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/Analysis/ValueTracking.h"
55 #include "llvm/Support/CFG.h"
56 #include "llvm/Support/CommandLine.h"
57 #include "llvm/Support/Debug.h"
58 #include "llvm/Support/raw_ostream.h"
59 #include "llvm/Transforms/Utils/Local.h"
60 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
61 #include "llvm/Target/TargetData.h"
62 #include "llvm/ADT/SmallVector.h"
63 #include "llvm/ADT/Statistic.h"
64 #include "llvm/ADT/STLExtras.h"
67 STATISTIC(NumRemoved , "Number of aux indvars removed");
68 STATISTIC(NumInserted, "Number of canonical indvars added");
69 STATISTIC(NumReplaced, "Number of exit values replaced");
70 STATISTIC(NumLFTR , "Number of loop exit tests replaced");
73 class IndVarSimplify : public LoopPass {
79 SmallVector<WeakVH, 16> DeadInsts;
83 static char ID; // Pass identification, replacement for typeid
84 IndVarSimplify() : LoopPass(ID) {
85 initializeIndVarSimplifyPass(*PassRegistry::getPassRegistry());
88 virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
90 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
91 AU.addRequired<DominatorTree>();
92 AU.addRequired<LoopInfo>();
93 AU.addRequired<ScalarEvolution>();
94 AU.addRequiredID(LoopSimplifyID);
95 AU.addRequiredID(LCSSAID);
96 AU.addRequired<IVUsers>();
97 AU.addPreserved<ScalarEvolution>();
98 AU.addPreservedID(LoopSimplifyID);
99 AU.addPreservedID(LCSSAID);
100 AU.addPreserved<IVUsers>();
101 AU.setPreservesCFG();
105 bool isValidRewrite(Value *FromVal, Value *ToVal);
107 void EliminateIVComparisons();
108 void EliminateIVRemainders();
109 void RewriteNonIntegerIVs(Loop *L);
111 ICmpInst *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
113 BasicBlock *ExitingBlock,
115 SCEVExpander &Rewriter);
116 void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter);
118 void RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter);
120 void SinkUnusedInvariants(Loop *L);
122 void HandleFloatingPointIV(Loop *L, PHINode *PH);
126 char IndVarSimplify::ID = 0;
127 INITIALIZE_PASS_BEGIN(IndVarSimplify, "indvars",
128 "Canonicalize Induction Variables", false, false)
129 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
130 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
131 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
132 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
133 INITIALIZE_PASS_DEPENDENCY(LCSSA)
134 INITIALIZE_PASS_DEPENDENCY(IVUsers)
135 INITIALIZE_PASS_END(IndVarSimplify, "indvars",
136 "Canonicalize Induction Variables", false, false)
138 Pass *llvm::createIndVarSimplifyPass() {
139 return new IndVarSimplify();
142 /// isValidRewrite - Return true if the SCEV expansion generated by the
143 /// rewriter can replace the original value. SCEV guarantees that it
144 /// produces the same value, but the way it is produced may be illegal IR.
145 /// Ideally, this function will only be called for verification.
146 bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) {
147 // If an SCEV expression subsumed multiple pointers, its expansion could
148 // reassociate the GEP changing the base pointer. This is illegal because the
149 // final address produced by a GEP chain must be inbounds relative to its
150 // underlying object. Otherwise basic alias analysis, among other things,
151 // could fail in a dangerous way. Ultimately, SCEV will be improved to avoid
152 // producing an expression involving multiple pointers. Until then, we must
155 // Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject
156 // because it understands lcssa phis while SCEV does not.
157 Value *FromPtr = FromVal;
158 Value *ToPtr = ToVal;
159 if (GEPOperator *GEP = dyn_cast<GEPOperator>(FromVal)) {
160 FromPtr = GEP->getPointerOperand();
162 if (GEPOperator *GEP = dyn_cast<GEPOperator>(ToVal)) {
163 ToPtr = GEP->getPointerOperand();
165 if (FromPtr != FromVal || ToPtr != ToVal) {
166 // Quickly check the common case
167 if (FromPtr == ToPtr)
170 // SCEV may have rewritten an expression that produces the GEP's pointer
171 // operand. That's ok as long as the pointer operand has the same base
172 // pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the
173 // base of a recurrence. This handles the case in which SCEV expansion
174 // converts a pointer type recurrence into a nonrecurrent pointer base
175 // indexed by an integer recurrence.
176 const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr));
177 const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr));
178 if (FromBase == ToBase)
181 DEBUG(dbgs() << "INDVARS: GEP rewrite bail out "
182 << *FromBase << " != " << *ToBase << "\n");
189 /// LinearFunctionTestReplace - This method rewrites the exit condition of the
190 /// loop to be a canonical != comparison against the incremented loop induction
191 /// variable. This pass is able to rewrite the exit tests of any loop where the
192 /// SCEV analysis can determine a loop-invariant trip count of the loop, which
193 /// is actually a much broader range than just linear tests.
194 ICmpInst *IndVarSimplify::LinearFunctionTestReplace(Loop *L,
195 const SCEV *BackedgeTakenCount,
197 BasicBlock *ExitingBlock,
199 SCEVExpander &Rewriter) {
200 // Special case: If the backedge-taken count is a UDiv, it's very likely a
201 // UDiv that ScalarEvolution produced in order to compute a precise
202 // expression, rather than a UDiv from the user's code. If we can't find a
203 // UDiv in the code with some simple searching, assume the former and forego
204 // rewriting the loop.
205 if (isa<SCEVUDivExpr>(BackedgeTakenCount)) {
206 ICmpInst *OrigCond = dyn_cast<ICmpInst>(BI->getCondition());
207 if (!OrigCond) return 0;
208 const SCEV *R = SE->getSCEV(OrigCond->getOperand(1));
209 R = SE->getMinusSCEV(R, SE->getConstant(R->getType(), 1));
210 if (R != BackedgeTakenCount) {
211 const SCEV *L = SE->getSCEV(OrigCond->getOperand(0));
212 L = SE->getMinusSCEV(L, SE->getConstant(L->getType(), 1));
213 if (L != BackedgeTakenCount)
218 // If the exiting block is not the same as the backedge block, we must compare
219 // against the preincremented value, otherwise we prefer to compare against
220 // the post-incremented value.
222 const SCEV *RHS = BackedgeTakenCount;
223 if (ExitingBlock == L->getLoopLatch()) {
224 // Add one to the "backedge-taken" count to get the trip count.
225 // If this addition may overflow, we have to be more pessimistic and
226 // cast the induction variable before doing the add.
227 const SCEV *Zero = SE->getConstant(BackedgeTakenCount->getType(), 0);
229 SE->getAddExpr(BackedgeTakenCount,
230 SE->getConstant(BackedgeTakenCount->getType(), 1));
231 if ((isa<SCEVConstant>(N) && !N->isZero()) ||
232 SE->isLoopEntryGuardedByCond(L, ICmpInst::ICMP_NE, N, Zero)) {
233 // No overflow. Cast the sum.
234 RHS = SE->getTruncateOrZeroExtend(N, IndVar->getType());
236 // Potential overflow. Cast before doing the add.
237 RHS = SE->getTruncateOrZeroExtend(BackedgeTakenCount,
239 RHS = SE->getAddExpr(RHS,
240 SE->getConstant(IndVar->getType(), 1));
243 // The BackedgeTaken expression contains the number of times that the
244 // backedge branches to the loop header. This is one less than the
245 // number of times the loop executes, so use the incremented indvar.
246 CmpIndVar = IndVar->getIncomingValueForBlock(ExitingBlock);
248 // We have to use the preincremented value...
249 RHS = SE->getTruncateOrZeroExtend(BackedgeTakenCount,
254 // Expand the code for the iteration count.
255 assert(SE->isLoopInvariant(RHS, L) &&
256 "Computed iteration count is not loop invariant!");
257 Value *ExitCnt = Rewriter.expandCodeFor(RHS, IndVar->getType(), BI);
259 // Insert a new icmp_ne or icmp_eq instruction before the branch.
260 ICmpInst::Predicate Opcode;
261 if (L->contains(BI->getSuccessor(0)))
262 Opcode = ICmpInst::ICMP_NE;
264 Opcode = ICmpInst::ICMP_EQ;
266 DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
267 << " LHS:" << *CmpIndVar << '\n'
269 << (Opcode == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
270 << " RHS:\t" << *RHS << "\n");
272 ICmpInst *Cond = new ICmpInst(BI, Opcode, CmpIndVar, ExitCnt, "exitcond");
274 Value *OrigCond = BI->getCondition();
275 // It's tempting to use replaceAllUsesWith here to fully replace the old
276 // comparison, but that's not immediately safe, since users of the old
277 // comparison may not be dominated by the new comparison. Instead, just
278 // update the branch to use the new comparison; in the common case this
279 // will make old comparison dead.
280 BI->setCondition(Cond);
281 RecursivelyDeleteTriviallyDeadInstructions(OrigCond);
288 /// RewriteLoopExitValues - Check to see if this loop has a computable
289 /// loop-invariant execution count. If so, this means that we can compute the
290 /// final value of any expressions that are recurrent in the loop, and
291 /// substitute the exit values from the loop into any instructions outside of
292 /// the loop that use the final values of the current expressions.
294 /// This is mostly redundant with the regular IndVarSimplify activities that
295 /// happen later, except that it's more powerful in some cases, because it's
296 /// able to brute-force evaluate arbitrary instructions as long as they have
297 /// constant operands at the beginning of the loop.
298 void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) {
299 // Verify the input to the pass in already in LCSSA form.
300 assert(L->isLCSSAForm(*DT));
302 SmallVector<BasicBlock*, 8> ExitBlocks;
303 L->getUniqueExitBlocks(ExitBlocks);
305 // Find all values that are computed inside the loop, but used outside of it.
306 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan
307 // the exit blocks of the loop to find them.
308 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
309 BasicBlock *ExitBB = ExitBlocks[i];
311 // If there are no PHI nodes in this exit block, then no values defined
312 // inside the loop are used on this path, skip it.
313 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
316 unsigned NumPreds = PN->getNumIncomingValues();
318 // Iterate over all of the PHI nodes.
319 BasicBlock::iterator BBI = ExitBB->begin();
320 while ((PN = dyn_cast<PHINode>(BBI++))) {
322 continue; // dead use, don't replace it
324 // SCEV only supports integer expressions for now.
325 if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy())
328 // It's necessary to tell ScalarEvolution about this explicitly so that
329 // it can walk the def-use list and forget all SCEVs, as it may not be
330 // watching the PHI itself. Once the new exit value is in place, there
331 // may not be a def-use connection between the loop and every instruction
332 // which got a SCEVAddRecExpr for that loop.
335 // Iterate over all of the values in all the PHI nodes.
336 for (unsigned i = 0; i != NumPreds; ++i) {
337 // If the value being merged in is not integer or is not defined
338 // in the loop, skip it.
339 Value *InVal = PN->getIncomingValue(i);
340 if (!isa<Instruction>(InVal))
343 // If this pred is for a subloop, not L itself, skip it.
344 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
345 continue; // The Block is in a subloop, skip it.
347 // Check that InVal is defined in the loop.
348 Instruction *Inst = cast<Instruction>(InVal);
349 if (!L->contains(Inst))
352 // Okay, this instruction has a user outside of the current loop
353 // and varies predictably *inside* the loop. Evaluate the value it
354 // contains when the loop exits, if possible.
355 const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
356 if (!SE->isLoopInvariant(ExitValue, L))
359 Value *ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), Inst);
361 DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
362 << " LoopVal = " << *Inst << "\n");
364 if (!isValidRewrite(Inst, ExitVal)) {
365 DeadInsts.push_back(ExitVal);
371 PN->setIncomingValue(i, ExitVal);
373 // If this instruction is dead now, delete it.
374 RecursivelyDeleteTriviallyDeadInstructions(Inst);
377 // Completely replace a single-pred PHI. This is safe, because the
378 // NewVal won't be variant in the loop, so we don't need an LCSSA phi
380 PN->replaceAllUsesWith(ExitVal);
381 RecursivelyDeleteTriviallyDeadInstructions(PN);
385 // Clone the PHI and delete the original one. This lets IVUsers and
386 // any other maps purge the original user from their records.
387 PHINode *NewPN = cast<PHINode>(PN->clone());
389 NewPN->insertBefore(PN);
390 PN->replaceAllUsesWith(NewPN);
391 PN->eraseFromParent();
396 // The insertion point instruction may have been deleted; clear it out
397 // so that the rewriter doesn't trip over it later.
398 Rewriter.clearInsertPoint();
401 void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
402 // First step. Check to see if there are any floating-point recurrences.
403 // If there are, change them into integer recurrences, permitting analysis by
404 // the SCEV routines.
406 BasicBlock *Header = L->getHeader();
408 SmallVector<WeakVH, 8> PHIs;
409 for (BasicBlock::iterator I = Header->begin();
410 PHINode *PN = dyn_cast<PHINode>(I); ++I)
413 for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
414 if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i]))
415 HandleFloatingPointIV(L, PN);
417 // If the loop previously had floating-point IV, ScalarEvolution
418 // may not have been able to compute a trip count. Now that we've done some
419 // re-writing, the trip count may be computable.
424 void IndVarSimplify::EliminateIVComparisons() {
425 // Look for ICmp users.
426 for (IVUsers::iterator I = IU->begin(), E = IU->end(); I != E; ++I) {
427 IVStrideUse &UI = *I;
428 ICmpInst *ICmp = dyn_cast<ICmpInst>(UI.getUser());
431 bool Swapped = UI.getOperandValToReplace() == ICmp->getOperand(1);
432 ICmpInst::Predicate Pred = ICmp->getPredicate();
433 if (Swapped) Pred = ICmpInst::getSwappedPredicate(Pred);
435 // Get the SCEVs for the ICmp operands.
436 const SCEV *S = IU->getReplacementExpr(UI);
437 const SCEV *X = SE->getSCEV(ICmp->getOperand(!Swapped));
439 // Simplify unnecessary loops away.
440 const Loop *ICmpLoop = LI->getLoopFor(ICmp->getParent());
441 S = SE->getSCEVAtScope(S, ICmpLoop);
442 X = SE->getSCEVAtScope(X, ICmpLoop);
444 // If the condition is always true or always false, replace it with
446 if (SE->isKnownPredicate(Pred, S, X))
447 ICmp->replaceAllUsesWith(ConstantInt::getTrue(ICmp->getContext()));
448 else if (SE->isKnownPredicate(ICmpInst::getInversePredicate(Pred), S, X))
449 ICmp->replaceAllUsesWith(ConstantInt::getFalse(ICmp->getContext()));
453 DEBUG(dbgs() << "INDVARS: Eliminated comparison: " << *ICmp << '\n');
454 DeadInsts.push_back(ICmp);
458 void IndVarSimplify::EliminateIVRemainders() {
459 // Look for SRem and URem users.
460 for (IVUsers::iterator I = IU->begin(), E = IU->end(); I != E; ++I) {
461 IVStrideUse &UI = *I;
462 BinaryOperator *Rem = dyn_cast<BinaryOperator>(UI.getUser());
465 bool isSigned = Rem->getOpcode() == Instruction::SRem;
466 if (!isSigned && Rem->getOpcode() != Instruction::URem)
469 // We're only interested in the case where we know something about
471 if (UI.getOperandValToReplace() != Rem->getOperand(0))
474 // Get the SCEVs for the ICmp operands.
475 const SCEV *S = SE->getSCEV(Rem->getOperand(0));
476 const SCEV *X = SE->getSCEV(Rem->getOperand(1));
478 // Simplify unnecessary loops away.
479 const Loop *ICmpLoop = LI->getLoopFor(Rem->getParent());
480 S = SE->getSCEVAtScope(S, ICmpLoop);
481 X = SE->getSCEVAtScope(X, ICmpLoop);
483 // i % n --> i if i is in [0,n).
484 if ((!isSigned || SE->isKnownNonNegative(S)) &&
485 SE->isKnownPredicate(isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
487 Rem->replaceAllUsesWith(Rem->getOperand(0));
489 // (i+1) % n --> (i+1)==n?0:(i+1) if i is in [0,n).
490 const SCEV *LessOne =
491 SE->getMinusSCEV(S, SE->getConstant(S->getType(), 1));
492 if ((!isSigned || SE->isKnownNonNegative(LessOne)) &&
493 SE->isKnownPredicate(isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
495 ICmpInst *ICmp = new ICmpInst(Rem, ICmpInst::ICMP_EQ,
496 Rem->getOperand(0), Rem->getOperand(1),
499 SelectInst::Create(ICmp,
500 ConstantInt::get(Rem->getType(), 0),
501 Rem->getOperand(0), "tmp", Rem);
502 Rem->replaceAllUsesWith(Sel);
507 // Inform IVUsers about the new users.
508 if (Instruction *I = dyn_cast<Instruction>(Rem->getOperand(0)))
509 IU->AddUsersIfInteresting(I);
511 DEBUG(dbgs() << "INDVARS: Simplified rem: " << *Rem << '\n');
512 DeadInsts.push_back(Rem);
516 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
517 // If LoopSimplify form is not available, stay out of trouble. Some notes:
518 // - LSR currently only supports LoopSimplify-form loops. Indvars'
519 // canonicalization can be a pessimization without LSR to "clean up"
521 // - We depend on having a preheader; in particular,
522 // Loop::getCanonicalInductionVariable only supports loops with preheaders,
523 // and we're in trouble if we can't find the induction variable even when
524 // we've manually inserted one.
525 if (!L->isLoopSimplifyForm())
528 IU = &getAnalysis<IVUsers>();
529 LI = &getAnalysis<LoopInfo>();
530 SE = &getAnalysis<ScalarEvolution>();
531 DT = &getAnalysis<DominatorTree>();
532 TD = getAnalysisIfAvailable<TargetData>();
536 // If there are any floating-point recurrences, attempt to
537 // transform them to use integer recurrences.
538 RewriteNonIntegerIVs(L);
540 BasicBlock *ExitingBlock = L->getExitingBlock(); // may be null
541 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
543 // Create a rewriter object which we'll use to transform the code with.
544 SCEVExpander Rewriter(*SE);
546 // Check to see if this loop has a computable loop-invariant execution count.
547 // If so, this means that we can compute the final value of any expressions
548 // that are recurrent in the loop, and substitute the exit values from the
549 // loop into any instructions outside of the loop that use the final values of
550 // the current expressions.
552 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount))
553 RewriteLoopExitValues(L, Rewriter);
555 // Simplify ICmp IV users.
556 EliminateIVComparisons();
558 // Simplify SRem and URem IV users.
559 EliminateIVRemainders();
561 // Compute the type of the largest recurrence expression, and decide whether
562 // a canonical induction variable should be inserted.
563 const Type *LargestType = 0;
564 bool NeedCannIV = false;
565 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount)) {
566 LargestType = BackedgeTakenCount->getType();
567 LargestType = SE->getEffectiveSCEVType(LargestType);
568 // If we have a known trip count and a single exit block, we'll be
569 // rewriting the loop exit test condition below, which requires a
570 // canonical induction variable.
574 for (IVUsers::const_iterator I = IU->begin(), E = IU->end(); I != E; ++I) {
576 SE->getEffectiveSCEVType(I->getOperandValToReplace()->getType());
578 SE->getTypeSizeInBits(Ty) >
579 SE->getTypeSizeInBits(LargestType))
584 // Now that we know the largest of the induction variable expressions
585 // in this loop, insert a canonical induction variable of the largest size.
588 // Check to see if the loop already has any canonical-looking induction
589 // variables. If any are present and wider than the planned canonical
590 // induction variable, temporarily remove them, so that the Rewriter
591 // doesn't attempt to reuse them.
592 SmallVector<PHINode *, 2> OldCannIVs;
593 while (PHINode *OldCannIV = L->getCanonicalInductionVariable()) {
594 if (SE->getTypeSizeInBits(OldCannIV->getType()) >
595 SE->getTypeSizeInBits(LargestType))
596 OldCannIV->removeFromParent();
599 OldCannIVs.push_back(OldCannIV);
602 IndVar = Rewriter.getOrInsertCanonicalInductionVariable(L, LargestType);
606 DEBUG(dbgs() << "INDVARS: New CanIV: " << *IndVar << '\n');
608 // Now that the official induction variable is established, reinsert
609 // any old canonical-looking variables after it so that the IR remains
610 // consistent. They will be deleted as part of the dead-PHI deletion at
611 // the end of the pass.
612 while (!OldCannIVs.empty()) {
613 PHINode *OldCannIV = OldCannIVs.pop_back_val();
614 OldCannIV->insertBefore(L->getHeader()->getFirstNonPHI());
618 // If we have a trip count expression, rewrite the loop's exit condition
619 // using it. We can currently only handle loops with a single exit.
620 ICmpInst *NewICmp = 0;
621 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount) &&
622 !BackedgeTakenCount->isZero() &&
625 "LinearFunctionTestReplace requires a canonical induction variable");
626 // Can't rewrite non-branch yet.
627 if (BranchInst *BI = dyn_cast<BranchInst>(ExitingBlock->getTerminator()))
628 NewICmp = LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar,
629 ExitingBlock, BI, Rewriter);
632 // Rewrite IV-derived expressions.
633 RewriteIVExpressions(L, Rewriter);
635 // Clear the rewriter cache, because values that are in the rewriter's cache
636 // can be deleted in the loop below, causing the AssertingVH in the cache to
640 // Now that we're done iterating through lists, clean up any instructions
641 // which are now dead.
642 while (!DeadInsts.empty())
643 if (Instruction *Inst =
644 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
645 RecursivelyDeleteTriviallyDeadInstructions(Inst);
647 // The Rewriter may not be used from this point on.
649 // Loop-invariant instructions in the preheader that aren't used in the
650 // loop may be sunk below the loop to reduce register pressure.
651 SinkUnusedInvariants(L);
653 // For completeness, inform IVUsers of the IV use in the newly-created
654 // loop exit test instruction.
656 IU->AddUsersIfInteresting(cast<Instruction>(NewICmp->getOperand(0)));
658 // Clean up dead instructions.
659 Changed |= DeleteDeadPHIs(L->getHeader());
660 // Check a post-condition.
661 assert(L->isLCSSAForm(*DT) && "Indvars did not leave the loop in lcssa form!");
665 // FIXME: It is an extremely bad idea to indvar substitute anything more
666 // complex than affine induction variables. Doing so will put expensive
667 // polynomial evaluations inside of the loop, and the str reduction pass
668 // currently can only reduce affine polynomials. For now just disable
669 // indvar subst on anything more complex than an affine addrec, unless
670 // it can be expanded to a trivial value.
671 static bool isSafe(const SCEV *S, const Loop *L, ScalarEvolution *SE) {
672 // Loop-invariant values are safe.
673 if (SE->isLoopInvariant(S, L)) return true;
675 // Affine addrecs are safe. Non-affine are not, because LSR doesn't know how
676 // to transform them into efficient code.
677 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
678 return AR->isAffine();
680 // An add is safe it all its operands are safe.
681 if (const SCEVCommutativeExpr *Commutative = dyn_cast<SCEVCommutativeExpr>(S)) {
682 for (SCEVCommutativeExpr::op_iterator I = Commutative->op_begin(),
683 E = Commutative->op_end(); I != E; ++I)
684 if (!isSafe(*I, L, SE)) return false;
688 // A cast is safe if its operand is.
689 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
690 return isSafe(C->getOperand(), L, SE);
692 // A udiv is safe if its operands are.
693 if (const SCEVUDivExpr *UD = dyn_cast<SCEVUDivExpr>(S))
694 return isSafe(UD->getLHS(), L, SE) &&
695 isSafe(UD->getRHS(), L, SE);
697 // SCEVUnknown is always safe.
698 if (isa<SCEVUnknown>(S))
701 // Nothing else is safe.
705 void IndVarSimplify::RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter) {
706 // Rewrite all induction variable expressions in terms of the canonical
707 // induction variable.
709 // If there were induction variables of other sizes or offsets, manually
710 // add the offsets to the primary induction variable and cast, avoiding
711 // the need for the code evaluation methods to insert induction variables
712 // of different sizes.
713 for (IVUsers::iterator UI = IU->begin(), E = IU->end(); UI != E; ++UI) {
714 Value *Op = UI->getOperandValToReplace();
715 const Type *UseTy = Op->getType();
716 Instruction *User = UI->getUser();
718 // Compute the final addrec to expand into code.
719 const SCEV *AR = IU->getReplacementExpr(*UI);
721 // Evaluate the expression out of the loop, if possible.
722 if (!L->contains(UI->getUser())) {
723 const SCEV *ExitVal = SE->getSCEVAtScope(AR, L->getParentLoop());
724 if (SE->isLoopInvariant(ExitVal, L))
728 // FIXME: It is an extremely bad idea to indvar substitute anything more
729 // complex than affine induction variables. Doing so will put expensive
730 // polynomial evaluations inside of the loop, and the str reduction pass
731 // currently can only reduce affine polynomials. For now just disable
732 // indvar subst on anything more complex than an affine addrec, unless
733 // it can be expanded to a trivial value.
734 if (!isSafe(AR, L, SE))
737 // Determine the insertion point for this user. By default, insert
738 // immediately before the user. The SCEVExpander class will automatically
739 // hoist loop invariants out of the loop. For PHI nodes, there may be
740 // multiple uses, so compute the nearest common dominator for the
742 Instruction *InsertPt = User;
743 if (PHINode *PHI = dyn_cast<PHINode>(InsertPt))
744 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
745 if (PHI->getIncomingValue(i) == Op) {
746 if (InsertPt == User)
747 InsertPt = PHI->getIncomingBlock(i)->getTerminator();
750 DT->findNearestCommonDominator(InsertPt->getParent(),
751 PHI->getIncomingBlock(i))
755 // Now expand it into actual Instructions and patch it into place.
756 Value *NewVal = Rewriter.expandCodeFor(AR, UseTy, InsertPt);
758 DEBUG(dbgs() << "INDVARS: Rewrote IV '" << *AR << "' " << *Op << '\n'
759 << " into = " << *NewVal << "\n");
761 if (!isValidRewrite(Op, NewVal)) {
762 DeadInsts.push_back(NewVal);
765 // Inform ScalarEvolution that this value is changing. The change doesn't
766 // affect its value, but it does potentially affect which use lists the
767 // value will be on after the replacement, which affects ScalarEvolution's
768 // ability to walk use lists and drop dangling pointers when a value is
770 SE->forgetValue(User);
772 // Patch the new value into place.
774 NewVal->takeName(Op);
775 User->replaceUsesOfWith(Op, NewVal);
776 UI->setOperandValToReplace(NewVal);
781 // The old value may be dead now.
782 DeadInsts.push_back(Op);
786 /// If there's a single exit block, sink any loop-invariant values that
787 /// were defined in the preheader but not used inside the loop into the
788 /// exit block to reduce register pressure in the loop.
789 void IndVarSimplify::SinkUnusedInvariants(Loop *L) {
790 BasicBlock *ExitBlock = L->getExitBlock();
791 if (!ExitBlock) return;
793 BasicBlock *Preheader = L->getLoopPreheader();
794 if (!Preheader) return;
796 Instruction *InsertPt = ExitBlock->getFirstNonPHI();
797 BasicBlock::iterator I = Preheader->getTerminator();
798 while (I != Preheader->begin()) {
800 // New instructions were inserted at the end of the preheader.
804 // Don't move instructions which might have side effects, since the side
805 // effects need to complete before instructions inside the loop. Also don't
806 // move instructions which might read memory, since the loop may modify
807 // memory. Note that it's okay if the instruction might have undefined
808 // behavior: LoopSimplify guarantees that the preheader dominates the exit
810 if (I->mayHaveSideEffects() || I->mayReadFromMemory())
813 // Skip debug info intrinsics.
814 if (isa<DbgInfoIntrinsic>(I))
817 // Don't sink static AllocaInsts out of the entry block, which would
818 // turn them into dynamic allocas!
819 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
820 if (AI->isStaticAlloca())
823 // Determine if there is a use in or before the loop (direct or
825 bool UsedInLoop = false;
826 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
829 BasicBlock *UseBB = cast<Instruction>(U)->getParent();
830 if (PHINode *P = dyn_cast<PHINode>(U)) {
832 PHINode::getIncomingValueNumForOperand(UI.getOperandNo());
833 UseBB = P->getIncomingBlock(i);
835 if (UseBB == Preheader || L->contains(UseBB)) {
841 // If there is, the def must remain in the preheader.
845 // Otherwise, sink it to the exit block.
846 Instruction *ToMove = I;
849 if (I != Preheader->begin()) {
850 // Skip debug info intrinsics.
853 } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin());
855 if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin())
861 ToMove->moveBefore(InsertPt);
867 /// ConvertToSInt - Convert APF to an integer, if possible.
868 static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) {
869 bool isExact = false;
870 if (&APF.getSemantics() == &APFloat::PPCDoubleDouble)
872 // See if we can convert this to an int64_t
874 if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero,
875 &isExact) != APFloat::opOK || !isExact)
881 /// HandleFloatingPointIV - If the loop has floating induction variable
882 /// then insert corresponding integer induction variable if possible.
884 /// for(double i = 0; i < 10000; ++i)
886 /// is converted into
887 /// for(int i = 0; i < 10000; ++i)
890 void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) {
891 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
892 unsigned BackEdge = IncomingEdge^1;
894 // Check incoming value.
895 ConstantFP *InitValueVal =
896 dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));
899 if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
902 // Check IV increment. Reject this PN if increment operation is not
903 // an add or increment value can not be represented by an integer.
904 BinaryOperator *Incr =
905 dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
906 if (Incr == 0 || Incr->getOpcode() != Instruction::FAdd) return;
908 // If this is not an add of the PHI with a constantfp, or if the constant fp
909 // is not an integer, bail out.
910 ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
912 if (IncValueVal == 0 || Incr->getOperand(0) != PN ||
913 !ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
916 // Check Incr uses. One user is PN and the other user is an exit condition
917 // used by the conditional terminator.
918 Value::use_iterator IncrUse = Incr->use_begin();
919 Instruction *U1 = cast<Instruction>(*IncrUse++);
920 if (IncrUse == Incr->use_end()) return;
921 Instruction *U2 = cast<Instruction>(*IncrUse++);
922 if (IncrUse != Incr->use_end()) return;
924 // Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't
925 // only used by a branch, we can't transform it.
926 FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
928 Compare = dyn_cast<FCmpInst>(U2);
929 if (Compare == 0 || !Compare->hasOneUse() ||
930 !isa<BranchInst>(Compare->use_back()))
933 BranchInst *TheBr = cast<BranchInst>(Compare->use_back());
935 // We need to verify that the branch actually controls the iteration count
936 // of the loop. If not, the new IV can overflow and no one will notice.
937 // The branch block must be in the loop and one of the successors must be out
939 assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
940 if (!L->contains(TheBr->getParent()) ||
941 (L->contains(TheBr->getSuccessor(0)) &&
942 L->contains(TheBr->getSuccessor(1))))
946 // If it isn't a comparison with an integer-as-fp (the exit value), we can't
948 ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
950 if (ExitValueVal == 0 ||
951 !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
954 // Find new predicate for integer comparison.
955 CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
956 switch (Compare->getPredicate()) {
957 default: return; // Unknown comparison.
958 case CmpInst::FCMP_OEQ:
959 case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
960 case CmpInst::FCMP_ONE:
961 case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
962 case CmpInst::FCMP_OGT:
963 case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
964 case CmpInst::FCMP_OGE:
965 case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
966 case CmpInst::FCMP_OLT:
967 case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
968 case CmpInst::FCMP_OLE:
969 case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
972 // We convert the floating point induction variable to a signed i32 value if
973 // we can. This is only safe if the comparison will not overflow in a way
974 // that won't be trapped by the integer equivalent operations. Check for this
976 // TODO: We could use i64 if it is native and the range requires it.
978 // The start/stride/exit values must all fit in signed i32.
979 if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
982 // If not actually striding (add x, 0.0), avoid touching the code.
986 // Positive and negative strides have different safety conditions.
988 // If we have a positive stride, we require the init to be less than the
989 // exit value and an equality or less than comparison.
990 if (InitValue >= ExitValue ||
991 NewPred == CmpInst::ICMP_SGT || NewPred == CmpInst::ICMP_SGE)
994 uint32_t Range = uint32_t(ExitValue-InitValue);
995 if (NewPred == CmpInst::ICMP_SLE) {
996 // Normalize SLE -> SLT, check for infinite loop.
997 if (++Range == 0) return; // Range overflows.
1000 unsigned Leftover = Range % uint32_t(IncValue);
1002 // If this is an equality comparison, we require that the strided value
1003 // exactly land on the exit value, otherwise the IV condition will wrap
1004 // around and do things the fp IV wouldn't.
1005 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
1009 // If the stride would wrap around the i32 before exiting, we can't
1010 // transform the IV.
1011 if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
1015 // If we have a negative stride, we require the init to be greater than the
1016 // exit value and an equality or greater than comparison.
1017 if (InitValue >= ExitValue ||
1018 NewPred == CmpInst::ICMP_SLT || NewPred == CmpInst::ICMP_SLE)
1021 uint32_t Range = uint32_t(InitValue-ExitValue);
1022 if (NewPred == CmpInst::ICMP_SGE) {
1023 // Normalize SGE -> SGT, check for infinite loop.
1024 if (++Range == 0) return; // Range overflows.
1027 unsigned Leftover = Range % uint32_t(-IncValue);
1029 // If this is an equality comparison, we require that the strided value
1030 // exactly land on the exit value, otherwise the IV condition will wrap
1031 // around and do things the fp IV wouldn't.
1032 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
1036 // If the stride would wrap around the i32 before exiting, we can't
1037 // transform the IV.
1038 if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
1042 const IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
1044 // Insert new integer induction variable.
1045 PHINode *NewPHI = PHINode::Create(Int32Ty, PN->getName()+".int", PN);
1046 NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
1047 PN->getIncomingBlock(IncomingEdge));
1050 BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
1051 Incr->getName()+".int", Incr);
1052 NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
1054 ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
1055 ConstantInt::get(Int32Ty, ExitValue),
1056 Compare->getName());
1058 // In the following deletions, PN may become dead and may be deleted.
1059 // Use a WeakVH to observe whether this happens.
1062 // Delete the old floating point exit comparison. The branch starts using the
1064 NewCompare->takeName(Compare);
1065 Compare->replaceAllUsesWith(NewCompare);
1066 RecursivelyDeleteTriviallyDeadInstructions(Compare);
1068 // Delete the old floating point increment.
1069 Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
1070 RecursivelyDeleteTriviallyDeadInstructions(Incr);
1072 // If the FP induction variable still has uses, this is because something else
1073 // in the loop uses its value. In order to canonicalize the induction
1074 // variable, we chose to eliminate the IV and rewrite it in terms of an
1077 // We give preference to sitofp over uitofp because it is faster on most
1080 Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
1081 PN->getParent()->getFirstNonPHI());
1082 PN->replaceAllUsesWith(Conv);
1083 RecursivelyDeleteTriviallyDeadInstructions(PN);
1086 // Add a new IVUsers entry for the newly-created integer PHI.
1087 IU->AddUsersIfInteresting(NewPHI);