1 //===-- LoopReroll.cpp - Loop rerolling pass ------------------------------===//
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 pass implements a simple loop reroller.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Transforms/Scalar.h"
15 #include "llvm/ADT/STLExtras.h"
16 #include "llvm/ADT/SmallSet.h"
17 #include "llvm/ADT/Statistic.h"
18 #include "llvm/Analysis/AliasAnalysis.h"
19 #include "llvm/Analysis/AliasSetTracker.h"
20 #include "llvm/Analysis/LoopPass.h"
21 #include "llvm/Analysis/ScalarEvolution.h"
22 #include "llvm/Analysis/ScalarEvolutionExpander.h"
23 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
24 #include "llvm/Analysis/ValueTracking.h"
25 #include "llvm/IR/DataLayout.h"
26 #include "llvm/IR/Dominators.h"
27 #include "llvm/IR/IntrinsicInst.h"
28 #include "llvm/Support/CommandLine.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Support/raw_ostream.h"
31 #include "llvm/Target/TargetLibraryInfo.h"
32 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
33 #include "llvm/Transforms/Utils/Local.h"
34 #include "llvm/Transforms/Utils/LoopUtils.h"
38 #define DEBUG_TYPE "loop-reroll"
40 STATISTIC(NumRerolledLoops, "Number of rerolled loops");
42 static cl::opt<unsigned>
43 MaxInc("max-reroll-increment", cl::init(2048), cl::Hidden,
44 cl::desc("The maximum increment for loop rerolling"));
46 // This loop re-rolling transformation aims to transform loops like this:
50 // for (int i = 0; i < 500; i += 3) {
57 // into a loop like this:
60 // for (int i = 0; i < 500; ++i)
64 // It does this by looking for loops that, besides the latch code, are composed
65 // of isomorphic DAGs of instructions, with each DAG rooted at some increment
66 // to the induction variable, and where each DAG is isomorphic to the DAG
67 // rooted at the induction variable (excepting the sub-DAGs which root the
68 // other induction-variable increments). In other words, we're looking for loop
69 // bodies of the form:
71 // %iv = phi [ (preheader, ...), (body, %iv.next) ]
73 // %iv.1 = add %iv, 1 <-- a root increment
75 // %iv.2 = add %iv, 2 <-- a root increment
77 // %iv.scale_m_1 = add %iv, scale-1 <-- a root increment
80 // %iv.next = add %iv, scale
81 // %cmp = icmp(%iv, ...)
82 // br %cmp, header, exit
84 // where each f(i) is a set of instructions that, collectively, are a function
85 // only of i (and other loop-invariant values).
87 // As a special case, we can also reroll loops like this:
91 // for (int i = 0; i < 500; ++i) {
100 // void bar(int *x) {
101 // for (int i = 0; i < 1500; ++i)
105 // in which case, we're looking for inputs like this:
107 // %iv = phi [ (preheader, ...), (body, %iv.next) ]
108 // %scaled.iv = mul %iv, scale
110 // %scaled.iv.1 = add %scaled.iv, 1
112 // %scaled.iv.2 = add %scaled.iv, 2
114 // %scaled.iv.scale_m_1 = add %scaled.iv, scale-1
115 // f(%scaled.iv.scale_m_1)
117 // %iv.next = add %iv, 1
118 // %cmp = icmp(%iv, ...)
119 // br %cmp, header, exit
122 class LoopReroll : public LoopPass {
124 static char ID; // Pass ID, replacement for typeid
125 LoopReroll() : LoopPass(ID) {
126 initializeLoopRerollPass(*PassRegistry::getPassRegistry());
129 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
131 void getAnalysisUsage(AnalysisUsage &AU) const override {
132 AU.addRequired<AliasAnalysis>();
133 AU.addRequired<LoopInfo>();
134 AU.addPreserved<LoopInfo>();
135 AU.addRequired<DominatorTreeWrapperPass>();
136 AU.addPreserved<DominatorTreeWrapperPass>();
137 AU.addRequired<ScalarEvolution>();
138 AU.addRequired<TargetLibraryInfo>();
145 const DataLayout *DL;
146 TargetLibraryInfo *TLI;
149 typedef SmallVector<Instruction *, 16> SmallInstructionVector;
150 typedef SmallSet<Instruction *, 16> SmallInstructionSet;
152 // A chain of isomorphic instructions, indentified by a single-use PHI,
153 // representing a reduction. Only the last value may be used outside the
155 struct SimpleLoopReduction {
156 SimpleLoopReduction(Instruction *P, Loop *L)
157 : Valid(false), Instructions(1, P) {
158 assert(isa<PHINode>(P) && "First reduction instruction must be a PHI");
166 Instruction *getPHI() const {
167 assert(Valid && "Using invalid reduction");
168 return Instructions.front();
171 Instruction *getReducedValue() const {
172 assert(Valid && "Using invalid reduction");
173 return Instructions.back();
176 Instruction *get(size_t i) const {
177 assert(Valid && "Using invalid reduction");
178 return Instructions[i+1];
181 Instruction *operator [] (size_t i) const { return get(i); }
183 // The size, ignoring the initial PHI.
184 size_t size() const {
185 assert(Valid && "Using invalid reduction");
186 return Instructions.size()-1;
189 typedef SmallInstructionVector::iterator iterator;
190 typedef SmallInstructionVector::const_iterator const_iterator;
193 assert(Valid && "Using invalid reduction");
194 return std::next(Instructions.begin());
197 const_iterator begin() const {
198 assert(Valid && "Using invalid reduction");
199 return std::next(Instructions.begin());
202 iterator end() { return Instructions.end(); }
203 const_iterator end() const { return Instructions.end(); }
207 SmallInstructionVector Instructions;
212 // The set of all reductions, and state tracking of possible reductions
213 // during loop instruction processing.
214 struct ReductionTracker {
215 typedef SmallVector<SimpleLoopReduction, 16> SmallReductionVector;
217 // Add a new possible reduction.
218 void addSLR(SimpleLoopReduction &SLR) {
219 PossibleReds.push_back(SLR);
222 // Setup to track possible reductions corresponding to the provided
223 // rerolling scale. Only reductions with a number of non-PHI instructions
224 // that is divisible by the scale are considered. Three instructions sets
226 // - A set of all possible instructions in eligible reductions.
227 // - A set of all PHIs in eligible reductions
228 // - A set of all reduced values (last instructions) in eligible reductions.
229 void restrictToScale(uint64_t Scale,
230 SmallInstructionSet &PossibleRedSet,
231 SmallInstructionSet &PossibleRedPHISet,
232 SmallInstructionSet &PossibleRedLastSet) {
233 PossibleRedIdx.clear();
234 PossibleRedIter.clear();
237 for (unsigned i = 0, e = PossibleReds.size(); i != e; ++i)
238 if (PossibleReds[i].size() % Scale == 0) {
239 PossibleRedLastSet.insert(PossibleReds[i].getReducedValue());
240 PossibleRedPHISet.insert(PossibleReds[i].getPHI());
242 PossibleRedSet.insert(PossibleReds[i].getPHI());
243 PossibleRedIdx[PossibleReds[i].getPHI()] = i;
244 for (Instruction *J : PossibleReds[i]) {
245 PossibleRedSet.insert(J);
246 PossibleRedIdx[J] = i;
251 // The functions below are used while processing the loop instructions.
253 // Are the two instructions both from reductions, and furthermore, from
254 // the same reduction?
255 bool isPairInSame(Instruction *J1, Instruction *J2) {
256 DenseMap<Instruction *, int>::iterator J1I = PossibleRedIdx.find(J1);
257 if (J1I != PossibleRedIdx.end()) {
258 DenseMap<Instruction *, int>::iterator J2I = PossibleRedIdx.find(J2);
259 if (J2I != PossibleRedIdx.end() && J1I->second == J2I->second)
266 // The two provided instructions, the first from the base iteration, and
267 // the second from iteration i, form a matched pair. If these are part of
268 // a reduction, record that fact.
269 void recordPair(Instruction *J1, Instruction *J2, unsigned i) {
270 if (PossibleRedIdx.count(J1)) {
271 assert(PossibleRedIdx.count(J2) &&
272 "Recording reduction vs. non-reduction instruction?");
274 PossibleRedIter[J1] = 0;
275 PossibleRedIter[J2] = i;
277 int Idx = PossibleRedIdx[J1];
278 assert(Idx == PossibleRedIdx[J2] &&
279 "Recording pair from different reductions?");
284 // The functions below can be called after we've finished processing all
285 // instructions in the loop, and we know which reductions were selected.
287 // Is the provided instruction the PHI of a reduction selected for
289 bool isSelectedPHI(Instruction *J) {
290 if (!isa<PHINode>(J))
293 for (DenseSet<int>::iterator RI = Reds.begin(), RIE = Reds.end();
296 if (cast<Instruction>(J) == PossibleReds[i].getPHI())
303 bool validateSelected();
304 void replaceSelected();
307 // The vector of all possible reductions (for any scale).
308 SmallReductionVector PossibleReds;
310 DenseMap<Instruction *, int> PossibleRedIdx;
311 DenseMap<Instruction *, int> PossibleRedIter;
315 void collectPossibleIVs(Loop *L, SmallInstructionVector &PossibleIVs);
316 void collectPossibleReductions(Loop *L,
317 ReductionTracker &Reductions);
318 void collectInLoopUserSet(Loop *L,
319 const SmallInstructionVector &Roots,
320 const SmallInstructionSet &Exclude,
321 const SmallInstructionSet &Final,
322 DenseSet<Instruction *> &Users);
323 void collectInLoopUserSet(Loop *L,
325 const SmallInstructionSet &Exclude,
326 const SmallInstructionSet &Final,
327 DenseSet<Instruction *> &Users);
328 bool findScaleFromMul(Instruction *RealIV, uint64_t &Scale,
330 SmallInstructionVector &LoopIncs);
331 bool collectAllRoots(Loop *L, uint64_t Inc, uint64_t Scale, Instruction *IV,
332 SmallVector<SmallInstructionVector, 32> &Roots,
333 SmallInstructionSet &AllRoots,
334 SmallInstructionVector &LoopIncs);
335 bool reroll(Instruction *IV, Loop *L, BasicBlock *Header, const SCEV *IterCount,
336 ReductionTracker &Reductions);
340 char LoopReroll::ID = 0;
341 INITIALIZE_PASS_BEGIN(LoopReroll, "loop-reroll", "Reroll loops", false, false)
342 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
343 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
344 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
345 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
346 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
347 INITIALIZE_PASS_END(LoopReroll, "loop-reroll", "Reroll loops", false, false)
349 Pass *llvm::createLoopRerollPass() {
350 return new LoopReroll;
353 // Returns true if the provided instruction is used outside the given loop.
354 // This operates like Instruction::isUsedOutsideOfBlock, but considers PHIs in
355 // non-loop blocks to be outside the loop.
356 static bool hasUsesOutsideLoop(Instruction *I, Loop *L) {
357 for (User *U : I->users())
358 if (!L->contains(cast<Instruction>(U)))
364 // Collect the list of loop induction variables with respect to which it might
365 // be possible to reroll the loop.
366 void LoopReroll::collectPossibleIVs(Loop *L,
367 SmallInstructionVector &PossibleIVs) {
368 BasicBlock *Header = L->getHeader();
369 for (BasicBlock::iterator I = Header->begin(),
370 IE = Header->getFirstInsertionPt(); I != IE; ++I) {
371 if (!isa<PHINode>(I))
373 if (!I->getType()->isIntegerTy())
376 if (const SCEVAddRecExpr *PHISCEV =
377 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(I))) {
378 if (PHISCEV->getLoop() != L)
380 if (!PHISCEV->isAffine())
382 if (const SCEVConstant *IncSCEV =
383 dyn_cast<SCEVConstant>(PHISCEV->getStepRecurrence(*SE))) {
384 if (!IncSCEV->getValue()->getValue().isStrictlyPositive())
386 if (IncSCEV->getValue()->uge(MaxInc))
389 DEBUG(dbgs() << "LRR: Possible IV: " << *I << " = " <<
391 PossibleIVs.push_back(I);
397 // Add the remainder of the reduction-variable chain to the instruction vector
398 // (the initial PHINode has already been added). If successful, the object is
400 void LoopReroll::SimpleLoopReduction::add(Loop *L) {
401 assert(!Valid && "Cannot add to an already-valid chain");
403 // The reduction variable must be a chain of single-use instructions
404 // (including the PHI), except for the last value (which is used by the PHI
405 // and also outside the loop).
406 Instruction *C = Instructions.front();
409 C = cast<Instruction>(*C->user_begin());
410 if (C->hasOneUse()) {
411 if (!C->isBinaryOp())
414 if (!(isa<PHINode>(Instructions.back()) ||
415 C->isSameOperationAs(Instructions.back())))
418 Instructions.push_back(C);
420 } while (C->hasOneUse());
422 if (Instructions.size() < 2 ||
423 !C->isSameOperationAs(Instructions.back()) ||
427 // C is now the (potential) last instruction in the reduction chain.
428 for (User *U : C->users())
429 // The only in-loop user can be the initial PHI.
430 if (L->contains(cast<Instruction>(U)))
431 if (cast<Instruction>(U) != Instructions.front())
434 Instructions.push_back(C);
438 // Collect the vector of possible reduction variables.
439 void LoopReroll::collectPossibleReductions(Loop *L,
440 ReductionTracker &Reductions) {
441 BasicBlock *Header = L->getHeader();
442 for (BasicBlock::iterator I = Header->begin(),
443 IE = Header->getFirstInsertionPt(); I != IE; ++I) {
444 if (!isa<PHINode>(I))
446 if (!I->getType()->isSingleValueType())
449 SimpleLoopReduction SLR(I, L);
453 DEBUG(dbgs() << "LRR: Possible reduction: " << *I << " (with " <<
454 SLR.size() << " chained instructions)\n");
455 Reductions.addSLR(SLR);
459 // Collect the set of all users of the provided root instruction. This set of
460 // users contains not only the direct users of the root instruction, but also
461 // all users of those users, and so on. There are two exceptions:
463 // 1. Instructions in the set of excluded instructions are never added to the
464 // use set (even if they are users). This is used, for example, to exclude
465 // including root increments in the use set of the primary IV.
467 // 2. Instructions in the set of final instructions are added to the use set
468 // if they are users, but their users are not added. This is used, for
469 // example, to prevent a reduction update from forcing all later reduction
470 // updates into the use set.
471 void LoopReroll::collectInLoopUserSet(Loop *L,
472 Instruction *Root, const SmallInstructionSet &Exclude,
473 const SmallInstructionSet &Final,
474 DenseSet<Instruction *> &Users) {
475 SmallInstructionVector Queue(1, Root);
476 while (!Queue.empty()) {
477 Instruction *I = Queue.pop_back_val();
478 if (!Users.insert(I).second)
482 for (Use &U : I->uses()) {
483 Instruction *User = cast<Instruction>(U.getUser());
484 if (PHINode *PN = dyn_cast<PHINode>(User)) {
485 // Ignore "wrap-around" uses to PHIs of this loop's header.
486 if (PN->getIncomingBlock(U) == L->getHeader())
490 if (L->contains(User) && !Exclude.count(User)) {
491 Queue.push_back(User);
495 // We also want to collect single-user "feeder" values.
496 for (User::op_iterator OI = I->op_begin(),
497 OIE = I->op_end(); OI != OIE; ++OI) {
498 if (Instruction *Op = dyn_cast<Instruction>(*OI))
499 if (Op->hasOneUse() && L->contains(Op) && !Exclude.count(Op) &&
506 // Collect all of the users of all of the provided root instructions (combined
507 // into a single set).
508 void LoopReroll::collectInLoopUserSet(Loop *L,
509 const SmallInstructionVector &Roots,
510 const SmallInstructionSet &Exclude,
511 const SmallInstructionSet &Final,
512 DenseSet<Instruction *> &Users) {
513 for (SmallInstructionVector::const_iterator I = Roots.begin(),
514 IE = Roots.end(); I != IE; ++I)
515 collectInLoopUserSet(L, *I, Exclude, Final, Users);
518 static bool isSimpleLoadStore(Instruction *I) {
519 if (LoadInst *LI = dyn_cast<LoadInst>(I))
520 return LI->isSimple();
521 if (StoreInst *SI = dyn_cast<StoreInst>(I))
522 return SI->isSimple();
523 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I))
524 return !MI->isVolatile();
528 // Recognize loops that are setup like this:
530 // %iv = phi [ (preheader, ...), (body, %iv.next) ]
531 // %scaled.iv = mul %iv, scale
533 // %scaled.iv.1 = add %scaled.iv, 1
535 // %scaled.iv.2 = add %scaled.iv, 2
537 // %scaled.iv.scale_m_1 = add %scaled.iv, scale-1
538 // f(%scaled.iv.scale_m_1)
540 // %iv.next = add %iv, 1
541 // %cmp = icmp(%iv, ...)
542 // br %cmp, header, exit
544 // and, if found, set IV = %scaled.iv, and add %iv.next to LoopIncs.
545 bool LoopReroll::findScaleFromMul(Instruction *RealIV, uint64_t &Scale,
547 SmallInstructionVector &LoopIncs) {
548 // This is a special case: here we're looking for all uses (except for
549 // the increment) to be multiplied by a common factor. The increment must
550 // be by one. This is to capture loops like:
551 // for (int i = 0; i < 500; ++i) {
552 // foo(3*i); foo(3*i+1); foo(3*i+2);
554 if (RealIV->getNumUses() != 2)
556 const SCEVAddRecExpr *RealIVSCEV = cast<SCEVAddRecExpr>(SE->getSCEV(RealIV));
557 Instruction *User1 = cast<Instruction>(*RealIV->user_begin()),
558 *User2 = cast<Instruction>(*std::next(RealIV->user_begin()));
559 if (!SE->isSCEVable(User1->getType()) || !SE->isSCEVable(User2->getType()))
561 const SCEVAddRecExpr *User1SCEV =
562 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(User1)),
564 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(User2));
565 if (!User1SCEV || !User1SCEV->isAffine() ||
566 !User2SCEV || !User2SCEV->isAffine())
569 // We assume below that User1 is the scale multiply and User2 is the
570 // increment. If this can't be true, then swap them.
571 if (User1SCEV == RealIVSCEV->getPostIncExpr(*SE)) {
572 std::swap(User1, User2);
573 std::swap(User1SCEV, User2SCEV);
576 if (User2SCEV != RealIVSCEV->getPostIncExpr(*SE))
578 assert(User2SCEV->getStepRecurrence(*SE)->isOne() &&
579 "Invalid non-unit step for multiplicative scaling");
580 LoopIncs.push_back(User2);
582 if (const SCEVConstant *MulScale =
583 dyn_cast<SCEVConstant>(User1SCEV->getStepRecurrence(*SE))) {
584 // Make sure that both the start and step have the same multiplier.
585 if (RealIVSCEV->getStart()->getType() != MulScale->getType())
587 if (SE->getMulExpr(RealIVSCEV->getStart(), MulScale) !=
588 User1SCEV->getStart())
591 ConstantInt *MulScaleCI = MulScale->getValue();
592 if (!MulScaleCI->uge(2) || MulScaleCI->uge(MaxInc))
594 Scale = MulScaleCI->getZExtValue();
599 DEBUG(dbgs() << "LRR: Found possible scaling " << *User1 << "\n");
603 // Collect all root increments with respect to the provided induction variable
604 // (normally the PHI, but sometimes a multiply). A root increment is an
605 // instruction, normally an add, with a positive constant less than Scale. In a
606 // rerollable loop, each of these increments is the root of an instruction
607 // graph isomorphic to the others. Also, we collect the final induction
608 // increment (the increment equal to the Scale), and its users in LoopIncs.
609 bool LoopReroll::collectAllRoots(Loop *L, uint64_t Inc, uint64_t Scale,
611 SmallVector<SmallInstructionVector, 32> &Roots,
612 SmallInstructionSet &AllRoots,
613 SmallInstructionVector &LoopIncs) {
614 for (User *U : IV->users()) {
615 Instruction *UI = cast<Instruction>(U);
616 if (!SE->isSCEVable(UI->getType()))
618 if (UI->getType() != IV->getType())
620 if (!L->contains(UI))
622 if (hasUsesOutsideLoop(UI, L))
625 if (const SCEVConstant *Diff = dyn_cast<SCEVConstant>(SE->getMinusSCEV(
626 SE->getSCEV(UI), SE->getSCEV(IV)))) {
627 uint64_t Idx = Diff->getValue()->getValue().getZExtValue();
628 if (Idx > 0 && Idx < Scale) {
629 Roots[Idx-1].push_back(UI);
631 } else if (Idx == Scale && Inc > 1) {
632 LoopIncs.push_back(UI);
637 if (Roots[0].empty())
640 for (unsigned i = 1; i < Scale-1; ++i)
641 if (Roots[i].size() != Roots[0].size()) {
652 // Validate the selected reductions. All iterations must have an isomorphic
653 // part of the reduction chain and, for non-associative reductions, the chain
654 // entries must appear in order.
655 bool LoopReroll::ReductionTracker::validateSelected() {
656 // For a non-associative reduction, the chain entries must appear in order.
657 for (DenseSet<int>::iterator RI = Reds.begin(), RIE = Reds.end();
660 int PrevIter = 0, BaseCount = 0, Count = 0;
661 for (Instruction *J : PossibleReds[i]) {
662 // Note that all instructions in the chain must have been found because
663 // all instructions in the function must have been assigned to some
665 int Iter = PossibleRedIter[J];
666 if (Iter != PrevIter && Iter != PrevIter + 1 &&
667 !PossibleReds[i].getReducedValue()->isAssociative()) {
668 DEBUG(dbgs() << "LRR: Out-of-order non-associative reduction: " <<
673 if (Iter != PrevIter) {
674 if (Count != BaseCount) {
675 DEBUG(dbgs() << "LRR: Iteration " << PrevIter <<
676 " reduction use count " << Count <<
677 " is not equal to the base use count " <<
696 // For all selected reductions, remove all parts except those in the first
697 // iteration (and the PHI). Replace outside uses of the reduced value with uses
698 // of the first-iteration reduced value (in other words, reroll the selected
700 void LoopReroll::ReductionTracker::replaceSelected() {
701 // Fixup reductions to refer to the last instruction associated with the
702 // first iteration (not the last).
703 for (DenseSet<int>::iterator RI = Reds.begin(), RIE = Reds.end();
707 for (int e = PossibleReds[i].size(); j != e; ++j)
708 if (PossibleRedIter[PossibleReds[i][j]] != 0) {
713 // Replace users with the new end-of-chain value.
714 SmallInstructionVector Users;
715 for (User *U : PossibleReds[i].getReducedValue()->users())
716 Users.push_back(cast<Instruction>(U));
718 for (SmallInstructionVector::iterator J = Users.begin(),
719 JE = Users.end(); J != JE; ++J)
720 (*J)->replaceUsesOfWith(PossibleReds[i].getReducedValue(),
725 // Reroll the provided loop with respect to the provided induction variable.
726 // Generally, we're looking for a loop like this:
728 // %iv = phi [ (preheader, ...), (body, %iv.next) ]
730 // %iv.1 = add %iv, 1 <-- a root increment
732 // %iv.2 = add %iv, 2 <-- a root increment
734 // %iv.scale_m_1 = add %iv, scale-1 <-- a root increment
737 // %iv.next = add %iv, scale
738 // %cmp = icmp(%iv, ...)
739 // br %cmp, header, exit
741 // Notably, we do not require that f(%iv), f(%iv.1), etc. be isolated groups of
742 // instructions. In other words, the instructions in f(%iv), f(%iv.1), etc. can
743 // be intermixed with eachother. The restriction imposed by this algorithm is
744 // that the relative order of the isomorphic instructions in f(%iv), f(%iv.1),
747 // First, we collect the use set of %iv, excluding the other increment roots.
748 // This gives us f(%iv). Then we iterate over the loop instructions (scale-1)
749 // times, having collected the use set of f(%iv.(i+1)), during which we:
750 // - Ensure that the next unmatched instruction in f(%iv) is isomorphic to
751 // the next unmatched instruction in f(%iv.(i+1)).
752 // - Ensure that both matched instructions don't have any external users
753 // (with the exception of last-in-chain reduction instructions).
754 // - Track the (aliasing) write set, and other side effects, of all
755 // instructions that belong to future iterations that come before the matched
756 // instructions. If the matched instructions read from that write set, then
757 // f(%iv) or f(%iv.(i+1)) has some dependency on instructions in
758 // f(%iv.(j+1)) for some j > i, and we cannot reroll the loop. Similarly,
759 // if any of these future instructions had side effects (could not be
760 // speculatively executed), and so do the matched instructions, when we
761 // cannot reorder those side-effect-producing instructions, and rerolling
764 // Finally, we make sure that all loop instructions are either loop increment
765 // roots, belong to simple latch code, parts of validated reductions, part of
766 // f(%iv) or part of some f(%iv.i). If all of that is true (and all reductions
767 // have been validated), then we reroll the loop.
768 bool LoopReroll::reroll(Instruction *IV, Loop *L, BasicBlock *Header,
769 const SCEV *IterCount,
770 ReductionTracker &Reductions) {
771 const SCEVAddRecExpr *RealIVSCEV = cast<SCEVAddRecExpr>(SE->getSCEV(IV));
772 uint64_t Inc = cast<SCEVConstant>(RealIVSCEV->getOperand(1))->
773 getValue()->getZExtValue();
774 // The collection of loop increment instructions.
775 SmallInstructionVector LoopIncs;
776 uint64_t Scale = Inc;
778 // The effective induction variable, IV, is normally also the real induction
779 // variable. When we're dealing with a loop like:
780 // for (int i = 0; i < 500; ++i)
784 // then the real IV is still i, but the effective IV is (3*i).
785 Instruction *RealIV = IV;
786 if (Inc == 1 && !findScaleFromMul(RealIV, Scale, IV, LoopIncs))
789 assert(Scale <= MaxInc && "Scale is too large");
790 assert(Scale > 1 && "Scale must be at least 2");
792 // The set of increment instructions for each increment value.
793 SmallVector<SmallInstructionVector, 32> Roots(Scale-1);
794 SmallInstructionSet AllRoots;
795 if (!collectAllRoots(L, Inc, Scale, IV, Roots, AllRoots, LoopIncs))
798 DEBUG(dbgs() << "LRR: Found all root induction increments for: " <<
801 // An array of just the possible reductions for this scale factor. When we
802 // collect the set of all users of some root instructions, these reduction
803 // instructions are treated as 'final' (their uses are not considered).
804 // This is important because we don't want the root use set to search down
805 // the reduction chain.
806 SmallInstructionSet PossibleRedSet;
807 SmallInstructionSet PossibleRedLastSet, PossibleRedPHISet;
808 Reductions.restrictToScale(Scale, PossibleRedSet, PossibleRedPHISet,
811 // We now need to check for equivalence of the use graph of each root with
812 // that of the primary induction variable (excluding the roots). Our goal
813 // here is not to solve the full graph isomorphism problem, but rather to
814 // catch common cases without a lot of work. As a result, we will assume
815 // that the relative order of the instructions in each unrolled iteration
816 // is the same (although we will not make an assumption about how the
817 // different iterations are intermixed). Note that while the order must be
818 // the same, the instructions may not be in the same basic block.
819 SmallInstructionSet Exclude(AllRoots);
820 Exclude.insert(LoopIncs.begin(), LoopIncs.end());
822 DenseSet<Instruction *> BaseUseSet;
823 collectInLoopUserSet(L, IV, Exclude, PossibleRedSet, BaseUseSet);
825 DenseSet<Instruction *> AllRootUses;
826 std::vector<DenseSet<Instruction *> > RootUseSets(Scale-1);
828 bool MatchFailed = false;
829 for (unsigned i = 0; i < Scale-1 && !MatchFailed; ++i) {
830 DenseSet<Instruction *> &RootUseSet = RootUseSets[i];
831 collectInLoopUserSet(L, Roots[i], SmallInstructionSet(),
832 PossibleRedSet, RootUseSet);
834 DEBUG(dbgs() << "LRR: base use set size: " << BaseUseSet.size() <<
835 " vs. iteration increment " << (i+1) <<
836 " use set size: " << RootUseSet.size() << "\n");
838 if (BaseUseSet.size() != RootUseSet.size()) {
843 // In addition to regular aliasing information, we need to look for
844 // instructions from later (future) iterations that have side effects
845 // preventing us from reordering them past other instructions with side
847 bool FutureSideEffects = false;
848 AliasSetTracker AST(*AA);
850 // The map between instructions in f(%iv.(i+1)) and f(%iv).
851 DenseMap<Value *, Value *> BaseMap;
853 assert(L->getNumBlocks() == 1 && "Cannot handle multi-block loops");
854 for (BasicBlock::iterator J1 = Header->begin(), J2 = Header->begin(),
855 JE = Header->end(); J1 != JE && !MatchFailed; ++J1) {
856 if (cast<Instruction>(J1) == RealIV)
858 if (cast<Instruction>(J1) == IV)
860 if (!BaseUseSet.count(J1))
862 if (PossibleRedPHISet.count(J1)) // Skip reduction PHIs.
865 while (J2 != JE && (!RootUseSet.count(J2) ||
866 std::find(Roots[i].begin(), Roots[i].end(), J2) !=
868 // As we iterate through the instructions, instructions that don't
869 // belong to previous iterations (or the base case), must belong to
870 // future iterations. We want to track the alias set of writes from
871 // previous iterations.
872 if (!isa<PHINode>(J2) && !BaseUseSet.count(J2) &&
873 !AllRootUses.count(J2)) {
874 if (J2->mayWriteToMemory())
877 // Note: This is specifically guarded by a check on isa<PHINode>,
878 // which while a valid (somewhat arbitrary) micro-optimization, is
879 // needed because otherwise isSafeToSpeculativelyExecute returns
880 // false on PHI nodes.
881 if (!isSimpleLoadStore(J2) && !isSafeToSpeculativelyExecute(J2, DL))
882 FutureSideEffects = true;
888 if (!J1->isSameOperationAs(J2)) {
889 DEBUG(dbgs() << "LRR: iteration root match failed at " << *J1 <<
890 " vs. " << *J2 << "\n");
895 // Make sure that this instruction, which is in the use set of this
896 // root instruction, does not also belong to the base set or the set of
897 // some previous root instruction.
898 if (BaseUseSet.count(J2) || AllRootUses.count(J2)) {
899 DEBUG(dbgs() << "LRR: iteration root match failed at " << *J1 <<
900 " vs. " << *J2 << " (prev. case overlap)\n");
905 // Make sure that we don't alias with any instruction in the alias set
906 // tracker. If we do, then we depend on a future iteration, and we
908 if (J2->mayReadFromMemory()) {
909 for (AliasSetTracker::iterator K = AST.begin(), KE = AST.end();
910 K != KE && !MatchFailed; ++K) {
911 if (K->aliasesUnknownInst(J2, *AA)) {
912 DEBUG(dbgs() << "LRR: iteration root match failed at " << *J1 <<
913 " vs. " << *J2 << " (depends on future store)\n");
920 // If we've past an instruction from a future iteration that may have
921 // side effects, and this instruction might also, then we can't reorder
922 // them, and this matching fails. As an exception, we allow the alias
923 // set tracker to handle regular (simple) load/store dependencies.
924 if (FutureSideEffects &&
925 ((!isSimpleLoadStore(J1) &&
926 !isSafeToSpeculativelyExecute(J1, DL)) ||
927 (!isSimpleLoadStore(J2) &&
928 !isSafeToSpeculativelyExecute(J2, DL)))) {
929 DEBUG(dbgs() << "LRR: iteration root match failed at " << *J1 <<
931 " (side effects prevent reordering)\n");
936 // For instructions that are part of a reduction, if the operation is
937 // associative, then don't bother matching the operands (because we
938 // already know that the instructions are isomorphic, and the order
939 // within the iteration does not matter). For non-associative reductions,
940 // we do need to match the operands, because we need to reject
941 // out-of-order instructions within an iteration!
942 // For example (assume floating-point addition), we need to reject this:
943 // x += a[i]; x += b[i];
944 // x += a[i+1]; x += b[i+1];
945 // x += b[i+2]; x += a[i+2];
946 bool InReduction = Reductions.isPairInSame(J1, J2);
948 if (!(InReduction && J1->isAssociative())) {
949 bool Swapped = false, SomeOpMatched = false;
950 for (unsigned j = 0; j < J1->getNumOperands() && !MatchFailed; ++j) {
951 Value *Op2 = J2->getOperand(j);
953 // If this is part of a reduction (and the operation is not
954 // associatve), then we match all operands, but not those that are
955 // part of the reduction.
957 if (Instruction *Op2I = dyn_cast<Instruction>(Op2))
958 if (Reductions.isPairInSame(J2, Op2I))
961 DenseMap<Value *, Value *>::iterator BMI = BaseMap.find(Op2);
962 if (BMI != BaseMap.end())
964 else if (std::find(Roots[i].begin(), Roots[i].end(),
965 (Instruction*) Op2) != Roots[i].end())
968 if (J1->getOperand(Swapped ? unsigned(!j) : j) != Op2) {
969 // If we've not already decided to swap the matched operands, and
970 // we've not already matched our first operand (note that we could
971 // have skipped matching the first operand because it is part of a
972 // reduction above), and the instruction is commutative, then try
973 // the swapped match.
974 if (!Swapped && J1->isCommutative() && !SomeOpMatched &&
975 J1->getOperand(!j) == Op2) {
978 DEBUG(dbgs() << "LRR: iteration root match failed at " << *J1 <<
979 " vs. " << *J2 << " (operand " << j << ")\n");
985 SomeOpMatched = true;
989 if ((!PossibleRedLastSet.count(J1) && hasUsesOutsideLoop(J1, L)) ||
990 (!PossibleRedLastSet.count(J2) && hasUsesOutsideLoop(J2, L))) {
991 DEBUG(dbgs() << "LRR: iteration root match failed at " << *J1 <<
992 " vs. " << *J2 << " (uses outside loop)\n");
998 BaseMap.insert(std::pair<Value *, Value *>(J2, J1));
1000 AllRootUses.insert(J2);
1001 Reductions.recordPair(J1, J2, i+1);
1010 DEBUG(dbgs() << "LRR: Matched all iteration increments for " <<
1013 DenseSet<Instruction *> LoopIncUseSet;
1014 collectInLoopUserSet(L, LoopIncs, SmallInstructionSet(),
1015 SmallInstructionSet(), LoopIncUseSet);
1016 DEBUG(dbgs() << "LRR: Loop increment set size: " <<
1017 LoopIncUseSet.size() << "\n");
1019 // Make sure that all instructions in the loop have been included in some
1021 for (BasicBlock::iterator J = Header->begin(), JE = Header->end();
1023 if (isa<DbgInfoIntrinsic>(J))
1025 if (cast<Instruction>(J) == RealIV)
1027 if (cast<Instruction>(J) == IV)
1029 if (BaseUseSet.count(J) || AllRootUses.count(J) ||
1030 (LoopIncUseSet.count(J) && (J->isTerminator() ||
1031 isSafeToSpeculativelyExecute(J, DL))))
1034 if (AllRoots.count(J))
1037 if (Reductions.isSelectedPHI(J))
1040 DEBUG(dbgs() << "LRR: aborting reroll based on " << *RealIV <<
1041 " unprocessed instruction found: " << *J << "\n");
1049 DEBUG(dbgs() << "LRR: all instructions processed from " <<
1052 if (!Reductions.validateSelected())
1055 // At this point, we've validated the rerolling, and we're committed to
1058 Reductions.replaceSelected();
1060 // Remove instructions associated with non-base iterations.
1061 for (BasicBlock::reverse_iterator J = Header->rbegin();
1062 J != Header->rend();) {
1063 if (AllRootUses.count(&*J)) {
1064 Instruction *D = &*J;
1065 DEBUG(dbgs() << "LRR: removing: " << *D << "\n");
1066 D->eraseFromParent();
1073 // Insert the new induction variable.
1074 const SCEV *Start = RealIVSCEV->getStart();
1076 Start = SE->getMulExpr(Start,
1077 SE->getConstant(Start->getType(), Scale));
1078 const SCEVAddRecExpr *H =
1079 cast<SCEVAddRecExpr>(SE->getAddRecExpr(Start,
1080 SE->getConstant(RealIVSCEV->getType(), 1),
1081 L, SCEV::FlagAnyWrap));
1082 { // Limit the lifetime of SCEVExpander.
1083 SCEVExpander Expander(*SE, "reroll");
1084 Value *NewIV = Expander.expandCodeFor(H, IV->getType(), Header->begin());
1086 for (DenseSet<Instruction *>::iterator J = BaseUseSet.begin(),
1087 JE = BaseUseSet.end(); J != JE; ++J)
1088 (*J)->replaceUsesOfWith(IV, NewIV);
1090 if (BranchInst *BI = dyn_cast<BranchInst>(Header->getTerminator())) {
1091 if (LoopIncUseSet.count(BI)) {
1092 const SCEV *ICSCEV = RealIVSCEV->evaluateAtIteration(IterCount, *SE);
1095 SE->getMulExpr(ICSCEV, SE->getConstant(ICSCEV->getType(), Scale));
1096 // Iteration count SCEV minus 1
1097 const SCEV *ICMinus1SCEV =
1098 SE->getMinusSCEV(ICSCEV, SE->getConstant(ICSCEV->getType(), 1));
1100 Value *ICMinus1; // Iteration count minus 1
1101 if (isa<SCEVConstant>(ICMinus1SCEV)) {
1102 ICMinus1 = Expander.expandCodeFor(ICMinus1SCEV, NewIV->getType(), BI);
1104 BasicBlock *Preheader = L->getLoopPreheader();
1106 Preheader = InsertPreheaderForLoop(L, this);
1108 ICMinus1 = Expander.expandCodeFor(ICMinus1SCEV, NewIV->getType(),
1109 Preheader->getTerminator());
1112 Value *Cond = new ICmpInst(BI, CmpInst::ICMP_EQ, NewIV, ICMinus1,
1114 BI->setCondition(Cond);
1116 if (BI->getSuccessor(1) != Header)
1117 BI->swapSuccessors();
1122 SimplifyInstructionsInBlock(Header, DL, TLI);
1123 DeleteDeadPHIs(Header, TLI);
1128 bool LoopReroll::runOnLoop(Loop *L, LPPassManager &LPM) {
1129 if (skipOptnoneFunction(L))
1132 AA = &getAnalysis<AliasAnalysis>();
1133 LI = &getAnalysis<LoopInfo>();
1134 SE = &getAnalysis<ScalarEvolution>();
1135 TLI = &getAnalysis<TargetLibraryInfo>();
1136 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1137 DL = DLP ? &DLP->getDataLayout() : nullptr;
1138 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1140 BasicBlock *Header = L->getHeader();
1141 DEBUG(dbgs() << "LRR: F[" << Header->getParent()->getName() <<
1142 "] Loop %" << Header->getName() << " (" <<
1143 L->getNumBlocks() << " block(s))\n");
1145 bool Changed = false;
1147 // For now, we'll handle only single BB loops.
1148 if (L->getNumBlocks() > 1)
1151 if (!SE->hasLoopInvariantBackedgeTakenCount(L))
1154 const SCEV *LIBETC = SE->getBackedgeTakenCount(L);
1155 const SCEV *IterCount =
1156 SE->getAddExpr(LIBETC, SE->getConstant(LIBETC->getType(), 1));
1157 DEBUG(dbgs() << "LRR: iteration count = " << *IterCount << "\n");
1159 // First, we need to find the induction variable with respect to which we can
1160 // reroll (there may be several possible options).
1161 SmallInstructionVector PossibleIVs;
1162 collectPossibleIVs(L, PossibleIVs);
1164 if (PossibleIVs.empty()) {
1165 DEBUG(dbgs() << "LRR: No possible IVs found\n");
1169 ReductionTracker Reductions;
1170 collectPossibleReductions(L, Reductions);
1172 // For each possible IV, collect the associated possible set of 'root' nodes
1173 // (i+1, i+2, etc.).
1174 for (SmallInstructionVector::iterator I = PossibleIVs.begin(),
1175 IE = PossibleIVs.end(); I != IE; ++I)
1176 if (reroll(*I, L, Header, IterCount, Reductions)) {