1 //===- IndVarSimplify.cpp - Induction Variable Elimination ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation analyzes and transforms the induction variables (and
11 // computations derived from them) into simpler forms suitable for subsequent
12 // analysis and transformation.
14 // This transformation makes the following changes to each loop with an
15 // identifiable induction variable:
16 // 1. All loops are transformed to have a SINGLE canonical induction variable
17 // which starts at zero and steps by one.
18 // 2. The canonical induction variable is guaranteed to be the first PHI node
19 // in the loop header block.
20 // 3. The canonical induction variable is guaranteed to be in a wide enough
21 // type so that IV expressions need not be (directly) zero-extended or
23 // 4. Any pointer arithmetic recurrences are raised to use array subscripts.
25 // If the trip count of a loop is computable, this pass also makes the following
27 // 1. The exit condition for the loop is canonicalized to compare the
28 // induction value against the exit value. This turns loops like:
29 // 'for (i = 7; i*i < 1000; ++i)' into 'for (i = 0; i != 25; ++i)'
30 // 2. Any use outside of the loop of an expression derived from the indvar
31 // is changed to compute the derived value outside of the loop, eliminating
32 // the dependence on the exit value of the induction variable. If the only
33 // purpose of the loop is to compute the exit value of some derived
34 // expression, this transformation will make the loop dead.
36 // This transformation should be followed by strength reduction after all of the
37 // desired loop transformations have been performed.
39 //===----------------------------------------------------------------------===//
41 #define DEBUG_TYPE "indvars"
42 #include "llvm/Transforms/Scalar.h"
43 #include "llvm/BasicBlock.h"
44 #include "llvm/Constants.h"
45 #include "llvm/Instructions.h"
46 #include "llvm/IntrinsicInst.h"
47 #include "llvm/LLVMContext.h"
48 #include "llvm/Type.h"
49 #include "llvm/Analysis/Dominators.h"
50 #include "llvm/Analysis/IVUsers.h"
51 #include "llvm/Analysis/ScalarEvolutionExpander.h"
52 #include "llvm/Analysis/LoopInfo.h"
53 #include "llvm/Analysis/LoopPass.h"
54 #include "llvm/Support/CFG.h"
55 #include "llvm/Support/CommandLine.h"
56 #include "llvm/Support/Debug.h"
57 #include "llvm/Support/raw_ostream.h"
58 #include "llvm/Transforms/Utils/Local.h"
59 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
60 #include "llvm/Target/TargetData.h"
61 #include "llvm/ADT/DenseMap.h"
62 #include "llvm/ADT/SmallVector.h"
63 #include "llvm/ADT/Statistic.h"
64 #include "llvm/ADT/STLExtras.h"
67 STATISTIC(NumRemoved , "Number of aux indvars removed");
68 STATISTIC(NumWidened , "Number of indvars widened");
69 STATISTIC(NumInserted , "Number of canonical indvars added");
70 STATISTIC(NumReplaced , "Number of exit values replaced");
71 STATISTIC(NumLFTR , "Number of loop exit tests replaced");
72 STATISTIC(NumElimIdentity, "Number of IV identities eliminated");
73 STATISTIC(NumElimExt , "Number of IV sign/zero extends eliminated");
74 STATISTIC(NumElimRem , "Number of IV remainder operations eliminated");
75 STATISTIC(NumElimCmp , "Number of IV comparisons eliminated");
76 STATISTIC(NumElimIV , "Number of congruent IVs eliminated");
78 static cl::opt<bool> DisableIVRewrite(
79 "disable-iv-rewrite", cl::Hidden,
80 cl::desc("Disable canonical induction variable rewriting"));
82 // Temporary flag for use with -disable-iv-rewrite to force a canonical IV for
84 static cl::opt<bool> ForceLFTR(
85 "force-lftr", cl::Hidden,
86 cl::desc("Enable forced linear function test replacement"));
89 class IndVarSimplify : public LoopPass {
96 SmallVector<WeakVH, 16> DeadInsts;
100 static char ID; // Pass identification, replacement for typeid
101 IndVarSimplify() : LoopPass(ID), IU(0), LI(0), SE(0), DT(0), TD(0),
103 initializeIndVarSimplifyPass(*PassRegistry::getPassRegistry());
106 virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
108 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
109 AU.addRequired<DominatorTree>();
110 AU.addRequired<LoopInfo>();
111 AU.addRequired<ScalarEvolution>();
112 AU.addRequiredID(LoopSimplifyID);
113 AU.addRequiredID(LCSSAID);
114 if (!DisableIVRewrite)
115 AU.addRequired<IVUsers>();
116 AU.addPreserved<ScalarEvolution>();
117 AU.addPreservedID(LoopSimplifyID);
118 AU.addPreservedID(LCSSAID);
119 if (!DisableIVRewrite)
120 AU.addPreserved<IVUsers>();
121 AU.setPreservesCFG();
125 virtual void releaseMemory() {
129 bool isValidRewrite(Value *FromVal, Value *ToVal);
131 void HandleFloatingPointIV(Loop *L, PHINode *PH);
132 void RewriteNonIntegerIVs(Loop *L);
134 void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter);
136 void SimplifyIVUsers(SCEVExpander &Rewriter);
137 void SimplifyIVUsersNoRewrite(Loop *L, SCEVExpander &Rewriter);
139 bool EliminateIVUser(Instruction *UseInst, Instruction *IVOperand);
140 void EliminateIVComparison(ICmpInst *ICmp, Value *IVOperand);
141 void EliminateIVRemainder(BinaryOperator *Rem,
145 void SimplifyCongruentIVs(Loop *L);
147 void RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter);
149 Value *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
150 PHINode *IndVar, SCEVExpander &Rewriter);
152 void SinkUnusedInvariants(Loop *L);
156 char IndVarSimplify::ID = 0;
157 INITIALIZE_PASS_BEGIN(IndVarSimplify, "indvars",
158 "Induction Variable Simplification", false, false)
159 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
160 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
161 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
162 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
163 INITIALIZE_PASS_DEPENDENCY(LCSSA)
164 INITIALIZE_PASS_DEPENDENCY(IVUsers)
165 INITIALIZE_PASS_END(IndVarSimplify, "indvars",
166 "Induction Variable Simplification", false, false)
168 Pass *llvm::createIndVarSimplifyPass() {
169 return new IndVarSimplify();
172 /// isValidRewrite - Return true if the SCEV expansion generated by the
173 /// rewriter can replace the original value. SCEV guarantees that it
174 /// produces the same value, but the way it is produced may be illegal IR.
175 /// Ideally, this function will only be called for verification.
176 bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) {
177 // If an SCEV expression subsumed multiple pointers, its expansion could
178 // reassociate the GEP changing the base pointer. This is illegal because the
179 // final address produced by a GEP chain must be inbounds relative to its
180 // underlying object. Otherwise basic alias analysis, among other things,
181 // could fail in a dangerous way. Ultimately, SCEV will be improved to avoid
182 // producing an expression involving multiple pointers. Until then, we must
185 // Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject
186 // because it understands lcssa phis while SCEV does not.
187 Value *FromPtr = FromVal;
188 Value *ToPtr = ToVal;
189 if (GEPOperator *GEP = dyn_cast<GEPOperator>(FromVal)) {
190 FromPtr = GEP->getPointerOperand();
192 if (GEPOperator *GEP = dyn_cast<GEPOperator>(ToVal)) {
193 ToPtr = GEP->getPointerOperand();
195 if (FromPtr != FromVal || ToPtr != ToVal) {
196 // Quickly check the common case
197 if (FromPtr == ToPtr)
200 // SCEV may have rewritten an expression that produces the GEP's pointer
201 // operand. That's ok as long as the pointer operand has the same base
202 // pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the
203 // base of a recurrence. This handles the case in which SCEV expansion
204 // converts a pointer type recurrence into a nonrecurrent pointer base
205 // indexed by an integer recurrence.
206 const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr));
207 const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr));
208 if (FromBase == ToBase)
211 DEBUG(dbgs() << "INDVARS: GEP rewrite bail out "
212 << *FromBase << " != " << *ToBase << "\n");
219 /// Determine the insertion point for this user. By default, insert immediately
220 /// before the user. SCEVExpander or LICM will hoist loop invariants out of the
221 /// loop. For PHI nodes, there may be multiple uses, so compute the nearest
222 /// common dominator for the incoming blocks.
223 static Instruction *getInsertPointForUses(Instruction *User, Value *Def,
225 PHINode *PHI = dyn_cast<PHINode>(User);
229 Instruction *InsertPt = 0;
230 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) {
231 if (PHI->getIncomingValue(i) != Def)
234 BasicBlock *InsertBB = PHI->getIncomingBlock(i);
236 InsertPt = InsertBB->getTerminator();
239 InsertBB = DT->findNearestCommonDominator(InsertPt->getParent(), InsertBB);
240 InsertPt = InsertBB->getTerminator();
242 assert(InsertPt && "Missing phi operand");
243 assert((!isa<Instruction>(Def) ||
244 DT->dominates(cast<Instruction>(Def), InsertPt)) &&
245 "def does not dominate all uses");
249 //===----------------------------------------------------------------------===//
250 // RewriteNonIntegerIVs and helpers. Prefer integer IVs.
251 //===----------------------------------------------------------------------===//
253 /// ConvertToSInt - Convert APF to an integer, if possible.
254 static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) {
255 bool isExact = false;
256 if (&APF.getSemantics() == &APFloat::PPCDoubleDouble)
258 // See if we can convert this to an int64_t
260 if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero,
261 &isExact) != APFloat::opOK || !isExact)
267 /// HandleFloatingPointIV - If the loop has floating induction variable
268 /// then insert corresponding integer induction variable if possible.
270 /// for(double i = 0; i < 10000; ++i)
272 /// is converted into
273 /// for(int i = 0; i < 10000; ++i)
276 void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) {
277 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
278 unsigned BackEdge = IncomingEdge^1;
280 // Check incoming value.
281 ConstantFP *InitValueVal =
282 dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));
285 if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
288 // Check IV increment. Reject this PN if increment operation is not
289 // an add or increment value can not be represented by an integer.
290 BinaryOperator *Incr =
291 dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
292 if (Incr == 0 || Incr->getOpcode() != Instruction::FAdd) return;
294 // If this is not an add of the PHI with a constantfp, or if the constant fp
295 // is not an integer, bail out.
296 ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
298 if (IncValueVal == 0 || Incr->getOperand(0) != PN ||
299 !ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
302 // Check Incr uses. One user is PN and the other user is an exit condition
303 // used by the conditional terminator.
304 Value::use_iterator IncrUse = Incr->use_begin();
305 Instruction *U1 = cast<Instruction>(*IncrUse++);
306 if (IncrUse == Incr->use_end()) return;
307 Instruction *U2 = cast<Instruction>(*IncrUse++);
308 if (IncrUse != Incr->use_end()) return;
310 // Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't
311 // only used by a branch, we can't transform it.
312 FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
314 Compare = dyn_cast<FCmpInst>(U2);
315 if (Compare == 0 || !Compare->hasOneUse() ||
316 !isa<BranchInst>(Compare->use_back()))
319 BranchInst *TheBr = cast<BranchInst>(Compare->use_back());
321 // We need to verify that the branch actually controls the iteration count
322 // of the loop. If not, the new IV can overflow and no one will notice.
323 // The branch block must be in the loop and one of the successors must be out
325 assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
326 if (!L->contains(TheBr->getParent()) ||
327 (L->contains(TheBr->getSuccessor(0)) &&
328 L->contains(TheBr->getSuccessor(1))))
332 // If it isn't a comparison with an integer-as-fp (the exit value), we can't
334 ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
336 if (ExitValueVal == 0 ||
337 !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
340 // Find new predicate for integer comparison.
341 CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
342 switch (Compare->getPredicate()) {
343 default: return; // Unknown comparison.
344 case CmpInst::FCMP_OEQ:
345 case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
346 case CmpInst::FCMP_ONE:
347 case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
348 case CmpInst::FCMP_OGT:
349 case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
350 case CmpInst::FCMP_OGE:
351 case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
352 case CmpInst::FCMP_OLT:
353 case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
354 case CmpInst::FCMP_OLE:
355 case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
358 // We convert the floating point induction variable to a signed i32 value if
359 // we can. This is only safe if the comparison will not overflow in a way
360 // that won't be trapped by the integer equivalent operations. Check for this
362 // TODO: We could use i64 if it is native and the range requires it.
364 // The start/stride/exit values must all fit in signed i32.
365 if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
368 // If not actually striding (add x, 0.0), avoid touching the code.
372 // Positive and negative strides have different safety conditions.
374 // If we have a positive stride, we require the init to be less than the
375 // exit value and an equality or less than comparison.
376 if (InitValue >= ExitValue ||
377 NewPred == CmpInst::ICMP_SGT || NewPred == CmpInst::ICMP_SGE)
380 uint32_t Range = uint32_t(ExitValue-InitValue);
381 if (NewPred == CmpInst::ICMP_SLE) {
382 // Normalize SLE -> SLT, check for infinite loop.
383 if (++Range == 0) return; // Range overflows.
386 unsigned Leftover = Range % uint32_t(IncValue);
388 // If this is an equality comparison, we require that the strided value
389 // exactly land on the exit value, otherwise the IV condition will wrap
390 // around and do things the fp IV wouldn't.
391 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
395 // If the stride would wrap around the i32 before exiting, we can't
397 if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
401 // If we have a negative stride, we require the init to be greater than the
402 // exit value and an equality or greater than comparison.
403 if (InitValue >= ExitValue ||
404 NewPred == CmpInst::ICMP_SLT || NewPred == CmpInst::ICMP_SLE)
407 uint32_t Range = uint32_t(InitValue-ExitValue);
408 if (NewPred == CmpInst::ICMP_SGE) {
409 // Normalize SGE -> SGT, check for infinite loop.
410 if (++Range == 0) return; // Range overflows.
413 unsigned Leftover = Range % uint32_t(-IncValue);
415 // If this is an equality comparison, we require that the strided value
416 // exactly land on the exit value, otherwise the IV condition will wrap
417 // around and do things the fp IV wouldn't.
418 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
422 // If the stride would wrap around the i32 before exiting, we can't
424 if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
428 IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
430 // Insert new integer induction variable.
431 PHINode *NewPHI = PHINode::Create(Int32Ty, 2, PN->getName()+".int", PN);
432 NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
433 PN->getIncomingBlock(IncomingEdge));
436 BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
437 Incr->getName()+".int", Incr);
438 NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
440 ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
441 ConstantInt::get(Int32Ty, ExitValue),
444 // In the following deletions, PN may become dead and may be deleted.
445 // Use a WeakVH to observe whether this happens.
448 // Delete the old floating point exit comparison. The branch starts using the
450 NewCompare->takeName(Compare);
451 Compare->replaceAllUsesWith(NewCompare);
452 RecursivelyDeleteTriviallyDeadInstructions(Compare);
454 // Delete the old floating point increment.
455 Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
456 RecursivelyDeleteTriviallyDeadInstructions(Incr);
458 // If the FP induction variable still has uses, this is because something else
459 // in the loop uses its value. In order to canonicalize the induction
460 // variable, we chose to eliminate the IV and rewrite it in terms of an
463 // We give preference to sitofp over uitofp because it is faster on most
466 Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
467 PN->getParent()->getFirstNonPHI());
468 PN->replaceAllUsesWith(Conv);
469 RecursivelyDeleteTriviallyDeadInstructions(PN);
472 // Add a new IVUsers entry for the newly-created integer PHI.
474 IU->AddUsersIfInteresting(NewPHI);
477 void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
478 // First step. Check to see if there are any floating-point recurrences.
479 // If there are, change them into integer recurrences, permitting analysis by
480 // the SCEV routines.
482 BasicBlock *Header = L->getHeader();
484 SmallVector<WeakVH, 8> PHIs;
485 for (BasicBlock::iterator I = Header->begin();
486 PHINode *PN = dyn_cast<PHINode>(I); ++I)
489 for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
490 if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i]))
491 HandleFloatingPointIV(L, PN);
493 // If the loop previously had floating-point IV, ScalarEvolution
494 // may not have been able to compute a trip count. Now that we've done some
495 // re-writing, the trip count may be computable.
500 //===----------------------------------------------------------------------===//
501 // RewriteLoopExitValues - Optimize IV users outside the loop.
502 // As a side effect, reduces the amount of IV processing within the loop.
503 //===----------------------------------------------------------------------===//
505 /// RewriteLoopExitValues - Check to see if this loop has a computable
506 /// loop-invariant execution count. If so, this means that we can compute the
507 /// final value of any expressions that are recurrent in the loop, and
508 /// substitute the exit values from the loop into any instructions outside of
509 /// the loop that use the final values of the current expressions.
511 /// This is mostly redundant with the regular IndVarSimplify activities that
512 /// happen later, except that it's more powerful in some cases, because it's
513 /// able to brute-force evaluate arbitrary instructions as long as they have
514 /// constant operands at the beginning of the loop.
515 void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) {
516 // Verify the input to the pass in already in LCSSA form.
517 assert(L->isLCSSAForm(*DT));
519 SmallVector<BasicBlock*, 8> ExitBlocks;
520 L->getUniqueExitBlocks(ExitBlocks);
522 // Find all values that are computed inside the loop, but used outside of it.
523 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan
524 // the exit blocks of the loop to find them.
525 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
526 BasicBlock *ExitBB = ExitBlocks[i];
528 // If there are no PHI nodes in this exit block, then no values defined
529 // inside the loop are used on this path, skip it.
530 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
533 unsigned NumPreds = PN->getNumIncomingValues();
535 // Iterate over all of the PHI nodes.
536 BasicBlock::iterator BBI = ExitBB->begin();
537 while ((PN = dyn_cast<PHINode>(BBI++))) {
539 continue; // dead use, don't replace it
541 // SCEV only supports integer expressions for now.
542 if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy())
545 // It's necessary to tell ScalarEvolution about this explicitly so that
546 // it can walk the def-use list and forget all SCEVs, as it may not be
547 // watching the PHI itself. Once the new exit value is in place, there
548 // may not be a def-use connection between the loop and every instruction
549 // which got a SCEVAddRecExpr for that loop.
552 // Iterate over all of the values in all the PHI nodes.
553 for (unsigned i = 0; i != NumPreds; ++i) {
554 // If the value being merged in is not integer or is not defined
555 // in the loop, skip it.
556 Value *InVal = PN->getIncomingValue(i);
557 if (!isa<Instruction>(InVal))
560 // If this pred is for a subloop, not L itself, skip it.
561 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
562 continue; // The Block is in a subloop, skip it.
564 // Check that InVal is defined in the loop.
565 Instruction *Inst = cast<Instruction>(InVal);
566 if (!L->contains(Inst))
569 // Okay, this instruction has a user outside of the current loop
570 // and varies predictably *inside* the loop. Evaluate the value it
571 // contains when the loop exits, if possible.
572 const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
573 if (!SE->isLoopInvariant(ExitValue, L))
576 Value *ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), Inst);
578 DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
579 << " LoopVal = " << *Inst << "\n");
581 if (!isValidRewrite(Inst, ExitVal)) {
582 DeadInsts.push_back(ExitVal);
588 PN->setIncomingValue(i, ExitVal);
590 // If this instruction is dead now, delete it.
591 RecursivelyDeleteTriviallyDeadInstructions(Inst);
594 // Completely replace a single-pred PHI. This is safe, because the
595 // NewVal won't be variant in the loop, so we don't need an LCSSA phi
597 PN->replaceAllUsesWith(ExitVal);
598 RecursivelyDeleteTriviallyDeadInstructions(PN);
602 // Clone the PHI and delete the original one. This lets IVUsers and
603 // any other maps purge the original user from their records.
604 PHINode *NewPN = cast<PHINode>(PN->clone());
606 NewPN->insertBefore(PN);
607 PN->replaceAllUsesWith(NewPN);
608 PN->eraseFromParent();
613 // The insertion point instruction may have been deleted; clear it out
614 // so that the rewriter doesn't trip over it later.
615 Rewriter.clearInsertPoint();
618 //===----------------------------------------------------------------------===//
619 // Rewrite IV users based on a canonical IV.
620 // To be replaced by -disable-iv-rewrite.
621 //===----------------------------------------------------------------------===//
623 /// SimplifyIVUsers - Iteratively perform simplification on IVUsers within this
624 /// loop. IVUsers is treated as a worklist. Each successive simplification may
625 /// push more users which may themselves be candidates for simplification.
627 /// This is the old approach to IV simplification to be replaced by
628 /// SimplifyIVUsersNoRewrite.
630 void IndVarSimplify::SimplifyIVUsers(SCEVExpander &Rewriter) {
631 // Each round of simplification involves a round of eliminating operations
632 // followed by a round of widening IVs. A single IVUsers worklist is used
633 // across all rounds. The inner loop advances the user. If widening exposes
634 // more uses, then another pass through the outer loop is triggered.
635 for (IVUsers::iterator I = IU->begin(); I != IU->end(); ++I) {
636 Instruction *UseInst = I->getUser();
637 Value *IVOperand = I->getOperandValToReplace();
639 if (ICmpInst *ICmp = dyn_cast<ICmpInst>(UseInst)) {
640 EliminateIVComparison(ICmp, IVOperand);
643 if (BinaryOperator *Rem = dyn_cast<BinaryOperator>(UseInst)) {
644 bool IsSigned = Rem->getOpcode() == Instruction::SRem;
645 if (IsSigned || Rem->getOpcode() == Instruction::URem) {
646 EliminateIVRemainder(Rem, IVOperand, IsSigned);
653 // FIXME: It is an extremely bad idea to indvar substitute anything more
654 // complex than affine induction variables. Doing so will put expensive
655 // polynomial evaluations inside of the loop, and the str reduction pass
656 // currently can only reduce affine polynomials. For now just disable
657 // indvar subst on anything more complex than an affine addrec, unless
658 // it can be expanded to a trivial value.
659 static bool isSafe(const SCEV *S, const Loop *L, ScalarEvolution *SE) {
660 // Loop-invariant values are safe.
661 if (SE->isLoopInvariant(S, L)) return true;
663 // Affine addrecs are safe. Non-affine are not, because LSR doesn't know how
664 // to transform them into efficient code.
665 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
666 return AR->isAffine();
668 // An add is safe it all its operands are safe.
669 if (const SCEVCommutativeExpr *Commutative = dyn_cast<SCEVCommutativeExpr>(S)) {
670 for (SCEVCommutativeExpr::op_iterator I = Commutative->op_begin(),
671 E = Commutative->op_end(); I != E; ++I)
672 if (!isSafe(*I, L, SE)) return false;
676 // A cast is safe if its operand is.
677 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
678 return isSafe(C->getOperand(), L, SE);
680 // A udiv is safe if its operands are.
681 if (const SCEVUDivExpr *UD = dyn_cast<SCEVUDivExpr>(S))
682 return isSafe(UD->getLHS(), L, SE) &&
683 isSafe(UD->getRHS(), L, SE);
685 // SCEVUnknown is always safe.
686 if (isa<SCEVUnknown>(S))
689 // Nothing else is safe.
693 void IndVarSimplify::RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter) {
694 // Rewrite all induction variable expressions in terms of the canonical
695 // induction variable.
697 // If there were induction variables of other sizes or offsets, manually
698 // add the offsets to the primary induction variable and cast, avoiding
699 // the need for the code evaluation methods to insert induction variables
700 // of different sizes.
701 for (IVUsers::iterator UI = IU->begin(), E = IU->end(); UI != E; ++UI) {
702 Value *Op = UI->getOperandValToReplace();
703 Type *UseTy = Op->getType();
704 Instruction *User = UI->getUser();
706 // Compute the final addrec to expand into code.
707 const SCEV *AR = IU->getReplacementExpr(*UI);
709 // Evaluate the expression out of the loop, if possible.
710 if (!L->contains(UI->getUser())) {
711 const SCEV *ExitVal = SE->getSCEVAtScope(AR, L->getParentLoop());
712 if (SE->isLoopInvariant(ExitVal, L))
716 // FIXME: It is an extremely bad idea to indvar substitute anything more
717 // complex than affine induction variables. Doing so will put expensive
718 // polynomial evaluations inside of the loop, and the str reduction pass
719 // currently can only reduce affine polynomials. For now just disable
720 // indvar subst on anything more complex than an affine addrec, unless
721 // it can be expanded to a trivial value.
722 if (!isSafe(AR, L, SE))
725 // Determine the insertion point for this user. By default, insert
726 // immediately before the user. The SCEVExpander class will automatically
727 // hoist loop invariants out of the loop. For PHI nodes, there may be
728 // multiple uses, so compute the nearest common dominator for the
730 Instruction *InsertPt = getInsertPointForUses(User, Op, DT);
732 // Now expand it into actual Instructions and patch it into place.
733 Value *NewVal = Rewriter.expandCodeFor(AR, UseTy, InsertPt);
735 DEBUG(dbgs() << "INDVARS: Rewrote IV '" << *AR << "' " << *Op << '\n'
736 << " into = " << *NewVal << "\n");
738 if (!isValidRewrite(Op, NewVal)) {
739 DeadInsts.push_back(NewVal);
742 // Inform ScalarEvolution that this value is changing. The change doesn't
743 // affect its value, but it does potentially affect which use lists the
744 // value will be on after the replacement, which affects ScalarEvolution's
745 // ability to walk use lists and drop dangling pointers when a value is
747 SE->forgetValue(User);
749 // Patch the new value into place.
751 NewVal->takeName(Op);
752 if (Instruction *NewValI = dyn_cast<Instruction>(NewVal))
753 NewValI->setDebugLoc(User->getDebugLoc());
754 User->replaceUsesOfWith(Op, NewVal);
755 UI->setOperandValToReplace(NewVal);
760 // The old value may be dead now.
761 DeadInsts.push_back(Op);
765 //===----------------------------------------------------------------------===//
766 // IV Widening - Extend the width of an IV to cover its widest uses.
767 //===----------------------------------------------------------------------===//
770 // Collect information about induction variables that are used by sign/zero
771 // extend operations. This information is recorded by CollectExtend and
772 // provides the input to WidenIV.
774 Type *WidestNativeType; // Widest integer type created [sz]ext
775 bool IsSigned; // Was an sext user seen before a zext?
777 WideIVInfo() : WidestNativeType(0), IsSigned(false) {}
781 /// CollectExtend - Update information about the induction variable that is
782 /// extended by this sign or zero extend operation. This is used to determine
783 /// the final width of the IV before actually widening it.
784 static void CollectExtend(CastInst *Cast, bool IsSigned, WideIVInfo &WI,
785 ScalarEvolution *SE, const TargetData *TD) {
786 Type *Ty = Cast->getType();
787 uint64_t Width = SE->getTypeSizeInBits(Ty);
788 if (TD && !TD->isLegalInteger(Width))
791 if (!WI.WidestNativeType) {
792 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
793 WI.IsSigned = IsSigned;
797 // We extend the IV to satisfy the sign of its first user, arbitrarily.
798 if (WI.IsSigned != IsSigned)
801 if (Width > SE->getTypeSizeInBits(WI.WidestNativeType))
802 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
807 /// NarrowIVDefUse - Record a link in the Narrow IV def-use chain along with the
808 /// WideIV that computes the same value as the Narrow IV def. This avoids
809 /// caching Use* pointers.
810 struct NarrowIVDefUse {
811 Instruction *NarrowDef;
812 Instruction *NarrowUse;
813 Instruction *WideDef;
815 NarrowIVDefUse(): NarrowDef(0), NarrowUse(0), WideDef(0) {}
817 NarrowIVDefUse(Instruction *ND, Instruction *NU, Instruction *WD):
818 NarrowDef(ND), NarrowUse(NU), WideDef(WD) {}
821 /// WidenIV - The goal of this transform is to remove sign and zero extends
822 /// without creating any new induction variables. To do this, it creates a new
823 /// phi of the wider type and redirects all users, either removing extends or
824 /// inserting truncs whenever we stop propagating the type.
840 Instruction *WideInc;
841 const SCEV *WideIncExpr;
842 SmallVectorImpl<WeakVH> &DeadInsts;
844 SmallPtrSet<Instruction*,16> Widened;
845 SmallVector<NarrowIVDefUse, 8> NarrowIVUsers;
848 WidenIV(PHINode *PN, const WideIVInfo &WI, LoopInfo *LInfo,
849 ScalarEvolution *SEv, DominatorTree *DTree,
850 SmallVectorImpl<WeakVH> &DI) :
852 WideType(WI.WidestNativeType),
853 IsSigned(WI.IsSigned),
855 L(LI->getLoopFor(OrigPhi->getParent())),
862 assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV");
865 PHINode *CreateWideIV(SCEVExpander &Rewriter);
868 Instruction *CloneIVUser(NarrowIVDefUse DU);
870 const SCEVAddRecExpr *GetWideRecurrence(Instruction *NarrowUse);
872 Instruction *WidenIVUse(NarrowIVDefUse DU);
874 void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef);
876 } // anonymous namespace
878 static Value *getExtend( Value *NarrowOper, Type *WideType,
879 bool IsSigned, IRBuilder<> &Builder) {
880 return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) :
881 Builder.CreateZExt(NarrowOper, WideType);
884 /// CloneIVUser - Instantiate a wide operation to replace a narrow
885 /// operation. This only needs to handle operations that can evaluation to
886 /// SCEVAddRec. It can safely return 0 for any operation we decide not to clone.
887 Instruction *WidenIV::CloneIVUser(NarrowIVDefUse DU) {
888 unsigned Opcode = DU.NarrowUse->getOpcode();
892 case Instruction::Add:
893 case Instruction::Mul:
894 case Instruction::UDiv:
895 case Instruction::Sub:
896 case Instruction::And:
897 case Instruction::Or:
898 case Instruction::Xor:
899 case Instruction::Shl:
900 case Instruction::LShr:
901 case Instruction::AShr:
902 DEBUG(dbgs() << "Cloning IVUser: " << *DU.NarrowUse << "\n");
904 IRBuilder<> Builder(DU.NarrowUse);
906 // Replace NarrowDef operands with WideDef. Otherwise, we don't know
907 // anything about the narrow operand yet so must insert a [sz]ext. It is
908 // probably loop invariant and will be folded or hoisted. If it actually
909 // comes from a widened IV, it should be removed during a future call to
911 Value *LHS = (DU.NarrowUse->getOperand(0) == DU.NarrowDef) ? DU.WideDef :
912 getExtend(DU.NarrowUse->getOperand(0), WideType, IsSigned, Builder);
913 Value *RHS = (DU.NarrowUse->getOperand(1) == DU.NarrowDef) ? DU.WideDef :
914 getExtend(DU.NarrowUse->getOperand(1), WideType, IsSigned, Builder);
916 BinaryOperator *NarrowBO = cast<BinaryOperator>(DU.NarrowUse);
917 BinaryOperator *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(),
919 NarrowBO->getName());
920 Builder.Insert(WideBO);
921 if (const OverflowingBinaryOperator *OBO =
922 dyn_cast<OverflowingBinaryOperator>(NarrowBO)) {
923 if (OBO->hasNoUnsignedWrap()) WideBO->setHasNoUnsignedWrap();
924 if (OBO->hasNoSignedWrap()) WideBO->setHasNoSignedWrap();
931 /// HoistStep - Attempt to hoist an IV increment above a potential use.
933 /// To successfully hoist, two criteria must be met:
934 /// - IncV operands dominate InsertPos and
935 /// - InsertPos dominates IncV
937 /// Meeting the second condition means that we don't need to check all of IncV's
938 /// existing uses (it's moving up in the domtree).
940 /// This does not yet recursively hoist the operands, although that would
941 /// not be difficult.
942 static bool HoistStep(Instruction *IncV, Instruction *InsertPos,
943 const DominatorTree *DT)
945 if (DT->dominates(IncV, InsertPos))
948 if (!DT->dominates(InsertPos->getParent(), IncV->getParent()))
951 if (IncV->mayHaveSideEffects())
954 // Attempt to hoist IncV
955 for (User::op_iterator OI = IncV->op_begin(), OE = IncV->op_end();
957 Instruction *OInst = dyn_cast<Instruction>(OI);
958 if (OInst && !DT->dominates(OInst, InsertPos))
961 IncV->moveBefore(InsertPos);
965 // GetWideRecurrence - Is this instruction potentially interesting from IVUsers'
966 // perspective after widening it's type? In other words, can the extend be
967 // safely hoisted out of the loop with SCEV reducing the value to a recurrence
968 // on the same loop. If so, return the sign or zero extended
969 // recurrence. Otherwise return NULL.
970 const SCEVAddRecExpr *WidenIV::GetWideRecurrence(Instruction *NarrowUse) {
971 if (!SE->isSCEVable(NarrowUse->getType()))
974 const SCEV *NarrowExpr = SE->getSCEV(NarrowUse);
975 if (SE->getTypeSizeInBits(NarrowExpr->getType())
976 >= SE->getTypeSizeInBits(WideType)) {
977 // NarrowUse implicitly widens its operand. e.g. a gep with a narrow
978 // index. So don't follow this use.
982 const SCEV *WideExpr = IsSigned ?
983 SE->getSignExtendExpr(NarrowExpr, WideType) :
984 SE->getZeroExtendExpr(NarrowExpr, WideType);
985 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr);
986 if (!AddRec || AddRec->getLoop() != L)
992 /// WidenIVUse - Determine whether an individual user of the narrow IV can be
993 /// widened. If so, return the wide clone of the user.
994 Instruction *WidenIV::WidenIVUse(NarrowIVDefUse DU) {
996 // Stop traversing the def-use chain at inner-loop phis or post-loop phis.
997 if (isa<PHINode>(DU.NarrowUse) &&
998 LI->getLoopFor(DU.NarrowUse->getParent()) != L)
1001 // Our raison d'etre! Eliminate sign and zero extension.
1002 if (IsSigned ? isa<SExtInst>(DU.NarrowUse) : isa<ZExtInst>(DU.NarrowUse)) {
1003 Value *NewDef = DU.WideDef;
1004 if (DU.NarrowUse->getType() != WideType) {
1005 unsigned CastWidth = SE->getTypeSizeInBits(DU.NarrowUse->getType());
1006 unsigned IVWidth = SE->getTypeSizeInBits(WideType);
1007 if (CastWidth < IVWidth) {
1008 // The cast isn't as wide as the IV, so insert a Trunc.
1009 IRBuilder<> Builder(DU.NarrowUse);
1010 NewDef = Builder.CreateTrunc(DU.WideDef, DU.NarrowUse->getType());
1013 // A wider extend was hidden behind a narrower one. This may induce
1014 // another round of IV widening in which the intermediate IV becomes
1015 // dead. It should be very rare.
1016 DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi
1017 << " not wide enough to subsume " << *DU.NarrowUse << "\n");
1018 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
1019 NewDef = DU.NarrowUse;
1022 if (NewDef != DU.NarrowUse) {
1023 DEBUG(dbgs() << "INDVARS: eliminating " << *DU.NarrowUse
1024 << " replaced by " << *DU.WideDef << "\n");
1026 DU.NarrowUse->replaceAllUsesWith(NewDef);
1027 DeadInsts.push_back(DU.NarrowUse);
1029 // Now that the extend is gone, we want to expose it's uses for potential
1030 // further simplification. We don't need to directly inform SimplifyIVUsers
1031 // of the new users, because their parent IV will be processed later as a
1032 // new loop phi. If we preserved IVUsers analysis, we would also want to
1033 // push the uses of WideDef here.
1035 // No further widening is needed. The deceased [sz]ext had done it for us.
1039 // Does this user itself evaluate to a recurrence after widening?
1040 const SCEVAddRecExpr *WideAddRec = GetWideRecurrence(DU.NarrowUse);
1042 // This user does not evaluate to a recurence after widening, so don't
1043 // follow it. Instead insert a Trunc to kill off the original use,
1044 // eventually isolating the original narrow IV so it can be removed.
1045 IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT));
1046 Value *Trunc = Builder.CreateTrunc(DU.WideDef, DU.NarrowDef->getType());
1047 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, Trunc);
1050 // Assume block terminators cannot evaluate to a recurrence. We can't to
1051 // insert a Trunc after a terminator if there happens to be a critical edge.
1052 assert(DU.NarrowUse != DU.NarrowUse->getParent()->getTerminator() &&
1053 "SCEV is not expected to evaluate a block terminator");
1055 // Reuse the IV increment that SCEVExpander created as long as it dominates
1057 Instruction *WideUse = 0;
1058 if (WideAddRec == WideIncExpr && HoistStep(WideInc, DU.NarrowUse, DT)) {
1062 WideUse = CloneIVUser(DU);
1066 // Evaluation of WideAddRec ensured that the narrow expression could be
1067 // extended outside the loop without overflow. This suggests that the wide use
1068 // evaluates to the same expression as the extended narrow use, but doesn't
1069 // absolutely guarantee it. Hence the following failsafe check. In rare cases
1070 // where it fails, we simply throw away the newly created wide use.
1071 if (WideAddRec != SE->getSCEV(WideUse)) {
1072 DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse
1073 << ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n");
1074 DeadInsts.push_back(WideUse);
1078 // Returning WideUse pushes it on the worklist.
1082 /// pushNarrowIVUsers - Add eligible users of NarrowDef to NarrowIVUsers.
1084 void WidenIV::pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef) {
1085 for (Value::use_iterator UI = NarrowDef->use_begin(),
1086 UE = NarrowDef->use_end(); UI != UE; ++UI) {
1087 Instruction *NarrowUse = cast<Instruction>(*UI);
1089 // Handle data flow merges and bizarre phi cycles.
1090 if (!Widened.insert(NarrowUse))
1093 NarrowIVUsers.push_back(NarrowIVDefUse(NarrowDef, NarrowUse, WideDef));
1097 /// CreateWideIV - Process a single induction variable. First use the
1098 /// SCEVExpander to create a wide induction variable that evaluates to the same
1099 /// recurrence as the original narrow IV. Then use a worklist to forward
1100 /// traverse the narrow IV's def-use chain. After WidenIVUse has processed all
1101 /// interesting IV users, the narrow IV will be isolated for removal by
1104 /// It would be simpler to delete uses as they are processed, but we must avoid
1105 /// invalidating SCEV expressions.
1107 PHINode *WidenIV::CreateWideIV(SCEVExpander &Rewriter) {
1108 // Is this phi an induction variable?
1109 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi));
1113 // Widen the induction variable expression.
1114 const SCEV *WideIVExpr = IsSigned ?
1115 SE->getSignExtendExpr(AddRec, WideType) :
1116 SE->getZeroExtendExpr(AddRec, WideType);
1118 assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType &&
1119 "Expect the new IV expression to preserve its type");
1121 // Can the IV be extended outside the loop without overflow?
1122 AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr);
1123 if (!AddRec || AddRec->getLoop() != L)
1126 // An AddRec must have loop-invariant operands. Since this AddRec is
1127 // materialized by a loop header phi, the expression cannot have any post-loop
1128 // operands, so they must dominate the loop header.
1129 assert(SE->properlyDominates(AddRec->getStart(), L->getHeader()) &&
1130 SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader())
1131 && "Loop header phi recurrence inputs do not dominate the loop");
1133 // The rewriter provides a value for the desired IV expression. This may
1134 // either find an existing phi or materialize a new one. Either way, we
1135 // expect a well-formed cyclic phi-with-increments. i.e. any operand not part
1136 // of the phi-SCC dominates the loop entry.
1137 Instruction *InsertPt = L->getHeader()->begin();
1138 WidePhi = cast<PHINode>(Rewriter.expandCodeFor(AddRec, WideType, InsertPt));
1140 // Remembering the WideIV increment generated by SCEVExpander allows
1141 // WidenIVUse to reuse it when widening the narrow IV's increment. We don't
1142 // employ a general reuse mechanism because the call above is the only call to
1143 // SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses.
1144 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1146 cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock));
1147 WideIncExpr = SE->getSCEV(WideInc);
1150 DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n");
1153 // Traverse the def-use chain using a worklist starting at the original IV.
1154 assert(Widened.empty() && NarrowIVUsers.empty() && "expect initial state" );
1156 Widened.insert(OrigPhi);
1157 pushNarrowIVUsers(OrigPhi, WidePhi);
1159 while (!NarrowIVUsers.empty()) {
1160 NarrowIVDefUse DU = NarrowIVUsers.pop_back_val();
1162 // Process a def-use edge. This may replace the use, so don't hold a
1163 // use_iterator across it.
1164 Instruction *WideUse = WidenIVUse(DU);
1166 // Follow all def-use edges from the previous narrow use.
1168 pushNarrowIVUsers(DU.NarrowUse, WideUse);
1170 // WidenIVUse may have removed the def-use edge.
1171 if (DU.NarrowDef->use_empty())
1172 DeadInsts.push_back(DU.NarrowDef);
1177 //===----------------------------------------------------------------------===//
1178 // Simplification of IV users based on SCEV evaluation.
1179 //===----------------------------------------------------------------------===//
1181 void IndVarSimplify::EliminateIVComparison(ICmpInst *ICmp, Value *IVOperand) {
1182 unsigned IVOperIdx = 0;
1183 ICmpInst::Predicate Pred = ICmp->getPredicate();
1184 if (IVOperand != ICmp->getOperand(0)) {
1186 assert(IVOperand == ICmp->getOperand(1) && "Can't find IVOperand");
1188 Pred = ICmpInst::getSwappedPredicate(Pred);
1191 // Get the SCEVs for the ICmp operands.
1192 const SCEV *S = SE->getSCEV(ICmp->getOperand(IVOperIdx));
1193 const SCEV *X = SE->getSCEV(ICmp->getOperand(1 - IVOperIdx));
1195 // Simplify unnecessary loops away.
1196 const Loop *ICmpLoop = LI->getLoopFor(ICmp->getParent());
1197 S = SE->getSCEVAtScope(S, ICmpLoop);
1198 X = SE->getSCEVAtScope(X, ICmpLoop);
1200 // If the condition is always true or always false, replace it with
1201 // a constant value.
1202 if (SE->isKnownPredicate(Pred, S, X))
1203 ICmp->replaceAllUsesWith(ConstantInt::getTrue(ICmp->getContext()));
1204 else if (SE->isKnownPredicate(ICmpInst::getInversePredicate(Pred), S, X))
1205 ICmp->replaceAllUsesWith(ConstantInt::getFalse(ICmp->getContext()));
1209 DEBUG(dbgs() << "INDVARS: Eliminated comparison: " << *ICmp << '\n');
1212 DeadInsts.push_back(ICmp);
1215 void IndVarSimplify::EliminateIVRemainder(BinaryOperator *Rem,
1218 // We're only interested in the case where we know something about
1220 if (IVOperand != Rem->getOperand(0))
1223 // Get the SCEVs for the ICmp operands.
1224 const SCEV *S = SE->getSCEV(Rem->getOperand(0));
1225 const SCEV *X = SE->getSCEV(Rem->getOperand(1));
1227 // Simplify unnecessary loops away.
1228 const Loop *ICmpLoop = LI->getLoopFor(Rem->getParent());
1229 S = SE->getSCEVAtScope(S, ICmpLoop);
1230 X = SE->getSCEVAtScope(X, ICmpLoop);
1232 // i % n --> i if i is in [0,n).
1233 if ((!IsSigned || SE->isKnownNonNegative(S)) &&
1234 SE->isKnownPredicate(IsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
1236 Rem->replaceAllUsesWith(Rem->getOperand(0));
1238 // (i+1) % n --> (i+1)==n?0:(i+1) if i is in [0,n).
1239 const SCEV *LessOne =
1240 SE->getMinusSCEV(S, SE->getConstant(S->getType(), 1));
1241 if (IsSigned && !SE->isKnownNonNegative(LessOne))
1244 if (!SE->isKnownPredicate(IsSigned ?
1245 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
1249 ICmpInst *ICmp = new ICmpInst(Rem, ICmpInst::ICMP_EQ,
1250 Rem->getOperand(0), Rem->getOperand(1),
1253 SelectInst::Create(ICmp,
1254 ConstantInt::get(Rem->getType(), 0),
1255 Rem->getOperand(0), "tmp", Rem);
1256 Rem->replaceAllUsesWith(Sel);
1259 // Inform IVUsers about the new users.
1261 if (Instruction *I = dyn_cast<Instruction>(Rem->getOperand(0)))
1262 IU->AddUsersIfInteresting(I);
1264 DEBUG(dbgs() << "INDVARS: Simplified rem: " << *Rem << '\n');
1267 DeadInsts.push_back(Rem);
1270 /// EliminateIVUser - Eliminate an operation that consumes a simple IV and has
1271 /// no observable side-effect given the range of IV values.
1272 bool IndVarSimplify::EliminateIVUser(Instruction *UseInst,
1273 Instruction *IVOperand) {
1274 if (ICmpInst *ICmp = dyn_cast<ICmpInst>(UseInst)) {
1275 EliminateIVComparison(ICmp, IVOperand);
1278 if (BinaryOperator *Rem = dyn_cast<BinaryOperator>(UseInst)) {
1279 bool IsSigned = Rem->getOpcode() == Instruction::SRem;
1280 if (IsSigned || Rem->getOpcode() == Instruction::URem) {
1281 EliminateIVRemainder(Rem, IVOperand, IsSigned);
1286 // Eliminate any operation that SCEV can prove is an identity function.
1287 if (!SE->isSCEVable(UseInst->getType()) ||
1288 (UseInst->getType() != IVOperand->getType()) ||
1289 (SE->getSCEV(UseInst) != SE->getSCEV(IVOperand)))
1292 DEBUG(dbgs() << "INDVARS: Eliminated identity: " << *UseInst << '\n');
1294 UseInst->replaceAllUsesWith(IVOperand);
1297 DeadInsts.push_back(UseInst);
1301 /// pushIVUsers - Add all uses of Def to the current IV's worklist.
1303 static void pushIVUsers(
1305 SmallPtrSet<Instruction*,16> &Simplified,
1306 SmallVectorImpl< std::pair<Instruction*,Instruction*> > &SimpleIVUsers) {
1308 for (Value::use_iterator UI = Def->use_begin(), E = Def->use_end();
1310 Instruction *User = cast<Instruction>(*UI);
1312 // Avoid infinite or exponential worklist processing.
1313 // Also ensure unique worklist users.
1314 // If Def is a LoopPhi, it may not be in the Simplified set, so check for
1315 // self edges first.
1316 if (User != Def && Simplified.insert(User))
1317 SimpleIVUsers.push_back(std::make_pair(User, Def));
1321 /// isSimpleIVUser - Return true if this instruction generates a simple SCEV
1322 /// expression in terms of that IV.
1324 /// This is similar to IVUsers' isInsteresting() but processes each instruction
1325 /// non-recursively when the operand is already known to be a simpleIVUser.
1327 static bool isSimpleIVUser(Instruction *I, const Loop *L, ScalarEvolution *SE) {
1328 if (!SE->isSCEVable(I->getType()))
1331 // Get the symbolic expression for this instruction.
1332 const SCEV *S = SE->getSCEV(I);
1334 // Only consider affine recurrences.
1335 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S);
1336 if (AR && AR->getLoop() == L)
1342 /// SimplifyIVUsersNoRewrite - Iteratively perform simplification on a worklist
1343 /// of IV users. Each successive simplification may push more users which may
1344 /// themselves be candidates for simplification.
1346 /// The "NoRewrite" algorithm does not require IVUsers analysis. Instead, it
1347 /// simplifies instructions in-place during analysis. Rather than rewriting
1348 /// induction variables bottom-up from their users, it transforms a chain of
1349 /// IVUsers top-down, updating the IR only when it encouters a clear
1350 /// optimization opportunitiy. A SCEVExpander "Rewriter" instance is still
1351 /// needed, but only used to generate a new IV (phi) of wider type for sign/zero
1352 /// extend elimination.
1354 /// Once DisableIVRewrite is default, LSR will be the only client of IVUsers.
1356 void IndVarSimplify::SimplifyIVUsersNoRewrite(Loop *L, SCEVExpander &Rewriter) {
1357 std::map<PHINode *, WideIVInfo> WideIVMap;
1359 SmallVector<PHINode*, 8> LoopPhis;
1360 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1361 LoopPhis.push_back(cast<PHINode>(I));
1363 // Each round of simplification iterates through the SimplifyIVUsers worklist
1364 // for all current phis, then determines whether any IVs can be
1365 // widened. Widening adds new phis to LoopPhis, inducing another round of
1366 // simplification on the wide IVs.
1367 while (!LoopPhis.empty()) {
1368 // Evaluate as many IV expressions as possible before widening any IVs. This
1369 // forces SCEV to set no-wrap flags before evaluating sign/zero
1370 // extension. The first time SCEV attempts to normalize sign/zero extension,
1371 // the result becomes final. So for the most predictable results, we delay
1372 // evaluation of sign/zero extend evaluation until needed, and avoid running
1373 // other SCEV based analysis prior to SimplifyIVUsersNoRewrite.
1375 PHINode *CurrIV = LoopPhis.pop_back_val();
1377 // Information about sign/zero extensions of CurrIV.
1380 // Instructions processed by SimplifyIVUsers for CurrIV.
1381 SmallPtrSet<Instruction*,16> Simplified;
1383 // Use-def pairs if IV users waiting to be processed for CurrIV.
1384 SmallVector<std::pair<Instruction*, Instruction*>, 8> SimpleIVUsers;
1386 // Push users of the current LoopPhi. In rare cases, pushIVUsers may be
1387 // called multiple times for the same LoopPhi. This is the proper thing to
1388 // do for loop header phis that use each other.
1389 pushIVUsers(CurrIV, Simplified, SimpleIVUsers);
1391 while (!SimpleIVUsers.empty()) {
1392 std::pair<Instruction*, Instruction*> UseOper =
1393 SimpleIVUsers.pop_back_val();
1394 // Bypass back edges to avoid extra work.
1395 if (UseOper.first == CurrIV) continue;
1397 if (EliminateIVUser(UseOper.first, UseOper.second)) {
1398 pushIVUsers(UseOper.second, Simplified, SimpleIVUsers);
1401 if (CastInst *Cast = dyn_cast<CastInst>(UseOper.first)) {
1402 bool IsSigned = Cast->getOpcode() == Instruction::SExt;
1403 if (IsSigned || Cast->getOpcode() == Instruction::ZExt) {
1404 CollectExtend(Cast, IsSigned, WI, SE, TD);
1408 if (isSimpleIVUser(UseOper.first, L, SE)) {
1409 pushIVUsers(UseOper.first, Simplified, SimpleIVUsers);
1412 if (WI.WidestNativeType) {
1413 WideIVMap[CurrIV] = WI;
1415 } while(!LoopPhis.empty());
1417 for (std::map<PHINode *, WideIVInfo>::const_iterator I = WideIVMap.begin(),
1418 E = WideIVMap.end(); I != E; ++I) {
1419 WidenIV Widener(I->first, I->second, LI, SE, DT, DeadInsts);
1420 if (PHINode *WidePhi = Widener.CreateWideIV(Rewriter)) {
1422 LoopPhis.push_back(WidePhi);
1429 /// SimplifyCongruentIVs - Check for congruent phis in this loop header and
1430 /// populate ExprToIVMap for use later.
1432 void IndVarSimplify::SimplifyCongruentIVs(Loop *L) {
1433 DenseMap<const SCEV *, PHINode *> ExprToIVMap;
1434 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1435 PHINode *Phi = cast<PHINode>(I);
1436 if (!SE->isSCEVable(Phi->getType()))
1439 const SCEV *S = SE->getSCEV(Phi);
1440 std::pair<DenseMap<const SCEV *, PHINode *>::const_iterator, bool> Tmp =
1441 ExprToIVMap.insert(std::make_pair(S, Phi));
1444 PHINode *OrigPhi = Tmp.first->second;
1446 // If one phi derives from the other via GEPs, types may differ.
1447 if (OrigPhi->getType() != Phi->getType())
1450 // Replacing the congruent phi is sufficient because acyclic redundancy
1451 // elimination, CSE/GVN, should handle the rest. However, once SCEV proves
1452 // that a phi is congruent, it's almost certain to be the head of an IV
1453 // user cycle that is isomorphic with the original phi. So it's worth
1454 // eagerly cleaning up the common case of a single IV increment.
1455 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1456 Instruction *OrigInc =
1457 cast<Instruction>(OrigPhi->getIncomingValueForBlock(LatchBlock));
1458 Instruction *IsomorphicInc =
1459 cast<Instruction>(Phi->getIncomingValueForBlock(LatchBlock));
1460 if (OrigInc != IsomorphicInc &&
1461 OrigInc->getType() == IsomorphicInc->getType() &&
1462 SE->getSCEV(OrigInc) == SE->getSCEV(IsomorphicInc) &&
1463 HoistStep(OrigInc, IsomorphicInc, DT)) {
1464 DEBUG(dbgs() << "INDVARS: Eliminated congruent iv.inc: "
1465 << *IsomorphicInc << '\n');
1466 IsomorphicInc->replaceAllUsesWith(OrigInc);
1467 DeadInsts.push_back(IsomorphicInc);
1470 DEBUG(dbgs() << "INDVARS: Eliminated congruent iv: " << *Phi << '\n');
1472 Phi->replaceAllUsesWith(OrigPhi);
1473 DeadInsts.push_back(Phi);
1477 //===----------------------------------------------------------------------===//
1478 // LinearFunctionTestReplace and its kin. Rewrite the loop exit condition.
1479 //===----------------------------------------------------------------------===//
1481 // Check for expressions that ScalarEvolution generates to compute
1482 // BackedgeTakenInfo. If these expressions have not been reduced, then expanding
1483 // them may incur additional cost (albeit in the loop preheader).
1484 static bool isHighCostExpansion(const SCEV *S, BranchInst *BI,
1485 ScalarEvolution *SE) {
1486 // If the backedge-taken count is a UDiv, it's very likely a UDiv that
1487 // ScalarEvolution's HowFarToZero or HowManyLessThans produced to compute a
1488 // precise expression, rather than a UDiv from the user's code. If we can't
1489 // find a UDiv in the code with some simple searching, assume the former and
1490 // forego rewriting the loop.
1491 if (isa<SCEVUDivExpr>(S)) {
1492 ICmpInst *OrigCond = dyn_cast<ICmpInst>(BI->getCondition());
1493 if (!OrigCond) return true;
1494 const SCEV *R = SE->getSCEV(OrigCond->getOperand(1));
1495 R = SE->getMinusSCEV(R, SE->getConstant(R->getType(), 1));
1497 const SCEV *L = SE->getSCEV(OrigCond->getOperand(0));
1498 L = SE->getMinusSCEV(L, SE->getConstant(L->getType(), 1));
1504 if (!DisableIVRewrite || ForceLFTR)
1507 // Recurse past add expressions, which commonly occur in the
1508 // BackedgeTakenCount. They may already exist in program code, and if not,
1509 // they are not too expensive rematerialize.
1510 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
1511 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
1513 if (isHighCostExpansion(*I, BI, SE))
1519 // HowManyLessThans uses a Max expression whenever the loop is not guarded by
1520 // the exit condition.
1521 if (isa<SCEVSMaxExpr>(S) || isa<SCEVUMaxExpr>(S))
1524 // If we haven't recognized an expensive SCEV patter, assume its an expression
1525 // produced by program code.
1529 /// canExpandBackedgeTakenCount - Return true if this loop's backedge taken
1530 /// count expression can be safely and cheaply expanded into an instruction
1531 /// sequence that can be used by LinearFunctionTestReplace.
1532 static bool canExpandBackedgeTakenCount(Loop *L, ScalarEvolution *SE) {
1533 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1534 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) ||
1535 BackedgeTakenCount->isZero())
1538 if (!L->getExitingBlock())
1541 // Can't rewrite non-branch yet.
1542 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
1546 if (isHighCostExpansion(BackedgeTakenCount, BI, SE))
1552 /// getBackedgeIVType - Get the widest type used by the loop test after peeking
1555 /// TODO: Unnecessary when ForceLFTR is removed.
1556 static Type *getBackedgeIVType(Loop *L) {
1557 if (!L->getExitingBlock())
1560 // Can't rewrite non-branch yet.
1561 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
1565 ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition());
1570 for(User::op_iterator OI = Cond->op_begin(), OE = Cond->op_end();
1572 assert((!Ty || Ty == (*OI)->getType()) && "bad icmp operand types");
1573 TruncInst *Trunc = dyn_cast<TruncInst>(*OI);
1577 return Trunc->getSrcTy();
1582 /// isLoopInvariant - Perform a quick domtree based check for loop invariance
1583 /// assuming that V is used within the loop. LoopInfo::isLoopInvariant() seems
1584 /// gratuitous for this purpose.
1585 static bool isLoopInvariant(Value *V, Loop *L, DominatorTree *DT) {
1586 Instruction *Inst = dyn_cast<Instruction>(V);
1590 return DT->properlyDominates(Inst->getParent(), L->getHeader());
1593 /// getLoopPhiForCounter - Return the loop header phi IFF IncV adds a loop
1594 /// invariant value to the phi.
1595 static PHINode *getLoopPhiForCounter(Value *IncV, Loop *L, DominatorTree *DT) {
1596 Instruction *IncI = dyn_cast<Instruction>(IncV);
1600 switch (IncI->getOpcode()) {
1601 case Instruction::Add:
1602 case Instruction::Sub:
1604 case Instruction::GetElementPtr:
1605 // An IV counter must preserve its type.
1606 if (IncI->getNumOperands() == 2)
1612 PHINode *Phi = dyn_cast<PHINode>(IncI->getOperand(0));
1613 if (Phi && Phi->getParent() == L->getHeader()) {
1614 if (isLoopInvariant(IncI->getOperand(1), L, DT))
1618 if (IncI->getOpcode() == Instruction::GetElementPtr)
1621 // Allow add/sub to be commuted.
1622 Phi = dyn_cast<PHINode>(IncI->getOperand(1));
1623 if (Phi && Phi->getParent() == L->getHeader()) {
1624 if (isLoopInvariant(IncI->getOperand(0), L, DT))
1630 /// needsLFTR - LinearFunctionTestReplace policy. Return true unless we can show
1631 /// that the current exit test is already sufficiently canonical.
1632 static bool needsLFTR(Loop *L, DominatorTree *DT) {
1633 assert(L->getExitingBlock() && "expected loop exit");
1635 BasicBlock *LatchBlock = L->getLoopLatch();
1636 // Don't bother with LFTR if the loop is not properly simplified.
1640 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
1641 assert(BI && "expected exit branch");
1643 // Do LFTR to simplify the exit condition to an ICMP.
1644 ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition());
1648 // Do LFTR to simplify the exit ICMP to EQ/NE
1649 ICmpInst::Predicate Pred = Cond->getPredicate();
1650 if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ)
1653 // Look for a loop invariant RHS
1654 Value *LHS = Cond->getOperand(0);
1655 Value *RHS = Cond->getOperand(1);
1656 if (!isLoopInvariant(RHS, L, DT)) {
1657 if (!isLoopInvariant(LHS, L, DT))
1659 std::swap(LHS, RHS);
1661 // Look for a simple IV counter LHS
1662 PHINode *Phi = dyn_cast<PHINode>(LHS);
1664 Phi = getLoopPhiForCounter(LHS, L, DT);
1669 // Do LFTR if the exit condition's IV is *not* a simple counter.
1670 Value *IncV = Phi->getIncomingValueForBlock(L->getLoopLatch());
1671 return Phi != getLoopPhiForCounter(IncV, L, DT);
1674 /// AlmostDeadIV - Return true if this IV has any uses other than the (soon to
1675 /// be rewritten) loop exit test.
1676 static bool AlmostDeadIV(PHINode *Phi, BasicBlock *LatchBlock, Value *Cond) {
1677 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1678 Value *IncV = Phi->getIncomingValue(LatchIdx);
1680 for (Value::use_iterator UI = Phi->use_begin(), UE = Phi->use_end();
1682 if (*UI != Cond && *UI != IncV) return false;
1685 for (Value::use_iterator UI = IncV->use_begin(), UE = IncV->use_end();
1687 if (*UI != Cond && *UI != Phi) return false;
1692 /// FindLoopCounter - Find an affine IV in canonical form.
1694 /// FIXME: Accept -1 stride and set IVLimit = IVInit - BECount
1696 /// FIXME: Accept non-unit stride as long as SCEV can reduce BECount * Stride.
1697 /// This is difficult in general for SCEV because of potential overflow. But we
1698 /// could at least handle constant BECounts.
1700 FindLoopCounter(Loop *L, const SCEV *BECount,
1701 ScalarEvolution *SE, DominatorTree *DT, const TargetData *TD) {
1702 // I'm not sure how BECount could be a pointer type, but we definitely don't
1703 // want to LFTR that.
1704 if (BECount->getType()->isPointerTy())
1707 uint64_t BCWidth = SE->getTypeSizeInBits(BECount->getType());
1710 cast<BranchInst>(L->getExitingBlock()->getTerminator())->getCondition();
1712 // Loop over all of the PHI nodes, looking for a simple counter.
1713 PHINode *BestPhi = 0;
1714 const SCEV *BestInit = 0;
1715 BasicBlock *LatchBlock = L->getLoopLatch();
1716 assert(LatchBlock && "needsLFTR should guarantee a loop latch");
1718 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1719 PHINode *Phi = cast<PHINode>(I);
1720 if (!SE->isSCEVable(Phi->getType()))
1723 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Phi));
1724 if (!AR || AR->getLoop() != L || !AR->isAffine())
1727 // AR may be a pointer type, while BECount is an integer type.
1728 // AR may be wider than BECount. With eq/ne tests overflow is immaterial.
1729 // AR may not be a narrower type, or we may never exit.
1730 uint64_t PhiWidth = SE->getTypeSizeInBits(AR->getType());
1731 if (PhiWidth < BCWidth || (TD && !TD->isLegalInteger(PhiWidth)))
1734 const SCEV *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
1735 if (!Step || !Step->isOne())
1738 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1739 Value *IncV = Phi->getIncomingValue(LatchIdx);
1740 if (getLoopPhiForCounter(IncV, L, DT) != Phi)
1743 const SCEV *Init = AR->getStart();
1745 if (BestPhi && !AlmostDeadIV(BestPhi, LatchBlock, Cond)) {
1746 // Don't force a live loop counter if another IV can be used.
1747 if (AlmostDeadIV(Phi, LatchBlock, Cond))
1750 // Prefer to count-from-zero. This is a more "canonical" counter form. It
1751 // also prefers integer to pointer IVs.
1752 if (BestInit->isZero() != Init->isZero()) {
1753 if (BestInit->isZero())
1756 // If two IVs both count from zero or both count from nonzero then the
1757 // narrower is likely a dead phi that has been widened. Use the wider phi
1758 // to allow the other to be eliminated.
1759 if (PhiWidth <= SE->getTypeSizeInBits(BestPhi->getType()))
1768 /// LinearFunctionTestReplace - This method rewrites the exit condition of the
1769 /// loop to be a canonical != comparison against the incremented loop induction
1770 /// variable. This pass is able to rewrite the exit tests of any loop where the
1771 /// SCEV analysis can determine a loop-invariant trip count of the loop, which
1772 /// is actually a much broader range than just linear tests.
1773 Value *IndVarSimplify::
1774 LinearFunctionTestReplace(Loop *L,
1775 const SCEV *BackedgeTakenCount,
1777 SCEVExpander &Rewriter) {
1778 assert(canExpandBackedgeTakenCount(L, SE) && "precondition");
1779 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1781 // In DisableIVRewrite mode, IndVar is not necessarily a canonical IV. In this
1782 // mode, LFTR can ignore IV overflow and truncate to the width of
1783 // BECount. This avoids materializing the add(zext(add)) expression.
1784 Type *CntTy = DisableIVRewrite ?
1785 BackedgeTakenCount->getType() : IndVar->getType();
1787 const SCEV *IVLimit = BackedgeTakenCount;
1789 // If the exiting block is not the same as the backedge block, we must compare
1790 // against the preincremented value, otherwise we prefer to compare against
1791 // the post-incremented value.
1793 if (L->getExitingBlock() == L->getLoopLatch()) {
1794 // Add one to the "backedge-taken" count to get the trip count.
1795 // If this addition may overflow, we have to be more pessimistic and
1796 // cast the induction variable before doing the add.
1798 SE->getAddExpr(IVLimit, SE->getConstant(IVLimit->getType(), 1));
1799 if (CntTy == IVLimit->getType())
1802 const SCEV *Zero = SE->getConstant(IVLimit->getType(), 0);
1803 if ((isa<SCEVConstant>(N) && !N->isZero()) ||
1804 SE->isLoopEntryGuardedByCond(L, ICmpInst::ICMP_NE, N, Zero)) {
1805 // No overflow. Cast the sum.
1806 IVLimit = SE->getTruncateOrZeroExtend(N, CntTy);
1808 // Potential overflow. Cast before doing the add.
1809 IVLimit = SE->getTruncateOrZeroExtend(IVLimit, CntTy);
1810 IVLimit = SE->getAddExpr(IVLimit, SE->getConstant(CntTy, 1));
1813 // The BackedgeTaken expression contains the number of times that the
1814 // backedge branches to the loop header. This is one less than the
1815 // number of times the loop executes, so use the incremented indvar.
1816 CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock());
1818 // We have to use the preincremented value...
1819 IVLimit = SE->getTruncateOrZeroExtend(IVLimit, CntTy);
1823 // For unit stride, IVLimit = Start + BECount with 2's complement overflow.
1824 // So for, non-zero start compute the IVLimit here.
1825 bool isPtrIV = false;
1826 Type *CmpTy = CntTy;
1827 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
1828 assert(AR && AR->getLoop() == L && AR->isAffine() && "bad loop counter");
1829 if (!AR->getStart()->isZero()) {
1830 assert(AR->getStepRecurrence(*SE)->isOne() && "only handles unit stride");
1831 const SCEV *IVInit = AR->getStart();
1833 // For pointer types, sign extend BECount in order to materialize a GEP.
1834 // Note that for DisableIVRewrite, we never run SCEVExpander on a
1835 // pointer type, because we must preserve the existing GEPs. Instead we
1836 // directly generate a GEP later.
1837 if (IVInit->getType()->isPointerTy()) {
1839 CmpTy = SE->getEffectiveSCEVType(IVInit->getType());
1840 IVLimit = SE->getTruncateOrSignExtend(IVLimit, CmpTy);
1842 // For integer types, truncate the IV before computing IVInit + BECount.
1844 if (SE->getTypeSizeInBits(IVInit->getType())
1845 > SE->getTypeSizeInBits(CmpTy))
1846 IVInit = SE->getTruncateExpr(IVInit, CmpTy);
1848 IVLimit = SE->getAddExpr(IVInit, IVLimit);
1851 // Expand the code for the iteration count.
1852 IRBuilder<> Builder(BI);
1854 assert(SE->isLoopInvariant(IVLimit, L) &&
1855 "Computed iteration count is not loop invariant!");
1856 Value *ExitCnt = Rewriter.expandCodeFor(IVLimit, CmpTy, BI);
1858 // Create a gep for IVInit + IVLimit from on an existing pointer base.
1859 assert(isPtrIV == IndVar->getType()->isPointerTy() &&
1860 "IndVar type must match IVInit type");
1862 Value *IVStart = IndVar->getIncomingValueForBlock(L->getLoopPreheader());
1863 assert(AR->getStart() == SE->getSCEV(IVStart) && "bad loop counter");
1864 assert(SE->getSizeOfExpr(
1865 cast<PointerType>(IVStart->getType())->getElementType())->isOne()
1866 && "unit stride pointer IV must be i8*");
1868 Builder.SetInsertPoint(L->getLoopPreheader()->getTerminator());
1869 ExitCnt = Builder.CreateGEP(IVStart, ExitCnt, "lftr.limit");
1870 Builder.SetInsertPoint(BI);
1873 // Insert a new icmp_ne or icmp_eq instruction before the branch.
1874 ICmpInst::Predicate P;
1875 if (L->contains(BI->getSuccessor(0)))
1876 P = ICmpInst::ICMP_NE;
1878 P = ICmpInst::ICMP_EQ;
1880 DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
1881 << " LHS:" << *CmpIndVar << '\n'
1883 << (P == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
1884 << " RHS:\t" << *ExitCnt << "\n"
1885 << " Expr:\t" << *IVLimit << "\n");
1887 if (SE->getTypeSizeInBits(CmpIndVar->getType())
1888 > SE->getTypeSizeInBits(CmpTy)) {
1889 CmpIndVar = Builder.CreateTrunc(CmpIndVar, CmpTy, "lftr.wideiv");
1892 Value *Cond = Builder.CreateICmp(P, CmpIndVar, ExitCnt, "exitcond");
1893 Value *OrigCond = BI->getCondition();
1894 // It's tempting to use replaceAllUsesWith here to fully replace the old
1895 // comparison, but that's not immediately safe, since users of the old
1896 // comparison may not be dominated by the new comparison. Instead, just
1897 // update the branch to use the new comparison; in the common case this
1898 // will make old comparison dead.
1899 BI->setCondition(Cond);
1900 DeadInsts.push_back(OrigCond);
1907 //===----------------------------------------------------------------------===//
1908 // SinkUnusedInvariants. A late subpass to cleanup loop preheaders.
1909 //===----------------------------------------------------------------------===//
1911 /// If there's a single exit block, sink any loop-invariant values that
1912 /// were defined in the preheader but not used inside the loop into the
1913 /// exit block to reduce register pressure in the loop.
1914 void IndVarSimplify::SinkUnusedInvariants(Loop *L) {
1915 BasicBlock *ExitBlock = L->getExitBlock();
1916 if (!ExitBlock) return;
1918 BasicBlock *Preheader = L->getLoopPreheader();
1919 if (!Preheader) return;
1921 Instruction *InsertPt = ExitBlock->getFirstNonPHI();
1922 BasicBlock::iterator I = Preheader->getTerminator();
1923 while (I != Preheader->begin()) {
1925 // New instructions were inserted at the end of the preheader.
1926 if (isa<PHINode>(I))
1929 // Don't move instructions which might have side effects, since the side
1930 // effects need to complete before instructions inside the loop. Also don't
1931 // move instructions which might read memory, since the loop may modify
1932 // memory. Note that it's okay if the instruction might have undefined
1933 // behavior: LoopSimplify guarantees that the preheader dominates the exit
1935 if (I->mayHaveSideEffects() || I->mayReadFromMemory())
1938 // Skip debug info intrinsics.
1939 if (isa<DbgInfoIntrinsic>(I))
1942 // Don't sink static AllocaInsts out of the entry block, which would
1943 // turn them into dynamic allocas!
1944 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
1945 if (AI->isStaticAlloca())
1948 // Determine if there is a use in or before the loop (direct or
1950 bool UsedInLoop = false;
1951 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
1954 BasicBlock *UseBB = cast<Instruction>(U)->getParent();
1955 if (PHINode *P = dyn_cast<PHINode>(U)) {
1957 PHINode::getIncomingValueNumForOperand(UI.getOperandNo());
1958 UseBB = P->getIncomingBlock(i);
1960 if (UseBB == Preheader || L->contains(UseBB)) {
1966 // If there is, the def must remain in the preheader.
1970 // Otherwise, sink it to the exit block.
1971 Instruction *ToMove = I;
1974 if (I != Preheader->begin()) {
1975 // Skip debug info intrinsics.
1978 } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin());
1980 if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin())
1986 ToMove->moveBefore(InsertPt);
1992 //===----------------------------------------------------------------------===//
1993 // IndVarSimplify driver. Manage several subpasses of IV simplification.
1994 //===----------------------------------------------------------------------===//
1996 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
1997 // If LoopSimplify form is not available, stay out of trouble. Some notes:
1998 // - LSR currently only supports LoopSimplify-form loops. Indvars'
1999 // canonicalization can be a pessimization without LSR to "clean up"
2001 // - We depend on having a preheader; in particular,
2002 // Loop::getCanonicalInductionVariable only supports loops with preheaders,
2003 // and we're in trouble if we can't find the induction variable even when
2004 // we've manually inserted one.
2005 if (!L->isLoopSimplifyForm())
2008 if (!DisableIVRewrite)
2009 IU = &getAnalysis<IVUsers>();
2010 LI = &getAnalysis<LoopInfo>();
2011 SE = &getAnalysis<ScalarEvolution>();
2012 DT = &getAnalysis<DominatorTree>();
2013 TD = getAnalysisIfAvailable<TargetData>();
2018 // If there are any floating-point recurrences, attempt to
2019 // transform them to use integer recurrences.
2020 RewriteNonIntegerIVs(L);
2022 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2024 // Create a rewriter object which we'll use to transform the code with.
2025 SCEVExpander Rewriter(*SE, "indvars");
2027 // Eliminate redundant IV users.
2029 // Simplification works best when run before other consumers of SCEV. We
2030 // attempt to avoid evaluating SCEVs for sign/zero extend operations until
2031 // other expressions involving loop IVs have been evaluated. This helps SCEV
2032 // set no-wrap flags before normalizing sign/zero extension.
2033 if (DisableIVRewrite) {
2034 Rewriter.disableCanonicalMode();
2035 SimplifyIVUsersNoRewrite(L, Rewriter);
2038 // Check to see if this loop has a computable loop-invariant execution count.
2039 // If so, this means that we can compute the final value of any expressions
2040 // that are recurrent in the loop, and substitute the exit values from the
2041 // loop into any instructions outside of the loop that use the final values of
2042 // the current expressions.
2044 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2045 RewriteLoopExitValues(L, Rewriter);
2047 // Eliminate redundant IV users.
2048 if (!DisableIVRewrite)
2049 SimplifyIVUsers(Rewriter);
2051 // Eliminate redundant IV cycles.
2052 if (DisableIVRewrite)
2053 SimplifyCongruentIVs(L);
2055 // Compute the type of the largest recurrence expression, and decide whether
2056 // a canonical induction variable should be inserted.
2057 Type *LargestType = 0;
2058 bool NeedCannIV = false;
2059 bool ReuseIVForExit = DisableIVRewrite && !ForceLFTR;
2060 bool ExpandBECount = canExpandBackedgeTakenCount(L, SE);
2061 if (ExpandBECount && !ReuseIVForExit) {
2062 // If we have a known trip count and a single exit block, we'll be
2063 // rewriting the loop exit test condition below, which requires a
2064 // canonical induction variable.
2066 Type *Ty = BackedgeTakenCount->getType();
2067 if (DisableIVRewrite) {
2068 // In this mode, SimplifyIVUsers may have already widened the IV used by
2069 // the backedge test and inserted a Trunc on the compare's operand. Get
2070 // the wider type to avoid creating a redundant narrow IV only used by the
2072 LargestType = getBackedgeIVType(L);
2075 SE->getTypeSizeInBits(Ty) >
2076 SE->getTypeSizeInBits(LargestType))
2077 LargestType = SE->getEffectiveSCEVType(Ty);
2079 if (!DisableIVRewrite) {
2080 for (IVUsers::const_iterator I = IU->begin(), E = IU->end(); I != E; ++I) {
2083 SE->getEffectiveSCEVType(I->getOperandValToReplace()->getType());
2085 SE->getTypeSizeInBits(Ty) >
2086 SE->getTypeSizeInBits(LargestType))
2091 // Now that we know the largest of the induction variable expressions
2092 // in this loop, insert a canonical induction variable of the largest size.
2093 PHINode *IndVar = 0;
2095 // Check to see if the loop already has any canonical-looking induction
2096 // variables. If any are present and wider than the planned canonical
2097 // induction variable, temporarily remove them, so that the Rewriter
2098 // doesn't attempt to reuse them.
2099 SmallVector<PHINode *, 2> OldCannIVs;
2100 while (PHINode *OldCannIV = L->getCanonicalInductionVariable()) {
2101 if (SE->getTypeSizeInBits(OldCannIV->getType()) >
2102 SE->getTypeSizeInBits(LargestType))
2103 OldCannIV->removeFromParent();
2106 OldCannIVs.push_back(OldCannIV);
2109 IndVar = Rewriter.getOrInsertCanonicalInductionVariable(L, LargestType);
2113 DEBUG(dbgs() << "INDVARS: New CanIV: " << *IndVar << '\n');
2115 // Now that the official induction variable is established, reinsert
2116 // any old canonical-looking variables after it so that the IR remains
2117 // consistent. They will be deleted as part of the dead-PHI deletion at
2118 // the end of the pass.
2119 while (!OldCannIVs.empty()) {
2120 PHINode *OldCannIV = OldCannIVs.pop_back_val();
2121 OldCannIV->insertBefore(L->getHeader()->getFirstNonPHI());
2124 else if (ExpandBECount && ReuseIVForExit && needsLFTR(L, DT)) {
2125 IndVar = FindLoopCounter(L, BackedgeTakenCount, SE, DT, TD);
2127 // If we have a trip count expression, rewrite the loop's exit condition
2128 // using it. We can currently only handle loops with a single exit.
2130 if (ExpandBECount && IndVar) {
2131 // Check preconditions for proper SCEVExpander operation. SCEV does not
2132 // express SCEVExpander's dependencies, such as LoopSimplify. Instead any
2133 // pass that uses the SCEVExpander must do it. This does not work well for
2134 // loop passes because SCEVExpander makes assumptions about all loops, while
2135 // LoopPassManager only forces the current loop to be simplified.
2137 // FIXME: SCEV expansion has no way to bail out, so the caller must
2138 // explicitly check any assumptions made by SCEV. Brittle.
2139 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(BackedgeTakenCount);
2140 if (!AR || AR->getLoop()->getLoopPreheader())
2142 LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar, Rewriter);
2144 // Rewrite IV-derived expressions.
2145 if (!DisableIVRewrite)
2146 RewriteIVExpressions(L, Rewriter);
2148 // Clear the rewriter cache, because values that are in the rewriter's cache
2149 // can be deleted in the loop below, causing the AssertingVH in the cache to
2153 // Now that we're done iterating through lists, clean up any instructions
2154 // which are now dead.
2155 while (!DeadInsts.empty())
2156 if (Instruction *Inst =
2157 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
2158 RecursivelyDeleteTriviallyDeadInstructions(Inst);
2160 // The Rewriter may not be used from this point on.
2162 // Loop-invariant instructions in the preheader that aren't used in the
2163 // loop may be sunk below the loop to reduce register pressure.
2164 SinkUnusedInvariants(L);
2166 // For completeness, inform IVUsers of the IV use in the newly-created
2167 // loop exit test instruction.
2168 if (IU && NewICmp) {
2169 ICmpInst *NewICmpInst = dyn_cast<ICmpInst>(NewICmp);
2171 IU->AddUsersIfInteresting(cast<Instruction>(NewICmpInst->getOperand(0)));
2173 // Clean up dead instructions.
2174 Changed |= DeleteDeadPHIs(L->getHeader());
2175 // Check a post-condition.
2176 assert(L->isLCSSAForm(*DT) &&
2177 "Indvars did not leave the loop in lcssa form!");
2179 // Verify that LFTR, and any other change have not interfered with SCEV's
2180 // ability to compute trip count.
2182 if (DisableIVRewrite && !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) {
2184 const SCEV *NewBECount = SE->getBackedgeTakenCount(L);
2185 if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) <
2186 SE->getTypeSizeInBits(NewBECount->getType()))
2187 NewBECount = SE->getTruncateOrNoop(NewBECount,
2188 BackedgeTakenCount->getType());
2190 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount,
2191 NewBECount->getType());
2192 assert(BackedgeTakenCount == NewBECount && "indvars must preserve SCEV");