1 //===-- LoopPredication.cpp - Guard based loop predication 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 // The LoopPredication pass tries to convert loop variant range checks to loop
10 // invariant by widening checks across loop iterations. For example, it will
11 // convert
12 //
13 //   for (i = 0; i < n; i++) {
14 //     guard(i < len);
15 //     ...
16 //   }
17 //
18 // to
19 //
20 //   for (i = 0; i < n; i++) {
21 //     guard(n - 1 < len);
22 //     ...
23 //   }
24 //
25 // After this transformation the condition of the guard is loop invariant, so
26 // loop-unswitch can later unswitch the loop by this condition which basically
27 // predicates the loop by the widened condition:
28 //
29 //   if (n - 1 < len)
30 //     for (i = 0; i < n; i++) {
31 //       ...
32 //     }
33 //   else
34 //     deoptimize
35 //
36 // It's tempting to rely on SCEV here, but it has proven to be problematic.
37 // Generally the facts SCEV provides about the increment step of add
38 // recurrences are true if the backedge of the loop is taken, which implicitly
39 // assumes that the guard doesn't fail. Using these facts to optimize the
40 // guard results in a circular logic where the guard is optimized under the
41 // assumption that it never fails.
42 //
43 // For example, in the loop below the induction variable will be marked as nuw
44 // basing on the guard. Basing on nuw the guard predicate will be considered
45 // monotonic. Given a monotonic condition it's tempting to replace the induction
46 // variable in the condition with its value on the last iteration. But this
47 // transformation is not correct, e.g. e = 4, b = 5 breaks the loop.
48 //
49 //   for (int i = b; i != e; i++)
50 //     guard(i u< len)
51 //
52 // One of the ways to reason about this problem is to use an inductive proof
53 // approach. Given the loop:
54 //
55 //   if (B(0)) {
56 //     do {
57 //       I = PHI(0, I.INC)
58 //       I.INC = I + Step
59 //       guard(G(I));
60 //     } while (B(I));
61 //   }
62 //
63 // where B(x) and G(x) are predicates that map integers to booleans, we want a
64 // loop invariant expression M such the following program has the same semantics
65 // as the above:
66 //
67 //   if (B(0)) {
68 //     do {
69 //       I = PHI(0, I.INC)
70 //       I.INC = I + Step
71 //       guard(G(0) && M);
72 //     } while (B(I));
73 //   }
74 //
75 // One solution for M is M = forall X . (G(X) && B(X)) => G(X + Step)
76 //
77 // Informal proof that the transformation above is correct:
78 //
79 //   By the definition of guards we can rewrite the guard condition to:
80 //     G(I) && G(0) && M
81 //
82 //   Let's prove that for each iteration of the loop:
83 //     G(0) && M => G(I)
84 //   And the condition above can be simplified to G(Start) && M.
85 //
86 //   Induction base.
87 //     G(0) && M => G(0)
88 //
89 //   Induction step. Assuming G(0) && M => G(I) on the subsequent
90 //   iteration:
91 //
92 //     B(I) is true because it's the backedge condition.
93 //     G(I) is true because the backedge is guarded by this condition.
94 //
95 //   So M = forall X . (G(X) && B(X)) => G(X + Step) implies G(I + Step).
96 //
97 // Note that we can use anything stronger than M, i.e. any condition which
98 // implies M.
99 //
100 // When S = 1 (i.e. forward iterating loop), the transformation is supported
101 // when:
102 //   * The loop has a single latch with the condition of the form:
103 //     B(X) = latchStart + X <pred> latchLimit,
104 //     where <pred> is u<, u<=, s<, or s<=.
105 //   * The guard condition is of the form
106 //     G(X) = guardStart + X u< guardLimit
107 //
108 //   For the ult latch comparison case M is:
109 //     forall X . guardStart + X u< guardLimit && latchStart + X <u latchLimit =>
110 //        guardStart + X + 1 u< guardLimit
111 //
112 //   The only way the antecedent can be true and the consequent can be false is
113 //   if
114 //     X == guardLimit - 1 - guardStart
115 //   (and guardLimit is non-zero, but we won't use this latter fact).
116 //   If X == guardLimit - 1 - guardStart then the second half of the antecedent is
117 //     latchStart + guardLimit - 1 - guardStart u< latchLimit
118 //   and its negation is
119 //     latchStart + guardLimit - 1 - guardStart u>= latchLimit
120 //
121 //   In other words, if
122 //     latchLimit u<= latchStart + guardLimit - 1 - guardStart
123 //   then:
124 //   (the ranges below are written in ConstantRange notation, where [A, B) is the
125 //   set for (I = A; I != B; I++ /*maywrap*/) yield(I);)
126 //
127 //      forall X . guardStart + X u< guardLimit &&
128 //                 latchStart + X u< latchLimit =>
129 //        guardStart + X + 1 u< guardLimit
130 //   == forall X . guardStart + X u< guardLimit &&
131 //                 latchStart + X u< latchStart + guardLimit - 1 - guardStart =>
132 //        guardStart + X + 1 u< guardLimit
133 //   == forall X . (guardStart + X) in [0, guardLimit) &&
134 //                 (latchStart + X) in [0, latchStart + guardLimit - 1 - guardStart) =>
135 //        (guardStart + X + 1) in [0, guardLimit)
136 //   == forall X . X in [-guardStart, guardLimit - guardStart) &&
137 //                 X in [-latchStart, guardLimit - 1 - guardStart) =>
138 //         X in [-guardStart - 1, guardLimit - guardStart - 1)
139 //   == true
140 //
141 //   So the widened condition is:
142 //     guardStart u< guardLimit &&
143 //     latchStart + guardLimit - 1 - guardStart u>= latchLimit
144 //   Similarly for ule condition the widened condition is:
145 //     guardStart u< guardLimit &&
146 //     latchStart + guardLimit - 1 - guardStart u> latchLimit
147 //   For slt condition the widened condition is:
148 //     guardStart u< guardLimit &&
149 //     latchStart + guardLimit - 1 - guardStart s>= latchLimit
150 //   For sle condition the widened condition is:
151 //     guardStart u< guardLimit &&
152 //     latchStart + guardLimit - 1 - guardStart s> latchLimit
153 //
154 // When S = -1 (i.e. reverse iterating loop), the transformation is supported
155 // when:
156 //   * The loop has a single latch with the condition of the form:
157 //     B(X) = X <pred> latchLimit, where <pred> is u>, u>=, s>, or s>=.
158 //   * The guard condition is of the form
159 //     G(X) = X - 1 u< guardLimit
160 //
161 //   For the ugt latch comparison case M is:
162 //     forall X. X-1 u< guardLimit and X u> latchLimit => X-2 u< guardLimit
163 //
164 //   The only way the antecedent can be true and the consequent can be false is if
165 //     X == 1.
166 //   If X == 1 then the second half of the antecedent is
167 //     1 u> latchLimit, and its negation is latchLimit u>= 1.
168 //
169 //   So the widened condition is:
170 //     guardStart u< guardLimit && latchLimit u>= 1.
171 //   Similarly for sgt condition the widened condition is:
172 //     guardStart u< guardLimit && latchLimit s>= 1.
173 //   For uge condition the widened condition is:
174 //     guardStart u< guardLimit && latchLimit u> 1.
175 //   For sge condition the widened condition is:
176 //     guardStart u< guardLimit && latchLimit s> 1.
177 //===----------------------------------------------------------------------===//
178 
179 #include "llvm/Transforms/Scalar/LoopPredication.h"
180 #include "llvm/ADT/Statistic.h"
181 #include "llvm/Analysis/AliasAnalysis.h"
182 #include "llvm/Analysis/BranchProbabilityInfo.h"
183 #include "llvm/Analysis/GuardUtils.h"
184 #include "llvm/Analysis/LoopInfo.h"
185 #include "llvm/Analysis/LoopPass.h"
186 #include "llvm/Analysis/ScalarEvolution.h"
187 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
188 #include "llvm/IR/Function.h"
189 #include "llvm/IR/GlobalValue.h"
190 #include "llvm/IR/IntrinsicInst.h"
191 #include "llvm/IR/Module.h"
192 #include "llvm/IR/PatternMatch.h"
193 #include "llvm/InitializePasses.h"
194 #include "llvm/Pass.h"
195 #include "llvm/Support/CommandLine.h"
196 #include "llvm/Support/Debug.h"
197 #include "llvm/Transforms/Scalar.h"
198 #include "llvm/Transforms/Utils/GuardUtils.h"
199 #include "llvm/Transforms/Utils/Local.h"
200 #include "llvm/Transforms/Utils/LoopUtils.h"
201 #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
202 
203 #define DEBUG_TYPE "loop-predication"
204 
205 STATISTIC(TotalConsidered, "Number of guards considered");
206 STATISTIC(TotalWidened, "Number of checks widened");
207 
208 using namespace llvm;
209 
210 static cl::opt<bool> EnableIVTruncation("loop-predication-enable-iv-truncation",
211                                         cl::Hidden, cl::init(true));
212 
213 static cl::opt<bool> EnableCountDownLoop("loop-predication-enable-count-down-loop",
214                                         cl::Hidden, cl::init(true));
215 
216 static cl::opt<bool>
217     SkipProfitabilityChecks("loop-predication-skip-profitability-checks",
218                             cl::Hidden, cl::init(false));
219 
220 // This is the scale factor for the latch probability. We use this during
221 // profitability analysis to find other exiting blocks that have a much higher
222 // probability of exiting the loop instead of loop exiting via latch.
223 // This value should be greater than 1 for a sane profitability check.
224 static cl::opt<float> LatchExitProbabilityScale(
225     "loop-predication-latch-probability-scale", cl::Hidden, cl::init(2.0),
226     cl::desc("scale factor for the latch probability. Value should be greater "
227              "than 1. Lower values are ignored"));
228 
229 static cl::opt<bool> PredicateWidenableBranchGuards(
230     "loop-predication-predicate-widenable-branches-to-deopt", cl::Hidden,
231     cl::desc("Whether or not we should predicate guards "
232              "expressed as widenable branches to deoptimize blocks"),
233     cl::init(true));
234 
235 namespace {
236 /// Represents an induction variable check:
237 ///   icmp Pred, <induction variable>, <loop invariant limit>
238 struct LoopICmp {
239   ICmpInst::Predicate Pred;
240   const SCEVAddRecExpr *IV;
241   const SCEV *Limit;
LoopICmp__anone9668e770111::LoopICmp242   LoopICmp(ICmpInst::Predicate Pred, const SCEVAddRecExpr *IV,
243            const SCEV *Limit)
244     : Pred(Pred), IV(IV), Limit(Limit) {}
LoopICmp__anone9668e770111::LoopICmp245   LoopICmp() {}
dump__anone9668e770111::LoopICmp246   void dump() {
247     dbgs() << "LoopICmp Pred = " << Pred << ", IV = " << *IV
248            << ", Limit = " << *Limit << "\n";
249   }
250 };
251 
252 class LoopPredication {
253   AliasAnalysis *AA;
254   DominatorTree *DT;
255   ScalarEvolution *SE;
256   LoopInfo *LI;
257   BranchProbabilityInfo *BPI;
258 
259   Loop *L;
260   const DataLayout *DL;
261   BasicBlock *Preheader;
262   LoopICmp LatchCheck;
263 
264   bool isSupportedStep(const SCEV* Step);
265   Optional<LoopICmp> parseLoopICmp(ICmpInst *ICI);
266   Optional<LoopICmp> parseLoopLatchICmp();
267 
268   /// Return an insertion point suitable for inserting a safe to speculate
269   /// instruction whose only user will be 'User' which has operands 'Ops'.  A
270   /// trivial result would be the at the User itself, but we try to return a
271   /// loop invariant location if possible.
272   Instruction *findInsertPt(Instruction *User, ArrayRef<Value*> Ops);
273   /// Same as above, *except* that this uses the SCEV definition of invariant
274   /// which is that an expression *can be made* invariant via SCEVExpander.
275   /// Thus, this version is only suitable for finding an insert point to be be
276   /// passed to SCEVExpander!
277   Instruction *findInsertPt(Instruction *User, ArrayRef<const SCEV*> Ops);
278 
279   /// Return true if the value is known to produce a single fixed value across
280   /// all iterations on which it executes.  Note that this does not imply
281   /// speculation safety.  That must be established separately.
282   bool isLoopInvariantValue(const SCEV* S);
283 
284   Value *expandCheck(SCEVExpander &Expander, Instruction *Guard,
285                      ICmpInst::Predicate Pred, const SCEV *LHS,
286                      const SCEV *RHS);
287 
288   Optional<Value *> widenICmpRangeCheck(ICmpInst *ICI, SCEVExpander &Expander,
289                                         Instruction *Guard);
290   Optional<Value *> widenICmpRangeCheckIncrementingLoop(LoopICmp LatchCheck,
291                                                         LoopICmp RangeCheck,
292                                                         SCEVExpander &Expander,
293                                                         Instruction *Guard);
294   Optional<Value *> widenICmpRangeCheckDecrementingLoop(LoopICmp LatchCheck,
295                                                         LoopICmp RangeCheck,
296                                                         SCEVExpander &Expander,
297                                                         Instruction *Guard);
298   unsigned collectChecks(SmallVectorImpl<Value *> &Checks, Value *Condition,
299                          SCEVExpander &Expander, Instruction *Guard);
300   bool widenGuardConditions(IntrinsicInst *II, SCEVExpander &Expander);
301   bool widenWidenableBranchGuardConditions(BranchInst *Guard, SCEVExpander &Expander);
302   // If the loop always exits through another block in the loop, we should not
303   // predicate based on the latch check. For example, the latch check can be a
304   // very coarse grained check and there can be more fine grained exit checks
305   // within the loop. We identify such unprofitable loops through BPI.
306   bool isLoopProfitableToPredicate();
307 
308   bool predicateLoopExits(Loop *L, SCEVExpander &Rewriter);
309 
310 public:
LoopPredication(AliasAnalysis * AA,DominatorTree * DT,ScalarEvolution * SE,LoopInfo * LI,BranchProbabilityInfo * BPI)311   LoopPredication(AliasAnalysis *AA, DominatorTree *DT,
312                   ScalarEvolution *SE, LoopInfo *LI,
313                   BranchProbabilityInfo *BPI)
314     : AA(AA), DT(DT), SE(SE), LI(LI), BPI(BPI) {};
315   bool runOnLoop(Loop *L);
316 };
317 
318 class LoopPredicationLegacyPass : public LoopPass {
319 public:
320   static char ID;
LoopPredicationLegacyPass()321   LoopPredicationLegacyPass() : LoopPass(ID) {
322     initializeLoopPredicationLegacyPassPass(*PassRegistry::getPassRegistry());
323   }
324 
getAnalysisUsage(AnalysisUsage & AU) const325   void getAnalysisUsage(AnalysisUsage &AU) const override {
326     AU.addRequired<BranchProbabilityInfoWrapperPass>();
327     getLoopAnalysisUsage(AU);
328   }
329 
runOnLoop(Loop * L,LPPassManager & LPM)330   bool runOnLoop(Loop *L, LPPassManager &LPM) override {
331     if (skipLoop(L))
332       return false;
333     auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
334     auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
335     auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
336     BranchProbabilityInfo &BPI =
337         getAnalysis<BranchProbabilityInfoWrapperPass>().getBPI();
338     auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
339     LoopPredication LP(AA, DT, SE, LI, &BPI);
340     return LP.runOnLoop(L);
341   }
342 };
343 
344 char LoopPredicationLegacyPass::ID = 0;
345 } // end namespace
346 
347 INITIALIZE_PASS_BEGIN(LoopPredicationLegacyPass, "loop-predication",
348                       "Loop predication", false, false)
INITIALIZE_PASS_DEPENDENCY(BranchProbabilityInfoWrapperPass)349 INITIALIZE_PASS_DEPENDENCY(BranchProbabilityInfoWrapperPass)
350 INITIALIZE_PASS_DEPENDENCY(LoopPass)
351 INITIALIZE_PASS_END(LoopPredicationLegacyPass, "loop-predication",
352                     "Loop predication", false, false)
353 
354 Pass *llvm::createLoopPredicationPass() {
355   return new LoopPredicationLegacyPass();
356 }
357 
run(Loop & L,LoopAnalysisManager & AM,LoopStandardAnalysisResults & AR,LPMUpdater & U)358 PreservedAnalyses LoopPredicationPass::run(Loop &L, LoopAnalysisManager &AM,
359                                            LoopStandardAnalysisResults &AR,
360                                            LPMUpdater &U) {
361   Function *F = L.getHeader()->getParent();
362   // For the new PM, we also can't use BranchProbabilityInfo as an analysis
363   // pass. Function analyses need to be preserved across loop transformations
364   // but BPI is not preserved, hence a newly built one is needed.
365   BranchProbabilityInfo BPI(*F, AR.LI, &AR.TLI, &AR.DT, nullptr);
366   LoopPredication LP(&AR.AA, &AR.DT, &AR.SE, &AR.LI, &BPI);
367   if (!LP.runOnLoop(&L))
368     return PreservedAnalyses::all();
369 
370   return getLoopPassPreservedAnalyses();
371 }
372 
373 Optional<LoopICmp>
parseLoopICmp(ICmpInst * ICI)374 LoopPredication::parseLoopICmp(ICmpInst *ICI) {
375   auto Pred = ICI->getPredicate();
376   auto *LHS = ICI->getOperand(0);
377   auto *RHS = ICI->getOperand(1);
378 
379   const SCEV *LHSS = SE->getSCEV(LHS);
380   if (isa<SCEVCouldNotCompute>(LHSS))
381     return None;
382   const SCEV *RHSS = SE->getSCEV(RHS);
383   if (isa<SCEVCouldNotCompute>(RHSS))
384     return None;
385 
386   // Canonicalize RHS to be loop invariant bound, LHS - a loop computable IV
387   if (SE->isLoopInvariant(LHSS, L)) {
388     std::swap(LHS, RHS);
389     std::swap(LHSS, RHSS);
390     Pred = ICmpInst::getSwappedPredicate(Pred);
391   }
392 
393   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHSS);
394   if (!AR || AR->getLoop() != L)
395     return None;
396 
397   return LoopICmp(Pred, AR, RHSS);
398 }
399 
expandCheck(SCEVExpander & Expander,Instruction * Guard,ICmpInst::Predicate Pred,const SCEV * LHS,const SCEV * RHS)400 Value *LoopPredication::expandCheck(SCEVExpander &Expander,
401                                     Instruction *Guard,
402                                     ICmpInst::Predicate Pred, const SCEV *LHS,
403                                     const SCEV *RHS) {
404   Type *Ty = LHS->getType();
405   assert(Ty == RHS->getType() && "expandCheck operands have different types?");
406 
407   if (SE->isLoopInvariant(LHS, L) && SE->isLoopInvariant(RHS, L)) {
408     IRBuilder<> Builder(Guard);
409     if (SE->isLoopEntryGuardedByCond(L, Pred, LHS, RHS))
410       return Builder.getTrue();
411     if (SE->isLoopEntryGuardedByCond(L, ICmpInst::getInversePredicate(Pred),
412                                      LHS, RHS))
413       return Builder.getFalse();
414   }
415 
416   Value *LHSV = Expander.expandCodeFor(LHS, Ty, findInsertPt(Guard, {LHS}));
417   Value *RHSV = Expander.expandCodeFor(RHS, Ty, findInsertPt(Guard, {RHS}));
418   IRBuilder<> Builder(findInsertPt(Guard, {LHSV, RHSV}));
419   return Builder.CreateICmp(Pred, LHSV, RHSV);
420 }
421 
422 
423 // Returns true if its safe to truncate the IV to RangeCheckType.
424 // When the IV type is wider than the range operand type, we can still do loop
425 // predication, by generating SCEVs for the range and latch that are of the
426 // same type. We achieve this by generating a SCEV truncate expression for the
427 // latch IV. This is done iff truncation of the IV is a safe operation,
428 // without loss of information.
429 // Another way to achieve this is by generating a wider type SCEV for the
430 // range check operand, however, this needs a more involved check that
431 // operands do not overflow. This can lead to loss of information when the
432 // range operand is of the form: add i32 %offset, %iv. We need to prove that
433 // sext(x + y) is same as sext(x) + sext(y).
434 // This function returns true if we can safely represent the IV type in
435 // the RangeCheckType without loss of information.
isSafeToTruncateWideIVType(const DataLayout & DL,ScalarEvolution & SE,const LoopICmp LatchCheck,Type * RangeCheckType)436 static bool isSafeToTruncateWideIVType(const DataLayout &DL,
437                                        ScalarEvolution &SE,
438                                        const LoopICmp LatchCheck,
439                                        Type *RangeCheckType) {
440   if (!EnableIVTruncation)
441     return false;
442   assert(DL.getTypeSizeInBits(LatchCheck.IV->getType()).getFixedSize() >
443              DL.getTypeSizeInBits(RangeCheckType).getFixedSize() &&
444          "Expected latch check IV type to be larger than range check operand "
445          "type!");
446   // The start and end values of the IV should be known. This is to guarantee
447   // that truncating the wide type will not lose information.
448   auto *Limit = dyn_cast<SCEVConstant>(LatchCheck.Limit);
449   auto *Start = dyn_cast<SCEVConstant>(LatchCheck.IV->getStart());
450   if (!Limit || !Start)
451     return false;
452   // This check makes sure that the IV does not change sign during loop
453   // iterations. Consider latchType = i64, LatchStart = 5, Pred = ICMP_SGE,
454   // LatchEnd = 2, rangeCheckType = i32. If it's not a monotonic predicate, the
455   // IV wraps around, and the truncation of the IV would lose the range of
456   // iterations between 2^32 and 2^64.
457   if (!SE.getMonotonicPredicateType(LatchCheck.IV, LatchCheck.Pred))
458     return false;
459   // The active bits should be less than the bits in the RangeCheckType. This
460   // guarantees that truncating the latch check to RangeCheckType is a safe
461   // operation.
462   auto RangeCheckTypeBitSize =
463       DL.getTypeSizeInBits(RangeCheckType).getFixedSize();
464   return Start->getAPInt().getActiveBits() < RangeCheckTypeBitSize &&
465          Limit->getAPInt().getActiveBits() < RangeCheckTypeBitSize;
466 }
467 
468 
469 // Return an LoopICmp describing a latch check equivlent to LatchCheck but with
470 // the requested type if safe to do so.  May involve the use of a new IV.
generateLoopLatchCheck(const DataLayout & DL,ScalarEvolution & SE,const LoopICmp LatchCheck,Type * RangeCheckType)471 static Optional<LoopICmp> generateLoopLatchCheck(const DataLayout &DL,
472                                                  ScalarEvolution &SE,
473                                                  const LoopICmp LatchCheck,
474                                                  Type *RangeCheckType) {
475 
476   auto *LatchType = LatchCheck.IV->getType();
477   if (RangeCheckType == LatchType)
478     return LatchCheck;
479   // For now, bail out if latch type is narrower than range type.
480   if (DL.getTypeSizeInBits(LatchType).getFixedSize() <
481       DL.getTypeSizeInBits(RangeCheckType).getFixedSize())
482     return None;
483   if (!isSafeToTruncateWideIVType(DL, SE, LatchCheck, RangeCheckType))
484     return None;
485   // We can now safely identify the truncated version of the IV and limit for
486   // RangeCheckType.
487   LoopICmp NewLatchCheck;
488   NewLatchCheck.Pred = LatchCheck.Pred;
489   NewLatchCheck.IV = dyn_cast<SCEVAddRecExpr>(
490       SE.getTruncateExpr(LatchCheck.IV, RangeCheckType));
491   if (!NewLatchCheck.IV)
492     return None;
493   NewLatchCheck.Limit = SE.getTruncateExpr(LatchCheck.Limit, RangeCheckType);
494   LLVM_DEBUG(dbgs() << "IV of type: " << *LatchType
495                     << "can be represented as range check type:"
496                     << *RangeCheckType << "\n");
497   LLVM_DEBUG(dbgs() << "LatchCheck.IV: " << *NewLatchCheck.IV << "\n");
498   LLVM_DEBUG(dbgs() << "LatchCheck.Limit: " << *NewLatchCheck.Limit << "\n");
499   return NewLatchCheck;
500 }
501 
isSupportedStep(const SCEV * Step)502 bool LoopPredication::isSupportedStep(const SCEV* Step) {
503   return Step->isOne() || (Step->isAllOnesValue() && EnableCountDownLoop);
504 }
505 
findInsertPt(Instruction * Use,ArrayRef<Value * > Ops)506 Instruction *LoopPredication::findInsertPt(Instruction *Use,
507                                            ArrayRef<Value*> Ops) {
508   for (Value *Op : Ops)
509     if (!L->isLoopInvariant(Op))
510       return Use;
511   return Preheader->getTerminator();
512 }
513 
findInsertPt(Instruction * Use,ArrayRef<const SCEV * > Ops)514 Instruction *LoopPredication::findInsertPt(Instruction *Use,
515                                            ArrayRef<const SCEV*> Ops) {
516   // Subtlety: SCEV considers things to be invariant if the value produced is
517   // the same across iterations.  This is not the same as being able to
518   // evaluate outside the loop, which is what we actually need here.
519   for (const SCEV *Op : Ops)
520     if (!SE->isLoopInvariant(Op, L) ||
521         !isSafeToExpandAt(Op, Preheader->getTerminator(), *SE))
522       return Use;
523   return Preheader->getTerminator();
524 }
525 
isLoopInvariantValue(const SCEV * S)526 bool LoopPredication::isLoopInvariantValue(const SCEV* S) {
527   // Handling expressions which produce invariant results, but *haven't* yet
528   // been removed from the loop serves two important purposes.
529   // 1) Most importantly, it resolves a pass ordering cycle which would
530   // otherwise need us to iteration licm, loop-predication, and either
531   // loop-unswitch or loop-peeling to make progress on examples with lots of
532   // predicable range checks in a row.  (Since, in the general case,  we can't
533   // hoist the length checks until the dominating checks have been discharged
534   // as we can't prove doing so is safe.)
535   // 2) As a nice side effect, this exposes the value of peeling or unswitching
536   // much more obviously in the IR.  Otherwise, the cost modeling for other
537   // transforms would end up needing to duplicate all of this logic to model a
538   // check which becomes predictable based on a modeled peel or unswitch.
539   //
540   // The cost of doing so in the worst case is an extra fill from the stack  in
541   // the loop to materialize the loop invariant test value instead of checking
542   // against the original IV which is presumable in a register inside the loop.
543   // Such cases are presumably rare, and hint at missing oppurtunities for
544   // other passes.
545 
546   if (SE->isLoopInvariant(S, L))
547     // Note: This the SCEV variant, so the original Value* may be within the
548     // loop even though SCEV has proven it is loop invariant.
549     return true;
550 
551   // Handle a particular important case which SCEV doesn't yet know about which
552   // shows up in range checks on arrays with immutable lengths.
553   // TODO: This should be sunk inside SCEV.
554   if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S))
555     if (const auto *LI = dyn_cast<LoadInst>(U->getValue()))
556       if (LI->isUnordered() && L->hasLoopInvariantOperands(LI))
557         if (AA->pointsToConstantMemory(LI->getOperand(0)) ||
558             LI->hasMetadata(LLVMContext::MD_invariant_load))
559           return true;
560   return false;
561 }
562 
widenICmpRangeCheckIncrementingLoop(LoopICmp LatchCheck,LoopICmp RangeCheck,SCEVExpander & Expander,Instruction * Guard)563 Optional<Value *> LoopPredication::widenICmpRangeCheckIncrementingLoop(
564     LoopICmp LatchCheck, LoopICmp RangeCheck,
565     SCEVExpander &Expander, Instruction *Guard) {
566   auto *Ty = RangeCheck.IV->getType();
567   // Generate the widened condition for the forward loop:
568   //   guardStart u< guardLimit &&
569   //   latchLimit <pred> guardLimit - 1 - guardStart + latchStart
570   // where <pred> depends on the latch condition predicate. See the file
571   // header comment for the reasoning.
572   // guardLimit - guardStart + latchStart - 1
573   const SCEV *GuardStart = RangeCheck.IV->getStart();
574   const SCEV *GuardLimit = RangeCheck.Limit;
575   const SCEV *LatchStart = LatchCheck.IV->getStart();
576   const SCEV *LatchLimit = LatchCheck.Limit;
577   // Subtlety: We need all the values to be *invariant* across all iterations,
578   // but we only need to check expansion safety for those which *aren't*
579   // already guaranteed to dominate the guard.
580   if (!isLoopInvariantValue(GuardStart) ||
581       !isLoopInvariantValue(GuardLimit) ||
582       !isLoopInvariantValue(LatchStart) ||
583       !isLoopInvariantValue(LatchLimit)) {
584     LLVM_DEBUG(dbgs() << "Can't expand limit check!\n");
585     return None;
586   }
587   if (!isSafeToExpandAt(LatchStart, Guard, *SE) ||
588       !isSafeToExpandAt(LatchLimit, Guard, *SE)) {
589     LLVM_DEBUG(dbgs() << "Can't expand limit check!\n");
590     return None;
591   }
592 
593   // guardLimit - guardStart + latchStart - 1
594   const SCEV *RHS =
595       SE->getAddExpr(SE->getMinusSCEV(GuardLimit, GuardStart),
596                      SE->getMinusSCEV(LatchStart, SE->getOne(Ty)));
597   auto LimitCheckPred =
598       ICmpInst::getFlippedStrictnessPredicate(LatchCheck.Pred);
599 
600   LLVM_DEBUG(dbgs() << "LHS: " << *LatchLimit << "\n");
601   LLVM_DEBUG(dbgs() << "RHS: " << *RHS << "\n");
602   LLVM_DEBUG(dbgs() << "Pred: " << LimitCheckPred << "\n");
603 
604   auto *LimitCheck =
605       expandCheck(Expander, Guard, LimitCheckPred, LatchLimit, RHS);
606   auto *FirstIterationCheck = expandCheck(Expander, Guard, RangeCheck.Pred,
607                                           GuardStart, GuardLimit);
608   IRBuilder<> Builder(findInsertPt(Guard, {FirstIterationCheck, LimitCheck}));
609   return Builder.CreateAnd(FirstIterationCheck, LimitCheck);
610 }
611 
widenICmpRangeCheckDecrementingLoop(LoopICmp LatchCheck,LoopICmp RangeCheck,SCEVExpander & Expander,Instruction * Guard)612 Optional<Value *> LoopPredication::widenICmpRangeCheckDecrementingLoop(
613     LoopICmp LatchCheck, LoopICmp RangeCheck,
614     SCEVExpander &Expander, Instruction *Guard) {
615   auto *Ty = RangeCheck.IV->getType();
616   const SCEV *GuardStart = RangeCheck.IV->getStart();
617   const SCEV *GuardLimit = RangeCheck.Limit;
618   const SCEV *LatchStart = LatchCheck.IV->getStart();
619   const SCEV *LatchLimit = LatchCheck.Limit;
620   // Subtlety: We need all the values to be *invariant* across all iterations,
621   // but we only need to check expansion safety for those which *aren't*
622   // already guaranteed to dominate the guard.
623   if (!isLoopInvariantValue(GuardStart) ||
624       !isLoopInvariantValue(GuardLimit) ||
625       !isLoopInvariantValue(LatchStart) ||
626       !isLoopInvariantValue(LatchLimit)) {
627     LLVM_DEBUG(dbgs() << "Can't expand limit check!\n");
628     return None;
629   }
630   if (!isSafeToExpandAt(LatchStart, Guard, *SE) ||
631       !isSafeToExpandAt(LatchLimit, Guard, *SE)) {
632     LLVM_DEBUG(dbgs() << "Can't expand limit check!\n");
633     return None;
634   }
635   // The decrement of the latch check IV should be the same as the
636   // rangeCheckIV.
637   auto *PostDecLatchCheckIV = LatchCheck.IV->getPostIncExpr(*SE);
638   if (RangeCheck.IV != PostDecLatchCheckIV) {
639     LLVM_DEBUG(dbgs() << "Not the same. PostDecLatchCheckIV: "
640                       << *PostDecLatchCheckIV
641                       << "  and RangeCheckIV: " << *RangeCheck.IV << "\n");
642     return None;
643   }
644 
645   // Generate the widened condition for CountDownLoop:
646   // guardStart u< guardLimit &&
647   // latchLimit <pred> 1.
648   // See the header comment for reasoning of the checks.
649   auto LimitCheckPred =
650       ICmpInst::getFlippedStrictnessPredicate(LatchCheck.Pred);
651   auto *FirstIterationCheck = expandCheck(Expander, Guard,
652                                           ICmpInst::ICMP_ULT,
653                                           GuardStart, GuardLimit);
654   auto *LimitCheck = expandCheck(Expander, Guard, LimitCheckPred, LatchLimit,
655                                  SE->getOne(Ty));
656   IRBuilder<> Builder(findInsertPt(Guard, {FirstIterationCheck, LimitCheck}));
657   return Builder.CreateAnd(FirstIterationCheck, LimitCheck);
658 }
659 
normalizePredicate(ScalarEvolution * SE,Loop * L,LoopICmp & RC)660 static void normalizePredicate(ScalarEvolution *SE, Loop *L,
661                                LoopICmp& RC) {
662   // LFTR canonicalizes checks to the ICMP_NE/EQ form; normalize back to the
663   // ULT/UGE form for ease of handling by our caller.
664   if (ICmpInst::isEquality(RC.Pred) &&
665       RC.IV->getStepRecurrence(*SE)->isOne() &&
666       SE->isKnownPredicate(ICmpInst::ICMP_ULE, RC.IV->getStart(), RC.Limit))
667     RC.Pred = RC.Pred == ICmpInst::ICMP_NE ?
668       ICmpInst::ICMP_ULT : ICmpInst::ICMP_UGE;
669 }
670 
671 
672 /// If ICI can be widened to a loop invariant condition emits the loop
673 /// invariant condition in the loop preheader and return it, otherwise
674 /// returns None.
widenICmpRangeCheck(ICmpInst * ICI,SCEVExpander & Expander,Instruction * Guard)675 Optional<Value *> LoopPredication::widenICmpRangeCheck(ICmpInst *ICI,
676                                                        SCEVExpander &Expander,
677                                                        Instruction *Guard) {
678   LLVM_DEBUG(dbgs() << "Analyzing ICmpInst condition:\n");
679   LLVM_DEBUG(ICI->dump());
680 
681   // parseLoopStructure guarantees that the latch condition is:
682   //   ++i <pred> latchLimit, where <pred> is u<, u<=, s<, or s<=.
683   // We are looking for the range checks of the form:
684   //   i u< guardLimit
685   auto RangeCheck = parseLoopICmp(ICI);
686   if (!RangeCheck) {
687     LLVM_DEBUG(dbgs() << "Failed to parse the loop latch condition!\n");
688     return None;
689   }
690   LLVM_DEBUG(dbgs() << "Guard check:\n");
691   LLVM_DEBUG(RangeCheck->dump());
692   if (RangeCheck->Pred != ICmpInst::ICMP_ULT) {
693     LLVM_DEBUG(dbgs() << "Unsupported range check predicate("
694                       << RangeCheck->Pred << ")!\n");
695     return None;
696   }
697   auto *RangeCheckIV = RangeCheck->IV;
698   if (!RangeCheckIV->isAffine()) {
699     LLVM_DEBUG(dbgs() << "Range check IV is not affine!\n");
700     return None;
701   }
702   auto *Step = RangeCheckIV->getStepRecurrence(*SE);
703   // We cannot just compare with latch IV step because the latch and range IVs
704   // may have different types.
705   if (!isSupportedStep(Step)) {
706     LLVM_DEBUG(dbgs() << "Range check and latch have IVs different steps!\n");
707     return None;
708   }
709   auto *Ty = RangeCheckIV->getType();
710   auto CurrLatchCheckOpt = generateLoopLatchCheck(*DL, *SE, LatchCheck, Ty);
711   if (!CurrLatchCheckOpt) {
712     LLVM_DEBUG(dbgs() << "Failed to generate a loop latch check "
713                          "corresponding to range type: "
714                       << *Ty << "\n");
715     return None;
716   }
717 
718   LoopICmp CurrLatchCheck = *CurrLatchCheckOpt;
719   // At this point, the range and latch step should have the same type, but need
720   // not have the same value (we support both 1 and -1 steps).
721   assert(Step->getType() ==
722              CurrLatchCheck.IV->getStepRecurrence(*SE)->getType() &&
723          "Range and latch steps should be of same type!");
724   if (Step != CurrLatchCheck.IV->getStepRecurrence(*SE)) {
725     LLVM_DEBUG(dbgs() << "Range and latch have different step values!\n");
726     return None;
727   }
728 
729   if (Step->isOne())
730     return widenICmpRangeCheckIncrementingLoop(CurrLatchCheck, *RangeCheck,
731                                                Expander, Guard);
732   else {
733     assert(Step->isAllOnesValue() && "Step should be -1!");
734     return widenICmpRangeCheckDecrementingLoop(CurrLatchCheck, *RangeCheck,
735                                                Expander, Guard);
736   }
737 }
738 
collectChecks(SmallVectorImpl<Value * > & Checks,Value * Condition,SCEVExpander & Expander,Instruction * Guard)739 unsigned LoopPredication::collectChecks(SmallVectorImpl<Value *> &Checks,
740                                         Value *Condition,
741                                         SCEVExpander &Expander,
742                                         Instruction *Guard) {
743   unsigned NumWidened = 0;
744   // The guard condition is expected to be in form of:
745   //   cond1 && cond2 && cond3 ...
746   // Iterate over subconditions looking for icmp conditions which can be
747   // widened across loop iterations. Widening these conditions remember the
748   // resulting list of subconditions in Checks vector.
749   SmallVector<Value *, 4> Worklist(1, Condition);
750   SmallPtrSet<Value *, 4> Visited;
751   Value *WideableCond = nullptr;
752   do {
753     Value *Condition = Worklist.pop_back_val();
754     if (!Visited.insert(Condition).second)
755       continue;
756 
757     Value *LHS, *RHS;
758     using namespace llvm::PatternMatch;
759     if (match(Condition, m_And(m_Value(LHS), m_Value(RHS)))) {
760       Worklist.push_back(LHS);
761       Worklist.push_back(RHS);
762       continue;
763     }
764 
765     if (match(Condition,
766               m_Intrinsic<Intrinsic::experimental_widenable_condition>())) {
767       // Pick any, we don't care which
768       WideableCond = Condition;
769       continue;
770     }
771 
772     if (ICmpInst *ICI = dyn_cast<ICmpInst>(Condition)) {
773       if (auto NewRangeCheck = widenICmpRangeCheck(ICI, Expander,
774                                                    Guard)) {
775         Checks.push_back(NewRangeCheck.getValue());
776         NumWidened++;
777         continue;
778       }
779     }
780 
781     // Save the condition as is if we can't widen it
782     Checks.push_back(Condition);
783   } while (!Worklist.empty());
784   // At the moment, our matching logic for wideable conditions implicitly
785   // assumes we preserve the form: (br (and Cond, WC())).  FIXME
786   // Note that if there were multiple calls to wideable condition in the
787   // traversal, we only need to keep one, and which one is arbitrary.
788   if (WideableCond)
789     Checks.push_back(WideableCond);
790   return NumWidened;
791 }
792 
widenGuardConditions(IntrinsicInst * Guard,SCEVExpander & Expander)793 bool LoopPredication::widenGuardConditions(IntrinsicInst *Guard,
794                                            SCEVExpander &Expander) {
795   LLVM_DEBUG(dbgs() << "Processing guard:\n");
796   LLVM_DEBUG(Guard->dump());
797 
798   TotalConsidered++;
799   SmallVector<Value *, 4> Checks;
800   unsigned NumWidened = collectChecks(Checks, Guard->getOperand(0), Expander,
801                                       Guard);
802   if (NumWidened == 0)
803     return false;
804 
805   TotalWidened += NumWidened;
806 
807   // Emit the new guard condition
808   IRBuilder<> Builder(findInsertPt(Guard, Checks));
809   Value *AllChecks = Builder.CreateAnd(Checks);
810   auto *OldCond = Guard->getOperand(0);
811   Guard->setOperand(0, AllChecks);
812   RecursivelyDeleteTriviallyDeadInstructions(OldCond);
813 
814   LLVM_DEBUG(dbgs() << "Widened checks = " << NumWidened << "\n");
815   return true;
816 }
817 
widenWidenableBranchGuardConditions(BranchInst * BI,SCEVExpander & Expander)818 bool LoopPredication::widenWidenableBranchGuardConditions(
819     BranchInst *BI, SCEVExpander &Expander) {
820   assert(isGuardAsWidenableBranch(BI) && "Must be!");
821   LLVM_DEBUG(dbgs() << "Processing guard:\n");
822   LLVM_DEBUG(BI->dump());
823 
824   TotalConsidered++;
825   SmallVector<Value *, 4> Checks;
826   unsigned NumWidened = collectChecks(Checks, BI->getCondition(),
827                                       Expander, BI);
828   if (NumWidened == 0)
829     return false;
830 
831   TotalWidened += NumWidened;
832 
833   // Emit the new guard condition
834   IRBuilder<> Builder(findInsertPt(BI, Checks));
835   Value *AllChecks = Builder.CreateAnd(Checks);
836   auto *OldCond = BI->getCondition();
837   BI->setCondition(AllChecks);
838   RecursivelyDeleteTriviallyDeadInstructions(OldCond);
839   assert(isGuardAsWidenableBranch(BI) &&
840          "Stopped being a guard after transform?");
841 
842   LLVM_DEBUG(dbgs() << "Widened checks = " << NumWidened << "\n");
843   return true;
844 }
845 
parseLoopLatchICmp()846 Optional<LoopICmp> LoopPredication::parseLoopLatchICmp() {
847   using namespace PatternMatch;
848 
849   BasicBlock *LoopLatch = L->getLoopLatch();
850   if (!LoopLatch) {
851     LLVM_DEBUG(dbgs() << "The loop doesn't have a single latch!\n");
852     return None;
853   }
854 
855   auto *BI = dyn_cast<BranchInst>(LoopLatch->getTerminator());
856   if (!BI || !BI->isConditional()) {
857     LLVM_DEBUG(dbgs() << "Failed to match the latch terminator!\n");
858     return None;
859   }
860   BasicBlock *TrueDest = BI->getSuccessor(0);
861   assert(
862       (TrueDest == L->getHeader() || BI->getSuccessor(1) == L->getHeader()) &&
863       "One of the latch's destinations must be the header");
864 
865   auto *ICI = dyn_cast<ICmpInst>(BI->getCondition());
866   if (!ICI) {
867     LLVM_DEBUG(dbgs() << "Failed to match the latch condition!\n");
868     return None;
869   }
870   auto Result = parseLoopICmp(ICI);
871   if (!Result) {
872     LLVM_DEBUG(dbgs() << "Failed to parse the loop latch condition!\n");
873     return None;
874   }
875 
876   if (TrueDest != L->getHeader())
877     Result->Pred = ICmpInst::getInversePredicate(Result->Pred);
878 
879   // Check affine first, so if it's not we don't try to compute the step
880   // recurrence.
881   if (!Result->IV->isAffine()) {
882     LLVM_DEBUG(dbgs() << "The induction variable is not affine!\n");
883     return None;
884   }
885 
886   auto *Step = Result->IV->getStepRecurrence(*SE);
887   if (!isSupportedStep(Step)) {
888     LLVM_DEBUG(dbgs() << "Unsupported loop stride(" << *Step << ")!\n");
889     return None;
890   }
891 
892   auto IsUnsupportedPredicate = [](const SCEV *Step, ICmpInst::Predicate Pred) {
893     if (Step->isOne()) {
894       return Pred != ICmpInst::ICMP_ULT && Pred != ICmpInst::ICMP_SLT &&
895              Pred != ICmpInst::ICMP_ULE && Pred != ICmpInst::ICMP_SLE;
896     } else {
897       assert(Step->isAllOnesValue() && "Step should be -1!");
898       return Pred != ICmpInst::ICMP_UGT && Pred != ICmpInst::ICMP_SGT &&
899              Pred != ICmpInst::ICMP_UGE && Pred != ICmpInst::ICMP_SGE;
900     }
901   };
902 
903   normalizePredicate(SE, L, *Result);
904   if (IsUnsupportedPredicate(Step, Result->Pred)) {
905     LLVM_DEBUG(dbgs() << "Unsupported loop latch predicate(" << Result->Pred
906                       << ")!\n");
907     return None;
908   }
909 
910   return Result;
911 }
912 
913 
isLoopProfitableToPredicate()914 bool LoopPredication::isLoopProfitableToPredicate() {
915   if (SkipProfitabilityChecks || !BPI)
916     return true;
917 
918   SmallVector<std::pair<BasicBlock *, BasicBlock *>, 8> ExitEdges;
919   L->getExitEdges(ExitEdges);
920   // If there is only one exiting edge in the loop, it is always profitable to
921   // predicate the loop.
922   if (ExitEdges.size() == 1)
923     return true;
924 
925   // Calculate the exiting probabilities of all exiting edges from the loop,
926   // starting with the LatchExitProbability.
927   // Heuristic for profitability: If any of the exiting blocks' probability of
928   // exiting the loop is larger than exiting through the latch block, it's not
929   // profitable to predicate the loop.
930   auto *LatchBlock = L->getLoopLatch();
931   assert(LatchBlock && "Should have a single latch at this point!");
932   auto *LatchTerm = LatchBlock->getTerminator();
933   assert(LatchTerm->getNumSuccessors() == 2 &&
934          "expected to be an exiting block with 2 succs!");
935   unsigned LatchBrExitIdx =
936       LatchTerm->getSuccessor(0) == L->getHeader() ? 1 : 0;
937   BranchProbability LatchExitProbability =
938       BPI->getEdgeProbability(LatchBlock, LatchBrExitIdx);
939 
940   // Protect against degenerate inputs provided by the user. Providing a value
941   // less than one, can invert the definition of profitable loop predication.
942   float ScaleFactor = LatchExitProbabilityScale;
943   if (ScaleFactor < 1) {
944     LLVM_DEBUG(
945         dbgs()
946         << "Ignored user setting for loop-predication-latch-probability-scale: "
947         << LatchExitProbabilityScale << "\n");
948     LLVM_DEBUG(dbgs() << "The value is set to 1.0\n");
949     ScaleFactor = 1.0;
950   }
951   const auto LatchProbabilityThreshold =
952       LatchExitProbability * ScaleFactor;
953 
954   for (const auto &ExitEdge : ExitEdges) {
955     BranchProbability ExitingBlockProbability =
956         BPI->getEdgeProbability(ExitEdge.first, ExitEdge.second);
957     // Some exiting edge has higher probability than the latch exiting edge.
958     // No longer profitable to predicate.
959     if (ExitingBlockProbability > LatchProbabilityThreshold)
960       return false;
961   }
962   // Using BPI, we have concluded that the most probable way to exit from the
963   // loop is through the latch (or there's no profile information and all
964   // exits are equally likely).
965   return true;
966 }
967 
968 /// If we can (cheaply) find a widenable branch which controls entry into the
969 /// loop, return it.
FindWidenableTerminatorAboveLoop(Loop * L,LoopInfo & LI)970 static BranchInst *FindWidenableTerminatorAboveLoop(Loop *L, LoopInfo &LI) {
971   // Walk back through any unconditional executed blocks and see if we can find
972   // a widenable condition which seems to control execution of this loop.  Note
973   // that we predict that maythrow calls are likely untaken and thus that it's
974   // profitable to widen a branch before a maythrow call with a condition
975   // afterwards even though that may cause the slow path to run in a case where
976   // it wouldn't have otherwise.
977   BasicBlock *BB = L->getLoopPreheader();
978   if (!BB)
979     return nullptr;
980   do {
981     if (BasicBlock *Pred = BB->getSinglePredecessor())
982       if (BB == Pred->getSingleSuccessor()) {
983         BB = Pred;
984         continue;
985       }
986     break;
987   } while (true);
988 
989   if (BasicBlock *Pred = BB->getSinglePredecessor()) {
990     auto *Term = Pred->getTerminator();
991 
992     Value *Cond, *WC;
993     BasicBlock *IfTrueBB, *IfFalseBB;
994     if (parseWidenableBranch(Term, Cond, WC, IfTrueBB, IfFalseBB) &&
995         IfTrueBB == BB)
996       return cast<BranchInst>(Term);
997   }
998   return nullptr;
999 }
1000 
1001 /// Return the minimum of all analyzeable exit counts.  This is an upper bound
1002 /// on the actual exit count.  If there are not at least two analyzeable exits,
1003 /// returns SCEVCouldNotCompute.
getMinAnalyzeableBackedgeTakenCount(ScalarEvolution & SE,DominatorTree & DT,Loop * L)1004 static const SCEV *getMinAnalyzeableBackedgeTakenCount(ScalarEvolution &SE,
1005                                                        DominatorTree &DT,
1006                                                        Loop *L) {
1007   SmallVector<BasicBlock *, 16> ExitingBlocks;
1008   L->getExitingBlocks(ExitingBlocks);
1009 
1010   SmallVector<const SCEV *, 4> ExitCounts;
1011   for (BasicBlock *ExitingBB : ExitingBlocks) {
1012     const SCEV *ExitCount = SE.getExitCount(L, ExitingBB);
1013     if (isa<SCEVCouldNotCompute>(ExitCount))
1014       continue;
1015     assert(DT.dominates(ExitingBB, L->getLoopLatch()) &&
1016            "We should only have known counts for exiting blocks that "
1017            "dominate latch!");
1018     ExitCounts.push_back(ExitCount);
1019   }
1020   if (ExitCounts.size() < 2)
1021     return SE.getCouldNotCompute();
1022   return SE.getUMinFromMismatchedTypes(ExitCounts);
1023 }
1024 
1025 /// This implements an analogous, but entirely distinct transform from the main
1026 /// loop predication transform.  This one is phrased in terms of using a
1027 /// widenable branch *outside* the loop to allow us to simplify loop exits in a
1028 /// following loop.  This is close in spirit to the IndVarSimplify transform
1029 /// of the same name, but is materially different widening loosens legality
1030 /// sharply.
predicateLoopExits(Loop * L,SCEVExpander & Rewriter)1031 bool LoopPredication::predicateLoopExits(Loop *L, SCEVExpander &Rewriter) {
1032   // The transformation performed here aims to widen a widenable condition
1033   // above the loop such that all analyzeable exit leading to deopt are dead.
1034   // It assumes that the latch is the dominant exit for profitability and that
1035   // exits branching to deoptimizing blocks are rarely taken. It relies on the
1036   // semantics of widenable expressions for legality. (i.e. being able to fall
1037   // down the widenable path spuriously allows us to ignore exit order,
1038   // unanalyzeable exits, side effects, exceptional exits, and other challenges
1039   // which restrict the applicability of the non-WC based version of this
1040   // transform in IndVarSimplify.)
1041   //
1042   // NOTE ON POISON/UNDEF - We're hoisting an expression above guards which may
1043   // imply flags on the expression being hoisted and inserting new uses (flags
1044   // are only correct for current uses).  The result is that we may be
1045   // inserting a branch on the value which can be either poison or undef.  In
1046   // this case, the branch can legally go either way; we just need to avoid
1047   // introducing UB.  This is achieved through the use of the freeze
1048   // instruction.
1049 
1050   SmallVector<BasicBlock *, 16> ExitingBlocks;
1051   L->getExitingBlocks(ExitingBlocks);
1052 
1053   if (ExitingBlocks.empty())
1054     return false; // Nothing to do.
1055 
1056   auto *Latch = L->getLoopLatch();
1057   if (!Latch)
1058     return false;
1059 
1060   auto *WidenableBR = FindWidenableTerminatorAboveLoop(L, *LI);
1061   if (!WidenableBR)
1062     return false;
1063 
1064   const SCEV *LatchEC = SE->getExitCount(L, Latch);
1065   if (isa<SCEVCouldNotCompute>(LatchEC))
1066     return false; // profitability - want hot exit in analyzeable set
1067 
1068   // At this point, we have found an analyzeable latch, and a widenable
1069   // condition above the loop.  If we have a widenable exit within the loop
1070   // (for which we can't compute exit counts), drop the ability to further
1071   // widen so that we gain ability to analyze it's exit count and perform this
1072   // transform.  TODO: It'd be nice to know for sure the exit became
1073   // analyzeable after dropping widenability.
1074   {
1075     bool Invalidate = false;
1076 
1077     for (auto *ExitingBB : ExitingBlocks) {
1078       if (LI->getLoopFor(ExitingBB) != L)
1079         continue;
1080 
1081       auto *BI = dyn_cast<BranchInst>(ExitingBB->getTerminator());
1082       if (!BI)
1083         continue;
1084 
1085       Use *Cond, *WC;
1086       BasicBlock *IfTrueBB, *IfFalseBB;
1087       if (parseWidenableBranch(BI, Cond, WC, IfTrueBB, IfFalseBB) &&
1088           L->contains(IfTrueBB)) {
1089         WC->set(ConstantInt::getTrue(IfTrueBB->getContext()));
1090         Invalidate = true;
1091       }
1092     }
1093     if (Invalidate)
1094       SE->forgetLoop(L);
1095   }
1096 
1097   // The use of umin(all analyzeable exits) instead of latch is subtle, but
1098   // important for profitability.  We may have a loop which hasn't been fully
1099   // canonicalized just yet.  If the exit we chose to widen is provably never
1100   // taken, we want the widened form to *also* be provably never taken.  We
1101   // can't guarantee this as a current unanalyzeable exit may later become
1102   // analyzeable, but we can at least avoid the obvious cases.
1103   const SCEV *MinEC = getMinAnalyzeableBackedgeTakenCount(*SE, *DT, L);
1104   if (isa<SCEVCouldNotCompute>(MinEC) || MinEC->getType()->isPointerTy() ||
1105       !SE->isLoopInvariant(MinEC, L) ||
1106       !isSafeToExpandAt(MinEC, WidenableBR, *SE))
1107     return false;
1108 
1109   // Subtlety: We need to avoid inserting additional uses of the WC.  We know
1110   // that it can only have one transitive use at the moment, and thus moving
1111   // that use to just before the branch and inserting code before it and then
1112   // modifying the operand is legal.
1113   auto *IP = cast<Instruction>(WidenableBR->getCondition());
1114   IP->moveBefore(WidenableBR);
1115   Rewriter.setInsertPoint(IP);
1116   IRBuilder<> B(IP);
1117 
1118   bool Changed = false;
1119   Value *MinECV = nullptr; // lazily generated if needed
1120   for (BasicBlock *ExitingBB : ExitingBlocks) {
1121     // If our exiting block exits multiple loops, we can only rewrite the
1122     // innermost one.  Otherwise, we're changing how many times the innermost
1123     // loop runs before it exits.
1124     if (LI->getLoopFor(ExitingBB) != L)
1125       continue;
1126 
1127     // Can't rewrite non-branch yet.
1128     auto *BI = dyn_cast<BranchInst>(ExitingBB->getTerminator());
1129     if (!BI)
1130       continue;
1131 
1132     // If already constant, nothing to do.
1133     if (isa<Constant>(BI->getCondition()))
1134       continue;
1135 
1136     const SCEV *ExitCount = SE->getExitCount(L, ExitingBB);
1137     if (isa<SCEVCouldNotCompute>(ExitCount) ||
1138         ExitCount->getType()->isPointerTy() ||
1139         !isSafeToExpandAt(ExitCount, WidenableBR, *SE))
1140       continue;
1141 
1142     const bool ExitIfTrue = !L->contains(*succ_begin(ExitingBB));
1143     BasicBlock *ExitBB = BI->getSuccessor(ExitIfTrue ? 0 : 1);
1144     if (!ExitBB->getPostdominatingDeoptimizeCall())
1145       continue;
1146 
1147     /// Here we can be fairly sure that executing this exit will most likely
1148     /// lead to executing llvm.experimental.deoptimize.
1149     /// This is a profitability heuristic, not a legality constraint.
1150 
1151     // If we found a widenable exit condition, do two things:
1152     // 1) fold the widened exit test into the widenable condition
1153     // 2) fold the branch to untaken - avoids infinite looping
1154 
1155     Value *ECV = Rewriter.expandCodeFor(ExitCount);
1156     if (!MinECV)
1157       MinECV = Rewriter.expandCodeFor(MinEC);
1158     Value *RHS = MinECV;
1159     if (ECV->getType() != RHS->getType()) {
1160       Type *WiderTy = SE->getWiderType(ECV->getType(), RHS->getType());
1161       ECV = B.CreateZExt(ECV, WiderTy);
1162       RHS = B.CreateZExt(RHS, WiderTy);
1163     }
1164     assert(!Latch || DT->dominates(ExitingBB, Latch));
1165     Value *NewCond = B.CreateICmp(ICmpInst::ICMP_UGT, ECV, RHS);
1166     // Freeze poison or undef to an arbitrary bit pattern to ensure we can
1167     // branch without introducing UB.  See NOTE ON POISON/UNDEF above for
1168     // context.
1169     NewCond = B.CreateFreeze(NewCond);
1170 
1171     widenWidenableBranch(WidenableBR, NewCond);
1172 
1173     Value *OldCond = BI->getCondition();
1174     BI->setCondition(ConstantInt::get(OldCond->getType(), !ExitIfTrue));
1175     Changed = true;
1176   }
1177 
1178   if (Changed)
1179     // We just mutated a bunch of loop exits changing there exit counts
1180     // widely.  We need to force recomputation of the exit counts given these
1181     // changes.  Note that all of the inserted exits are never taken, and
1182     // should be removed next time the CFG is modified.
1183     SE->forgetLoop(L);
1184   return Changed;
1185 }
1186 
runOnLoop(Loop * Loop)1187 bool LoopPredication::runOnLoop(Loop *Loop) {
1188   L = Loop;
1189 
1190   LLVM_DEBUG(dbgs() << "Analyzing ");
1191   LLVM_DEBUG(L->dump());
1192 
1193   Module *M = L->getHeader()->getModule();
1194 
1195   // There is nothing to do if the module doesn't use guards
1196   auto *GuardDecl =
1197       M->getFunction(Intrinsic::getName(Intrinsic::experimental_guard));
1198   bool HasIntrinsicGuards = GuardDecl && !GuardDecl->use_empty();
1199   auto *WCDecl = M->getFunction(
1200       Intrinsic::getName(Intrinsic::experimental_widenable_condition));
1201   bool HasWidenableConditions =
1202       PredicateWidenableBranchGuards && WCDecl && !WCDecl->use_empty();
1203   if (!HasIntrinsicGuards && !HasWidenableConditions)
1204     return false;
1205 
1206   DL = &M->getDataLayout();
1207 
1208   Preheader = L->getLoopPreheader();
1209   if (!Preheader)
1210     return false;
1211 
1212   auto LatchCheckOpt = parseLoopLatchICmp();
1213   if (!LatchCheckOpt)
1214     return false;
1215   LatchCheck = *LatchCheckOpt;
1216 
1217   LLVM_DEBUG(dbgs() << "Latch check:\n");
1218   LLVM_DEBUG(LatchCheck.dump());
1219 
1220   if (!isLoopProfitableToPredicate()) {
1221     LLVM_DEBUG(dbgs() << "Loop not profitable to predicate!\n");
1222     return false;
1223   }
1224   // Collect all the guards into a vector and process later, so as not
1225   // to invalidate the instruction iterator.
1226   SmallVector<IntrinsicInst *, 4> Guards;
1227   SmallVector<BranchInst *, 4> GuardsAsWidenableBranches;
1228   for (const auto BB : L->blocks()) {
1229     for (auto &I : *BB)
1230       if (isGuard(&I))
1231         Guards.push_back(cast<IntrinsicInst>(&I));
1232     if (PredicateWidenableBranchGuards &&
1233         isGuardAsWidenableBranch(BB->getTerminator()))
1234       GuardsAsWidenableBranches.push_back(
1235           cast<BranchInst>(BB->getTerminator()));
1236   }
1237 
1238   SCEVExpander Expander(*SE, *DL, "loop-predication");
1239   bool Changed = false;
1240   for (auto *Guard : Guards)
1241     Changed |= widenGuardConditions(Guard, Expander);
1242   for (auto *Guard : GuardsAsWidenableBranches)
1243     Changed |= widenWidenableBranchGuardConditions(Guard, Expander);
1244   Changed |= predicateLoopExits(L, Expander);
1245   return Changed;
1246 }
1247