1//===- llvm/Analysis/ScalarEvolution.h - Scalar Evolution -------*- C++ -*-===//
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 ScalarEvolution class is an LLVM pass which can be used to analyze and
10// categorize scalar expressions in loops. It specializes in recognizing
11// general induction variables, representing them with the abstract and opaque
12// SCEV class. Given this analysis, trip counts of loops and other important
13// properties can be obtained.
14//
15// This analysis is primarily useful for induction variable substitution and
16// strength reduction.
17//
18//===----------------------------------------------------------------------===//
19
20#ifndef LLVM_ANALYSIS_SCALAREVOLUTION_H
21#define LLVM_ANALYSIS_SCALAREVOLUTION_H
22
23#include "llvm/ADT/APInt.h"
24#include "llvm/ADT/ArrayRef.h"
25#include "llvm/ADT/DenseMap.h"
26#include "llvm/ADT/DenseMapInfo.h"
27#include "llvm/ADT/FoldingSet.h"
28#include "llvm/ADT/PointerIntPair.h"
29#include "llvm/ADT/SetVector.h"
30#include "llvm/ADT/SmallPtrSet.h"
31#include "llvm/ADT/SmallVector.h"
32#include "llvm/IR/ConstantRange.h"
33#include "llvm/IR/InstrTypes.h"
34#include "llvm/IR/Instructions.h"
35#include "llvm/IR/PassManager.h"
36#include "llvm/IR/ValueHandle.h"
37#include "llvm/IR/ValueMap.h"
38#include "llvm/Pass.h"
39#include <cassert>
40#include <cstdint>
41#include <memory>
42#include <optional>
43#include <utility>
44
45namespace llvm {
46
47class OverflowingBinaryOperator;
48class AssumptionCache;
49class BasicBlock;
50class Constant;
51class ConstantInt;
52class DataLayout;
53class DominatorTree;
54class Function;
55class GEPOperator;
56class Instruction;
57class LLVMContext;
58class Loop;
59class LoopInfo;
60class raw_ostream;
61class ScalarEvolution;
62class SCEVAddRecExpr;
63class SCEVUnknown;
64class StructType;
65class TargetLibraryInfo;
66class Type;
67class Value;
68enum SCEVTypes : unsigned short;
69
70extern bool VerifySCEV;
71
72/// This class represents an analyzed expression in the program. These are
73/// opaque objects that the client is not allowed to do much with directly.
74///
75class SCEV : public FoldingSetNode {
76 friend struct FoldingSetTrait<SCEV>;
77
78 /// A reference to an Interned FoldingSetNodeID for this node. The
79 /// ScalarEvolution's BumpPtrAllocator holds the data.
80 FoldingSetNodeIDRef FastID;
81
82 // The SCEV baseclass this node corresponds to
83 const SCEVTypes SCEVType;
84
85protected:
86 // Estimated complexity of this node's expression tree size.
87 const unsigned short ExpressionSize;
88
89 /// This field is initialized to zero and may be used in subclasses to store
90 /// miscellaneous information.
91 unsigned short SubclassData = 0;
92
93public:
94 /// NoWrapFlags are bitfield indices into SubclassData.
95 ///
96 /// Add and Mul expressions may have no-unsigned-wrap <NUW> or
97 /// no-signed-wrap <NSW> properties, which are derived from the IR
98 /// operator. NSW is a misnomer that we use to mean no signed overflow or
99 /// underflow.
100 ///
101 /// AddRec expressions may have a no-self-wraparound <NW> property if, in
102 /// the integer domain, abs(step) * max-iteration(loop) <=
103 /// unsigned-max(bitwidth). This means that the recurrence will never reach
104 /// its start value if the step is non-zero. Computing the same value on
105 /// each iteration is not considered wrapping, and recurrences with step = 0
106 /// are trivially <NW>. <NW> is independent of the sign of step and the
107 /// value the add recurrence starts with.
108 ///
109 /// Note that NUW and NSW are also valid properties of a recurrence, and
110 /// either implies NW. For convenience, NW will be set for a recurrence
111 /// whenever either NUW or NSW are set.
112 ///
113 /// We require that the flag on a SCEV apply to the entire scope in which
114 /// that SCEV is defined. A SCEV's scope is set of locations dominated by
115 /// a defining location, which is in turn described by the following rules:
116 /// * A SCEVUnknown is at the point of definition of the Value.
117 /// * A SCEVConstant is defined at all points.
118 /// * A SCEVAddRec is defined starting with the header of the associated
119 /// loop.
120 /// * All other SCEVs are defined at the earlest point all operands are
121 /// defined.
122 ///
123 /// The above rules describe a maximally hoisted form (without regards to
124 /// potential control dependence). A SCEV is defined anywhere a
125 /// corresponding instruction could be defined in said maximally hoisted
126 /// form. Note that SCEVUDivExpr (currently the only expression type which
127 /// can trap) can be defined per these rules in regions where it would trap
128 /// at runtime. A SCEV being defined does not require the existence of any
129 /// instruction within the defined scope.
130 enum NoWrapFlags {
131 FlagAnyWrap = 0, // No guarantee.
132 FlagNW = (1 << 0), // No self-wrap.
133 FlagNUW = (1 << 1), // No unsigned wrap.
134 FlagNSW = (1 << 2), // No signed wrap.
135 NoWrapMask = (1 << 3) - 1
136 };
137
138 explicit SCEV(const FoldingSetNodeIDRef ID, SCEVTypes SCEVTy,
139 unsigned short ExpressionSize)
140 : FastID(ID), SCEVType(SCEVTy), ExpressionSize(ExpressionSize) {}
141 SCEV(const SCEV &) = delete;
142 SCEV &operator=(const SCEV &) = delete;
143
144 SCEVTypes getSCEVType() const { return SCEVType; }
145
146 /// Return the LLVM type of this SCEV expression.
147 Type *getType() const;
148
149 /// Return operands of this SCEV expression.
150 ArrayRef<const SCEV *> operands() const;
151
152 /// Return true if the expression is a constant zero.
153 bool isZero() const;
154
155 /// Return true if the expression is a constant one.
156 bool isOne() const;
157
158 /// Return true if the expression is a constant all-ones value.
159 bool isAllOnesValue() const;
160
161 /// Return true if the specified scev is negated, but not a constant.
162 bool isNonConstantNegative() const;
163
164 // Returns estimated size of the mathematical expression represented by this
165 // SCEV. The rules of its calculation are following:
166 // 1) Size of a SCEV without operands (like constants and SCEVUnknown) is 1;
167 // 2) Size SCEV with operands Op1, Op2, ..., OpN is calculated by formula:
168 // (1 + Size(Op1) + ... + Size(OpN)).
169 // This value gives us an estimation of time we need to traverse through this
170 // SCEV and all its operands recursively. We may use it to avoid performing
171 // heavy transformations on SCEVs of excessive size for sake of saving the
172 // compilation time.
173 unsigned short getExpressionSize() const {
174 return ExpressionSize;
175 }
176
177 /// Print out the internal representation of this scalar to the specified
178 /// stream. This should really only be used for debugging purposes.
179 void print(raw_ostream &OS) const;
180
181 /// This method is used for debugging.
182 void dump() const;
183};
184
185// Specialize FoldingSetTrait for SCEV to avoid needing to compute
186// temporary FoldingSetNodeID values.
187template <> struct FoldingSetTrait<SCEV> : DefaultFoldingSetTrait<SCEV> {
188 static void Profile(const SCEV &X, FoldingSetNodeID &ID) { ID = X.FastID; }
189
190 static bool Equals(const SCEV &X, const FoldingSetNodeID &ID, unsigned IDHash,
191 FoldingSetNodeID &TempID) {
192 return ID == X.FastID;
193 }
194
195 static unsigned ComputeHash(const SCEV &X, FoldingSetNodeID &TempID) {
196 return X.FastID.ComputeHash();
197 }
198};
199
200inline raw_ostream &operator<<(raw_ostream &OS, const SCEV &S) {
201 S.print(OS);
202 return OS;
203}
204
205/// An object of this class is returned by queries that could not be answered.
206/// For example, if you ask for the number of iterations of a linked-list
207/// traversal loop, you will get one of these. None of the standard SCEV
208/// operations are valid on this class, it is just a marker.
209struct SCEVCouldNotCompute : public SCEV {
210 SCEVCouldNotCompute();
211
212 /// Methods for support type inquiry through isa, cast, and dyn_cast:
213 static bool classof(const SCEV *S);
214};
215
216/// This class represents an assumption made using SCEV expressions which can
217/// be checked at run-time.
218class SCEVPredicate : public FoldingSetNode {
219 friend struct FoldingSetTrait<SCEVPredicate>;
220
221 /// A reference to an Interned FoldingSetNodeID for this node. The
222 /// ScalarEvolution's BumpPtrAllocator holds the data.
223 FoldingSetNodeIDRef FastID;
224
225public:
226 enum SCEVPredicateKind { P_Union, P_Compare, P_Wrap };
227
228protected:
229 SCEVPredicateKind Kind;
230 ~SCEVPredicate() = default;
231 SCEVPredicate(const SCEVPredicate &) = default;
232 SCEVPredicate &operator=(const SCEVPredicate &) = default;
233
234public:
235 SCEVPredicate(const FoldingSetNodeIDRef ID, SCEVPredicateKind Kind);
236
237 SCEVPredicateKind getKind() const { return Kind; }
238
239 /// Returns the estimated complexity of this predicate. This is roughly
240 /// measured in the number of run-time checks required.
241 virtual unsigned getComplexity() const { return 1; }
242
243 /// Returns true if the predicate is always true. This means that no
244 /// assumptions were made and nothing needs to be checked at run-time.
245 virtual bool isAlwaysTrue() const = 0;
246
247 /// Returns true if this predicate implies \p N.
248 virtual bool implies(const SCEVPredicate *N) const = 0;
249
250 /// Prints a textual representation of this predicate with an indentation of
251 /// \p Depth.
252 virtual void print(raw_ostream &OS, unsigned Depth = 0) const = 0;
253};
254
255inline raw_ostream &operator<<(raw_ostream &OS, const SCEVPredicate &P) {
256 P.print(OS);
257 return OS;
258}
259
260// Specialize FoldingSetTrait for SCEVPredicate to avoid needing to compute
261// temporary FoldingSetNodeID values.
262template <>
263struct FoldingSetTrait<SCEVPredicate> : DefaultFoldingSetTrait<SCEVPredicate> {
264 static void Profile(const SCEVPredicate &X, FoldingSetNodeID &ID) {
265 ID = X.FastID;
266 }
267
268 static bool Equals(const SCEVPredicate &X, const FoldingSetNodeID &ID,
269 unsigned IDHash, FoldingSetNodeID &TempID) {
270 return ID == X.FastID;
271 }
272
273 static unsigned ComputeHash(const SCEVPredicate &X,
274 FoldingSetNodeID &TempID) {
275 return X.FastID.ComputeHash();
276 }
277};
278
279/// This class represents an assumption that the expression LHS Pred RHS
280/// evaluates to true, and this can be checked at run-time.
281class SCEVComparePredicate final : public SCEVPredicate {
282 /// We assume that LHS Pred RHS is true.
283 const ICmpInst::Predicate Pred;
284 const SCEV *LHS;
285 const SCEV *RHS;
286
287public:
288 SCEVComparePredicate(const FoldingSetNodeIDRef ID,
289 const ICmpInst::Predicate Pred,
290 const SCEV *LHS, const SCEV *RHS);
291
292 /// Implementation of the SCEVPredicate interface
293 bool implies(const SCEVPredicate *N) const override;
294 void print(raw_ostream &OS, unsigned Depth = 0) const override;
295 bool isAlwaysTrue() const override;
296
297 ICmpInst::Predicate getPredicate() const { return Pred; }
298
299 /// Returns the left hand side of the predicate.
300 const SCEV *getLHS() const { return LHS; }
301
302 /// Returns the right hand side of the predicate.
303 const SCEV *getRHS() const { return RHS; }
304
305 /// Methods for support type inquiry through isa, cast, and dyn_cast:
306 static bool classof(const SCEVPredicate *P) {
307 return P->getKind() == P_Compare;
308 }
309};
310
311/// This class represents an assumption made on an AddRec expression. Given an
312/// affine AddRec expression {a,+,b}, we assume that it has the nssw or nusw
313/// flags (defined below) in the first X iterations of the loop, where X is a
314/// SCEV expression returned by getPredicatedBackedgeTakenCount).
315///
316/// Note that this does not imply that X is equal to the backedge taken
317/// count. This means that if we have a nusw predicate for i32 {0,+,1} with a
318/// predicated backedge taken count of X, we only guarantee that {0,+,1} has
319/// nusw in the first X iterations. {0,+,1} may still wrap in the loop if we
320/// have more than X iterations.
321class SCEVWrapPredicate final : public SCEVPredicate {
322public:
323 /// Similar to SCEV::NoWrapFlags, but with slightly different semantics
324 /// for FlagNUSW. The increment is considered to be signed, and a + b
325 /// (where b is the increment) is considered to wrap if:
326 /// zext(a + b) != zext(a) + sext(b)
327 ///
328 /// If Signed is a function that takes an n-bit tuple and maps to the
329 /// integer domain as the tuples value interpreted as twos complement,
330 /// and Unsigned a function that takes an n-bit tuple and maps to the
331 /// integer domain as the base two value of input tuple, then a + b
332 /// has IncrementNUSW iff:
333 ///
334 /// 0 <= Unsigned(a) + Signed(b) < 2^n
335 ///
336 /// The IncrementNSSW flag has identical semantics with SCEV::FlagNSW.
337 ///
338 /// Note that the IncrementNUSW flag is not commutative: if base + inc
339 /// has IncrementNUSW, then inc + base doesn't neccessarily have this
340 /// property. The reason for this is that this is used for sign/zero
341 /// extending affine AddRec SCEV expressions when a SCEVWrapPredicate is
342 /// assumed. A {base,+,inc} expression is already non-commutative with
343 /// regards to base and inc, since it is interpreted as:
344 /// (((base + inc) + inc) + inc) ...
345 enum IncrementWrapFlags {
346 IncrementAnyWrap = 0, // No guarantee.
347 IncrementNUSW = (1 << 0), // No unsigned with signed increment wrap.
348 IncrementNSSW = (1 << 1), // No signed with signed increment wrap
349 // (equivalent with SCEV::NSW)
350 IncrementNoWrapMask = (1 << 2) - 1
351 };
352
353 /// Convenient IncrementWrapFlags manipulation methods.
354 [[nodiscard]] static SCEVWrapPredicate::IncrementWrapFlags
355 clearFlags(SCEVWrapPredicate::IncrementWrapFlags Flags,
356 SCEVWrapPredicate::IncrementWrapFlags OffFlags) {
357 assert((Flags & IncrementNoWrapMask) == Flags && "Invalid flags value!");
358 assert((OffFlags & IncrementNoWrapMask) == OffFlags &&
359 "Invalid flags value!");
360 return (SCEVWrapPredicate::IncrementWrapFlags)(Flags & ~OffFlags);
361 }
362
363 [[nodiscard]] static SCEVWrapPredicate::IncrementWrapFlags
364 maskFlags(SCEVWrapPredicate::IncrementWrapFlags Flags, int Mask) {
365 assert((Flags & IncrementNoWrapMask) == Flags && "Invalid flags value!");
366 assert((Mask & IncrementNoWrapMask) == Mask && "Invalid mask value!");
367
368 return (SCEVWrapPredicate::IncrementWrapFlags)(Flags & Mask);
369 }
370
371 [[nodiscard]] static SCEVWrapPredicate::IncrementWrapFlags
372 setFlags(SCEVWrapPredicate::IncrementWrapFlags Flags,
373 SCEVWrapPredicate::IncrementWrapFlags OnFlags) {
374 assert((Flags & IncrementNoWrapMask) == Flags && "Invalid flags value!");
375 assert((OnFlags & IncrementNoWrapMask) == OnFlags &&
376 "Invalid flags value!");
377
378 return (SCEVWrapPredicate::IncrementWrapFlags)(Flags | OnFlags);
379 }
380
381 /// Returns the set of SCEVWrapPredicate no wrap flags implied by a
382 /// SCEVAddRecExpr.
383 [[nodiscard]] static SCEVWrapPredicate::IncrementWrapFlags
384 getImpliedFlags(const SCEVAddRecExpr *AR, ScalarEvolution &SE);
385
386private:
387 const SCEVAddRecExpr *AR;
388 IncrementWrapFlags Flags;
389
390public:
391 explicit SCEVWrapPredicate(const FoldingSetNodeIDRef ID,
392 const SCEVAddRecExpr *AR,
393 IncrementWrapFlags Flags);
394
395 /// Returns the set assumed no overflow flags.
396 IncrementWrapFlags getFlags() const { return Flags; }
397
398 /// Implementation of the SCEVPredicate interface
399 const SCEVAddRecExpr *getExpr() const;
400 bool implies(const SCEVPredicate *N) const override;
401 void print(raw_ostream &OS, unsigned Depth = 0) const override;
402 bool isAlwaysTrue() const override;
403
404 /// Methods for support type inquiry through isa, cast, and dyn_cast:
405 static bool classof(const SCEVPredicate *P) {
406 return P->getKind() == P_Wrap;
407 }
408};
409
410/// This class represents a composition of other SCEV predicates, and is the
411/// class that most clients will interact with. This is equivalent to a
412/// logical "AND" of all the predicates in the union.
413///
414/// NB! Unlike other SCEVPredicate sub-classes this class does not live in the
415/// ScalarEvolution::Preds folding set. This is why the \c add function is sound.
416class SCEVUnionPredicate final : public SCEVPredicate {
417private:
418 using PredicateMap =
419 DenseMap<const SCEV *, SmallVector<const SCEVPredicate *, 4>>;
420
421 /// Vector with references to all predicates in this union.
422 SmallVector<const SCEVPredicate *, 16> Preds;
423
424 /// Adds a predicate to this union.
425 void add(const SCEVPredicate *N);
426
427public:
428 SCEVUnionPredicate(ArrayRef<const SCEVPredicate *> Preds);
429
430 const SmallVectorImpl<const SCEVPredicate *> &getPredicates() const {
431 return Preds;
432 }
433
434 /// Implementation of the SCEVPredicate interface
435 bool isAlwaysTrue() const override;
436 bool implies(const SCEVPredicate *N) const override;
437 void print(raw_ostream &OS, unsigned Depth) const override;
438
439 /// We estimate the complexity of a union predicate as the size number of
440 /// predicates in the union.
441 unsigned getComplexity() const override { return Preds.size(); }
442
443 /// Methods for support type inquiry through isa, cast, and dyn_cast:
444 static bool classof(const SCEVPredicate *P) {
445 return P->getKind() == P_Union;
446 }
447};
448
449/// The main scalar evolution driver. Because client code (intentionally)
450/// can't do much with the SCEV objects directly, they must ask this class
451/// for services.
452class ScalarEvolution {
453 friend class ScalarEvolutionsTest;
454
455public:
456 /// An enum describing the relationship between a SCEV and a loop.
457 enum LoopDisposition {
458 LoopVariant, ///< The SCEV is loop-variant (unknown).
459 LoopInvariant, ///< The SCEV is loop-invariant.
460 LoopComputable ///< The SCEV varies predictably with the loop.
461 };
462
463 /// An enum describing the relationship between a SCEV and a basic block.
464 enum BlockDisposition {
465 DoesNotDominateBlock, ///< The SCEV does not dominate the block.
466 DominatesBlock, ///< The SCEV dominates the block.
467 ProperlyDominatesBlock ///< The SCEV properly dominates the block.
468 };
469
470 /// Convenient NoWrapFlags manipulation that hides enum casts and is
471 /// visible in the ScalarEvolution name space.
472 [[nodiscard]] static SCEV::NoWrapFlags maskFlags(SCEV::NoWrapFlags Flags,
473 int Mask) {
474 return (SCEV::NoWrapFlags)(Flags & Mask);
475 }
476 [[nodiscard]] static SCEV::NoWrapFlags setFlags(SCEV::NoWrapFlags Flags,
477 SCEV::NoWrapFlags OnFlags) {
478 return (SCEV::NoWrapFlags)(Flags | OnFlags);
479 }
480 [[nodiscard]] static SCEV::NoWrapFlags
481 clearFlags(SCEV::NoWrapFlags Flags, SCEV::NoWrapFlags OffFlags) {
482 return (SCEV::NoWrapFlags)(Flags & ~OffFlags);
483 }
484 [[nodiscard]] static bool hasFlags(SCEV::NoWrapFlags Flags,
485 SCEV::NoWrapFlags TestFlags) {
486 return TestFlags == maskFlags(Flags, Mask: TestFlags);
487 };
488
489 ScalarEvolution(Function &F, TargetLibraryInfo &TLI, AssumptionCache &AC,
490 DominatorTree &DT, LoopInfo &LI);
491 ScalarEvolution(ScalarEvolution &&Arg);
492 ~ScalarEvolution();
493
494 LLVMContext &getContext() const { return F.getContext(); }
495
496 /// Test if values of the given type are analyzable within the SCEV
497 /// framework. This primarily includes integer types, and it can optionally
498 /// include pointer types if the ScalarEvolution class has access to
499 /// target-specific information.
500 bool isSCEVable(Type *Ty) const;
501
502 /// Return the size in bits of the specified type, for which isSCEVable must
503 /// return true.
504 uint64_t getTypeSizeInBits(Type *Ty) const;
505
506 /// Return a type with the same bitwidth as the given type and which
507 /// represents how SCEV will treat the given type, for which isSCEVable must
508 /// return true. For pointer types, this is the pointer-sized integer type.
509 Type *getEffectiveSCEVType(Type *Ty) const;
510
511 // Returns a wider type among {Ty1, Ty2}.
512 Type *getWiderType(Type *Ty1, Type *Ty2) const;
513
514 /// Return true if there exists a point in the program at which both
515 /// A and B could be operands to the same instruction.
516 /// SCEV expressions are generally assumed to correspond to instructions
517 /// which could exists in IR. In general, this requires that there exists
518 /// a use point in the program where all operands dominate the use.
519 ///
520 /// Example:
521 /// loop {
522 /// if
523 /// loop { v1 = load @global1; }
524 /// else
525 /// loop { v2 = load @global2; }
526 /// }
527 /// No SCEV with operand V1, and v2 can exist in this program.
528 bool instructionCouldExistWithOperands(const SCEV *A, const SCEV *B);
529
530 /// Return true if the SCEV is a scAddRecExpr or it contains
531 /// scAddRecExpr. The result will be cached in HasRecMap.
532 bool containsAddRecurrence(const SCEV *S);
533
534 /// Is operation \p BinOp between \p LHS and \p RHS provably does not have
535 /// a signed/unsigned overflow (\p Signed)? If \p CtxI is specified, the
536 /// no-overflow fact should be true in the context of this instruction.
537 bool willNotOverflow(Instruction::BinaryOps BinOp, bool Signed,
538 const SCEV *LHS, const SCEV *RHS,
539 const Instruction *CtxI = nullptr);
540
541 /// Parse NSW/NUW flags from add/sub/mul IR binary operation \p Op into
542 /// SCEV no-wrap flags, and deduce flag[s] that aren't known yet.
543 /// Does not mutate the original instruction. Returns std::nullopt if it could
544 /// not deduce more precise flags than the instruction already has, otherwise
545 /// returns proven flags.
546 std::optional<SCEV::NoWrapFlags>
547 getStrengthenedNoWrapFlagsFromBinOp(const OverflowingBinaryOperator *OBO);
548
549 /// Notify this ScalarEvolution that \p User directly uses SCEVs in \p Ops.
550 void registerUser(const SCEV *User, ArrayRef<const SCEV *> Ops);
551
552 /// Return true if the SCEV expression contains an undef value.
553 bool containsUndefs(const SCEV *S) const;
554
555 /// Return true if the SCEV expression contains a Value that has been
556 /// optimised out and is now a nullptr.
557 bool containsErasedValue(const SCEV *S) const;
558
559 /// Return a SCEV expression for the full generality of the specified
560 /// expression.
561 const SCEV *getSCEV(Value *V);
562
563 /// Return an existing SCEV for V if there is one, otherwise return nullptr.
564 const SCEV *getExistingSCEV(Value *V);
565
566 const SCEV *getConstant(ConstantInt *V);
567 const SCEV *getConstant(const APInt &Val);
568 const SCEV *getConstant(Type *Ty, uint64_t V, bool isSigned = false);
569 const SCEV *getLosslessPtrToIntExpr(const SCEV *Op, unsigned Depth = 0);
570 const SCEV *getPtrToIntExpr(const SCEV *Op, Type *Ty);
571 const SCEV *getTruncateExpr(const SCEV *Op, Type *Ty, unsigned Depth = 0);
572 const SCEV *getVScale(Type *Ty);
573 const SCEV *getElementCount(Type *Ty, ElementCount EC);
574 const SCEV *getZeroExtendExpr(const SCEV *Op, Type *Ty, unsigned Depth = 0);
575 const SCEV *getZeroExtendExprImpl(const SCEV *Op, Type *Ty,
576 unsigned Depth = 0);
577 const SCEV *getSignExtendExpr(const SCEV *Op, Type *Ty, unsigned Depth = 0);
578 const SCEV *getSignExtendExprImpl(const SCEV *Op, Type *Ty,
579 unsigned Depth = 0);
580 const SCEV *getCastExpr(SCEVTypes Kind, const SCEV *Op, Type *Ty);
581 const SCEV *getAnyExtendExpr(const SCEV *Op, Type *Ty);
582 const SCEV *getAddExpr(SmallVectorImpl<const SCEV *> &Ops,
583 SCEV::NoWrapFlags Flags = SCEV::FlagAnyWrap,
584 unsigned Depth = 0);
585 const SCEV *getAddExpr(const SCEV *LHS, const SCEV *RHS,
586 SCEV::NoWrapFlags Flags = SCEV::FlagAnyWrap,
587 unsigned Depth = 0) {
588 SmallVector<const SCEV *, 2> Ops = {LHS, RHS};
589 return getAddExpr(Ops, Flags, Depth);
590 }
591 const SCEV *getAddExpr(const SCEV *Op0, const SCEV *Op1, const SCEV *Op2,
592 SCEV::NoWrapFlags Flags = SCEV::FlagAnyWrap,
593 unsigned Depth = 0) {
594 SmallVector<const SCEV *, 3> Ops = {Op0, Op1, Op2};
595 return getAddExpr(Ops, Flags, Depth);
596 }
597 const SCEV *getMulExpr(SmallVectorImpl<const SCEV *> &Ops,
598 SCEV::NoWrapFlags Flags = SCEV::FlagAnyWrap,
599 unsigned Depth = 0);
600 const SCEV *getMulExpr(const SCEV *LHS, const SCEV *RHS,
601 SCEV::NoWrapFlags Flags = SCEV::FlagAnyWrap,
602 unsigned Depth = 0) {
603 SmallVector<const SCEV *, 2> Ops = {LHS, RHS};
604 return getMulExpr(Ops, Flags, Depth);
605 }
606 const SCEV *getMulExpr(const SCEV *Op0, const SCEV *Op1, const SCEV *Op2,
607 SCEV::NoWrapFlags Flags = SCEV::FlagAnyWrap,
608 unsigned Depth = 0) {
609 SmallVector<const SCEV *, 3> Ops = {Op0, Op1, Op2};
610 return getMulExpr(Ops, Flags, Depth);
611 }
612 const SCEV *getUDivExpr(const SCEV *LHS, const SCEV *RHS);
613 const SCEV *getUDivExactExpr(const SCEV *LHS, const SCEV *RHS);
614 const SCEV *getURemExpr(const SCEV *LHS, const SCEV *RHS);
615 const SCEV *getAddRecExpr(const SCEV *Start, const SCEV *Step, const Loop *L,
616 SCEV::NoWrapFlags Flags);
617 const SCEV *getAddRecExpr(SmallVectorImpl<const SCEV *> &Operands,
618 const Loop *L, SCEV::NoWrapFlags Flags);
619 const SCEV *getAddRecExpr(const SmallVectorImpl<const SCEV *> &Operands,
620 const Loop *L, SCEV::NoWrapFlags Flags) {
621 SmallVector<const SCEV *, 4> NewOp(Operands.begin(), Operands.end());
622 return getAddRecExpr(Operands&: NewOp, L, Flags);
623 }
624
625 /// Checks if \p SymbolicPHI can be rewritten as an AddRecExpr under some
626 /// Predicates. If successful return these <AddRecExpr, Predicates>;
627 /// The function is intended to be called from PSCEV (the caller will decide
628 /// whether to actually add the predicates and carry out the rewrites).
629 std::optional<std::pair<const SCEV *, SmallVector<const SCEVPredicate *, 3>>>
630 createAddRecFromPHIWithCasts(const SCEVUnknown *SymbolicPHI);
631
632 /// Returns an expression for a GEP
633 ///
634 /// \p GEP The GEP. The indices contained in the GEP itself are ignored,
635 /// instead we use IndexExprs.
636 /// \p IndexExprs The expressions for the indices.
637 const SCEV *getGEPExpr(GEPOperator *GEP,
638 const SmallVectorImpl<const SCEV *> &IndexExprs);
639 const SCEV *getAbsExpr(const SCEV *Op, bool IsNSW);
640 const SCEV *getMinMaxExpr(SCEVTypes Kind,
641 SmallVectorImpl<const SCEV *> &Operands);
642 const SCEV *getSequentialMinMaxExpr(SCEVTypes Kind,
643 SmallVectorImpl<const SCEV *> &Operands);
644 const SCEV *getSMaxExpr(const SCEV *LHS, const SCEV *RHS);
645 const SCEV *getSMaxExpr(SmallVectorImpl<const SCEV *> &Operands);
646 const SCEV *getUMaxExpr(const SCEV *LHS, const SCEV *RHS);
647 const SCEV *getUMaxExpr(SmallVectorImpl<const SCEV *> &Operands);
648 const SCEV *getSMinExpr(const SCEV *LHS, const SCEV *RHS);
649 const SCEV *getSMinExpr(SmallVectorImpl<const SCEV *> &Operands);
650 const SCEV *getUMinExpr(const SCEV *LHS, const SCEV *RHS,
651 bool Sequential = false);
652 const SCEV *getUMinExpr(SmallVectorImpl<const SCEV *> &Operands,
653 bool Sequential = false);
654 const SCEV *getUnknown(Value *V);
655 const SCEV *getCouldNotCompute();
656
657 /// Return a SCEV for the constant 0 of a specific type.
658 const SCEV *getZero(Type *Ty) { return getConstant(Ty, V: 0); }
659
660 /// Return a SCEV for the constant 1 of a specific type.
661 const SCEV *getOne(Type *Ty) { return getConstant(Ty, V: 1); }
662
663 /// Return a SCEV for the constant \p Power of two.
664 const SCEV *getPowerOfTwo(Type *Ty, unsigned Power) {
665 assert(Power < getTypeSizeInBits(Ty) && "Power out of range");
666 return getConstant(Val: APInt::getOneBitSet(numBits: getTypeSizeInBits(Ty), BitNo: Power));
667 }
668
669 /// Return a SCEV for the constant -1 of a specific type.
670 const SCEV *getMinusOne(Type *Ty) {
671 return getConstant(Ty, V: -1, /*isSigned=*/isSigned: true);
672 }
673
674 /// Return an expression for a TypeSize.
675 const SCEV *getSizeOfExpr(Type *IntTy, TypeSize Size);
676
677 /// Return an expression for the alloc size of AllocTy that is type IntTy
678 const SCEV *getSizeOfExpr(Type *IntTy, Type *AllocTy);
679
680 /// Return an expression for the store size of StoreTy that is type IntTy
681 const SCEV *getStoreSizeOfExpr(Type *IntTy, Type *StoreTy);
682
683 /// Return an expression for offsetof on the given field with type IntTy
684 const SCEV *getOffsetOfExpr(Type *IntTy, StructType *STy, unsigned FieldNo);
685
686 /// Return the SCEV object corresponding to -V.
687 const SCEV *getNegativeSCEV(const SCEV *V,
688 SCEV::NoWrapFlags Flags = SCEV::FlagAnyWrap);
689
690 /// Return the SCEV object corresponding to ~V.
691 const SCEV *getNotSCEV(const SCEV *V);
692
693 /// Return LHS-RHS. Minus is represented in SCEV as A+B*-1.
694 ///
695 /// If the LHS and RHS are pointers which don't share a common base
696 /// (according to getPointerBase()), this returns a SCEVCouldNotCompute.
697 /// To compute the difference between two unrelated pointers, you can
698 /// explicitly convert the arguments using getPtrToIntExpr(), for pointer
699 /// types that support it.
700 const SCEV *getMinusSCEV(const SCEV *LHS, const SCEV *RHS,
701 SCEV::NoWrapFlags Flags = SCEV::FlagAnyWrap,
702 unsigned Depth = 0);
703
704 /// Compute ceil(N / D). N and D are treated as unsigned values.
705 ///
706 /// Since SCEV doesn't have native ceiling division, this generates a
707 /// SCEV expression of the following form:
708 ///
709 /// umin(N, 1) + floor((N - umin(N, 1)) / D)
710 ///
711 /// A denominator of zero or poison is handled the same way as getUDivExpr().
712 const SCEV *getUDivCeilSCEV(const SCEV *N, const SCEV *D);
713
714 /// Return a SCEV corresponding to a conversion of the input value to the
715 /// specified type. If the type must be extended, it is zero extended.
716 const SCEV *getTruncateOrZeroExtend(const SCEV *V, Type *Ty,
717 unsigned Depth = 0);
718
719 /// Return a SCEV corresponding to a conversion of the input value to the
720 /// specified type. If the type must be extended, it is sign extended.
721 const SCEV *getTruncateOrSignExtend(const SCEV *V, Type *Ty,
722 unsigned Depth = 0);
723
724 /// Return a SCEV corresponding to a conversion of the input value to the
725 /// specified type. If the type must be extended, it is zero extended. The
726 /// conversion must not be narrowing.
727 const SCEV *getNoopOrZeroExtend(const SCEV *V, Type *Ty);
728
729 /// Return a SCEV corresponding to a conversion of the input value to the
730 /// specified type. If the type must be extended, it is sign extended. The
731 /// conversion must not be narrowing.
732 const SCEV *getNoopOrSignExtend(const SCEV *V, Type *Ty);
733
734 /// Return a SCEV corresponding to a conversion of the input value to the
735 /// specified type. If the type must be extended, it is extended with
736 /// unspecified bits. The conversion must not be narrowing.
737 const SCEV *getNoopOrAnyExtend(const SCEV *V, Type *Ty);
738
739 /// Return a SCEV corresponding to a conversion of the input value to the
740 /// specified type. The conversion must not be widening.
741 const SCEV *getTruncateOrNoop(const SCEV *V, Type *Ty);
742
743 /// Promote the operands to the wider of the types using zero-extension, and
744 /// then perform a umax operation with them.
745 const SCEV *getUMaxFromMismatchedTypes(const SCEV *LHS, const SCEV *RHS);
746
747 /// Promote the operands to the wider of the types using zero-extension, and
748 /// then perform a umin operation with them.
749 const SCEV *getUMinFromMismatchedTypes(const SCEV *LHS, const SCEV *RHS,
750 bool Sequential = false);
751
752 /// Promote the operands to the wider of the types using zero-extension, and
753 /// then perform a umin operation with them. N-ary function.
754 const SCEV *getUMinFromMismatchedTypes(SmallVectorImpl<const SCEV *> &Ops,
755 bool Sequential = false);
756
757 /// Transitively follow the chain of pointer-type operands until reaching a
758 /// SCEV that does not have a single pointer operand. This returns a
759 /// SCEVUnknown pointer for well-formed pointer-type expressions, but corner
760 /// cases do exist.
761 const SCEV *getPointerBase(const SCEV *V);
762
763 /// Compute an expression equivalent to S - getPointerBase(S).
764 const SCEV *removePointerBase(const SCEV *S);
765
766 /// Return a SCEV expression for the specified value at the specified scope
767 /// in the program. The L value specifies a loop nest to evaluate the
768 /// expression at, where null is the top-level or a specified loop is
769 /// immediately inside of the loop.
770 ///
771 /// This method can be used to compute the exit value for a variable defined
772 /// in a loop by querying what the value will hold in the parent loop.
773 ///
774 /// In the case that a relevant loop exit value cannot be computed, the
775 /// original value V is returned.
776 const SCEV *getSCEVAtScope(const SCEV *S, const Loop *L);
777
778 /// This is a convenience function which does getSCEVAtScope(getSCEV(V), L).
779 const SCEV *getSCEVAtScope(Value *V, const Loop *L);
780
781 /// Test whether entry to the loop is protected by a conditional between LHS
782 /// and RHS. This is used to help avoid max expressions in loop trip
783 /// counts, and to eliminate casts.
784 bool isLoopEntryGuardedByCond(const Loop *L, ICmpInst::Predicate Pred,
785 const SCEV *LHS, const SCEV *RHS);
786
787 /// Test whether entry to the basic block is protected by a conditional
788 /// between LHS and RHS.
789 bool isBasicBlockEntryGuardedByCond(const BasicBlock *BB,
790 ICmpInst::Predicate Pred, const SCEV *LHS,
791 const SCEV *RHS);
792
793 /// Test whether the backedge of the loop is protected by a conditional
794 /// between LHS and RHS. This is used to eliminate casts.
795 bool isLoopBackedgeGuardedByCond(const Loop *L, ICmpInst::Predicate Pred,
796 const SCEV *LHS, const SCEV *RHS);
797
798 /// A version of getTripCountFromExitCount below which always picks an
799 /// evaluation type which can not result in overflow.
800 const SCEV *getTripCountFromExitCount(const SCEV *ExitCount);
801
802 /// Convert from an "exit count" (i.e. "backedge taken count") to a "trip
803 /// count". A "trip count" is the number of times the header of the loop
804 /// will execute if an exit is taken after the specified number of backedges
805 /// have been taken. (e.g. TripCount = ExitCount + 1). Note that the
806 /// expression can overflow if ExitCount = UINT_MAX. If EvalTy is not wide
807 /// enough to hold the result without overflow, result unsigned wraps with
808 /// 2s-complement semantics. ex: EC = 255 (i8), TC = 0 (i8)
809 const SCEV *getTripCountFromExitCount(const SCEV *ExitCount, Type *EvalTy,
810 const Loop *L);
811
812 /// Returns the exact trip count of the loop if we can compute it, and
813 /// the result is a small constant. '0' is used to represent an unknown
814 /// or non-constant trip count. Note that a trip count is simply one more
815 /// than the backedge taken count for the loop.
816 unsigned getSmallConstantTripCount(const Loop *L);
817
818 /// Return the exact trip count for this loop if we exit through ExitingBlock.
819 /// '0' is used to represent an unknown or non-constant trip count. Note
820 /// that a trip count is simply one more than the backedge taken count for
821 /// the same exit.
822 /// This "trip count" assumes that control exits via ExitingBlock. More
823 /// precisely, it is the number of times that control will reach ExitingBlock
824 /// before taking the branch. For loops with multiple exits, it may not be
825 /// the number times that the loop header executes if the loop exits
826 /// prematurely via another branch.
827 unsigned getSmallConstantTripCount(const Loop *L,
828 const BasicBlock *ExitingBlock);
829
830 /// Returns the upper bound of the loop trip count as a normal unsigned
831 /// value.
832 /// Returns 0 if the trip count is unknown or not constant.
833 unsigned getSmallConstantMaxTripCount(const Loop *L);
834
835 /// Returns the largest constant divisor of the trip count as a normal
836 /// unsigned value, if possible. This means that the actual trip count is
837 /// always a multiple of the returned value. Returns 1 if the trip count is
838 /// unknown or not guaranteed to be the multiple of a constant., Will also
839 /// return 1 if the trip count is very large (>= 2^32).
840 /// Note that the argument is an exit count for loop L, NOT a trip count.
841 unsigned getSmallConstantTripMultiple(const Loop *L,
842 const SCEV *ExitCount);
843
844 /// Returns the largest constant divisor of the trip count of the
845 /// loop. Will return 1 if no trip count could be computed, or if a
846 /// divisor could not be found.
847 unsigned getSmallConstantTripMultiple(const Loop *L);
848
849 /// Returns the largest constant divisor of the trip count of this loop as a
850 /// normal unsigned value, if possible. This means that the actual trip
851 /// count is always a multiple of the returned value (don't forget the trip
852 /// count could very well be zero as well!). As explained in the comments
853 /// for getSmallConstantTripCount, this assumes that control exits the loop
854 /// via ExitingBlock.
855 unsigned getSmallConstantTripMultiple(const Loop *L,
856 const BasicBlock *ExitingBlock);
857
858 /// The terms "backedge taken count" and "exit count" are used
859 /// interchangeably to refer to the number of times the backedge of a loop
860 /// has executed before the loop is exited.
861 enum ExitCountKind {
862 /// An expression exactly describing the number of times the backedge has
863 /// executed when a loop is exited.
864 Exact,
865 /// A constant which provides an upper bound on the exact trip count.
866 ConstantMaximum,
867 /// An expression which provides an upper bound on the exact trip count.
868 SymbolicMaximum,
869 };
870
871 /// Return the number of times the backedge executes before the given exit
872 /// would be taken; if not exactly computable, return SCEVCouldNotCompute.
873 /// For a single exit loop, this value is equivelent to the result of
874 /// getBackedgeTakenCount. The loop is guaranteed to exit (via *some* exit)
875 /// before the backedge is executed (ExitCount + 1) times. Note that there
876 /// is no guarantee about *which* exit is taken on the exiting iteration.
877 const SCEV *getExitCount(const Loop *L, const BasicBlock *ExitingBlock,
878 ExitCountKind Kind = Exact);
879
880 /// If the specified loop has a predictable backedge-taken count, return it,
881 /// otherwise return a SCEVCouldNotCompute object. The backedge-taken count is
882 /// the number of times the loop header will be branched to from within the
883 /// loop, assuming there are no abnormal exists like exception throws. This is
884 /// one less than the trip count of the loop, since it doesn't count the first
885 /// iteration, when the header is branched to from outside the loop.
886 ///
887 /// Note that it is not valid to call this method on a loop without a
888 /// loop-invariant backedge-taken count (see
889 /// hasLoopInvariantBackedgeTakenCount).
890 const SCEV *getBackedgeTakenCount(const Loop *L, ExitCountKind Kind = Exact);
891
892 /// Similar to getBackedgeTakenCount, except it will add a set of
893 /// SCEV predicates to Predicates that are required to be true in order for
894 /// the answer to be correct. Predicates can be checked with run-time
895 /// checks and can be used to perform loop versioning.
896 const SCEV *getPredicatedBackedgeTakenCount(const Loop *L,
897 SmallVector<const SCEVPredicate *, 4> &Predicates);
898
899 /// When successful, this returns a SCEVConstant that is greater than or equal
900 /// to (i.e. a "conservative over-approximation") of the value returend by
901 /// getBackedgeTakenCount. If such a value cannot be computed, it returns the
902 /// SCEVCouldNotCompute object.
903 const SCEV *getConstantMaxBackedgeTakenCount(const Loop *L) {
904 return getBackedgeTakenCount(L, Kind: ConstantMaximum);
905 }
906
907 /// When successful, this returns a SCEV that is greater than or equal
908 /// to (i.e. a "conservative over-approximation") of the value returend by
909 /// getBackedgeTakenCount. If such a value cannot be computed, it returns the
910 /// SCEVCouldNotCompute object.
911 const SCEV *getSymbolicMaxBackedgeTakenCount(const Loop *L) {
912 return getBackedgeTakenCount(L, Kind: SymbolicMaximum);
913 }
914
915 /// Return true if the backedge taken count is either the value returned by
916 /// getConstantMaxBackedgeTakenCount or zero.
917 bool isBackedgeTakenCountMaxOrZero(const Loop *L);
918
919 /// Return true if the specified loop has an analyzable loop-invariant
920 /// backedge-taken count.
921 bool hasLoopInvariantBackedgeTakenCount(const Loop *L);
922
923 // This method should be called by the client when it made any change that
924 // would invalidate SCEV's answers, and the client wants to remove all loop
925 // information held internally by ScalarEvolution. This is intended to be used
926 // when the alternative to forget a loop is too expensive (i.e. large loop
927 // bodies).
928 void forgetAllLoops();
929
930 /// This method should be called by the client when it has changed a loop in
931 /// a way that may effect ScalarEvolution's ability to compute a trip count,
932 /// or if the loop is deleted. This call is potentially expensive for large
933 /// loop bodies.
934 void forgetLoop(const Loop *L);
935
936 // This method invokes forgetLoop for the outermost loop of the given loop
937 // \p L, making ScalarEvolution forget about all this subtree. This needs to
938 // be done whenever we make a transform that may affect the parameters of the
939 // outer loop, such as exit counts for branches.
940 void forgetTopmostLoop(const Loop *L);
941
942 /// This method should be called by the client when it has changed a value
943 /// in a way that may effect its value, or which may disconnect it from a
944 /// def-use chain linking it to a loop.
945 void forgetValue(Value *V);
946
947 /// Forget LCSSA phi node V of loop L to which a new predecessor was added,
948 /// such that it may no longer be trivial.
949 void forgetLcssaPhiWithNewPredecessor(Loop *L, PHINode *V);
950
951 /// Called when the client has changed the disposition of values in
952 /// this loop.
953 ///
954 /// We don't have a way to invalidate per-loop dispositions. Clear and
955 /// recompute is simpler.
956 void forgetLoopDispositions();
957
958 /// Called when the client has changed the disposition of values in
959 /// a loop or block.
960 ///
961 /// We don't have a way to invalidate per-loop/per-block dispositions. Clear
962 /// and recompute is simpler.
963 void forgetBlockAndLoopDispositions(Value *V = nullptr);
964
965 /// Determine the minimum number of zero bits that S is guaranteed to end in
966 /// (at every loop iteration). It is, at the same time, the minimum number
967 /// of times S is divisible by 2. For example, given {4,+,8} it returns 2.
968 /// If S is guaranteed to be 0, it returns the bitwidth of S.
969 uint32_t getMinTrailingZeros(const SCEV *S);
970
971 /// Returns the max constant multiple of S.
972 APInt getConstantMultiple(const SCEV *S);
973
974 // Returns the max constant multiple of S. If S is exactly 0, return 1.
975 APInt getNonZeroConstantMultiple(const SCEV *S);
976
977 /// Determine the unsigned range for a particular SCEV.
978 /// NOTE: This returns a copy of the reference returned by getRangeRef.
979 ConstantRange getUnsignedRange(const SCEV *S) {
980 return getRangeRef(S, Hint: HINT_RANGE_UNSIGNED);
981 }
982
983 /// Determine the min of the unsigned range for a particular SCEV.
984 APInt getUnsignedRangeMin(const SCEV *S) {
985 return getRangeRef(S, Hint: HINT_RANGE_UNSIGNED).getUnsignedMin();
986 }
987
988 /// Determine the max of the unsigned range for a particular SCEV.
989 APInt getUnsignedRangeMax(const SCEV *S) {
990 return getRangeRef(S, Hint: HINT_RANGE_UNSIGNED).getUnsignedMax();
991 }
992
993 /// Determine the signed range for a particular SCEV.
994 /// NOTE: This returns a copy of the reference returned by getRangeRef.
995 ConstantRange getSignedRange(const SCEV *S) {
996 return getRangeRef(S, Hint: HINT_RANGE_SIGNED);
997 }
998
999 /// Determine the min of the signed range for a particular SCEV.
1000 APInt getSignedRangeMin(const SCEV *S) {
1001 return getRangeRef(S, Hint: HINT_RANGE_SIGNED).getSignedMin();
1002 }
1003
1004 /// Determine the max of the signed range for a particular SCEV.
1005 APInt getSignedRangeMax(const SCEV *S) {
1006 return getRangeRef(S, Hint: HINT_RANGE_SIGNED).getSignedMax();
1007 }
1008
1009 /// Test if the given expression is known to be negative.
1010 bool isKnownNegative(const SCEV *S);
1011
1012 /// Test if the given expression is known to be positive.
1013 bool isKnownPositive(const SCEV *S);
1014
1015 /// Test if the given expression is known to be non-negative.
1016 bool isKnownNonNegative(const SCEV *S);
1017
1018 /// Test if the given expression is known to be non-positive.
1019 bool isKnownNonPositive(const SCEV *S);
1020
1021 /// Test if the given expression is known to be non-zero.
1022 bool isKnownNonZero(const SCEV *S);
1023
1024 /// Splits SCEV expression \p S into two SCEVs. One of them is obtained from
1025 /// \p S by substitution of all AddRec sub-expression related to loop \p L
1026 /// with initial value of that SCEV. The second is obtained from \p S by
1027 /// substitution of all AddRec sub-expressions related to loop \p L with post
1028 /// increment of this AddRec in the loop \p L. In both cases all other AddRec
1029 /// sub-expressions (not related to \p L) remain the same.
1030 /// If the \p S contains non-invariant unknown SCEV the function returns
1031 /// CouldNotCompute SCEV in both values of std::pair.
1032 /// For example, for SCEV S={0, +, 1}<L1> + {0, +, 1}<L2> and loop L=L1
1033 /// the function returns pair:
1034 /// first = {0, +, 1}<L2>
1035 /// second = {1, +, 1}<L1> + {0, +, 1}<L2>
1036 /// We can see that for the first AddRec sub-expression it was replaced with
1037 /// 0 (initial value) for the first element and to {1, +, 1}<L1> (post
1038 /// increment value) for the second one. In both cases AddRec expression
1039 /// related to L2 remains the same.
1040 std::pair<const SCEV *, const SCEV *> SplitIntoInitAndPostInc(const Loop *L,
1041 const SCEV *S);
1042
1043 /// We'd like to check the predicate on every iteration of the most dominated
1044 /// loop between loops used in LHS and RHS.
1045 /// To do this we use the following list of steps:
1046 /// 1. Collect set S all loops on which either LHS or RHS depend.
1047 /// 2. If S is non-empty
1048 /// a. Let PD be the element of S which is dominated by all other elements.
1049 /// b. Let E(LHS) be value of LHS on entry of PD.
1050 /// To get E(LHS), we should just take LHS and replace all AddRecs that are
1051 /// attached to PD on with their entry values.
1052 /// Define E(RHS) in the same way.
1053 /// c. Let B(LHS) be value of L on backedge of PD.
1054 /// To get B(LHS), we should just take LHS and replace all AddRecs that are
1055 /// attached to PD on with their backedge values.
1056 /// Define B(RHS) in the same way.
1057 /// d. Note that E(LHS) and E(RHS) are automatically available on entry of PD,
1058 /// so we can assert on that.
1059 /// e. Return true if isLoopEntryGuardedByCond(Pred, E(LHS), E(RHS)) &&
1060 /// isLoopBackedgeGuardedByCond(Pred, B(LHS), B(RHS))
1061 bool isKnownViaInduction(ICmpInst::Predicate Pred, const SCEV *LHS,
1062 const SCEV *RHS);
1063
1064 /// Test if the given expression is known to satisfy the condition described
1065 /// by Pred, LHS, and RHS.
1066 bool isKnownPredicate(ICmpInst::Predicate Pred, const SCEV *LHS,
1067 const SCEV *RHS);
1068
1069 /// Check whether the condition described by Pred, LHS, and RHS is true or
1070 /// false. If we know it, return the evaluation of this condition. If neither
1071 /// is proved, return std::nullopt.
1072 std::optional<bool> evaluatePredicate(ICmpInst::Predicate Pred,
1073 const SCEV *LHS, const SCEV *RHS);
1074
1075 /// Test if the given expression is known to satisfy the condition described
1076 /// by Pred, LHS, and RHS in the given Context.
1077 bool isKnownPredicateAt(ICmpInst::Predicate Pred, const SCEV *LHS,
1078 const SCEV *RHS, const Instruction *CtxI);
1079
1080 /// Check whether the condition described by Pred, LHS, and RHS is true or
1081 /// false in the given \p Context. If we know it, return the evaluation of
1082 /// this condition. If neither is proved, return std::nullopt.
1083 std::optional<bool> evaluatePredicateAt(ICmpInst::Predicate Pred,
1084 const SCEV *LHS, const SCEV *RHS,
1085 const Instruction *CtxI);
1086
1087 /// Test if the condition described by Pred, LHS, RHS is known to be true on
1088 /// every iteration of the loop of the recurrency LHS.
1089 bool isKnownOnEveryIteration(ICmpInst::Predicate Pred,
1090 const SCEVAddRecExpr *LHS, const SCEV *RHS);
1091
1092 /// Information about the number of loop iterations for which a loop exit's
1093 /// branch condition evaluates to the not-taken path. This is a temporary
1094 /// pair of exact and max expressions that are eventually summarized in
1095 /// ExitNotTakenInfo and BackedgeTakenInfo.
1096 struct ExitLimit {
1097 const SCEV *ExactNotTaken; // The exit is not taken exactly this many times
1098 const SCEV *ConstantMaxNotTaken; // The exit is not taken at most this many
1099 // times
1100 const SCEV *SymbolicMaxNotTaken;
1101
1102 // Not taken either exactly ConstantMaxNotTaken or zero times
1103 bool MaxOrZero = false;
1104
1105 /// A set of predicate guards for this ExitLimit. The result is only valid
1106 /// if all of the predicates in \c Predicates evaluate to 'true' at
1107 /// run-time.
1108 SmallPtrSet<const SCEVPredicate *, 4> Predicates;
1109
1110 void addPredicate(const SCEVPredicate *P) {
1111 assert(!isa<SCEVUnionPredicate>(P) && "Only add leaf predicates here!");
1112 Predicates.insert(Ptr: P);
1113 }
1114
1115 /// Construct either an exact exit limit from a constant, or an unknown
1116 /// one from a SCEVCouldNotCompute. No other types of SCEVs are allowed
1117 /// as arguments and asserts enforce that internally.
1118 /*implicit*/ ExitLimit(const SCEV *E);
1119
1120 ExitLimit(
1121 const SCEV *E, const SCEV *ConstantMaxNotTaken,
1122 const SCEV *SymbolicMaxNotTaken, bool MaxOrZero,
1123 ArrayRef<const SmallPtrSetImpl<const SCEVPredicate *> *> PredSetList =
1124 std::nullopt);
1125
1126 ExitLimit(const SCEV *E, const SCEV *ConstantMaxNotTaken,
1127 const SCEV *SymbolicMaxNotTaken, bool MaxOrZero,
1128 const SmallPtrSetImpl<const SCEVPredicate *> &PredSet);
1129
1130 /// Test whether this ExitLimit contains any computed information, or
1131 /// whether it's all SCEVCouldNotCompute values.
1132 bool hasAnyInfo() const {
1133 return !isa<SCEVCouldNotCompute>(Val: ExactNotTaken) ||
1134 !isa<SCEVCouldNotCompute>(Val: ConstantMaxNotTaken);
1135 }
1136
1137 /// Test whether this ExitLimit contains all information.
1138 bool hasFullInfo() const {
1139 return !isa<SCEVCouldNotCompute>(Val: ExactNotTaken);
1140 }
1141 };
1142
1143 /// Compute the number of times the backedge of the specified loop will
1144 /// execute if its exit condition were a conditional branch of ExitCond.
1145 ///
1146 /// \p ControlsOnlyExit is true if ExitCond directly controls the only exit
1147 /// branch. In this case, we can assume that the loop exits only if the
1148 /// condition is true and can infer that failing to meet the condition prior
1149 /// to integer wraparound results in undefined behavior.
1150 ///
1151 /// If \p AllowPredicates is set, this call will try to use a minimal set of
1152 /// SCEV predicates in order to return an exact answer.
1153 ExitLimit computeExitLimitFromCond(const Loop *L, Value *ExitCond,
1154 bool ExitIfTrue, bool ControlsOnlyExit,
1155 bool AllowPredicates = false);
1156
1157 /// A predicate is said to be monotonically increasing if may go from being
1158 /// false to being true as the loop iterates, but never the other way
1159 /// around. A predicate is said to be monotonically decreasing if may go
1160 /// from being true to being false as the loop iterates, but never the other
1161 /// way around.
1162 enum MonotonicPredicateType {
1163 MonotonicallyIncreasing,
1164 MonotonicallyDecreasing
1165 };
1166
1167 /// If, for all loop invariant X, the predicate "LHS `Pred` X" is
1168 /// monotonically increasing or decreasing, returns
1169 /// Some(MonotonicallyIncreasing) and Some(MonotonicallyDecreasing)
1170 /// respectively. If we could not prove either of these facts, returns
1171 /// std::nullopt.
1172 std::optional<MonotonicPredicateType>
1173 getMonotonicPredicateType(const SCEVAddRecExpr *LHS,
1174 ICmpInst::Predicate Pred);
1175
1176 struct LoopInvariantPredicate {
1177 ICmpInst::Predicate Pred;
1178 const SCEV *LHS;
1179 const SCEV *RHS;
1180
1181 LoopInvariantPredicate(ICmpInst::Predicate Pred, const SCEV *LHS,
1182 const SCEV *RHS)
1183 : Pred(Pred), LHS(LHS), RHS(RHS) {}
1184 };
1185 /// If the result of the predicate LHS `Pred` RHS is loop invariant with
1186 /// respect to L, return a LoopInvariantPredicate with LHS and RHS being
1187 /// invariants, available at L's entry. Otherwise, return std::nullopt.
1188 std::optional<LoopInvariantPredicate>
1189 getLoopInvariantPredicate(ICmpInst::Predicate Pred, const SCEV *LHS,
1190 const SCEV *RHS, const Loop *L,
1191 const Instruction *CtxI = nullptr);
1192
1193 /// If the result of the predicate LHS `Pred` RHS is loop invariant with
1194 /// respect to L at given Context during at least first MaxIter iterations,
1195 /// return a LoopInvariantPredicate with LHS and RHS being invariants,
1196 /// available at L's entry. Otherwise, return std::nullopt. The predicate
1197 /// should be the loop's exit condition.
1198 std::optional<LoopInvariantPredicate>
1199 getLoopInvariantExitCondDuringFirstIterations(ICmpInst::Predicate Pred,
1200 const SCEV *LHS,
1201 const SCEV *RHS, const Loop *L,
1202 const Instruction *CtxI,
1203 const SCEV *MaxIter);
1204
1205 std::optional<LoopInvariantPredicate>
1206 getLoopInvariantExitCondDuringFirstIterationsImpl(
1207 ICmpInst::Predicate Pred, const SCEV *LHS, const SCEV *RHS, const Loop *L,
1208 const Instruction *CtxI, const SCEV *MaxIter);
1209
1210 /// Simplify LHS and RHS in a comparison with predicate Pred. Return true
1211 /// iff any changes were made. If the operands are provably equal or
1212 /// unequal, LHS and RHS are set to the same value and Pred is set to either
1213 /// ICMP_EQ or ICMP_NE.
1214 bool SimplifyICmpOperands(ICmpInst::Predicate &Pred, const SCEV *&LHS,
1215 const SCEV *&RHS, unsigned Depth = 0);
1216
1217 /// Return the "disposition" of the given SCEV with respect to the given
1218 /// loop.
1219 LoopDisposition getLoopDisposition(const SCEV *S, const Loop *L);
1220
1221 /// Return true if the value of the given SCEV is unchanging in the
1222 /// specified loop.
1223 bool isLoopInvariant(const SCEV *S, const Loop *L);
1224
1225 /// Determine if the SCEV can be evaluated at loop's entry. It is true if it
1226 /// doesn't depend on a SCEVUnknown of an instruction which is dominated by
1227 /// the header of loop L.
1228 bool isAvailableAtLoopEntry(const SCEV *S, const Loop *L);
1229
1230 /// Return true if the given SCEV changes value in a known way in the
1231 /// specified loop. This property being true implies that the value is
1232 /// variant in the loop AND that we can emit an expression to compute the
1233 /// value of the expression at any particular loop iteration.
1234 bool hasComputableLoopEvolution(const SCEV *S, const Loop *L);
1235
1236 /// Return the "disposition" of the given SCEV with respect to the given
1237 /// block.
1238 BlockDisposition getBlockDisposition(const SCEV *S, const BasicBlock *BB);
1239
1240 /// Return true if elements that makes up the given SCEV dominate the
1241 /// specified basic block.
1242 bool dominates(const SCEV *S, const BasicBlock *BB);
1243
1244 /// Return true if elements that makes up the given SCEV properly dominate
1245 /// the specified basic block.
1246 bool properlyDominates(const SCEV *S, const BasicBlock *BB);
1247
1248 /// Test whether the given SCEV has Op as a direct or indirect operand.
1249 bool hasOperand(const SCEV *S, const SCEV *Op) const;
1250
1251 /// Return the size of an element read or written by Inst.
1252 const SCEV *getElementSize(Instruction *Inst);
1253
1254 void print(raw_ostream &OS) const;
1255 void verify() const;
1256 bool invalidate(Function &F, const PreservedAnalyses &PA,
1257 FunctionAnalysisManager::Invalidator &Inv);
1258
1259 /// Return the DataLayout associated with the module this SCEV instance is
1260 /// operating on.
1261 const DataLayout &getDataLayout() const {
1262 return F.getParent()->getDataLayout();
1263 }
1264
1265 const SCEVPredicate *getEqualPredicate(const SCEV *LHS, const SCEV *RHS);
1266 const SCEVPredicate *getComparePredicate(ICmpInst::Predicate Pred,
1267 const SCEV *LHS, const SCEV *RHS);
1268
1269 const SCEVPredicate *
1270 getWrapPredicate(const SCEVAddRecExpr *AR,
1271 SCEVWrapPredicate::IncrementWrapFlags AddedFlags);
1272
1273 /// Re-writes the SCEV according to the Predicates in \p A.
1274 const SCEV *rewriteUsingPredicate(const SCEV *S, const Loop *L,
1275 const SCEVPredicate &A);
1276 /// Tries to convert the \p S expression to an AddRec expression,
1277 /// adding additional predicates to \p Preds as required.
1278 const SCEVAddRecExpr *convertSCEVToAddRecWithPredicates(
1279 const SCEV *S, const Loop *L,
1280 SmallPtrSetImpl<const SCEVPredicate *> &Preds);
1281
1282 /// Compute \p LHS - \p RHS and returns the result as an APInt if it is a
1283 /// constant, and std::nullopt if it isn't.
1284 ///
1285 /// This is intended to be a cheaper version of getMinusSCEV. We can be
1286 /// frugal here since we just bail out of actually constructing and
1287 /// canonicalizing an expression in the cases where the result isn't going
1288 /// to be a constant.
1289 std::optional<APInt> computeConstantDifference(const SCEV *LHS,
1290 const SCEV *RHS);
1291
1292 /// Update no-wrap flags of an AddRec. This may drop the cached info about
1293 /// this AddRec (such as range info) in case if new flags may potentially
1294 /// sharpen it.
1295 void setNoWrapFlags(SCEVAddRecExpr *AddRec, SCEV::NoWrapFlags Flags);
1296
1297 /// Try to apply information from loop guards for \p L to \p Expr.
1298 const SCEV *applyLoopGuards(const SCEV *Expr, const Loop *L);
1299
1300 /// Return true if the loop has no abnormal exits. That is, if the loop
1301 /// is not infinite, it must exit through an explicit edge in the CFG.
1302 /// (As opposed to either a) throwing out of the function or b) entering a
1303 /// well defined infinite loop in some callee.)
1304 bool loopHasNoAbnormalExits(const Loop *L) {
1305 return getLoopProperties(L).HasNoAbnormalExits;
1306 }
1307
1308 /// Return true if this loop is finite by assumption. That is,
1309 /// to be infinite, it must also be undefined.
1310 bool loopIsFiniteByAssumption(const Loop *L);
1311
1312 /// Return the set of Values that, if poison, will definitively result in S
1313 /// being poison as well. The returned set may be incomplete, i.e. there can
1314 /// be additional Values that also result in S being poison.
1315 void getPoisonGeneratingValues(SmallPtrSetImpl<const Value *> &Result,
1316 const SCEV *S);
1317
1318 /// Check whether it is poison-safe to represent the expression S using the
1319 /// instruction I. If such a replacement is performed, the poison flags of
1320 /// instructions in DropPoisonGeneratingInsts must be dropped.
1321 bool canReuseInstruction(
1322 const SCEV *S, Instruction *I,
1323 SmallVectorImpl<Instruction *> &DropPoisonGeneratingInsts);
1324
1325 class FoldID {
1326 const SCEV *Op = nullptr;
1327 const Type *Ty = nullptr;
1328 unsigned short C;
1329
1330 public:
1331 FoldID(SCEVTypes C, const SCEV *Op, const Type *Ty) : Op(Op), Ty(Ty), C(C) {
1332 assert(Op);
1333 assert(Ty);
1334 }
1335
1336 FoldID(unsigned short C) : C(C) {}
1337
1338 unsigned computeHash() const {
1339 return detail::combineHashValue(
1340 a: C, b: detail::combineHashValue(a: reinterpret_cast<uintptr_t>(Op),
1341 b: reinterpret_cast<uintptr_t>(Ty)));
1342 }
1343
1344 bool operator==(const FoldID &RHS) const {
1345 return std::tie(args: Op, args: Ty, args: C) == std::tie(args: RHS.Op, args: RHS.Ty, args: RHS.C);
1346 }
1347 };
1348
1349private:
1350 /// A CallbackVH to arrange for ScalarEvolution to be notified whenever a
1351 /// Value is deleted.
1352 class SCEVCallbackVH final : public CallbackVH {
1353 ScalarEvolution *SE;
1354
1355 void deleted() override;
1356 void allUsesReplacedWith(Value *New) override;
1357
1358 public:
1359 SCEVCallbackVH(Value *V, ScalarEvolution *SE = nullptr);
1360 };
1361
1362 friend class SCEVCallbackVH;
1363 friend class SCEVExpander;
1364 friend class SCEVUnknown;
1365
1366 /// The function we are analyzing.
1367 Function &F;
1368
1369 /// Does the module have any calls to the llvm.experimental.guard intrinsic
1370 /// at all? If this is false, we avoid doing work that will only help if
1371 /// thare are guards present in the IR.
1372 bool HasGuards;
1373
1374 /// The target library information for the target we are targeting.
1375 TargetLibraryInfo &TLI;
1376
1377 /// The tracker for \@llvm.assume intrinsics in this function.
1378 AssumptionCache &AC;
1379
1380 /// The dominator tree.
1381 DominatorTree &DT;
1382
1383 /// The loop information for the function we are currently analyzing.
1384 LoopInfo &LI;
1385
1386 /// This SCEV is used to represent unknown trip counts and things.
1387 std::unique_ptr<SCEVCouldNotCompute> CouldNotCompute;
1388
1389 /// The type for HasRecMap.
1390 using HasRecMapType = DenseMap<const SCEV *, bool>;
1391
1392 /// This is a cache to record whether a SCEV contains any scAddRecExpr.
1393 HasRecMapType HasRecMap;
1394
1395 /// The type for ExprValueMap.
1396 using ValueSetVector = SmallSetVector<Value *, 4>;
1397 using ExprValueMapType = DenseMap<const SCEV *, ValueSetVector>;
1398
1399 /// ExprValueMap -- This map records the original values from which
1400 /// the SCEV expr is generated from.
1401 ExprValueMapType ExprValueMap;
1402
1403 /// The type for ValueExprMap.
1404 using ValueExprMapType =
1405 DenseMap<SCEVCallbackVH, const SCEV *, DenseMapInfo<Value *>>;
1406
1407 /// This is a cache of the values we have analyzed so far.
1408 ValueExprMapType ValueExprMap;
1409
1410 /// This is a cache for expressions that got folded to a different existing
1411 /// SCEV.
1412 DenseMap<FoldID, const SCEV *> FoldCache;
1413 DenseMap<const SCEV *, SmallVector<FoldID, 2>> FoldCacheUser;
1414
1415 /// Mark predicate values currently being processed by isImpliedCond.
1416 SmallPtrSet<const Value *, 6> PendingLoopPredicates;
1417
1418 /// Mark SCEVUnknown Phis currently being processed by getRangeRef.
1419 SmallPtrSet<const PHINode *, 6> PendingPhiRanges;
1420
1421 /// Mark SCEVUnknown Phis currently being processed by getRangeRefIter.
1422 SmallPtrSet<const PHINode *, 6> PendingPhiRangesIter;
1423
1424 // Mark SCEVUnknown Phis currently being processed by isImpliedViaMerge.
1425 SmallPtrSet<const PHINode *, 6> PendingMerges;
1426
1427 /// Set to true by isLoopBackedgeGuardedByCond when we're walking the set of
1428 /// conditions dominating the backedge of a loop.
1429 bool WalkingBEDominatingConds = false;
1430
1431 /// Set to true by isKnownPredicateViaSplitting when we're trying to prove a
1432 /// predicate by splitting it into a set of independent predicates.
1433 bool ProvingSplitPredicate = false;
1434
1435 /// Memoized values for the getConstantMultiple
1436 DenseMap<const SCEV *, APInt> ConstantMultipleCache;
1437
1438 /// Return the Value set from which the SCEV expr is generated.
1439 ArrayRef<Value *> getSCEVValues(const SCEV *S);
1440
1441 /// Private helper method for the getConstantMultiple method.
1442 APInt getConstantMultipleImpl(const SCEV *S);
1443
1444 /// Information about the number of times a particular loop exit may be
1445 /// reached before exiting the loop.
1446 struct ExitNotTakenInfo {
1447 PoisoningVH<BasicBlock> ExitingBlock;
1448 const SCEV *ExactNotTaken;
1449 const SCEV *ConstantMaxNotTaken;
1450 const SCEV *SymbolicMaxNotTaken;
1451 SmallPtrSet<const SCEVPredicate *, 4> Predicates;
1452
1453 explicit ExitNotTakenInfo(
1454 PoisoningVH<BasicBlock> ExitingBlock, const SCEV *ExactNotTaken,
1455 const SCEV *ConstantMaxNotTaken, const SCEV *SymbolicMaxNotTaken,
1456 const SmallPtrSet<const SCEVPredicate *, 4> &Predicates)
1457 : ExitingBlock(ExitingBlock), ExactNotTaken(ExactNotTaken),
1458 ConstantMaxNotTaken(ConstantMaxNotTaken),
1459 SymbolicMaxNotTaken(SymbolicMaxNotTaken), Predicates(Predicates) {}
1460
1461 bool hasAlwaysTruePredicate() const {
1462 return Predicates.empty();
1463 }
1464 };
1465
1466 /// Information about the backedge-taken count of a loop. This currently
1467 /// includes an exact count and a maximum count.
1468 ///
1469 class BackedgeTakenInfo {
1470 friend class ScalarEvolution;
1471
1472 /// A list of computable exits and their not-taken counts. Loops almost
1473 /// never have more than one computable exit.
1474 SmallVector<ExitNotTakenInfo, 1> ExitNotTaken;
1475
1476 /// Expression indicating the least constant maximum backedge-taken count of
1477 /// the loop that is known, or a SCEVCouldNotCompute. This expression is
1478 /// only valid if the redicates associated with all loop exits are true.
1479 const SCEV *ConstantMax = nullptr;
1480
1481 /// Indicating if \c ExitNotTaken has an element for every exiting block in
1482 /// the loop.
1483 bool IsComplete = false;
1484
1485 /// Expression indicating the least maximum backedge-taken count of the loop
1486 /// that is known, or a SCEVCouldNotCompute. Lazily computed on first query.
1487 const SCEV *SymbolicMax = nullptr;
1488
1489 /// True iff the backedge is taken either exactly Max or zero times.
1490 bool MaxOrZero = false;
1491
1492 bool isComplete() const { return IsComplete; }
1493 const SCEV *getConstantMax() const { return ConstantMax; }
1494
1495 public:
1496 BackedgeTakenInfo() = default;
1497 BackedgeTakenInfo(BackedgeTakenInfo &&) = default;
1498 BackedgeTakenInfo &operator=(BackedgeTakenInfo &&) = default;
1499
1500 using EdgeExitInfo = std::pair<BasicBlock *, ExitLimit>;
1501
1502 /// Initialize BackedgeTakenInfo from a list of exact exit counts.
1503 BackedgeTakenInfo(ArrayRef<EdgeExitInfo> ExitCounts, bool IsComplete,
1504 const SCEV *ConstantMax, bool MaxOrZero);
1505
1506 /// Test whether this BackedgeTakenInfo contains any computed information,
1507 /// or whether it's all SCEVCouldNotCompute values.
1508 bool hasAnyInfo() const {
1509 return !ExitNotTaken.empty() ||
1510 !isa<SCEVCouldNotCompute>(Val: getConstantMax());
1511 }
1512
1513 /// Test whether this BackedgeTakenInfo contains complete information.
1514 bool hasFullInfo() const { return isComplete(); }
1515
1516 /// Return an expression indicating the exact *backedge-taken*
1517 /// count of the loop if it is known or SCEVCouldNotCompute
1518 /// otherwise. If execution makes it to the backedge on every
1519 /// iteration (i.e. there are no abnormal exists like exception
1520 /// throws and thread exits) then this is the number of times the
1521 /// loop header will execute minus one.
1522 ///
1523 /// If the SCEV predicate associated with the answer can be different
1524 /// from AlwaysTrue, we must add a (non null) Predicates argument.
1525 /// The SCEV predicate associated with the answer will be added to
1526 /// Predicates. A run-time check needs to be emitted for the SCEV
1527 /// predicate in order for the answer to be valid.
1528 ///
1529 /// Note that we should always know if we need to pass a predicate
1530 /// argument or not from the way the ExitCounts vector was computed.
1531 /// If we allowed SCEV predicates to be generated when populating this
1532 /// vector, this information can contain them and therefore a
1533 /// SCEVPredicate argument should be added to getExact.
1534 const SCEV *getExact(const Loop *L, ScalarEvolution *SE,
1535 SmallVector<const SCEVPredicate *, 4> *Predicates = nullptr) const;
1536
1537 /// Return the number of times this loop exit may fall through to the back
1538 /// edge, or SCEVCouldNotCompute. The loop is guaranteed not to exit via
1539 /// this block before this number of iterations, but may exit via another
1540 /// block.
1541 const SCEV *getExact(const BasicBlock *ExitingBlock,
1542 ScalarEvolution *SE) const;
1543
1544 /// Get the constant max backedge taken count for the loop.
1545 const SCEV *getConstantMax(ScalarEvolution *SE) const;
1546
1547 /// Get the constant max backedge taken count for the particular loop exit.
1548 const SCEV *getConstantMax(const BasicBlock *ExitingBlock,
1549 ScalarEvolution *SE) const;
1550
1551 /// Get the symbolic max backedge taken count for the loop.
1552 const SCEV *getSymbolicMax(const Loop *L, ScalarEvolution *SE);
1553
1554 /// Get the symbolic max backedge taken count for the particular loop exit.
1555 const SCEV *getSymbolicMax(const BasicBlock *ExitingBlock,
1556 ScalarEvolution *SE) const;
1557
1558 /// Return true if the number of times this backedge is taken is either the
1559 /// value returned by getConstantMax or zero.
1560 bool isConstantMaxOrZero(ScalarEvolution *SE) const;
1561 };
1562
1563 /// Cache the backedge-taken count of the loops for this function as they
1564 /// are computed.
1565 DenseMap<const Loop *, BackedgeTakenInfo> BackedgeTakenCounts;
1566
1567 /// Cache the predicated backedge-taken count of the loops for this
1568 /// function as they are computed.
1569 DenseMap<const Loop *, BackedgeTakenInfo> PredicatedBackedgeTakenCounts;
1570
1571 /// Loops whose backedge taken counts directly use this non-constant SCEV.
1572 DenseMap<const SCEV *, SmallPtrSet<PointerIntPair<const Loop *, 1, bool>, 4>>
1573 BECountUsers;
1574
1575 /// This map contains entries for all of the PHI instructions that we
1576 /// attempt to compute constant evolutions for. This allows us to avoid
1577 /// potentially expensive recomputation of these properties. An instruction
1578 /// maps to null if we are unable to compute its exit value.
1579 DenseMap<PHINode *, Constant *> ConstantEvolutionLoopExitValue;
1580
1581 /// This map contains entries for all the expressions that we attempt to
1582 /// compute getSCEVAtScope information for, which can be expensive in
1583 /// extreme cases.
1584 DenseMap<const SCEV *, SmallVector<std::pair<const Loop *, const SCEV *>, 2>>
1585 ValuesAtScopes;
1586
1587 /// Reverse map for invalidation purposes: Stores of which SCEV and which
1588 /// loop this is the value-at-scope of.
1589 DenseMap<const SCEV *, SmallVector<std::pair<const Loop *, const SCEV *>, 2>>
1590 ValuesAtScopesUsers;
1591
1592 /// Memoized computeLoopDisposition results.
1593 DenseMap<const SCEV *,
1594 SmallVector<PointerIntPair<const Loop *, 2, LoopDisposition>, 2>>
1595 LoopDispositions;
1596
1597 struct LoopProperties {
1598 /// Set to true if the loop contains no instruction that can abnormally exit
1599 /// the loop (i.e. via throwing an exception, by terminating the thread
1600 /// cleanly or by infinite looping in a called function). Strictly
1601 /// speaking, the last one is not leaving the loop, but is identical to
1602 /// leaving the loop for reasoning about undefined behavior.
1603 bool HasNoAbnormalExits;
1604
1605 /// Set to true if the loop contains no instruction that can have side
1606 /// effects (i.e. via throwing an exception, volatile or atomic access).
1607 bool HasNoSideEffects;
1608 };
1609
1610 /// Cache for \c getLoopProperties.
1611 DenseMap<const Loop *, LoopProperties> LoopPropertiesCache;
1612
1613 /// Return a \c LoopProperties instance for \p L, creating one if necessary.
1614 LoopProperties getLoopProperties(const Loop *L);
1615
1616 bool loopHasNoSideEffects(const Loop *L) {
1617 return getLoopProperties(L).HasNoSideEffects;
1618 }
1619
1620 /// Compute a LoopDisposition value.
1621 LoopDisposition computeLoopDisposition(const SCEV *S, const Loop *L);
1622
1623 /// Memoized computeBlockDisposition results.
1624 DenseMap<
1625 const SCEV *,
1626 SmallVector<PointerIntPair<const BasicBlock *, 2, BlockDisposition>, 2>>
1627 BlockDispositions;
1628
1629 /// Compute a BlockDisposition value.
1630 BlockDisposition computeBlockDisposition(const SCEV *S, const BasicBlock *BB);
1631
1632 /// Stores all SCEV that use a given SCEV as its direct operand.
1633 DenseMap<const SCEV *, SmallPtrSet<const SCEV *, 8> > SCEVUsers;
1634
1635 /// Memoized results from getRange
1636 DenseMap<const SCEV *, ConstantRange> UnsignedRanges;
1637
1638 /// Memoized results from getRange
1639 DenseMap<const SCEV *, ConstantRange> SignedRanges;
1640
1641 /// Used to parameterize getRange
1642 enum RangeSignHint { HINT_RANGE_UNSIGNED, HINT_RANGE_SIGNED };
1643
1644 /// Set the memoized range for the given SCEV.
1645 const ConstantRange &setRange(const SCEV *S, RangeSignHint Hint,
1646 ConstantRange CR) {
1647 DenseMap<const SCEV *, ConstantRange> &Cache =
1648 Hint == HINT_RANGE_UNSIGNED ? UnsignedRanges : SignedRanges;
1649
1650 auto Pair = Cache.try_emplace(Key: S, Args: std::move(CR));
1651 if (!Pair.second)
1652 Pair.first->second = std::move(CR);
1653 return Pair.first->second;
1654 }
1655
1656 /// Determine the range for a particular SCEV.
1657 /// NOTE: This returns a reference to an entry in a cache. It must be
1658 /// copied if its needed for longer.
1659 const ConstantRange &getRangeRef(const SCEV *S, RangeSignHint Hint,
1660 unsigned Depth = 0);
1661
1662 /// Determine the range for a particular SCEV, but evaluates ranges for
1663 /// operands iteratively first.
1664 const ConstantRange &getRangeRefIter(const SCEV *S, RangeSignHint Hint);
1665
1666 /// Determines the range for the affine SCEVAddRecExpr {\p Start,+,\p Step}.
1667 /// Helper for \c getRange.
1668 ConstantRange getRangeForAffineAR(const SCEV *Start, const SCEV *Step,
1669 const APInt &MaxBECount);
1670
1671 /// Determines the range for the affine non-self-wrapping SCEVAddRecExpr {\p
1672 /// Start,+,\p Step}<nw>.
1673 ConstantRange getRangeForAffineNoSelfWrappingAR(const SCEVAddRecExpr *AddRec,
1674 const SCEV *MaxBECount,
1675 unsigned BitWidth,
1676 RangeSignHint SignHint);
1677
1678 /// Try to compute a range for the affine SCEVAddRecExpr {\p Start,+,\p
1679 /// Step} by "factoring out" a ternary expression from the add recurrence.
1680 /// Helper called by \c getRange.
1681 ConstantRange getRangeViaFactoring(const SCEV *Start, const SCEV *Step,
1682 const APInt &MaxBECount);
1683
1684 /// If the unknown expression U corresponds to a simple recurrence, return
1685 /// a constant range which represents the entire recurrence. Note that
1686 /// *add* recurrences with loop invariant steps aren't represented by
1687 /// SCEVUnknowns and thus don't use this mechanism.
1688 ConstantRange getRangeForUnknownRecurrence(const SCEVUnknown *U);
1689
1690 /// We know that there is no SCEV for the specified value. Analyze the
1691 /// expression recursively.
1692 const SCEV *createSCEV(Value *V);
1693
1694 /// We know that there is no SCEV for the specified value. Create a new SCEV
1695 /// for \p V iteratively.
1696 const SCEV *createSCEVIter(Value *V);
1697 /// Collect operands of \p V for which SCEV expressions should be constructed
1698 /// first. Returns a SCEV directly if it can be constructed trivially for \p
1699 /// V.
1700 const SCEV *getOperandsToCreate(Value *V, SmallVectorImpl<Value *> &Ops);
1701
1702 /// Provide the special handling we need to analyze PHI SCEVs.
1703 const SCEV *createNodeForPHI(PHINode *PN);
1704
1705 /// Helper function called from createNodeForPHI.
1706 const SCEV *createAddRecFromPHI(PHINode *PN);
1707
1708 /// A helper function for createAddRecFromPHI to handle simple cases.
1709 const SCEV *createSimpleAffineAddRec(PHINode *PN, Value *BEValueV,
1710 Value *StartValueV);
1711
1712 /// Helper function called from createNodeForPHI.
1713 const SCEV *createNodeFromSelectLikePHI(PHINode *PN);
1714
1715 /// Provide special handling for a select-like instruction (currently this
1716 /// is either a select instruction or a phi node). \p Ty is the type of the
1717 /// instruction being processed, that is assumed equivalent to
1718 /// "Cond ? TrueVal : FalseVal".
1719 std::optional<const SCEV *>
1720 createNodeForSelectOrPHIInstWithICmpInstCond(Type *Ty, ICmpInst *Cond,
1721 Value *TrueVal, Value *FalseVal);
1722
1723 /// See if we can model this select-like instruction via umin_seq expression.
1724 const SCEV *createNodeForSelectOrPHIViaUMinSeq(Value *I, Value *Cond,
1725 Value *TrueVal,
1726 Value *FalseVal);
1727
1728 /// Given a value \p V, which is a select-like instruction (currently this is
1729 /// either a select instruction or a phi node), which is assumed equivalent to
1730 /// Cond ? TrueVal : FalseVal
1731 /// see if we can model it as a SCEV expression.
1732 const SCEV *createNodeForSelectOrPHI(Value *V, Value *Cond, Value *TrueVal,
1733 Value *FalseVal);
1734
1735 /// Provide the special handling we need to analyze GEP SCEVs.
1736 const SCEV *createNodeForGEP(GEPOperator *GEP);
1737
1738 /// Implementation code for getSCEVAtScope; called at most once for each
1739 /// SCEV+Loop pair.
1740 const SCEV *computeSCEVAtScope(const SCEV *S, const Loop *L);
1741
1742 /// Return the BackedgeTakenInfo for the given loop, lazily computing new
1743 /// values if the loop hasn't been analyzed yet. The returned result is
1744 /// guaranteed not to be predicated.
1745 BackedgeTakenInfo &getBackedgeTakenInfo(const Loop *L);
1746
1747 /// Similar to getBackedgeTakenInfo, but will add predicates as required
1748 /// with the purpose of returning complete information.
1749 const BackedgeTakenInfo &getPredicatedBackedgeTakenInfo(const Loop *L);
1750
1751 /// Compute the number of times the specified loop will iterate.
1752 /// If AllowPredicates is set, we will create new SCEV predicates as
1753 /// necessary in order to return an exact answer.
1754 BackedgeTakenInfo computeBackedgeTakenCount(const Loop *L,
1755 bool AllowPredicates = false);
1756
1757 /// Compute the number of times the backedge of the specified loop will
1758 /// execute if it exits via the specified block. If AllowPredicates is set,
1759 /// this call will try to use a minimal set of SCEV predicates in order to
1760 /// return an exact answer.
1761 ExitLimit computeExitLimit(const Loop *L, BasicBlock *ExitingBlock,
1762 bool AllowPredicates = false);
1763
1764 /// Return a symbolic upper bound for the backedge taken count of the loop.
1765 /// This is more general than getConstantMaxBackedgeTakenCount as it returns
1766 /// an arbitrary expression as opposed to only constants.
1767 const SCEV *computeSymbolicMaxBackedgeTakenCount(const Loop *L);
1768
1769 // Helper functions for computeExitLimitFromCond to avoid exponential time
1770 // complexity.
1771
1772 class ExitLimitCache {
1773 // It may look like we need key on the whole (L, ExitIfTrue,
1774 // ControlsOnlyExit, AllowPredicates) tuple, but recursive calls to
1775 // computeExitLimitFromCondCached from computeExitLimitFromCondImpl only
1776 // vary the in \c ExitCond and \c ControlsOnlyExit parameters. We remember
1777 // the initial values of the other values to assert our assumption.
1778 SmallDenseMap<PointerIntPair<Value *, 1>, ExitLimit> TripCountMap;
1779
1780 const Loop *L;
1781 bool ExitIfTrue;
1782 bool AllowPredicates;
1783
1784 public:
1785 ExitLimitCache(const Loop *L, bool ExitIfTrue, bool AllowPredicates)
1786 : L(L), ExitIfTrue(ExitIfTrue), AllowPredicates(AllowPredicates) {}
1787
1788 std::optional<ExitLimit> find(const Loop *L, Value *ExitCond,
1789 bool ExitIfTrue, bool ControlsOnlyExit,
1790 bool AllowPredicates);
1791
1792 void insert(const Loop *L, Value *ExitCond, bool ExitIfTrue,
1793 bool ControlsOnlyExit, bool AllowPredicates,
1794 const ExitLimit &EL);
1795 };
1796
1797 using ExitLimitCacheTy = ExitLimitCache;
1798
1799 ExitLimit computeExitLimitFromCondCached(ExitLimitCacheTy &Cache,
1800 const Loop *L, Value *ExitCond,
1801 bool ExitIfTrue,
1802 bool ControlsOnlyExit,
1803 bool AllowPredicates);
1804 ExitLimit computeExitLimitFromCondImpl(ExitLimitCacheTy &Cache, const Loop *L,
1805 Value *ExitCond, bool ExitIfTrue,
1806 bool ControlsOnlyExit,
1807 bool AllowPredicates);
1808 std::optional<ScalarEvolution::ExitLimit> computeExitLimitFromCondFromBinOp(
1809 ExitLimitCacheTy &Cache, const Loop *L, Value *ExitCond, bool ExitIfTrue,
1810 bool ControlsOnlyExit, bool AllowPredicates);
1811
1812 /// Compute the number of times the backedge of the specified loop will
1813 /// execute if its exit condition were a conditional branch of the ICmpInst
1814 /// ExitCond and ExitIfTrue. If AllowPredicates is set, this call will try
1815 /// to use a minimal set of SCEV predicates in order to return an exact
1816 /// answer.
1817 ExitLimit computeExitLimitFromICmp(const Loop *L, ICmpInst *ExitCond,
1818 bool ExitIfTrue,
1819 bool IsSubExpr,
1820 bool AllowPredicates = false);
1821
1822 /// Variant of previous which takes the components representing an ICmp
1823 /// as opposed to the ICmpInst itself. Note that the prior version can
1824 /// return more precise results in some cases and is preferred when caller
1825 /// has a materialized ICmp.
1826 ExitLimit computeExitLimitFromICmp(const Loop *L, ICmpInst::Predicate Pred,
1827 const SCEV *LHS, const SCEV *RHS,
1828 bool IsSubExpr,
1829 bool AllowPredicates = false);
1830
1831 /// Compute the number of times the backedge of the specified loop will
1832 /// execute if its exit condition were a switch with a single exiting case
1833 /// to ExitingBB.
1834 ExitLimit computeExitLimitFromSingleExitSwitch(const Loop *L,
1835 SwitchInst *Switch,
1836 BasicBlock *ExitingBB,
1837 bool IsSubExpr);
1838
1839 /// Compute the exit limit of a loop that is controlled by a
1840 /// "(IV >> 1) != 0" type comparison. We cannot compute the exact trip
1841 /// count in these cases (since SCEV has no way of expressing them), but we
1842 /// can still sometimes compute an upper bound.
1843 ///
1844 /// Return an ExitLimit for a loop whose backedge is guarded by `LHS Pred
1845 /// RHS`.
1846 ExitLimit computeShiftCompareExitLimit(Value *LHS, Value *RHS, const Loop *L,
1847 ICmpInst::Predicate Pred);
1848
1849 /// If the loop is known to execute a constant number of times (the
1850 /// condition evolves only from constants), try to evaluate a few iterations
1851 /// of the loop until we get the exit condition gets a value of ExitWhen
1852 /// (true or false). If we cannot evaluate the exit count of the loop,
1853 /// return CouldNotCompute.
1854 const SCEV *computeExitCountExhaustively(const Loop *L, Value *Cond,
1855 bool ExitWhen);
1856
1857 /// Return the number of times an exit condition comparing the specified
1858 /// value to zero will execute. If not computable, return CouldNotCompute.
1859 /// If AllowPredicates is set, this call will try to use a minimal set of
1860 /// SCEV predicates in order to return an exact answer.
1861 ExitLimit howFarToZero(const SCEV *V, const Loop *L, bool IsSubExpr,
1862 bool AllowPredicates = false);
1863
1864 /// Return the number of times an exit condition checking the specified
1865 /// value for nonzero will execute. If not computable, return
1866 /// CouldNotCompute.
1867 ExitLimit howFarToNonZero(const SCEV *V, const Loop *L);
1868
1869 /// Return the number of times an exit condition containing the specified
1870 /// less-than comparison will execute. If not computable, return
1871 /// CouldNotCompute.
1872 ///
1873 /// \p isSigned specifies whether the less-than is signed.
1874 ///
1875 /// \p ControlsOnlyExit is true when the LHS < RHS condition directly controls
1876 /// the branch (loops exits only if condition is true). In this case, we can
1877 /// use NoWrapFlags to skip overflow checks.
1878 ///
1879 /// If \p AllowPredicates is set, this call will try to use a minimal set of
1880 /// SCEV predicates in order to return an exact answer.
1881 ExitLimit howManyLessThans(const SCEV *LHS, const SCEV *RHS, const Loop *L,
1882 bool isSigned, bool ControlsOnlyExit,
1883 bool AllowPredicates = false);
1884
1885 ExitLimit howManyGreaterThans(const SCEV *LHS, const SCEV *RHS, const Loop *L,
1886 bool isSigned, bool IsSubExpr,
1887 bool AllowPredicates = false);
1888
1889 /// Return a predecessor of BB (which may not be an immediate predecessor)
1890 /// which has exactly one successor from which BB is reachable, or null if
1891 /// no such block is found.
1892 std::pair<const BasicBlock *, const BasicBlock *>
1893 getPredecessorWithUniqueSuccessorForBB(const BasicBlock *BB) const;
1894
1895 /// Test whether the condition described by Pred, LHS, and RHS is true
1896 /// whenever the given FoundCondValue value evaluates to true in given
1897 /// Context. If Context is nullptr, then the found predicate is true
1898 /// everywhere. LHS and FoundLHS may have different type width.
1899 bool isImpliedCond(ICmpInst::Predicate Pred, const SCEV *LHS, const SCEV *RHS,
1900 const Value *FoundCondValue, bool Inverse,
1901 const Instruction *Context = nullptr);
1902
1903 /// Test whether the condition described by Pred, LHS, and RHS is true
1904 /// whenever the given FoundCondValue value evaluates to true in given
1905 /// Context. If Context is nullptr, then the found predicate is true
1906 /// everywhere. LHS and FoundLHS must have same type width.
1907 bool isImpliedCondBalancedTypes(ICmpInst::Predicate Pred, const SCEV *LHS,
1908 const SCEV *RHS,
1909 ICmpInst::Predicate FoundPred,
1910 const SCEV *FoundLHS, const SCEV *FoundRHS,
1911 const Instruction *CtxI);
1912
1913 /// Test whether the condition described by Pred, LHS, and RHS is true
1914 /// whenever the condition described by FoundPred, FoundLHS, FoundRHS is
1915 /// true in given Context. If Context is nullptr, then the found predicate is
1916 /// true everywhere.
1917 bool isImpliedCond(ICmpInst::Predicate Pred, const SCEV *LHS, const SCEV *RHS,
1918 ICmpInst::Predicate FoundPred, const SCEV *FoundLHS,
1919 const SCEV *FoundRHS,
1920 const Instruction *Context = nullptr);
1921
1922 /// Test whether the condition described by Pred, LHS, and RHS is true
1923 /// whenever the condition described by Pred, FoundLHS, and FoundRHS is
1924 /// true in given Context. If Context is nullptr, then the found predicate is
1925 /// true everywhere.
1926 bool isImpliedCondOperands(ICmpInst::Predicate Pred, const SCEV *LHS,
1927 const SCEV *RHS, const SCEV *FoundLHS,
1928 const SCEV *FoundRHS,
1929 const Instruction *Context = nullptr);
1930
1931 /// Test whether the condition described by Pred, LHS, and RHS is true
1932 /// whenever the condition described by Pred, FoundLHS, and FoundRHS is
1933 /// true. Here LHS is an operation that includes FoundLHS as one of its
1934 /// arguments.
1935 bool isImpliedViaOperations(ICmpInst::Predicate Pred,
1936 const SCEV *LHS, const SCEV *RHS,
1937 const SCEV *FoundLHS, const SCEV *FoundRHS,
1938 unsigned Depth = 0);
1939
1940 /// Test whether the condition described by Pred, LHS, and RHS is true.
1941 /// Use only simple non-recursive types of checks, such as range analysis etc.
1942 bool isKnownViaNonRecursiveReasoning(ICmpInst::Predicate Pred,
1943 const SCEV *LHS, const SCEV *RHS);
1944
1945 /// Test whether the condition described by Pred, LHS, and RHS is true
1946 /// whenever the condition described by Pred, FoundLHS, and FoundRHS is
1947 /// true.
1948 bool isImpliedCondOperandsHelper(ICmpInst::Predicate Pred, const SCEV *LHS,
1949 const SCEV *RHS, const SCEV *FoundLHS,
1950 const SCEV *FoundRHS);
1951
1952 /// Test whether the condition described by Pred, LHS, and RHS is true
1953 /// whenever the condition described by Pred, FoundLHS, and FoundRHS is
1954 /// true. Utility function used by isImpliedCondOperands. Tries to get
1955 /// cases like "X `sgt` 0 => X - 1 `sgt` -1".
1956 bool isImpliedCondOperandsViaRanges(ICmpInst::Predicate Pred, const SCEV *LHS,
1957 const SCEV *RHS,
1958 ICmpInst::Predicate FoundPred,
1959 const SCEV *FoundLHS,
1960 const SCEV *FoundRHS);
1961
1962 /// Return true if the condition denoted by \p LHS \p Pred \p RHS is implied
1963 /// by a call to @llvm.experimental.guard in \p BB.
1964 bool isImpliedViaGuard(const BasicBlock *BB, ICmpInst::Predicate Pred,
1965 const SCEV *LHS, const SCEV *RHS);
1966
1967 /// Test whether the condition described by Pred, LHS, and RHS is true
1968 /// whenever the condition described by Pred, FoundLHS, and FoundRHS is
1969 /// true.
1970 ///
1971 /// This routine tries to rule out certain kinds of integer overflow, and
1972 /// then tries to reason about arithmetic properties of the predicates.
1973 bool isImpliedCondOperandsViaNoOverflow(ICmpInst::Predicate Pred,
1974 const SCEV *LHS, const SCEV *RHS,
1975 const SCEV *FoundLHS,
1976 const SCEV *FoundRHS);
1977
1978 /// Test whether the condition described by Pred, LHS, and RHS is true
1979 /// whenever the condition described by Pred, FoundLHS, and FoundRHS is
1980 /// true.
1981 ///
1982 /// This routine tries to weaken the known condition basing on fact that
1983 /// FoundLHS is an AddRec.
1984 bool isImpliedCondOperandsViaAddRecStart(ICmpInst::Predicate Pred,
1985 const SCEV *LHS, const SCEV *RHS,
1986 const SCEV *FoundLHS,
1987 const SCEV *FoundRHS,
1988 const Instruction *CtxI);
1989
1990 /// Test whether the condition described by Pred, LHS, and RHS is true
1991 /// whenever the condition described by Pred, FoundLHS, and FoundRHS is
1992 /// true.
1993 ///
1994 /// This routine tries to figure out predicate for Phis which are SCEVUnknown
1995 /// if it is true for every possible incoming value from their respective
1996 /// basic blocks.
1997 bool isImpliedViaMerge(ICmpInst::Predicate Pred,
1998 const SCEV *LHS, const SCEV *RHS,
1999 const SCEV *FoundLHS, const SCEV *FoundRHS,
2000 unsigned Depth);
2001
2002 /// Test whether the condition described by Pred, LHS, and RHS is true
2003 /// whenever the condition described by Pred, FoundLHS, and FoundRHS is
2004 /// true.
2005 ///
2006 /// This routine tries to reason about shifts.
2007 bool isImpliedCondOperandsViaShift(ICmpInst::Predicate Pred, const SCEV *LHS,
2008 const SCEV *RHS, const SCEV *FoundLHS,
2009 const SCEV *FoundRHS);
2010
2011 /// If we know that the specified Phi is in the header of its containing
2012 /// loop, we know the loop executes a constant number of times, and the PHI
2013 /// node is just a recurrence involving constants, fold it.
2014 Constant *getConstantEvolutionLoopExitValue(PHINode *PN, const APInt &BEs,
2015 const Loop *L);
2016
2017 /// Test if the given expression is known to satisfy the condition described
2018 /// by Pred and the known constant ranges of LHS and RHS.
2019 bool isKnownPredicateViaConstantRanges(ICmpInst::Predicate Pred,
2020 const SCEV *LHS, const SCEV *RHS);
2021
2022 /// Try to prove the condition described by "LHS Pred RHS" by ruling out
2023 /// integer overflow.
2024 ///
2025 /// For instance, this will return true for "A s< (A + C)<nsw>" if C is
2026 /// positive.
2027 bool isKnownPredicateViaNoOverflow(ICmpInst::Predicate Pred, const SCEV *LHS,
2028 const SCEV *RHS);
2029
2030 /// Try to split Pred LHS RHS into logical conjunctions (and's) and try to
2031 /// prove them individually.
2032 bool isKnownPredicateViaSplitting(ICmpInst::Predicate Pred, const SCEV *LHS,
2033 const SCEV *RHS);
2034
2035 /// Try to match the Expr as "(L + R)<Flags>".
2036 bool splitBinaryAdd(const SCEV *Expr, const SCEV *&L, const SCEV *&R,
2037 SCEV::NoWrapFlags &Flags);
2038
2039 /// Forget predicated/non-predicated backedge taken counts for the given loop.
2040 void forgetBackedgeTakenCounts(const Loop *L, bool Predicated);
2041
2042 /// Drop memoized information for all \p SCEVs.
2043 void forgetMemoizedResults(ArrayRef<const SCEV *> SCEVs);
2044
2045 /// Helper for forgetMemoizedResults.
2046 void forgetMemoizedResultsImpl(const SCEV *S);
2047
2048 /// Iterate over instructions in \p Worklist and their users. Erase entries
2049 /// from ValueExprMap and collect SCEV expressions in \p ToForget
2050 void visitAndClearUsers(SmallVectorImpl<Instruction *> &Worklist,
2051 SmallPtrSetImpl<Instruction *> &Visited,
2052 SmallVectorImpl<const SCEV *> &ToForget);
2053
2054 /// Erase Value from ValueExprMap and ExprValueMap.
2055 void eraseValueFromMap(Value *V);
2056
2057 /// Insert V to S mapping into ValueExprMap and ExprValueMap.
2058 void insertValueToMap(Value *V, const SCEV *S);
2059
2060 /// Return false iff given SCEV contains a SCEVUnknown with NULL value-
2061 /// pointer.
2062 bool checkValidity(const SCEV *S) const;
2063
2064 /// Return true if `ExtendOpTy`({`Start`,+,`Step`}) can be proved to be
2065 /// equal to {`ExtendOpTy`(`Start`),+,`ExtendOpTy`(`Step`)}. This is
2066 /// equivalent to proving no signed (resp. unsigned) wrap in
2067 /// {`Start`,+,`Step`} if `ExtendOpTy` is `SCEVSignExtendExpr`
2068 /// (resp. `SCEVZeroExtendExpr`).
2069 template <typename ExtendOpTy>
2070 bool proveNoWrapByVaryingStart(const SCEV *Start, const SCEV *Step,
2071 const Loop *L);
2072
2073 /// Try to prove NSW or NUW on \p AR relying on ConstantRange manipulation.
2074 SCEV::NoWrapFlags proveNoWrapViaConstantRanges(const SCEVAddRecExpr *AR);
2075
2076 /// Try to prove NSW on \p AR by proving facts about conditions known on
2077 /// entry and backedge.
2078 SCEV::NoWrapFlags proveNoSignedWrapViaInduction(const SCEVAddRecExpr *AR);
2079
2080 /// Try to prove NUW on \p AR by proving facts about conditions known on
2081 /// entry and backedge.
2082 SCEV::NoWrapFlags proveNoUnsignedWrapViaInduction(const SCEVAddRecExpr *AR);
2083
2084 std::optional<MonotonicPredicateType>
2085 getMonotonicPredicateTypeImpl(const SCEVAddRecExpr *LHS,
2086 ICmpInst::Predicate Pred);
2087
2088 /// Return SCEV no-wrap flags that can be proven based on reasoning about
2089 /// how poison produced from no-wrap flags on this value (e.g. a nuw add)
2090 /// would trigger undefined behavior on overflow.
2091 SCEV::NoWrapFlags getNoWrapFlagsFromUB(const Value *V);
2092
2093 /// Return a scope which provides an upper bound on the defining scope of
2094 /// 'S'. Specifically, return the first instruction in said bounding scope.
2095 /// Return nullptr if the scope is trivial (function entry).
2096 /// (See scope definition rules associated with flag discussion above)
2097 const Instruction *getNonTrivialDefiningScopeBound(const SCEV *S);
2098
2099 /// Return a scope which provides an upper bound on the defining scope for
2100 /// a SCEV with the operands in Ops. The outparam Precise is set if the
2101 /// bound found is a precise bound (i.e. must be the defining scope.)
2102 const Instruction *getDefiningScopeBound(ArrayRef<const SCEV *> Ops,
2103 bool &Precise);
2104
2105 /// Wrapper around the above for cases which don't care if the bound
2106 /// is precise.
2107 const Instruction *getDefiningScopeBound(ArrayRef<const SCEV *> Ops);
2108
2109 /// Given two instructions in the same function, return true if we can
2110 /// prove B must execute given A executes.
2111 bool isGuaranteedToTransferExecutionTo(const Instruction *A,
2112 const Instruction *B);
2113
2114 /// Return true if the SCEV corresponding to \p I is never poison. Proving
2115 /// this is more complex than proving that just \p I is never poison, since
2116 /// SCEV commons expressions across control flow, and you can have cases
2117 /// like:
2118 ///
2119 /// idx0 = a + b;
2120 /// ptr[idx0] = 100;
2121 /// if (<condition>) {
2122 /// idx1 = a +nsw b;
2123 /// ptr[idx1] = 200;
2124 /// }
2125 ///
2126 /// where the SCEV expression (+ a b) is guaranteed to not be poison (and
2127 /// hence not sign-overflow) only if "<condition>" is true. Since both
2128 /// `idx0` and `idx1` will be mapped to the same SCEV expression, (+ a b),
2129 /// it is not okay to annotate (+ a b) with <nsw> in the above example.
2130 bool isSCEVExprNeverPoison(const Instruction *I);
2131
2132 /// This is like \c isSCEVExprNeverPoison but it specifically works for
2133 /// instructions that will get mapped to SCEV add recurrences. Return true
2134 /// if \p I will never generate poison under the assumption that \p I is an
2135 /// add recurrence on the loop \p L.
2136 bool isAddRecNeverPoison(const Instruction *I, const Loop *L);
2137
2138 /// Similar to createAddRecFromPHI, but with the additional flexibility of
2139 /// suggesting runtime overflow checks in case casts are encountered.
2140 /// If successful, the analysis records that for this loop, \p SymbolicPHI,
2141 /// which is the UnknownSCEV currently representing the PHI, can be rewritten
2142 /// into an AddRec, assuming some predicates; The function then returns the
2143 /// AddRec and the predicates as a pair, and caches this pair in
2144 /// PredicatedSCEVRewrites.
2145 /// If the analysis is not successful, a mapping from the \p SymbolicPHI to
2146 /// itself (with no predicates) is recorded, and a nullptr with an empty
2147 /// predicates vector is returned as a pair.
2148 std::optional<std::pair<const SCEV *, SmallVector<const SCEVPredicate *, 3>>>
2149 createAddRecFromPHIWithCastsImpl(const SCEVUnknown *SymbolicPHI);
2150
2151 /// Compute the maximum backedge count based on the range of values
2152 /// permitted by Start, End, and Stride. This is for loops of the form
2153 /// {Start, +, Stride} LT End.
2154 ///
2155 /// Preconditions:
2156 /// * the induction variable is known to be positive.
2157 /// * the induction variable is assumed not to overflow (i.e. either it
2158 /// actually doesn't, or we'd have to immediately execute UB)
2159 /// We *don't* assert these preconditions so please be careful.
2160 const SCEV *computeMaxBECountForLT(const SCEV *Start, const SCEV *Stride,
2161 const SCEV *End, unsigned BitWidth,
2162 bool IsSigned);
2163
2164 /// Verify if an linear IV with positive stride can overflow when in a
2165 /// less-than comparison, knowing the invariant term of the comparison,
2166 /// the stride.
2167 bool canIVOverflowOnLT(const SCEV *RHS, const SCEV *Stride, bool IsSigned);
2168
2169 /// Verify if an linear IV with negative stride can overflow when in a
2170 /// greater-than comparison, knowing the invariant term of the comparison,
2171 /// the stride.
2172 bool canIVOverflowOnGT(const SCEV *RHS, const SCEV *Stride, bool IsSigned);
2173
2174 /// Get add expr already created or create a new one.
2175 const SCEV *getOrCreateAddExpr(ArrayRef<const SCEV *> Ops,
2176 SCEV::NoWrapFlags Flags);
2177
2178 /// Get mul expr already created or create a new one.
2179 const SCEV *getOrCreateMulExpr(ArrayRef<const SCEV *> Ops,
2180 SCEV::NoWrapFlags Flags);
2181
2182 // Get addrec expr already created or create a new one.
2183 const SCEV *getOrCreateAddRecExpr(ArrayRef<const SCEV *> Ops,
2184 const Loop *L, SCEV::NoWrapFlags Flags);
2185
2186 /// Return x if \p Val is f(x) where f is a 1-1 function.
2187 const SCEV *stripInjectiveFunctions(const SCEV *Val) const;
2188
2189 /// Find all of the loops transitively used in \p S, and fill \p LoopsUsed.
2190 /// A loop is considered "used" by an expression if it contains
2191 /// an add rec on said loop.
2192 void getUsedLoops(const SCEV *S, SmallPtrSetImpl<const Loop *> &LoopsUsed);
2193
2194 /// Try to match the pattern generated by getURemExpr(A, B). If successful,
2195 /// Assign A and B to LHS and RHS, respectively.
2196 bool matchURem(const SCEV *Expr, const SCEV *&LHS, const SCEV *&RHS);
2197
2198 /// Look for a SCEV expression with type `SCEVType` and operands `Ops` in
2199 /// `UniqueSCEVs`. Return if found, else nullptr.
2200 SCEV *findExistingSCEVInCache(SCEVTypes SCEVType, ArrayRef<const SCEV *> Ops);
2201
2202 /// Get reachable blocks in this function, making limited use of SCEV
2203 /// reasoning about conditions.
2204 void getReachableBlocks(SmallPtrSetImpl<BasicBlock *> &Reachable,
2205 Function &F);
2206
2207 /// Return the given SCEV expression with a new set of operands.
2208 /// This preserves the origial nowrap flags.
2209 const SCEV *getWithOperands(const SCEV *S,
2210 SmallVectorImpl<const SCEV *> &NewOps);
2211
2212 FoldingSet<SCEV> UniqueSCEVs;
2213 FoldingSet<SCEVPredicate> UniquePreds;
2214 BumpPtrAllocator SCEVAllocator;
2215
2216 /// This maps loops to a list of addrecs that directly use said loop.
2217 DenseMap<const Loop *, SmallVector<const SCEVAddRecExpr *, 4>> LoopUsers;
2218
2219 /// Cache tentative mappings from UnknownSCEVs in a Loop, to a SCEV expression
2220 /// they can be rewritten into under certain predicates.
2221 DenseMap<std::pair<const SCEVUnknown *, const Loop *>,
2222 std::pair<const SCEV *, SmallVector<const SCEVPredicate *, 3>>>
2223 PredicatedSCEVRewrites;
2224
2225 /// Set of AddRecs for which proving NUW via an induction has already been
2226 /// tried.
2227 SmallPtrSet<const SCEVAddRecExpr *, 16> UnsignedWrapViaInductionTried;
2228
2229 /// Set of AddRecs for which proving NSW via an induction has already been
2230 /// tried.
2231 SmallPtrSet<const SCEVAddRecExpr *, 16> SignedWrapViaInductionTried;
2232
2233 /// The head of a linked list of all SCEVUnknown values that have been
2234 /// allocated. This is used by releaseMemory to locate them all and call
2235 /// their destructors.
2236 SCEVUnknown *FirstUnknown = nullptr;
2237};
2238
2239/// Analysis pass that exposes the \c ScalarEvolution for a function.
2240class ScalarEvolutionAnalysis
2241 : public AnalysisInfoMixin<ScalarEvolutionAnalysis> {
2242 friend AnalysisInfoMixin<ScalarEvolutionAnalysis>;
2243
2244 static AnalysisKey Key;
2245
2246public:
2247 using Result = ScalarEvolution;
2248
2249 ScalarEvolution run(Function &F, FunctionAnalysisManager &AM);
2250};
2251
2252/// Verifier pass for the \c ScalarEvolutionAnalysis results.
2253class ScalarEvolutionVerifierPass
2254 : public PassInfoMixin<ScalarEvolutionVerifierPass> {
2255public:
2256 PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM);
2257 static bool isRequired() { return true; }
2258};
2259
2260/// Printer pass for the \c ScalarEvolutionAnalysis results.
2261class ScalarEvolutionPrinterPass
2262 : public PassInfoMixin<ScalarEvolutionPrinterPass> {
2263 raw_ostream &OS;
2264
2265public:
2266 explicit ScalarEvolutionPrinterPass(raw_ostream &OS) : OS(OS) {}
2267
2268 PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM);
2269
2270 static bool isRequired() { return true; }
2271};
2272
2273class ScalarEvolutionWrapperPass : public FunctionPass {
2274 std::unique_ptr<ScalarEvolution> SE;
2275
2276public:
2277 static char ID;
2278
2279 ScalarEvolutionWrapperPass();
2280
2281 ScalarEvolution &getSE() { return *SE; }
2282 const ScalarEvolution &getSE() const { return *SE; }
2283
2284 bool runOnFunction(Function &F) override;
2285 void releaseMemory() override;
2286 void getAnalysisUsage(AnalysisUsage &AU) const override;
2287 void print(raw_ostream &OS, const Module * = nullptr) const override;
2288 void verifyAnalysis() const override;
2289};
2290
2291/// An interface layer with SCEV used to manage how we see SCEV expressions
2292/// for values in the context of existing predicates. We can add new
2293/// predicates, but we cannot remove them.
2294///
2295/// This layer has multiple purposes:
2296/// - provides a simple interface for SCEV versioning.
2297/// - guarantees that the order of transformations applied on a SCEV
2298/// expression for a single Value is consistent across two different
2299/// getSCEV calls. This means that, for example, once we've obtained
2300/// an AddRec expression for a certain value through expression
2301/// rewriting, we will continue to get an AddRec expression for that
2302/// Value.
2303/// - lowers the number of expression rewrites.
2304class PredicatedScalarEvolution {
2305public:
2306 PredicatedScalarEvolution(ScalarEvolution &SE, Loop &L);
2307
2308 const SCEVPredicate &getPredicate() const;
2309
2310 /// Returns the SCEV expression of V, in the context of the current SCEV
2311 /// predicate. The order of transformations applied on the expression of V
2312 /// returned by ScalarEvolution is guaranteed to be preserved, even when
2313 /// adding new predicates.
2314 const SCEV *getSCEV(Value *V);
2315
2316 /// Get the (predicated) backedge count for the analyzed loop.
2317 const SCEV *getBackedgeTakenCount();
2318
2319 /// Adds a new predicate.
2320 void addPredicate(const SCEVPredicate &Pred);
2321
2322 /// Attempts to produce an AddRecExpr for V by adding additional SCEV
2323 /// predicates. If we can't transform the expression into an AddRecExpr we
2324 /// return nullptr and not add additional SCEV predicates to the current
2325 /// context.
2326 const SCEVAddRecExpr *getAsAddRec(Value *V);
2327
2328 /// Proves that V doesn't overflow by adding SCEV predicate.
2329 void setNoOverflow(Value *V, SCEVWrapPredicate::IncrementWrapFlags Flags);
2330
2331 /// Returns true if we've proved that V doesn't wrap by means of a SCEV
2332 /// predicate.
2333 bool hasNoOverflow(Value *V, SCEVWrapPredicate::IncrementWrapFlags Flags);
2334
2335 /// Returns the ScalarEvolution analysis used.
2336 ScalarEvolution *getSE() const { return &SE; }
2337
2338 /// We need to explicitly define the copy constructor because of FlagsMap.
2339 PredicatedScalarEvolution(const PredicatedScalarEvolution &);
2340
2341 /// Print the SCEV mappings done by the Predicated Scalar Evolution.
2342 /// The printed text is indented by \p Depth.
2343 void print(raw_ostream &OS, unsigned Depth) const;
2344
2345 /// Check if \p AR1 and \p AR2 are equal, while taking into account
2346 /// Equal predicates in Preds.
2347 bool areAddRecsEqualWithPreds(const SCEVAddRecExpr *AR1,
2348 const SCEVAddRecExpr *AR2) const;
2349
2350private:
2351 /// Increments the version number of the predicate. This needs to be called
2352 /// every time the SCEV predicate changes.
2353 void updateGeneration();
2354
2355 /// Holds a SCEV and the version number of the SCEV predicate used to
2356 /// perform the rewrite of the expression.
2357 using RewriteEntry = std::pair<unsigned, const SCEV *>;
2358
2359 /// Maps a SCEV to the rewrite result of that SCEV at a certain version
2360 /// number. If this number doesn't match the current Generation, we will
2361 /// need to do a rewrite. To preserve the transformation order of previous
2362 /// rewrites, we will rewrite the previous result instead of the original
2363 /// SCEV.
2364 DenseMap<const SCEV *, RewriteEntry> RewriteMap;
2365
2366 /// Records what NoWrap flags we've added to a Value *.
2367 ValueMap<Value *, SCEVWrapPredicate::IncrementWrapFlags> FlagsMap;
2368
2369 /// The ScalarEvolution analysis.
2370 ScalarEvolution &SE;
2371
2372 /// The analyzed Loop.
2373 const Loop &L;
2374
2375 /// The SCEVPredicate that forms our context. We will rewrite all
2376 /// expressions assuming that this predicate true.
2377 std::unique_ptr<SCEVUnionPredicate> Preds;
2378
2379 /// Marks the version of the SCEV predicate used. When rewriting a SCEV
2380 /// expression we mark it with the version of the predicate. We use this to
2381 /// figure out if the predicate has changed from the last rewrite of the
2382 /// SCEV. If so, we need to perform a new rewrite.
2383 unsigned Generation = 0;
2384
2385 /// The backedge taken count.
2386 const SCEV *BackedgeCount = nullptr;
2387};
2388
2389template <> struct DenseMapInfo<ScalarEvolution::FoldID> {
2390 static inline ScalarEvolution::FoldID getEmptyKey() {
2391 ScalarEvolution::FoldID ID(0);
2392 return ID;
2393 }
2394 static inline ScalarEvolution::FoldID getTombstoneKey() {
2395 ScalarEvolution::FoldID ID(1);
2396 return ID;
2397 }
2398
2399 static unsigned getHashValue(const ScalarEvolution::FoldID &Val) {
2400 return Val.computeHash();
2401 }
2402
2403 static bool isEqual(const ScalarEvolution::FoldID &LHS,
2404 const ScalarEvolution::FoldID &RHS) {
2405 return LHS == RHS;
2406 }
2407};
2408
2409} // end namespace llvm
2410
2411#endif // LLVM_ANALYSIS_SCALAREVOLUTION_H
2412

source code of llvm/include/llvm/Analysis/ScalarEvolution.h