xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Scalar/LoopPredication.cpp (revision 7fdf597e96a02165cfe22ff357b857d5fa15ed8a)
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/MemorySSA.h"
187 #include "llvm/Analysis/MemorySSAUpdater.h"
188 #include "llvm/Analysis/ScalarEvolution.h"
189 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
190 #include "llvm/IR/Function.h"
191 #include "llvm/IR/IntrinsicInst.h"
192 #include "llvm/IR/Module.h"
193 #include "llvm/IR/PatternMatch.h"
194 #include "llvm/IR/ProfDataUtils.h"
195 #include "llvm/InitializePasses.h"
196 #include "llvm/Pass.h"
197 #include "llvm/Support/CommandLine.h"
198 #include "llvm/Support/Debug.h"
199 #include "llvm/Transforms/Scalar.h"
200 #include "llvm/Transforms/Utils/GuardUtils.h"
201 #include "llvm/Transforms/Utils/Local.h"
202 #include "llvm/Transforms/Utils/LoopUtils.h"
203 #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
204 #include <optional>
205 
206 #define DEBUG_TYPE "loop-predication"
207 
208 STATISTIC(TotalConsidered, "Number of guards considered");
209 STATISTIC(TotalWidened, "Number of checks widened");
210 
211 using namespace llvm;
212 
213 static cl::opt<bool> EnableIVTruncation("loop-predication-enable-iv-truncation",
214                                         cl::Hidden, cl::init(true));
215 
216 static cl::opt<bool> EnableCountDownLoop("loop-predication-enable-count-down-loop",
217                                         cl::Hidden, cl::init(true));
218 
219 static cl::opt<bool>
220     SkipProfitabilityChecks("loop-predication-skip-profitability-checks",
221                             cl::Hidden, cl::init(false));
222 
223 // This is the scale factor for the latch probability. We use this during
224 // profitability analysis to find other exiting blocks that have a much higher
225 // probability of exiting the loop instead of loop exiting via latch.
226 // This value should be greater than 1 for a sane profitability check.
227 static cl::opt<float> LatchExitProbabilityScale(
228     "loop-predication-latch-probability-scale", cl::Hidden, cl::init(2.0),
229     cl::desc("scale factor for the latch probability. Value should be greater "
230              "than 1. Lower values are ignored"));
231 
232 static cl::opt<bool> PredicateWidenableBranchGuards(
233     "loop-predication-predicate-widenable-branches-to-deopt", cl::Hidden,
234     cl::desc("Whether or not we should predicate guards "
235              "expressed as widenable branches to deoptimize blocks"),
236     cl::init(true));
237 
238 static cl::opt<bool> InsertAssumesOfPredicatedGuardsConditions(
239     "loop-predication-insert-assumes-of-predicated-guards-conditions",
240     cl::Hidden,
241     cl::desc("Whether or not we should insert assumes of conditions of "
242              "predicated guards"),
243     cl::init(true));
244 
245 namespace {
246 /// Represents an induction variable check:
247 ///   icmp Pred, <induction variable>, <loop invariant limit>
248 struct LoopICmp {
249   ICmpInst::Predicate Pred;
250   const SCEVAddRecExpr *IV;
251   const SCEV *Limit;
252   LoopICmp(ICmpInst::Predicate Pred, const SCEVAddRecExpr *IV,
253            const SCEV *Limit)
254     : Pred(Pred), IV(IV), Limit(Limit) {}
255   LoopICmp() = default;
256   void dump() {
257     dbgs() << "LoopICmp Pred = " << Pred << ", IV = " << *IV
258            << ", Limit = " << *Limit << "\n";
259   }
260 };
261 
262 class LoopPredication {
263   AliasAnalysis *AA;
264   DominatorTree *DT;
265   ScalarEvolution *SE;
266   LoopInfo *LI;
267   MemorySSAUpdater *MSSAU;
268 
269   Loop *L;
270   const DataLayout *DL;
271   BasicBlock *Preheader;
272   LoopICmp LatchCheck;
273 
274   bool isSupportedStep(const SCEV* Step);
275   std::optional<LoopICmp> parseLoopICmp(ICmpInst *ICI);
276   std::optional<LoopICmp> parseLoopLatchICmp();
277 
278   /// Return an insertion point suitable for inserting a safe to speculate
279   /// instruction whose only user will be 'User' which has operands 'Ops'.  A
280   /// trivial result would be the at the User itself, but we try to return a
281   /// loop invariant location if possible.
282   Instruction *findInsertPt(Instruction *User, ArrayRef<Value*> Ops);
283   /// Same as above, *except* that this uses the SCEV definition of invariant
284   /// which is that an expression *can be made* invariant via SCEVExpander.
285   /// Thus, this version is only suitable for finding an insert point to be
286   /// passed to SCEVExpander!
287   Instruction *findInsertPt(const SCEVExpander &Expander, Instruction *User,
288                             ArrayRef<const SCEV *> Ops);
289 
290   /// Return true if the value is known to produce a single fixed value across
291   /// all iterations on which it executes.  Note that this does not imply
292   /// speculation safety.  That must be established separately.
293   bool isLoopInvariantValue(const SCEV* S);
294 
295   Value *expandCheck(SCEVExpander &Expander, Instruction *Guard,
296                      ICmpInst::Predicate Pred, const SCEV *LHS,
297                      const SCEV *RHS);
298 
299   std::optional<Value *> widenICmpRangeCheck(ICmpInst *ICI,
300                                              SCEVExpander &Expander,
301                                              Instruction *Guard);
302   std::optional<Value *>
303   widenICmpRangeCheckIncrementingLoop(LoopICmp LatchCheck, LoopICmp RangeCheck,
304                                       SCEVExpander &Expander,
305                                       Instruction *Guard);
306   std::optional<Value *>
307   widenICmpRangeCheckDecrementingLoop(LoopICmp LatchCheck, LoopICmp RangeCheck,
308                                       SCEVExpander &Expander,
309                                       Instruction *Guard);
310   void widenChecks(SmallVectorImpl<Value *> &Checks,
311                    SmallVectorImpl<Value *> &WidenedChecks,
312                    SCEVExpander &Expander, Instruction *Guard);
313   bool widenGuardConditions(IntrinsicInst *II, SCEVExpander &Expander);
314   bool widenWidenableBranchGuardConditions(BranchInst *Guard, SCEVExpander &Expander);
315   // If the loop always exits through another block in the loop, we should not
316   // predicate based on the latch check. For example, the latch check can be a
317   // very coarse grained check and there can be more fine grained exit checks
318   // within the loop.
319   bool isLoopProfitableToPredicate();
320 
321   bool predicateLoopExits(Loop *L, SCEVExpander &Rewriter);
322 
323 public:
324   LoopPredication(AliasAnalysis *AA, DominatorTree *DT, ScalarEvolution *SE,
325                   LoopInfo *LI, MemorySSAUpdater *MSSAU)
326       : AA(AA), DT(DT), SE(SE), LI(LI), MSSAU(MSSAU){};
327   bool runOnLoop(Loop *L);
328 };
329 
330 } // end namespace
331 
332 PreservedAnalyses LoopPredicationPass::run(Loop &L, LoopAnalysisManager &AM,
333                                            LoopStandardAnalysisResults &AR,
334                                            LPMUpdater &U) {
335   std::unique_ptr<MemorySSAUpdater> MSSAU;
336   if (AR.MSSA)
337     MSSAU = std::make_unique<MemorySSAUpdater>(AR.MSSA);
338   LoopPredication LP(&AR.AA, &AR.DT, &AR.SE, &AR.LI,
339                      MSSAU ? MSSAU.get() : nullptr);
340   if (!LP.runOnLoop(&L))
341     return PreservedAnalyses::all();
342 
343   auto PA = getLoopPassPreservedAnalyses();
344   if (AR.MSSA)
345     PA.preserve<MemorySSAAnalysis>();
346   return PA;
347 }
348 
349 std::optional<LoopICmp> LoopPredication::parseLoopICmp(ICmpInst *ICI) {
350   auto Pred = ICI->getPredicate();
351   auto *LHS = ICI->getOperand(0);
352   auto *RHS = ICI->getOperand(1);
353 
354   const SCEV *LHSS = SE->getSCEV(LHS);
355   if (isa<SCEVCouldNotCompute>(LHSS))
356     return std::nullopt;
357   const SCEV *RHSS = SE->getSCEV(RHS);
358   if (isa<SCEVCouldNotCompute>(RHSS))
359     return std::nullopt;
360 
361   // Canonicalize RHS to be loop invariant bound, LHS - a loop computable IV
362   if (SE->isLoopInvariant(LHSS, L)) {
363     std::swap(LHS, RHS);
364     std::swap(LHSS, RHSS);
365     Pred = ICmpInst::getSwappedPredicate(Pred);
366   }
367 
368   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHSS);
369   if (!AR || AR->getLoop() != L)
370     return std::nullopt;
371 
372   return LoopICmp(Pred, AR, RHSS);
373 }
374 
375 Value *LoopPredication::expandCheck(SCEVExpander &Expander,
376                                     Instruction *Guard,
377                                     ICmpInst::Predicate Pred, const SCEV *LHS,
378                                     const SCEV *RHS) {
379   Type *Ty = LHS->getType();
380   assert(Ty == RHS->getType() && "expandCheck operands have different types?");
381 
382   if (SE->isLoopInvariant(LHS, L) && SE->isLoopInvariant(RHS, L)) {
383     IRBuilder<> Builder(Guard);
384     if (SE->isLoopEntryGuardedByCond(L, Pred, LHS, RHS))
385       return Builder.getTrue();
386     if (SE->isLoopEntryGuardedByCond(L, ICmpInst::getInversePredicate(Pred),
387                                      LHS, RHS))
388       return Builder.getFalse();
389   }
390 
391   Value *LHSV =
392       Expander.expandCodeFor(LHS, Ty, findInsertPt(Expander, Guard, {LHS}));
393   Value *RHSV =
394       Expander.expandCodeFor(RHS, Ty, findInsertPt(Expander, Guard, {RHS}));
395   IRBuilder<> Builder(findInsertPt(Guard, {LHSV, RHSV}));
396   return Builder.CreateICmp(Pred, LHSV, RHSV);
397 }
398 
399 // Returns true if its safe to truncate the IV to RangeCheckType.
400 // When the IV type is wider than the range operand type, we can still do loop
401 // predication, by generating SCEVs for the range and latch that are of the
402 // same type. We achieve this by generating a SCEV truncate expression for the
403 // latch IV. This is done iff truncation of the IV is a safe operation,
404 // without loss of information.
405 // Another way to achieve this is by generating a wider type SCEV for the
406 // range check operand, however, this needs a more involved check that
407 // operands do not overflow. This can lead to loss of information when the
408 // range operand is of the form: add i32 %offset, %iv. We need to prove that
409 // sext(x + y) is same as sext(x) + sext(y).
410 // This function returns true if we can safely represent the IV type in
411 // the RangeCheckType without loss of information.
412 static bool isSafeToTruncateWideIVType(const DataLayout &DL,
413                                        ScalarEvolution &SE,
414                                        const LoopICmp LatchCheck,
415                                        Type *RangeCheckType) {
416   if (!EnableIVTruncation)
417     return false;
418   assert(DL.getTypeSizeInBits(LatchCheck.IV->getType()).getFixedValue() >
419              DL.getTypeSizeInBits(RangeCheckType).getFixedValue() &&
420          "Expected latch check IV type to be larger than range check operand "
421          "type!");
422   // The start and end values of the IV should be known. This is to guarantee
423   // that truncating the wide type will not lose information.
424   auto *Limit = dyn_cast<SCEVConstant>(LatchCheck.Limit);
425   auto *Start = dyn_cast<SCEVConstant>(LatchCheck.IV->getStart());
426   if (!Limit || !Start)
427     return false;
428   // This check makes sure that the IV does not change sign during loop
429   // iterations. Consider latchType = i64, LatchStart = 5, Pred = ICMP_SGE,
430   // LatchEnd = 2, rangeCheckType = i32. If it's not a monotonic predicate, the
431   // IV wraps around, and the truncation of the IV would lose the range of
432   // iterations between 2^32 and 2^64.
433   if (!SE.getMonotonicPredicateType(LatchCheck.IV, LatchCheck.Pred))
434     return false;
435   // The active bits should be less than the bits in the RangeCheckType. This
436   // guarantees that truncating the latch check to RangeCheckType is a safe
437   // operation.
438   auto RangeCheckTypeBitSize =
439       DL.getTypeSizeInBits(RangeCheckType).getFixedValue();
440   return Start->getAPInt().getActiveBits() < RangeCheckTypeBitSize &&
441          Limit->getAPInt().getActiveBits() < RangeCheckTypeBitSize;
442 }
443 
444 
445 // Return an LoopICmp describing a latch check equivlent to LatchCheck but with
446 // the requested type if safe to do so.  May involve the use of a new IV.
447 static std::optional<LoopICmp> generateLoopLatchCheck(const DataLayout &DL,
448                                                       ScalarEvolution &SE,
449                                                       const LoopICmp LatchCheck,
450                                                       Type *RangeCheckType) {
451 
452   auto *LatchType = LatchCheck.IV->getType();
453   if (RangeCheckType == LatchType)
454     return LatchCheck;
455   // For now, bail out if latch type is narrower than range type.
456   if (DL.getTypeSizeInBits(LatchType).getFixedValue() <
457       DL.getTypeSizeInBits(RangeCheckType).getFixedValue())
458     return std::nullopt;
459   if (!isSafeToTruncateWideIVType(DL, SE, LatchCheck, RangeCheckType))
460     return std::nullopt;
461   // We can now safely identify the truncated version of the IV and limit for
462   // RangeCheckType.
463   LoopICmp NewLatchCheck;
464   NewLatchCheck.Pred = LatchCheck.Pred;
465   NewLatchCheck.IV = dyn_cast<SCEVAddRecExpr>(
466       SE.getTruncateExpr(LatchCheck.IV, RangeCheckType));
467   if (!NewLatchCheck.IV)
468     return std::nullopt;
469   NewLatchCheck.Limit = SE.getTruncateExpr(LatchCheck.Limit, RangeCheckType);
470   LLVM_DEBUG(dbgs() << "IV of type: " << *LatchType
471                     << "can be represented as range check type:"
472                     << *RangeCheckType << "\n");
473   LLVM_DEBUG(dbgs() << "LatchCheck.IV: " << *NewLatchCheck.IV << "\n");
474   LLVM_DEBUG(dbgs() << "LatchCheck.Limit: " << *NewLatchCheck.Limit << "\n");
475   return NewLatchCheck;
476 }
477 
478 bool LoopPredication::isSupportedStep(const SCEV* Step) {
479   return Step->isOne() || (Step->isAllOnesValue() && EnableCountDownLoop);
480 }
481 
482 Instruction *LoopPredication::findInsertPt(Instruction *Use,
483                                            ArrayRef<Value*> Ops) {
484   for (Value *Op : Ops)
485     if (!L->isLoopInvariant(Op))
486       return Use;
487   return Preheader->getTerminator();
488 }
489 
490 Instruction *LoopPredication::findInsertPt(const SCEVExpander &Expander,
491                                            Instruction *Use,
492                                            ArrayRef<const SCEV *> Ops) {
493   // Subtlety: SCEV considers things to be invariant if the value produced is
494   // the same across iterations.  This is not the same as being able to
495   // evaluate outside the loop, which is what we actually need here.
496   for (const SCEV *Op : Ops)
497     if (!SE->isLoopInvariant(Op, L) ||
498         !Expander.isSafeToExpandAt(Op, Preheader->getTerminator()))
499       return Use;
500   return Preheader->getTerminator();
501 }
502 
503 bool LoopPredication::isLoopInvariantValue(const SCEV* S) {
504   // Handling expressions which produce invariant results, but *haven't* yet
505   // been removed from the loop serves two important purposes.
506   // 1) Most importantly, it resolves a pass ordering cycle which would
507   // otherwise need us to iteration licm, loop-predication, and either
508   // loop-unswitch or loop-peeling to make progress on examples with lots of
509   // predicable range checks in a row.  (Since, in the general case,  we can't
510   // hoist the length checks until the dominating checks have been discharged
511   // as we can't prove doing so is safe.)
512   // 2) As a nice side effect, this exposes the value of peeling or unswitching
513   // much more obviously in the IR.  Otherwise, the cost modeling for other
514   // transforms would end up needing to duplicate all of this logic to model a
515   // check which becomes predictable based on a modeled peel or unswitch.
516   //
517   // The cost of doing so in the worst case is an extra fill from the stack  in
518   // the loop to materialize the loop invariant test value instead of checking
519   // against the original IV which is presumable in a register inside the loop.
520   // Such cases are presumably rare, and hint at missing oppurtunities for
521   // other passes.
522 
523   if (SE->isLoopInvariant(S, L))
524     // Note: This the SCEV variant, so the original Value* may be within the
525     // loop even though SCEV has proven it is loop invariant.
526     return true;
527 
528   // Handle a particular important case which SCEV doesn't yet know about which
529   // shows up in range checks on arrays with immutable lengths.
530   // TODO: This should be sunk inside SCEV.
531   if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S))
532     if (const auto *LI = dyn_cast<LoadInst>(U->getValue()))
533       if (LI->isUnordered() && L->hasLoopInvariantOperands(LI))
534         if (!isModSet(AA->getModRefInfoMask(LI->getOperand(0))) ||
535             LI->hasMetadata(LLVMContext::MD_invariant_load))
536           return true;
537   return false;
538 }
539 
540 std::optional<Value *> LoopPredication::widenICmpRangeCheckIncrementingLoop(
541     LoopICmp LatchCheck, LoopICmp RangeCheck, SCEVExpander &Expander,
542     Instruction *Guard) {
543   auto *Ty = RangeCheck.IV->getType();
544   // Generate the widened condition for the forward loop:
545   //   guardStart u< guardLimit &&
546   //   latchLimit <pred> guardLimit - 1 - guardStart + latchStart
547   // where <pred> depends on the latch condition predicate. See the file
548   // header comment for the reasoning.
549   // guardLimit - guardStart + latchStart - 1
550   const SCEV *GuardStart = RangeCheck.IV->getStart();
551   const SCEV *GuardLimit = RangeCheck.Limit;
552   const SCEV *LatchStart = LatchCheck.IV->getStart();
553   const SCEV *LatchLimit = LatchCheck.Limit;
554   // Subtlety: We need all the values to be *invariant* across all iterations,
555   // but we only need to check expansion safety for those which *aren't*
556   // already guaranteed to dominate the guard.
557   if (!isLoopInvariantValue(GuardStart) ||
558       !isLoopInvariantValue(GuardLimit) ||
559       !isLoopInvariantValue(LatchStart) ||
560       !isLoopInvariantValue(LatchLimit)) {
561     LLVM_DEBUG(dbgs() << "Can't expand limit check!\n");
562     return std::nullopt;
563   }
564   if (!Expander.isSafeToExpandAt(LatchStart, Guard) ||
565       !Expander.isSafeToExpandAt(LatchLimit, Guard)) {
566     LLVM_DEBUG(dbgs() << "Can't expand limit check!\n");
567     return std::nullopt;
568   }
569 
570   // guardLimit - guardStart + latchStart - 1
571   const SCEV *RHS =
572       SE->getAddExpr(SE->getMinusSCEV(GuardLimit, GuardStart),
573                      SE->getMinusSCEV(LatchStart, SE->getOne(Ty)));
574   auto LimitCheckPred =
575       ICmpInst::getFlippedStrictnessPredicate(LatchCheck.Pred);
576 
577   LLVM_DEBUG(dbgs() << "LHS: " << *LatchLimit << "\n");
578   LLVM_DEBUG(dbgs() << "RHS: " << *RHS << "\n");
579   LLVM_DEBUG(dbgs() << "Pred: " << LimitCheckPred << "\n");
580 
581   auto *LimitCheck =
582       expandCheck(Expander, Guard, LimitCheckPred, LatchLimit, RHS);
583   auto *FirstIterationCheck = expandCheck(Expander, Guard, RangeCheck.Pred,
584                                           GuardStart, GuardLimit);
585   IRBuilder<> Builder(findInsertPt(Guard, {FirstIterationCheck, LimitCheck}));
586   return Builder.CreateFreeze(
587       Builder.CreateAnd(FirstIterationCheck, LimitCheck));
588 }
589 
590 std::optional<Value *> LoopPredication::widenICmpRangeCheckDecrementingLoop(
591     LoopICmp LatchCheck, LoopICmp RangeCheck, SCEVExpander &Expander,
592     Instruction *Guard) {
593   auto *Ty = RangeCheck.IV->getType();
594   const SCEV *GuardStart = RangeCheck.IV->getStart();
595   const SCEV *GuardLimit = RangeCheck.Limit;
596   const SCEV *LatchStart = LatchCheck.IV->getStart();
597   const SCEV *LatchLimit = LatchCheck.Limit;
598   // Subtlety: We need all the values to be *invariant* across all iterations,
599   // but we only need to check expansion safety for those which *aren't*
600   // already guaranteed to dominate the guard.
601   if (!isLoopInvariantValue(GuardStart) ||
602       !isLoopInvariantValue(GuardLimit) ||
603       !isLoopInvariantValue(LatchStart) ||
604       !isLoopInvariantValue(LatchLimit)) {
605     LLVM_DEBUG(dbgs() << "Can't expand limit check!\n");
606     return std::nullopt;
607   }
608   if (!Expander.isSafeToExpandAt(LatchStart, Guard) ||
609       !Expander.isSafeToExpandAt(LatchLimit, Guard)) {
610     LLVM_DEBUG(dbgs() << "Can't expand limit check!\n");
611     return std::nullopt;
612   }
613   // The decrement of the latch check IV should be the same as the
614   // rangeCheckIV.
615   auto *PostDecLatchCheckIV = LatchCheck.IV->getPostIncExpr(*SE);
616   if (RangeCheck.IV != PostDecLatchCheckIV) {
617     LLVM_DEBUG(dbgs() << "Not the same. PostDecLatchCheckIV: "
618                       << *PostDecLatchCheckIV
619                       << "  and RangeCheckIV: " << *RangeCheck.IV << "\n");
620     return std::nullopt;
621   }
622 
623   // Generate the widened condition for CountDownLoop:
624   // guardStart u< guardLimit &&
625   // latchLimit <pred> 1.
626   // See the header comment for reasoning of the checks.
627   auto LimitCheckPred =
628       ICmpInst::getFlippedStrictnessPredicate(LatchCheck.Pred);
629   auto *FirstIterationCheck = expandCheck(Expander, Guard,
630                                           ICmpInst::ICMP_ULT,
631                                           GuardStart, GuardLimit);
632   auto *LimitCheck = expandCheck(Expander, Guard, LimitCheckPred, LatchLimit,
633                                  SE->getOne(Ty));
634   IRBuilder<> Builder(findInsertPt(Guard, {FirstIterationCheck, LimitCheck}));
635   return Builder.CreateFreeze(
636       Builder.CreateAnd(FirstIterationCheck, LimitCheck));
637 }
638 
639 static void normalizePredicate(ScalarEvolution *SE, Loop *L,
640                                LoopICmp& RC) {
641   // LFTR canonicalizes checks to the ICMP_NE/EQ form; normalize back to the
642   // ULT/UGE form for ease of handling by our caller.
643   if (ICmpInst::isEquality(RC.Pred) &&
644       RC.IV->getStepRecurrence(*SE)->isOne() &&
645       SE->isKnownPredicate(ICmpInst::ICMP_ULE, RC.IV->getStart(), RC.Limit))
646     RC.Pred = RC.Pred == ICmpInst::ICMP_NE ?
647       ICmpInst::ICMP_ULT : ICmpInst::ICMP_UGE;
648 }
649 
650 /// If ICI can be widened to a loop invariant condition emits the loop
651 /// invariant condition in the loop preheader and return it, otherwise
652 /// returns std::nullopt.
653 std::optional<Value *>
654 LoopPredication::widenICmpRangeCheck(ICmpInst *ICI, SCEVExpander &Expander,
655                                      Instruction *Guard) {
656   LLVM_DEBUG(dbgs() << "Analyzing ICmpInst condition:\n");
657   LLVM_DEBUG(ICI->dump());
658 
659   // parseLoopStructure guarantees that the latch condition is:
660   //   ++i <pred> latchLimit, where <pred> is u<, u<=, s<, or s<=.
661   // We are looking for the range checks of the form:
662   //   i u< guardLimit
663   auto RangeCheck = parseLoopICmp(ICI);
664   if (!RangeCheck) {
665     LLVM_DEBUG(dbgs() << "Failed to parse the loop latch condition!\n");
666     return std::nullopt;
667   }
668   LLVM_DEBUG(dbgs() << "Guard check:\n");
669   LLVM_DEBUG(RangeCheck->dump());
670   if (RangeCheck->Pred != ICmpInst::ICMP_ULT) {
671     LLVM_DEBUG(dbgs() << "Unsupported range check predicate("
672                       << RangeCheck->Pred << ")!\n");
673     return std::nullopt;
674   }
675   auto *RangeCheckIV = RangeCheck->IV;
676   if (!RangeCheckIV->isAffine()) {
677     LLVM_DEBUG(dbgs() << "Range check IV is not affine!\n");
678     return std::nullopt;
679   }
680   auto *Step = RangeCheckIV->getStepRecurrence(*SE);
681   // We cannot just compare with latch IV step because the latch and range IVs
682   // may have different types.
683   if (!isSupportedStep(Step)) {
684     LLVM_DEBUG(dbgs() << "Range check and latch have IVs different steps!\n");
685     return std::nullopt;
686   }
687   auto *Ty = RangeCheckIV->getType();
688   auto CurrLatchCheckOpt = generateLoopLatchCheck(*DL, *SE, LatchCheck, Ty);
689   if (!CurrLatchCheckOpt) {
690     LLVM_DEBUG(dbgs() << "Failed to generate a loop latch check "
691                          "corresponding to range type: "
692                       << *Ty << "\n");
693     return std::nullopt;
694   }
695 
696   LoopICmp CurrLatchCheck = *CurrLatchCheckOpt;
697   // At this point, the range and latch step should have the same type, but need
698   // not have the same value (we support both 1 and -1 steps).
699   assert(Step->getType() ==
700              CurrLatchCheck.IV->getStepRecurrence(*SE)->getType() &&
701          "Range and latch steps should be of same type!");
702   if (Step != CurrLatchCheck.IV->getStepRecurrence(*SE)) {
703     LLVM_DEBUG(dbgs() << "Range and latch have different step values!\n");
704     return std::nullopt;
705   }
706 
707   if (Step->isOne())
708     return widenICmpRangeCheckIncrementingLoop(CurrLatchCheck, *RangeCheck,
709                                                Expander, Guard);
710   else {
711     assert(Step->isAllOnesValue() && "Step should be -1!");
712     return widenICmpRangeCheckDecrementingLoop(CurrLatchCheck, *RangeCheck,
713                                                Expander, Guard);
714   }
715 }
716 
717 void LoopPredication::widenChecks(SmallVectorImpl<Value *> &Checks,
718                                   SmallVectorImpl<Value *> &WidenedChecks,
719                                   SCEVExpander &Expander, Instruction *Guard) {
720   for (auto &Check : Checks)
721     if (ICmpInst *ICI = dyn_cast<ICmpInst>(Check))
722       if (auto NewRangeCheck = widenICmpRangeCheck(ICI, Expander, Guard)) {
723         WidenedChecks.push_back(Check);
724         Check = *NewRangeCheck;
725       }
726 }
727 
728 bool LoopPredication::widenGuardConditions(IntrinsicInst *Guard,
729                                            SCEVExpander &Expander) {
730   LLVM_DEBUG(dbgs() << "Processing guard:\n");
731   LLVM_DEBUG(Guard->dump());
732 
733   TotalConsidered++;
734   SmallVector<Value *, 4> Checks;
735   SmallVector<Value *> WidenedChecks;
736   parseWidenableGuard(Guard, Checks);
737   widenChecks(Checks, WidenedChecks, Expander, Guard);
738   if (WidenedChecks.empty())
739     return false;
740 
741   TotalWidened += WidenedChecks.size();
742 
743   // Emit the new guard condition
744   IRBuilder<> Builder(findInsertPt(Guard, Checks));
745   Value *AllChecks = Builder.CreateAnd(Checks);
746   auto *OldCond = Guard->getOperand(0);
747   Guard->setOperand(0, AllChecks);
748   if (InsertAssumesOfPredicatedGuardsConditions) {
749     Builder.SetInsertPoint(&*++BasicBlock::iterator(Guard));
750     Builder.CreateAssumption(OldCond);
751   }
752   RecursivelyDeleteTriviallyDeadInstructions(OldCond, nullptr /* TLI */, MSSAU);
753 
754   LLVM_DEBUG(dbgs() << "Widened checks = " << WidenedChecks.size() << "\n");
755   return true;
756 }
757 
758 bool LoopPredication::widenWidenableBranchGuardConditions(
759     BranchInst *BI, SCEVExpander &Expander) {
760   assert(isGuardAsWidenableBranch(BI) && "Must be!");
761   LLVM_DEBUG(dbgs() << "Processing guard:\n");
762   LLVM_DEBUG(BI->dump());
763 
764   TotalConsidered++;
765   SmallVector<Value *, 4> Checks;
766   SmallVector<Value *> WidenedChecks;
767   parseWidenableGuard(BI, Checks);
768   // At the moment, our matching logic for wideable conditions implicitly
769   // assumes we preserve the form: (br (and Cond, WC())).  FIXME
770   auto WC = extractWidenableCondition(BI);
771   Checks.push_back(WC);
772   widenChecks(Checks, WidenedChecks, Expander, BI);
773   if (WidenedChecks.empty())
774     return false;
775 
776   TotalWidened += WidenedChecks.size();
777 
778   // Emit the new guard condition
779   IRBuilder<> Builder(findInsertPt(BI, Checks));
780   Value *AllChecks = Builder.CreateAnd(Checks);
781   auto *OldCond = BI->getCondition();
782   BI->setCondition(AllChecks);
783   if (InsertAssumesOfPredicatedGuardsConditions) {
784     BasicBlock *IfTrueBB = BI->getSuccessor(0);
785     Builder.SetInsertPoint(IfTrueBB, IfTrueBB->getFirstInsertionPt());
786     // If this block has other predecessors, we might not be able to use Cond.
787     // In this case, create a Phi where every other input is `true` and input
788     // from guard block is Cond.
789     Value *AssumeCond = Builder.CreateAnd(WidenedChecks);
790     if (!IfTrueBB->getUniquePredecessor()) {
791       auto *GuardBB = BI->getParent();
792       auto *PN = Builder.CreatePHI(AssumeCond->getType(), pred_size(IfTrueBB),
793                                    "assume.cond");
794       for (auto *Pred : predecessors(IfTrueBB))
795         PN->addIncoming(Pred == GuardBB ? AssumeCond : Builder.getTrue(), Pred);
796       AssumeCond = PN;
797     }
798     Builder.CreateAssumption(AssumeCond);
799   }
800   RecursivelyDeleteTriviallyDeadInstructions(OldCond, nullptr /* TLI */, MSSAU);
801   assert(isGuardAsWidenableBranch(BI) &&
802          "Stopped being a guard after transform?");
803 
804   LLVM_DEBUG(dbgs() << "Widened checks = " << WidenedChecks.size() << "\n");
805   return true;
806 }
807 
808 std::optional<LoopICmp> LoopPredication::parseLoopLatchICmp() {
809   using namespace PatternMatch;
810 
811   BasicBlock *LoopLatch = L->getLoopLatch();
812   if (!LoopLatch) {
813     LLVM_DEBUG(dbgs() << "The loop doesn't have a single latch!\n");
814     return std::nullopt;
815   }
816 
817   auto *BI = dyn_cast<BranchInst>(LoopLatch->getTerminator());
818   if (!BI || !BI->isConditional()) {
819     LLVM_DEBUG(dbgs() << "Failed to match the latch terminator!\n");
820     return std::nullopt;
821   }
822   BasicBlock *TrueDest = BI->getSuccessor(0);
823   assert(
824       (TrueDest == L->getHeader() || BI->getSuccessor(1) == L->getHeader()) &&
825       "One of the latch's destinations must be the header");
826 
827   auto *ICI = dyn_cast<ICmpInst>(BI->getCondition());
828   if (!ICI) {
829     LLVM_DEBUG(dbgs() << "Failed to match the latch condition!\n");
830     return std::nullopt;
831   }
832   auto Result = parseLoopICmp(ICI);
833   if (!Result) {
834     LLVM_DEBUG(dbgs() << "Failed to parse the loop latch condition!\n");
835     return std::nullopt;
836   }
837 
838   if (TrueDest != L->getHeader())
839     Result->Pred = ICmpInst::getInversePredicate(Result->Pred);
840 
841   // Check affine first, so if it's not we don't try to compute the step
842   // recurrence.
843   if (!Result->IV->isAffine()) {
844     LLVM_DEBUG(dbgs() << "The induction variable is not affine!\n");
845     return std::nullopt;
846   }
847 
848   auto *Step = Result->IV->getStepRecurrence(*SE);
849   if (!isSupportedStep(Step)) {
850     LLVM_DEBUG(dbgs() << "Unsupported loop stride(" << *Step << ")!\n");
851     return std::nullopt;
852   }
853 
854   auto IsUnsupportedPredicate = [](const SCEV *Step, ICmpInst::Predicate Pred) {
855     if (Step->isOne()) {
856       return Pred != ICmpInst::ICMP_ULT && Pred != ICmpInst::ICMP_SLT &&
857              Pred != ICmpInst::ICMP_ULE && Pred != ICmpInst::ICMP_SLE;
858     } else {
859       assert(Step->isAllOnesValue() && "Step should be -1!");
860       return Pred != ICmpInst::ICMP_UGT && Pred != ICmpInst::ICMP_SGT &&
861              Pred != ICmpInst::ICMP_UGE && Pred != ICmpInst::ICMP_SGE;
862     }
863   };
864 
865   normalizePredicate(SE, L, *Result);
866   if (IsUnsupportedPredicate(Step, Result->Pred)) {
867     LLVM_DEBUG(dbgs() << "Unsupported loop latch predicate(" << Result->Pred
868                       << ")!\n");
869     return std::nullopt;
870   }
871 
872   return Result;
873 }
874 
875 bool LoopPredication::isLoopProfitableToPredicate() {
876   if (SkipProfitabilityChecks)
877     return true;
878 
879   SmallVector<std::pair<BasicBlock *, BasicBlock *>, 8> ExitEdges;
880   L->getExitEdges(ExitEdges);
881   // If there is only one exiting edge in the loop, it is always profitable to
882   // predicate the loop.
883   if (ExitEdges.size() == 1)
884     return true;
885 
886   // Calculate the exiting probabilities of all exiting edges from the loop,
887   // starting with the LatchExitProbability.
888   // Heuristic for profitability: If any of the exiting blocks' probability of
889   // exiting the loop is larger than exiting through the latch block, it's not
890   // profitable to predicate the loop.
891   auto *LatchBlock = L->getLoopLatch();
892   assert(LatchBlock && "Should have a single latch at this point!");
893   auto *LatchTerm = LatchBlock->getTerminator();
894   assert(LatchTerm->getNumSuccessors() == 2 &&
895          "expected to be an exiting block with 2 succs!");
896   unsigned LatchBrExitIdx =
897       LatchTerm->getSuccessor(0) == L->getHeader() ? 1 : 0;
898   // We compute branch probabilities without BPI. We do not rely on BPI since
899   // Loop predication is usually run in an LPM and BPI is only preserved
900   // lossily within loop pass managers, while BPI has an inherent notion of
901   // being complete for an entire function.
902 
903   // If the latch exits into a deoptimize or an unreachable block, do not
904   // predicate on that latch check.
905   auto *LatchExitBlock = LatchTerm->getSuccessor(LatchBrExitIdx);
906   if (isa<UnreachableInst>(LatchTerm) ||
907       LatchExitBlock->getTerminatingDeoptimizeCall())
908     return false;
909 
910   // Latch terminator has no valid profile data, so nothing to check
911   // profitability on.
912   if (!hasValidBranchWeightMD(*LatchTerm))
913     return true;
914 
915   auto ComputeBranchProbability =
916       [&](const BasicBlock *ExitingBlock,
917           const BasicBlock *ExitBlock) -> BranchProbability {
918     auto *Term = ExitingBlock->getTerminator();
919     unsigned NumSucc = Term->getNumSuccessors();
920     if (MDNode *ProfileData = getValidBranchWeightMDNode(*Term)) {
921       SmallVector<uint32_t> Weights;
922       extractBranchWeights(ProfileData, Weights);
923       uint64_t Numerator = 0, Denominator = 0;
924       for (auto [i, Weight] : llvm::enumerate(Weights)) {
925         if (Term->getSuccessor(i) == ExitBlock)
926           Numerator += Weight;
927         Denominator += Weight;
928       }
929       // If all weights are zero act as if there was no profile data
930       if (Denominator == 0)
931         return BranchProbability::getBranchProbability(1, NumSucc);
932       return BranchProbability::getBranchProbability(Numerator, Denominator);
933     } else {
934       assert(LatchBlock != ExitingBlock &&
935              "Latch term should always have profile data!");
936       // No profile data, so we choose the weight as 1/num_of_succ(Src)
937       return BranchProbability::getBranchProbability(1, NumSucc);
938     }
939   };
940 
941   BranchProbability LatchExitProbability =
942       ComputeBranchProbability(LatchBlock, LatchExitBlock);
943 
944   // Protect against degenerate inputs provided by the user. Providing a value
945   // less than one, can invert the definition of profitable loop predication.
946   float ScaleFactor = LatchExitProbabilityScale;
947   if (ScaleFactor < 1) {
948     LLVM_DEBUG(
949         dbgs()
950         << "Ignored user setting for loop-predication-latch-probability-scale: "
951         << LatchExitProbabilityScale << "\n");
952     LLVM_DEBUG(dbgs() << "The value is set to 1.0\n");
953     ScaleFactor = 1.0;
954   }
955   const auto LatchProbabilityThreshold = LatchExitProbability * ScaleFactor;
956 
957   for (const auto &ExitEdge : ExitEdges) {
958     BranchProbability ExitingBlockProbability =
959         ComputeBranchProbability(ExitEdge.first, ExitEdge.second);
960     // Some exiting edge has higher probability than the latch exiting edge.
961     // No longer profitable to predicate.
962     if (ExitingBlockProbability > LatchProbabilityThreshold)
963       return false;
964   }
965 
966   // We have concluded that the most probable way to exit from the
967   // loop is through the latch (or there's no profile information and all
968   // exits are equally likely).
969   return true;
970 }
971 
972 /// If we can (cheaply) find a widenable branch which controls entry into the
973 /// loop, return it.
974 static BranchInst *FindWidenableTerminatorAboveLoop(Loop *L, LoopInfo &LI) {
975   // Walk back through any unconditional executed blocks and see if we can find
976   // a widenable condition which seems to control execution of this loop.  Note
977   // that we predict that maythrow calls are likely untaken and thus that it's
978   // profitable to widen a branch before a maythrow call with a condition
979   // afterwards even though that may cause the slow path to run in a case where
980   // it wouldn't have otherwise.
981   BasicBlock *BB = L->getLoopPreheader();
982   if (!BB)
983     return nullptr;
984   do {
985     if (BasicBlock *Pred = BB->getSinglePredecessor())
986       if (BB == Pred->getSingleSuccessor()) {
987         BB = Pred;
988         continue;
989       }
990     break;
991   } while (true);
992 
993   if (BasicBlock *Pred = BB->getSinglePredecessor()) {
994     if (auto *BI = dyn_cast<BranchInst>(Pred->getTerminator()))
995       if (BI->getSuccessor(0) == BB && isWidenableBranch(BI))
996         return BI;
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.
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.
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   bool ChangedLoop = false;
1075 
1076   for (auto *ExitingBB : ExitingBlocks) {
1077     if (LI->getLoopFor(ExitingBB) != L)
1078       continue;
1079 
1080     auto *BI = dyn_cast<BranchInst>(ExitingBB->getTerminator());
1081     if (!BI)
1082       continue;
1083 
1084     if (auto WC = extractWidenableCondition(BI))
1085       if (L->contains(BI->getSuccessor(0))) {
1086         assert(WC->hasOneUse() && "Not appropriate widenable branch!");
1087         WC->user_back()->replaceUsesOfWith(
1088             WC, ConstantInt::getTrue(BI->getContext()));
1089         ChangedLoop = true;
1090       }
1091   }
1092   if (ChangedLoop)
1093     SE->forgetLoop(L);
1094 
1095   // The insertion point for the widening should be at the widenably call, not
1096   // at the WidenableBR. If we do this at the widenableBR, we can incorrectly
1097   // change a loop-invariant condition to a loop-varying one.
1098   auto *IP = cast<Instruction>(WidenableBR->getCondition());
1099 
1100   // The use of umin(all analyzeable exits) instead of latch is subtle, but
1101   // important for profitability.  We may have a loop which hasn't been fully
1102   // canonicalized just yet.  If the exit we chose to widen is provably never
1103   // taken, we want the widened form to *also* be provably never taken.  We
1104   // can't guarantee this as a current unanalyzeable exit may later become
1105   // analyzeable, but we can at least avoid the obvious cases.
1106   const SCEV *MinEC = getMinAnalyzeableBackedgeTakenCount(*SE, *DT, L);
1107   if (isa<SCEVCouldNotCompute>(MinEC) || MinEC->getType()->isPointerTy() ||
1108       !SE->isLoopInvariant(MinEC, L) ||
1109       !Rewriter.isSafeToExpandAt(MinEC, IP))
1110     return ChangedLoop;
1111 
1112   Rewriter.setInsertPoint(IP);
1113   IRBuilder<> B(IP);
1114 
1115   bool InvalidateLoop = false;
1116   Value *MinECV = nullptr; // lazily generated if needed
1117   for (BasicBlock *ExitingBB : ExitingBlocks) {
1118     // If our exiting block exits multiple loops, we can only rewrite the
1119     // innermost one.  Otherwise, we're changing how many times the innermost
1120     // loop runs before it exits.
1121     if (LI->getLoopFor(ExitingBB) != L)
1122       continue;
1123 
1124     // Can't rewrite non-branch yet.
1125     auto *BI = dyn_cast<BranchInst>(ExitingBB->getTerminator());
1126     if (!BI)
1127       continue;
1128 
1129     // If already constant, nothing to do.
1130     if (isa<Constant>(BI->getCondition()))
1131       continue;
1132 
1133     const SCEV *ExitCount = SE->getExitCount(L, ExitingBB);
1134     if (isa<SCEVCouldNotCompute>(ExitCount) ||
1135         ExitCount->getType()->isPointerTy() ||
1136         !Rewriter.isSafeToExpandAt(ExitCount, WidenableBR))
1137       continue;
1138 
1139     const bool ExitIfTrue = !L->contains(*succ_begin(ExitingBB));
1140     BasicBlock *ExitBB = BI->getSuccessor(ExitIfTrue ? 0 : 1);
1141     if (!ExitBB->getPostdominatingDeoptimizeCall())
1142       continue;
1143 
1144     /// Here we can be fairly sure that executing this exit will most likely
1145     /// lead to executing llvm.experimental.deoptimize.
1146     /// This is a profitability heuristic, not a legality constraint.
1147 
1148     // If we found a widenable exit condition, do two things:
1149     // 1) fold the widened exit test into the widenable condition
1150     // 2) fold the branch to untaken - avoids infinite looping
1151 
1152     Value *ECV = Rewriter.expandCodeFor(ExitCount);
1153     if (!MinECV)
1154       MinECV = Rewriter.expandCodeFor(MinEC);
1155     Value *RHS = MinECV;
1156     if (ECV->getType() != RHS->getType()) {
1157       Type *WiderTy = SE->getWiderType(ECV->getType(), RHS->getType());
1158       ECV = B.CreateZExt(ECV, WiderTy);
1159       RHS = B.CreateZExt(RHS, WiderTy);
1160     }
1161     assert(!Latch || DT->dominates(ExitingBB, Latch));
1162     Value *NewCond = B.CreateICmp(ICmpInst::ICMP_UGT, ECV, RHS);
1163     // Freeze poison or undef to an arbitrary bit pattern to ensure we can
1164     // branch without introducing UB.  See NOTE ON POISON/UNDEF above for
1165     // context.
1166     NewCond = B.CreateFreeze(NewCond);
1167 
1168     widenWidenableBranch(WidenableBR, NewCond);
1169 
1170     Value *OldCond = BI->getCondition();
1171     BI->setCondition(ConstantInt::get(OldCond->getType(), !ExitIfTrue));
1172     InvalidateLoop = true;
1173   }
1174 
1175   if (InvalidateLoop)
1176     // We just mutated a bunch of loop exits changing there exit counts
1177     // widely.  We need to force recomputation of the exit counts given these
1178     // changes.  Note that all of the inserted exits are never taken, and
1179     // should be removed next time the CFG is modified.
1180     SE->forgetLoop(L);
1181 
1182   // Always return `true` since we have moved the WidenableBR's condition.
1183   return true;
1184 }
1185 
1186 bool LoopPredication::runOnLoop(Loop *Loop) {
1187   L = Loop;
1188 
1189   LLVM_DEBUG(dbgs() << "Analyzing ");
1190   LLVM_DEBUG(L->dump());
1191 
1192   Module *M = L->getHeader()->getModule();
1193 
1194   // There is nothing to do if the module doesn't use guards
1195   auto *GuardDecl =
1196       M->getFunction(Intrinsic::getName(Intrinsic::experimental_guard));
1197   bool HasIntrinsicGuards = GuardDecl && !GuardDecl->use_empty();
1198   auto *WCDecl = M->getFunction(
1199       Intrinsic::getName(Intrinsic::experimental_widenable_condition));
1200   bool HasWidenableConditions =
1201       PredicateWidenableBranchGuards && WCDecl && !WCDecl->use_empty();
1202   if (!HasIntrinsicGuards && !HasWidenableConditions)
1203     return false;
1204 
1205   DL = &M->getDataLayout();
1206 
1207   Preheader = L->getLoopPreheader();
1208   if (!Preheader)
1209     return false;
1210 
1211   auto LatchCheckOpt = parseLoopLatchICmp();
1212   if (!LatchCheckOpt)
1213     return false;
1214   LatchCheck = *LatchCheckOpt;
1215 
1216   LLVM_DEBUG(dbgs() << "Latch check:\n");
1217   LLVM_DEBUG(LatchCheck.dump());
1218 
1219   if (!isLoopProfitableToPredicate()) {
1220     LLVM_DEBUG(dbgs() << "Loop not profitable to predicate!\n");
1221     return false;
1222   }
1223   // Collect all the guards into a vector and process later, so as not
1224   // to invalidate the instruction iterator.
1225   SmallVector<IntrinsicInst *, 4> Guards;
1226   SmallVector<BranchInst *, 4> GuardsAsWidenableBranches;
1227   for (const auto BB : L->blocks()) {
1228     for (auto &I : *BB)
1229       if (isGuard(&I))
1230         Guards.push_back(cast<IntrinsicInst>(&I));
1231     if (PredicateWidenableBranchGuards &&
1232         isGuardAsWidenableBranch(BB->getTerminator()))
1233       GuardsAsWidenableBranches.push_back(
1234           cast<BranchInst>(BB->getTerminator()));
1235   }
1236 
1237   SCEVExpander Expander(*SE, *DL, "loop-predication");
1238   bool Changed = false;
1239   for (auto *Guard : Guards)
1240     Changed |= widenGuardConditions(Guard, Expander);
1241   for (auto *Guard : GuardsAsWidenableBranches)
1242     Changed |= widenWidenableBranchGuardConditions(Guard, Expander);
1243   Changed |= predicateLoopExits(L, Expander);
1244 
1245   if (MSSAU && VerifyMemorySSA)
1246     MSSAU->getMemorySSA()->verifyMemorySSA();
1247   return Changed;
1248 }
1249