1 //===- LoopReroll.cpp - Loop rerolling pass -------------------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This pass implements a simple loop reroller.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "llvm/ADT/APInt.h"
15 #include "llvm/ADT/BitVector.h"
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/DenseSet.h"
18 #include "llvm/ADT/MapVector.h"
19 #include "llvm/ADT/STLExtras.h"
20 #include "llvm/ADT/SmallPtrSet.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/Analysis/AliasAnalysis.h"
24 #include "llvm/Analysis/AliasSetTracker.h"
25 #include "llvm/Analysis/LoopInfo.h"
26 #include "llvm/Analysis/LoopPass.h"
27 #include "llvm/Analysis/ScalarEvolution.h"
28 #include "llvm/Analysis/ScalarEvolutionExpander.h"
29 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
30 #include "llvm/Analysis/TargetLibraryInfo.h"
31 #include "llvm/Transforms/Utils/Local.h"
32 #include "llvm/Analysis/ValueTracking.h"
33 #include "llvm/IR/BasicBlock.h"
34 #include "llvm/IR/Constants.h"
35 #include "llvm/IR/DataLayout.h"
36 #include "llvm/IR/DerivedTypes.h"
37 #include "llvm/IR/Dominators.h"
38 #include "llvm/IR/IRBuilder.h"
39 #include "llvm/IR/InstrTypes.h"
40 #include "llvm/IR/Instruction.h"
41 #include "llvm/IR/Instructions.h"
42 #include "llvm/IR/IntrinsicInst.h"
43 #include "llvm/IR/Intrinsics.h"
44 #include "llvm/IR/Module.h"
45 #include "llvm/IR/Type.h"
46 #include "llvm/IR/Use.h"
47 #include "llvm/IR/User.h"
48 #include "llvm/IR/Value.h"
49 #include "llvm/Pass.h"
50 #include "llvm/Support/Casting.h"
51 #include "llvm/Support/CommandLine.h"
52 #include "llvm/Support/Debug.h"
53 #include "llvm/Support/raw_ostream.h"
54 #include "llvm/Transforms/Scalar.h"
55 #include "llvm/Transforms/Utils.h"
56 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
57 #include "llvm/Transforms/Utils/LoopUtils.h"
58 #include <cassert>
59 #include <cstddef>
60 #include <cstdint>
61 #include <cstdlib>
62 #include <iterator>
63 #include <map>
64 #include <utility>
65
66 using namespace llvm;
67
68 #define DEBUG_TYPE "loop-reroll"
69
70 STATISTIC(NumRerolledLoops, "Number of rerolled loops");
71
72 static cl::opt<unsigned>
73 NumToleratedFailedMatches("reroll-num-tolerated-failed-matches", cl::init(400),
74 cl::Hidden,
75 cl::desc("The maximum number of failures to tolerate"
76 " during fuzzy matching. (default: 400)"));
77
78 // This loop re-rolling transformation aims to transform loops like this:
79 //
80 // int foo(int a);
81 // void bar(int *x) {
82 // for (int i = 0; i < 500; i += 3) {
83 // foo(i);
84 // foo(i+1);
85 // foo(i+2);
86 // }
87 // }
88 //
89 // into a loop like this:
90 //
91 // void bar(int *x) {
92 // for (int i = 0; i < 500; ++i)
93 // foo(i);
94 // }
95 //
96 // It does this by looking for loops that, besides the latch code, are composed
97 // of isomorphic DAGs of instructions, with each DAG rooted at some increment
98 // to the induction variable, and where each DAG is isomorphic to the DAG
99 // rooted at the induction variable (excepting the sub-DAGs which root the
100 // other induction-variable increments). In other words, we're looking for loop
101 // bodies of the form:
102 //
103 // %iv = phi [ (preheader, ...), (body, %iv.next) ]
104 // f(%iv)
105 // %iv.1 = add %iv, 1 <-- a root increment
106 // f(%iv.1)
107 // %iv.2 = add %iv, 2 <-- a root increment
108 // f(%iv.2)
109 // %iv.scale_m_1 = add %iv, scale-1 <-- a root increment
110 // f(%iv.scale_m_1)
111 // ...
112 // %iv.next = add %iv, scale
113 // %cmp = icmp(%iv, ...)
114 // br %cmp, header, exit
115 //
116 // where each f(i) is a set of instructions that, collectively, are a function
117 // only of i (and other loop-invariant values).
118 //
119 // As a special case, we can also reroll loops like this:
120 //
121 // int foo(int);
122 // void bar(int *x) {
123 // for (int i = 0; i < 500; ++i) {
124 // x[3*i] = foo(0);
125 // x[3*i+1] = foo(0);
126 // x[3*i+2] = foo(0);
127 // }
128 // }
129 //
130 // into this:
131 //
132 // void bar(int *x) {
133 // for (int i = 0; i < 1500; ++i)
134 // x[i] = foo(0);
135 // }
136 //
137 // in which case, we're looking for inputs like this:
138 //
139 // %iv = phi [ (preheader, ...), (body, %iv.next) ]
140 // %scaled.iv = mul %iv, scale
141 // f(%scaled.iv)
142 // %scaled.iv.1 = add %scaled.iv, 1
143 // f(%scaled.iv.1)
144 // %scaled.iv.2 = add %scaled.iv, 2
145 // f(%scaled.iv.2)
146 // %scaled.iv.scale_m_1 = add %scaled.iv, scale-1
147 // f(%scaled.iv.scale_m_1)
148 // ...
149 // %iv.next = add %iv, 1
150 // %cmp = icmp(%iv, ...)
151 // br %cmp, header, exit
152
153 namespace {
154
155 enum IterationLimits {
156 /// The maximum number of iterations that we'll try and reroll.
157 IL_MaxRerollIterations = 32,
158 /// The bitvector index used by loop induction variables and other
159 /// instructions that belong to all iterations.
160 IL_All,
161 IL_End
162 };
163
164 class LoopReroll : public LoopPass {
165 public:
166 static char ID; // Pass ID, replacement for typeid
167
LoopReroll()168 LoopReroll() : LoopPass(ID) {
169 initializeLoopRerollPass(*PassRegistry::getPassRegistry());
170 }
171
172 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
173
getAnalysisUsage(AnalysisUsage & AU) const174 void getAnalysisUsage(AnalysisUsage &AU) const override {
175 AU.addRequired<TargetLibraryInfoWrapperPass>();
176 getLoopAnalysisUsage(AU);
177 }
178
179 protected:
180 AliasAnalysis *AA;
181 LoopInfo *LI;
182 ScalarEvolution *SE;
183 TargetLibraryInfo *TLI;
184 DominatorTree *DT;
185 bool PreserveLCSSA;
186
187 using SmallInstructionVector = SmallVector<Instruction *, 16>;
188 using SmallInstructionSet = SmallPtrSet<Instruction *, 16>;
189
190 // Map between induction variable and its increment
191 DenseMap<Instruction *, int64_t> IVToIncMap;
192
193 // For loop with multiple induction variable, remember the one used only to
194 // control the loop.
195 Instruction *LoopControlIV;
196
197 // A chain of isomorphic instructions, identified by a single-use PHI
198 // representing a reduction. Only the last value may be used outside the
199 // loop.
200 struct SimpleLoopReduction {
SimpleLoopReduction__anon39f464390111::LoopReroll::SimpleLoopReduction201 SimpleLoopReduction(Instruction *P, Loop *L) : Instructions(1, P) {
202 assert(isa<PHINode>(P) && "First reduction instruction must be a PHI");
203 add(L);
204 }
205
valid__anon39f464390111::LoopReroll::SimpleLoopReduction206 bool valid() const {
207 return Valid;
208 }
209
getPHI__anon39f464390111::LoopReroll::SimpleLoopReduction210 Instruction *getPHI() const {
211 assert(Valid && "Using invalid reduction");
212 return Instructions.front();
213 }
214
getReducedValue__anon39f464390111::LoopReroll::SimpleLoopReduction215 Instruction *getReducedValue() const {
216 assert(Valid && "Using invalid reduction");
217 return Instructions.back();
218 }
219
get__anon39f464390111::LoopReroll::SimpleLoopReduction220 Instruction *get(size_t i) const {
221 assert(Valid && "Using invalid reduction");
222 return Instructions[i+1];
223 }
224
operator []__anon39f464390111::LoopReroll::SimpleLoopReduction225 Instruction *operator [] (size_t i) const { return get(i); }
226
227 // The size, ignoring the initial PHI.
size__anon39f464390111::LoopReroll::SimpleLoopReduction228 size_t size() const {
229 assert(Valid && "Using invalid reduction");
230 return Instructions.size()-1;
231 }
232
233 using iterator = SmallInstructionVector::iterator;
234 using const_iterator = SmallInstructionVector::const_iterator;
235
begin__anon39f464390111::LoopReroll::SimpleLoopReduction236 iterator begin() {
237 assert(Valid && "Using invalid reduction");
238 return std::next(Instructions.begin());
239 }
240
begin__anon39f464390111::LoopReroll::SimpleLoopReduction241 const_iterator begin() const {
242 assert(Valid && "Using invalid reduction");
243 return std::next(Instructions.begin());
244 }
245
end__anon39f464390111::LoopReroll::SimpleLoopReduction246 iterator end() { return Instructions.end(); }
end__anon39f464390111::LoopReroll::SimpleLoopReduction247 const_iterator end() const { return Instructions.end(); }
248
249 protected:
250 bool Valid = false;
251 SmallInstructionVector Instructions;
252
253 void add(Loop *L);
254 };
255
256 // The set of all reductions, and state tracking of possible reductions
257 // during loop instruction processing.
258 struct ReductionTracker {
259 using SmallReductionVector = SmallVector<SimpleLoopReduction, 16>;
260
261 // Add a new possible reduction.
addSLR__anon39f464390111::LoopReroll::ReductionTracker262 void addSLR(SimpleLoopReduction &SLR) { PossibleReds.push_back(SLR); }
263
264 // Setup to track possible reductions corresponding to the provided
265 // rerolling scale. Only reductions with a number of non-PHI instructions
266 // that is divisible by the scale are considered. Three instructions sets
267 // are filled in:
268 // - A set of all possible instructions in eligible reductions.
269 // - A set of all PHIs in eligible reductions
270 // - A set of all reduced values (last instructions) in eligible
271 // reductions.
restrictToScale__anon39f464390111::LoopReroll::ReductionTracker272 void restrictToScale(uint64_t Scale,
273 SmallInstructionSet &PossibleRedSet,
274 SmallInstructionSet &PossibleRedPHISet,
275 SmallInstructionSet &PossibleRedLastSet) {
276 PossibleRedIdx.clear();
277 PossibleRedIter.clear();
278 Reds.clear();
279
280 for (unsigned i = 0, e = PossibleReds.size(); i != e; ++i)
281 if (PossibleReds[i].size() % Scale == 0) {
282 PossibleRedLastSet.insert(PossibleReds[i].getReducedValue());
283 PossibleRedPHISet.insert(PossibleReds[i].getPHI());
284
285 PossibleRedSet.insert(PossibleReds[i].getPHI());
286 PossibleRedIdx[PossibleReds[i].getPHI()] = i;
287 for (Instruction *J : PossibleReds[i]) {
288 PossibleRedSet.insert(J);
289 PossibleRedIdx[J] = i;
290 }
291 }
292 }
293
294 // The functions below are used while processing the loop instructions.
295
296 // Are the two instructions both from reductions, and furthermore, from
297 // the same reduction?
isPairInSame__anon39f464390111::LoopReroll::ReductionTracker298 bool isPairInSame(Instruction *J1, Instruction *J2) {
299 DenseMap<Instruction *, int>::iterator J1I = PossibleRedIdx.find(J1);
300 if (J1I != PossibleRedIdx.end()) {
301 DenseMap<Instruction *, int>::iterator J2I = PossibleRedIdx.find(J2);
302 if (J2I != PossibleRedIdx.end() && J1I->second == J2I->second)
303 return true;
304 }
305
306 return false;
307 }
308
309 // The two provided instructions, the first from the base iteration, and
310 // the second from iteration i, form a matched pair. If these are part of
311 // a reduction, record that fact.
recordPair__anon39f464390111::LoopReroll::ReductionTracker312 void recordPair(Instruction *J1, Instruction *J2, unsigned i) {
313 if (PossibleRedIdx.count(J1)) {
314 assert(PossibleRedIdx.count(J2) &&
315 "Recording reduction vs. non-reduction instruction?");
316
317 PossibleRedIter[J1] = 0;
318 PossibleRedIter[J2] = i;
319
320 int Idx = PossibleRedIdx[J1];
321 assert(Idx == PossibleRedIdx[J2] &&
322 "Recording pair from different reductions?");
323 Reds.insert(Idx);
324 }
325 }
326
327 // The functions below can be called after we've finished processing all
328 // instructions in the loop, and we know which reductions were selected.
329
330 bool validateSelected();
331 void replaceSelected();
332
333 protected:
334 // The vector of all possible reductions (for any scale).
335 SmallReductionVector PossibleReds;
336
337 DenseMap<Instruction *, int> PossibleRedIdx;
338 DenseMap<Instruction *, int> PossibleRedIter;
339 DenseSet<int> Reds;
340 };
341
342 // A DAGRootSet models an induction variable being used in a rerollable
343 // loop. For example,
344 //
345 // x[i*3+0] = y1
346 // x[i*3+1] = y2
347 // x[i*3+2] = y3
348 //
349 // Base instruction -> i*3
350 // +---+----+
351 // / | \
352 // ST[y1] +1 +2 <-- Roots
353 // | |
354 // ST[y2] ST[y3]
355 //
356 // There may be multiple DAGRoots, for example:
357 //
358 // x[i*2+0] = ... (1)
359 // x[i*2+1] = ... (1)
360 // x[i*2+4] = ... (2)
361 // x[i*2+5] = ... (2)
362 // x[(i+1234)*2+5678] = ... (3)
363 // x[(i+1234)*2+5679] = ... (3)
364 //
365 // The loop will be rerolled by adding a new loop induction variable,
366 // one for the Base instruction in each DAGRootSet.
367 //
368 struct DAGRootSet {
369 Instruction *BaseInst;
370 SmallInstructionVector Roots;
371
372 // The instructions between IV and BaseInst (but not including BaseInst).
373 SmallInstructionSet SubsumedInsts;
374 };
375
376 // The set of all DAG roots, and state tracking of all roots
377 // for a particular induction variable.
378 struct DAGRootTracker {
DAGRootTracker__anon39f464390111::LoopReroll::DAGRootTracker379 DAGRootTracker(LoopReroll *Parent, Loop *L, Instruction *IV,
380 ScalarEvolution *SE, AliasAnalysis *AA,
381 TargetLibraryInfo *TLI, DominatorTree *DT, LoopInfo *LI,
382 bool PreserveLCSSA,
383 DenseMap<Instruction *, int64_t> &IncrMap,
384 Instruction *LoopCtrlIV)
385 : Parent(Parent), L(L), SE(SE), AA(AA), TLI(TLI), DT(DT), LI(LI),
386 PreserveLCSSA(PreserveLCSSA), IV(IV), IVToIncMap(IncrMap),
387 LoopControlIV(LoopCtrlIV) {}
388
389 /// Stage 1: Find all the DAG roots for the induction variable.
390 bool findRoots();
391
392 /// Stage 2: Validate if the found roots are valid.
393 bool validate(ReductionTracker &Reductions);
394
395 /// Stage 3: Assuming validate() returned true, perform the
396 /// replacement.
397 /// @param BackedgeTakenCount The backedge-taken count of L.
398 void replace(const SCEV *BackedgeTakenCount);
399
400 protected:
401 using UsesTy = MapVector<Instruction *, BitVector>;
402
403 void findRootsRecursive(Instruction *IVU,
404 SmallInstructionSet SubsumedInsts);
405 bool findRootsBase(Instruction *IVU, SmallInstructionSet SubsumedInsts);
406 bool collectPossibleRoots(Instruction *Base,
407 std::map<int64_t,Instruction*> &Roots);
408 bool validateRootSet(DAGRootSet &DRS);
409
410 bool collectUsedInstructions(SmallInstructionSet &PossibleRedSet);
411 void collectInLoopUserSet(const SmallInstructionVector &Roots,
412 const SmallInstructionSet &Exclude,
413 const SmallInstructionSet &Final,
414 DenseSet<Instruction *> &Users);
415 void collectInLoopUserSet(Instruction *Root,
416 const SmallInstructionSet &Exclude,
417 const SmallInstructionSet &Final,
418 DenseSet<Instruction *> &Users);
419
420 UsesTy::iterator nextInstr(int Val, UsesTy &In,
421 const SmallInstructionSet &Exclude,
422 UsesTy::iterator *StartI=nullptr);
423 bool isBaseInst(Instruction *I);
424 bool isRootInst(Instruction *I);
425 bool instrDependsOn(Instruction *I,
426 UsesTy::iterator Start,
427 UsesTy::iterator End);
428 void replaceIV(DAGRootSet &DRS, const SCEV *Start, const SCEV *IncrExpr);
429
430 LoopReroll *Parent;
431
432 // Members of Parent, replicated here for brevity.
433 Loop *L;
434 ScalarEvolution *SE;
435 AliasAnalysis *AA;
436 TargetLibraryInfo *TLI;
437 DominatorTree *DT;
438 LoopInfo *LI;
439 bool PreserveLCSSA;
440
441 // The loop induction variable.
442 Instruction *IV;
443
444 // Loop step amount.
445 int64_t Inc;
446
447 // Loop reroll count; if Inc == 1, this records the scaling applied
448 // to the indvar: a[i*2+0] = ...; a[i*2+1] = ... ;
449 // If Inc is not 1, Scale = Inc.
450 uint64_t Scale;
451
452 // The roots themselves.
453 SmallVector<DAGRootSet,16> RootSets;
454
455 // All increment instructions for IV.
456 SmallInstructionVector LoopIncs;
457
458 // Map of all instructions in the loop (in order) to the iterations
459 // they are used in (or specially, IL_All for instructions
460 // used in the loop increment mechanism).
461 UsesTy Uses;
462
463 // Map between induction variable and its increment
464 DenseMap<Instruction *, int64_t> &IVToIncMap;
465
466 Instruction *LoopControlIV;
467 };
468
469 // Check if it is a compare-like instruction whose user is a branch
isCompareUsedByBranch(Instruction * I)470 bool isCompareUsedByBranch(Instruction *I) {
471 auto *TI = I->getParent()->getTerminator();
472 if (!isa<BranchInst>(TI) || !isa<CmpInst>(I))
473 return false;
474 return I->hasOneUse() && TI->getOperand(0) == I;
475 };
476
477 bool isLoopControlIV(Loop *L, Instruction *IV);
478 void collectPossibleIVs(Loop *L, SmallInstructionVector &PossibleIVs);
479 void collectPossibleReductions(Loop *L,
480 ReductionTracker &Reductions);
481 bool reroll(Instruction *IV, Loop *L, BasicBlock *Header,
482 const SCEV *BackedgeTakenCount, ReductionTracker &Reductions);
483 };
484
485 } // end anonymous namespace
486
487 char LoopReroll::ID = 0;
488
489 INITIALIZE_PASS_BEGIN(LoopReroll, "loop-reroll", "Reroll loops", false, false)
INITIALIZE_PASS_DEPENDENCY(LoopPass)490 INITIALIZE_PASS_DEPENDENCY(LoopPass)
491 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
492 INITIALIZE_PASS_END(LoopReroll, "loop-reroll", "Reroll loops", false, false)
493
494 Pass *llvm::createLoopRerollPass() {
495 return new LoopReroll;
496 }
497
498 // Returns true if the provided instruction is used outside the given loop.
499 // This operates like Instruction::isUsedOutsideOfBlock, but considers PHIs in
500 // non-loop blocks to be outside the loop.
hasUsesOutsideLoop(Instruction * I,Loop * L)501 static bool hasUsesOutsideLoop(Instruction *I, Loop *L) {
502 for (User *U : I->users()) {
503 if (!L->contains(cast<Instruction>(U)))
504 return true;
505 }
506 return false;
507 }
508
509 // Check if an IV is only used to control the loop. There are two cases:
510 // 1. It only has one use which is loop increment, and the increment is only
511 // used by comparison and the PHI (could has sext with nsw in between), and the
512 // comparison is only used by branch.
513 // 2. It is used by loop increment and the comparison, the loop increment is
514 // only used by the PHI, and the comparison is used only by the branch.
isLoopControlIV(Loop * L,Instruction * IV)515 bool LoopReroll::isLoopControlIV(Loop *L, Instruction *IV) {
516 unsigned IVUses = IV->getNumUses();
517 if (IVUses != 2 && IVUses != 1)
518 return false;
519
520 for (auto *User : IV->users()) {
521 int32_t IncOrCmpUses = User->getNumUses();
522 bool IsCompInst = isCompareUsedByBranch(cast<Instruction>(User));
523
524 // User can only have one or two uses.
525 if (IncOrCmpUses != 2 && IncOrCmpUses != 1)
526 return false;
527
528 // Case 1
529 if (IVUses == 1) {
530 // The only user must be the loop increment.
531 // The loop increment must have two uses.
532 if (IsCompInst || IncOrCmpUses != 2)
533 return false;
534 }
535
536 // Case 2
537 if (IVUses == 2 && IncOrCmpUses != 1)
538 return false;
539
540 // The users of the IV must be a binary operation or a comparison
541 if (auto *BO = dyn_cast<BinaryOperator>(User)) {
542 if (BO->getOpcode() == Instruction::Add) {
543 // Loop Increment
544 // User of Loop Increment should be either PHI or CMP
545 for (auto *UU : User->users()) {
546 if (PHINode *PN = dyn_cast<PHINode>(UU)) {
547 if (PN != IV)
548 return false;
549 }
550 // Must be a CMP or an ext (of a value with nsw) then CMP
551 else {
552 Instruction *UUser = dyn_cast<Instruction>(UU);
553 // Skip SExt if we are extending an nsw value
554 // TODO: Allow ZExt too
555 if (BO->hasNoSignedWrap() && UUser && UUser->hasOneUse() &&
556 isa<SExtInst>(UUser))
557 UUser = dyn_cast<Instruction>(*(UUser->user_begin()));
558 if (!isCompareUsedByBranch(UUser))
559 return false;
560 }
561 }
562 } else
563 return false;
564 // Compare : can only have one use, and must be branch
565 } else if (!IsCompInst)
566 return false;
567 }
568 return true;
569 }
570
571 // Collect the list of loop induction variables with respect to which it might
572 // be possible to reroll the loop.
collectPossibleIVs(Loop * L,SmallInstructionVector & PossibleIVs)573 void LoopReroll::collectPossibleIVs(Loop *L,
574 SmallInstructionVector &PossibleIVs) {
575 BasicBlock *Header = L->getHeader();
576 for (BasicBlock::iterator I = Header->begin(),
577 IE = Header->getFirstInsertionPt(); I != IE; ++I) {
578 if (!isa<PHINode>(I))
579 continue;
580 if (!I->getType()->isIntegerTy() && !I->getType()->isPointerTy())
581 continue;
582
583 if (const SCEVAddRecExpr *PHISCEV =
584 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(&*I))) {
585 if (PHISCEV->getLoop() != L)
586 continue;
587 if (!PHISCEV->isAffine())
588 continue;
589 auto IncSCEV = dyn_cast<SCEVConstant>(PHISCEV->getStepRecurrence(*SE));
590 if (IncSCEV) {
591 IVToIncMap[&*I] = IncSCEV->getValue()->getSExtValue();
592 LLVM_DEBUG(dbgs() << "LRR: Possible IV: " << *I << " = " << *PHISCEV
593 << "\n");
594
595 if (isLoopControlIV(L, &*I)) {
596 assert(!LoopControlIV && "Found two loop control only IV");
597 LoopControlIV = &(*I);
598 LLVM_DEBUG(dbgs() << "LRR: Possible loop control only IV: " << *I
599 << " = " << *PHISCEV << "\n");
600 } else
601 PossibleIVs.push_back(&*I);
602 }
603 }
604 }
605 }
606
607 // Add the remainder of the reduction-variable chain to the instruction vector
608 // (the initial PHINode has already been added). If successful, the object is
609 // marked as valid.
add(Loop * L)610 void LoopReroll::SimpleLoopReduction::add(Loop *L) {
611 assert(!Valid && "Cannot add to an already-valid chain");
612
613 // The reduction variable must be a chain of single-use instructions
614 // (including the PHI), except for the last value (which is used by the PHI
615 // and also outside the loop).
616 Instruction *C = Instructions.front();
617 if (C->user_empty())
618 return;
619
620 do {
621 C = cast<Instruction>(*C->user_begin());
622 if (C->hasOneUse()) {
623 if (!C->isBinaryOp())
624 return;
625
626 if (!(isa<PHINode>(Instructions.back()) ||
627 C->isSameOperationAs(Instructions.back())))
628 return;
629
630 Instructions.push_back(C);
631 }
632 } while (C->hasOneUse());
633
634 if (Instructions.size() < 2 ||
635 !C->isSameOperationAs(Instructions.back()) ||
636 C->use_empty())
637 return;
638
639 // C is now the (potential) last instruction in the reduction chain.
640 for (User *U : C->users()) {
641 // The only in-loop user can be the initial PHI.
642 if (L->contains(cast<Instruction>(U)))
643 if (cast<Instruction>(U) != Instructions.front())
644 return;
645 }
646
647 Instructions.push_back(C);
648 Valid = true;
649 }
650
651 // Collect the vector of possible reduction variables.
collectPossibleReductions(Loop * L,ReductionTracker & Reductions)652 void LoopReroll::collectPossibleReductions(Loop *L,
653 ReductionTracker &Reductions) {
654 BasicBlock *Header = L->getHeader();
655 for (BasicBlock::iterator I = Header->begin(),
656 IE = Header->getFirstInsertionPt(); I != IE; ++I) {
657 if (!isa<PHINode>(I))
658 continue;
659 if (!I->getType()->isSingleValueType())
660 continue;
661
662 SimpleLoopReduction SLR(&*I, L);
663 if (!SLR.valid())
664 continue;
665
666 LLVM_DEBUG(dbgs() << "LRR: Possible reduction: " << *I << " (with "
667 << SLR.size() << " chained instructions)\n");
668 Reductions.addSLR(SLR);
669 }
670 }
671
672 // Collect the set of all users of the provided root instruction. This set of
673 // users contains not only the direct users of the root instruction, but also
674 // all users of those users, and so on. There are two exceptions:
675 //
676 // 1. Instructions in the set of excluded instructions are never added to the
677 // use set (even if they are users). This is used, for example, to exclude
678 // including root increments in the use set of the primary IV.
679 //
680 // 2. Instructions in the set of final instructions are added to the use set
681 // if they are users, but their users are not added. This is used, for
682 // example, to prevent a reduction update from forcing all later reduction
683 // updates into the use set.
collectInLoopUserSet(Instruction * Root,const SmallInstructionSet & Exclude,const SmallInstructionSet & Final,DenseSet<Instruction * > & Users)684 void LoopReroll::DAGRootTracker::collectInLoopUserSet(
685 Instruction *Root, const SmallInstructionSet &Exclude,
686 const SmallInstructionSet &Final,
687 DenseSet<Instruction *> &Users) {
688 SmallInstructionVector Queue(1, Root);
689 while (!Queue.empty()) {
690 Instruction *I = Queue.pop_back_val();
691 if (!Users.insert(I).second)
692 continue;
693
694 if (!Final.count(I))
695 for (Use &U : I->uses()) {
696 Instruction *User = cast<Instruction>(U.getUser());
697 if (PHINode *PN = dyn_cast<PHINode>(User)) {
698 // Ignore "wrap-around" uses to PHIs of this loop's header.
699 if (PN->getIncomingBlock(U) == L->getHeader())
700 continue;
701 }
702
703 if (L->contains(User) && !Exclude.count(User)) {
704 Queue.push_back(User);
705 }
706 }
707
708 // We also want to collect single-user "feeder" values.
709 for (User::op_iterator OI = I->op_begin(),
710 OIE = I->op_end(); OI != OIE; ++OI) {
711 if (Instruction *Op = dyn_cast<Instruction>(*OI))
712 if (Op->hasOneUse() && L->contains(Op) && !Exclude.count(Op) &&
713 !Final.count(Op))
714 Queue.push_back(Op);
715 }
716 }
717 }
718
719 // Collect all of the users of all of the provided root instructions (combined
720 // into a single set).
collectInLoopUserSet(const SmallInstructionVector & Roots,const SmallInstructionSet & Exclude,const SmallInstructionSet & Final,DenseSet<Instruction * > & Users)721 void LoopReroll::DAGRootTracker::collectInLoopUserSet(
722 const SmallInstructionVector &Roots,
723 const SmallInstructionSet &Exclude,
724 const SmallInstructionSet &Final,
725 DenseSet<Instruction *> &Users) {
726 for (Instruction *Root : Roots)
727 collectInLoopUserSet(Root, Exclude, Final, Users);
728 }
729
isUnorderedLoadStore(Instruction * I)730 static bool isUnorderedLoadStore(Instruction *I) {
731 if (LoadInst *LI = dyn_cast<LoadInst>(I))
732 return LI->isUnordered();
733 if (StoreInst *SI = dyn_cast<StoreInst>(I))
734 return SI->isUnordered();
735 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I))
736 return !MI->isVolatile();
737 return false;
738 }
739
740 /// Return true if IVU is a "simple" arithmetic operation.
741 /// This is used for narrowing the search space for DAGRoots; only arithmetic
742 /// and GEPs can be part of a DAGRoot.
isSimpleArithmeticOp(User * IVU)743 static bool isSimpleArithmeticOp(User *IVU) {
744 if (Instruction *I = dyn_cast<Instruction>(IVU)) {
745 switch (I->getOpcode()) {
746 default: return false;
747 case Instruction::Add:
748 case Instruction::Sub:
749 case Instruction::Mul:
750 case Instruction::Shl:
751 case Instruction::AShr:
752 case Instruction::LShr:
753 case Instruction::GetElementPtr:
754 case Instruction::Trunc:
755 case Instruction::ZExt:
756 case Instruction::SExt:
757 return true;
758 }
759 }
760 return false;
761 }
762
isLoopIncrement(User * U,Instruction * IV)763 static bool isLoopIncrement(User *U, Instruction *IV) {
764 BinaryOperator *BO = dyn_cast<BinaryOperator>(U);
765
766 if ((BO && BO->getOpcode() != Instruction::Add) ||
767 (!BO && !isa<GetElementPtrInst>(U)))
768 return false;
769
770 for (auto *UU : U->users()) {
771 PHINode *PN = dyn_cast<PHINode>(UU);
772 if (PN && PN == IV)
773 return true;
774 }
775 return false;
776 }
777
778 bool LoopReroll::DAGRootTracker::
collectPossibleRoots(Instruction * Base,std::map<int64_t,Instruction * > & Roots)779 collectPossibleRoots(Instruction *Base, std::map<int64_t,Instruction*> &Roots) {
780 SmallInstructionVector BaseUsers;
781
782 for (auto *I : Base->users()) {
783 ConstantInt *CI = nullptr;
784
785 if (isLoopIncrement(I, IV)) {
786 LoopIncs.push_back(cast<Instruction>(I));
787 continue;
788 }
789
790 // The root nodes must be either GEPs, ORs or ADDs.
791 if (auto *BO = dyn_cast<BinaryOperator>(I)) {
792 if (BO->getOpcode() == Instruction::Add ||
793 BO->getOpcode() == Instruction::Or)
794 CI = dyn_cast<ConstantInt>(BO->getOperand(1));
795 } else if (auto *GEP = dyn_cast<GetElementPtrInst>(I)) {
796 Value *LastOperand = GEP->getOperand(GEP->getNumOperands()-1);
797 CI = dyn_cast<ConstantInt>(LastOperand);
798 }
799
800 if (!CI) {
801 if (Instruction *II = dyn_cast<Instruction>(I)) {
802 BaseUsers.push_back(II);
803 continue;
804 } else {
805 LLVM_DEBUG(dbgs() << "LRR: Aborting due to non-instruction: " << *I
806 << "\n");
807 return false;
808 }
809 }
810
811 int64_t V = std::abs(CI->getValue().getSExtValue());
812 if (Roots.find(V) != Roots.end())
813 // No duplicates, please.
814 return false;
815
816 Roots[V] = cast<Instruction>(I);
817 }
818
819 // Make sure we have at least two roots.
820 if (Roots.empty() || (Roots.size() == 1 && BaseUsers.empty()))
821 return false;
822
823 // If we found non-loop-inc, non-root users of Base, assume they are
824 // for the zeroth root index. This is because "add %a, 0" gets optimized
825 // away.
826 if (BaseUsers.size()) {
827 if (Roots.find(0) != Roots.end()) {
828 LLVM_DEBUG(dbgs() << "LRR: Multiple roots found for base - aborting!\n");
829 return false;
830 }
831 Roots[0] = Base;
832 }
833
834 // Calculate the number of users of the base, or lowest indexed, iteration.
835 unsigned NumBaseUses = BaseUsers.size();
836 if (NumBaseUses == 0)
837 NumBaseUses = Roots.begin()->second->getNumUses();
838
839 // Check that every node has the same number of users.
840 for (auto &KV : Roots) {
841 if (KV.first == 0)
842 continue;
843 if (!KV.second->hasNUses(NumBaseUses)) {
844 LLVM_DEBUG(dbgs() << "LRR: Aborting - Root and Base #users not the same: "
845 << "#Base=" << NumBaseUses
846 << ", #Root=" << KV.second->getNumUses() << "\n");
847 return false;
848 }
849 }
850
851 return true;
852 }
853
854 void LoopReroll::DAGRootTracker::
findRootsRecursive(Instruction * I,SmallInstructionSet SubsumedInsts)855 findRootsRecursive(Instruction *I, SmallInstructionSet SubsumedInsts) {
856 // Does the user look like it could be part of a root set?
857 // All its users must be simple arithmetic ops.
858 if (I->hasNUsesOrMore(IL_MaxRerollIterations + 1))
859 return;
860
861 if (I != IV && findRootsBase(I, SubsumedInsts))
862 return;
863
864 SubsumedInsts.insert(I);
865
866 for (User *V : I->users()) {
867 Instruction *I = cast<Instruction>(V);
868 if (is_contained(LoopIncs, I))
869 continue;
870
871 if (!isSimpleArithmeticOp(I))
872 continue;
873
874 // The recursive call makes a copy of SubsumedInsts.
875 findRootsRecursive(I, SubsumedInsts);
876 }
877 }
878
validateRootSet(DAGRootSet & DRS)879 bool LoopReroll::DAGRootTracker::validateRootSet(DAGRootSet &DRS) {
880 if (DRS.Roots.empty())
881 return false;
882
883 // Consider a DAGRootSet with N-1 roots (so N different values including
884 // BaseInst).
885 // Define d = Roots[0] - BaseInst, which should be the same as
886 // Roots[I] - Roots[I-1] for all I in [1..N).
887 // Define D = BaseInst@J - BaseInst@J-1, where "@J" means the value at the
888 // loop iteration J.
889 //
890 // Now, For the loop iterations to be consecutive:
891 // D = d * N
892 const auto *ADR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(DRS.BaseInst));
893 if (!ADR)
894 return false;
895 unsigned N = DRS.Roots.size() + 1;
896 const SCEV *StepSCEV = SE->getMinusSCEV(SE->getSCEV(DRS.Roots[0]), ADR);
897 const SCEV *ScaleSCEV = SE->getConstant(StepSCEV->getType(), N);
898 if (ADR->getStepRecurrence(*SE) != SE->getMulExpr(StepSCEV, ScaleSCEV))
899 return false;
900
901 return true;
902 }
903
904 bool LoopReroll::DAGRootTracker::
findRootsBase(Instruction * IVU,SmallInstructionSet SubsumedInsts)905 findRootsBase(Instruction *IVU, SmallInstructionSet SubsumedInsts) {
906 // The base of a RootSet must be an AddRec, so it can be erased.
907 const auto *IVU_ADR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IVU));
908 if (!IVU_ADR || IVU_ADR->getLoop() != L)
909 return false;
910
911 std::map<int64_t, Instruction*> V;
912 if (!collectPossibleRoots(IVU, V))
913 return false;
914
915 // If we didn't get a root for index zero, then IVU must be
916 // subsumed.
917 if (V.find(0) == V.end())
918 SubsumedInsts.insert(IVU);
919
920 // Partition the vector into monotonically increasing indexes.
921 DAGRootSet DRS;
922 DRS.BaseInst = nullptr;
923
924 SmallVector<DAGRootSet, 16> PotentialRootSets;
925
926 for (auto &KV : V) {
927 if (!DRS.BaseInst) {
928 DRS.BaseInst = KV.second;
929 DRS.SubsumedInsts = SubsumedInsts;
930 } else if (DRS.Roots.empty()) {
931 DRS.Roots.push_back(KV.second);
932 } else if (V.find(KV.first - 1) != V.end()) {
933 DRS.Roots.push_back(KV.second);
934 } else {
935 // Linear sequence terminated.
936 if (!validateRootSet(DRS))
937 return false;
938
939 // Construct a new DAGRootSet with the next sequence.
940 PotentialRootSets.push_back(DRS);
941 DRS.BaseInst = KV.second;
942 DRS.Roots.clear();
943 }
944 }
945
946 if (!validateRootSet(DRS))
947 return false;
948
949 PotentialRootSets.push_back(DRS);
950
951 RootSets.append(PotentialRootSets.begin(), PotentialRootSets.end());
952
953 return true;
954 }
955
findRoots()956 bool LoopReroll::DAGRootTracker::findRoots() {
957 Inc = IVToIncMap[IV];
958
959 assert(RootSets.empty() && "Unclean state!");
960 if (std::abs(Inc) == 1) {
961 for (auto *IVU : IV->users()) {
962 if (isLoopIncrement(IVU, IV))
963 LoopIncs.push_back(cast<Instruction>(IVU));
964 }
965 findRootsRecursive(IV, SmallInstructionSet());
966 LoopIncs.push_back(IV);
967 } else {
968 if (!findRootsBase(IV, SmallInstructionSet()))
969 return false;
970 }
971
972 // Ensure all sets have the same size.
973 if (RootSets.empty()) {
974 LLVM_DEBUG(dbgs() << "LRR: Aborting because no root sets found!\n");
975 return false;
976 }
977 for (auto &V : RootSets) {
978 if (V.Roots.empty() || V.Roots.size() != RootSets[0].Roots.size()) {
979 LLVM_DEBUG(
980 dbgs()
981 << "LRR: Aborting because not all root sets have the same size\n");
982 return false;
983 }
984 }
985
986 Scale = RootSets[0].Roots.size() + 1;
987
988 if (Scale > IL_MaxRerollIterations) {
989 LLVM_DEBUG(dbgs() << "LRR: Aborting - too many iterations found. "
990 << "#Found=" << Scale
991 << ", #Max=" << IL_MaxRerollIterations << "\n");
992 return false;
993 }
994
995 LLVM_DEBUG(dbgs() << "LRR: Successfully found roots: Scale=" << Scale
996 << "\n");
997
998 return true;
999 }
1000
collectUsedInstructions(SmallInstructionSet & PossibleRedSet)1001 bool LoopReroll::DAGRootTracker::collectUsedInstructions(SmallInstructionSet &PossibleRedSet) {
1002 // Populate the MapVector with all instructions in the block, in order first,
1003 // so we can iterate over the contents later in perfect order.
1004 for (auto &I : *L->getHeader()) {
1005 Uses[&I].resize(IL_End);
1006 }
1007
1008 SmallInstructionSet Exclude;
1009 for (auto &DRS : RootSets) {
1010 Exclude.insert(DRS.Roots.begin(), DRS.Roots.end());
1011 Exclude.insert(DRS.SubsumedInsts.begin(), DRS.SubsumedInsts.end());
1012 Exclude.insert(DRS.BaseInst);
1013 }
1014 Exclude.insert(LoopIncs.begin(), LoopIncs.end());
1015
1016 for (auto &DRS : RootSets) {
1017 DenseSet<Instruction*> VBase;
1018 collectInLoopUserSet(DRS.BaseInst, Exclude, PossibleRedSet, VBase);
1019 for (auto *I : VBase) {
1020 Uses[I].set(0);
1021 }
1022
1023 unsigned Idx = 1;
1024 for (auto *Root : DRS.Roots) {
1025 DenseSet<Instruction*> V;
1026 collectInLoopUserSet(Root, Exclude, PossibleRedSet, V);
1027
1028 // While we're here, check the use sets are the same size.
1029 if (V.size() != VBase.size()) {
1030 LLVM_DEBUG(dbgs() << "LRR: Aborting - use sets are different sizes\n");
1031 return false;
1032 }
1033
1034 for (auto *I : V) {
1035 Uses[I].set(Idx);
1036 }
1037 ++Idx;
1038 }
1039
1040 // Make sure our subsumed instructions are remembered too.
1041 for (auto *I : DRS.SubsumedInsts) {
1042 Uses[I].set(IL_All);
1043 }
1044 }
1045
1046 // Make sure the loop increments are also accounted for.
1047
1048 Exclude.clear();
1049 for (auto &DRS : RootSets) {
1050 Exclude.insert(DRS.Roots.begin(), DRS.Roots.end());
1051 Exclude.insert(DRS.SubsumedInsts.begin(), DRS.SubsumedInsts.end());
1052 Exclude.insert(DRS.BaseInst);
1053 }
1054
1055 DenseSet<Instruction*> V;
1056 collectInLoopUserSet(LoopIncs, Exclude, PossibleRedSet, V);
1057 for (auto *I : V) {
1058 Uses[I].set(IL_All);
1059 }
1060
1061 return true;
1062 }
1063
1064 /// Get the next instruction in "In" that is a member of set Val.
1065 /// Start searching from StartI, and do not return anything in Exclude.
1066 /// If StartI is not given, start from In.begin().
1067 LoopReroll::DAGRootTracker::UsesTy::iterator
nextInstr(int Val,UsesTy & In,const SmallInstructionSet & Exclude,UsesTy::iterator * StartI)1068 LoopReroll::DAGRootTracker::nextInstr(int Val, UsesTy &In,
1069 const SmallInstructionSet &Exclude,
1070 UsesTy::iterator *StartI) {
1071 UsesTy::iterator I = StartI ? *StartI : In.begin();
1072 while (I != In.end() && (I->second.test(Val) == 0 ||
1073 Exclude.count(I->first) != 0))
1074 ++I;
1075 return I;
1076 }
1077
isBaseInst(Instruction * I)1078 bool LoopReroll::DAGRootTracker::isBaseInst(Instruction *I) {
1079 for (auto &DRS : RootSets) {
1080 if (DRS.BaseInst == I)
1081 return true;
1082 }
1083 return false;
1084 }
1085
isRootInst(Instruction * I)1086 bool LoopReroll::DAGRootTracker::isRootInst(Instruction *I) {
1087 for (auto &DRS : RootSets) {
1088 if (is_contained(DRS.Roots, I))
1089 return true;
1090 }
1091 return false;
1092 }
1093
1094 /// Return true if instruction I depends on any instruction between
1095 /// Start and End.
instrDependsOn(Instruction * I,UsesTy::iterator Start,UsesTy::iterator End)1096 bool LoopReroll::DAGRootTracker::instrDependsOn(Instruction *I,
1097 UsesTy::iterator Start,
1098 UsesTy::iterator End) {
1099 for (auto *U : I->users()) {
1100 for (auto It = Start; It != End; ++It)
1101 if (U == It->first)
1102 return true;
1103 }
1104 return false;
1105 }
1106
isIgnorableInst(const Instruction * I)1107 static bool isIgnorableInst(const Instruction *I) {
1108 if (isa<DbgInfoIntrinsic>(I))
1109 return true;
1110 const IntrinsicInst* II = dyn_cast<IntrinsicInst>(I);
1111 if (!II)
1112 return false;
1113 switch (II->getIntrinsicID()) {
1114 default:
1115 return false;
1116 case Intrinsic::annotation:
1117 case Intrinsic::ptr_annotation:
1118 case Intrinsic::var_annotation:
1119 // TODO: the following intrinsics may also be whitelisted:
1120 // lifetime_start, lifetime_end, invariant_start, invariant_end
1121 return true;
1122 }
1123 return false;
1124 }
1125
validate(ReductionTracker & Reductions)1126 bool LoopReroll::DAGRootTracker::validate(ReductionTracker &Reductions) {
1127 // We now need to check for equivalence of the use graph of each root with
1128 // that of the primary induction variable (excluding the roots). Our goal
1129 // here is not to solve the full graph isomorphism problem, but rather to
1130 // catch common cases without a lot of work. As a result, we will assume
1131 // that the relative order of the instructions in each unrolled iteration
1132 // is the same (although we will not make an assumption about how the
1133 // different iterations are intermixed). Note that while the order must be
1134 // the same, the instructions may not be in the same basic block.
1135
1136 // An array of just the possible reductions for this scale factor. When we
1137 // collect the set of all users of some root instructions, these reduction
1138 // instructions are treated as 'final' (their uses are not considered).
1139 // This is important because we don't want the root use set to search down
1140 // the reduction chain.
1141 SmallInstructionSet PossibleRedSet;
1142 SmallInstructionSet PossibleRedLastSet;
1143 SmallInstructionSet PossibleRedPHISet;
1144 Reductions.restrictToScale(Scale, PossibleRedSet,
1145 PossibleRedPHISet, PossibleRedLastSet);
1146
1147 // Populate "Uses" with where each instruction is used.
1148 if (!collectUsedInstructions(PossibleRedSet))
1149 return false;
1150
1151 // Make sure we mark the reduction PHIs as used in all iterations.
1152 for (auto *I : PossibleRedPHISet) {
1153 Uses[I].set(IL_All);
1154 }
1155
1156 // Make sure we mark loop-control-only PHIs as used in all iterations. See
1157 // comment above LoopReroll::isLoopControlIV for more information.
1158 BasicBlock *Header = L->getHeader();
1159 if (LoopControlIV && LoopControlIV != IV) {
1160 for (auto *U : LoopControlIV->users()) {
1161 Instruction *IVUser = dyn_cast<Instruction>(U);
1162 // IVUser could be loop increment or compare
1163 Uses[IVUser].set(IL_All);
1164 for (auto *UU : IVUser->users()) {
1165 Instruction *UUser = dyn_cast<Instruction>(UU);
1166 // UUser could be compare, PHI or branch
1167 Uses[UUser].set(IL_All);
1168 // Skip SExt
1169 if (isa<SExtInst>(UUser)) {
1170 UUser = dyn_cast<Instruction>(*(UUser->user_begin()));
1171 Uses[UUser].set(IL_All);
1172 }
1173 // Is UUser a compare instruction?
1174 if (UU->hasOneUse()) {
1175 Instruction *BI = dyn_cast<BranchInst>(*UUser->user_begin());
1176 if (BI == cast<BranchInst>(Header->getTerminator()))
1177 Uses[BI].set(IL_All);
1178 }
1179 }
1180 }
1181 }
1182
1183 // Make sure all instructions in the loop are in one and only one
1184 // set.
1185 for (auto &KV : Uses) {
1186 if (KV.second.count() != 1 && !isIgnorableInst(KV.first)) {
1187 LLVM_DEBUG(
1188 dbgs() << "LRR: Aborting - instruction is not used in 1 iteration: "
1189 << *KV.first << " (#uses=" << KV.second.count() << ")\n");
1190 return false;
1191 }
1192 }
1193
1194 LLVM_DEBUG(for (auto &KV
1195 : Uses) {
1196 dbgs() << "LRR: " << KV.second.find_first() << "\t" << *KV.first << "\n";
1197 });
1198
1199 for (unsigned Iter = 1; Iter < Scale; ++Iter) {
1200 // In addition to regular aliasing information, we need to look for
1201 // instructions from later (future) iterations that have side effects
1202 // preventing us from reordering them past other instructions with side
1203 // effects.
1204 bool FutureSideEffects = false;
1205 AliasSetTracker AST(*AA);
1206 // The map between instructions in f(%iv.(i+1)) and f(%iv).
1207 DenseMap<Value *, Value *> BaseMap;
1208
1209 // Compare iteration Iter to the base.
1210 SmallInstructionSet Visited;
1211 auto BaseIt = nextInstr(0, Uses, Visited);
1212 auto RootIt = nextInstr(Iter, Uses, Visited);
1213 auto LastRootIt = Uses.begin();
1214
1215 while (BaseIt != Uses.end() && RootIt != Uses.end()) {
1216 Instruction *BaseInst = BaseIt->first;
1217 Instruction *RootInst = RootIt->first;
1218
1219 // Skip over the IV or root instructions; only match their users.
1220 bool Continue = false;
1221 if (isBaseInst(BaseInst)) {
1222 Visited.insert(BaseInst);
1223 BaseIt = nextInstr(0, Uses, Visited);
1224 Continue = true;
1225 }
1226 if (isRootInst(RootInst)) {
1227 LastRootIt = RootIt;
1228 Visited.insert(RootInst);
1229 RootIt = nextInstr(Iter, Uses, Visited);
1230 Continue = true;
1231 }
1232 if (Continue) continue;
1233
1234 if (!BaseInst->isSameOperationAs(RootInst)) {
1235 // Last chance saloon. We don't try and solve the full isomorphism
1236 // problem, but try and at least catch the case where two instructions
1237 // *of different types* are round the wrong way. We won't be able to
1238 // efficiently tell, given two ADD instructions, which way around we
1239 // should match them, but given an ADD and a SUB, we can at least infer
1240 // which one is which.
1241 //
1242 // This should allow us to deal with a greater subset of the isomorphism
1243 // problem. It does however change a linear algorithm into a quadratic
1244 // one, so limit the number of probes we do.
1245 auto TryIt = RootIt;
1246 unsigned N = NumToleratedFailedMatches;
1247 while (TryIt != Uses.end() &&
1248 !BaseInst->isSameOperationAs(TryIt->first) &&
1249 N--) {
1250 ++TryIt;
1251 TryIt = nextInstr(Iter, Uses, Visited, &TryIt);
1252 }
1253
1254 if (TryIt == Uses.end() || TryIt == RootIt ||
1255 instrDependsOn(TryIt->first, RootIt, TryIt)) {
1256 LLVM_DEBUG(dbgs() << "LRR: iteration root match failed at "
1257 << *BaseInst << " vs. " << *RootInst << "\n");
1258 return false;
1259 }
1260
1261 RootIt = TryIt;
1262 RootInst = TryIt->first;
1263 }
1264
1265 // All instructions between the last root and this root
1266 // may belong to some other iteration. If they belong to a
1267 // future iteration, then they're dangerous to alias with.
1268 //
1269 // Note that because we allow a limited amount of flexibility in the order
1270 // that we visit nodes, LastRootIt might be *before* RootIt, in which
1271 // case we've already checked this set of instructions so we shouldn't
1272 // do anything.
1273 for (; LastRootIt < RootIt; ++LastRootIt) {
1274 Instruction *I = LastRootIt->first;
1275 if (LastRootIt->second.find_first() < (int)Iter)
1276 continue;
1277 if (I->mayWriteToMemory())
1278 AST.add(I);
1279 // Note: This is specifically guarded by a check on isa<PHINode>,
1280 // which while a valid (somewhat arbitrary) micro-optimization, is
1281 // needed because otherwise isSafeToSpeculativelyExecute returns
1282 // false on PHI nodes.
1283 if (!isa<PHINode>(I) && !isUnorderedLoadStore(I) &&
1284 !isSafeToSpeculativelyExecute(I))
1285 // Intervening instructions cause side effects.
1286 FutureSideEffects = true;
1287 }
1288
1289 // Make sure that this instruction, which is in the use set of this
1290 // root instruction, does not also belong to the base set or the set of
1291 // some other root instruction.
1292 if (RootIt->second.count() > 1) {
1293 LLVM_DEBUG(dbgs() << "LRR: iteration root match failed at " << *BaseInst
1294 << " vs. " << *RootInst << " (prev. case overlap)\n");
1295 return false;
1296 }
1297
1298 // Make sure that we don't alias with any instruction in the alias set
1299 // tracker. If we do, then we depend on a future iteration, and we
1300 // can't reroll.
1301 if (RootInst->mayReadFromMemory())
1302 for (auto &K : AST) {
1303 if (K.aliasesUnknownInst(RootInst, *AA)) {
1304 LLVM_DEBUG(dbgs() << "LRR: iteration root match failed at "
1305 << *BaseInst << " vs. " << *RootInst
1306 << " (depends on future store)\n");
1307 return false;
1308 }
1309 }
1310
1311 // If we've past an instruction from a future iteration that may have
1312 // side effects, and this instruction might also, then we can't reorder
1313 // them, and this matching fails. As an exception, we allow the alias
1314 // set tracker to handle regular (unordered) load/store dependencies.
1315 if (FutureSideEffects && ((!isUnorderedLoadStore(BaseInst) &&
1316 !isSafeToSpeculativelyExecute(BaseInst)) ||
1317 (!isUnorderedLoadStore(RootInst) &&
1318 !isSafeToSpeculativelyExecute(RootInst)))) {
1319 LLVM_DEBUG(dbgs() << "LRR: iteration root match failed at " << *BaseInst
1320 << " vs. " << *RootInst
1321 << " (side effects prevent reordering)\n");
1322 return false;
1323 }
1324
1325 // For instructions that are part of a reduction, if the operation is
1326 // associative, then don't bother matching the operands (because we
1327 // already know that the instructions are isomorphic, and the order
1328 // within the iteration does not matter). For non-associative reductions,
1329 // we do need to match the operands, because we need to reject
1330 // out-of-order instructions within an iteration!
1331 // For example (assume floating-point addition), we need to reject this:
1332 // x += a[i]; x += b[i];
1333 // x += a[i+1]; x += b[i+1];
1334 // x += b[i+2]; x += a[i+2];
1335 bool InReduction = Reductions.isPairInSame(BaseInst, RootInst);
1336
1337 if (!(InReduction && BaseInst->isAssociative())) {
1338 bool Swapped = false, SomeOpMatched = false;
1339 for (unsigned j = 0; j < BaseInst->getNumOperands(); ++j) {
1340 Value *Op2 = RootInst->getOperand(j);
1341
1342 // If this is part of a reduction (and the operation is not
1343 // associatve), then we match all operands, but not those that are
1344 // part of the reduction.
1345 if (InReduction)
1346 if (Instruction *Op2I = dyn_cast<Instruction>(Op2))
1347 if (Reductions.isPairInSame(RootInst, Op2I))
1348 continue;
1349
1350 DenseMap<Value *, Value *>::iterator BMI = BaseMap.find(Op2);
1351 if (BMI != BaseMap.end()) {
1352 Op2 = BMI->second;
1353 } else {
1354 for (auto &DRS : RootSets) {
1355 if (DRS.Roots[Iter-1] == (Instruction*) Op2) {
1356 Op2 = DRS.BaseInst;
1357 break;
1358 }
1359 }
1360 }
1361
1362 if (BaseInst->getOperand(Swapped ? unsigned(!j) : j) != Op2) {
1363 // If we've not already decided to swap the matched operands, and
1364 // we've not already matched our first operand (note that we could
1365 // have skipped matching the first operand because it is part of a
1366 // reduction above), and the instruction is commutative, then try
1367 // the swapped match.
1368 if (!Swapped && BaseInst->isCommutative() && !SomeOpMatched &&
1369 BaseInst->getOperand(!j) == Op2) {
1370 Swapped = true;
1371 } else {
1372 LLVM_DEBUG(dbgs()
1373 << "LRR: iteration root match failed at " << *BaseInst
1374 << " vs. " << *RootInst << " (operand " << j << ")\n");
1375 return false;
1376 }
1377 }
1378
1379 SomeOpMatched = true;
1380 }
1381 }
1382
1383 if ((!PossibleRedLastSet.count(BaseInst) &&
1384 hasUsesOutsideLoop(BaseInst, L)) ||
1385 (!PossibleRedLastSet.count(RootInst) &&
1386 hasUsesOutsideLoop(RootInst, L))) {
1387 LLVM_DEBUG(dbgs() << "LRR: iteration root match failed at " << *BaseInst
1388 << " vs. " << *RootInst << " (uses outside loop)\n");
1389 return false;
1390 }
1391
1392 Reductions.recordPair(BaseInst, RootInst, Iter);
1393 BaseMap.insert(std::make_pair(RootInst, BaseInst));
1394
1395 LastRootIt = RootIt;
1396 Visited.insert(BaseInst);
1397 Visited.insert(RootInst);
1398 BaseIt = nextInstr(0, Uses, Visited);
1399 RootIt = nextInstr(Iter, Uses, Visited);
1400 }
1401 assert(BaseIt == Uses.end() && RootIt == Uses.end() &&
1402 "Mismatched set sizes!");
1403 }
1404
1405 LLVM_DEBUG(dbgs() << "LRR: Matched all iteration increments for " << *IV
1406 << "\n");
1407
1408 return true;
1409 }
1410
replace(const SCEV * BackedgeTakenCount)1411 void LoopReroll::DAGRootTracker::replace(const SCEV *BackedgeTakenCount) {
1412 BasicBlock *Header = L->getHeader();
1413
1414 // Compute the start and increment for each BaseInst before we start erasing
1415 // instructions.
1416 SmallVector<const SCEV *, 8> StartExprs;
1417 SmallVector<const SCEV *, 8> IncrExprs;
1418 for (auto &DRS : RootSets) {
1419 const SCEVAddRecExpr *IVSCEV =
1420 cast<SCEVAddRecExpr>(SE->getSCEV(DRS.BaseInst));
1421 StartExprs.push_back(IVSCEV->getStart());
1422 IncrExprs.push_back(SE->getMinusSCEV(SE->getSCEV(DRS.Roots[0]), IVSCEV));
1423 }
1424
1425 // Remove instructions associated with non-base iterations.
1426 for (BasicBlock::reverse_iterator J = Header->rbegin(), JE = Header->rend();
1427 J != JE;) {
1428 unsigned I = Uses[&*J].find_first();
1429 if (I > 0 && I < IL_All) {
1430 LLVM_DEBUG(dbgs() << "LRR: removing: " << *J << "\n");
1431 J++->eraseFromParent();
1432 continue;
1433 }
1434
1435 ++J;
1436 }
1437
1438 // Rewrite each BaseInst using SCEV.
1439 for (size_t i = 0, e = RootSets.size(); i != e; ++i)
1440 // Insert the new induction variable.
1441 replaceIV(RootSets[i], StartExprs[i], IncrExprs[i]);
1442
1443 { // Limit the lifetime of SCEVExpander.
1444 BranchInst *BI = cast<BranchInst>(Header->getTerminator());
1445 const DataLayout &DL = Header->getModule()->getDataLayout();
1446 SCEVExpander Expander(*SE, DL, "reroll");
1447 auto Zero = SE->getZero(BackedgeTakenCount->getType());
1448 auto One = SE->getOne(BackedgeTakenCount->getType());
1449 auto NewIVSCEV = SE->getAddRecExpr(Zero, One, L, SCEV::FlagAnyWrap);
1450 Value *NewIV =
1451 Expander.expandCodeFor(NewIVSCEV, BackedgeTakenCount->getType(),
1452 Header->getFirstNonPHIOrDbg());
1453 // FIXME: This arithmetic can overflow.
1454 auto TripCount = SE->getAddExpr(BackedgeTakenCount, One);
1455 auto ScaledTripCount = SE->getMulExpr(
1456 TripCount, SE->getConstant(BackedgeTakenCount->getType(), Scale));
1457 auto ScaledBECount = SE->getMinusSCEV(ScaledTripCount, One);
1458 Value *TakenCount =
1459 Expander.expandCodeFor(ScaledBECount, BackedgeTakenCount->getType(),
1460 Header->getFirstNonPHIOrDbg());
1461 Value *Cond =
1462 new ICmpInst(BI, CmpInst::ICMP_EQ, NewIV, TakenCount, "exitcond");
1463 BI->setCondition(Cond);
1464
1465 if (BI->getSuccessor(1) != Header)
1466 BI->swapSuccessors();
1467 }
1468
1469 SimplifyInstructionsInBlock(Header, TLI);
1470 DeleteDeadPHIs(Header, TLI);
1471 }
1472
replaceIV(DAGRootSet & DRS,const SCEV * Start,const SCEV * IncrExpr)1473 void LoopReroll::DAGRootTracker::replaceIV(DAGRootSet &DRS,
1474 const SCEV *Start,
1475 const SCEV *IncrExpr) {
1476 BasicBlock *Header = L->getHeader();
1477 Instruction *Inst = DRS.BaseInst;
1478
1479 const SCEV *NewIVSCEV =
1480 SE->getAddRecExpr(Start, IncrExpr, L, SCEV::FlagAnyWrap);
1481
1482 { // Limit the lifetime of SCEVExpander.
1483 const DataLayout &DL = Header->getModule()->getDataLayout();
1484 SCEVExpander Expander(*SE, DL, "reroll");
1485 Value *NewIV = Expander.expandCodeFor(NewIVSCEV, Inst->getType(),
1486 Header->getFirstNonPHIOrDbg());
1487
1488 for (auto &KV : Uses)
1489 if (KV.second.find_first() == 0)
1490 KV.first->replaceUsesOfWith(Inst, NewIV);
1491 }
1492 }
1493
1494 // Validate the selected reductions. All iterations must have an isomorphic
1495 // part of the reduction chain and, for non-associative reductions, the chain
1496 // entries must appear in order.
validateSelected()1497 bool LoopReroll::ReductionTracker::validateSelected() {
1498 // For a non-associative reduction, the chain entries must appear in order.
1499 for (int i : Reds) {
1500 int PrevIter = 0, BaseCount = 0, Count = 0;
1501 for (Instruction *J : PossibleReds[i]) {
1502 // Note that all instructions in the chain must have been found because
1503 // all instructions in the function must have been assigned to some
1504 // iteration.
1505 int Iter = PossibleRedIter[J];
1506 if (Iter != PrevIter && Iter != PrevIter + 1 &&
1507 !PossibleReds[i].getReducedValue()->isAssociative()) {
1508 LLVM_DEBUG(dbgs() << "LRR: Out-of-order non-associative reduction: "
1509 << J << "\n");
1510 return false;
1511 }
1512
1513 if (Iter != PrevIter) {
1514 if (Count != BaseCount) {
1515 LLVM_DEBUG(dbgs()
1516 << "LRR: Iteration " << PrevIter << " reduction use count "
1517 << Count << " is not equal to the base use count "
1518 << BaseCount << "\n");
1519 return false;
1520 }
1521
1522 Count = 0;
1523 }
1524
1525 ++Count;
1526 if (Iter == 0)
1527 ++BaseCount;
1528
1529 PrevIter = Iter;
1530 }
1531 }
1532
1533 return true;
1534 }
1535
1536 // For all selected reductions, remove all parts except those in the first
1537 // iteration (and the PHI). Replace outside uses of the reduced value with uses
1538 // of the first-iteration reduced value (in other words, reroll the selected
1539 // reductions).
replaceSelected()1540 void LoopReroll::ReductionTracker::replaceSelected() {
1541 // Fixup reductions to refer to the last instruction associated with the
1542 // first iteration (not the last).
1543 for (int i : Reds) {
1544 int j = 0;
1545 for (int e = PossibleReds[i].size(); j != e; ++j)
1546 if (PossibleRedIter[PossibleReds[i][j]] != 0) {
1547 --j;
1548 break;
1549 }
1550
1551 // Replace users with the new end-of-chain value.
1552 SmallInstructionVector Users;
1553 for (User *U : PossibleReds[i].getReducedValue()->users()) {
1554 Users.push_back(cast<Instruction>(U));
1555 }
1556
1557 for (Instruction *User : Users)
1558 User->replaceUsesOfWith(PossibleReds[i].getReducedValue(),
1559 PossibleReds[i][j]);
1560 }
1561 }
1562
1563 // Reroll the provided loop with respect to the provided induction variable.
1564 // Generally, we're looking for a loop like this:
1565 //
1566 // %iv = phi [ (preheader, ...), (body, %iv.next) ]
1567 // f(%iv)
1568 // %iv.1 = add %iv, 1 <-- a root increment
1569 // f(%iv.1)
1570 // %iv.2 = add %iv, 2 <-- a root increment
1571 // f(%iv.2)
1572 // %iv.scale_m_1 = add %iv, scale-1 <-- a root increment
1573 // f(%iv.scale_m_1)
1574 // ...
1575 // %iv.next = add %iv, scale
1576 // %cmp = icmp(%iv, ...)
1577 // br %cmp, header, exit
1578 //
1579 // Notably, we do not require that f(%iv), f(%iv.1), etc. be isolated groups of
1580 // instructions. In other words, the instructions in f(%iv), f(%iv.1), etc. can
1581 // be intermixed with eachother. The restriction imposed by this algorithm is
1582 // that the relative order of the isomorphic instructions in f(%iv), f(%iv.1),
1583 // etc. be the same.
1584 //
1585 // First, we collect the use set of %iv, excluding the other increment roots.
1586 // This gives us f(%iv). Then we iterate over the loop instructions (scale-1)
1587 // times, having collected the use set of f(%iv.(i+1)), during which we:
1588 // - Ensure that the next unmatched instruction in f(%iv) is isomorphic to
1589 // the next unmatched instruction in f(%iv.(i+1)).
1590 // - Ensure that both matched instructions don't have any external users
1591 // (with the exception of last-in-chain reduction instructions).
1592 // - Track the (aliasing) write set, and other side effects, of all
1593 // instructions that belong to future iterations that come before the matched
1594 // instructions. If the matched instructions read from that write set, then
1595 // f(%iv) or f(%iv.(i+1)) has some dependency on instructions in
1596 // f(%iv.(j+1)) for some j > i, and we cannot reroll the loop. Similarly,
1597 // if any of these future instructions had side effects (could not be
1598 // speculatively executed), and so do the matched instructions, when we
1599 // cannot reorder those side-effect-producing instructions, and rerolling
1600 // fails.
1601 //
1602 // Finally, we make sure that all loop instructions are either loop increment
1603 // roots, belong to simple latch code, parts of validated reductions, part of
1604 // f(%iv) or part of some f(%iv.i). If all of that is true (and all reductions
1605 // have been validated), then we reroll the loop.
reroll(Instruction * IV,Loop * L,BasicBlock * Header,const SCEV * BackedgeTakenCount,ReductionTracker & Reductions)1606 bool LoopReroll::reroll(Instruction *IV, Loop *L, BasicBlock *Header,
1607 const SCEV *BackedgeTakenCount,
1608 ReductionTracker &Reductions) {
1609 DAGRootTracker DAGRoots(this, L, IV, SE, AA, TLI, DT, LI, PreserveLCSSA,
1610 IVToIncMap, LoopControlIV);
1611
1612 if (!DAGRoots.findRoots())
1613 return false;
1614 LLVM_DEBUG(dbgs() << "LRR: Found all root induction increments for: " << *IV
1615 << "\n");
1616
1617 if (!DAGRoots.validate(Reductions))
1618 return false;
1619 if (!Reductions.validateSelected())
1620 return false;
1621 // At this point, we've validated the rerolling, and we're committed to
1622 // making changes!
1623
1624 Reductions.replaceSelected();
1625 DAGRoots.replace(BackedgeTakenCount);
1626
1627 ++NumRerolledLoops;
1628 return true;
1629 }
1630
runOnLoop(Loop * L,LPPassManager & LPM)1631 bool LoopReroll::runOnLoop(Loop *L, LPPassManager &LPM) {
1632 if (skipLoop(L))
1633 return false;
1634
1635 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
1636 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1637 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1638 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
1639 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1640 PreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
1641
1642 BasicBlock *Header = L->getHeader();
1643 LLVM_DEBUG(dbgs() << "LRR: F[" << Header->getParent()->getName() << "] Loop %"
1644 << Header->getName() << " (" << L->getNumBlocks()
1645 << " block(s))\n");
1646
1647 // For now, we'll handle only single BB loops.
1648 if (L->getNumBlocks() > 1)
1649 return false;
1650
1651 if (!SE->hasLoopInvariantBackedgeTakenCount(L))
1652 return false;
1653
1654 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1655 LLVM_DEBUG(dbgs() << "\n Before Reroll:\n" << *(L->getHeader()) << "\n");
1656 LLVM_DEBUG(dbgs() << "LRR: backedge-taken count = " << *BackedgeTakenCount
1657 << "\n");
1658
1659 // First, we need to find the induction variable with respect to which we can
1660 // reroll (there may be several possible options).
1661 SmallInstructionVector PossibleIVs;
1662 IVToIncMap.clear();
1663 LoopControlIV = nullptr;
1664 collectPossibleIVs(L, PossibleIVs);
1665
1666 if (PossibleIVs.empty()) {
1667 LLVM_DEBUG(dbgs() << "LRR: No possible IVs found\n");
1668 return false;
1669 }
1670
1671 ReductionTracker Reductions;
1672 collectPossibleReductions(L, Reductions);
1673 bool Changed = false;
1674
1675 // For each possible IV, collect the associated possible set of 'root' nodes
1676 // (i+1, i+2, etc.).
1677 for (Instruction *PossibleIV : PossibleIVs)
1678 if (reroll(PossibleIV, L, Header, BackedgeTakenCount, Reductions)) {
1679 Changed = true;
1680 break;
1681 }
1682 LLVM_DEBUG(dbgs() << "\n After Reroll:\n" << *(L->getHeader()) << "\n");
1683
1684 // Trip count of L has changed so SE must be re-evaluated.
1685 if (Changed)
1686 SE->forgetLoop(L);
1687
1688 return Changed;
1689 }
1690