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/Target/TargetData.h"
47 #include "llvm/ADT/DenseMap.h"
48 #include "llvm/ADT/SmallVector.h"
49 #include "llvm/ADT/Statistic.h"
52 STATISTIC(NumWidened , "Number of indvars widened");
53 STATISTIC(NumReplaced , "Number of exit values replaced");
54 STATISTIC(NumLFTR , "Number of loop exit tests replaced");
55 STATISTIC(NumElimExt , "Number of IV sign/zero extends eliminated");
56 STATISTIC(NumElimIV , "Number of congruent IVs eliminated");
58 // Trip count verification can be enabled by default under NDEBUG if we
59 // implement a strong expression equivalence checker in SCEV. Until then, we
60 // use the verify-indvars flag, which may assert in some cases.
61 static cl::opt<bool> VerifyIndvars(
62 "verify-indvars", cl::Hidden,
63 cl::desc("Verify the ScalarEvolution result after running indvars"));
66 class IndVarSimplify : public LoopPass {
72 SmallVector<WeakVH, 16> DeadInsts;
76 static char ID; // Pass identification, replacement for typeid
77 IndVarSimplify() : LoopPass(ID), LI(0), SE(0), DT(0), TD(0),
79 initializeIndVarSimplifyPass(*PassRegistry::getPassRegistry());
82 virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
84 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
85 AU.addRequired<DominatorTree>();
86 AU.addRequired<LoopInfo>();
87 AU.addRequired<ScalarEvolution>();
88 AU.addRequiredID(LoopSimplifyID);
89 AU.addRequiredID(LCSSAID);
90 AU.addPreserved<ScalarEvolution>();
91 AU.addPreservedID(LoopSimplifyID);
92 AU.addPreservedID(LCSSAID);
97 virtual void releaseMemory() {
101 bool isValidRewrite(Value *FromVal, Value *ToVal);
103 void HandleFloatingPointIV(Loop *L, PHINode *PH);
104 void RewriteNonIntegerIVs(Loop *L);
106 void SimplifyAndExtend(Loop *L, SCEVExpander &Rewriter, LPPassManager &LPM);
108 void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter);
110 Value *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
111 PHINode *IndVar, SCEVExpander &Rewriter);
113 void SinkUnusedInvariants(Loop *L);
117 char IndVarSimplify::ID = 0;
118 INITIALIZE_PASS_BEGIN(IndVarSimplify, "indvars",
119 "Induction Variable Simplification", false, false)
120 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
121 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
122 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
123 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
124 INITIALIZE_PASS_DEPENDENCY(LCSSA)
125 INITIALIZE_PASS_END(IndVarSimplify, "indvars",
126 "Induction Variable Simplification", false, false)
128 Pass *llvm::createIndVarSimplifyPass() {
129 return new IndVarSimplify();
132 /// isValidRewrite - Return true if the SCEV expansion generated by the
133 /// rewriter can replace the original value. SCEV guarantees that it
134 /// produces the same value, but the way it is produced may be illegal IR.
135 /// Ideally, this function will only be called for verification.
136 bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) {
137 // If an SCEV expression subsumed multiple pointers, its expansion could
138 // reassociate the GEP changing the base pointer. This is illegal because the
139 // final address produced by a GEP chain must be inbounds relative to its
140 // underlying object. Otherwise basic alias analysis, among other things,
141 // could fail in a dangerous way. Ultimately, SCEV will be improved to avoid
142 // producing an expression involving multiple pointers. Until then, we must
145 // Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject
146 // because it understands lcssa phis while SCEV does not.
147 Value *FromPtr = FromVal;
148 Value *ToPtr = ToVal;
149 if (GEPOperator *GEP = dyn_cast<GEPOperator>(FromVal)) {
150 FromPtr = GEP->getPointerOperand();
152 if (GEPOperator *GEP = dyn_cast<GEPOperator>(ToVal)) {
153 ToPtr = GEP->getPointerOperand();
155 if (FromPtr != FromVal || ToPtr != ToVal) {
156 // Quickly check the common case
157 if (FromPtr == ToPtr)
160 // SCEV may have rewritten an expression that produces the GEP's pointer
161 // operand. That's ok as long as the pointer operand has the same base
162 // pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the
163 // base of a recurrence. This handles the case in which SCEV expansion
164 // converts a pointer type recurrence into a nonrecurrent pointer base
165 // indexed by an integer recurrence.
167 // If the GEP base pointer is a vector of pointers, abort.
168 if (!FromPtr->getType()->isPointerTy() || !ToPtr->getType()->isPointerTy())
171 const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr));
172 const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr));
173 if (FromBase == ToBase)
176 DEBUG(dbgs() << "INDVARS: GEP rewrite bail out "
177 << *FromBase << " != " << *ToBase << "\n");
184 /// Determine the insertion point for this user. By default, insert immediately
185 /// before the user. SCEVExpander or LICM will hoist loop invariants out of the
186 /// loop. For PHI nodes, there may be multiple uses, so compute the nearest
187 /// common dominator for the incoming blocks.
188 static Instruction *getInsertPointForUses(Instruction *User, Value *Def,
190 PHINode *PHI = dyn_cast<PHINode>(User);
194 Instruction *InsertPt = 0;
195 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) {
196 if (PHI->getIncomingValue(i) != Def)
199 BasicBlock *InsertBB = PHI->getIncomingBlock(i);
201 InsertPt = InsertBB->getTerminator();
204 InsertBB = DT->findNearestCommonDominator(InsertPt->getParent(), InsertBB);
205 InsertPt = InsertBB->getTerminator();
207 assert(InsertPt && "Missing phi operand");
208 assert((!isa<Instruction>(Def) ||
209 DT->dominates(cast<Instruction>(Def), InsertPt)) &&
210 "def does not dominate all uses");
214 //===----------------------------------------------------------------------===//
215 // RewriteNonIntegerIVs and helpers. Prefer integer IVs.
216 //===----------------------------------------------------------------------===//
218 /// ConvertToSInt - Convert APF to an integer, if possible.
219 static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) {
220 bool isExact = false;
221 if (&APF.getSemantics() == &APFloat::PPCDoubleDouble)
223 // See if we can convert this to an int64_t
225 if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero,
226 &isExact) != APFloat::opOK || !isExact)
232 /// HandleFloatingPointIV - If the loop has floating induction variable
233 /// then insert corresponding integer induction variable if possible.
235 /// for(double i = 0; i < 10000; ++i)
237 /// is converted into
238 /// for(int i = 0; i < 10000; ++i)
241 void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) {
242 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
243 unsigned BackEdge = IncomingEdge^1;
245 // Check incoming value.
246 ConstantFP *InitValueVal =
247 dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));
250 if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
253 // Check IV increment. Reject this PN if increment operation is not
254 // an add or increment value can not be represented by an integer.
255 BinaryOperator *Incr =
256 dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
257 if (Incr == 0 || Incr->getOpcode() != Instruction::FAdd) return;
259 // If this is not an add of the PHI with a constantfp, or if the constant fp
260 // is not an integer, bail out.
261 ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
263 if (IncValueVal == 0 || Incr->getOperand(0) != PN ||
264 !ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
267 // Check Incr uses. One user is PN and the other user is an exit condition
268 // used by the conditional terminator.
269 Value::use_iterator IncrUse = Incr->use_begin();
270 Instruction *U1 = cast<Instruction>(*IncrUse++);
271 if (IncrUse == Incr->use_end()) return;
272 Instruction *U2 = cast<Instruction>(*IncrUse++);
273 if (IncrUse != Incr->use_end()) return;
275 // Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't
276 // only used by a branch, we can't transform it.
277 FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
279 Compare = dyn_cast<FCmpInst>(U2);
280 if (Compare == 0 || !Compare->hasOneUse() ||
281 !isa<BranchInst>(Compare->use_back()))
284 BranchInst *TheBr = cast<BranchInst>(Compare->use_back());
286 // We need to verify that the branch actually controls the iteration count
287 // of the loop. If not, the new IV can overflow and no one will notice.
288 // The branch block must be in the loop and one of the successors must be out
290 assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
291 if (!L->contains(TheBr->getParent()) ||
292 (L->contains(TheBr->getSuccessor(0)) &&
293 L->contains(TheBr->getSuccessor(1))))
297 // If it isn't a comparison with an integer-as-fp (the exit value), we can't
299 ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
301 if (ExitValueVal == 0 ||
302 !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
305 // Find new predicate for integer comparison.
306 CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
307 switch (Compare->getPredicate()) {
308 default: return; // Unknown comparison.
309 case CmpInst::FCMP_OEQ:
310 case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
311 case CmpInst::FCMP_ONE:
312 case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
313 case CmpInst::FCMP_OGT:
314 case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
315 case CmpInst::FCMP_OGE:
316 case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
317 case CmpInst::FCMP_OLT:
318 case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
319 case CmpInst::FCMP_OLE:
320 case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
323 // We convert the floating point induction variable to a signed i32 value if
324 // we can. This is only safe if the comparison will not overflow in a way
325 // that won't be trapped by the integer equivalent operations. Check for this
327 // TODO: We could use i64 if it is native and the range requires it.
329 // The start/stride/exit values must all fit in signed i32.
330 if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
333 // If not actually striding (add x, 0.0), avoid touching the code.
337 // Positive and negative strides have different safety conditions.
339 // If we have a positive stride, we require the init to be less than the
341 if (InitValue >= ExitValue)
344 uint32_t Range = uint32_t(ExitValue-InitValue);
345 // Check for infinite loop, either:
346 // while (i <= Exit) or until (i > Exit)
347 if (NewPred == CmpInst::ICMP_SLE || NewPred == CmpInst::ICMP_SGT) {
348 if (++Range == 0) return; // Range overflows.
351 unsigned Leftover = Range % uint32_t(IncValue);
353 // If this is an equality comparison, we require that the strided value
354 // exactly land on the exit value, otherwise the IV condition will wrap
355 // around and do things the fp IV wouldn't.
356 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
360 // If the stride would wrap around the i32 before exiting, we can't
362 if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
366 // If we have a negative stride, we require the init to be greater than the
368 if (InitValue <= ExitValue)
371 uint32_t Range = uint32_t(InitValue-ExitValue);
372 // Check for infinite loop, either:
373 // while (i >= Exit) or until (i < Exit)
374 if (NewPred == CmpInst::ICMP_SGE || NewPred == CmpInst::ICMP_SLT) {
375 if (++Range == 0) return; // Range overflows.
378 unsigned Leftover = Range % uint32_t(-IncValue);
380 // If this is an equality comparison, we require that the strided value
381 // exactly land on the exit value, otherwise the IV condition will wrap
382 // around and do things the fp IV wouldn't.
383 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
387 // If the stride would wrap around the i32 before exiting, we can't
389 if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
393 IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
395 // Insert new integer induction variable.
396 PHINode *NewPHI = PHINode::Create(Int32Ty, 2, PN->getName()+".int", PN);
397 NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
398 PN->getIncomingBlock(IncomingEdge));
401 BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
402 Incr->getName()+".int", Incr);
403 NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
405 ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
406 ConstantInt::get(Int32Ty, ExitValue),
409 // In the following deletions, PN may become dead and may be deleted.
410 // Use a WeakVH to observe whether this happens.
413 // Delete the old floating point exit comparison. The branch starts using the
415 NewCompare->takeName(Compare);
416 Compare->replaceAllUsesWith(NewCompare);
417 RecursivelyDeleteTriviallyDeadInstructions(Compare);
419 // Delete the old floating point increment.
420 Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
421 RecursivelyDeleteTriviallyDeadInstructions(Incr);
423 // If the FP induction variable still has uses, this is because something else
424 // in the loop uses its value. In order to canonicalize the induction
425 // variable, we chose to eliminate the IV and rewrite it in terms of an
428 // We give preference to sitofp over uitofp because it is faster on most
431 Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
432 PN->getParent()->getFirstInsertionPt());
433 PN->replaceAllUsesWith(Conv);
434 RecursivelyDeleteTriviallyDeadInstructions(PN);
439 void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
440 // First step. Check to see if there are any floating-point recurrences.
441 // If there are, change them into integer recurrences, permitting analysis by
442 // the SCEV routines.
444 BasicBlock *Header = L->getHeader();
446 SmallVector<WeakVH, 8> PHIs;
447 for (BasicBlock::iterator I = Header->begin();
448 PHINode *PN = dyn_cast<PHINode>(I); ++I)
451 for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
452 if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i]))
453 HandleFloatingPointIV(L, PN);
455 // If the loop previously had floating-point IV, ScalarEvolution
456 // may not have been able to compute a trip count. Now that we've done some
457 // re-writing, the trip count may be computable.
462 //===----------------------------------------------------------------------===//
463 // RewriteLoopExitValues - Optimize IV users outside the loop.
464 // As a side effect, reduces the amount of IV processing within the loop.
465 //===----------------------------------------------------------------------===//
467 /// RewriteLoopExitValues - Check to see if this loop has a computable
468 /// loop-invariant execution count. If so, this means that we can compute the
469 /// final value of any expressions that are recurrent in the loop, and
470 /// substitute the exit values from the loop into any instructions outside of
471 /// the loop that use the final values of the current expressions.
473 /// This is mostly redundant with the regular IndVarSimplify activities that
474 /// happen later, except that it's more powerful in some cases, because it's
475 /// able to brute-force evaluate arbitrary instructions as long as they have
476 /// constant operands at the beginning of the loop.
477 void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) {
478 // Verify the input to the pass in already in LCSSA form.
479 assert(L->isLCSSAForm(*DT));
481 SmallVector<BasicBlock*, 8> ExitBlocks;
482 L->getUniqueExitBlocks(ExitBlocks);
484 // Find all values that are computed inside the loop, but used outside of it.
485 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan
486 // the exit blocks of the loop to find them.
487 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
488 BasicBlock *ExitBB = ExitBlocks[i];
490 // If there are no PHI nodes in this exit block, then no values defined
491 // inside the loop are used on this path, skip it.
492 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
495 unsigned NumPreds = PN->getNumIncomingValues();
497 // Iterate over all of the PHI nodes.
498 BasicBlock::iterator BBI = ExitBB->begin();
499 while ((PN = dyn_cast<PHINode>(BBI++))) {
501 continue; // dead use, don't replace it
503 // SCEV only supports integer expressions for now.
504 if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy())
507 // It's necessary to tell ScalarEvolution about this explicitly so that
508 // it can walk the def-use list and forget all SCEVs, as it may not be
509 // watching the PHI itself. Once the new exit value is in place, there
510 // may not be a def-use connection between the loop and every instruction
511 // which got a SCEVAddRecExpr for that loop.
514 // Iterate over all of the values in all the PHI nodes.
515 for (unsigned i = 0; i != NumPreds; ++i) {
516 // If the value being merged in is not integer or is not defined
517 // in the loop, skip it.
518 Value *InVal = PN->getIncomingValue(i);
519 if (!isa<Instruction>(InVal))
522 // If this pred is for a subloop, not L itself, skip it.
523 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
524 continue; // The Block is in a subloop, skip it.
526 // Check that InVal is defined in the loop.
527 Instruction *Inst = cast<Instruction>(InVal);
528 if (!L->contains(Inst))
531 // Okay, this instruction has a user outside of the current loop
532 // and varies predictably *inside* the loop. Evaluate the value it
533 // contains when the loop exits, if possible.
534 const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
535 if (!SE->isLoopInvariant(ExitValue, L))
538 Value *ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), Inst);
540 DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
541 << " LoopVal = " << *Inst << "\n");
543 if (!isValidRewrite(Inst, ExitVal)) {
544 DeadInsts.push_back(ExitVal);
550 PN->setIncomingValue(i, ExitVal);
552 // If this instruction is dead now, delete it.
553 RecursivelyDeleteTriviallyDeadInstructions(Inst);
556 // Completely replace a single-pred PHI. This is safe, because the
557 // NewVal won't be variant in the loop, so we don't need an LCSSA phi
559 PN->replaceAllUsesWith(ExitVal);
560 RecursivelyDeleteTriviallyDeadInstructions(PN);
564 // Clone the PHI and delete the original one. This lets IVUsers and
565 // any other maps purge the original user from their records.
566 PHINode *NewPN = cast<PHINode>(PN->clone());
568 NewPN->insertBefore(PN);
569 PN->replaceAllUsesWith(NewPN);
570 PN->eraseFromParent();
575 // The insertion point instruction may have been deleted; clear it out
576 // so that the rewriter doesn't trip over it later.
577 Rewriter.clearInsertPoint();
580 //===----------------------------------------------------------------------===//
581 // IV Widening - Extend the width of an IV to cover its widest uses.
582 //===----------------------------------------------------------------------===//
585 // Collect information about induction variables that are used by sign/zero
586 // extend operations. This information is recorded by CollectExtend and
587 // provides the input to WidenIV.
590 Type *WidestNativeType; // Widest integer type created [sz]ext
591 bool IsSigned; // Was an sext user seen before a zext?
593 WideIVInfo() : NarrowIV(0), WidestNativeType(0), IsSigned(false) {}
596 class WideIVVisitor : public IVVisitor {
598 const TargetData *TD;
603 WideIVVisitor(PHINode *NarrowIV, ScalarEvolution *SCEV,
604 const TargetData *TData) :
605 SE(SCEV), TD(TData) { WI.NarrowIV = NarrowIV; }
607 // Implement the interface used by simplifyUsersOfIV.
608 virtual void visitCast(CastInst *Cast);
612 /// visitCast - Update information about the induction variable that is
613 /// extended by this sign or zero extend operation. This is used to determine
614 /// the final width of the IV before actually widening it.
615 void WideIVVisitor::visitCast(CastInst *Cast) {
616 bool IsSigned = Cast->getOpcode() == Instruction::SExt;
617 if (!IsSigned && Cast->getOpcode() != Instruction::ZExt)
620 Type *Ty = Cast->getType();
621 uint64_t Width = SE->getTypeSizeInBits(Ty);
622 if (TD && !TD->isLegalInteger(Width))
625 if (!WI.WidestNativeType) {
626 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
627 WI.IsSigned = IsSigned;
631 // We extend the IV to satisfy the sign of its first user, arbitrarily.
632 if (WI.IsSigned != IsSigned)
635 if (Width > SE->getTypeSizeInBits(WI.WidestNativeType))
636 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
641 /// NarrowIVDefUse - Record a link in the Narrow IV def-use chain along with the
642 /// WideIV that computes the same value as the Narrow IV def. This avoids
643 /// caching Use* pointers.
644 struct NarrowIVDefUse {
645 Instruction *NarrowDef;
646 Instruction *NarrowUse;
647 Instruction *WideDef;
649 NarrowIVDefUse(): NarrowDef(0), NarrowUse(0), WideDef(0) {}
651 NarrowIVDefUse(Instruction *ND, Instruction *NU, Instruction *WD):
652 NarrowDef(ND), NarrowUse(NU), WideDef(WD) {}
655 /// WidenIV - The goal of this transform is to remove sign and zero extends
656 /// without creating any new induction variables. To do this, it creates a new
657 /// phi of the wider type and redirects all users, either removing extends or
658 /// inserting truncs whenever we stop propagating the type.
674 Instruction *WideInc;
675 const SCEV *WideIncExpr;
676 SmallVectorImpl<WeakVH> &DeadInsts;
678 SmallPtrSet<Instruction*,16> Widened;
679 SmallVector<NarrowIVDefUse, 8> NarrowIVUsers;
682 WidenIV(const WideIVInfo &WI, LoopInfo *LInfo,
683 ScalarEvolution *SEv, DominatorTree *DTree,
684 SmallVectorImpl<WeakVH> &DI) :
685 OrigPhi(WI.NarrowIV),
686 WideType(WI.WidestNativeType),
687 IsSigned(WI.IsSigned),
689 L(LI->getLoopFor(OrigPhi->getParent())),
696 assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV");
699 PHINode *CreateWideIV(SCEVExpander &Rewriter);
702 Value *getExtend(Value *NarrowOper, Type *WideType, bool IsSigned,
705 Instruction *CloneIVUser(NarrowIVDefUse DU);
707 const SCEVAddRecExpr *GetWideRecurrence(Instruction *NarrowUse);
709 const SCEVAddRecExpr* GetExtendedOperandRecurrence(NarrowIVDefUse DU);
711 Instruction *WidenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter);
713 void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef);
715 } // anonymous namespace
717 /// isLoopInvariant - Perform a quick domtree based check for loop invariance
718 /// assuming that V is used within the loop. LoopInfo::isLoopInvariant() seems
719 /// gratuitous for this purpose.
720 static bool isLoopInvariant(Value *V, const Loop *L, const DominatorTree *DT) {
721 Instruction *Inst = dyn_cast<Instruction>(V);
725 return DT->properlyDominates(Inst->getParent(), L->getHeader());
728 Value *WidenIV::getExtend(Value *NarrowOper, Type *WideType, bool IsSigned,
730 // Set the debug location and conservative insertion point.
731 IRBuilder<> Builder(Use);
732 // Hoist the insertion point into loop preheaders as far as possible.
733 for (const Loop *L = LI->getLoopFor(Use->getParent());
734 L && L->getLoopPreheader() && isLoopInvariant(NarrowOper, L, DT);
735 L = L->getParentLoop())
736 Builder.SetInsertPoint(L->getLoopPreheader()->getTerminator());
738 return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) :
739 Builder.CreateZExt(NarrowOper, WideType);
742 /// CloneIVUser - Instantiate a wide operation to replace a narrow
743 /// operation. This only needs to handle operations that can evaluation to
744 /// SCEVAddRec. It can safely return 0 for any operation we decide not to clone.
745 Instruction *WidenIV::CloneIVUser(NarrowIVDefUse DU) {
746 unsigned Opcode = DU.NarrowUse->getOpcode();
750 case Instruction::Add:
751 case Instruction::Mul:
752 case Instruction::UDiv:
753 case Instruction::Sub:
754 case Instruction::And:
755 case Instruction::Or:
756 case Instruction::Xor:
757 case Instruction::Shl:
758 case Instruction::LShr:
759 case Instruction::AShr:
760 DEBUG(dbgs() << "Cloning IVUser: " << *DU.NarrowUse << "\n");
762 // Replace NarrowDef operands with WideDef. Otherwise, we don't know
763 // anything about the narrow operand yet so must insert a [sz]ext. It is
764 // probably loop invariant and will be folded or hoisted. If it actually
765 // comes from a widened IV, it should be removed during a future call to
767 Value *LHS = (DU.NarrowUse->getOperand(0) == DU.NarrowDef) ? DU.WideDef :
768 getExtend(DU.NarrowUse->getOperand(0), WideType, IsSigned, DU.NarrowUse);
769 Value *RHS = (DU.NarrowUse->getOperand(1) == DU.NarrowDef) ? DU.WideDef :
770 getExtend(DU.NarrowUse->getOperand(1), WideType, IsSigned, DU.NarrowUse);
772 BinaryOperator *NarrowBO = cast<BinaryOperator>(DU.NarrowUse);
773 BinaryOperator *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(),
775 NarrowBO->getName());
776 IRBuilder<> Builder(DU.NarrowUse);
777 Builder.Insert(WideBO);
778 if (const OverflowingBinaryOperator *OBO =
779 dyn_cast<OverflowingBinaryOperator>(NarrowBO)) {
780 if (OBO->hasNoUnsignedWrap()) WideBO->setHasNoUnsignedWrap();
781 if (OBO->hasNoSignedWrap()) WideBO->setHasNoSignedWrap();
787 /// No-wrap operations can transfer sign extension of their result to their
788 /// operands. Generate the SCEV value for the widened operation without
789 /// actually modifying the IR yet. If the expression after extending the
790 /// operands is an AddRec for this loop, return it.
791 const SCEVAddRecExpr* WidenIV::GetExtendedOperandRecurrence(NarrowIVDefUse DU) {
792 // Handle the common case of add<nsw/nuw>
793 if (DU.NarrowUse->getOpcode() != Instruction::Add)
796 // One operand (NarrowDef) has already been extended to WideDef. Now determine
797 // if extending the other will lead to a recurrence.
798 unsigned ExtendOperIdx = DU.NarrowUse->getOperand(0) == DU.NarrowDef ? 1 : 0;
799 assert(DU.NarrowUse->getOperand(1-ExtendOperIdx) == DU.NarrowDef && "bad DU");
801 const SCEV *ExtendOperExpr = 0;
802 const OverflowingBinaryOperator *OBO =
803 cast<OverflowingBinaryOperator>(DU.NarrowUse);
804 if (IsSigned && OBO->hasNoSignedWrap())
805 ExtendOperExpr = SE->getSignExtendExpr(
806 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
807 else if(!IsSigned && OBO->hasNoUnsignedWrap())
808 ExtendOperExpr = SE->getZeroExtendExpr(
809 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
813 // When creating this AddExpr, don't apply the current operations NSW or NUW
814 // flags. This instruction may be guarded by control flow that the no-wrap
815 // behavior depends on. Non-control-equivalent instructions can be mapped to
816 // the same SCEV expression, and it would be incorrect to transfer NSW/NUW
817 // semantics to those operations.
818 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(
819 SE->getAddExpr(SE->getSCEV(DU.WideDef), ExtendOperExpr));
821 if (!AddRec || AddRec->getLoop() != L)
826 /// GetWideRecurrence - Is this instruction potentially interesting from
827 /// IVUsers' perspective after widening it's type? In other words, can the
828 /// extend be safely hoisted out of the loop with SCEV reducing the value to a
829 /// recurrence on the same loop. If so, return the sign or zero extended
830 /// recurrence. Otherwise return NULL.
831 const SCEVAddRecExpr *WidenIV::GetWideRecurrence(Instruction *NarrowUse) {
832 if (!SE->isSCEVable(NarrowUse->getType()))
835 const SCEV *NarrowExpr = SE->getSCEV(NarrowUse);
836 if (SE->getTypeSizeInBits(NarrowExpr->getType())
837 >= SE->getTypeSizeInBits(WideType)) {
838 // NarrowUse implicitly widens its operand. e.g. a gep with a narrow
839 // index. So don't follow this use.
843 const SCEV *WideExpr = IsSigned ?
844 SE->getSignExtendExpr(NarrowExpr, WideType) :
845 SE->getZeroExtendExpr(NarrowExpr, WideType);
846 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr);
847 if (!AddRec || AddRec->getLoop() != L)
852 /// WidenIVUse - Determine whether an individual user of the narrow IV can be
853 /// widened. If so, return the wide clone of the user.
854 Instruction *WidenIV::WidenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter) {
856 // Stop traversing the def-use chain at inner-loop phis or post-loop phis.
857 if (isa<PHINode>(DU.NarrowUse) &&
858 LI->getLoopFor(DU.NarrowUse->getParent()) != L)
861 // Our raison d'etre! Eliminate sign and zero extension.
862 if (IsSigned ? isa<SExtInst>(DU.NarrowUse) : isa<ZExtInst>(DU.NarrowUse)) {
863 Value *NewDef = DU.WideDef;
864 if (DU.NarrowUse->getType() != WideType) {
865 unsigned CastWidth = SE->getTypeSizeInBits(DU.NarrowUse->getType());
866 unsigned IVWidth = SE->getTypeSizeInBits(WideType);
867 if (CastWidth < IVWidth) {
868 // The cast isn't as wide as the IV, so insert a Trunc.
869 IRBuilder<> Builder(DU.NarrowUse);
870 NewDef = Builder.CreateTrunc(DU.WideDef, DU.NarrowUse->getType());
873 // A wider extend was hidden behind a narrower one. This may induce
874 // another round of IV widening in which the intermediate IV becomes
875 // dead. It should be very rare.
876 DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi
877 << " not wide enough to subsume " << *DU.NarrowUse << "\n");
878 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
879 NewDef = DU.NarrowUse;
882 if (NewDef != DU.NarrowUse) {
883 DEBUG(dbgs() << "INDVARS: eliminating " << *DU.NarrowUse
884 << " replaced by " << *DU.WideDef << "\n");
886 DU.NarrowUse->replaceAllUsesWith(NewDef);
887 DeadInsts.push_back(DU.NarrowUse);
889 // Now that the extend is gone, we want to expose it's uses for potential
890 // further simplification. We don't need to directly inform SimplifyIVUsers
891 // of the new users, because their parent IV will be processed later as a
892 // new loop phi. If we preserved IVUsers analysis, we would also want to
893 // push the uses of WideDef here.
895 // No further widening is needed. The deceased [sz]ext had done it for us.
899 // Does this user itself evaluate to a recurrence after widening?
900 const SCEVAddRecExpr *WideAddRec = GetWideRecurrence(DU.NarrowUse);
902 WideAddRec = GetExtendedOperandRecurrence(DU);
905 // This user does not evaluate to a recurence after widening, so don't
906 // follow it. Instead insert a Trunc to kill off the original use,
907 // eventually isolating the original narrow IV so it can be removed.
908 IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT));
909 Value *Trunc = Builder.CreateTrunc(DU.WideDef, DU.NarrowDef->getType());
910 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, Trunc);
913 // Assume block terminators cannot evaluate to a recurrence. We can't to
914 // insert a Trunc after a terminator if there happens to be a critical edge.
915 assert(DU.NarrowUse != DU.NarrowUse->getParent()->getTerminator() &&
916 "SCEV is not expected to evaluate a block terminator");
918 // Reuse the IV increment that SCEVExpander created as long as it dominates
920 Instruction *WideUse = 0;
921 if (WideAddRec == WideIncExpr
922 && Rewriter.hoistIVInc(WideInc, DU.NarrowUse))
925 WideUse = CloneIVUser(DU);
929 // Evaluation of WideAddRec ensured that the narrow expression could be
930 // extended outside the loop without overflow. This suggests that the wide use
931 // evaluates to the same expression as the extended narrow use, but doesn't
932 // absolutely guarantee it. Hence the following failsafe check. In rare cases
933 // where it fails, we simply throw away the newly created wide use.
934 if (WideAddRec != SE->getSCEV(WideUse)) {
935 DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse
936 << ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n");
937 DeadInsts.push_back(WideUse);
941 // Returning WideUse pushes it on the worklist.
945 /// pushNarrowIVUsers - Add eligible users of NarrowDef to NarrowIVUsers.
947 void WidenIV::pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef) {
948 for (Value::use_iterator UI = NarrowDef->use_begin(),
949 UE = NarrowDef->use_end(); UI != UE; ++UI) {
950 Instruction *NarrowUse = cast<Instruction>(*UI);
952 // Handle data flow merges and bizarre phi cycles.
953 if (!Widened.insert(NarrowUse))
956 NarrowIVUsers.push_back(NarrowIVDefUse(NarrowDef, NarrowUse, WideDef));
960 /// CreateWideIV - Process a single induction variable. First use the
961 /// SCEVExpander to create a wide induction variable that evaluates to the same
962 /// recurrence as the original narrow IV. Then use a worklist to forward
963 /// traverse the narrow IV's def-use chain. After WidenIVUse has processed all
964 /// interesting IV users, the narrow IV will be isolated for removal by
967 /// It would be simpler to delete uses as they are processed, but we must avoid
968 /// invalidating SCEV expressions.
970 PHINode *WidenIV::CreateWideIV(SCEVExpander &Rewriter) {
971 // Is this phi an induction variable?
972 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi));
976 // Widen the induction variable expression.
977 const SCEV *WideIVExpr = IsSigned ?
978 SE->getSignExtendExpr(AddRec, WideType) :
979 SE->getZeroExtendExpr(AddRec, WideType);
981 assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType &&
982 "Expect the new IV expression to preserve its type");
984 // Can the IV be extended outside the loop without overflow?
985 AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr);
986 if (!AddRec || AddRec->getLoop() != L)
989 // An AddRec must have loop-invariant operands. Since this AddRec is
990 // materialized by a loop header phi, the expression cannot have any post-loop
991 // operands, so they must dominate the loop header.
992 assert(SE->properlyDominates(AddRec->getStart(), L->getHeader()) &&
993 SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader())
994 && "Loop header phi recurrence inputs do not dominate the loop");
996 // The rewriter provides a value for the desired IV expression. This may
997 // either find an existing phi or materialize a new one. Either way, we
998 // expect a well-formed cyclic phi-with-increments. i.e. any operand not part
999 // of the phi-SCC dominates the loop entry.
1000 Instruction *InsertPt = L->getHeader()->begin();
1001 WidePhi = cast<PHINode>(Rewriter.expandCodeFor(AddRec, WideType, InsertPt));
1003 // Remembering the WideIV increment generated by SCEVExpander allows
1004 // WidenIVUse to reuse it when widening the narrow IV's increment. We don't
1005 // employ a general reuse mechanism because the call above is the only call to
1006 // SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses.
1007 if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1009 cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock));
1010 WideIncExpr = SE->getSCEV(WideInc);
1013 DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n");
1016 // Traverse the def-use chain using a worklist starting at the original IV.
1017 assert(Widened.empty() && NarrowIVUsers.empty() && "expect initial state" );
1019 Widened.insert(OrigPhi);
1020 pushNarrowIVUsers(OrigPhi, WidePhi);
1022 while (!NarrowIVUsers.empty()) {
1023 NarrowIVDefUse DU = NarrowIVUsers.pop_back_val();
1025 // Process a def-use edge. This may replace the use, so don't hold a
1026 // use_iterator across it.
1027 Instruction *WideUse = WidenIVUse(DU, Rewriter);
1029 // Follow all def-use edges from the previous narrow use.
1031 pushNarrowIVUsers(DU.NarrowUse, WideUse);
1033 // WidenIVUse may have removed the def-use edge.
1034 if (DU.NarrowDef->use_empty())
1035 DeadInsts.push_back(DU.NarrowDef);
1040 //===----------------------------------------------------------------------===//
1041 // Simplification of IV users based on SCEV evaluation.
1042 //===----------------------------------------------------------------------===//
1045 /// SimplifyAndExtend - Iteratively perform simplification on a worklist of IV
1046 /// users. Each successive simplification may push more users which may
1047 /// themselves be candidates for simplification.
1049 /// Sign/Zero extend elimination is interleaved with IV simplification.
1051 void IndVarSimplify::SimplifyAndExtend(Loop *L,
1052 SCEVExpander &Rewriter,
1053 LPPassManager &LPM) {
1054 SmallVector<WideIVInfo, 8> WideIVs;
1056 SmallVector<PHINode*, 8> LoopPhis;
1057 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1058 LoopPhis.push_back(cast<PHINode>(I));
1060 // Each round of simplification iterates through the SimplifyIVUsers worklist
1061 // for all current phis, then determines whether any IVs can be
1062 // widened. Widening adds new phis to LoopPhis, inducing another round of
1063 // simplification on the wide IVs.
1064 while (!LoopPhis.empty()) {
1065 // Evaluate as many IV expressions as possible before widening any IVs. This
1066 // forces SCEV to set no-wrap flags before evaluating sign/zero
1067 // extension. The first time SCEV attempts to normalize sign/zero extension,
1068 // the result becomes final. So for the most predictable results, we delay
1069 // evaluation of sign/zero extend evaluation until needed, and avoid running
1070 // other SCEV based analysis prior to SimplifyAndExtend.
1072 PHINode *CurrIV = LoopPhis.pop_back_val();
1074 // Information about sign/zero extensions of CurrIV.
1075 WideIVVisitor WIV(CurrIV, SE, TD);
1077 Changed |= simplifyUsersOfIV(CurrIV, SE, &LPM, DeadInsts, &WIV);
1079 if (WIV.WI.WidestNativeType) {
1080 WideIVs.push_back(WIV.WI);
1082 } while(!LoopPhis.empty());
1084 for (; !WideIVs.empty(); WideIVs.pop_back()) {
1085 WidenIV Widener(WideIVs.back(), LI, SE, DT, DeadInsts);
1086 if (PHINode *WidePhi = Widener.CreateWideIV(Rewriter)) {
1088 LoopPhis.push_back(WidePhi);
1094 //===----------------------------------------------------------------------===//
1095 // LinearFunctionTestReplace and its kin. Rewrite the loop exit condition.
1096 //===----------------------------------------------------------------------===//
1098 /// Check for expressions that ScalarEvolution generates to compute
1099 /// BackedgeTakenInfo. If these expressions have not been reduced, then
1100 /// expanding them may incur additional cost (albeit in the loop preheader).
1101 static bool isHighCostExpansion(const SCEV *S, BranchInst *BI,
1102 SmallPtrSet<const SCEV*, 8> &Processed,
1103 ScalarEvolution *SE) {
1104 if (!Processed.insert(S))
1107 // If the backedge-taken count is a UDiv, it's very likely a UDiv that
1108 // ScalarEvolution's HowFarToZero or HowManyLessThans produced to compute a
1109 // precise expression, rather than a UDiv from the user's code. If we can't
1110 // find a UDiv in the code with some simple searching, assume the former and
1111 // forego rewriting the loop.
1112 if (isa<SCEVUDivExpr>(S)) {
1113 ICmpInst *OrigCond = dyn_cast<ICmpInst>(BI->getCondition());
1114 if (!OrigCond) return true;
1115 const SCEV *R = SE->getSCEV(OrigCond->getOperand(1));
1116 R = SE->getMinusSCEV(R, SE->getConstant(R->getType(), 1));
1118 const SCEV *L = SE->getSCEV(OrigCond->getOperand(0));
1119 L = SE->getMinusSCEV(L, SE->getConstant(L->getType(), 1));
1125 // Recurse past add expressions, which commonly occur in the
1126 // BackedgeTakenCount. They may already exist in program code, and if not,
1127 // they are not too expensive rematerialize.
1128 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
1129 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
1131 if (isHighCostExpansion(*I, BI, Processed, SE))
1137 // HowManyLessThans uses a Max expression whenever the loop is not guarded by
1138 // the exit condition.
1139 if (isa<SCEVSMaxExpr>(S) || isa<SCEVUMaxExpr>(S))
1142 // If we haven't recognized an expensive SCEV pattern, assume it's an
1143 // expression produced by program code.
1147 /// canExpandBackedgeTakenCount - Return true if this loop's backedge taken
1148 /// count expression can be safely and cheaply expanded into an instruction
1149 /// sequence that can be used by LinearFunctionTestReplace.
1151 /// TODO: This fails for pointer-type loop counters with greater than one byte
1152 /// strides, consequently preventing LFTR from running. For the purpose of LFTR
1153 /// we could skip this check in the case that the LFTR loop counter (chosen by
1154 /// FindLoopCounter) is also pointer type. Instead, we could directly convert
1155 /// the loop test to an inequality test by checking the target data's alignment
1156 /// of element types (given that the initial pointer value originates from or is
1157 /// used by ABI constrained operation, as opposed to inttoptr/ptrtoint).
1158 /// However, we don't yet have a strong motivation for converting loop tests
1159 /// into inequality tests.
1160 static bool canExpandBackedgeTakenCount(Loop *L, ScalarEvolution *SE) {
1161 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1162 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) ||
1163 BackedgeTakenCount->isZero())
1166 if (!L->getExitingBlock())
1169 // Can't rewrite non-branch yet.
1170 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
1174 SmallPtrSet<const SCEV*, 8> Processed;
1175 if (isHighCostExpansion(BackedgeTakenCount, BI, Processed, SE))
1181 /// getLoopPhiForCounter - Return the loop header phi IFF IncV adds a loop
1182 /// invariant value to the phi.
1183 static PHINode *getLoopPhiForCounter(Value *IncV, Loop *L, DominatorTree *DT) {
1184 Instruction *IncI = dyn_cast<Instruction>(IncV);
1188 switch (IncI->getOpcode()) {
1189 case Instruction::Add:
1190 case Instruction::Sub:
1192 case Instruction::GetElementPtr:
1193 // An IV counter must preserve its type.
1194 if (IncI->getNumOperands() == 2)
1200 PHINode *Phi = dyn_cast<PHINode>(IncI->getOperand(0));
1201 if (Phi && Phi->getParent() == L->getHeader()) {
1202 if (isLoopInvariant(IncI->getOperand(1), L, DT))
1206 if (IncI->getOpcode() == Instruction::GetElementPtr)
1209 // Allow add/sub to be commuted.
1210 Phi = dyn_cast<PHINode>(IncI->getOperand(1));
1211 if (Phi && Phi->getParent() == L->getHeader()) {
1212 if (isLoopInvariant(IncI->getOperand(0), L, DT))
1218 /// needsLFTR - LinearFunctionTestReplace policy. Return true unless we can show
1219 /// that the current exit test is already sufficiently canonical.
1220 static bool needsLFTR(Loop *L, DominatorTree *DT) {
1221 assert(L->getExitingBlock() && "expected loop exit");
1223 BasicBlock *LatchBlock = L->getLoopLatch();
1224 // Don't bother with LFTR if the loop is not properly simplified.
1228 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
1229 assert(BI && "expected exit branch");
1231 // Do LFTR to simplify the exit condition to an ICMP.
1232 ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition());
1236 // Do LFTR to simplify the exit ICMP to EQ/NE
1237 ICmpInst::Predicate Pred = Cond->getPredicate();
1238 if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ)
1241 // Look for a loop invariant RHS
1242 Value *LHS = Cond->getOperand(0);
1243 Value *RHS = Cond->getOperand(1);
1244 if (!isLoopInvariant(RHS, L, DT)) {
1245 if (!isLoopInvariant(LHS, L, DT))
1247 std::swap(LHS, RHS);
1249 // Look for a simple IV counter LHS
1250 PHINode *Phi = dyn_cast<PHINode>(LHS);
1252 Phi = getLoopPhiForCounter(LHS, L, DT);
1257 // Do LFTR if the exit condition's IV is *not* a simple counter.
1258 Value *IncV = Phi->getIncomingValueForBlock(L->getLoopLatch());
1259 return Phi != getLoopPhiForCounter(IncV, L, DT);
1262 /// AlmostDeadIV - Return true if this IV has any uses other than the (soon to
1263 /// be rewritten) loop exit test.
1264 static bool AlmostDeadIV(PHINode *Phi, BasicBlock *LatchBlock, Value *Cond) {
1265 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1266 Value *IncV = Phi->getIncomingValue(LatchIdx);
1268 for (Value::use_iterator UI = Phi->use_begin(), UE = Phi->use_end();
1270 if (*UI != Cond && *UI != IncV) return false;
1273 for (Value::use_iterator UI = IncV->use_begin(), UE = IncV->use_end();
1275 if (*UI != Cond && *UI != Phi) return false;
1280 /// FindLoopCounter - Find an affine IV in canonical form.
1282 /// BECount may be an i8* pointer type. The pointer difference is already
1283 /// valid count without scaling the address stride, so it remains a pointer
1284 /// expression as far as SCEV is concerned.
1286 /// FIXME: Accept -1 stride and set IVLimit = IVInit - BECount
1288 /// FIXME: Accept non-unit stride as long as SCEV can reduce BECount * Stride.
1289 /// This is difficult in general for SCEV because of potential overflow. But we
1290 /// could at least handle constant BECounts.
1292 FindLoopCounter(Loop *L, const SCEV *BECount,
1293 ScalarEvolution *SE, DominatorTree *DT, const TargetData *TD) {
1294 uint64_t BCWidth = SE->getTypeSizeInBits(BECount->getType());
1297 cast<BranchInst>(L->getExitingBlock()->getTerminator())->getCondition();
1299 // Loop over all of the PHI nodes, looking for a simple counter.
1300 PHINode *BestPhi = 0;
1301 const SCEV *BestInit = 0;
1302 BasicBlock *LatchBlock = L->getLoopLatch();
1303 assert(LatchBlock && "needsLFTR should guarantee a loop latch");
1305 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1306 PHINode *Phi = cast<PHINode>(I);
1307 if (!SE->isSCEVable(Phi->getType()))
1310 // Avoid comparing an integer IV against a pointer Limit.
1311 if (BECount->getType()->isPointerTy() && !Phi->getType()->isPointerTy())
1314 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Phi));
1315 if (!AR || AR->getLoop() != L || !AR->isAffine())
1318 // AR may be a pointer type, while BECount is an integer type.
1319 // AR may be wider than BECount. With eq/ne tests overflow is immaterial.
1320 // AR may not be a narrower type, or we may never exit.
1321 uint64_t PhiWidth = SE->getTypeSizeInBits(AR->getType());
1322 if (PhiWidth < BCWidth || (TD && !TD->isLegalInteger(PhiWidth)))
1325 const SCEV *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
1326 if (!Step || !Step->isOne())
1329 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1330 Value *IncV = Phi->getIncomingValue(LatchIdx);
1331 if (getLoopPhiForCounter(IncV, L, DT) != Phi)
1334 const SCEV *Init = AR->getStart();
1336 if (BestPhi && !AlmostDeadIV(BestPhi, LatchBlock, Cond)) {
1337 // Don't force a live loop counter if another IV can be used.
1338 if (AlmostDeadIV(Phi, LatchBlock, Cond))
1341 // Prefer to count-from-zero. This is a more "canonical" counter form. It
1342 // also prefers integer to pointer IVs.
1343 if (BestInit->isZero() != Init->isZero()) {
1344 if (BestInit->isZero())
1347 // If two IVs both count from zero or both count from nonzero then the
1348 // narrower is likely a dead phi that has been widened. Use the wider phi
1349 // to allow the other to be eliminated.
1350 if (PhiWidth <= SE->getTypeSizeInBits(BestPhi->getType()))
1359 /// genLoopLimit - Help LinearFunctionTestReplace by generating a value that
1360 /// holds the RHS of the new loop test.
1361 static Value *genLoopLimit(PHINode *IndVar, const SCEV *IVCount, Loop *L,
1362 SCEVExpander &Rewriter, ScalarEvolution *SE) {
1363 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
1364 assert(AR && AR->getLoop() == L && AR->isAffine() && "bad loop counter");
1365 const SCEV *IVInit = AR->getStart();
1367 // IVInit may be a pointer while IVCount is an integer when FindLoopCounter
1368 // finds a valid pointer IV. Sign extend BECount in order to materialize a
1369 // GEP. Avoid running SCEVExpander on a new pointer value, instead reusing
1370 // the existing GEPs whenever possible.
1371 if (IndVar->getType()->isPointerTy()
1372 && !IVCount->getType()->isPointerTy()) {
1374 Type *OfsTy = SE->getEffectiveSCEVType(IVInit->getType());
1375 const SCEV *IVOffset = SE->getTruncateOrSignExtend(IVCount, OfsTy);
1377 // Expand the code for the iteration count.
1378 assert(SE->isLoopInvariant(IVOffset, L) &&
1379 "Computed iteration count is not loop invariant!");
1380 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1381 Value *GEPOffset = Rewriter.expandCodeFor(IVOffset, OfsTy, BI);
1383 Value *GEPBase = IndVar->getIncomingValueForBlock(L->getLoopPreheader());
1384 assert(AR->getStart() == SE->getSCEV(GEPBase) && "bad loop counter");
1385 // We could handle pointer IVs other than i8*, but we need to compensate for
1386 // gep index scaling. See canExpandBackedgeTakenCount comments.
1387 assert(SE->getSizeOfExpr(
1388 cast<PointerType>(GEPBase->getType())->getElementType())->isOne()
1389 && "unit stride pointer IV must be i8*");
1391 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
1392 return Builder.CreateGEP(GEPBase, GEPOffset, "lftr.limit");
1395 // In any other case, convert both IVInit and IVCount to integers before
1396 // comparing. This may result in SCEV expension of pointers, but in practice
1397 // SCEV will fold the pointer arithmetic away as such:
1398 // BECount = (IVEnd - IVInit - 1) => IVLimit = IVInit (postinc).
1400 // Valid Cases: (1) both integers is most common; (2) both may be pointers
1401 // for simple memset-style loops; (3) IVInit is an integer and IVCount is a
1402 // pointer may occur when enable-iv-rewrite generates a canonical IV on top
1405 const SCEV *IVLimit = 0;
1406 // For unit stride, IVCount = Start + BECount with 2's complement overflow.
1407 // For non-zero Start, compute IVCount here.
1408 if (AR->getStart()->isZero())
1411 assert(AR->getStepRecurrence(*SE)->isOne() && "only handles unit stride");
1412 const SCEV *IVInit = AR->getStart();
1414 // For integer IVs, truncate the IV before computing IVInit + BECount.
1415 if (SE->getTypeSizeInBits(IVInit->getType())
1416 > SE->getTypeSizeInBits(IVCount->getType()))
1417 IVInit = SE->getTruncateExpr(IVInit, IVCount->getType());
1419 IVLimit = SE->getAddExpr(IVInit, IVCount);
1421 // Expand the code for the iteration count.
1422 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1423 IRBuilder<> Builder(BI);
1424 assert(SE->isLoopInvariant(IVLimit, L) &&
1425 "Computed iteration count is not loop invariant!");
1426 // Ensure that we generate the same type as IndVar, or a smaller integer
1427 // type. In the presence of null pointer values, we have an integer type
1428 // SCEV expression (IVInit) for a pointer type IV value (IndVar).
1429 Type *LimitTy = IVCount->getType()->isPointerTy() ?
1430 IndVar->getType() : IVCount->getType();
1431 return Rewriter.expandCodeFor(IVLimit, LimitTy, BI);
1435 /// LinearFunctionTestReplace - This method rewrites the exit condition of the
1436 /// loop to be a canonical != comparison against the incremented loop induction
1437 /// variable. This pass is able to rewrite the exit tests of any loop where the
1438 /// SCEV analysis can determine a loop-invariant trip count of the loop, which
1439 /// is actually a much broader range than just linear tests.
1440 Value *IndVarSimplify::
1441 LinearFunctionTestReplace(Loop *L,
1442 const SCEV *BackedgeTakenCount,
1444 SCEVExpander &Rewriter) {
1445 assert(canExpandBackedgeTakenCount(L, SE) && "precondition");
1447 // LFTR can ignore IV overflow and truncate to the width of
1448 // BECount. This avoids materializing the add(zext(add)) expression.
1449 Type *CntTy = BackedgeTakenCount->getType();
1451 const SCEV *IVCount = BackedgeTakenCount;
1453 // If the exiting block is the same as the backedge block, we prefer to
1454 // compare against the post-incremented value, otherwise we must compare
1455 // against the preincremented value.
1457 if (L->getExitingBlock() == L->getLoopLatch()) {
1458 // Add one to the "backedge-taken" count to get the trip count.
1459 // If this addition may overflow, we have to be more pessimistic and
1460 // cast the induction variable before doing the add.
1462 SE->getAddExpr(IVCount, SE->getConstant(IVCount->getType(), 1));
1463 if (CntTy == IVCount->getType())
1466 const SCEV *Zero = SE->getConstant(IVCount->getType(), 0);
1467 if ((isa<SCEVConstant>(N) && !N->isZero()) ||
1468 SE->isLoopEntryGuardedByCond(L, ICmpInst::ICMP_NE, N, Zero)) {
1469 // No overflow. Cast the sum.
1470 IVCount = SE->getTruncateOrZeroExtend(N, CntTy);
1472 // Potential overflow. Cast before doing the add.
1473 IVCount = SE->getTruncateOrZeroExtend(IVCount, CntTy);
1474 IVCount = SE->getAddExpr(IVCount, SE->getConstant(CntTy, 1));
1477 // The BackedgeTaken expression contains the number of times that the
1478 // backedge branches to the loop header. This is one less than the
1479 // number of times the loop executes, so use the incremented indvar.
1480 CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock());
1482 // We must use the preincremented value...
1483 IVCount = SE->getTruncateOrZeroExtend(IVCount, CntTy);
1487 Value *ExitCnt = genLoopLimit(IndVar, IVCount, L, Rewriter, SE);
1488 assert(ExitCnt->getType()->isPointerTy() == IndVar->getType()->isPointerTy()
1489 && "genLoopLimit missed a cast");
1491 // Insert a new icmp_ne or icmp_eq instruction before the branch.
1492 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1493 ICmpInst::Predicate P;
1494 if (L->contains(BI->getSuccessor(0)))
1495 P = ICmpInst::ICMP_NE;
1497 P = ICmpInst::ICMP_EQ;
1499 DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
1500 << " LHS:" << *CmpIndVar << '\n'
1502 << (P == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
1503 << " RHS:\t" << *ExitCnt << "\n"
1504 << " IVCount:\t" << *IVCount << "\n");
1506 IRBuilder<> Builder(BI);
1507 if (SE->getTypeSizeInBits(CmpIndVar->getType())
1508 > SE->getTypeSizeInBits(ExitCnt->getType())) {
1509 CmpIndVar = Builder.CreateTrunc(CmpIndVar, ExitCnt->getType(),
1513 Value *Cond = Builder.CreateICmp(P, CmpIndVar, ExitCnt, "exitcond");
1514 Value *OrigCond = BI->getCondition();
1515 // It's tempting to use replaceAllUsesWith here to fully replace the old
1516 // comparison, but that's not immediately safe, since users of the old
1517 // comparison may not be dominated by the new comparison. Instead, just
1518 // update the branch to use the new comparison; in the common case this
1519 // will make old comparison dead.
1520 BI->setCondition(Cond);
1521 DeadInsts.push_back(OrigCond);
1528 //===----------------------------------------------------------------------===//
1529 // SinkUnusedInvariants. A late subpass to cleanup loop preheaders.
1530 //===----------------------------------------------------------------------===//
1532 /// If there's a single exit block, sink any loop-invariant values that
1533 /// were defined in the preheader but not used inside the loop into the
1534 /// exit block to reduce register pressure in the loop.
1535 void IndVarSimplify::SinkUnusedInvariants(Loop *L) {
1536 BasicBlock *ExitBlock = L->getExitBlock();
1537 if (!ExitBlock) return;
1539 BasicBlock *Preheader = L->getLoopPreheader();
1540 if (!Preheader) return;
1542 Instruction *InsertPt = ExitBlock->getFirstInsertionPt();
1543 BasicBlock::iterator I = Preheader->getTerminator();
1544 while (I != Preheader->begin()) {
1546 // New instructions were inserted at the end of the preheader.
1547 if (isa<PHINode>(I))
1550 // Don't move instructions which might have side effects, since the side
1551 // effects need to complete before instructions inside the loop. Also don't
1552 // move instructions which might read memory, since the loop may modify
1553 // memory. Note that it's okay if the instruction might have undefined
1554 // behavior: LoopSimplify guarantees that the preheader dominates the exit
1556 if (I->mayHaveSideEffects() || I->mayReadFromMemory())
1559 // Skip debug info intrinsics.
1560 if (isa<DbgInfoIntrinsic>(I))
1563 // Skip landingpad instructions.
1564 if (isa<LandingPadInst>(I))
1567 // Don't sink alloca: we never want to sink static alloca's out of the
1568 // entry block, and correctly sinking dynamic alloca's requires
1569 // checks for stacksave/stackrestore intrinsics.
1570 // FIXME: Refactor this check somehow?
1571 if (isa<AllocaInst>(I))
1574 // Determine if there is a use in or before the loop (direct or
1576 bool UsedInLoop = false;
1577 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
1580 BasicBlock *UseBB = cast<Instruction>(U)->getParent();
1581 if (PHINode *P = dyn_cast<PHINode>(U)) {
1583 PHINode::getIncomingValueNumForOperand(UI.getOperandNo());
1584 UseBB = P->getIncomingBlock(i);
1586 if (UseBB == Preheader || L->contains(UseBB)) {
1592 // If there is, the def must remain in the preheader.
1596 // Otherwise, sink it to the exit block.
1597 Instruction *ToMove = I;
1600 if (I != Preheader->begin()) {
1601 // Skip debug info intrinsics.
1604 } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin());
1606 if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin())
1612 ToMove->moveBefore(InsertPt);
1618 //===----------------------------------------------------------------------===//
1619 // IndVarSimplify driver. Manage several subpasses of IV simplification.
1620 //===----------------------------------------------------------------------===//
1622 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
1623 // If LoopSimplify form is not available, stay out of trouble. Some notes:
1624 // - LSR currently only supports LoopSimplify-form loops. Indvars'
1625 // canonicalization can be a pessimization without LSR to "clean up"
1627 // - We depend on having a preheader; in particular,
1628 // Loop::getCanonicalInductionVariable only supports loops with preheaders,
1629 // and we're in trouble if we can't find the induction variable even when
1630 // we've manually inserted one.
1631 if (!L->isLoopSimplifyForm())
1634 LI = &getAnalysis<LoopInfo>();
1635 SE = &getAnalysis<ScalarEvolution>();
1636 DT = &getAnalysis<DominatorTree>();
1637 TD = getAnalysisIfAvailable<TargetData>();
1642 // If there are any floating-point recurrences, attempt to
1643 // transform them to use integer recurrences.
1644 RewriteNonIntegerIVs(L);
1646 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1648 // Create a rewriter object which we'll use to transform the code with.
1649 SCEVExpander Rewriter(*SE, "indvars");
1651 Rewriter.setDebugType(DEBUG_TYPE);
1654 // Eliminate redundant IV users.
1656 // Simplification works best when run before other consumers of SCEV. We
1657 // attempt to avoid evaluating SCEVs for sign/zero extend operations until
1658 // other expressions involving loop IVs have been evaluated. This helps SCEV
1659 // set no-wrap flags before normalizing sign/zero extension.
1660 Rewriter.disableCanonicalMode();
1661 SimplifyAndExtend(L, Rewriter, LPM);
1663 // Check to see if this loop has a computable loop-invariant execution count.
1664 // If so, this means that we can compute the final value of any expressions
1665 // that are recurrent in the loop, and substitute the exit values from the
1666 // loop into any instructions outside of the loop that use the final values of
1667 // the current expressions.
1669 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1670 RewriteLoopExitValues(L, Rewriter);
1672 // Eliminate redundant IV cycles.
1673 NumElimIV += Rewriter.replaceCongruentIVs(L, DT, DeadInsts);
1675 // If we have a trip count expression, rewrite the loop's exit condition
1676 // using it. We can currently only handle loops with a single exit.
1677 if (canExpandBackedgeTakenCount(L, SE) && needsLFTR(L, DT)) {
1678 PHINode *IndVar = FindLoopCounter(L, BackedgeTakenCount, SE, DT, TD);
1680 // Check preconditions for proper SCEVExpander operation. SCEV does not
1681 // express SCEVExpander's dependencies, such as LoopSimplify. Instead any
1682 // pass that uses the SCEVExpander must do it. This does not work well for
1683 // loop passes because SCEVExpander makes assumptions about all loops, while
1684 // LoopPassManager only forces the current loop to be simplified.
1686 // FIXME: SCEV expansion has no way to bail out, so the caller must
1687 // explicitly check any assumptions made by SCEV. Brittle.
1688 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(BackedgeTakenCount);
1689 if (!AR || AR->getLoop()->getLoopPreheader())
1690 (void)LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar,
1694 // Clear the rewriter cache, because values that are in the rewriter's cache
1695 // can be deleted in the loop below, causing the AssertingVH in the cache to
1699 // Now that we're done iterating through lists, clean up any instructions
1700 // which are now dead.
1701 while (!DeadInsts.empty())
1702 if (Instruction *Inst =
1703 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
1704 RecursivelyDeleteTriviallyDeadInstructions(Inst);
1706 // The Rewriter may not be used from this point on.
1708 // Loop-invariant instructions in the preheader that aren't used in the
1709 // loop may be sunk below the loop to reduce register pressure.
1710 SinkUnusedInvariants(L);
1712 // Clean up dead instructions.
1713 Changed |= DeleteDeadPHIs(L->getHeader());
1714 // Check a post-condition.
1715 assert(L->isLCSSAForm(*DT) &&
1716 "Indvars did not leave the loop in lcssa form!");
1718 // Verify that LFTR, and any other change have not interfered with SCEV's
1719 // ability to compute trip count.
1721 if (VerifyIndvars && !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) {
1723 const SCEV *NewBECount = SE->getBackedgeTakenCount(L);
1724 if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) <
1725 SE->getTypeSizeInBits(NewBECount->getType()))
1726 NewBECount = SE->getTruncateOrNoop(NewBECount,
1727 BackedgeTakenCount->getType());
1729 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount,
1730 NewBECount->getType());
1731 assert(BackedgeTakenCount == NewBECount && "indvars must preserve SCEV");