1//===- InductiveRangeCheckElimination.cpp - -------------------------------===//
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 InductiveRangeCheckElimination pass splits a loop's iteration space into
10// three disjoint ranges.  It does that in a way such that the loop running in
11// the middle loop provably does not need range checks. As an example, it will
12// convert
13//
14//   len = < known positive >
15//   for (i = 0; i < n; i++) {
16//     if (0 <= i && i < len) {
17//       do_something();
18//     } else {
19//       throw_out_of_bounds();
20//     }
21//   }
22//
23// to
24//
25//   len = < known positive >
26//   limit = smin(n, len)
27//   // no first segment
28//   for (i = 0; i < limit; i++) {
29//     if (0 <= i && i < len) { // this check is fully redundant
30//       do_something();
31//     } else {
32//       throw_out_of_bounds();
33//     }
34//   }
35//   for (i = limit; i < n; i++) {
36//     if (0 <= i && i < len) {
37//       do_something();
38//     } else {
39//       throw_out_of_bounds();
40//     }
41//   }
42//
43//===----------------------------------------------------------------------===//
44
45#include "llvm/Transforms/Scalar/InductiveRangeCheckElimination.h"
46#include "llvm/ADT/APInt.h"
47#include "llvm/ADT/ArrayRef.h"
48#include "llvm/ADT/PriorityWorklist.h"
49#include "llvm/ADT/SmallPtrSet.h"
50#include "llvm/ADT/SmallVector.h"
51#include "llvm/ADT/StringRef.h"
52#include "llvm/ADT/Twine.h"
53#include "llvm/Analysis/BlockFrequencyInfo.h"
54#include "llvm/Analysis/BranchProbabilityInfo.h"
55#include "llvm/Analysis/LoopAnalysisManager.h"
56#include "llvm/Analysis/LoopInfo.h"
57#include "llvm/Analysis/ScalarEvolution.h"
58#include "llvm/Analysis/ScalarEvolutionExpressions.h"
59#include "llvm/IR/BasicBlock.h"
60#include "llvm/IR/CFG.h"
61#include "llvm/IR/Constants.h"
62#include "llvm/IR/DerivedTypes.h"
63#include "llvm/IR/Dominators.h"
64#include "llvm/IR/Function.h"
65#include "llvm/IR/IRBuilder.h"
66#include "llvm/IR/InstrTypes.h"
67#include "llvm/IR/Instructions.h"
68#include "llvm/IR/Metadata.h"
69#include "llvm/IR/Module.h"
70#include "llvm/IR/PatternMatch.h"
71#include "llvm/IR/Type.h"
72#include "llvm/IR/Use.h"
73#include "llvm/IR/User.h"
74#include "llvm/IR/Value.h"
75#include "llvm/Support/BranchProbability.h"
76#include "llvm/Support/Casting.h"
77#include "llvm/Support/CommandLine.h"
78#include "llvm/Support/Compiler.h"
79#include "llvm/Support/Debug.h"
80#include "llvm/Support/ErrorHandling.h"
81#include "llvm/Support/raw_ostream.h"
82#include "llvm/Transforms/Utils/BasicBlockUtils.h"
83#include "llvm/Transforms/Utils/Cloning.h"
84#include "llvm/Transforms/Utils/LoopConstrainer.h"
85#include "llvm/Transforms/Utils/LoopSimplify.h"
86#include "llvm/Transforms/Utils/LoopUtils.h"
87#include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
88#include "llvm/Transforms/Utils/ValueMapper.h"
89#include <algorithm>
90#include <cassert>
91#include <iterator>
92#include <optional>
93#include <utility>
94
95using namespace llvm;
96using namespace llvm::PatternMatch;
97
98static cl::opt<unsigned> LoopSizeCutoff("irce-loop-size-cutoff", cl::Hidden,
99                                        cl::init(64));
100
101static cl::opt<bool> PrintChangedLoops("irce-print-changed-loops", cl::Hidden,
102                                       cl::init(false));
103
104static cl::opt<bool> PrintRangeChecks("irce-print-range-checks", cl::Hidden,
105                                      cl::init(false));
106
107static cl::opt<bool> SkipProfitabilityChecks("irce-skip-profitability-checks",
108                                             cl::Hidden, cl::init(false));
109
110static cl::opt<unsigned> MinRuntimeIterations("irce-min-runtime-iterations",
111                                              cl::Hidden, cl::init(10));
112
113static cl::opt<bool> AllowUnsignedLatchCondition("irce-allow-unsigned-latch",
114                                                 cl::Hidden, cl::init(true));
115
116static cl::opt<bool> AllowNarrowLatchCondition(
117    "irce-allow-narrow-latch", cl::Hidden, cl::init(true),
118    cl::desc("If set to true, IRCE may eliminate wide range checks in loops "
119             "with narrow latch condition."));
120
121static cl::opt<unsigned> MaxTypeSizeForOverflowCheck(
122    "irce-max-type-size-for-overflow-check", cl::Hidden, cl::init(32),
123    cl::desc(
124        "Maximum size of range check type for which can be produced runtime "
125        "overflow check of its limit's computation"));
126
127static cl::opt<bool>
128    PrintScaledBoundaryRangeChecks("irce-print-scaled-boundary-range-checks",
129                                   cl::Hidden, cl::init(false));
130
131#define DEBUG_TYPE "irce"
132
133namespace {
134
135/// An inductive range check is conditional branch in a loop with
136///
137///  1. a very cold successor (i.e. the branch jumps to that successor very
138///     rarely)
139///
140///  and
141///
142///  2. a condition that is provably true for some contiguous range of values
143///     taken by the containing loop's induction variable.
144///
145class InductiveRangeCheck {
146
147  const SCEV *Begin = nullptr;
148  const SCEV *Step = nullptr;
149  const SCEV *End = nullptr;
150  Use *CheckUse = nullptr;
151
152  static bool parseRangeCheckICmp(Loop *L, ICmpInst *ICI, ScalarEvolution &SE,
153                                  const SCEVAddRecExpr *&Index,
154                                  const SCEV *&End);
155
156  static void
157  extractRangeChecksFromCond(Loop *L, ScalarEvolution &SE, Use &ConditionUse,
158                             SmallVectorImpl<InductiveRangeCheck> &Checks,
159                             SmallPtrSetImpl<Value *> &Visited);
160
161  static bool parseIvAgaisntLimit(Loop *L, Value *LHS, Value *RHS,
162                                  ICmpInst::Predicate Pred, ScalarEvolution &SE,
163                                  const SCEVAddRecExpr *&Index,
164                                  const SCEV *&End);
165
166  static bool reassociateSubLHS(Loop *L, Value *VariantLHS, Value *InvariantRHS,
167                                ICmpInst::Predicate Pred, ScalarEvolution &SE,
168                                const SCEVAddRecExpr *&Index, const SCEV *&End);
169
170public:
171  const SCEV *getBegin() const { return Begin; }
172  const SCEV *getStep() const { return Step; }
173  const SCEV *getEnd() const { return End; }
174
175  void print(raw_ostream &OS) const {
176    OS << "InductiveRangeCheck:\n";
177    OS << "  Begin: ";
178    Begin->print(OS);
179    OS << "  Step: ";
180    Step->print(OS);
181    OS << "  End: ";
182    End->print(OS);
183    OS << "\n  CheckUse: ";
184    getCheckUse()->getUser()->print(OS);
185    OS << " Operand: " << getCheckUse()->getOperandNo() << "\n";
186  }
187
188  LLVM_DUMP_METHOD
189  void dump() {
190    print(dbgs());
191  }
192
193  Use *getCheckUse() const { return CheckUse; }
194
195  /// Represents an signed integer range [Range.getBegin(), Range.getEnd()).  If
196  /// R.getEnd() le R.getBegin(), then R denotes the empty range.
197
198  class Range {
199    const SCEV *Begin;
200    const SCEV *End;
201
202  public:
203    Range(const SCEV *Begin, const SCEV *End) : Begin(Begin), End(End) {
204      assert(Begin->getType() == End->getType() && "ill-typed range!");
205    }
206
207    Type *getType() const { return Begin->getType(); }
208    const SCEV *getBegin() const { return Begin; }
209    const SCEV *getEnd() const { return End; }
210    bool isEmpty(ScalarEvolution &SE, bool IsSigned) const {
211      if (Begin == End)
212        return true;
213      if (IsSigned)
214        return SE.isKnownPredicate(ICmpInst::ICMP_SGE, Begin, End);
215      else
216        return SE.isKnownPredicate(ICmpInst::ICMP_UGE, Begin, End);
217    }
218  };
219
220  /// This is the value the condition of the branch needs to evaluate to for the
221  /// branch to take the hot successor (see (1) above).
222  bool getPassingDirection() { return true; }
223
224  /// Computes a range for the induction variable (IndVar) in which the range
225  /// check is redundant and can be constant-folded away.  The induction
226  /// variable is not required to be the canonical {0,+,1} induction variable.
227  std::optional<Range> computeSafeIterationSpace(ScalarEvolution &SE,
228                                                 const SCEVAddRecExpr *IndVar,
229                                                 bool IsLatchSigned) const;
230
231  /// Parse out a set of inductive range checks from \p BI and append them to \p
232  /// Checks.
233  ///
234  /// NB! There may be conditions feeding into \p BI that aren't inductive range
235  /// checks, and hence don't end up in \p Checks.
236  static void extractRangeChecksFromBranch(
237      BranchInst *BI, Loop *L, ScalarEvolution &SE, BranchProbabilityInfo *BPI,
238      SmallVectorImpl<InductiveRangeCheck> &Checks, bool &Changed);
239};
240
241class InductiveRangeCheckElimination {
242  ScalarEvolution &SE;
243  BranchProbabilityInfo *BPI;
244  DominatorTree &DT;
245  LoopInfo &LI;
246
247  using GetBFIFunc =
248      std::optional<llvm::function_ref<llvm::BlockFrequencyInfo &()>>;
249  GetBFIFunc GetBFI;
250
251  // Returns true if it is profitable to do a transform basing on estimation of
252  // number of iterations.
253  bool isProfitableToTransform(const Loop &L, LoopStructure &LS);
254
255public:
256  InductiveRangeCheckElimination(ScalarEvolution &SE,
257                                 BranchProbabilityInfo *BPI, DominatorTree &DT,
258                                 LoopInfo &LI, GetBFIFunc GetBFI = std::nullopt)
259      : SE(SE), BPI(BPI), DT(DT), LI(LI), GetBFI(GetBFI) {}
260
261  bool run(Loop *L, function_ref<void(Loop *, bool)> LPMAddNewLoop);
262};
263
264} // end anonymous namespace
265
266/// Parse a single ICmp instruction, `ICI`, into a range check.  If `ICI` cannot
267/// be interpreted as a range check, return false.  Otherwise set `Index` to the
268/// SCEV being range checked, and set `End` to the upper or lower limit `Index`
269/// is being range checked.
270bool InductiveRangeCheck::parseRangeCheckICmp(Loop *L, ICmpInst *ICI,
271                                              ScalarEvolution &SE,
272                                              const SCEVAddRecExpr *&Index,
273                                              const SCEV *&End) {
274  auto IsLoopInvariant = [&SE, L](Value *V) {
275    return SE.isLoopInvariant(SE.getSCEV(V), L);
276  };
277
278  ICmpInst::Predicate Pred = ICI->getPredicate();
279  Value *LHS = ICI->getOperand(0);
280  Value *RHS = ICI->getOperand(1);
281
282  if (!LHS->getType()->isIntegerTy())
283    return false;
284
285  // Canonicalize to the `Index Pred Invariant` comparison
286  if (IsLoopInvariant(LHS)) {
287    std::swap(LHS, RHS);
288    Pred = CmpInst::getSwappedPredicate(Pred);
289  } else if (!IsLoopInvariant(RHS))
290    // Both LHS and RHS are loop variant
291    return false;
292
293  if (parseIvAgaisntLimit(L, LHS, RHS, Pred, SE, Index, End))
294    return true;
295
296  if (reassociateSubLHS(L, LHS, RHS, Pred, SE, Index, End))
297    return true;
298
299  // TODO: support ReassociateAddLHS
300  return false;
301}
302
303// Try to parse range check in the form of "IV vs Limit"
304bool InductiveRangeCheck::parseIvAgaisntLimit(Loop *L, Value *LHS, Value *RHS,
305                                              ICmpInst::Predicate Pred,
306                                              ScalarEvolution &SE,
307                                              const SCEVAddRecExpr *&Index,
308                                              const SCEV *&End) {
309
310  auto SIntMaxSCEV = [&](Type *T) {
311    unsigned BitWidth = cast<IntegerType>(T)->getBitWidth();
312    return SE.getConstant(APInt::getSignedMaxValue(BitWidth));
313  };
314
315  const auto *AddRec = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(LHS));
316  if (!AddRec)
317    return false;
318
319  // We strengthen "0 <= I" to "0 <= I < INT_SMAX" and "I < L" to "0 <= I < L".
320  // We can potentially do much better here.
321  // If we want to adjust upper bound for the unsigned range check as we do it
322  // for signed one, we will need to pick Unsigned max
323  switch (Pred) {
324  default:
325    return false;
326
327  case ICmpInst::ICMP_SGE:
328    if (match(RHS, m_ConstantInt<0>())) {
329      Index = AddRec;
330      End = SIntMaxSCEV(Index->getType());
331      return true;
332    }
333    return false;
334
335  case ICmpInst::ICMP_SGT:
336    if (match(RHS, m_ConstantInt<-1>())) {
337      Index = AddRec;
338      End = SIntMaxSCEV(Index->getType());
339      return true;
340    }
341    return false;
342
343  case ICmpInst::ICMP_SLT:
344  case ICmpInst::ICMP_ULT:
345    Index = AddRec;
346    End = SE.getSCEV(RHS);
347    return true;
348
349  case ICmpInst::ICMP_SLE:
350  case ICmpInst::ICMP_ULE:
351    const SCEV *One = SE.getOne(RHS->getType());
352    const SCEV *RHSS = SE.getSCEV(RHS);
353    bool Signed = Pred == ICmpInst::ICMP_SLE;
354    if (SE.willNotOverflow(Instruction::BinaryOps::Add, Signed, RHSS, One)) {
355      Index = AddRec;
356      End = SE.getAddExpr(RHSS, One);
357      return true;
358    }
359    return false;
360  }
361
362  llvm_unreachable("default clause returns!");
363}
364
365// Try to parse range check in the form of "IV - Offset vs Limit" or "Offset -
366// IV vs Limit"
367bool InductiveRangeCheck::reassociateSubLHS(
368    Loop *L, Value *VariantLHS, Value *InvariantRHS, ICmpInst::Predicate Pred,
369    ScalarEvolution &SE, const SCEVAddRecExpr *&Index, const SCEV *&End) {
370  Value *LHS, *RHS;
371  if (!match(VariantLHS, m_Sub(m_Value(LHS), m_Value(RHS))))
372    return false;
373
374  const SCEV *IV = SE.getSCEV(LHS);
375  const SCEV *Offset = SE.getSCEV(RHS);
376  const SCEV *Limit = SE.getSCEV(InvariantRHS);
377
378  bool OffsetSubtracted = false;
379  if (SE.isLoopInvariant(IV, L))
380    // "Offset - IV vs Limit"
381    std::swap(IV, Offset);
382  else if (SE.isLoopInvariant(Offset, L))
383    // "IV - Offset vs Limit"
384    OffsetSubtracted = true;
385  else
386    return false;
387
388  const auto *AddRec = dyn_cast<SCEVAddRecExpr>(IV);
389  if (!AddRec)
390    return false;
391
392  // In order to turn "IV - Offset < Limit" into "IV < Limit + Offset", we need
393  // to be able to freely move values from left side of inequality to right side
394  // (just as in normal linear arithmetics). Overflows make things much more
395  // complicated, so we want to avoid this.
396  //
397  // Let's prove that the initial subtraction doesn't overflow with all IV's
398  // values from the safe range constructed for that check.
399  //
400  // [Case 1] IV - Offset < Limit
401  // It doesn't overflow if:
402  //     SINT_MIN <= IV - Offset <= SINT_MAX
403  // In terms of scaled SINT we need to prove:
404  //     SINT_MIN + Offset <= IV <= SINT_MAX + Offset
405  // Safe range will be constructed:
406  //     0 <= IV < Limit + Offset
407  // It means that 'IV - Offset' doesn't underflow, because:
408  //     SINT_MIN + Offset < 0 <= IV
409  // and doesn't overflow:
410  //     IV < Limit + Offset <= SINT_MAX + Offset
411  //
412  // [Case 2] Offset - IV > Limit
413  // It doesn't overflow if:
414  //     SINT_MIN <= Offset - IV <= SINT_MAX
415  // In terms of scaled SINT we need to prove:
416  //     -SINT_MIN >= IV - Offset >= -SINT_MAX
417  //     Offset - SINT_MIN >= IV >= Offset - SINT_MAX
418  // Safe range will be constructed:
419  //     0 <= IV < Offset - Limit
420  // It means that 'Offset - IV' doesn't underflow, because
421  //     Offset - SINT_MAX < 0 <= IV
422  // and doesn't overflow:
423  //     IV < Offset - Limit <= Offset - SINT_MIN
424  //
425  // For the computed upper boundary of the IV's range (Offset +/- Limit) we
426  // don't know exactly whether it overflows or not. So if we can't prove this
427  // fact at compile time, we scale boundary computations to a wider type with
428  // the intention to add runtime overflow check.
429
430  auto getExprScaledIfOverflow = [&](Instruction::BinaryOps BinOp,
431                                     const SCEV *LHS,
432                                     const SCEV *RHS) -> const SCEV * {
433    const SCEV *(ScalarEvolution::*Operation)(const SCEV *, const SCEV *,
434                                              SCEV::NoWrapFlags, unsigned);
435    switch (BinOp) {
436    default:
437      llvm_unreachable("Unsupported binary op");
438    case Instruction::Add:
439      Operation = &ScalarEvolution::getAddExpr;
440      break;
441    case Instruction::Sub:
442      Operation = &ScalarEvolution::getMinusSCEV;
443      break;
444    }
445
446    if (SE.willNotOverflow(BinOp, ICmpInst::isSigned(Pred), LHS, RHS,
447                           cast<Instruction>(VariantLHS)))
448      return (SE.*Operation)(LHS, RHS, SCEV::FlagAnyWrap, 0);
449
450    // We couldn't prove that the expression does not overflow.
451    // Than scale it to a wider type to check overflow at runtime.
452    auto *Ty = cast<IntegerType>(LHS->getType());
453    if (Ty->getBitWidth() > MaxTypeSizeForOverflowCheck)
454      return nullptr;
455
456    auto WideTy = IntegerType::get(Ty->getContext(), Ty->getBitWidth() * 2);
457    return (SE.*Operation)(SE.getSignExtendExpr(LHS, WideTy),
458                           SE.getSignExtendExpr(RHS, WideTy), SCEV::FlagAnyWrap,
459                           0);
460  };
461
462  if (OffsetSubtracted)
463    // "IV - Offset < Limit" -> "IV" < Offset + Limit
464    Limit = getExprScaledIfOverflow(Instruction::BinaryOps::Add, Offset, Limit);
465  else {
466    // "Offset - IV > Limit" -> "IV" < Offset - Limit
467    Limit = getExprScaledIfOverflow(Instruction::BinaryOps::Sub, Offset, Limit);
468    Pred = ICmpInst::getSwappedPredicate(Pred);
469  }
470
471  if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) {
472    // "Expr <= Limit" -> "Expr < Limit + 1"
473    if (Pred == ICmpInst::ICMP_SLE && Limit)
474      Limit = getExprScaledIfOverflow(Instruction::BinaryOps::Add, Limit,
475                                      SE.getOne(Limit->getType()));
476    if (Limit) {
477      Index = AddRec;
478      End = Limit;
479      return true;
480    }
481  }
482  return false;
483}
484
485void InductiveRangeCheck::extractRangeChecksFromCond(
486    Loop *L, ScalarEvolution &SE, Use &ConditionUse,
487    SmallVectorImpl<InductiveRangeCheck> &Checks,
488    SmallPtrSetImpl<Value *> &Visited) {
489  Value *Condition = ConditionUse.get();
490  if (!Visited.insert(Condition).second)
491    return;
492
493  // TODO: Do the same for OR, XOR, NOT etc?
494  if (match(Condition, m_LogicalAnd(m_Value(), m_Value()))) {
495    extractRangeChecksFromCond(L, SE, cast<User>(Condition)->getOperandUse(0),
496                               Checks, Visited);
497    extractRangeChecksFromCond(L, SE, cast<User>(Condition)->getOperandUse(1),
498                               Checks, Visited);
499    return;
500  }
501
502  ICmpInst *ICI = dyn_cast<ICmpInst>(Condition);
503  if (!ICI)
504    return;
505
506  const SCEV *End = nullptr;
507  const SCEVAddRecExpr *IndexAddRec = nullptr;
508  if (!parseRangeCheckICmp(L, ICI, SE, IndexAddRec, End))
509    return;
510
511  assert(IndexAddRec && "IndexAddRec was not computed");
512  assert(End && "End was not computed");
513
514  if ((IndexAddRec->getLoop() != L) || !IndexAddRec->isAffine())
515    return;
516
517  InductiveRangeCheck IRC;
518  IRC.End = End;
519  IRC.Begin = IndexAddRec->getStart();
520  IRC.Step = IndexAddRec->getStepRecurrence(SE);
521  IRC.CheckUse = &ConditionUse;
522  Checks.push_back(IRC);
523}
524
525void InductiveRangeCheck::extractRangeChecksFromBranch(
526    BranchInst *BI, Loop *L, ScalarEvolution &SE, BranchProbabilityInfo *BPI,
527    SmallVectorImpl<InductiveRangeCheck> &Checks, bool &Changed) {
528  if (BI->isUnconditional() || BI->getParent() == L->getLoopLatch())
529    return;
530
531  unsigned IndexLoopSucc = L->contains(BI->getSuccessor(0)) ? 0 : 1;
532  assert(L->contains(BI->getSuccessor(IndexLoopSucc)) &&
533         "No edges coming to loop?");
534  BranchProbability LikelyTaken(15, 16);
535
536  if (!SkipProfitabilityChecks && BPI &&
537      BPI->getEdgeProbability(BI->getParent(), IndexLoopSucc) < LikelyTaken)
538    return;
539
540  // IRCE expects branch's true edge comes to loop. Invert branch for opposite
541  // case.
542  if (IndexLoopSucc != 0) {
543    IRBuilder<> Builder(BI);
544    InvertBranch(BI, Builder);
545    if (BPI)
546      BPI->swapSuccEdgesProbabilities(BI->getParent());
547    Changed = true;
548  }
549
550  SmallPtrSet<Value *, 8> Visited;
551  InductiveRangeCheck::extractRangeChecksFromCond(L, SE, BI->getOperandUse(0),
552                                                  Checks, Visited);
553}
554
555/// If the type of \p S matches with \p Ty, return \p S. Otherwise, return
556/// signed or unsigned extension of \p S to type \p Ty.
557static const SCEV *NoopOrExtend(const SCEV *S, Type *Ty, ScalarEvolution &SE,
558                                bool Signed) {
559  return Signed ? SE.getNoopOrSignExtend(S, Ty) : SE.getNoopOrZeroExtend(S, Ty);
560}
561
562// Compute a safe set of limits for the main loop to run in -- effectively the
563// intersection of `Range' and the iteration space of the original loop.
564// Return std::nullopt if unable to compute the set of subranges.
565static std::optional<LoopConstrainer::SubRanges>
566calculateSubRanges(ScalarEvolution &SE, const Loop &L,
567                   InductiveRangeCheck::Range &Range,
568                   const LoopStructure &MainLoopStructure) {
569  auto *RTy = cast<IntegerType>(Range.getType());
570  // We only support wide range checks and narrow latches.
571  if (!AllowNarrowLatchCondition && RTy != MainLoopStructure.ExitCountTy)
572    return std::nullopt;
573  if (RTy->getBitWidth() < MainLoopStructure.ExitCountTy->getBitWidth())
574    return std::nullopt;
575
576  LoopConstrainer::SubRanges Result;
577
578  bool IsSignedPredicate = MainLoopStructure.IsSignedPredicate;
579  // I think we can be more aggressive here and make this nuw / nsw if the
580  // addition that feeds into the icmp for the latch's terminating branch is nuw
581  // / nsw.  In any case, a wrapping 2's complement addition is safe.
582  const SCEV *Start = NoopOrExtend(SE.getSCEV(MainLoopStructure.IndVarStart),
583                                   RTy, SE, IsSignedPredicate);
584  const SCEV *End = NoopOrExtend(SE.getSCEV(MainLoopStructure.LoopExitAt), RTy,
585                                 SE, IsSignedPredicate);
586
587  bool Increasing = MainLoopStructure.IndVarIncreasing;
588
589  // We compute `Smallest` and `Greatest` such that [Smallest, Greatest), or
590  // [Smallest, GreatestSeen] is the range of values the induction variable
591  // takes.
592
593  const SCEV *Smallest = nullptr, *Greatest = nullptr, *GreatestSeen = nullptr;
594
595  const SCEV *One = SE.getOne(RTy);
596  if (Increasing) {
597    Smallest = Start;
598    Greatest = End;
599    // No overflow, because the range [Smallest, GreatestSeen] is not empty.
600    GreatestSeen = SE.getMinusSCEV(End, One);
601  } else {
602    // These two computations may sign-overflow.  Here is why that is okay:
603    //
604    // We know that the induction variable does not sign-overflow on any
605    // iteration except the last one, and it starts at `Start` and ends at
606    // `End`, decrementing by one every time.
607    //
608    //  * if `Smallest` sign-overflows we know `End` is `INT_SMAX`. Since the
609    //    induction variable is decreasing we know that the smallest value
610    //    the loop body is actually executed with is `INT_SMIN` == `Smallest`.
611    //
612    //  * if `Greatest` sign-overflows, we know it can only be `INT_SMIN`.  In
613    //    that case, `Clamp` will always return `Smallest` and
614    //    [`Result.LowLimit`, `Result.HighLimit`) = [`Smallest`, `Smallest`)
615    //    will be an empty range.  Returning an empty range is always safe.
616
617    Smallest = SE.getAddExpr(End, One);
618    Greatest = SE.getAddExpr(Start, One);
619    GreatestSeen = Start;
620  }
621
622  auto Clamp = [&SE, Smallest, Greatest, IsSignedPredicate](const SCEV *S) {
623    return IsSignedPredicate
624               ? SE.getSMaxExpr(Smallest, SE.getSMinExpr(Greatest, S))
625               : SE.getUMaxExpr(Smallest, SE.getUMinExpr(Greatest, S));
626  };
627
628  // In some cases we can prove that we don't need a pre or post loop.
629  ICmpInst::Predicate PredLE =
630      IsSignedPredicate ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
631  ICmpInst::Predicate PredLT =
632      IsSignedPredicate ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
633
634  bool ProvablyNoPreloop =
635      SE.isKnownPredicate(PredLE, Range.getBegin(), Smallest);
636  if (!ProvablyNoPreloop)
637    Result.LowLimit = Clamp(Range.getBegin());
638
639  bool ProvablyNoPostLoop =
640      SE.isKnownPredicate(PredLT, GreatestSeen, Range.getEnd());
641  if (!ProvablyNoPostLoop)
642    Result.HighLimit = Clamp(Range.getEnd());
643
644  return Result;
645}
646
647/// Computes and returns a range of values for the induction variable (IndVar)
648/// in which the range check can be safely elided.  If it cannot compute such a
649/// range, returns std::nullopt.
650std::optional<InductiveRangeCheck::Range>
651InductiveRangeCheck::computeSafeIterationSpace(ScalarEvolution &SE,
652                                               const SCEVAddRecExpr *IndVar,
653                                               bool IsLatchSigned) const {
654  // We can deal when types of latch check and range checks don't match in case
655  // if latch check is more narrow.
656  auto *IVType = dyn_cast<IntegerType>(IndVar->getType());
657  auto *RCType = dyn_cast<IntegerType>(getBegin()->getType());
658  auto *EndType = dyn_cast<IntegerType>(getEnd()->getType());
659  // Do not work with pointer types.
660  if (!IVType || !RCType)
661    return std::nullopt;
662  if (IVType->getBitWidth() > RCType->getBitWidth())
663    return std::nullopt;
664
665  // IndVar is of the form "A + B * I" (where "I" is the canonical induction
666  // variable, that may or may not exist as a real llvm::Value in the loop) and
667  // this inductive range check is a range check on the "C + D * I" ("C" is
668  // getBegin() and "D" is getStep()).  We rewrite the value being range
669  // checked to "M + N * IndVar" where "N" = "D * B^(-1)" and "M" = "C - NA".
670  //
671  // The actual inequalities we solve are of the form
672  //
673  //   0 <= M + 1 * IndVar < L given L >= 0  (i.e. N == 1)
674  //
675  // Here L stands for upper limit of the safe iteration space.
676  // The inequality is satisfied by (0 - M) <= IndVar < (L - M). To avoid
677  // overflows when calculating (0 - M) and (L - M) we, depending on type of
678  // IV's iteration space, limit the calculations by borders of the iteration
679  // space. For example, if IndVar is unsigned, (0 - M) overflows for any M > 0.
680  // If we figured out that "anything greater than (-M) is safe", we strengthen
681  // this to "everything greater than 0 is safe", assuming that values between
682  // -M and 0 just do not exist in unsigned iteration space, and we don't want
683  // to deal with overflown values.
684
685  if (!IndVar->isAffine())
686    return std::nullopt;
687
688  const SCEV *A = NoopOrExtend(IndVar->getStart(), RCType, SE, IsLatchSigned);
689  const SCEVConstant *B = dyn_cast<SCEVConstant>(
690      NoopOrExtend(IndVar->getStepRecurrence(SE), RCType, SE, IsLatchSigned));
691  if (!B)
692    return std::nullopt;
693  assert(!B->isZero() && "Recurrence with zero step?");
694
695  const SCEV *C = getBegin();
696  const SCEVConstant *D = dyn_cast<SCEVConstant>(getStep());
697  if (D != B)
698    return std::nullopt;
699
700  assert(!D->getValue()->isZero() && "Recurrence with zero step?");
701  unsigned BitWidth = RCType->getBitWidth();
702  const SCEV *SIntMax = SE.getConstant(APInt::getSignedMaxValue(BitWidth));
703  const SCEV *SIntMin = SE.getConstant(APInt::getSignedMinValue(BitWidth));
704
705  // Subtract Y from X so that it does not go through border of the IV
706  // iteration space. Mathematically, it is equivalent to:
707  //
708  //    ClampedSubtract(X, Y) = min(max(X - Y, INT_MIN), INT_MAX).        [1]
709  //
710  // In [1], 'X - Y' is a mathematical subtraction (result is not bounded to
711  // any width of bit grid). But after we take min/max, the result is
712  // guaranteed to be within [INT_MIN, INT_MAX].
713  //
714  // In [1], INT_MAX and INT_MIN are respectively signed and unsigned max/min
715  // values, depending on type of latch condition that defines IV iteration
716  // space.
717  auto ClampedSubtract = [&](const SCEV *X, const SCEV *Y) {
718    // FIXME: The current implementation assumes that X is in [0, SINT_MAX].
719    // This is required to ensure that SINT_MAX - X does not overflow signed and
720    // that X - Y does not overflow unsigned if Y is negative. Can we lift this
721    // restriction and make it work for negative X either?
722    if (IsLatchSigned) {
723      // X is a number from signed range, Y is interpreted as signed.
724      // Even if Y is SINT_MAX, (X - Y) does not reach SINT_MIN. So the only
725      // thing we should care about is that we didn't cross SINT_MAX.
726      // So, if Y is positive, we subtract Y safely.
727      //   Rule 1: Y > 0 ---> Y.
728      // If 0 <= -Y <= (SINT_MAX - X), we subtract Y safely.
729      //   Rule 2: Y >=s (X - SINT_MAX) ---> Y.
730      // If 0 <= (SINT_MAX - X) < -Y, we can only subtract (X - SINT_MAX).
731      //   Rule 3: Y <s (X - SINT_MAX) ---> (X - SINT_MAX).
732      // It gives us smax(Y, X - SINT_MAX) to subtract in all cases.
733      const SCEV *XMinusSIntMax = SE.getMinusSCEV(X, SIntMax);
734      return SE.getMinusSCEV(X, SE.getSMaxExpr(Y, XMinusSIntMax),
735                             SCEV::FlagNSW);
736    } else
737      // X is a number from unsigned range, Y is interpreted as signed.
738      // Even if Y is SINT_MIN, (X - Y) does not reach UINT_MAX. So the only
739      // thing we should care about is that we didn't cross zero.
740      // So, if Y is negative, we subtract Y safely.
741      //   Rule 1: Y <s 0 ---> Y.
742      // If 0 <= Y <= X, we subtract Y safely.
743      //   Rule 2: Y <=s X ---> Y.
744      // If 0 <= X < Y, we should stop at 0 and can only subtract X.
745      //   Rule 3: Y >s X ---> X.
746      // It gives us smin(X, Y) to subtract in all cases.
747      return SE.getMinusSCEV(X, SE.getSMinExpr(X, Y), SCEV::FlagNUW);
748  };
749  const SCEV *M = SE.getMinusSCEV(C, A);
750  const SCEV *Zero = SE.getZero(M->getType());
751
752  // This function returns SCEV equal to 1 if X is non-negative 0 otherwise.
753  auto SCEVCheckNonNegative = [&](const SCEV *X) {
754    const Loop *L = IndVar->getLoop();
755    const SCEV *Zero = SE.getZero(X->getType());
756    const SCEV *One = SE.getOne(X->getType());
757    // Can we trivially prove that X is a non-negative or negative value?
758    if (isKnownNonNegativeInLoop(X, L, SE))
759      return One;
760    else if (isKnownNegativeInLoop(X, L, SE))
761      return Zero;
762    // If not, we will have to figure it out during the execution.
763    // Function smax(smin(X, 0), -1) + 1 equals to 1 if X >= 0 and 0 if X < 0.
764    const SCEV *NegOne = SE.getNegativeSCEV(One);
765    return SE.getAddExpr(SE.getSMaxExpr(SE.getSMinExpr(X, Zero), NegOne), One);
766  };
767
768  // This function returns SCEV equal to 1 if X will not overflow in terms of
769  // range check type, 0 otherwise.
770  auto SCEVCheckWillNotOverflow = [&](const SCEV *X) {
771    // X doesn't overflow if SINT_MAX >= X.
772    // Then if (SINT_MAX - X) >= 0, X doesn't overflow
773    const SCEV *SIntMaxExt = SE.getSignExtendExpr(SIntMax, X->getType());
774    const SCEV *OverflowCheck =
775        SCEVCheckNonNegative(SE.getMinusSCEV(SIntMaxExt, X));
776
777    // X doesn't underflow if X >= SINT_MIN.
778    // Then if (X - SINT_MIN) >= 0, X doesn't underflow
779    const SCEV *SIntMinExt = SE.getSignExtendExpr(SIntMin, X->getType());
780    const SCEV *UnderflowCheck =
781        SCEVCheckNonNegative(SE.getMinusSCEV(X, SIntMinExt));
782
783    return SE.getMulExpr(OverflowCheck, UnderflowCheck);
784  };
785
786  // FIXME: Current implementation of ClampedSubtract implicitly assumes that
787  // X is non-negative (in sense of a signed value). We need to re-implement
788  // this function in a way that it will correctly handle negative X as well.
789  // We use it twice: for X = 0 everything is fine, but for X = getEnd() we can
790  // end up with a negative X and produce wrong results. So currently we ensure
791  // that if getEnd() is negative then both ends of the safe range are zero.
792  // Note that this may pessimize elimination of unsigned range checks against
793  // negative values.
794  const SCEV *REnd = getEnd();
795  const SCEV *EndWillNotOverflow = SE.getOne(RCType);
796
797  auto PrintRangeCheck = [&](raw_ostream &OS) {
798    auto L = IndVar->getLoop();
799    OS << "irce: in function ";
800    OS << L->getHeader()->getParent()->getName();
801    OS << ", in ";
802    L->print(OS);
803    OS << "there is range check with scaled boundary:\n";
804    print(OS);
805  };
806
807  if (EndType->getBitWidth() > RCType->getBitWidth()) {
808    assert(EndType->getBitWidth() == RCType->getBitWidth() * 2);
809    if (PrintScaledBoundaryRangeChecks)
810      PrintRangeCheck(errs());
811    // End is computed with extended type but will be truncated to a narrow one
812    // type of range check. Therefore we need a check that the result will not
813    // overflow in terms of narrow type.
814    EndWillNotOverflow =
815        SE.getTruncateExpr(SCEVCheckWillNotOverflow(REnd), RCType);
816    REnd = SE.getTruncateExpr(REnd, RCType);
817  }
818
819  const SCEV *RuntimeChecks =
820      SE.getMulExpr(SCEVCheckNonNegative(REnd), EndWillNotOverflow);
821  const SCEV *Begin = SE.getMulExpr(ClampedSubtract(Zero, M), RuntimeChecks);
822  const SCEV *End = SE.getMulExpr(ClampedSubtract(REnd, M), RuntimeChecks);
823
824  return InductiveRangeCheck::Range(Begin, End);
825}
826
827static std::optional<InductiveRangeCheck::Range>
828IntersectSignedRange(ScalarEvolution &SE,
829                     const std::optional<InductiveRangeCheck::Range> &R1,
830                     const InductiveRangeCheck::Range &R2) {
831  if (R2.isEmpty(SE, /* IsSigned */ true))
832    return std::nullopt;
833  if (!R1)
834    return R2;
835  auto &R1Value = *R1;
836  // We never return empty ranges from this function, and R1 is supposed to be
837  // a result of intersection. Thus, R1 is never empty.
838  assert(!R1Value.isEmpty(SE, /* IsSigned */ true) &&
839         "We should never have empty R1!");
840
841  // TODO: we could widen the smaller range and have this work; but for now we
842  // bail out to keep things simple.
843  if (R1Value.getType() != R2.getType())
844    return std::nullopt;
845
846  const SCEV *NewBegin = SE.getSMaxExpr(R1Value.getBegin(), R2.getBegin());
847  const SCEV *NewEnd = SE.getSMinExpr(R1Value.getEnd(), R2.getEnd());
848
849  // If the resulting range is empty, just return std::nullopt.
850  auto Ret = InductiveRangeCheck::Range(NewBegin, NewEnd);
851  if (Ret.isEmpty(SE, /* IsSigned */ true))
852    return std::nullopt;
853  return Ret;
854}
855
856static std::optional<InductiveRangeCheck::Range>
857IntersectUnsignedRange(ScalarEvolution &SE,
858                       const std::optional<InductiveRangeCheck::Range> &R1,
859                       const InductiveRangeCheck::Range &R2) {
860  if (R2.isEmpty(SE, /* IsSigned */ false))
861    return std::nullopt;
862  if (!R1)
863    return R2;
864  auto &R1Value = *R1;
865  // We never return empty ranges from this function, and R1 is supposed to be
866  // a result of intersection. Thus, R1 is never empty.
867  assert(!R1Value.isEmpty(SE, /* IsSigned */ false) &&
868         "We should never have empty R1!");
869
870  // TODO: we could widen the smaller range and have this work; but for now we
871  // bail out to keep things simple.
872  if (R1Value.getType() != R2.getType())
873    return std::nullopt;
874
875  const SCEV *NewBegin = SE.getUMaxExpr(R1Value.getBegin(), R2.getBegin());
876  const SCEV *NewEnd = SE.getUMinExpr(R1Value.getEnd(), R2.getEnd());
877
878  // If the resulting range is empty, just return std::nullopt.
879  auto Ret = InductiveRangeCheck::Range(NewBegin, NewEnd);
880  if (Ret.isEmpty(SE, /* IsSigned */ false))
881    return std::nullopt;
882  return Ret;
883}
884
885PreservedAnalyses IRCEPass::run(Function &F, FunctionAnalysisManager &AM) {
886  auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
887  LoopInfo &LI = AM.getResult<LoopAnalysis>(F);
888  // There are no loops in the function. Return before computing other expensive
889  // analyses.
890  if (LI.empty())
891    return PreservedAnalyses::all();
892  auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F);
893  auto &BPI = AM.getResult<BranchProbabilityAnalysis>(F);
894
895  // Get BFI analysis result on demand. Please note that modification of
896  // CFG invalidates this analysis and we should handle it.
897  auto getBFI = [&F, &AM ]()->BlockFrequencyInfo & {
898    return AM.getResult<BlockFrequencyAnalysis>(F);
899  };
900  InductiveRangeCheckElimination IRCE(SE, &BPI, DT, LI, { getBFI });
901
902  bool Changed = false;
903  {
904    bool CFGChanged = false;
905    for (const auto &L : LI) {
906      CFGChanged |= simplifyLoop(L, &DT, &LI, &SE, nullptr, nullptr,
907                                 /*PreserveLCSSA=*/false);
908      Changed |= formLCSSARecursively(*L, DT, &LI, &SE);
909    }
910    Changed |= CFGChanged;
911
912    if (CFGChanged && !SkipProfitabilityChecks) {
913      PreservedAnalyses PA = PreservedAnalyses::all();
914      PA.abandon<BlockFrequencyAnalysis>();
915      AM.invalidate(F, PA);
916    }
917  }
918
919  SmallPriorityWorklist<Loop *, 4> Worklist;
920  appendLoopsToWorklist(LI, Worklist);
921  auto LPMAddNewLoop = [&Worklist](Loop *NL, bool IsSubloop) {
922    if (!IsSubloop)
923      appendLoopsToWorklist(*NL, Worklist);
924  };
925
926  while (!Worklist.empty()) {
927    Loop *L = Worklist.pop_back_val();
928    if (IRCE.run(L, LPMAddNewLoop)) {
929      Changed = true;
930      if (!SkipProfitabilityChecks) {
931        PreservedAnalyses PA = PreservedAnalyses::all();
932        PA.abandon<BlockFrequencyAnalysis>();
933        AM.invalidate(F, PA);
934      }
935    }
936  }
937
938  if (!Changed)
939    return PreservedAnalyses::all();
940  return getLoopPassPreservedAnalyses();
941}
942
943bool
944InductiveRangeCheckElimination::isProfitableToTransform(const Loop &L,
945                                                        LoopStructure &LS) {
946  if (SkipProfitabilityChecks)
947    return true;
948  if (GetBFI) {
949    BlockFrequencyInfo &BFI = (*GetBFI)();
950    uint64_t hFreq = BFI.getBlockFreq(LS.Header).getFrequency();
951    uint64_t phFreq = BFI.getBlockFreq(L.getLoopPreheader()).getFrequency();
952    if (phFreq != 0 && hFreq != 0 && (hFreq / phFreq < MinRuntimeIterations)) {
953      LLVM_DEBUG(dbgs() << "irce: could not prove profitability: "
954                        << "the estimated number of iterations basing on "
955                           "frequency info is " << (hFreq / phFreq) << "\n";);
956      return false;
957    }
958    return true;
959  }
960
961  if (!BPI)
962    return true;
963  BranchProbability ExitProbability =
964      BPI->getEdgeProbability(LS.Latch, LS.LatchBrExitIdx);
965  if (ExitProbability > BranchProbability(1, MinRuntimeIterations)) {
966    LLVM_DEBUG(dbgs() << "irce: could not prove profitability: "
967                      << "the exit probability is too big " << ExitProbability
968                      << "\n";);
969    return false;
970  }
971  return true;
972}
973
974bool InductiveRangeCheckElimination::run(
975    Loop *L, function_ref<void(Loop *, bool)> LPMAddNewLoop) {
976  if (L->getBlocks().size() >= LoopSizeCutoff) {
977    LLVM_DEBUG(dbgs() << "irce: giving up constraining loop, too large\n");
978    return false;
979  }
980
981  BasicBlock *Preheader = L->getLoopPreheader();
982  if (!Preheader) {
983    LLVM_DEBUG(dbgs() << "irce: loop has no preheader, leaving\n");
984    return false;
985  }
986
987  LLVMContext &Context = Preheader->getContext();
988  SmallVector<InductiveRangeCheck, 16> RangeChecks;
989  bool Changed = false;
990
991  for (auto *BBI : L->getBlocks())
992    if (BranchInst *TBI = dyn_cast<BranchInst>(BBI->getTerminator()))
993      InductiveRangeCheck::extractRangeChecksFromBranch(TBI, L, SE, BPI,
994                                                        RangeChecks, Changed);
995
996  if (RangeChecks.empty())
997    return Changed;
998
999  auto PrintRecognizedRangeChecks = [&](raw_ostream &OS) {
1000    OS << "irce: looking at loop "; L->print(OS);
1001    OS << "irce: loop has " << RangeChecks.size()
1002       << " inductive range checks: \n";
1003    for (InductiveRangeCheck &IRC : RangeChecks)
1004      IRC.print(OS);
1005  };
1006
1007  LLVM_DEBUG(PrintRecognizedRangeChecks(dbgs()));
1008
1009  if (PrintRangeChecks)
1010    PrintRecognizedRangeChecks(errs());
1011
1012  const char *FailureReason = nullptr;
1013  std::optional<LoopStructure> MaybeLoopStructure =
1014      LoopStructure::parseLoopStructure(SE, *L, AllowUnsignedLatchCondition,
1015                                        FailureReason);
1016  if (!MaybeLoopStructure) {
1017    LLVM_DEBUG(dbgs() << "irce: could not parse loop structure: "
1018                      << FailureReason << "\n";);
1019    return Changed;
1020  }
1021  LoopStructure LS = *MaybeLoopStructure;
1022  if (!isProfitableToTransform(*L, LS))
1023    return Changed;
1024  const SCEVAddRecExpr *IndVar =
1025      cast<SCEVAddRecExpr>(SE.getMinusSCEV(SE.getSCEV(LS.IndVarBase), SE.getSCEV(LS.IndVarStep)));
1026
1027  std::optional<InductiveRangeCheck::Range> SafeIterRange;
1028
1029  SmallVector<InductiveRangeCheck, 4> RangeChecksToEliminate;
1030  // Basing on the type of latch predicate, we interpret the IV iteration range
1031  // as signed or unsigned range. We use different min/max functions (signed or
1032  // unsigned) when intersecting this range with safe iteration ranges implied
1033  // by range checks.
1034  auto IntersectRange =
1035      LS.IsSignedPredicate ? IntersectSignedRange : IntersectUnsignedRange;
1036
1037  for (InductiveRangeCheck &IRC : RangeChecks) {
1038    auto Result = IRC.computeSafeIterationSpace(SE, IndVar,
1039                                                LS.IsSignedPredicate);
1040    if (Result) {
1041      auto MaybeSafeIterRange = IntersectRange(SE, SafeIterRange, *Result);
1042      if (MaybeSafeIterRange) {
1043        assert(!MaybeSafeIterRange->isEmpty(SE, LS.IsSignedPredicate) &&
1044               "We should never return empty ranges!");
1045        RangeChecksToEliminate.push_back(IRC);
1046        SafeIterRange = *MaybeSafeIterRange;
1047      }
1048    }
1049  }
1050
1051  if (!SafeIterRange)
1052    return Changed;
1053
1054  std::optional<LoopConstrainer::SubRanges> MaybeSR =
1055      calculateSubRanges(SE, *L, *SafeIterRange, LS);
1056  if (!MaybeSR) {
1057    LLVM_DEBUG(dbgs() << "irce: could not compute subranges\n");
1058    return false;
1059  }
1060
1061  LoopConstrainer LC(*L, LI, LPMAddNewLoop, LS, SE, DT,
1062                     SafeIterRange->getBegin()->getType(), *MaybeSR);
1063
1064  if (LC.run()) {
1065    Changed = true;
1066
1067    auto PrintConstrainedLoopInfo = [L]() {
1068      dbgs() << "irce: in function ";
1069      dbgs() << L->getHeader()->getParent()->getName() << ": ";
1070      dbgs() << "constrained ";
1071      L->print(dbgs());
1072    };
1073
1074    LLVM_DEBUG(PrintConstrainedLoopInfo());
1075
1076    if (PrintChangedLoops)
1077      PrintConstrainedLoopInfo();
1078
1079    // Optimize away the now-redundant range checks.
1080
1081    for (InductiveRangeCheck &IRC : RangeChecksToEliminate) {
1082      ConstantInt *FoldedRangeCheck = IRC.getPassingDirection()
1083                                          ? ConstantInt::getTrue(Context)
1084                                          : ConstantInt::getFalse(Context);
1085      IRC.getCheckUse()->set(FoldedRangeCheck);
1086    }
1087  }
1088
1089  return Changed;
1090}
1091