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