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 {
76 SmallVector<WeakVH, 16> DeadInsts;
80 static char ID; // Pass identification, replacement for typeid
81 IndVarSimplify() : LoopPass(ID) {
82 initializeIndVarSimplifyPass(*PassRegistry::getPassRegistry());
85 virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
87 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
88 AU.addRequired<DominatorTree>();
89 AU.addRequired<LoopInfo>();
90 AU.addRequired<ScalarEvolution>();
91 AU.addRequiredID(LoopSimplifyID);
92 AU.addRequiredID(LCSSAID);
93 AU.addRequired<IVUsers>();
94 AU.addPreserved<ScalarEvolution>();
95 AU.addPreservedID(LoopSimplifyID);
96 AU.addPreservedID(LCSSAID);
97 AU.addPreserved<IVUsers>();
102 bool isValidRewrite(Value *FromVal, Value *ToVal);
104 void EliminateIVComparisons();
105 void EliminateIVRemainders();
106 void RewriteNonIntegerIVs(Loop *L);
108 bool canExpandBackedgeTakenCount(Loop *L,
109 const SCEV *BackedgeTakenCount);
111 ICmpInst *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
113 SCEVExpander &Rewriter);
114 void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter);
116 void RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter);
118 void SinkUnusedInvariants(Loop *L);
120 void HandleFloatingPointIV(Loop *L, PHINode *PH);
124 char IndVarSimplify::ID = 0;
125 INITIALIZE_PASS_BEGIN(IndVarSimplify, "indvars",
126 "Canonicalize Induction Variables", false, false)
127 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
128 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
129 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
130 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
131 INITIALIZE_PASS_DEPENDENCY(LCSSA)
132 INITIALIZE_PASS_DEPENDENCY(IVUsers)
133 INITIALIZE_PASS_END(IndVarSimplify, "indvars",
134 "Canonicalize Induction Variables", false, false)
136 Pass *llvm::createIndVarSimplifyPass() {
137 return new IndVarSimplify();
140 /// isValidRewrite - Return true if the SCEV expansion generated by the
141 /// rewriter can replace the original value. SCEV guarantees that it
142 /// produces the same value, but the way it is produced may be illegal IR.
143 /// Ideally, this function will only be called for verification.
144 bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) {
145 // If an SCEV expression subsumed multiple pointers, its expansion could
146 // reassociate the GEP changing the base pointer. This is illegal because the
147 // final address produced by a GEP chain must be inbounds relative to its
148 // underlying object. Otherwise basic alias analysis, among other things,
149 // could fail in a dangerous way. Ultimately, SCEV will be improved to avoid
150 // producing an expression involving multiple pointers. Until then, we must
153 // Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject
154 // because it understands lcssa phis while SCEV does not.
155 Value *FromPtr = FromVal;
156 Value *ToPtr = ToVal;
157 if (GEPOperator *GEP = dyn_cast<GEPOperator>(FromVal)) {
158 FromPtr = GEP->getPointerOperand();
160 if (GEPOperator *GEP = dyn_cast<GEPOperator>(ToVal)) {
161 ToPtr = GEP->getPointerOperand();
163 if (FromPtr != FromVal || ToPtr != ToVal) {
164 // Quickly check the common case
165 if (FromPtr == ToPtr)
168 // SCEV may have rewritten an expression that produces the GEP's pointer
169 // operand. That's ok as long as the pointer operand has the same base
170 // pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the
171 // base of a recurrence. This handles the case in which SCEV expansion
172 // converts a pointer type recurrence into a nonrecurrent pointer base
173 // indexed by an integer recurrence.
174 const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr));
175 const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr));
176 if (FromBase == ToBase)
179 DEBUG(dbgs() << "INDVARS: GEP rewrite bail out "
180 << *FromBase << " != " << *ToBase << "\n");
187 /// canExpandBackedgeTakenCount - Return true if this loop's backedge taken
188 /// count expression can be safely and cheaply expanded into an instruction
189 /// sequence that can be used by LinearFunctionTestReplace.
190 bool IndVarSimplify::
191 canExpandBackedgeTakenCount(Loop *L,
192 const SCEV *BackedgeTakenCount) {
193 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) ||
194 BackedgeTakenCount->isZero())
197 if (!L->getExitingBlock())
200 // Can't rewrite non-branch yet.
201 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
205 // Special case: If the backedge-taken count is a UDiv, it's very likely a
206 // UDiv that ScalarEvolution produced in order to compute a precise
207 // expression, rather than a UDiv from the user's code. If we can't find a
208 // UDiv in the code with some simple searching, assume the former and forego
209 // rewriting the loop.
210 if (isa<SCEVUDivExpr>(BackedgeTakenCount)) {
211 ICmpInst *OrigCond = dyn_cast<ICmpInst>(BI->getCondition());
212 if (!OrigCond) return 0;
213 const SCEV *R = SE->getSCEV(OrigCond->getOperand(1));
214 R = SE->getMinusSCEV(R, SE->getConstant(R->getType(), 1));
215 if (R != BackedgeTakenCount) {
216 const SCEV *L = SE->getSCEV(OrigCond->getOperand(0));
217 L = SE->getMinusSCEV(L, SE->getConstant(L->getType(), 1));
218 if (L != BackedgeTakenCount)
225 /// LinearFunctionTestReplace - This method rewrites the exit condition of the
226 /// loop to be a canonical != comparison against the incremented loop induction
227 /// variable. This pass is able to rewrite the exit tests of any loop where the
228 /// SCEV analysis can determine a loop-invariant trip count of the loop, which
229 /// is actually a much broader range than just linear tests.
230 ICmpInst *IndVarSimplify::
231 LinearFunctionTestReplace(Loop *L,
232 const SCEV *BackedgeTakenCount,
234 SCEVExpander &Rewriter) {
235 assert(canExpandBackedgeTakenCount(L, BackedgeTakenCount) && "precondition");
236 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
238 // If the exiting block is not the same as the backedge block, we must compare
239 // against the preincremented value, otherwise we prefer to compare against
240 // the post-incremented value.
242 const SCEV *RHS = BackedgeTakenCount;
243 if (L->getExitingBlock() == L->getLoopLatch()) {
244 // Add one to the "backedge-taken" count to get the trip count.
245 // If this addition may overflow, we have to be more pessimistic and
246 // cast the induction variable before doing the add.
247 const SCEV *Zero = SE->getConstant(BackedgeTakenCount->getType(), 0);
249 SE->getAddExpr(BackedgeTakenCount,
250 SE->getConstant(BackedgeTakenCount->getType(), 1));
251 if ((isa<SCEVConstant>(N) && !N->isZero()) ||
252 SE->isLoopEntryGuardedByCond(L, ICmpInst::ICMP_NE, N, Zero)) {
253 // No overflow. Cast the sum.
254 RHS = SE->getTruncateOrZeroExtend(N, IndVar->getType());
256 // Potential overflow. Cast before doing the add.
257 RHS = SE->getTruncateOrZeroExtend(BackedgeTakenCount,
259 RHS = SE->getAddExpr(RHS,
260 SE->getConstant(IndVar->getType(), 1));
263 // The BackedgeTaken expression contains the number of times that the
264 // backedge branches to the loop header. This is one less than the
265 // number of times the loop executes, so use the incremented indvar.
266 CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock());
268 // We have to use the preincremented value...
269 RHS = SE->getTruncateOrZeroExtend(BackedgeTakenCount,
274 // Expand the code for the iteration count.
275 assert(SE->isLoopInvariant(RHS, L) &&
276 "Computed iteration count is not loop invariant!");
277 Value *ExitCnt = Rewriter.expandCodeFor(RHS, IndVar->getType(), BI);
279 // Insert a new icmp_ne or icmp_eq instruction before the branch.
280 ICmpInst::Predicate Opcode;
281 if (L->contains(BI->getSuccessor(0)))
282 Opcode = ICmpInst::ICMP_NE;
284 Opcode = ICmpInst::ICMP_EQ;
286 DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
287 << " LHS:" << *CmpIndVar << '\n'
289 << (Opcode == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
290 << " RHS:\t" << *RHS << "\n");
292 ICmpInst *Cond = new ICmpInst(BI, Opcode, CmpIndVar, ExitCnt, "exitcond");
294 Value *OrigCond = BI->getCondition();
295 // It's tempting to use replaceAllUsesWith here to fully replace the old
296 // comparison, but that's not immediately safe, since users of the old
297 // comparison may not be dominated by the new comparison. Instead, just
298 // update the branch to use the new comparison; in the common case this
299 // will make old comparison dead.
300 BI->setCondition(Cond);
301 DeadInsts.push_back(OrigCond);
308 /// RewriteLoopExitValues - Check to see if this loop has a computable
309 /// loop-invariant execution count. If so, this means that we can compute the
310 /// final value of any expressions that are recurrent in the loop, and
311 /// substitute the exit values from the loop into any instructions outside of
312 /// the loop that use the final values of the current expressions.
314 /// This is mostly redundant with the regular IndVarSimplify activities that
315 /// happen later, except that it's more powerful in some cases, because it's
316 /// able to brute-force evaluate arbitrary instructions as long as they have
317 /// constant operands at the beginning of the loop.
318 void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) {
319 // Verify the input to the pass in already in LCSSA form.
320 assert(L->isLCSSAForm(*DT));
322 SmallVector<BasicBlock*, 8> ExitBlocks;
323 L->getUniqueExitBlocks(ExitBlocks);
325 // Find all values that are computed inside the loop, but used outside of it.
326 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan
327 // the exit blocks of the loop to find them.
328 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
329 BasicBlock *ExitBB = ExitBlocks[i];
331 // If there are no PHI nodes in this exit block, then no values defined
332 // inside the loop are used on this path, skip it.
333 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
336 unsigned NumPreds = PN->getNumIncomingValues();
338 // Iterate over all of the PHI nodes.
339 BasicBlock::iterator BBI = ExitBB->begin();
340 while ((PN = dyn_cast<PHINode>(BBI++))) {
342 continue; // dead use, don't replace it
344 // SCEV only supports integer expressions for now.
345 if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy())
348 // It's necessary to tell ScalarEvolution about this explicitly so that
349 // it can walk the def-use list and forget all SCEVs, as it may not be
350 // watching the PHI itself. Once the new exit value is in place, there
351 // may not be a def-use connection between the loop and every instruction
352 // which got a SCEVAddRecExpr for that loop.
355 // Iterate over all of the values in all the PHI nodes.
356 for (unsigned i = 0; i != NumPreds; ++i) {
357 // If the value being merged in is not integer or is not defined
358 // in the loop, skip it.
359 Value *InVal = PN->getIncomingValue(i);
360 if (!isa<Instruction>(InVal))
363 // If this pred is for a subloop, not L itself, skip it.
364 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
365 continue; // The Block is in a subloop, skip it.
367 // Check that InVal is defined in the loop.
368 Instruction *Inst = cast<Instruction>(InVal);
369 if (!L->contains(Inst))
372 // Okay, this instruction has a user outside of the current loop
373 // and varies predictably *inside* the loop. Evaluate the value it
374 // contains when the loop exits, if possible.
375 const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
376 if (!SE->isLoopInvariant(ExitValue, L))
379 Value *ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), Inst);
381 DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
382 << " LoopVal = " << *Inst << "\n");
384 if (!isValidRewrite(Inst, ExitVal)) {
385 DeadInsts.push_back(ExitVal);
391 PN->setIncomingValue(i, ExitVal);
393 // If this instruction is dead now, delete it.
394 RecursivelyDeleteTriviallyDeadInstructions(Inst);
397 // Completely replace a single-pred PHI. This is safe, because the
398 // NewVal won't be variant in the loop, so we don't need an LCSSA phi
400 PN->replaceAllUsesWith(ExitVal);
401 RecursivelyDeleteTriviallyDeadInstructions(PN);
405 // Clone the PHI and delete the original one. This lets IVUsers and
406 // any other maps purge the original user from their records.
407 PHINode *NewPN = cast<PHINode>(PN->clone());
409 NewPN->insertBefore(PN);
410 PN->replaceAllUsesWith(NewPN);
411 PN->eraseFromParent();
416 // The insertion point instruction may have been deleted; clear it out
417 // so that the rewriter doesn't trip over it later.
418 Rewriter.clearInsertPoint();
421 void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
422 // First step. Check to see if there are any floating-point recurrences.
423 // If there are, change them into integer recurrences, permitting analysis by
424 // the SCEV routines.
426 BasicBlock *Header = L->getHeader();
428 SmallVector<WeakVH, 8> PHIs;
429 for (BasicBlock::iterator I = Header->begin();
430 PHINode *PN = dyn_cast<PHINode>(I); ++I)
433 for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
434 if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i]))
435 HandleFloatingPointIV(L, PN);
437 // If the loop previously had floating-point IV, ScalarEvolution
438 // may not have been able to compute a trip count. Now that we've done some
439 // re-writing, the trip count may be computable.
444 void IndVarSimplify::EliminateIVComparisons() {
445 // Look for ICmp users.
446 for (IVUsers::iterator I = IU->begin(), E = IU->end(); I != E; ++I) {
447 IVStrideUse &UI = *I;
448 ICmpInst *ICmp = dyn_cast<ICmpInst>(UI.getUser());
451 bool Swapped = UI.getOperandValToReplace() == ICmp->getOperand(1);
452 ICmpInst::Predicate Pred = ICmp->getPredicate();
453 if (Swapped) Pred = ICmpInst::getSwappedPredicate(Pred);
455 // Get the SCEVs for the ICmp operands.
456 const SCEV *S = IU->getReplacementExpr(UI);
457 const SCEV *X = SE->getSCEV(ICmp->getOperand(!Swapped));
459 // Simplify unnecessary loops away.
460 const Loop *ICmpLoop = LI->getLoopFor(ICmp->getParent());
461 S = SE->getSCEVAtScope(S, ICmpLoop);
462 X = SE->getSCEVAtScope(X, ICmpLoop);
464 // If the condition is always true or always false, replace it with
466 if (SE->isKnownPredicate(Pred, S, X))
467 ICmp->replaceAllUsesWith(ConstantInt::getTrue(ICmp->getContext()));
468 else if (SE->isKnownPredicate(ICmpInst::getInversePredicate(Pred), S, X))
469 ICmp->replaceAllUsesWith(ConstantInt::getFalse(ICmp->getContext()));
473 DEBUG(dbgs() << "INDVARS: Eliminated comparison: " << *ICmp << '\n');
474 DeadInsts.push_back(ICmp);
478 void IndVarSimplify::EliminateIVRemainders() {
479 // Look for SRem and URem users.
480 for (IVUsers::iterator I = IU->begin(), E = IU->end(); I != E; ++I) {
481 IVStrideUse &UI = *I;
482 BinaryOperator *Rem = dyn_cast<BinaryOperator>(UI.getUser());
485 bool isSigned = Rem->getOpcode() == Instruction::SRem;
486 if (!isSigned && Rem->getOpcode() != Instruction::URem)
489 // We're only interested in the case where we know something about
491 if (UI.getOperandValToReplace() != Rem->getOperand(0))
494 // Get the SCEVs for the ICmp operands.
495 const SCEV *S = SE->getSCEV(Rem->getOperand(0));
496 const SCEV *X = SE->getSCEV(Rem->getOperand(1));
498 // Simplify unnecessary loops away.
499 const Loop *ICmpLoop = LI->getLoopFor(Rem->getParent());
500 S = SE->getSCEVAtScope(S, ICmpLoop);
501 X = SE->getSCEVAtScope(X, ICmpLoop);
503 // i % n --> i if i is in [0,n).
504 if ((!isSigned || SE->isKnownNonNegative(S)) &&
505 SE->isKnownPredicate(isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
507 Rem->replaceAllUsesWith(Rem->getOperand(0));
509 // (i+1) % n --> (i+1)==n?0:(i+1) if i is in [0,n).
510 const SCEV *LessOne =
511 SE->getMinusSCEV(S, SE->getConstant(S->getType(), 1));
512 if ((!isSigned || SE->isKnownNonNegative(LessOne)) &&
513 SE->isKnownPredicate(isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
515 ICmpInst *ICmp = new ICmpInst(Rem, ICmpInst::ICMP_EQ,
516 Rem->getOperand(0), Rem->getOperand(1),
519 SelectInst::Create(ICmp,
520 ConstantInt::get(Rem->getType(), 0),
521 Rem->getOperand(0), "tmp", Rem);
522 Rem->replaceAllUsesWith(Sel);
527 // Inform IVUsers about the new users.
528 if (Instruction *I = dyn_cast<Instruction>(Rem->getOperand(0)))
529 IU->AddUsersIfInteresting(I);
531 DEBUG(dbgs() << "INDVARS: Simplified rem: " << *Rem << '\n');
532 DeadInsts.push_back(Rem);
536 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
537 // If LoopSimplify form is not available, stay out of trouble. Some notes:
538 // - LSR currently only supports LoopSimplify-form loops. Indvars'
539 // canonicalization can be a pessimization without LSR to "clean up"
541 // - We depend on having a preheader; in particular,
542 // Loop::getCanonicalInductionVariable only supports loops with preheaders,
543 // and we're in trouble if we can't find the induction variable even when
544 // we've manually inserted one.
545 if (!L->isLoopSimplifyForm())
548 IU = &getAnalysis<IVUsers>();
549 LI = &getAnalysis<LoopInfo>();
550 SE = &getAnalysis<ScalarEvolution>();
551 DT = &getAnalysis<DominatorTree>();
555 // If there are any floating-point recurrences, attempt to
556 // transform them to use integer recurrences.
557 RewriteNonIntegerIVs(L);
559 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
561 // Create a rewriter object which we'll use to transform the code with.
562 SCEVExpander Rewriter(*SE);
564 // Check to see if this loop has a computable loop-invariant execution count.
565 // If so, this means that we can compute the final value of any expressions
566 // that are recurrent in the loop, and substitute the exit values from the
567 // loop into any instructions outside of the loop that use the final values of
568 // the current expressions.
570 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount))
571 RewriteLoopExitValues(L, Rewriter);
573 // Simplify ICmp IV users.
574 EliminateIVComparisons();
576 // Simplify SRem and URem IV users.
577 EliminateIVRemainders();
579 // Compute the type of the largest recurrence expression, and decide whether
580 // a canonical induction variable should be inserted.
581 const Type *LargestType = 0;
582 bool NeedCannIV = false;
583 bool ExpandBECount = canExpandBackedgeTakenCount(L, BackedgeTakenCount);
585 // If we have a known trip count and a single exit block, we'll be
586 // rewriting the loop exit test condition below, which requires a
587 // canonical induction variable.
589 const Type *Ty = BackedgeTakenCount->getType();
591 SE->getTypeSizeInBits(Ty) >
592 SE->getTypeSizeInBits(LargestType))
593 LargestType = SE->getEffectiveSCEVType(Ty);
595 for (IVUsers::const_iterator I = IU->begin(), E = IU->end(); I != E; ++I) {
598 SE->getEffectiveSCEVType(I->getOperandValToReplace()->getType());
600 SE->getTypeSizeInBits(Ty) >
601 SE->getTypeSizeInBits(LargestType))
605 // Now that we know the largest of the induction variable expressions
606 // in this loop, insert a canonical induction variable of the largest size.
609 // Check to see if the loop already has any canonical-looking induction
610 // variables. If any are present and wider than the planned canonical
611 // induction variable, temporarily remove them, so that the Rewriter
612 // doesn't attempt to reuse them.
613 SmallVector<PHINode *, 2> OldCannIVs;
614 while (PHINode *OldCannIV = L->getCanonicalInductionVariable()) {
615 if (SE->getTypeSizeInBits(OldCannIV->getType()) >
616 SE->getTypeSizeInBits(LargestType))
617 OldCannIV->removeFromParent();
620 OldCannIVs.push_back(OldCannIV);
623 IndVar = Rewriter.getOrInsertCanonicalInductionVariable(L, LargestType);
627 DEBUG(dbgs() << "INDVARS: New CanIV: " << *IndVar << '\n');
629 // Now that the official induction variable is established, reinsert
630 // any old canonical-looking variables after it so that the IR remains
631 // consistent. They will be deleted as part of the dead-PHI deletion at
632 // the end of the pass.
633 while (!OldCannIVs.empty()) {
634 PHINode *OldCannIV = OldCannIVs.pop_back_val();
635 OldCannIV->insertBefore(L->getHeader()->getFirstNonPHI());
639 // If we have a trip count expression, rewrite the loop's exit condition
640 // using it. We can currently only handle loops with a single exit.
641 ICmpInst *NewICmp = 0;
643 assert(canExpandBackedgeTakenCount(L, BackedgeTakenCount) &&
644 "canonical IV disrupted BackedgeTaken expansion");
646 "LinearFunctionTestReplace requires a canonical induction variable");
647 NewICmp = LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar,
651 // Rewrite IV-derived expressions.
652 RewriteIVExpressions(L, Rewriter);
654 // Clear the rewriter cache, because values that are in the rewriter's cache
655 // can be deleted in the loop below, causing the AssertingVH in the cache to
659 // Now that we're done iterating through lists, clean up any instructions
660 // which are now dead.
661 while (!DeadInsts.empty())
662 if (Instruction *Inst =
663 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
664 RecursivelyDeleteTriviallyDeadInstructions(Inst);
666 // The Rewriter may not be used from this point on.
668 // Loop-invariant instructions in the preheader that aren't used in the
669 // loop may be sunk below the loop to reduce register pressure.
670 SinkUnusedInvariants(L);
672 // For completeness, inform IVUsers of the IV use in the newly-created
673 // loop exit test instruction.
675 IU->AddUsersIfInteresting(cast<Instruction>(NewICmp->getOperand(0)));
677 // Clean up dead instructions.
678 Changed |= DeleteDeadPHIs(L->getHeader());
679 // Check a post-condition.
680 assert(L->isLCSSAForm(*DT) && "Indvars did not leave the loop in lcssa form!");
684 // FIXME: It is an extremely bad idea to indvar substitute anything more
685 // complex than affine induction variables. Doing so will put expensive
686 // polynomial evaluations inside of the loop, and the str reduction pass
687 // currently can only reduce affine polynomials. For now just disable
688 // indvar subst on anything more complex than an affine addrec, unless
689 // it can be expanded to a trivial value.
690 static bool isSafe(const SCEV *S, const Loop *L, ScalarEvolution *SE) {
691 // Loop-invariant values are safe.
692 if (SE->isLoopInvariant(S, L)) return true;
694 // Affine addrecs are safe. Non-affine are not, because LSR doesn't know how
695 // to transform them into efficient code.
696 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
697 return AR->isAffine();
699 // An add is safe it all its operands are safe.
700 if (const SCEVCommutativeExpr *Commutative = dyn_cast<SCEVCommutativeExpr>(S)) {
701 for (SCEVCommutativeExpr::op_iterator I = Commutative->op_begin(),
702 E = Commutative->op_end(); I != E; ++I)
703 if (!isSafe(*I, L, SE)) return false;
707 // A cast is safe if its operand is.
708 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
709 return isSafe(C->getOperand(), L, SE);
711 // A udiv is safe if its operands are.
712 if (const SCEVUDivExpr *UD = dyn_cast<SCEVUDivExpr>(S))
713 return isSafe(UD->getLHS(), L, SE) &&
714 isSafe(UD->getRHS(), L, SE);
716 // SCEVUnknown is always safe.
717 if (isa<SCEVUnknown>(S))
720 // Nothing else is safe.
724 void IndVarSimplify::RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter) {
725 // Rewrite all induction variable expressions in terms of the canonical
726 // induction variable.
728 // If there were induction variables of other sizes or offsets, manually
729 // add the offsets to the primary induction variable and cast, avoiding
730 // the need for the code evaluation methods to insert induction variables
731 // of different sizes.
732 for (IVUsers::iterator UI = IU->begin(), E = IU->end(); UI != E; ++UI) {
733 Value *Op = UI->getOperandValToReplace();
734 const Type *UseTy = Op->getType();
735 Instruction *User = UI->getUser();
737 // Compute the final addrec to expand into code.
738 const SCEV *AR = IU->getReplacementExpr(*UI);
740 // Evaluate the expression out of the loop, if possible.
741 if (!L->contains(UI->getUser())) {
742 const SCEV *ExitVal = SE->getSCEVAtScope(AR, L->getParentLoop());
743 if (SE->isLoopInvariant(ExitVal, L))
747 // FIXME: It is an extremely bad idea to indvar substitute anything more
748 // complex than affine induction variables. Doing so will put expensive
749 // polynomial evaluations inside of the loop, and the str reduction pass
750 // currently can only reduce affine polynomials. For now just disable
751 // indvar subst on anything more complex than an affine addrec, unless
752 // it can be expanded to a trivial value.
753 if (!isSafe(AR, L, SE))
756 // Determine the insertion point for this user. By default, insert
757 // immediately before the user. The SCEVExpander class will automatically
758 // hoist loop invariants out of the loop. For PHI nodes, there may be
759 // multiple uses, so compute the nearest common dominator for the
761 Instruction *InsertPt = User;
762 if (PHINode *PHI = dyn_cast<PHINode>(InsertPt))
763 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
764 if (PHI->getIncomingValue(i) == Op) {
765 if (InsertPt == User)
766 InsertPt = PHI->getIncomingBlock(i)->getTerminator();
769 DT->findNearestCommonDominator(InsertPt->getParent(),
770 PHI->getIncomingBlock(i))
774 // Now expand it into actual Instructions and patch it into place.
775 Value *NewVal = Rewriter.expandCodeFor(AR, UseTy, InsertPt);
777 DEBUG(dbgs() << "INDVARS: Rewrote IV '" << *AR << "' " << *Op << '\n'
778 << " into = " << *NewVal << "\n");
780 if (!isValidRewrite(Op, NewVal)) {
781 DeadInsts.push_back(NewVal);
784 // Inform ScalarEvolution that this value is changing. The change doesn't
785 // affect its value, but it does potentially affect which use lists the
786 // value will be on after the replacement, which affects ScalarEvolution's
787 // ability to walk use lists and drop dangling pointers when a value is
789 SE->forgetValue(User);
791 // Patch the new value into place.
793 NewVal->takeName(Op);
794 User->replaceUsesOfWith(Op, NewVal);
795 UI->setOperandValToReplace(NewVal);
800 // The old value may be dead now.
801 DeadInsts.push_back(Op);
805 /// If there's a single exit block, sink any loop-invariant values that
806 /// were defined in the preheader but not used inside the loop into the
807 /// exit block to reduce register pressure in the loop.
808 void IndVarSimplify::SinkUnusedInvariants(Loop *L) {
809 BasicBlock *ExitBlock = L->getExitBlock();
810 if (!ExitBlock) return;
812 BasicBlock *Preheader = L->getLoopPreheader();
813 if (!Preheader) return;
815 Instruction *InsertPt = ExitBlock->getFirstNonPHI();
816 BasicBlock::iterator I = Preheader->getTerminator();
817 while (I != Preheader->begin()) {
819 // New instructions were inserted at the end of the preheader.
823 // Don't move instructions which might have side effects, since the side
824 // effects need to complete before instructions inside the loop. Also don't
825 // move instructions which might read memory, since the loop may modify
826 // memory. Note that it's okay if the instruction might have undefined
827 // behavior: LoopSimplify guarantees that the preheader dominates the exit
829 if (I->mayHaveSideEffects() || I->mayReadFromMemory())
832 // Skip debug info intrinsics.
833 if (isa<DbgInfoIntrinsic>(I))
836 // Don't sink static AllocaInsts out of the entry block, which would
837 // turn them into dynamic allocas!
838 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
839 if (AI->isStaticAlloca())
842 // Determine if there is a use in or before the loop (direct or
844 bool UsedInLoop = false;
845 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
848 BasicBlock *UseBB = cast<Instruction>(U)->getParent();
849 if (PHINode *P = dyn_cast<PHINode>(U)) {
851 PHINode::getIncomingValueNumForOperand(UI.getOperandNo());
852 UseBB = P->getIncomingBlock(i);
854 if (UseBB == Preheader || L->contains(UseBB)) {
860 // If there is, the def must remain in the preheader.
864 // Otherwise, sink it to the exit block.
865 Instruction *ToMove = I;
868 if (I != Preheader->begin()) {
869 // Skip debug info intrinsics.
872 } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin());
874 if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin())
880 ToMove->moveBefore(InsertPt);
886 /// ConvertToSInt - Convert APF to an integer, if possible.
887 static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) {
888 bool isExact = false;
889 if (&APF.getSemantics() == &APFloat::PPCDoubleDouble)
891 // See if we can convert this to an int64_t
893 if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero,
894 &isExact) != APFloat::opOK || !isExact)
900 /// HandleFloatingPointIV - If the loop has floating induction variable
901 /// then insert corresponding integer induction variable if possible.
903 /// for(double i = 0; i < 10000; ++i)
905 /// is converted into
906 /// for(int i = 0; i < 10000; ++i)
909 void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) {
910 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
911 unsigned BackEdge = IncomingEdge^1;
913 // Check incoming value.
914 ConstantFP *InitValueVal =
915 dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));
918 if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
921 // Check IV increment. Reject this PN if increment operation is not
922 // an add or increment value can not be represented by an integer.
923 BinaryOperator *Incr =
924 dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
925 if (Incr == 0 || Incr->getOpcode() != Instruction::FAdd) return;
927 // If this is not an add of the PHI with a constantfp, or if the constant fp
928 // is not an integer, bail out.
929 ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
931 if (IncValueVal == 0 || Incr->getOperand(0) != PN ||
932 !ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
935 // Check Incr uses. One user is PN and the other user is an exit condition
936 // used by the conditional terminator.
937 Value::use_iterator IncrUse = Incr->use_begin();
938 Instruction *U1 = cast<Instruction>(*IncrUse++);
939 if (IncrUse == Incr->use_end()) return;
940 Instruction *U2 = cast<Instruction>(*IncrUse++);
941 if (IncrUse != Incr->use_end()) return;
943 // Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't
944 // only used by a branch, we can't transform it.
945 FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
947 Compare = dyn_cast<FCmpInst>(U2);
948 if (Compare == 0 || !Compare->hasOneUse() ||
949 !isa<BranchInst>(Compare->use_back()))
952 BranchInst *TheBr = cast<BranchInst>(Compare->use_back());
954 // We need to verify that the branch actually controls the iteration count
955 // of the loop. If not, the new IV can overflow and no one will notice.
956 // The branch block must be in the loop and one of the successors must be out
958 assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
959 if (!L->contains(TheBr->getParent()) ||
960 (L->contains(TheBr->getSuccessor(0)) &&
961 L->contains(TheBr->getSuccessor(1))))
965 // If it isn't a comparison with an integer-as-fp (the exit value), we can't
967 ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
969 if (ExitValueVal == 0 ||
970 !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
973 // Find new predicate for integer comparison.
974 CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
975 switch (Compare->getPredicate()) {
976 default: return; // Unknown comparison.
977 case CmpInst::FCMP_OEQ:
978 case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
979 case CmpInst::FCMP_ONE:
980 case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
981 case CmpInst::FCMP_OGT:
982 case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
983 case CmpInst::FCMP_OGE:
984 case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
985 case CmpInst::FCMP_OLT:
986 case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
987 case CmpInst::FCMP_OLE:
988 case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
991 // We convert the floating point induction variable to a signed i32 value if
992 // we can. This is only safe if the comparison will not overflow in a way
993 // that won't be trapped by the integer equivalent operations. Check for this
995 // TODO: We could use i64 if it is native and the range requires it.
997 // The start/stride/exit values must all fit in signed i32.
998 if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
1001 // If not actually striding (add x, 0.0), avoid touching the code.
1005 // Positive and negative strides have different safety conditions.
1007 // If we have a positive stride, we require the init to be less than the
1008 // exit value and an equality or less than comparison.
1009 if (InitValue >= ExitValue ||
1010 NewPred == CmpInst::ICMP_SGT || NewPred == CmpInst::ICMP_SGE)
1013 uint32_t Range = uint32_t(ExitValue-InitValue);
1014 if (NewPred == CmpInst::ICMP_SLE) {
1015 // Normalize SLE -> SLT, check for infinite loop.
1016 if (++Range == 0) return; // Range overflows.
1019 unsigned Leftover = Range % uint32_t(IncValue);
1021 // If this is an equality comparison, we require that the strided value
1022 // exactly land on the exit value, otherwise the IV condition will wrap
1023 // around and do things the fp IV wouldn't.
1024 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
1028 // If the stride would wrap around the i32 before exiting, we can't
1029 // transform the IV.
1030 if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
1034 // If we have a negative stride, we require the init to be greater than the
1035 // exit value and an equality or greater than comparison.
1036 if (InitValue >= ExitValue ||
1037 NewPred == CmpInst::ICMP_SLT || NewPred == CmpInst::ICMP_SLE)
1040 uint32_t Range = uint32_t(InitValue-ExitValue);
1041 if (NewPred == CmpInst::ICMP_SGE) {
1042 // Normalize SGE -> SGT, check for infinite loop.
1043 if (++Range == 0) return; // Range overflows.
1046 unsigned Leftover = Range % uint32_t(-IncValue);
1048 // If this is an equality comparison, we require that the strided value
1049 // exactly land on the exit value, otherwise the IV condition will wrap
1050 // around and do things the fp IV wouldn't.
1051 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
1055 // If the stride would wrap around the i32 before exiting, we can't
1056 // transform the IV.
1057 if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
1061 const IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
1063 // Insert new integer induction variable.
1064 PHINode *NewPHI = PHINode::Create(Int32Ty, 2, PN->getName()+".int", PN);
1065 NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
1066 PN->getIncomingBlock(IncomingEdge));
1069 BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
1070 Incr->getName()+".int", Incr);
1071 NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
1073 ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
1074 ConstantInt::get(Int32Ty, ExitValue),
1075 Compare->getName());
1077 // In the following deletions, PN may become dead and may be deleted.
1078 // Use a WeakVH to observe whether this happens.
1081 // Delete the old floating point exit comparison. The branch starts using the
1083 NewCompare->takeName(Compare);
1084 Compare->replaceAllUsesWith(NewCompare);
1085 RecursivelyDeleteTriviallyDeadInstructions(Compare);
1087 // Delete the old floating point increment.
1088 Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
1089 RecursivelyDeleteTriviallyDeadInstructions(Incr);
1091 // If the FP induction variable still has uses, this is because something else
1092 // in the loop uses its value. In order to canonicalize the induction
1093 // variable, we chose to eliminate the IV and rewrite it in terms of an
1096 // We give preference to sitofp over uitofp because it is faster on most
1099 Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
1100 PN->getParent()->getFirstNonPHI());
1101 PN->replaceAllUsesWith(Conv);
1102 RecursivelyDeleteTriviallyDeadInstructions(PN);
1105 // Add a new IVUsers entry for the newly-created integer PHI.
1106 IU->AddUsersIfInteresting(NewPHI);