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/ScalarEvolutionExpander.h"
37 #include "llvm/Analysis/LoopInfo.h"
38 #include "llvm/Analysis/LoopPass.h"
39 #include "llvm/Support/CFG.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/raw_ostream.h"
43 #include "llvm/Transforms/Utils/Local.h"
44 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
45 #include "llvm/Transforms/Utils/SimplifyIndVar.h"
46 #include "llvm/DataLayout.h"
47 #include "llvm/Target/TargetLibraryInfo.h"
48 #include "llvm/ADT/DenseMap.h"
49 #include "llvm/ADT/SmallVector.h"
50 #include "llvm/ADT/Statistic.h"
53 STATISTIC(NumWidened , "Number of indvars widened");
54 STATISTIC(NumReplaced , "Number of exit values replaced");
55 STATISTIC(NumLFTR , "Number of loop exit tests replaced");
56 STATISTIC(NumElimExt , "Number of IV sign/zero extends eliminated");
57 STATISTIC(NumElimIV , "Number of congruent IVs eliminated");
59 // Trip count verification can be enabled by default under NDEBUG if we
60 // implement a strong expression equivalence checker in SCEV. Until then, we
61 // use the verify-indvars flag, which may assert in some cases.
62 static cl::opt<bool> VerifyIndvars(
63 "verify-indvars", cl::Hidden,
64 cl::desc("Verify the ScalarEvolution result after running indvars"));
67 class IndVarSimplify : public LoopPass {
72 TargetLibraryInfo *TLI;
74 SmallVector<WeakVH, 16> DeadInsts;
78 static char ID; // Pass identification, replacement for typeid
79 IndVarSimplify() : LoopPass(ID), LI(0), SE(0), DT(0), TD(0),
81 initializeIndVarSimplifyPass(*PassRegistry::getPassRegistry());
84 virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
86 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
87 AU.addRequired<DominatorTree>();
88 AU.addRequired<LoopInfo>();
89 AU.addRequired<ScalarEvolution>();
90 AU.addRequiredID(LoopSimplifyID);
91 AU.addRequiredID(LCSSAID);
92 AU.addPreserved<ScalarEvolution>();
93 AU.addPreservedID(LoopSimplifyID);
94 AU.addPreservedID(LCSSAID);
99 virtual void releaseMemory() {
103 bool isValidRewrite(Value *FromVal, Value *ToVal);
105 void HandleFloatingPointIV(Loop *L, PHINode *PH);
106 void RewriteNonIntegerIVs(Loop *L);
108 void SimplifyAndExtend(Loop *L, SCEVExpander &Rewriter, LPPassManager &LPM);
110 void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter);
112 Value *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
113 PHINode *IndVar, SCEVExpander &Rewriter);
115 void SinkUnusedInvariants(Loop *L);
119 char IndVarSimplify::ID = 0;
120 INITIALIZE_PASS_BEGIN(IndVarSimplify, "indvars",
121 "Induction Variable Simplification", false, false)
122 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
123 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
124 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
125 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
126 INITIALIZE_PASS_DEPENDENCY(LCSSA)
127 INITIALIZE_PASS_END(IndVarSimplify, "indvars",
128 "Induction Variable Simplification", false, false)
130 Pass *llvm::createIndVarSimplifyPass() {
131 return new IndVarSimplify();
134 /// isValidRewrite - Return true if the SCEV expansion generated by the
135 /// rewriter can replace the original value. SCEV guarantees that it
136 /// produces the same value, but the way it is produced may be illegal IR.
137 /// Ideally, this function will only be called for verification.
138 bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) {
139 // If an SCEV expression subsumed multiple pointers, its expansion could
140 // reassociate the GEP changing the base pointer. This is illegal because the
141 // final address produced by a GEP chain must be inbounds relative to its
142 // underlying object. Otherwise basic alias analysis, among other things,
143 // could fail in a dangerous way. Ultimately, SCEV will be improved to avoid
144 // producing an expression involving multiple pointers. Until then, we must
147 // Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject
148 // because it understands lcssa phis while SCEV does not.
149 Value *FromPtr = FromVal;
150 Value *ToPtr = ToVal;
151 if (GEPOperator *GEP = dyn_cast<GEPOperator>(FromVal)) {
152 FromPtr = GEP->getPointerOperand();
154 if (GEPOperator *GEP = dyn_cast<GEPOperator>(ToVal)) {
155 ToPtr = GEP->getPointerOperand();
157 if (FromPtr != FromVal || ToPtr != ToVal) {
158 // Quickly check the common case
159 if (FromPtr == ToPtr)
162 // SCEV may have rewritten an expression that produces the GEP's pointer
163 // operand. That's ok as long as the pointer operand has the same base
164 // pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the
165 // base of a recurrence. This handles the case in which SCEV expansion
166 // converts a pointer type recurrence into a nonrecurrent pointer base
167 // indexed by an integer recurrence.
169 // If the GEP base pointer is a vector of pointers, abort.
170 if (!FromPtr->getType()->isPointerTy() || !ToPtr->getType()->isPointerTy())
173 const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr));
174 const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr));
175 if (FromBase == ToBase)
178 DEBUG(dbgs() << "INDVARS: GEP rewrite bail out "
179 << *FromBase << " != " << *ToBase << "\n");
186 /// Determine the insertion point for this user. By default, insert immediately
187 /// before the user. SCEVExpander or LICM will hoist loop invariants out of the
188 /// loop. For PHI nodes, there may be multiple uses, so compute the nearest
189 /// common dominator for the incoming blocks.
190 static Instruction *getInsertPointForUses(Instruction *User, Value *Def,
192 PHINode *PHI = dyn_cast<PHINode>(User);
196 Instruction *InsertPt = 0;
197 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) {
198 if (PHI->getIncomingValue(i) != Def)
201 BasicBlock *InsertBB = PHI->getIncomingBlock(i);
203 InsertPt = InsertBB->getTerminator();
206 InsertBB = DT->findNearestCommonDominator(InsertPt->getParent(), InsertBB);
207 InsertPt = InsertBB->getTerminator();
209 assert(InsertPt && "Missing phi operand");
210 assert((!isa<Instruction>(Def) ||
211 DT->dominates(cast<Instruction>(Def), InsertPt)) &&
212 "def does not dominate all uses");
216 //===----------------------------------------------------------------------===//
217 // RewriteNonIntegerIVs and helpers. Prefer integer IVs.
218 //===----------------------------------------------------------------------===//
220 /// ConvertToSInt - Convert APF to an integer, if possible.
221 static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) {
222 bool isExact = false;
223 if (&APF.getSemantics() == &APFloat::PPCDoubleDouble)
225 // See if we can convert this to an int64_t
227 if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero,
228 &isExact) != APFloat::opOK || !isExact)
234 /// HandleFloatingPointIV - If the loop has floating induction variable
235 /// then insert corresponding integer induction variable if possible.
237 /// for(double i = 0; i < 10000; ++i)
239 /// is converted into
240 /// for(int i = 0; i < 10000; ++i)
243 void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) {
244 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
245 unsigned BackEdge = IncomingEdge^1;
247 // Check incoming value.
248 ConstantFP *InitValueVal =
249 dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));
252 if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
255 // Check IV increment. Reject this PN if increment operation is not
256 // an add or increment value can not be represented by an integer.
257 BinaryOperator *Incr =
258 dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
259 if (Incr == 0 || Incr->getOpcode() != Instruction::FAdd) return;
261 // If this is not an add of the PHI with a constantfp, or if the constant fp
262 // is not an integer, bail out.
263 ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
265 if (IncValueVal == 0 || Incr->getOperand(0) != PN ||
266 !ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
269 // Check Incr uses. One user is PN and the other user is an exit condition
270 // used by the conditional terminator.
271 Value::use_iterator IncrUse = Incr->use_begin();
272 Instruction *U1 = cast<Instruction>(*IncrUse++);
273 if (IncrUse == Incr->use_end()) return;
274 Instruction *U2 = cast<Instruction>(*IncrUse++);
275 if (IncrUse != Incr->use_end()) return;
277 // Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't
278 // only used by a branch, we can't transform it.
279 FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
281 Compare = dyn_cast<FCmpInst>(U2);
282 if (Compare == 0 || !Compare->hasOneUse() ||
283 !isa<BranchInst>(Compare->use_back()))
286 BranchInst *TheBr = cast<BranchInst>(Compare->use_back());
288 // We need to verify that the branch actually controls the iteration count
289 // of the loop. If not, the new IV can overflow and no one will notice.
290 // The branch block must be in the loop and one of the successors must be out
292 assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
293 if (!L->contains(TheBr->getParent()) ||
294 (L->contains(TheBr->getSuccessor(0)) &&
295 L->contains(TheBr->getSuccessor(1))))
299 // If it isn't a comparison with an integer-as-fp (the exit value), we can't
301 ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
303 if (ExitValueVal == 0 ||
304 !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
307 // Find new predicate for integer comparison.
308 CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
309 switch (Compare->getPredicate()) {
310 default: return; // Unknown comparison.
311 case CmpInst::FCMP_OEQ:
312 case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
313 case CmpInst::FCMP_ONE:
314 case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
315 case CmpInst::FCMP_OGT:
316 case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
317 case CmpInst::FCMP_OGE:
318 case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
319 case CmpInst::FCMP_OLT:
320 case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
321 case CmpInst::FCMP_OLE:
322 case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
325 // We convert the floating point induction variable to a signed i32 value if
326 // we can. This is only safe if the comparison will not overflow in a way
327 // that won't be trapped by the integer equivalent operations. Check for this
329 // TODO: We could use i64 if it is native and the range requires it.
331 // The start/stride/exit values must all fit in signed i32.
332 if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
335 // If not actually striding (add x, 0.0), avoid touching the code.
339 // Positive and negative strides have different safety conditions.
341 // If we have a positive stride, we require the init to be less than the
343 if (InitValue >= ExitValue)
346 uint32_t Range = uint32_t(ExitValue-InitValue);
347 // Check for infinite loop, either:
348 // while (i <= Exit) or until (i > Exit)
349 if (NewPred == CmpInst::ICMP_SLE || NewPred == CmpInst::ICMP_SGT) {
350 if (++Range == 0) return; // Range overflows.
353 unsigned Leftover = Range % uint32_t(IncValue);
355 // If this is an equality comparison, we require that the strided value
356 // exactly land on the exit value, otherwise the IV condition will wrap
357 // around and do things the fp IV wouldn't.
358 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
362 // If the stride would wrap around the i32 before exiting, we can't
364 if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
368 // If we have a negative stride, we require the init to be greater than the
370 if (InitValue <= ExitValue)
373 uint32_t Range = uint32_t(InitValue-ExitValue);
374 // Check for infinite loop, either:
375 // while (i >= Exit) or until (i < Exit)
376 if (NewPred == CmpInst::ICMP_SGE || NewPred == CmpInst::ICMP_SLT) {
377 if (++Range == 0) return; // Range overflows.
380 unsigned Leftover = Range % uint32_t(-IncValue);
382 // If this is an equality comparison, we require that the strided value
383 // exactly land on the exit value, otherwise the IV condition will wrap
384 // around and do things the fp IV wouldn't.
385 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
389 // If the stride would wrap around the i32 before exiting, we can't
391 if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
395 IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
397 // Insert new integer induction variable.
398 PHINode *NewPHI = PHINode::Create(Int32Ty, 2, PN->getName()+".int", PN);
399 NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
400 PN->getIncomingBlock(IncomingEdge));
403 BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
404 Incr->getName()+".int", Incr);
405 NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
407 ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
408 ConstantInt::get(Int32Ty, ExitValue),
411 // In the following deletions, PN may become dead and may be deleted.
412 // Use a WeakVH to observe whether this happens.
415 // Delete the old floating point exit comparison. The branch starts using the
417 NewCompare->takeName(Compare);
418 Compare->replaceAllUsesWith(NewCompare);
419 RecursivelyDeleteTriviallyDeadInstructions(Compare, TLI);
421 // Delete the old floating point increment.
422 Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
423 RecursivelyDeleteTriviallyDeadInstructions(Incr, TLI);
425 // If the FP induction variable still has uses, this is because something else
426 // in the loop uses its value. In order to canonicalize the induction
427 // variable, we chose to eliminate the IV and rewrite it in terms of an
430 // We give preference to sitofp over uitofp because it is faster on most
433 Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
434 PN->getParent()->getFirstInsertionPt());
435 PN->replaceAllUsesWith(Conv);
436 RecursivelyDeleteTriviallyDeadInstructions(PN, TLI);
441 void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
442 // First step. Check to see if there are any floating-point recurrences.
443 // If there are, change them into integer recurrences, permitting analysis by
444 // the SCEV routines.
446 BasicBlock *Header = L->getHeader();
448 SmallVector<WeakVH, 8> PHIs;
449 for (BasicBlock::iterator I = Header->begin();
450 PHINode *PN = dyn_cast<PHINode>(I); ++I)
453 for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
454 if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i]))
455 HandleFloatingPointIV(L, PN);
457 // If the loop previously had floating-point IV, ScalarEvolution
458 // may not have been able to compute a trip count. Now that we've done some
459 // re-writing, the trip count may be computable.
464 //===----------------------------------------------------------------------===//
465 // RewriteLoopExitValues - Optimize IV users outside the loop.
466 // As a side effect, reduces the amount of IV processing within the loop.
467 //===----------------------------------------------------------------------===//
469 /// RewriteLoopExitValues - Check to see if this loop has a computable
470 /// loop-invariant execution count. If so, this means that we can compute the
471 /// final value of any expressions that are recurrent in the loop, and
472 /// substitute the exit values from the loop into any instructions outside of
473 /// the loop that use the final values of the current expressions.
475 /// This is mostly redundant with the regular IndVarSimplify activities that
476 /// happen later, except that it's more powerful in some cases, because it's
477 /// able to brute-force evaluate arbitrary instructions as long as they have
478 /// constant operands at the beginning of the loop.
479 void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) {
480 // Verify the input to the pass in already in LCSSA form.
481 assert(L->isLCSSAForm(*DT));
483 SmallVector<BasicBlock*, 8> ExitBlocks;
484 L->getUniqueExitBlocks(ExitBlocks);
486 // Find all values that are computed inside the loop, but used outside of it.
487 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan
488 // the exit blocks of the loop to find them.
489 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
490 BasicBlock *ExitBB = ExitBlocks[i];
492 // If there are no PHI nodes in this exit block, then no values defined
493 // inside the loop are used on this path, skip it.
494 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
497 unsigned NumPreds = PN->getNumIncomingValues();
499 // Iterate over all of the PHI nodes.
500 BasicBlock::iterator BBI = ExitBB->begin();
501 while ((PN = dyn_cast<PHINode>(BBI++))) {
503 continue; // dead use, don't replace it
505 // SCEV only supports integer expressions for now.
506 if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy())
509 // It's necessary to tell ScalarEvolution about this explicitly so that
510 // it can walk the def-use list and forget all SCEVs, as it may not be
511 // watching the PHI itself. Once the new exit value is in place, there
512 // may not be a def-use connection between the loop and every instruction
513 // which got a SCEVAddRecExpr for that loop.
516 // Iterate over all of the values in all the PHI nodes.
517 for (unsigned i = 0; i != NumPreds; ++i) {
518 // If the value being merged in is not integer or is not defined
519 // in the loop, skip it.
520 Value *InVal = PN->getIncomingValue(i);
521 if (!isa<Instruction>(InVal))
524 // If this pred is for a subloop, not L itself, skip it.
525 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
526 continue; // The Block is in a subloop, skip it.
528 // Check that InVal is defined in the loop.
529 Instruction *Inst = cast<Instruction>(InVal);
530 if (!L->contains(Inst))
533 // Okay, this instruction has a user outside of the current loop
534 // and varies predictably *inside* the loop. Evaluate the value it
535 // contains when the loop exits, if possible.
536 const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
537 if (!SE->isLoopInvariant(ExitValue, L))
540 Value *ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), Inst);
542 DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
543 << " LoopVal = " << *Inst << "\n");
545 if (!isValidRewrite(Inst, ExitVal)) {
546 DeadInsts.push_back(ExitVal);
552 PN->setIncomingValue(i, ExitVal);
554 // If this instruction is dead now, delete it. Don't do it now to avoid
555 // invalidating iterators.
556 if (isInstructionTriviallyDead(Inst, TLI))
557 DeadInsts.push_back(Inst);
560 // Completely replace a single-pred PHI. This is safe, because the
561 // NewVal won't be variant in the loop, so we don't need an LCSSA phi
563 PN->replaceAllUsesWith(ExitVal);
564 PN->eraseFromParent();
568 // Clone the PHI and delete the original one. This lets IVUsers and
569 // any other maps purge the original user from their records.
570 PHINode *NewPN = cast<PHINode>(PN->clone());
572 NewPN->insertBefore(PN);
573 PN->replaceAllUsesWith(NewPN);
574 PN->eraseFromParent();
579 // The insertion point instruction may have been deleted; clear it out
580 // so that the rewriter doesn't trip over it later.
581 Rewriter.clearInsertPoint();
584 //===----------------------------------------------------------------------===//
585 // IV Widening - Extend the width of an IV to cover its widest uses.
586 //===----------------------------------------------------------------------===//
589 // Collect information about induction variables that are used by sign/zero
590 // extend operations. This information is recorded by CollectExtend and
591 // provides the input to WidenIV.
594 Type *WidestNativeType; // Widest integer type created [sz]ext
595 bool IsSigned; // Was an sext user seen before a zext?
597 WideIVInfo() : NarrowIV(0), WidestNativeType(0), IsSigned(false) {}
600 class WideIVVisitor : public IVVisitor {
602 const DataLayout *TD;
607 WideIVVisitor(PHINode *NarrowIV, ScalarEvolution *SCEV,
608 const DataLayout *TData) :
609 SE(SCEV), TD(TData) { WI.NarrowIV = NarrowIV; }
611 // Implement the interface used by simplifyUsersOfIV.
612 virtual void visitCast(CastInst *Cast);
616 /// visitCast - Update information about the induction variable that is
617 /// extended by this sign or zero extend operation. This is used to determine
618 /// the final width of the IV before actually widening it.
619 void WideIVVisitor::visitCast(CastInst *Cast) {
620 bool IsSigned = Cast->getOpcode() == Instruction::SExt;
621 if (!IsSigned && Cast->getOpcode() != Instruction::ZExt)
624 Type *Ty = Cast->getType();
625 uint64_t Width = SE->getTypeSizeInBits(Ty);
626 if (TD && !TD->isLegalInteger(Width))
629 if (!WI.WidestNativeType) {
630 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
631 WI.IsSigned = IsSigned;
635 // We extend the IV to satisfy the sign of its first user, arbitrarily.
636 if (WI.IsSigned != IsSigned)
639 if (Width > SE->getTypeSizeInBits(WI.WidestNativeType))
640 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
645 /// NarrowIVDefUse - Record a link in the Narrow IV def-use chain along with the
646 /// WideIV that computes the same value as the Narrow IV def. This avoids
647 /// caching Use* pointers.
648 struct NarrowIVDefUse {
649 Instruction *NarrowDef;
650 Instruction *NarrowUse;
651 Instruction *WideDef;
653 NarrowIVDefUse(): NarrowDef(0), NarrowUse(0), WideDef(0) {}
655 NarrowIVDefUse(Instruction *ND, Instruction *NU, Instruction *WD):
656 NarrowDef(ND), NarrowUse(NU), WideDef(WD) {}
659 /// WidenIV - The goal of this transform is to remove sign and zero extends
660 /// without creating any new induction variables. To do this, it creates a new
661 /// phi of the wider type and redirects all users, either removing extends or
662 /// inserting truncs whenever we stop propagating the type.
678 Instruction *WideInc;
679 const SCEV *WideIncExpr;
680 SmallVectorImpl<WeakVH> &DeadInsts;
682 SmallPtrSet<Instruction*,16> Widened;
683 SmallVector<NarrowIVDefUse, 8> NarrowIVUsers;
686 WidenIV(const WideIVInfo &WI, LoopInfo *LInfo,
687 ScalarEvolution *SEv, DominatorTree *DTree,
688 SmallVectorImpl<WeakVH> &DI) :
689 OrigPhi(WI.NarrowIV),
690 WideType(WI.WidestNativeType),
691 IsSigned(WI.IsSigned),
693 L(LI->getLoopFor(OrigPhi->getParent())),
700 assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV");
703 PHINode *CreateWideIV(SCEVExpander &Rewriter);
706 Value *getExtend(Value *NarrowOper, Type *WideType, bool IsSigned,
709 Instruction *CloneIVUser(NarrowIVDefUse DU);
711 const SCEVAddRecExpr *GetWideRecurrence(Instruction *NarrowUse);
713 const SCEVAddRecExpr* GetExtendedOperandRecurrence(NarrowIVDefUse DU);
715 Instruction *WidenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter);
717 void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef);
719 } // anonymous namespace
721 /// isLoopInvariant - Perform a quick domtree based check for loop invariance
722 /// assuming that V is used within the loop. LoopInfo::isLoopInvariant() seems
723 /// gratuitous for this purpose.
724 static bool isLoopInvariant(Value *V, const Loop *L, const DominatorTree *DT) {
725 Instruction *Inst = dyn_cast<Instruction>(V);
729 return DT->properlyDominates(Inst->getParent(), L->getHeader());
732 Value *WidenIV::getExtend(Value *NarrowOper, Type *WideType, bool IsSigned,
734 // Set the debug location and conservative insertion point.
735 IRBuilder<> Builder(Use);
736 // Hoist the insertion point into loop preheaders as far as possible.
737 for (const Loop *L = LI->getLoopFor(Use->getParent());
738 L && L->getLoopPreheader() && isLoopInvariant(NarrowOper, L, DT);
739 L = L->getParentLoop())
740 Builder.SetInsertPoint(L->getLoopPreheader()->getTerminator());
742 return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) :
743 Builder.CreateZExt(NarrowOper, WideType);
746 /// CloneIVUser - Instantiate a wide operation to replace a narrow
747 /// operation. This only needs to handle operations that can evaluation to
748 /// SCEVAddRec. It can safely return 0 for any operation we decide not to clone.
749 Instruction *WidenIV::CloneIVUser(NarrowIVDefUse DU) {
750 unsigned Opcode = DU.NarrowUse->getOpcode();
754 case Instruction::Add:
755 case Instruction::Mul:
756 case Instruction::UDiv:
757 case Instruction::Sub:
758 case Instruction::And:
759 case Instruction::Or:
760 case Instruction::Xor:
761 case Instruction::Shl:
762 case Instruction::LShr:
763 case Instruction::AShr:
764 DEBUG(dbgs() << "Cloning IVUser: " << *DU.NarrowUse << "\n");
766 // Replace NarrowDef operands with WideDef. Otherwise, we don't know
767 // anything about the narrow operand yet so must insert a [sz]ext. It is
768 // probably loop invariant and will be folded or hoisted. If it actually
769 // comes from a widened IV, it should be removed during a future call to
771 Value *LHS = (DU.NarrowUse->getOperand(0) == DU.NarrowDef) ? DU.WideDef :
772 getExtend(DU.NarrowUse->getOperand(0), WideType, IsSigned, DU.NarrowUse);
773 Value *RHS = (DU.NarrowUse->getOperand(1) == DU.NarrowDef) ? DU.WideDef :
774 getExtend(DU.NarrowUse->getOperand(1), WideType, IsSigned, DU.NarrowUse);
776 BinaryOperator *NarrowBO = cast<BinaryOperator>(DU.NarrowUse);
777 BinaryOperator *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(),
779 NarrowBO->getName());
780 IRBuilder<> Builder(DU.NarrowUse);
781 Builder.Insert(WideBO);
782 if (const OverflowingBinaryOperator *OBO =
783 dyn_cast<OverflowingBinaryOperator>(NarrowBO)) {
784 if (OBO->hasNoUnsignedWrap()) WideBO->setHasNoUnsignedWrap();
785 if (OBO->hasNoSignedWrap()) WideBO->setHasNoSignedWrap();
791 /// No-wrap operations can transfer sign extension of their result to their
792 /// operands. Generate the SCEV value for the widened operation without
793 /// actually modifying the IR yet. If the expression after extending the
794 /// operands is an AddRec for this loop, return it.
795 const SCEVAddRecExpr* WidenIV::GetExtendedOperandRecurrence(NarrowIVDefUse DU) {
796 // Handle the common case of add<nsw/nuw>
797 if (DU.NarrowUse->getOpcode() != Instruction::Add)
800 // One operand (NarrowDef) has already been extended to WideDef. Now determine
801 // if extending the other will lead to a recurrence.
802 unsigned ExtendOperIdx = DU.NarrowUse->getOperand(0) == DU.NarrowDef ? 1 : 0;
803 assert(DU.NarrowUse->getOperand(1-ExtendOperIdx) == DU.NarrowDef && "bad DU");
805 const SCEV *ExtendOperExpr = 0;
806 const OverflowingBinaryOperator *OBO =
807 cast<OverflowingBinaryOperator>(DU.NarrowUse);
808 if (IsSigned && OBO->hasNoSignedWrap())
809 ExtendOperExpr = SE->getSignExtendExpr(
810 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
811 else if(!IsSigned && OBO->hasNoUnsignedWrap())
812 ExtendOperExpr = SE->getZeroExtendExpr(
813 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
817 // When creating this AddExpr, don't apply the current operations NSW or NUW
818 // flags. This instruction may be guarded by control flow that the no-wrap
819 // behavior depends on. Non-control-equivalent instructions can be mapped to
820 // the same SCEV expression, and it would be incorrect to transfer NSW/NUW
821 // semantics to those operations.
822 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(
823 SE->getAddExpr(SE->getSCEV(DU.WideDef), ExtendOperExpr));
825 if (!AddRec || AddRec->getLoop() != L)
830 /// GetWideRecurrence - Is this instruction potentially interesting from
831 /// IVUsers' perspective after widening it's type? In other words, can the
832 /// extend be safely hoisted out of the loop with SCEV reducing the value to a
833 /// recurrence on the same loop. If so, return the sign or zero extended
834 /// recurrence. Otherwise return NULL.
835 const SCEVAddRecExpr *WidenIV::GetWideRecurrence(Instruction *NarrowUse) {
836 if (!SE->isSCEVable(NarrowUse->getType()))
839 const SCEV *NarrowExpr = SE->getSCEV(NarrowUse);
840 if (SE->getTypeSizeInBits(NarrowExpr->getType())
841 >= SE->getTypeSizeInBits(WideType)) {
842 // NarrowUse implicitly widens its operand. e.g. a gep with a narrow
843 // index. So don't follow this use.
847 const SCEV *WideExpr = IsSigned ?
848 SE->getSignExtendExpr(NarrowExpr, WideType) :
849 SE->getZeroExtendExpr(NarrowExpr, WideType);
850 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr);
851 if (!AddRec || AddRec->getLoop() != L)
856 /// WidenIVUse - Determine whether an individual user of the narrow IV can be
857 /// widened. If so, return the wide clone of the user.
858 Instruction *WidenIV::WidenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter) {
860 // Stop traversing the def-use chain at inner-loop phis or post-loop phis.
861 if (isa<PHINode>(DU.NarrowUse) &&
862 LI->getLoopFor(DU.NarrowUse->getParent()) != L)
865 // Our raison d'etre! Eliminate sign and zero extension.
866 if (IsSigned ? isa<SExtInst>(DU.NarrowUse) : isa<ZExtInst>(DU.NarrowUse)) {
867 Value *NewDef = DU.WideDef;
868 if (DU.NarrowUse->getType() != WideType) {
869 unsigned CastWidth = SE->getTypeSizeInBits(DU.NarrowUse->getType());
870 unsigned IVWidth = SE->getTypeSizeInBits(WideType);
871 if (CastWidth < IVWidth) {
872 // The cast isn't as wide as the IV, so insert a Trunc.
873 IRBuilder<> Builder(DU.NarrowUse);
874 NewDef = Builder.CreateTrunc(DU.WideDef, DU.NarrowUse->getType());
877 // A wider extend was hidden behind a narrower one. This may induce
878 // another round of IV widening in which the intermediate IV becomes
879 // dead. It should be very rare.
880 DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi
881 << " not wide enough to subsume " << *DU.NarrowUse << "\n");
882 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
883 NewDef = DU.NarrowUse;
886 if (NewDef != DU.NarrowUse) {
887 DEBUG(dbgs() << "INDVARS: eliminating " << *DU.NarrowUse
888 << " replaced by " << *DU.WideDef << "\n");
890 DU.NarrowUse->replaceAllUsesWith(NewDef);
891 DeadInsts.push_back(DU.NarrowUse);
893 // Now that the extend is gone, we want to expose it's uses for potential
894 // further simplification. We don't need to directly inform SimplifyIVUsers
895 // of the new users, because their parent IV will be processed later as a
896 // new loop phi. If we preserved IVUsers analysis, we would also want to
897 // push the uses of WideDef here.
899 // No further widening is needed. The deceased [sz]ext had done it for us.
903 // Does this user itself evaluate to a recurrence after widening?
904 const SCEVAddRecExpr *WideAddRec = GetWideRecurrence(DU.NarrowUse);
906 WideAddRec = GetExtendedOperandRecurrence(DU);
909 // This user does not evaluate to a recurence after widening, so don't
910 // follow it. Instead insert a Trunc to kill off the original use,
911 // eventually isolating the original narrow IV so it can be removed.
912 IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT));
913 Value *Trunc = Builder.CreateTrunc(DU.WideDef, DU.NarrowDef->getType());
914 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, Trunc);
917 // Assume block terminators cannot evaluate to a recurrence. We can't to
918 // insert a Trunc after a terminator if there happens to be a critical edge.
919 assert(DU.NarrowUse != DU.NarrowUse->getParent()->getTerminator() &&
920 "SCEV is not expected to evaluate a block terminator");
922 // Reuse the IV increment that SCEVExpander created as long as it dominates
924 Instruction *WideUse = 0;
925 if (WideAddRec == WideIncExpr
926 && Rewriter.hoistIVInc(WideInc, DU.NarrowUse))
929 WideUse = CloneIVUser(DU);
933 // Evaluation of WideAddRec ensured that the narrow expression could be
934 // extended outside the loop without overflow. This suggests that the wide use
935 // evaluates to the same expression as the extended narrow use, but doesn't
936 // absolutely guarantee it. Hence the following failsafe check. In rare cases
937 // where it fails, we simply throw away the newly created wide use.
938 if (WideAddRec != SE->getSCEV(WideUse)) {
939 DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse
940 << ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n");
941 DeadInsts.push_back(WideUse);
945 // Returning WideUse pushes it on the worklist.
949 /// pushNarrowIVUsers - Add eligible users of NarrowDef to NarrowIVUsers.
951 void WidenIV::pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef) {
952 for (Value::use_iterator UI = NarrowDef->use_begin(),
953 UE = NarrowDef->use_end(); UI != UE; ++UI) {
954 Instruction *NarrowUse = cast<Instruction>(*UI);
956 // Handle data flow merges and bizarre phi cycles.
957 if (!Widened.insert(NarrowUse))
960 NarrowIVUsers.push_back(NarrowIVDefUse(NarrowDef, NarrowUse, WideDef));
964 /// CreateWideIV - Process a single induction variable. First use the
965 /// SCEVExpander to create a wide induction variable that evaluates to the same
966 /// recurrence as the original narrow IV. Then use a worklist to forward
967 /// traverse the narrow IV's def-use chain. After WidenIVUse has processed all
968 /// interesting IV users, the narrow IV will be isolated for removal by
971 /// It would be simpler to delete uses as they are processed, but we must avoid
972 /// invalidating SCEV expressions.
974 PHINode *WidenIV::CreateWideIV(SCEVExpander &Rewriter) {
975 // Is this phi an induction variable?
976 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi));
980 // Widen the induction variable expression.
981 const SCEV *WideIVExpr = IsSigned ?
982 SE->getSignExtendExpr(AddRec, WideType) :
983 SE->getZeroExtendExpr(AddRec, WideType);
985 assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType &&
986 "Expect the new IV expression to preserve its type");
988 // Can the IV be extended outside the loop without overflow?
989 AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr);
990 if (!AddRec || AddRec->getLoop() != L)
993 // An AddRec must have loop-invariant operands. Since this AddRec is
994 // materialized by a loop header phi, the expression cannot have any post-loop
995 // operands, so they must dominate the loop header.
996 assert(SE->properlyDominates(AddRec->getStart(), L->getHeader()) &&
997 SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader())
998 && "Loop header phi recurrence inputs do not dominate the loop");
1000 // The rewriter provides a value for the desired IV expression. This may
1001 // either find an existing phi or materialize a new one. Either way, we
1002 // expect a well-formed cyclic phi-with-increments. i.e. any operand not part
1003 // of the phi-SCC dominates the loop entry.
1004 Instruction *InsertPt = L->getHeader()->begin();
1005 WidePhi = cast<PHINode>(Rewriter.expandCodeFor(AddRec, WideType, InsertPt));
1007 // Remembering the WideIV increment generated by SCEVExpander allows
1008 // WidenIVUse to reuse it when widening the narrow IV's increment. We don't
1009 // employ a general reuse mechanism because the call above is the only call to
1010 // SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses.
1011 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1013 cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock));
1014 WideIncExpr = SE->getSCEV(WideInc);
1017 DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n");
1020 // Traverse the def-use chain using a worklist starting at the original IV.
1021 assert(Widened.empty() && NarrowIVUsers.empty() && "expect initial state" );
1023 Widened.insert(OrigPhi);
1024 pushNarrowIVUsers(OrigPhi, WidePhi);
1026 while (!NarrowIVUsers.empty()) {
1027 NarrowIVDefUse DU = NarrowIVUsers.pop_back_val();
1029 // Process a def-use edge. This may replace the use, so don't hold a
1030 // use_iterator across it.
1031 Instruction *WideUse = WidenIVUse(DU, Rewriter);
1033 // Follow all def-use edges from the previous narrow use.
1035 pushNarrowIVUsers(DU.NarrowUse, WideUse);
1037 // WidenIVUse may have removed the def-use edge.
1038 if (DU.NarrowDef->use_empty())
1039 DeadInsts.push_back(DU.NarrowDef);
1044 //===----------------------------------------------------------------------===//
1045 // Simplification of IV users based on SCEV evaluation.
1046 //===----------------------------------------------------------------------===//
1049 /// SimplifyAndExtend - Iteratively perform simplification on a worklist of IV
1050 /// users. Each successive simplification may push more users which may
1051 /// themselves be candidates for simplification.
1053 /// Sign/Zero extend elimination is interleaved with IV simplification.
1055 void IndVarSimplify::SimplifyAndExtend(Loop *L,
1056 SCEVExpander &Rewriter,
1057 LPPassManager &LPM) {
1058 SmallVector<WideIVInfo, 8> WideIVs;
1060 SmallVector<PHINode*, 8> LoopPhis;
1061 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1062 LoopPhis.push_back(cast<PHINode>(I));
1064 // Each round of simplification iterates through the SimplifyIVUsers worklist
1065 // for all current phis, then determines whether any IVs can be
1066 // widened. Widening adds new phis to LoopPhis, inducing another round of
1067 // simplification on the wide IVs.
1068 while (!LoopPhis.empty()) {
1069 // Evaluate as many IV expressions as possible before widening any IVs. This
1070 // forces SCEV to set no-wrap flags before evaluating sign/zero
1071 // extension. The first time SCEV attempts to normalize sign/zero extension,
1072 // the result becomes final. So for the most predictable results, we delay
1073 // evaluation of sign/zero extend evaluation until needed, and avoid running
1074 // other SCEV based analysis prior to SimplifyAndExtend.
1076 PHINode *CurrIV = LoopPhis.pop_back_val();
1078 // Information about sign/zero extensions of CurrIV.
1079 WideIVVisitor WIV(CurrIV, SE, TD);
1081 Changed |= simplifyUsersOfIV(CurrIV, SE, &LPM, DeadInsts, &WIV);
1083 if (WIV.WI.WidestNativeType) {
1084 WideIVs.push_back(WIV.WI);
1086 } while(!LoopPhis.empty());
1088 for (; !WideIVs.empty(); WideIVs.pop_back()) {
1089 WidenIV Widener(WideIVs.back(), LI, SE, DT, DeadInsts);
1090 if (PHINode *WidePhi = Widener.CreateWideIV(Rewriter)) {
1092 LoopPhis.push_back(WidePhi);
1098 //===----------------------------------------------------------------------===//
1099 // LinearFunctionTestReplace and its kin. Rewrite the loop exit condition.
1100 //===----------------------------------------------------------------------===//
1102 /// Check for expressions that ScalarEvolution generates to compute
1103 /// BackedgeTakenInfo. If these expressions have not been reduced, then
1104 /// expanding them may incur additional cost (albeit in the loop preheader).
1105 static bool isHighCostExpansion(const SCEV *S, BranchInst *BI,
1106 SmallPtrSet<const SCEV*, 8> &Processed,
1107 ScalarEvolution *SE) {
1108 if (!Processed.insert(S))
1111 // If the backedge-taken count is a UDiv, it's very likely a UDiv that
1112 // ScalarEvolution's HowFarToZero or HowManyLessThans produced to compute a
1113 // precise expression, rather than a UDiv from the user's code. If we can't
1114 // find a UDiv in the code with some simple searching, assume the former and
1115 // forego rewriting the loop.
1116 if (isa<SCEVUDivExpr>(S)) {
1117 ICmpInst *OrigCond = dyn_cast<ICmpInst>(BI->getCondition());
1118 if (!OrigCond) return true;
1119 const SCEV *R = SE->getSCEV(OrigCond->getOperand(1));
1120 R = SE->getMinusSCEV(R, SE->getConstant(R->getType(), 1));
1122 const SCEV *L = SE->getSCEV(OrigCond->getOperand(0));
1123 L = SE->getMinusSCEV(L, SE->getConstant(L->getType(), 1));
1129 // Recurse past add expressions, which commonly occur in the
1130 // BackedgeTakenCount. They may already exist in program code, and if not,
1131 // they are not too expensive rematerialize.
1132 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
1133 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
1135 if (isHighCostExpansion(*I, BI, Processed, SE))
1141 // HowManyLessThans uses a Max expression whenever the loop is not guarded by
1142 // the exit condition.
1143 if (isa<SCEVSMaxExpr>(S) || isa<SCEVUMaxExpr>(S))
1146 // If we haven't recognized an expensive SCEV pattern, assume it's an
1147 // expression produced by program code.
1151 /// canExpandBackedgeTakenCount - Return true if this loop's backedge taken
1152 /// count expression can be safely and cheaply expanded into an instruction
1153 /// sequence that can be used by LinearFunctionTestReplace.
1155 /// TODO: This fails for pointer-type loop counters with greater than one byte
1156 /// strides, consequently preventing LFTR from running. For the purpose of LFTR
1157 /// we could skip this check in the case that the LFTR loop counter (chosen by
1158 /// FindLoopCounter) is also pointer type. Instead, we could directly convert
1159 /// the loop test to an inequality test by checking the target data's alignment
1160 /// of element types (given that the initial pointer value originates from or is
1161 /// used by ABI constrained operation, as opposed to inttoptr/ptrtoint).
1162 /// However, we don't yet have a strong motivation for converting loop tests
1163 /// into inequality tests.
1164 static bool canExpandBackedgeTakenCount(Loop *L, ScalarEvolution *SE) {
1165 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1166 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) ||
1167 BackedgeTakenCount->isZero())
1170 if (!L->getExitingBlock())
1173 // Can't rewrite non-branch yet.
1174 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
1178 SmallPtrSet<const SCEV*, 8> Processed;
1179 if (isHighCostExpansion(BackedgeTakenCount, BI, Processed, SE))
1185 /// getLoopPhiForCounter - Return the loop header phi IFF IncV adds a loop
1186 /// invariant value to the phi.
1187 static PHINode *getLoopPhiForCounter(Value *IncV, Loop *L, DominatorTree *DT) {
1188 Instruction *IncI = dyn_cast<Instruction>(IncV);
1192 switch (IncI->getOpcode()) {
1193 case Instruction::Add:
1194 case Instruction::Sub:
1196 case Instruction::GetElementPtr:
1197 // An IV counter must preserve its type.
1198 if (IncI->getNumOperands() == 2)
1204 PHINode *Phi = dyn_cast<PHINode>(IncI->getOperand(0));
1205 if (Phi && Phi->getParent() == L->getHeader()) {
1206 if (isLoopInvariant(IncI->getOperand(1), L, DT))
1210 if (IncI->getOpcode() == Instruction::GetElementPtr)
1213 // Allow add/sub to be commuted.
1214 Phi = dyn_cast<PHINode>(IncI->getOperand(1));
1215 if (Phi && Phi->getParent() == L->getHeader()) {
1216 if (isLoopInvariant(IncI->getOperand(0), L, DT))
1222 /// Return the compare guarding the loop latch, or NULL for unrecognized tests.
1223 static ICmpInst *getLoopTest(Loop *L) {
1224 assert(L->getExitingBlock() && "expected loop exit");
1226 BasicBlock *LatchBlock = L->getLoopLatch();
1227 // Don't bother with LFTR if the loop is not properly simplified.
1231 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
1232 assert(BI && "expected exit branch");
1234 return dyn_cast<ICmpInst>(BI->getCondition());
1237 /// needsLFTR - LinearFunctionTestReplace policy. Return true unless we can show
1238 /// that the current exit test is already sufficiently canonical.
1239 static bool needsLFTR(Loop *L, DominatorTree *DT) {
1240 // Do LFTR to simplify the exit condition to an ICMP.
1241 ICmpInst *Cond = getLoopTest(L);
1245 // Do LFTR to simplify the exit ICMP to EQ/NE
1246 ICmpInst::Predicate Pred = Cond->getPredicate();
1247 if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ)
1250 // Look for a loop invariant RHS
1251 Value *LHS = Cond->getOperand(0);
1252 Value *RHS = Cond->getOperand(1);
1253 if (!isLoopInvariant(RHS, L, DT)) {
1254 if (!isLoopInvariant(LHS, L, DT))
1256 std::swap(LHS, RHS);
1258 // Look for a simple IV counter LHS
1259 PHINode *Phi = dyn_cast<PHINode>(LHS);
1261 Phi = getLoopPhiForCounter(LHS, L, DT);
1266 // Do LFTR if PHI node is defined in the loop, but is *not* a counter.
1267 int Idx = Phi->getBasicBlockIndex(L->getLoopLatch());
1271 // Do LFTR if the exit condition's IV is *not* a simple counter.
1272 Value *IncV = Phi->getIncomingValue(Idx);
1273 return Phi != getLoopPhiForCounter(IncV, L, DT);
1276 /// Recursive helper for hasConcreteDef(). Unfortunately, this currently boils
1277 /// down to checking that all operands are constant and listing instructions
1278 /// that may hide undef.
1279 static bool hasConcreteDefImpl(Value *V, SmallPtrSet<Value*, 8> &Visited,
1281 if (isa<Constant>(V))
1282 return !isa<UndefValue>(V);
1287 // Conservatively handle non-constant non-instructions. For example, Arguments
1289 Instruction *I = dyn_cast<Instruction>(V);
1293 // Load and return values may be undef.
1294 if(I->mayReadFromMemory() || isa<CallInst>(I) || isa<InvokeInst>(I))
1297 // Optimistically handle other instructions.
1298 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) {
1299 if (!Visited.insert(*OI))
1301 if (!hasConcreteDefImpl(*OI, Visited, Depth+1))
1307 /// Return true if the given value is concrete. We must prove that undef can
1310 /// TODO: If we decide that this is a good approach to checking for undef, we
1311 /// may factor it into a common location.
1312 static bool hasConcreteDef(Value *V) {
1313 SmallPtrSet<Value*, 8> Visited;
1315 return hasConcreteDefImpl(V, Visited, 0);
1318 /// AlmostDeadIV - Return true if this IV has any uses other than the (soon to
1319 /// be rewritten) loop exit test.
1320 static bool AlmostDeadIV(PHINode *Phi, BasicBlock *LatchBlock, Value *Cond) {
1321 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1322 Value *IncV = Phi->getIncomingValue(LatchIdx);
1324 for (Value::use_iterator UI = Phi->use_begin(), UE = Phi->use_end();
1326 if (*UI != Cond && *UI != IncV) return false;
1329 for (Value::use_iterator UI = IncV->use_begin(), UE = IncV->use_end();
1331 if (*UI != Cond && *UI != Phi) return false;
1336 /// FindLoopCounter - Find an affine IV in canonical form.
1338 /// BECount may be an i8* pointer type. The pointer difference is already
1339 /// valid count without scaling the address stride, so it remains a pointer
1340 /// expression as far as SCEV is concerned.
1342 /// Currently only valid for LFTR. See the comments on hasConcreteDef below.
1344 /// FIXME: Accept -1 stride and set IVLimit = IVInit - BECount
1346 /// FIXME: Accept non-unit stride as long as SCEV can reduce BECount * Stride.
1347 /// This is difficult in general for SCEV because of potential overflow. But we
1348 /// could at least handle constant BECounts.
1350 FindLoopCounter(Loop *L, const SCEV *BECount,
1351 ScalarEvolution *SE, DominatorTree *DT, const DataLayout *TD) {
1352 uint64_t BCWidth = SE->getTypeSizeInBits(BECount->getType());
1355 cast<BranchInst>(L->getExitingBlock()->getTerminator())->getCondition();
1357 // Loop over all of the PHI nodes, looking for a simple counter.
1358 PHINode *BestPhi = 0;
1359 const SCEV *BestInit = 0;
1360 BasicBlock *LatchBlock = L->getLoopLatch();
1361 assert(LatchBlock && "needsLFTR should guarantee a loop latch");
1363 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1364 PHINode *Phi = cast<PHINode>(I);
1365 if (!SE->isSCEVable(Phi->getType()))
1368 // Avoid comparing an integer IV against a pointer Limit.
1369 if (BECount->getType()->isPointerTy() && !Phi->getType()->isPointerTy())
1372 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Phi));
1373 if (!AR || AR->getLoop() != L || !AR->isAffine())
1376 // AR may be a pointer type, while BECount is an integer type.
1377 // AR may be wider than BECount. With eq/ne tests overflow is immaterial.
1378 // AR may not be a narrower type, or we may never exit.
1379 uint64_t PhiWidth = SE->getTypeSizeInBits(AR->getType());
1380 if (PhiWidth < BCWidth || (TD && !TD->isLegalInteger(PhiWidth)))
1383 const SCEV *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
1384 if (!Step || !Step->isOne())
1387 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1388 Value *IncV = Phi->getIncomingValue(LatchIdx);
1389 if (getLoopPhiForCounter(IncV, L, DT) != Phi)
1392 // Avoid reusing a potentially undef value to compute other values that may
1393 // have originally had a concrete definition.
1394 if (!hasConcreteDef(Phi)) {
1395 // We explicitly allow unknown phis as long as they are already used by
1396 // the loop test. In this case we assume that performing LFTR could not
1397 // increase the number of undef users.
1398 if (ICmpInst *Cond = getLoopTest(L)) {
1399 if (Phi != getLoopPhiForCounter(Cond->getOperand(0), L, DT)
1400 && Phi != getLoopPhiForCounter(Cond->getOperand(1), L, DT)) {
1405 const SCEV *Init = AR->getStart();
1407 if (BestPhi && !AlmostDeadIV(BestPhi, LatchBlock, Cond)) {
1408 // Don't force a live loop counter if another IV can be used.
1409 if (AlmostDeadIV(Phi, LatchBlock, Cond))
1412 // Prefer to count-from-zero. This is a more "canonical" counter form. It
1413 // also prefers integer to pointer IVs.
1414 if (BestInit->isZero() != Init->isZero()) {
1415 if (BestInit->isZero())
1418 // If two IVs both count from zero or both count from nonzero then the
1419 // narrower is likely a dead phi that has been widened. Use the wider phi
1420 // to allow the other to be eliminated.
1421 else if (PhiWidth <= SE->getTypeSizeInBits(BestPhi->getType()))
1430 /// genLoopLimit - Help LinearFunctionTestReplace by generating a value that
1431 /// holds the RHS of the new loop test.
1432 static Value *genLoopLimit(PHINode *IndVar, const SCEV *IVCount, Loop *L,
1433 SCEVExpander &Rewriter, ScalarEvolution *SE) {
1434 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
1435 assert(AR && AR->getLoop() == L && AR->isAffine() && "bad loop counter");
1436 const SCEV *IVInit = AR->getStart();
1438 // IVInit may be a pointer while IVCount is an integer when FindLoopCounter
1439 // finds a valid pointer IV. Sign extend BECount in order to materialize a
1440 // GEP. Avoid running SCEVExpander on a new pointer value, instead reusing
1441 // the existing GEPs whenever possible.
1442 if (IndVar->getType()->isPointerTy()
1443 && !IVCount->getType()->isPointerTy()) {
1445 Type *OfsTy = SE->getEffectiveSCEVType(IVInit->getType());
1446 const SCEV *IVOffset = SE->getTruncateOrSignExtend(IVCount, OfsTy);
1448 // Expand the code for the iteration count.
1449 assert(SE->isLoopInvariant(IVOffset, L) &&
1450 "Computed iteration count is not loop invariant!");
1451 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1452 Value *GEPOffset = Rewriter.expandCodeFor(IVOffset, OfsTy, BI);
1454 Value *GEPBase = IndVar->getIncomingValueForBlock(L->getLoopPreheader());
1455 assert(AR->getStart() == SE->getSCEV(GEPBase) && "bad loop counter");
1456 // We could handle pointer IVs other than i8*, but we need to compensate for
1457 // gep index scaling. See canExpandBackedgeTakenCount comments.
1458 assert(SE->getSizeOfExpr(
1459 cast<PointerType>(GEPBase->getType())->getElementType())->isOne()
1460 && "unit stride pointer IV must be i8*");
1462 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
1463 return Builder.CreateGEP(GEPBase, GEPOffset, "lftr.limit");
1466 // In any other case, convert both IVInit and IVCount to integers before
1467 // comparing. This may result in SCEV expension of pointers, but in practice
1468 // SCEV will fold the pointer arithmetic away as such:
1469 // BECount = (IVEnd - IVInit - 1) => IVLimit = IVInit (postinc).
1471 // Valid Cases: (1) both integers is most common; (2) both may be pointers
1472 // for simple memset-style loops; (3) IVInit is an integer and IVCount is a
1473 // pointer may occur when enable-iv-rewrite generates a canonical IV on top
1476 const SCEV *IVLimit = 0;
1477 // For unit stride, IVCount = Start + BECount with 2's complement overflow.
1478 // For non-zero Start, compute IVCount here.
1479 if (AR->getStart()->isZero())
1482 assert(AR->getStepRecurrence(*SE)->isOne() && "only handles unit stride");
1483 const SCEV *IVInit = AR->getStart();
1485 // For integer IVs, truncate the IV before computing IVInit + BECount.
1486 if (SE->getTypeSizeInBits(IVInit->getType())
1487 > SE->getTypeSizeInBits(IVCount->getType()))
1488 IVInit = SE->getTruncateExpr(IVInit, IVCount->getType());
1490 IVLimit = SE->getAddExpr(IVInit, IVCount);
1492 // Expand the code for the iteration count.
1493 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1494 IRBuilder<> Builder(BI);
1495 assert(SE->isLoopInvariant(IVLimit, L) &&
1496 "Computed iteration count is not loop invariant!");
1497 // Ensure that we generate the same type as IndVar, or a smaller integer
1498 // type. In the presence of null pointer values, we have an integer type
1499 // SCEV expression (IVInit) for a pointer type IV value (IndVar).
1500 Type *LimitTy = IVCount->getType()->isPointerTy() ?
1501 IndVar->getType() : IVCount->getType();
1502 return Rewriter.expandCodeFor(IVLimit, LimitTy, BI);
1506 /// LinearFunctionTestReplace - This method rewrites the exit condition of the
1507 /// loop to be a canonical != comparison against the incremented loop induction
1508 /// variable. This pass is able to rewrite the exit tests of any loop where the
1509 /// SCEV analysis can determine a loop-invariant trip count of the loop, which
1510 /// is actually a much broader range than just linear tests.
1511 Value *IndVarSimplify::
1512 LinearFunctionTestReplace(Loop *L,
1513 const SCEV *BackedgeTakenCount,
1515 SCEVExpander &Rewriter) {
1516 assert(canExpandBackedgeTakenCount(L, SE) && "precondition");
1518 // LFTR can ignore IV overflow and truncate to the width of
1519 // BECount. This avoids materializing the add(zext(add)) expression.
1520 Type *CntTy = BackedgeTakenCount->getType();
1522 const SCEV *IVCount = BackedgeTakenCount;
1524 // If the exiting block is the same as the backedge block, we prefer to
1525 // compare against the post-incremented value, otherwise we must compare
1526 // against the preincremented value.
1528 if (L->getExitingBlock() == L->getLoopLatch()) {
1529 // Add one to the "backedge-taken" count to get the trip count.
1530 // If this addition may overflow, we have to be more pessimistic and
1531 // cast the induction variable before doing the add.
1533 SE->getAddExpr(IVCount, SE->getConstant(IVCount->getType(), 1));
1534 if (CntTy == IVCount->getType())
1537 const SCEV *Zero = SE->getConstant(IVCount->getType(), 0);
1538 if ((isa<SCEVConstant>(N) && !N->isZero()) ||
1539 SE->isLoopEntryGuardedByCond(L, ICmpInst::ICMP_NE, N, Zero)) {
1540 // No overflow. Cast the sum.
1541 IVCount = SE->getTruncateOrZeroExtend(N, CntTy);
1543 // Potential overflow. Cast before doing the add.
1544 IVCount = SE->getTruncateOrZeroExtend(IVCount, CntTy);
1545 IVCount = SE->getAddExpr(IVCount, SE->getConstant(CntTy, 1));
1548 // The BackedgeTaken expression contains the number of times that the
1549 // backedge branches to the loop header. This is one less than the
1550 // number of times the loop executes, so use the incremented indvar.
1551 CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock());
1553 // We must use the preincremented value...
1554 IVCount = SE->getTruncateOrZeroExtend(IVCount, CntTy);
1558 Value *ExitCnt = genLoopLimit(IndVar, IVCount, L, Rewriter, SE);
1559 assert(ExitCnt->getType()->isPointerTy() == IndVar->getType()->isPointerTy()
1560 && "genLoopLimit missed a cast");
1562 // Insert a new icmp_ne or icmp_eq instruction before the branch.
1563 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1564 ICmpInst::Predicate P;
1565 if (L->contains(BI->getSuccessor(0)))
1566 P = ICmpInst::ICMP_NE;
1568 P = ICmpInst::ICMP_EQ;
1570 DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
1571 << " LHS:" << *CmpIndVar << '\n'
1573 << (P == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
1574 << " RHS:\t" << *ExitCnt << "\n"
1575 << " IVCount:\t" << *IVCount << "\n");
1577 IRBuilder<> Builder(BI);
1578 if (SE->getTypeSizeInBits(CmpIndVar->getType())
1579 > SE->getTypeSizeInBits(ExitCnt->getType())) {
1580 CmpIndVar = Builder.CreateTrunc(CmpIndVar, ExitCnt->getType(),
1584 Value *Cond = Builder.CreateICmp(P, CmpIndVar, ExitCnt, "exitcond");
1585 Value *OrigCond = BI->getCondition();
1586 // It's tempting to use replaceAllUsesWith here to fully replace the old
1587 // comparison, but that's not immediately safe, since users of the old
1588 // comparison may not be dominated by the new comparison. Instead, just
1589 // update the branch to use the new comparison; in the common case this
1590 // will make old comparison dead.
1591 BI->setCondition(Cond);
1592 DeadInsts.push_back(OrigCond);
1599 //===----------------------------------------------------------------------===//
1600 // SinkUnusedInvariants. A late subpass to cleanup loop preheaders.
1601 //===----------------------------------------------------------------------===//
1603 /// If there's a single exit block, sink any loop-invariant values that
1604 /// were defined in the preheader but not used inside the loop into the
1605 /// exit block to reduce register pressure in the loop.
1606 void IndVarSimplify::SinkUnusedInvariants(Loop *L) {
1607 BasicBlock *ExitBlock = L->getExitBlock();
1608 if (!ExitBlock) return;
1610 BasicBlock *Preheader = L->getLoopPreheader();
1611 if (!Preheader) return;
1613 Instruction *InsertPt = ExitBlock->getFirstInsertionPt();
1614 BasicBlock::iterator I = Preheader->getTerminator();
1615 while (I != Preheader->begin()) {
1617 // New instructions were inserted at the end of the preheader.
1618 if (isa<PHINode>(I))
1621 // Don't move instructions which might have side effects, since the side
1622 // effects need to complete before instructions inside the loop. Also don't
1623 // move instructions which might read memory, since the loop may modify
1624 // memory. Note that it's okay if the instruction might have undefined
1625 // behavior: LoopSimplify guarantees that the preheader dominates the exit
1627 if (I->mayHaveSideEffects() || I->mayReadFromMemory())
1630 // Skip debug info intrinsics.
1631 if (isa<DbgInfoIntrinsic>(I))
1634 // Skip landingpad instructions.
1635 if (isa<LandingPadInst>(I))
1638 // Don't sink alloca: we never want to sink static alloca's out of the
1639 // entry block, and correctly sinking dynamic alloca's requires
1640 // checks for stacksave/stackrestore intrinsics.
1641 // FIXME: Refactor this check somehow?
1642 if (isa<AllocaInst>(I))
1645 // Determine if there is a use in or before the loop (direct or
1647 bool UsedInLoop = false;
1648 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
1651 BasicBlock *UseBB = cast<Instruction>(U)->getParent();
1652 if (PHINode *P = dyn_cast<PHINode>(U)) {
1654 PHINode::getIncomingValueNumForOperand(UI.getOperandNo());
1655 UseBB = P->getIncomingBlock(i);
1657 if (UseBB == Preheader || L->contains(UseBB)) {
1663 // If there is, the def must remain in the preheader.
1667 // Otherwise, sink it to the exit block.
1668 Instruction *ToMove = I;
1671 if (I != Preheader->begin()) {
1672 // Skip debug info intrinsics.
1675 } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin());
1677 if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin())
1683 ToMove->moveBefore(InsertPt);
1689 //===----------------------------------------------------------------------===//
1690 // IndVarSimplify driver. Manage several subpasses of IV simplification.
1691 //===----------------------------------------------------------------------===//
1693 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
1694 // If LoopSimplify form is not available, stay out of trouble. Some notes:
1695 // - LSR currently only supports LoopSimplify-form loops. Indvars'
1696 // canonicalization can be a pessimization without LSR to "clean up"
1698 // - We depend on having a preheader; in particular,
1699 // Loop::getCanonicalInductionVariable only supports loops with preheaders,
1700 // and we're in trouble if we can't find the induction variable even when
1701 // we've manually inserted one.
1702 if (!L->isLoopSimplifyForm())
1705 LI = &getAnalysis<LoopInfo>();
1706 SE = &getAnalysis<ScalarEvolution>();
1707 DT = &getAnalysis<DominatorTree>();
1708 TD = getAnalysisIfAvailable<DataLayout>();
1709 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
1714 // If there are any floating-point recurrences, attempt to
1715 // transform them to use integer recurrences.
1716 RewriteNonIntegerIVs(L);
1718 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1720 // Create a rewriter object which we'll use to transform the code with.
1721 SCEVExpander Rewriter(*SE, "indvars");
1723 Rewriter.setDebugType(DEBUG_TYPE);
1726 // Eliminate redundant IV users.
1728 // Simplification works best when run before other consumers of SCEV. We
1729 // attempt to avoid evaluating SCEVs for sign/zero extend operations until
1730 // other expressions involving loop IVs have been evaluated. This helps SCEV
1731 // set no-wrap flags before normalizing sign/zero extension.
1732 Rewriter.disableCanonicalMode();
1733 SimplifyAndExtend(L, Rewriter, LPM);
1735 // Check to see if this loop has a computable loop-invariant execution count.
1736 // If so, this means that we can compute the final value of any expressions
1737 // that are recurrent in the loop, and substitute the exit values from the
1738 // loop into any instructions outside of the loop that use the final values of
1739 // the current expressions.
1741 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1742 RewriteLoopExitValues(L, Rewriter);
1744 // Eliminate redundant IV cycles.
1745 NumElimIV += Rewriter.replaceCongruentIVs(L, DT, DeadInsts);
1747 // If we have a trip count expression, rewrite the loop's exit condition
1748 // using it. We can currently only handle loops with a single exit.
1749 if (canExpandBackedgeTakenCount(L, SE) && needsLFTR(L, DT)) {
1750 PHINode *IndVar = FindLoopCounter(L, BackedgeTakenCount, SE, DT, TD);
1752 // Check preconditions for proper SCEVExpander operation. SCEV does not
1753 // express SCEVExpander's dependencies, such as LoopSimplify. Instead any
1754 // pass that uses the SCEVExpander must do it. This does not work well for
1755 // loop passes because SCEVExpander makes assumptions about all loops, while
1756 // LoopPassManager only forces the current loop to be simplified.
1758 // FIXME: SCEV expansion has no way to bail out, so the caller must
1759 // explicitly check any assumptions made by SCEV. Brittle.
1760 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(BackedgeTakenCount);
1761 if (!AR || AR->getLoop()->getLoopPreheader())
1762 (void)LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar,
1766 // Clear the rewriter cache, because values that are in the rewriter's cache
1767 // can be deleted in the loop below, causing the AssertingVH in the cache to
1771 // Now that we're done iterating through lists, clean up any instructions
1772 // which are now dead.
1773 while (!DeadInsts.empty())
1774 if (Instruction *Inst =
1775 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
1776 RecursivelyDeleteTriviallyDeadInstructions(Inst, TLI);
1778 // The Rewriter may not be used from this point on.
1780 // Loop-invariant instructions in the preheader that aren't used in the
1781 // loop may be sunk below the loop to reduce register pressure.
1782 SinkUnusedInvariants(L);
1784 // Clean up dead instructions.
1785 Changed |= DeleteDeadPHIs(L->getHeader(), TLI);
1786 // Check a post-condition.
1787 assert(L->isLCSSAForm(*DT) &&
1788 "Indvars did not leave the loop in lcssa form!");
1790 // Verify that LFTR, and any other change have not interfered with SCEV's
1791 // ability to compute trip count.
1793 if (VerifyIndvars && !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) {
1795 const SCEV *NewBECount = SE->getBackedgeTakenCount(L);
1796 if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) <
1797 SE->getTypeSizeInBits(NewBECount->getType()))
1798 NewBECount = SE->getTruncateOrNoop(NewBECount,
1799 BackedgeTakenCount->getType());
1801 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount,
1802 NewBECount->getType());
1803 assert(BackedgeTakenCount == NewBECount && "indvars must preserve SCEV");