1 //===-- LoopPredication.cpp - Guard based loop predication pass -----------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // The LoopPredication pass tries to convert loop variant range checks to loop
11 // invariant by widening checks across loop iterations. For example, it will
12 // convert
13 //
14 //   for (i = 0; i < n; i++) {
15 //     guard(i < len);
16 //     ...
17 //   }
18 //
19 // to
20 //
21 //   for (i = 0; i < n; i++) {
22 //     guard(n - 1 < len);
23 //     ...
24 //   }
25 //
26 // After this transformation the condition of the guard is loop invariant, so
27 // loop-unswitch can later unswitch the loop by this condition which basically
28 // predicates the loop by the widened condition:
29 //
30 //   if (n - 1 < len)
31 //     for (i = 0; i < n; i++) {
32 //       ...
33 //     }
34 //   else
35 //     deoptimize
36 //
37 // It's tempting to rely on SCEV here, but it has proven to be problematic.
38 // Generally the facts SCEV provides about the increment step of add
39 // recurrences are true if the backedge of the loop is taken, which implicitly
40 // assumes that the guard doesn't fail. Using these facts to optimize the
41 // guard results in a circular logic where the guard is optimized under the
42 // assumption that it never fails.
43 //
44 // For example, in the loop below the induction variable will be marked as nuw
45 // basing on the guard. Basing on nuw the guard predicate will be considered
46 // monotonic. Given a monotonic condition it's tempting to replace the induction
47 // variable in the condition with its value on the last iteration. But this
48 // transformation is not correct, e.g. e = 4, b = 5 breaks the loop.
49 //
50 //   for (int i = b; i != e; i++)
51 //     guard(i u< len)
52 //
53 // One of the ways to reason about this problem is to use an inductive proof
54 // approach. Given the loop:
55 //
56 //   if (B(0)) {
57 //     do {
58 //       I = PHI(0, I.INC)
59 //       I.INC = I + Step
60 //       guard(G(I));
61 //     } while (B(I));
62 //   }
63 //
64 // where B(x) and G(x) are predicates that map integers to booleans, we want a
65 // loop invariant expression M such the following program has the same semantics
66 // as the above:
67 //
68 //   if (B(0)) {
69 //     do {
70 //       I = PHI(0, I.INC)
71 //       I.INC = I + Step
72 //       guard(G(0) && M);
73 //     } while (B(I));
74 //   }
75 //
76 // One solution for M is M = forall X . (G(X) && B(X)) => G(X + Step)
77 //
78 // Informal proof that the transformation above is correct:
79 //
80 //   By the definition of guards we can rewrite the guard condition to:
81 //     G(I) && G(0) && M
82 //
83 //   Let's prove that for each iteration of the loop:
84 //     G(0) && M => G(I)
85 //   And the condition above can be simplified to G(Start) && M.
86 //
87 //   Induction base.
88 //     G(0) && M => G(0)
89 //
90 //   Induction step. Assuming G(0) && M => G(I) on the subsequent
91 //   iteration:
92 //
93 //     B(I) is true because it's the backedge condition.
94 //     G(I) is true because the backedge is guarded by this condition.
95 //
96 //   So M = forall X . (G(X) && B(X)) => G(X + Step) implies G(I + Step).
97 //
98 // Note that we can use anything stronger than M, i.e. any condition which
99 // implies M.
100 //
101 // When S = 1 (i.e. forward iterating loop), the transformation is supported
102 // when:
103 //   * The loop has a single latch with the condition of the form:
104 //     B(X) = latchStart + X <pred> latchLimit,
105 //     where <pred> is u<, u<=, s<, or s<=.
106 //   * The guard condition is of the form
107 //     G(X) = guardStart + X u< guardLimit
108 //
109 //   For the ult latch comparison case M is:
110 //     forall X . guardStart + X u< guardLimit && latchStart + X <u latchLimit =>
111 //        guardStart + X + 1 u< guardLimit
112 //
113 //   The only way the antecedent can be true and the consequent can be false is
114 //   if
115 //     X == guardLimit - 1 - guardStart
116 //   (and guardLimit is non-zero, but we won't use this latter fact).
117 //   If X == guardLimit - 1 - guardStart then the second half of the antecedent is
118 //     latchStart + guardLimit - 1 - guardStart u< latchLimit
119 //   and its negation is
120 //     latchStart + guardLimit - 1 - guardStart u>= latchLimit
121 //
122 //   In other words, if
123 //     latchLimit u<= latchStart + guardLimit - 1 - guardStart
124 //   then:
125 //   (the ranges below are written in ConstantRange notation, where [A, B) is the
126 //   set for (I = A; I != B; I++ /*maywrap*/) yield(I);)
127 //
128 //      forall X . guardStart + X u< guardLimit &&
129 //                 latchStart + X u< latchLimit =>
130 //        guardStart + X + 1 u< guardLimit
131 //   == forall X . guardStart + X u< guardLimit &&
132 //                 latchStart + X u< latchStart + guardLimit - 1 - guardStart =>
133 //        guardStart + X + 1 u< guardLimit
134 //   == forall X . (guardStart + X) in [0, guardLimit) &&
135 //                 (latchStart + X) in [0, latchStart + guardLimit - 1 - guardStart) =>
136 //        (guardStart + X + 1) in [0, guardLimit)
137 //   == forall X . X in [-guardStart, guardLimit - guardStart) &&
138 //                 X in [-latchStart, guardLimit - 1 - guardStart) =>
139 //         X in [-guardStart - 1, guardLimit - guardStart - 1)
140 //   == true
141 //
142 //   So the widened condition is:
143 //     guardStart u< guardLimit &&
144 //     latchStart + guardLimit - 1 - guardStart u>= latchLimit
145 //   Similarly for ule condition the widened condition is:
146 //     guardStart u< guardLimit &&
147 //     latchStart + guardLimit - 1 - guardStart u> latchLimit
148 //   For slt condition the widened condition is:
149 //     guardStart u< guardLimit &&
150 //     latchStart + guardLimit - 1 - guardStart s>= latchLimit
151 //   For sle condition the widened condition is:
152 //     guardStart u< guardLimit &&
153 //     latchStart + guardLimit - 1 - guardStart s> latchLimit
154 //
155 // When S = -1 (i.e. reverse iterating loop), the transformation is supported
156 // when:
157 //   * The loop has a single latch with the condition of the form:
158 //     B(X) = X <pred> latchLimit, where <pred> is u>, u>=, s>, or s>=.
159 //   * The guard condition is of the form
160 //     G(X) = X - 1 u< guardLimit
161 //
162 //   For the ugt latch comparison case M is:
163 //     forall X. X-1 u< guardLimit and X u> latchLimit => X-2 u< guardLimit
164 //
165 //   The only way the antecedent can be true and the consequent can be false is if
166 //     X == 1.
167 //   If X == 1 then the second half of the antecedent is
168 //     1 u> latchLimit, and its negation is latchLimit u>= 1.
169 //
170 //   So the widened condition is:
171 //     guardStart u< guardLimit && latchLimit u>= 1.
172 //   Similarly for sgt condition the widened condition is:
173 //     guardStart u< guardLimit && latchLimit s>= 1.
174 //   For uge condition the widened condition is:
175 //     guardStart u< guardLimit && latchLimit u> 1.
176 //   For sge condition the widened condition is:
177 //     guardStart u< guardLimit && latchLimit s> 1.
178 //===----------------------------------------------------------------------===//
179 
180 #include "llvm/Transforms/Scalar/LoopPredication.h"
181 #include "llvm/ADT/Statistic.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/ScalarEvolutionExpander.h"
188 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
189 #include "llvm/IR/Function.h"
190 #include "llvm/IR/GlobalValue.h"
191 #include "llvm/IR/IntrinsicInst.h"
192 #include "llvm/IR/Module.h"
193 #include "llvm/IR/PatternMatch.h"
194 #include "llvm/Pass.h"
195 #include "llvm/Support/Debug.h"
196 #include "llvm/Transforms/Scalar.h"
197 #include "llvm/Transforms/Utils/LoopUtils.h"
198 
199 #define DEBUG_TYPE "loop-predication"
200 
201 STATISTIC(TotalConsidered, "Number of guards considered");
202 STATISTIC(TotalWidened, "Number of checks widened");
203 
204 using namespace llvm;
205 
206 static cl::opt<bool> EnableIVTruncation("loop-predication-enable-iv-truncation",
207                                         cl::Hidden, cl::init(true));
208 
209 static cl::opt<bool> EnableCountDownLoop("loop-predication-enable-count-down-loop",
210                                         cl::Hidden, cl::init(true));
211 
212 static cl::opt<bool>
213     SkipProfitabilityChecks("loop-predication-skip-profitability-checks",
214                             cl::Hidden, cl::init(false));
215 
216 // This is the scale factor for the latch probability. We use this during
217 // profitability analysis to find other exiting blocks that have a much higher
218 // probability of exiting the loop instead of loop exiting via latch.
219 // This value should be greater than 1 for a sane profitability check.
220 static cl::opt<float> LatchExitProbabilityScale(
221     "loop-predication-latch-probability-scale", cl::Hidden, cl::init(2.0),
222     cl::desc("scale factor for the latch probability. Value should be greater "
223              "than 1. Lower values are ignored"));
224 
225 namespace {
226 class LoopPredication {
227   /// Represents an induction variable check:
228   ///   icmp Pred, <induction variable>, <loop invariant limit>
229   struct LoopICmp {
230     ICmpInst::Predicate Pred;
231     const SCEVAddRecExpr *IV;
232     const SCEV *Limit;
LoopICmp__anonaf4e6ee30111::LoopPredication::LoopICmp233     LoopICmp(ICmpInst::Predicate Pred, const SCEVAddRecExpr *IV,
234              const SCEV *Limit)
235         : Pred(Pred), IV(IV), Limit(Limit) {}
LoopICmp__anonaf4e6ee30111::LoopPredication::LoopICmp236     LoopICmp() {}
dump__anonaf4e6ee30111::LoopPredication::LoopICmp237     void dump() {
238       dbgs() << "LoopICmp Pred = " << Pred << ", IV = " << *IV
239              << ", Limit = " << *Limit << "\n";
240     }
241   };
242 
243   ScalarEvolution *SE;
244   BranchProbabilityInfo *BPI;
245 
246   Loop *L;
247   const DataLayout *DL;
248   BasicBlock *Preheader;
249   LoopICmp LatchCheck;
250 
251   bool isSupportedStep(const SCEV* Step);
parseLoopICmp(ICmpInst * ICI)252   Optional<LoopICmp> parseLoopICmp(ICmpInst *ICI) {
253     return parseLoopICmp(ICI->getPredicate(), ICI->getOperand(0),
254                          ICI->getOperand(1));
255   }
256   Optional<LoopICmp> parseLoopICmp(ICmpInst::Predicate Pred, Value *LHS,
257                                    Value *RHS);
258 
259   Optional<LoopICmp> parseLoopLatchICmp();
260 
261   bool CanExpand(const SCEV* S);
262   Value *expandCheck(SCEVExpander &Expander, IRBuilder<> &Builder,
263                      ICmpInst::Predicate Pred, const SCEV *LHS, const SCEV *RHS,
264                      Instruction *InsertAt);
265 
266   Optional<Value *> widenICmpRangeCheck(ICmpInst *ICI, SCEVExpander &Expander,
267                                         IRBuilder<> &Builder);
268   Optional<Value *> widenICmpRangeCheckIncrementingLoop(LoopICmp LatchCheck,
269                                                         LoopICmp RangeCheck,
270                                                         SCEVExpander &Expander,
271                                                         IRBuilder<> &Builder);
272   Optional<Value *> widenICmpRangeCheckDecrementingLoop(LoopICmp LatchCheck,
273                                                         LoopICmp RangeCheck,
274                                                         SCEVExpander &Expander,
275                                                         IRBuilder<> &Builder);
276   bool widenGuardConditions(IntrinsicInst *II, SCEVExpander &Expander);
277 
278   // If the loop always exits through another block in the loop, we should not
279   // predicate based on the latch check. For example, the latch check can be a
280   // very coarse grained check and there can be more fine grained exit checks
281   // within the loop. We identify such unprofitable loops through BPI.
282   bool isLoopProfitableToPredicate();
283 
284   // When the IV type is wider than the range operand type, we can still do loop
285   // predication, by generating SCEVs for the range and latch that are of the
286   // same type. We achieve this by generating a SCEV truncate expression for the
287   // latch IV. This is done iff truncation of the IV is a safe operation,
288   // without loss of information.
289   // Another way to achieve this is by generating a wider type SCEV for the
290   // range check operand, however, this needs a more involved check that
291   // operands do not overflow. This can lead to loss of information when the
292   // range operand is of the form: add i32 %offset, %iv. We need to prove that
293   // sext(x + y) is same as sext(x) + sext(y).
294   // This function returns true if we can safely represent the IV type in
295   // the RangeCheckType without loss of information.
296   bool isSafeToTruncateWideIVType(Type *RangeCheckType);
297   // Return the loopLatchCheck corresponding to the RangeCheckType if safe to do
298   // so.
299   Optional<LoopICmp> generateLoopLatchCheck(Type *RangeCheckType);
300 
301 public:
LoopPredication(ScalarEvolution * SE,BranchProbabilityInfo * BPI)302   LoopPredication(ScalarEvolution *SE, BranchProbabilityInfo *BPI)
303       : SE(SE), BPI(BPI){};
304   bool runOnLoop(Loop *L);
305 };
306 
307 class LoopPredicationLegacyPass : public LoopPass {
308 public:
309   static char ID;
LoopPredicationLegacyPass()310   LoopPredicationLegacyPass() : LoopPass(ID) {
311     initializeLoopPredicationLegacyPassPass(*PassRegistry::getPassRegistry());
312   }
313 
getAnalysisUsage(AnalysisUsage & AU) const314   void getAnalysisUsage(AnalysisUsage &AU) const override {
315     AU.addRequired<BranchProbabilityInfoWrapperPass>();
316     getLoopAnalysisUsage(AU);
317   }
318 
runOnLoop(Loop * L,LPPassManager & LPM)319   bool runOnLoop(Loop *L, LPPassManager &LPM) override {
320     if (skipLoop(L))
321       return false;
322     auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
323     BranchProbabilityInfo &BPI =
324         getAnalysis<BranchProbabilityInfoWrapperPass>().getBPI();
325     LoopPredication LP(SE, &BPI);
326     return LP.runOnLoop(L);
327   }
328 };
329 
330 char LoopPredicationLegacyPass::ID = 0;
331 } // end namespace llvm
332 
333 INITIALIZE_PASS_BEGIN(LoopPredicationLegacyPass, "loop-predication",
334                       "Loop predication", false, false)
INITIALIZE_PASS_DEPENDENCY(BranchProbabilityInfoWrapperPass)335 INITIALIZE_PASS_DEPENDENCY(BranchProbabilityInfoWrapperPass)
336 INITIALIZE_PASS_DEPENDENCY(LoopPass)
337 INITIALIZE_PASS_END(LoopPredicationLegacyPass, "loop-predication",
338                     "Loop predication", false, false)
339 
340 Pass *llvm::createLoopPredicationPass() {
341   return new LoopPredicationLegacyPass();
342 }
343 
run(Loop & L,LoopAnalysisManager & AM,LoopStandardAnalysisResults & AR,LPMUpdater & U)344 PreservedAnalyses LoopPredicationPass::run(Loop &L, LoopAnalysisManager &AM,
345                                            LoopStandardAnalysisResults &AR,
346                                            LPMUpdater &U) {
347   const auto &FAM =
348       AM.getResult<FunctionAnalysisManagerLoopProxy>(L, AR).getManager();
349   Function *F = L.getHeader()->getParent();
350   auto *BPI = FAM.getCachedResult<BranchProbabilityAnalysis>(*F);
351   LoopPredication LP(&AR.SE, BPI);
352   if (!LP.runOnLoop(&L))
353     return PreservedAnalyses::all();
354 
355   return getLoopPassPreservedAnalyses();
356 }
357 
358 Optional<LoopPredication::LoopICmp>
parseLoopICmp(ICmpInst::Predicate Pred,Value * LHS,Value * RHS)359 LoopPredication::parseLoopICmp(ICmpInst::Predicate Pred, Value *LHS,
360                                Value *RHS) {
361   const SCEV *LHSS = SE->getSCEV(LHS);
362   if (isa<SCEVCouldNotCompute>(LHSS))
363     return None;
364   const SCEV *RHSS = SE->getSCEV(RHS);
365   if (isa<SCEVCouldNotCompute>(RHSS))
366     return None;
367 
368   // Canonicalize RHS to be loop invariant bound, LHS - a loop computable IV
369   if (SE->isLoopInvariant(LHSS, L)) {
370     std::swap(LHS, RHS);
371     std::swap(LHSS, RHSS);
372     Pred = ICmpInst::getSwappedPredicate(Pred);
373   }
374 
375   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHSS);
376   if (!AR || AR->getLoop() != L)
377     return None;
378 
379   return LoopICmp(Pred, AR, RHSS);
380 }
381 
expandCheck(SCEVExpander & Expander,IRBuilder<> & Builder,ICmpInst::Predicate Pred,const SCEV * LHS,const SCEV * RHS,Instruction * InsertAt)382 Value *LoopPredication::expandCheck(SCEVExpander &Expander,
383                                     IRBuilder<> &Builder,
384                                     ICmpInst::Predicate Pred, const SCEV *LHS,
385                                     const SCEV *RHS, Instruction *InsertAt) {
386   // TODO: we can check isLoopEntryGuardedByCond before emitting the check
387 
388   Type *Ty = LHS->getType();
389   assert(Ty == RHS->getType() && "expandCheck operands have different types?");
390 
391   if (SE->isLoopEntryGuardedByCond(L, Pred, LHS, RHS))
392     return Builder.getTrue();
393 
394   Value *LHSV = Expander.expandCodeFor(LHS, Ty, InsertAt);
395   Value *RHSV = Expander.expandCodeFor(RHS, Ty, InsertAt);
396   return Builder.CreateICmp(Pred, LHSV, RHSV);
397 }
398 
399 Optional<LoopPredication::LoopICmp>
generateLoopLatchCheck(Type * RangeCheckType)400 LoopPredication::generateLoopLatchCheck(Type *RangeCheckType) {
401 
402   auto *LatchType = LatchCheck.IV->getType();
403   if (RangeCheckType == LatchType)
404     return LatchCheck;
405   // For now, bail out if latch type is narrower than range type.
406   if (DL->getTypeSizeInBits(LatchType) < DL->getTypeSizeInBits(RangeCheckType))
407     return None;
408   if (!isSafeToTruncateWideIVType(RangeCheckType))
409     return None;
410   // We can now safely identify the truncated version of the IV and limit for
411   // RangeCheckType.
412   LoopICmp NewLatchCheck;
413   NewLatchCheck.Pred = LatchCheck.Pred;
414   NewLatchCheck.IV = dyn_cast<SCEVAddRecExpr>(
415       SE->getTruncateExpr(LatchCheck.IV, RangeCheckType));
416   if (!NewLatchCheck.IV)
417     return None;
418   NewLatchCheck.Limit = SE->getTruncateExpr(LatchCheck.Limit, RangeCheckType);
419   LLVM_DEBUG(dbgs() << "IV of type: " << *LatchType
420                     << "can be represented as range check type:"
421                     << *RangeCheckType << "\n");
422   LLVM_DEBUG(dbgs() << "LatchCheck.IV: " << *NewLatchCheck.IV << "\n");
423   LLVM_DEBUG(dbgs() << "LatchCheck.Limit: " << *NewLatchCheck.Limit << "\n");
424   return NewLatchCheck;
425 }
426 
isSupportedStep(const SCEV * Step)427 bool LoopPredication::isSupportedStep(const SCEV* Step) {
428   return Step->isOne() || (Step->isAllOnesValue() && EnableCountDownLoop);
429 }
430 
CanExpand(const SCEV * S)431 bool LoopPredication::CanExpand(const SCEV* S) {
432   return SE->isLoopInvariant(S, L) && isSafeToExpand(S, *SE);
433 }
434 
widenICmpRangeCheckIncrementingLoop(LoopPredication::LoopICmp LatchCheck,LoopPredication::LoopICmp RangeCheck,SCEVExpander & Expander,IRBuilder<> & Builder)435 Optional<Value *> LoopPredication::widenICmpRangeCheckIncrementingLoop(
436     LoopPredication::LoopICmp LatchCheck, LoopPredication::LoopICmp RangeCheck,
437     SCEVExpander &Expander, IRBuilder<> &Builder) {
438   auto *Ty = RangeCheck.IV->getType();
439   // Generate the widened condition for the forward loop:
440   //   guardStart u< guardLimit &&
441   //   latchLimit <pred> guardLimit - 1 - guardStart + latchStart
442   // where <pred> depends on the latch condition predicate. See the file
443   // header comment for the reasoning.
444   // guardLimit - guardStart + latchStart - 1
445   const SCEV *GuardStart = RangeCheck.IV->getStart();
446   const SCEV *GuardLimit = RangeCheck.Limit;
447   const SCEV *LatchStart = LatchCheck.IV->getStart();
448   const SCEV *LatchLimit = LatchCheck.Limit;
449 
450   // guardLimit - guardStart + latchStart - 1
451   const SCEV *RHS =
452       SE->getAddExpr(SE->getMinusSCEV(GuardLimit, GuardStart),
453                      SE->getMinusSCEV(LatchStart, SE->getOne(Ty)));
454   if (!CanExpand(GuardStart) || !CanExpand(GuardLimit) ||
455       !CanExpand(LatchLimit) || !CanExpand(RHS)) {
456     LLVM_DEBUG(dbgs() << "Can't expand limit check!\n");
457     return None;
458   }
459   auto LimitCheckPred =
460       ICmpInst::getFlippedStrictnessPredicate(LatchCheck.Pred);
461 
462   LLVM_DEBUG(dbgs() << "LHS: " << *LatchLimit << "\n");
463   LLVM_DEBUG(dbgs() << "RHS: " << *RHS << "\n");
464   LLVM_DEBUG(dbgs() << "Pred: " << LimitCheckPred << "\n");
465 
466   Instruction *InsertAt = Preheader->getTerminator();
467   auto *LimitCheck =
468       expandCheck(Expander, Builder, LimitCheckPred, LatchLimit, RHS, InsertAt);
469   auto *FirstIterationCheck = expandCheck(Expander, Builder, RangeCheck.Pred,
470                                           GuardStart, GuardLimit, InsertAt);
471   return Builder.CreateAnd(FirstIterationCheck, LimitCheck);
472 }
473 
widenICmpRangeCheckDecrementingLoop(LoopPredication::LoopICmp LatchCheck,LoopPredication::LoopICmp RangeCheck,SCEVExpander & Expander,IRBuilder<> & Builder)474 Optional<Value *> LoopPredication::widenICmpRangeCheckDecrementingLoop(
475     LoopPredication::LoopICmp LatchCheck, LoopPredication::LoopICmp RangeCheck,
476     SCEVExpander &Expander, IRBuilder<> &Builder) {
477   auto *Ty = RangeCheck.IV->getType();
478   const SCEV *GuardStart = RangeCheck.IV->getStart();
479   const SCEV *GuardLimit = RangeCheck.Limit;
480   const SCEV *LatchLimit = LatchCheck.Limit;
481   if (!CanExpand(GuardStart) || !CanExpand(GuardLimit) ||
482       !CanExpand(LatchLimit)) {
483     LLVM_DEBUG(dbgs() << "Can't expand limit check!\n");
484     return None;
485   }
486   // The decrement of the latch check IV should be the same as the
487   // rangeCheckIV.
488   auto *PostDecLatchCheckIV = LatchCheck.IV->getPostIncExpr(*SE);
489   if (RangeCheck.IV != PostDecLatchCheckIV) {
490     LLVM_DEBUG(dbgs() << "Not the same. PostDecLatchCheckIV: "
491                       << *PostDecLatchCheckIV
492                       << "  and RangeCheckIV: " << *RangeCheck.IV << "\n");
493     return None;
494   }
495 
496   // Generate the widened condition for CountDownLoop:
497   // guardStart u< guardLimit &&
498   // latchLimit <pred> 1.
499   // See the header comment for reasoning of the checks.
500   Instruction *InsertAt = Preheader->getTerminator();
501   auto LimitCheckPred =
502       ICmpInst::getFlippedStrictnessPredicate(LatchCheck.Pred);
503   auto *FirstIterationCheck = expandCheck(Expander, Builder, ICmpInst::ICMP_ULT,
504                                           GuardStart, GuardLimit, InsertAt);
505   auto *LimitCheck = expandCheck(Expander, Builder, LimitCheckPred, LatchLimit,
506                                  SE->getOne(Ty), InsertAt);
507   return Builder.CreateAnd(FirstIterationCheck, LimitCheck);
508 }
509 
510 /// If ICI can be widened to a loop invariant condition emits the loop
511 /// invariant condition in the loop preheader and return it, otherwise
512 /// returns None.
widenICmpRangeCheck(ICmpInst * ICI,SCEVExpander & Expander,IRBuilder<> & Builder)513 Optional<Value *> LoopPredication::widenICmpRangeCheck(ICmpInst *ICI,
514                                                        SCEVExpander &Expander,
515                                                        IRBuilder<> &Builder) {
516   LLVM_DEBUG(dbgs() << "Analyzing ICmpInst condition:\n");
517   LLVM_DEBUG(ICI->dump());
518 
519   // parseLoopStructure guarantees that the latch condition is:
520   //   ++i <pred> latchLimit, where <pred> is u<, u<=, s<, or s<=.
521   // We are looking for the range checks of the form:
522   //   i u< guardLimit
523   auto RangeCheck = parseLoopICmp(ICI);
524   if (!RangeCheck) {
525     LLVM_DEBUG(dbgs() << "Failed to parse the loop latch condition!\n");
526     return None;
527   }
528   LLVM_DEBUG(dbgs() << "Guard check:\n");
529   LLVM_DEBUG(RangeCheck->dump());
530   if (RangeCheck->Pred != ICmpInst::ICMP_ULT) {
531     LLVM_DEBUG(dbgs() << "Unsupported range check predicate("
532                       << RangeCheck->Pred << ")!\n");
533     return None;
534   }
535   auto *RangeCheckIV = RangeCheck->IV;
536   if (!RangeCheckIV->isAffine()) {
537     LLVM_DEBUG(dbgs() << "Range check IV is not affine!\n");
538     return None;
539   }
540   auto *Step = RangeCheckIV->getStepRecurrence(*SE);
541   // We cannot just compare with latch IV step because the latch and range IVs
542   // may have different types.
543   if (!isSupportedStep(Step)) {
544     LLVM_DEBUG(dbgs() << "Range check and latch have IVs different steps!\n");
545     return None;
546   }
547   auto *Ty = RangeCheckIV->getType();
548   auto CurrLatchCheckOpt = generateLoopLatchCheck(Ty);
549   if (!CurrLatchCheckOpt) {
550     LLVM_DEBUG(dbgs() << "Failed to generate a loop latch check "
551                          "corresponding to range type: "
552                       << *Ty << "\n");
553     return None;
554   }
555 
556   LoopICmp CurrLatchCheck = *CurrLatchCheckOpt;
557   // At this point, the range and latch step should have the same type, but need
558   // not have the same value (we support both 1 and -1 steps).
559   assert(Step->getType() ==
560              CurrLatchCheck.IV->getStepRecurrence(*SE)->getType() &&
561          "Range and latch steps should be of same type!");
562   if (Step != CurrLatchCheck.IV->getStepRecurrence(*SE)) {
563     LLVM_DEBUG(dbgs() << "Range and latch have different step values!\n");
564     return None;
565   }
566 
567   if (Step->isOne())
568     return widenICmpRangeCheckIncrementingLoop(CurrLatchCheck, *RangeCheck,
569                                                Expander, Builder);
570   else {
571     assert(Step->isAllOnesValue() && "Step should be -1!");
572     return widenICmpRangeCheckDecrementingLoop(CurrLatchCheck, *RangeCheck,
573                                                Expander, Builder);
574   }
575 }
576 
widenGuardConditions(IntrinsicInst * Guard,SCEVExpander & Expander)577 bool LoopPredication::widenGuardConditions(IntrinsicInst *Guard,
578                                            SCEVExpander &Expander) {
579   LLVM_DEBUG(dbgs() << "Processing guard:\n");
580   LLVM_DEBUG(Guard->dump());
581 
582   TotalConsidered++;
583 
584   IRBuilder<> Builder(cast<Instruction>(Preheader->getTerminator()));
585 
586   // The guard condition is expected to be in form of:
587   //   cond1 && cond2 && cond3 ...
588   // Iterate over subconditions looking for icmp conditions which can be
589   // widened across loop iterations. Widening these conditions remember the
590   // resulting list of subconditions in Checks vector.
591   SmallVector<Value *, 4> Worklist(1, Guard->getOperand(0));
592   SmallPtrSet<Value *, 4> Visited;
593 
594   SmallVector<Value *, 4> Checks;
595 
596   unsigned NumWidened = 0;
597   do {
598     Value *Condition = Worklist.pop_back_val();
599     if (!Visited.insert(Condition).second)
600       continue;
601 
602     Value *LHS, *RHS;
603     using namespace llvm::PatternMatch;
604     if (match(Condition, m_And(m_Value(LHS), m_Value(RHS)))) {
605       Worklist.push_back(LHS);
606       Worklist.push_back(RHS);
607       continue;
608     }
609 
610     if (ICmpInst *ICI = dyn_cast<ICmpInst>(Condition)) {
611       if (auto NewRangeCheck = widenICmpRangeCheck(ICI, Expander, Builder)) {
612         Checks.push_back(NewRangeCheck.getValue());
613         NumWidened++;
614         continue;
615       }
616     }
617 
618     // Save the condition as is if we can't widen it
619     Checks.push_back(Condition);
620   } while (Worklist.size() != 0);
621 
622   if (NumWidened == 0)
623     return false;
624 
625   TotalWidened += NumWidened;
626 
627   // Emit the new guard condition
628   Builder.SetInsertPoint(Guard);
629   Value *LastCheck = nullptr;
630   for (auto *Check : Checks)
631     if (!LastCheck)
632       LastCheck = Check;
633     else
634       LastCheck = Builder.CreateAnd(LastCheck, Check);
635   Guard->setOperand(0, LastCheck);
636 
637   LLVM_DEBUG(dbgs() << "Widened checks = " << NumWidened << "\n");
638   return true;
639 }
640 
parseLoopLatchICmp()641 Optional<LoopPredication::LoopICmp> LoopPredication::parseLoopLatchICmp() {
642   using namespace PatternMatch;
643 
644   BasicBlock *LoopLatch = L->getLoopLatch();
645   if (!LoopLatch) {
646     LLVM_DEBUG(dbgs() << "The loop doesn't have a single latch!\n");
647     return None;
648   }
649 
650   ICmpInst::Predicate Pred;
651   Value *LHS, *RHS;
652   BasicBlock *TrueDest, *FalseDest;
653 
654   if (!match(LoopLatch->getTerminator(),
655              m_Br(m_ICmp(Pred, m_Value(LHS), m_Value(RHS)), TrueDest,
656                   FalseDest))) {
657     LLVM_DEBUG(dbgs() << "Failed to match the latch terminator!\n");
658     return None;
659   }
660   assert((TrueDest == L->getHeader() || FalseDest == L->getHeader()) &&
661          "One of the latch's destinations must be the header");
662   if (TrueDest != L->getHeader())
663     Pred = ICmpInst::getInversePredicate(Pred);
664 
665   auto Result = parseLoopICmp(Pred, LHS, RHS);
666   if (!Result) {
667     LLVM_DEBUG(dbgs() << "Failed to parse the loop latch condition!\n");
668     return None;
669   }
670 
671   // Check affine first, so if it's not we don't try to compute the step
672   // recurrence.
673   if (!Result->IV->isAffine()) {
674     LLVM_DEBUG(dbgs() << "The induction variable is not affine!\n");
675     return None;
676   }
677 
678   auto *Step = Result->IV->getStepRecurrence(*SE);
679   if (!isSupportedStep(Step)) {
680     LLVM_DEBUG(dbgs() << "Unsupported loop stride(" << *Step << ")!\n");
681     return None;
682   }
683 
684   auto IsUnsupportedPredicate = [](const SCEV *Step, ICmpInst::Predicate Pred) {
685     if (Step->isOne()) {
686       return Pred != ICmpInst::ICMP_ULT && Pred != ICmpInst::ICMP_SLT &&
687              Pred != ICmpInst::ICMP_ULE && Pred != ICmpInst::ICMP_SLE;
688     } else {
689       assert(Step->isAllOnesValue() && "Step should be -1!");
690       return Pred != ICmpInst::ICMP_UGT && Pred != ICmpInst::ICMP_SGT &&
691              Pred != ICmpInst::ICMP_UGE && Pred != ICmpInst::ICMP_SGE;
692     }
693   };
694 
695   if (IsUnsupportedPredicate(Step, Result->Pred)) {
696     LLVM_DEBUG(dbgs() << "Unsupported loop latch predicate(" << Result->Pred
697                       << ")!\n");
698     return None;
699   }
700   return Result;
701 }
702 
703 // Returns true if its safe to truncate the IV to RangeCheckType.
isSafeToTruncateWideIVType(Type * RangeCheckType)704 bool LoopPredication::isSafeToTruncateWideIVType(Type *RangeCheckType) {
705   if (!EnableIVTruncation)
706     return false;
707   assert(DL->getTypeSizeInBits(LatchCheck.IV->getType()) >
708              DL->getTypeSizeInBits(RangeCheckType) &&
709          "Expected latch check IV type to be larger than range check operand "
710          "type!");
711   // The start and end values of the IV should be known. This is to guarantee
712   // that truncating the wide type will not lose information.
713   auto *Limit = dyn_cast<SCEVConstant>(LatchCheck.Limit);
714   auto *Start = dyn_cast<SCEVConstant>(LatchCheck.IV->getStart());
715   if (!Limit || !Start)
716     return false;
717   // This check makes sure that the IV does not change sign during loop
718   // iterations. Consider latchType = i64, LatchStart = 5, Pred = ICMP_SGE,
719   // LatchEnd = 2, rangeCheckType = i32. If it's not a monotonic predicate, the
720   // IV wraps around, and the truncation of the IV would lose the range of
721   // iterations between 2^32 and 2^64.
722   bool Increasing;
723   if (!SE->isMonotonicPredicate(LatchCheck.IV, LatchCheck.Pred, Increasing))
724     return false;
725   // The active bits should be less than the bits in the RangeCheckType. This
726   // guarantees that truncating the latch check to RangeCheckType is a safe
727   // operation.
728   auto RangeCheckTypeBitSize = DL->getTypeSizeInBits(RangeCheckType);
729   return Start->getAPInt().getActiveBits() < RangeCheckTypeBitSize &&
730          Limit->getAPInt().getActiveBits() < RangeCheckTypeBitSize;
731 }
732 
isLoopProfitableToPredicate()733 bool LoopPredication::isLoopProfitableToPredicate() {
734   if (SkipProfitabilityChecks || !BPI)
735     return true;
736 
737   SmallVector<std::pair<const BasicBlock *, const BasicBlock *>, 8> ExitEdges;
738   L->getExitEdges(ExitEdges);
739   // If there is only one exiting edge in the loop, it is always profitable to
740   // predicate the loop.
741   if (ExitEdges.size() == 1)
742     return true;
743 
744   // Calculate the exiting probabilities of all exiting edges from the loop,
745   // starting with the LatchExitProbability.
746   // Heuristic for profitability: If any of the exiting blocks' probability of
747   // exiting the loop is larger than exiting through the latch block, it's not
748   // profitable to predicate the loop.
749   auto *LatchBlock = L->getLoopLatch();
750   assert(LatchBlock && "Should have a single latch at this point!");
751   auto *LatchTerm = LatchBlock->getTerminator();
752   assert(LatchTerm->getNumSuccessors() == 2 &&
753          "expected to be an exiting block with 2 succs!");
754   unsigned LatchBrExitIdx =
755       LatchTerm->getSuccessor(0) == L->getHeader() ? 1 : 0;
756   BranchProbability LatchExitProbability =
757       BPI->getEdgeProbability(LatchBlock, LatchBrExitIdx);
758 
759   // Protect against degenerate inputs provided by the user. Providing a value
760   // less than one, can invert the definition of profitable loop predication.
761   float ScaleFactor = LatchExitProbabilityScale;
762   if (ScaleFactor < 1) {
763     LLVM_DEBUG(
764         dbgs()
765         << "Ignored user setting for loop-predication-latch-probability-scale: "
766         << LatchExitProbabilityScale << "\n");
767     LLVM_DEBUG(dbgs() << "The value is set to 1.0\n");
768     ScaleFactor = 1.0;
769   }
770   const auto LatchProbabilityThreshold =
771       LatchExitProbability * ScaleFactor;
772 
773   for (const auto &ExitEdge : ExitEdges) {
774     BranchProbability ExitingBlockProbability =
775         BPI->getEdgeProbability(ExitEdge.first, ExitEdge.second);
776     // Some exiting edge has higher probability than the latch exiting edge.
777     // No longer profitable to predicate.
778     if (ExitingBlockProbability > LatchProbabilityThreshold)
779       return false;
780   }
781   // Using BPI, we have concluded that the most probable way to exit from the
782   // loop is through the latch (or there's no profile information and all
783   // exits are equally likely).
784   return true;
785 }
786 
runOnLoop(Loop * Loop)787 bool LoopPredication::runOnLoop(Loop *Loop) {
788   L = Loop;
789 
790   LLVM_DEBUG(dbgs() << "Analyzing ");
791   LLVM_DEBUG(L->dump());
792 
793   Module *M = L->getHeader()->getModule();
794 
795   // There is nothing to do if the module doesn't use guards
796   auto *GuardDecl =
797       M->getFunction(Intrinsic::getName(Intrinsic::experimental_guard));
798   if (!GuardDecl || GuardDecl->use_empty())
799     return false;
800 
801   DL = &M->getDataLayout();
802 
803   Preheader = L->getLoopPreheader();
804   if (!Preheader)
805     return false;
806 
807   auto LatchCheckOpt = parseLoopLatchICmp();
808   if (!LatchCheckOpt)
809     return false;
810   LatchCheck = *LatchCheckOpt;
811 
812   LLVM_DEBUG(dbgs() << "Latch check:\n");
813   LLVM_DEBUG(LatchCheck.dump());
814 
815   if (!isLoopProfitableToPredicate()) {
816     LLVM_DEBUG(dbgs() << "Loop not profitable to predicate!\n");
817     return false;
818   }
819   // Collect all the guards into a vector and process later, so as not
820   // to invalidate the instruction iterator.
821   SmallVector<IntrinsicInst *, 4> Guards;
822   for (const auto BB : L->blocks())
823     for (auto &I : *BB)
824       if (isGuard(&I))
825         Guards.push_back(cast<IntrinsicInst>(&I));
826 
827   if (Guards.empty())
828     return false;
829 
830   SCEVExpander Expander(*SE, *DL, "loop-predication");
831 
832   bool Changed = false;
833   for (auto *Guard : Guards)
834     Changed |= widenGuardConditions(Guard, Expander);
835 
836   return Changed;
837 }
838