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");
70 STATISTIC(NumElimIdentity, "Number of IV identities eliminated");
71 STATISTIC(NumElimExt , "Number of IV sign/zero extends eliminated");
72 STATISTIC(NumElimRem , "Number of IV remainder operations eliminated");
73 STATISTIC(NumElimCmp , "Number of IV comparisons eliminated");
75 // DisableIVRewrite mode currently affects IVUsers, so is defined in libAnalysis
76 // and referenced here.
78 extern bool DisableIVRewrite;
82 class IndVarSimplify : public LoopPass {
89 PHINode *CurrIV; // Current IV being simplified.
91 // Instructions processed by SimplifyIVUsers for CurrIV.
92 SmallPtrSet<Instruction*,16> Simplified;
94 // Use-def pairs if IVUsers waiting to be processed for CurrIV.
95 SmallVector<std::pair<Instruction*, Instruction*>, 8> SimpleIVUsers;
97 SmallVector<WeakVH, 16> DeadInsts;
101 static char ID; // Pass identification, replacement for typeid
102 IndVarSimplify() : LoopPass(ID), IU(0), LI(0), SE(0), DT(0), TD(0),
103 CurrIV(0), Changed(false) {
104 initializeIndVarSimplifyPass(*PassRegistry::getPassRegistry());
107 virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
109 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
110 AU.addRequired<DominatorTree>();
111 AU.addRequired<LoopInfo>();
112 AU.addRequired<ScalarEvolution>();
113 AU.addRequiredID(LoopSimplifyID);
114 AU.addRequiredID(LCSSAID);
115 AU.addRequired<IVUsers>();
116 AU.addPreserved<ScalarEvolution>();
117 AU.addPreservedID(LoopSimplifyID);
118 AU.addPreservedID(LCSSAID);
119 if (!DisableIVRewrite)
120 AU.addPreserved<IVUsers>();
121 AU.setPreservesCFG();
125 bool isValidRewrite(Value *FromVal, Value *ToVal);
127 void SimplifyIVUsers(SCEVExpander &Rewriter);
128 void SimplifyIVUsersNoRewrite(Loop *L, SCEVExpander &Rewriter);
130 bool EliminateIVUser(Instruction *UseInst, Instruction *IVOperand);
131 void EliminateIVComparison(ICmpInst *ICmp, Value *IVOperand);
132 void EliminateIVRemainder(BinaryOperator *Rem,
136 void pushIVUsers(Instruction *Def);
137 bool isSimpleIVUser(Instruction *I, const Loop *L);
138 void RewriteNonIntegerIVs(Loop *L);
140 ICmpInst *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
142 SCEVExpander &Rewriter);
144 void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter);
146 void RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter);
148 void SinkUnusedInvariants(Loop *L);
150 void HandleFloatingPointIV(Loop *L, PHINode *PH);
154 char IndVarSimplify::ID = 0;
155 INITIALIZE_PASS_BEGIN(IndVarSimplify, "indvars",
156 "Induction Variable Simplification", false, false)
157 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
158 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
159 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
160 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
161 INITIALIZE_PASS_DEPENDENCY(LCSSA)
162 INITIALIZE_PASS_DEPENDENCY(IVUsers)
163 INITIALIZE_PASS_END(IndVarSimplify, "indvars",
164 "Induction Variable Simplification", false, false)
166 Pass *llvm::createIndVarSimplifyPass() {
167 return new IndVarSimplify();
170 /// isValidRewrite - Return true if the SCEV expansion generated by the
171 /// rewriter can replace the original value. SCEV guarantees that it
172 /// produces the same value, but the way it is produced may be illegal IR.
173 /// Ideally, this function will only be called for verification.
174 bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) {
175 // If an SCEV expression subsumed multiple pointers, its expansion could
176 // reassociate the GEP changing the base pointer. This is illegal because the
177 // final address produced by a GEP chain must be inbounds relative to its
178 // underlying object. Otherwise basic alias analysis, among other things,
179 // could fail in a dangerous way. Ultimately, SCEV will be improved to avoid
180 // producing an expression involving multiple pointers. Until then, we must
183 // Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject
184 // because it understands lcssa phis while SCEV does not.
185 Value *FromPtr = FromVal;
186 Value *ToPtr = ToVal;
187 if (GEPOperator *GEP = dyn_cast<GEPOperator>(FromVal)) {
188 FromPtr = GEP->getPointerOperand();
190 if (GEPOperator *GEP = dyn_cast<GEPOperator>(ToVal)) {
191 ToPtr = GEP->getPointerOperand();
193 if (FromPtr != FromVal || ToPtr != ToVal) {
194 // Quickly check the common case
195 if (FromPtr == ToPtr)
198 // SCEV may have rewritten an expression that produces the GEP's pointer
199 // operand. That's ok as long as the pointer operand has the same base
200 // pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the
201 // base of a recurrence. This handles the case in which SCEV expansion
202 // converts a pointer type recurrence into a nonrecurrent pointer base
203 // indexed by an integer recurrence.
204 const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr));
205 const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr));
206 if (FromBase == ToBase)
209 DEBUG(dbgs() << "INDVARS: GEP rewrite bail out "
210 << *FromBase << " != " << *ToBase << "\n");
217 /// canExpandBackedgeTakenCount - Return true if this loop's backedge taken
218 /// count expression can be safely and cheaply expanded into an instruction
219 /// sequence that can be used by LinearFunctionTestReplace.
220 static bool canExpandBackedgeTakenCount(Loop *L, ScalarEvolution *SE) {
221 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
222 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) ||
223 BackedgeTakenCount->isZero())
226 if (!L->getExitingBlock())
229 // Can't rewrite non-branch yet.
230 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
234 // Special case: If the backedge-taken count is a UDiv, it's very likely a
235 // UDiv that ScalarEvolution produced in order to compute a precise
236 // expression, rather than a UDiv from the user's code. If we can't find a
237 // UDiv in the code with some simple searching, assume the former and forego
238 // rewriting the loop.
239 if (isa<SCEVUDivExpr>(BackedgeTakenCount)) {
240 ICmpInst *OrigCond = dyn_cast<ICmpInst>(BI->getCondition());
241 if (!OrigCond) return false;
242 const SCEV *R = SE->getSCEV(OrigCond->getOperand(1));
243 R = SE->getMinusSCEV(R, SE->getConstant(R->getType(), 1));
244 if (R != BackedgeTakenCount) {
245 const SCEV *L = SE->getSCEV(OrigCond->getOperand(0));
246 L = SE->getMinusSCEV(L, SE->getConstant(L->getType(), 1));
247 if (L != BackedgeTakenCount)
254 /// getBackedgeIVType - Get the widest type used by the loop test after peeking
257 /// TODO: Unnecessary once LinearFunctionTestReplace is removed.
258 static const Type *getBackedgeIVType(Loop *L) {
259 if (!L->getExitingBlock())
262 // Can't rewrite non-branch yet.
263 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
267 ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition());
272 for(User::op_iterator OI = Cond->op_begin(), OE = Cond->op_end();
274 assert((!Ty || Ty == (*OI)->getType()) && "bad icmp operand types");
275 TruncInst *Trunc = dyn_cast<TruncInst>(*OI);
279 return Trunc->getSrcTy();
284 /// LinearFunctionTestReplace - This method rewrites the exit condition of the
285 /// loop to be a canonical != comparison against the incremented loop induction
286 /// variable. This pass is able to rewrite the exit tests of any loop where the
287 /// SCEV analysis can determine a loop-invariant trip count of the loop, which
288 /// is actually a much broader range than just linear tests.
289 ICmpInst *IndVarSimplify::
290 LinearFunctionTestReplace(Loop *L,
291 const SCEV *BackedgeTakenCount,
293 SCEVExpander &Rewriter) {
294 assert(canExpandBackedgeTakenCount(L, SE) && "precondition");
295 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
297 // If the exiting block is not the same as the backedge block, we must compare
298 // against the preincremented value, otherwise we prefer to compare against
299 // the post-incremented value.
301 const SCEV *RHS = BackedgeTakenCount;
302 if (L->getExitingBlock() == L->getLoopLatch()) {
303 // Add one to the "backedge-taken" count to get the trip count.
304 // If this addition may overflow, we have to be more pessimistic and
305 // cast the induction variable before doing the add.
306 const SCEV *Zero = SE->getConstant(BackedgeTakenCount->getType(), 0);
308 SE->getAddExpr(BackedgeTakenCount,
309 SE->getConstant(BackedgeTakenCount->getType(), 1));
310 if ((isa<SCEVConstant>(N) && !N->isZero()) ||
311 SE->isLoopEntryGuardedByCond(L, ICmpInst::ICMP_NE, N, Zero)) {
312 // No overflow. Cast the sum.
313 RHS = SE->getTruncateOrZeroExtend(N, IndVar->getType());
315 // Potential overflow. Cast before doing the add.
316 RHS = SE->getTruncateOrZeroExtend(BackedgeTakenCount,
318 RHS = SE->getAddExpr(RHS,
319 SE->getConstant(IndVar->getType(), 1));
322 // The BackedgeTaken expression contains the number of times that the
323 // backedge branches to the loop header. This is one less than the
324 // number of times the loop executes, so use the incremented indvar.
325 CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock());
327 // We have to use the preincremented value...
328 RHS = SE->getTruncateOrZeroExtend(BackedgeTakenCount,
333 // Expand the code for the iteration count.
334 assert(SE->isLoopInvariant(RHS, L) &&
335 "Computed iteration count is not loop invariant!");
336 Value *ExitCnt = Rewriter.expandCodeFor(RHS, IndVar->getType(), BI);
338 // Insert a new icmp_ne or icmp_eq instruction before the branch.
339 ICmpInst::Predicate Opcode;
340 if (L->contains(BI->getSuccessor(0)))
341 Opcode = ICmpInst::ICMP_NE;
343 Opcode = ICmpInst::ICMP_EQ;
345 DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
346 << " LHS:" << *CmpIndVar << '\n'
348 << (Opcode == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
349 << " RHS:\t" << *RHS << "\n");
351 ICmpInst *Cond = new ICmpInst(BI, Opcode, CmpIndVar, ExitCnt, "exitcond");
353 Value *OrigCond = BI->getCondition();
354 // It's tempting to use replaceAllUsesWith here to fully replace the old
355 // comparison, but that's not immediately safe, since users of the old
356 // comparison may not be dominated by the new comparison. Instead, just
357 // update the branch to use the new comparison; in the common case this
358 // will make old comparison dead.
359 BI->setCondition(Cond);
360 DeadInsts.push_back(OrigCond);
367 /// RewriteLoopExitValues - Check to see if this loop has a computable
368 /// loop-invariant execution count. If so, this means that we can compute the
369 /// final value of any expressions that are recurrent in the loop, and
370 /// substitute the exit values from the loop into any instructions outside of
371 /// the loop that use the final values of the current expressions.
373 /// This is mostly redundant with the regular IndVarSimplify activities that
374 /// happen later, except that it's more powerful in some cases, because it's
375 /// able to brute-force evaluate arbitrary instructions as long as they have
376 /// constant operands at the beginning of the loop.
377 void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) {
378 // Verify the input to the pass in already in LCSSA form.
379 assert(L->isLCSSAForm(*DT));
381 SmallVector<BasicBlock*, 8> ExitBlocks;
382 L->getUniqueExitBlocks(ExitBlocks);
384 // Find all values that are computed inside the loop, but used outside of it.
385 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan
386 // the exit blocks of the loop to find them.
387 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
388 BasicBlock *ExitBB = ExitBlocks[i];
390 // If there are no PHI nodes in this exit block, then no values defined
391 // inside the loop are used on this path, skip it.
392 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
395 unsigned NumPreds = PN->getNumIncomingValues();
397 // Iterate over all of the PHI nodes.
398 BasicBlock::iterator BBI = ExitBB->begin();
399 while ((PN = dyn_cast<PHINode>(BBI++))) {
401 continue; // dead use, don't replace it
403 // SCEV only supports integer expressions for now.
404 if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy())
407 // It's necessary to tell ScalarEvolution about this explicitly so that
408 // it can walk the def-use list and forget all SCEVs, as it may not be
409 // watching the PHI itself. Once the new exit value is in place, there
410 // may not be a def-use connection between the loop and every instruction
411 // which got a SCEVAddRecExpr for that loop.
414 // Iterate over all of the values in all the PHI nodes.
415 for (unsigned i = 0; i != NumPreds; ++i) {
416 // If the value being merged in is not integer or is not defined
417 // in the loop, skip it.
418 Value *InVal = PN->getIncomingValue(i);
419 if (!isa<Instruction>(InVal))
422 // If this pred is for a subloop, not L itself, skip it.
423 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
424 continue; // The Block is in a subloop, skip it.
426 // Check that InVal is defined in the loop.
427 Instruction *Inst = cast<Instruction>(InVal);
428 if (!L->contains(Inst))
431 // Okay, this instruction has a user outside of the current loop
432 // and varies predictably *inside* the loop. Evaluate the value it
433 // contains when the loop exits, if possible.
434 const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
435 if (!SE->isLoopInvariant(ExitValue, L))
438 Value *ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), Inst);
440 DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
441 << " LoopVal = " << *Inst << "\n");
443 if (!isValidRewrite(Inst, ExitVal)) {
444 DeadInsts.push_back(ExitVal);
450 PN->setIncomingValue(i, ExitVal);
452 // If this instruction is dead now, delete it.
453 RecursivelyDeleteTriviallyDeadInstructions(Inst);
456 // Completely replace a single-pred PHI. This is safe, because the
457 // NewVal won't be variant in the loop, so we don't need an LCSSA phi
459 PN->replaceAllUsesWith(ExitVal);
460 RecursivelyDeleteTriviallyDeadInstructions(PN);
464 // Clone the PHI and delete the original one. This lets IVUsers and
465 // any other maps purge the original user from their records.
466 PHINode *NewPN = cast<PHINode>(PN->clone());
468 NewPN->insertBefore(PN);
469 PN->replaceAllUsesWith(NewPN);
470 PN->eraseFromParent();
475 // The insertion point instruction may have been deleted; clear it out
476 // so that the rewriter doesn't trip over it later.
477 Rewriter.clearInsertPoint();
480 void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
481 // First step. Check to see if there are any floating-point recurrences.
482 // If there are, change them into integer recurrences, permitting analysis by
483 // the SCEV routines.
485 BasicBlock *Header = L->getHeader();
487 SmallVector<WeakVH, 8> PHIs;
488 for (BasicBlock::iterator I = Header->begin();
489 PHINode *PN = dyn_cast<PHINode>(I); ++I)
492 for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
493 if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i]))
494 HandleFloatingPointIV(L, PN);
496 // If the loop previously had floating-point IV, ScalarEvolution
497 // may not have been able to compute a trip count. Now that we've done some
498 // re-writing, the trip count may be computable.
503 /// SimplifyIVUsers - Iteratively perform simplification on IVUsers within this
504 /// loop. IVUsers is treated as a worklist. Each successive simplification may
505 /// push more users which may themselves be candidates for simplification.
507 /// This is the old approach to IV simplification to be replaced by
508 /// SimplifyIVUsersNoRewrite.
510 void IndVarSimplify::SimplifyIVUsers(SCEVExpander &Rewriter) {
511 // Each round of simplification involves a round of eliminating operations
512 // followed by a round of widening IVs. A single IVUsers worklist is used
513 // across all rounds. The inner loop advances the user. If widening exposes
514 // more uses, then another pass through the outer loop is triggered.
515 for (IVUsers::iterator I = IU->begin(); I != IU->end(); ++I) {
516 Instruction *UseInst = I->getUser();
517 Value *IVOperand = I->getOperandValToReplace();
519 if (ICmpInst *ICmp = dyn_cast<ICmpInst>(UseInst)) {
520 EliminateIVComparison(ICmp, IVOperand);
523 if (BinaryOperator *Rem = dyn_cast<BinaryOperator>(UseInst)) {
524 bool IsSigned = Rem->getOpcode() == Instruction::SRem;
525 if (IsSigned || Rem->getOpcode() == Instruction::URem) {
526 EliminateIVRemainder(Rem, IVOperand, IsSigned, I->getPhi());
534 // Collect information about induction variables that are used by sign/zero
535 // extend operations. This information is recorded by CollectExtend and
536 // provides the input to WidenIV.
538 const Type *WidestNativeType; // Widest integer type created [sz]ext
539 bool IsSigned; // Was an sext user seen before a zext?
541 WideIVInfo() : WidestNativeType(0), IsSigned(false) {}
545 /// CollectExtend - Update information about the induction variable that is
546 /// extended by this sign or zero extend operation. This is used to determine
547 /// the final width of the IV before actually widening it.
548 static void CollectExtend(CastInst *Cast, bool IsSigned, WideIVInfo &WI,
549 ScalarEvolution *SE, const TargetData *TD) {
550 const Type *Ty = Cast->getType();
551 uint64_t Width = SE->getTypeSizeInBits(Ty);
552 if (TD && !TD->isLegalInteger(Width))
555 if (!WI.WidestNativeType) {
556 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
557 WI.IsSigned = IsSigned;
561 // We extend the IV to satisfy the sign of its first user, arbitrarily.
562 if (WI.IsSigned != IsSigned)
565 if (Width > SE->getTypeSizeInBits(WI.WidestNativeType))
566 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
570 /// WidenIV - The goal of this transform is to remove sign and zero extends
571 /// without creating any new induction variables. To do this, it creates a new
572 /// phi of the wider type and redirects all users, either removing extends or
573 /// inserting truncs whenever we stop propagating the type.
578 const Type *WideType;
589 Instruction *WideInc;
590 const SCEV *WideIncExpr;
591 SmallVectorImpl<WeakVH> &DeadInsts;
593 SmallPtrSet<Instruction*,16> Widened;
596 WidenIV(PHINode *PN, const WideIVInfo &WI, LoopInfo *LInfo,
597 ScalarEvolution *SEv, DominatorTree *DTree,
598 SmallVectorImpl<WeakVH> &DI) :
600 WideType(WI.WidestNativeType),
601 IsSigned(WI.IsSigned),
603 L(LI->getLoopFor(OrigPhi->getParent())),
610 assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV");
613 PHINode *CreateWideIV(SCEVExpander &Rewriter);
616 Instruction *CloneIVUser(Instruction *NarrowUse,
617 Instruction *NarrowDef,
618 Instruction *WideDef);
620 const SCEVAddRecExpr *GetWideRecurrence(Instruction *NarrowUse);
622 Instruction *WidenIVUse(Instruction *NarrowUse,
623 Instruction *NarrowDef,
624 Instruction *WideDef);
626 } // anonymous namespace
628 static Value *getExtend( Value *NarrowOper, const Type *WideType,
629 bool IsSigned, IRBuilder<> &Builder) {
630 return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) :
631 Builder.CreateZExt(NarrowOper, WideType);
634 /// CloneIVUser - Instantiate a wide operation to replace a narrow
635 /// operation. This only needs to handle operations that can evaluation to
636 /// SCEVAddRec. It can safely return 0 for any operation we decide not to clone.
637 Instruction *WidenIV::CloneIVUser(Instruction *NarrowUse,
638 Instruction *NarrowDef,
639 Instruction *WideDef) {
640 unsigned Opcode = NarrowUse->getOpcode();
644 case Instruction::Add:
645 case Instruction::Mul:
646 case Instruction::UDiv:
647 case Instruction::Sub:
648 case Instruction::And:
649 case Instruction::Or:
650 case Instruction::Xor:
651 case Instruction::Shl:
652 case Instruction::LShr:
653 case Instruction::AShr:
654 DEBUG(dbgs() << "Cloning IVUser: " << *NarrowUse << "\n");
656 IRBuilder<> Builder(NarrowUse);
658 // Replace NarrowDef operands with WideDef. Otherwise, we don't know
659 // anything about the narrow operand yet so must insert a [sz]ext. It is
660 // probably loop invariant and will be folded or hoisted. If it actually
661 // comes from a widened IV, it should be removed during a future call to
663 Value *LHS = (NarrowUse->getOperand(0) == NarrowDef) ? WideDef :
664 getExtend(NarrowUse->getOperand(0), WideType, IsSigned, Builder);
665 Value *RHS = (NarrowUse->getOperand(1) == NarrowDef) ? WideDef :
666 getExtend(NarrowUse->getOperand(1), WideType, IsSigned, Builder);
668 BinaryOperator *NarrowBO = cast<BinaryOperator>(NarrowUse);
669 BinaryOperator *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(),
671 NarrowBO->getName());
672 Builder.Insert(WideBO);
673 if (NarrowBO->hasNoUnsignedWrap()) WideBO->setHasNoUnsignedWrap();
674 if (NarrowBO->hasNoSignedWrap()) WideBO->setHasNoSignedWrap();
681 // GetWideRecurrence - Is this instruction potentially interesting from IVUsers'
682 // perspective after widening it's type? In other words, can the extend be
683 // safely hoisted out of the loop with SCEV reducing the value to a recurrence
684 // on the same loop. If so, return the sign or zero extended
685 // recurrence. Otherwise return NULL.
686 const SCEVAddRecExpr *WidenIV::GetWideRecurrence(Instruction *NarrowUse) {
687 if (!SE->isSCEVable(NarrowUse->getType()))
690 const SCEV *NarrowExpr = SE->getSCEV(NarrowUse);
691 const SCEV *WideExpr = IsSigned ?
692 SE->getSignExtendExpr(NarrowExpr, WideType) :
693 SE->getZeroExtendExpr(NarrowExpr, WideType);
694 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr);
695 if (!AddRec || AddRec->getLoop() != L)
701 /// HoistStep - Attempt to hoist an IV increment above a potential use.
703 /// To successfully hoist, two criteria must be met:
704 /// - IncV operands dominate InsertPos and
705 /// - InsertPos dominates IncV
707 /// Meeting the second condition means that we don't need to check all of IncV's
708 /// existing uses (it's moving up in the domtree).
710 /// This does not yet recursively hoist the operands, although that would
711 /// not be difficult.
712 static bool HoistStep(Instruction *IncV, Instruction *InsertPos,
713 const DominatorTree *DT)
715 if (DT->dominates(IncV, InsertPos))
718 if (!DT->dominates(InsertPos->getParent(), IncV->getParent()))
721 if (IncV->mayHaveSideEffects())
724 // Attempt to hoist IncV
725 for (User::op_iterator OI = IncV->op_begin(), OE = IncV->op_end();
727 Instruction *OInst = dyn_cast<Instruction>(OI);
728 if (OInst && !DT->dominates(OInst, InsertPos))
731 IncV->moveBefore(InsertPos);
735 /// WidenIVUse - Determine whether an individual user of the narrow IV can be
736 /// widened. If so, return the wide clone of the user.
737 Instruction *WidenIV::WidenIVUse(Instruction *NarrowUse,
738 Instruction *NarrowDef,
739 Instruction *WideDef) {
740 // To be consistent with IVUsers, stop traversing the def-use chain at
741 // inner-loop phis or post-loop phis.
742 if (isa<PHINode>(NarrowUse) && LI->getLoopFor(NarrowUse->getParent()) != L)
745 // Handle data flow merges and bizarre phi cycles.
746 if (!Widened.insert(NarrowUse))
749 // Our raison d'etre! Eliminate sign and zero extension.
750 if (IsSigned ? isa<SExtInst>(NarrowUse) : isa<ZExtInst>(NarrowUse)) {
751 Value *NewDef = WideDef;
752 if (NarrowUse->getType() != WideType) {
753 unsigned CastWidth = SE->getTypeSizeInBits(NarrowUse->getType());
754 unsigned IVWidth = SE->getTypeSizeInBits(WideType);
755 if (CastWidth < IVWidth) {
756 // The cast isn't as wide as the IV, so insert a Trunc.
757 IRBuilder<> Builder(NarrowUse);
758 NewDef = Builder.CreateTrunc(WideDef, NarrowUse->getType());
761 // A wider extend was hidden behind a narrower one. This may induce
762 // another round of IV widening in which the intermediate IV becomes
763 // dead. It should be very rare.
764 DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi
765 << " not wide enough to subsume " << *NarrowUse << "\n");
766 NarrowUse->replaceUsesOfWith(NarrowDef, WideDef);
770 if (NewDef != NarrowUse) {
771 DEBUG(dbgs() << "INDVARS: eliminating " << *NarrowUse
772 << " replaced by " << *WideDef << "\n");
774 NarrowUse->replaceAllUsesWith(NewDef);
775 DeadInsts.push_back(NarrowUse);
777 // Now that the extend is gone, we want to expose it's uses for potential
778 // further simplification. We don't need to directly inform SimplifyIVUsers
779 // of the new users, because their parent IV will be processed later as a
780 // new loop phi. If we preserved IVUsers analysis, we would also want to
781 // push the uses of WideDef here.
783 // No further widening is needed. The deceased [sz]ext had done it for us.
786 const SCEVAddRecExpr *WideAddRec = GetWideRecurrence(NarrowUse);
788 // This user does not evaluate to a recurence after widening, so don't
789 // follow it. Instead insert a Trunc to kill off the original use,
790 // eventually isolating the original narrow IV so it can be removed.
791 IRBuilder<> Builder(NarrowUse);
792 Value *Trunc = Builder.CreateTrunc(WideDef, NarrowDef->getType());
793 NarrowUse->replaceUsesOfWith(NarrowDef, Trunc);
796 // Reuse the IV increment that SCEVExpander created as long as it dominates
798 Instruction *WideUse = 0;
799 if (WideAddRec == WideIncExpr && HoistStep(WideInc, NarrowUse, DT)) {
803 WideUse = CloneIVUser(NarrowUse, NarrowDef, WideDef);
807 // GetWideRecurrence ensured that the narrow expression could be extended
808 // outside the loop without overflow. This suggests that the wide use
809 // evaluates to the same expression as the extended narrow use, but doesn't
810 // absolutely guarantee it. Hence the following failsafe check. In rare cases
811 // where it fails, we simply throw away the newly created wide use.
812 if (WideAddRec != SE->getSCEV(WideUse)) {
813 DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse
814 << ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n");
815 DeadInsts.push_back(WideUse);
819 // Returning WideUse pushes it on the worklist.
823 /// CreateWideIV - Process a single induction variable. First use the
824 /// SCEVExpander to create a wide induction variable that evaluates to the same
825 /// recurrence as the original narrow IV. Then use a worklist to forward
826 /// traverse the narrow IV's def-use chain. After WidenIVUse has processed all
827 /// interesting IV users, the narrow IV will be isolated for removal by
830 /// It would be simpler to delete uses as they are processed, but we must avoid
831 /// invalidating SCEV expressions.
833 PHINode *WidenIV::CreateWideIV(SCEVExpander &Rewriter) {
834 // Is this phi an induction variable?
835 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi));
839 // Widen the induction variable expression.
840 const SCEV *WideIVExpr = IsSigned ?
841 SE->getSignExtendExpr(AddRec, WideType) :
842 SE->getZeroExtendExpr(AddRec, WideType);
844 assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType &&
845 "Expect the new IV expression to preserve its type");
847 // Can the IV be extended outside the loop without overflow?
848 AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr);
849 if (!AddRec || AddRec->getLoop() != L)
852 // An AddRec must have loop-invariant operands. Since this AddRec is
853 // materialized by a loop header phi, the expression cannot have any post-loop
854 // operands, so they must dominate the loop header.
855 assert(SE->properlyDominates(AddRec->getStart(), L->getHeader()) &&
856 SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader())
857 && "Loop header phi recurrence inputs do not dominate the loop");
859 // The rewriter provides a value for the desired IV expression. This may
860 // either find an existing phi or materialize a new one. Either way, we
861 // expect a well-formed cyclic phi-with-increments. i.e. any operand not part
862 // of the phi-SCC dominates the loop entry.
863 Instruction *InsertPt = L->getHeader()->begin();
864 WidePhi = cast<PHINode>(Rewriter.expandCodeFor(AddRec, WideType, InsertPt));
866 // Remembering the WideIV increment generated by SCEVExpander allows
867 // WidenIVUse to reuse it when widening the narrow IV's increment. We don't
868 // employ a general reuse mechanism because the call above is the only call to
869 // SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses.
870 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
872 cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock));
873 WideIncExpr = SE->getSCEV(WideInc);
876 DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n");
879 // Traverse the def-use chain using a worklist starting at the original IV.
880 assert(Widened.empty() && "expect initial state" );
882 // Each worklist entry has a Narrow def-use link and Wide def.
883 SmallVector<std::pair<Use *, Instruction *>, 8> NarrowIVUsers;
884 for (Value::use_iterator UI = OrigPhi->use_begin(),
885 UE = OrigPhi->use_end(); UI != UE; ++UI) {
886 NarrowIVUsers.push_back(std::make_pair(&UI.getUse(), WidePhi));
888 while (!NarrowIVUsers.empty()) {
890 Instruction *WideDef;
891 tie(NarrowDefUse, WideDef) = NarrowIVUsers.pop_back_val();
893 // Process a def-use edge. This may replace the use, so don't hold a
894 // use_iterator across it.
895 Instruction *NarrowDef = cast<Instruction>(NarrowDefUse->get());
896 Instruction *NarrowUse = cast<Instruction>(NarrowDefUse->getUser());
897 Instruction *WideUse = WidenIVUse(NarrowUse, NarrowDef, WideDef);
899 // Follow all def-use edges from the previous narrow use.
901 for (Value::use_iterator UI = NarrowUse->use_begin(),
902 UE = NarrowUse->use_end(); UI != UE; ++UI) {
903 NarrowIVUsers.push_back(std::make_pair(&UI.getUse(), WideUse));
906 // WidenIVUse may have removed the def-use edge.
907 if (NarrowDef->use_empty())
908 DeadInsts.push_back(NarrowDef);
913 void IndVarSimplify::EliminateIVComparison(ICmpInst *ICmp, Value *IVOperand) {
914 unsigned IVOperIdx = 0;
915 ICmpInst::Predicate Pred = ICmp->getPredicate();
916 if (IVOperand != ICmp->getOperand(0)) {
918 assert(IVOperand == ICmp->getOperand(1) && "Can't find IVOperand");
920 Pred = ICmpInst::getSwappedPredicate(Pred);
923 // Get the SCEVs for the ICmp operands.
924 const SCEV *S = SE->getSCEV(ICmp->getOperand(IVOperIdx));
925 const SCEV *X = SE->getSCEV(ICmp->getOperand(1 - IVOperIdx));
927 // Simplify unnecessary loops away.
928 const Loop *ICmpLoop = LI->getLoopFor(ICmp->getParent());
929 S = SE->getSCEVAtScope(S, ICmpLoop);
930 X = SE->getSCEVAtScope(X, ICmpLoop);
932 // If the condition is always true or always false, replace it with
934 if (SE->isKnownPredicate(Pred, S, X))
935 ICmp->replaceAllUsesWith(ConstantInt::getTrue(ICmp->getContext()));
936 else if (SE->isKnownPredicate(ICmpInst::getInversePredicate(Pred), S, X))
937 ICmp->replaceAllUsesWith(ConstantInt::getFalse(ICmp->getContext()));
941 DEBUG(dbgs() << "INDVARS: Eliminated comparison: " << *ICmp << '\n');
944 DeadInsts.push_back(ICmp);
947 void IndVarSimplify::EliminateIVRemainder(BinaryOperator *Rem,
951 // We're only interested in the case where we know something about
953 if (IVOperand != Rem->getOperand(0))
956 // Get the SCEVs for the ICmp operands.
957 const SCEV *S = SE->getSCEV(Rem->getOperand(0));
958 const SCEV *X = SE->getSCEV(Rem->getOperand(1));
960 // Simplify unnecessary loops away.
961 const Loop *ICmpLoop = LI->getLoopFor(Rem->getParent());
962 S = SE->getSCEVAtScope(S, ICmpLoop);
963 X = SE->getSCEVAtScope(X, ICmpLoop);
965 // i % n --> i if i is in [0,n).
966 if ((!IsSigned || SE->isKnownNonNegative(S)) &&
967 SE->isKnownPredicate(IsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
969 Rem->replaceAllUsesWith(Rem->getOperand(0));
971 // (i+1) % n --> (i+1)==n?0:(i+1) if i is in [0,n).
972 const SCEV *LessOne =
973 SE->getMinusSCEV(S, SE->getConstant(S->getType(), 1));
974 if (IsSigned && !SE->isKnownNonNegative(LessOne))
977 if (!SE->isKnownPredicate(IsSigned ?
978 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
982 ICmpInst *ICmp = new ICmpInst(Rem, ICmpInst::ICMP_EQ,
983 Rem->getOperand(0), Rem->getOperand(1),
986 SelectInst::Create(ICmp,
987 ConstantInt::get(Rem->getType(), 0),
988 Rem->getOperand(0), "tmp", Rem);
989 Rem->replaceAllUsesWith(Sel);
992 // Inform IVUsers about the new users.
994 if (Instruction *I = dyn_cast<Instruction>(Rem->getOperand(0)))
995 IU->AddUsersIfInteresting(I, IVPhi);
997 DEBUG(dbgs() << "INDVARS: Simplified rem: " << *Rem << '\n');
1000 DeadInsts.push_back(Rem);
1003 /// EliminateIVUser - Eliminate an operation that consumes a simple IV and has
1004 /// no observable side-effect given the range of IV values.
1005 bool IndVarSimplify::EliminateIVUser(Instruction *UseInst,
1006 Instruction *IVOperand) {
1007 if (ICmpInst *ICmp = dyn_cast<ICmpInst>(UseInst)) {
1008 EliminateIVComparison(ICmp, IVOperand);
1011 if (BinaryOperator *Rem = dyn_cast<BinaryOperator>(UseInst)) {
1012 bool IsSigned = Rem->getOpcode() == Instruction::SRem;
1013 if (IsSigned || Rem->getOpcode() == Instruction::URem) {
1014 EliminateIVRemainder(Rem, IVOperand, IsSigned, CurrIV);
1019 // Eliminate any operation that SCEV can prove is an identity function.
1020 if (!SE->isSCEVable(UseInst->getType()) ||
1021 (SE->getSCEV(UseInst) != SE->getSCEV(IVOperand)))
1024 UseInst->replaceAllUsesWith(IVOperand);
1026 DEBUG(dbgs() << "INDVARS: Eliminated identity: " << *UseInst << '\n');
1029 DeadInsts.push_back(UseInst);
1033 /// pushIVUsers - Add all uses of Def to the current IV's worklist.
1035 void IndVarSimplify::pushIVUsers(Instruction *Def) {
1037 for (Value::use_iterator UI = Def->use_begin(), E = Def->use_end();
1039 Instruction *User = cast<Instruction>(*UI);
1041 // Avoid infinite or exponential worklist processing.
1042 // Also ensure unique worklist users.
1043 if (Simplified.insert(User))
1044 SimpleIVUsers.push_back(std::make_pair(User, Def));
1048 /// isSimpleIVUser - Return true if this instruction generates a simple SCEV
1049 /// expression in terms of that IV.
1051 /// This is similar to IVUsers' isInsteresting() but processes each instruction
1052 /// non-recursively when the operand is already known to be a simpleIVUser.
1054 bool IndVarSimplify::isSimpleIVUser(Instruction *I, const Loop *L) {
1055 if (!SE->isSCEVable(I->getType()))
1058 // Get the symbolic expression for this instruction.
1059 const SCEV *S = SE->getSCEV(I);
1061 // Only consider affine recurrences.
1062 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S);
1063 if (AR && AR->getLoop() == L)
1069 /// SimplifyIVUsersNoRewrite - Iteratively perform simplification on a worklist
1070 /// of IV users. Each successive simplification may push more users which may
1071 /// themselves be candidates for simplification.
1073 /// The "NoRewrite" algorithm does not require IVUsers analysis. Instead, it
1074 /// simplifies instructions in-place during analysis. Rather than rewriting
1075 /// induction variables bottom-up from their users, it transforms a chain of
1076 /// IVUsers top-down, updating the IR only when it encouters a clear
1077 /// optimization opportunitiy. A SCEVExpander "Rewriter" instance is still
1078 /// needed, but only used to generate a new IV (phi) of wider type for sign/zero
1079 /// extend elimination.
1081 /// Once DisableIVRewrite is default, LSR will be the only client of IVUsers.
1083 void IndVarSimplify::SimplifyIVUsersNoRewrite(Loop *L, SCEVExpander &Rewriter) {
1084 // Simplification is performed independently for each IV, as represented by a
1085 // loop header phi. Each round of simplification first iterates through the
1086 // SimplifyIVUsers worklist, then determines whether the current IV should be
1087 // widened. Widening adds a new phi to LoopPhis, inducing another round of
1088 // simplification on the wide IV.
1089 SmallVector<PHINode*, 8> LoopPhis;
1090 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1091 LoopPhis.push_back(cast<PHINode>(I));
1093 while (!LoopPhis.empty()) {
1094 CurrIV = LoopPhis.pop_back_val();
1096 assert(SimpleIVUsers.empty() && "expect empty IV users list");
1100 pushIVUsers(CurrIV);
1102 while (!SimpleIVUsers.empty()) {
1103 Instruction *UseInst, *Operand;
1104 tie(UseInst, Operand) = SimpleIVUsers.pop_back_val();
1106 if (EliminateIVUser(UseInst, Operand)) {
1107 pushIVUsers(Operand);
1110 if (CastInst *Cast = dyn_cast<CastInst>(UseInst)) {
1111 bool IsSigned = Cast->getOpcode() == Instruction::SExt;
1112 if (IsSigned || Cast->getOpcode() == Instruction::ZExt) {
1113 CollectExtend(Cast, IsSigned, WI, SE, TD);
1117 if (isSimpleIVUser(UseInst, L)) {
1118 pushIVUsers(UseInst);
1121 if (WI.WidestNativeType) {
1122 WidenIV Widener(CurrIV, WI, LI, SE, DT, DeadInsts);
1123 if (PHINode *WidePhi = Widener.CreateWideIV(Rewriter)) {
1125 LoopPhis.push_back(WidePhi);
1131 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
1132 // If LoopSimplify form is not available, stay out of trouble. Some notes:
1133 // - LSR currently only supports LoopSimplify-form loops. Indvars'
1134 // canonicalization can be a pessimization without LSR to "clean up"
1136 // - We depend on having a preheader; in particular,
1137 // Loop::getCanonicalInductionVariable only supports loops with preheaders,
1138 // and we're in trouble if we can't find the induction variable even when
1139 // we've manually inserted one.
1140 if (!L->isLoopSimplifyForm())
1143 if (!DisableIVRewrite)
1144 IU = &getAnalysis<IVUsers>();
1145 LI = &getAnalysis<LoopInfo>();
1146 SE = &getAnalysis<ScalarEvolution>();
1147 DT = &getAnalysis<DominatorTree>();
1148 TD = getAnalysisIfAvailable<TargetData>();
1155 // If there are any floating-point recurrences, attempt to
1156 // transform them to use integer recurrences.
1157 RewriteNonIntegerIVs(L);
1159 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1161 // Create a rewriter object which we'll use to transform the code with.
1162 SCEVExpander Rewriter(*SE);
1163 if (DisableIVRewrite)
1164 Rewriter.disableCanonicalMode();
1166 // Check to see if this loop has a computable loop-invariant execution count.
1167 // If so, this means that we can compute the final value of any expressions
1168 // that are recurrent in the loop, and substitute the exit values from the
1169 // loop into any instructions outside of the loop that use the final values of
1170 // the current expressions.
1172 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1173 RewriteLoopExitValues(L, Rewriter);
1175 // Eliminate redundant IV users.
1176 if (DisableIVRewrite)
1177 SimplifyIVUsersNoRewrite(L, Rewriter);
1179 SimplifyIVUsers(Rewriter);
1181 // Compute the type of the largest recurrence expression, and decide whether
1182 // a canonical induction variable should be inserted.
1183 const Type *LargestType = 0;
1184 bool NeedCannIV = false;
1185 bool ExpandBECount = canExpandBackedgeTakenCount(L, SE);
1186 if (ExpandBECount) {
1187 // If we have a known trip count and a single exit block, we'll be
1188 // rewriting the loop exit test condition below, which requires a
1189 // canonical induction variable.
1191 const Type *Ty = BackedgeTakenCount->getType();
1192 if (DisableIVRewrite) {
1193 // In this mode, SimplifyIVUsers may have already widened the IV used by
1194 // the backedge test and inserted a Trunc on the compare's operand. Get
1195 // the wider type to avoid creating a redundant narrow IV only used by the
1197 LargestType = getBackedgeIVType(L);
1200 SE->getTypeSizeInBits(Ty) >
1201 SE->getTypeSizeInBits(LargestType))
1202 LargestType = SE->getEffectiveSCEVType(Ty);
1204 if (!DisableIVRewrite) {
1205 for (IVUsers::const_iterator I = IU->begin(), E = IU->end(); I != E; ++I) {
1208 SE->getEffectiveSCEVType(I->getOperandValToReplace()->getType());
1210 SE->getTypeSizeInBits(Ty) >
1211 SE->getTypeSizeInBits(LargestType))
1216 // Now that we know the largest of the induction variable expressions
1217 // in this loop, insert a canonical induction variable of the largest size.
1218 PHINode *IndVar = 0;
1220 // Check to see if the loop already has any canonical-looking induction
1221 // variables. If any are present and wider than the planned canonical
1222 // induction variable, temporarily remove them, so that the Rewriter
1223 // doesn't attempt to reuse them.
1224 SmallVector<PHINode *, 2> OldCannIVs;
1225 while (PHINode *OldCannIV = L->getCanonicalInductionVariable()) {
1226 if (SE->getTypeSizeInBits(OldCannIV->getType()) >
1227 SE->getTypeSizeInBits(LargestType))
1228 OldCannIV->removeFromParent();
1231 OldCannIVs.push_back(OldCannIV);
1234 IndVar = Rewriter.getOrInsertCanonicalInductionVariable(L, LargestType);
1238 DEBUG(dbgs() << "INDVARS: New CanIV: " << *IndVar << '\n');
1240 // Now that the official induction variable is established, reinsert
1241 // any old canonical-looking variables after it so that the IR remains
1242 // consistent. They will be deleted as part of the dead-PHI deletion at
1243 // the end of the pass.
1244 while (!OldCannIVs.empty()) {
1245 PHINode *OldCannIV = OldCannIVs.pop_back_val();
1246 OldCannIV->insertBefore(L->getHeader()->getFirstNonPHI());
1250 // If we have a trip count expression, rewrite the loop's exit condition
1251 // using it. We can currently only handle loops with a single exit.
1252 ICmpInst *NewICmp = 0;
1253 if (ExpandBECount) {
1254 assert(canExpandBackedgeTakenCount(L, SE) &&
1255 "canonical IV disrupted BackedgeTaken expansion");
1256 assert(NeedCannIV &&
1257 "LinearFunctionTestReplace requires a canonical induction variable");
1258 NewICmp = LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar,
1261 // Rewrite IV-derived expressions.
1262 if (!DisableIVRewrite)
1263 RewriteIVExpressions(L, Rewriter);
1265 // Clear the rewriter cache, because values that are in the rewriter's cache
1266 // can be deleted in the loop below, causing the AssertingVH in the cache to
1270 // Now that we're done iterating through lists, clean up any instructions
1271 // which are now dead.
1272 while (!DeadInsts.empty())
1273 if (Instruction *Inst =
1274 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
1275 RecursivelyDeleteTriviallyDeadInstructions(Inst);
1277 // The Rewriter may not be used from this point on.
1279 // Loop-invariant instructions in the preheader that aren't used in the
1280 // loop may be sunk below the loop to reduce register pressure.
1281 SinkUnusedInvariants(L);
1283 // For completeness, inform IVUsers of the IV use in the newly-created
1284 // loop exit test instruction.
1286 IU->AddUsersIfInteresting(cast<Instruction>(NewICmp->getOperand(0)),
1289 // Clean up dead instructions.
1290 Changed |= DeleteDeadPHIs(L->getHeader());
1291 // Check a post-condition.
1292 assert(L->isLCSSAForm(*DT) && "Indvars did not leave the loop in lcssa form!");
1296 // FIXME: It is an extremely bad idea to indvar substitute anything more
1297 // complex than affine induction variables. Doing so will put expensive
1298 // polynomial evaluations inside of the loop, and the str reduction pass
1299 // currently can only reduce affine polynomials. For now just disable
1300 // indvar subst on anything more complex than an affine addrec, unless
1301 // it can be expanded to a trivial value.
1302 static bool isSafe(const SCEV *S, const Loop *L, ScalarEvolution *SE) {
1303 // Loop-invariant values are safe.
1304 if (SE->isLoopInvariant(S, L)) return true;
1306 // Affine addrecs are safe. Non-affine are not, because LSR doesn't know how
1307 // to transform them into efficient code.
1308 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
1309 return AR->isAffine();
1311 // An add is safe it all its operands are safe.
1312 if (const SCEVCommutativeExpr *Commutative = dyn_cast<SCEVCommutativeExpr>(S)) {
1313 for (SCEVCommutativeExpr::op_iterator I = Commutative->op_begin(),
1314 E = Commutative->op_end(); I != E; ++I)
1315 if (!isSafe(*I, L, SE)) return false;
1319 // A cast is safe if its operand is.
1320 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
1321 return isSafe(C->getOperand(), L, SE);
1323 // A udiv is safe if its operands are.
1324 if (const SCEVUDivExpr *UD = dyn_cast<SCEVUDivExpr>(S))
1325 return isSafe(UD->getLHS(), L, SE) &&
1326 isSafe(UD->getRHS(), L, SE);
1328 // SCEVUnknown is always safe.
1329 if (isa<SCEVUnknown>(S))
1332 // Nothing else is safe.
1336 void IndVarSimplify::RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter) {
1337 // Rewrite all induction variable expressions in terms of the canonical
1338 // induction variable.
1340 // If there were induction variables of other sizes or offsets, manually
1341 // add the offsets to the primary induction variable and cast, avoiding
1342 // the need for the code evaluation methods to insert induction variables
1343 // of different sizes.
1344 for (IVUsers::iterator UI = IU->begin(), E = IU->end(); UI != E; ++UI) {
1345 Value *Op = UI->getOperandValToReplace();
1346 const Type *UseTy = Op->getType();
1347 Instruction *User = UI->getUser();
1349 // Compute the final addrec to expand into code.
1350 const SCEV *AR = IU->getReplacementExpr(*UI);
1352 // Evaluate the expression out of the loop, if possible.
1353 if (!L->contains(UI->getUser())) {
1354 const SCEV *ExitVal = SE->getSCEVAtScope(AR, L->getParentLoop());
1355 if (SE->isLoopInvariant(ExitVal, L))
1359 // FIXME: It is an extremely bad idea to indvar substitute anything more
1360 // complex than affine induction variables. Doing so will put expensive
1361 // polynomial evaluations inside of the loop, and the str reduction pass
1362 // currently can only reduce affine polynomials. For now just disable
1363 // indvar subst on anything more complex than an affine addrec, unless
1364 // it can be expanded to a trivial value.
1365 if (!isSafe(AR, L, SE))
1368 // Determine the insertion point for this user. By default, insert
1369 // immediately before the user. The SCEVExpander class will automatically
1370 // hoist loop invariants out of the loop. For PHI nodes, there may be
1371 // multiple uses, so compute the nearest common dominator for the
1373 Instruction *InsertPt = User;
1374 if (PHINode *PHI = dyn_cast<PHINode>(InsertPt))
1375 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
1376 if (PHI->getIncomingValue(i) == Op) {
1377 if (InsertPt == User)
1378 InsertPt = PHI->getIncomingBlock(i)->getTerminator();
1381 DT->findNearestCommonDominator(InsertPt->getParent(),
1382 PHI->getIncomingBlock(i))
1386 // Now expand it into actual Instructions and patch it into place.
1387 Value *NewVal = Rewriter.expandCodeFor(AR, UseTy, InsertPt);
1389 DEBUG(dbgs() << "INDVARS: Rewrote IV '" << *AR << "' " << *Op << '\n'
1390 << " into = " << *NewVal << "\n");
1392 if (!isValidRewrite(Op, NewVal)) {
1393 DeadInsts.push_back(NewVal);
1396 // Inform ScalarEvolution that this value is changing. The change doesn't
1397 // affect its value, but it does potentially affect which use lists the
1398 // value will be on after the replacement, which affects ScalarEvolution's
1399 // ability to walk use lists and drop dangling pointers when a value is
1401 SE->forgetValue(User);
1403 // Patch the new value into place.
1405 NewVal->takeName(Op);
1406 User->replaceUsesOfWith(Op, NewVal);
1407 UI->setOperandValToReplace(NewVal);
1412 // The old value may be dead now.
1413 DeadInsts.push_back(Op);
1417 /// If there's a single exit block, sink any loop-invariant values that
1418 /// were defined in the preheader but not used inside the loop into the
1419 /// exit block to reduce register pressure in the loop.
1420 void IndVarSimplify::SinkUnusedInvariants(Loop *L) {
1421 BasicBlock *ExitBlock = L->getExitBlock();
1422 if (!ExitBlock) return;
1424 BasicBlock *Preheader = L->getLoopPreheader();
1425 if (!Preheader) return;
1427 Instruction *InsertPt = ExitBlock->getFirstNonPHI();
1428 BasicBlock::iterator I = Preheader->getTerminator();
1429 while (I != Preheader->begin()) {
1431 // New instructions were inserted at the end of the preheader.
1432 if (isa<PHINode>(I))
1435 // Don't move instructions which might have side effects, since the side
1436 // effects need to complete before instructions inside the loop. Also don't
1437 // move instructions which might read memory, since the loop may modify
1438 // memory. Note that it's okay if the instruction might have undefined
1439 // behavior: LoopSimplify guarantees that the preheader dominates the exit
1441 if (I->mayHaveSideEffects() || I->mayReadFromMemory())
1444 // Skip debug info intrinsics.
1445 if (isa<DbgInfoIntrinsic>(I))
1448 // Don't sink static AllocaInsts out of the entry block, which would
1449 // turn them into dynamic allocas!
1450 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
1451 if (AI->isStaticAlloca())
1454 // Determine if there is a use in or before the loop (direct or
1456 bool UsedInLoop = false;
1457 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
1460 BasicBlock *UseBB = cast<Instruction>(U)->getParent();
1461 if (PHINode *P = dyn_cast<PHINode>(U)) {
1463 PHINode::getIncomingValueNumForOperand(UI.getOperandNo());
1464 UseBB = P->getIncomingBlock(i);
1466 if (UseBB == Preheader || L->contains(UseBB)) {
1472 // If there is, the def must remain in the preheader.
1476 // Otherwise, sink it to the exit block.
1477 Instruction *ToMove = I;
1480 if (I != Preheader->begin()) {
1481 // Skip debug info intrinsics.
1484 } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin());
1486 if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin())
1492 ToMove->moveBefore(InsertPt);
1498 /// ConvertToSInt - Convert APF to an integer, if possible.
1499 static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) {
1500 bool isExact = false;
1501 if (&APF.getSemantics() == &APFloat::PPCDoubleDouble)
1503 // See if we can convert this to an int64_t
1505 if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero,
1506 &isExact) != APFloat::opOK || !isExact)
1512 /// HandleFloatingPointIV - If the loop has floating induction variable
1513 /// then insert corresponding integer induction variable if possible.
1515 /// for(double i = 0; i < 10000; ++i)
1517 /// is converted into
1518 /// for(int i = 0; i < 10000; ++i)
1521 void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) {
1522 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
1523 unsigned BackEdge = IncomingEdge^1;
1525 // Check incoming value.
1526 ConstantFP *InitValueVal =
1527 dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));
1530 if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
1533 // Check IV increment. Reject this PN if increment operation is not
1534 // an add or increment value can not be represented by an integer.
1535 BinaryOperator *Incr =
1536 dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
1537 if (Incr == 0 || Incr->getOpcode() != Instruction::FAdd) return;
1539 // If this is not an add of the PHI with a constantfp, or if the constant fp
1540 // is not an integer, bail out.
1541 ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
1543 if (IncValueVal == 0 || Incr->getOperand(0) != PN ||
1544 !ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
1547 // Check Incr uses. One user is PN and the other user is an exit condition
1548 // used by the conditional terminator.
1549 Value::use_iterator IncrUse = Incr->use_begin();
1550 Instruction *U1 = cast<Instruction>(*IncrUse++);
1551 if (IncrUse == Incr->use_end()) return;
1552 Instruction *U2 = cast<Instruction>(*IncrUse++);
1553 if (IncrUse != Incr->use_end()) return;
1555 // Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't
1556 // only used by a branch, we can't transform it.
1557 FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
1559 Compare = dyn_cast<FCmpInst>(U2);
1560 if (Compare == 0 || !Compare->hasOneUse() ||
1561 !isa<BranchInst>(Compare->use_back()))
1564 BranchInst *TheBr = cast<BranchInst>(Compare->use_back());
1566 // We need to verify that the branch actually controls the iteration count
1567 // of the loop. If not, the new IV can overflow and no one will notice.
1568 // The branch block must be in the loop and one of the successors must be out
1570 assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
1571 if (!L->contains(TheBr->getParent()) ||
1572 (L->contains(TheBr->getSuccessor(0)) &&
1573 L->contains(TheBr->getSuccessor(1))))
1577 // If it isn't a comparison with an integer-as-fp (the exit value), we can't
1579 ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
1581 if (ExitValueVal == 0 ||
1582 !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
1585 // Find new predicate for integer comparison.
1586 CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
1587 switch (Compare->getPredicate()) {
1588 default: return; // Unknown comparison.
1589 case CmpInst::FCMP_OEQ:
1590 case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
1591 case CmpInst::FCMP_ONE:
1592 case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
1593 case CmpInst::FCMP_OGT:
1594 case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
1595 case CmpInst::FCMP_OGE:
1596 case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
1597 case CmpInst::FCMP_OLT:
1598 case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
1599 case CmpInst::FCMP_OLE:
1600 case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
1603 // We convert the floating point induction variable to a signed i32 value if
1604 // we can. This is only safe if the comparison will not overflow in a way
1605 // that won't be trapped by the integer equivalent operations. Check for this
1607 // TODO: We could use i64 if it is native and the range requires it.
1609 // The start/stride/exit values must all fit in signed i32.
1610 if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
1613 // If not actually striding (add x, 0.0), avoid touching the code.
1617 // Positive and negative strides have different safety conditions.
1619 // If we have a positive stride, we require the init to be less than the
1620 // exit value and an equality or less than comparison.
1621 if (InitValue >= ExitValue ||
1622 NewPred == CmpInst::ICMP_SGT || NewPred == CmpInst::ICMP_SGE)
1625 uint32_t Range = uint32_t(ExitValue-InitValue);
1626 if (NewPred == CmpInst::ICMP_SLE) {
1627 // Normalize SLE -> SLT, check for infinite loop.
1628 if (++Range == 0) return; // Range overflows.
1631 unsigned Leftover = Range % uint32_t(IncValue);
1633 // If this is an equality comparison, we require that the strided value
1634 // exactly land on the exit value, otherwise the IV condition will wrap
1635 // around and do things the fp IV wouldn't.
1636 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
1640 // If the stride would wrap around the i32 before exiting, we can't
1641 // transform the IV.
1642 if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
1646 // If we have a negative stride, we require the init to be greater than the
1647 // exit value and an equality or greater than comparison.
1648 if (InitValue >= ExitValue ||
1649 NewPred == CmpInst::ICMP_SLT || NewPred == CmpInst::ICMP_SLE)
1652 uint32_t Range = uint32_t(InitValue-ExitValue);
1653 if (NewPred == CmpInst::ICMP_SGE) {
1654 // Normalize SGE -> SGT, check for infinite loop.
1655 if (++Range == 0) return; // Range overflows.
1658 unsigned Leftover = Range % uint32_t(-IncValue);
1660 // If this is an equality comparison, we require that the strided value
1661 // exactly land on the exit value, otherwise the IV condition will wrap
1662 // around and do things the fp IV wouldn't.
1663 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
1667 // If the stride would wrap around the i32 before exiting, we can't
1668 // transform the IV.
1669 if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
1673 const IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
1675 // Insert new integer induction variable.
1676 PHINode *NewPHI = PHINode::Create(Int32Ty, 2, PN->getName()+".int", PN);
1677 NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
1678 PN->getIncomingBlock(IncomingEdge));
1681 BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
1682 Incr->getName()+".int", Incr);
1683 NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
1685 ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
1686 ConstantInt::get(Int32Ty, ExitValue),
1687 Compare->getName());
1689 // In the following deletions, PN may become dead and may be deleted.
1690 // Use a WeakVH to observe whether this happens.
1693 // Delete the old floating point exit comparison. The branch starts using the
1695 NewCompare->takeName(Compare);
1696 Compare->replaceAllUsesWith(NewCompare);
1697 RecursivelyDeleteTriviallyDeadInstructions(Compare);
1699 // Delete the old floating point increment.
1700 Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
1701 RecursivelyDeleteTriviallyDeadInstructions(Incr);
1703 // If the FP induction variable still has uses, this is because something else
1704 // in the loop uses its value. In order to canonicalize the induction
1705 // variable, we chose to eliminate the IV and rewrite it in terms of an
1708 // We give preference to sitofp over uitofp because it is faster on most
1711 Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
1712 PN->getParent()->getFirstNonPHI());
1713 PN->replaceAllUsesWith(Conv);
1714 RecursivelyDeleteTriviallyDeadInstructions(PN);
1717 // Add a new IVUsers entry for the newly-created integer PHI.
1719 IU->AddUsersIfInteresting(NewPHI, NewPHI);