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/BasicBlock.h"
30 #include "llvm/Constants.h"
31 #include "llvm/Instructions.h"
32 #include "llvm/IntrinsicInst.h"
33 #include "llvm/LLVMContext.h"
34 #include "llvm/Type.h"
35 #include "llvm/Analysis/Dominators.h"
36 #include "llvm/Analysis/IVUsers.h"
37 #include "llvm/Analysis/ScalarEvolutionExpander.h"
38 #include "llvm/Analysis/LoopInfo.h"
39 #include "llvm/Analysis/LoopPass.h"
40 #include "llvm/Support/CFG.h"
41 #include "llvm/Support/CommandLine.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Support/raw_ostream.h"
44 #include "llvm/Transforms/Utils/Local.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/SimplifyIndVar.h"
47 #include "llvm/Target/TargetData.h"
48 #include "llvm/ADT/DenseMap.h"
49 #include "llvm/ADT/SmallVector.h"
50 #include "llvm/ADT/Statistic.h"
53 STATISTIC(NumRemoved , "Number of aux indvars removed");
54 STATISTIC(NumWidened , "Number of indvars widened");
55 STATISTIC(NumInserted , "Number of canonical indvars added");
56 STATISTIC(NumReplaced , "Number of exit values replaced");
57 STATISTIC(NumLFTR , "Number of loop exit tests replaced");
58 STATISTIC(NumElimExt , "Number of IV sign/zero extends eliminated");
59 STATISTIC(NumElimIV , "Number of congruent IVs eliminated");
62 cl::opt<bool> EnableIVRewrite(
63 "enable-iv-rewrite", cl::Hidden,
64 cl::desc("Enable canonical induction variable rewriting"));
66 // Trip count verification can be enabled by default under NDEBUG if we
67 // implement a strong expression equivalence checker in SCEV. Until then, we
68 // use the verify-indvars flag, which may assert in some cases.
69 cl::opt<bool> VerifyIndvars(
70 "verify-indvars", cl::Hidden,
71 cl::desc("Verify the ScalarEvolution result after running indvars"));
75 class IndVarSimplify : public LoopPass {
82 SmallVector<WeakVH, 16> DeadInsts;
86 static char ID; // Pass identification, replacement for typeid
87 IndVarSimplify() : LoopPass(ID), IU(0), LI(0), SE(0), DT(0), TD(0),
89 initializeIndVarSimplifyPass(*PassRegistry::getPassRegistry());
92 virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
94 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
95 AU.addRequired<DominatorTree>();
96 AU.addRequired<LoopInfo>();
97 AU.addRequired<ScalarEvolution>();
98 AU.addRequiredID(LoopSimplifyID);
99 AU.addRequiredID(LCSSAID);
101 AU.addRequired<IVUsers>();
102 AU.addPreserved<ScalarEvolution>();
103 AU.addPreservedID(LoopSimplifyID);
104 AU.addPreservedID(LCSSAID);
106 AU.addPreserved<IVUsers>();
107 AU.setPreservesCFG();
111 virtual void releaseMemory() {
115 bool isValidRewrite(Value *FromVal, Value *ToVal);
117 void HandleFloatingPointIV(Loop *L, PHINode *PH);
118 void RewriteNonIntegerIVs(Loop *L);
120 void SimplifyAndExtend(Loop *L, SCEVExpander &Rewriter, LPPassManager &LPM);
122 void SimplifyCongruentIVs(Loop *L);
124 void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter);
126 void RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter);
128 Value *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
129 PHINode *IndVar, SCEVExpander &Rewriter);
131 void SinkUnusedInvariants(Loop *L);
135 char IndVarSimplify::ID = 0;
136 INITIALIZE_PASS_BEGIN(IndVarSimplify, "indvars",
137 "Induction Variable Simplification", false, false)
138 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
139 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
140 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
141 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
142 INITIALIZE_PASS_DEPENDENCY(LCSSA)
143 INITIALIZE_PASS_DEPENDENCY(IVUsers)
144 INITIALIZE_PASS_END(IndVarSimplify, "indvars",
145 "Induction Variable Simplification", false, false)
147 Pass *llvm::createIndVarSimplifyPass() {
148 return new IndVarSimplify();
151 /// isValidRewrite - Return true if the SCEV expansion generated by the
152 /// rewriter can replace the original value. SCEV guarantees that it
153 /// produces the same value, but the way it is produced may be illegal IR.
154 /// Ideally, this function will only be called for verification.
155 bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) {
156 // If an SCEV expression subsumed multiple pointers, its expansion could
157 // reassociate the GEP changing the base pointer. This is illegal because the
158 // final address produced by a GEP chain must be inbounds relative to its
159 // underlying object. Otherwise basic alias analysis, among other things,
160 // could fail in a dangerous way. Ultimately, SCEV will be improved to avoid
161 // producing an expression involving multiple pointers. Until then, we must
164 // Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject
165 // because it understands lcssa phis while SCEV does not.
166 Value *FromPtr = FromVal;
167 Value *ToPtr = ToVal;
168 if (GEPOperator *GEP = dyn_cast<GEPOperator>(FromVal)) {
169 FromPtr = GEP->getPointerOperand();
171 if (GEPOperator *GEP = dyn_cast<GEPOperator>(ToVal)) {
172 ToPtr = GEP->getPointerOperand();
174 if (FromPtr != FromVal || ToPtr != ToVal) {
175 // Quickly check the common case
176 if (FromPtr == ToPtr)
179 // SCEV may have rewritten an expression that produces the GEP's pointer
180 // operand. That's ok as long as the pointer operand has the same base
181 // pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the
182 // base of a recurrence. This handles the case in which SCEV expansion
183 // converts a pointer type recurrence into a nonrecurrent pointer base
184 // indexed by an integer recurrence.
185 const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr));
186 const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr));
187 if (FromBase == ToBase)
190 DEBUG(dbgs() << "INDVARS: GEP rewrite bail out "
191 << *FromBase << " != " << *ToBase << "\n");
198 /// Determine the insertion point for this user. By default, insert immediately
199 /// before the user. SCEVExpander or LICM will hoist loop invariants out of the
200 /// loop. For PHI nodes, there may be multiple uses, so compute the nearest
201 /// common dominator for the incoming blocks.
202 static Instruction *getInsertPointForUses(Instruction *User, Value *Def,
204 PHINode *PHI = dyn_cast<PHINode>(User);
208 Instruction *InsertPt = 0;
209 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) {
210 if (PHI->getIncomingValue(i) != Def)
213 BasicBlock *InsertBB = PHI->getIncomingBlock(i);
215 InsertPt = InsertBB->getTerminator();
218 InsertBB = DT->findNearestCommonDominator(InsertPt->getParent(), InsertBB);
219 InsertPt = InsertBB->getTerminator();
221 assert(InsertPt && "Missing phi operand");
222 assert((!isa<Instruction>(Def) ||
223 DT->dominates(cast<Instruction>(Def), InsertPt)) &&
224 "def does not dominate all uses");
228 //===----------------------------------------------------------------------===//
229 // RewriteNonIntegerIVs and helpers. Prefer integer IVs.
230 //===----------------------------------------------------------------------===//
232 /// ConvertToSInt - Convert APF to an integer, if possible.
233 static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) {
234 bool isExact = false;
235 if (&APF.getSemantics() == &APFloat::PPCDoubleDouble)
237 // See if we can convert this to an int64_t
239 if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero,
240 &isExact) != APFloat::opOK || !isExact)
246 /// HandleFloatingPointIV - If the loop has floating induction variable
247 /// then insert corresponding integer induction variable if possible.
249 /// for(double i = 0; i < 10000; ++i)
251 /// is converted into
252 /// for(int i = 0; i < 10000; ++i)
255 void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) {
256 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
257 unsigned BackEdge = IncomingEdge^1;
259 // Check incoming value.
260 ConstantFP *InitValueVal =
261 dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));
264 if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
267 // Check IV increment. Reject this PN if increment operation is not
268 // an add or increment value can not be represented by an integer.
269 BinaryOperator *Incr =
270 dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
271 if (Incr == 0 || Incr->getOpcode() != Instruction::FAdd) return;
273 // If this is not an add of the PHI with a constantfp, or if the constant fp
274 // is not an integer, bail out.
275 ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
277 if (IncValueVal == 0 || Incr->getOperand(0) != PN ||
278 !ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
281 // Check Incr uses. One user is PN and the other user is an exit condition
282 // used by the conditional terminator.
283 Value::use_iterator IncrUse = Incr->use_begin();
284 Instruction *U1 = cast<Instruction>(*IncrUse++);
285 if (IncrUse == Incr->use_end()) return;
286 Instruction *U2 = cast<Instruction>(*IncrUse++);
287 if (IncrUse != Incr->use_end()) return;
289 // Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't
290 // only used by a branch, we can't transform it.
291 FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
293 Compare = dyn_cast<FCmpInst>(U2);
294 if (Compare == 0 || !Compare->hasOneUse() ||
295 !isa<BranchInst>(Compare->use_back()))
298 BranchInst *TheBr = cast<BranchInst>(Compare->use_back());
300 // We need to verify that the branch actually controls the iteration count
301 // of the loop. If not, the new IV can overflow and no one will notice.
302 // The branch block must be in the loop and one of the successors must be out
304 assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
305 if (!L->contains(TheBr->getParent()) ||
306 (L->contains(TheBr->getSuccessor(0)) &&
307 L->contains(TheBr->getSuccessor(1))))
311 // If it isn't a comparison with an integer-as-fp (the exit value), we can't
313 ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
315 if (ExitValueVal == 0 ||
316 !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
319 // Find new predicate for integer comparison.
320 CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
321 switch (Compare->getPredicate()) {
322 default: return; // Unknown comparison.
323 case CmpInst::FCMP_OEQ:
324 case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
325 case CmpInst::FCMP_ONE:
326 case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
327 case CmpInst::FCMP_OGT:
328 case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
329 case CmpInst::FCMP_OGE:
330 case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
331 case CmpInst::FCMP_OLT:
332 case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
333 case CmpInst::FCMP_OLE:
334 case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
337 // We convert the floating point induction variable to a signed i32 value if
338 // we can. This is only safe if the comparison will not overflow in a way
339 // that won't be trapped by the integer equivalent operations. Check for this
341 // TODO: We could use i64 if it is native and the range requires it.
343 // The start/stride/exit values must all fit in signed i32.
344 if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
347 // If not actually striding (add x, 0.0), avoid touching the code.
351 // Positive and negative strides have different safety conditions.
353 // If we have a positive stride, we require the init to be less than the
355 if (InitValue >= ExitValue)
358 uint32_t Range = uint32_t(ExitValue-InitValue);
359 // Check for infinite loop, either:
360 // while (i <= Exit) or until (i > Exit)
361 if (NewPred == CmpInst::ICMP_SLE || NewPred == CmpInst::ICMP_SGT) {
362 if (++Range == 0) return; // Range overflows.
365 unsigned Leftover = Range % uint32_t(IncValue);
367 // If this is an equality comparison, we require that the strided value
368 // exactly land on the exit value, otherwise the IV condition will wrap
369 // around and do things the fp IV wouldn't.
370 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
374 // If the stride would wrap around the i32 before exiting, we can't
376 if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
380 // If we have a negative stride, we require the init to be greater than the
382 if (InitValue <= ExitValue)
385 uint32_t Range = uint32_t(InitValue-ExitValue);
386 // Check for infinite loop, either:
387 // while (i >= Exit) or until (i < Exit)
388 if (NewPred == CmpInst::ICMP_SGE || NewPred == CmpInst::ICMP_SLT) {
389 if (++Range == 0) return; // Range overflows.
392 unsigned Leftover = Range % uint32_t(-IncValue);
394 // If this is an equality comparison, we require that the strided value
395 // exactly land on the exit value, otherwise the IV condition will wrap
396 // around and do things the fp IV wouldn't.
397 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
401 // If the stride would wrap around the i32 before exiting, we can't
403 if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
407 IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
409 // Insert new integer induction variable.
410 PHINode *NewPHI = PHINode::Create(Int32Ty, 2, PN->getName()+".int", PN);
411 NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
412 PN->getIncomingBlock(IncomingEdge));
415 BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
416 Incr->getName()+".int", Incr);
417 NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
419 ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
420 ConstantInt::get(Int32Ty, ExitValue),
423 // In the following deletions, PN may become dead and may be deleted.
424 // Use a WeakVH to observe whether this happens.
427 // Delete the old floating point exit comparison. The branch starts using the
429 NewCompare->takeName(Compare);
430 Compare->replaceAllUsesWith(NewCompare);
431 RecursivelyDeleteTriviallyDeadInstructions(Compare);
433 // Delete the old floating point increment.
434 Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
435 RecursivelyDeleteTriviallyDeadInstructions(Incr);
437 // If the FP induction variable still has uses, this is because something else
438 // in the loop uses its value. In order to canonicalize the induction
439 // variable, we chose to eliminate the IV and rewrite it in terms of an
442 // We give preference to sitofp over uitofp because it is faster on most
445 Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
446 PN->getParent()->getFirstInsertionPt());
447 PN->replaceAllUsesWith(Conv);
448 RecursivelyDeleteTriviallyDeadInstructions(PN);
451 // Add a new IVUsers entry for the newly-created integer PHI.
453 IU->AddUsersIfInteresting(NewPHI);
458 void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
459 // First step. Check to see if there are any floating-point recurrences.
460 // If there are, change them into integer recurrences, permitting analysis by
461 // the SCEV routines.
463 BasicBlock *Header = L->getHeader();
465 SmallVector<WeakVH, 8> PHIs;
466 for (BasicBlock::iterator I = Header->begin();
467 PHINode *PN = dyn_cast<PHINode>(I); ++I)
470 for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
471 if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i]))
472 HandleFloatingPointIV(L, PN);
474 // If the loop previously had floating-point IV, ScalarEvolution
475 // may not have been able to compute a trip count. Now that we've done some
476 // re-writing, the trip count may be computable.
481 //===----------------------------------------------------------------------===//
482 // RewriteLoopExitValues - Optimize IV users outside the loop.
483 // As a side effect, reduces the amount of IV processing within the loop.
484 //===----------------------------------------------------------------------===//
486 /// RewriteLoopExitValues - Check to see if this loop has a computable
487 /// loop-invariant execution count. If so, this means that we can compute the
488 /// final value of any expressions that are recurrent in the loop, and
489 /// substitute the exit values from the loop into any instructions outside of
490 /// the loop that use the final values of the current expressions.
492 /// This is mostly redundant with the regular IndVarSimplify activities that
493 /// happen later, except that it's more powerful in some cases, because it's
494 /// able to brute-force evaluate arbitrary instructions as long as they have
495 /// constant operands at the beginning of the loop.
496 void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) {
497 // Verify the input to the pass in already in LCSSA form.
498 assert(L->isLCSSAForm(*DT));
500 SmallVector<BasicBlock*, 8> ExitBlocks;
501 L->getUniqueExitBlocks(ExitBlocks);
503 // Find all values that are computed inside the loop, but used outside of it.
504 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan
505 // the exit blocks of the loop to find them.
506 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
507 BasicBlock *ExitBB = ExitBlocks[i];
509 // If there are no PHI nodes in this exit block, then no values defined
510 // inside the loop are used on this path, skip it.
511 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
514 unsigned NumPreds = PN->getNumIncomingValues();
516 // Iterate over all of the PHI nodes.
517 BasicBlock::iterator BBI = ExitBB->begin();
518 while ((PN = dyn_cast<PHINode>(BBI++))) {
520 continue; // dead use, don't replace it
522 // SCEV only supports integer expressions for now.
523 if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy())
526 // It's necessary to tell ScalarEvolution about this explicitly so that
527 // it can walk the def-use list and forget all SCEVs, as it may not be
528 // watching the PHI itself. Once the new exit value is in place, there
529 // may not be a def-use connection between the loop and every instruction
530 // which got a SCEVAddRecExpr for that loop.
533 // Iterate over all of the values in all the PHI nodes.
534 for (unsigned i = 0; i != NumPreds; ++i) {
535 // If the value being merged in is not integer or is not defined
536 // in the loop, skip it.
537 Value *InVal = PN->getIncomingValue(i);
538 if (!isa<Instruction>(InVal))
541 // If this pred is for a subloop, not L itself, skip it.
542 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
543 continue; // The Block is in a subloop, skip it.
545 // Check that InVal is defined in the loop.
546 Instruction *Inst = cast<Instruction>(InVal);
547 if (!L->contains(Inst))
550 // Okay, this instruction has a user outside of the current loop
551 // and varies predictably *inside* the loop. Evaluate the value it
552 // contains when the loop exits, if possible.
553 const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
554 if (!SE->isLoopInvariant(ExitValue, L))
557 Value *ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), Inst);
559 DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
560 << " LoopVal = " << *Inst << "\n");
562 if (!isValidRewrite(Inst, ExitVal)) {
563 DeadInsts.push_back(ExitVal);
569 PN->setIncomingValue(i, ExitVal);
571 // If this instruction is dead now, delete it.
572 RecursivelyDeleteTriviallyDeadInstructions(Inst);
575 // Completely replace a single-pred PHI. This is safe, because the
576 // NewVal won't be variant in the loop, so we don't need an LCSSA phi
578 PN->replaceAllUsesWith(ExitVal);
579 RecursivelyDeleteTriviallyDeadInstructions(PN);
583 // Clone the PHI and delete the original one. This lets IVUsers and
584 // any other maps purge the original user from their records.
585 PHINode *NewPN = cast<PHINode>(PN->clone());
587 NewPN->insertBefore(PN);
588 PN->replaceAllUsesWith(NewPN);
589 PN->eraseFromParent();
594 // The insertion point instruction may have been deleted; clear it out
595 // so that the rewriter doesn't trip over it later.
596 Rewriter.clearInsertPoint();
599 //===----------------------------------------------------------------------===//
600 // Rewrite IV users based on a canonical IV.
601 // Only for use with -enable-iv-rewrite.
602 //===----------------------------------------------------------------------===//
604 /// FIXME: It is an extremely bad idea to indvar substitute anything more
605 /// complex than affine induction variables. Doing so will put expensive
606 /// polynomial evaluations inside of the loop, and the str reduction pass
607 /// currently can only reduce affine polynomials. For now just disable
608 /// indvar subst on anything more complex than an affine addrec, unless
609 /// it can be expanded to a trivial value.
610 static bool isSafe(const SCEV *S, const Loop *L, ScalarEvolution *SE) {
611 // Loop-invariant values are safe.
612 if (SE->isLoopInvariant(S, L)) return true;
614 // Affine addrecs are safe. Non-affine are not, because LSR doesn't know how
615 // to transform them into efficient code.
616 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
617 return AR->isAffine();
619 // An add is safe it all its operands are safe.
620 if (const SCEVCommutativeExpr *Commutative
621 = dyn_cast<SCEVCommutativeExpr>(S)) {
622 for (SCEVCommutativeExpr::op_iterator I = Commutative->op_begin(),
623 E = Commutative->op_end(); I != E; ++I)
624 if (!isSafe(*I, L, SE)) return false;
628 // A cast is safe if its operand is.
629 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
630 return isSafe(C->getOperand(), L, SE);
632 // A udiv is safe if its operands are.
633 if (const SCEVUDivExpr *UD = dyn_cast<SCEVUDivExpr>(S))
634 return isSafe(UD->getLHS(), L, SE) &&
635 isSafe(UD->getRHS(), L, SE);
637 // SCEVUnknown is always safe.
638 if (isa<SCEVUnknown>(S))
641 // Nothing else is safe.
645 void IndVarSimplify::RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter) {
646 // Rewrite all induction variable expressions in terms of the canonical
647 // induction variable.
649 // If there were induction variables of other sizes or offsets, manually
650 // add the offsets to the primary induction variable and cast, avoiding
651 // the need for the code evaluation methods to insert induction variables
652 // of different sizes.
653 for (IVUsers::iterator UI = IU->begin(), E = IU->end(); UI != E; ++UI) {
654 Value *Op = UI->getOperandValToReplace();
655 Type *UseTy = Op->getType();
656 Instruction *User = UI->getUser();
658 // Compute the final addrec to expand into code.
659 const SCEV *AR = IU->getReplacementExpr(*UI);
661 // Evaluate the expression out of the loop, if possible.
662 if (!L->contains(UI->getUser())) {
663 const SCEV *ExitVal = SE->getSCEVAtScope(AR, L->getParentLoop());
664 if (SE->isLoopInvariant(ExitVal, L))
668 // FIXME: It is an extremely bad idea to indvar substitute anything more
669 // complex than affine induction variables. Doing so will put expensive
670 // polynomial evaluations inside of the loop, and the str reduction pass
671 // currently can only reduce affine polynomials. For now just disable
672 // indvar subst on anything more complex than an affine addrec, unless
673 // it can be expanded to a trivial value.
674 if (!isSafe(AR, L, SE))
677 // Determine the insertion point for this user. By default, insert
678 // immediately before the user. The SCEVExpander class will automatically
679 // hoist loop invariants out of the loop. For PHI nodes, there may be
680 // multiple uses, so compute the nearest common dominator for the
682 Instruction *InsertPt = getInsertPointForUses(User, Op, DT);
684 // Now expand it into actual Instructions and patch it into place.
685 Value *NewVal = Rewriter.expandCodeFor(AR, UseTy, InsertPt);
687 DEBUG(dbgs() << "INDVARS: Rewrote IV '" << *AR << "' " << *Op << '\n'
688 << " into = " << *NewVal << "\n");
690 if (!isValidRewrite(Op, NewVal)) {
691 DeadInsts.push_back(NewVal);
694 // Inform ScalarEvolution that this value is changing. The change doesn't
695 // affect its value, but it does potentially affect which use lists the
696 // value will be on after the replacement, which affects ScalarEvolution's
697 // ability to walk use lists and drop dangling pointers when a value is
699 SE->forgetValue(User);
701 // Patch the new value into place.
703 NewVal->takeName(Op);
704 if (Instruction *NewValI = dyn_cast<Instruction>(NewVal))
705 NewValI->setDebugLoc(User->getDebugLoc());
706 User->replaceUsesOfWith(Op, NewVal);
707 UI->setOperandValToReplace(NewVal);
712 // The old value may be dead now.
713 DeadInsts.push_back(Op);
717 //===----------------------------------------------------------------------===//
718 // IV Widening - Extend the width of an IV to cover its widest uses.
719 //===----------------------------------------------------------------------===//
722 // Collect information about induction variables that are used by sign/zero
723 // extend operations. This information is recorded by CollectExtend and
724 // provides the input to WidenIV.
726 Type *WidestNativeType; // Widest integer type created [sz]ext
727 bool IsSigned; // Was an sext user seen before a zext?
729 WideIVInfo() : WidestNativeType(0), IsSigned(false) {}
732 class WideIVVisitor : public IVVisitor {
734 const TargetData *TD;
739 WideIVVisitor(ScalarEvolution *SCEV, const TargetData *TData) :
740 SE(SCEV), TD(TData) {}
742 // Implement the interface used by simplifyUsersOfIV.
743 virtual void visitCast(CastInst *Cast);
747 /// visitCast - Update information about the induction variable that is
748 /// extended by this sign or zero extend operation. This is used to determine
749 /// the final width of the IV before actually widening it.
750 void WideIVVisitor::visitCast(CastInst *Cast) {
751 bool IsSigned = Cast->getOpcode() == Instruction::SExt;
752 if (!IsSigned && Cast->getOpcode() != Instruction::ZExt)
755 Type *Ty = Cast->getType();
756 uint64_t Width = SE->getTypeSizeInBits(Ty);
757 if (TD && !TD->isLegalInteger(Width))
760 if (!WI.WidestNativeType) {
761 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
762 WI.IsSigned = IsSigned;
766 // We extend the IV to satisfy the sign of its first user, arbitrarily.
767 if (WI.IsSigned != IsSigned)
770 if (Width > SE->getTypeSizeInBits(WI.WidestNativeType))
771 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
776 /// NarrowIVDefUse - Record a link in the Narrow IV def-use chain along with the
777 /// WideIV that computes the same value as the Narrow IV def. This avoids
778 /// caching Use* pointers.
779 struct NarrowIVDefUse {
780 Instruction *NarrowDef;
781 Instruction *NarrowUse;
782 Instruction *WideDef;
784 NarrowIVDefUse(): NarrowDef(0), NarrowUse(0), WideDef(0) {}
786 NarrowIVDefUse(Instruction *ND, Instruction *NU, Instruction *WD):
787 NarrowDef(ND), NarrowUse(NU), WideDef(WD) {}
790 /// WidenIV - The goal of this transform is to remove sign and zero extends
791 /// without creating any new induction variables. To do this, it creates a new
792 /// phi of the wider type and redirects all users, either removing extends or
793 /// inserting truncs whenever we stop propagating the type.
809 Instruction *WideInc;
810 const SCEV *WideIncExpr;
811 SmallVectorImpl<WeakVH> &DeadInsts;
813 SmallPtrSet<Instruction*,16> Widened;
814 SmallVector<NarrowIVDefUse, 8> NarrowIVUsers;
817 WidenIV(PHINode *PN, const WideIVInfo &WI, LoopInfo *LInfo,
818 ScalarEvolution *SEv, DominatorTree *DTree,
819 SmallVectorImpl<WeakVH> &DI) :
821 WideType(WI.WidestNativeType),
822 IsSigned(WI.IsSigned),
824 L(LI->getLoopFor(OrigPhi->getParent())),
831 assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV");
834 PHINode *CreateWideIV(SCEVExpander &Rewriter);
837 Instruction *CloneIVUser(NarrowIVDefUse DU);
839 const SCEVAddRecExpr *GetWideRecurrence(Instruction *NarrowUse);
841 const SCEVAddRecExpr* GetExtendedOperandRecurrence(NarrowIVDefUse DU);
843 Instruction *WidenIVUse(NarrowIVDefUse DU);
845 void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef);
847 } // anonymous namespace
849 static Value *getExtend( Value *NarrowOper, Type *WideType,
850 bool IsSigned, IRBuilder<> &Builder) {
851 return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) :
852 Builder.CreateZExt(NarrowOper, WideType);
855 /// CloneIVUser - Instantiate a wide operation to replace a narrow
856 /// operation. This only needs to handle operations that can evaluation to
857 /// SCEVAddRec. It can safely return 0 for any operation we decide not to clone.
858 Instruction *WidenIV::CloneIVUser(NarrowIVDefUse DU) {
859 unsigned Opcode = DU.NarrowUse->getOpcode();
863 case Instruction::Add:
864 case Instruction::Mul:
865 case Instruction::UDiv:
866 case Instruction::Sub:
867 case Instruction::And:
868 case Instruction::Or:
869 case Instruction::Xor:
870 case Instruction::Shl:
871 case Instruction::LShr:
872 case Instruction::AShr:
873 DEBUG(dbgs() << "Cloning IVUser: " << *DU.NarrowUse << "\n");
875 IRBuilder<> Builder(DU.NarrowUse);
877 // Replace NarrowDef operands with WideDef. Otherwise, we don't know
878 // anything about the narrow operand yet so must insert a [sz]ext. It is
879 // probably loop invariant and will be folded or hoisted. If it actually
880 // comes from a widened IV, it should be removed during a future call to
882 Value *LHS = (DU.NarrowUse->getOperand(0) == DU.NarrowDef) ? DU.WideDef :
883 getExtend(DU.NarrowUse->getOperand(0), WideType, IsSigned, Builder);
884 Value *RHS = (DU.NarrowUse->getOperand(1) == DU.NarrowDef) ? DU.WideDef :
885 getExtend(DU.NarrowUse->getOperand(1), WideType, IsSigned, Builder);
887 BinaryOperator *NarrowBO = cast<BinaryOperator>(DU.NarrowUse);
888 BinaryOperator *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(),
890 NarrowBO->getName());
891 Builder.Insert(WideBO);
892 if (const OverflowingBinaryOperator *OBO =
893 dyn_cast<OverflowingBinaryOperator>(NarrowBO)) {
894 if (OBO->hasNoUnsignedWrap()) WideBO->setHasNoUnsignedWrap();
895 if (OBO->hasNoSignedWrap()) WideBO->setHasNoSignedWrap();
902 /// HoistStep - Attempt to hoist an IV increment above a potential use.
904 /// To successfully hoist, two criteria must be met:
905 /// - IncV operands dominate InsertPos and
906 /// - InsertPos dominates IncV
908 /// Meeting the second condition means that we don't need to check all of IncV's
909 /// existing uses (it's moving up in the domtree).
911 /// This does not yet recursively hoist the operands, although that would
912 /// not be difficult.
913 static bool HoistStep(Instruction *IncV, Instruction *InsertPos,
914 const DominatorTree *DT)
916 if (DT->dominates(IncV, InsertPos))
919 if (!DT->dominates(InsertPos->getParent(), IncV->getParent()))
922 if (IncV->mayHaveSideEffects())
925 // Attempt to hoist IncV
926 for (User::op_iterator OI = IncV->op_begin(), OE = IncV->op_end();
928 Instruction *OInst = dyn_cast<Instruction>(OI);
929 if (OInst && !DT->dominates(OInst, InsertPos))
932 IncV->moveBefore(InsertPos);
936 /// No-wrap operations can transfer sign extension of their result to their
937 /// operands. Generate the SCEV value for the widened operation without
938 /// actually modifying the IR yet. If the expression after extending the
939 /// operands is an AddRec for this loop, return it.
940 const SCEVAddRecExpr* WidenIV::GetExtendedOperandRecurrence(NarrowIVDefUse DU) {
941 // Handle the common case of add<nsw/nuw>
942 if (DU.NarrowUse->getOpcode() != Instruction::Add)
945 // One operand (NarrowDef) has already been extended to WideDef. Now determine
946 // if extending the other will lead to a recurrence.
947 unsigned ExtendOperIdx = DU.NarrowUse->getOperand(0) == DU.NarrowDef ? 1 : 0;
948 assert(DU.NarrowUse->getOperand(1-ExtendOperIdx) == DU.NarrowDef && "bad DU");
950 const SCEV *ExtendOperExpr = 0;
951 const OverflowingBinaryOperator *OBO =
952 cast<OverflowingBinaryOperator>(DU.NarrowUse);
953 if (IsSigned && OBO->hasNoSignedWrap())
954 ExtendOperExpr = SE->getSignExtendExpr(
955 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
956 else if(!IsSigned && OBO->hasNoUnsignedWrap())
957 ExtendOperExpr = SE->getZeroExtendExpr(
958 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
962 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(
963 SE->getAddExpr(SE->getSCEV(DU.WideDef), ExtendOperExpr,
964 IsSigned ? SCEV::FlagNSW : SCEV::FlagNUW));
966 if (!AddRec || AddRec->getLoop() != L)
971 /// GetWideRecurrence - Is this instruction potentially interesting from
972 /// IVUsers' perspective after widening it's type? In other words, can the
973 /// extend be safely hoisted out of the loop with SCEV reducing the value to a
974 /// recurrence on the same loop. If so, return the sign or zero extended
975 /// recurrence. Otherwise return NULL.
976 const SCEVAddRecExpr *WidenIV::GetWideRecurrence(Instruction *NarrowUse) {
977 if (!SE->isSCEVable(NarrowUse->getType()))
980 const SCEV *NarrowExpr = SE->getSCEV(NarrowUse);
981 if (SE->getTypeSizeInBits(NarrowExpr->getType())
982 >= SE->getTypeSizeInBits(WideType)) {
983 // NarrowUse implicitly widens its operand. e.g. a gep with a narrow
984 // index. So don't follow this use.
988 const SCEV *WideExpr = IsSigned ?
989 SE->getSignExtendExpr(NarrowExpr, WideType) :
990 SE->getZeroExtendExpr(NarrowExpr, WideType);
991 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr);
992 if (!AddRec || AddRec->getLoop() != L)
997 /// WidenIVUse - Determine whether an individual user of the narrow IV can be
998 /// widened. If so, return the wide clone of the user.
999 Instruction *WidenIV::WidenIVUse(NarrowIVDefUse DU) {
1001 // Stop traversing the def-use chain at inner-loop phis or post-loop phis.
1002 if (isa<PHINode>(DU.NarrowUse) &&
1003 LI->getLoopFor(DU.NarrowUse->getParent()) != L)
1006 // Our raison d'etre! Eliminate sign and zero extension.
1007 if (IsSigned ? isa<SExtInst>(DU.NarrowUse) : isa<ZExtInst>(DU.NarrowUse)) {
1008 Value *NewDef = DU.WideDef;
1009 if (DU.NarrowUse->getType() != WideType) {
1010 unsigned CastWidth = SE->getTypeSizeInBits(DU.NarrowUse->getType());
1011 unsigned IVWidth = SE->getTypeSizeInBits(WideType);
1012 if (CastWidth < IVWidth) {
1013 // The cast isn't as wide as the IV, so insert a Trunc.
1014 IRBuilder<> Builder(DU.NarrowUse);
1015 NewDef = Builder.CreateTrunc(DU.WideDef, DU.NarrowUse->getType());
1018 // A wider extend was hidden behind a narrower one. This may induce
1019 // another round of IV widening in which the intermediate IV becomes
1020 // dead. It should be very rare.
1021 DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi
1022 << " not wide enough to subsume " << *DU.NarrowUse << "\n");
1023 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
1024 NewDef = DU.NarrowUse;
1027 if (NewDef != DU.NarrowUse) {
1028 DEBUG(dbgs() << "INDVARS: eliminating " << *DU.NarrowUse
1029 << " replaced by " << *DU.WideDef << "\n");
1031 DU.NarrowUse->replaceAllUsesWith(NewDef);
1032 DeadInsts.push_back(DU.NarrowUse);
1034 // Now that the extend is gone, we want to expose it's uses for potential
1035 // further simplification. We don't need to directly inform SimplifyIVUsers
1036 // of the new users, because their parent IV will be processed later as a
1037 // new loop phi. If we preserved IVUsers analysis, we would also want to
1038 // push the uses of WideDef here.
1040 // No further widening is needed. The deceased [sz]ext had done it for us.
1044 // Does this user itself evaluate to a recurrence after widening?
1045 const SCEVAddRecExpr *WideAddRec = GetWideRecurrence(DU.NarrowUse);
1047 WideAddRec = GetExtendedOperandRecurrence(DU);
1050 // This user does not evaluate to a recurence after widening, so don't
1051 // follow it. Instead insert a Trunc to kill off the original use,
1052 // eventually isolating the original narrow IV so it can be removed.
1053 IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT));
1054 Value *Trunc = Builder.CreateTrunc(DU.WideDef, DU.NarrowDef->getType());
1055 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, Trunc);
1058 // Assume block terminators cannot evaluate to a recurrence. We can't to
1059 // insert a Trunc after a terminator if there happens to be a critical edge.
1060 assert(DU.NarrowUse != DU.NarrowUse->getParent()->getTerminator() &&
1061 "SCEV is not expected to evaluate a block terminator");
1063 // Reuse the IV increment that SCEVExpander created as long as it dominates
1065 Instruction *WideUse = 0;
1066 if (WideAddRec == WideIncExpr && HoistStep(WideInc, DU.NarrowUse, DT)) {
1070 WideUse = CloneIVUser(DU);
1074 // Evaluation of WideAddRec ensured that the narrow expression could be
1075 // extended outside the loop without overflow. This suggests that the wide use
1076 // evaluates to the same expression as the extended narrow use, but doesn't
1077 // absolutely guarantee it. Hence the following failsafe check. In rare cases
1078 // where it fails, we simply throw away the newly created wide use.
1079 if (WideAddRec != SE->getSCEV(WideUse)) {
1080 DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse
1081 << ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n");
1082 DeadInsts.push_back(WideUse);
1086 // Returning WideUse pushes it on the worklist.
1090 /// pushNarrowIVUsers - Add eligible users of NarrowDef to NarrowIVUsers.
1092 void WidenIV::pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef) {
1093 for (Value::use_iterator UI = NarrowDef->use_begin(),
1094 UE = NarrowDef->use_end(); UI != UE; ++UI) {
1095 Instruction *NarrowUse = cast<Instruction>(*UI);
1097 // Handle data flow merges and bizarre phi cycles.
1098 if (!Widened.insert(NarrowUse))
1101 NarrowIVUsers.push_back(NarrowIVDefUse(NarrowDef, NarrowUse, WideDef));
1105 /// CreateWideIV - Process a single induction variable. First use the
1106 /// SCEVExpander to create a wide induction variable that evaluates to the same
1107 /// recurrence as the original narrow IV. Then use a worklist to forward
1108 /// traverse the narrow IV's def-use chain. After WidenIVUse has processed all
1109 /// interesting IV users, the narrow IV will be isolated for removal by
1112 /// It would be simpler to delete uses as they are processed, but we must avoid
1113 /// invalidating SCEV expressions.
1115 PHINode *WidenIV::CreateWideIV(SCEVExpander &Rewriter) {
1116 // Is this phi an induction variable?
1117 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi));
1121 // Widen the induction variable expression.
1122 const SCEV *WideIVExpr = IsSigned ?
1123 SE->getSignExtendExpr(AddRec, WideType) :
1124 SE->getZeroExtendExpr(AddRec, WideType);
1126 assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType &&
1127 "Expect the new IV expression to preserve its type");
1129 // Can the IV be extended outside the loop without overflow?
1130 AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr);
1131 if (!AddRec || AddRec->getLoop() != L)
1134 // An AddRec must have loop-invariant operands. Since this AddRec is
1135 // materialized by a loop header phi, the expression cannot have any post-loop
1136 // operands, so they must dominate the loop header.
1137 assert(SE->properlyDominates(AddRec->getStart(), L->getHeader()) &&
1138 SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader())
1139 && "Loop header phi recurrence inputs do not dominate the loop");
1141 // The rewriter provides a value for the desired IV expression. This may
1142 // either find an existing phi or materialize a new one. Either way, we
1143 // expect a well-formed cyclic phi-with-increments. i.e. any operand not part
1144 // of the phi-SCC dominates the loop entry.
1145 Instruction *InsertPt = L->getHeader()->begin();
1146 WidePhi = cast<PHINode>(Rewriter.expandCodeFor(AddRec, WideType, InsertPt));
1148 // Remembering the WideIV increment generated by SCEVExpander allows
1149 // WidenIVUse to reuse it when widening the narrow IV's increment. We don't
1150 // employ a general reuse mechanism because the call above is the only call to
1151 // SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses.
1152 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1154 cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock));
1155 WideIncExpr = SE->getSCEV(WideInc);
1158 DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n");
1161 // Traverse the def-use chain using a worklist starting at the original IV.
1162 assert(Widened.empty() && NarrowIVUsers.empty() && "expect initial state" );
1164 Widened.insert(OrigPhi);
1165 pushNarrowIVUsers(OrigPhi, WidePhi);
1167 while (!NarrowIVUsers.empty()) {
1168 NarrowIVDefUse DU = NarrowIVUsers.pop_back_val();
1170 // Process a def-use edge. This may replace the use, so don't hold a
1171 // use_iterator across it.
1172 Instruction *WideUse = WidenIVUse(DU);
1174 // Follow all def-use edges from the previous narrow use.
1176 pushNarrowIVUsers(DU.NarrowUse, WideUse);
1178 // WidenIVUse may have removed the def-use edge.
1179 if (DU.NarrowDef->use_empty())
1180 DeadInsts.push_back(DU.NarrowDef);
1185 //===----------------------------------------------------------------------===//
1186 // Simplification of IV users based on SCEV evaluation.
1187 //===----------------------------------------------------------------------===//
1190 /// SimplifyAndExtend - Iteratively perform simplification on a worklist of IV
1191 /// users. Each successive simplification may push more users which may
1192 /// themselves be candidates for simplification.
1194 /// Sign/Zero extend elimination is interleaved with IV simplification.
1196 void IndVarSimplify::SimplifyAndExtend(Loop *L,
1197 SCEVExpander &Rewriter,
1198 LPPassManager &LPM) {
1199 std::map<PHINode *, WideIVInfo> WideIVMap;
1201 SmallVector<PHINode*, 8> LoopPhis;
1202 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1203 LoopPhis.push_back(cast<PHINode>(I));
1205 // Each round of simplification iterates through the SimplifyIVUsers worklist
1206 // for all current phis, then determines whether any IVs can be
1207 // widened. Widening adds new phis to LoopPhis, inducing another round of
1208 // simplification on the wide IVs.
1209 while (!LoopPhis.empty()) {
1210 // Evaluate as many IV expressions as possible before widening any IVs. This
1211 // forces SCEV to set no-wrap flags before evaluating sign/zero
1212 // extension. The first time SCEV attempts to normalize sign/zero extension,
1213 // the result becomes final. So for the most predictable results, we delay
1214 // evaluation of sign/zero extend evaluation until needed, and avoid running
1215 // other SCEV based analysis prior to SimplifyAndExtend.
1217 PHINode *CurrIV = LoopPhis.pop_back_val();
1219 // Information about sign/zero extensions of CurrIV.
1220 WideIVVisitor WIV(SE, TD);
1222 Changed |= simplifyUsersOfIV(CurrIV, SE, &LPM, DeadInsts, &WIV);
1224 if (WIV.WI.WidestNativeType) {
1225 WideIVMap[CurrIV] = WIV.WI;
1227 } while(!LoopPhis.empty());
1229 for (std::map<PHINode *, WideIVInfo>::const_iterator I = WideIVMap.begin(),
1230 E = WideIVMap.end(); I != E; ++I) {
1231 WidenIV Widener(I->first, I->second, LI, SE, DT, DeadInsts);
1232 if (PHINode *WidePhi = Widener.CreateWideIV(Rewriter)) {
1234 LoopPhis.push_back(WidePhi);
1241 /// SimplifyCongruentIVs - Check for congruent phis in this loop header and
1242 /// replace them with their chosen representative.
1244 void IndVarSimplify::SimplifyCongruentIVs(Loop *L) {
1245 DenseMap<const SCEV *, PHINode *> ExprToIVMap;
1246 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1247 PHINode *Phi = cast<PHINode>(I);
1248 if (!SE->isSCEVable(Phi->getType()))
1251 const SCEV *S = SE->getSCEV(Phi);
1252 std::pair<DenseMap<const SCEV *, PHINode *>::const_iterator, bool> Tmp =
1253 ExprToIVMap.insert(std::make_pair(S, Phi));
1256 PHINode *OrigPhi = Tmp.first->second;
1258 // If one phi derives from the other via GEPs, types may differ.
1259 if (OrigPhi->getType() != Phi->getType())
1262 // Replacing the congruent phi is sufficient because acyclic redundancy
1263 // elimination, CSE/GVN, should handle the rest. However, once SCEV proves
1264 // that a phi is congruent, it's almost certain to be the head of an IV
1265 // user cycle that is isomorphic with the original phi. So it's worth
1266 // eagerly cleaning up the common case of a single IV increment.
1267 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1268 Instruction *OrigInc =
1269 cast<Instruction>(OrigPhi->getIncomingValueForBlock(LatchBlock));
1270 Instruction *IsomorphicInc =
1271 cast<Instruction>(Phi->getIncomingValueForBlock(LatchBlock));
1272 if (OrigInc != IsomorphicInc &&
1273 OrigInc->getType() == IsomorphicInc->getType() &&
1274 SE->getSCEV(OrigInc) == SE->getSCEV(IsomorphicInc) &&
1275 HoistStep(OrigInc, IsomorphicInc, DT)) {
1276 DEBUG(dbgs() << "INDVARS: Eliminated congruent iv.inc: "
1277 << *IsomorphicInc << '\n');
1278 IsomorphicInc->replaceAllUsesWith(OrigInc);
1279 DeadInsts.push_back(IsomorphicInc);
1282 DEBUG(dbgs() << "INDVARS: Eliminated congruent iv: " << *Phi << '\n');
1284 Phi->replaceAllUsesWith(OrigPhi);
1285 DeadInsts.push_back(Phi);
1289 //===----------------------------------------------------------------------===//
1290 // LinearFunctionTestReplace and its kin. Rewrite the loop exit condition.
1291 //===----------------------------------------------------------------------===//
1293 /// Check for expressions that ScalarEvolution generates to compute
1294 /// BackedgeTakenInfo. If these expressions have not been reduced, then
1295 /// expanding them may incur additional cost (albeit in the loop preheader).
1296 static bool isHighCostExpansion(const SCEV *S, BranchInst *BI,
1297 ScalarEvolution *SE) {
1298 // If the backedge-taken count is a UDiv, it's very likely a UDiv that
1299 // ScalarEvolution's HowFarToZero or HowManyLessThans produced to compute a
1300 // precise expression, rather than a UDiv from the user's code. If we can't
1301 // find a UDiv in the code with some simple searching, assume the former and
1302 // forego rewriting the loop.
1303 if (isa<SCEVUDivExpr>(S)) {
1304 ICmpInst *OrigCond = dyn_cast<ICmpInst>(BI->getCondition());
1305 if (!OrigCond) return true;
1306 const SCEV *R = SE->getSCEV(OrigCond->getOperand(1));
1307 R = SE->getMinusSCEV(R, SE->getConstant(R->getType(), 1));
1309 const SCEV *L = SE->getSCEV(OrigCond->getOperand(0));
1310 L = SE->getMinusSCEV(L, SE->getConstant(L->getType(), 1));
1316 if (EnableIVRewrite)
1319 // Recurse past add expressions, which commonly occur in the
1320 // BackedgeTakenCount. They may already exist in program code, and if not,
1321 // they are not too expensive rematerialize.
1322 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
1323 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
1325 if (isHighCostExpansion(*I, BI, SE))
1331 // HowManyLessThans uses a Max expression whenever the loop is not guarded by
1332 // the exit condition.
1333 if (isa<SCEVSMaxExpr>(S) || isa<SCEVUMaxExpr>(S))
1336 // If we haven't recognized an expensive SCEV patter, assume its an expression
1337 // produced by program code.
1341 /// canExpandBackedgeTakenCount - Return true if this loop's backedge taken
1342 /// count expression can be safely and cheaply expanded into an instruction
1343 /// sequence that can be used by LinearFunctionTestReplace.
1344 static bool canExpandBackedgeTakenCount(Loop *L, ScalarEvolution *SE) {
1345 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1346 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) ||
1347 BackedgeTakenCount->isZero())
1350 if (!L->getExitingBlock())
1353 // Can't rewrite non-branch yet.
1354 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
1358 if (isHighCostExpansion(BackedgeTakenCount, BI, SE))
1364 /// getBackedgeIVType - Get the widest type used by the loop test after peeking
1367 /// TODO: Unnecessary when ForceLFTR is removed.
1368 static Type *getBackedgeIVType(Loop *L) {
1369 if (!L->getExitingBlock())
1372 // Can't rewrite non-branch yet.
1373 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
1377 ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition());
1382 for(User::op_iterator OI = Cond->op_begin(), OE = Cond->op_end();
1384 assert((!Ty || Ty == (*OI)->getType()) && "bad icmp operand types");
1385 TruncInst *Trunc = dyn_cast<TruncInst>(*OI);
1389 return Trunc->getSrcTy();
1394 /// isLoopInvariant - Perform a quick domtree based check for loop invariance
1395 /// assuming that V is used within the loop. LoopInfo::isLoopInvariant() seems
1396 /// gratuitous for this purpose.
1397 static bool isLoopInvariant(Value *V, Loop *L, DominatorTree *DT) {
1398 Instruction *Inst = dyn_cast<Instruction>(V);
1402 return DT->properlyDominates(Inst->getParent(), L->getHeader());
1405 /// getLoopPhiForCounter - Return the loop header phi IFF IncV adds a loop
1406 /// invariant value to the phi.
1407 static PHINode *getLoopPhiForCounter(Value *IncV, Loop *L, DominatorTree *DT) {
1408 Instruction *IncI = dyn_cast<Instruction>(IncV);
1412 switch (IncI->getOpcode()) {
1413 case Instruction::Add:
1414 case Instruction::Sub:
1416 case Instruction::GetElementPtr:
1417 // An IV counter must preserve its type.
1418 if (IncI->getNumOperands() == 2)
1424 PHINode *Phi = dyn_cast<PHINode>(IncI->getOperand(0));
1425 if (Phi && Phi->getParent() == L->getHeader()) {
1426 if (isLoopInvariant(IncI->getOperand(1), L, DT))
1430 if (IncI->getOpcode() == Instruction::GetElementPtr)
1433 // Allow add/sub to be commuted.
1434 Phi = dyn_cast<PHINode>(IncI->getOperand(1));
1435 if (Phi && Phi->getParent() == L->getHeader()) {
1436 if (isLoopInvariant(IncI->getOperand(0), L, DT))
1442 /// needsLFTR - LinearFunctionTestReplace policy. Return true unless we can show
1443 /// that the current exit test is already sufficiently canonical.
1444 static bool needsLFTR(Loop *L, DominatorTree *DT) {
1445 assert(L->getExitingBlock() && "expected loop exit");
1447 BasicBlock *LatchBlock = L->getLoopLatch();
1448 // Don't bother with LFTR if the loop is not properly simplified.
1452 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
1453 assert(BI && "expected exit branch");
1455 // Do LFTR to simplify the exit condition to an ICMP.
1456 ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition());
1460 // Do LFTR to simplify the exit ICMP to EQ/NE
1461 ICmpInst::Predicate Pred = Cond->getPredicate();
1462 if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ)
1465 // Look for a loop invariant RHS
1466 Value *LHS = Cond->getOperand(0);
1467 Value *RHS = Cond->getOperand(1);
1468 if (!isLoopInvariant(RHS, L, DT)) {
1469 if (!isLoopInvariant(LHS, L, DT))
1471 std::swap(LHS, RHS);
1473 // Look for a simple IV counter LHS
1474 PHINode *Phi = dyn_cast<PHINode>(LHS);
1476 Phi = getLoopPhiForCounter(LHS, L, DT);
1481 // Do LFTR if the exit condition's IV is *not* a simple counter.
1482 Value *IncV = Phi->getIncomingValueForBlock(L->getLoopLatch());
1483 return Phi != getLoopPhiForCounter(IncV, L, DT);
1486 /// AlmostDeadIV - Return true if this IV has any uses other than the (soon to
1487 /// be rewritten) loop exit test.
1488 static bool AlmostDeadIV(PHINode *Phi, BasicBlock *LatchBlock, Value *Cond) {
1489 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1490 Value *IncV = Phi->getIncomingValue(LatchIdx);
1492 for (Value::use_iterator UI = Phi->use_begin(), UE = Phi->use_end();
1494 if (*UI != Cond && *UI != IncV) return false;
1497 for (Value::use_iterator UI = IncV->use_begin(), UE = IncV->use_end();
1499 if (*UI != Cond && *UI != Phi) return false;
1504 /// FindLoopCounter - Find an affine IV in canonical form.
1506 /// FIXME: Accept -1 stride and set IVLimit = IVInit - BECount
1508 /// FIXME: Accept non-unit stride as long as SCEV can reduce BECount * Stride.
1509 /// This is difficult in general for SCEV because of potential overflow. But we
1510 /// could at least handle constant BECounts.
1512 FindLoopCounter(Loop *L, const SCEV *BECount,
1513 ScalarEvolution *SE, DominatorTree *DT, const TargetData *TD) {
1514 // I'm not sure how BECount could be a pointer type, but we definitely don't
1515 // want to LFTR that.
1516 if (BECount->getType()->isPointerTy())
1519 uint64_t BCWidth = SE->getTypeSizeInBits(BECount->getType());
1522 cast<BranchInst>(L->getExitingBlock()->getTerminator())->getCondition();
1524 // Loop over all of the PHI nodes, looking for a simple counter.
1525 PHINode *BestPhi = 0;
1526 const SCEV *BestInit = 0;
1527 BasicBlock *LatchBlock = L->getLoopLatch();
1528 assert(LatchBlock && "needsLFTR should guarantee a loop latch");
1530 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1531 PHINode *Phi = cast<PHINode>(I);
1532 if (!SE->isSCEVable(Phi->getType()))
1535 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Phi));
1536 if (!AR || AR->getLoop() != L || !AR->isAffine())
1539 // AR may be a pointer type, while BECount is an integer type.
1540 // AR may be wider than BECount. With eq/ne tests overflow is immaterial.
1541 // AR may not be a narrower type, or we may never exit.
1542 uint64_t PhiWidth = SE->getTypeSizeInBits(AR->getType());
1543 if (PhiWidth < BCWidth || (TD && !TD->isLegalInteger(PhiWidth)))
1546 const SCEV *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
1547 if (!Step || !Step->isOne())
1550 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1551 Value *IncV = Phi->getIncomingValue(LatchIdx);
1552 if (getLoopPhiForCounter(IncV, L, DT) != Phi)
1555 const SCEV *Init = AR->getStart();
1557 if (BestPhi && !AlmostDeadIV(BestPhi, LatchBlock, Cond)) {
1558 // Don't force a live loop counter if another IV can be used.
1559 if (AlmostDeadIV(Phi, LatchBlock, Cond))
1562 // Prefer to count-from-zero. This is a more "canonical" counter form. It
1563 // also prefers integer to pointer IVs.
1564 if (BestInit->isZero() != Init->isZero()) {
1565 if (BestInit->isZero())
1568 // If two IVs both count from zero or both count from nonzero then the
1569 // narrower is likely a dead phi that has been widened. Use the wider phi
1570 // to allow the other to be eliminated.
1571 if (PhiWidth <= SE->getTypeSizeInBits(BestPhi->getType()))
1580 /// LinearFunctionTestReplace - This method rewrites the exit condition of the
1581 /// loop to be a canonical != comparison against the incremented loop induction
1582 /// variable. This pass is able to rewrite the exit tests of any loop where the
1583 /// SCEV analysis can determine a loop-invariant trip count of the loop, which
1584 /// is actually a much broader range than just linear tests.
1585 Value *IndVarSimplify::
1586 LinearFunctionTestReplace(Loop *L,
1587 const SCEV *BackedgeTakenCount,
1589 SCEVExpander &Rewriter) {
1590 assert(canExpandBackedgeTakenCount(L, SE) && "precondition");
1591 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1593 // LFTR can ignore IV overflow and truncate to the width of
1594 // BECount. This avoids materializing the add(zext(add)) expression.
1595 Type *CntTy = !EnableIVRewrite ?
1596 BackedgeTakenCount->getType() : IndVar->getType();
1598 const SCEV *IVLimit = BackedgeTakenCount;
1600 // If the exiting block is not the same as the backedge block, we must compare
1601 // against the preincremented value, otherwise we prefer to compare against
1602 // the post-incremented value.
1604 if (L->getExitingBlock() == L->getLoopLatch()) {
1605 // Add one to the "backedge-taken" count to get the trip count.
1606 // If this addition may overflow, we have to be more pessimistic and
1607 // cast the induction variable before doing the add.
1609 SE->getAddExpr(IVLimit, SE->getConstant(IVLimit->getType(), 1));
1610 if (CntTy == IVLimit->getType())
1613 const SCEV *Zero = SE->getConstant(IVLimit->getType(), 0);
1614 if ((isa<SCEVConstant>(N) && !N->isZero()) ||
1615 SE->isLoopEntryGuardedByCond(L, ICmpInst::ICMP_NE, N, Zero)) {
1616 // No overflow. Cast the sum.
1617 IVLimit = SE->getTruncateOrZeroExtend(N, CntTy);
1619 // Potential overflow. Cast before doing the add.
1620 IVLimit = SE->getTruncateOrZeroExtend(IVLimit, CntTy);
1621 IVLimit = SE->getAddExpr(IVLimit, SE->getConstant(CntTy, 1));
1624 // The BackedgeTaken expression contains the number of times that the
1625 // backedge branches to the loop header. This is one less than the
1626 // number of times the loop executes, so use the incremented indvar.
1627 CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock());
1629 // We have to use the preincremented value...
1630 IVLimit = SE->getTruncateOrZeroExtend(IVLimit, CntTy);
1634 // For unit stride, IVLimit = Start + BECount with 2's complement overflow.
1635 // So for, non-zero start compute the IVLimit here.
1636 bool isPtrIV = false;
1637 Type *CmpTy = CntTy;
1638 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
1639 assert(AR && AR->getLoop() == L && AR->isAffine() && "bad loop counter");
1640 if (!AR->getStart()->isZero()) {
1641 assert(AR->getStepRecurrence(*SE)->isOne() && "only handles unit stride");
1642 const SCEV *IVInit = AR->getStart();
1644 // For pointer types, sign extend BECount in order to materialize a GEP.
1645 // Note that for without EnableIVRewrite, we never run SCEVExpander on a
1646 // pointer type, because we must preserve the existing GEPs. Instead we
1647 // directly generate a GEP later.
1648 if (IVInit->getType()->isPointerTy()) {
1650 CmpTy = SE->getEffectiveSCEVType(IVInit->getType());
1651 IVLimit = SE->getTruncateOrSignExtend(IVLimit, CmpTy);
1653 // For integer types, truncate the IV before computing IVInit + BECount.
1655 if (SE->getTypeSizeInBits(IVInit->getType())
1656 > SE->getTypeSizeInBits(CmpTy))
1657 IVInit = SE->getTruncateExpr(IVInit, CmpTy);
1659 IVLimit = SE->getAddExpr(IVInit, IVLimit);
1662 // Expand the code for the iteration count.
1663 IRBuilder<> Builder(BI);
1665 assert(SE->isLoopInvariant(IVLimit, L) &&
1666 "Computed iteration count is not loop invariant!");
1667 Value *ExitCnt = Rewriter.expandCodeFor(IVLimit, CmpTy, BI);
1669 // Create a gep for IVInit + IVLimit from on an existing pointer base.
1670 assert(isPtrIV == IndVar->getType()->isPointerTy() &&
1671 "IndVar type must match IVInit type");
1673 Value *IVStart = IndVar->getIncomingValueForBlock(L->getLoopPreheader());
1674 assert(AR->getStart() == SE->getSCEV(IVStart) && "bad loop counter");
1675 assert(SE->getSizeOfExpr(
1676 cast<PointerType>(IVStart->getType())->getElementType())->isOne()
1677 && "unit stride pointer IV must be i8*");
1679 Builder.SetInsertPoint(L->getLoopPreheader()->getTerminator());
1680 ExitCnt = Builder.CreateGEP(IVStart, ExitCnt, "lftr.limit");
1681 Builder.SetInsertPoint(BI);
1684 // Insert a new icmp_ne or icmp_eq instruction before the branch.
1685 ICmpInst::Predicate P;
1686 if (L->contains(BI->getSuccessor(0)))
1687 P = ICmpInst::ICMP_NE;
1689 P = ICmpInst::ICMP_EQ;
1691 DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
1692 << " LHS:" << *CmpIndVar << '\n'
1694 << (P == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
1695 << " RHS:\t" << *ExitCnt << "\n"
1696 << " Expr:\t" << *IVLimit << "\n");
1698 if (SE->getTypeSizeInBits(CmpIndVar->getType())
1699 > SE->getTypeSizeInBits(CmpTy)) {
1700 CmpIndVar = Builder.CreateTrunc(CmpIndVar, CmpTy, "lftr.wideiv");
1703 Value *Cond = Builder.CreateICmp(P, CmpIndVar, ExitCnt, "exitcond");
1704 Value *OrigCond = BI->getCondition();
1705 // It's tempting to use replaceAllUsesWith here to fully replace the old
1706 // comparison, but that's not immediately safe, since users of the old
1707 // comparison may not be dominated by the new comparison. Instead, just
1708 // update the branch to use the new comparison; in the common case this
1709 // will make old comparison dead.
1710 BI->setCondition(Cond);
1711 DeadInsts.push_back(OrigCond);
1718 //===----------------------------------------------------------------------===//
1719 // SinkUnusedInvariants. A late subpass to cleanup loop preheaders.
1720 //===----------------------------------------------------------------------===//
1722 /// If there's a single exit block, sink any loop-invariant values that
1723 /// were defined in the preheader but not used inside the loop into the
1724 /// exit block to reduce register pressure in the loop.
1725 void IndVarSimplify::SinkUnusedInvariants(Loop *L) {
1726 BasicBlock *ExitBlock = L->getExitBlock();
1727 if (!ExitBlock) return;
1729 BasicBlock *Preheader = L->getLoopPreheader();
1730 if (!Preheader) return;
1732 Instruction *InsertPt = ExitBlock->getFirstInsertionPt();
1733 BasicBlock::iterator I = Preheader->getTerminator();
1734 while (I != Preheader->begin()) {
1736 // New instructions were inserted at the end of the preheader.
1737 if (isa<PHINode>(I))
1740 // Don't move instructions which might have side effects, since the side
1741 // effects need to complete before instructions inside the loop. Also don't
1742 // move instructions which might read memory, since the loop may modify
1743 // memory. Note that it's okay if the instruction might have undefined
1744 // behavior: LoopSimplify guarantees that the preheader dominates the exit
1746 if (I->mayHaveSideEffects() || I->mayReadFromMemory())
1749 // Skip debug info intrinsics.
1750 if (isa<DbgInfoIntrinsic>(I))
1753 // Skip landingpad instructions.
1754 if (isa<LandingPadInst>(I))
1757 // Don't sink static AllocaInsts out of the entry block, which would
1758 // turn them into dynamic allocas!
1759 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
1760 if (AI->isStaticAlloca())
1763 // Determine if there is a use in or before the loop (direct or
1765 bool UsedInLoop = false;
1766 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
1769 BasicBlock *UseBB = cast<Instruction>(U)->getParent();
1770 if (PHINode *P = dyn_cast<PHINode>(U)) {
1772 PHINode::getIncomingValueNumForOperand(UI.getOperandNo());
1773 UseBB = P->getIncomingBlock(i);
1775 if (UseBB == Preheader || L->contains(UseBB)) {
1781 // If there is, the def must remain in the preheader.
1785 // Otherwise, sink it to the exit block.
1786 Instruction *ToMove = I;
1789 if (I != Preheader->begin()) {
1790 // Skip debug info intrinsics.
1793 } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin());
1795 if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin())
1801 ToMove->moveBefore(InsertPt);
1807 //===----------------------------------------------------------------------===//
1808 // IndVarSimplify driver. Manage several subpasses of IV simplification.
1809 //===----------------------------------------------------------------------===//
1811 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
1812 // If LoopSimplify form is not available, stay out of trouble. Some notes:
1813 // - LSR currently only supports LoopSimplify-form loops. Indvars'
1814 // canonicalization can be a pessimization without LSR to "clean up"
1816 // - We depend on having a preheader; in particular,
1817 // Loop::getCanonicalInductionVariable only supports loops with preheaders,
1818 // and we're in trouble if we can't find the induction variable even when
1819 // we've manually inserted one.
1820 if (!L->isLoopSimplifyForm())
1823 if (EnableIVRewrite)
1824 IU = &getAnalysis<IVUsers>();
1825 LI = &getAnalysis<LoopInfo>();
1826 SE = &getAnalysis<ScalarEvolution>();
1827 DT = &getAnalysis<DominatorTree>();
1828 TD = getAnalysisIfAvailable<TargetData>();
1833 // If there are any floating-point recurrences, attempt to
1834 // transform them to use integer recurrences.
1835 RewriteNonIntegerIVs(L);
1837 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1839 // Create a rewriter object which we'll use to transform the code with.
1840 SCEVExpander Rewriter(*SE, "indvars");
1842 // Eliminate redundant IV users.
1844 // Simplification works best when run before other consumers of SCEV. We
1845 // attempt to avoid evaluating SCEVs for sign/zero extend operations until
1846 // other expressions involving loop IVs have been evaluated. This helps SCEV
1847 // set no-wrap flags before normalizing sign/zero extension.
1848 if (!EnableIVRewrite) {
1849 Rewriter.disableCanonicalMode();
1850 SimplifyAndExtend(L, Rewriter, LPM);
1853 // Check to see if this loop has a computable loop-invariant execution count.
1854 // If so, this means that we can compute the final value of any expressions
1855 // that are recurrent in the loop, and substitute the exit values from the
1856 // loop into any instructions outside of the loop that use the final values of
1857 // the current expressions.
1859 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1860 RewriteLoopExitValues(L, Rewriter);
1862 // Eliminate redundant IV users.
1863 if (EnableIVRewrite)
1864 Changed |= simplifyIVUsers(IU, SE, &LPM, DeadInsts);
1866 // Eliminate redundant IV cycles.
1867 if (!EnableIVRewrite)
1868 SimplifyCongruentIVs(L);
1870 // Compute the type of the largest recurrence expression, and decide whether
1871 // a canonical induction variable should be inserted.
1872 Type *LargestType = 0;
1873 bool NeedCannIV = false;
1874 bool ExpandBECount = canExpandBackedgeTakenCount(L, SE);
1875 if (EnableIVRewrite && ExpandBECount) {
1876 // If we have a known trip count and a single exit block, we'll be
1877 // rewriting the loop exit test condition below, which requires a
1878 // canonical induction variable.
1880 Type *Ty = BackedgeTakenCount->getType();
1881 if (!EnableIVRewrite) {
1882 // In this mode, SimplifyIVUsers may have already widened the IV used by
1883 // the backedge test and inserted a Trunc on the compare's operand. Get
1884 // the wider type to avoid creating a redundant narrow IV only used by the
1886 LargestType = getBackedgeIVType(L);
1889 SE->getTypeSizeInBits(Ty) >
1890 SE->getTypeSizeInBits(LargestType))
1891 LargestType = SE->getEffectiveSCEVType(Ty);
1893 if (EnableIVRewrite) {
1894 for (IVUsers::const_iterator I = IU->begin(), E = IU->end(); I != E; ++I) {
1897 SE->getEffectiveSCEVType(I->getOperandValToReplace()->getType());
1899 SE->getTypeSizeInBits(Ty) >
1900 SE->getTypeSizeInBits(LargestType))
1905 // Now that we know the largest of the induction variable expressions
1906 // in this loop, insert a canonical induction variable of the largest size.
1907 PHINode *IndVar = 0;
1909 // Check to see if the loop already has any canonical-looking induction
1910 // variables. If any are present and wider than the planned canonical
1911 // induction variable, temporarily remove them, so that the Rewriter
1912 // doesn't attempt to reuse them.
1913 SmallVector<PHINode *, 2> OldCannIVs;
1914 while (PHINode *OldCannIV = L->getCanonicalInductionVariable()) {
1915 if (SE->getTypeSizeInBits(OldCannIV->getType()) >
1916 SE->getTypeSizeInBits(LargestType))
1917 OldCannIV->removeFromParent();
1920 OldCannIVs.push_back(OldCannIV);
1923 IndVar = Rewriter.getOrInsertCanonicalInductionVariable(L, LargestType);
1927 DEBUG(dbgs() << "INDVARS: New CanIV: " << *IndVar << '\n');
1929 // Now that the official induction variable is established, reinsert
1930 // any old canonical-looking variables after it so that the IR remains
1931 // consistent. They will be deleted as part of the dead-PHI deletion at
1932 // the end of the pass.
1933 while (!OldCannIVs.empty()) {
1934 PHINode *OldCannIV = OldCannIVs.pop_back_val();
1935 OldCannIV->insertBefore(L->getHeader()->getFirstInsertionPt());
1938 else if (!EnableIVRewrite && ExpandBECount && needsLFTR(L, DT)) {
1939 IndVar = FindLoopCounter(L, BackedgeTakenCount, SE, DT, TD);
1941 // If we have a trip count expression, rewrite the loop's exit condition
1942 // using it. We can currently only handle loops with a single exit.
1944 if (ExpandBECount && IndVar) {
1945 // Check preconditions for proper SCEVExpander operation. SCEV does not
1946 // express SCEVExpander's dependencies, such as LoopSimplify. Instead any
1947 // pass that uses the SCEVExpander must do it. This does not work well for
1948 // loop passes because SCEVExpander makes assumptions about all loops, while
1949 // LoopPassManager only forces the current loop to be simplified.
1951 // FIXME: SCEV expansion has no way to bail out, so the caller must
1952 // explicitly check any assumptions made by SCEV. Brittle.
1953 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(BackedgeTakenCount);
1954 if (!AR || AR->getLoop()->getLoopPreheader())
1956 LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar, Rewriter);
1958 // Rewrite IV-derived expressions.
1959 if (EnableIVRewrite)
1960 RewriteIVExpressions(L, Rewriter);
1962 // Clear the rewriter cache, because values that are in the rewriter's cache
1963 // can be deleted in the loop below, causing the AssertingVH in the cache to
1967 // Now that we're done iterating through lists, clean up any instructions
1968 // which are now dead.
1969 while (!DeadInsts.empty())
1970 if (Instruction *Inst =
1971 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
1972 RecursivelyDeleteTriviallyDeadInstructions(Inst);
1974 // The Rewriter may not be used from this point on.
1976 // Loop-invariant instructions in the preheader that aren't used in the
1977 // loop may be sunk below the loop to reduce register pressure.
1978 SinkUnusedInvariants(L);
1980 // For completeness, inform IVUsers of the IV use in the newly-created
1981 // loop exit test instruction.
1982 if (IU && NewICmp) {
1983 ICmpInst *NewICmpInst = dyn_cast<ICmpInst>(NewICmp);
1985 IU->AddUsersIfInteresting(cast<Instruction>(NewICmpInst->getOperand(0)));
1987 // Clean up dead instructions.
1988 Changed |= DeleteDeadPHIs(L->getHeader());
1989 // Check a post-condition.
1990 assert(L->isLCSSAForm(*DT) &&
1991 "Indvars did not leave the loop in lcssa form!");
1993 // Verify that LFTR, and any other change have not interfered with SCEV's
1994 // ability to compute trip count.
1996 if (!EnableIVRewrite && VerifyIndvars &&
1997 !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) {
1999 const SCEV *NewBECount = SE->getBackedgeTakenCount(L);
2000 if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) <
2001 SE->getTypeSizeInBits(NewBECount->getType()))
2002 NewBECount = SE->getTruncateOrNoop(NewBECount,
2003 BackedgeTakenCount->getType());
2005 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount,
2006 NewBECount->getType());
2007 assert(BackedgeTakenCount == NewBECount && "indvars must preserve SCEV");