1 //===- LoopReroll.cpp - Loop rerolling pass -------------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This pass implements a simple loop reroller.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "llvm/ADT/APInt.h"
14 #include "llvm/ADT/BitVector.h"
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/DenseSet.h"
17 #include "llvm/ADT/MapVector.h"
18 #include "llvm/ADT/STLExtras.h"
19 #include "llvm/ADT/SmallPtrSet.h"
20 #include "llvm/ADT/SmallVector.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/Analysis/AliasAnalysis.h"
23 #include "llvm/Analysis/AliasSetTracker.h"
24 #include "llvm/Analysis/LoopInfo.h"
25 #include "llvm/Analysis/LoopPass.h"
26 #include "llvm/Analysis/ScalarEvolution.h"
27 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
28 #include "llvm/Analysis/TargetLibraryInfo.h"
29 #include "llvm/Analysis/ValueTracking.h"
30 #include "llvm/IR/BasicBlock.h"
31 #include "llvm/IR/Constants.h"
32 #include "llvm/IR/Dominators.h"
33 #include "llvm/IR/InstrTypes.h"
34 #include "llvm/IR/Instruction.h"
35 #include "llvm/IR/Instructions.h"
36 #include "llvm/IR/IntrinsicInst.h"
37 #include "llvm/IR/Module.h"
38 #include "llvm/IR/Type.h"
39 #include "llvm/IR/Use.h"
40 #include "llvm/IR/User.h"
41 #include "llvm/IR/Value.h"
42 #include "llvm/InitializePasses.h"
43 #include "llvm/Pass.h"
44 #include "llvm/Support/Casting.h"
45 #include "llvm/Support/CommandLine.h"
46 #include "llvm/Support/Debug.h"
47 #include "llvm/Support/raw_ostream.h"
48 #include "llvm/Transforms/Scalar.h"
49 #include "llvm/Transforms/Scalar/LoopReroll.h"
50 #include "llvm/Transforms/Utils.h"
51 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
52 #include "llvm/Transforms/Utils/Local.h"
53 #include "llvm/Transforms/Utils/LoopUtils.h"
54 #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
55 #include <cassert>
56 #include <cstddef>
57 #include <cstdint>
58 #include <iterator>
59 #include <map>
60 #include <utility>
61 
62 using namespace llvm;
63 
64 #define DEBUG_TYPE "loop-reroll"
65 
66 STATISTIC(NumRerolledLoops, "Number of rerolled loops");
67 
68 static cl::opt<unsigned>
69 NumToleratedFailedMatches("reroll-num-tolerated-failed-matches", cl::init(400),
70                           cl::Hidden,
71                           cl::desc("The maximum number of failures to tolerate"
72                                    " during fuzzy matching. (default: 400)"));
73 
74 // This loop re-rolling transformation aims to transform loops like this:
75 //
76 // int foo(int a);
77 // void bar(int *x) {
78 //   for (int i = 0; i < 500; i += 3) {
79 //     foo(i);
80 //     foo(i+1);
81 //     foo(i+2);
82 //   }
83 // }
84 //
85 // into a loop like this:
86 //
87 // void bar(int *x) {
88 //   for (int i = 0; i < 500; ++i)
89 //     foo(i);
90 // }
91 //
92 // It does this by looking for loops that, besides the latch code, are composed
93 // of isomorphic DAGs of instructions, with each DAG rooted at some increment
94 // to the induction variable, and where each DAG is isomorphic to the DAG
95 // rooted at the induction variable (excepting the sub-DAGs which root the
96 // other induction-variable increments). In other words, we're looking for loop
97 // bodies of the form:
98 //
99 // %iv = phi [ (preheader, ...), (body, %iv.next) ]
100 // f(%iv)
101 // %iv.1 = add %iv, 1                <-- a root increment
102 // f(%iv.1)
103 // %iv.2 = add %iv, 2                <-- a root increment
104 // f(%iv.2)
105 // %iv.scale_m_1 = add %iv, scale-1  <-- a root increment
106 // f(%iv.scale_m_1)
107 // ...
108 // %iv.next = add %iv, scale
109 // %cmp = icmp(%iv, ...)
110 // br %cmp, header, exit
111 //
112 // where each f(i) is a set of instructions that, collectively, are a function
113 // only of i (and other loop-invariant values).
114 //
115 // As a special case, we can also reroll loops like this:
116 //
117 // int foo(int);
118 // void bar(int *x) {
119 //   for (int i = 0; i < 500; ++i) {
120 //     x[3*i] = foo(0);
121 //     x[3*i+1] = foo(0);
122 //     x[3*i+2] = foo(0);
123 //   }
124 // }
125 //
126 // into this:
127 //
128 // void bar(int *x) {
129 //   for (int i = 0; i < 1500; ++i)
130 //     x[i] = foo(0);
131 // }
132 //
133 // in which case, we're looking for inputs like this:
134 //
135 // %iv = phi [ (preheader, ...), (body, %iv.next) ]
136 // %scaled.iv = mul %iv, scale
137 // f(%scaled.iv)
138 // %scaled.iv.1 = add %scaled.iv, 1
139 // f(%scaled.iv.1)
140 // %scaled.iv.2 = add %scaled.iv, 2
141 // f(%scaled.iv.2)
142 // %scaled.iv.scale_m_1 = add %scaled.iv, scale-1
143 // f(%scaled.iv.scale_m_1)
144 // ...
145 // %iv.next = add %iv, 1
146 // %cmp = icmp(%iv, ...)
147 // br %cmp, header, exit
148 
149 namespace {
150 
151   enum IterationLimits {
152     /// The maximum number of iterations that we'll try and reroll.
153     IL_MaxRerollIterations = 32,
154     /// The bitvector index used by loop induction variables and other
155     /// instructions that belong to all iterations.
156     IL_All,
157     IL_End
158   };
159 
160   class LoopRerollLegacyPass : public LoopPass {
161   public:
162     static char ID; // Pass ID, replacement for typeid
163 
164     LoopRerollLegacyPass() : LoopPass(ID) {
165       initializeLoopRerollLegacyPassPass(*PassRegistry::getPassRegistry());
166     }
167 
168     bool runOnLoop(Loop *L, LPPassManager &LPM) override;
169 
170     void getAnalysisUsage(AnalysisUsage &AU) const override {
171       AU.addRequired<TargetLibraryInfoWrapperPass>();
172       getLoopAnalysisUsage(AU);
173     }
174   };
175 
176   class LoopReroll {
177   public:
178     LoopReroll(AliasAnalysis *AA, LoopInfo *LI, ScalarEvolution *SE,
179                TargetLibraryInfo *TLI, DominatorTree *DT, bool PreserveLCSSA)
180         : AA(AA), LI(LI), SE(SE), TLI(TLI), DT(DT),
181           PreserveLCSSA(PreserveLCSSA) {}
182     bool runOnLoop(Loop *L);
183 
184   protected:
185     AliasAnalysis *AA;
186     LoopInfo *LI;
187     ScalarEvolution *SE;
188     TargetLibraryInfo *TLI;
189     DominatorTree *DT;
190     bool PreserveLCSSA;
191 
192     using SmallInstructionVector = SmallVector<Instruction *, 16>;
193     using SmallInstructionSet = SmallPtrSet<Instruction *, 16>;
194 
195     // Map between induction variable and its increment
196     DenseMap<Instruction *, int64_t> IVToIncMap;
197 
198     // For loop with multiple induction variable, remember the one used only to
199     // control the loop.
200     Instruction *LoopControlIV;
201 
202     // A chain of isomorphic instructions, identified by a single-use PHI
203     // representing a reduction. Only the last value may be used outside the
204     // loop.
205     struct SimpleLoopReduction {
206       SimpleLoopReduction(Instruction *P, Loop *L) : Instructions(1, P) {
207         assert(isa<PHINode>(P) && "First reduction instruction must be a PHI");
208         add(L);
209       }
210 
211       bool valid() const {
212         return Valid;
213       }
214 
215       Instruction *getPHI() const {
216         assert(Valid && "Using invalid reduction");
217         return Instructions.front();
218       }
219 
220       Instruction *getReducedValue() const {
221         assert(Valid && "Using invalid reduction");
222         return Instructions.back();
223       }
224 
225       Instruction *get(size_t i) const {
226         assert(Valid && "Using invalid reduction");
227         return Instructions[i+1];
228       }
229 
230       Instruction *operator [] (size_t i) const { return get(i); }
231 
232       // The size, ignoring the initial PHI.
233       size_t size() const {
234         assert(Valid && "Using invalid reduction");
235         return Instructions.size()-1;
236       }
237 
238       using iterator = SmallInstructionVector::iterator;
239       using const_iterator = SmallInstructionVector::const_iterator;
240 
241       iterator begin() {
242         assert(Valid && "Using invalid reduction");
243         return std::next(Instructions.begin());
244       }
245 
246       const_iterator begin() const {
247         assert(Valid && "Using invalid reduction");
248         return std::next(Instructions.begin());
249       }
250 
251       iterator end() { return Instructions.end(); }
252       const_iterator end() const { return Instructions.end(); }
253 
254     protected:
255       bool Valid = false;
256       SmallInstructionVector Instructions;
257 
258       void add(Loop *L);
259     };
260 
261     // The set of all reductions, and state tracking of possible reductions
262     // during loop instruction processing.
263     struct ReductionTracker {
264       using SmallReductionVector = SmallVector<SimpleLoopReduction, 16>;
265 
266       // Add a new possible reduction.
267       void addSLR(SimpleLoopReduction &SLR) { PossibleReds.push_back(SLR); }
268 
269       // Setup to track possible reductions corresponding to the provided
270       // rerolling scale. Only reductions with a number of non-PHI instructions
271       // that is divisible by the scale are considered. Three instructions sets
272       // are filled in:
273       //   - A set of all possible instructions in eligible reductions.
274       //   - A set of all PHIs in eligible reductions
275       //   - A set of all reduced values (last instructions) in eligible
276       //     reductions.
277       void restrictToScale(uint64_t Scale,
278                            SmallInstructionSet &PossibleRedSet,
279                            SmallInstructionSet &PossibleRedPHISet,
280                            SmallInstructionSet &PossibleRedLastSet) {
281         PossibleRedIdx.clear();
282         PossibleRedIter.clear();
283         Reds.clear();
284 
285         for (unsigned i = 0, e = PossibleReds.size(); i != e; ++i)
286           if (PossibleReds[i].size() % Scale == 0) {
287             PossibleRedLastSet.insert(PossibleReds[i].getReducedValue());
288             PossibleRedPHISet.insert(PossibleReds[i].getPHI());
289 
290             PossibleRedSet.insert(PossibleReds[i].getPHI());
291             PossibleRedIdx[PossibleReds[i].getPHI()] = i;
292             for (Instruction *J : PossibleReds[i]) {
293               PossibleRedSet.insert(J);
294               PossibleRedIdx[J] = i;
295             }
296           }
297       }
298 
299       // The functions below are used while processing the loop instructions.
300 
301       // Are the two instructions both from reductions, and furthermore, from
302       // the same reduction?
303       bool isPairInSame(Instruction *J1, Instruction *J2) {
304         DenseMap<Instruction *, int>::iterator J1I = PossibleRedIdx.find(J1);
305         if (J1I != PossibleRedIdx.end()) {
306           DenseMap<Instruction *, int>::iterator J2I = PossibleRedIdx.find(J2);
307           if (J2I != PossibleRedIdx.end() && J1I->second == J2I->second)
308             return true;
309         }
310 
311         return false;
312       }
313 
314       // The two provided instructions, the first from the base iteration, and
315       // the second from iteration i, form a matched pair. If these are part of
316       // a reduction, record that fact.
317       void recordPair(Instruction *J1, Instruction *J2, unsigned i) {
318         if (PossibleRedIdx.count(J1)) {
319           assert(PossibleRedIdx.count(J2) &&
320                  "Recording reduction vs. non-reduction instruction?");
321 
322           PossibleRedIter[J1] = 0;
323           PossibleRedIter[J2] = i;
324 
325           int Idx = PossibleRedIdx[J1];
326           assert(Idx == PossibleRedIdx[J2] &&
327                  "Recording pair from different reductions?");
328           Reds.insert(Idx);
329         }
330       }
331 
332       // The functions below can be called after we've finished processing all
333       // instructions in the loop, and we know which reductions were selected.
334 
335       bool validateSelected();
336       void replaceSelected();
337 
338     protected:
339       // The vector of all possible reductions (for any scale).
340       SmallReductionVector PossibleReds;
341 
342       DenseMap<Instruction *, int> PossibleRedIdx;
343       DenseMap<Instruction *, int> PossibleRedIter;
344       DenseSet<int> Reds;
345     };
346 
347     // A DAGRootSet models an induction variable being used in a rerollable
348     // loop. For example,
349     //
350     //   x[i*3+0] = y1
351     //   x[i*3+1] = y2
352     //   x[i*3+2] = y3
353     //
354     //   Base instruction -> i*3
355     //                    +---+----+
356     //                   /    |     \
357     //               ST[y1]  +1     +2  <-- Roots
358     //                        |      |
359     //                      ST[y2] ST[y3]
360     //
361     // There may be multiple DAGRoots, for example:
362     //
363     //   x[i*2+0] = ...   (1)
364     //   x[i*2+1] = ...   (1)
365     //   x[i*2+4] = ...   (2)
366     //   x[i*2+5] = ...   (2)
367     //   x[(i+1234)*2+5678] = ... (3)
368     //   x[(i+1234)*2+5679] = ... (3)
369     //
370     // The loop will be rerolled by adding a new loop induction variable,
371     // one for the Base instruction in each DAGRootSet.
372     //
373     struct DAGRootSet {
374       Instruction *BaseInst;
375       SmallInstructionVector Roots;
376 
377       // The instructions between IV and BaseInst (but not including BaseInst).
378       SmallInstructionSet SubsumedInsts;
379     };
380 
381     // The set of all DAG roots, and state tracking of all roots
382     // for a particular induction variable.
383     struct DAGRootTracker {
384       DAGRootTracker(LoopReroll *Parent, Loop *L, Instruction *IV,
385                      ScalarEvolution *SE, AliasAnalysis *AA,
386                      TargetLibraryInfo *TLI, DominatorTree *DT, LoopInfo *LI,
387                      bool PreserveLCSSA,
388                      DenseMap<Instruction *, int64_t> &IncrMap,
389                      Instruction *LoopCtrlIV)
390           : Parent(Parent), L(L), SE(SE), AA(AA), TLI(TLI), DT(DT), LI(LI),
391             PreserveLCSSA(PreserveLCSSA), IV(IV), IVToIncMap(IncrMap),
392             LoopControlIV(LoopCtrlIV) {}
393 
394       /// Stage 1: Find all the DAG roots for the induction variable.
395       bool findRoots();
396 
397       /// Stage 2: Validate if the found roots are valid.
398       bool validate(ReductionTracker &Reductions);
399 
400       /// Stage 3: Assuming validate() returned true, perform the
401       /// replacement.
402       /// @param BackedgeTakenCount The backedge-taken count of L.
403       void replace(const SCEV *BackedgeTakenCount);
404 
405     protected:
406       using UsesTy = MapVector<Instruction *, BitVector>;
407 
408       void findRootsRecursive(Instruction *IVU,
409                               SmallInstructionSet SubsumedInsts);
410       bool findRootsBase(Instruction *IVU, SmallInstructionSet SubsumedInsts);
411       bool collectPossibleRoots(Instruction *Base,
412                                 std::map<int64_t,Instruction*> &Roots);
413       bool validateRootSet(DAGRootSet &DRS);
414 
415       bool collectUsedInstructions(SmallInstructionSet &PossibleRedSet);
416       void collectInLoopUserSet(const SmallInstructionVector &Roots,
417                                 const SmallInstructionSet &Exclude,
418                                 const SmallInstructionSet &Final,
419                                 DenseSet<Instruction *> &Users);
420       void collectInLoopUserSet(Instruction *Root,
421                                 const SmallInstructionSet &Exclude,
422                                 const SmallInstructionSet &Final,
423                                 DenseSet<Instruction *> &Users);
424 
425       UsesTy::iterator nextInstr(int Val, UsesTy &In,
426                                  const SmallInstructionSet &Exclude,
427                                  UsesTy::iterator *StartI=nullptr);
428       bool isBaseInst(Instruction *I);
429       bool isRootInst(Instruction *I);
430       bool instrDependsOn(Instruction *I,
431                           UsesTy::iterator Start,
432                           UsesTy::iterator End);
433       void replaceIV(DAGRootSet &DRS, const SCEV *Start, const SCEV *IncrExpr);
434 
435       LoopReroll *Parent;
436 
437       // Members of Parent, replicated here for brevity.
438       Loop *L;
439       ScalarEvolution *SE;
440       AliasAnalysis *AA;
441       TargetLibraryInfo *TLI;
442       DominatorTree *DT;
443       LoopInfo *LI;
444       bool PreserveLCSSA;
445 
446       // The loop induction variable.
447       Instruction *IV;
448 
449       // Loop step amount.
450       int64_t Inc;
451 
452       // Loop reroll count; if Inc == 1, this records the scaling applied
453       // to the indvar: a[i*2+0] = ...; a[i*2+1] = ... ;
454       // If Inc is not 1, Scale = Inc.
455       uint64_t Scale;
456 
457       // The roots themselves.
458       SmallVector<DAGRootSet,16> RootSets;
459 
460       // All increment instructions for IV.
461       SmallInstructionVector LoopIncs;
462 
463       // Map of all instructions in the loop (in order) to the iterations
464       // they are used in (or specially, IL_All for instructions
465       // used in the loop increment mechanism).
466       UsesTy Uses;
467 
468       // Map between induction variable and its increment
469       DenseMap<Instruction *, int64_t> &IVToIncMap;
470 
471       Instruction *LoopControlIV;
472     };
473 
474     // Check if it is a compare-like instruction whose user is a branch
475     bool isCompareUsedByBranch(Instruction *I) {
476       auto *TI = I->getParent()->getTerminator();
477       if (!isa<BranchInst>(TI) || !isa<CmpInst>(I))
478         return false;
479       return I->hasOneUse() && TI->getOperand(0) == I;
480     };
481 
482     bool isLoopControlIV(Loop *L, Instruction *IV);
483     void collectPossibleIVs(Loop *L, SmallInstructionVector &PossibleIVs);
484     void collectPossibleReductions(Loop *L,
485            ReductionTracker &Reductions);
486     bool reroll(Instruction *IV, Loop *L, BasicBlock *Header,
487                 const SCEV *BackedgeTakenCount, ReductionTracker &Reductions);
488   };
489 
490 } // end anonymous namespace
491 
492 char LoopRerollLegacyPass::ID = 0;
493 
494 INITIALIZE_PASS_BEGIN(LoopRerollLegacyPass, "loop-reroll", "Reroll loops",
495                       false, false)
496 INITIALIZE_PASS_DEPENDENCY(LoopPass)
497 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
498 INITIALIZE_PASS_END(LoopRerollLegacyPass, "loop-reroll", "Reroll loops", false,
499                     false)
500 
501 Pass *llvm::createLoopRerollPass() { return new LoopRerollLegacyPass; }
502 
503 // Returns true if the provided instruction is used outside the given loop.
504 // This operates like Instruction::isUsedOutsideOfBlock, but considers PHIs in
505 // non-loop blocks to be outside the loop.
506 static bool hasUsesOutsideLoop(Instruction *I, Loop *L) {
507   for (User *U : I->users()) {
508     if (!L->contains(cast<Instruction>(U)))
509       return true;
510   }
511   return false;
512 }
513 
514 // Check if an IV is only used to control the loop. There are two cases:
515 // 1. It only has one use which is loop increment, and the increment is only
516 // used by comparison and the PHI (could has sext with nsw in between), and the
517 // comparison is only used by branch.
518 // 2. It is used by loop increment and the comparison, the loop increment is
519 // only used by the PHI, and the comparison is used only by the branch.
520 bool LoopReroll::isLoopControlIV(Loop *L, Instruction *IV) {
521   unsigned IVUses = IV->getNumUses();
522   if (IVUses != 2 && IVUses != 1)
523     return false;
524 
525   for (auto *User : IV->users()) {
526     int32_t IncOrCmpUses = User->getNumUses();
527     bool IsCompInst = isCompareUsedByBranch(cast<Instruction>(User));
528 
529     // User can only have one or two uses.
530     if (IncOrCmpUses != 2 && IncOrCmpUses != 1)
531       return false;
532 
533     // Case 1
534     if (IVUses == 1) {
535       // The only user must be the loop increment.
536       // The loop increment must have two uses.
537       if (IsCompInst || IncOrCmpUses != 2)
538         return false;
539     }
540 
541     // Case 2
542     if (IVUses == 2 && IncOrCmpUses != 1)
543       return false;
544 
545     // The users of the IV must be a binary operation or a comparison
546     if (auto *BO = dyn_cast<BinaryOperator>(User)) {
547       if (BO->getOpcode() == Instruction::Add) {
548         // Loop Increment
549         // User of Loop Increment should be either PHI or CMP
550         for (auto *UU : User->users()) {
551           if (PHINode *PN = dyn_cast<PHINode>(UU)) {
552             if (PN != IV)
553               return false;
554           }
555           // Must be a CMP or an ext (of a value with nsw) then CMP
556           else {
557             auto *UUser = cast<Instruction>(UU);
558             // Skip SExt if we are extending an nsw value
559             // TODO: Allow ZExt too
560             if (BO->hasNoSignedWrap() && UUser->hasOneUse() &&
561                 isa<SExtInst>(UUser))
562               UUser = cast<Instruction>(*(UUser->user_begin()));
563             if (!isCompareUsedByBranch(UUser))
564               return false;
565           }
566         }
567       } else
568         return false;
569       // Compare : can only have one use, and must be branch
570     } else if (!IsCompInst)
571       return false;
572   }
573   return true;
574 }
575 
576 // Collect the list of loop induction variables with respect to which it might
577 // be possible to reroll the loop.
578 void LoopReroll::collectPossibleIVs(Loop *L,
579                                     SmallInstructionVector &PossibleIVs) {
580   BasicBlock *Header = L->getHeader();
581   for (BasicBlock::iterator I = Header->begin(),
582        IE = Header->getFirstInsertionPt(); I != IE; ++I) {
583     if (!isa<PHINode>(I))
584       continue;
585     if (!I->getType()->isIntegerTy() && !I->getType()->isPointerTy())
586       continue;
587 
588     if (const SCEVAddRecExpr *PHISCEV =
589             dyn_cast<SCEVAddRecExpr>(SE->getSCEV(&*I))) {
590       if (PHISCEV->getLoop() != L)
591         continue;
592       if (!PHISCEV->isAffine())
593         continue;
594       auto IncSCEV = dyn_cast<SCEVConstant>(PHISCEV->getStepRecurrence(*SE));
595       if (IncSCEV) {
596         IVToIncMap[&*I] = IncSCEV->getValue()->getSExtValue();
597         LLVM_DEBUG(dbgs() << "LRR: Possible IV: " << *I << " = " << *PHISCEV
598                           << "\n");
599 
600         if (isLoopControlIV(L, &*I)) {
601           assert(!LoopControlIV && "Found two loop control only IV");
602           LoopControlIV = &(*I);
603           LLVM_DEBUG(dbgs() << "LRR: Possible loop control only IV: " << *I
604                             << " = " << *PHISCEV << "\n");
605         } else
606           PossibleIVs.push_back(&*I);
607       }
608     }
609   }
610 }
611 
612 // Add the remainder of the reduction-variable chain to the instruction vector
613 // (the initial PHINode has already been added). If successful, the object is
614 // marked as valid.
615 void LoopReroll::SimpleLoopReduction::add(Loop *L) {
616   assert(!Valid && "Cannot add to an already-valid chain");
617 
618   // The reduction variable must be a chain of single-use instructions
619   // (including the PHI), except for the last value (which is used by the PHI
620   // and also outside the loop).
621   Instruction *C = Instructions.front();
622   if (C->user_empty())
623     return;
624 
625   do {
626     C = cast<Instruction>(*C->user_begin());
627     if (C->hasOneUse()) {
628       if (!C->isBinaryOp())
629         return;
630 
631       if (!(isa<PHINode>(Instructions.back()) ||
632             C->isSameOperationAs(Instructions.back())))
633         return;
634 
635       Instructions.push_back(C);
636     }
637   } while (C->hasOneUse());
638 
639   if (Instructions.size() < 2 ||
640       !C->isSameOperationAs(Instructions.back()) ||
641       C->use_empty())
642     return;
643 
644   // C is now the (potential) last instruction in the reduction chain.
645   for (User *U : C->users()) {
646     // The only in-loop user can be the initial PHI.
647     if (L->contains(cast<Instruction>(U)))
648       if (cast<Instruction>(U) != Instructions.front())
649         return;
650   }
651 
652   Instructions.push_back(C);
653   Valid = true;
654 }
655 
656 // Collect the vector of possible reduction variables.
657 void LoopReroll::collectPossibleReductions(Loop *L,
658   ReductionTracker &Reductions) {
659   BasicBlock *Header = L->getHeader();
660   for (BasicBlock::iterator I = Header->begin(),
661        IE = Header->getFirstInsertionPt(); I != IE; ++I) {
662     if (!isa<PHINode>(I))
663       continue;
664     if (!I->getType()->isSingleValueType())
665       continue;
666 
667     SimpleLoopReduction SLR(&*I, L);
668     if (!SLR.valid())
669       continue;
670 
671     LLVM_DEBUG(dbgs() << "LRR: Possible reduction: " << *I << " (with "
672                       << SLR.size() << " chained instructions)\n");
673     Reductions.addSLR(SLR);
674   }
675 }
676 
677 // Collect the set of all users of the provided root instruction. This set of
678 // users contains not only the direct users of the root instruction, but also
679 // all users of those users, and so on. There are two exceptions:
680 //
681 //   1. Instructions in the set of excluded instructions are never added to the
682 //   use set (even if they are users). This is used, for example, to exclude
683 //   including root increments in the use set of the primary IV.
684 //
685 //   2. Instructions in the set of final instructions are added to the use set
686 //   if they are users, but their users are not added. This is used, for
687 //   example, to prevent a reduction update from forcing all later reduction
688 //   updates into the use set.
689 void LoopReroll::DAGRootTracker::collectInLoopUserSet(
690   Instruction *Root, const SmallInstructionSet &Exclude,
691   const SmallInstructionSet &Final,
692   DenseSet<Instruction *> &Users) {
693   SmallInstructionVector Queue(1, Root);
694   while (!Queue.empty()) {
695     Instruction *I = Queue.pop_back_val();
696     if (!Users.insert(I).second)
697       continue;
698 
699     if (!Final.count(I))
700       for (Use &U : I->uses()) {
701         Instruction *User = cast<Instruction>(U.getUser());
702         if (PHINode *PN = dyn_cast<PHINode>(User)) {
703           // Ignore "wrap-around" uses to PHIs of this loop's header.
704           if (PN->getIncomingBlock(U) == L->getHeader())
705             continue;
706         }
707 
708         if (L->contains(User) && !Exclude.count(User)) {
709           Queue.push_back(User);
710         }
711       }
712 
713     // We also want to collect single-user "feeder" values.
714     for (Use &U : I->operands()) {
715       if (Instruction *Op = dyn_cast<Instruction>(U))
716         if (Op->hasOneUse() && L->contains(Op) && !Exclude.count(Op) &&
717             !Final.count(Op))
718           Queue.push_back(Op);
719     }
720   }
721 }
722 
723 // Collect all of the users of all of the provided root instructions (combined
724 // into a single set).
725 void LoopReroll::DAGRootTracker::collectInLoopUserSet(
726   const SmallInstructionVector &Roots,
727   const SmallInstructionSet &Exclude,
728   const SmallInstructionSet &Final,
729   DenseSet<Instruction *> &Users) {
730   for (Instruction *Root : Roots)
731     collectInLoopUserSet(Root, Exclude, Final, Users);
732 }
733 
734 static bool isUnorderedLoadStore(Instruction *I) {
735   if (LoadInst *LI = dyn_cast<LoadInst>(I))
736     return LI->isUnordered();
737   if (StoreInst *SI = dyn_cast<StoreInst>(I))
738     return SI->isUnordered();
739   if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I))
740     return !MI->isVolatile();
741   return false;
742 }
743 
744 /// Return true if IVU is a "simple" arithmetic operation.
745 /// This is used for narrowing the search space for DAGRoots; only arithmetic
746 /// and GEPs can be part of a DAGRoot.
747 static bool isSimpleArithmeticOp(User *IVU) {
748   if (Instruction *I = dyn_cast<Instruction>(IVU)) {
749     switch (I->getOpcode()) {
750     default: return false;
751     case Instruction::Add:
752     case Instruction::Sub:
753     case Instruction::Mul:
754     case Instruction::Shl:
755     case Instruction::AShr:
756     case Instruction::LShr:
757     case Instruction::GetElementPtr:
758     case Instruction::Trunc:
759     case Instruction::ZExt:
760     case Instruction::SExt:
761       return true;
762     }
763   }
764   return false;
765 }
766 
767 static bool isLoopIncrement(User *U, Instruction *IV) {
768   BinaryOperator *BO = dyn_cast<BinaryOperator>(U);
769 
770   if ((BO && BO->getOpcode() != Instruction::Add) ||
771       (!BO && !isa<GetElementPtrInst>(U)))
772     return false;
773 
774   for (auto *UU : U->users()) {
775     PHINode *PN = dyn_cast<PHINode>(UU);
776     if (PN && PN == IV)
777       return true;
778   }
779   return false;
780 }
781 
782 bool LoopReroll::DAGRootTracker::
783 collectPossibleRoots(Instruction *Base, std::map<int64_t,Instruction*> &Roots) {
784   SmallInstructionVector BaseUsers;
785 
786   for (auto *I : Base->users()) {
787     ConstantInt *CI = nullptr;
788 
789     if (isLoopIncrement(I, IV)) {
790       LoopIncs.push_back(cast<Instruction>(I));
791       continue;
792     }
793 
794     // The root nodes must be either GEPs, ORs or ADDs.
795     if (auto *BO = dyn_cast<BinaryOperator>(I)) {
796       if (BO->getOpcode() == Instruction::Add ||
797           BO->getOpcode() == Instruction::Or)
798         CI = dyn_cast<ConstantInt>(BO->getOperand(1));
799     } else if (auto *GEP = dyn_cast<GetElementPtrInst>(I)) {
800       Value *LastOperand = GEP->getOperand(GEP->getNumOperands()-1);
801       CI = dyn_cast<ConstantInt>(LastOperand);
802     }
803 
804     if (!CI) {
805       if (Instruction *II = dyn_cast<Instruction>(I)) {
806         BaseUsers.push_back(II);
807         continue;
808       } else {
809         LLVM_DEBUG(dbgs() << "LRR: Aborting due to non-instruction: " << *I
810                           << "\n");
811         return false;
812       }
813     }
814 
815     int64_t V = std::abs(CI->getValue().getSExtValue());
816     if (Roots.find(V) != Roots.end())
817       // No duplicates, please.
818       return false;
819 
820     Roots[V] = cast<Instruction>(I);
821   }
822 
823   // Make sure we have at least two roots.
824   if (Roots.empty() || (Roots.size() == 1 && BaseUsers.empty()))
825     return false;
826 
827   // If we found non-loop-inc, non-root users of Base, assume they are
828   // for the zeroth root index. This is because "add %a, 0" gets optimized
829   // away.
830   if (BaseUsers.size()) {
831     if (Roots.find(0) != Roots.end()) {
832       LLVM_DEBUG(dbgs() << "LRR: Multiple roots found for base - aborting!\n");
833       return false;
834     }
835     Roots[0] = Base;
836   }
837 
838   // Calculate the number of users of the base, or lowest indexed, iteration.
839   unsigned NumBaseUses = BaseUsers.size();
840   if (NumBaseUses == 0)
841     NumBaseUses = Roots.begin()->second->getNumUses();
842 
843   // Check that every node has the same number of users.
844   for (auto &KV : Roots) {
845     if (KV.first == 0)
846       continue;
847     if (!KV.second->hasNUses(NumBaseUses)) {
848       LLVM_DEBUG(dbgs() << "LRR: Aborting - Root and Base #users not the same: "
849                         << "#Base=" << NumBaseUses
850                         << ", #Root=" << KV.second->getNumUses() << "\n");
851       return false;
852     }
853   }
854 
855   return true;
856 }
857 
858 void LoopReroll::DAGRootTracker::
859 findRootsRecursive(Instruction *I, SmallInstructionSet SubsumedInsts) {
860   // Does the user look like it could be part of a root set?
861   // All its users must be simple arithmetic ops.
862   if (I->hasNUsesOrMore(IL_MaxRerollIterations + 1))
863     return;
864 
865   if (I != IV && findRootsBase(I, SubsumedInsts))
866     return;
867 
868   SubsumedInsts.insert(I);
869 
870   for (User *V : I->users()) {
871     Instruction *I = cast<Instruction>(V);
872     if (is_contained(LoopIncs, I))
873       continue;
874 
875     if (!isSimpleArithmeticOp(I))
876       continue;
877 
878     // The recursive call makes a copy of SubsumedInsts.
879     findRootsRecursive(I, SubsumedInsts);
880   }
881 }
882 
883 bool LoopReroll::DAGRootTracker::validateRootSet(DAGRootSet &DRS) {
884   if (DRS.Roots.empty())
885     return false;
886 
887   // If the value of the base instruction is used outside the loop, we cannot
888   // reroll the loop. Check for other root instructions is unnecessary because
889   // they don't match any base instructions if their values are used outside.
890   if (hasUsesOutsideLoop(DRS.BaseInst, L))
891     return false;
892 
893   // Consider a DAGRootSet with N-1 roots (so N different values including
894   //   BaseInst).
895   // Define d = Roots[0] - BaseInst, which should be the same as
896   //   Roots[I] - Roots[I-1] for all I in [1..N).
897   // Define D = BaseInst@J - BaseInst@J-1, where "@J" means the value at the
898   //   loop iteration J.
899   //
900   // Now, For the loop iterations to be consecutive:
901   //   D = d * N
902   const auto *ADR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(DRS.BaseInst));
903   if (!ADR)
904     return false;
905 
906   // Check that the first root is evenly spaced.
907   unsigned N = DRS.Roots.size() + 1;
908   const SCEV *StepSCEV = SE->getMinusSCEV(SE->getSCEV(DRS.Roots[0]), ADR);
909   if (isa<SCEVCouldNotCompute>(StepSCEV) || StepSCEV->getType()->isPointerTy())
910     return false;
911   const SCEV *ScaleSCEV = SE->getConstant(StepSCEV->getType(), N);
912   if (ADR->getStepRecurrence(*SE) != SE->getMulExpr(StepSCEV, ScaleSCEV))
913     return false;
914 
915   // Check that the remainling roots are evenly spaced.
916   for (unsigned i = 1; i < N - 1; ++i) {
917     const SCEV *NewStepSCEV = SE->getMinusSCEV(SE->getSCEV(DRS.Roots[i]),
918                                                SE->getSCEV(DRS.Roots[i-1]));
919     if (NewStepSCEV != StepSCEV)
920       return false;
921   }
922 
923   return true;
924 }
925 
926 bool LoopReroll::DAGRootTracker::
927 findRootsBase(Instruction *IVU, SmallInstructionSet SubsumedInsts) {
928   // The base of a RootSet must be an AddRec, so it can be erased.
929   const auto *IVU_ADR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IVU));
930   if (!IVU_ADR || IVU_ADR->getLoop() != L)
931     return false;
932 
933   std::map<int64_t, Instruction*> V;
934   if (!collectPossibleRoots(IVU, V))
935     return false;
936 
937   // If we didn't get a root for index zero, then IVU must be
938   // subsumed.
939   if (V.find(0) == V.end())
940     SubsumedInsts.insert(IVU);
941 
942   // Partition the vector into monotonically increasing indexes.
943   DAGRootSet DRS;
944   DRS.BaseInst = nullptr;
945 
946   SmallVector<DAGRootSet, 16> PotentialRootSets;
947 
948   for (auto &KV : V) {
949     if (!DRS.BaseInst) {
950       DRS.BaseInst = KV.second;
951       DRS.SubsumedInsts = SubsumedInsts;
952     } else if (DRS.Roots.empty()) {
953       DRS.Roots.push_back(KV.second);
954     } else if (V.find(KV.first - 1) != V.end()) {
955       DRS.Roots.push_back(KV.second);
956     } else {
957       // Linear sequence terminated.
958       if (!validateRootSet(DRS))
959         return false;
960 
961       // Construct a new DAGRootSet with the next sequence.
962       PotentialRootSets.push_back(DRS);
963       DRS.BaseInst = KV.second;
964       DRS.Roots.clear();
965     }
966   }
967 
968   if (!validateRootSet(DRS))
969     return false;
970 
971   PotentialRootSets.push_back(DRS);
972 
973   RootSets.append(PotentialRootSets.begin(), PotentialRootSets.end());
974 
975   return true;
976 }
977 
978 bool LoopReroll::DAGRootTracker::findRoots() {
979   Inc = IVToIncMap[IV];
980 
981   assert(RootSets.empty() && "Unclean state!");
982   if (std::abs(Inc) == 1) {
983     for (auto *IVU : IV->users()) {
984       if (isLoopIncrement(IVU, IV))
985         LoopIncs.push_back(cast<Instruction>(IVU));
986     }
987     findRootsRecursive(IV, SmallInstructionSet());
988     LoopIncs.push_back(IV);
989   } else {
990     if (!findRootsBase(IV, SmallInstructionSet()))
991       return false;
992   }
993 
994   // Ensure all sets have the same size.
995   if (RootSets.empty()) {
996     LLVM_DEBUG(dbgs() << "LRR: Aborting because no root sets found!\n");
997     return false;
998   }
999   for (auto &V : RootSets) {
1000     if (V.Roots.empty() || V.Roots.size() != RootSets[0].Roots.size()) {
1001       LLVM_DEBUG(
1002           dbgs()
1003           << "LRR: Aborting because not all root sets have the same size\n");
1004       return false;
1005     }
1006   }
1007 
1008   Scale = RootSets[0].Roots.size() + 1;
1009 
1010   if (Scale > IL_MaxRerollIterations) {
1011     LLVM_DEBUG(dbgs() << "LRR: Aborting - too many iterations found. "
1012                       << "#Found=" << Scale
1013                       << ", #Max=" << IL_MaxRerollIterations << "\n");
1014     return false;
1015   }
1016 
1017   LLVM_DEBUG(dbgs() << "LRR: Successfully found roots: Scale=" << Scale
1018                     << "\n");
1019 
1020   return true;
1021 }
1022 
1023 bool LoopReroll::DAGRootTracker::collectUsedInstructions(SmallInstructionSet &PossibleRedSet) {
1024   // Populate the MapVector with all instructions in the block, in order first,
1025   // so we can iterate over the contents later in perfect order.
1026   for (auto &I : *L->getHeader()) {
1027     Uses[&I].resize(IL_End);
1028   }
1029 
1030   SmallInstructionSet Exclude;
1031   for (auto &DRS : RootSets) {
1032     Exclude.insert(DRS.Roots.begin(), DRS.Roots.end());
1033     Exclude.insert(DRS.SubsumedInsts.begin(), DRS.SubsumedInsts.end());
1034     Exclude.insert(DRS.BaseInst);
1035   }
1036   Exclude.insert(LoopIncs.begin(), LoopIncs.end());
1037 
1038   for (auto &DRS : RootSets) {
1039     DenseSet<Instruction*> VBase;
1040     collectInLoopUserSet(DRS.BaseInst, Exclude, PossibleRedSet, VBase);
1041     for (auto *I : VBase) {
1042       Uses[I].set(0);
1043     }
1044 
1045     unsigned Idx = 1;
1046     for (auto *Root : DRS.Roots) {
1047       DenseSet<Instruction*> V;
1048       collectInLoopUserSet(Root, Exclude, PossibleRedSet, V);
1049 
1050       // While we're here, check the use sets are the same size.
1051       if (V.size() != VBase.size()) {
1052         LLVM_DEBUG(dbgs() << "LRR: Aborting - use sets are different sizes\n");
1053         return false;
1054       }
1055 
1056       for (auto *I : V) {
1057         Uses[I].set(Idx);
1058       }
1059       ++Idx;
1060     }
1061 
1062     // Make sure our subsumed instructions are remembered too.
1063     for (auto *I : DRS.SubsumedInsts) {
1064       Uses[I].set(IL_All);
1065     }
1066   }
1067 
1068   // Make sure the loop increments are also accounted for.
1069 
1070   Exclude.clear();
1071   for (auto &DRS : RootSets) {
1072     Exclude.insert(DRS.Roots.begin(), DRS.Roots.end());
1073     Exclude.insert(DRS.SubsumedInsts.begin(), DRS.SubsumedInsts.end());
1074     Exclude.insert(DRS.BaseInst);
1075   }
1076 
1077   DenseSet<Instruction*> V;
1078   collectInLoopUserSet(LoopIncs, Exclude, PossibleRedSet, V);
1079   for (auto *I : V) {
1080     if (I->mayHaveSideEffects()) {
1081       LLVM_DEBUG(dbgs() << "LRR: Aborting - "
1082                         << "An instruction which does not belong to any root "
1083                         << "sets must not have side effects: " << *I);
1084       return false;
1085     }
1086     Uses[I].set(IL_All);
1087   }
1088 
1089   return true;
1090 }
1091 
1092 /// Get the next instruction in "In" that is a member of set Val.
1093 /// Start searching from StartI, and do not return anything in Exclude.
1094 /// If StartI is not given, start from In.begin().
1095 LoopReroll::DAGRootTracker::UsesTy::iterator
1096 LoopReroll::DAGRootTracker::nextInstr(int Val, UsesTy &In,
1097                                       const SmallInstructionSet &Exclude,
1098                                       UsesTy::iterator *StartI) {
1099   UsesTy::iterator I = StartI ? *StartI : In.begin();
1100   while (I != In.end() && (I->second.test(Val) == 0 ||
1101                            Exclude.contains(I->first)))
1102     ++I;
1103   return I;
1104 }
1105 
1106 bool LoopReroll::DAGRootTracker::isBaseInst(Instruction *I) {
1107   for (auto &DRS : RootSets) {
1108     if (DRS.BaseInst == I)
1109       return true;
1110   }
1111   return false;
1112 }
1113 
1114 bool LoopReroll::DAGRootTracker::isRootInst(Instruction *I) {
1115   for (auto &DRS : RootSets) {
1116     if (is_contained(DRS.Roots, I))
1117       return true;
1118   }
1119   return false;
1120 }
1121 
1122 /// Return true if instruction I depends on any instruction between
1123 /// Start and End.
1124 bool LoopReroll::DAGRootTracker::instrDependsOn(Instruction *I,
1125                                                 UsesTy::iterator Start,
1126                                                 UsesTy::iterator End) {
1127   for (auto *U : I->users()) {
1128     for (auto It = Start; It != End; ++It)
1129       if (U == It->first)
1130         return true;
1131   }
1132   return false;
1133 }
1134 
1135 static bool isIgnorableInst(const Instruction *I) {
1136   if (isa<DbgInfoIntrinsic>(I))
1137     return true;
1138   const IntrinsicInst* II = dyn_cast<IntrinsicInst>(I);
1139   if (!II)
1140     return false;
1141   switch (II->getIntrinsicID()) {
1142     default:
1143       return false;
1144     case Intrinsic::annotation:
1145     case Intrinsic::ptr_annotation:
1146     case Intrinsic::var_annotation:
1147     // TODO: the following intrinsics may also be allowed:
1148     //   lifetime_start, lifetime_end, invariant_start, invariant_end
1149       return true;
1150   }
1151   return false;
1152 }
1153 
1154 bool LoopReroll::DAGRootTracker::validate(ReductionTracker &Reductions) {
1155   // We now need to check for equivalence of the use graph of each root with
1156   // that of the primary induction variable (excluding the roots). Our goal
1157   // here is not to solve the full graph isomorphism problem, but rather to
1158   // catch common cases without a lot of work. As a result, we will assume
1159   // that the relative order of the instructions in each unrolled iteration
1160   // is the same (although we will not make an assumption about how the
1161   // different iterations are intermixed). Note that while the order must be
1162   // the same, the instructions may not be in the same basic block.
1163 
1164   // An array of just the possible reductions for this scale factor. When we
1165   // collect the set of all users of some root instructions, these reduction
1166   // instructions are treated as 'final' (their uses are not considered).
1167   // This is important because we don't want the root use set to search down
1168   // the reduction chain.
1169   SmallInstructionSet PossibleRedSet;
1170   SmallInstructionSet PossibleRedLastSet;
1171   SmallInstructionSet PossibleRedPHISet;
1172   Reductions.restrictToScale(Scale, PossibleRedSet,
1173                              PossibleRedPHISet, PossibleRedLastSet);
1174 
1175   // Populate "Uses" with where each instruction is used.
1176   if (!collectUsedInstructions(PossibleRedSet))
1177     return false;
1178 
1179   // Make sure we mark the reduction PHIs as used in all iterations.
1180   for (auto *I : PossibleRedPHISet) {
1181     Uses[I].set(IL_All);
1182   }
1183 
1184   // Make sure we mark loop-control-only PHIs as used in all iterations. See
1185   // comment above LoopReroll::isLoopControlIV for more information.
1186   BasicBlock *Header = L->getHeader();
1187   if (LoopControlIV && LoopControlIV != IV) {
1188     for (auto *U : LoopControlIV->users()) {
1189       Instruction *IVUser = dyn_cast<Instruction>(U);
1190       // IVUser could be loop increment or compare
1191       Uses[IVUser].set(IL_All);
1192       for (auto *UU : IVUser->users()) {
1193         Instruction *UUser = dyn_cast<Instruction>(UU);
1194         // UUser could be compare, PHI or branch
1195         Uses[UUser].set(IL_All);
1196         // Skip SExt
1197         if (isa<SExtInst>(UUser)) {
1198           UUser = dyn_cast<Instruction>(*(UUser->user_begin()));
1199           Uses[UUser].set(IL_All);
1200         }
1201         // Is UUser a compare instruction?
1202         if (UU->hasOneUse()) {
1203           Instruction *BI = dyn_cast<BranchInst>(*UUser->user_begin());
1204           if (BI == cast<BranchInst>(Header->getTerminator()))
1205             Uses[BI].set(IL_All);
1206         }
1207       }
1208     }
1209   }
1210 
1211   // Make sure all instructions in the loop are in one and only one
1212   // set.
1213   for (auto &KV : Uses) {
1214     if (KV.second.count() != 1 && !isIgnorableInst(KV.first)) {
1215       LLVM_DEBUG(
1216           dbgs() << "LRR: Aborting - instruction is not used in 1 iteration: "
1217                  << *KV.first << " (#uses=" << KV.second.count() << ")\n");
1218       return false;
1219     }
1220   }
1221 
1222   LLVM_DEBUG(for (auto &KV
1223                   : Uses) {
1224     dbgs() << "LRR: " << KV.second.find_first() << "\t" << *KV.first << "\n";
1225   });
1226 
1227   for (unsigned Iter = 1; Iter < Scale; ++Iter) {
1228     // In addition to regular aliasing information, we need to look for
1229     // instructions from later (future) iterations that have side effects
1230     // preventing us from reordering them past other instructions with side
1231     // effects.
1232     bool FutureSideEffects = false;
1233     AliasSetTracker AST(*AA);
1234     // The map between instructions in f(%iv.(i+1)) and f(%iv).
1235     DenseMap<Value *, Value *> BaseMap;
1236 
1237     // Compare iteration Iter to the base.
1238     SmallInstructionSet Visited;
1239     auto BaseIt = nextInstr(0, Uses, Visited);
1240     auto RootIt = nextInstr(Iter, Uses, Visited);
1241     auto LastRootIt = Uses.begin();
1242 
1243     while (BaseIt != Uses.end() && RootIt != Uses.end()) {
1244       Instruction *BaseInst = BaseIt->first;
1245       Instruction *RootInst = RootIt->first;
1246 
1247       // Skip over the IV or root instructions; only match their users.
1248       bool Continue = false;
1249       if (isBaseInst(BaseInst)) {
1250         Visited.insert(BaseInst);
1251         BaseIt = nextInstr(0, Uses, Visited);
1252         Continue = true;
1253       }
1254       if (isRootInst(RootInst)) {
1255         LastRootIt = RootIt;
1256         Visited.insert(RootInst);
1257         RootIt = nextInstr(Iter, Uses, Visited);
1258         Continue = true;
1259       }
1260       if (Continue) continue;
1261 
1262       if (!BaseInst->isSameOperationAs(RootInst)) {
1263         // Last chance saloon. We don't try and solve the full isomorphism
1264         // problem, but try and at least catch the case where two instructions
1265         // *of different types* are round the wrong way. We won't be able to
1266         // efficiently tell, given two ADD instructions, which way around we
1267         // should match them, but given an ADD and a SUB, we can at least infer
1268         // which one is which.
1269         //
1270         // This should allow us to deal with a greater subset of the isomorphism
1271         // problem. It does however change a linear algorithm into a quadratic
1272         // one, so limit the number of probes we do.
1273         auto TryIt = RootIt;
1274         unsigned N = NumToleratedFailedMatches;
1275         while (TryIt != Uses.end() &&
1276                !BaseInst->isSameOperationAs(TryIt->first) &&
1277                N--) {
1278           ++TryIt;
1279           TryIt = nextInstr(Iter, Uses, Visited, &TryIt);
1280         }
1281 
1282         if (TryIt == Uses.end() || TryIt == RootIt ||
1283             instrDependsOn(TryIt->first, RootIt, TryIt)) {
1284           LLVM_DEBUG(dbgs() << "LRR: iteration root match failed at "
1285                             << *BaseInst << " vs. " << *RootInst << "\n");
1286           return false;
1287         }
1288 
1289         RootIt = TryIt;
1290         RootInst = TryIt->first;
1291       }
1292 
1293       // All instructions between the last root and this root
1294       // may belong to some other iteration. If they belong to a
1295       // future iteration, then they're dangerous to alias with.
1296       //
1297       // Note that because we allow a limited amount of flexibility in the order
1298       // that we visit nodes, LastRootIt might be *before* RootIt, in which
1299       // case we've already checked this set of instructions so we shouldn't
1300       // do anything.
1301       for (; LastRootIt < RootIt; ++LastRootIt) {
1302         Instruction *I = LastRootIt->first;
1303         if (LastRootIt->second.find_first() < (int)Iter)
1304           continue;
1305         if (I->mayWriteToMemory())
1306           AST.add(I);
1307         // Note: This is specifically guarded by a check on isa<PHINode>,
1308         // which while a valid (somewhat arbitrary) micro-optimization, is
1309         // needed because otherwise isSafeToSpeculativelyExecute returns
1310         // false on PHI nodes.
1311         if (!isa<PHINode>(I) && !isUnorderedLoadStore(I) &&
1312             !isSafeToSpeculativelyExecute(I))
1313           // Intervening instructions cause side effects.
1314           FutureSideEffects = true;
1315       }
1316 
1317       // Make sure that this instruction, which is in the use set of this
1318       // root instruction, does not also belong to the base set or the set of
1319       // some other root instruction.
1320       if (RootIt->second.count() > 1) {
1321         LLVM_DEBUG(dbgs() << "LRR: iteration root match failed at " << *BaseInst
1322                           << " vs. " << *RootInst << " (prev. case overlap)\n");
1323         return false;
1324       }
1325 
1326       // Make sure that we don't alias with any instruction in the alias set
1327       // tracker. If we do, then we depend on a future iteration, and we
1328       // can't reroll.
1329       if (RootInst->mayReadFromMemory())
1330         for (auto &K : AST) {
1331           if (K.aliasesUnknownInst(RootInst, *AA)) {
1332             LLVM_DEBUG(dbgs() << "LRR: iteration root match failed at "
1333                               << *BaseInst << " vs. " << *RootInst
1334                               << " (depends on future store)\n");
1335             return false;
1336           }
1337         }
1338 
1339       // If we've past an instruction from a future iteration that may have
1340       // side effects, and this instruction might also, then we can't reorder
1341       // them, and this matching fails. As an exception, we allow the alias
1342       // set tracker to handle regular (unordered) load/store dependencies.
1343       if (FutureSideEffects && ((!isUnorderedLoadStore(BaseInst) &&
1344                                  !isSafeToSpeculativelyExecute(BaseInst)) ||
1345                                 (!isUnorderedLoadStore(RootInst) &&
1346                                  !isSafeToSpeculativelyExecute(RootInst)))) {
1347         LLVM_DEBUG(dbgs() << "LRR: iteration root match failed at " << *BaseInst
1348                           << " vs. " << *RootInst
1349                           << " (side effects prevent reordering)\n");
1350         return false;
1351       }
1352 
1353       // For instructions that are part of a reduction, if the operation is
1354       // associative, then don't bother matching the operands (because we
1355       // already know that the instructions are isomorphic, and the order
1356       // within the iteration does not matter). For non-associative reductions,
1357       // we do need to match the operands, because we need to reject
1358       // out-of-order instructions within an iteration!
1359       // For example (assume floating-point addition), we need to reject this:
1360       //   x += a[i]; x += b[i];
1361       //   x += a[i+1]; x += b[i+1];
1362       //   x += b[i+2]; x += a[i+2];
1363       bool InReduction = Reductions.isPairInSame(BaseInst, RootInst);
1364 
1365       if (!(InReduction && BaseInst->isAssociative())) {
1366         bool Swapped = false, SomeOpMatched = false;
1367         for (unsigned j = 0; j < BaseInst->getNumOperands(); ++j) {
1368           Value *Op2 = RootInst->getOperand(j);
1369 
1370           // If this is part of a reduction (and the operation is not
1371           // associatve), then we match all operands, but not those that are
1372           // part of the reduction.
1373           if (InReduction)
1374             if (Instruction *Op2I = dyn_cast<Instruction>(Op2))
1375               if (Reductions.isPairInSame(RootInst, Op2I))
1376                 continue;
1377 
1378           DenseMap<Value *, Value *>::iterator BMI = BaseMap.find(Op2);
1379           if (BMI != BaseMap.end()) {
1380             Op2 = BMI->second;
1381           } else {
1382             for (auto &DRS : RootSets) {
1383               if (DRS.Roots[Iter-1] == (Instruction*) Op2) {
1384                 Op2 = DRS.BaseInst;
1385                 break;
1386               }
1387             }
1388           }
1389 
1390           if (BaseInst->getOperand(Swapped ? unsigned(!j) : j) != Op2) {
1391             // If we've not already decided to swap the matched operands, and
1392             // we've not already matched our first operand (note that we could
1393             // have skipped matching the first operand because it is part of a
1394             // reduction above), and the instruction is commutative, then try
1395             // the swapped match.
1396             if (!Swapped && BaseInst->isCommutative() && !SomeOpMatched &&
1397                 BaseInst->getOperand(!j) == Op2) {
1398               Swapped = true;
1399             } else {
1400               LLVM_DEBUG(dbgs()
1401                          << "LRR: iteration root match failed at " << *BaseInst
1402                          << " vs. " << *RootInst << " (operand " << j << ")\n");
1403               return false;
1404             }
1405           }
1406 
1407           SomeOpMatched = true;
1408         }
1409       }
1410 
1411       if ((!PossibleRedLastSet.count(BaseInst) &&
1412            hasUsesOutsideLoop(BaseInst, L)) ||
1413           (!PossibleRedLastSet.count(RootInst) &&
1414            hasUsesOutsideLoop(RootInst, L))) {
1415         LLVM_DEBUG(dbgs() << "LRR: iteration root match failed at " << *BaseInst
1416                           << " vs. " << *RootInst << " (uses outside loop)\n");
1417         return false;
1418       }
1419 
1420       Reductions.recordPair(BaseInst, RootInst, Iter);
1421       BaseMap.insert(std::make_pair(RootInst, BaseInst));
1422 
1423       LastRootIt = RootIt;
1424       Visited.insert(BaseInst);
1425       Visited.insert(RootInst);
1426       BaseIt = nextInstr(0, Uses, Visited);
1427       RootIt = nextInstr(Iter, Uses, Visited);
1428     }
1429     assert(BaseIt == Uses.end() && RootIt == Uses.end() &&
1430            "Mismatched set sizes!");
1431   }
1432 
1433   LLVM_DEBUG(dbgs() << "LRR: Matched all iteration increments for " << *IV
1434                     << "\n");
1435 
1436   return true;
1437 }
1438 
1439 void LoopReroll::DAGRootTracker::replace(const SCEV *BackedgeTakenCount) {
1440   BasicBlock *Header = L->getHeader();
1441 
1442   // Compute the start and increment for each BaseInst before we start erasing
1443   // instructions.
1444   SmallVector<const SCEV *, 8> StartExprs;
1445   SmallVector<const SCEV *, 8> IncrExprs;
1446   for (auto &DRS : RootSets) {
1447     const SCEVAddRecExpr *IVSCEV =
1448         cast<SCEVAddRecExpr>(SE->getSCEV(DRS.BaseInst));
1449     StartExprs.push_back(IVSCEV->getStart());
1450     IncrExprs.push_back(SE->getMinusSCEV(SE->getSCEV(DRS.Roots[0]), IVSCEV));
1451   }
1452 
1453   // Remove instructions associated with non-base iterations.
1454   for (Instruction &Inst : llvm::make_early_inc_range(llvm::reverse(*Header))) {
1455     unsigned I = Uses[&Inst].find_first();
1456     if (I > 0 && I < IL_All) {
1457       LLVM_DEBUG(dbgs() << "LRR: removing: " << Inst << "\n");
1458       Inst.eraseFromParent();
1459     }
1460   }
1461 
1462   // Rewrite each BaseInst using SCEV.
1463   for (size_t i = 0, e = RootSets.size(); i != e; ++i)
1464     // Insert the new induction variable.
1465     replaceIV(RootSets[i], StartExprs[i], IncrExprs[i]);
1466 
1467   { // Limit the lifetime of SCEVExpander.
1468     BranchInst *BI = cast<BranchInst>(Header->getTerminator());
1469     const DataLayout &DL = Header->getModule()->getDataLayout();
1470     SCEVExpander Expander(*SE, DL, "reroll");
1471     auto Zero = SE->getZero(BackedgeTakenCount->getType());
1472     auto One = SE->getOne(BackedgeTakenCount->getType());
1473     auto NewIVSCEV = SE->getAddRecExpr(Zero, One, L, SCEV::FlagAnyWrap);
1474     Value *NewIV =
1475         Expander.expandCodeFor(NewIVSCEV, BackedgeTakenCount->getType(),
1476                                Header->getFirstNonPHIOrDbg());
1477     // FIXME: This arithmetic can overflow.
1478     auto TripCount = SE->getAddExpr(BackedgeTakenCount, One);
1479     auto ScaledTripCount = SE->getMulExpr(
1480         TripCount, SE->getConstant(BackedgeTakenCount->getType(), Scale));
1481     auto ScaledBECount = SE->getMinusSCEV(ScaledTripCount, One);
1482     Value *TakenCount =
1483         Expander.expandCodeFor(ScaledBECount, BackedgeTakenCount->getType(),
1484                                Header->getFirstNonPHIOrDbg());
1485     Value *Cond =
1486         new ICmpInst(BI, CmpInst::ICMP_EQ, NewIV, TakenCount, "exitcond");
1487     BI->setCondition(Cond);
1488 
1489     if (BI->getSuccessor(1) != Header)
1490       BI->swapSuccessors();
1491   }
1492 
1493   SimplifyInstructionsInBlock(Header, TLI);
1494   DeleteDeadPHIs(Header, TLI);
1495 }
1496 
1497 void LoopReroll::DAGRootTracker::replaceIV(DAGRootSet &DRS,
1498                                            const SCEV *Start,
1499                                            const SCEV *IncrExpr) {
1500   BasicBlock *Header = L->getHeader();
1501   Instruction *Inst = DRS.BaseInst;
1502 
1503   const SCEV *NewIVSCEV =
1504       SE->getAddRecExpr(Start, IncrExpr, L, SCEV::FlagAnyWrap);
1505 
1506   { // Limit the lifetime of SCEVExpander.
1507     const DataLayout &DL = Header->getModule()->getDataLayout();
1508     SCEVExpander Expander(*SE, DL, "reroll");
1509     Value *NewIV = Expander.expandCodeFor(NewIVSCEV, Inst->getType(),
1510                                           Header->getFirstNonPHIOrDbg());
1511 
1512     for (auto &KV : Uses)
1513       if (KV.second.find_first() == 0)
1514         KV.first->replaceUsesOfWith(Inst, NewIV);
1515   }
1516 }
1517 
1518 // Validate the selected reductions. All iterations must have an isomorphic
1519 // part of the reduction chain and, for non-associative reductions, the chain
1520 // entries must appear in order.
1521 bool LoopReroll::ReductionTracker::validateSelected() {
1522   // For a non-associative reduction, the chain entries must appear in order.
1523   for (int i : Reds) {
1524     int PrevIter = 0, BaseCount = 0, Count = 0;
1525     for (Instruction *J : PossibleReds[i]) {
1526       // Note that all instructions in the chain must have been found because
1527       // all instructions in the function must have been assigned to some
1528       // iteration.
1529       int Iter = PossibleRedIter[J];
1530       if (Iter != PrevIter && Iter != PrevIter + 1 &&
1531           !PossibleReds[i].getReducedValue()->isAssociative()) {
1532         LLVM_DEBUG(dbgs() << "LRR: Out-of-order non-associative reduction: "
1533                           << J << "\n");
1534         return false;
1535       }
1536 
1537       if (Iter != PrevIter) {
1538         if (Count != BaseCount) {
1539           LLVM_DEBUG(dbgs()
1540                      << "LRR: Iteration " << PrevIter << " reduction use count "
1541                      << Count << " is not equal to the base use count "
1542                      << BaseCount << "\n");
1543           return false;
1544         }
1545 
1546         Count = 0;
1547       }
1548 
1549       ++Count;
1550       if (Iter == 0)
1551         ++BaseCount;
1552 
1553       PrevIter = Iter;
1554     }
1555   }
1556 
1557   return true;
1558 }
1559 
1560 // For all selected reductions, remove all parts except those in the first
1561 // iteration (and the PHI). Replace outside uses of the reduced value with uses
1562 // of the first-iteration reduced value (in other words, reroll the selected
1563 // reductions).
1564 void LoopReroll::ReductionTracker::replaceSelected() {
1565   // Fixup reductions to refer to the last instruction associated with the
1566   // first iteration (not the last).
1567   for (int i : Reds) {
1568     int j = 0;
1569     for (int e = PossibleReds[i].size(); j != e; ++j)
1570       if (PossibleRedIter[PossibleReds[i][j]] != 0) {
1571         --j;
1572         break;
1573       }
1574 
1575     // Replace users with the new end-of-chain value.
1576     SmallInstructionVector Users;
1577     for (User *U : PossibleReds[i].getReducedValue()->users()) {
1578       Users.push_back(cast<Instruction>(U));
1579     }
1580 
1581     for (Instruction *User : Users)
1582       User->replaceUsesOfWith(PossibleReds[i].getReducedValue(),
1583                               PossibleReds[i][j]);
1584   }
1585 }
1586 
1587 // Reroll the provided loop with respect to the provided induction variable.
1588 // Generally, we're looking for a loop like this:
1589 //
1590 // %iv = phi [ (preheader, ...), (body, %iv.next) ]
1591 // f(%iv)
1592 // %iv.1 = add %iv, 1                <-- a root increment
1593 // f(%iv.1)
1594 // %iv.2 = add %iv, 2                <-- a root increment
1595 // f(%iv.2)
1596 // %iv.scale_m_1 = add %iv, scale-1  <-- a root increment
1597 // f(%iv.scale_m_1)
1598 // ...
1599 // %iv.next = add %iv, scale
1600 // %cmp = icmp(%iv, ...)
1601 // br %cmp, header, exit
1602 //
1603 // Notably, we do not require that f(%iv), f(%iv.1), etc. be isolated groups of
1604 // instructions. In other words, the instructions in f(%iv), f(%iv.1), etc. can
1605 // be intermixed with eachother. The restriction imposed by this algorithm is
1606 // that the relative order of the isomorphic instructions in f(%iv), f(%iv.1),
1607 // etc. be the same.
1608 //
1609 // First, we collect the use set of %iv, excluding the other increment roots.
1610 // This gives us f(%iv). Then we iterate over the loop instructions (scale-1)
1611 // times, having collected the use set of f(%iv.(i+1)), during which we:
1612 //   - Ensure that the next unmatched instruction in f(%iv) is isomorphic to
1613 //     the next unmatched instruction in f(%iv.(i+1)).
1614 //   - Ensure that both matched instructions don't have any external users
1615 //     (with the exception of last-in-chain reduction instructions).
1616 //   - Track the (aliasing) write set, and other side effects, of all
1617 //     instructions that belong to future iterations that come before the matched
1618 //     instructions. If the matched instructions read from that write set, then
1619 //     f(%iv) or f(%iv.(i+1)) has some dependency on instructions in
1620 //     f(%iv.(j+1)) for some j > i, and we cannot reroll the loop. Similarly,
1621 //     if any of these future instructions had side effects (could not be
1622 //     speculatively executed), and so do the matched instructions, when we
1623 //     cannot reorder those side-effect-producing instructions, and rerolling
1624 //     fails.
1625 //
1626 // Finally, we make sure that all loop instructions are either loop increment
1627 // roots, belong to simple latch code, parts of validated reductions, part of
1628 // f(%iv) or part of some f(%iv.i). If all of that is true (and all reductions
1629 // have been validated), then we reroll the loop.
1630 bool LoopReroll::reroll(Instruction *IV, Loop *L, BasicBlock *Header,
1631                         const SCEV *BackedgeTakenCount,
1632                         ReductionTracker &Reductions) {
1633   DAGRootTracker DAGRoots(this, L, IV, SE, AA, TLI, DT, LI, PreserveLCSSA,
1634                           IVToIncMap, LoopControlIV);
1635 
1636   if (!DAGRoots.findRoots())
1637     return false;
1638   LLVM_DEBUG(dbgs() << "LRR: Found all root induction increments for: " << *IV
1639                     << "\n");
1640 
1641   if (!DAGRoots.validate(Reductions))
1642     return false;
1643   if (!Reductions.validateSelected())
1644     return false;
1645   // At this point, we've validated the rerolling, and we're committed to
1646   // making changes!
1647 
1648   Reductions.replaceSelected();
1649   DAGRoots.replace(BackedgeTakenCount);
1650 
1651   ++NumRerolledLoops;
1652   return true;
1653 }
1654 
1655 bool LoopReroll::runOnLoop(Loop *L) {
1656   BasicBlock *Header = L->getHeader();
1657   LLVM_DEBUG(dbgs() << "LRR: F[" << Header->getParent()->getName() << "] Loop %"
1658                     << Header->getName() << " (" << L->getNumBlocks()
1659                     << " block(s))\n");
1660 
1661   // For now, we'll handle only single BB loops.
1662   if (L->getNumBlocks() > 1)
1663     return false;
1664 
1665   if (!SE->hasLoopInvariantBackedgeTakenCount(L))
1666     return false;
1667 
1668   const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1669   LLVM_DEBUG(dbgs() << "\n Before Reroll:\n" << *(L->getHeader()) << "\n");
1670   LLVM_DEBUG(dbgs() << "LRR: backedge-taken count = " << *BackedgeTakenCount
1671                << "\n");
1672 
1673   // First, we need to find the induction variable with respect to which we can
1674   // reroll (there may be several possible options).
1675   SmallInstructionVector PossibleIVs;
1676   IVToIncMap.clear();
1677   LoopControlIV = nullptr;
1678   collectPossibleIVs(L, PossibleIVs);
1679 
1680   if (PossibleIVs.empty()) {
1681     LLVM_DEBUG(dbgs() << "LRR: No possible IVs found\n");
1682     return false;
1683   }
1684 
1685   ReductionTracker Reductions;
1686   collectPossibleReductions(L, Reductions);
1687   bool Changed = false;
1688 
1689   // For each possible IV, collect the associated possible set of 'root' nodes
1690   // (i+1, i+2, etc.).
1691   for (Instruction *PossibleIV : PossibleIVs)
1692     if (reroll(PossibleIV, L, Header, BackedgeTakenCount, Reductions)) {
1693       Changed = true;
1694       break;
1695     }
1696   LLVM_DEBUG(dbgs() << "\n After Reroll:\n" << *(L->getHeader()) << "\n");
1697 
1698   // Trip count of L has changed so SE must be re-evaluated.
1699   if (Changed)
1700     SE->forgetLoop(L);
1701 
1702   return Changed;
1703 }
1704 
1705 bool LoopRerollLegacyPass::runOnLoop(Loop *L, LPPassManager &LPM) {
1706   if (skipLoop(L))
1707     return false;
1708 
1709   auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
1710   auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1711   auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1712   auto *TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(
1713       *L->getHeader()->getParent());
1714   auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1715   bool PreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
1716 
1717   return LoopReroll(AA, LI, SE, TLI, DT, PreserveLCSSA).runOnLoop(L);
1718 }
1719 
1720 PreservedAnalyses LoopRerollPass::run(Loop &L, LoopAnalysisManager &AM,
1721                                       LoopStandardAnalysisResults &AR,
1722                                       LPMUpdater &U) {
1723   return LoopReroll(&AR.AA, &AR.LI, &AR.SE, &AR.TLI, &AR.DT, true).runOnLoop(&L)
1724              ? getLoopPassPreservedAnalyses()
1725              : PreservedAnalyses::all();
1726 }
1727