // Nothing else is safe.
497 void IndVarSimplify::RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter) {
498 SmallVector<WeakVH, 16> DeadInsts;
500 // Rewrite all induction variable expressions in terms of the canonical
501 // induction variable.
503 // If there were induction variables of other sizes or offsets, manually
504 // add the offsets to the primary induction variable and cast, avoiding
505 // the need for the code evaluation methods to insert induction variables
506 // of different sizes.
507 for (IVUsers::iterator UI = IU->begin(), E = IU->end(); UI != E; ++UI) {
508 Value *Op = UI->getOperandValToReplace();
509 const Type *UseTy = Op->getType();
510 Instruction *User = UI->getUser();
512 // Compute the final addrec to expand into code.
513 const SCEV *AR = IU->getReplacementExpr(*UI);
515 // Evaluate the expression out of the loop, if possible.
516 if (!L->contains(UI->getUser())) {
517 const SCEV *ExitVal = SE->getSCEVAtScope(AR, L->getParentLoop());
518 if (ExitVal->isLoopInvariant(L))
522 // FIXME: It is an extremely bad idea to indvar substitute anything more
523 // complex than affine induction variables. Doing so will put expensive
524 // polynomial evaluations inside of the loop, and the str reduction pass
525 // currently can only reduce affine polynomials. For now just disable
526 // indvar subst on anything more complex than an affine addrec, unless
527 // it can be expanded to a trivial value.
531 // Determine the insertion point for this user. By default, insert
532 // immediately before the user. The SCEVExpander class will automatically
533 // hoist loop invariants out of the loop. For PHI nodes, there may be
534 // multiple uses, so compute the nearest common dominator for the
536 Instruction *InsertPt = User;
537 if (PHINode *PHI = dyn_cast<PHINode>(InsertPt))
538 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
539 if (PHI->getIncomingValue(i) == Op) {
540 if (InsertPt == User)
541 InsertPt = PHI->getIncomingBlock(i)->getTerminator();
544 DT->findNearestCommonDominator(InsertPt->getParent(),
545 PHI->getIncomingBlock(i))
549 // Now expand it into actual Instructions and patch it into place.
550 Value *NewVal = Rewriter.expandCodeFor(AR, UseTy, InsertPt);
552 // Inform ScalarEvolution that this value is changing. The change doesn't
553 // affect its value, but it does potentially affect which use lists the
554 // value will be on after the replacement, which affects ScalarEvolution's
555 // ability to walk use lists and drop dangling pointers when a value is
557 SE->forgetValue(User);
559 // Patch the new value into place.
561 NewVal->takeName(Op);
562 User->replaceUsesOfWith(Op, NewVal);
563 UI->setOperandValToReplace(NewVal);
564 DEBUG(dbgs() << "INDVARS: Rewrote IV '" << *AR << "' " << *Op << '\n'
565 << " into = " << *NewVal << "\n");
569 // The old value may be dead now.
570 DeadInsts.push_back(Op);
573 // Clear the rewriter cache, because values that are in the rewriter's cache
574 // can be deleted in the loop below, causing the AssertingVH in the cache to
577 // Now that we're done iterating through lists, clean up any instructions
578 // which are now dead.
579 while (!DeadInsts.empty())
580 if (Instruction *Inst =
581 dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val()))
582 RecursivelyDeleteTriviallyDeadInstructions(Inst);
585 /// If there's a single exit block, sink any loop-invariant values that
586 /// were defined in the preheader but not used inside the loop into the
587 /// exit block to reduce register pressure in the loop.
588 void IndVarSimplify::SinkUnusedInvariants(Loop *L) {
589 BasicBlock *ExitBlock = L->getExitBlock();
590 if (!ExitBlock) return;
592 BasicBlock *Preheader = L->getLoopPreheader();
593 if (!Preheader) return;
595 Instruction *InsertPt = ExitBlock->getFirstNonPHI();
596 BasicBlock::iterator I = Preheader->getTerminator();
597 while (I != Preheader->begin()) {
599 // New instructions were inserted at the end of the preheader.
603 // Don't move instructions which might have side effects, since the side
604 // effects need to complete before instructions inside the loop. Also don't
605 // move instructions which might read memory, since the loop may modify
606 // memory. Note that it's okay if the instruction might have undefined
607 // behavior: LoopSimplify guarantees that the preheader dominates the exit
609 if (I->mayHaveSideEffects() || I->mayReadFromMemory())
612 // Skip debug info intrinsics.
613 if (isa<DbgInfoIntrinsic>(I))
616 // Don't sink static AllocaInsts out of the entry block, which would
617 // turn them into dynamic allocas!
618 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
619 if (AI->isStaticAlloca())
622 // Determine if there is a use in or before the loop (direct or
624 bool UsedInLoop = false;
625 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
627 BasicBlock *UseBB = cast<Instruction>(UI)->getParent();
628 if (PHINode *P = dyn_cast<PHINode>(UI)) {
630 PHINode::getIncomingValueNumForOperand(UI.getOperandNo());
631 UseBB = P->getIncomingBlock(i);
633 if (UseBB == Preheader || L->contains(UseBB)) {
639 // If there is, the def must remain in the preheader.
643 // Otherwise, sink it to the exit block.
644 Instruction *ToMove = I;
647 if (I != Preheader->begin()) {
648 // Skip debug info intrinsics.
651 } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin());
653 if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin())
659 ToMove->moveBefore(InsertPt);
665 /// ConvertToSInt - Convert APF to an integer, if possible.
666 static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) {
667 bool isExact = false;
668 if (&APF.getSemantics() == &APFloat::PPCDoubleDouble)
670 // See if we can convert this to an int64_t
672 if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero,
673 &isExact) != APFloat::opOK || !isExact)
679 /// HandleFloatingPointIV - If the loop has floating induction variable
680 /// then insert corresponding integer induction variable if possible.
682 /// for(double i = 0; i < 10000; ++i)
684 /// is converted into
685 /// for(int i = 0; i < 10000; ++i)
688 void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) {
689 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
690 unsigned BackEdge = IncomingEdge^1;
692 // Check incoming value.
693 ConstantFP *InitValueVal =
694 dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));
697 if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
700 // Check IV increment. Reject this PN if increment operation is not
701 // an add or increment value can not be represented by an integer.
702 BinaryOperator *Incr =
703 dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
704 if (Incr == 0 || Incr->getOpcode() != Instruction::FAdd) return;
706 // If this is not an add of the PHI with a constantfp, or if the constant fp
707 // is not an integer, bail out.
708 ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
710 if (IncValueVal == 0 || Incr->getOperand(0) != PN ||
711 !ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
714 // Check Incr uses. One user is PN and the other user is an exit condition
715 // used by the conditional terminator.
716 Value::use_iterator IncrUse = Incr->use_begin();
717 Instruction *U1 = cast<Instruction>(IncrUse++);
718 if (IncrUse == Incr->use_end()) return;
719 Instruction *U2 = cast<Instruction>(IncrUse++);
720 if (IncrUse != Incr->use_end()) return;
722 // Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't
723 // only used by a branch, we can't transform it.
724 FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
726 Compare = dyn_cast<FCmpInst>(U2);
727 if (Compare == 0 || !Compare->hasOneUse() ||
728 !isa<BranchInst>(Compare->use_back()))
731 BranchInst *TheBr = cast<BranchInst>(Compare->use_back());
733 // We need to verify that the branch actually controls the iteration count
734 // of the loop. If not, the new IV can overflow and no one will notice.
735 // The branch block must be in the loop and one of the successors must be out
737 assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
738 if (!L->contains(TheBr->getParent()) ||
739 (L->contains(TheBr->getSuccessor(0)) &&
740 L->contains(TheBr->getSuccessor(1))))
744 // If it isn't a comparison with an integer-as-fp (the exit value), we can't
746 ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
748 if (ExitValueVal == 0 ||
749 !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
752 // Find new predicate for integer comparison.
753 CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
754 switch (Compare->getPredicate()) {
755 default: return; // Unknown comparison.
756 case CmpInst::FCMP_OEQ:
757 case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
758 case CmpInst::FCMP_ONE:
759 case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
760 case CmpInst::FCMP_OGT:
761 case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
762 case CmpInst::FCMP_OGE:
763 case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
764 case CmpInst::FCMP_OLT:
765 case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
766 case CmpInst::FCMP_OLE:
767 case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
770 // We convert the floating point induction variable to a signed i32 value if
771 // we can. This is only safe if the comparison will not overflow in a way
772 // that won't be trapped by the integer equivalent operations. Check for this
774 // TODO: We could use i64 if it is native and the range requires it.
776 // The start/stride/exit values must all fit in signed i32.
777 if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
780 // If not actually striding (add x, 0.0), avoid touching the code.
784 // Positive and negative strides have different safety conditions.
786 // If we have a positive stride, we require the init to be less than the
787 // exit value and an equality or less than comparison.
788 if (InitValue >= ExitValue ||
789 NewPred == CmpInst::ICMP_SGT || NewPred == CmpInst::ICMP_SGE)
792 uint32_t Range = uint32_t(ExitValue-InitValue);
793 if (NewPred == CmpInst::ICMP_SLE) {
794 // Normalize SLE -> SLT, check for infinite loop.
795 if (++Range == 0) return; // Range overflows.
798 unsigned Leftover = Range % uint32_t(IncValue);
800 // If this is an equality comparison, we require that the strided value
801 // exactly land on the exit value, otherwise the IV condition will wrap
802 // around and do things the fp IV wouldn't.
803 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
807 // If the stride would wrap around the i32 before exiting, we can't
809 if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
813 // If we have a negative stride, we require the init to be greater than the
814 // exit value and an equality or greater than comparison.
815 if (InitValue >= ExitValue ||
816 NewPred == CmpInst::ICMP_SLT || NewPred == CmpInst::ICMP_SLE)
819 uint32_t Range = uint32_t(InitValue-ExitValue);
820 if (NewPred == CmpInst::ICMP_SGE) {
821 // Normalize SGE -> SGT, check for infinite loop.
822 if (++Range == 0) return; // Range overflows.
825 unsigned Leftover = Range % uint32_t(-IncValue);
827 // If this is an equality comparison, we require that the strided value
828 // exactly land on the exit value, otherwise the IV condition will wrap
829 // around and do things the fp IV wouldn't.
830 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
834 // If the stride would wrap around the i32 before exiting, we can't
836 if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
840 const IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
842 // Insert new integer induction variable.
843 PHINode *NewPHI = PHINode::Create(Int32Ty, PN->getName()+".int", PN);
844 NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
845 PN->getIncomingBlock(IncomingEdge));
848 BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
849 Incr->getName()+".int", Incr);
850 NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
852 ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
853 ConstantInt::get(Int32Ty, ExitValue),
856 // In the following deletions, PN may become dead and may be deleted.
857 // Use a WeakVH to observe whether this happens.
860 // Delete the old floating point exit comparison. The branch starts using the
862 NewCompare->takeName(Compare);
863 Compare->replaceAllUsesWith(NewCompare);
864 RecursivelyDeleteTriviallyDeadInstructions(Compare);
866 // Delete the old floating point increment.
867 Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
868 RecursivelyDeleteTriviallyDeadInstructions(Incr);
870 // If the FP induction variable still has uses, this is because something else
871 // in the loop uses its value. In order to canonicalize the induction
872 // variable, we chose to eliminate the IV and rewrite it in terms of an
875 // We give preference to sitofp over uitofp because it is faster on most
878 Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
879 PN->getParent()->getFirstNonPHI());
880 PN->replaceAllUsesWith(Conv);
881 RecursivelyDeleteTriviallyDeadInstructions(PN);
884 // Add a new IVUsers entry for the newly-created integer PHI.
885 IU->AddUsersIfInteresting(NewPHI);