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/None.h"
49 #include "llvm/ADT/Optional.h"
50 #include "llvm/ADT/PriorityWorklist.h"
51 #include "llvm/ADT/SmallPtrSet.h"
52 #include "llvm/ADT/SmallVector.h"
53 #include "llvm/ADT/StringRef.h"
54 #include "llvm/ADT/Twine.h"
55 #include "llvm/Analysis/BlockFrequencyInfo.h"
56 #include "llvm/Analysis/BranchProbabilityInfo.h"
57 #include "llvm/Analysis/LoopAnalysisManager.h"
58 #include "llvm/Analysis/LoopInfo.h"
59 #include "llvm/Analysis/LoopPass.h"
60 #include "llvm/Analysis/PostDominators.h"
61 #include "llvm/Analysis/ScalarEvolution.h"
62 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
63 #include "llvm/IR/BasicBlock.h"
64 #include "llvm/IR/CFG.h"
65 #include "llvm/IR/Constants.h"
66 #include "llvm/IR/DerivedTypes.h"
67 #include "llvm/IR/Dominators.h"
68 #include "llvm/IR/Function.h"
69 #include "llvm/IR/IRBuilder.h"
70 #include "llvm/IR/InstrTypes.h"
71 #include "llvm/IR/Instructions.h"
72 #include "llvm/IR/Metadata.h"
73 #include "llvm/IR/Module.h"
74 #include "llvm/IR/PatternMatch.h"
75 #include "llvm/IR/Type.h"
76 #include "llvm/IR/Use.h"
77 #include "llvm/IR/User.h"
78 #include "llvm/IR/Value.h"
79 #include "llvm/InitializePasses.h"
80 #include "llvm/Pass.h"
81 #include "llvm/Support/BranchProbability.h"
82 #include "llvm/Support/Casting.h"
83 #include "llvm/Support/CommandLine.h"
84 #include "llvm/Support/Compiler.h"
85 #include "llvm/Support/Debug.h"
86 #include "llvm/Support/ErrorHandling.h"
87 #include "llvm/Support/raw_ostream.h"
88 #include "llvm/Transforms/Scalar.h"
89 #include "llvm/Transforms/Utils/Cloning.h"
90 #include "llvm/Transforms/Utils/LoopSimplify.h"
91 #include "llvm/Transforms/Utils/LoopUtils.h"
92 #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
93 #include "llvm/Transforms/Utils/ValueMapper.h"
94 #include <algorithm>
95 #include <cassert>
96 #include <iterator>
97 #include <limits>
98 #include <utility>
99 #include <vector>
100 
101 using namespace llvm;
102 using namespace llvm::PatternMatch;
103 
104 static cl::opt<unsigned> LoopSizeCutoff("irce-loop-size-cutoff", cl::Hidden,
105                                         cl::init(64));
106 
107 static cl::opt<bool> PrintChangedLoops("irce-print-changed-loops", cl::Hidden,
108                                        cl::init(false));
109 
110 static cl::opt<bool> PrintRangeChecks("irce-print-range-checks", cl::Hidden,
111                                       cl::init(false));
112 
113 static cl::opt<bool> SkipProfitabilityChecks("irce-skip-profitability-checks",
114                                              cl::Hidden, cl::init(false));
115 
116 static cl::opt<unsigned> MinRuntimeIterations("irce-min-runtime-iterations",
117                                               cl::Hidden, cl::init(10));
118 
119 static cl::opt<bool> AllowUnsignedLatchCondition("irce-allow-unsigned-latch",
120                                                  cl::Hidden, cl::init(true));
121 
122 static cl::opt<bool> AllowNarrowLatchCondition(
123     "irce-allow-narrow-latch", cl::Hidden, cl::init(true),
124     cl::desc("If set to true, IRCE may eliminate wide range checks in loops "
125              "with narrow latch condition."));
126 
127 static const char *ClonedLoopTag = "irce.loop.clone";
128 
129 #define DEBUG_TYPE "irce"
130 
131 namespace {
132 
133 /// An inductive range check is conditional branch in a loop with
134 ///
135 ///  1. a very cold successor (i.e. the branch jumps to that successor very
136 ///     rarely)
137 ///
138 ///  and
139 ///
140 ///  2. a condition that is provably true for some contiguous range of values
141 ///     taken by the containing loop's induction variable.
142 ///
143 class InductiveRangeCheck {
144 
145   const SCEV *Begin = nullptr;
146   const SCEV *Step = nullptr;
147   const SCEV *End = nullptr;
148   Use *CheckUse = nullptr;
149 
150   static bool parseRangeCheckICmp(Loop *L, ICmpInst *ICI, ScalarEvolution &SE,
151                                   Value *&Index, Value *&Length,
152                                   bool &IsSigned);
153 
154   static void
155   extractRangeChecksFromCond(Loop *L, ScalarEvolution &SE, Use &ConditionUse,
156                              SmallVectorImpl<InductiveRangeCheck> &Checks,
157                              SmallPtrSetImpl<Value *> &Visited);
158 
159 public:
160   const SCEV *getBegin() const { return Begin; }
161   const SCEV *getStep() const { return Step; }
162   const SCEV *getEnd() const { return End; }
163 
164   void print(raw_ostream &OS) const {
165     OS << "InductiveRangeCheck:\n";
166     OS << "  Begin: ";
167     Begin->print(OS);
168     OS << "  Step: ";
169     Step->print(OS);
170     OS << "  End: ";
171     End->print(OS);
172     OS << "\n  CheckUse: ";
173     getCheckUse()->getUser()->print(OS);
174     OS << " Operand: " << getCheckUse()->getOperandNo() << "\n";
175   }
176 
177   LLVM_DUMP_METHOD
178   void dump() {
179     print(dbgs());
180   }
181 
182   Use *getCheckUse() const { return CheckUse; }
183 
184   /// Represents an signed integer range [Range.getBegin(), Range.getEnd()).  If
185   /// R.getEnd() le R.getBegin(), then R denotes the empty range.
186 
187   class Range {
188     const SCEV *Begin;
189     const SCEV *End;
190 
191   public:
192     Range(const SCEV *Begin, const SCEV *End) : Begin(Begin), End(End) {
193       assert(Begin->getType() == End->getType() && "ill-typed range!");
194     }
195 
196     Type *getType() const { return Begin->getType(); }
197     const SCEV *getBegin() const { return Begin; }
198     const SCEV *getEnd() const { return End; }
199     bool isEmpty(ScalarEvolution &SE, bool IsSigned) const {
200       if (Begin == End)
201         return true;
202       if (IsSigned)
203         return SE.isKnownPredicate(ICmpInst::ICMP_SGE, Begin, End);
204       else
205         return SE.isKnownPredicate(ICmpInst::ICMP_UGE, Begin, End);
206     }
207   };
208 
209   /// This is the value the condition of the branch needs to evaluate to for the
210   /// branch to take the hot successor (see (1) above).
211   bool getPassingDirection() { return true; }
212 
213   /// Computes a range for the induction variable (IndVar) in which the range
214   /// check is redundant and can be constant-folded away.  The induction
215   /// variable is not required to be the canonical {0,+,1} induction variable.
216   Optional<Range> computeSafeIterationSpace(ScalarEvolution &SE,
217                                             const SCEVAddRecExpr *IndVar,
218                                             bool IsLatchSigned) const;
219 
220   /// Parse out a set of inductive range checks from \p BI and append them to \p
221   /// Checks.
222   ///
223   /// NB! There may be conditions feeding into \p BI that aren't inductive range
224   /// checks, and hence don't end up in \p Checks.
225   static void
226   extractRangeChecksFromBranch(BranchInst *BI, Loop *L, ScalarEvolution &SE,
227                                BranchProbabilityInfo *BPI,
228                                SmallVectorImpl<InductiveRangeCheck> &Checks);
229 };
230 
231 struct LoopStructure;
232 
233 class InductiveRangeCheckElimination {
234   ScalarEvolution &SE;
235   BranchProbabilityInfo *BPI;
236   DominatorTree &DT;
237   LoopInfo &LI;
238 
239   using GetBFIFunc =
240       llvm::Optional<llvm::function_ref<llvm::BlockFrequencyInfo &()> >;
241   GetBFIFunc GetBFI;
242 
243   // Returns true if it is profitable to do a transform basing on estimation of
244   // number of iterations.
245   bool isProfitableToTransform(const Loop &L, LoopStructure &LS);
246 
247 public:
248   InductiveRangeCheckElimination(ScalarEvolution &SE,
249                                  BranchProbabilityInfo *BPI, DominatorTree &DT,
250                                  LoopInfo &LI, GetBFIFunc GetBFI = None)
251       : SE(SE), BPI(BPI), DT(DT), LI(LI), GetBFI(GetBFI) {}
252 
253   bool run(Loop *L, function_ref<void(Loop *, bool)> LPMAddNewLoop);
254 };
255 
256 class IRCELegacyPass : public FunctionPass {
257 public:
258   static char ID;
259 
260   IRCELegacyPass() : FunctionPass(ID) {
261     initializeIRCELegacyPassPass(*PassRegistry::getPassRegistry());
262   }
263 
264   void getAnalysisUsage(AnalysisUsage &AU) const override {
265     AU.addRequired<BranchProbabilityInfoWrapperPass>();
266     AU.addRequired<DominatorTreeWrapperPass>();
267     AU.addPreserved<DominatorTreeWrapperPass>();
268     AU.addRequired<LoopInfoWrapperPass>();
269     AU.addPreserved<LoopInfoWrapperPass>();
270     AU.addRequired<ScalarEvolutionWrapperPass>();
271     AU.addPreserved<ScalarEvolutionWrapperPass>();
272   }
273 
274   bool runOnFunction(Function &F) override;
275 };
276 
277 } // end anonymous namespace
278 
279 char IRCELegacyPass::ID = 0;
280 
281 INITIALIZE_PASS_BEGIN(IRCELegacyPass, "irce",
282                       "Inductive range check elimination", false, false)
283 INITIALIZE_PASS_DEPENDENCY(BranchProbabilityInfoWrapperPass)
284 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
285 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
286 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
287 INITIALIZE_PASS_END(IRCELegacyPass, "irce", "Inductive range check elimination",
288                     false, false)
289 
290 /// Parse a single ICmp instruction, `ICI`, into a range check.  If `ICI` cannot
291 /// be interpreted as a range check, return false and set `Index` and `Length`
292 /// to `nullptr`.  Otherwise set `Index` to the value being range checked, and
293 /// set `Length` to the upper limit `Index` is being range checked.
294 bool
295 InductiveRangeCheck::parseRangeCheckICmp(Loop *L, ICmpInst *ICI,
296                                          ScalarEvolution &SE, Value *&Index,
297                                          Value *&Length, bool &IsSigned) {
298   auto IsLoopInvariant = [&SE, L](Value *V) {
299     return SE.isLoopInvariant(SE.getSCEV(V), L);
300   };
301 
302   ICmpInst::Predicate Pred = ICI->getPredicate();
303   Value *LHS = ICI->getOperand(0);
304   Value *RHS = ICI->getOperand(1);
305 
306   switch (Pred) {
307   default:
308     return false;
309 
310   case ICmpInst::ICMP_SLE:
311     std::swap(LHS, RHS);
312     LLVM_FALLTHROUGH;
313   case ICmpInst::ICMP_SGE:
314     IsSigned = true;
315     if (match(RHS, m_ConstantInt<0>())) {
316       Index = LHS;
317       return true; // Lower.
318     }
319     return false;
320 
321   case ICmpInst::ICMP_SLT:
322     std::swap(LHS, RHS);
323     LLVM_FALLTHROUGH;
324   case ICmpInst::ICMP_SGT:
325     IsSigned = true;
326     if (match(RHS, m_ConstantInt<-1>())) {
327       Index = LHS;
328       return true; // Lower.
329     }
330 
331     if (IsLoopInvariant(LHS)) {
332       Index = RHS;
333       Length = LHS;
334       return true; // Upper.
335     }
336     return false;
337 
338   case ICmpInst::ICMP_ULT:
339     std::swap(LHS, RHS);
340     LLVM_FALLTHROUGH;
341   case ICmpInst::ICMP_UGT:
342     IsSigned = false;
343     if (IsLoopInvariant(LHS)) {
344       Index = RHS;
345       Length = LHS;
346       return true; // Both lower and upper.
347     }
348     return false;
349   }
350 
351   llvm_unreachable("default clause returns!");
352 }
353 
354 void InductiveRangeCheck::extractRangeChecksFromCond(
355     Loop *L, ScalarEvolution &SE, Use &ConditionUse,
356     SmallVectorImpl<InductiveRangeCheck> &Checks,
357     SmallPtrSetImpl<Value *> &Visited) {
358   Value *Condition = ConditionUse.get();
359   if (!Visited.insert(Condition).second)
360     return;
361 
362   // TODO: Do the same for OR, XOR, NOT etc?
363   if (match(Condition, m_And(m_Value(), m_Value()))) {
364     extractRangeChecksFromCond(L, SE, cast<User>(Condition)->getOperandUse(0),
365                                Checks, Visited);
366     extractRangeChecksFromCond(L, SE, cast<User>(Condition)->getOperandUse(1),
367                                Checks, Visited);
368     return;
369   }
370 
371   ICmpInst *ICI = dyn_cast<ICmpInst>(Condition);
372   if (!ICI)
373     return;
374 
375   Value *Length = nullptr, *Index;
376   bool IsSigned;
377   if (!parseRangeCheckICmp(L, ICI, SE, Index, Length, IsSigned))
378     return;
379 
380   const auto *IndexAddRec = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Index));
381   bool IsAffineIndex =
382       IndexAddRec && (IndexAddRec->getLoop() == L) && IndexAddRec->isAffine();
383 
384   if (!IsAffineIndex)
385     return;
386 
387   const SCEV *End = nullptr;
388   // We strengthen "0 <= I" to "0 <= I < INT_SMAX" and "I < L" to "0 <= I < L".
389   // We can potentially do much better here.
390   if (Length)
391     End = SE.getSCEV(Length);
392   else {
393     // So far we can only reach this point for Signed range check. This may
394     // change in future. In this case we will need to pick Unsigned max for the
395     // unsigned range check.
396     unsigned BitWidth = cast<IntegerType>(IndexAddRec->getType())->getBitWidth();
397     const SCEV *SIntMax = SE.getConstant(APInt::getSignedMaxValue(BitWidth));
398     End = SIntMax;
399   }
400 
401   InductiveRangeCheck IRC;
402   IRC.End = End;
403   IRC.Begin = IndexAddRec->getStart();
404   IRC.Step = IndexAddRec->getStepRecurrence(SE);
405   IRC.CheckUse = &ConditionUse;
406   Checks.push_back(IRC);
407 }
408 
409 void InductiveRangeCheck::extractRangeChecksFromBranch(
410     BranchInst *BI, Loop *L, ScalarEvolution &SE, BranchProbabilityInfo *BPI,
411     SmallVectorImpl<InductiveRangeCheck> &Checks) {
412   if (BI->isUnconditional() || BI->getParent() == L->getLoopLatch())
413     return;
414 
415   BranchProbability LikelyTaken(15, 16);
416 
417   if (!SkipProfitabilityChecks && BPI &&
418       BPI->getEdgeProbability(BI->getParent(), (unsigned)0) < LikelyTaken)
419     return;
420 
421   SmallPtrSet<Value *, 8> Visited;
422   InductiveRangeCheck::extractRangeChecksFromCond(L, SE, BI->getOperandUse(0),
423                                                   Checks, Visited);
424 }
425 
426 // Add metadata to the loop L to disable loop optimizations. Callers need to
427 // confirm that optimizing loop L is not beneficial.
428 static void DisableAllLoopOptsOnLoop(Loop &L) {
429   // We do not care about any existing loopID related metadata for L, since we
430   // are setting all loop metadata to false.
431   LLVMContext &Context = L.getHeader()->getContext();
432   // Reserve first location for self reference to the LoopID metadata node.
433   MDNode *Dummy = MDNode::get(Context, {});
434   MDNode *DisableUnroll = MDNode::get(
435       Context, {MDString::get(Context, "llvm.loop.unroll.disable")});
436   Metadata *FalseVal =
437       ConstantAsMetadata::get(ConstantInt::get(Type::getInt1Ty(Context), 0));
438   MDNode *DisableVectorize = MDNode::get(
439       Context,
440       {MDString::get(Context, "llvm.loop.vectorize.enable"), FalseVal});
441   MDNode *DisableLICMVersioning = MDNode::get(
442       Context, {MDString::get(Context, "llvm.loop.licm_versioning.disable")});
443   MDNode *DisableDistribution= MDNode::get(
444       Context,
445       {MDString::get(Context, "llvm.loop.distribute.enable"), FalseVal});
446   MDNode *NewLoopID =
447       MDNode::get(Context, {Dummy, DisableUnroll, DisableVectorize,
448                             DisableLICMVersioning, DisableDistribution});
449   // Set operand 0 to refer to the loop id itself.
450   NewLoopID->replaceOperandWith(0, NewLoopID);
451   L.setLoopID(NewLoopID);
452 }
453 
454 namespace {
455 
456 // Keeps track of the structure of a loop.  This is similar to llvm::Loop,
457 // except that it is more lightweight and can track the state of a loop through
458 // changing and potentially invalid IR.  This structure also formalizes the
459 // kinds of loops we can deal with -- ones that have a single latch that is also
460 // an exiting block *and* have a canonical induction variable.
461 struct LoopStructure {
462   const char *Tag = "";
463 
464   BasicBlock *Header = nullptr;
465   BasicBlock *Latch = nullptr;
466 
467   // `Latch's terminator instruction is `LatchBr', and it's `LatchBrExitIdx'th
468   // successor is `LatchExit', the exit block of the loop.
469   BranchInst *LatchBr = nullptr;
470   BasicBlock *LatchExit = nullptr;
471   unsigned LatchBrExitIdx = std::numeric_limits<unsigned>::max();
472 
473   // The loop represented by this instance of LoopStructure is semantically
474   // equivalent to:
475   //
476   // intN_ty inc = IndVarIncreasing ? 1 : -1;
477   // pred_ty predicate = IndVarIncreasing ? ICMP_SLT : ICMP_SGT;
478   //
479   // for (intN_ty iv = IndVarStart; predicate(iv, LoopExitAt); iv = IndVarBase)
480   //   ... body ...
481 
482   Value *IndVarBase = nullptr;
483   Value *IndVarStart = nullptr;
484   Value *IndVarStep = nullptr;
485   Value *LoopExitAt = nullptr;
486   bool IndVarIncreasing = false;
487   bool IsSignedPredicate = true;
488 
489   LoopStructure() = default;
490 
491   template <typename M> LoopStructure map(M Map) const {
492     LoopStructure Result;
493     Result.Tag = Tag;
494     Result.Header = cast<BasicBlock>(Map(Header));
495     Result.Latch = cast<BasicBlock>(Map(Latch));
496     Result.LatchBr = cast<BranchInst>(Map(LatchBr));
497     Result.LatchExit = cast<BasicBlock>(Map(LatchExit));
498     Result.LatchBrExitIdx = LatchBrExitIdx;
499     Result.IndVarBase = Map(IndVarBase);
500     Result.IndVarStart = Map(IndVarStart);
501     Result.IndVarStep = Map(IndVarStep);
502     Result.LoopExitAt = Map(LoopExitAt);
503     Result.IndVarIncreasing = IndVarIncreasing;
504     Result.IsSignedPredicate = IsSignedPredicate;
505     return Result;
506   }
507 
508   static Optional<LoopStructure> parseLoopStructure(ScalarEvolution &, Loop &,
509                                                     const char *&);
510 };
511 
512 /// This class is used to constrain loops to run within a given iteration space.
513 /// The algorithm this class implements is given a Loop and a range [Begin,
514 /// End).  The algorithm then tries to break out a "main loop" out of the loop
515 /// it is given in a way that the "main loop" runs with the induction variable
516 /// in a subset of [Begin, End).  The algorithm emits appropriate pre and post
517 /// loops to run any remaining iterations.  The pre loop runs any iterations in
518 /// which the induction variable is < Begin, and the post loop runs any
519 /// iterations in which the induction variable is >= End.
520 class LoopConstrainer {
521   // The representation of a clone of the original loop we started out with.
522   struct ClonedLoop {
523     // The cloned blocks
524     std::vector<BasicBlock *> Blocks;
525 
526     // `Map` maps values in the clonee into values in the cloned version
527     ValueToValueMapTy Map;
528 
529     // An instance of `LoopStructure` for the cloned loop
530     LoopStructure Structure;
531   };
532 
533   // Result of rewriting the range of a loop.  See changeIterationSpaceEnd for
534   // more details on what these fields mean.
535   struct RewrittenRangeInfo {
536     BasicBlock *PseudoExit = nullptr;
537     BasicBlock *ExitSelector = nullptr;
538     std::vector<PHINode *> PHIValuesAtPseudoExit;
539     PHINode *IndVarEnd = nullptr;
540 
541     RewrittenRangeInfo() = default;
542   };
543 
544   // Calculated subranges we restrict the iteration space of the main loop to.
545   // See the implementation of `calculateSubRanges' for more details on how
546   // these fields are computed.  `LowLimit` is None if there is no restriction
547   // on low end of the restricted iteration space of the main loop.  `HighLimit`
548   // is None if there is no restriction on high end of the restricted iteration
549   // space of the main loop.
550 
551   struct SubRanges {
552     Optional<const SCEV *> LowLimit;
553     Optional<const SCEV *> HighLimit;
554   };
555 
556   // Compute a safe set of limits for the main loop to run in -- effectively the
557   // intersection of `Range' and the iteration space of the original loop.
558   // Return None if unable to compute the set of subranges.
559   Optional<SubRanges> calculateSubRanges(bool IsSignedPredicate) const;
560 
561   // Clone `OriginalLoop' and return the result in CLResult.  The IR after
562   // running `cloneLoop' is well formed except for the PHI nodes in CLResult --
563   // the PHI nodes say that there is an incoming edge from `OriginalPreheader`
564   // but there is no such edge.
565   void cloneLoop(ClonedLoop &CLResult, const char *Tag) const;
566 
567   // Create the appropriate loop structure needed to describe a cloned copy of
568   // `Original`.  The clone is described by `VM`.
569   Loop *createClonedLoopStructure(Loop *Original, Loop *Parent,
570                                   ValueToValueMapTy &VM, bool IsSubloop);
571 
572   // Rewrite the iteration space of the loop denoted by (LS, Preheader). The
573   // iteration space of the rewritten loop ends at ExitLoopAt.  The start of the
574   // iteration space is not changed.  `ExitLoopAt' is assumed to be slt
575   // `OriginalHeaderCount'.
576   //
577   // If there are iterations left to execute, control is made to jump to
578   // `ContinuationBlock', otherwise they take the normal loop exit.  The
579   // returned `RewrittenRangeInfo' object is populated as follows:
580   //
581   //  .PseudoExit is a basic block that unconditionally branches to
582   //      `ContinuationBlock'.
583   //
584   //  .ExitSelector is a basic block that decides, on exit from the loop,
585   //      whether to branch to the "true" exit or to `PseudoExit'.
586   //
587   //  .PHIValuesAtPseudoExit are PHINodes in `PseudoExit' that compute the value
588   //      for each PHINode in the loop header on taking the pseudo exit.
589   //
590   // After changeIterationSpaceEnd, `Preheader' is no longer a legitimate
591   // preheader because it is made to branch to the loop header only
592   // conditionally.
593   RewrittenRangeInfo
594   changeIterationSpaceEnd(const LoopStructure &LS, BasicBlock *Preheader,
595                           Value *ExitLoopAt,
596                           BasicBlock *ContinuationBlock) const;
597 
598   // The loop denoted by `LS' has `OldPreheader' as its preheader.  This
599   // function creates a new preheader for `LS' and returns it.
600   BasicBlock *createPreheader(const LoopStructure &LS, BasicBlock *OldPreheader,
601                               const char *Tag) const;
602 
603   // `ContinuationBlockAndPreheader' was the continuation block for some call to
604   // `changeIterationSpaceEnd' and is the preheader to the loop denoted by `LS'.
605   // This function rewrites the PHI nodes in `LS.Header' to start with the
606   // correct value.
607   void rewriteIncomingValuesForPHIs(
608       LoopStructure &LS, BasicBlock *ContinuationBlockAndPreheader,
609       const LoopConstrainer::RewrittenRangeInfo &RRI) const;
610 
611   // Even though we do not preserve any passes at this time, we at least need to
612   // keep the parent loop structure consistent.  The `LPPassManager' seems to
613   // verify this after running a loop pass.  This function adds the list of
614   // blocks denoted by BBs to this loops parent loop if required.
615   void addToParentLoopIfNeeded(ArrayRef<BasicBlock *> BBs);
616 
617   // Some global state.
618   Function &F;
619   LLVMContext &Ctx;
620   ScalarEvolution &SE;
621   DominatorTree &DT;
622   LoopInfo &LI;
623   function_ref<void(Loop *, bool)> LPMAddNewLoop;
624 
625   // Information about the original loop we started out with.
626   Loop &OriginalLoop;
627 
628   const SCEV *LatchTakenCount = nullptr;
629   BasicBlock *OriginalPreheader = nullptr;
630 
631   // The preheader of the main loop.  This may or may not be different from
632   // `OriginalPreheader'.
633   BasicBlock *MainLoopPreheader = nullptr;
634 
635   // The range we need to run the main loop in.
636   InductiveRangeCheck::Range Range;
637 
638   // The structure of the main loop (see comment at the beginning of this class
639   // for a definition)
640   LoopStructure MainLoopStructure;
641 
642 public:
643   LoopConstrainer(Loop &L, LoopInfo &LI,
644                   function_ref<void(Loop *, bool)> LPMAddNewLoop,
645                   const LoopStructure &LS, ScalarEvolution &SE,
646                   DominatorTree &DT, InductiveRangeCheck::Range R)
647       : F(*L.getHeader()->getParent()), Ctx(L.getHeader()->getContext()),
648         SE(SE), DT(DT), LI(LI), LPMAddNewLoop(LPMAddNewLoop), OriginalLoop(L),
649         Range(R), MainLoopStructure(LS) {}
650 
651   // Entry point for the algorithm.  Returns true on success.
652   bool run();
653 };
654 
655 } // end anonymous namespace
656 
657 /// Given a loop with an deccreasing induction variable, is it possible to
658 /// safely calculate the bounds of a new loop using the given Predicate.
659 static bool isSafeDecreasingBound(const SCEV *Start,
660                                   const SCEV *BoundSCEV, const SCEV *Step,
661                                   ICmpInst::Predicate Pred,
662                                   unsigned LatchBrExitIdx,
663                                   Loop *L, ScalarEvolution &SE) {
664   if (Pred != ICmpInst::ICMP_SLT && Pred != ICmpInst::ICMP_SGT &&
665       Pred != ICmpInst::ICMP_ULT && Pred != ICmpInst::ICMP_UGT)
666     return false;
667 
668   if (!SE.isAvailableAtLoopEntry(BoundSCEV, L))
669     return false;
670 
671   assert(SE.isKnownNegative(Step) && "expecting negative step");
672 
673   LLVM_DEBUG(dbgs() << "irce: isSafeDecreasingBound with:\n");
674   LLVM_DEBUG(dbgs() << "irce: Start: " << *Start << "\n");
675   LLVM_DEBUG(dbgs() << "irce: Step: " << *Step << "\n");
676   LLVM_DEBUG(dbgs() << "irce: BoundSCEV: " << *BoundSCEV << "\n");
677   LLVM_DEBUG(dbgs() << "irce: Pred: " << ICmpInst::getPredicateName(Pred)
678                     << "\n");
679   LLVM_DEBUG(dbgs() << "irce: LatchExitBrIdx: " << LatchBrExitIdx << "\n");
680 
681   bool IsSigned = ICmpInst::isSigned(Pred);
682   // The predicate that we need to check that the induction variable lies
683   // within bounds.
684   ICmpInst::Predicate BoundPred =
685     IsSigned ? CmpInst::ICMP_SGT : CmpInst::ICMP_UGT;
686 
687   if (LatchBrExitIdx == 1)
688     return SE.isLoopEntryGuardedByCond(L, BoundPred, Start, BoundSCEV);
689 
690   assert(LatchBrExitIdx == 0 &&
691          "LatchBrExitIdx should be either 0 or 1");
692 
693   const SCEV *StepPlusOne = SE.getAddExpr(Step, SE.getOne(Step->getType()));
694   unsigned BitWidth = cast<IntegerType>(BoundSCEV->getType())->getBitWidth();
695   APInt Min = IsSigned ? APInt::getSignedMinValue(BitWidth) :
696     APInt::getMinValue(BitWidth);
697   const SCEV *Limit = SE.getMinusSCEV(SE.getConstant(Min), StepPlusOne);
698 
699   const SCEV *MinusOne =
700     SE.getMinusSCEV(BoundSCEV, SE.getOne(BoundSCEV->getType()));
701 
702   return SE.isLoopEntryGuardedByCond(L, BoundPred, Start, MinusOne) &&
703          SE.isLoopEntryGuardedByCond(L, BoundPred, BoundSCEV, Limit);
704 
705 }
706 
707 /// Given a loop with an increasing induction variable, is it possible to
708 /// safely calculate the bounds of a new loop using the given Predicate.
709 static bool isSafeIncreasingBound(const SCEV *Start,
710                                   const SCEV *BoundSCEV, const SCEV *Step,
711                                   ICmpInst::Predicate Pred,
712                                   unsigned LatchBrExitIdx,
713                                   Loop *L, ScalarEvolution &SE) {
714   if (Pred != ICmpInst::ICMP_SLT && Pred != ICmpInst::ICMP_SGT &&
715       Pred != ICmpInst::ICMP_ULT && Pred != ICmpInst::ICMP_UGT)
716     return false;
717 
718   if (!SE.isAvailableAtLoopEntry(BoundSCEV, L))
719     return false;
720 
721   LLVM_DEBUG(dbgs() << "irce: isSafeIncreasingBound with:\n");
722   LLVM_DEBUG(dbgs() << "irce: Start: " << *Start << "\n");
723   LLVM_DEBUG(dbgs() << "irce: Step: " << *Step << "\n");
724   LLVM_DEBUG(dbgs() << "irce: BoundSCEV: " << *BoundSCEV << "\n");
725   LLVM_DEBUG(dbgs() << "irce: Pred: " << ICmpInst::getPredicateName(Pred)
726                     << "\n");
727   LLVM_DEBUG(dbgs() << "irce: LatchExitBrIdx: " << LatchBrExitIdx << "\n");
728 
729   bool IsSigned = ICmpInst::isSigned(Pred);
730   // The predicate that we need to check that the induction variable lies
731   // within bounds.
732   ICmpInst::Predicate BoundPred =
733       IsSigned ? CmpInst::ICMP_SLT : CmpInst::ICMP_ULT;
734 
735   if (LatchBrExitIdx == 1)
736     return SE.isLoopEntryGuardedByCond(L, BoundPred, Start, BoundSCEV);
737 
738   assert(LatchBrExitIdx == 0 && "LatchBrExitIdx should be 0 or 1");
739 
740   const SCEV *StepMinusOne =
741     SE.getMinusSCEV(Step, SE.getOne(Step->getType()));
742   unsigned BitWidth = cast<IntegerType>(BoundSCEV->getType())->getBitWidth();
743   APInt Max = IsSigned ? APInt::getSignedMaxValue(BitWidth) :
744     APInt::getMaxValue(BitWidth);
745   const SCEV *Limit = SE.getMinusSCEV(SE.getConstant(Max), StepMinusOne);
746 
747   return (SE.isLoopEntryGuardedByCond(L, BoundPred, Start,
748                                       SE.getAddExpr(BoundSCEV, Step)) &&
749           SE.isLoopEntryGuardedByCond(L, BoundPred, BoundSCEV, Limit));
750 }
751 
752 Optional<LoopStructure>
753 LoopStructure::parseLoopStructure(ScalarEvolution &SE, Loop &L,
754                                   const char *&FailureReason) {
755   if (!L.isLoopSimplifyForm()) {
756     FailureReason = "loop not in LoopSimplify form";
757     return None;
758   }
759 
760   BasicBlock *Latch = L.getLoopLatch();
761   assert(Latch && "Simplified loops only have one latch!");
762 
763   if (Latch->getTerminator()->getMetadata(ClonedLoopTag)) {
764     FailureReason = "loop has already been cloned";
765     return None;
766   }
767 
768   if (!L.isLoopExiting(Latch)) {
769     FailureReason = "no loop latch";
770     return None;
771   }
772 
773   BasicBlock *Header = L.getHeader();
774   BasicBlock *Preheader = L.getLoopPreheader();
775   if (!Preheader) {
776     FailureReason = "no preheader";
777     return None;
778   }
779 
780   BranchInst *LatchBr = dyn_cast<BranchInst>(Latch->getTerminator());
781   if (!LatchBr || LatchBr->isUnconditional()) {
782     FailureReason = "latch terminator not conditional branch";
783     return None;
784   }
785 
786   unsigned LatchBrExitIdx = LatchBr->getSuccessor(0) == Header ? 1 : 0;
787 
788   ICmpInst *ICI = dyn_cast<ICmpInst>(LatchBr->getCondition());
789   if (!ICI || !isa<IntegerType>(ICI->getOperand(0)->getType())) {
790     FailureReason = "latch terminator branch not conditional on integral icmp";
791     return None;
792   }
793 
794   const SCEV *LatchCount = SE.getExitCount(&L, Latch);
795   if (isa<SCEVCouldNotCompute>(LatchCount)) {
796     FailureReason = "could not compute latch count";
797     return None;
798   }
799 
800   ICmpInst::Predicate Pred = ICI->getPredicate();
801   Value *LeftValue = ICI->getOperand(0);
802   const SCEV *LeftSCEV = SE.getSCEV(LeftValue);
803   IntegerType *IndVarTy = cast<IntegerType>(LeftValue->getType());
804 
805   Value *RightValue = ICI->getOperand(1);
806   const SCEV *RightSCEV = SE.getSCEV(RightValue);
807 
808   // We canonicalize `ICI` such that `LeftSCEV` is an add recurrence.
809   if (!isa<SCEVAddRecExpr>(LeftSCEV)) {
810     if (isa<SCEVAddRecExpr>(RightSCEV)) {
811       std::swap(LeftSCEV, RightSCEV);
812       std::swap(LeftValue, RightValue);
813       Pred = ICmpInst::getSwappedPredicate(Pred);
814     } else {
815       FailureReason = "no add recurrences in the icmp";
816       return None;
817     }
818   }
819 
820   auto HasNoSignedWrap = [&](const SCEVAddRecExpr *AR) {
821     if (AR->getNoWrapFlags(SCEV::FlagNSW))
822       return true;
823 
824     IntegerType *Ty = cast<IntegerType>(AR->getType());
825     IntegerType *WideTy =
826         IntegerType::get(Ty->getContext(), Ty->getBitWidth() * 2);
827 
828     const SCEVAddRecExpr *ExtendAfterOp =
829         dyn_cast<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
830     if (ExtendAfterOp) {
831       const SCEV *ExtendedStart = SE.getSignExtendExpr(AR->getStart(), WideTy);
832       const SCEV *ExtendedStep =
833           SE.getSignExtendExpr(AR->getStepRecurrence(SE), WideTy);
834 
835       bool NoSignedWrap = ExtendAfterOp->getStart() == ExtendedStart &&
836                           ExtendAfterOp->getStepRecurrence(SE) == ExtendedStep;
837 
838       if (NoSignedWrap)
839         return true;
840     }
841 
842     // We may have proved this when computing the sign extension above.
843     return AR->getNoWrapFlags(SCEV::FlagNSW) != SCEV::FlagAnyWrap;
844   };
845 
846   // `ICI` is interpreted as taking the backedge if the *next* value of the
847   // induction variable satisfies some constraint.
848 
849   const SCEVAddRecExpr *IndVarBase = cast<SCEVAddRecExpr>(LeftSCEV);
850   if (!IndVarBase->isAffine()) {
851     FailureReason = "LHS in icmp not induction variable";
852     return None;
853   }
854   const SCEV* StepRec = IndVarBase->getStepRecurrence(SE);
855   if (!isa<SCEVConstant>(StepRec)) {
856     FailureReason = "LHS in icmp not induction variable";
857     return None;
858   }
859   ConstantInt *StepCI = cast<SCEVConstant>(StepRec)->getValue();
860 
861   if (ICI->isEquality() && !HasNoSignedWrap(IndVarBase)) {
862     FailureReason = "LHS in icmp needs nsw for equality predicates";
863     return None;
864   }
865 
866   assert(!StepCI->isZero() && "Zero step?");
867   bool IsIncreasing = !StepCI->isNegative();
868   bool IsSignedPredicate;
869   const SCEV *StartNext = IndVarBase->getStart();
870   const SCEV *Addend = SE.getNegativeSCEV(IndVarBase->getStepRecurrence(SE));
871   const SCEV *IndVarStart = SE.getAddExpr(StartNext, Addend);
872   const SCEV *Step = SE.getSCEV(StepCI);
873 
874   const SCEV *FixedRightSCEV = nullptr;
875 
876   // If RightValue resides within loop (but still being loop invariant),
877   // regenerate it as preheader.
878   if (auto *I = dyn_cast<Instruction>(RightValue))
879     if (L.contains(I->getParent()))
880       FixedRightSCEV = RightSCEV;
881 
882   if (IsIncreasing) {
883     bool DecreasedRightValueByOne = false;
884     if (StepCI->isOne()) {
885       // Try to turn eq/ne predicates to those we can work with.
886       if (Pred == ICmpInst::ICMP_NE && LatchBrExitIdx == 1)
887         // while (++i != len) {         while (++i < len) {
888         //   ...                 --->     ...
889         // }                            }
890         // If both parts are known non-negative, it is profitable to use
891         // unsigned comparison in increasing loop. This allows us to make the
892         // comparison check against "RightSCEV + 1" more optimistic.
893         if (isKnownNonNegativeInLoop(IndVarStart, &L, SE) &&
894             isKnownNonNegativeInLoop(RightSCEV, &L, SE))
895           Pred = ICmpInst::ICMP_ULT;
896         else
897           Pred = ICmpInst::ICMP_SLT;
898       else if (Pred == ICmpInst::ICMP_EQ && LatchBrExitIdx == 0) {
899         // while (true) {               while (true) {
900         //   if (++i == len)     --->     if (++i > len - 1)
901         //     break;                       break;
902         //   ...                          ...
903         // }                            }
904         if (IndVarBase->getNoWrapFlags(SCEV::FlagNUW) &&
905             cannotBeMinInLoop(RightSCEV, &L, SE, /*Signed*/false)) {
906           Pred = ICmpInst::ICMP_UGT;
907           RightSCEV = SE.getMinusSCEV(RightSCEV,
908                                       SE.getOne(RightSCEV->getType()));
909           DecreasedRightValueByOne = true;
910         } else if (cannotBeMinInLoop(RightSCEV, &L, SE, /*Signed*/true)) {
911           Pred = ICmpInst::ICMP_SGT;
912           RightSCEV = SE.getMinusSCEV(RightSCEV,
913                                       SE.getOne(RightSCEV->getType()));
914           DecreasedRightValueByOne = true;
915         }
916       }
917     }
918 
919     bool LTPred = (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_ULT);
920     bool GTPred = (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_UGT);
921     bool FoundExpectedPred =
922         (LTPred && LatchBrExitIdx == 1) || (GTPred && LatchBrExitIdx == 0);
923 
924     if (!FoundExpectedPred) {
925       FailureReason = "expected icmp slt semantically, found something else";
926       return None;
927     }
928 
929     IsSignedPredicate = ICmpInst::isSigned(Pred);
930     if (!IsSignedPredicate && !AllowUnsignedLatchCondition) {
931       FailureReason = "unsigned latch conditions are explicitly prohibited";
932       return None;
933     }
934 
935     if (!isSafeIncreasingBound(IndVarStart, RightSCEV, Step, Pred,
936                                LatchBrExitIdx, &L, SE)) {
937       FailureReason = "Unsafe loop bounds";
938       return None;
939     }
940     if (LatchBrExitIdx == 0) {
941       // We need to increase the right value unless we have already decreased
942       // it virtually when we replaced EQ with SGT.
943       if (!DecreasedRightValueByOne)
944         FixedRightSCEV =
945             SE.getAddExpr(RightSCEV, SE.getOne(RightSCEV->getType()));
946     } else {
947       assert(!DecreasedRightValueByOne &&
948              "Right value can be decreased only for LatchBrExitIdx == 0!");
949     }
950   } else {
951     bool IncreasedRightValueByOne = false;
952     if (StepCI->isMinusOne()) {
953       // Try to turn eq/ne predicates to those we can work with.
954       if (Pred == ICmpInst::ICMP_NE && LatchBrExitIdx == 1)
955         // while (--i != len) {         while (--i > len) {
956         //   ...                 --->     ...
957         // }                            }
958         // We intentionally don't turn the predicate into UGT even if we know
959         // that both operands are non-negative, because it will only pessimize
960         // our check against "RightSCEV - 1".
961         Pred = ICmpInst::ICMP_SGT;
962       else if (Pred == ICmpInst::ICMP_EQ && LatchBrExitIdx == 0) {
963         // while (true) {               while (true) {
964         //   if (--i == len)     --->     if (--i < len + 1)
965         //     break;                       break;
966         //   ...                          ...
967         // }                            }
968         if (IndVarBase->getNoWrapFlags(SCEV::FlagNUW) &&
969             cannotBeMaxInLoop(RightSCEV, &L, SE, /* Signed */ false)) {
970           Pred = ICmpInst::ICMP_ULT;
971           RightSCEV = SE.getAddExpr(RightSCEV, SE.getOne(RightSCEV->getType()));
972           IncreasedRightValueByOne = true;
973         } else if (cannotBeMaxInLoop(RightSCEV, &L, SE, /* Signed */ true)) {
974           Pred = ICmpInst::ICMP_SLT;
975           RightSCEV = SE.getAddExpr(RightSCEV, SE.getOne(RightSCEV->getType()));
976           IncreasedRightValueByOne = true;
977         }
978       }
979     }
980 
981     bool LTPred = (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_ULT);
982     bool GTPred = (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_UGT);
983 
984     bool FoundExpectedPred =
985         (GTPred && LatchBrExitIdx == 1) || (LTPred && LatchBrExitIdx == 0);
986 
987     if (!FoundExpectedPred) {
988       FailureReason = "expected icmp sgt semantically, found something else";
989       return None;
990     }
991 
992     IsSignedPredicate =
993         Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGT;
994 
995     if (!IsSignedPredicate && !AllowUnsignedLatchCondition) {
996       FailureReason = "unsigned latch conditions are explicitly prohibited";
997       return None;
998     }
999 
1000     if (!isSafeDecreasingBound(IndVarStart, RightSCEV, Step, Pred,
1001                                LatchBrExitIdx, &L, SE)) {
1002       FailureReason = "Unsafe bounds";
1003       return None;
1004     }
1005 
1006     if (LatchBrExitIdx == 0) {
1007       // We need to decrease the right value unless we have already increased
1008       // it virtually when we replaced EQ with SLT.
1009       if (!IncreasedRightValueByOne)
1010         FixedRightSCEV =
1011             SE.getMinusSCEV(RightSCEV, SE.getOne(RightSCEV->getType()));
1012     } else {
1013       assert(!IncreasedRightValueByOne &&
1014              "Right value can be increased only for LatchBrExitIdx == 0!");
1015     }
1016   }
1017   BasicBlock *LatchExit = LatchBr->getSuccessor(LatchBrExitIdx);
1018 
1019   assert(SE.getLoopDisposition(LatchCount, &L) ==
1020              ScalarEvolution::LoopInvariant &&
1021          "loop variant exit count doesn't make sense!");
1022 
1023   assert(!L.contains(LatchExit) && "expected an exit block!");
1024   const DataLayout &DL = Preheader->getModule()->getDataLayout();
1025   SCEVExpander Expander(SE, DL, "irce");
1026   Instruction *Ins = Preheader->getTerminator();
1027 
1028   if (FixedRightSCEV)
1029     RightValue =
1030         Expander.expandCodeFor(FixedRightSCEV, FixedRightSCEV->getType(), Ins);
1031 
1032   Value *IndVarStartV = Expander.expandCodeFor(IndVarStart, IndVarTy, Ins);
1033   IndVarStartV->setName("indvar.start");
1034 
1035   LoopStructure Result;
1036 
1037   Result.Tag = "main";
1038   Result.Header = Header;
1039   Result.Latch = Latch;
1040   Result.LatchBr = LatchBr;
1041   Result.LatchExit = LatchExit;
1042   Result.LatchBrExitIdx = LatchBrExitIdx;
1043   Result.IndVarStart = IndVarStartV;
1044   Result.IndVarStep = StepCI;
1045   Result.IndVarBase = LeftValue;
1046   Result.IndVarIncreasing = IsIncreasing;
1047   Result.LoopExitAt = RightValue;
1048   Result.IsSignedPredicate = IsSignedPredicate;
1049 
1050   FailureReason = nullptr;
1051 
1052   return Result;
1053 }
1054 
1055 /// If the type of \p S matches with \p Ty, return \p S. Otherwise, return
1056 /// signed or unsigned extension of \p S to type \p Ty.
1057 static const SCEV *NoopOrExtend(const SCEV *S, Type *Ty, ScalarEvolution &SE,
1058                                 bool Signed) {
1059   return Signed ? SE.getNoopOrSignExtend(S, Ty) : SE.getNoopOrZeroExtend(S, Ty);
1060 }
1061 
1062 Optional<LoopConstrainer::SubRanges>
1063 LoopConstrainer::calculateSubRanges(bool IsSignedPredicate) const {
1064   IntegerType *Ty = cast<IntegerType>(LatchTakenCount->getType());
1065 
1066   auto *RTy = cast<IntegerType>(Range.getType());
1067 
1068   // We only support wide range checks and narrow latches.
1069   if (!AllowNarrowLatchCondition && RTy != Ty)
1070     return None;
1071   if (RTy->getBitWidth() < Ty->getBitWidth())
1072     return None;
1073 
1074   LoopConstrainer::SubRanges Result;
1075 
1076   // I think we can be more aggressive here and make this nuw / nsw if the
1077   // addition that feeds into the icmp for the latch's terminating branch is nuw
1078   // / nsw.  In any case, a wrapping 2's complement addition is safe.
1079   const SCEV *Start = NoopOrExtend(SE.getSCEV(MainLoopStructure.IndVarStart),
1080                                    RTy, SE, IsSignedPredicate);
1081   const SCEV *End = NoopOrExtend(SE.getSCEV(MainLoopStructure.LoopExitAt), RTy,
1082                                  SE, IsSignedPredicate);
1083 
1084   bool Increasing = MainLoopStructure.IndVarIncreasing;
1085 
1086   // We compute `Smallest` and `Greatest` such that [Smallest, Greatest), or
1087   // [Smallest, GreatestSeen] is the range of values the induction variable
1088   // takes.
1089 
1090   const SCEV *Smallest = nullptr, *Greatest = nullptr, *GreatestSeen = nullptr;
1091 
1092   const SCEV *One = SE.getOne(RTy);
1093   if (Increasing) {
1094     Smallest = Start;
1095     Greatest = End;
1096     // No overflow, because the range [Smallest, GreatestSeen] is not empty.
1097     GreatestSeen = SE.getMinusSCEV(End, One);
1098   } else {
1099     // These two computations may sign-overflow.  Here is why that is okay:
1100     //
1101     // We know that the induction variable does not sign-overflow on any
1102     // iteration except the last one, and it starts at `Start` and ends at
1103     // `End`, decrementing by one every time.
1104     //
1105     //  * if `Smallest` sign-overflows we know `End` is `INT_SMAX`. Since the
1106     //    induction variable is decreasing we know that that the smallest value
1107     //    the loop body is actually executed with is `INT_SMIN` == `Smallest`.
1108     //
1109     //  * if `Greatest` sign-overflows, we know it can only be `INT_SMIN`.  In
1110     //    that case, `Clamp` will always return `Smallest` and
1111     //    [`Result.LowLimit`, `Result.HighLimit`) = [`Smallest`, `Smallest`)
1112     //    will be an empty range.  Returning an empty range is always safe.
1113 
1114     Smallest = SE.getAddExpr(End, One);
1115     Greatest = SE.getAddExpr(Start, One);
1116     GreatestSeen = Start;
1117   }
1118 
1119   auto Clamp = [this, Smallest, Greatest, IsSignedPredicate](const SCEV *S) {
1120     return IsSignedPredicate
1121                ? SE.getSMaxExpr(Smallest, SE.getSMinExpr(Greatest, S))
1122                : SE.getUMaxExpr(Smallest, SE.getUMinExpr(Greatest, S));
1123   };
1124 
1125   // In some cases we can prove that we don't need a pre or post loop.
1126   ICmpInst::Predicate PredLE =
1127       IsSignedPredicate ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
1128   ICmpInst::Predicate PredLT =
1129       IsSignedPredicate ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1130 
1131   bool ProvablyNoPreloop =
1132       SE.isKnownPredicate(PredLE, Range.getBegin(), Smallest);
1133   if (!ProvablyNoPreloop)
1134     Result.LowLimit = Clamp(Range.getBegin());
1135 
1136   bool ProvablyNoPostLoop =
1137       SE.isKnownPredicate(PredLT, GreatestSeen, Range.getEnd());
1138   if (!ProvablyNoPostLoop)
1139     Result.HighLimit = Clamp(Range.getEnd());
1140 
1141   return Result;
1142 }
1143 
1144 void LoopConstrainer::cloneLoop(LoopConstrainer::ClonedLoop &Result,
1145                                 const char *Tag) const {
1146   for (BasicBlock *BB : OriginalLoop.getBlocks()) {
1147     BasicBlock *Clone = CloneBasicBlock(BB, Result.Map, Twine(".") + Tag, &F);
1148     Result.Blocks.push_back(Clone);
1149     Result.Map[BB] = Clone;
1150   }
1151 
1152   auto GetClonedValue = [&Result](Value *V) {
1153     assert(V && "null values not in domain!");
1154     auto It = Result.Map.find(V);
1155     if (It == Result.Map.end())
1156       return V;
1157     return static_cast<Value *>(It->second);
1158   };
1159 
1160   auto *ClonedLatch =
1161       cast<BasicBlock>(GetClonedValue(OriginalLoop.getLoopLatch()));
1162   ClonedLatch->getTerminator()->setMetadata(ClonedLoopTag,
1163                                             MDNode::get(Ctx, {}));
1164 
1165   Result.Structure = MainLoopStructure.map(GetClonedValue);
1166   Result.Structure.Tag = Tag;
1167 
1168   for (unsigned i = 0, e = Result.Blocks.size(); i != e; ++i) {
1169     BasicBlock *ClonedBB = Result.Blocks[i];
1170     BasicBlock *OriginalBB = OriginalLoop.getBlocks()[i];
1171 
1172     assert(Result.Map[OriginalBB] == ClonedBB && "invariant!");
1173 
1174     for (Instruction &I : *ClonedBB)
1175       RemapInstruction(&I, Result.Map,
1176                        RF_NoModuleLevelChanges | RF_IgnoreMissingLocals);
1177 
1178     // Exit blocks will now have one more predecessor and their PHI nodes need
1179     // to be edited to reflect that.  No phi nodes need to be introduced because
1180     // the loop is in LCSSA.
1181 
1182     for (auto *SBB : successors(OriginalBB)) {
1183       if (OriginalLoop.contains(SBB))
1184         continue; // not an exit block
1185 
1186       for (PHINode &PN : SBB->phis()) {
1187         Value *OldIncoming = PN.getIncomingValueForBlock(OriginalBB);
1188         PN.addIncoming(GetClonedValue(OldIncoming), ClonedBB);
1189       }
1190     }
1191   }
1192 }
1193 
1194 LoopConstrainer::RewrittenRangeInfo LoopConstrainer::changeIterationSpaceEnd(
1195     const LoopStructure &LS, BasicBlock *Preheader, Value *ExitSubloopAt,
1196     BasicBlock *ContinuationBlock) const {
1197   // We start with a loop with a single latch:
1198   //
1199   //    +--------------------+
1200   //    |                    |
1201   //    |     preheader      |
1202   //    |                    |
1203   //    +--------+-----------+
1204   //             |      ----------------\
1205   //             |     /                |
1206   //    +--------v----v------+          |
1207   //    |                    |          |
1208   //    |      header        |          |
1209   //    |                    |          |
1210   //    +--------------------+          |
1211   //                                    |
1212   //            .....                   |
1213   //                                    |
1214   //    +--------------------+          |
1215   //    |                    |          |
1216   //    |       latch        >----------/
1217   //    |                    |
1218   //    +-------v------------+
1219   //            |
1220   //            |
1221   //            |   +--------------------+
1222   //            |   |                    |
1223   //            +--->   original exit    |
1224   //                |                    |
1225   //                +--------------------+
1226   //
1227   // We change the control flow to look like
1228   //
1229   //
1230   //    +--------------------+
1231   //    |                    |
1232   //    |     preheader      >-------------------------+
1233   //    |                    |                         |
1234   //    +--------v-----------+                         |
1235   //             |    /-------------+                  |
1236   //             |   /              |                  |
1237   //    +--------v--v--------+      |                  |
1238   //    |                    |      |                  |
1239   //    |      header        |      |   +--------+     |
1240   //    |                    |      |   |        |     |
1241   //    +--------------------+      |   |  +-----v-----v-----------+
1242   //                                |   |  |                       |
1243   //                                |   |  |     .pseudo.exit      |
1244   //                                |   |  |                       |
1245   //                                |   |  +-----------v-----------+
1246   //                                |   |              |
1247   //            .....               |   |              |
1248   //                                |   |     +--------v-------------+
1249   //    +--------------------+      |   |     |                      |
1250   //    |                    |      |   |     |   ContinuationBlock  |
1251   //    |       latch        >------+   |     |                      |
1252   //    |                    |          |     +----------------------+
1253   //    +---------v----------+          |
1254   //              |                     |
1255   //              |                     |
1256   //              |     +---------------^-----+
1257   //              |     |                     |
1258   //              +----->    .exit.selector   |
1259   //                    |                     |
1260   //                    +----------v----------+
1261   //                               |
1262   //     +--------------------+    |
1263   //     |                    |    |
1264   //     |   original exit    <----+
1265   //     |                    |
1266   //     +--------------------+
1267 
1268   RewrittenRangeInfo RRI;
1269 
1270   BasicBlock *BBInsertLocation = LS.Latch->getNextNode();
1271   RRI.ExitSelector = BasicBlock::Create(Ctx, Twine(LS.Tag) + ".exit.selector",
1272                                         &F, BBInsertLocation);
1273   RRI.PseudoExit = BasicBlock::Create(Ctx, Twine(LS.Tag) + ".pseudo.exit", &F,
1274                                       BBInsertLocation);
1275 
1276   BranchInst *PreheaderJump = cast<BranchInst>(Preheader->getTerminator());
1277   bool Increasing = LS.IndVarIncreasing;
1278   bool IsSignedPredicate = LS.IsSignedPredicate;
1279 
1280   IRBuilder<> B(PreheaderJump);
1281   auto *RangeTy = Range.getBegin()->getType();
1282   auto NoopOrExt = [&](Value *V) {
1283     if (V->getType() == RangeTy)
1284       return V;
1285     return IsSignedPredicate ? B.CreateSExt(V, RangeTy, "wide." + V->getName())
1286                              : B.CreateZExt(V, RangeTy, "wide." + V->getName());
1287   };
1288 
1289   // EnterLoopCond - is it okay to start executing this `LS'?
1290   Value *EnterLoopCond = nullptr;
1291   auto Pred =
1292       Increasing
1293           ? (IsSignedPredicate ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT)
1294           : (IsSignedPredicate ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
1295   Value *IndVarStart = NoopOrExt(LS.IndVarStart);
1296   EnterLoopCond = B.CreateICmp(Pred, IndVarStart, ExitSubloopAt);
1297 
1298   B.CreateCondBr(EnterLoopCond, LS.Header, RRI.PseudoExit);
1299   PreheaderJump->eraseFromParent();
1300 
1301   LS.LatchBr->setSuccessor(LS.LatchBrExitIdx, RRI.ExitSelector);
1302   B.SetInsertPoint(LS.LatchBr);
1303   Value *IndVarBase = NoopOrExt(LS.IndVarBase);
1304   Value *TakeBackedgeLoopCond = B.CreateICmp(Pred, IndVarBase, ExitSubloopAt);
1305 
1306   Value *CondForBranch = LS.LatchBrExitIdx == 1
1307                              ? TakeBackedgeLoopCond
1308                              : B.CreateNot(TakeBackedgeLoopCond);
1309 
1310   LS.LatchBr->setCondition(CondForBranch);
1311 
1312   B.SetInsertPoint(RRI.ExitSelector);
1313 
1314   // IterationsLeft - are there any more iterations left, given the original
1315   // upper bound on the induction variable?  If not, we branch to the "real"
1316   // exit.
1317   Value *LoopExitAt = NoopOrExt(LS.LoopExitAt);
1318   Value *IterationsLeft = B.CreateICmp(Pred, IndVarBase, LoopExitAt);
1319   B.CreateCondBr(IterationsLeft, RRI.PseudoExit, LS.LatchExit);
1320 
1321   BranchInst *BranchToContinuation =
1322       BranchInst::Create(ContinuationBlock, RRI.PseudoExit);
1323 
1324   // We emit PHI nodes into `RRI.PseudoExit' that compute the "latest" value of
1325   // each of the PHI nodes in the loop header.  This feeds into the initial
1326   // value of the same PHI nodes if/when we continue execution.
1327   for (PHINode &PN : LS.Header->phis()) {
1328     PHINode *NewPHI = PHINode::Create(PN.getType(), 2, PN.getName() + ".copy",
1329                                       BranchToContinuation);
1330 
1331     NewPHI->addIncoming(PN.getIncomingValueForBlock(Preheader), Preheader);
1332     NewPHI->addIncoming(PN.getIncomingValueForBlock(LS.Latch),
1333                         RRI.ExitSelector);
1334     RRI.PHIValuesAtPseudoExit.push_back(NewPHI);
1335   }
1336 
1337   RRI.IndVarEnd = PHINode::Create(IndVarBase->getType(), 2, "indvar.end",
1338                                   BranchToContinuation);
1339   RRI.IndVarEnd->addIncoming(IndVarStart, Preheader);
1340   RRI.IndVarEnd->addIncoming(IndVarBase, RRI.ExitSelector);
1341 
1342   // The latch exit now has a branch from `RRI.ExitSelector' instead of
1343   // `LS.Latch'.  The PHI nodes need to be updated to reflect that.
1344   LS.LatchExit->replacePhiUsesWith(LS.Latch, RRI.ExitSelector);
1345 
1346   return RRI;
1347 }
1348 
1349 void LoopConstrainer::rewriteIncomingValuesForPHIs(
1350     LoopStructure &LS, BasicBlock *ContinuationBlock,
1351     const LoopConstrainer::RewrittenRangeInfo &RRI) const {
1352   unsigned PHIIndex = 0;
1353   for (PHINode &PN : LS.Header->phis())
1354     PN.setIncomingValueForBlock(ContinuationBlock,
1355                                 RRI.PHIValuesAtPseudoExit[PHIIndex++]);
1356 
1357   LS.IndVarStart = RRI.IndVarEnd;
1358 }
1359 
1360 BasicBlock *LoopConstrainer::createPreheader(const LoopStructure &LS,
1361                                              BasicBlock *OldPreheader,
1362                                              const char *Tag) const {
1363   BasicBlock *Preheader = BasicBlock::Create(Ctx, Tag, &F, LS.Header);
1364   BranchInst::Create(LS.Header, Preheader);
1365 
1366   LS.Header->replacePhiUsesWith(OldPreheader, Preheader);
1367 
1368   return Preheader;
1369 }
1370 
1371 void LoopConstrainer::addToParentLoopIfNeeded(ArrayRef<BasicBlock *> BBs) {
1372   Loop *ParentLoop = OriginalLoop.getParentLoop();
1373   if (!ParentLoop)
1374     return;
1375 
1376   for (BasicBlock *BB : BBs)
1377     ParentLoop->addBasicBlockToLoop(BB, LI);
1378 }
1379 
1380 Loop *LoopConstrainer::createClonedLoopStructure(Loop *Original, Loop *Parent,
1381                                                  ValueToValueMapTy &VM,
1382                                                  bool IsSubloop) {
1383   Loop &New = *LI.AllocateLoop();
1384   if (Parent)
1385     Parent->addChildLoop(&New);
1386   else
1387     LI.addTopLevelLoop(&New);
1388   LPMAddNewLoop(&New, IsSubloop);
1389 
1390   // Add all of the blocks in Original to the new loop.
1391   for (auto *BB : Original->blocks())
1392     if (LI.getLoopFor(BB) == Original)
1393       New.addBasicBlockToLoop(cast<BasicBlock>(VM[BB]), LI);
1394 
1395   // Add all of the subloops to the new loop.
1396   for (Loop *SubLoop : *Original)
1397     createClonedLoopStructure(SubLoop, &New, VM, /* IsSubloop */ true);
1398 
1399   return &New;
1400 }
1401 
1402 bool LoopConstrainer::run() {
1403   BasicBlock *Preheader = nullptr;
1404   LatchTakenCount = SE.getExitCount(&OriginalLoop, MainLoopStructure.Latch);
1405   Preheader = OriginalLoop.getLoopPreheader();
1406   assert(!isa<SCEVCouldNotCompute>(LatchTakenCount) && Preheader != nullptr &&
1407          "preconditions!");
1408 
1409   OriginalPreheader = Preheader;
1410   MainLoopPreheader = Preheader;
1411 
1412   bool IsSignedPredicate = MainLoopStructure.IsSignedPredicate;
1413   Optional<SubRanges> MaybeSR = calculateSubRanges(IsSignedPredicate);
1414   if (!MaybeSR.hasValue()) {
1415     LLVM_DEBUG(dbgs() << "irce: could not compute subranges\n");
1416     return false;
1417   }
1418 
1419   SubRanges SR = MaybeSR.getValue();
1420   bool Increasing = MainLoopStructure.IndVarIncreasing;
1421   IntegerType *IVTy =
1422       cast<IntegerType>(Range.getBegin()->getType());
1423 
1424   SCEVExpander Expander(SE, F.getParent()->getDataLayout(), "irce");
1425   Instruction *InsertPt = OriginalPreheader->getTerminator();
1426 
1427   // It would have been better to make `PreLoop' and `PostLoop'
1428   // `Optional<ClonedLoop>'s, but `ValueToValueMapTy' does not have a copy
1429   // constructor.
1430   ClonedLoop PreLoop, PostLoop;
1431   bool NeedsPreLoop =
1432       Increasing ? SR.LowLimit.hasValue() : SR.HighLimit.hasValue();
1433   bool NeedsPostLoop =
1434       Increasing ? SR.HighLimit.hasValue() : SR.LowLimit.hasValue();
1435 
1436   Value *ExitPreLoopAt = nullptr;
1437   Value *ExitMainLoopAt = nullptr;
1438   const SCEVConstant *MinusOneS =
1439       cast<SCEVConstant>(SE.getConstant(IVTy, -1, true /* isSigned */));
1440 
1441   if (NeedsPreLoop) {
1442     const SCEV *ExitPreLoopAtSCEV = nullptr;
1443 
1444     if (Increasing)
1445       ExitPreLoopAtSCEV = *SR.LowLimit;
1446     else if (cannotBeMinInLoop(*SR.HighLimit, &OriginalLoop, SE,
1447                                IsSignedPredicate))
1448       ExitPreLoopAtSCEV = SE.getAddExpr(*SR.HighLimit, MinusOneS);
1449     else {
1450       LLVM_DEBUG(dbgs() << "irce: could not prove no-overflow when computing "
1451                         << "preloop exit limit.  HighLimit = "
1452                         << *(*SR.HighLimit) << "\n");
1453       return false;
1454     }
1455 
1456     if (!isSafeToExpandAt(ExitPreLoopAtSCEV, InsertPt, SE)) {
1457       LLVM_DEBUG(dbgs() << "irce: could not prove that it is safe to expand the"
1458                         << " preloop exit limit " << *ExitPreLoopAtSCEV
1459                         << " at block " << InsertPt->getParent()->getName()
1460                         << "\n");
1461       return false;
1462     }
1463 
1464     ExitPreLoopAt = Expander.expandCodeFor(ExitPreLoopAtSCEV, IVTy, InsertPt);
1465     ExitPreLoopAt->setName("exit.preloop.at");
1466   }
1467 
1468   if (NeedsPostLoop) {
1469     const SCEV *ExitMainLoopAtSCEV = nullptr;
1470 
1471     if (Increasing)
1472       ExitMainLoopAtSCEV = *SR.HighLimit;
1473     else if (cannotBeMinInLoop(*SR.LowLimit, &OriginalLoop, SE,
1474                                IsSignedPredicate))
1475       ExitMainLoopAtSCEV = SE.getAddExpr(*SR.LowLimit, MinusOneS);
1476     else {
1477       LLVM_DEBUG(dbgs() << "irce: could not prove no-overflow when computing "
1478                         << "mainloop exit limit.  LowLimit = "
1479                         << *(*SR.LowLimit) << "\n");
1480       return false;
1481     }
1482 
1483     if (!isSafeToExpandAt(ExitMainLoopAtSCEV, InsertPt, SE)) {
1484       LLVM_DEBUG(dbgs() << "irce: could not prove that it is safe to expand the"
1485                         << " main loop exit limit " << *ExitMainLoopAtSCEV
1486                         << " at block " << InsertPt->getParent()->getName()
1487                         << "\n");
1488       return false;
1489     }
1490 
1491     ExitMainLoopAt = Expander.expandCodeFor(ExitMainLoopAtSCEV, IVTy, InsertPt);
1492     ExitMainLoopAt->setName("exit.mainloop.at");
1493   }
1494 
1495   // We clone these ahead of time so that we don't have to deal with changing
1496   // and temporarily invalid IR as we transform the loops.
1497   if (NeedsPreLoop)
1498     cloneLoop(PreLoop, "preloop");
1499   if (NeedsPostLoop)
1500     cloneLoop(PostLoop, "postloop");
1501 
1502   RewrittenRangeInfo PreLoopRRI;
1503 
1504   if (NeedsPreLoop) {
1505     Preheader->getTerminator()->replaceUsesOfWith(MainLoopStructure.Header,
1506                                                   PreLoop.Structure.Header);
1507 
1508     MainLoopPreheader =
1509         createPreheader(MainLoopStructure, Preheader, "mainloop");
1510     PreLoopRRI = changeIterationSpaceEnd(PreLoop.Structure, Preheader,
1511                                          ExitPreLoopAt, MainLoopPreheader);
1512     rewriteIncomingValuesForPHIs(MainLoopStructure, MainLoopPreheader,
1513                                  PreLoopRRI);
1514   }
1515 
1516   BasicBlock *PostLoopPreheader = nullptr;
1517   RewrittenRangeInfo PostLoopRRI;
1518 
1519   if (NeedsPostLoop) {
1520     PostLoopPreheader =
1521         createPreheader(PostLoop.Structure, Preheader, "postloop");
1522     PostLoopRRI = changeIterationSpaceEnd(MainLoopStructure, MainLoopPreheader,
1523                                           ExitMainLoopAt, PostLoopPreheader);
1524     rewriteIncomingValuesForPHIs(PostLoop.Structure, PostLoopPreheader,
1525                                  PostLoopRRI);
1526   }
1527 
1528   BasicBlock *NewMainLoopPreheader =
1529       MainLoopPreheader != Preheader ? MainLoopPreheader : nullptr;
1530   BasicBlock *NewBlocks[] = {PostLoopPreheader,        PreLoopRRI.PseudoExit,
1531                              PreLoopRRI.ExitSelector,  PostLoopRRI.PseudoExit,
1532                              PostLoopRRI.ExitSelector, NewMainLoopPreheader};
1533 
1534   // Some of the above may be nullptr, filter them out before passing to
1535   // addToParentLoopIfNeeded.
1536   auto NewBlocksEnd =
1537       std::remove(std::begin(NewBlocks), std::end(NewBlocks), nullptr);
1538 
1539   addToParentLoopIfNeeded(makeArrayRef(std::begin(NewBlocks), NewBlocksEnd));
1540 
1541   DT.recalculate(F);
1542 
1543   // We need to first add all the pre and post loop blocks into the loop
1544   // structures (as part of createClonedLoopStructure), and then update the
1545   // LCSSA form and LoopSimplifyForm. This is necessary for correctly updating
1546   // LI when LoopSimplifyForm is generated.
1547   Loop *PreL = nullptr, *PostL = nullptr;
1548   if (!PreLoop.Blocks.empty()) {
1549     PreL = createClonedLoopStructure(&OriginalLoop,
1550                                      OriginalLoop.getParentLoop(), PreLoop.Map,
1551                                      /* IsSubLoop */ false);
1552   }
1553 
1554   if (!PostLoop.Blocks.empty()) {
1555     PostL =
1556         createClonedLoopStructure(&OriginalLoop, OriginalLoop.getParentLoop(),
1557                                   PostLoop.Map, /* IsSubLoop */ false);
1558   }
1559 
1560   // This function canonicalizes the loop into Loop-Simplify and LCSSA forms.
1561   auto CanonicalizeLoop = [&] (Loop *L, bool IsOriginalLoop) {
1562     formLCSSARecursively(*L, DT, &LI, &SE);
1563     simplifyLoop(L, &DT, &LI, &SE, nullptr, nullptr, true);
1564     // Pre/post loops are slow paths, we do not need to perform any loop
1565     // optimizations on them.
1566     if (!IsOriginalLoop)
1567       DisableAllLoopOptsOnLoop(*L);
1568   };
1569   if (PreL)
1570     CanonicalizeLoop(PreL, false);
1571   if (PostL)
1572     CanonicalizeLoop(PostL, false);
1573   CanonicalizeLoop(&OriginalLoop, true);
1574 
1575   return true;
1576 }
1577 
1578 /// Computes and returns a range of values for the induction variable (IndVar)
1579 /// in which the range check can be safely elided.  If it cannot compute such a
1580 /// range, returns None.
1581 Optional<InductiveRangeCheck::Range>
1582 InductiveRangeCheck::computeSafeIterationSpace(
1583     ScalarEvolution &SE, const SCEVAddRecExpr *IndVar,
1584     bool IsLatchSigned) const {
1585   // We can deal when types of latch check and range checks don't match in case
1586   // if latch check is more narrow.
1587   auto *IVType = cast<IntegerType>(IndVar->getType());
1588   auto *RCType = cast<IntegerType>(getBegin()->getType());
1589   if (IVType->getBitWidth() > RCType->getBitWidth())
1590     return None;
1591   // IndVar is of the form "A + B * I" (where "I" is the canonical induction
1592   // variable, that may or may not exist as a real llvm::Value in the loop) and
1593   // this inductive range check is a range check on the "C + D * I" ("C" is
1594   // getBegin() and "D" is getStep()).  We rewrite the value being range
1595   // checked to "M + N * IndVar" where "N" = "D * B^(-1)" and "M" = "C - NA".
1596   //
1597   // The actual inequalities we solve are of the form
1598   //
1599   //   0 <= M + 1 * IndVar < L given L >= 0  (i.e. N == 1)
1600   //
1601   // Here L stands for upper limit of the safe iteration space.
1602   // The inequality is satisfied by (0 - M) <= IndVar < (L - M). To avoid
1603   // overflows when calculating (0 - M) and (L - M) we, depending on type of
1604   // IV's iteration space, limit the calculations by borders of the iteration
1605   // space. For example, if IndVar is unsigned, (0 - M) overflows for any M > 0.
1606   // If we figured out that "anything greater than (-M) is safe", we strengthen
1607   // this to "everything greater than 0 is safe", assuming that values between
1608   // -M and 0 just do not exist in unsigned iteration space, and we don't want
1609   // to deal with overflown values.
1610 
1611   if (!IndVar->isAffine())
1612     return None;
1613 
1614   const SCEV *A = NoopOrExtend(IndVar->getStart(), RCType, SE, IsLatchSigned);
1615   const SCEVConstant *B = dyn_cast<SCEVConstant>(
1616       NoopOrExtend(IndVar->getStepRecurrence(SE), RCType, SE, IsLatchSigned));
1617   if (!B)
1618     return None;
1619   assert(!B->isZero() && "Recurrence with zero step?");
1620 
1621   const SCEV *C = getBegin();
1622   const SCEVConstant *D = dyn_cast<SCEVConstant>(getStep());
1623   if (D != B)
1624     return None;
1625 
1626   assert(!D->getValue()->isZero() && "Recurrence with zero step?");
1627   unsigned BitWidth = RCType->getBitWidth();
1628   const SCEV *SIntMax = SE.getConstant(APInt::getSignedMaxValue(BitWidth));
1629 
1630   // Subtract Y from X so that it does not go through border of the IV
1631   // iteration space. Mathematically, it is equivalent to:
1632   //
1633   //    ClampedSubtract(X, Y) = min(max(X - Y, INT_MIN), INT_MAX).        [1]
1634   //
1635   // In [1], 'X - Y' is a mathematical subtraction (result is not bounded to
1636   // any width of bit grid). But after we take min/max, the result is
1637   // guaranteed to be within [INT_MIN, INT_MAX].
1638   //
1639   // In [1], INT_MAX and INT_MIN are respectively signed and unsigned max/min
1640   // values, depending on type of latch condition that defines IV iteration
1641   // space.
1642   auto ClampedSubtract = [&](const SCEV *X, const SCEV *Y) {
1643     // FIXME: The current implementation assumes that X is in [0, SINT_MAX].
1644     // This is required to ensure that SINT_MAX - X does not overflow signed and
1645     // that X - Y does not overflow unsigned if Y is negative. Can we lift this
1646     // restriction and make it work for negative X either?
1647     if (IsLatchSigned) {
1648       // X is a number from signed range, Y is interpreted as signed.
1649       // Even if Y is SINT_MAX, (X - Y) does not reach SINT_MIN. So the only
1650       // thing we should care about is that we didn't cross SINT_MAX.
1651       // So, if Y is positive, we subtract Y safely.
1652       //   Rule 1: Y > 0 ---> Y.
1653       // If 0 <= -Y <= (SINT_MAX - X), we subtract Y safely.
1654       //   Rule 2: Y >=s (X - SINT_MAX) ---> Y.
1655       // If 0 <= (SINT_MAX - X) < -Y, we can only subtract (X - SINT_MAX).
1656       //   Rule 3: Y <s (X - SINT_MAX) ---> (X - SINT_MAX).
1657       // It gives us smax(Y, X - SINT_MAX) to subtract in all cases.
1658       const SCEV *XMinusSIntMax = SE.getMinusSCEV(X, SIntMax);
1659       return SE.getMinusSCEV(X, SE.getSMaxExpr(Y, XMinusSIntMax),
1660                              SCEV::FlagNSW);
1661     } else
1662       // X is a number from unsigned range, Y is interpreted as signed.
1663       // Even if Y is SINT_MIN, (X - Y) does not reach UINT_MAX. So the only
1664       // thing we should care about is that we didn't cross zero.
1665       // So, if Y is negative, we subtract Y safely.
1666       //   Rule 1: Y <s 0 ---> Y.
1667       // If 0 <= Y <= X, we subtract Y safely.
1668       //   Rule 2: Y <=s X ---> Y.
1669       // If 0 <= X < Y, we should stop at 0 and can only subtract X.
1670       //   Rule 3: Y >s X ---> X.
1671       // It gives us smin(X, Y) to subtract in all cases.
1672       return SE.getMinusSCEV(X, SE.getSMinExpr(X, Y), SCEV::FlagNUW);
1673   };
1674   const SCEV *M = SE.getMinusSCEV(C, A);
1675   const SCEV *Zero = SE.getZero(M->getType());
1676 
1677   // This function returns SCEV equal to 1 if X is non-negative 0 otherwise.
1678   auto SCEVCheckNonNegative = [&](const SCEV *X) {
1679     const Loop *L = IndVar->getLoop();
1680     const SCEV *One = SE.getOne(X->getType());
1681     // Can we trivially prove that X is a non-negative or negative value?
1682     if (isKnownNonNegativeInLoop(X, L, SE))
1683       return One;
1684     else if (isKnownNegativeInLoop(X, L, SE))
1685       return Zero;
1686     // If not, we will have to figure it out during the execution.
1687     // Function smax(smin(X, 0), -1) + 1 equals to 1 if X >= 0 and 0 if X < 0.
1688     const SCEV *NegOne = SE.getNegativeSCEV(One);
1689     return SE.getAddExpr(SE.getSMaxExpr(SE.getSMinExpr(X, Zero), NegOne), One);
1690   };
1691   // FIXME: Current implementation of ClampedSubtract implicitly assumes that
1692   // X is non-negative (in sense of a signed value). We need to re-implement
1693   // this function in a way that it will correctly handle negative X as well.
1694   // We use it twice: for X = 0 everything is fine, but for X = getEnd() we can
1695   // end up with a negative X and produce wrong results. So currently we ensure
1696   // that if getEnd() is negative then both ends of the safe range are zero.
1697   // Note that this may pessimize elimination of unsigned range checks against
1698   // negative values.
1699   const SCEV *REnd = getEnd();
1700   const SCEV *EndIsNonNegative = SCEVCheckNonNegative(REnd);
1701 
1702   const SCEV *Begin = SE.getMulExpr(ClampedSubtract(Zero, M), EndIsNonNegative);
1703   const SCEV *End = SE.getMulExpr(ClampedSubtract(REnd, M), EndIsNonNegative);
1704   return InductiveRangeCheck::Range(Begin, End);
1705 }
1706 
1707 static Optional<InductiveRangeCheck::Range>
1708 IntersectSignedRange(ScalarEvolution &SE,
1709                      const Optional<InductiveRangeCheck::Range> &R1,
1710                      const InductiveRangeCheck::Range &R2) {
1711   if (R2.isEmpty(SE, /* IsSigned */ true))
1712     return None;
1713   if (!R1.hasValue())
1714     return R2;
1715   auto &R1Value = R1.getValue();
1716   // We never return empty ranges from this function, and R1 is supposed to be
1717   // a result of intersection. Thus, R1 is never empty.
1718   assert(!R1Value.isEmpty(SE, /* IsSigned */ true) &&
1719          "We should never have empty R1!");
1720 
1721   // TODO: we could widen the smaller range and have this work; but for now we
1722   // bail out to keep things simple.
1723   if (R1Value.getType() != R2.getType())
1724     return None;
1725 
1726   const SCEV *NewBegin = SE.getSMaxExpr(R1Value.getBegin(), R2.getBegin());
1727   const SCEV *NewEnd = SE.getSMinExpr(R1Value.getEnd(), R2.getEnd());
1728 
1729   // If the resulting range is empty, just return None.
1730   auto Ret = InductiveRangeCheck::Range(NewBegin, NewEnd);
1731   if (Ret.isEmpty(SE, /* IsSigned */ true))
1732     return None;
1733   return Ret;
1734 }
1735 
1736 static Optional<InductiveRangeCheck::Range>
1737 IntersectUnsignedRange(ScalarEvolution &SE,
1738                        const Optional<InductiveRangeCheck::Range> &R1,
1739                        const InductiveRangeCheck::Range &R2) {
1740   if (R2.isEmpty(SE, /* IsSigned */ false))
1741     return None;
1742   if (!R1.hasValue())
1743     return R2;
1744   auto &R1Value = R1.getValue();
1745   // We never return empty ranges from this function, and R1 is supposed to be
1746   // a result of intersection. Thus, R1 is never empty.
1747   assert(!R1Value.isEmpty(SE, /* IsSigned */ false) &&
1748          "We should never have empty R1!");
1749 
1750   // TODO: we could widen the smaller range and have this work; but for now we
1751   // bail out to keep things simple.
1752   if (R1Value.getType() != R2.getType())
1753     return None;
1754 
1755   const SCEV *NewBegin = SE.getUMaxExpr(R1Value.getBegin(), R2.getBegin());
1756   const SCEV *NewEnd = SE.getUMinExpr(R1Value.getEnd(), R2.getEnd());
1757 
1758   // If the resulting range is empty, just return None.
1759   auto Ret = InductiveRangeCheck::Range(NewBegin, NewEnd);
1760   if (Ret.isEmpty(SE, /* IsSigned */ false))
1761     return None;
1762   return Ret;
1763 }
1764 
1765 PreservedAnalyses IRCEPass::run(Function &F, FunctionAnalysisManager &AM) {
1766   auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F);
1767   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
1768   auto &BPI = AM.getResult<BranchProbabilityAnalysis>(F);
1769   LoopInfo &LI = AM.getResult<LoopAnalysis>(F);
1770 
1771   // Get BFI analysis result on demand. Please note that modification of
1772   // CFG invalidates this analysis and we should handle it.
1773   auto getBFI = [&F, &AM ]()->BlockFrequencyInfo & {
1774     return AM.getResult<BlockFrequencyAnalysis>(F);
1775   };
1776   InductiveRangeCheckElimination IRCE(SE, &BPI, DT, LI, { getBFI });
1777 
1778   bool Changed = false;
1779   {
1780     bool CFGChanged = false;
1781     for (const auto &L : LI) {
1782       CFGChanged |= simplifyLoop(L, &DT, &LI, &SE, nullptr, nullptr,
1783                                  /*PreserveLCSSA=*/false);
1784       Changed |= formLCSSARecursively(*L, DT, &LI, &SE);
1785     }
1786     Changed |= CFGChanged;
1787 
1788     if (CFGChanged && !SkipProfitabilityChecks)
1789       AM.invalidate<BlockFrequencyAnalysis>(F);
1790   }
1791 
1792   SmallPriorityWorklist<Loop *, 4> Worklist;
1793   appendLoopsToWorklist(LI, Worklist);
1794   auto LPMAddNewLoop = [&Worklist](Loop *NL, bool IsSubloop) {
1795     if (!IsSubloop)
1796       appendLoopsToWorklist(*NL, Worklist);
1797   };
1798 
1799   while (!Worklist.empty()) {
1800     Loop *L = Worklist.pop_back_val();
1801     if (IRCE.run(L, LPMAddNewLoop)) {
1802       Changed = true;
1803       if (!SkipProfitabilityChecks)
1804         AM.invalidate<BlockFrequencyAnalysis>(F);
1805     }
1806   }
1807 
1808   if (!Changed)
1809     return PreservedAnalyses::all();
1810   return getLoopPassPreservedAnalyses();
1811 }
1812 
1813 bool IRCELegacyPass::runOnFunction(Function &F) {
1814   if (skipFunction(F))
1815     return false;
1816 
1817   ScalarEvolution &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
1818   BranchProbabilityInfo &BPI =
1819       getAnalysis<BranchProbabilityInfoWrapperPass>().getBPI();
1820   auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1821   auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1822   InductiveRangeCheckElimination IRCE(SE, &BPI, DT, LI);
1823 
1824   bool Changed = false;
1825 
1826   for (const auto &L : LI) {
1827     Changed |= simplifyLoop(L, &DT, &LI, &SE, nullptr, nullptr,
1828                             /*PreserveLCSSA=*/false);
1829     Changed |= formLCSSARecursively(*L, DT, &LI, &SE);
1830   }
1831 
1832   SmallPriorityWorklist<Loop *, 4> Worklist;
1833   appendLoopsToWorklist(LI, Worklist);
1834   auto LPMAddNewLoop = [&](Loop *NL, bool IsSubloop) {
1835     if (!IsSubloop)
1836       appendLoopsToWorklist(*NL, Worklist);
1837   };
1838 
1839   while (!Worklist.empty()) {
1840     Loop *L = Worklist.pop_back_val();
1841     Changed |= IRCE.run(L, LPMAddNewLoop);
1842   }
1843   return Changed;
1844 }
1845 
1846 bool
1847 InductiveRangeCheckElimination::isProfitableToTransform(const Loop &L,
1848                                                         LoopStructure &LS) {
1849   if (SkipProfitabilityChecks)
1850     return true;
1851   if (GetBFI.hasValue()) {
1852     BlockFrequencyInfo &BFI = (*GetBFI)();
1853     uint64_t hFreq = BFI.getBlockFreq(LS.Header).getFrequency();
1854     uint64_t phFreq = BFI.getBlockFreq(L.getLoopPreheader()).getFrequency();
1855     if (phFreq != 0 && hFreq != 0 && (hFreq / phFreq < MinRuntimeIterations)) {
1856       LLVM_DEBUG(dbgs() << "irce: could not prove profitability: "
1857                         << "the estimated number of iterations basing on "
1858                            "frequency info is " << (hFreq / phFreq) << "\n";);
1859       return false;
1860     }
1861     return true;
1862   }
1863 
1864   if (!BPI)
1865     return true;
1866   BranchProbability ExitProbability =
1867       BPI->getEdgeProbability(LS.Latch, LS.LatchBrExitIdx);
1868   if (ExitProbability > BranchProbability(1, MinRuntimeIterations)) {
1869     LLVM_DEBUG(dbgs() << "irce: could not prove profitability: "
1870                       << "the exit probability is too big " << ExitProbability
1871                       << "\n";);
1872     return false;
1873   }
1874   return true;
1875 }
1876 
1877 bool InductiveRangeCheckElimination::run(
1878     Loop *L, function_ref<void(Loop *, bool)> LPMAddNewLoop) {
1879   if (L->getBlocks().size() >= LoopSizeCutoff) {
1880     LLVM_DEBUG(dbgs() << "irce: giving up constraining loop, too large\n");
1881     return false;
1882   }
1883 
1884   BasicBlock *Preheader = L->getLoopPreheader();
1885   if (!Preheader) {
1886     LLVM_DEBUG(dbgs() << "irce: loop has no preheader, leaving\n");
1887     return false;
1888   }
1889 
1890   LLVMContext &Context = Preheader->getContext();
1891   SmallVector<InductiveRangeCheck, 16> RangeChecks;
1892 
1893   for (auto BBI : L->getBlocks())
1894     if (BranchInst *TBI = dyn_cast<BranchInst>(BBI->getTerminator()))
1895       InductiveRangeCheck::extractRangeChecksFromBranch(TBI, L, SE, BPI,
1896                                                         RangeChecks);
1897 
1898   if (RangeChecks.empty())
1899     return false;
1900 
1901   auto PrintRecognizedRangeChecks = [&](raw_ostream &OS) {
1902     OS << "irce: looking at loop "; L->print(OS);
1903     OS << "irce: loop has " << RangeChecks.size()
1904        << " inductive range checks: \n";
1905     for (InductiveRangeCheck &IRC : RangeChecks)
1906       IRC.print(OS);
1907   };
1908 
1909   LLVM_DEBUG(PrintRecognizedRangeChecks(dbgs()));
1910 
1911   if (PrintRangeChecks)
1912     PrintRecognizedRangeChecks(errs());
1913 
1914   const char *FailureReason = nullptr;
1915   Optional<LoopStructure> MaybeLoopStructure =
1916       LoopStructure::parseLoopStructure(SE, *L, FailureReason);
1917   if (!MaybeLoopStructure.hasValue()) {
1918     LLVM_DEBUG(dbgs() << "irce: could not parse loop structure: "
1919                       << FailureReason << "\n";);
1920     return false;
1921   }
1922   LoopStructure LS = MaybeLoopStructure.getValue();
1923   if (!isProfitableToTransform(*L, LS))
1924     return false;
1925   const SCEVAddRecExpr *IndVar =
1926       cast<SCEVAddRecExpr>(SE.getMinusSCEV(SE.getSCEV(LS.IndVarBase), SE.getSCEV(LS.IndVarStep)));
1927 
1928   Optional<InductiveRangeCheck::Range> SafeIterRange;
1929   Instruction *ExprInsertPt = Preheader->getTerminator();
1930 
1931   SmallVector<InductiveRangeCheck, 4> RangeChecksToEliminate;
1932   // Basing on the type of latch predicate, we interpret the IV iteration range
1933   // as signed or unsigned range. We use different min/max functions (signed or
1934   // unsigned) when intersecting this range with safe iteration ranges implied
1935   // by range checks.
1936   auto IntersectRange =
1937       LS.IsSignedPredicate ? IntersectSignedRange : IntersectUnsignedRange;
1938 
1939   IRBuilder<> B(ExprInsertPt);
1940   for (InductiveRangeCheck &IRC : RangeChecks) {
1941     auto Result = IRC.computeSafeIterationSpace(SE, IndVar,
1942                                                 LS.IsSignedPredicate);
1943     if (Result.hasValue()) {
1944       auto MaybeSafeIterRange =
1945           IntersectRange(SE, SafeIterRange, Result.getValue());
1946       if (MaybeSafeIterRange.hasValue()) {
1947         assert(
1948             !MaybeSafeIterRange.getValue().isEmpty(SE, LS.IsSignedPredicate) &&
1949             "We should never return empty ranges!");
1950         RangeChecksToEliminate.push_back(IRC);
1951         SafeIterRange = MaybeSafeIterRange.getValue();
1952       }
1953     }
1954   }
1955 
1956   if (!SafeIterRange.hasValue())
1957     return false;
1958 
1959   LoopConstrainer LC(*L, LI, LPMAddNewLoop, LS, SE, DT,
1960                      SafeIterRange.getValue());
1961   bool Changed = LC.run();
1962 
1963   if (Changed) {
1964     auto PrintConstrainedLoopInfo = [L]() {
1965       dbgs() << "irce: in function ";
1966       dbgs() << L->getHeader()->getParent()->getName() << ": ";
1967       dbgs() << "constrained ";
1968       L->print(dbgs());
1969     };
1970 
1971     LLVM_DEBUG(PrintConstrainedLoopInfo());
1972 
1973     if (PrintChangedLoops)
1974       PrintConstrainedLoopInfo();
1975 
1976     // Optimize away the now-redundant range checks.
1977 
1978     for (InductiveRangeCheck &IRC : RangeChecksToEliminate) {
1979       ConstantInt *FoldedRangeCheck = IRC.getPassingDirection()
1980                                           ? ConstantInt::getTrue(Context)
1981                                           : ConstantInt::getFalse(Context);
1982       IRC.getCheckUse()->set(FoldedRangeCheck);
1983     }
1984   }
1985 
1986   return Changed;
1987 }
1988 
1989 Pass *llvm::createInductiveRangeCheckEliminationPass() {
1990   return new IRCELegacyPass();
1991 }
1992