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(NumElimExt , "Number of IV sign/zero extends eliminated");
71 STATISTIC(NumElimRem , "Number of IV remainder operations eliminated");
72 STATISTIC(NumElimCmp , "Number of IV comparisons eliminated");
74 // DisableIVRewrite mode currently affects IVUsers, so is defined in libAnalysis
75 // and referenced here.
77 extern bool DisableIVRewrite;
81 class IndVarSimplify : public LoopPass {
87 SmallVector<WeakVH, 16> DeadInsts;
91 static char ID; // Pass identification, replacement for typeid
92 IndVarSimplify() : LoopPass(ID), IU(0), LI(0), SE(0), DT(0), TD(0) {
93 initializeIndVarSimplifyPass(*PassRegistry::getPassRegistry());
96 virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
98 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
99 AU.addRequired<DominatorTree>();
100 AU.addRequired<LoopInfo>();
101 AU.addRequired<ScalarEvolution>();
102 AU.addRequiredID(LoopSimplifyID);
103 AU.addRequiredID(LCSSAID);
104 AU.addRequired<IVUsers>();
105 AU.addPreserved<ScalarEvolution>();
106 AU.addPreservedID(LoopSimplifyID);
107 AU.addPreservedID(LCSSAID);
108 AU.addPreserved<IVUsers>();
109 AU.setPreservesCFG();
113 bool isValidRewrite(Value *FromVal, Value *ToVal);
115 void SimplifyIVUsers(SCEVExpander &Rewriter);
116 void EliminateIVComparison(ICmpInst *ICmp, Value *IVOperand);
117 void EliminateIVRemainder(BinaryOperator *Rem,
121 void RewriteNonIntegerIVs(Loop *L);
123 ICmpInst *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
125 SCEVExpander &Rewriter);
127 void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter);
129 void RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter);
131 void SinkUnusedInvariants(Loop *L);
133 void HandleFloatingPointIV(Loop *L, PHINode *PH);
137 char IndVarSimplify::ID = 0;
138 INITIALIZE_PASS_BEGIN(IndVarSimplify, "indvars",
139 "Induction Variable Simplification", false, false)
140 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
141 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
142 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
143 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
144 INITIALIZE_PASS_DEPENDENCY(LCSSA)
145 INITIALIZE_PASS_DEPENDENCY(IVUsers)
146 INITIALIZE_PASS_END(IndVarSimplify, "indvars",
147 "Induction Variable Simplification", false, false)
149 Pass *llvm::createIndVarSimplifyPass() {
150 return new IndVarSimplify();
153 /// isValidRewrite - Return true if the SCEV expansion generated by the
154 /// rewriter can replace the original value. SCEV guarantees that it
155 /// produces the same value, but the way it is produced may be illegal IR.
156 /// Ideally, this function will only be called for verification.
157 bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) {
158 // If an SCEV expression subsumed multiple pointers, its expansion could
159 // reassociate the GEP changing the base pointer. This is illegal because the
160 // final address produced by a GEP chain must be inbounds relative to its
161 // underlying object. Otherwise basic alias analysis, among other things,
162 // could fail in a dangerous way. Ultimately, SCEV will be improved to avoid
163 // producing an expression involving multiple pointers. Until then, we must
166 // Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject
167 // because it understands lcssa phis while SCEV does not.
168 Value *FromPtr = FromVal;
169 Value *ToPtr = ToVal;
170 if (GEPOperator *GEP = dyn_cast<GEPOperator>(FromVal)) {
171 FromPtr = GEP->getPointerOperand();
173 if (GEPOperator *GEP = dyn_cast<GEPOperator>(ToVal)) {
174 ToPtr = GEP->getPointerOperand();
176 if (FromPtr != FromVal || ToPtr != ToVal) {
177 // Quickly check the common case
178 if (FromPtr == ToPtr)
181 // SCEV may have rewritten an expression that produces the GEP's pointer
182 // operand. That's ok as long as the pointer operand has the same base
183 // pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the
184 // base of a recurrence. This handles the case in which SCEV expansion
185 // converts a pointer type recurrence into a nonrecurrent pointer base
186 // indexed by an integer recurrence.
187 const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr));
188 const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr));
189 if (FromBase == ToBase)
192 DEBUG(dbgs() << "INDVARS: GEP rewrite bail out "
193 << *FromBase << " != " << *ToBase << "\n");
200 /// canExpandBackedgeTakenCount - Return true if this loop's backedge taken
201 /// count expression can be safely and cheaply expanded into an instruction
202 /// sequence that can be used by LinearFunctionTestReplace.
203 static bool canExpandBackedgeTakenCount(Loop *L, ScalarEvolution *SE) {
204 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
205 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) ||
206 BackedgeTakenCount->isZero())
209 if (!L->getExitingBlock())
212 // Can't rewrite non-branch yet.
213 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
217 // Special case: If the backedge-taken count is a UDiv, it's very likely a
218 // UDiv that ScalarEvolution produced in order to compute a precise
219 // expression, rather than a UDiv from the user's code. If we can't find a
220 // UDiv in the code with some simple searching, assume the former and forego
221 // rewriting the loop.
222 if (isa<SCEVUDivExpr>(BackedgeTakenCount)) {
223 ICmpInst *OrigCond = dyn_cast<ICmpInst>(BI->getCondition());
224 if (!OrigCond) return false;
225 const SCEV *R = SE->getSCEV(OrigCond->getOperand(1));
226 R = SE->getMinusSCEV(R, SE->getConstant(R->getType(), 1));
227 if (R != BackedgeTakenCount) {
228 const SCEV *L = SE->getSCEV(OrigCond->getOperand(0));
229 L = SE->getMinusSCEV(L, SE->getConstant(L->getType(), 1));
230 if (L != BackedgeTakenCount)
237 /// getBackedgeIVType - Get the widest type used by the loop test after peeking
240 /// TODO: Unnecessary once LinearFunctionTestReplace is removed.
241 static const Type *getBackedgeIVType(Loop *L) {
242 if (!L->getExitingBlock())
245 // Can't rewrite non-branch yet.
246 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
250 ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition());
255 for(User::op_iterator OI = Cond->op_begin(), OE = Cond->op_end();
257 assert((!Ty || Ty == (*OI)->getType()) && "bad icmp operand types");
258 TruncInst *Trunc = dyn_cast<TruncInst>(*OI);
262 return Trunc->getSrcTy();
267 /// LinearFunctionTestReplace - This method rewrites the exit condition of the
268 /// loop to be a canonical != comparison against the incremented loop induction
269 /// variable. This pass is able to rewrite the exit tests of any loop where the
270 /// SCEV analysis can determine a loop-invariant trip count of the loop, which
271 /// is actually a much broader range than just linear tests.
272 ICmpInst *IndVarSimplify::
273 LinearFunctionTestReplace(Loop *L,
274 const SCEV *BackedgeTakenCount,
276 SCEVExpander &Rewriter) {
277 assert(canExpandBackedgeTakenCount(L, SE) && "precondition");
278 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
280 // If the exiting block is not the same as the backedge block, we must compare
281 // against the preincremented value, otherwise we prefer to compare against
282 // the post-incremented value.
284 const SCEV *RHS = BackedgeTakenCount;
285 if (L->getExitingBlock() == L->getLoopLatch()) {
286 // Add one to the "backedge-taken" count to get the trip count.
287 // If this addition may overflow, we have to be more pessimistic and
288 // cast the induction variable before doing the add.
289 const SCEV *Zero = SE->getConstant(BackedgeTakenCount->getType(), 0);
291 SE->getAddExpr(BackedgeTakenCount,
292 SE->getConstant(BackedgeTakenCount->getType(), 1));
293 if ((isa<SCEVConstant>(N) && !N->isZero()) ||
294 SE->isLoopEntryGuardedByCond(L, ICmpInst::ICMP_NE, N, Zero)) {
295 // No overflow. Cast the sum.
296 RHS = SE->getTruncateOrZeroExtend(N, IndVar->getType());
298 // Potential overflow. Cast before doing the add.
299 RHS = SE->getTruncateOrZeroExtend(BackedgeTakenCount,
301 RHS = SE->getAddExpr(RHS,
302 SE->getConstant(IndVar->getType(), 1));
305 // The BackedgeTaken expression contains the number of times that the
306 // backedge branches to the loop header. This is one less than the
307 // number of times the loop executes, so use the incremented indvar.
308 CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock());
310 // We have to use the preincremented value...
311 RHS = SE->getTruncateOrZeroExtend(BackedgeTakenCount,
316 // Expand the code for the iteration count.
317 assert(SE->isLoopInvariant(RHS, L) &&
318 "Computed iteration count is not loop invariant!");
319 Value *ExitCnt = Rewriter.expandCodeFor(RHS, IndVar->getType(), BI);
321 // Insert a new icmp_ne or icmp_eq instruction before the branch.
322 ICmpInst::Predicate Opcode;
323 if (L->contains(BI->getSuccessor(0)))
324 Opcode = ICmpInst::ICMP_NE;
326 Opcode = ICmpInst::ICMP_EQ;
328 DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
329 << " LHS:" << *CmpIndVar << '\n'
331 << (Opcode == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
332 << " RHS:\t" << *RHS << "\n");
334 ICmpInst *Cond = new ICmpInst(BI, Opcode, CmpIndVar, ExitCnt, "exitcond");
336 Value *OrigCond = BI->getCondition();
337 // It's tempting to use replaceAllUsesWith here to fully replace the old
338 // comparison, but that's not immediately safe, since users of the old
339 // comparison may not be dominated by the new comparison. Instead, just
340 // update the branch to use the new comparison; in the common case this
341 // will make old comparison dead.
342 BI->setCondition(Cond);
343 DeadInsts.push_back(OrigCond);
350 /// RewriteLoopExitValues - Check to see if this loop has a computable
351 /// loop-invariant execution count. If so, this means that we can compute the
352 /// final value of any expressions that are recurrent in the loop, and
353 /// substitute the exit values from the loop into any instructions outside of
354 /// the loop that use the final values of the current expressions.
356 /// This is mostly redundant with the regular IndVarSimplify activities that
357 /// happen later, except that it's more powerful in some cases, because it's
358 /// able to brute-force evaluate arbitrary instructions as long as they have
359 /// constant operands at the beginning of the loop.
360 void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) {
361 // Verify the input to the pass in already in LCSSA form.
362 assert(L->isLCSSAForm(*DT));
364 SmallVector<BasicBlock*, 8> ExitBlocks;
365 L->getUniqueExitBlocks(ExitBlocks);
367 // Find all values that are computed inside the loop, but used outside of it.
368 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan
369 // the exit blocks of the loop to find them.
370 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
371 BasicBlock *ExitBB = ExitBlocks[i];
373 // If there are no PHI nodes in this exit block, then no values defined
374 // inside the loop are used on this path, skip it.
375 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
378 unsigned NumPreds = PN->getNumIncomingValues();
380 // Iterate over all of the PHI nodes.
381 BasicBlock::iterator BBI = ExitBB->begin();
382 while ((PN = dyn_cast<PHINode>(BBI++))) {
384 continue; // dead use, don't replace it
386 // SCEV only supports integer expressions for now.
387 if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy())
390 // It's necessary to tell ScalarEvolution about this explicitly so that
391 // it can walk the def-use list and forget all SCEVs, as it may not be
392 // watching the PHI itself. Once the new exit value is in place, there
393 // may not be a def-use connection between the loop and every instruction
394 // which got a SCEVAddRecExpr for that loop.
397 // Iterate over all of the values in all the PHI nodes.
398 for (unsigned i = 0; i != NumPreds; ++i) {
399 // If the value being merged in is not integer or is not defined
400 // in the loop, skip it.
401 Value *InVal = PN->getIncomingValue(i);
402 if (!isa<Instruction>(InVal))
405 // If this pred is for a subloop, not L itself, skip it.
406 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
407 continue; // The Block is in a subloop, skip it.
409 // Check that InVal is defined in the loop.
410 Instruction *Inst = cast<Instruction>(InVal);
411 if (!L->contains(Inst))
414 // Okay, this instruction has a user outside of the current loop
415 // and varies predictably *inside* the loop. Evaluate the value it
416 // contains when the loop exits, if possible.
417 const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
418 if (!SE->isLoopInvariant(ExitValue, L))
421 Value *ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), Inst);
423 DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
424 << " LoopVal = " << *Inst << "\n");
426 if (!isValidRewrite(Inst, ExitVal)) {
427 DeadInsts.push_back(ExitVal);
433 PN->setIncomingValue(i, ExitVal);
435 // If this instruction is dead now, delete it.
436 RecursivelyDeleteTriviallyDeadInstructions(Inst);
439 // Completely replace a single-pred PHI. This is safe, because the
440 // NewVal won't be variant in the loop, so we don't need an LCSSA phi
442 PN->replaceAllUsesWith(ExitVal);
443 RecursivelyDeleteTriviallyDeadInstructions(PN);
447 // Clone the PHI and delete the original one. This lets IVUsers and
448 // any other maps purge the original user from their records.
449 PHINode *NewPN = cast<PHINode>(PN->clone());
451 NewPN->insertBefore(PN);
452 PN->replaceAllUsesWith(NewPN);
453 PN->eraseFromParent();
458 // The insertion point instruction may have been deleted; clear it out
459 // so that the rewriter doesn't trip over it later.
460 Rewriter.clearInsertPoint();
463 void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
464 // First step. Check to see if there are any floating-point recurrences.
465 // If there are, change them into integer recurrences, permitting analysis by
466 // the SCEV routines.
468 BasicBlock *Header = L->getHeader();
470 SmallVector<WeakVH, 8> PHIs;
471 for (BasicBlock::iterator I = Header->begin();
472 PHINode *PN = dyn_cast<PHINode>(I); ++I)
475 for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
476 if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i]))
477 HandleFloatingPointIV(L, PN);
479 // If the loop previously had floating-point IV, ScalarEvolution
480 // may not have been able to compute a trip count. Now that we've done some
481 // re-writing, the trip count may be computable.
487 // Collect information about induction variables that are used by sign/zero
488 // extend operations. This information is recorded by CollectExtend and
489 // provides the input to WidenIV.
491 const Type *WidestNativeType; // Widest integer type created [sz]ext
492 bool IsSigned; // Was an sext user seen before a zext?
494 WideIVInfo() : WidestNativeType(0), IsSigned(false) {}
496 typedef std::map<PHINode *, WideIVInfo> WideIVMap;
499 /// CollectExtend - Update information about the induction variable that is
500 /// extended by this sign or zero extend operation. This is used to determine
501 /// the final width of the IV before actually widening it.
502 static void CollectExtend(CastInst *Cast, PHINode *Phi, bool IsSigned,
503 WideIVMap &IVMap, ScalarEvolution *SE,
504 const TargetData *TD) {
505 const Type *Ty = Cast->getType();
506 uint64_t Width = SE->getTypeSizeInBits(Ty);
507 if (TD && !TD->isLegalInteger(Width))
510 WideIVInfo &IVInfo = IVMap[Phi];
511 if (!IVInfo.WidestNativeType) {
512 IVInfo.WidestNativeType = SE->getEffectiveSCEVType(Ty);
513 IVInfo.IsSigned = IsSigned;
517 // We extend the IV to satisfy the sign of its first user, arbitrarily.
518 if (IVInfo.IsSigned != IsSigned)
521 if (Width > SE->getTypeSizeInBits(IVInfo.WidestNativeType))
522 IVInfo.WidestNativeType = SE->getEffectiveSCEVType(Ty);
526 /// WidenIV - The goal of this transform is to remove sign and zero extends
527 /// without creating any new induction variables. To do this, it creates a new
528 /// phi of the wider type and redirects all users, either removing extends or
529 /// inserting truncs whenever we stop propagating the type.
533 const Type *WideType;
541 SmallVectorImpl<WeakVH> &DeadInsts;
544 Instruction *WideInc;
545 const SCEV *WideIncExpr;
547 SmallPtrSet<Instruction*,16> Processed;
550 WidenIV(PHINode *PN, const WideIVInfo &IVInfo, IVUsers *IUsers,
551 LoopInfo *LInfo, ScalarEvolution *SEv, DominatorTree *DTree,
552 SmallVectorImpl<WeakVH> &DI) :
554 WideType(IVInfo.WidestNativeType),
555 IsSigned(IVInfo.IsSigned),
558 L(LI->getLoopFor(OrigPhi->getParent())),
565 assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV");
568 bool CreateWideIV(SCEVExpander &Rewriter);
571 Instruction *CloneIVUser(Instruction *NarrowUse,
572 Instruction *NarrowDef,
573 Instruction *WideDef);
575 const SCEVAddRecExpr *GetWideRecurrence(Instruction *NarrowUse);
577 Instruction *WidenIVUse(Instruction *NarrowUse,
578 Instruction *NarrowDef,
579 Instruction *WideDef);
581 } // anonymous namespace
583 /// SimplifyIVUsers - Iteratively perform simplification on IVUsers within this
584 /// loop. IVUsers is treated as a worklist. Each successive simplification may
585 /// push more users which may themselves be candidates for simplification.
587 void IndVarSimplify::SimplifyIVUsers(SCEVExpander &Rewriter) {
590 // Each round of simplification involves a round of eliminating operations
591 // followed by a round of widening IVs. A single IVUsers worklist is used
592 // across all rounds. The inner loop advances the user. If widening exposes
593 // more uses, then another pass through the outer loop is triggered.
594 for (IVUsers::iterator I = IU->begin(), E = IU->end(); I != E;) {
596 Instruction *UseInst = I->getUser();
597 Value *IVOperand = I->getOperandValToReplace();
599 if (DisableIVRewrite) {
600 if (CastInst *Cast = dyn_cast<CastInst>(UseInst)) {
601 bool IsSigned = Cast->getOpcode() == Instruction::SExt;
602 if (IsSigned || Cast->getOpcode() == Instruction::ZExt) {
603 CollectExtend(Cast, I->getPhi(), IsSigned, IVMap, SE, TD);
608 if (ICmpInst *ICmp = dyn_cast<ICmpInst>(UseInst)) {
609 EliminateIVComparison(ICmp, IVOperand);
612 if (BinaryOperator *Rem = dyn_cast<BinaryOperator>(UseInst)) {
613 bool IsSigned = Rem->getOpcode() == Instruction::SRem;
614 if (IsSigned || Rem->getOpcode() == Instruction::URem) {
615 EliminateIVRemainder(Rem, IVOperand, IsSigned, I->getPhi());
620 for (WideIVMap::const_iterator I = IVMap.begin(), E = IVMap.end();
622 WidenIV Widener(I->first, I->second, IU, LI, SE, DT, DeadInsts);
623 if (Widener.CreateWideIV(Rewriter))
629 static Value *getExtend( Value *NarrowOper, const Type *WideType,
630 bool IsSigned, IRBuilder<> &Builder) {
631 return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) :
632 Builder.CreateZExt(NarrowOper, WideType);
635 /// CloneIVUser - Instantiate a wide operation to replace a narrow
636 /// operation. This only needs to handle operations that can evaluation to
637 /// SCEVAddRec. It can safely return 0 for any operation we decide not to clone.
638 Instruction *WidenIV::CloneIVUser(Instruction *NarrowUse,
639 Instruction *NarrowDef,
640 Instruction *WideDef) {
641 unsigned Opcode = NarrowUse->getOpcode();
645 case Instruction::Add:
646 case Instruction::Mul:
647 case Instruction::UDiv:
648 case Instruction::Sub:
649 case Instruction::And:
650 case Instruction::Or:
651 case Instruction::Xor:
652 case Instruction::Shl:
653 case Instruction::LShr:
654 case Instruction::AShr:
655 DEBUG(dbgs() << "Cloning IVUser: " << *NarrowUse << "\n");
657 IRBuilder<> Builder(NarrowUse);
659 // Replace NarrowDef operands with WideDef. Otherwise, we don't know
660 // anything about the narrow operand yet so must insert a [sz]ext. It is
661 // probably loop invariant and will be folded or hoisted. If it actually
662 // comes from a widened IV, it should be removed during a future call to
664 Value *LHS = (NarrowUse->getOperand(0) == NarrowDef) ? WideDef :
665 getExtend(NarrowUse->getOperand(0), WideType, IsSigned, Builder);
666 Value *RHS = (NarrowUse->getOperand(1) == NarrowDef) ? WideDef :
667 getExtend(NarrowUse->getOperand(1), WideType, IsSigned, Builder);
669 BinaryOperator *NarrowBO = cast<BinaryOperator>(NarrowUse);
670 BinaryOperator *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(),
672 NarrowBO->getName());
673 Builder.Insert(WideBO);
674 if (NarrowBO->hasNoUnsignedWrap()) WideBO->setHasNoUnsignedWrap();
675 if (NarrowBO->hasNoSignedWrap()) WideBO->setHasNoSignedWrap();
682 // GetWideRecurrence - Is this instruction potentially interesting from IVUsers'
683 // perspective after widening it's type? In other words, can the extend be
684 // safely hoisted out of the loop with SCEV reducing the value to a recurrence
685 // on the same loop. If so, return the sign or zero extended
686 // recurrence. Otherwise return NULL.
687 const SCEVAddRecExpr *WidenIV::GetWideRecurrence(Instruction *NarrowUse) {
688 if (!SE->isSCEVable(NarrowUse->getType()))
691 const SCEV *NarrowExpr = SE->getSCEV(NarrowUse);
692 const SCEV *WideExpr = IsSigned ?
693 SE->getSignExtendExpr(NarrowExpr, WideType) :
694 SE->getZeroExtendExpr(NarrowExpr, WideType);
695 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr);
696 if (!AddRec || AddRec->getLoop() != L)
702 /// HoistStep - Attempt to hoist an IV increment above a potential use.
704 /// To successfully hoist, two criteria must be met:
705 /// - IncV operands dominate InsertPos and
706 /// - InsertPos dominates IncV
708 /// Meeting the second condition means that we don't need to check all of IncV's
709 /// existing uses (it's moving up in the domtree).
711 /// This does not yet recursively hoist the operands, although that would
712 /// not be difficult.
713 static bool HoistStep(Instruction *IncV, Instruction *InsertPos,
714 const DominatorTree *DT)
716 if (DT->dominates(IncV, InsertPos))
719 if (!DT->dominates(InsertPos->getParent(), IncV->getParent()))
722 if (IncV->mayHaveSideEffects())
725 // Attempt to hoist IncV
726 for (User::op_iterator OI = IncV->op_begin(), OE = IncV->op_end();
728 Instruction *OInst = dyn_cast<Instruction>(OI);
729 if (OInst && !DT->dominates(OInst, InsertPos))
732 IncV->moveBefore(InsertPos);
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(Instruction *NarrowUse,
739 Instruction *NarrowDef,
740 Instruction *WideDef) {
741 // To be consistent with IVUsers, stop traversing the def-use chain at
742 // inner-loop phis or post-loop phis.
743 if (isa<PHINode>(NarrowUse) && LI->getLoopFor(NarrowUse->getParent()) != L)
746 // Handle data flow merges and bizarre phi cycles.
747 if (!Processed.insert(NarrowUse))
750 // Our raison d'etre! Eliminate sign and zero extension.
751 if (IsSigned ? isa<SExtInst>(NarrowUse) : isa<ZExtInst>(NarrowUse)) {
752 Value *NewDef = WideDef;
753 if (NarrowUse->getType() != WideType) {
754 unsigned CastWidth = SE->getTypeSizeInBits(NarrowUse->getType());
755 unsigned IVWidth = SE->getTypeSizeInBits(WideType);
756 if (CastWidth < IVWidth) {
757 // The cast isn't as wide as the IV, so insert a Trunc.
758 IRBuilder<> Builder(NarrowUse);
759 NewDef = Builder.CreateTrunc(WideDef, NarrowUse->getType());
762 // A wider extend was hidden behind a narrower one. This may induce
763 // another round of IV widening in which the intermediate IV becomes
764 // dead. It should be very rare.
765 DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi
766 << " not wide enough to subsume " << *NarrowUse << "\n");
767 NarrowUse->replaceUsesOfWith(NarrowDef, WideDef);
771 if (NewDef != NarrowUse) {
772 DEBUG(dbgs() << "INDVARS: eliminating " << *NarrowUse
773 << " replaced by " << *WideDef << "\n");
775 NarrowUse->replaceAllUsesWith(NewDef);
776 DeadInsts.push_back(NarrowUse);
778 // Now that the extend is gone, expose it's uses to IVUsers for potential
779 // further simplification within SimplifyIVUsers.
780 IU->AddUsersIfInteresting(WideDef, WidePhi);
782 // No further widening is needed. The deceased [sz]ext had done it for us.
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(NarrowUse);
791 Value *Trunc = Builder.CreateTrunc(WideDef, NarrowDef->getType());
792 NarrowUse->replaceUsesOfWith(NarrowDef, Trunc);
795 // Reuse the IV increment that SCEVExpander created as long as it dominates
797 Instruction *WideUse = 0;
798 if (WideAddRec == WideIncExpr && HoistStep(WideInc, NarrowUse, DT)) {
802 WideUse = CloneIVUser(NarrowUse, NarrowDef, WideDef);
806 // GetWideRecurrence ensured that the narrow expression could be extended
807 // outside the loop without overflow. This suggests that the wide use
808 // evaluates to the same expression as the extended narrow use, but doesn't
809 // absolutely guarantee it. Hence the following failsafe check. In rare cases
810 // where it fails, we simple throw away the newly created wide use.
811 if (WideAddRec != SE->getSCEV(WideUse)) {
812 DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse
813 << ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n");
814 DeadInsts.push_back(WideUse);
818 // Returning WideUse pushes it on the worklist.
822 /// CreateWideIV - Process a single induction variable. First use the
823 /// SCEVExpander to create a wide induction variable that evaluates to the same
824 /// recurrence as the original narrow IV. Then use a worklist to forward
825 /// traverse the narrow IV's def-use chain. After WidenIVUse as processed all
826 /// interesting IV users, the narrow IV will be isolated for removal by
829 /// It would be simpler to delete uses as they are processed, but we must avoid
830 /// invalidating SCEV expressions.
832 bool WidenIV::CreateWideIV(SCEVExpander &Rewriter) {
833 // Is this phi an induction variable?
834 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi));
838 // Widen the induction variable expression.
839 const SCEV *WideIVExpr = IsSigned ?
840 SE->getSignExtendExpr(AddRec, WideType) :
841 SE->getZeroExtendExpr(AddRec, WideType);
843 assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType &&
844 "Expect the new IV expression to preserve its type");
846 // Can the IV be extended outside the loop without overflow?
847 AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr);
848 if (!AddRec || AddRec->getLoop() != L)
851 // An AddRec must have loop-invariant operands. Since this AddRec it
852 // materialized by a loop header phi, the expression cannot have any post-loop
853 // operands, so they must dominate the loop header.
854 assert(SE->properlyDominates(AddRec->getStart(), L->getHeader()) &&
855 SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader())
856 && "Loop header phi recurrence inputs do not dominate the loop");
858 // The rewriter provides a value for the desired IV expression. This may
859 // either find an existing phi or materialize a new one. Either way, we
860 // expect a well-formed cyclic phi-with-increments. i.e. any operand not part
861 // of the phi-SCC dominates the loop entry.
862 Instruction *InsertPt = L->getHeader()->begin();
863 WidePhi = cast<PHINode>(Rewriter.expandCodeFor(AddRec, WideType, InsertPt));
865 // Remembering the WideIV increment generated by SCEVExpander allows
866 // WidenIVUse to reuse it when widening the narrow IV's increment. We don't
867 // employ a general reuse mechanism because the call above is the only call to
868 // SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses.
869 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
871 cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock));
872 WideIncExpr = SE->getSCEV(WideInc);
875 DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n");
878 // Traverse the def-use chain using a worklist starting at the original IV.
879 assert(Processed.empty() && "expect initial state" );
881 // Each worklist entry has a Narrow def-use link and Wide def.
882 SmallVector<std::pair<Use *, Instruction *>, 8> NarrowIVUsers;
883 for (Value::use_iterator UI = OrigPhi->use_begin(),
884 UE = OrigPhi->use_end(); UI != UE; ++UI) {
885 NarrowIVUsers.push_back(std::make_pair(&UI.getUse(), WidePhi));
887 while (!NarrowIVUsers.empty()) {
889 Instruction *WideDef;
890 tie(NarrowDefUse, WideDef) = NarrowIVUsers.pop_back_val();
892 // Process a def-use edge. This may replace the use, so don't hold a
893 // use_iterator across it.
894 Instruction *NarrowDef = cast<Instruction>(NarrowDefUse->get());
895 Instruction *NarrowUse = cast<Instruction>(NarrowDefUse->getUser());
896 Instruction *WideUse = WidenIVUse(NarrowUse, NarrowDef, WideDef);
898 // Follow all def-use edges from the previous narrow use.
900 for (Value::use_iterator UI = NarrowUse->use_begin(),
901 UE = NarrowUse->use_end(); UI != UE; ++UI) {
902 NarrowIVUsers.push_back(std::make_pair(&UI.getUse(), WideUse));
905 // WidenIVUse may have removed the def-use edge.
906 if (NarrowDef->use_empty())
907 DeadInsts.push_back(NarrowDef);
912 void IndVarSimplify::EliminateIVComparison(ICmpInst *ICmp, Value *IVOperand) {
913 unsigned IVOperIdx = 0;
914 ICmpInst::Predicate Pred = ICmp->getPredicate();
915 if (IVOperand != ICmp->getOperand(0)) {
917 assert(IVOperand == ICmp->getOperand(1) && "Can't find IVOperand");
919 Pred = ICmpInst::getSwappedPredicate(Pred);
922 // Get the SCEVs for the ICmp operands.
923 const SCEV *S = SE->getSCEV(ICmp->getOperand(IVOperIdx));
924 const SCEV *X = SE->getSCEV(ICmp->getOperand(1 - IVOperIdx));
926 // Simplify unnecessary loops away.
927 const Loop *ICmpLoop = LI->getLoopFor(ICmp->getParent());
928 S = SE->getSCEVAtScope(S, ICmpLoop);
929 X = SE->getSCEVAtScope(X, ICmpLoop);
931 // If the condition is always true or always false, replace it with
933 if (SE->isKnownPredicate(Pred, S, X))
934 ICmp->replaceAllUsesWith(ConstantInt::getTrue(ICmp->getContext()));
935 else if (SE->isKnownPredicate(ICmpInst::getInversePredicate(Pred), S, X))
936 ICmp->replaceAllUsesWith(ConstantInt::getFalse(ICmp->getContext()));
940 DEBUG(dbgs() << "INDVARS: Eliminated comparison: " << *ICmp << '\n');
943 DeadInsts.push_back(ICmp);
946 void IndVarSimplify::EliminateIVRemainder(BinaryOperator *Rem,
950 // We're only interested in the case where we know something about
952 if (IVOperand != Rem->getOperand(0))
955 // Get the SCEVs for the ICmp operands.
956 const SCEV *S = SE->getSCEV(Rem->getOperand(0));
957 const SCEV *X = SE->getSCEV(Rem->getOperand(1));
959 // Simplify unnecessary loops away.
960 const Loop *ICmpLoop = LI->getLoopFor(Rem->getParent());
961 S = SE->getSCEVAtScope(S, ICmpLoop);
962 X = SE->getSCEVAtScope(X, ICmpLoop);
964 // i % n --> i if i is in [0,n).
965 if ((!IsSigned || SE->isKnownNonNegative(S)) &&
966 SE->isKnownPredicate(IsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
968 Rem->replaceAllUsesWith(Rem->getOperand(0));
970 // (i+1) % n --> (i+1)==n?0:(i+1) if i is in [0,n).
971 const SCEV *LessOne =
972 SE->getMinusSCEV(S, SE->getConstant(S->getType(), 1));
973 if (IsSigned && !SE->isKnownNonNegative(LessOne))
976 if (!SE->isKnownPredicate(IsSigned ?
977 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
981 ICmpInst *ICmp = new ICmpInst(Rem, ICmpInst::ICMP_EQ,
982 Rem->getOperand(0), Rem->getOperand(1),
985 SelectInst::Create(ICmp,
986 ConstantInt::get(Rem->getType(), 0),
987 Rem->getOperand(0), "tmp", Rem);
988 Rem->replaceAllUsesWith(Sel);
991 // Inform IVUsers about the new users.
992 if (Instruction *I = dyn_cast<Instruction>(Rem->getOperand(0)))
993 IU->AddUsersIfInteresting(I, IVPhi);
995 DEBUG(dbgs() << "INDVARS: Simplified rem: " << *Rem << '\n');
998 DeadInsts.push_back(Rem);
1001 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
1002 // If LoopSimplify form is not available, stay out of trouble. Some notes:
1003 // - LSR currently only supports LoopSimplify-form loops. Indvars'
1004 // canonicalization can be a pessimization without LSR to "clean up"
1006 // - We depend on having a preheader; in particular,
1007 // Loop::getCanonicalInductionVariable only supports loops with preheaders,
1008 // and we're in trouble if we can't find the induction variable even when
1009 // we've manually inserted one.
1010 if (!L->isLoopSimplifyForm())
1013 IU = &getAnalysis<IVUsers>();
1014 LI = &getAnalysis<LoopInfo>();
1015 SE = &getAnalysis<ScalarEvolution>();
1016 DT = &getAnalysis<DominatorTree>();
1017 TD = getAnalysisIfAvailable<TargetData>();
1022 // If there are any floating-point recurrences, attempt to
1023 // transform them to use integer recurrences.
1024 RewriteNonIntegerIVs(L);
1026 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1028 // Create a rewriter object which we'll use to transform the code with.
1029 SCEVExpander Rewriter(*SE);
1030 if (DisableIVRewrite)
1031 Rewriter.disableCanonicalMode();
1033 // Check to see if this loop has a computable loop-invariant execution count.
1034 // If so, this means that we can compute the final value of any expressions
1035 // that are recurrent in the loop, and substitute the exit values from the
1036 // loop into any instructions outside of the loop that use the final values of
1037 // the current expressions.
1039 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1040 RewriteLoopExitValues(L, Rewriter);
1042 // Eliminate redundant IV users.
1043 SimplifyIVUsers(Rewriter);
1045 // Compute the type of the largest recurrence expression, and decide whether
1046 // a canonical induction variable should be inserted.
1047 const Type *LargestType = 0;
1048 bool NeedCannIV = false;
1049 bool ExpandBECount = canExpandBackedgeTakenCount(L, SE);
1050 if (ExpandBECount) {
1051 // If we have a known trip count and a single exit block, we'll be
1052 // rewriting the loop exit test condition below, which requires a
1053 // canonical induction variable.
1055 const Type *Ty = BackedgeTakenCount->getType();
1056 if (DisableIVRewrite) {
1057 // In this mode, SimplifyIVUsers may have already widened the IV used by
1058 // the backedge test and inserted a Trunc on the compare's operand. Get
1059 // the wider type to avoid creating a redundant narrow IV only used by the
1061 LargestType = getBackedgeIVType(L);
1064 SE->getTypeSizeInBits(Ty) >
1065 SE->getTypeSizeInBits(LargestType))
1066 LargestType = SE->getEffectiveSCEVType(Ty);
1068 if (!DisableIVRewrite) {
1069 for (IVUsers::const_iterator I = IU->begin(), E = IU->end(); I != E; ++I) {
1072 SE->getEffectiveSCEVType(I->getOperandValToReplace()->getType());
1074 SE->getTypeSizeInBits(Ty) >
1075 SE->getTypeSizeInBits(LargestType))
1080 // Now that we know the largest of the induction variable expressions
1081 // in this loop, insert a canonical induction variable of the largest size.
1082 PHINode *IndVar = 0;
1084 // Check to see if the loop already has any canonical-looking induction
1085 // variables. If any are present and wider than the planned canonical
1086 // induction variable, temporarily remove them, so that the Rewriter
1087 // doesn't attempt to reuse them.
1088 SmallVector<PHINode *, 2> OldCannIVs;
1089 while (PHINode *OldCannIV = L->getCanonicalInductionVariable()) {
1090 if (SE->getTypeSizeInBits(OldCannIV->getType()) >
1091 SE->getTypeSizeInBits(LargestType))
1092 OldCannIV->removeFromParent();
1095 OldCannIVs.push_back(OldCannIV);
1098 IndVar = Rewriter.getOrInsertCanonicalInductionVariable(L, LargestType);
1102 DEBUG(dbgs() << "INDVARS: New CanIV: " << *IndVar << '\n');
1104 // Now that the official induction variable is established, reinsert
1105 // any old canonical-looking variables after it so that the IR remains
1106 // consistent. They will be deleted as part of the dead-PHI deletion at
1107 // the end of the pass.
1108 while (!OldCannIVs.empty()) {
1109 PHINode *OldCannIV = OldCannIVs.pop_back_val();
1110 OldCannIV->insertBefore(L->getHeader()->getFirstNonPHI());
1114 // If we have a trip count expression, rewrite the loop's exit condition
1115 // using it. We can currently only handle loops with a single exit.
1116 ICmpInst *NewICmp = 0;
1117 if (ExpandBECount) {
1118 assert(canExpandBackedgeTakenCount(L, SE) &&
1119 "canonical IV disrupted BackedgeTaken expansion");
1120 assert(NeedCannIV &&
1121 "LinearFunctionTestReplace requires a canonical induction variable");
1122 NewICmp = LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar,
1125 // Rewrite IV-derived expressions.
1126 if (!DisableIVRewrite)
1127 RewriteIVExpressions(L, Rewriter);
1129 // Clear the rewriter cache, because values that are in the rewriter's cache
1130 // can be deleted in the loop below, causing the AssertingVH in the cache to
1134 // Now that we're done iterating through lists, clean up any instructions
1135 // which are now dead.
1136 while (!DeadInsts.empty())
1137 if (Instruction *Inst =
1138 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
1139 RecursivelyDeleteTriviallyDeadInstructions(Inst);
1141 // The Rewriter may not be used from this point on.
1143 // Loop-invariant instructions in the preheader that aren't used in the
1144 // loop may be sunk below the loop to reduce register pressure.
1145 SinkUnusedInvariants(L);
1147 // For completeness, inform IVUsers of the IV use in the newly-created
1148 // loop exit test instruction.
1150 IU->AddUsersIfInteresting(cast<Instruction>(NewICmp->getOperand(0)),
1153 // Clean up dead instructions.
1154 Changed |= DeleteDeadPHIs(L->getHeader());
1155 // Check a post-condition.
1156 assert(L->isLCSSAForm(*DT) && "Indvars did not leave the loop in lcssa form!");
1160 // FIXME: It is an extremely bad idea to indvar substitute anything more
1161 // complex than affine induction variables. Doing so will put expensive
1162 // polynomial evaluations inside of the loop, and the str reduction pass
1163 // currently can only reduce affine polynomials. For now just disable
1164 // indvar subst on anything more complex than an affine addrec, unless
1165 // it can be expanded to a trivial value.
1166 static bool isSafe(const SCEV *S, const Loop *L, ScalarEvolution *SE) {
1167 // Loop-invariant values are safe.
1168 if (SE->isLoopInvariant(S, L)) return true;
1170 // Affine addrecs are safe. Non-affine are not, because LSR doesn't know how
1171 // to transform them into efficient code.
1172 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
1173 return AR->isAffine();
1175 // An add is safe it all its operands are safe.
1176 if (const SCEVCommutativeExpr *Commutative = dyn_cast<SCEVCommutativeExpr>(S)) {
1177 for (SCEVCommutativeExpr::op_iterator I = Commutative->op_begin(),
1178 E = Commutative->op_end(); I != E; ++I)
1179 if (!isSafe(*I, L, SE)) return false;
1183 // A cast is safe if its operand is.
1184 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
1185 return isSafe(C->getOperand(), L, SE);
1187 // A udiv is safe if its operands are.
1188 if (const SCEVUDivExpr *UD = dyn_cast<SCEVUDivExpr>(S))
1189 return isSafe(UD->getLHS(), L, SE) &&
1190 isSafe(UD->getRHS(), L, SE);
1192 // SCEVUnknown is always safe.
1193 if (isa<SCEVUnknown>(S))
1196 // Nothing else is safe.
1200 void IndVarSimplify::RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter) {
1201 // Rewrite all induction variable expressions in terms of the canonical
1202 // induction variable.
1204 // If there were induction variables of other sizes or offsets, manually
1205 // add the offsets to the primary induction variable and cast, avoiding
1206 // the need for the code evaluation methods to insert induction variables
1207 // of different sizes.
1208 for (IVUsers::iterator UI = IU->begin(), E = IU->end(); UI != E; ++UI) {
1209 Value *Op = UI->getOperandValToReplace();
1210 const Type *UseTy = Op->getType();
1211 Instruction *User = UI->getUser();
1213 // Compute the final addrec to expand into code.
1214 const SCEV *AR = IU->getReplacementExpr(*UI);
1216 // Evaluate the expression out of the loop, if possible.
1217 if (!L->contains(UI->getUser())) {
1218 const SCEV *ExitVal = SE->getSCEVAtScope(AR, L->getParentLoop());
1219 if (SE->isLoopInvariant(ExitVal, L))
1223 // FIXME: It is an extremely bad idea to indvar substitute anything more
1224 // complex than affine induction variables. Doing so will put expensive
1225 // polynomial evaluations inside of the loop, and the str reduction pass
1226 // currently can only reduce affine polynomials. For now just disable
1227 // indvar subst on anything more complex than an affine addrec, unless
1228 // it can be expanded to a trivial value.
1229 if (!isSafe(AR, L, SE))
1232 // Determine the insertion point for this user. By default, insert
1233 // immediately before the user. The SCEVExpander class will automatically
1234 // hoist loop invariants out of the loop. For PHI nodes, there may be
1235 // multiple uses, so compute the nearest common dominator for the
1237 Instruction *InsertPt = User;
1238 if (PHINode *PHI = dyn_cast<PHINode>(InsertPt))
1239 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
1240 if (PHI->getIncomingValue(i) == Op) {
1241 if (InsertPt == User)
1242 InsertPt = PHI->getIncomingBlock(i)->getTerminator();
1245 DT->findNearestCommonDominator(InsertPt->getParent(),
1246 PHI->getIncomingBlock(i))
1250 // Now expand it into actual Instructions and patch it into place.
1251 Value *NewVal = Rewriter.expandCodeFor(AR, UseTy, InsertPt);
1253 DEBUG(dbgs() << "INDVARS: Rewrote IV '" << *AR << "' " << *Op << '\n'
1254 << " into = " << *NewVal << "\n");
1256 if (!isValidRewrite(Op, NewVal)) {
1257 DeadInsts.push_back(NewVal);
1260 // Inform ScalarEvolution that this value is changing. The change doesn't
1261 // affect its value, but it does potentially affect which use lists the
1262 // value will be on after the replacement, which affects ScalarEvolution's
1263 // ability to walk use lists and drop dangling pointers when a value is
1265 SE->forgetValue(User);
1267 // Patch the new value into place.
1269 NewVal->takeName(Op);
1270 User->replaceUsesOfWith(Op, NewVal);
1271 UI->setOperandValToReplace(NewVal);
1276 // The old value may be dead now.
1277 DeadInsts.push_back(Op);
1281 /// If there's a single exit block, sink any loop-invariant values that
1282 /// were defined in the preheader but not used inside the loop into the
1283 /// exit block to reduce register pressure in the loop.
1284 void IndVarSimplify::SinkUnusedInvariants(Loop *L) {
1285 BasicBlock *ExitBlock = L->getExitBlock();
1286 if (!ExitBlock) return;
1288 BasicBlock *Preheader = L->getLoopPreheader();
1289 if (!Preheader) return;
1291 Instruction *InsertPt = ExitBlock->getFirstNonPHI();
1292 BasicBlock::iterator I = Preheader->getTerminator();
1293 while (I != Preheader->begin()) {
1295 // New instructions were inserted at the end of the preheader.
1296 if (isa<PHINode>(I))
1299 // Don't move instructions which might have side effects, since the side
1300 // effects need to complete before instructions inside the loop. Also don't
1301 // move instructions which might read memory, since the loop may modify
1302 // memory. Note that it's okay if the instruction might have undefined
1303 // behavior: LoopSimplify guarantees that the preheader dominates the exit
1305 if (I->mayHaveSideEffects() || I->mayReadFromMemory())
1308 // Skip debug info intrinsics.
1309 if (isa<DbgInfoIntrinsic>(I))
1312 // Don't sink static AllocaInsts out of the entry block, which would
1313 // turn them into dynamic allocas!
1314 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
1315 if (AI->isStaticAlloca())
1318 // Determine if there is a use in or before the loop (direct or
1320 bool UsedInLoop = false;
1321 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
1324 BasicBlock *UseBB = cast<Instruction>(U)->getParent();
1325 if (PHINode *P = dyn_cast<PHINode>(U)) {
1327 PHINode::getIncomingValueNumForOperand(UI.getOperandNo());
1328 UseBB = P->getIncomingBlock(i);
1330 if (UseBB == Preheader || L->contains(UseBB)) {
1336 // If there is, the def must remain in the preheader.
1340 // Otherwise, sink it to the exit block.
1341 Instruction *ToMove = I;
1344 if (I != Preheader->begin()) {
1345 // Skip debug info intrinsics.
1348 } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin());
1350 if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin())
1356 ToMove->moveBefore(InsertPt);
1362 /// ConvertToSInt - Convert APF to an integer, if possible.
1363 static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) {
1364 bool isExact = false;
1365 if (&APF.getSemantics() == &APFloat::PPCDoubleDouble)
1367 // See if we can convert this to an int64_t
1369 if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero,
1370 &isExact) != APFloat::opOK || !isExact)
1376 /// HandleFloatingPointIV - If the loop has floating induction variable
1377 /// then insert corresponding integer induction variable if possible.
1379 /// for(double i = 0; i < 10000; ++i)
1381 /// is converted into
1382 /// for(int i = 0; i < 10000; ++i)
1385 void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) {
1386 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
1387 unsigned BackEdge = IncomingEdge^1;
1389 // Check incoming value.
1390 ConstantFP *InitValueVal =
1391 dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));
1394 if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
1397 // Check IV increment. Reject this PN if increment operation is not
1398 // an add or increment value can not be represented by an integer.
1399 BinaryOperator *Incr =
1400 dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
1401 if (Incr == 0 || Incr->getOpcode() != Instruction::FAdd) return;
1403 // If this is not an add of the PHI with a constantfp, or if the constant fp
1404 // is not an integer, bail out.
1405 ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
1407 if (IncValueVal == 0 || Incr->getOperand(0) != PN ||
1408 !ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
1411 // Check Incr uses. One user is PN and the other user is an exit condition
1412 // used by the conditional terminator.
1413 Value::use_iterator IncrUse = Incr->use_begin();
1414 Instruction *U1 = cast<Instruction>(*IncrUse++);
1415 if (IncrUse == Incr->use_end()) return;
1416 Instruction *U2 = cast<Instruction>(*IncrUse++);
1417 if (IncrUse != Incr->use_end()) return;
1419 // Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't
1420 // only used by a branch, we can't transform it.
1421 FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
1423 Compare = dyn_cast<FCmpInst>(U2);
1424 if (Compare == 0 || !Compare->hasOneUse() ||
1425 !isa<BranchInst>(Compare->use_back()))
1428 BranchInst *TheBr = cast<BranchInst>(Compare->use_back());
1430 // We need to verify that the branch actually controls the iteration count
1431 // of the loop. If not, the new IV can overflow and no one will notice.
1432 // The branch block must be in the loop and one of the successors must be out
1434 assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
1435 if (!L->contains(TheBr->getParent()) ||
1436 (L->contains(TheBr->getSuccessor(0)) &&
1437 L->contains(TheBr->getSuccessor(1))))
1441 // If it isn't a comparison with an integer-as-fp (the exit value), we can't
1443 ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
1445 if (ExitValueVal == 0 ||
1446 !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
1449 // Find new predicate for integer comparison.
1450 CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
1451 switch (Compare->getPredicate()) {
1452 default: return; // Unknown comparison.
1453 case CmpInst::FCMP_OEQ:
1454 case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
1455 case CmpInst::FCMP_ONE:
1456 case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
1457 case CmpInst::FCMP_OGT:
1458 case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
1459 case CmpInst::FCMP_OGE:
1460 case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
1461 case CmpInst::FCMP_OLT:
1462 case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
1463 case CmpInst::FCMP_OLE:
1464 case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
1467 // We convert the floating point induction variable to a signed i32 value if
1468 // we can. This is only safe if the comparison will not overflow in a way
1469 // that won't be trapped by the integer equivalent operations. Check for this
1471 // TODO: We could use i64 if it is native and the range requires it.
1473 // The start/stride/exit values must all fit in signed i32.
1474 if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
1477 // If not actually striding (add x, 0.0), avoid touching the code.
1481 // Positive and negative strides have different safety conditions.
1483 // If we have a positive stride, we require the init to be less than the
1484 // exit value and an equality or less than comparison.
1485 if (InitValue >= ExitValue ||
1486 NewPred == CmpInst::ICMP_SGT || NewPred == CmpInst::ICMP_SGE)
1489 uint32_t Range = uint32_t(ExitValue-InitValue);
1490 if (NewPred == CmpInst::ICMP_SLE) {
1491 // Normalize SLE -> SLT, check for infinite loop.
1492 if (++Range == 0) return; // Range overflows.
1495 unsigned Leftover = Range % uint32_t(IncValue);
1497 // If this is an equality comparison, we require that the strided value
1498 // exactly land on the exit value, otherwise the IV condition will wrap
1499 // around and do things the fp IV wouldn't.
1500 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
1504 // If the stride would wrap around the i32 before exiting, we can't
1505 // transform the IV.
1506 if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
1510 // If we have a negative stride, we require the init to be greater than the
1511 // exit value and an equality or greater than comparison.
1512 if (InitValue >= ExitValue ||
1513 NewPred == CmpInst::ICMP_SLT || NewPred == CmpInst::ICMP_SLE)
1516 uint32_t Range = uint32_t(InitValue-ExitValue);
1517 if (NewPred == CmpInst::ICMP_SGE) {
1518 // Normalize SGE -> SGT, check for infinite loop.
1519 if (++Range == 0) return; // Range overflows.
1522 unsigned Leftover = Range % uint32_t(-IncValue);
1524 // If this is an equality comparison, we require that the strided value
1525 // exactly land on the exit value, otherwise the IV condition will wrap
1526 // around and do things the fp IV wouldn't.
1527 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
1531 // If the stride would wrap around the i32 before exiting, we can't
1532 // transform the IV.
1533 if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
1537 const IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
1539 // Insert new integer induction variable.
1540 PHINode *NewPHI = PHINode::Create(Int32Ty, 2, PN->getName()+".int", PN);
1541 NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
1542 PN->getIncomingBlock(IncomingEdge));
1545 BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
1546 Incr->getName()+".int", Incr);
1547 NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
1549 ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
1550 ConstantInt::get(Int32Ty, ExitValue),
1551 Compare->getName());
1553 // In the following deletions, PN may become dead and may be deleted.
1554 // Use a WeakVH to observe whether this happens.
1557 // Delete the old floating point exit comparison. The branch starts using the
1559 NewCompare->takeName(Compare);
1560 Compare->replaceAllUsesWith(NewCompare);
1561 RecursivelyDeleteTriviallyDeadInstructions(Compare);
1563 // Delete the old floating point increment.
1564 Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
1565 RecursivelyDeleteTriviallyDeadInstructions(Incr);
1567 // If the FP induction variable still has uses, this is because something else
1568 // in the loop uses its value. In order to canonicalize the induction
1569 // variable, we chose to eliminate the IV and rewrite it in terms of an
1572 // We give preference to sitofp over uitofp because it is faster on most
1575 Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
1576 PN->getParent()->getFirstNonPHI());
1577 PN->replaceAllUsesWith(Conv);
1578 RecursivelyDeleteTriviallyDeadInstructions(PN);
1581 // Add a new IVUsers entry for the newly-created integer PHI.
1582 IU->AddUsersIfInteresting(NewPHI, NewPHI);