1 //===- IndVarSimplify.cpp - Induction Variable Elimination ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation analyzes and transforms the induction variables (and
11 // computations derived from them) into simpler forms suitable for subsequent
12 // analysis and transformation.
14 // This transformation makes the following changes to each loop with an
15 // identifiable induction variable:
16 // 1. All loops are transformed to have a SINGLE canonical induction variable
17 // which starts at zero and steps by one.
18 // 2. The canonical induction variable is guaranteed to be the first PHI node
19 // in the loop header block.
20 // 3. The canonical induction variable is guaranteed to be in a wide enough
21 // type so that IV expressions need not be (directly) zero-extended or
23 // 4. Any pointer arithmetic recurrences are raised to use array subscripts.
25 // If the trip count of a loop is computable, this pass also makes the following
27 // 1. The exit condition for the loop is canonicalized to compare the
28 // induction value against the exit value. This turns loops like:
29 // 'for (i = 7; i*i < 1000; ++i)' into 'for (i = 0; i != 25; ++i)'
30 // 2. Any use outside of the loop of an expression derived from the indvar
31 // is changed to compute the derived value outside of the loop, eliminating
32 // the dependence on the exit value of the induction variable. If the only
33 // purpose of the loop is to compute the exit value of some derived
34 // expression, this transformation will make the loop dead.
36 // This transformation should be followed by strength reduction after all of the
37 // desired loop transformations have been performed.
39 //===----------------------------------------------------------------------===//
41 #define DEBUG_TYPE "indvars"
42 #include "llvm/Transforms/Scalar.h"
43 #include "llvm/BasicBlock.h"
44 #include "llvm/Constants.h"
45 #include "llvm/Instructions.h"
46 #include "llvm/IntrinsicInst.h"
47 #include "llvm/LLVMContext.h"
48 #include "llvm/Type.h"
49 #include "llvm/Analysis/Dominators.h"
50 #include "llvm/Analysis/IVUsers.h"
51 #include "llvm/Analysis/ScalarEvolutionExpander.h"
52 #include "llvm/Analysis/LoopInfo.h"
53 #include "llvm/Analysis/LoopPass.h"
54 #include "llvm/Support/CFG.h"
55 #include "llvm/Support/CommandLine.h"
56 #include "llvm/Support/Debug.h"
57 #include "llvm/Support/raw_ostream.h"
58 #include "llvm/Transforms/Utils/Local.h"
59 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
60 #include "llvm/Target/TargetData.h"
61 #include "llvm/ADT/SmallVector.h"
62 #include "llvm/ADT/Statistic.h"
63 #include "llvm/ADT/STLExtras.h"
66 STATISTIC(NumRemoved , "Number of aux indvars removed");
67 STATISTIC(NumWidened , "Number of indvars widened");
68 STATISTIC(NumInserted , "Number of canonical indvars added");
69 STATISTIC(NumReplaced , "Number of exit values replaced");
70 STATISTIC(NumLFTR , "Number of loop exit tests replaced");
71 STATISTIC(NumElimIdentity, "Number of IV identities eliminated");
72 STATISTIC(NumElimExt , "Number of IV sign/zero extends eliminated");
73 STATISTIC(NumElimRem , "Number of IV remainder operations eliminated");
74 STATISTIC(NumElimCmp , "Number of IV comparisons eliminated");
76 static cl::opt<bool> DisableIVRewrite(
77 "disable-iv-rewrite", cl::Hidden,
78 cl::desc("Disable canonical induction variable rewriting"));
81 class IndVarSimplify : public LoopPass {
88 SmallVector<WeakVH, 16> DeadInsts;
92 static char ID; // Pass identification, replacement for typeid
93 IndVarSimplify() : LoopPass(ID), IU(0), LI(0), SE(0), DT(0), TD(0),
95 initializeIndVarSimplifyPass(*PassRegistry::getPassRegistry());
98 virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
100 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
101 AU.addRequired<DominatorTree>();
102 AU.addRequired<LoopInfo>();
103 AU.addRequired<ScalarEvolution>();
104 AU.addRequiredID(LoopSimplifyID);
105 AU.addRequiredID(LCSSAID);
106 if (!DisableIVRewrite)
107 AU.addRequired<IVUsers>();
108 AU.addPreserved<ScalarEvolution>();
109 AU.addPreservedID(LoopSimplifyID);
110 AU.addPreservedID(LCSSAID);
111 if (!DisableIVRewrite)
112 AU.addPreserved<IVUsers>();
113 AU.setPreservesCFG();
117 bool isValidRewrite(Value *FromVal, Value *ToVal);
119 void SimplifyIVUsers(SCEVExpander &Rewriter);
120 void SimplifyIVUsersNoRewrite(Loop *L, SCEVExpander &Rewriter);
122 bool EliminateIVUser(Instruction *UseInst, Instruction *IVOperand);
123 void EliminateIVComparison(ICmpInst *ICmp, Value *IVOperand);
124 void EliminateIVRemainder(BinaryOperator *Rem,
127 bool isSimpleIVUser(Instruction *I, const Loop *L);
128 void RewriteNonIntegerIVs(Loop *L);
130 ICmpInst *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
132 SCEVExpander &Rewriter);
134 void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter);
136 void RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter);
138 void SinkUnusedInvariants(Loop *L);
140 void HandleFloatingPointIV(Loop *L, PHINode *PH);
144 char IndVarSimplify::ID = 0;
145 INITIALIZE_PASS_BEGIN(IndVarSimplify, "indvars",
146 "Induction Variable Simplification", false, false)
147 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
148 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
149 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
150 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
151 INITIALIZE_PASS_DEPENDENCY(LCSSA)
152 INITIALIZE_PASS_DEPENDENCY(IVUsers)
153 INITIALIZE_PASS_END(IndVarSimplify, "indvars",
154 "Induction Variable Simplification", false, false)
156 Pass *llvm::createIndVarSimplifyPass() {
157 return new IndVarSimplify();
160 /// isValidRewrite - Return true if the SCEV expansion generated by the
161 /// rewriter can replace the original value. SCEV guarantees that it
162 /// produces the same value, but the way it is produced may be illegal IR.
163 /// Ideally, this function will only be called for verification.
164 bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) {
165 // If an SCEV expression subsumed multiple pointers, its expansion could
166 // reassociate the GEP changing the base pointer. This is illegal because the
167 // final address produced by a GEP chain must be inbounds relative to its
168 // underlying object. Otherwise basic alias analysis, among other things,
169 // could fail in a dangerous way. Ultimately, SCEV will be improved to avoid
170 // producing an expression involving multiple pointers. Until then, we must
173 // Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject
174 // because it understands lcssa phis while SCEV does not.
175 Value *FromPtr = FromVal;
176 Value *ToPtr = ToVal;
177 if (GEPOperator *GEP = dyn_cast<GEPOperator>(FromVal)) {
178 FromPtr = GEP->getPointerOperand();
180 if (GEPOperator *GEP = dyn_cast<GEPOperator>(ToVal)) {
181 ToPtr = GEP->getPointerOperand();
183 if (FromPtr != FromVal || ToPtr != ToVal) {
184 // Quickly check the common case
185 if (FromPtr == ToPtr)
188 // SCEV may have rewritten an expression that produces the GEP's pointer
189 // operand. That's ok as long as the pointer operand has the same base
190 // pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the
191 // base of a recurrence. This handles the case in which SCEV expansion
192 // converts a pointer type recurrence into a nonrecurrent pointer base
193 // indexed by an integer recurrence.
194 const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr));
195 const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr));
196 if (FromBase == ToBase)
199 DEBUG(dbgs() << "INDVARS: GEP rewrite bail out "
200 << *FromBase << " != " << *ToBase << "\n");
207 /// canExpandBackedgeTakenCount - Return true if this loop's backedge taken
208 /// count expression can be safely and cheaply expanded into an instruction
209 /// sequence that can be used by LinearFunctionTestReplace.
210 static bool canExpandBackedgeTakenCount(Loop *L, ScalarEvolution *SE) {
211 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
212 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) ||
213 BackedgeTakenCount->isZero())
216 if (!L->getExitingBlock())
219 // Can't rewrite non-branch yet.
220 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
224 // Special case: If the backedge-taken count is a UDiv, it's very likely a
225 // UDiv that ScalarEvolution produced in order to compute a precise
226 // expression, rather than a UDiv from the user's code. If we can't find a
227 // UDiv in the code with some simple searching, assume the former and forego
228 // rewriting the loop.
229 if (isa<SCEVUDivExpr>(BackedgeTakenCount)) {
230 ICmpInst *OrigCond = dyn_cast<ICmpInst>(BI->getCondition());
231 if (!OrigCond) return false;
232 const SCEV *R = SE->getSCEV(OrigCond->getOperand(1));
233 R = SE->getMinusSCEV(R, SE->getConstant(R->getType(), 1));
234 if (R != BackedgeTakenCount) {
235 const SCEV *L = SE->getSCEV(OrigCond->getOperand(0));
236 L = SE->getMinusSCEV(L, SE->getConstant(L->getType(), 1));
237 if (L != BackedgeTakenCount)
244 /// getBackedgeIVType - Get the widest type used by the loop test after peeking
247 /// TODO: Unnecessary once LinearFunctionTestReplace is removed.
248 static const Type *getBackedgeIVType(Loop *L) {
249 if (!L->getExitingBlock())
252 // Can't rewrite non-branch yet.
253 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
257 ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition());
262 for(User::op_iterator OI = Cond->op_begin(), OE = Cond->op_end();
264 assert((!Ty || Ty == (*OI)->getType()) && "bad icmp operand types");
265 TruncInst *Trunc = dyn_cast<TruncInst>(*OI);
269 return Trunc->getSrcTy();
274 /// LinearFunctionTestReplace - This method rewrites the exit condition of the
275 /// loop to be a canonical != comparison against the incremented loop induction
276 /// variable. This pass is able to rewrite the exit tests of any loop where the
277 /// SCEV analysis can determine a loop-invariant trip count of the loop, which
278 /// is actually a much broader range than just linear tests.
279 ICmpInst *IndVarSimplify::
280 LinearFunctionTestReplace(Loop *L,
281 const SCEV *BackedgeTakenCount,
283 SCEVExpander &Rewriter) {
284 assert(canExpandBackedgeTakenCount(L, SE) && "precondition");
285 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
287 // If the exiting block is not the same as the backedge block, we must compare
288 // against the preincremented value, otherwise we prefer to compare against
289 // the post-incremented value.
291 const SCEV *RHS = BackedgeTakenCount;
292 if (L->getExitingBlock() == L->getLoopLatch()) {
293 // Add one to the "backedge-taken" count to get the trip count.
294 // If this addition may overflow, we have to be more pessimistic and
295 // cast the induction variable before doing the add.
296 const SCEV *Zero = SE->getConstant(BackedgeTakenCount->getType(), 0);
298 SE->getAddExpr(BackedgeTakenCount,
299 SE->getConstant(BackedgeTakenCount->getType(), 1));
300 if ((isa<SCEVConstant>(N) && !N->isZero()) ||
301 SE->isLoopEntryGuardedByCond(L, ICmpInst::ICMP_NE, N, Zero)) {
302 // No overflow. Cast the sum.
303 RHS = SE->getTruncateOrZeroExtend(N, IndVar->getType());
305 // Potential overflow. Cast before doing the add.
306 RHS = SE->getTruncateOrZeroExtend(BackedgeTakenCount,
308 RHS = SE->getAddExpr(RHS,
309 SE->getConstant(IndVar->getType(), 1));
312 // The BackedgeTaken expression contains the number of times that the
313 // backedge branches to the loop header. This is one less than the
314 // number of times the loop executes, so use the incremented indvar.
315 CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock());
317 // We have to use the preincremented value...
318 RHS = SE->getTruncateOrZeroExtend(BackedgeTakenCount,
323 // Expand the code for the iteration count.
324 assert(SE->isLoopInvariant(RHS, L) &&
325 "Computed iteration count is not loop invariant!");
326 Value *ExitCnt = Rewriter.expandCodeFor(RHS, IndVar->getType(), BI);
328 // Insert a new icmp_ne or icmp_eq instruction before the branch.
329 ICmpInst::Predicate Opcode;
330 if (L->contains(BI->getSuccessor(0)))
331 Opcode = ICmpInst::ICMP_NE;
333 Opcode = ICmpInst::ICMP_EQ;
335 DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
336 << " LHS:" << *CmpIndVar << '\n'
338 << (Opcode == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
339 << " RHS:\t" << *RHS << "\n");
341 ICmpInst *Cond = new ICmpInst(BI, Opcode, CmpIndVar, ExitCnt, "exitcond");
342 Cond->setDebugLoc(BI->getDebugLoc());
343 Value *OrigCond = BI->getCondition();
344 // It's tempting to use replaceAllUsesWith here to fully replace the old
345 // comparison, but that's not immediately safe, since users of the old
346 // comparison may not be dominated by the new comparison. Instead, just
347 // update the branch to use the new comparison; in the common case this
348 // will make old comparison dead.
349 BI->setCondition(Cond);
350 DeadInsts.push_back(OrigCond);
357 /// RewriteLoopExitValues - Check to see if this loop has a computable
358 /// loop-invariant execution count. If so, this means that we can compute the
359 /// final value of any expressions that are recurrent in the loop, and
360 /// substitute the exit values from the loop into any instructions outside of
361 /// the loop that use the final values of the current expressions.
363 /// This is mostly redundant with the regular IndVarSimplify activities that
364 /// happen later, except that it's more powerful in some cases, because it's
365 /// able to brute-force evaluate arbitrary instructions as long as they have
366 /// constant operands at the beginning of the loop.
367 void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) {
368 // Verify the input to the pass in already in LCSSA form.
369 assert(L->isLCSSAForm(*DT));
371 SmallVector<BasicBlock*, 8> ExitBlocks;
372 L->getUniqueExitBlocks(ExitBlocks);
374 // Find all values that are computed inside the loop, but used outside of it.
375 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan
376 // the exit blocks of the loop to find them.
377 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
378 BasicBlock *ExitBB = ExitBlocks[i];
380 // If there are no PHI nodes in this exit block, then no values defined
381 // inside the loop are used on this path, skip it.
382 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
385 unsigned NumPreds = PN->getNumIncomingValues();
387 // Iterate over all of the PHI nodes.
388 BasicBlock::iterator BBI = ExitBB->begin();
389 while ((PN = dyn_cast<PHINode>(BBI++))) {
391 continue; // dead use, don't replace it
393 // SCEV only supports integer expressions for now.
394 if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy())
397 // It's necessary to tell ScalarEvolution about this explicitly so that
398 // it can walk the def-use list and forget all SCEVs, as it may not be
399 // watching the PHI itself. Once the new exit value is in place, there
400 // may not be a def-use connection between the loop and every instruction
401 // which got a SCEVAddRecExpr for that loop.
404 // Iterate over all of the values in all the PHI nodes.
405 for (unsigned i = 0; i != NumPreds; ++i) {
406 // If the value being merged in is not integer or is not defined
407 // in the loop, skip it.
408 Value *InVal = PN->getIncomingValue(i);
409 if (!isa<Instruction>(InVal))
412 // If this pred is for a subloop, not L itself, skip it.
413 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
414 continue; // The Block is in a subloop, skip it.
416 // Check that InVal is defined in the loop.
417 Instruction *Inst = cast<Instruction>(InVal);
418 if (!L->contains(Inst))
421 // Okay, this instruction has a user outside of the current loop
422 // and varies predictably *inside* the loop. Evaluate the value it
423 // contains when the loop exits, if possible.
424 const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
425 if (!SE->isLoopInvariant(ExitValue, L))
428 Value *ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), Inst);
430 DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
431 << " LoopVal = " << *Inst << "\n");
433 if (!isValidRewrite(Inst, ExitVal)) {
434 DeadInsts.push_back(ExitVal);
440 PN->setIncomingValue(i, ExitVal);
442 // If this instruction is dead now, delete it.
443 RecursivelyDeleteTriviallyDeadInstructions(Inst);
446 // Completely replace a single-pred PHI. This is safe, because the
447 // NewVal won't be variant in the loop, so we don't need an LCSSA phi
449 PN->replaceAllUsesWith(ExitVal);
450 RecursivelyDeleteTriviallyDeadInstructions(PN);
454 // Clone the PHI and delete the original one. This lets IVUsers and
455 // any other maps purge the original user from their records.
456 PHINode *NewPN = cast<PHINode>(PN->clone());
458 NewPN->insertBefore(PN);
459 PN->replaceAllUsesWith(NewPN);
460 PN->eraseFromParent();
465 // The insertion point instruction may have been deleted; clear it out
466 // so that the rewriter doesn't trip over it later.
467 Rewriter.clearInsertPoint();
470 void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
471 // First step. Check to see if there are any floating-point recurrences.
472 // If there are, change them into integer recurrences, permitting analysis by
473 // the SCEV routines.
475 BasicBlock *Header = L->getHeader();
477 SmallVector<WeakVH, 8> PHIs;
478 for (BasicBlock::iterator I = Header->begin();
479 PHINode *PN = dyn_cast<PHINode>(I); ++I)
482 for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
483 if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i]))
484 HandleFloatingPointIV(L, PN);
486 // If the loop previously had floating-point IV, ScalarEvolution
487 // may not have been able to compute a trip count. Now that we've done some
488 // re-writing, the trip count may be computable.
493 /// SimplifyIVUsers - Iteratively perform simplification on IVUsers within this
494 /// loop. IVUsers is treated as a worklist. Each successive simplification may
495 /// push more users which may themselves be candidates for simplification.
497 /// This is the old approach to IV simplification to be replaced by
498 /// SimplifyIVUsersNoRewrite.
500 void IndVarSimplify::SimplifyIVUsers(SCEVExpander &Rewriter) {
501 // Each round of simplification involves a round of eliminating operations
502 // followed by a round of widening IVs. A single IVUsers worklist is used
503 // across all rounds. The inner loop advances the user. If widening exposes
504 // more uses, then another pass through the outer loop is triggered.
505 for (IVUsers::iterator I = IU->begin(); I != IU->end(); ++I) {
506 Instruction *UseInst = I->getUser();
507 Value *IVOperand = I->getOperandValToReplace();
509 if (ICmpInst *ICmp = dyn_cast<ICmpInst>(UseInst)) {
510 EliminateIVComparison(ICmp, IVOperand);
513 if (BinaryOperator *Rem = dyn_cast<BinaryOperator>(UseInst)) {
514 bool IsSigned = Rem->getOpcode() == Instruction::SRem;
515 if (IsSigned || Rem->getOpcode() == Instruction::URem) {
516 EliminateIVRemainder(Rem, IVOperand, IsSigned);
524 // Collect information about induction variables that are used by sign/zero
525 // extend operations. This information is recorded by CollectExtend and
526 // provides the input to WidenIV.
528 const Type *WidestNativeType; // Widest integer type created [sz]ext
529 bool IsSigned; // Was an sext user seen before a zext?
531 WideIVInfo() : WidestNativeType(0), IsSigned(false) {}
535 /// CollectExtend - Update information about the induction variable that is
536 /// extended by this sign or zero extend operation. This is used to determine
537 /// the final width of the IV before actually widening it.
538 static void CollectExtend(CastInst *Cast, bool IsSigned, WideIVInfo &WI,
539 ScalarEvolution *SE, const TargetData *TD) {
540 const Type *Ty = Cast->getType();
541 uint64_t Width = SE->getTypeSizeInBits(Ty);
542 if (TD && !TD->isLegalInteger(Width))
545 if (!WI.WidestNativeType) {
546 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
547 WI.IsSigned = IsSigned;
551 // We extend the IV to satisfy the sign of its first user, arbitrarily.
552 if (WI.IsSigned != IsSigned)
555 if (Width > SE->getTypeSizeInBits(WI.WidestNativeType))
556 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
560 /// WidenIV - The goal of this transform is to remove sign and zero extends
561 /// without creating any new induction variables. To do this, it creates a new
562 /// phi of the wider type and redirects all users, either removing extends or
563 /// inserting truncs whenever we stop propagating the type.
568 const Type *WideType;
579 Instruction *WideInc;
580 const SCEV *WideIncExpr;
581 SmallVectorImpl<WeakVH> &DeadInsts;
583 SmallPtrSet<Instruction*,16> Widened;
584 SmallVector<std::pair<Use *, Instruction *>, 8> NarrowIVUsers;
587 WidenIV(PHINode *PN, const WideIVInfo &WI, LoopInfo *LInfo,
588 ScalarEvolution *SEv, DominatorTree *DTree,
589 SmallVectorImpl<WeakVH> &DI) :
591 WideType(WI.WidestNativeType),
592 IsSigned(WI.IsSigned),
594 L(LI->getLoopFor(OrigPhi->getParent())),
601 assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV");
604 PHINode *CreateWideIV(SCEVExpander &Rewriter);
607 Instruction *CloneIVUser(Instruction *NarrowUse,
608 Instruction *NarrowDef,
609 Instruction *WideDef);
611 const SCEVAddRecExpr *GetWideRecurrence(Instruction *NarrowUse);
613 Instruction *WidenIVUse(Use &NarrowDefUse, Instruction *NarrowDef,
614 Instruction *WideDef);
616 void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef);
618 } // anonymous namespace
620 static Value *getExtend( Value *NarrowOper, const Type *WideType,
621 bool IsSigned, IRBuilder<> &Builder) {
622 return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) :
623 Builder.CreateZExt(NarrowOper, WideType);
626 /// CloneIVUser - Instantiate a wide operation to replace a narrow
627 /// operation. This only needs to handle operations that can evaluation to
628 /// SCEVAddRec. It can safely return 0 for any operation we decide not to clone.
629 Instruction *WidenIV::CloneIVUser(Instruction *NarrowUse,
630 Instruction *NarrowDef,
631 Instruction *WideDef) {
632 unsigned Opcode = NarrowUse->getOpcode();
636 case Instruction::Add:
637 case Instruction::Mul:
638 case Instruction::UDiv:
639 case Instruction::Sub:
640 case Instruction::And:
641 case Instruction::Or:
642 case Instruction::Xor:
643 case Instruction::Shl:
644 case Instruction::LShr:
645 case Instruction::AShr:
646 DEBUG(dbgs() << "Cloning IVUser: " << *NarrowUse << "\n");
648 IRBuilder<> Builder(NarrowUse);
650 // Replace NarrowDef operands with WideDef. Otherwise, we don't know
651 // anything about the narrow operand yet so must insert a [sz]ext. It is
652 // probably loop invariant and will be folded or hoisted. If it actually
653 // comes from a widened IV, it should be removed during a future call to
655 Value *LHS = (NarrowUse->getOperand(0) == NarrowDef) ? WideDef :
656 getExtend(NarrowUse->getOperand(0), WideType, IsSigned, Builder);
657 Value *RHS = (NarrowUse->getOperand(1) == NarrowDef) ? WideDef :
658 getExtend(NarrowUse->getOperand(1), WideType, IsSigned, Builder);
660 BinaryOperator *NarrowBO = cast<BinaryOperator>(NarrowUse);
661 BinaryOperator *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(),
663 NarrowBO->getName());
664 Builder.Insert(WideBO);
665 if (const OverflowingBinaryOperator *OBO =
666 dyn_cast<OverflowingBinaryOperator>(NarrowBO)) {
667 if (OBO->hasNoUnsignedWrap()) WideBO->setHasNoUnsignedWrap();
668 if (OBO->hasNoSignedWrap()) WideBO->setHasNoSignedWrap();
675 /// HoistStep - Attempt to hoist an IV increment above a potential use.
677 /// To successfully hoist, two criteria must be met:
678 /// - IncV operands dominate InsertPos and
679 /// - InsertPos dominates IncV
681 /// Meeting the second condition means that we don't need to check all of IncV's
682 /// existing uses (it's moving up in the domtree).
684 /// This does not yet recursively hoist the operands, although that would
685 /// not be difficult.
686 static bool HoistStep(Instruction *IncV, Instruction *InsertPos,
687 const DominatorTree *DT)
689 if (DT->dominates(IncV, InsertPos))
692 if (!DT->dominates(InsertPos->getParent(), IncV->getParent()))
695 if (IncV->mayHaveSideEffects())
698 // Attempt to hoist IncV
699 for (User::op_iterator OI = IncV->op_begin(), OE = IncV->op_end();
701 Instruction *OInst = dyn_cast<Instruction>(OI);
702 if (OInst && !DT->dominates(OInst, InsertPos))
705 IncV->moveBefore(InsertPos);
709 // GetWideRecurrence - Is this instruction potentially interesting from IVUsers'
710 // perspective after widening it's type? In other words, can the extend be
711 // safely hoisted out of the loop with SCEV reducing the value to a recurrence
712 // on the same loop. If so, return the sign or zero extended
713 // recurrence. Otherwise return NULL.
714 const SCEVAddRecExpr *WidenIV::GetWideRecurrence(Instruction *NarrowUse) {
715 if (!SE->isSCEVable(NarrowUse->getType()))
718 const SCEV *NarrowExpr = SE->getSCEV(NarrowUse);
719 if (SE->getTypeSizeInBits(NarrowExpr->getType())
720 >= SE->getTypeSizeInBits(WideType)) {
721 // NarrowUse implicitly widens its operand. e.g. a gep with a narrow
722 // index. So don't follow this use.
726 const SCEV *WideExpr = IsSigned ?
727 SE->getSignExtendExpr(NarrowExpr, WideType) :
728 SE->getZeroExtendExpr(NarrowExpr, WideType);
729 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr);
730 if (!AddRec || AddRec->getLoop() != L)
736 /// WidenIVUse - Determine whether an individual user of the narrow IV can be
737 /// widened. If so, return the wide clone of the user.
738 Instruction *WidenIV::WidenIVUse(Use &NarrowDefUse, Instruction *NarrowDef,
739 Instruction *WideDef) {
740 Instruction *NarrowUse = cast<Instruction>(NarrowDefUse.getUser());
742 // Stop traversing the def-use chain at inner-loop phis or post-loop phis.
743 if (isa<PHINode>(NarrowUse) && LI->getLoopFor(NarrowUse->getParent()) != L)
746 // Our raison d'etre! Eliminate sign and zero extension.
747 if (IsSigned ? isa<SExtInst>(NarrowUse) : isa<ZExtInst>(NarrowUse)) {
748 Value *NewDef = WideDef;
749 if (NarrowUse->getType() != WideType) {
750 unsigned CastWidth = SE->getTypeSizeInBits(NarrowUse->getType());
751 unsigned IVWidth = SE->getTypeSizeInBits(WideType);
752 if (CastWidth < IVWidth) {
753 // The cast isn't as wide as the IV, so insert a Trunc.
754 IRBuilder<> Builder(NarrowDefUse);
755 NewDef = Builder.CreateTrunc(WideDef, NarrowUse->getType());
758 // A wider extend was hidden behind a narrower one. This may induce
759 // another round of IV widening in which the intermediate IV becomes
760 // dead. It should be very rare.
761 DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi
762 << " not wide enough to subsume " << *NarrowUse << "\n");
763 NarrowUse->replaceUsesOfWith(NarrowDef, WideDef);
767 if (NewDef != NarrowUse) {
768 DEBUG(dbgs() << "INDVARS: eliminating " << *NarrowUse
769 << " replaced by " << *WideDef << "\n");
771 NarrowUse->replaceAllUsesWith(NewDef);
772 DeadInsts.push_back(NarrowUse);
774 // Now that the extend is gone, we want to expose it's uses for potential
775 // further simplification. We don't need to directly inform SimplifyIVUsers
776 // of the new users, because their parent IV will be processed later as a
777 // new loop phi. If we preserved IVUsers analysis, we would also want to
778 // push the uses of WideDef here.
780 // No further widening is needed. The deceased [sz]ext had done it for us.
784 // Does this user itself evaluate to a recurrence after widening?
785 const SCEVAddRecExpr *WideAddRec = GetWideRecurrence(NarrowUse);
787 // This user does not evaluate to a recurence after widening, so don't
788 // follow it. Instead insert a Trunc to kill off the original use,
789 // eventually isolating the original narrow IV so it can be removed.
790 IRBuilder<> Builder(NarrowDefUse);
791 Value *Trunc = Builder.CreateTrunc(WideDef, NarrowDef->getType());
792 NarrowUse->replaceUsesOfWith(NarrowDef, Trunc);
795 // We assume that block terminators are not SCEVable. We wouldn't want to
796 // insert a Trunc after a terminator if there happens to be a critical edge.
797 assert(NarrowUse != NarrowUse->getParent()->getTerminator() &&
798 "SCEV is not expected to evaluate a block terminator");
800 // Reuse the IV increment that SCEVExpander created as long as it dominates
802 Instruction *WideUse = 0;
803 if (WideAddRec == WideIncExpr && HoistStep(WideInc, NarrowUse, DT)) {
807 WideUse = CloneIVUser(NarrowUse, NarrowDef, WideDef);
811 // Evaluation of WideAddRec ensured that the narrow expression could be
812 // extended outside the loop without overflow. This suggests that the wide use
813 // evaluates to the same expression as the extended narrow use, but doesn't
814 // absolutely guarantee it. Hence the following failsafe check. In rare cases
815 // where it fails, we simply throw away the newly created wide use.
816 if (WideAddRec != SE->getSCEV(WideUse)) {
817 DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse
818 << ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n");
819 DeadInsts.push_back(WideUse);
823 // Returning WideUse pushes it on the worklist.
827 /// pushNarrowIVUsers - Add eligible users of NarrowDef to NarrowIVUsers.
829 void WidenIV::pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef) {
830 for (Value::use_iterator UI = NarrowDef->use_begin(),
831 UE = NarrowDef->use_end(); UI != UE; ++UI) {
832 Use &U = UI.getUse();
834 // Handle data flow merges and bizarre phi cycles.
835 if (!Widened.insert(cast<Instruction>(U.getUser())))
838 NarrowIVUsers.push_back(std::make_pair(&UI.getUse(), WideDef));
842 /// CreateWideIV - Process a single induction variable. First use the
843 /// SCEVExpander to create a wide induction variable that evaluates to the same
844 /// recurrence as the original narrow IV. Then use a worklist to forward
845 /// traverse the narrow IV's def-use chain. After WidenIVUse has processed all
846 /// interesting IV users, the narrow IV will be isolated for removal by
849 /// It would be simpler to delete uses as they are processed, but we must avoid
850 /// invalidating SCEV expressions.
852 PHINode *WidenIV::CreateWideIV(SCEVExpander &Rewriter) {
853 // Is this phi an induction variable?
854 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi));
858 // Widen the induction variable expression.
859 const SCEV *WideIVExpr = IsSigned ?
860 SE->getSignExtendExpr(AddRec, WideType) :
861 SE->getZeroExtendExpr(AddRec, WideType);
863 assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType &&
864 "Expect the new IV expression to preserve its type");
866 // Can the IV be extended outside the loop without overflow?
867 AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr);
868 if (!AddRec || AddRec->getLoop() != L)
871 // An AddRec must have loop-invariant operands. Since this AddRec is
872 // materialized by a loop header phi, the expression cannot have any post-loop
873 // operands, so they must dominate the loop header.
874 assert(SE->properlyDominates(AddRec->getStart(), L->getHeader()) &&
875 SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader())
876 && "Loop header phi recurrence inputs do not dominate the loop");
878 // The rewriter provides a value for the desired IV expression. This may
879 // either find an existing phi or materialize a new one. Either way, we
880 // expect a well-formed cyclic phi-with-increments. i.e. any operand not part
881 // of the phi-SCC dominates the loop entry.
882 Instruction *InsertPt = L->getHeader()->begin();
883 WidePhi = cast<PHINode>(Rewriter.expandCodeFor(AddRec, WideType, InsertPt));
885 // Remembering the WideIV increment generated by SCEVExpander allows
886 // WidenIVUse to reuse it when widening the narrow IV's increment. We don't
887 // employ a general reuse mechanism because the call above is the only call to
888 // SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses.
889 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
891 cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock));
892 WideIncExpr = SE->getSCEV(WideInc);
895 DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n");
898 // Traverse the def-use chain using a worklist starting at the original IV.
899 assert(Widened.empty() && NarrowIVUsers.empty() && "expect initial state" );
901 Widened.insert(OrigPhi);
902 pushNarrowIVUsers(OrigPhi, WidePhi);
904 while (!NarrowIVUsers.empty()) {
906 Instruction *WideDef;
907 tie(UsePtr, WideDef) = NarrowIVUsers.pop_back_val();
908 Use &NarrowDefUse = *UsePtr;
910 // Process a def-use edge. This may replace the use, so don't hold a
911 // use_iterator across it.
912 Instruction *NarrowDef = cast<Instruction>(NarrowDefUse.get());
913 Instruction *WideUse = WidenIVUse(NarrowDefUse, NarrowDef, WideDef);
915 // Follow all def-use edges from the previous narrow use.
917 pushNarrowIVUsers(cast<Instruction>(NarrowDefUse.getUser()), WideUse);
919 // WidenIVUse may have removed the def-use edge.
920 if (NarrowDef->use_empty())
921 DeadInsts.push_back(NarrowDef);
926 void IndVarSimplify::EliminateIVComparison(ICmpInst *ICmp, Value *IVOperand) {
927 unsigned IVOperIdx = 0;
928 ICmpInst::Predicate Pred = ICmp->getPredicate();
929 if (IVOperand != ICmp->getOperand(0)) {
931 assert(IVOperand == ICmp->getOperand(1) && "Can't find IVOperand");
933 Pred = ICmpInst::getSwappedPredicate(Pred);
936 // Get the SCEVs for the ICmp operands.
937 const SCEV *S = SE->getSCEV(ICmp->getOperand(IVOperIdx));
938 const SCEV *X = SE->getSCEV(ICmp->getOperand(1 - IVOperIdx));
940 // Simplify unnecessary loops away.
941 const Loop *ICmpLoop = LI->getLoopFor(ICmp->getParent());
942 S = SE->getSCEVAtScope(S, ICmpLoop);
943 X = SE->getSCEVAtScope(X, ICmpLoop);
945 // If the condition is always true or always false, replace it with
947 if (SE->isKnownPredicate(Pred, S, X))
948 ICmp->replaceAllUsesWith(ConstantInt::getTrue(ICmp->getContext()));
949 else if (SE->isKnownPredicate(ICmpInst::getInversePredicate(Pred), S, X))
950 ICmp->replaceAllUsesWith(ConstantInt::getFalse(ICmp->getContext()));
954 DEBUG(dbgs() << "INDVARS: Eliminated comparison: " << *ICmp << '\n');
957 DeadInsts.push_back(ICmp);
960 void IndVarSimplify::EliminateIVRemainder(BinaryOperator *Rem,
963 // We're only interested in the case where we know something about
965 if (IVOperand != Rem->getOperand(0))
968 // Get the SCEVs for the ICmp operands.
969 const SCEV *S = SE->getSCEV(Rem->getOperand(0));
970 const SCEV *X = SE->getSCEV(Rem->getOperand(1));
972 // Simplify unnecessary loops away.
973 const Loop *ICmpLoop = LI->getLoopFor(Rem->getParent());
974 S = SE->getSCEVAtScope(S, ICmpLoop);
975 X = SE->getSCEVAtScope(X, ICmpLoop);
977 // i % n --> i if i is in [0,n).
978 if ((!IsSigned || SE->isKnownNonNegative(S)) &&
979 SE->isKnownPredicate(IsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
981 Rem->replaceAllUsesWith(Rem->getOperand(0));
983 // (i+1) % n --> (i+1)==n?0:(i+1) if i is in [0,n).
984 const SCEV *LessOne =
985 SE->getMinusSCEV(S, SE->getConstant(S->getType(), 1));
986 if (IsSigned && !SE->isKnownNonNegative(LessOne))
989 if (!SE->isKnownPredicate(IsSigned ?
990 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
994 ICmpInst *ICmp = new ICmpInst(Rem, ICmpInst::ICMP_EQ,
995 Rem->getOperand(0), Rem->getOperand(1),
998 SelectInst::Create(ICmp,
999 ConstantInt::get(Rem->getType(), 0),
1000 Rem->getOperand(0), "tmp", Rem);
1001 Rem->replaceAllUsesWith(Sel);
1004 // Inform IVUsers about the new users.
1006 if (Instruction *I = dyn_cast<Instruction>(Rem->getOperand(0)))
1007 IU->AddUsersIfInteresting(I);
1009 DEBUG(dbgs() << "INDVARS: Simplified rem: " << *Rem << '\n');
1012 DeadInsts.push_back(Rem);
1015 /// EliminateIVUser - Eliminate an operation that consumes a simple IV and has
1016 /// no observable side-effect given the range of IV values.
1017 bool IndVarSimplify::EliminateIVUser(Instruction *UseInst,
1018 Instruction *IVOperand) {
1019 if (ICmpInst *ICmp = dyn_cast<ICmpInst>(UseInst)) {
1020 EliminateIVComparison(ICmp, IVOperand);
1023 if (BinaryOperator *Rem = dyn_cast<BinaryOperator>(UseInst)) {
1024 bool IsSigned = Rem->getOpcode() == Instruction::SRem;
1025 if (IsSigned || Rem->getOpcode() == Instruction::URem) {
1026 EliminateIVRemainder(Rem, IVOperand, IsSigned);
1031 // Eliminate any operation that SCEV can prove is an identity function.
1032 if (!SE->isSCEVable(UseInst->getType()) ||
1033 (UseInst->getType() != IVOperand->getType()) ||
1034 (SE->getSCEV(UseInst) != SE->getSCEV(IVOperand)))
1037 DEBUG(dbgs() << "INDVARS: Eliminated identity: " << *UseInst << '\n');
1039 UseInst->replaceAllUsesWith(IVOperand);
1042 DeadInsts.push_back(UseInst);
1046 /// pushIVUsers - Add all uses of Def to the current IV's worklist.
1048 static void pushIVUsers(
1050 SmallPtrSet<Instruction*,16> &Simplified,
1051 SmallVectorImpl< std::pair<Instruction*,Instruction*> > &SimpleIVUsers) {
1053 for (Value::use_iterator UI = Def->use_begin(), E = Def->use_end();
1055 Instruction *User = cast<Instruction>(*UI);
1057 // Avoid infinite or exponential worklist processing.
1058 // Also ensure unique worklist users.
1059 // If Def is a LoopPhi, it may not be in the Simplified set, so check for
1060 // self edges first.
1061 if (User != Def && Simplified.insert(User))
1062 SimpleIVUsers.push_back(std::make_pair(User, Def));
1066 /// isSimpleIVUser - Return true if this instruction generates a simple SCEV
1067 /// expression in terms of that IV.
1069 /// This is similar to IVUsers' isInsteresting() but processes each instruction
1070 /// non-recursively when the operand is already known to be a simpleIVUser.
1072 bool IndVarSimplify::isSimpleIVUser(Instruction *I, const Loop *L) {
1073 if (!SE->isSCEVable(I->getType()))
1076 // Get the symbolic expression for this instruction.
1077 const SCEV *S = SE->getSCEV(I);
1079 // We assume that terminators are not SCEVable.
1080 assert((!S || I != I->getParent()->getTerminator()) &&
1081 "can't fold terminators");
1083 // Only consider affine recurrences.
1084 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S);
1085 if (AR && AR->getLoop() == L)
1091 /// SimplifyIVUsersNoRewrite - Iteratively perform simplification on a worklist
1092 /// of IV users. Each successive simplification may push more users which may
1093 /// themselves be candidates for simplification.
1095 /// The "NoRewrite" algorithm does not require IVUsers analysis. Instead, it
1096 /// simplifies instructions in-place during analysis. Rather than rewriting
1097 /// induction variables bottom-up from their users, it transforms a chain of
1098 /// IVUsers top-down, updating the IR only when it encouters a clear
1099 /// optimization opportunitiy. A SCEVExpander "Rewriter" instance is still
1100 /// needed, but only used to generate a new IV (phi) of wider type for sign/zero
1101 /// extend elimination.
1103 /// Once DisableIVRewrite is default, LSR will be the only client of IVUsers.
1105 void IndVarSimplify::SimplifyIVUsersNoRewrite(Loop *L, SCEVExpander &Rewriter) {
1106 std::map<PHINode *, WideIVInfo> WideIVMap;
1108 SmallVector<PHINode*, 8> LoopPhis;
1109 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1110 LoopPhis.push_back(cast<PHINode>(I));
1112 // Each round of simplification iterates through the SimplifyIVUsers worklist
1113 // for all current phis, then determines whether any IVs can be
1114 // widened. Widening adds new phis to LoopPhis, inducing another round of
1115 // simplification on the wide IVs.
1116 while (!LoopPhis.empty()) {
1117 // Evaluate as many IV expressions as possible before widening any IVs. This
1118 // forces SCEV to set no-wrap flags before evaluating sign/zero
1119 // extension. The first time SCEV attempts to normalize sign/zero extension,
1120 // the result becomes final. So for the most predictable results, we delay
1121 // evaluation of sign/zero extend evaluation until needed, and avoid running
1122 // other SCEV based analysis prior to SimplifyIVUsersNoRewrite.
1124 PHINode *CurrIV = LoopPhis.pop_back_val();
1126 // Information about sign/zero extensions of CurrIV.
1129 // Instructions processed by SimplifyIVUsers for CurrIV.
1130 SmallPtrSet<Instruction*,16> Simplified;
1132 // Use-def pairs if IVUsers waiting to be processed for CurrIV.
1133 SmallVector<std::pair<Instruction*, Instruction*>, 8> SimpleIVUsers;
1135 // Push users of the current LoopPhi. In rare cases, pushIVUsers may be
1136 // called multiple times for the same LoopPhi. This is the proper thing to
1137 // do for loop header phis that use each other.
1138 pushIVUsers(CurrIV, Simplified, SimpleIVUsers);
1140 while (!SimpleIVUsers.empty()) {
1141 Instruction *UseInst, *Operand;
1142 tie(UseInst, Operand) = SimpleIVUsers.pop_back_val();
1143 // Bypass back edges to avoid extra work.
1144 if (UseInst == CurrIV) continue;
1146 if (EliminateIVUser(UseInst, Operand)) {
1147 pushIVUsers(Operand, Simplified, SimpleIVUsers);
1150 if (CastInst *Cast = dyn_cast<CastInst>(UseInst)) {
1151 bool IsSigned = Cast->getOpcode() == Instruction::SExt;
1152 if (IsSigned || Cast->getOpcode() == Instruction::ZExt) {
1153 CollectExtend(Cast, IsSigned, WI, SE, TD);
1157 if (isSimpleIVUser(UseInst, L)) {
1158 pushIVUsers(UseInst, Simplified, SimpleIVUsers);
1161 if (WI.WidestNativeType) {
1162 WideIVMap[CurrIV] = WI;
1164 } while(!LoopPhis.empty());
1166 for (std::map<PHINode *, WideIVInfo>::const_iterator I = WideIVMap.begin(),
1167 E = WideIVMap.end(); I != E; ++I) {
1168 WidenIV Widener(I->first, I->second, LI, SE, DT, DeadInsts);
1169 if (PHINode *WidePhi = Widener.CreateWideIV(Rewriter)) {
1171 LoopPhis.push_back(WidePhi);
1178 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
1179 // If LoopSimplify form is not available, stay out of trouble. Some notes:
1180 // - LSR currently only supports LoopSimplify-form loops. Indvars'
1181 // canonicalization can be a pessimization without LSR to "clean up"
1183 // - We depend on having a preheader; in particular,
1184 // Loop::getCanonicalInductionVariable only supports loops with preheaders,
1185 // and we're in trouble if we can't find the induction variable even when
1186 // we've manually inserted one.
1187 if (!L->isLoopSimplifyForm())
1190 if (!DisableIVRewrite)
1191 IU = &getAnalysis<IVUsers>();
1192 LI = &getAnalysis<LoopInfo>();
1193 SE = &getAnalysis<ScalarEvolution>();
1194 DT = &getAnalysis<DominatorTree>();
1195 TD = getAnalysisIfAvailable<TargetData>();
1200 // If there are any floating-point recurrences, attempt to
1201 // transform them to use integer recurrences.
1202 RewriteNonIntegerIVs(L);
1204 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1206 // Create a rewriter object which we'll use to transform the code with.
1207 SCEVExpander Rewriter(*SE, "indvars");
1209 // Eliminate redundant IV users.
1211 // Simplification works best when run before other consumers of SCEV. We
1212 // attempt to avoid evaluating SCEVs for sign/zero extend operations until
1213 // other expressions involving loop IVs have been evaluated. This helps SCEV
1214 // set no-wrap flags before normalizing sign/zero extension.
1215 if (DisableIVRewrite) {
1216 Rewriter.disableCanonicalMode();
1217 SimplifyIVUsersNoRewrite(L, Rewriter);
1220 // Check to see if this loop has a computable loop-invariant execution count.
1221 // If so, this means that we can compute the final value of any expressions
1222 // that are recurrent in the loop, and substitute the exit values from the
1223 // loop into any instructions outside of the loop that use the final values of
1224 // the current expressions.
1226 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1227 RewriteLoopExitValues(L, Rewriter);
1229 // Eliminate redundant IV users.
1230 if (!DisableIVRewrite)
1231 SimplifyIVUsers(Rewriter);
1233 // Compute the type of the largest recurrence expression, and decide whether
1234 // a canonical induction variable should be inserted.
1235 const Type *LargestType = 0;
1236 bool NeedCannIV = false;
1237 bool ExpandBECount = canExpandBackedgeTakenCount(L, SE);
1238 if (ExpandBECount) {
1239 // If we have a known trip count and a single exit block, we'll be
1240 // rewriting the loop exit test condition below, which requires a
1241 // canonical induction variable.
1243 const Type *Ty = BackedgeTakenCount->getType();
1244 if (DisableIVRewrite) {
1245 // In this mode, SimplifyIVUsers may have already widened the IV used by
1246 // the backedge test and inserted a Trunc on the compare's operand. Get
1247 // the wider type to avoid creating a redundant narrow IV only used by the
1249 LargestType = getBackedgeIVType(L);
1252 SE->getTypeSizeInBits(Ty) >
1253 SE->getTypeSizeInBits(LargestType))
1254 LargestType = SE->getEffectiveSCEVType(Ty);
1256 if (!DisableIVRewrite) {
1257 for (IVUsers::const_iterator I = IU->begin(), E = IU->end(); I != E; ++I) {
1260 SE->getEffectiveSCEVType(I->getOperandValToReplace()->getType());
1262 SE->getTypeSizeInBits(Ty) >
1263 SE->getTypeSizeInBits(LargestType))
1268 // Now that we know the largest of the induction variable expressions
1269 // in this loop, insert a canonical induction variable of the largest size.
1270 PHINode *IndVar = 0;
1272 // Check to see if the loop already has any canonical-looking induction
1273 // variables. If any are present and wider than the planned canonical
1274 // induction variable, temporarily remove them, so that the Rewriter
1275 // doesn't attempt to reuse them.
1276 SmallVector<PHINode *, 2> OldCannIVs;
1277 while (PHINode *OldCannIV = L->getCanonicalInductionVariable()) {
1278 if (SE->getTypeSizeInBits(OldCannIV->getType()) >
1279 SE->getTypeSizeInBits(LargestType))
1280 OldCannIV->removeFromParent();
1283 OldCannIVs.push_back(OldCannIV);
1286 IndVar = Rewriter.getOrInsertCanonicalInductionVariable(L, LargestType);
1290 DEBUG(dbgs() << "INDVARS: New CanIV: " << *IndVar << '\n');
1292 // Now that the official induction variable is established, reinsert
1293 // any old canonical-looking variables after it so that the IR remains
1294 // consistent. They will be deleted as part of the dead-PHI deletion at
1295 // the end of the pass.
1296 while (!OldCannIVs.empty()) {
1297 PHINode *OldCannIV = OldCannIVs.pop_back_val();
1298 OldCannIV->insertBefore(L->getHeader()->getFirstNonPHI());
1302 // If we have a trip count expression, rewrite the loop's exit condition
1303 // using it. We can currently only handle loops with a single exit.
1304 ICmpInst *NewICmp = 0;
1305 if (ExpandBECount) {
1306 assert(canExpandBackedgeTakenCount(L, SE) &&
1307 "canonical IV disrupted BackedgeTaken expansion");
1308 assert(NeedCannIV &&
1309 "LinearFunctionTestReplace requires a canonical induction variable");
1310 NewICmp = LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar,
1313 // Rewrite IV-derived expressions.
1314 if (!DisableIVRewrite)
1315 RewriteIVExpressions(L, Rewriter);
1317 // Clear the rewriter cache, because values that are in the rewriter's cache
1318 // can be deleted in the loop below, causing the AssertingVH in the cache to
1322 // Now that we're done iterating through lists, clean up any instructions
1323 // which are now dead.
1324 while (!DeadInsts.empty())
1325 if (Instruction *Inst =
1326 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
1327 RecursivelyDeleteTriviallyDeadInstructions(Inst);
1329 // The Rewriter may not be used from this point on.
1331 // Loop-invariant instructions in the preheader that aren't used in the
1332 // loop may be sunk below the loop to reduce register pressure.
1333 SinkUnusedInvariants(L);
1335 // For completeness, inform IVUsers of the IV use in the newly-created
1336 // loop exit test instruction.
1338 IU->AddUsersIfInteresting(cast<Instruction>(NewICmp->getOperand(0)));
1340 // Clean up dead instructions.
1341 Changed |= DeleteDeadPHIs(L->getHeader());
1342 // Check a post-condition.
1343 assert(L->isLCSSAForm(*DT) && "Indvars did not leave the loop in lcssa form!");
1347 // FIXME: It is an extremely bad idea to indvar substitute anything more
1348 // complex than affine induction variables. Doing so will put expensive
1349 // polynomial evaluations inside of the loop, and the str reduction pass
1350 // currently can only reduce affine polynomials. For now just disable
1351 // indvar subst on anything more complex than an affine addrec, unless
1352 // it can be expanded to a trivial value.
1353 static bool isSafe(const SCEV *S, const Loop *L, ScalarEvolution *SE) {
1354 // Loop-invariant values are safe.
1355 if (SE->isLoopInvariant(S, L)) return true;
1357 // Affine addrecs are safe. Non-affine are not, because LSR doesn't know how
1358 // to transform them into efficient code.
1359 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
1360 return AR->isAffine();
1362 // An add is safe it all its operands are safe.
1363 if (const SCEVCommutativeExpr *Commutative = dyn_cast<SCEVCommutativeExpr>(S)) {
1364 for (SCEVCommutativeExpr::op_iterator I = Commutative->op_begin(),
1365 E = Commutative->op_end(); I != E; ++I)
1366 if (!isSafe(*I, L, SE)) return false;
1370 // A cast is safe if its operand is.
1371 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
1372 return isSafe(C->getOperand(), L, SE);
1374 // A udiv is safe if its operands are.
1375 if (const SCEVUDivExpr *UD = dyn_cast<SCEVUDivExpr>(S))
1376 return isSafe(UD->getLHS(), L, SE) &&
1377 isSafe(UD->getRHS(), L, SE);
1379 // SCEVUnknown is always safe.
1380 if (isa<SCEVUnknown>(S))
1383 // Nothing else is safe.
1387 void IndVarSimplify::RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter) {
1388 // Rewrite all induction variable expressions in terms of the canonical
1389 // induction variable.
1391 // If there were induction variables of other sizes or offsets, manually
1392 // add the offsets to the primary induction variable and cast, avoiding
1393 // the need for the code evaluation methods to insert induction variables
1394 // of different sizes.
1395 for (IVUsers::iterator UI = IU->begin(), E = IU->end(); UI != E; ++UI) {
1396 Value *Op = UI->getOperandValToReplace();
1397 const Type *UseTy = Op->getType();
1398 Instruction *User = UI->getUser();
1400 // Compute the final addrec to expand into code.
1401 const SCEV *AR = IU->getReplacementExpr(*UI);
1403 // Evaluate the expression out of the loop, if possible.
1404 if (!L->contains(UI->getUser())) {
1405 const SCEV *ExitVal = SE->getSCEVAtScope(AR, L->getParentLoop());
1406 if (SE->isLoopInvariant(ExitVal, L))
1410 // FIXME: It is an extremely bad idea to indvar substitute anything more
1411 // complex than affine induction variables. Doing so will put expensive
1412 // polynomial evaluations inside of the loop, and the str reduction pass
1413 // currently can only reduce affine polynomials. For now just disable
1414 // indvar subst on anything more complex than an affine addrec, unless
1415 // it can be expanded to a trivial value.
1416 if (!isSafe(AR, L, SE))
1419 // Determine the insertion point for this user. By default, insert
1420 // immediately before the user. The SCEVExpander class will automatically
1421 // hoist loop invariants out of the loop. For PHI nodes, there may be
1422 // multiple uses, so compute the nearest common dominator for the
1424 Instruction *InsertPt = User;
1425 if (PHINode *PHI = dyn_cast<PHINode>(InsertPt))
1426 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
1427 if (PHI->getIncomingValue(i) == Op) {
1428 if (InsertPt == User)
1429 InsertPt = PHI->getIncomingBlock(i)->getTerminator();
1432 DT->findNearestCommonDominator(InsertPt->getParent(),
1433 PHI->getIncomingBlock(i))
1437 // Now expand it into actual Instructions and patch it into place.
1438 Value *NewVal = Rewriter.expandCodeFor(AR, UseTy, InsertPt);
1440 DEBUG(dbgs() << "INDVARS: Rewrote IV '" << *AR << "' " << *Op << '\n'
1441 << " into = " << *NewVal << "\n");
1443 if (!isValidRewrite(Op, NewVal)) {
1444 DeadInsts.push_back(NewVal);
1447 // Inform ScalarEvolution that this value is changing. The change doesn't
1448 // affect its value, but it does potentially affect which use lists the
1449 // value will be on after the replacement, which affects ScalarEvolution's
1450 // ability to walk use lists and drop dangling pointers when a value is
1452 SE->forgetValue(User);
1454 // Patch the new value into place.
1456 NewVal->takeName(Op);
1457 if (Instruction *NewValI = dyn_cast<Instruction>(NewVal))
1458 NewValI->setDebugLoc(User->getDebugLoc());
1459 User->replaceUsesOfWith(Op, NewVal);
1460 UI->setOperandValToReplace(NewVal);
1465 // The old value may be dead now.
1466 DeadInsts.push_back(Op);
1470 /// If there's a single exit block, sink any loop-invariant values that
1471 /// were defined in the preheader but not used inside the loop into the
1472 /// exit block to reduce register pressure in the loop.
1473 void IndVarSimplify::SinkUnusedInvariants(Loop *L) {
1474 BasicBlock *ExitBlock = L->getExitBlock();
1475 if (!ExitBlock) return;
1477 BasicBlock *Preheader = L->getLoopPreheader();
1478 if (!Preheader) return;
1480 Instruction *InsertPt = ExitBlock->getFirstNonPHI();
1481 BasicBlock::iterator I = Preheader->getTerminator();
1482 while (I != Preheader->begin()) {
1484 // New instructions were inserted at the end of the preheader.
1485 if (isa<PHINode>(I))
1488 // Don't move instructions which might have side effects, since the side
1489 // effects need to complete before instructions inside the loop. Also don't
1490 // move instructions which might read memory, since the loop may modify
1491 // memory. Note that it's okay if the instruction might have undefined
1492 // behavior: LoopSimplify guarantees that the preheader dominates the exit
1494 if (I->mayHaveSideEffects() || I->mayReadFromMemory())
1497 // Skip debug info intrinsics.
1498 if (isa<DbgInfoIntrinsic>(I))
1501 // Don't sink static AllocaInsts out of the entry block, which would
1502 // turn them into dynamic allocas!
1503 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
1504 if (AI->isStaticAlloca())
1507 // Determine if there is a use in or before the loop (direct or
1509 bool UsedInLoop = false;
1510 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
1513 BasicBlock *UseBB = cast<Instruction>(U)->getParent();
1514 if (PHINode *P = dyn_cast<PHINode>(U)) {
1516 PHINode::getIncomingValueNumForOperand(UI.getOperandNo());
1517 UseBB = P->getIncomingBlock(i);
1519 if (UseBB == Preheader || L->contains(UseBB)) {
1525 // If there is, the def must remain in the preheader.
1529 // Otherwise, sink it to the exit block.
1530 Instruction *ToMove = I;
1533 if (I != Preheader->begin()) {
1534 // Skip debug info intrinsics.
1537 } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin());
1539 if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin())
1545 ToMove->moveBefore(InsertPt);
1551 /// ConvertToSInt - Convert APF to an integer, if possible.
1552 static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) {
1553 bool isExact = false;
1554 if (&APF.getSemantics() == &APFloat::PPCDoubleDouble)
1556 // See if we can convert this to an int64_t
1558 if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero,
1559 &isExact) != APFloat::opOK || !isExact)
1565 /// HandleFloatingPointIV - If the loop has floating induction variable
1566 /// then insert corresponding integer induction variable if possible.
1568 /// for(double i = 0; i < 10000; ++i)
1570 /// is converted into
1571 /// for(int i = 0; i < 10000; ++i)
1574 void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) {
1575 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
1576 unsigned BackEdge = IncomingEdge^1;
1578 // Check incoming value.
1579 ConstantFP *InitValueVal =
1580 dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));
1583 if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
1586 // Check IV increment. Reject this PN if increment operation is not
1587 // an add or increment value can not be represented by an integer.
1588 BinaryOperator *Incr =
1589 dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
1590 if (Incr == 0 || Incr->getOpcode() != Instruction::FAdd) return;
1592 // If this is not an add of the PHI with a constantfp, or if the constant fp
1593 // is not an integer, bail out.
1594 ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
1596 if (IncValueVal == 0 || Incr->getOperand(0) != PN ||
1597 !ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
1600 // Check Incr uses. One user is PN and the other user is an exit condition
1601 // used by the conditional terminator.
1602 Value::use_iterator IncrUse = Incr->use_begin();
1603 Instruction *U1 = cast<Instruction>(*IncrUse++);
1604 if (IncrUse == Incr->use_end()) return;
1605 Instruction *U2 = cast<Instruction>(*IncrUse++);
1606 if (IncrUse != Incr->use_end()) return;
1608 // Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't
1609 // only used by a branch, we can't transform it.
1610 FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
1612 Compare = dyn_cast<FCmpInst>(U2);
1613 if (Compare == 0 || !Compare->hasOneUse() ||
1614 !isa<BranchInst>(Compare->use_back()))
1617 BranchInst *TheBr = cast<BranchInst>(Compare->use_back());
1619 // We need to verify that the branch actually controls the iteration count
1620 // of the loop. If not, the new IV can overflow and no one will notice.
1621 // The branch block must be in the loop and one of the successors must be out
1623 assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
1624 if (!L->contains(TheBr->getParent()) ||
1625 (L->contains(TheBr->getSuccessor(0)) &&
1626 L->contains(TheBr->getSuccessor(1))))
1630 // If it isn't a comparison with an integer-as-fp (the exit value), we can't
1632 ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
1634 if (ExitValueVal == 0 ||
1635 !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
1638 // Find new predicate for integer comparison.
1639 CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
1640 switch (Compare->getPredicate()) {
1641 default: return; // Unknown comparison.
1642 case CmpInst::FCMP_OEQ:
1643 case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
1644 case CmpInst::FCMP_ONE:
1645 case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
1646 case CmpInst::FCMP_OGT:
1647 case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
1648 case CmpInst::FCMP_OGE:
1649 case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
1650 case CmpInst::FCMP_OLT:
1651 case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
1652 case CmpInst::FCMP_OLE:
1653 case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
1656 // We convert the floating point induction variable to a signed i32 value if
1657 // we can. This is only safe if the comparison will not overflow in a way
1658 // that won't be trapped by the integer equivalent operations. Check for this
1660 // TODO: We could use i64 if it is native and the range requires it.
1662 // The start/stride/exit values must all fit in signed i32.
1663 if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
1666 // If not actually striding (add x, 0.0), avoid touching the code.
1670 // Positive and negative strides have different safety conditions.
1672 // If we have a positive stride, we require the init to be less than the
1673 // exit value and an equality or less than comparison.
1674 if (InitValue >= ExitValue ||
1675 NewPred == CmpInst::ICMP_SGT || NewPred == CmpInst::ICMP_SGE)
1678 uint32_t Range = uint32_t(ExitValue-InitValue);
1679 if (NewPred == CmpInst::ICMP_SLE) {
1680 // Normalize SLE -> SLT, check for infinite loop.
1681 if (++Range == 0) return; // Range overflows.
1684 unsigned Leftover = Range % uint32_t(IncValue);
1686 // If this is an equality comparison, we require that the strided value
1687 // exactly land on the exit value, otherwise the IV condition will wrap
1688 // around and do things the fp IV wouldn't.
1689 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
1693 // If the stride would wrap around the i32 before exiting, we can't
1694 // transform the IV.
1695 if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
1699 // If we have a negative stride, we require the init to be greater than the
1700 // exit value and an equality or greater than comparison.
1701 if (InitValue >= ExitValue ||
1702 NewPred == CmpInst::ICMP_SLT || NewPred == CmpInst::ICMP_SLE)
1705 uint32_t Range = uint32_t(InitValue-ExitValue);
1706 if (NewPred == CmpInst::ICMP_SGE) {
1707 // Normalize SGE -> SGT, check for infinite loop.
1708 if (++Range == 0) return; // Range overflows.
1711 unsigned Leftover = Range % uint32_t(-IncValue);
1713 // If this is an equality comparison, we require that the strided value
1714 // exactly land on the exit value, otherwise the IV condition will wrap
1715 // around and do things the fp IV wouldn't.
1716 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
1720 // If the stride would wrap around the i32 before exiting, we can't
1721 // transform the IV.
1722 if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
1726 const IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
1728 // Insert new integer induction variable.
1729 PHINode *NewPHI = PHINode::Create(Int32Ty, 2, PN->getName()+".int", PN);
1730 NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
1731 PN->getIncomingBlock(IncomingEdge));
1734 BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
1735 Incr->getName()+".int", Incr);
1736 NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
1738 ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
1739 ConstantInt::get(Int32Ty, ExitValue),
1740 Compare->getName());
1742 // In the following deletions, PN may become dead and may be deleted.
1743 // Use a WeakVH to observe whether this happens.
1746 // Delete the old floating point exit comparison. The branch starts using the
1748 NewCompare->takeName(Compare);
1749 Compare->replaceAllUsesWith(NewCompare);
1750 RecursivelyDeleteTriviallyDeadInstructions(Compare);
1752 // Delete the old floating point increment.
1753 Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
1754 RecursivelyDeleteTriviallyDeadInstructions(Incr);
1756 // If the FP induction variable still has uses, this is because something else
1757 // in the loop uses its value. In order to canonicalize the induction
1758 // variable, we chose to eliminate the IV and rewrite it in terms of an
1761 // We give preference to sitofp over uitofp because it is faster on most
1764 Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
1765 PN->getParent()->getFirstNonPHI());
1766 PN->replaceAllUsesWith(Conv);
1767 RecursivelyDeleteTriviallyDeadInstructions(PN);
1770 // Add a new IVUsers entry for the newly-created integer PHI.
1772 IU->AddUsersIfInteresting(NewPHI);