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/Debug.h"
56 #include "llvm/Support/raw_ostream.h"
57 #include "llvm/Transforms/Utils/Local.h"
58 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
59 #include "llvm/Target/TargetData.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(NumWidened , "Number of indvars widened");
67 STATISTIC(NumInserted, "Number of canonical indvars added");
68 STATISTIC(NumReplaced, "Number of exit values replaced");
69 STATISTIC(NumLFTR , "Number of loop exit tests replaced");
71 // DisableIVRewrite mode currently affects IVUsers, so is defined in libAnalysis
72 // and referenced here.
74 extern bool DisableIVRewrite;
78 class IndVarSimplify : public LoopPass {
84 SmallVector<WeakVH, 16> DeadInsts;
88 static char ID; // Pass identification, replacement for typeid
89 IndVarSimplify() : LoopPass(ID), IU(0), LI(0), SE(0), DT(0), TD(0) {
90 initializeIndVarSimplifyPass(*PassRegistry::getPassRegistry());
93 virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
95 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
96 AU.addRequired<DominatorTree>();
97 AU.addRequired<LoopInfo>();
98 AU.addRequired<ScalarEvolution>();
99 AU.addRequiredID(LoopSimplifyID);
100 AU.addRequiredID(LCSSAID);
101 AU.addRequired<IVUsers>();
102 AU.addPreserved<ScalarEvolution>();
103 AU.addPreservedID(LoopSimplifyID);
104 AU.addPreservedID(LCSSAID);
105 AU.addPreserved<IVUsers>();
106 AU.setPreservesCFG();
110 bool isValidRewrite(Value *FromVal, Value *ToVal);
112 void SimplifyIVUsers();
113 void EliminateIVComparison(ICmpInst *ICmp, Value *IVOperand);
114 void EliminateIVRemainder(BinaryOperator *Rem,
117 void RewriteNonIntegerIVs(Loop *L);
118 const Type *WidenIVs(Loop *L, SCEVExpander &Rewriter);
120 bool canExpandBackedgeTakenCount(Loop *L,
121 const SCEV *BackedgeTakenCount);
123 ICmpInst *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
125 SCEVExpander &Rewriter);
127 void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter);
129 void RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter);
131 void SinkUnusedInvariants(Loop *L);
133 void HandleFloatingPointIV(Loop *L, PHINode *PH);
137 char IndVarSimplify::ID = 0;
138 INITIALIZE_PASS_BEGIN(IndVarSimplify, "indvars",
139 "Induction Variable Simplification", false, false)
140 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
141 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
142 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
143 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
144 INITIALIZE_PASS_DEPENDENCY(LCSSA)
145 INITIALIZE_PASS_DEPENDENCY(IVUsers)
146 INITIALIZE_PASS_END(IndVarSimplify, "indvars",
147 "Induction Variable Simplification", false, false)
149 Pass *llvm::createIndVarSimplifyPass() {
150 return new IndVarSimplify();
153 /// isValidRewrite - Return true if the SCEV expansion generated by the
154 /// rewriter can replace the original value. SCEV guarantees that it
155 /// produces the same value, but the way it is produced may be illegal IR.
156 /// Ideally, this function will only be called for verification.
157 bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) {
158 // If an SCEV expression subsumed multiple pointers, its expansion could
159 // reassociate the GEP changing the base pointer. This is illegal because the
160 // final address produced by a GEP chain must be inbounds relative to its
161 // underlying object. Otherwise basic alias analysis, among other things,
162 // could fail in a dangerous way. Ultimately, SCEV will be improved to avoid
163 // producing an expression involving multiple pointers. Until then, we must
166 // Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject
167 // because it understands lcssa phis while SCEV does not.
168 Value *FromPtr = FromVal;
169 Value *ToPtr = ToVal;
170 if (GEPOperator *GEP = dyn_cast<GEPOperator>(FromVal)) {
171 FromPtr = GEP->getPointerOperand();
173 if (GEPOperator *GEP = dyn_cast<GEPOperator>(ToVal)) {
174 ToPtr = GEP->getPointerOperand();
176 if (FromPtr != FromVal || ToPtr != ToVal) {
177 // Quickly check the common case
178 if (FromPtr == ToPtr)
181 // SCEV may have rewritten an expression that produces the GEP's pointer
182 // operand. That's ok as long as the pointer operand has the same base
183 // pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the
184 // base of a recurrence. This handles the case in which SCEV expansion
185 // converts a pointer type recurrence into a nonrecurrent pointer base
186 // indexed by an integer recurrence.
187 const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr));
188 const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr));
189 if (FromBase == ToBase)
192 DEBUG(dbgs() << "INDVARS: GEP rewrite bail out "
193 << *FromBase << " != " << *ToBase << "\n");
200 /// canExpandBackedgeTakenCount - Return true if this loop's backedge taken
201 /// count expression can be safely and cheaply expanded into an instruction
202 /// sequence that can be used by LinearFunctionTestReplace.
203 bool IndVarSimplify::
204 canExpandBackedgeTakenCount(Loop *L,
205 const SCEV *BackedgeTakenCount) {
206 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) ||
207 BackedgeTakenCount->isZero())
210 if (!L->getExitingBlock())
213 // Can't rewrite non-branch yet.
214 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
218 // Special case: If the backedge-taken count is a UDiv, it's very likely a
219 // UDiv that ScalarEvolution produced in order to compute a precise
220 // expression, rather than a UDiv from the user's code. If we can't find a
221 // UDiv in the code with some simple searching, assume the former and forego
222 // rewriting the loop.
223 if (isa<SCEVUDivExpr>(BackedgeTakenCount)) {
224 ICmpInst *OrigCond = dyn_cast<ICmpInst>(BI->getCondition());
225 if (!OrigCond) return false;
226 const SCEV *R = SE->getSCEV(OrigCond->getOperand(1));
227 R = SE->getMinusSCEV(R, SE->getConstant(R->getType(), 1));
228 if (R != BackedgeTakenCount) {
229 const SCEV *L = SE->getSCEV(OrigCond->getOperand(0));
230 L = SE->getMinusSCEV(L, SE->getConstant(L->getType(), 1));
231 if (L != BackedgeTakenCount)
238 /// LinearFunctionTestReplace - This method rewrites the exit condition of the
239 /// loop to be a canonical != comparison against the incremented loop induction
240 /// variable. This pass is able to rewrite the exit tests of any loop where the
241 /// SCEV analysis can determine a loop-invariant trip count of the loop, which
242 /// is actually a much broader range than just linear tests.
243 ICmpInst *IndVarSimplify::
244 LinearFunctionTestReplace(Loop *L,
245 const SCEV *BackedgeTakenCount,
247 SCEVExpander &Rewriter) {
248 assert(canExpandBackedgeTakenCount(L, BackedgeTakenCount) && "precondition");
249 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
251 // If the exiting block is not the same as the backedge block, we must compare
252 // against the preincremented value, otherwise we prefer to compare against
253 // the post-incremented value.
255 const SCEV *RHS = BackedgeTakenCount;
256 if (L->getExitingBlock() == L->getLoopLatch()) {
257 // Add one to the "backedge-taken" count to get the trip count.
258 // If this addition may overflow, we have to be more pessimistic and
259 // cast the induction variable before doing the add.
260 const SCEV *Zero = SE->getConstant(BackedgeTakenCount->getType(), 0);
262 SE->getAddExpr(BackedgeTakenCount,
263 SE->getConstant(BackedgeTakenCount->getType(), 1));
264 if ((isa<SCEVConstant>(N) && !N->isZero()) ||
265 SE->isLoopEntryGuardedByCond(L, ICmpInst::ICMP_NE, N, Zero)) {
266 // No overflow. Cast the sum.
267 RHS = SE->getTruncateOrZeroExtend(N, IndVar->getType());
269 // Potential overflow. Cast before doing the add.
270 RHS = SE->getTruncateOrZeroExtend(BackedgeTakenCount,
272 RHS = SE->getAddExpr(RHS,
273 SE->getConstant(IndVar->getType(), 1));
276 // The BackedgeTaken expression contains the number of times that the
277 // backedge branches to the loop header. This is one less than the
278 // number of times the loop executes, so use the incremented indvar.
279 CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock());
281 // We have to use the preincremented value...
282 RHS = SE->getTruncateOrZeroExtend(BackedgeTakenCount,
287 // Expand the code for the iteration count.
288 assert(SE->isLoopInvariant(RHS, L) &&
289 "Computed iteration count is not loop invariant!");
290 Value *ExitCnt = Rewriter.expandCodeFor(RHS, IndVar->getType(), BI);
292 // Insert a new icmp_ne or icmp_eq instruction before the branch.
293 ICmpInst::Predicate Opcode;
294 if (L->contains(BI->getSuccessor(0)))
295 Opcode = ICmpInst::ICMP_NE;
297 Opcode = ICmpInst::ICMP_EQ;
299 DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
300 << " LHS:" << *CmpIndVar << '\n'
302 << (Opcode == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
303 << " RHS:\t" << *RHS << "\n");
305 ICmpInst *Cond = new ICmpInst(BI, Opcode, CmpIndVar, ExitCnt, "exitcond");
307 Value *OrigCond = BI->getCondition();
308 // It's tempting to use replaceAllUsesWith here to fully replace the old
309 // comparison, but that's not immediately safe, since users of the old
310 // comparison may not be dominated by the new comparison. Instead, just
311 // update the branch to use the new comparison; in the common case this
312 // will make old comparison dead.
313 BI->setCondition(Cond);
314 DeadInsts.push_back(OrigCond);
321 /// RewriteLoopExitValues - Check to see if this loop has a computable
322 /// loop-invariant execution count. If so, this means that we can compute the
323 /// final value of any expressions that are recurrent in the loop, and
324 /// substitute the exit values from the loop into any instructions outside of
325 /// the loop that use the final values of the current expressions.
327 /// This is mostly redundant with the regular IndVarSimplify activities that
328 /// happen later, except that it's more powerful in some cases, because it's
329 /// able to brute-force evaluate arbitrary instructions as long as they have
330 /// constant operands at the beginning of the loop.
331 void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) {
332 // Verify the input to the pass in already in LCSSA form.
333 assert(L->isLCSSAForm(*DT));
335 SmallVector<BasicBlock*, 8> ExitBlocks;
336 L->getUniqueExitBlocks(ExitBlocks);
338 // Find all values that are computed inside the loop, but used outside of it.
339 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan
340 // the exit blocks of the loop to find them.
341 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
342 BasicBlock *ExitBB = ExitBlocks[i];
344 // If there are no PHI nodes in this exit block, then no values defined
345 // inside the loop are used on this path, skip it.
346 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
349 unsigned NumPreds = PN->getNumIncomingValues();
351 // Iterate over all of the PHI nodes.
352 BasicBlock::iterator BBI = ExitBB->begin();
353 while ((PN = dyn_cast<PHINode>(BBI++))) {
355 continue; // dead use, don't replace it
357 // SCEV only supports integer expressions for now.
358 if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy())
361 // It's necessary to tell ScalarEvolution about this explicitly so that
362 // it can walk the def-use list and forget all SCEVs, as it may not be
363 // watching the PHI itself. Once the new exit value is in place, there
364 // may not be a def-use connection between the loop and every instruction
365 // which got a SCEVAddRecExpr for that loop.
368 // Iterate over all of the values in all the PHI nodes.
369 for (unsigned i = 0; i != NumPreds; ++i) {
370 // If the value being merged in is not integer or is not defined
371 // in the loop, skip it.
372 Value *InVal = PN->getIncomingValue(i);
373 if (!isa<Instruction>(InVal))
376 // If this pred is for a subloop, not L itself, skip it.
377 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
378 continue; // The Block is in a subloop, skip it.
380 // Check that InVal is defined in the loop.
381 Instruction *Inst = cast<Instruction>(InVal);
382 if (!L->contains(Inst))
385 // Okay, this instruction has a user outside of the current loop
386 // and varies predictably *inside* the loop. Evaluate the value it
387 // contains when the loop exits, if possible.
388 const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
389 if (!SE->isLoopInvariant(ExitValue, L))
392 Value *ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), Inst);
394 DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
395 << " LoopVal = " << *Inst << "\n");
397 if (!isValidRewrite(Inst, ExitVal)) {
398 DeadInsts.push_back(ExitVal);
404 PN->setIncomingValue(i, ExitVal);
406 // If this instruction is dead now, delete it.
407 RecursivelyDeleteTriviallyDeadInstructions(Inst);
410 // Completely replace a single-pred PHI. This is safe, because the
411 // NewVal won't be variant in the loop, so we don't need an LCSSA phi
413 PN->replaceAllUsesWith(ExitVal);
414 RecursivelyDeleteTriviallyDeadInstructions(PN);
418 // Clone the PHI and delete the original one. This lets IVUsers and
419 // any other maps purge the original user from their records.
420 PHINode *NewPN = cast<PHINode>(PN->clone());
422 NewPN->insertBefore(PN);
423 PN->replaceAllUsesWith(NewPN);
424 PN->eraseFromParent();
429 // The insertion point instruction may have been deleted; clear it out
430 // so that the rewriter doesn't trip over it later.
431 Rewriter.clearInsertPoint();
434 void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
435 // First step. Check to see if there are any floating-point recurrences.
436 // If there are, change them into integer recurrences, permitting analysis by
437 // the SCEV routines.
439 BasicBlock *Header = L->getHeader();
441 SmallVector<WeakVH, 8> PHIs;
442 for (BasicBlock::iterator I = Header->begin();
443 PHINode *PN = dyn_cast<PHINode>(I); ++I)
446 for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
447 if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i]))
448 HandleFloatingPointIV(L, PN);
450 // If the loop previously had floating-point IV, ScalarEvolution
451 // may not have been able to compute a trip count. Now that we've done some
452 // re-writing, the trip count may be computable.
457 /// SimplifyIVUsers - Iteratively perform simplification on IVUsers within this
458 /// loop. IVUsers is treated as a worklist. Each successive simplification may
459 /// push more users which may themselves be candidates for simplification.
460 void IndVarSimplify::SimplifyIVUsers() {
461 for (IVUsers::iterator I = IU->begin(); I != IU->end(); ++I) {
462 Instruction *UseInst = I->getUser();
463 Value *IVOperand = I->getOperandValToReplace();
465 if (ICmpInst *ICmp = dyn_cast<ICmpInst>(UseInst)) {
466 EliminateIVComparison(ICmp, IVOperand);
470 if (BinaryOperator *Rem = dyn_cast<BinaryOperator>(UseInst)) {
471 bool isSigned = Rem->getOpcode() == Instruction::SRem;
472 if (isSigned || Rem->getOpcode() == Instruction::URem) {
473 EliminateIVRemainder(Rem, IVOperand, isSigned);
480 void IndVarSimplify::EliminateIVComparison(ICmpInst *ICmp, Value *IVOperand) {
481 unsigned IVOperIdx = 0;
482 ICmpInst::Predicate Pred = ICmp->getPredicate();
483 if (IVOperand != ICmp->getOperand(0)) {
485 assert(IVOperand == ICmp->getOperand(1) && "Can't find IVOperand");
487 Pred = ICmpInst::getSwappedPredicate(Pred);
490 // Get the SCEVs for the ICmp operands.
491 const SCEV *S = SE->getSCEV(ICmp->getOperand(IVOperIdx));
492 const SCEV *X = SE->getSCEV(ICmp->getOperand(1 - IVOperIdx));
494 // Simplify unnecessary loops away.
495 const Loop *ICmpLoop = LI->getLoopFor(ICmp->getParent());
496 S = SE->getSCEVAtScope(S, ICmpLoop);
497 X = SE->getSCEVAtScope(X, ICmpLoop);
499 // If the condition is always true or always false, replace it with
501 if (SE->isKnownPredicate(Pred, S, X))
502 ICmp->replaceAllUsesWith(ConstantInt::getTrue(ICmp->getContext()));
503 else if (SE->isKnownPredicate(ICmpInst::getInversePredicate(Pred), S, X))
504 ICmp->replaceAllUsesWith(ConstantInt::getFalse(ICmp->getContext()));
508 DEBUG(dbgs() << "INDVARS: Eliminated comparison: " << *ICmp << '\n');
509 DeadInsts.push_back(ICmp);
512 void IndVarSimplify::EliminateIVRemainder(BinaryOperator *Rem,
515 // We're only interested in the case where we know something about
517 if (IVOperand != Rem->getOperand(0))
520 // Get the SCEVs for the ICmp operands.
521 const SCEV *S = SE->getSCEV(Rem->getOperand(0));
522 const SCEV *X = SE->getSCEV(Rem->getOperand(1));
524 // Simplify unnecessary loops away.
525 const Loop *ICmpLoop = LI->getLoopFor(Rem->getParent());
526 S = SE->getSCEVAtScope(S, ICmpLoop);
527 X = SE->getSCEVAtScope(X, ICmpLoop);
529 // i % n --> i if i is in [0,n).
530 if ((!isSigned || SE->isKnownNonNegative(S)) &&
531 SE->isKnownPredicate(isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
533 Rem->replaceAllUsesWith(Rem->getOperand(0));
535 // (i+1) % n --> (i+1)==n?0:(i+1) if i is in [0,n).
536 const SCEV *LessOne =
537 SE->getMinusSCEV(S, SE->getConstant(S->getType(), 1));
538 if (isSigned && !SE->isKnownNonNegative(LessOne))
541 if (!SE->isKnownPredicate(isSigned ?
542 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
546 ICmpInst *ICmp = new ICmpInst(Rem, ICmpInst::ICMP_EQ,
547 Rem->getOperand(0), Rem->getOperand(1),
550 SelectInst::Create(ICmp,
551 ConstantInt::get(Rem->getType(), 0),
552 Rem->getOperand(0), "tmp", Rem);
553 Rem->replaceAllUsesWith(Sel);
556 // Inform IVUsers about the new users.
557 if (Instruction *I = dyn_cast<Instruction>(Rem->getOperand(0)))
558 IU->AddUsersIfInteresting(I);
560 DEBUG(dbgs() << "INDVARS: Simplified rem: " << *Rem << '\n');
561 DeadInsts.push_back(Rem);
564 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
565 // If LoopSimplify form is not available, stay out of trouble. Some notes:
566 // - LSR currently only supports LoopSimplify-form loops. Indvars'
567 // canonicalization can be a pessimization without LSR to "clean up"
569 // - We depend on having a preheader; in particular,
570 // Loop::getCanonicalInductionVariable only supports loops with preheaders,
571 // and we're in trouble if we can't find the induction variable even when
572 // we've manually inserted one.
573 if (!L->isLoopSimplifyForm())
576 IU = &getAnalysis<IVUsers>();
577 LI = &getAnalysis<LoopInfo>();
578 SE = &getAnalysis<ScalarEvolution>();
579 DT = &getAnalysis<DominatorTree>();
580 TD = getAnalysisIfAvailable<TargetData>();
585 // If there are any floating-point recurrences, attempt to
586 // transform them to use integer recurrences.
587 RewriteNonIntegerIVs(L);
589 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
591 // Create a rewriter object which we'll use to transform the code with.
592 SCEVExpander Rewriter(*SE);
593 if (DisableIVRewrite)
594 Rewriter.disableCanonicalMode();
596 const Type *LargestType = 0;
597 if (DisableIVRewrite) {
598 LargestType = WidenIVs(L, Rewriter);
601 // Check to see if this loop has a computable loop-invariant execution count.
602 // If so, this means that we can compute the final value of any expressions
603 // that are recurrent in the loop, and substitute the exit values from the
604 // loop into any instructions outside of the loop that use the final values of
605 // the current expressions.
607 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount))
608 RewriteLoopExitValues(L, Rewriter);
612 // Compute the type of the largest recurrence expression, and decide whether
613 // a canonical induction variable should be inserted.
614 bool NeedCannIV = false;
615 bool ExpandBECount = canExpandBackedgeTakenCount(L, BackedgeTakenCount);
617 // If we have a known trip count and a single exit block, we'll be
618 // rewriting the loop exit test condition below, which requires a
619 // canonical induction variable.
621 const Type *Ty = BackedgeTakenCount->getType();
623 SE->getTypeSizeInBits(Ty) >
624 SE->getTypeSizeInBits(LargestType))
625 LargestType = SE->getEffectiveSCEVType(Ty);
627 for (IVUsers::const_iterator I = IU->begin(), E = IU->end(); I != E; ++I) {
630 SE->getEffectiveSCEVType(I->getOperandValToReplace()->getType());
632 SE->getTypeSizeInBits(Ty) >
633 SE->getTypeSizeInBits(LargestType))
634 LargestType = SE->getEffectiveSCEVType(Ty);
636 if (!DisableIVRewrite) {
637 for (IVUsers::const_iterator I = IU->begin(), E = IU->end(); I != E; ++I) {
640 SE->getEffectiveSCEVType(I->getOperandValToReplace()->getType());
642 SE->getTypeSizeInBits(Ty) >
643 SE->getTypeSizeInBits(LargestType))
648 // Now that we know the largest of the induction variable expressions
649 // in this loop, insert a canonical induction variable of the largest size.
652 // Check to see if the loop already has any canonical-looking induction
653 // variables. If any are present and wider than the planned canonical
654 // induction variable, temporarily remove them, so that the Rewriter
655 // doesn't attempt to reuse them.
656 SmallVector<PHINode *, 2> OldCannIVs;
657 while (PHINode *OldCannIV = L->getCanonicalInductionVariable()) {
658 if (SE->getTypeSizeInBits(OldCannIV->getType()) >
659 SE->getTypeSizeInBits(LargestType))
660 OldCannIV->removeFromParent();
663 OldCannIVs.push_back(OldCannIV);
666 IndVar = Rewriter.getOrInsertCanonicalInductionVariable(L, LargestType);
670 DEBUG(dbgs() << "INDVARS: New CanIV: " << *IndVar << '\n');
672 // Now that the official induction variable is established, reinsert
673 // any old canonical-looking variables after it so that the IR remains
674 // consistent. They will be deleted as part of the dead-PHI deletion at
675 // the end of the pass.
676 while (!OldCannIVs.empty()) {
677 PHINode *OldCannIV = OldCannIVs.pop_back_val();
678 OldCannIV->insertBefore(L->getHeader()->getFirstNonPHI());
682 // If we have a trip count expression, rewrite the loop's exit condition
683 // using it. We can currently only handle loops with a single exit.
684 ICmpInst *NewICmp = 0;
686 assert(canExpandBackedgeTakenCount(L, BackedgeTakenCount) &&
687 "canonical IV disrupted BackedgeTaken expansion");
689 "LinearFunctionTestReplace requires a canonical induction variable");
690 NewICmp = LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar,
693 // Rewrite IV-derived expressions.
694 if (!DisableIVRewrite)
695 RewriteIVExpressions(L, Rewriter);
697 // Clear the rewriter cache, because values that are in the rewriter's cache
698 // can be deleted in the loop below, causing the AssertingVH in the cache to
702 // Now that we're done iterating through lists, clean up any instructions
703 // which are now dead.
704 while (!DeadInsts.empty())
705 if (Instruction *Inst =
706 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
707 RecursivelyDeleteTriviallyDeadInstructions(Inst);
709 // The Rewriter may not be used from this point on.
711 // Loop-invariant instructions in the preheader that aren't used in the
712 // loop may be sunk below the loop to reduce register pressure.
713 SinkUnusedInvariants(L);
715 // For completeness, inform IVUsers of the IV use in the newly-created
716 // loop exit test instruction.
718 IU->AddUsersIfInteresting(cast<Instruction>(NewICmp->getOperand(0)));
720 // Clean up dead instructions.
721 Changed |= DeleteDeadPHIs(L->getHeader());
722 // Check a post-condition.
723 assert(L->isLCSSAForm(*DT) && "Indvars did not leave the loop in lcssa form!");
727 // FIXME: It is an extremely bad idea to indvar substitute anything more
728 // complex than affine induction variables. Doing so will put expensive
729 // polynomial evaluations inside of the loop, and the str reduction pass
730 // currently can only reduce affine polynomials. For now just disable
731 // indvar subst on anything more complex than an affine addrec, unless
732 // it can be expanded to a trivial value.
733 static bool isSafe(const SCEV *S, const Loop *L, ScalarEvolution *SE) {
734 // Loop-invariant values are safe.
735 if (SE->isLoopInvariant(S, L)) return true;
737 // Affine addrecs are safe. Non-affine are not, because LSR doesn't know how
738 // to transform them into efficient code.
739 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
740 return AR->isAffine();
742 // An add is safe it all its operands are safe.
743 if (const SCEVCommutativeExpr *Commutative = dyn_cast<SCEVCommutativeExpr>(S)) {
744 for (SCEVCommutativeExpr::op_iterator I = Commutative->op_begin(),
745 E = Commutative->op_end(); I != E; ++I)
746 if (!isSafe(*I, L, SE)) return false;
750 // A cast is safe if its operand is.
751 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
752 return isSafe(C->getOperand(), L, SE);
754 // A udiv is safe if its operands are.
755 if (const SCEVUDivExpr *UD = dyn_cast<SCEVUDivExpr>(S))
756 return isSafe(UD->getLHS(), L, SE) &&
757 isSafe(UD->getRHS(), L, SE);
759 // SCEVUnknown is always safe.
760 if (isa<SCEVUnknown>(S))
763 // Nothing else is safe.
767 /// Widen the type of any induction variables that are sign/zero extended and
768 /// remove the [sz]ext uses.
770 /// FIXME: This may currently create extra IVs which could increase regpressure
771 /// (without LSR to cleanup).
773 /// FIXME: may factor this with RewriteIVExpressions once it stabilizes.
774 const Type *IndVarSimplify::WidenIVs(Loop *L, SCEVExpander &Rewriter) {
775 const Type *LargestType = 0;
776 for (IVUsers::iterator UI = IU->begin(), E = IU->end(); UI != E; ++UI) {
777 Instruction *ExtInst = UI->getUser();
778 if (!isa<SExtInst>(ExtInst) && !isa<ZExtInst>(ExtInst))
780 const SCEV *AR = SE->getSCEV(ExtInst);
781 // Only widen this IV is SCEV tells us it's safe.
782 if (!isa<SCEVAddRecExpr>(AR) && !isa<SCEVAddExpr>(AR))
785 if (!L->contains(UI->getUser())) {
786 const SCEV *ExitVal = SE->getSCEVAtScope(AR, L->getParentLoop());
787 if (SE->isLoopInvariant(ExitVal, L))
791 // Only expand affine recurences.
792 if (!isSafe(AR, L, SE))
796 SE->getEffectiveSCEVType(ExtInst->getType());
798 // Only remove [sz]ext if the wide IV is still a native type.
800 // FIXME: We may be able to remove the copy of this logic in
801 // IVUsers::AddUsersIfInteresting.
802 uint64_t Width = SE->getTypeSizeInBits(Ty);
803 if (Width > 64 || (TD && !TD->isLegalInteger(Width)))
806 // Now expand it into actual Instructions and patch it into place.
808 // FIXME: avoid creating a new IV.
809 Value *NewVal = Rewriter.expandCodeFor(AR, Ty, ExtInst);
811 DEBUG(dbgs() << "INDVARS: Widened IV '" << *AR << "' " << *ExtInst << '\n'
812 << " into = " << *NewVal << "\n");
814 if (!isValidRewrite(ExtInst, NewVal)) {
815 DeadInsts.push_back(NewVal);
823 SE->getTypeSizeInBits(Ty) >
824 SE->getTypeSizeInBits(LargestType))
827 SE->forgetValue(ExtInst);
829 // Patch the new value into place.
830 if (ExtInst->hasName())
831 NewVal->takeName(ExtInst);
832 ExtInst->replaceAllUsesWith(NewVal);
834 // The old value may be dead now.
835 DeadInsts.push_back(ExtInst);
837 // UI is a linked list iterator, so AddUsersIfInteresting effectively pushes
838 // nodes on the worklist.
839 IU->AddUsersIfInteresting(ExtInst);
844 void IndVarSimplify::RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter) {
845 // Rewrite all induction variable expressions in terms of the canonical
846 // induction variable.
848 // If there were induction variables of other sizes or offsets, manually
849 // add the offsets to the primary induction variable and cast, avoiding
850 // the need for the code evaluation methods to insert induction variables
851 // of different sizes.
852 for (IVUsers::iterator UI = IU->begin(), E = IU->end(); UI != E; ++UI) {
853 Value *Op = UI->getOperandValToReplace();
854 const Type *UseTy = Op->getType();
855 Instruction *User = UI->getUser();
857 // Compute the final addrec to expand into code.
858 const SCEV *AR = IU->getReplacementExpr(*UI);
860 // Evaluate the expression out of the loop, if possible.
861 if (!L->contains(UI->getUser())) {
862 const SCEV *ExitVal = SE->getSCEVAtScope(AR, L->getParentLoop());
863 if (SE->isLoopInvariant(ExitVal, L))
867 // FIXME: It is an extremely bad idea to indvar substitute anything more
868 // complex than affine induction variables. Doing so will put expensive
869 // polynomial evaluations inside of the loop, and the str reduction pass
870 // currently can only reduce affine polynomials. For now just disable
871 // indvar subst on anything more complex than an affine addrec, unless
872 // it can be expanded to a trivial value.
873 if (!isSafe(AR, L, SE))
876 // Determine the insertion point for this user. By default, insert
877 // immediately before the user. The SCEVExpander class will automatically
878 // hoist loop invariants out of the loop. For PHI nodes, there may be
879 // multiple uses, so compute the nearest common dominator for the
881 Instruction *InsertPt = User;
882 if (PHINode *PHI = dyn_cast<PHINode>(InsertPt))
883 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
884 if (PHI->getIncomingValue(i) == Op) {
885 if (InsertPt == User)
886 InsertPt = PHI->getIncomingBlock(i)->getTerminator();
889 DT->findNearestCommonDominator(InsertPt->getParent(),
890 PHI->getIncomingBlock(i))
894 // Now expand it into actual Instructions and patch it into place.
895 Value *NewVal = Rewriter.expandCodeFor(AR, UseTy, InsertPt);
897 DEBUG(dbgs() << "INDVARS: Rewrote IV '" << *AR << "' " << *Op << '\n'
898 << " into = " << *NewVal << "\n");
900 if (!isValidRewrite(Op, NewVal)) {
901 DeadInsts.push_back(NewVal);
904 // Inform ScalarEvolution that this value is changing. The change doesn't
905 // affect its value, but it does potentially affect which use lists the
906 // value will be on after the replacement, which affects ScalarEvolution's
907 // ability to walk use lists and drop dangling pointers when a value is
909 SE->forgetValue(User);
911 // Patch the new value into place.
913 NewVal->takeName(Op);
914 User->replaceUsesOfWith(Op, NewVal);
915 UI->setOperandValToReplace(NewVal);
920 // The old value may be dead now.
921 DeadInsts.push_back(Op);
925 /// If there's a single exit block, sink any loop-invariant values that
926 /// were defined in the preheader but not used inside the loop into the
927 /// exit block to reduce register pressure in the loop.
928 void IndVarSimplify::SinkUnusedInvariants(Loop *L) {
929 BasicBlock *ExitBlock = L->getExitBlock();
930 if (!ExitBlock) return;
932 BasicBlock *Preheader = L->getLoopPreheader();
933 if (!Preheader) return;
935 Instruction *InsertPt = ExitBlock->getFirstNonPHI();
936 BasicBlock::iterator I = Preheader->getTerminator();
937 while (I != Preheader->begin()) {
939 // New instructions were inserted at the end of the preheader.
943 // Don't move instructions which might have side effects, since the side
944 // effects need to complete before instructions inside the loop. Also don't
945 // move instructions which might read memory, since the loop may modify
946 // memory. Note that it's okay if the instruction might have undefined
947 // behavior: LoopSimplify guarantees that the preheader dominates the exit
949 if (I->mayHaveSideEffects() || I->mayReadFromMemory())
952 // Skip debug info intrinsics.
953 if (isa<DbgInfoIntrinsic>(I))
956 // Don't sink static AllocaInsts out of the entry block, which would
957 // turn them into dynamic allocas!
958 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
959 if (AI->isStaticAlloca())
962 // Determine if there is a use in or before the loop (direct or
964 bool UsedInLoop = false;
965 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
968 BasicBlock *UseBB = cast<Instruction>(U)->getParent();
969 if (PHINode *P = dyn_cast<PHINode>(U)) {
971 PHINode::getIncomingValueNumForOperand(UI.getOperandNo());
972 UseBB = P->getIncomingBlock(i);
974 if (UseBB == Preheader || L->contains(UseBB)) {
980 // If there is, the def must remain in the preheader.
984 // Otherwise, sink it to the exit block.
985 Instruction *ToMove = I;
988 if (I != Preheader->begin()) {
989 // Skip debug info intrinsics.
992 } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin());
994 if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin())
1000 ToMove->moveBefore(InsertPt);
1006 /// ConvertToSInt - Convert APF to an integer, if possible.
1007 static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) {
1008 bool isExact = false;
1009 if (&APF.getSemantics() == &APFloat::PPCDoubleDouble)
1011 // See if we can convert this to an int64_t
1013 if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero,
1014 &isExact) != APFloat::opOK || !isExact)
1020 /// HandleFloatingPointIV - If the loop has floating induction variable
1021 /// then insert corresponding integer induction variable if possible.
1023 /// for(double i = 0; i < 10000; ++i)
1025 /// is converted into
1026 /// for(int i = 0; i < 10000; ++i)
1029 void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) {
1030 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
1031 unsigned BackEdge = IncomingEdge^1;
1033 // Check incoming value.
1034 ConstantFP *InitValueVal =
1035 dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));
1038 if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
1041 // Check IV increment. Reject this PN if increment operation is not
1042 // an add or increment value can not be represented by an integer.
1043 BinaryOperator *Incr =
1044 dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
1045 if (Incr == 0 || Incr->getOpcode() != Instruction::FAdd) return;
1047 // If this is not an add of the PHI with a constantfp, or if the constant fp
1048 // is not an integer, bail out.
1049 ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
1051 if (IncValueVal == 0 || Incr->getOperand(0) != PN ||
1052 !ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
1055 // Check Incr uses. One user is PN and the other user is an exit condition
1056 // used by the conditional terminator.
1057 Value::use_iterator IncrUse = Incr->use_begin();
1058 Instruction *U1 = cast<Instruction>(*IncrUse++);
1059 if (IncrUse == Incr->use_end()) return;
1060 Instruction *U2 = cast<Instruction>(*IncrUse++);
1061 if (IncrUse != Incr->use_end()) return;
1063 // Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't
1064 // only used by a branch, we can't transform it.
1065 FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
1067 Compare = dyn_cast<FCmpInst>(U2);
1068 if (Compare == 0 || !Compare->hasOneUse() ||
1069 !isa<BranchInst>(Compare->use_back()))
1072 BranchInst *TheBr = cast<BranchInst>(Compare->use_back());
1074 // We need to verify that the branch actually controls the iteration count
1075 // of the loop. If not, the new IV can overflow and no one will notice.
1076 // The branch block must be in the loop and one of the successors must be out
1078 assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
1079 if (!L->contains(TheBr->getParent()) ||
1080 (L->contains(TheBr->getSuccessor(0)) &&
1081 L->contains(TheBr->getSuccessor(1))))
1085 // If it isn't a comparison with an integer-as-fp (the exit value), we can't
1087 ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
1089 if (ExitValueVal == 0 ||
1090 !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
1093 // Find new predicate for integer comparison.
1094 CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
1095 switch (Compare->getPredicate()) {
1096 default: return; // Unknown comparison.
1097 case CmpInst::FCMP_OEQ:
1098 case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
1099 case CmpInst::FCMP_ONE:
1100 case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
1101 case CmpInst::FCMP_OGT:
1102 case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
1103 case CmpInst::FCMP_OGE:
1104 case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
1105 case CmpInst::FCMP_OLT:
1106 case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
1107 case CmpInst::FCMP_OLE:
1108 case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
1111 // We convert the floating point induction variable to a signed i32 value if
1112 // we can. This is only safe if the comparison will not overflow in a way
1113 // that won't be trapped by the integer equivalent operations. Check for this
1115 // TODO: We could use i64 if it is native and the range requires it.
1117 // The start/stride/exit values must all fit in signed i32.
1118 if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
1121 // If not actually striding (add x, 0.0), avoid touching the code.
1125 // Positive and negative strides have different safety conditions.
1127 // If we have a positive stride, we require the init to be less than the
1128 // exit value and an equality or less than comparison.
1129 if (InitValue >= ExitValue ||
1130 NewPred == CmpInst::ICMP_SGT || NewPred == CmpInst::ICMP_SGE)
1133 uint32_t Range = uint32_t(ExitValue-InitValue);
1134 if (NewPred == CmpInst::ICMP_SLE) {
1135 // Normalize SLE -> SLT, check for infinite loop.
1136 if (++Range == 0) return; // Range overflows.
1139 unsigned Leftover = Range % uint32_t(IncValue);
1141 // If this is an equality comparison, we require that the strided value
1142 // exactly land on the exit value, otherwise the IV condition will wrap
1143 // around and do things the fp IV wouldn't.
1144 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
1148 // If the stride would wrap around the i32 before exiting, we can't
1149 // transform the IV.
1150 if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
1154 // If we have a negative stride, we require the init to be greater than the
1155 // exit value and an equality or greater than comparison.
1156 if (InitValue >= ExitValue ||
1157 NewPred == CmpInst::ICMP_SLT || NewPred == CmpInst::ICMP_SLE)
1160 uint32_t Range = uint32_t(InitValue-ExitValue);
1161 if (NewPred == CmpInst::ICMP_SGE) {
1162 // Normalize SGE -> SGT, check for infinite loop.
1163 if (++Range == 0) return; // Range overflows.
1166 unsigned Leftover = Range % uint32_t(-IncValue);
1168 // If this is an equality comparison, we require that the strided value
1169 // exactly land on the exit value, otherwise the IV condition will wrap
1170 // around and do things the fp IV wouldn't.
1171 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
1175 // If the stride would wrap around the i32 before exiting, we can't
1176 // transform the IV.
1177 if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
1181 const IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
1183 // Insert new integer induction variable.
1184 PHINode *NewPHI = PHINode::Create(Int32Ty, 2, PN->getName()+".int", PN);
1185 NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
1186 PN->getIncomingBlock(IncomingEdge));
1189 BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
1190 Incr->getName()+".int", Incr);
1191 NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
1193 ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
1194 ConstantInt::get(Int32Ty, ExitValue),
1195 Compare->getName());
1197 // In the following deletions, PN may become dead and may be deleted.
1198 // Use a WeakVH to observe whether this happens.
1201 // Delete the old floating point exit comparison. The branch starts using the
1203 NewCompare->takeName(Compare);
1204 Compare->replaceAllUsesWith(NewCompare);
1205 RecursivelyDeleteTriviallyDeadInstructions(Compare);
1207 // Delete the old floating point increment.
1208 Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
1209 RecursivelyDeleteTriviallyDeadInstructions(Incr);
1211 // If the FP induction variable still has uses, this is because something else
1212 // in the loop uses its value. In order to canonicalize the induction
1213 // variable, we chose to eliminate the IV and rewrite it in terms of an
1216 // We give preference to sitofp over uitofp because it is faster on most
1219 Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
1220 PN->getParent()->getFirstNonPHI());
1221 PN->replaceAllUsesWith(Conv);
1222 RecursivelyDeleteTriviallyDeadInstructions(PN);
1225 // Add a new IVUsers entry for the newly-created integer PHI.
1226 IU->AddUsersIfInteresting(NewPHI);