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 #define DEBUG_TYPE "loop-reroll"
15 #include "llvm/Transforms/Scalar.h"
16 #include "llvm/ADT/STLExtras.h"
17 #include "llvm/ADT/SmallSet.h"
18 #include "llvm/ADT/Statistic.h"
19 #include "llvm/Analysis/AliasAnalysis.h"
20 #include "llvm/Analysis/AliasSetTracker.h"
21 #include "llvm/Analysis/LoopPass.h"
22 #include "llvm/Analysis/ScalarEvolution.h"
23 #include "llvm/Analysis/ScalarEvolutionExpander.h"
24 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
25 #include "llvm/Analysis/ValueTracking.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/Dominators.h"
28 #include "llvm/IR/IntrinsicInst.h"
29 #include "llvm/Support/CommandLine.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/raw_ostream.h"
32 #include "llvm/Target/TargetLibraryInfo.h"
33 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
34 #include "llvm/Transforms/Utils/Local.h"
35 #include "llvm/Transforms/Utils/LoopUtils.h"
39 STATISTIC(NumRerolledLoops, "Number of rerolled loops");
41 static cl::opt<unsigned>
42 MaxInc("max-reroll-increment", cl::init(2048), cl::Hidden,
43 cl::desc("The maximum increment for loop rerolling"));
45 // This loop re-rolling transformation aims to transform loops like this:
49 // for (int i = 0; i < 500; i += 3) {
56 // into a loop like this:
59 // for (int i = 0; i < 500; ++i)
63 // It does this by looking for loops that, besides the latch code, are composed
64 // of isomorphic DAGs of instructions, with each DAG rooted at some increment
65 // to the induction variable, and where each DAG is isomorphic to the DAG
66 // rooted at the induction variable (excepting the sub-DAGs which root the
67 // other induction-variable increments). In other words, we're looking for loop
68 // bodies of the form:
70 // %iv = phi [ (preheader, ...), (body, %iv.next) ]
72 // %iv.1 = add %iv, 1 <-- a root increment
74 // %iv.2 = add %iv, 2 <-- a root increment
76 // %iv.scale_m_1 = add %iv, scale-1 <-- a root increment
79 // %iv.next = add %iv, scale
80 // %cmp = icmp(%iv, ...)
81 // br %cmp, header, exit
83 // where each f(i) is a set of instructions that, collectively, are a function
84 // only of i (and other loop-invariant values).
86 // As a special case, we can also reroll loops like this:
90 // for (int i = 0; i < 500; ++i) {
100 // for (int i = 0; i < 1500; ++i)
104 // in which case, we're looking for inputs like this:
106 // %iv = phi [ (preheader, ...), (body, %iv.next) ]
107 // %scaled.iv = mul %iv, scale
109 // %scaled.iv.1 = add %scaled.iv, 1
111 // %scaled.iv.2 = add %scaled.iv, 2
113 // %scaled.iv.scale_m_1 = add %scaled.iv, scale-1
114 // f(%scaled.iv.scale_m_1)
116 // %iv.next = add %iv, 1
117 // %cmp = icmp(%iv, ...)
118 // br %cmp, header, exit
121 class LoopReroll : public LoopPass {
123 static char ID; // Pass ID, replacement for typeid
124 LoopReroll() : LoopPass(ID) {
125 initializeLoopRerollPass(*PassRegistry::getPassRegistry());
128 bool runOnLoop(Loop *L, LPPassManager &LPM);
130 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
131 AU.addRequired<AliasAnalysis>();
132 AU.addRequired<LoopInfo>();
133 AU.addPreserved<LoopInfo>();
134 AU.addRequired<DominatorTreeWrapperPass>();
135 AU.addPreserved<DominatorTreeWrapperPass>();
136 AU.addRequired<ScalarEvolution>();
137 AU.addRequired<TargetLibraryInfo>();
145 TargetLibraryInfo *TLI;
148 typedef SmallVector<Instruction *, 16> SmallInstructionVector;
149 typedef SmallSet<Instruction *, 16> SmallInstructionSet;
151 // A chain of isomorphic instructions, indentified by a single-use PHI,
152 // representing a reduction. Only the last value may be used outside the
154 struct SimpleLoopReduction {
155 SimpleLoopReduction(Instruction *P, Loop *L)
156 : Valid(false), Instructions(1, P) {
157 assert(isa<PHINode>(P) && "First reduction instruction must be a PHI");
165 Instruction *getPHI() const {
166 assert(Valid && "Using invalid reduction");
167 return Instructions.front();
170 Instruction *getReducedValue() const {
171 assert(Valid && "Using invalid reduction");
172 return Instructions.back();
175 Instruction *get(size_t i) const {
176 assert(Valid && "Using invalid reduction");
177 return Instructions[i+1];
180 Instruction *operator [] (size_t i) const { return get(i); }
182 // The size, ignoring the initial PHI.
183 size_t size() const {
184 assert(Valid && "Using invalid reduction");
185 return Instructions.size()-1;
188 typedef SmallInstructionVector::iterator iterator;
189 typedef SmallInstructionVector::const_iterator const_iterator;
192 assert(Valid && "Using invalid reduction");
193 return llvm::next(Instructions.begin());
196 const_iterator begin() const {
197 assert(Valid && "Using invalid reduction");
198 return llvm::next(Instructions.begin());
201 iterator end() { return Instructions.end(); }
202 const_iterator end() const { return Instructions.end(); }
206 SmallInstructionVector Instructions;
211 // The set of all reductions, and state tracking of possible reductions
212 // during loop instruction processing.
213 struct ReductionTracker {
214 typedef SmallVector<SimpleLoopReduction, 16> SmallReductionVector;
216 // Add a new possible reduction.
217 void addSLR(SimpleLoopReduction &SLR) {
218 PossibleReds.push_back(SLR);
221 // Setup to track possible reductions corresponding to the provided
222 // rerolling scale. Only reductions with a number of non-PHI instructions
223 // that is divisible by the scale are considered. Three instructions sets
225 // - A set of all possible instructions in eligible reductions.
226 // - A set of all PHIs in eligible reductions
227 // - A set of all reduced values (last instructions) in eligible reductions.
228 void restrictToScale(uint64_t Scale,
229 SmallInstructionSet &PossibleRedSet,
230 SmallInstructionSet &PossibleRedPHISet,
231 SmallInstructionSet &PossibleRedLastSet) {
232 PossibleRedIdx.clear();
233 PossibleRedIter.clear();
236 for (unsigned i = 0, e = PossibleReds.size(); i != e; ++i)
237 if (PossibleReds[i].size() % Scale == 0) {
238 PossibleRedLastSet.insert(PossibleReds[i].getReducedValue());
239 PossibleRedPHISet.insert(PossibleReds[i].getPHI());
241 PossibleRedSet.insert(PossibleReds[i].getPHI());
242 PossibleRedIdx[PossibleReds[i].getPHI()] = i;
243 for (SimpleLoopReduction::iterator J = PossibleReds[i].begin(),
244 JE = PossibleReds[i].end(); J != JE; ++J) {
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 (Value::use_iterator UI = I->use_begin(),
358 UIE = I->use_end(); UI != UIE; ++UI) {
359 Instruction *User = cast<Instruction>(*UI);
360 if (!L->contains(User))
367 // Collect the list of loop induction variables with respect to which it might
368 // be possible to reroll the loop.
369 void LoopReroll::collectPossibleIVs(Loop *L,
370 SmallInstructionVector &PossibleIVs) {
371 BasicBlock *Header = L->getHeader();
372 for (BasicBlock::iterator I = Header->begin(),
373 IE = Header->getFirstInsertionPt(); I != IE; ++I) {
374 if (!isa<PHINode>(I))
376 if (!I->getType()->isIntegerTy())
379 if (const SCEVAddRecExpr *PHISCEV =
380 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(I))) {
381 if (PHISCEV->getLoop() != L)
383 if (!PHISCEV->isAffine())
385 if (const SCEVConstant *IncSCEV =
386 dyn_cast<SCEVConstant>(PHISCEV->getStepRecurrence(*SE))) {
387 if (!IncSCEV->getValue()->getValue().isStrictlyPositive())
389 if (IncSCEV->getValue()->uge(MaxInc))
392 DEBUG(dbgs() << "LRR: Possible IV: " << *I << " = " <<
394 PossibleIVs.push_back(I);
400 // Add the remainder of the reduction-variable chain to the instruction vector
401 // (the initial PHINode has already been added). If successful, the object is
403 void LoopReroll::SimpleLoopReduction::add(Loop *L) {
404 assert(!Valid && "Cannot add to an already-valid chain");
406 // The reduction variable must be a chain of single-use instructions
407 // (including the PHI), except for the last value (which is used by the PHI
408 // and also outside the loop).
409 Instruction *C = Instructions.front();
412 C = cast<Instruction>(*C->use_begin());
413 if (C->hasOneUse()) {
414 if (!C->isBinaryOp())
417 if (!(isa<PHINode>(Instructions.back()) ||
418 C->isSameOperationAs(Instructions.back())))
421 Instructions.push_back(C);
423 } while (C->hasOneUse());
425 if (Instructions.size() < 2 ||
426 !C->isSameOperationAs(Instructions.back()) ||
427 C->use_begin() == C->use_end())
430 // C is now the (potential) last instruction in the reduction chain.
431 for (Value::use_iterator UI = C->use_begin(), UIE = C->use_end();
433 // The only in-loop user can be the initial PHI.
434 if (L->contains(cast<Instruction>(*UI)))
435 if (cast<Instruction>(*UI ) != Instructions.front())
439 Instructions.push_back(C);
443 // Collect the vector of possible reduction variables.
444 void LoopReroll::collectPossibleReductions(Loop *L,
445 ReductionTracker &Reductions) {
446 BasicBlock *Header = L->getHeader();
447 for (BasicBlock::iterator I = Header->begin(),
448 IE = Header->getFirstInsertionPt(); I != IE; ++I) {
449 if (!isa<PHINode>(I))
451 if (!I->getType()->isSingleValueType())
454 SimpleLoopReduction SLR(I, L);
458 DEBUG(dbgs() << "LRR: Possible reduction: " << *I << " (with " <<
459 SLR.size() << " chained instructions)\n");
460 Reductions.addSLR(SLR);
464 // Collect the set of all users of the provided root instruction. This set of
465 // users contains not only the direct users of the root instruction, but also
466 // all users of those users, and so on. There are two exceptions:
468 // 1. Instructions in the set of excluded instructions are never added to the
469 // use set (even if they are users). This is used, for example, to exclude
470 // including root increments in the use set of the primary IV.
472 // 2. Instructions in the set of final instructions are added to the use set
473 // if they are users, but their users are not added. This is used, for
474 // example, to prevent a reduction update from forcing all later reduction
475 // updates into the use set.
476 void LoopReroll::collectInLoopUserSet(Loop *L,
477 Instruction *Root, const SmallInstructionSet &Exclude,
478 const SmallInstructionSet &Final,
479 DenseSet<Instruction *> &Users) {
480 SmallInstructionVector Queue(1, Root);
481 while (!Queue.empty()) {
482 Instruction *I = Queue.pop_back_val();
483 if (!Users.insert(I).second)
487 for (Value::use_iterator UI = I->use_begin(),
488 UIE = I->use_end(); UI != UIE; ++UI) {
489 Instruction *User = cast<Instruction>(*UI);
490 if (PHINode *PN = dyn_cast<PHINode>(User)) {
491 // Ignore "wrap-around" uses to PHIs of this loop's header.
492 if (PN->getIncomingBlock(UI) == L->getHeader())
496 if (L->contains(User) && !Exclude.count(User)) {
497 Queue.push_back(User);
501 // We also want to collect single-user "feeder" values.
502 for (User::op_iterator OI = I->op_begin(),
503 OIE = I->op_end(); OI != OIE; ++OI) {
504 if (Instruction *Op = dyn_cast<Instruction>(*OI))
505 if (Op->hasOneUse() && L->contains(Op) && !Exclude.count(Op) &&
512 // Collect all of the users of all of the provided root instructions (combined
513 // into a single set).
514 void LoopReroll::collectInLoopUserSet(Loop *L,
515 const SmallInstructionVector &Roots,
516 const SmallInstructionSet &Exclude,
517 const SmallInstructionSet &Final,
518 DenseSet<Instruction *> &Users) {
519 for (SmallInstructionVector::const_iterator I = Roots.begin(),
520 IE = Roots.end(); I != IE; ++I)
521 collectInLoopUserSet(L, *I, Exclude, Final, Users);
524 static bool isSimpleLoadStore(Instruction *I) {
525 if (LoadInst *LI = dyn_cast<LoadInst>(I))
526 return LI->isSimple();
527 if (StoreInst *SI = dyn_cast<StoreInst>(I))
528 return SI->isSimple();
529 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I))
530 return !MI->isVolatile();
534 // Recognize loops that are setup like this:
536 // %iv = phi [ (preheader, ...), (body, %iv.next) ]
537 // %scaled.iv = mul %iv, scale
539 // %scaled.iv.1 = add %scaled.iv, 1
541 // %scaled.iv.2 = add %scaled.iv, 2
543 // %scaled.iv.scale_m_1 = add %scaled.iv, scale-1
544 // f(%scaled.iv.scale_m_1)
546 // %iv.next = add %iv, 1
547 // %cmp = icmp(%iv, ...)
548 // br %cmp, header, exit
550 // and, if found, set IV = %scaled.iv, and add %iv.next to LoopIncs.
551 bool LoopReroll::findScaleFromMul(Instruction *RealIV, uint64_t &Scale,
553 SmallInstructionVector &LoopIncs) {
554 // This is a special case: here we're looking for all uses (except for
555 // the increment) to be multiplied by a common factor. The increment must
556 // be by one. This is to capture loops like:
557 // for (int i = 0; i < 500; ++i) {
558 // foo(3*i); foo(3*i+1); foo(3*i+2);
560 if (RealIV->getNumUses() != 2)
562 const SCEVAddRecExpr *RealIVSCEV = cast<SCEVAddRecExpr>(SE->getSCEV(RealIV));
563 Instruction *User1 = cast<Instruction>(*RealIV->use_begin()),
564 *User2 = cast<Instruction>(*llvm::next(RealIV->use_begin()));
565 if (!SE->isSCEVable(User1->getType()) || !SE->isSCEVable(User2->getType()))
567 const SCEVAddRecExpr *User1SCEV =
568 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(User1)),
570 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(User2));
571 if (!User1SCEV || !User1SCEV->isAffine() ||
572 !User2SCEV || !User2SCEV->isAffine())
575 // We assume below that User1 is the scale multiply and User2 is the
576 // increment. If this can't be true, then swap them.
577 if (User1SCEV == RealIVSCEV->getPostIncExpr(*SE)) {
578 std::swap(User1, User2);
579 std::swap(User1SCEV, User2SCEV);
582 if (User2SCEV != RealIVSCEV->getPostIncExpr(*SE))
584 assert(User2SCEV->getStepRecurrence(*SE)->isOne() &&
585 "Invalid non-unit step for multiplicative scaling");
586 LoopIncs.push_back(User2);
588 if (const SCEVConstant *MulScale =
589 dyn_cast<SCEVConstant>(User1SCEV->getStepRecurrence(*SE))) {
590 // Make sure that both the start and step have the same multiplier.
591 if (RealIVSCEV->getStart()->getType() != MulScale->getType())
593 if (SE->getMulExpr(RealIVSCEV->getStart(), MulScale) !=
594 User1SCEV->getStart())
597 ConstantInt *MulScaleCI = MulScale->getValue();
598 if (!MulScaleCI->uge(2) || MulScaleCI->uge(MaxInc))
600 Scale = MulScaleCI->getZExtValue();
605 DEBUG(dbgs() << "LRR: Found possible scaling " << *User1 << "\n");
609 // Collect all root increments with respect to the provided induction variable
610 // (normally the PHI, but sometimes a multiply). A root increment is an
611 // instruction, normally an add, with a positive constant less than Scale. In a
612 // rerollable loop, each of these increments is the root of an instruction
613 // graph isomorphic to the others. Also, we collect the final induction
614 // increment (the increment equal to the Scale), and its users in LoopIncs.
615 bool LoopReroll::collectAllRoots(Loop *L, uint64_t Inc, uint64_t Scale,
617 SmallVector<SmallInstructionVector, 32> &Roots,
618 SmallInstructionSet &AllRoots,
619 SmallInstructionVector &LoopIncs) {
620 for (Value::use_iterator UI = IV->use_begin(),
621 UIE = IV->use_end(); UI != UIE; ++UI) {
622 Instruction *User = cast<Instruction>(*UI);
623 if (!SE->isSCEVable(User->getType()))
625 if (User->getType() != IV->getType())
627 if (!L->contains(User))
629 if (hasUsesOutsideLoop(User, L))
632 if (const SCEVConstant *Diff = dyn_cast<SCEVConstant>(SE->getMinusSCEV(
633 SE->getSCEV(User), SE->getSCEV(IV)))) {
634 uint64_t Idx = Diff->getValue()->getValue().getZExtValue();
635 if (Idx > 0 && Idx < Scale) {
636 Roots[Idx-1].push_back(User);
637 AllRoots.insert(User);
638 } else if (Idx == Scale && Inc > 1) {
639 LoopIncs.push_back(User);
644 if (Roots[0].empty())
647 for (unsigned i = 1; i < Scale-1; ++i)
648 if (Roots[i].size() != Roots[0].size()) {
659 // Validate the selected reductions. All iterations must have an isomorphic
660 // part of the reduction chain and, for non-associative reductions, the chain
661 // entries must appear in order.
662 bool LoopReroll::ReductionTracker::validateSelected() {
663 // For a non-associative reduction, the chain entries must appear in order.
664 for (DenseSet<int>::iterator RI = Reds.begin(), RIE = Reds.end();
667 int PrevIter = 0, BaseCount = 0, Count = 0;
668 for (SimpleLoopReduction::iterator J = PossibleReds[i].begin(),
669 JE = PossibleReds[i].end(); J != JE; ++J) {
670 // Note that all instructions in the chain must have been found because
671 // all instructions in the function must have been assigned to some
673 int Iter = PossibleRedIter[*J];
674 if (Iter != PrevIter && Iter != PrevIter + 1 &&
675 !PossibleReds[i].getReducedValue()->isAssociative()) {
676 DEBUG(dbgs() << "LRR: Out-of-order non-associative reduction: " <<
681 if (Iter != PrevIter) {
682 if (Count != BaseCount) {
683 DEBUG(dbgs() << "LRR: Iteration " << PrevIter <<
684 " reduction use count " << Count <<
685 " is not equal to the base use count " <<
704 // For all selected reductions, remove all parts except those in the first
705 // iteration (and the PHI). Replace outside uses of the reduced value with uses
706 // of the first-iteration reduced value (in other words, reroll the selected
708 void LoopReroll::ReductionTracker::replaceSelected() {
709 // Fixup reductions to refer to the last instruction associated with the
710 // first iteration (not the last).
711 for (DenseSet<int>::iterator RI = Reds.begin(), RIE = Reds.end();
715 for (int e = PossibleReds[i].size(); j != e; ++j)
716 if (PossibleRedIter[PossibleReds[i][j]] != 0) {
721 // Replace users with the new end-of-chain value.
722 SmallInstructionVector Users;
723 for (Value::use_iterator UI =
724 PossibleReds[i].getReducedValue()->use_begin(),
725 UIE = PossibleReds[i].getReducedValue()->use_end(); UI != UIE; ++UI)
726 Users.push_back(cast<Instruction>(*UI));
728 for (SmallInstructionVector::iterator J = Users.begin(),
729 JE = Users.end(); J != JE; ++J)
730 (*J)->replaceUsesOfWith(PossibleReds[i].getReducedValue(),
735 // Reroll the provided loop with respect to the provided induction variable.
736 // Generally, we're looking for a loop like this:
738 // %iv = phi [ (preheader, ...), (body, %iv.next) ]
740 // %iv.1 = add %iv, 1 <-- a root increment
742 // %iv.2 = add %iv, 2 <-- a root increment
744 // %iv.scale_m_1 = add %iv, scale-1 <-- a root increment
747 // %iv.next = add %iv, scale
748 // %cmp = icmp(%iv, ...)
749 // br %cmp, header, exit
751 // Notably, we do not require that f(%iv), f(%iv.1), etc. be isolated groups of
752 // instructions. In other words, the instructions in f(%iv), f(%iv.1), etc. can
753 // be intermixed with eachother. The restriction imposed by this algorithm is
754 // that the relative order of the isomorphic instructions in f(%iv), f(%iv.1),
757 // First, we collect the use set of %iv, excluding the other increment roots.
758 // This gives us f(%iv). Then we iterate over the loop instructions (scale-1)
759 // times, having collected the use set of f(%iv.(i+1)), during which we:
760 // - Ensure that the next unmatched instruction in f(%iv) is isomorphic to
761 // the next unmatched instruction in f(%iv.(i+1)).
762 // - Ensure that both matched instructions don't have any external users
763 // (with the exception of last-in-chain reduction instructions).
764 // - Track the (aliasing) write set, and other side effects, of all
765 // instructions that belong to future iterations that come before the matched
766 // instructions. If the matched instructions read from that write set, then
767 // f(%iv) or f(%iv.(i+1)) has some dependency on instructions in
768 // f(%iv.(j+1)) for some j > i, and we cannot reroll the loop. Similarly,
769 // if any of these future instructions had side effects (could not be
770 // speculatively executed), and so do the matched instructions, when we
771 // cannot reorder those side-effect-producing instructions, and rerolling
774 // Finally, we make sure that all loop instructions are either loop increment
775 // roots, belong to simple latch code, parts of validated reductions, part of
776 // f(%iv) or part of some f(%iv.i). If all of that is true (and all reductions
777 // have been validated), then we reroll the loop.
778 bool LoopReroll::reroll(Instruction *IV, Loop *L, BasicBlock *Header,
779 const SCEV *IterCount,
780 ReductionTracker &Reductions) {
781 const SCEVAddRecExpr *RealIVSCEV = cast<SCEVAddRecExpr>(SE->getSCEV(IV));
782 uint64_t Inc = cast<SCEVConstant>(RealIVSCEV->getOperand(1))->
783 getValue()->getZExtValue();
784 // The collection of loop increment instructions.
785 SmallInstructionVector LoopIncs;
786 uint64_t Scale = Inc;
788 // The effective induction variable, IV, is normally also the real induction
789 // variable. When we're dealing with a loop like:
790 // for (int i = 0; i < 500; ++i)
794 // then the real IV is still i, but the effective IV is (3*i).
795 Instruction *RealIV = IV;
796 if (Inc == 1 && !findScaleFromMul(RealIV, Scale, IV, LoopIncs))
799 assert(Scale <= MaxInc && "Scale is too large");
800 assert(Scale > 1 && "Scale must be at least 2");
802 // The set of increment instructions for each increment value.
803 SmallVector<SmallInstructionVector, 32> Roots(Scale-1);
804 SmallInstructionSet AllRoots;
805 if (!collectAllRoots(L, Inc, Scale, IV, Roots, AllRoots, LoopIncs))
808 DEBUG(dbgs() << "LRR: Found all root induction increments for: " <<
811 // An array of just the possible reductions for this scale factor. When we
812 // collect the set of all users of some root instructions, these reduction
813 // instructions are treated as 'final' (their uses are not considered).
814 // This is important because we don't want the root use set to search down
815 // the reduction chain.
816 SmallInstructionSet PossibleRedSet;
817 SmallInstructionSet PossibleRedLastSet, PossibleRedPHISet;
818 Reductions.restrictToScale(Scale, PossibleRedSet, PossibleRedPHISet,
821 // We now need to check for equivalence of the use graph of each root with
822 // that of the primary induction variable (excluding the roots). Our goal
823 // here is not to solve the full graph isomorphism problem, but rather to
824 // catch common cases without a lot of work. As a result, we will assume
825 // that the relative order of the instructions in each unrolled iteration
826 // is the same (although we will not make an assumption about how the
827 // different iterations are intermixed). Note that while the order must be
828 // the same, the instructions may not be in the same basic block.
829 SmallInstructionSet Exclude(AllRoots);
830 Exclude.insert(LoopIncs.begin(), LoopIncs.end());
832 DenseSet<Instruction *> BaseUseSet;
833 collectInLoopUserSet(L, IV, Exclude, PossibleRedSet, BaseUseSet);
835 DenseSet<Instruction *> AllRootUses;
836 std::vector<DenseSet<Instruction *> > RootUseSets(Scale-1);
838 bool MatchFailed = false;
839 for (unsigned i = 0; i < Scale-1 && !MatchFailed; ++i) {
840 DenseSet<Instruction *> &RootUseSet = RootUseSets[i];
841 collectInLoopUserSet(L, Roots[i], SmallInstructionSet(),
842 PossibleRedSet, RootUseSet);
844 DEBUG(dbgs() << "LRR: base use set size: " << BaseUseSet.size() <<
845 " vs. iteration increment " << (i+1) <<
846 " use set size: " << RootUseSet.size() << "\n");
848 if (BaseUseSet.size() != RootUseSet.size()) {
853 // In addition to regular aliasing information, we need to look for
854 // instructions from later (future) iterations that have side effects
855 // preventing us from reordering them past other instructions with side
857 bool FutureSideEffects = false;
858 AliasSetTracker AST(*AA);
860 // The map between instructions in f(%iv.(i+1)) and f(%iv).
861 DenseMap<Value *, Value *> BaseMap;
863 assert(L->getNumBlocks() == 1 && "Cannot handle multi-block loops");
864 for (BasicBlock::iterator J1 = Header->begin(), J2 = Header->begin(),
865 JE = Header->end(); J1 != JE && !MatchFailed; ++J1) {
866 if (cast<Instruction>(J1) == RealIV)
868 if (cast<Instruction>(J1) == IV)
870 if (!BaseUseSet.count(J1))
872 if (PossibleRedPHISet.count(J1)) // Skip reduction PHIs.
875 while (J2 != JE && (!RootUseSet.count(J2) ||
876 std::find(Roots[i].begin(), Roots[i].end(), J2) !=
878 // As we iterate through the instructions, instructions that don't
879 // belong to previous iterations (or the base case), must belong to
880 // future iterations. We want to track the alias set of writes from
881 // previous iterations.
882 if (!isa<PHINode>(J2) && !BaseUseSet.count(J2) &&
883 !AllRootUses.count(J2)) {
884 if (J2->mayWriteToMemory())
887 // Note: This is specifically guarded by a check on isa<PHINode>,
888 // which while a valid (somewhat arbitrary) micro-optimization, is
889 // needed because otherwise isSafeToSpeculativelyExecute returns
890 // false on PHI nodes.
891 if (!isSimpleLoadStore(J2) && !isSafeToSpeculativelyExecute(J2, DL))
892 FutureSideEffects = true;
898 if (!J1->isSameOperationAs(J2)) {
899 DEBUG(dbgs() << "LRR: iteration root match failed at " << *J1 <<
900 " vs. " << *J2 << "\n");
905 // Make sure that this instruction, which is in the use set of this
906 // root instruction, does not also belong to the base set or the set of
907 // some previous root instruction.
908 if (BaseUseSet.count(J2) || AllRootUses.count(J2)) {
909 DEBUG(dbgs() << "LRR: iteration root match failed at " << *J1 <<
910 " vs. " << *J2 << " (prev. case overlap)\n");
915 // Make sure that we don't alias with any instruction in the alias set
916 // tracker. If we do, then we depend on a future iteration, and we
918 if (J2->mayReadFromMemory()) {
919 for (AliasSetTracker::iterator K = AST.begin(), KE = AST.end();
920 K != KE && !MatchFailed; ++K) {
921 if (K->aliasesUnknownInst(J2, *AA)) {
922 DEBUG(dbgs() << "LRR: iteration root match failed at " << *J1 <<
923 " vs. " << *J2 << " (depends on future store)\n");
930 // If we've past an instruction from a future iteration that may have
931 // side effects, and this instruction might also, then we can't reorder
932 // them, and this matching fails. As an exception, we allow the alias
933 // set tracker to handle regular (simple) load/store dependencies.
934 if (FutureSideEffects &&
935 ((!isSimpleLoadStore(J1) && !isSafeToSpeculativelyExecute(J1)) ||
936 (!isSimpleLoadStore(J2) && !isSafeToSpeculativelyExecute(J2)))) {
937 DEBUG(dbgs() << "LRR: iteration root match failed at " << *J1 <<
939 " (side effects prevent reordering)\n");
944 // For instructions that are part of a reduction, if the operation is
945 // associative, then don't bother matching the operands (because we
946 // already know that the instructions are isomorphic, and the order
947 // within the iteration does not matter). For non-associative reductions,
948 // we do need to match the operands, because we need to reject
949 // out-of-order instructions within an iteration!
950 // For example (assume floating-point addition), we need to reject this:
951 // x += a[i]; x += b[i];
952 // x += a[i+1]; x += b[i+1];
953 // x += b[i+2]; x += a[i+2];
954 bool InReduction = Reductions.isPairInSame(J1, J2);
956 if (!(InReduction && J1->isAssociative())) {
957 bool Swapped = false, SomeOpMatched = false;;
958 for (unsigned j = 0; j < J1->getNumOperands() && !MatchFailed; ++j) {
959 Value *Op2 = J2->getOperand(j);
961 // If this is part of a reduction (and the operation is not
962 // associatve), then we match all operands, but not those that are
963 // part of the reduction.
965 if (Instruction *Op2I = dyn_cast<Instruction>(Op2))
966 if (Reductions.isPairInSame(J2, Op2I))
969 DenseMap<Value *, Value *>::iterator BMI = BaseMap.find(Op2);
970 if (BMI != BaseMap.end())
972 else if (std::find(Roots[i].begin(), Roots[i].end(),
973 (Instruction*) Op2) != Roots[i].end())
976 if (J1->getOperand(Swapped ? unsigned(!j) : j) != Op2) {
977 // If we've not already decided to swap the matched operands, and
978 // we've not already matched our first operand (note that we could
979 // have skipped matching the first operand because it is part of a
980 // reduction above), and the instruction is commutative, then try
981 // the swapped match.
982 if (!Swapped && J1->isCommutative() && !SomeOpMatched &&
983 J1->getOperand(!j) == Op2) {
986 DEBUG(dbgs() << "LRR: iteration root match failed at " << *J1 <<
987 " vs. " << *J2 << " (operand " << j << ")\n");
993 SomeOpMatched = true;
997 if ((!PossibleRedLastSet.count(J1) && hasUsesOutsideLoop(J1, L)) ||
998 (!PossibleRedLastSet.count(J2) && hasUsesOutsideLoop(J2, L))) {
999 DEBUG(dbgs() << "LRR: iteration root match failed at " << *J1 <<
1000 " vs. " << *J2 << " (uses outside loop)\n");
1006 BaseMap.insert(std::pair<Value *, Value *>(J2, J1));
1008 AllRootUses.insert(J2);
1009 Reductions.recordPair(J1, J2, i+1);
1018 DEBUG(dbgs() << "LRR: Matched all iteration increments for " <<
1021 DenseSet<Instruction *> LoopIncUseSet;
1022 collectInLoopUserSet(L, LoopIncs, SmallInstructionSet(),
1023 SmallInstructionSet(), LoopIncUseSet);
1024 DEBUG(dbgs() << "LRR: Loop increment set size: " <<
1025 LoopIncUseSet.size() << "\n");
1027 // Make sure that all instructions in the loop have been included in some
1029 for (BasicBlock::iterator J = Header->begin(), JE = Header->end();
1031 if (isa<DbgInfoIntrinsic>(J))
1033 if (cast<Instruction>(J) == RealIV)
1035 if (cast<Instruction>(J) == IV)
1037 if (BaseUseSet.count(J) || AllRootUses.count(J) ||
1038 (LoopIncUseSet.count(J) && (J->isTerminator() ||
1039 isSafeToSpeculativelyExecute(J, DL))))
1042 if (AllRoots.count(J))
1045 if (Reductions.isSelectedPHI(J))
1048 DEBUG(dbgs() << "LRR: aborting reroll based on " << *RealIV <<
1049 " unprocessed instruction found: " << *J << "\n");
1057 DEBUG(dbgs() << "LRR: all instructions processed from " <<
1060 if (!Reductions.validateSelected())
1063 // At this point, we've validated the rerolling, and we're committed to
1066 Reductions.replaceSelected();
1068 // Remove instructions associated with non-base iterations.
1069 for (BasicBlock::reverse_iterator J = Header->rbegin();
1070 J != Header->rend();) {
1071 if (AllRootUses.count(&*J)) {
1072 Instruction *D = &*J;
1073 DEBUG(dbgs() << "LRR: removing: " << *D << "\n");
1074 D->eraseFromParent();
1081 // Insert the new induction variable.
1082 const SCEV *Start = RealIVSCEV->getStart();
1084 Start = SE->getMulExpr(Start,
1085 SE->getConstant(Start->getType(), Scale));
1086 const SCEVAddRecExpr *H =
1087 cast<SCEVAddRecExpr>(SE->getAddRecExpr(Start,
1088 SE->getConstant(RealIVSCEV->getType(), 1),
1089 L, SCEV::FlagAnyWrap));
1090 { // Limit the lifetime of SCEVExpander.
1091 SCEVExpander Expander(*SE, "reroll");
1092 Value *NewIV = Expander.expandCodeFor(H, IV->getType(), Header->begin());
1094 for (DenseSet<Instruction *>::iterator J = BaseUseSet.begin(),
1095 JE = BaseUseSet.end(); J != JE; ++J)
1096 (*J)->replaceUsesOfWith(IV, NewIV);
1098 if (BranchInst *BI = dyn_cast<BranchInst>(Header->getTerminator())) {
1099 if (LoopIncUseSet.count(BI)) {
1100 const SCEV *ICSCEV = RealIVSCEV->evaluateAtIteration(IterCount, *SE);
1103 SE->getMulExpr(ICSCEV, SE->getConstant(ICSCEV->getType(), Scale));
1104 // Iteration count SCEV minus 1
1105 const SCEV *ICMinus1SCEV =
1106 SE->getMinusSCEV(ICSCEV, SE->getConstant(ICSCEV->getType(), 1));
1108 Value *ICMinus1; // Iteration count minus 1
1109 if (isa<SCEVConstant>(ICMinus1SCEV)) {
1110 ICMinus1 = Expander.expandCodeFor(ICMinus1SCEV, NewIV->getType(), BI);
1112 BasicBlock *Preheader = L->getLoopPreheader();
1114 Preheader = InsertPreheaderForLoop(L, this);
1116 ICMinus1 = Expander.expandCodeFor(ICMinus1SCEV, NewIV->getType(),
1117 Preheader->getTerminator());
1120 Value *Cond = new ICmpInst(BI, CmpInst::ICMP_EQ, NewIV, ICMinus1,
1122 BI->setCondition(Cond);
1124 if (BI->getSuccessor(1) != Header)
1125 BI->swapSuccessors();
1130 SimplifyInstructionsInBlock(Header, DL, TLI);
1131 DeleteDeadPHIs(Header, TLI);
1136 bool LoopReroll::runOnLoop(Loop *L, LPPassManager &LPM) {
1137 AA = &getAnalysis<AliasAnalysis>();
1138 LI = &getAnalysis<LoopInfo>();
1139 SE = &getAnalysis<ScalarEvolution>();
1140 TLI = &getAnalysis<TargetLibraryInfo>();
1141 DL = getAnalysisIfAvailable<DataLayout>();
1142 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1144 BasicBlock *Header = L->getHeader();
1145 DEBUG(dbgs() << "LRR: F[" << Header->getParent()->getName() <<
1146 "] Loop %" << Header->getName() << " (" <<
1147 L->getNumBlocks() << " block(s))\n");
1149 bool Changed = false;
1151 // For now, we'll handle only single BB loops.
1152 if (L->getNumBlocks() > 1)
1155 if (!SE->hasLoopInvariantBackedgeTakenCount(L))
1158 const SCEV *LIBETC = SE->getBackedgeTakenCount(L);
1159 const SCEV *IterCount =
1160 SE->getAddExpr(LIBETC, SE->getConstant(LIBETC->getType(), 1));
1161 DEBUG(dbgs() << "LRR: iteration count = " << *IterCount << "\n");
1163 // First, we need to find the induction variable with respect to which we can
1164 // reroll (there may be several possible options).
1165 SmallInstructionVector PossibleIVs;
1166 collectPossibleIVs(L, PossibleIVs);
1168 if (PossibleIVs.empty()) {
1169 DEBUG(dbgs() << "LRR: No possible IVs found\n");
1173 ReductionTracker Reductions;
1174 collectPossibleReductions(L, Reductions);
1176 // For each possible IV, collect the associated possible set of 'root' nodes
1177 // (i+1, i+2, etc.).
1178 for (SmallInstructionVector::iterator I = PossibleIVs.begin(),
1179 IE = PossibleIVs.end(); I != IE; ++I)
1180 if (reroll(*I, L, Header, IterCount, Reductions)) {