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