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 // If the trip count of a loop is computable, this pass also makes the following
16 // 1. The exit condition for the loop is canonicalized to compare the
17 // induction value against the exit value. This turns loops like:
18 // 'for (i = 7; i*i < 1000; ++i)' into 'for (i = 0; i != 25; ++i)'
19 // 2. Any use outside of the loop of an expression derived from the indvar
20 // is changed to compute the derived value outside of the loop, eliminating
21 // the dependence on the exit value of the induction variable. If the only
22 // purpose of the loop is to compute the exit value of some derived
23 // expression, this transformation will make the loop dead.
25 //===----------------------------------------------------------------------===//
27 #define DEBUG_TYPE "indvars"
28 #include "llvm/Transforms/Scalar.h"
29 #include "llvm/ADT/DenseMap.h"
30 #include "llvm/ADT/SmallVector.h"
31 #include "llvm/ADT/Statistic.h"
32 #include "llvm/Analysis/Dominators.h"
33 #include "llvm/Analysis/LoopInfo.h"
34 #include "llvm/Analysis/LoopPass.h"
35 #include "llvm/Analysis/ScalarEvolutionExpander.h"
36 #include "llvm/IR/BasicBlock.h"
37 #include "llvm/IR/Constants.h"
38 #include "llvm/IR/DataLayout.h"
39 #include "llvm/IR/Instructions.h"
40 #include "llvm/IR/IntrinsicInst.h"
41 #include "llvm/IR/LLVMContext.h"
42 #include "llvm/IR/Type.h"
43 #include "llvm/Support/CFG.h"
44 #include "llvm/Support/CommandLine.h"
45 #include "llvm/Support/Debug.h"
46 #include "llvm/Support/raw_ostream.h"
47 #include "llvm/Target/TargetLibraryInfo.h"
48 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
49 #include "llvm/Transforms/Utils/Local.h"
50 #include "llvm/Transforms/Utils/SimplifyIndVar.h"
53 STATISTIC(NumWidened , "Number of indvars widened");
54 STATISTIC(NumReplaced , "Number of exit values replaced");
55 STATISTIC(NumLFTR , "Number of loop exit tests replaced");
56 STATISTIC(NumElimExt , "Number of IV sign/zero extends eliminated");
57 STATISTIC(NumElimIV , "Number of congruent IVs eliminated");
59 // Trip count verification can be enabled by default under NDEBUG if we
60 // implement a strong expression equivalence checker in SCEV. Until then, we
61 // use the verify-indvars flag, which may assert in some cases.
62 static cl::opt<bool> VerifyIndvars(
63 "verify-indvars", cl::Hidden,
64 cl::desc("Verify the ScalarEvolution result after running indvars"));
67 class IndVarSimplify : public LoopPass {
72 TargetLibraryInfo *TLI;
74 SmallVector<WeakVH, 16> DeadInsts;
78 static char ID; // Pass identification, replacement for typeid
79 IndVarSimplify() : LoopPass(ID), LI(0), SE(0), DT(0), TD(0),
81 initializeIndVarSimplifyPass(*PassRegistry::getPassRegistry());
84 virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
86 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
87 AU.addRequired<DominatorTree>();
88 AU.addRequired<LoopInfo>();
89 AU.addRequired<ScalarEvolution>();
90 AU.addRequiredID(LoopSimplifyID);
91 AU.addRequiredID(LCSSAID);
92 AU.addPreserved<ScalarEvolution>();
93 AU.addPreservedID(LoopSimplifyID);
94 AU.addPreservedID(LCSSAID);
99 virtual void releaseMemory() {
103 bool isValidRewrite(Value *FromVal, Value *ToVal);
105 void HandleFloatingPointIV(Loop *L, PHINode *PH);
106 void RewriteNonIntegerIVs(Loop *L);
108 void SimplifyAndExtend(Loop *L, SCEVExpander &Rewriter, LPPassManager &LPM);
110 void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter);
112 Value *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
113 PHINode *IndVar, SCEVExpander &Rewriter);
115 void SinkUnusedInvariants(Loop *L);
119 char IndVarSimplify::ID = 0;
120 INITIALIZE_PASS_BEGIN(IndVarSimplify, "indvars",
121 "Induction Variable Simplification", false, false)
122 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
123 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
124 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
125 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
126 INITIALIZE_PASS_DEPENDENCY(LCSSA)
127 INITIALIZE_PASS_END(IndVarSimplify, "indvars",
128 "Induction Variable Simplification", false, false)
130 Pass *llvm::createIndVarSimplifyPass() {
131 return new IndVarSimplify();
134 /// isValidRewrite - Return true if the SCEV expansion generated by the
135 /// rewriter can replace the original value. SCEV guarantees that it
136 /// produces the same value, but the way it is produced may be illegal IR.
137 /// Ideally, this function will only be called for verification.
138 bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) {
139 // If an SCEV expression subsumed multiple pointers, its expansion could
140 // reassociate the GEP changing the base pointer. This is illegal because the
141 // final address produced by a GEP chain must be inbounds relative to its
142 // underlying object. Otherwise basic alias analysis, among other things,
143 // could fail in a dangerous way. Ultimately, SCEV will be improved to avoid
144 // producing an expression involving multiple pointers. Until then, we must
147 // Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject
148 // because it understands lcssa phis while SCEV does not.
149 Value *FromPtr = FromVal;
150 Value *ToPtr = ToVal;
151 if (GEPOperator *GEP = dyn_cast<GEPOperator>(FromVal)) {
152 FromPtr = GEP->getPointerOperand();
154 if (GEPOperator *GEP = dyn_cast<GEPOperator>(ToVal)) {
155 ToPtr = GEP->getPointerOperand();
157 if (FromPtr != FromVal || ToPtr != ToVal) {
158 // Quickly check the common case
159 if (FromPtr == ToPtr)
162 // SCEV may have rewritten an expression that produces the GEP's pointer
163 // operand. That's ok as long as the pointer operand has the same base
164 // pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the
165 // base of a recurrence. This handles the case in which SCEV expansion
166 // converts a pointer type recurrence into a nonrecurrent pointer base
167 // indexed by an integer recurrence.
169 // If the GEP base pointer is a vector of pointers, abort.
170 if (!FromPtr->getType()->isPointerTy() || !ToPtr->getType()->isPointerTy())
173 const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr));
174 const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr));
175 if (FromBase == ToBase)
178 DEBUG(dbgs() << "INDVARS: GEP rewrite bail out "
179 << *FromBase << " != " << *ToBase << "\n");
186 /// Determine the insertion point for this user. By default, insert immediately
187 /// before the user. SCEVExpander or LICM will hoist loop invariants out of the
188 /// loop. For PHI nodes, there may be multiple uses, so compute the nearest
189 /// common dominator for the incoming blocks.
190 static Instruction *getInsertPointForUses(Instruction *User, Value *Def,
192 PHINode *PHI = dyn_cast<PHINode>(User);
196 Instruction *InsertPt = 0;
197 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) {
198 if (PHI->getIncomingValue(i) != Def)
201 BasicBlock *InsertBB = PHI->getIncomingBlock(i);
203 InsertPt = InsertBB->getTerminator();
206 InsertBB = DT->findNearestCommonDominator(InsertPt->getParent(), InsertBB);
207 InsertPt = InsertBB->getTerminator();
209 assert(InsertPt && "Missing phi operand");
210 assert((!isa<Instruction>(Def) ||
211 DT->dominates(cast<Instruction>(Def), InsertPt)) &&
212 "def does not dominate all uses");
216 //===----------------------------------------------------------------------===//
217 // RewriteNonIntegerIVs and helpers. Prefer integer IVs.
218 //===----------------------------------------------------------------------===//
220 /// ConvertToSInt - Convert APF to an integer, if possible.
221 static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) {
222 bool isExact = false;
223 // See if we can convert this to an int64_t
225 if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero,
226 &isExact) != APFloat::opOK || !isExact)
232 /// HandleFloatingPointIV - If the loop has floating induction variable
233 /// then insert corresponding integer induction variable if possible.
235 /// for(double i = 0; i < 10000; ++i)
237 /// is converted into
238 /// for(int i = 0; i < 10000; ++i)
241 void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) {
242 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
243 unsigned BackEdge = IncomingEdge^1;
245 // Check incoming value.
246 ConstantFP *InitValueVal =
247 dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));
250 if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
253 // Check IV increment. Reject this PN if increment operation is not
254 // an add or increment value can not be represented by an integer.
255 BinaryOperator *Incr =
256 dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
257 if (Incr == 0 || Incr->getOpcode() != Instruction::FAdd) return;
259 // If this is not an add of the PHI with a constantfp, or if the constant fp
260 // is not an integer, bail out.
261 ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
263 if (IncValueVal == 0 || Incr->getOperand(0) != PN ||
264 !ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
267 // Check Incr uses. One user is PN and the other user is an exit condition
268 // used by the conditional terminator.
269 Value::use_iterator IncrUse = Incr->use_begin();
270 Instruction *U1 = cast<Instruction>(*IncrUse++);
271 if (IncrUse == Incr->use_end()) return;
272 Instruction *U2 = cast<Instruction>(*IncrUse++);
273 if (IncrUse != Incr->use_end()) return;
275 // Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't
276 // only used by a branch, we can't transform it.
277 FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
279 Compare = dyn_cast<FCmpInst>(U2);
280 if (Compare == 0 || !Compare->hasOneUse() ||
281 !isa<BranchInst>(Compare->use_back()))
284 BranchInst *TheBr = cast<BranchInst>(Compare->use_back());
286 // We need to verify that the branch actually controls the iteration count
287 // of the loop. If not, the new IV can overflow and no one will notice.
288 // The branch block must be in the loop and one of the successors must be out
290 assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
291 if (!L->contains(TheBr->getParent()) ||
292 (L->contains(TheBr->getSuccessor(0)) &&
293 L->contains(TheBr->getSuccessor(1))))
297 // If it isn't a comparison with an integer-as-fp (the exit value), we can't
299 ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
301 if (ExitValueVal == 0 ||
302 !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
305 // Find new predicate for integer comparison.
306 CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
307 switch (Compare->getPredicate()) {
308 default: return; // Unknown comparison.
309 case CmpInst::FCMP_OEQ:
310 case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
311 case CmpInst::FCMP_ONE:
312 case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
313 case CmpInst::FCMP_OGT:
314 case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
315 case CmpInst::FCMP_OGE:
316 case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
317 case CmpInst::FCMP_OLT:
318 case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
319 case CmpInst::FCMP_OLE:
320 case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
323 // We convert the floating point induction variable to a signed i32 value if
324 // we can. This is only safe if the comparison will not overflow in a way
325 // that won't be trapped by the integer equivalent operations. Check for this
327 // TODO: We could use i64 if it is native and the range requires it.
329 // The start/stride/exit values must all fit in signed i32.
330 if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
333 // If not actually striding (add x, 0.0), avoid touching the code.
337 // Positive and negative strides have different safety conditions.
339 // If we have a positive stride, we require the init to be less than the
341 if (InitValue >= ExitValue)
344 uint32_t Range = uint32_t(ExitValue-InitValue);
345 // Check for infinite loop, either:
346 // while (i <= Exit) or until (i > Exit)
347 if (NewPred == CmpInst::ICMP_SLE || NewPred == CmpInst::ICMP_SGT) {
348 if (++Range == 0) return; // Range overflows.
351 unsigned Leftover = Range % uint32_t(IncValue);
353 // If this is an equality comparison, we require that the strided value
354 // exactly land on the exit value, otherwise the IV condition will wrap
355 // around and do things the fp IV wouldn't.
356 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
360 // If the stride would wrap around the i32 before exiting, we can't
362 if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
366 // If we have a negative stride, we require the init to be greater than the
368 if (InitValue <= ExitValue)
371 uint32_t Range = uint32_t(InitValue-ExitValue);
372 // Check for infinite loop, either:
373 // while (i >= Exit) or until (i < Exit)
374 if (NewPred == CmpInst::ICMP_SGE || NewPred == CmpInst::ICMP_SLT) {
375 if (++Range == 0) return; // Range overflows.
378 unsigned Leftover = Range % uint32_t(-IncValue);
380 // If this is an equality comparison, we require that the strided value
381 // exactly land on the exit value, otherwise the IV condition will wrap
382 // around and do things the fp IV wouldn't.
383 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
387 // If the stride would wrap around the i32 before exiting, we can't
389 if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
393 IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
395 // Insert new integer induction variable.
396 PHINode *NewPHI = PHINode::Create(Int32Ty, 2, PN->getName()+".int", PN);
397 NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
398 PN->getIncomingBlock(IncomingEdge));
401 BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
402 Incr->getName()+".int", Incr);
403 NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
405 ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
406 ConstantInt::get(Int32Ty, ExitValue),
409 // In the following deletions, PN may become dead and may be deleted.
410 // Use a WeakVH to observe whether this happens.
413 // Delete the old floating point exit comparison. The branch starts using the
415 NewCompare->takeName(Compare);
416 Compare->replaceAllUsesWith(NewCompare);
417 RecursivelyDeleteTriviallyDeadInstructions(Compare, TLI);
419 // Delete the old floating point increment.
420 Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
421 RecursivelyDeleteTriviallyDeadInstructions(Incr, TLI);
423 // If the FP induction variable still has uses, this is because something else
424 // in the loop uses its value. In order to canonicalize the induction
425 // variable, we chose to eliminate the IV and rewrite it in terms of an
428 // We give preference to sitofp over uitofp because it is faster on most
431 Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
432 PN->getParent()->getFirstInsertionPt());
433 PN->replaceAllUsesWith(Conv);
434 RecursivelyDeleteTriviallyDeadInstructions(PN, TLI);
439 void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
440 // First step. Check to see if there are any floating-point recurrences.
441 // If there are, change them into integer recurrences, permitting analysis by
442 // the SCEV routines.
444 BasicBlock *Header = L->getHeader();
446 SmallVector<WeakVH, 8> PHIs;
447 for (BasicBlock::iterator I = Header->begin();
448 PHINode *PN = dyn_cast<PHINode>(I); ++I)
451 for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
452 if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i]))
453 HandleFloatingPointIV(L, PN);
455 // If the loop previously had floating-point IV, ScalarEvolution
456 // may not have been able to compute a trip count. Now that we've done some
457 // re-writing, the trip count may be computable.
462 //===----------------------------------------------------------------------===//
463 // RewriteLoopExitValues - Optimize IV users outside the loop.
464 // As a side effect, reduces the amount of IV processing within the loop.
465 //===----------------------------------------------------------------------===//
467 /// RewriteLoopExitValues - Check to see if this loop has a computable
468 /// loop-invariant execution count. If so, this means that we can compute the
469 /// final value of any expressions that are recurrent in the loop, and
470 /// substitute the exit values from the loop into any instructions outside of
471 /// the loop that use the final values of the current expressions.
473 /// This is mostly redundant with the regular IndVarSimplify activities that
474 /// happen later, except that it's more powerful in some cases, because it's
475 /// able to brute-force evaluate arbitrary instructions as long as they have
476 /// constant operands at the beginning of the loop.
477 void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) {
478 // Verify the input to the pass in already in LCSSA form.
479 assert(L->isLCSSAForm(*DT));
481 SmallVector<BasicBlock*, 8> ExitBlocks;
482 L->getUniqueExitBlocks(ExitBlocks);
484 // Find all values that are computed inside the loop, but used outside of it.
485 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan
486 // the exit blocks of the loop to find them.
487 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
488 BasicBlock *ExitBB = ExitBlocks[i];
490 // If there are no PHI nodes in this exit block, then no values defined
491 // inside the loop are used on this path, skip it.
492 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
495 unsigned NumPreds = PN->getNumIncomingValues();
497 // Iterate over all of the PHI nodes.
498 BasicBlock::iterator BBI = ExitBB->begin();
499 while ((PN = dyn_cast<PHINode>(BBI++))) {
501 continue; // dead use, don't replace it
503 // SCEV only supports integer expressions for now.
504 if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy())
507 // It's necessary to tell ScalarEvolution about this explicitly so that
508 // it can walk the def-use list and forget all SCEVs, as it may not be
509 // watching the PHI itself. Once the new exit value is in place, there
510 // may not be a def-use connection between the loop and every instruction
511 // which got a SCEVAddRecExpr for that loop.
514 // Iterate over all of the values in all the PHI nodes.
515 for (unsigned i = 0; i != NumPreds; ++i) {
516 // If the value being merged in is not integer or is not defined
517 // in the loop, skip it.
518 Value *InVal = PN->getIncomingValue(i);
519 if (!isa<Instruction>(InVal))
522 // If this pred is for a subloop, not L itself, skip it.
523 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
524 continue; // The Block is in a subloop, skip it.
526 // Check that InVal is defined in the loop.
527 Instruction *Inst = cast<Instruction>(InVal);
528 if (!L->contains(Inst))
531 // Okay, this instruction has a user outside of the current loop
532 // and varies predictably *inside* the loop. Evaluate the value it
533 // contains when the loop exits, if possible.
534 const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
535 if (!SE->isLoopInvariant(ExitValue, L) || !isSafeToExpand(ExitValue))
538 // Computing the value outside of the loop brings no benefit if :
539 // - it is definitely used inside the loop in a way which can not be
541 // - no use outside of the loop can take advantage of hoisting the
542 // computation out of the loop
543 if (ExitValue->getSCEVType()>=scMulExpr) {
544 unsigned NumHardInternalUses = 0;
545 unsigned NumSoftExternalUses = 0;
546 unsigned NumUses = 0;
547 for (Value::use_iterator IB=Inst->use_begin(), IE=Inst->use_end();
548 IB!=IE && NumUses<=6 ; ++IB) {
549 Instruction *UseInstr = cast<Instruction>(*IB);
550 unsigned Opc = UseInstr->getOpcode();
552 if (L->contains(UseInstr)) {
553 if (Opc == Instruction::Call || Opc == Instruction::Ret)
554 NumHardInternalUses++;
556 if (Opc == Instruction::PHI) {
557 // Do not count the Phi as a use. LCSSA may have inserted
558 // plenty of trivial ones.
560 for (Value::use_iterator PB=UseInstr->use_begin(),
561 PE=UseInstr->use_end();
562 PB!=PE && NumUses<=6 ; ++PB, ++NumUses) {
563 unsigned PhiOpc = cast<Instruction>(*PB)->getOpcode();
564 if (PhiOpc != Instruction::Call && PhiOpc != Instruction::Ret)
565 NumSoftExternalUses++;
569 if (Opc != Instruction::Call && Opc != Instruction::Ret)
570 NumSoftExternalUses++;
573 if (NumUses <= 6 && NumHardInternalUses && !NumSoftExternalUses)
577 Value *ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), Inst);
579 DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
580 << " LoopVal = " << *Inst << "\n");
582 if (!isValidRewrite(Inst, ExitVal)) {
583 DeadInsts.push_back(ExitVal);
589 PN->setIncomingValue(i, ExitVal);
591 // If this instruction is dead now, delete it. Don't do it now to avoid
592 // invalidating iterators.
593 if (isInstructionTriviallyDead(Inst, TLI))
594 DeadInsts.push_back(Inst);
597 // Completely replace a single-pred PHI. This is safe, because the
598 // NewVal won't be variant in the loop, so we don't need an LCSSA phi
600 PN->replaceAllUsesWith(ExitVal);
601 PN->eraseFromParent();
605 // Clone the PHI and delete the original one. This lets IVUsers and
606 // any other maps purge the original user from their records.
607 PHINode *NewPN = cast<PHINode>(PN->clone());
609 NewPN->insertBefore(PN);
610 PN->replaceAllUsesWith(NewPN);
611 PN->eraseFromParent();
616 // The insertion point instruction may have been deleted; clear it out
617 // so that the rewriter doesn't trip over it later.
618 Rewriter.clearInsertPoint();
621 //===----------------------------------------------------------------------===//
622 // IV Widening - Extend the width of an IV to cover its widest uses.
623 //===----------------------------------------------------------------------===//
626 // Collect information about induction variables that are used by sign/zero
627 // extend operations. This information is recorded by CollectExtend and
628 // provides the input to WidenIV.
631 Type *WidestNativeType; // Widest integer type created [sz]ext
632 bool IsSigned; // Was an sext user seen before a zext?
634 WideIVInfo() : NarrowIV(0), WidestNativeType(0), IsSigned(false) {}
637 class WideIVVisitor : public IVVisitor {
639 const DataLayout *TD;
644 WideIVVisitor(PHINode *NarrowIV, ScalarEvolution *SCEV,
645 const DataLayout *TData) :
646 SE(SCEV), TD(TData) { WI.NarrowIV = NarrowIV; }
648 // Implement the interface used by simplifyUsersOfIV.
649 virtual void visitCast(CastInst *Cast);
653 /// visitCast - Update information about the induction variable that is
654 /// extended by this sign or zero extend operation. This is used to determine
655 /// the final width of the IV before actually widening it.
656 void WideIVVisitor::visitCast(CastInst *Cast) {
657 bool IsSigned = Cast->getOpcode() == Instruction::SExt;
658 if (!IsSigned && Cast->getOpcode() != Instruction::ZExt)
661 Type *Ty = Cast->getType();
662 uint64_t Width = SE->getTypeSizeInBits(Ty);
663 if (TD && !TD->isLegalInteger(Width))
666 if (!WI.WidestNativeType) {
667 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
668 WI.IsSigned = IsSigned;
672 // We extend the IV to satisfy the sign of its first user, arbitrarily.
673 if (WI.IsSigned != IsSigned)
676 if (Width > SE->getTypeSizeInBits(WI.WidestNativeType))
677 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
682 /// NarrowIVDefUse - Record a link in the Narrow IV def-use chain along with the
683 /// WideIV that computes the same value as the Narrow IV def. This avoids
684 /// caching Use* pointers.
685 struct NarrowIVDefUse {
686 Instruction *NarrowDef;
687 Instruction *NarrowUse;
688 Instruction *WideDef;
690 NarrowIVDefUse(): NarrowDef(0), NarrowUse(0), WideDef(0) {}
692 NarrowIVDefUse(Instruction *ND, Instruction *NU, Instruction *WD):
693 NarrowDef(ND), NarrowUse(NU), WideDef(WD) {}
696 /// WidenIV - The goal of this transform is to remove sign and zero extends
697 /// without creating any new induction variables. To do this, it creates a new
698 /// phi of the wider type and redirects all users, either removing extends or
699 /// inserting truncs whenever we stop propagating the type.
715 Instruction *WideInc;
716 const SCEV *WideIncExpr;
717 SmallVectorImpl<WeakVH> &DeadInsts;
719 SmallPtrSet<Instruction*,16> Widened;
720 SmallVector<NarrowIVDefUse, 8> NarrowIVUsers;
723 WidenIV(const WideIVInfo &WI, LoopInfo *LInfo,
724 ScalarEvolution *SEv, DominatorTree *DTree,
725 SmallVectorImpl<WeakVH> &DI) :
726 OrigPhi(WI.NarrowIV),
727 WideType(WI.WidestNativeType),
728 IsSigned(WI.IsSigned),
730 L(LI->getLoopFor(OrigPhi->getParent())),
737 assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV");
740 PHINode *CreateWideIV(SCEVExpander &Rewriter);
743 Value *getExtend(Value *NarrowOper, Type *WideType, bool IsSigned,
746 Instruction *CloneIVUser(NarrowIVDefUse DU);
748 const SCEVAddRecExpr *GetWideRecurrence(Instruction *NarrowUse);
750 const SCEVAddRecExpr* GetExtendedOperandRecurrence(NarrowIVDefUse DU);
752 Instruction *WidenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter);
754 void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef);
756 } // anonymous namespace
758 /// isLoopInvariant - Perform a quick domtree based check for loop invariance
759 /// assuming that V is used within the loop. LoopInfo::isLoopInvariant() seems
760 /// gratuitous for this purpose.
761 static bool isLoopInvariant(Value *V, const Loop *L, const DominatorTree *DT) {
762 Instruction *Inst = dyn_cast<Instruction>(V);
766 return DT->properlyDominates(Inst->getParent(), L->getHeader());
769 Value *WidenIV::getExtend(Value *NarrowOper, Type *WideType, bool IsSigned,
771 // Set the debug location and conservative insertion point.
772 IRBuilder<> Builder(Use);
773 // Hoist the insertion point into loop preheaders as far as possible.
774 for (const Loop *L = LI->getLoopFor(Use->getParent());
775 L && L->getLoopPreheader() && isLoopInvariant(NarrowOper, L, DT);
776 L = L->getParentLoop())
777 Builder.SetInsertPoint(L->getLoopPreheader()->getTerminator());
779 return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) :
780 Builder.CreateZExt(NarrowOper, WideType);
783 /// CloneIVUser - Instantiate a wide operation to replace a narrow
784 /// operation. This only needs to handle operations that can evaluation to
785 /// SCEVAddRec. It can safely return 0 for any operation we decide not to clone.
786 Instruction *WidenIV::CloneIVUser(NarrowIVDefUse DU) {
787 unsigned Opcode = DU.NarrowUse->getOpcode();
791 case Instruction::Add:
792 case Instruction::Mul:
793 case Instruction::UDiv:
794 case Instruction::Sub:
795 case Instruction::And:
796 case Instruction::Or:
797 case Instruction::Xor:
798 case Instruction::Shl:
799 case Instruction::LShr:
800 case Instruction::AShr:
801 DEBUG(dbgs() << "Cloning IVUser: " << *DU.NarrowUse << "\n");
803 // Replace NarrowDef operands with WideDef. Otherwise, we don't know
804 // anything about the narrow operand yet so must insert a [sz]ext. It is
805 // probably loop invariant and will be folded or hoisted. If it actually
806 // comes from a widened IV, it should be removed during a future call to
808 Value *LHS = (DU.NarrowUse->getOperand(0) == DU.NarrowDef) ? DU.WideDef :
809 getExtend(DU.NarrowUse->getOperand(0), WideType, IsSigned, DU.NarrowUse);
810 Value *RHS = (DU.NarrowUse->getOperand(1) == DU.NarrowDef) ? DU.WideDef :
811 getExtend(DU.NarrowUse->getOperand(1), WideType, IsSigned, DU.NarrowUse);
813 BinaryOperator *NarrowBO = cast<BinaryOperator>(DU.NarrowUse);
814 BinaryOperator *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(),
816 NarrowBO->getName());
817 IRBuilder<> Builder(DU.NarrowUse);
818 Builder.Insert(WideBO);
819 if (const OverflowingBinaryOperator *OBO =
820 dyn_cast<OverflowingBinaryOperator>(NarrowBO)) {
821 if (OBO->hasNoUnsignedWrap()) WideBO->setHasNoUnsignedWrap();
822 if (OBO->hasNoSignedWrap()) WideBO->setHasNoSignedWrap();
828 /// No-wrap operations can transfer sign extension of their result to their
829 /// operands. Generate the SCEV value for the widened operation without
830 /// actually modifying the IR yet. If the expression after extending the
831 /// operands is an AddRec for this loop, return it.
832 const SCEVAddRecExpr* WidenIV::GetExtendedOperandRecurrence(NarrowIVDefUse DU) {
833 // Handle the common case of add<nsw/nuw>
834 if (DU.NarrowUse->getOpcode() != Instruction::Add)
837 // One operand (NarrowDef) has already been extended to WideDef. Now determine
838 // if extending the other will lead to a recurrence.
839 unsigned ExtendOperIdx = DU.NarrowUse->getOperand(0) == DU.NarrowDef ? 1 : 0;
840 assert(DU.NarrowUse->getOperand(1-ExtendOperIdx) == DU.NarrowDef && "bad DU");
842 const SCEV *ExtendOperExpr = 0;
843 const OverflowingBinaryOperator *OBO =
844 cast<OverflowingBinaryOperator>(DU.NarrowUse);
845 if (IsSigned && OBO->hasNoSignedWrap())
846 ExtendOperExpr = SE->getSignExtendExpr(
847 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
848 else if(!IsSigned && OBO->hasNoUnsignedWrap())
849 ExtendOperExpr = SE->getZeroExtendExpr(
850 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
854 // When creating this AddExpr, don't apply the current operations NSW or NUW
855 // flags. This instruction may be guarded by control flow that the no-wrap
856 // behavior depends on. Non-control-equivalent instructions can be mapped to
857 // the same SCEV expression, and it would be incorrect to transfer NSW/NUW
858 // semantics to those operations.
859 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(
860 SE->getAddExpr(SE->getSCEV(DU.WideDef), ExtendOperExpr));
862 if (!AddRec || AddRec->getLoop() != L)
867 /// GetWideRecurrence - Is this instruction potentially interesting from
868 /// IVUsers' perspective after widening it's type? In other words, can the
869 /// extend be safely hoisted out of the loop with SCEV reducing the value to a
870 /// recurrence on the same loop. If so, return the sign or zero extended
871 /// recurrence. Otherwise return NULL.
872 const SCEVAddRecExpr *WidenIV::GetWideRecurrence(Instruction *NarrowUse) {
873 if (!SE->isSCEVable(NarrowUse->getType()))
876 const SCEV *NarrowExpr = SE->getSCEV(NarrowUse);
877 if (SE->getTypeSizeInBits(NarrowExpr->getType())
878 >= SE->getTypeSizeInBits(WideType)) {
879 // NarrowUse implicitly widens its operand. e.g. a gep with a narrow
880 // index. So don't follow this use.
884 const SCEV *WideExpr = IsSigned ?
885 SE->getSignExtendExpr(NarrowExpr, WideType) :
886 SE->getZeroExtendExpr(NarrowExpr, WideType);
887 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr);
888 if (!AddRec || AddRec->getLoop() != L)
893 /// WidenIVUse - Determine whether an individual user of the narrow IV can be
894 /// widened. If so, return the wide clone of the user.
895 Instruction *WidenIV::WidenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter) {
897 // Stop traversing the def-use chain at inner-loop phis or post-loop phis.
898 if (isa<PHINode>(DU.NarrowUse) &&
899 LI->getLoopFor(DU.NarrowUse->getParent()) != L)
902 // Our raison d'etre! Eliminate sign and zero extension.
903 if (IsSigned ? isa<SExtInst>(DU.NarrowUse) : isa<ZExtInst>(DU.NarrowUse)) {
904 Value *NewDef = DU.WideDef;
905 if (DU.NarrowUse->getType() != WideType) {
906 unsigned CastWidth = SE->getTypeSizeInBits(DU.NarrowUse->getType());
907 unsigned IVWidth = SE->getTypeSizeInBits(WideType);
908 if (CastWidth < IVWidth) {
909 // The cast isn't as wide as the IV, so insert a Trunc.
910 IRBuilder<> Builder(DU.NarrowUse);
911 NewDef = Builder.CreateTrunc(DU.WideDef, DU.NarrowUse->getType());
914 // A wider extend was hidden behind a narrower one. This may induce
915 // another round of IV widening in which the intermediate IV becomes
916 // dead. It should be very rare.
917 DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi
918 << " not wide enough to subsume " << *DU.NarrowUse << "\n");
919 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
920 NewDef = DU.NarrowUse;
923 if (NewDef != DU.NarrowUse) {
924 DEBUG(dbgs() << "INDVARS: eliminating " << *DU.NarrowUse
925 << " replaced by " << *DU.WideDef << "\n");
927 DU.NarrowUse->replaceAllUsesWith(NewDef);
928 DeadInsts.push_back(DU.NarrowUse);
930 // Now that the extend is gone, we want to expose it's uses for potential
931 // further simplification. We don't need to directly inform SimplifyIVUsers
932 // of the new users, because their parent IV will be processed later as a
933 // new loop phi. If we preserved IVUsers analysis, we would also want to
934 // push the uses of WideDef here.
936 // No further widening is needed. The deceased [sz]ext had done it for us.
940 // Does this user itself evaluate to a recurrence after widening?
941 const SCEVAddRecExpr *WideAddRec = GetWideRecurrence(DU.NarrowUse);
943 WideAddRec = GetExtendedOperandRecurrence(DU);
946 // This user does not evaluate to a recurence after widening, so don't
947 // follow it. Instead insert a Trunc to kill off the original use,
948 // eventually isolating the original narrow IV so it can be removed.
949 IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT));
950 Value *Trunc = Builder.CreateTrunc(DU.WideDef, DU.NarrowDef->getType());
951 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, Trunc);
954 // Assume block terminators cannot evaluate to a recurrence. We can't to
955 // insert a Trunc after a terminator if there happens to be a critical edge.
956 assert(DU.NarrowUse != DU.NarrowUse->getParent()->getTerminator() &&
957 "SCEV is not expected to evaluate a block terminator");
959 // Reuse the IV increment that SCEVExpander created as long as it dominates
961 Instruction *WideUse = 0;
962 if (WideAddRec == WideIncExpr
963 && Rewriter.hoistIVInc(WideInc, DU.NarrowUse))
966 WideUse = CloneIVUser(DU);
970 // Evaluation of WideAddRec ensured that the narrow expression could be
971 // extended outside the loop without overflow. This suggests that the wide use
972 // evaluates to the same expression as the extended narrow use, but doesn't
973 // absolutely guarantee it. Hence the following failsafe check. In rare cases
974 // where it fails, we simply throw away the newly created wide use.
975 if (WideAddRec != SE->getSCEV(WideUse)) {
976 DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse
977 << ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n");
978 DeadInsts.push_back(WideUse);
982 // Returning WideUse pushes it on the worklist.
986 /// pushNarrowIVUsers - Add eligible users of NarrowDef to NarrowIVUsers.
988 void WidenIV::pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef) {
989 for (Value::use_iterator UI = NarrowDef->use_begin(),
990 UE = NarrowDef->use_end(); UI != UE; ++UI) {
991 Instruction *NarrowUse = cast<Instruction>(*UI);
993 // Handle data flow merges and bizarre phi cycles.
994 if (!Widened.insert(NarrowUse))
997 NarrowIVUsers.push_back(NarrowIVDefUse(NarrowDef, NarrowUse, WideDef));
1001 /// CreateWideIV - Process a single induction variable. First use the
1002 /// SCEVExpander to create a wide induction variable that evaluates to the same
1003 /// recurrence as the original narrow IV. Then use a worklist to forward
1004 /// traverse the narrow IV's def-use chain. After WidenIVUse has processed all
1005 /// interesting IV users, the narrow IV will be isolated for removal by
1008 /// It would be simpler to delete uses as they are processed, but we must avoid
1009 /// invalidating SCEV expressions.
1011 PHINode *WidenIV::CreateWideIV(SCEVExpander &Rewriter) {
1012 // Is this phi an induction variable?
1013 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi));
1017 // Widen the induction variable expression.
1018 const SCEV *WideIVExpr = IsSigned ?
1019 SE->getSignExtendExpr(AddRec, WideType) :
1020 SE->getZeroExtendExpr(AddRec, WideType);
1022 assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType &&
1023 "Expect the new IV expression to preserve its type");
1025 // Can the IV be extended outside the loop without overflow?
1026 AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr);
1027 if (!AddRec || AddRec->getLoop() != L)
1030 // An AddRec must have loop-invariant operands. Since this AddRec is
1031 // materialized by a loop header phi, the expression cannot have any post-loop
1032 // operands, so they must dominate the loop header.
1033 assert(SE->properlyDominates(AddRec->getStart(), L->getHeader()) &&
1034 SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader())
1035 && "Loop header phi recurrence inputs do not dominate the loop");
1037 // The rewriter provides a value for the desired IV expression. This may
1038 // either find an existing phi or materialize a new one. Either way, we
1039 // expect a well-formed cyclic phi-with-increments. i.e. any operand not part
1040 // of the phi-SCC dominates the loop entry.
1041 Instruction *InsertPt = L->getHeader()->begin();
1042 WidePhi = cast<PHINode>(Rewriter.expandCodeFor(AddRec, WideType, InsertPt));
1044 // Remembering the WideIV increment generated by SCEVExpander allows
1045 // WidenIVUse to reuse it when widening the narrow IV's increment. We don't
1046 // employ a general reuse mechanism because the call above is the only call to
1047 // SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses.
1048 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1050 cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock));
1051 WideIncExpr = SE->getSCEV(WideInc);
1054 DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n");
1057 // Traverse the def-use chain using a worklist starting at the original IV.
1058 assert(Widened.empty() && NarrowIVUsers.empty() && "expect initial state" );
1060 Widened.insert(OrigPhi);
1061 pushNarrowIVUsers(OrigPhi, WidePhi);
1063 while (!NarrowIVUsers.empty()) {
1064 NarrowIVDefUse DU = NarrowIVUsers.pop_back_val();
1066 // Process a def-use edge. This may replace the use, so don't hold a
1067 // use_iterator across it.
1068 Instruction *WideUse = WidenIVUse(DU, Rewriter);
1070 // Follow all def-use edges from the previous narrow use.
1072 pushNarrowIVUsers(DU.NarrowUse, WideUse);
1074 // WidenIVUse may have removed the def-use edge.
1075 if (DU.NarrowDef->use_empty())
1076 DeadInsts.push_back(DU.NarrowDef);
1081 //===----------------------------------------------------------------------===//
1082 // Simplification of IV users based on SCEV evaluation.
1083 //===----------------------------------------------------------------------===//
1086 /// SimplifyAndExtend - Iteratively perform simplification on a worklist of IV
1087 /// users. Each successive simplification may push more users which may
1088 /// themselves be candidates for simplification.
1090 /// Sign/Zero extend elimination is interleaved with IV simplification.
1092 void IndVarSimplify::SimplifyAndExtend(Loop *L,
1093 SCEVExpander &Rewriter,
1094 LPPassManager &LPM) {
1095 SmallVector<WideIVInfo, 8> WideIVs;
1097 SmallVector<PHINode*, 8> LoopPhis;
1098 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1099 LoopPhis.push_back(cast<PHINode>(I));
1101 // Each round of simplification iterates through the SimplifyIVUsers worklist
1102 // for all current phis, then determines whether any IVs can be
1103 // widened. Widening adds new phis to LoopPhis, inducing another round of
1104 // simplification on the wide IVs.
1105 while (!LoopPhis.empty()) {
1106 // Evaluate as many IV expressions as possible before widening any IVs. This
1107 // forces SCEV to set no-wrap flags before evaluating sign/zero
1108 // extension. The first time SCEV attempts to normalize sign/zero extension,
1109 // the result becomes final. So for the most predictable results, we delay
1110 // evaluation of sign/zero extend evaluation until needed, and avoid running
1111 // other SCEV based analysis prior to SimplifyAndExtend.
1113 PHINode *CurrIV = LoopPhis.pop_back_val();
1115 // Information about sign/zero extensions of CurrIV.
1116 WideIVVisitor WIV(CurrIV, SE, TD);
1118 Changed |= simplifyUsersOfIV(CurrIV, SE, &LPM, DeadInsts, &WIV);
1120 if (WIV.WI.WidestNativeType) {
1121 WideIVs.push_back(WIV.WI);
1123 } while(!LoopPhis.empty());
1125 for (; !WideIVs.empty(); WideIVs.pop_back()) {
1126 WidenIV Widener(WideIVs.back(), LI, SE, DT, DeadInsts);
1127 if (PHINode *WidePhi = Widener.CreateWideIV(Rewriter)) {
1129 LoopPhis.push_back(WidePhi);
1135 //===----------------------------------------------------------------------===//
1136 // LinearFunctionTestReplace and its kin. Rewrite the loop exit condition.
1137 //===----------------------------------------------------------------------===//
1139 /// Check for expressions that ScalarEvolution generates to compute
1140 /// BackedgeTakenInfo. If these expressions have not been reduced, then
1141 /// expanding them may incur additional cost (albeit in the loop preheader).
1142 static bool isHighCostExpansion(const SCEV *S, BranchInst *BI,
1143 SmallPtrSet<const SCEV*, 8> &Processed,
1144 ScalarEvolution *SE) {
1145 if (!Processed.insert(S))
1148 // If the backedge-taken count is a UDiv, it's very likely a UDiv that
1149 // ScalarEvolution's HowFarToZero or HowManyLessThans produced to compute a
1150 // precise expression, rather than a UDiv from the user's code. If we can't
1151 // find a UDiv in the code with some simple searching, assume the former and
1152 // forego rewriting the loop.
1153 if (isa<SCEVUDivExpr>(S)) {
1154 ICmpInst *OrigCond = dyn_cast<ICmpInst>(BI->getCondition());
1155 if (!OrigCond) return true;
1156 const SCEV *R = SE->getSCEV(OrigCond->getOperand(1));
1157 R = SE->getMinusSCEV(R, SE->getConstant(R->getType(), 1));
1159 const SCEV *L = SE->getSCEV(OrigCond->getOperand(0));
1160 L = SE->getMinusSCEV(L, SE->getConstant(L->getType(), 1));
1166 // Recurse past add expressions, which commonly occur in the
1167 // BackedgeTakenCount. They may already exist in program code, and if not,
1168 // they are not too expensive rematerialize.
1169 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
1170 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
1172 if (isHighCostExpansion(*I, BI, Processed, SE))
1178 // HowManyLessThans uses a Max expression whenever the loop is not guarded by
1179 // the exit condition.
1180 if (isa<SCEVSMaxExpr>(S) || isa<SCEVUMaxExpr>(S))
1183 // If we haven't recognized an expensive SCEV pattern, assume it's an
1184 // expression produced by program code.
1188 /// canExpandBackedgeTakenCount - Return true if this loop's backedge taken
1189 /// count expression can be safely and cheaply expanded into an instruction
1190 /// sequence that can be used by LinearFunctionTestReplace.
1192 /// TODO: This fails for pointer-type loop counters with greater than one byte
1193 /// strides, consequently preventing LFTR from running. For the purpose of LFTR
1194 /// we could skip this check in the case that the LFTR loop counter (chosen by
1195 /// FindLoopCounter) is also pointer type. Instead, we could directly convert
1196 /// the loop test to an inequality test by checking the target data's alignment
1197 /// of element types (given that the initial pointer value originates from or is
1198 /// used by ABI constrained operation, as opposed to inttoptr/ptrtoint).
1199 /// However, we don't yet have a strong motivation for converting loop tests
1200 /// into inequality tests.
1201 static bool canExpandBackedgeTakenCount(Loop *L, ScalarEvolution *SE) {
1202 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1203 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) ||
1204 BackedgeTakenCount->isZero())
1207 if (!L->getExitingBlock())
1210 // Can't rewrite non-branch yet.
1211 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
1215 SmallPtrSet<const SCEV*, 8> Processed;
1216 if (isHighCostExpansion(BackedgeTakenCount, BI, Processed, SE))
1222 /// getLoopPhiForCounter - Return the loop header phi IFF IncV adds a loop
1223 /// invariant value to the phi.
1224 static PHINode *getLoopPhiForCounter(Value *IncV, Loop *L, DominatorTree *DT) {
1225 Instruction *IncI = dyn_cast<Instruction>(IncV);
1229 switch (IncI->getOpcode()) {
1230 case Instruction::Add:
1231 case Instruction::Sub:
1233 case Instruction::GetElementPtr:
1234 // An IV counter must preserve its type.
1235 if (IncI->getNumOperands() == 2)
1241 PHINode *Phi = dyn_cast<PHINode>(IncI->getOperand(0));
1242 if (Phi && Phi->getParent() == L->getHeader()) {
1243 if (isLoopInvariant(IncI->getOperand(1), L, DT))
1247 if (IncI->getOpcode() == Instruction::GetElementPtr)
1250 // Allow add/sub to be commuted.
1251 Phi = dyn_cast<PHINode>(IncI->getOperand(1));
1252 if (Phi && Phi->getParent() == L->getHeader()) {
1253 if (isLoopInvariant(IncI->getOperand(0), L, DT))
1259 /// Return the compare guarding the loop latch, or NULL for unrecognized tests.
1260 static ICmpInst *getLoopTest(Loop *L) {
1261 assert(L->getExitingBlock() && "expected loop exit");
1263 BasicBlock *LatchBlock = L->getLoopLatch();
1264 // Don't bother with LFTR if the loop is not properly simplified.
1268 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
1269 assert(BI && "expected exit branch");
1271 return dyn_cast<ICmpInst>(BI->getCondition());
1274 /// needsLFTR - LinearFunctionTestReplace policy. Return true unless we can show
1275 /// that the current exit test is already sufficiently canonical.
1276 static bool needsLFTR(Loop *L, DominatorTree *DT) {
1277 // Do LFTR to simplify the exit condition to an ICMP.
1278 ICmpInst *Cond = getLoopTest(L);
1282 // Do LFTR to simplify the exit ICMP to EQ/NE
1283 ICmpInst::Predicate Pred = Cond->getPredicate();
1284 if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ)
1287 // Look for a loop invariant RHS
1288 Value *LHS = Cond->getOperand(0);
1289 Value *RHS = Cond->getOperand(1);
1290 if (!isLoopInvariant(RHS, L, DT)) {
1291 if (!isLoopInvariant(LHS, L, DT))
1293 std::swap(LHS, RHS);
1295 // Look for a simple IV counter LHS
1296 PHINode *Phi = dyn_cast<PHINode>(LHS);
1298 Phi = getLoopPhiForCounter(LHS, L, DT);
1303 // Do LFTR if PHI node is defined in the loop, but is *not* a counter.
1304 int Idx = Phi->getBasicBlockIndex(L->getLoopLatch());
1308 // Do LFTR if the exit condition's IV is *not* a simple counter.
1309 Value *IncV = Phi->getIncomingValue(Idx);
1310 return Phi != getLoopPhiForCounter(IncV, L, DT);
1313 /// Recursive helper for hasConcreteDef(). Unfortunately, this currently boils
1314 /// down to checking that all operands are constant and listing instructions
1315 /// that may hide undef.
1316 static bool hasConcreteDefImpl(Value *V, SmallPtrSet<Value*, 8> &Visited,
1318 if (isa<Constant>(V))
1319 return !isa<UndefValue>(V);
1324 // Conservatively handle non-constant non-instructions. For example, Arguments
1326 Instruction *I = dyn_cast<Instruction>(V);
1330 // Load and return values may be undef.
1331 if(I->mayReadFromMemory() || isa<CallInst>(I) || isa<InvokeInst>(I))
1334 // Optimistically handle other instructions.
1335 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) {
1336 if (!Visited.insert(*OI))
1338 if (!hasConcreteDefImpl(*OI, Visited, Depth+1))
1344 /// Return true if the given value is concrete. We must prove that undef can
1347 /// TODO: If we decide that this is a good approach to checking for undef, we
1348 /// may factor it into a common location.
1349 static bool hasConcreteDef(Value *V) {
1350 SmallPtrSet<Value*, 8> Visited;
1352 return hasConcreteDefImpl(V, Visited, 0);
1355 /// AlmostDeadIV - Return true if this IV has any uses other than the (soon to
1356 /// be rewritten) loop exit test.
1357 static bool AlmostDeadIV(PHINode *Phi, BasicBlock *LatchBlock, Value *Cond) {
1358 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1359 Value *IncV = Phi->getIncomingValue(LatchIdx);
1361 for (Value::use_iterator UI = Phi->use_begin(), UE = Phi->use_end();
1363 if (*UI != Cond && *UI != IncV) return false;
1366 for (Value::use_iterator UI = IncV->use_begin(), UE = IncV->use_end();
1368 if (*UI != Cond && *UI != Phi) return false;
1373 /// FindLoopCounter - Find an affine IV in canonical form.
1375 /// BECount may be an i8* pointer type. The pointer difference is already
1376 /// valid count without scaling the address stride, so it remains a pointer
1377 /// expression as far as SCEV is concerned.
1379 /// Currently only valid for LFTR. See the comments on hasConcreteDef below.
1381 /// FIXME: Accept -1 stride and set IVLimit = IVInit - BECount
1383 /// FIXME: Accept non-unit stride as long as SCEV can reduce BECount * Stride.
1384 /// This is difficult in general for SCEV because of potential overflow. But we
1385 /// could at least handle constant BECounts.
1387 FindLoopCounter(Loop *L, const SCEV *BECount,
1388 ScalarEvolution *SE, DominatorTree *DT, const DataLayout *TD) {
1389 uint64_t BCWidth = SE->getTypeSizeInBits(BECount->getType());
1392 cast<BranchInst>(L->getExitingBlock()->getTerminator())->getCondition();
1394 // Loop over all of the PHI nodes, looking for a simple counter.
1395 PHINode *BestPhi = 0;
1396 const SCEV *BestInit = 0;
1397 BasicBlock *LatchBlock = L->getLoopLatch();
1398 assert(LatchBlock && "needsLFTR should guarantee a loop latch");
1400 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1401 PHINode *Phi = cast<PHINode>(I);
1402 if (!SE->isSCEVable(Phi->getType()))
1405 // Avoid comparing an integer IV against a pointer Limit.
1406 if (BECount->getType()->isPointerTy() && !Phi->getType()->isPointerTy())
1409 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Phi));
1410 if (!AR || AR->getLoop() != L || !AR->isAffine())
1413 // AR may be a pointer type, while BECount is an integer type.
1414 // AR may be wider than BECount. With eq/ne tests overflow is immaterial.
1415 // AR may not be a narrower type, or we may never exit.
1416 uint64_t PhiWidth = SE->getTypeSizeInBits(AR->getType());
1417 if (PhiWidth < BCWidth || (TD && !TD->isLegalInteger(PhiWidth)))
1420 const SCEV *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
1421 if (!Step || !Step->isOne())
1424 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1425 Value *IncV = Phi->getIncomingValue(LatchIdx);
1426 if (getLoopPhiForCounter(IncV, L, DT) != Phi)
1429 // Avoid reusing a potentially undef value to compute other values that may
1430 // have originally had a concrete definition.
1431 if (!hasConcreteDef(Phi)) {
1432 // We explicitly allow unknown phis as long as they are already used by
1433 // the loop test. In this case we assume that performing LFTR could not
1434 // increase the number of undef users.
1435 if (ICmpInst *Cond = getLoopTest(L)) {
1436 if (Phi != getLoopPhiForCounter(Cond->getOperand(0), L, DT)
1437 && Phi != getLoopPhiForCounter(Cond->getOperand(1), L, DT)) {
1442 const SCEV *Init = AR->getStart();
1444 if (BestPhi && !AlmostDeadIV(BestPhi, LatchBlock, Cond)) {
1445 // Don't force a live loop counter if another IV can be used.
1446 if (AlmostDeadIV(Phi, LatchBlock, Cond))
1449 // Prefer to count-from-zero. This is a more "canonical" counter form. It
1450 // also prefers integer to pointer IVs.
1451 if (BestInit->isZero() != Init->isZero()) {
1452 if (BestInit->isZero())
1455 // If two IVs both count from zero or both count from nonzero then the
1456 // narrower is likely a dead phi that has been widened. Use the wider phi
1457 // to allow the other to be eliminated.
1458 else if (PhiWidth <= SE->getTypeSizeInBits(BestPhi->getType()))
1467 /// genLoopLimit - Help LinearFunctionTestReplace by generating a value that
1468 /// holds the RHS of the new loop test.
1469 static Value *genLoopLimit(PHINode *IndVar, const SCEV *IVCount, Loop *L,
1470 SCEVExpander &Rewriter, ScalarEvolution *SE) {
1471 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
1472 assert(AR && AR->getLoop() == L && AR->isAffine() && "bad loop counter");
1473 const SCEV *IVInit = AR->getStart();
1475 // IVInit may be a pointer while IVCount is an integer when FindLoopCounter
1476 // finds a valid pointer IV. Sign extend BECount in order to materialize a
1477 // GEP. Avoid running SCEVExpander on a new pointer value, instead reusing
1478 // the existing GEPs whenever possible.
1479 if (IndVar->getType()->isPointerTy()
1480 && !IVCount->getType()->isPointerTy()) {
1482 Type *OfsTy = SE->getEffectiveSCEVType(IVInit->getType());
1483 const SCEV *IVOffset = SE->getTruncateOrSignExtend(IVCount, OfsTy);
1485 // Expand the code for the iteration count.
1486 assert(SE->isLoopInvariant(IVOffset, L) &&
1487 "Computed iteration count is not loop invariant!");
1488 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1489 Value *GEPOffset = Rewriter.expandCodeFor(IVOffset, OfsTy, BI);
1491 Value *GEPBase = IndVar->getIncomingValueForBlock(L->getLoopPreheader());
1492 assert(AR->getStart() == SE->getSCEV(GEPBase) && "bad loop counter");
1493 // We could handle pointer IVs other than i8*, but we need to compensate for
1494 // gep index scaling. See canExpandBackedgeTakenCount comments.
1495 assert(SE->getSizeOfExpr(
1496 cast<PointerType>(GEPBase->getType())->getElementType())->isOne()
1497 && "unit stride pointer IV must be i8*");
1499 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
1500 return Builder.CreateGEP(GEPBase, GEPOffset, "lftr.limit");
1503 // In any other case, convert both IVInit and IVCount to integers before
1504 // comparing. This may result in SCEV expension of pointers, but in practice
1505 // SCEV will fold the pointer arithmetic away as such:
1506 // BECount = (IVEnd - IVInit - 1) => IVLimit = IVInit (postinc).
1508 // Valid Cases: (1) both integers is most common; (2) both may be pointers
1509 // for simple memset-style loops; (3) IVInit is an integer and IVCount is a
1510 // pointer may occur when enable-iv-rewrite generates a canonical IV on top
1513 const SCEV *IVLimit = 0;
1514 // For unit stride, IVCount = Start + BECount with 2's complement overflow.
1515 // For non-zero Start, compute IVCount here.
1516 if (AR->getStart()->isZero())
1519 assert(AR->getStepRecurrence(*SE)->isOne() && "only handles unit stride");
1520 const SCEV *IVInit = AR->getStart();
1522 // For integer IVs, truncate the IV before computing IVInit + BECount.
1523 if (SE->getTypeSizeInBits(IVInit->getType())
1524 > SE->getTypeSizeInBits(IVCount->getType()))
1525 IVInit = SE->getTruncateExpr(IVInit, IVCount->getType());
1527 IVLimit = SE->getAddExpr(IVInit, IVCount);
1529 // Expand the code for the iteration count.
1530 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1531 IRBuilder<> Builder(BI);
1532 assert(SE->isLoopInvariant(IVLimit, L) &&
1533 "Computed iteration count is not loop invariant!");
1534 // Ensure that we generate the same type as IndVar, or a smaller integer
1535 // type. In the presence of null pointer values, we have an integer type
1536 // SCEV expression (IVInit) for a pointer type IV value (IndVar).
1537 Type *LimitTy = IVCount->getType()->isPointerTy() ?
1538 IndVar->getType() : IVCount->getType();
1539 return Rewriter.expandCodeFor(IVLimit, LimitTy, BI);
1543 /// LinearFunctionTestReplace - This method rewrites the exit condition of the
1544 /// loop to be a canonical != comparison against the incremented loop induction
1545 /// variable. This pass is able to rewrite the exit tests of any loop where the
1546 /// SCEV analysis can determine a loop-invariant trip count of the loop, which
1547 /// is actually a much broader range than just linear tests.
1548 Value *IndVarSimplify::
1549 LinearFunctionTestReplace(Loop *L,
1550 const SCEV *BackedgeTakenCount,
1552 SCEVExpander &Rewriter) {
1553 assert(canExpandBackedgeTakenCount(L, SE) && "precondition");
1555 // Initialize CmpIndVar and IVCount to their preincremented values.
1556 Value *CmpIndVar = IndVar;
1557 const SCEV *IVCount = BackedgeTakenCount;
1559 // If the exiting block is the same as the backedge block, we prefer to
1560 // compare against the post-incremented value, otherwise we must compare
1561 // against the preincremented value.
1562 if (L->getExitingBlock() == L->getLoopLatch()) {
1563 // Add one to the "backedge-taken" count to get the trip count.
1564 // This addition may overflow, which is valid as long as the comparison is
1565 // truncated to BackedgeTakenCount->getType().
1566 IVCount = SE->getAddExpr(BackedgeTakenCount,
1567 SE->getConstant(BackedgeTakenCount->getType(), 1));
1568 // The BackedgeTaken expression contains the number of times that the
1569 // backedge branches to the loop header. This is one less than the
1570 // number of times the loop executes, so use the incremented indvar.
1571 CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock());
1574 Value *ExitCnt = genLoopLimit(IndVar, IVCount, L, Rewriter, SE);
1575 assert(ExitCnt->getType()->isPointerTy() == IndVar->getType()->isPointerTy()
1576 && "genLoopLimit missed a cast");
1578 // Insert a new icmp_ne or icmp_eq instruction before the branch.
1579 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1580 ICmpInst::Predicate P;
1581 if (L->contains(BI->getSuccessor(0)))
1582 P = ICmpInst::ICMP_NE;
1584 P = ICmpInst::ICMP_EQ;
1586 DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
1587 << " LHS:" << *CmpIndVar << '\n'
1589 << (P == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
1590 << " RHS:\t" << *ExitCnt << "\n"
1591 << " IVCount:\t" << *IVCount << "\n");
1593 IRBuilder<> Builder(BI);
1595 // LFTR can ignore IV overflow and truncate to the width of
1596 // BECount. This avoids materializing the add(zext(add)) expression.
1597 unsigned CmpIndVarSize = SE->getTypeSizeInBits(CmpIndVar->getType());
1598 unsigned ExitCntSize = SE->getTypeSizeInBits(ExitCnt->getType());
1599 if (CmpIndVarSize > ExitCntSize) {
1600 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
1601 const SCEV *ARStart = AR->getStart();
1602 const SCEV *ARStep = AR->getStepRecurrence(*SE);
1603 // For constant IVCount, avoid truncation.
1604 if (isa<SCEVConstant>(ARStart) && isa<SCEVConstant>(IVCount)) {
1605 const APInt &Start = cast<SCEVConstant>(ARStart)->getValue()->getValue();
1606 APInt Count = cast<SCEVConstant>(IVCount)->getValue()->getValue();
1607 // Note that the post-inc value of BackedgeTakenCount may have overflowed
1608 // above such that IVCount is now zero.
1609 if (IVCount != BackedgeTakenCount && Count == 0) {
1610 Count = APInt::getMaxValue(Count.getBitWidth()).zext(CmpIndVarSize);
1614 Count = Count.zext(CmpIndVarSize);
1616 if (cast<SCEVConstant>(ARStep)->getValue()->isNegative())
1617 NewLimit = Start - Count;
1619 NewLimit = Start + Count;
1620 ExitCnt = ConstantInt::get(CmpIndVar->getType(), NewLimit);
1622 DEBUG(dbgs() << " Widen RHS:\t" << *ExitCnt << "\n");
1624 CmpIndVar = Builder.CreateTrunc(CmpIndVar, ExitCnt->getType(),
1628 Value *Cond = Builder.CreateICmp(P, CmpIndVar, ExitCnt, "exitcond");
1629 Value *OrigCond = BI->getCondition();
1630 // It's tempting to use replaceAllUsesWith here to fully replace the old
1631 // comparison, but that's not immediately safe, since users of the old
1632 // comparison may not be dominated by the new comparison. Instead, just
1633 // update the branch to use the new comparison; in the common case this
1634 // will make old comparison dead.
1635 BI->setCondition(Cond);
1636 DeadInsts.push_back(OrigCond);
1643 //===----------------------------------------------------------------------===//
1644 // SinkUnusedInvariants. A late subpass to cleanup loop preheaders.
1645 //===----------------------------------------------------------------------===//
1647 /// If there's a single exit block, sink any loop-invariant values that
1648 /// were defined in the preheader but not used inside the loop into the
1649 /// exit block to reduce register pressure in the loop.
1650 void IndVarSimplify::SinkUnusedInvariants(Loop *L) {
1651 BasicBlock *ExitBlock = L->getExitBlock();
1652 if (!ExitBlock) return;
1654 BasicBlock *Preheader = L->getLoopPreheader();
1655 if (!Preheader) return;
1657 Instruction *InsertPt = ExitBlock->getFirstInsertionPt();
1658 BasicBlock::iterator I = Preheader->getTerminator();
1659 while (I != Preheader->begin()) {
1661 // New instructions were inserted at the end of the preheader.
1662 if (isa<PHINode>(I))
1665 // Don't move instructions which might have side effects, since the side
1666 // effects need to complete before instructions inside the loop. Also don't
1667 // move instructions which might read memory, since the loop may modify
1668 // memory. Note that it's okay if the instruction might have undefined
1669 // behavior: LoopSimplify guarantees that the preheader dominates the exit
1671 if (I->mayHaveSideEffects() || I->mayReadFromMemory())
1674 // Skip debug info intrinsics.
1675 if (isa<DbgInfoIntrinsic>(I))
1678 // Skip landingpad instructions.
1679 if (isa<LandingPadInst>(I))
1682 // Don't sink alloca: we never want to sink static alloca's out of the
1683 // entry block, and correctly sinking dynamic alloca's requires
1684 // checks for stacksave/stackrestore intrinsics.
1685 // FIXME: Refactor this check somehow?
1686 if (isa<AllocaInst>(I))
1689 // Determine if there is a use in or before the loop (direct or
1691 bool UsedInLoop = false;
1692 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
1695 BasicBlock *UseBB = cast<Instruction>(U)->getParent();
1696 if (PHINode *P = dyn_cast<PHINode>(U)) {
1698 PHINode::getIncomingValueNumForOperand(UI.getOperandNo());
1699 UseBB = P->getIncomingBlock(i);
1701 if (UseBB == Preheader || L->contains(UseBB)) {
1707 // If there is, the def must remain in the preheader.
1711 // Otherwise, sink it to the exit block.
1712 Instruction *ToMove = I;
1715 if (I != Preheader->begin()) {
1716 // Skip debug info intrinsics.
1719 } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin());
1721 if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin())
1727 ToMove->moveBefore(InsertPt);
1733 //===----------------------------------------------------------------------===//
1734 // IndVarSimplify driver. Manage several subpasses of IV simplification.
1735 //===----------------------------------------------------------------------===//
1737 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
1738 // If LoopSimplify form is not available, stay out of trouble. Some notes:
1739 // - LSR currently only supports LoopSimplify-form loops. Indvars'
1740 // canonicalization can be a pessimization without LSR to "clean up"
1742 // - We depend on having a preheader; in particular,
1743 // Loop::getCanonicalInductionVariable only supports loops with preheaders,
1744 // and we're in trouble if we can't find the induction variable even when
1745 // we've manually inserted one.
1746 if (!L->isLoopSimplifyForm())
1749 LI = &getAnalysis<LoopInfo>();
1750 SE = &getAnalysis<ScalarEvolution>();
1751 DT = &getAnalysis<DominatorTree>();
1752 TD = getAnalysisIfAvailable<DataLayout>();
1753 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
1758 // If there are any floating-point recurrences, attempt to
1759 // transform them to use integer recurrences.
1760 RewriteNonIntegerIVs(L);
1762 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1764 // Create a rewriter object which we'll use to transform the code with.
1765 SCEVExpander Rewriter(*SE, "indvars");
1767 Rewriter.setDebugType(DEBUG_TYPE);
1770 // Eliminate redundant IV users.
1772 // Simplification works best when run before other consumers of SCEV. We
1773 // attempt to avoid evaluating SCEVs for sign/zero extend operations until
1774 // other expressions involving loop IVs have been evaluated. This helps SCEV
1775 // set no-wrap flags before normalizing sign/zero extension.
1776 Rewriter.disableCanonicalMode();
1777 SimplifyAndExtend(L, Rewriter, LPM);
1779 // Check to see if this loop has a computable loop-invariant execution count.
1780 // If so, this means that we can compute the final value of any expressions
1781 // that are recurrent in the loop, and substitute the exit values from the
1782 // loop into any instructions outside of the loop that use the final values of
1783 // the current expressions.
1785 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1786 RewriteLoopExitValues(L, Rewriter);
1788 // Eliminate redundant IV cycles.
1789 NumElimIV += Rewriter.replaceCongruentIVs(L, DT, DeadInsts);
1791 // If we have a trip count expression, rewrite the loop's exit condition
1792 // using it. We can currently only handle loops with a single exit.
1793 if (canExpandBackedgeTakenCount(L, SE) && needsLFTR(L, DT)) {
1794 PHINode *IndVar = FindLoopCounter(L, BackedgeTakenCount, SE, DT, TD);
1796 // Check preconditions for proper SCEVExpander operation. SCEV does not
1797 // express SCEVExpander's dependencies, such as LoopSimplify. Instead any
1798 // pass that uses the SCEVExpander must do it. This does not work well for
1799 // loop passes because SCEVExpander makes assumptions about all loops, while
1800 // LoopPassManager only forces the current loop to be simplified.
1802 // FIXME: SCEV expansion has no way to bail out, so the caller must
1803 // explicitly check any assumptions made by SCEV. Brittle.
1804 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(BackedgeTakenCount);
1805 if (!AR || AR->getLoop()->getLoopPreheader())
1806 (void)LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar,
1810 // Clear the rewriter cache, because values that are in the rewriter's cache
1811 // can be deleted in the loop below, causing the AssertingVH in the cache to
1815 // Now that we're done iterating through lists, clean up any instructions
1816 // which are now dead.
1817 while (!DeadInsts.empty())
1818 if (Instruction *Inst =
1819 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
1820 RecursivelyDeleteTriviallyDeadInstructions(Inst, TLI);
1822 // The Rewriter may not be used from this point on.
1824 // Loop-invariant instructions in the preheader that aren't used in the
1825 // loop may be sunk below the loop to reduce register pressure.
1826 SinkUnusedInvariants(L);
1828 // Clean up dead instructions.
1829 Changed |= DeleteDeadPHIs(L->getHeader(), TLI);
1830 // Check a post-condition.
1831 assert(L->isLCSSAForm(*DT) &&
1832 "Indvars did not leave the loop in lcssa form!");
1834 // Verify that LFTR, and any other change have not interfered with SCEV's
1835 // ability to compute trip count.
1837 if (VerifyIndvars && !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) {
1839 const SCEV *NewBECount = SE->getBackedgeTakenCount(L);
1840 if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) <
1841 SE->getTypeSizeInBits(NewBECount->getType()))
1842 NewBECount = SE->getTruncateOrNoop(NewBECount,
1843 BackedgeTakenCount->getType());
1845 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount,
1846 NewBECount->getType());
1847 assert(BackedgeTakenCount == NewBECount && "indvars must preserve SCEV");