| 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/Instructions.h" |
| 34 | #include "llvm/IR/PassManager.h" |
| 35 | #include "llvm/IR/ValueHandle.h" |
| 36 | #include "llvm/IR/ValueMap.h" |
| 37 | #include "llvm/Pass.h" |
| 38 | #include "llvm/Support/Compiler.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 GEPOperator; |
| 55 | class LLVMContext; |
| 56 | class Loop; |
| 57 | class LoopInfo; |
| 58 | class raw_ostream; |
| 59 | class ScalarEvolution; |
| 60 | class SCEVAddRecExpr; |
| 61 | class SCEVUnknown; |
| 62 | class StructType; |
| 63 | class TargetLibraryInfo; |
| 64 | class Type; |
| 65 | enum SCEVTypes : unsigned short; |
| 66 | |
| 67 | LLVM_ABI extern bool VerifySCEV; |
| 68 | |
| 69 | /// This class represents an analyzed expression in the program. These are |
| 70 | /// opaque objects that the client is not allowed to do much with directly. |
| 71 | /// |
| 72 | class SCEV : public FoldingSetNode { |
| 73 | friend struct FoldingSetTrait<SCEV>; |
| 74 | |
| 75 | /// A reference to an Interned FoldingSetNodeID for this node. The |
| 76 | /// ScalarEvolution's BumpPtrAllocator holds the data. |
| 77 | FoldingSetNodeIDRef FastID; |
| 78 | |
| 79 | // The SCEV baseclass this node corresponds to |
| 80 | const SCEVTypes SCEVType; |
| 81 | |
| 82 | protected: |
| 83 | // Estimated complexity of this node's expression tree size. |
| 84 | const unsigned short ExpressionSize; |
| 85 | |
| 86 | /// This field is initialized to zero and may be used in subclasses to store |
| 87 | /// miscellaneous information. |
| 88 | unsigned short SubclassData = 0; |
| 89 | |
| 90 | public: |
| 91 | /// NoWrapFlags are bitfield indices into SubclassData. |
| 92 | /// |
| 93 | /// Add and Mul expressions may have no-unsigned-wrap <NUW> or |
| 94 | /// no-signed-wrap <NSW> properties, which are derived from the IR |
| 95 | /// operator. NSW is a misnomer that we use to mean no signed overflow or |
| 96 | /// underflow. |
| 97 | /// |
| 98 | /// AddRec expressions may have a no-self-wraparound <NW> property if, in |
| 99 | /// the integer domain, abs(step) * max-iteration(loop) <= |
| 100 | /// unsigned-max(bitwidth). This means that the recurrence will never reach |
| 101 | /// its start value if the step is non-zero. Computing the same value on |
| 102 | /// each iteration is not considered wrapping, and recurrences with step = 0 |
| 103 | /// are trivially <NW>. <NW> is independent of the sign of step and the |
| 104 | /// value the add recurrence starts with. |
| 105 | /// |
| 106 | /// Note that NUW and NSW are also valid properties of a recurrence, and |
| 107 | /// either implies NW. For convenience, NW will be set for a recurrence |
| 108 | /// whenever either NUW or NSW are set. |
| 109 | /// |
| 110 | /// We require that the flag on a SCEV apply to the entire scope in which |
| 111 | /// that SCEV is defined. A SCEV's scope is set of locations dominated by |
| 112 | /// a defining location, which is in turn described by the following rules: |
| 113 | /// * A SCEVUnknown is at the point of definition of the Value. |
| 114 | /// * A SCEVConstant is defined at all points. |
| 115 | /// * A SCEVAddRec is defined starting with the header of the associated |
| 116 | /// loop. |
| 117 | /// * All other SCEVs are defined at the earlest point all operands are |
| 118 | /// defined. |
| 119 | /// |
| 120 | /// The above rules describe a maximally hoisted form (without regards to |
| 121 | /// potential control dependence). A SCEV is defined anywhere a |
| 122 | /// corresponding instruction could be defined in said maximally hoisted |
| 123 | /// form. Note that SCEVUDivExpr (currently the only expression type which |
| 124 | /// can trap) can be defined per these rules in regions where it would trap |
| 125 | /// at runtime. A SCEV being defined does not require the existence of any |
| 126 | /// instruction within the defined scope. |
| 127 | enum NoWrapFlags { |
| 128 | FlagAnyWrap = 0, // No guarantee. |
| 129 | FlagNW = (1 << 0), // No self-wrap. |
| 130 | FlagNUW = (1 << 1), // No unsigned wrap. |
| 131 | FlagNSW = (1 << 2), // No signed wrap. |
| 132 | NoWrapMask = (1 << 3) - 1 |
| 133 | }; |
| 134 | |
| 135 | explicit SCEV(const FoldingSetNodeIDRef ID, SCEVTypes SCEVTy, |
| 136 | unsigned short ExpressionSize) |
| 137 | : FastID(ID), SCEVType(SCEVTy), ExpressionSize(ExpressionSize) {} |
| 138 | SCEV(const SCEV &) = delete; |
| 139 | SCEV &operator=(const SCEV &) = delete; |
| 140 | |
| 141 | SCEVTypes getSCEVType() const { return SCEVType; } |
| 142 | |
| 143 | /// Return the LLVM type of this SCEV expression. |
| 144 | LLVM_ABI Type *getType() const; |
| 145 | |
| 146 | /// Return operands of this SCEV expression. |
| 147 | LLVM_ABI ArrayRef<const SCEV *> operands() const; |
| 148 | |
| 149 | /// Return true if the expression is a constant zero. |
| 150 | LLVM_ABI bool isZero() const; |
| 151 | |
| 152 | /// Return true if the expression is a constant one. |
| 153 | LLVM_ABI bool isOne() const; |
| 154 | |
| 155 | /// Return true if the expression is a constant all-ones value. |
| 156 | LLVM_ABI bool isAllOnesValue() const; |
| 157 | |
| 158 | /// Return true if the specified scev is negated, but not a constant. |
| 159 | LLVM_ABI bool isNonConstantNegative() const; |
| 160 | |
| 161 | // Returns estimated size of the mathematical expression represented by this |
| 162 | // SCEV. The rules of its calculation are following: |
| 163 | // 1) Size of a SCEV without operands (like constants and SCEVUnknown) is 1; |
| 164 | // 2) Size SCEV with operands Op1, Op2, ..., OpN is calculated by formula: |
| 165 | // (1 + Size(Op1) + ... + Size(OpN)). |
| 166 | // This value gives us an estimation of time we need to traverse through this |
| 167 | // SCEV and all its operands recursively. We may use it to avoid performing |
| 168 | // heavy transformations on SCEVs of excessive size for sake of saving the |
| 169 | // compilation time. |
| 170 | unsigned short getExpressionSize() const { |
| 171 | return ExpressionSize; |
| 172 | } |
| 173 | |
| 174 | /// Print out the internal representation of this scalar to the specified |
| 175 | /// stream. This should really only be used for debugging purposes. |
| 176 | LLVM_ABI void print(raw_ostream &OS) const; |
| 177 | |
| 178 | /// This method is used for debugging. |
| 179 | LLVM_ABI void dump() const; |
| 180 | }; |
| 181 | |
| 182 | // Specialize FoldingSetTrait for SCEV to avoid needing to compute |
| 183 | // temporary FoldingSetNodeID values. |
| 184 | template <> struct FoldingSetTrait<SCEV> : DefaultFoldingSetTrait<SCEV> { |
| 185 | static void Profile(const SCEV &X, FoldingSetNodeID &ID) { ID = X.FastID; } |
| 186 | |
| 187 | static bool Equals(const SCEV &X, const FoldingSetNodeID &ID, unsigned IDHash, |
| 188 | FoldingSetNodeID &TempID) { |
| 189 | return ID == X.FastID; |
| 190 | } |
| 191 | |
| 192 | static unsigned ComputeHash(const SCEV &X, FoldingSetNodeID &TempID) { |
| 193 | return X.FastID.ComputeHash(); |
| 194 | } |
| 195 | }; |
| 196 | |
| 197 | inline raw_ostream &operator<<(raw_ostream &OS, const SCEV &S) { |
| 198 | S.print(OS); |
| 199 | return OS; |
| 200 | } |
| 201 | |
| 202 | /// An object of this class is returned by queries that could not be answered. |
| 203 | /// For example, if you ask for the number of iterations of a linked-list |
| 204 | /// traversal loop, you will get one of these. None of the standard SCEV |
| 205 | /// operations are valid on this class, it is just a marker. |
| 206 | struct SCEVCouldNotCompute : public SCEV { |
| 207 | LLVM_ABI SCEVCouldNotCompute(); |
| 208 | |
| 209 | /// Methods for support type inquiry through isa, cast, and dyn_cast: |
| 210 | LLVM_ABI static bool classof(const SCEV *S); |
| 211 | }; |
| 212 | |
| 213 | /// This class represents an assumption made using SCEV expressions which can |
| 214 | /// be checked at run-time. |
| 215 | class SCEVPredicate : public FoldingSetNode { |
| 216 | friend struct FoldingSetTrait<SCEVPredicate>; |
| 217 | |
| 218 | /// A reference to an Interned FoldingSetNodeID for this node. The |
| 219 | /// ScalarEvolution's BumpPtrAllocator holds the data. |
| 220 | FoldingSetNodeIDRef FastID; |
| 221 | |
| 222 | public: |
| 223 | enum SCEVPredicateKind { P_Union, P_Compare, P_Wrap }; |
| 224 | |
| 225 | protected: |
| 226 | SCEVPredicateKind Kind; |
| 227 | ~SCEVPredicate() = default; |
| 228 | SCEVPredicate(const SCEVPredicate &) = default; |
| 229 | SCEVPredicate &operator=(const SCEVPredicate &) = default; |
| 230 | |
| 231 | public: |
| 232 | LLVM_ABI SCEVPredicate(const FoldingSetNodeIDRef ID, SCEVPredicateKind Kind); |
| 233 | |
| 234 | SCEVPredicateKind getKind() const { return Kind; } |
| 235 | |
| 236 | /// Returns the estimated complexity of this predicate. This is roughly |
| 237 | /// measured in the number of run-time checks required. |
| 238 | virtual unsigned getComplexity() const { return 1; } |
| 239 | |
| 240 | /// Returns true if the predicate is always true. This means that no |
| 241 | /// assumptions were made and nothing needs to be checked at run-time. |
| 242 | virtual bool isAlwaysTrue() const = 0; |
| 243 | |
| 244 | /// Returns true if this predicate implies \p N. |
| 245 | virtual bool implies(const SCEVPredicate *N, ScalarEvolution &SE) const = 0; |
| 246 | |
| 247 | /// Prints a textual representation of this predicate with an indentation of |
| 248 | /// \p Depth. |
| 249 | virtual void print(raw_ostream &OS, unsigned Depth = 0) const = 0; |
| 250 | }; |
| 251 | |
| 252 | inline raw_ostream &operator<<(raw_ostream &OS, const SCEVPredicate &P) { |
| 253 | P.print(OS); |
| 254 | return OS; |
| 255 | } |
| 256 | |
| 257 | // Specialize FoldingSetTrait for SCEVPredicate to avoid needing to compute |
| 258 | // temporary FoldingSetNodeID values. |
| 259 | template <> |
| 260 | struct FoldingSetTrait<SCEVPredicate> : DefaultFoldingSetTrait<SCEVPredicate> { |
| 261 | static void Profile(const SCEVPredicate &X, FoldingSetNodeID &ID) { |
| 262 | ID = X.FastID; |
| 263 | } |
| 264 | |
| 265 | static bool Equals(const SCEVPredicate &X, const FoldingSetNodeID &ID, |
| 266 | unsigned IDHash, FoldingSetNodeID &TempID) { |
| 267 | return ID == X.FastID; |
| 268 | } |
| 269 | |
| 270 | static unsigned ComputeHash(const SCEVPredicate &X, |
| 271 | FoldingSetNodeID &TempID) { |
| 272 | return X.FastID.ComputeHash(); |
| 273 | } |
| 274 | }; |
| 275 | |
| 276 | /// This class represents an assumption that the expression LHS Pred RHS |
| 277 | /// evaluates to true, and this can be checked at run-time. |
| 278 | class LLVM_ABI SCEVComparePredicate final : public SCEVPredicate { |
| 279 | /// We assume that LHS Pred RHS is true. |
| 280 | const ICmpInst::Predicate Pred; |
| 281 | const SCEV *LHS; |
| 282 | const SCEV *RHS; |
| 283 | |
| 284 | public: |
| 285 | SCEVComparePredicate(const FoldingSetNodeIDRef ID, |
| 286 | const ICmpInst::Predicate Pred, |
| 287 | const SCEV *LHS, const SCEV *RHS); |
| 288 | |
| 289 | /// Implementation of the SCEVPredicate interface |
| 290 | bool implies(const SCEVPredicate *N, ScalarEvolution &SE) const override; |
| 291 | void print(raw_ostream &OS, unsigned Depth = 0) const override; |
| 292 | bool isAlwaysTrue() const override; |
| 293 | |
| 294 | ICmpInst::Predicate getPredicate() const { return Pred; } |
| 295 | |
| 296 | /// Returns the left hand side of the predicate. |
| 297 | const SCEV *getLHS() const { return LHS; } |
| 298 | |
| 299 | /// Returns the right hand side of the predicate. |
| 300 | const SCEV *getRHS() const { return RHS; } |
| 301 | |
| 302 | /// Methods for support type inquiry through isa, cast, and dyn_cast: |
| 303 | static bool classof(const SCEVPredicate *P) { |
| 304 | return P->getKind() == P_Compare; |
| 305 | } |
| 306 | }; |
| 307 | |
| 308 | /// This class represents an assumption made on an AddRec expression. Given an |
| 309 | /// affine AddRec expression {a,+,b}, we assume that it has the nssw or nusw |
| 310 | /// flags (defined below) in the first X iterations of the loop, where X is a |
| 311 | /// SCEV expression returned by getPredicatedBackedgeTakenCount). |
| 312 | /// |
| 313 | /// Note that this does not imply that X is equal to the backedge taken |
| 314 | /// count. This means that if we have a nusw predicate for i32 {0,+,1} with a |
| 315 | /// predicated backedge taken count of X, we only guarantee that {0,+,1} has |
| 316 | /// nusw in the first X iterations. {0,+,1} may still wrap in the loop if we |
| 317 | /// have more than X iterations. |
| 318 | class LLVM_ABI SCEVWrapPredicate final : public SCEVPredicate { |
| 319 | public: |
| 320 | /// Similar to SCEV::NoWrapFlags, but with slightly different semantics |
| 321 | /// for FlagNUSW. The increment is considered to be signed, and a + b |
| 322 | /// (where b is the increment) is considered to wrap if: |
| 323 | /// zext(a + b) != zext(a) + sext(b) |
| 324 | /// |
| 325 | /// If Signed is a function that takes an n-bit tuple and maps to the |
| 326 | /// integer domain as the tuples value interpreted as twos complement, |
| 327 | /// and Unsigned a function that takes an n-bit tuple and maps to the |
| 328 | /// integer domain as the base two value of input tuple, then a + b |
| 329 | /// has IncrementNUSW iff: |
| 330 | /// |
| 331 | /// 0 <= Unsigned(a) + Signed(b) < 2^n |
| 332 | /// |
| 333 | /// The IncrementNSSW flag has identical semantics with SCEV::FlagNSW. |
| 334 | /// |
| 335 | /// Note that the IncrementNUSW flag is not commutative: if base + inc |
| 336 | /// has IncrementNUSW, then inc + base doesn't neccessarily have this |
| 337 | /// property. The reason for this is that this is used for sign/zero |
| 338 | /// extending affine AddRec SCEV expressions when a SCEVWrapPredicate is |
| 339 | /// assumed. A {base,+,inc} expression is already non-commutative with |
| 340 | /// regards to base and inc, since it is interpreted as: |
| 341 | /// (((base + inc) + inc) + inc) ... |
| 342 | enum IncrementWrapFlags { |
| 343 | IncrementAnyWrap = 0, // No guarantee. |
| 344 | IncrementNUSW = (1 << 0), // No unsigned with signed increment wrap. |
| 345 | IncrementNSSW = (1 << 1), // No signed with signed increment wrap |
| 346 | // (equivalent with SCEV::NSW) |
| 347 | IncrementNoWrapMask = (1 << 2) - 1 |
| 348 | }; |
| 349 | |
| 350 | /// Convenient IncrementWrapFlags manipulation methods. |
| 351 | [[nodiscard]] static SCEVWrapPredicate::IncrementWrapFlags |
| 352 | clearFlags(SCEVWrapPredicate::IncrementWrapFlags Flags, |
| 353 | SCEVWrapPredicate::IncrementWrapFlags OffFlags) { |
| 354 | assert((Flags & IncrementNoWrapMask) == Flags && "Invalid flags value!" ); |
| 355 | assert((OffFlags & IncrementNoWrapMask) == OffFlags && |
| 356 | "Invalid flags value!" ); |
| 357 | return (SCEVWrapPredicate::IncrementWrapFlags)(Flags & ~OffFlags); |
| 358 | } |
| 359 | |
| 360 | [[nodiscard]] static SCEVWrapPredicate::IncrementWrapFlags |
| 361 | maskFlags(SCEVWrapPredicate::IncrementWrapFlags Flags, int Mask) { |
| 362 | assert((Flags & IncrementNoWrapMask) == Flags && "Invalid flags value!" ); |
| 363 | assert((Mask & IncrementNoWrapMask) == Mask && "Invalid mask value!" ); |
| 364 | |
| 365 | return (SCEVWrapPredicate::IncrementWrapFlags)(Flags & Mask); |
| 366 | } |
| 367 | |
| 368 | [[nodiscard]] static SCEVWrapPredicate::IncrementWrapFlags |
| 369 | setFlags(SCEVWrapPredicate::IncrementWrapFlags Flags, |
| 370 | SCEVWrapPredicate::IncrementWrapFlags OnFlags) { |
| 371 | assert((Flags & IncrementNoWrapMask) == Flags && "Invalid flags value!" ); |
| 372 | assert((OnFlags & IncrementNoWrapMask) == OnFlags && |
| 373 | "Invalid flags value!" ); |
| 374 | |
| 375 | return (SCEVWrapPredicate::IncrementWrapFlags)(Flags | OnFlags); |
| 376 | } |
| 377 | |
| 378 | /// Returns the set of SCEVWrapPredicate no wrap flags implied by a |
| 379 | /// SCEVAddRecExpr. |
| 380 | [[nodiscard]] static SCEVWrapPredicate::IncrementWrapFlags |
| 381 | getImpliedFlags(const SCEVAddRecExpr *AR, ScalarEvolution &SE); |
| 382 | |
| 383 | private: |
| 384 | const SCEVAddRecExpr *AR; |
| 385 | IncrementWrapFlags Flags; |
| 386 | |
| 387 | public: |
| 388 | explicit SCEVWrapPredicate(const FoldingSetNodeIDRef ID, |
| 389 | const SCEVAddRecExpr *AR, |
| 390 | IncrementWrapFlags Flags); |
| 391 | |
| 392 | /// Returns the set assumed no overflow flags. |
| 393 | IncrementWrapFlags getFlags() const { return Flags; } |
| 394 | |
| 395 | /// Implementation of the SCEVPredicate interface |
| 396 | const SCEVAddRecExpr *getExpr() const; |
| 397 | bool implies(const SCEVPredicate *N, ScalarEvolution &SE) const override; |
| 398 | void print(raw_ostream &OS, unsigned Depth = 0) const override; |
| 399 | bool isAlwaysTrue() const override; |
| 400 | |
| 401 | /// Methods for support type inquiry through isa, cast, and dyn_cast: |
| 402 | static bool classof(const SCEVPredicate *P) { |
| 403 | return P->getKind() == P_Wrap; |
| 404 | } |
| 405 | }; |
| 406 | |
| 407 | /// This class represents a composition of other SCEV predicates, and is the |
| 408 | /// class that most clients will interact with. This is equivalent to a |
| 409 | /// logical "AND" of all the predicates in the union. |
| 410 | /// |
| 411 | /// NB! Unlike other SCEVPredicate sub-classes this class does not live in the |
| 412 | /// ScalarEvolution::Preds folding set. This is why the \c add function is sound. |
| 413 | class LLVM_ABI SCEVUnionPredicate final : public SCEVPredicate { |
| 414 | private: |
| 415 | using PredicateMap = |
| 416 | DenseMap<const SCEV *, SmallVector<const SCEVPredicate *, 4>>; |
| 417 | |
| 418 | /// Vector with references to all predicates in this union. |
| 419 | SmallVector<const SCEVPredicate *, 16> Preds; |
| 420 | |
| 421 | /// Adds a predicate to this union. |
| 422 | void add(const SCEVPredicate *N, ScalarEvolution &SE); |
| 423 | |
| 424 | public: |
| 425 | SCEVUnionPredicate(ArrayRef<const SCEVPredicate *> Preds, |
| 426 | ScalarEvolution &SE); |
| 427 | |
| 428 | ArrayRef<const SCEVPredicate *> getPredicates() const { return Preds; } |
| 429 | |
| 430 | /// Implementation of the SCEVPredicate interface |
| 431 | bool isAlwaysTrue() const override; |
| 432 | bool implies(const SCEVPredicate *N, ScalarEvolution &SE) const override; |
| 433 | void print(raw_ostream &OS, unsigned Depth) const override; |
| 434 | |
| 435 | /// We estimate the complexity of a union predicate as the size number of |
| 436 | /// predicates in the union. |
| 437 | unsigned getComplexity() const override { return Preds.size(); } |
| 438 | |
| 439 | /// Methods for support type inquiry through isa, cast, and dyn_cast: |
| 440 | static bool classof(const SCEVPredicate *P) { |
| 441 | return P->getKind() == P_Union; |
| 442 | } |
| 443 | }; |
| 444 | |
| 445 | /// The main scalar evolution driver. Because client code (intentionally) |
| 446 | /// can't do much with the SCEV objects directly, they must ask this class |
| 447 | /// for services. |
| 448 | class ScalarEvolution { |
| 449 | friend class ScalarEvolutionsTest; |
| 450 | |
| 451 | public: |
| 452 | /// An enum describing the relationship between a SCEV and a loop. |
| 453 | enum LoopDisposition { |
| 454 | LoopVariant, ///< The SCEV is loop-variant (unknown). |
| 455 | LoopInvariant, ///< The SCEV is loop-invariant. |
| 456 | LoopComputable ///< The SCEV varies predictably with the loop. |
| 457 | }; |
| 458 | |
| 459 | /// An enum describing the relationship between a SCEV and a basic block. |
| 460 | enum BlockDisposition { |
| 461 | DoesNotDominateBlock, ///< The SCEV does not dominate the block. |
| 462 | DominatesBlock, ///< The SCEV dominates the block. |
| 463 | ProperlyDominatesBlock ///< The SCEV properly dominates the block. |
| 464 | }; |
| 465 | |
| 466 | /// Convenient NoWrapFlags manipulation that hides enum casts and is |
| 467 | /// visible in the ScalarEvolution name space. |
| 468 | [[nodiscard]] static SCEV::NoWrapFlags maskFlags(SCEV::NoWrapFlags Flags, |
| 469 | int Mask) { |
| 470 | return (SCEV::NoWrapFlags)(Flags & Mask); |
| 471 | } |
| 472 | [[nodiscard]] static SCEV::NoWrapFlags setFlags(SCEV::NoWrapFlags Flags, |
| 473 | SCEV::NoWrapFlags OnFlags) { |
| 474 | return (SCEV::NoWrapFlags)(Flags | OnFlags); |
| 475 | } |
| 476 | [[nodiscard]] static SCEV::NoWrapFlags |
| 477 | clearFlags(SCEV::NoWrapFlags Flags, SCEV::NoWrapFlags OffFlags) { |
| 478 | return (SCEV::NoWrapFlags)(Flags & ~OffFlags); |
| 479 | } |
| 480 | [[nodiscard]] static bool hasFlags(SCEV::NoWrapFlags Flags, |
| 481 | SCEV::NoWrapFlags TestFlags) { |
| 482 | return TestFlags == maskFlags(Flags, Mask: TestFlags); |
| 483 | }; |
| 484 | |
| 485 | LLVM_ABI ScalarEvolution(Function &F, TargetLibraryInfo &TLI, |
| 486 | AssumptionCache &AC, DominatorTree &DT, |
| 487 | LoopInfo &LI); |
| 488 | LLVM_ABI ScalarEvolution(ScalarEvolution &&Arg); |
| 489 | LLVM_ABI ~ScalarEvolution(); |
| 490 | |
| 491 | LLVMContext &getContext() const { return F.getContext(); } |
| 492 | |
| 493 | /// Test if values of the given type are analyzable within the SCEV |
| 494 | /// framework. This primarily includes integer types, and it can optionally |
| 495 | /// include pointer types if the ScalarEvolution class has access to |
| 496 | /// target-specific information. |
| 497 | LLVM_ABI bool isSCEVable(Type *Ty) const; |
| 498 | |
| 499 | /// Return the size in bits of the specified type, for which isSCEVable must |
| 500 | /// return true. |
| 501 | LLVM_ABI uint64_t getTypeSizeInBits(Type *Ty) const; |
| 502 | |
| 503 | /// Return a type with the same bitwidth as the given type and which |
| 504 | /// represents how SCEV will treat the given type, for which isSCEVable must |
| 505 | /// return true. For pointer types, this is the pointer-sized integer type. |
| 506 | LLVM_ABI Type *getEffectiveSCEVType(Type *Ty) const; |
| 507 | |
| 508 | // Returns a wider type among {Ty1, Ty2}. |
| 509 | LLVM_ABI Type *getWiderType(Type *Ty1, Type *Ty2) const; |
| 510 | |
| 511 | /// Return true if there exists a point in the program at which both |
| 512 | /// A and B could be operands to the same instruction. |
| 513 | /// SCEV expressions are generally assumed to correspond to instructions |
| 514 | /// which could exists in IR. In general, this requires that there exists |
| 515 | /// a use point in the program where all operands dominate the use. |
| 516 | /// |
| 517 | /// Example: |
| 518 | /// loop { |
| 519 | /// if |
| 520 | /// loop { v1 = load @global1; } |
| 521 | /// else |
| 522 | /// loop { v2 = load @global2; } |
| 523 | /// } |
| 524 | /// No SCEV with operand V1, and v2 can exist in this program. |
| 525 | LLVM_ABI bool instructionCouldExistWithOperands(const SCEV *A, const SCEV *B); |
| 526 | |
| 527 | /// Return true if the SCEV is a scAddRecExpr or it contains |
| 528 | /// scAddRecExpr. The result will be cached in HasRecMap. |
| 529 | LLVM_ABI bool containsAddRecurrence(const SCEV *S); |
| 530 | |
| 531 | /// Is operation \p BinOp between \p LHS and \p RHS provably does not have |
| 532 | /// a signed/unsigned overflow (\p Signed)? If \p CtxI is specified, the |
| 533 | /// no-overflow fact should be true in the context of this instruction. |
| 534 | LLVM_ABI bool willNotOverflow(Instruction::BinaryOps BinOp, bool Signed, |
| 535 | const SCEV *LHS, const SCEV *RHS, |
| 536 | const Instruction *CtxI = nullptr); |
| 537 | |
| 538 | /// Parse NSW/NUW flags from add/sub/mul IR binary operation \p Op into |
| 539 | /// SCEV no-wrap flags, and deduce flag[s] that aren't known yet. |
| 540 | /// Does not mutate the original instruction. Returns std::nullopt if it could |
| 541 | /// not deduce more precise flags than the instruction already has, otherwise |
| 542 | /// returns proven flags. |
| 543 | LLVM_ABI std::optional<SCEV::NoWrapFlags> |
| 544 | getStrengthenedNoWrapFlagsFromBinOp(const OverflowingBinaryOperator *OBO); |
| 545 | |
| 546 | /// Notify this ScalarEvolution that \p User directly uses SCEVs in \p Ops. |
| 547 | LLVM_ABI void registerUser(const SCEV *User, ArrayRef<const SCEV *> Ops); |
| 548 | |
| 549 | /// Return true if the SCEV expression contains an undef value. |
| 550 | LLVM_ABI bool containsUndefs(const SCEV *S) const; |
| 551 | |
| 552 | /// Return true if the SCEV expression contains a Value that has been |
| 553 | /// optimised out and is now a nullptr. |
| 554 | LLVM_ABI bool containsErasedValue(const SCEV *S) const; |
| 555 | |
| 556 | /// Return a SCEV expression for the full generality of the specified |
| 557 | /// expression. |
| 558 | LLVM_ABI const SCEV *getSCEV(Value *V); |
| 559 | |
| 560 | /// Return an existing SCEV for V if there is one, otherwise return nullptr. |
| 561 | LLVM_ABI const SCEV *getExistingSCEV(Value *V); |
| 562 | |
| 563 | LLVM_ABI const SCEV *getConstant(ConstantInt *V); |
| 564 | LLVM_ABI const SCEV *getConstant(const APInt &Val); |
| 565 | LLVM_ABI const SCEV *getConstant(Type *Ty, uint64_t V, bool isSigned = false); |
| 566 | LLVM_ABI const SCEV *getLosslessPtrToIntExpr(const SCEV *Op, |
| 567 | unsigned Depth = 0); |
| 568 | LLVM_ABI const SCEV *getPtrToIntExpr(const SCEV *Op, Type *Ty); |
| 569 | LLVM_ABI const SCEV *getTruncateExpr(const SCEV *Op, Type *Ty, |
| 570 | unsigned Depth = 0); |
| 571 | LLVM_ABI const SCEV *getVScale(Type *Ty); |
| 572 | LLVM_ABI const SCEV *getElementCount(Type *Ty, ElementCount EC); |
| 573 | LLVM_ABI const SCEV *getZeroExtendExpr(const SCEV *Op, Type *Ty, |
| 574 | unsigned Depth = 0); |
| 575 | LLVM_ABI const SCEV *getZeroExtendExprImpl(const SCEV *Op, Type *Ty, |
| 576 | unsigned Depth = 0); |
| 577 | LLVM_ABI const SCEV *getSignExtendExpr(const SCEV *Op, Type *Ty, |
| 578 | unsigned Depth = 0); |
| 579 | LLVM_ABI const SCEV *getSignExtendExprImpl(const SCEV *Op, Type *Ty, |
| 580 | unsigned Depth = 0); |
| 581 | LLVM_ABI const SCEV *getCastExpr(SCEVTypes Kind, const SCEV *Op, Type *Ty); |
| 582 | LLVM_ABI const SCEV *getAnyExtendExpr(const SCEV *Op, Type *Ty); |
| 583 | LLVM_ABI const SCEV *getAddExpr(SmallVectorImpl<const SCEV *> &Ops, |
| 584 | SCEV::NoWrapFlags Flags = SCEV::FlagAnyWrap, |
| 585 | unsigned Depth = 0); |
| 586 | const SCEV *getAddExpr(const SCEV *LHS, const SCEV *RHS, |
| 587 | SCEV::NoWrapFlags Flags = SCEV::FlagAnyWrap, |
| 588 | unsigned Depth = 0) { |
| 589 | SmallVector<const SCEV *, 2> Ops = {LHS, RHS}; |
| 590 | return getAddExpr(Ops, Flags, Depth); |
| 591 | } |
| 592 | const SCEV *getAddExpr(const SCEV *Op0, const SCEV *Op1, const SCEV *Op2, |
| 593 | SCEV::NoWrapFlags Flags = SCEV::FlagAnyWrap, |
| 594 | unsigned Depth = 0) { |
| 595 | SmallVector<const SCEV *, 3> Ops = {Op0, Op1, Op2}; |
| 596 | return getAddExpr(Ops, Flags, Depth); |
| 597 | } |
| 598 | LLVM_ABI const SCEV *getMulExpr(SmallVectorImpl<const SCEV *> &Ops, |
| 599 | SCEV::NoWrapFlags Flags = SCEV::FlagAnyWrap, |
| 600 | unsigned Depth = 0); |
| 601 | const SCEV *getMulExpr(const SCEV *LHS, const SCEV *RHS, |
| 602 | SCEV::NoWrapFlags Flags = SCEV::FlagAnyWrap, |
| 603 | unsigned Depth = 0) { |
| 604 | SmallVector<const SCEV *, 2> Ops = {LHS, RHS}; |
| 605 | return getMulExpr(Ops, Flags, Depth); |
| 606 | } |
| 607 | const SCEV *getMulExpr(const SCEV *Op0, const SCEV *Op1, const SCEV *Op2, |
| 608 | SCEV::NoWrapFlags Flags = SCEV::FlagAnyWrap, |
| 609 | unsigned Depth = 0) { |
| 610 | SmallVector<const SCEV *, 3> Ops = {Op0, Op1, Op2}; |
| 611 | return getMulExpr(Ops, Flags, Depth); |
| 612 | } |
| 613 | LLVM_ABI const SCEV *getUDivExpr(const SCEV *LHS, const SCEV *RHS); |
| 614 | LLVM_ABI const SCEV *getUDivExactExpr(const SCEV *LHS, const SCEV *RHS); |
| 615 | LLVM_ABI const SCEV *getURemExpr(const SCEV *LHS, const SCEV *RHS); |
| 616 | LLVM_ABI const SCEV *getAddRecExpr(const SCEV *Start, const SCEV *Step, |
| 617 | const Loop *L, SCEV::NoWrapFlags Flags); |
| 618 | LLVM_ABI const SCEV *getAddRecExpr(SmallVectorImpl<const SCEV *> &Operands, |
| 619 | const Loop *L, SCEV::NoWrapFlags Flags); |
| 620 | const SCEV *getAddRecExpr(const SmallVectorImpl<const SCEV *> &Operands, |
| 621 | const Loop *L, SCEV::NoWrapFlags Flags) { |
| 622 | SmallVector<const SCEV *, 4> NewOp(Operands.begin(), Operands.end()); |
| 623 | return getAddRecExpr(Operands&: NewOp, L, Flags); |
| 624 | } |
| 625 | |
| 626 | /// Checks if \p SymbolicPHI can be rewritten as an AddRecExpr under some |
| 627 | /// Predicates. If successful return these <AddRecExpr, Predicates>; |
| 628 | /// The function is intended to be called from PSCEV (the caller will decide |
| 629 | /// whether to actually add the predicates and carry out the rewrites). |
| 630 | LLVM_ABI std::optional< |
| 631 | std::pair<const SCEV *, SmallVector<const SCEVPredicate *, 3>>> |
| 632 | createAddRecFromPHIWithCasts(const SCEVUnknown *SymbolicPHI); |
| 633 | |
| 634 | /// Returns an expression for a GEP |
| 635 | /// |
| 636 | /// \p GEP The GEP. The indices contained in the GEP itself are ignored, |
| 637 | /// instead we use IndexExprs. |
| 638 | /// \p IndexExprs The expressions for the indices. |
| 639 | LLVM_ABI const SCEV * |
| 640 | getGEPExpr(GEPOperator *GEP, const SmallVectorImpl<const SCEV *> &IndexExprs); |
| 641 | LLVM_ABI const SCEV *getAbsExpr(const SCEV *Op, bool IsNSW); |
| 642 | LLVM_ABI const SCEV *getMinMaxExpr(SCEVTypes Kind, |
| 643 | SmallVectorImpl<const SCEV *> &Operands); |
| 644 | LLVM_ABI const SCEV * |
| 645 | getSequentialMinMaxExpr(SCEVTypes Kind, |
| 646 | SmallVectorImpl<const SCEV *> &Operands); |
| 647 | LLVM_ABI const SCEV *getSMaxExpr(const SCEV *LHS, const SCEV *RHS); |
| 648 | LLVM_ABI const SCEV *getSMaxExpr(SmallVectorImpl<const SCEV *> &Operands); |
| 649 | LLVM_ABI const SCEV *getUMaxExpr(const SCEV *LHS, const SCEV *RHS); |
| 650 | LLVM_ABI const SCEV *getUMaxExpr(SmallVectorImpl<const SCEV *> &Operands); |
| 651 | LLVM_ABI const SCEV *getSMinExpr(const SCEV *LHS, const SCEV *RHS); |
| 652 | LLVM_ABI const SCEV *getSMinExpr(SmallVectorImpl<const SCEV *> &Operands); |
| 653 | LLVM_ABI const SCEV *getUMinExpr(const SCEV *LHS, const SCEV *RHS, |
| 654 | bool Sequential = false); |
| 655 | LLVM_ABI const SCEV *getUMinExpr(SmallVectorImpl<const SCEV *> &Operands, |
| 656 | bool Sequential = false); |
| 657 | LLVM_ABI const SCEV *getUnknown(Value *V); |
| 658 | LLVM_ABI const SCEV *getCouldNotCompute(); |
| 659 | |
| 660 | /// Return a SCEV for the constant 0 of a specific type. |
| 661 | const SCEV *getZero(Type *Ty) { return getConstant(Ty, V: 0); } |
| 662 | |
| 663 | /// Return a SCEV for the constant 1 of a specific type. |
| 664 | const SCEV *getOne(Type *Ty) { return getConstant(Ty, V: 1); } |
| 665 | |
| 666 | /// Return a SCEV for the constant \p Power of two. |
| 667 | const SCEV *getPowerOfTwo(Type *Ty, unsigned Power) { |
| 668 | assert(Power < getTypeSizeInBits(Ty) && "Power out of range" ); |
| 669 | return getConstant(Val: APInt::getOneBitSet(numBits: getTypeSizeInBits(Ty), BitNo: Power)); |
| 670 | } |
| 671 | |
| 672 | /// Return a SCEV for the constant -1 of a specific type. |
| 673 | const SCEV *getMinusOne(Type *Ty) { |
| 674 | return getConstant(Ty, V: -1, /*isSigned=*/isSigned: true); |
| 675 | } |
| 676 | |
| 677 | /// Return an expression for a TypeSize. |
| 678 | LLVM_ABI const SCEV *getSizeOfExpr(Type *IntTy, TypeSize Size); |
| 679 | |
| 680 | /// Return an expression for the alloc size of AllocTy that is type IntTy |
| 681 | LLVM_ABI const SCEV *getSizeOfExpr(Type *IntTy, Type *AllocTy); |
| 682 | |
| 683 | /// Return an expression for the store size of StoreTy that is type IntTy |
| 684 | LLVM_ABI const SCEV *getStoreSizeOfExpr(Type *IntTy, Type *StoreTy); |
| 685 | |
| 686 | /// Return an expression for offsetof on the given field with type IntTy |
| 687 | LLVM_ABI const SCEV *getOffsetOfExpr(Type *IntTy, StructType *STy, |
| 688 | unsigned FieldNo); |
| 689 | |
| 690 | /// Return the SCEV object corresponding to -V. |
| 691 | LLVM_ABI const SCEV * |
| 692 | getNegativeSCEV(const SCEV *V, SCEV::NoWrapFlags Flags = SCEV::FlagAnyWrap); |
| 693 | |
| 694 | /// Return the SCEV object corresponding to ~V. |
| 695 | LLVM_ABI const SCEV *getNotSCEV(const SCEV *V); |
| 696 | |
| 697 | /// Return LHS-RHS. Minus is represented in SCEV as A+B*-1. |
| 698 | /// |
| 699 | /// If the LHS and RHS are pointers which don't share a common base |
| 700 | /// (according to getPointerBase()), this returns a SCEVCouldNotCompute. |
| 701 | /// To compute the difference between two unrelated pointers, you can |
| 702 | /// explicitly convert the arguments using getPtrToIntExpr(), for pointer |
| 703 | /// types that support it. |
| 704 | LLVM_ABI const SCEV *getMinusSCEV(const SCEV *LHS, const SCEV *RHS, |
| 705 | SCEV::NoWrapFlags Flags = SCEV::FlagAnyWrap, |
| 706 | unsigned Depth = 0); |
| 707 | |
| 708 | /// Compute ceil(N / D). N and D are treated as unsigned values. |
| 709 | /// |
| 710 | /// Since SCEV doesn't have native ceiling division, this generates a |
| 711 | /// SCEV expression of the following form: |
| 712 | /// |
| 713 | /// umin(N, 1) + floor((N - umin(N, 1)) / D) |
| 714 | /// |
| 715 | /// A denominator of zero or poison is handled the same way as getUDivExpr(). |
| 716 | LLVM_ABI const SCEV *getUDivCeilSCEV(const SCEV *N, const SCEV *D); |
| 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 zero extended. |
| 720 | LLVM_ABI const SCEV *getTruncateOrZeroExtend(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 sign extended. |
| 725 | LLVM_ABI const SCEV *getTruncateOrSignExtend(const SCEV *V, Type *Ty, |
| 726 | unsigned Depth = 0); |
| 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 zero extended. The |
| 730 | /// conversion must not be narrowing. |
| 731 | LLVM_ABI const SCEV *getNoopOrZeroExtend(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 sign extended. The |
| 735 | /// conversion must not be narrowing. |
| 736 | LLVM_ABI const SCEV *getNoopOrSignExtend(const SCEV *V, Type *Ty); |
| 737 | |
| 738 | /// Return a SCEV corresponding to a conversion of the input value to the |
| 739 | /// specified type. If the type must be extended, it is extended with |
| 740 | /// unspecified bits. The conversion must not be narrowing. |
| 741 | LLVM_ABI const SCEV *getNoopOrAnyExtend(const SCEV *V, Type *Ty); |
| 742 | |
| 743 | /// Return a SCEV corresponding to a conversion of the input value to the |
| 744 | /// specified type. The conversion must not be widening. |
| 745 | LLVM_ABI const SCEV *getTruncateOrNoop(const SCEV *V, Type *Ty); |
| 746 | |
| 747 | /// Promote the operands to the wider of the types using zero-extension, and |
| 748 | /// then perform a umax operation with them. |
| 749 | LLVM_ABI const SCEV *getUMaxFromMismatchedTypes(const SCEV *LHS, |
| 750 | const SCEV *RHS); |
| 751 | |
| 752 | /// Promote the operands to the wider of the types using zero-extension, and |
| 753 | /// then perform a umin operation with them. |
| 754 | LLVM_ABI const SCEV *getUMinFromMismatchedTypes(const SCEV *LHS, |
| 755 | const SCEV *RHS, |
| 756 | bool Sequential = false); |
| 757 | |
| 758 | /// Promote the operands to the wider of the types using zero-extension, and |
| 759 | /// then perform a umin operation with them. N-ary function. |
| 760 | LLVM_ABI const SCEV * |
| 761 | getUMinFromMismatchedTypes(SmallVectorImpl<const SCEV *> &Ops, |
| 762 | bool Sequential = false); |
| 763 | |
| 764 | /// Transitively follow the chain of pointer-type operands until reaching a |
| 765 | /// SCEV that does not have a single pointer operand. This returns a |
| 766 | /// SCEVUnknown pointer for well-formed pointer-type expressions, but corner |
| 767 | /// cases do exist. |
| 768 | LLVM_ABI const SCEV *getPointerBase(const SCEV *V); |
| 769 | |
| 770 | /// Compute an expression equivalent to S - getPointerBase(S). |
| 771 | LLVM_ABI const SCEV *removePointerBase(const SCEV *S); |
| 772 | |
| 773 | /// Return a SCEV expression for the specified value at the specified scope |
| 774 | /// in the program. The L value specifies a loop nest to evaluate the |
| 775 | /// expression at, where null is the top-level or a specified loop is |
| 776 | /// immediately inside of the loop. |
| 777 | /// |
| 778 | /// This method can be used to compute the exit value for a variable defined |
| 779 | /// in a loop by querying what the value will hold in the parent loop. |
| 780 | /// |
| 781 | /// In the case that a relevant loop exit value cannot be computed, the |
| 782 | /// original value V is returned. |
| 783 | LLVM_ABI const SCEV *getSCEVAtScope(const SCEV *S, const Loop *L); |
| 784 | |
| 785 | /// This is a convenience function which does getSCEVAtScope(getSCEV(V), L). |
| 786 | LLVM_ABI const SCEV *getSCEVAtScope(Value *V, const Loop *L); |
| 787 | |
| 788 | /// Test whether entry to the loop is protected by a conditional between LHS |
| 789 | /// and RHS. This is used to help avoid max expressions in loop trip |
| 790 | /// counts, and to eliminate casts. |
| 791 | LLVM_ABI bool isLoopEntryGuardedByCond(const Loop *L, CmpPredicate Pred, |
| 792 | const SCEV *LHS, const SCEV *RHS); |
| 793 | |
| 794 | /// Test whether entry to the basic block is protected by a conditional |
| 795 | /// between LHS and RHS. |
| 796 | LLVM_ABI bool isBasicBlockEntryGuardedByCond(const BasicBlock *BB, |
| 797 | CmpPredicate Pred, |
| 798 | const SCEV *LHS, |
| 799 | const SCEV *RHS); |
| 800 | |
| 801 | /// Test whether the backedge of the loop is protected by a conditional |
| 802 | /// between LHS and RHS. This is used to eliminate casts. |
| 803 | LLVM_ABI bool isLoopBackedgeGuardedByCond(const Loop *L, CmpPredicate Pred, |
| 804 | const SCEV *LHS, const SCEV *RHS); |
| 805 | |
| 806 | /// A version of getTripCountFromExitCount below which always picks an |
| 807 | /// evaluation type which can not result in overflow. |
| 808 | LLVM_ABI const SCEV *getTripCountFromExitCount(const SCEV *ExitCount); |
| 809 | |
| 810 | /// Convert from an "exit count" (i.e. "backedge taken count") to a "trip |
| 811 | /// count". A "trip count" is the number of times the header of the loop |
| 812 | /// will execute if an exit is taken after the specified number of backedges |
| 813 | /// have been taken. (e.g. TripCount = ExitCount + 1). Note that the |
| 814 | /// expression can overflow if ExitCount = UINT_MAX. If EvalTy is not wide |
| 815 | /// enough to hold the result without overflow, result unsigned wraps with |
| 816 | /// 2s-complement semantics. ex: EC = 255 (i8), TC = 0 (i8) |
| 817 | LLVM_ABI const SCEV *getTripCountFromExitCount(const SCEV *ExitCount, |
| 818 | Type *EvalTy, const Loop *L); |
| 819 | |
| 820 | /// Returns the exact trip count of the loop if we can compute it, and |
| 821 | /// the result is a small constant. '0' is used to represent an unknown |
| 822 | /// or non-constant trip count. Note that a trip count is simply one more |
| 823 | /// than the backedge taken count for the loop. |
| 824 | LLVM_ABI unsigned getSmallConstantTripCount(const Loop *L); |
| 825 | |
| 826 | /// Return the exact trip count for this loop if we exit through ExitingBlock. |
| 827 | /// '0' is used to represent an unknown or non-constant trip count. Note |
| 828 | /// that a trip count is simply one more than the backedge taken count for |
| 829 | /// the same exit. |
| 830 | /// This "trip count" assumes that control exits via ExitingBlock. More |
| 831 | /// precisely, it is the number of times that control will reach ExitingBlock |
| 832 | /// before taking the branch. For loops with multiple exits, it may not be |
| 833 | /// the number times that the loop header executes if the loop exits |
| 834 | /// prematurely via another branch. |
| 835 | LLVM_ABI unsigned getSmallConstantTripCount(const Loop *L, |
| 836 | const BasicBlock *ExitingBlock); |
| 837 | |
| 838 | /// Returns the upper bound of the loop trip count as a normal unsigned |
| 839 | /// value. |
| 840 | /// Returns 0 if the trip count is unknown, not constant or requires |
| 841 | /// SCEV predicates and \p Predicates is nullptr. |
| 842 | LLVM_ABI unsigned getSmallConstantMaxTripCount( |
| 843 | const Loop *L, |
| 844 | SmallVectorImpl<const SCEVPredicate *> *Predicates = nullptr); |
| 845 | |
| 846 | /// Returns the largest constant divisor of the trip count as a normal |
| 847 | /// unsigned value, if possible. This means that the actual trip count is |
| 848 | /// always a multiple of the returned value. Returns 1 if the trip count is |
| 849 | /// unknown or not guaranteed to be the multiple of a constant., Will also |
| 850 | /// return 1 if the trip count is very large (>= 2^32). |
| 851 | /// Note that the argument is an exit count for loop L, NOT a trip count. |
| 852 | LLVM_ABI unsigned getSmallConstantTripMultiple(const Loop *L, |
| 853 | const SCEV *ExitCount); |
| 854 | |
| 855 | /// Returns the largest constant divisor of the trip count of the |
| 856 | /// loop. Will return 1 if no trip count could be computed, or if a |
| 857 | /// divisor could not be found. |
| 858 | LLVM_ABI unsigned getSmallConstantTripMultiple(const Loop *L); |
| 859 | |
| 860 | /// Returns the largest constant divisor of the trip count of this loop as a |
| 861 | /// normal unsigned value, if possible. This means that the actual trip |
| 862 | /// count is always a multiple of the returned value (don't forget the trip |
| 863 | /// count could very well be zero as well!). As explained in the comments |
| 864 | /// for getSmallConstantTripCount, this assumes that control exits the loop |
| 865 | /// via ExitingBlock. |
| 866 | LLVM_ABI unsigned |
| 867 | getSmallConstantTripMultiple(const Loop *L, const BasicBlock *ExitingBlock); |
| 868 | |
| 869 | /// The terms "backedge taken count" and "exit count" are used |
| 870 | /// interchangeably to refer to the number of times the backedge of a loop |
| 871 | /// has executed before the loop is exited. |
| 872 | enum ExitCountKind { |
| 873 | /// An expression exactly describing the number of times the backedge has |
| 874 | /// executed when a loop is exited. |
| 875 | Exact, |
| 876 | /// A constant which provides an upper bound on the exact trip count. |
| 877 | ConstantMaximum, |
| 878 | /// An expression which provides an upper bound on the exact trip count. |
| 879 | SymbolicMaximum, |
| 880 | }; |
| 881 | |
| 882 | /// Return the number of times the backedge executes before the given exit |
| 883 | /// would be taken; if not exactly computable, return SCEVCouldNotCompute. |
| 884 | /// For a single exit loop, this value is equivelent to the result of |
| 885 | /// getBackedgeTakenCount. The loop is guaranteed to exit (via *some* exit) |
| 886 | /// before the backedge is executed (ExitCount + 1) times. Note that there |
| 887 | /// is no guarantee about *which* exit is taken on the exiting iteration. |
| 888 | LLVM_ABI const SCEV *getExitCount(const Loop *L, |
| 889 | const BasicBlock *ExitingBlock, |
| 890 | ExitCountKind Kind = Exact); |
| 891 | |
| 892 | /// Same as above except this uses the predicated backedge taken info and |
| 893 | /// may require predicates. |
| 894 | LLVM_ABI const SCEV * |
| 895 | getPredicatedExitCount(const Loop *L, const BasicBlock *ExitingBlock, |
| 896 | SmallVectorImpl<const SCEVPredicate *> *Predicates, |
| 897 | ExitCountKind Kind = Exact); |
| 898 | |
| 899 | /// If the specified loop has a predictable backedge-taken count, return it, |
| 900 | /// otherwise return a SCEVCouldNotCompute object. The backedge-taken count is |
| 901 | /// the number of times the loop header will be branched to from within the |
| 902 | /// loop, assuming there are no abnormal exists like exception throws. This is |
| 903 | /// one less than the trip count of the loop, since it doesn't count the first |
| 904 | /// iteration, when the header is branched to from outside the loop. |
| 905 | /// |
| 906 | /// Note that it is not valid to call this method on a loop without a |
| 907 | /// loop-invariant backedge-taken count (see |
| 908 | /// hasLoopInvariantBackedgeTakenCount). |
| 909 | LLVM_ABI const SCEV *getBackedgeTakenCount(const Loop *L, |
| 910 | ExitCountKind Kind = Exact); |
| 911 | |
| 912 | /// Similar to getBackedgeTakenCount, except it will add a set of |
| 913 | /// SCEV predicates to Predicates that are required to be true in order for |
| 914 | /// the answer to be correct. Predicates can be checked with run-time |
| 915 | /// checks and can be used to perform loop versioning. |
| 916 | LLVM_ABI const SCEV *getPredicatedBackedgeTakenCount( |
| 917 | const Loop *L, SmallVectorImpl<const SCEVPredicate *> &Predicates); |
| 918 | |
| 919 | /// When successful, this returns a SCEVConstant that is greater than or equal |
| 920 | /// to (i.e. a "conservative over-approximation") of the value returend by |
| 921 | /// getBackedgeTakenCount. If such a value cannot be computed, it returns the |
| 922 | /// SCEVCouldNotCompute object. |
| 923 | const SCEV *getConstantMaxBackedgeTakenCount(const Loop *L) { |
| 924 | return getBackedgeTakenCount(L, Kind: ConstantMaximum); |
| 925 | } |
| 926 | |
| 927 | /// Similar to getConstantMaxBackedgeTakenCount, except it will add a set of |
| 928 | /// SCEV predicates to Predicates that are required to be true in order for |
| 929 | /// the answer to be correct. Predicates can be checked with run-time |
| 930 | /// checks and can be used to perform loop versioning. |
| 931 | LLVM_ABI const SCEV *getPredicatedConstantMaxBackedgeTakenCount( |
| 932 | const Loop *L, SmallVectorImpl<const SCEVPredicate *> &Predicates); |
| 933 | |
| 934 | /// When successful, this returns a SCEV that is greater than or equal |
| 935 | /// to (i.e. a "conservative over-approximation") of the value returend by |
| 936 | /// getBackedgeTakenCount. If such a value cannot be computed, it returns the |
| 937 | /// SCEVCouldNotCompute object. |
| 938 | const SCEV *getSymbolicMaxBackedgeTakenCount(const Loop *L) { |
| 939 | return getBackedgeTakenCount(L, Kind: SymbolicMaximum); |
| 940 | } |
| 941 | |
| 942 | /// Similar to getSymbolicMaxBackedgeTakenCount, except it will add a set of |
| 943 | /// SCEV predicates to Predicates that are required to be true in order for |
| 944 | /// the answer to be correct. Predicates can be checked with run-time |
| 945 | /// checks and can be used to perform loop versioning. |
| 946 | LLVM_ABI const SCEV *getPredicatedSymbolicMaxBackedgeTakenCount( |
| 947 | const Loop *L, SmallVectorImpl<const SCEVPredicate *> &Predicates); |
| 948 | |
| 949 | /// Return true if the backedge taken count is either the value returned by |
| 950 | /// getConstantMaxBackedgeTakenCount or zero. |
| 951 | LLVM_ABI bool isBackedgeTakenCountMaxOrZero(const Loop *L); |
| 952 | |
| 953 | /// Return true if the specified loop has an analyzable loop-invariant |
| 954 | /// backedge-taken count. |
| 955 | LLVM_ABI bool hasLoopInvariantBackedgeTakenCount(const Loop *L); |
| 956 | |
| 957 | // This method should be called by the client when it made any change that |
| 958 | // would invalidate SCEV's answers, and the client wants to remove all loop |
| 959 | // information held internally by ScalarEvolution. This is intended to be used |
| 960 | // when the alternative to forget a loop is too expensive (i.e. large loop |
| 961 | // bodies). |
| 962 | LLVM_ABI void forgetAllLoops(); |
| 963 | |
| 964 | /// This method should be called by the client when it has changed a loop in |
| 965 | /// a way that may effect ScalarEvolution's ability to compute a trip count, |
| 966 | /// or if the loop is deleted. This call is potentially expensive for large |
| 967 | /// loop bodies. |
| 968 | LLVM_ABI void forgetLoop(const Loop *L); |
| 969 | |
| 970 | // This method invokes forgetLoop for the outermost loop of the given loop |
| 971 | // \p L, making ScalarEvolution forget about all this subtree. This needs to |
| 972 | // be done whenever we make a transform that may affect the parameters of the |
| 973 | // outer loop, such as exit counts for branches. |
| 974 | LLVM_ABI void forgetTopmostLoop(const Loop *L); |
| 975 | |
| 976 | /// This method should be called by the client when it has changed a value |
| 977 | /// in a way that may effect its value, or which may disconnect it from a |
| 978 | /// def-use chain linking it to a loop. |
| 979 | LLVM_ABI void forgetValue(Value *V); |
| 980 | |
| 981 | /// Forget LCSSA phi node V of loop L to which a new predecessor was added, |
| 982 | /// such that it may no longer be trivial. |
| 983 | LLVM_ABI void forgetLcssaPhiWithNewPredecessor(Loop *L, PHINode *V); |
| 984 | |
| 985 | /// Called when the client has changed the disposition of values in |
| 986 | /// this loop. |
| 987 | /// |
| 988 | /// We don't have a way to invalidate per-loop dispositions. Clear and |
| 989 | /// recompute is simpler. |
| 990 | LLVM_ABI void forgetLoopDispositions(); |
| 991 | |
| 992 | /// Called when the client has changed the disposition of values in |
| 993 | /// a loop or block. |
| 994 | /// |
| 995 | /// We don't have a way to invalidate per-loop/per-block dispositions. Clear |
| 996 | /// and recompute is simpler. |
| 997 | LLVM_ABI void forgetBlockAndLoopDispositions(Value *V = nullptr); |
| 998 | |
| 999 | /// Determine the minimum number of zero bits that S is guaranteed to end in |
| 1000 | /// (at every loop iteration). It is, at the same time, the minimum number |
| 1001 | /// of times S is divisible by 2. For example, given {4,+,8} it returns 2. |
| 1002 | /// If S is guaranteed to be 0, it returns the bitwidth of S. |
| 1003 | LLVM_ABI uint32_t getMinTrailingZeros(const SCEV *S); |
| 1004 | |
| 1005 | /// Returns the max constant multiple of S. |
| 1006 | LLVM_ABI APInt getConstantMultiple(const SCEV *S); |
| 1007 | |
| 1008 | // Returns the max constant multiple of S. If S is exactly 0, return 1. |
| 1009 | LLVM_ABI APInt getNonZeroConstantMultiple(const SCEV *S); |
| 1010 | |
| 1011 | /// Determine the unsigned range for a particular SCEV. |
| 1012 | /// NOTE: This returns a copy of the reference returned by getRangeRef. |
| 1013 | ConstantRange getUnsignedRange(const SCEV *S) { |
| 1014 | return getRangeRef(S, Hint: HINT_RANGE_UNSIGNED); |
| 1015 | } |
| 1016 | |
| 1017 | /// Determine the min of the unsigned range for a particular SCEV. |
| 1018 | APInt getUnsignedRangeMin(const SCEV *S) { |
| 1019 | return getRangeRef(S, Hint: HINT_RANGE_UNSIGNED).getUnsignedMin(); |
| 1020 | } |
| 1021 | |
| 1022 | /// Determine the max of the unsigned range for a particular SCEV. |
| 1023 | APInt getUnsignedRangeMax(const SCEV *S) { |
| 1024 | return getRangeRef(S, Hint: HINT_RANGE_UNSIGNED).getUnsignedMax(); |
| 1025 | } |
| 1026 | |
| 1027 | /// Determine the signed range for a particular SCEV. |
| 1028 | /// NOTE: This returns a copy of the reference returned by getRangeRef. |
| 1029 | ConstantRange getSignedRange(const SCEV *S) { |
| 1030 | return getRangeRef(S, Hint: HINT_RANGE_SIGNED); |
| 1031 | } |
| 1032 | |
| 1033 | /// Determine the min of the signed range for a particular SCEV. |
| 1034 | APInt getSignedRangeMin(const SCEV *S) { |
| 1035 | return getRangeRef(S, Hint: HINT_RANGE_SIGNED).getSignedMin(); |
| 1036 | } |
| 1037 | |
| 1038 | /// Determine the max of the signed range for a particular SCEV. |
| 1039 | APInt getSignedRangeMax(const SCEV *S) { |
| 1040 | return getRangeRef(S, Hint: HINT_RANGE_SIGNED).getSignedMax(); |
| 1041 | } |
| 1042 | |
| 1043 | /// Test if the given expression is known to be negative. |
| 1044 | LLVM_ABI bool isKnownNegative(const SCEV *S); |
| 1045 | |
| 1046 | /// Test if the given expression is known to be positive. |
| 1047 | LLVM_ABI bool isKnownPositive(const SCEV *S); |
| 1048 | |
| 1049 | /// Test if the given expression is known to be non-negative. |
| 1050 | LLVM_ABI bool isKnownNonNegative(const SCEV *S); |
| 1051 | |
| 1052 | /// Test if the given expression is known to be non-positive. |
| 1053 | LLVM_ABI bool isKnownNonPositive(const SCEV *S); |
| 1054 | |
| 1055 | /// Test if the given expression is known to be non-zero. |
| 1056 | LLVM_ABI bool isKnownNonZero(const SCEV *S); |
| 1057 | |
| 1058 | /// Test if the given expression is known to be a power of 2. OrNegative |
| 1059 | /// allows matching negative power of 2s, and OrZero allows matching 0. |
| 1060 | LLVM_ABI bool isKnownToBeAPowerOfTwo(const SCEV *S, bool OrZero = false, |
| 1061 | bool OrNegative = false); |
| 1062 | |
| 1063 | /// Check that \p S is a multiple of \p M. When \p S is an AddRecExpr, \p S is |
| 1064 | /// a multiple of \p M if \p S starts with a multiple of \p M and at every |
| 1065 | /// iteration step \p S only adds multiples of \p M. \p Assumptions records |
| 1066 | /// the runtime predicates under which \p S is a multiple of \p M. |
| 1067 | LLVM_ABI bool |
| 1068 | isKnownMultipleOf(const SCEV *S, uint64_t M, |
| 1069 | SmallVectorImpl<const SCEVPredicate *> &Assumptions); |
| 1070 | |
| 1071 | /// Splits SCEV expression \p S into two SCEVs. One of them is obtained from |
| 1072 | /// \p S by substitution of all AddRec sub-expression related to loop \p L |
| 1073 | /// with initial value of that SCEV. The second is obtained from \p S by |
| 1074 | /// substitution of all AddRec sub-expressions related to loop \p L with post |
| 1075 | /// increment of this AddRec in the loop \p L. In both cases all other AddRec |
| 1076 | /// sub-expressions (not related to \p L) remain the same. |
| 1077 | /// If the \p S contains non-invariant unknown SCEV the function returns |
| 1078 | /// CouldNotCompute SCEV in both values of std::pair. |
| 1079 | /// For example, for SCEV S={0, +, 1}<L1> + {0, +, 1}<L2> and loop L=L1 |
| 1080 | /// the function returns pair: |
| 1081 | /// first = {0, +, 1}<L2> |
| 1082 | /// second = {1, +, 1}<L1> + {0, +, 1}<L2> |
| 1083 | /// We can see that for the first AddRec sub-expression it was replaced with |
| 1084 | /// 0 (initial value) for the first element and to {1, +, 1}<L1> (post |
| 1085 | /// increment value) for the second one. In both cases AddRec expression |
| 1086 | /// related to L2 remains the same. |
| 1087 | LLVM_ABI std::pair<const SCEV *, const SCEV *> |
| 1088 | SplitIntoInitAndPostInc(const Loop *L, const SCEV *S); |
| 1089 | |
| 1090 | /// We'd like to check the predicate on every iteration of the most dominated |
| 1091 | /// loop between loops used in LHS and RHS. |
| 1092 | /// To do this we use the following list of steps: |
| 1093 | /// 1. Collect set S all loops on which either LHS or RHS depend. |
| 1094 | /// 2. If S is non-empty |
| 1095 | /// a. Let PD be the element of S which is dominated by all other elements. |
| 1096 | /// b. Let E(LHS) be value of LHS on entry of PD. |
| 1097 | /// To get E(LHS), we should just take LHS and replace all AddRecs that are |
| 1098 | /// attached to PD on with their entry values. |
| 1099 | /// Define E(RHS) in the same way. |
| 1100 | /// c. Let B(LHS) be value of L on backedge of PD. |
| 1101 | /// To get B(LHS), we should just take LHS and replace all AddRecs that are |
| 1102 | /// attached to PD on with their backedge values. |
| 1103 | /// Define B(RHS) in the same way. |
| 1104 | /// d. Note that E(LHS) and E(RHS) are automatically available on entry of PD, |
| 1105 | /// so we can assert on that. |
| 1106 | /// e. Return true if isLoopEntryGuardedByCond(Pred, E(LHS), E(RHS)) && |
| 1107 | /// isLoopBackedgeGuardedByCond(Pred, B(LHS), B(RHS)) |
| 1108 | LLVM_ABI bool isKnownViaInduction(CmpPredicate Pred, const SCEV *LHS, |
| 1109 | const SCEV *RHS); |
| 1110 | |
| 1111 | /// Test if the given expression is known to satisfy the condition described |
| 1112 | /// by Pred, LHS, and RHS. |
| 1113 | LLVM_ABI bool isKnownPredicate(CmpPredicate Pred, const SCEV *LHS, |
| 1114 | const SCEV *RHS); |
| 1115 | |
| 1116 | /// Check whether the condition described by Pred, LHS, and RHS is true or |
| 1117 | /// false. If we know it, return the evaluation of this condition. If neither |
| 1118 | /// is proved, return std::nullopt. |
| 1119 | LLVM_ABI std::optional<bool> |
| 1120 | evaluatePredicate(CmpPredicate Pred, const SCEV *LHS, const SCEV *RHS); |
| 1121 | |
| 1122 | /// Test if the given expression is known to satisfy the condition described |
| 1123 | /// by Pred, LHS, and RHS in the given Context. |
| 1124 | LLVM_ABI bool isKnownPredicateAt(CmpPredicate Pred, const SCEV *LHS, |
| 1125 | const SCEV *RHS, const Instruction *CtxI); |
| 1126 | |
| 1127 | /// Check whether the condition described by Pred, LHS, and RHS is true or |
| 1128 | /// false in the given \p Context. If we know it, return the evaluation of |
| 1129 | /// this condition. If neither is proved, return std::nullopt. |
| 1130 | LLVM_ABI std::optional<bool> evaluatePredicateAt(CmpPredicate Pred, |
| 1131 | const SCEV *LHS, |
| 1132 | const SCEV *RHS, |
| 1133 | const Instruction *CtxI); |
| 1134 | |
| 1135 | /// Test if the condition described by Pred, LHS, RHS is known to be true on |
| 1136 | /// every iteration of the loop of the recurrency LHS. |
| 1137 | LLVM_ABI bool isKnownOnEveryIteration(CmpPredicate Pred, |
| 1138 | const SCEVAddRecExpr *LHS, |
| 1139 | const SCEV *RHS); |
| 1140 | |
| 1141 | /// Information about the number of loop iterations for which a loop exit's |
| 1142 | /// branch condition evaluates to the not-taken path. This is a temporary |
| 1143 | /// pair of exact and max expressions that are eventually summarized in |
| 1144 | /// ExitNotTakenInfo and BackedgeTakenInfo. |
| 1145 | struct ExitLimit { |
| 1146 | const SCEV *ExactNotTaken; // The exit is not taken exactly this many times |
| 1147 | const SCEV *ConstantMaxNotTaken; // The exit is not taken at most this many |
| 1148 | // times |
| 1149 | const SCEV *SymbolicMaxNotTaken; |
| 1150 | |
| 1151 | // Not taken either exactly ConstantMaxNotTaken or zero times |
| 1152 | bool MaxOrZero = false; |
| 1153 | |
| 1154 | /// A vector of predicate guards for this ExitLimit. The result is only |
| 1155 | /// valid if all of the predicates in \c Predicates evaluate to 'true' at |
| 1156 | /// run-time. |
| 1157 | SmallVector<const SCEVPredicate *, 4> Predicates; |
| 1158 | |
| 1159 | /// Construct either an exact exit limit from a constant, or an unknown |
| 1160 | /// one from a SCEVCouldNotCompute. No other types of SCEVs are allowed |
| 1161 | /// as arguments and asserts enforce that internally. |
| 1162 | /*implicit*/ LLVM_ABI ExitLimit(const SCEV *E); |
| 1163 | |
| 1164 | LLVM_ABI |
| 1165 | ExitLimit(const SCEV *E, const SCEV *ConstantMaxNotTaken, |
| 1166 | const SCEV *SymbolicMaxNotTaken, bool MaxOrZero, |
| 1167 | ArrayRef<ArrayRef<const SCEVPredicate *>> PredLists = {}); |
| 1168 | |
| 1169 | LLVM_ABI ExitLimit(const SCEV *E, const SCEV *ConstantMaxNotTaken, |
| 1170 | const SCEV *SymbolicMaxNotTaken, bool MaxOrZero, |
| 1171 | ArrayRef<const SCEVPredicate *> PredList); |
| 1172 | |
| 1173 | /// Test whether this ExitLimit contains any computed information, or |
| 1174 | /// whether it's all SCEVCouldNotCompute values. |
| 1175 | bool hasAnyInfo() const { |
| 1176 | return !isa<SCEVCouldNotCompute>(Val: ExactNotTaken) || |
| 1177 | !isa<SCEVCouldNotCompute>(Val: ConstantMaxNotTaken); |
| 1178 | } |
| 1179 | |
| 1180 | /// Test whether this ExitLimit contains all information. |
| 1181 | bool hasFullInfo() const { |
| 1182 | return !isa<SCEVCouldNotCompute>(Val: ExactNotTaken); |
| 1183 | } |
| 1184 | }; |
| 1185 | |
| 1186 | /// Compute the number of times the backedge of the specified loop will |
| 1187 | /// execute if its exit condition were a conditional branch of ExitCond. |
| 1188 | /// |
| 1189 | /// \p ControlsOnlyExit is true if ExitCond directly controls the only exit |
| 1190 | /// branch. In this case, we can assume that the loop exits only if the |
| 1191 | /// condition is true and can infer that failing to meet the condition prior |
| 1192 | /// to integer wraparound results in undefined behavior. |
| 1193 | /// |
| 1194 | /// If \p AllowPredicates is set, this call will try to use a minimal set of |
| 1195 | /// SCEV predicates in order to return an exact answer. |
| 1196 | LLVM_ABI ExitLimit computeExitLimitFromCond(const Loop *L, Value *ExitCond, |
| 1197 | bool ExitIfTrue, |
| 1198 | bool ControlsOnlyExit, |
| 1199 | bool AllowPredicates = false); |
| 1200 | |
| 1201 | /// A predicate is said to be monotonically increasing if may go from being |
| 1202 | /// false to being true as the loop iterates, but never the other way |
| 1203 | /// around. A predicate is said to be monotonically decreasing if may go |
| 1204 | /// from being true to being false as the loop iterates, but never the other |
| 1205 | /// way around. |
| 1206 | enum MonotonicPredicateType { |
| 1207 | MonotonicallyIncreasing, |
| 1208 | MonotonicallyDecreasing |
| 1209 | }; |
| 1210 | |
| 1211 | /// If, for all loop invariant X, the predicate "LHS `Pred` X" is |
| 1212 | /// monotonically increasing or decreasing, returns |
| 1213 | /// Some(MonotonicallyIncreasing) and Some(MonotonicallyDecreasing) |
| 1214 | /// respectively. If we could not prove either of these facts, returns |
| 1215 | /// std::nullopt. |
| 1216 | LLVM_ABI std::optional<MonotonicPredicateType> |
| 1217 | getMonotonicPredicateType(const SCEVAddRecExpr *LHS, |
| 1218 | ICmpInst::Predicate Pred); |
| 1219 | |
| 1220 | struct LoopInvariantPredicate { |
| 1221 | CmpPredicate Pred; |
| 1222 | const SCEV *LHS; |
| 1223 | const SCEV *RHS; |
| 1224 | |
| 1225 | LoopInvariantPredicate(CmpPredicate Pred, const SCEV *LHS, const SCEV *RHS) |
| 1226 | : Pred(Pred), LHS(LHS), RHS(RHS) {} |
| 1227 | }; |
| 1228 | /// If the result of the predicate LHS `Pred` RHS is loop invariant with |
| 1229 | /// respect to L, return a LoopInvariantPredicate with LHS and RHS being |
| 1230 | /// invariants, available at L's entry. Otherwise, return std::nullopt. |
| 1231 | LLVM_ABI std::optional<LoopInvariantPredicate> |
| 1232 | getLoopInvariantPredicate(CmpPredicate Pred, const SCEV *LHS, const SCEV *RHS, |
| 1233 | const Loop *L, const Instruction *CtxI = nullptr); |
| 1234 | |
| 1235 | /// If the result of the predicate LHS `Pred` RHS is loop invariant with |
| 1236 | /// respect to L at given Context during at least first MaxIter iterations, |
| 1237 | /// return a LoopInvariantPredicate with LHS and RHS being invariants, |
| 1238 | /// available at L's entry. Otherwise, return std::nullopt. The predicate |
| 1239 | /// should be the loop's exit condition. |
| 1240 | LLVM_ABI std::optional<LoopInvariantPredicate> |
| 1241 | getLoopInvariantExitCondDuringFirstIterations(CmpPredicate Pred, |
| 1242 | const SCEV *LHS, |
| 1243 | const SCEV *RHS, const Loop *L, |
| 1244 | const Instruction *CtxI, |
| 1245 | const SCEV *MaxIter); |
| 1246 | |
| 1247 | LLVM_ABI std::optional<LoopInvariantPredicate> |
| 1248 | getLoopInvariantExitCondDuringFirstIterationsImpl( |
| 1249 | CmpPredicate Pred, const SCEV *LHS, const SCEV *RHS, const Loop *L, |
| 1250 | const Instruction *CtxI, const SCEV *MaxIter); |
| 1251 | |
| 1252 | /// Simplify LHS and RHS in a comparison with predicate Pred. Return true |
| 1253 | /// iff any changes were made. If the operands are provably equal or |
| 1254 | /// unequal, LHS and RHS are set to the same value and Pred is set to either |
| 1255 | /// ICMP_EQ or ICMP_NE. |
| 1256 | LLVM_ABI bool SimplifyICmpOperands(CmpPredicate &Pred, const SCEV *&LHS, |
| 1257 | const SCEV *&RHS, unsigned Depth = 0); |
| 1258 | |
| 1259 | /// Return the "disposition" of the given SCEV with respect to the given |
| 1260 | /// loop. |
| 1261 | LLVM_ABI LoopDisposition getLoopDisposition(const SCEV *S, const Loop *L); |
| 1262 | |
| 1263 | /// Return true if the value of the given SCEV is unchanging in the |
| 1264 | /// specified loop. |
| 1265 | LLVM_ABI bool isLoopInvariant(const SCEV *S, const Loop *L); |
| 1266 | |
| 1267 | /// Determine if the SCEV can be evaluated at loop's entry. It is true if it |
| 1268 | /// doesn't depend on a SCEVUnknown of an instruction which is dominated by |
| 1269 | /// the header of loop L. |
| 1270 | LLVM_ABI bool isAvailableAtLoopEntry(const SCEV *S, const Loop *L); |
| 1271 | |
| 1272 | /// Return true if the given SCEV changes value in a known way in the |
| 1273 | /// specified loop. This property being true implies that the value is |
| 1274 | /// variant in the loop AND that we can emit an expression to compute the |
| 1275 | /// value of the expression at any particular loop iteration. |
| 1276 | LLVM_ABI bool hasComputableLoopEvolution(const SCEV *S, const Loop *L); |
| 1277 | |
| 1278 | /// Return the "disposition" of the given SCEV with respect to the given |
| 1279 | /// block. |
| 1280 | LLVM_ABI BlockDisposition getBlockDisposition(const SCEV *S, |
| 1281 | const BasicBlock *BB); |
| 1282 | |
| 1283 | /// Return true if elements that makes up the given SCEV dominate the |
| 1284 | /// specified basic block. |
| 1285 | LLVM_ABI bool dominates(const SCEV *S, const BasicBlock *BB); |
| 1286 | |
| 1287 | /// Return true if elements that makes up the given SCEV properly dominate |
| 1288 | /// the specified basic block. |
| 1289 | LLVM_ABI bool properlyDominates(const SCEV *S, const BasicBlock *BB); |
| 1290 | |
| 1291 | /// Test whether the given SCEV has Op as a direct or indirect operand. |
| 1292 | LLVM_ABI bool hasOperand(const SCEV *S, const SCEV *Op) const; |
| 1293 | |
| 1294 | /// Return the size of an element read or written by Inst. |
| 1295 | LLVM_ABI const SCEV *getElementSize(Instruction *Inst); |
| 1296 | |
| 1297 | LLVM_ABI void print(raw_ostream &OS) const; |
| 1298 | LLVM_ABI void verify() const; |
| 1299 | LLVM_ABI bool invalidate(Function &F, const PreservedAnalyses &PA, |
| 1300 | FunctionAnalysisManager::Invalidator &Inv); |
| 1301 | |
| 1302 | /// Return the DataLayout associated with the module this SCEV instance is |
| 1303 | /// operating on. |
| 1304 | const DataLayout &getDataLayout() const { return DL; } |
| 1305 | |
| 1306 | LLVM_ABI const SCEVPredicate *getEqualPredicate(const SCEV *LHS, |
| 1307 | const SCEV *RHS); |
| 1308 | LLVM_ABI const SCEVPredicate *getComparePredicate(ICmpInst::Predicate Pred, |
| 1309 | const SCEV *LHS, |
| 1310 | const SCEV *RHS); |
| 1311 | |
| 1312 | LLVM_ABI const SCEVPredicate * |
| 1313 | getWrapPredicate(const SCEVAddRecExpr *AR, |
| 1314 | SCEVWrapPredicate::IncrementWrapFlags AddedFlags); |
| 1315 | |
| 1316 | /// Re-writes the SCEV according to the Predicates in \p A. |
| 1317 | LLVM_ABI const SCEV *rewriteUsingPredicate(const SCEV *S, const Loop *L, |
| 1318 | const SCEVPredicate &A); |
| 1319 | /// Tries to convert the \p S expression to an AddRec expression, |
| 1320 | /// adding additional predicates to \p Preds as required. |
| 1321 | LLVM_ABI const SCEVAddRecExpr *convertSCEVToAddRecWithPredicates( |
| 1322 | const SCEV *S, const Loop *L, |
| 1323 | SmallVectorImpl<const SCEVPredicate *> &Preds); |
| 1324 | |
| 1325 | /// Compute \p LHS - \p RHS and returns the result as an APInt if it is a |
| 1326 | /// constant, and std::nullopt if it isn't. |
| 1327 | /// |
| 1328 | /// This is intended to be a cheaper version of getMinusSCEV. We can be |
| 1329 | /// frugal here since we just bail out of actually constructing and |
| 1330 | /// canonicalizing an expression in the cases where the result isn't going |
| 1331 | /// to be a constant. |
| 1332 | LLVM_ABI std::optional<APInt> computeConstantDifference(const SCEV *LHS, |
| 1333 | const SCEV *RHS); |
| 1334 | |
| 1335 | /// Update no-wrap flags of an AddRec. This may drop the cached info about |
| 1336 | /// this AddRec (such as range info) in case if new flags may potentially |
| 1337 | /// sharpen it. |
| 1338 | LLVM_ABI void setNoWrapFlags(SCEVAddRecExpr *AddRec, SCEV::NoWrapFlags Flags); |
| 1339 | |
| 1340 | class LoopGuards { |
| 1341 | DenseMap<const SCEV *, const SCEV *> RewriteMap; |
| 1342 | bool PreserveNUW = false; |
| 1343 | bool PreserveNSW = false; |
| 1344 | ScalarEvolution &SE; |
| 1345 | |
| 1346 | LoopGuards(ScalarEvolution &SE) : SE(SE) {} |
| 1347 | |
| 1348 | /// Recursively collect loop guards in \p Guards, starting from |
| 1349 | /// block \p Block with predecessor \p Pred. The intended starting point |
| 1350 | /// is to collect from a loop header and its predecessor. |
| 1351 | static void |
| 1352 | collectFromBlock(ScalarEvolution &SE, ScalarEvolution::LoopGuards &Guards, |
| 1353 | const BasicBlock *Block, const BasicBlock *Pred, |
| 1354 | SmallPtrSetImpl<const BasicBlock *> &VisitedBlocks, |
| 1355 | unsigned Depth = 0); |
| 1356 | |
| 1357 | /// Collect loop guards in \p Guards, starting from PHINode \p |
| 1358 | /// Phi, by calling \p collectFromBlock on the incoming blocks of |
| 1359 | /// \Phi and trying to merge the found constraints into a single |
| 1360 | /// combined one for \p Phi. |
| 1361 | static void collectFromPHI( |
| 1362 | ScalarEvolution &SE, ScalarEvolution::LoopGuards &Guards, |
| 1363 | const PHINode &Phi, SmallPtrSetImpl<const BasicBlock *> &VisitedBlocks, |
| 1364 | SmallDenseMap<const BasicBlock *, LoopGuards> &IncomingGuards, |
| 1365 | unsigned Depth); |
| 1366 | |
| 1367 | public: |
| 1368 | /// Collect rewrite map for loop guards for loop \p L, together with flags |
| 1369 | /// indicating if NUW and NSW can be preserved during rewriting. |
| 1370 | LLVM_ABI static LoopGuards collect(const Loop *L, ScalarEvolution &SE); |
| 1371 | |
| 1372 | /// Try to apply the collected loop guards to \p Expr. |
| 1373 | LLVM_ABI const SCEV *rewrite(const SCEV *Expr) const; |
| 1374 | }; |
| 1375 | |
| 1376 | /// Try to apply information from loop guards for \p L to \p Expr. |
| 1377 | LLVM_ABI const SCEV *applyLoopGuards(const SCEV *Expr, const Loop *L); |
| 1378 | LLVM_ABI const SCEV *applyLoopGuards(const SCEV *Expr, |
| 1379 | const LoopGuards &Guards); |
| 1380 | |
| 1381 | /// Return true if the loop has no abnormal exits. That is, if the loop |
| 1382 | /// is not infinite, it must exit through an explicit edge in the CFG. |
| 1383 | /// (As opposed to either a) throwing out of the function or b) entering a |
| 1384 | /// well defined infinite loop in some callee.) |
| 1385 | bool loopHasNoAbnormalExits(const Loop *L) { |
| 1386 | return getLoopProperties(L).HasNoAbnormalExits; |
| 1387 | } |
| 1388 | |
| 1389 | /// Return true if this loop is finite by assumption. That is, |
| 1390 | /// to be infinite, it must also be undefined. |
| 1391 | LLVM_ABI bool loopIsFiniteByAssumption(const Loop *L); |
| 1392 | |
| 1393 | /// Return the set of Values that, if poison, will definitively result in S |
| 1394 | /// being poison as well. The returned set may be incomplete, i.e. there can |
| 1395 | /// be additional Values that also result in S being poison. |
| 1396 | LLVM_ABI void |
| 1397 | getPoisonGeneratingValues(SmallPtrSetImpl<const Value *> &Result, |
| 1398 | const SCEV *S); |
| 1399 | |
| 1400 | /// Check whether it is poison-safe to represent the expression S using the |
| 1401 | /// instruction I. If such a replacement is performed, the poison flags of |
| 1402 | /// instructions in DropPoisonGeneratingInsts must be dropped. |
| 1403 | LLVM_ABI bool canReuseInstruction( |
| 1404 | const SCEV *S, Instruction *I, |
| 1405 | SmallVectorImpl<Instruction *> &DropPoisonGeneratingInsts); |
| 1406 | |
| 1407 | class FoldID { |
| 1408 | const SCEV *Op = nullptr; |
| 1409 | const Type *Ty = nullptr; |
| 1410 | unsigned short C; |
| 1411 | |
| 1412 | public: |
| 1413 | FoldID(SCEVTypes C, const SCEV *Op, const Type *Ty) : Op(Op), Ty(Ty), C(C) { |
| 1414 | assert(Op); |
| 1415 | assert(Ty); |
| 1416 | } |
| 1417 | |
| 1418 | FoldID(unsigned short C) : C(C) {} |
| 1419 | |
| 1420 | unsigned computeHash() const { |
| 1421 | return detail::combineHashValue( |
| 1422 | a: C, b: detail::combineHashValue(a: reinterpret_cast<uintptr_t>(Op), |
| 1423 | b: reinterpret_cast<uintptr_t>(Ty))); |
| 1424 | } |
| 1425 | |
| 1426 | bool operator==(const FoldID &RHS) const { |
| 1427 | return std::tie(args: Op, args: Ty, args: C) == std::tie(args: RHS.Op, args: RHS.Ty, args: RHS.C); |
| 1428 | } |
| 1429 | }; |
| 1430 | |
| 1431 | private: |
| 1432 | /// A CallbackVH to arrange for ScalarEvolution to be notified whenever a |
| 1433 | /// Value is deleted. |
| 1434 | class LLVM_ABI SCEVCallbackVH final : public CallbackVH { |
| 1435 | ScalarEvolution *SE; |
| 1436 | |
| 1437 | void deleted() override; |
| 1438 | void allUsesReplacedWith(Value *New) override; |
| 1439 | |
| 1440 | public: |
| 1441 | SCEVCallbackVH(Value *V, ScalarEvolution *SE = nullptr); |
| 1442 | }; |
| 1443 | |
| 1444 | friend class SCEVCallbackVH; |
| 1445 | friend class SCEVExpander; |
| 1446 | friend class SCEVUnknown; |
| 1447 | |
| 1448 | /// The function we are analyzing. |
| 1449 | Function &F; |
| 1450 | |
| 1451 | /// Data layout of the module. |
| 1452 | const DataLayout &DL; |
| 1453 | |
| 1454 | /// Does the module have any calls to the llvm.experimental.guard intrinsic |
| 1455 | /// at all? If this is false, we avoid doing work that will only help if |
| 1456 | /// thare are guards present in the IR. |
| 1457 | bool HasGuards; |
| 1458 | |
| 1459 | /// The target library information for the target we are targeting. |
| 1460 | TargetLibraryInfo &TLI; |
| 1461 | |
| 1462 | /// The tracker for \@llvm.assume intrinsics in this function. |
| 1463 | AssumptionCache &AC; |
| 1464 | |
| 1465 | /// The dominator tree. |
| 1466 | DominatorTree &DT; |
| 1467 | |
| 1468 | /// The loop information for the function we are currently analyzing. |
| 1469 | LoopInfo &LI; |
| 1470 | |
| 1471 | /// This SCEV is used to represent unknown trip counts and things. |
| 1472 | std::unique_ptr<SCEVCouldNotCompute> CouldNotCompute; |
| 1473 | |
| 1474 | /// The type for HasRecMap. |
| 1475 | using HasRecMapType = DenseMap<const SCEV *, bool>; |
| 1476 | |
| 1477 | /// This is a cache to record whether a SCEV contains any scAddRecExpr. |
| 1478 | HasRecMapType HasRecMap; |
| 1479 | |
| 1480 | /// The type for ExprValueMap. |
| 1481 | using ValueSetVector = SmallSetVector<Value *, 4>; |
| 1482 | using ExprValueMapType = DenseMap<const SCEV *, ValueSetVector>; |
| 1483 | |
| 1484 | /// ExprValueMap -- This map records the original values from which |
| 1485 | /// the SCEV expr is generated from. |
| 1486 | ExprValueMapType ExprValueMap; |
| 1487 | |
| 1488 | /// The type for ValueExprMap. |
| 1489 | using ValueExprMapType = |
| 1490 | DenseMap<SCEVCallbackVH, const SCEV *, DenseMapInfo<Value *>>; |
| 1491 | |
| 1492 | /// This is a cache of the values we have analyzed so far. |
| 1493 | ValueExprMapType ValueExprMap; |
| 1494 | |
| 1495 | /// This is a cache for expressions that got folded to a different existing |
| 1496 | /// SCEV. |
| 1497 | DenseMap<FoldID, const SCEV *> FoldCache; |
| 1498 | DenseMap<const SCEV *, SmallVector<FoldID, 2>> FoldCacheUser; |
| 1499 | |
| 1500 | /// Mark predicate values currently being processed by isImpliedCond. |
| 1501 | SmallPtrSet<const Value *, 6> PendingLoopPredicates; |
| 1502 | |
| 1503 | /// Mark SCEVUnknown Phis currently being processed by getRangeRef. |
| 1504 | SmallPtrSet<const PHINode *, 6> PendingPhiRanges; |
| 1505 | |
| 1506 | /// Mark SCEVUnknown Phis currently being processed by getRangeRefIter. |
| 1507 | SmallPtrSet<const PHINode *, 6> PendingPhiRangesIter; |
| 1508 | |
| 1509 | // Mark SCEVUnknown Phis currently being processed by isImpliedViaMerge. |
| 1510 | SmallPtrSet<const PHINode *, 6> PendingMerges; |
| 1511 | |
| 1512 | /// Set to true by isLoopBackedgeGuardedByCond when we're walking the set of |
| 1513 | /// conditions dominating the backedge of a loop. |
| 1514 | bool WalkingBEDominatingConds = false; |
| 1515 | |
| 1516 | /// Set to true by isKnownPredicateViaSplitting when we're trying to prove a |
| 1517 | /// predicate by splitting it into a set of independent predicates. |
| 1518 | bool ProvingSplitPredicate = false; |
| 1519 | |
| 1520 | /// Memoized values for the getConstantMultiple |
| 1521 | DenseMap<const SCEV *, APInt> ConstantMultipleCache; |
| 1522 | |
| 1523 | /// Return the Value set from which the SCEV expr is generated. |
| 1524 | ArrayRef<Value *> getSCEVValues(const SCEV *S); |
| 1525 | |
| 1526 | /// Private helper method for the getConstantMultiple method. |
| 1527 | APInt getConstantMultipleImpl(const SCEV *S); |
| 1528 | |
| 1529 | /// Information about the number of times a particular loop exit may be |
| 1530 | /// reached before exiting the loop. |
| 1531 | struct ExitNotTakenInfo { |
| 1532 | PoisoningVH<BasicBlock> ExitingBlock; |
| 1533 | const SCEV *ExactNotTaken; |
| 1534 | const SCEV *ConstantMaxNotTaken; |
| 1535 | const SCEV *SymbolicMaxNotTaken; |
| 1536 | SmallVector<const SCEVPredicate *, 4> Predicates; |
| 1537 | |
| 1538 | explicit ExitNotTakenInfo(PoisoningVH<BasicBlock> ExitingBlock, |
| 1539 | const SCEV *ExactNotTaken, |
| 1540 | const SCEV *ConstantMaxNotTaken, |
| 1541 | const SCEV *SymbolicMaxNotTaken, |
| 1542 | ArrayRef<const SCEVPredicate *> Predicates) |
| 1543 | : ExitingBlock(ExitingBlock), ExactNotTaken(ExactNotTaken), |
| 1544 | ConstantMaxNotTaken(ConstantMaxNotTaken), |
| 1545 | SymbolicMaxNotTaken(SymbolicMaxNotTaken), Predicates(Predicates) {} |
| 1546 | |
| 1547 | bool hasAlwaysTruePredicate() const { |
| 1548 | return Predicates.empty(); |
| 1549 | } |
| 1550 | }; |
| 1551 | |
| 1552 | /// Information about the backedge-taken count of a loop. This currently |
| 1553 | /// includes an exact count and a maximum count. |
| 1554 | /// |
| 1555 | class BackedgeTakenInfo { |
| 1556 | friend class ScalarEvolution; |
| 1557 | |
| 1558 | /// A list of computable exits and their not-taken counts. Loops almost |
| 1559 | /// never have more than one computable exit. |
| 1560 | SmallVector<ExitNotTakenInfo, 1> ExitNotTaken; |
| 1561 | |
| 1562 | /// Expression indicating the least constant maximum backedge-taken count of |
| 1563 | /// the loop that is known, or a SCEVCouldNotCompute. This expression is |
| 1564 | /// only valid if the predicates associated with all loop exits are true. |
| 1565 | const SCEV *ConstantMax = nullptr; |
| 1566 | |
| 1567 | /// Indicating if \c ExitNotTaken has an element for every exiting block in |
| 1568 | /// the loop. |
| 1569 | bool IsComplete = false; |
| 1570 | |
| 1571 | /// Expression indicating the least maximum backedge-taken count of the loop |
| 1572 | /// that is known, or a SCEVCouldNotCompute. Lazily computed on first query. |
| 1573 | const SCEV *SymbolicMax = nullptr; |
| 1574 | |
| 1575 | /// True iff the backedge is taken either exactly Max or zero times. |
| 1576 | bool MaxOrZero = false; |
| 1577 | |
| 1578 | bool isComplete() const { return IsComplete; } |
| 1579 | const SCEV *getConstantMax() const { return ConstantMax; } |
| 1580 | |
| 1581 | LLVM_ABI const ExitNotTakenInfo *getExitNotTaken( |
| 1582 | const BasicBlock *ExitingBlock, |
| 1583 | SmallVectorImpl<const SCEVPredicate *> *Predicates = nullptr) const; |
| 1584 | |
| 1585 | public: |
| 1586 | BackedgeTakenInfo() = default; |
| 1587 | BackedgeTakenInfo(BackedgeTakenInfo &&) = default; |
| 1588 | BackedgeTakenInfo &operator=(BackedgeTakenInfo &&) = default; |
| 1589 | |
| 1590 | using EdgeExitInfo = std::pair<BasicBlock *, ExitLimit>; |
| 1591 | |
| 1592 | /// Initialize BackedgeTakenInfo from a list of exact exit counts. |
| 1593 | LLVM_ABI BackedgeTakenInfo(ArrayRef<EdgeExitInfo> ExitCounts, |
| 1594 | bool IsComplete, const SCEV *ConstantMax, |
| 1595 | bool MaxOrZero); |
| 1596 | |
| 1597 | /// Test whether this BackedgeTakenInfo contains any computed information, |
| 1598 | /// or whether it's all SCEVCouldNotCompute values. |
| 1599 | bool hasAnyInfo() const { |
| 1600 | return !ExitNotTaken.empty() || |
| 1601 | !isa<SCEVCouldNotCompute>(Val: getConstantMax()); |
| 1602 | } |
| 1603 | |
| 1604 | /// Test whether this BackedgeTakenInfo contains complete information. |
| 1605 | bool hasFullInfo() const { return isComplete(); } |
| 1606 | |
| 1607 | /// Return an expression indicating the exact *backedge-taken* |
| 1608 | /// count of the loop if it is known or SCEVCouldNotCompute |
| 1609 | /// otherwise. If execution makes it to the backedge on every |
| 1610 | /// iteration (i.e. there are no abnormal exists like exception |
| 1611 | /// throws and thread exits) then this is the number of times the |
| 1612 | /// loop header will execute minus one. |
| 1613 | /// |
| 1614 | /// If the SCEV predicate associated with the answer can be different |
| 1615 | /// from AlwaysTrue, we must add a (non null) Predicates argument. |
| 1616 | /// The SCEV predicate associated with the answer will be added to |
| 1617 | /// Predicates. A run-time check needs to be emitted for the SCEV |
| 1618 | /// predicate in order for the answer to be valid. |
| 1619 | /// |
| 1620 | /// Note that we should always know if we need to pass a predicate |
| 1621 | /// argument or not from the way the ExitCounts vector was computed. |
| 1622 | /// If we allowed SCEV predicates to be generated when populating this |
| 1623 | /// vector, this information can contain them and therefore a |
| 1624 | /// SCEVPredicate argument should be added to getExact. |
| 1625 | LLVM_ABI const SCEV *getExact( |
| 1626 | const Loop *L, ScalarEvolution *SE, |
| 1627 | SmallVectorImpl<const SCEVPredicate *> *Predicates = nullptr) const; |
| 1628 | |
| 1629 | /// Return the number of times this loop exit may fall through to the back |
| 1630 | /// edge, or SCEVCouldNotCompute. The loop is guaranteed not to exit via |
| 1631 | /// this block before this number of iterations, but may exit via another |
| 1632 | /// block. If \p Predicates is null the function returns CouldNotCompute if |
| 1633 | /// predicates are required, otherwise it fills in the required predicates. |
| 1634 | const SCEV *getExact( |
| 1635 | const BasicBlock *ExitingBlock, ScalarEvolution *SE, |
| 1636 | SmallVectorImpl<const SCEVPredicate *> *Predicates = nullptr) const { |
| 1637 | if (auto *ENT = getExitNotTaken(ExitingBlock, Predicates)) |
| 1638 | return ENT->ExactNotTaken; |
| 1639 | else |
| 1640 | return SE->getCouldNotCompute(); |
| 1641 | } |
| 1642 | |
| 1643 | /// Get the constant max backedge taken count for the loop. |
| 1644 | LLVM_ABI const SCEV *getConstantMax( |
| 1645 | ScalarEvolution *SE, |
| 1646 | SmallVectorImpl<const SCEVPredicate *> *Predicates = nullptr) const; |
| 1647 | |
| 1648 | /// Get the constant max backedge taken count for the particular loop exit. |
| 1649 | const SCEV *getConstantMax( |
| 1650 | const BasicBlock *ExitingBlock, ScalarEvolution *SE, |
| 1651 | SmallVectorImpl<const SCEVPredicate *> *Predicates = nullptr) const { |
| 1652 | if (auto *ENT = getExitNotTaken(ExitingBlock, Predicates)) |
| 1653 | return ENT->ConstantMaxNotTaken; |
| 1654 | else |
| 1655 | return SE->getCouldNotCompute(); |
| 1656 | } |
| 1657 | |
| 1658 | /// Get the symbolic max backedge taken count for the loop. |
| 1659 | LLVM_ABI const SCEV *getSymbolicMax( |
| 1660 | const Loop *L, ScalarEvolution *SE, |
| 1661 | SmallVectorImpl<const SCEVPredicate *> *Predicates = nullptr); |
| 1662 | |
| 1663 | /// Get the symbolic max backedge taken count for the particular loop exit. |
| 1664 | const SCEV *getSymbolicMax( |
| 1665 | const BasicBlock *ExitingBlock, ScalarEvolution *SE, |
| 1666 | SmallVectorImpl<const SCEVPredicate *> *Predicates = nullptr) const { |
| 1667 | if (auto *ENT = getExitNotTaken(ExitingBlock, Predicates)) |
| 1668 | return ENT->SymbolicMaxNotTaken; |
| 1669 | else |
| 1670 | return SE->getCouldNotCompute(); |
| 1671 | } |
| 1672 | |
| 1673 | /// Return true if the number of times this backedge is taken is either the |
| 1674 | /// value returned by getConstantMax or zero. |
| 1675 | LLVM_ABI bool isConstantMaxOrZero(ScalarEvolution *SE) const; |
| 1676 | }; |
| 1677 | |
| 1678 | /// Cache the backedge-taken count of the loops for this function as they |
| 1679 | /// are computed. |
| 1680 | DenseMap<const Loop *, BackedgeTakenInfo> BackedgeTakenCounts; |
| 1681 | |
| 1682 | /// Cache the predicated backedge-taken count of the loops for this |
| 1683 | /// function as they are computed. |
| 1684 | DenseMap<const Loop *, BackedgeTakenInfo> PredicatedBackedgeTakenCounts; |
| 1685 | |
| 1686 | /// Loops whose backedge taken counts directly use this non-constant SCEV. |
| 1687 | DenseMap<const SCEV *, SmallPtrSet<PointerIntPair<const Loop *, 1, bool>, 4>> |
| 1688 | BECountUsers; |
| 1689 | |
| 1690 | /// This map contains entries for all of the PHI instructions that we |
| 1691 | /// attempt to compute constant evolutions for. This allows us to avoid |
| 1692 | /// potentially expensive recomputation of these properties. An instruction |
| 1693 | /// maps to null if we are unable to compute its exit value. |
| 1694 | DenseMap<PHINode *, Constant *> ConstantEvolutionLoopExitValue; |
| 1695 | |
| 1696 | /// This map contains entries for all the expressions that we attempt to |
| 1697 | /// compute getSCEVAtScope information for, which can be expensive in |
| 1698 | /// extreme cases. |
| 1699 | DenseMap<const SCEV *, SmallVector<std::pair<const Loop *, const SCEV *>, 2>> |
| 1700 | ValuesAtScopes; |
| 1701 | |
| 1702 | /// Reverse map for invalidation purposes: Stores of which SCEV and which |
| 1703 | /// loop this is the value-at-scope of. |
| 1704 | DenseMap<const SCEV *, SmallVector<std::pair<const Loop *, const SCEV *>, 2>> |
| 1705 | ValuesAtScopesUsers; |
| 1706 | |
| 1707 | /// Memoized computeLoopDisposition results. |
| 1708 | DenseMap<const SCEV *, |
| 1709 | SmallVector<PointerIntPair<const Loop *, 2, LoopDisposition>, 2>> |
| 1710 | LoopDispositions; |
| 1711 | |
| 1712 | struct LoopProperties { |
| 1713 | /// Set to true if the loop contains no instruction that can abnormally exit |
| 1714 | /// the loop (i.e. via throwing an exception, by terminating the thread |
| 1715 | /// cleanly or by infinite looping in a called function). Strictly |
| 1716 | /// speaking, the last one is not leaving the loop, but is identical to |
| 1717 | /// leaving the loop for reasoning about undefined behavior. |
| 1718 | bool HasNoAbnormalExits; |
| 1719 | |
| 1720 | /// Set to true if the loop contains no instruction that can have side |
| 1721 | /// effects (i.e. via throwing an exception, volatile or atomic access). |
| 1722 | bool HasNoSideEffects; |
| 1723 | }; |
| 1724 | |
| 1725 | /// Cache for \c getLoopProperties. |
| 1726 | DenseMap<const Loop *, LoopProperties> LoopPropertiesCache; |
| 1727 | |
| 1728 | /// Return a \c LoopProperties instance for \p L, creating one if necessary. |
| 1729 | LLVM_ABI LoopProperties getLoopProperties(const Loop *L); |
| 1730 | |
| 1731 | bool loopHasNoSideEffects(const Loop *L) { |
| 1732 | return getLoopProperties(L).HasNoSideEffects; |
| 1733 | } |
| 1734 | |
| 1735 | /// Compute a LoopDisposition value. |
| 1736 | LoopDisposition computeLoopDisposition(const SCEV *S, const Loop *L); |
| 1737 | |
| 1738 | /// Memoized computeBlockDisposition results. |
| 1739 | DenseMap< |
| 1740 | const SCEV *, |
| 1741 | SmallVector<PointerIntPair<const BasicBlock *, 2, BlockDisposition>, 2>> |
| 1742 | BlockDispositions; |
| 1743 | |
| 1744 | /// Compute a BlockDisposition value. |
| 1745 | BlockDisposition computeBlockDisposition(const SCEV *S, const BasicBlock *BB); |
| 1746 | |
| 1747 | /// Stores all SCEV that use a given SCEV as its direct operand. |
| 1748 | DenseMap<const SCEV *, SmallPtrSet<const SCEV *, 8> > SCEVUsers; |
| 1749 | |
| 1750 | /// Memoized results from getRange |
| 1751 | DenseMap<const SCEV *, ConstantRange> UnsignedRanges; |
| 1752 | |
| 1753 | /// Memoized results from getRange |
| 1754 | DenseMap<const SCEV *, ConstantRange> SignedRanges; |
| 1755 | |
| 1756 | /// Used to parameterize getRange |
| 1757 | enum RangeSignHint { HINT_RANGE_UNSIGNED, HINT_RANGE_SIGNED }; |
| 1758 | |
| 1759 | /// Set the memoized range for the given SCEV. |
| 1760 | const ConstantRange &setRange(const SCEV *S, RangeSignHint Hint, |
| 1761 | ConstantRange CR) { |
| 1762 | DenseMap<const SCEV *, ConstantRange> &Cache = |
| 1763 | Hint == HINT_RANGE_UNSIGNED ? UnsignedRanges : SignedRanges; |
| 1764 | |
| 1765 | auto Pair = Cache.insert_or_assign(Key: S, Val: std::move(CR)); |
| 1766 | return Pair.first->second; |
| 1767 | } |
| 1768 | |
| 1769 | /// Determine the range for a particular SCEV. |
| 1770 | /// NOTE: This returns a reference to an entry in a cache. It must be |
| 1771 | /// copied if its needed for longer. |
| 1772 | LLVM_ABI const ConstantRange &getRangeRef(const SCEV *S, RangeSignHint Hint, |
| 1773 | unsigned Depth = 0); |
| 1774 | |
| 1775 | /// Determine the range for a particular SCEV, but evaluates ranges for |
| 1776 | /// operands iteratively first. |
| 1777 | const ConstantRange &getRangeRefIter(const SCEV *S, RangeSignHint Hint); |
| 1778 | |
| 1779 | /// Determines the range for the affine SCEVAddRecExpr {\p Start,+,\p Step}. |
| 1780 | /// Helper for \c getRange. |
| 1781 | ConstantRange getRangeForAffineAR(const SCEV *Start, const SCEV *Step, |
| 1782 | const APInt &MaxBECount); |
| 1783 | |
| 1784 | /// Determines the range for the affine non-self-wrapping SCEVAddRecExpr {\p |
| 1785 | /// Start,+,\p Step}<nw>. |
| 1786 | ConstantRange getRangeForAffineNoSelfWrappingAR(const SCEVAddRecExpr *AddRec, |
| 1787 | const SCEV *MaxBECount, |
| 1788 | unsigned BitWidth, |
| 1789 | RangeSignHint SignHint); |
| 1790 | |
| 1791 | /// Try to compute a range for the affine SCEVAddRecExpr {\p Start,+,\p |
| 1792 | /// Step} by "factoring out" a ternary expression from the add recurrence. |
| 1793 | /// Helper called by \c getRange. |
| 1794 | ConstantRange getRangeViaFactoring(const SCEV *Start, const SCEV *Step, |
| 1795 | const APInt &MaxBECount); |
| 1796 | |
| 1797 | /// If the unknown expression U corresponds to a simple recurrence, return |
| 1798 | /// a constant range which represents the entire recurrence. Note that |
| 1799 | /// *add* recurrences with loop invariant steps aren't represented by |
| 1800 | /// SCEVUnknowns and thus don't use this mechanism. |
| 1801 | ConstantRange getRangeForUnknownRecurrence(const SCEVUnknown *U); |
| 1802 | |
| 1803 | /// We know that there is no SCEV for the specified value. Analyze the |
| 1804 | /// expression recursively. |
| 1805 | const SCEV *createSCEV(Value *V); |
| 1806 | |
| 1807 | /// We know that there is no SCEV for the specified value. Create a new SCEV |
| 1808 | /// for \p V iteratively. |
| 1809 | const SCEV *createSCEVIter(Value *V); |
| 1810 | /// Collect operands of \p V for which SCEV expressions should be constructed |
| 1811 | /// first. Returns a SCEV directly if it can be constructed trivially for \p |
| 1812 | /// V. |
| 1813 | const SCEV *getOperandsToCreate(Value *V, SmallVectorImpl<Value *> &Ops); |
| 1814 | |
| 1815 | /// Returns SCEV for the first operand of a phi if all phi operands have |
| 1816 | /// identical opcodes and operands. |
| 1817 | const SCEV *createNodeForPHIWithIdenticalOperands(PHINode *PN); |
| 1818 | |
| 1819 | /// Provide the special handling we need to analyze PHI SCEVs. |
| 1820 | const SCEV *createNodeForPHI(PHINode *PN); |
| 1821 | |
| 1822 | /// Helper function called from createNodeForPHI. |
| 1823 | const SCEV *createAddRecFromPHI(PHINode *PN); |
| 1824 | |
| 1825 | /// A helper function for createAddRecFromPHI to handle simple cases. |
| 1826 | const SCEV *createSimpleAffineAddRec(PHINode *PN, Value *BEValueV, |
| 1827 | Value *StartValueV); |
| 1828 | |
| 1829 | /// Helper function called from createNodeForPHI. |
| 1830 | const SCEV *createNodeFromSelectLikePHI(PHINode *PN); |
| 1831 | |
| 1832 | /// Provide special handling for a select-like instruction (currently this |
| 1833 | /// is either a select instruction or a phi node). \p Ty is the type of the |
| 1834 | /// instruction being processed, that is assumed equivalent to |
| 1835 | /// "Cond ? TrueVal : FalseVal". |
| 1836 | std::optional<const SCEV *> |
| 1837 | createNodeForSelectOrPHIInstWithICmpInstCond(Type *Ty, ICmpInst *Cond, |
| 1838 | Value *TrueVal, Value *FalseVal); |
| 1839 | |
| 1840 | /// See if we can model this select-like instruction via umin_seq expression. |
| 1841 | const SCEV *createNodeForSelectOrPHIViaUMinSeq(Value *I, Value *Cond, |
| 1842 | Value *TrueVal, |
| 1843 | Value *FalseVal); |
| 1844 | |
| 1845 | /// Given a value \p V, which is a select-like instruction (currently this is |
| 1846 | /// either a select instruction or a phi node), which is assumed equivalent to |
| 1847 | /// Cond ? TrueVal : FalseVal |
| 1848 | /// see if we can model it as a SCEV expression. |
| 1849 | const SCEV *createNodeForSelectOrPHI(Value *V, Value *Cond, Value *TrueVal, |
| 1850 | Value *FalseVal); |
| 1851 | |
| 1852 | /// Provide the special handling we need to analyze GEP SCEVs. |
| 1853 | const SCEV *createNodeForGEP(GEPOperator *GEP); |
| 1854 | |
| 1855 | /// Implementation code for getSCEVAtScope; called at most once for each |
| 1856 | /// SCEV+Loop pair. |
| 1857 | const SCEV *computeSCEVAtScope(const SCEV *S, const Loop *L); |
| 1858 | |
| 1859 | /// Return the BackedgeTakenInfo for the given loop, lazily computing new |
| 1860 | /// values if the loop hasn't been analyzed yet. The returned result is |
| 1861 | /// guaranteed not to be predicated. |
| 1862 | BackedgeTakenInfo &getBackedgeTakenInfo(const Loop *L); |
| 1863 | |
| 1864 | /// Similar to getBackedgeTakenInfo, but will add predicates as required |
| 1865 | /// with the purpose of returning complete information. |
| 1866 | BackedgeTakenInfo &getPredicatedBackedgeTakenInfo(const Loop *L); |
| 1867 | |
| 1868 | /// Compute the number of times the specified loop will iterate. |
| 1869 | /// If AllowPredicates is set, we will create new SCEV predicates as |
| 1870 | /// necessary in order to return an exact answer. |
| 1871 | BackedgeTakenInfo computeBackedgeTakenCount(const Loop *L, |
| 1872 | bool AllowPredicates = false); |
| 1873 | |
| 1874 | /// Compute the number of times the backedge of the specified loop will |
| 1875 | /// execute if it exits via the specified block. If AllowPredicates is set, |
| 1876 | /// this call will try to use a minimal set of SCEV predicates in order to |
| 1877 | /// return an exact answer. |
| 1878 | ExitLimit computeExitLimit(const Loop *L, BasicBlock *ExitingBlock, |
| 1879 | bool IsOnlyExit, bool AllowPredicates = false); |
| 1880 | |
| 1881 | // Helper functions for computeExitLimitFromCond to avoid exponential time |
| 1882 | // complexity. |
| 1883 | |
| 1884 | class ExitLimitCache { |
| 1885 | // It may look like we need key on the whole (L, ExitIfTrue, |
| 1886 | // ControlsOnlyExit, AllowPredicates) tuple, but recursive calls to |
| 1887 | // computeExitLimitFromCondCached from computeExitLimitFromCondImpl only |
| 1888 | // vary the in \c ExitCond and \c ControlsOnlyExit parameters. We remember |
| 1889 | // the initial values of the other values to assert our assumption. |
| 1890 | SmallDenseMap<PointerIntPair<Value *, 1>, ExitLimit> TripCountMap; |
| 1891 | |
| 1892 | const Loop *L; |
| 1893 | bool ExitIfTrue; |
| 1894 | bool AllowPredicates; |
| 1895 | |
| 1896 | public: |
| 1897 | ExitLimitCache(const Loop *L, bool ExitIfTrue, bool AllowPredicates) |
| 1898 | : L(L), ExitIfTrue(ExitIfTrue), AllowPredicates(AllowPredicates) {} |
| 1899 | |
| 1900 | LLVM_ABI std::optional<ExitLimit> find(const Loop *L, Value *ExitCond, |
| 1901 | bool ExitIfTrue, |
| 1902 | bool ControlsOnlyExit, |
| 1903 | bool AllowPredicates); |
| 1904 | |
| 1905 | LLVM_ABI void insert(const Loop *L, Value *ExitCond, bool ExitIfTrue, |
| 1906 | bool ControlsOnlyExit, bool AllowPredicates, |
| 1907 | const ExitLimit &EL); |
| 1908 | }; |
| 1909 | |
| 1910 | using ExitLimitCacheTy = ExitLimitCache; |
| 1911 | |
| 1912 | ExitLimit computeExitLimitFromCondCached(ExitLimitCacheTy &Cache, |
| 1913 | const Loop *L, Value *ExitCond, |
| 1914 | bool ExitIfTrue, |
| 1915 | bool ControlsOnlyExit, |
| 1916 | bool AllowPredicates); |
| 1917 | ExitLimit computeExitLimitFromCondImpl(ExitLimitCacheTy &Cache, const Loop *L, |
| 1918 | Value *ExitCond, bool ExitIfTrue, |
| 1919 | bool ControlsOnlyExit, |
| 1920 | bool AllowPredicates); |
| 1921 | std::optional<ScalarEvolution::ExitLimit> computeExitLimitFromCondFromBinOp( |
| 1922 | ExitLimitCacheTy &Cache, const Loop *L, Value *ExitCond, bool ExitIfTrue, |
| 1923 | bool ControlsOnlyExit, bool AllowPredicates); |
| 1924 | |
| 1925 | /// Compute the number of times the backedge of the specified loop will |
| 1926 | /// execute if its exit condition were a conditional branch of the ICmpInst |
| 1927 | /// ExitCond and ExitIfTrue. If AllowPredicates is set, this call will try |
| 1928 | /// to use a minimal set of SCEV predicates in order to return an exact |
| 1929 | /// answer. |
| 1930 | ExitLimit computeExitLimitFromICmp(const Loop *L, ICmpInst *ExitCond, |
| 1931 | bool ExitIfTrue, |
| 1932 | bool IsSubExpr, |
| 1933 | bool AllowPredicates = false); |
| 1934 | |
| 1935 | /// Variant of previous which takes the components representing an ICmp |
| 1936 | /// as opposed to the ICmpInst itself. Note that the prior version can |
| 1937 | /// return more precise results in some cases and is preferred when caller |
| 1938 | /// has a materialized ICmp. |
| 1939 | ExitLimit computeExitLimitFromICmp(const Loop *L, CmpPredicate Pred, |
| 1940 | const SCEV *LHS, const SCEV *RHS, |
| 1941 | bool IsSubExpr, |
| 1942 | bool AllowPredicates = false); |
| 1943 | |
| 1944 | /// Compute the number of times the backedge of the specified loop will |
| 1945 | /// execute if its exit condition were a switch with a single exiting case |
| 1946 | /// to ExitingBB. |
| 1947 | ExitLimit computeExitLimitFromSingleExitSwitch(const Loop *L, |
| 1948 | SwitchInst *Switch, |
| 1949 | BasicBlock *ExitingBB, |
| 1950 | bool IsSubExpr); |
| 1951 | |
| 1952 | /// Compute the exit limit of a loop that is controlled by a |
| 1953 | /// "(IV >> 1) != 0" type comparison. We cannot compute the exact trip |
| 1954 | /// count in these cases (since SCEV has no way of expressing them), but we |
| 1955 | /// can still sometimes compute an upper bound. |
| 1956 | /// |
| 1957 | /// Return an ExitLimit for a loop whose backedge is guarded by `LHS Pred |
| 1958 | /// RHS`. |
| 1959 | ExitLimit computeShiftCompareExitLimit(Value *LHS, Value *RHS, const Loop *L, |
| 1960 | ICmpInst::Predicate Pred); |
| 1961 | |
| 1962 | /// If the loop is known to execute a constant number of times (the |
| 1963 | /// condition evolves only from constants), try to evaluate a few iterations |
| 1964 | /// of the loop until we get the exit condition gets a value of ExitWhen |
| 1965 | /// (true or false). If we cannot evaluate the exit count of the loop, |
| 1966 | /// return CouldNotCompute. |
| 1967 | const SCEV *computeExitCountExhaustively(const Loop *L, Value *Cond, |
| 1968 | bool ExitWhen); |
| 1969 | |
| 1970 | /// Return the number of times an exit condition comparing the specified |
| 1971 | /// value to zero will execute. If not computable, return CouldNotCompute. |
| 1972 | /// If AllowPredicates is set, this call will try to use a minimal set of |
| 1973 | /// SCEV predicates in order to return an exact answer. |
| 1974 | ExitLimit howFarToZero(const SCEV *V, const Loop *L, bool IsSubExpr, |
| 1975 | bool AllowPredicates = false); |
| 1976 | |
| 1977 | /// Return the number of times an exit condition checking the specified |
| 1978 | /// value for nonzero will execute. If not computable, return |
| 1979 | /// CouldNotCompute. |
| 1980 | ExitLimit howFarToNonZero(const SCEV *V, const Loop *L); |
| 1981 | |
| 1982 | /// Return the number of times an exit condition containing the specified |
| 1983 | /// less-than comparison will execute. If not computable, return |
| 1984 | /// CouldNotCompute. |
| 1985 | /// |
| 1986 | /// \p isSigned specifies whether the less-than is signed. |
| 1987 | /// |
| 1988 | /// \p ControlsOnlyExit is true when the LHS < RHS condition directly controls |
| 1989 | /// the branch (loops exits only if condition is true). In this case, we can |
| 1990 | /// use NoWrapFlags to skip overflow checks. |
| 1991 | /// |
| 1992 | /// If \p AllowPredicates is set, this call will try to use a minimal set of |
| 1993 | /// SCEV predicates in order to return an exact answer. |
| 1994 | ExitLimit howManyLessThans(const SCEV *LHS, const SCEV *RHS, const Loop *L, |
| 1995 | bool isSigned, bool ControlsOnlyExit, |
| 1996 | bool AllowPredicates = false); |
| 1997 | |
| 1998 | ExitLimit howManyGreaterThans(const SCEV *LHS, const SCEV *RHS, const Loop *L, |
| 1999 | bool isSigned, bool IsSubExpr, |
| 2000 | bool AllowPredicates = false); |
| 2001 | |
| 2002 | /// Return a predecessor of BB (which may not be an immediate predecessor) |
| 2003 | /// which has exactly one successor from which BB is reachable, or null if |
| 2004 | /// no such block is found. |
| 2005 | std::pair<const BasicBlock *, const BasicBlock *> |
| 2006 | getPredecessorWithUniqueSuccessorForBB(const BasicBlock *BB) const; |
| 2007 | |
| 2008 | /// Test whether the condition described by Pred, LHS, and RHS is true |
| 2009 | /// whenever the given FoundCondValue value evaluates to true in given |
| 2010 | /// Context. If Context is nullptr, then the found predicate is true |
| 2011 | /// everywhere. LHS and FoundLHS may have different type width. |
| 2012 | LLVM_ABI bool isImpliedCond(CmpPredicate Pred, const SCEV *LHS, |
| 2013 | const SCEV *RHS, const Value *FoundCondValue, |
| 2014 | bool Inverse, |
| 2015 | const Instruction *Context = nullptr); |
| 2016 | |
| 2017 | /// Test whether the condition described by Pred, LHS, and RHS is true |
| 2018 | /// whenever the given FoundCondValue value evaluates to true in given |
| 2019 | /// Context. If Context is nullptr, then the found predicate is true |
| 2020 | /// everywhere. LHS and FoundLHS must have same type width. |
| 2021 | LLVM_ABI bool isImpliedCondBalancedTypes(CmpPredicate Pred, const SCEV *LHS, |
| 2022 | const SCEV *RHS, |
| 2023 | CmpPredicate FoundPred, |
| 2024 | const SCEV *FoundLHS, |
| 2025 | const SCEV *FoundRHS, |
| 2026 | const Instruction *CtxI); |
| 2027 | |
| 2028 | /// Test whether the condition described by Pred, LHS, and RHS is true |
| 2029 | /// whenever the condition described by FoundPred, FoundLHS, FoundRHS is |
| 2030 | /// true in given Context. If Context is nullptr, then the found predicate is |
| 2031 | /// true everywhere. |
| 2032 | LLVM_ABI bool isImpliedCond(CmpPredicate Pred, const SCEV *LHS, |
| 2033 | const SCEV *RHS, CmpPredicate FoundPred, |
| 2034 | const SCEV *FoundLHS, const SCEV *FoundRHS, |
| 2035 | const Instruction *Context = nullptr); |
| 2036 | |
| 2037 | /// Test whether the condition described by Pred, LHS, and RHS is true |
| 2038 | /// whenever the condition described by Pred, FoundLHS, and FoundRHS is |
| 2039 | /// true in given Context. If Context is nullptr, then the found predicate is |
| 2040 | /// true everywhere. |
| 2041 | bool isImpliedCondOperands(CmpPredicate Pred, const SCEV *LHS, |
| 2042 | const SCEV *RHS, const SCEV *FoundLHS, |
| 2043 | const SCEV *FoundRHS, |
| 2044 | const Instruction *Context = nullptr); |
| 2045 | |
| 2046 | /// Test whether the condition described by Pred, LHS, and RHS is true |
| 2047 | /// whenever the condition described by Pred, FoundLHS, and FoundRHS is |
| 2048 | /// true. Here LHS is an operation that includes FoundLHS as one of its |
| 2049 | /// arguments. |
| 2050 | bool isImpliedViaOperations(CmpPredicate Pred, const SCEV *LHS, |
| 2051 | const SCEV *RHS, const SCEV *FoundLHS, |
| 2052 | const SCEV *FoundRHS, unsigned Depth = 0); |
| 2053 | |
| 2054 | /// Test whether the condition described by Pred, LHS, and RHS is true. |
| 2055 | /// Use only simple non-recursive types of checks, such as range analysis etc. |
| 2056 | bool isKnownViaNonRecursiveReasoning(CmpPredicate Pred, const SCEV *LHS, |
| 2057 | const SCEV *RHS); |
| 2058 | |
| 2059 | /// Test whether the condition described by Pred, LHS, and RHS is true |
| 2060 | /// whenever the condition described by Pred, FoundLHS, and FoundRHS is |
| 2061 | /// true. |
| 2062 | bool isImpliedCondOperandsHelper(CmpPredicate Pred, const SCEV *LHS, |
| 2063 | const SCEV *RHS, const SCEV *FoundLHS, |
| 2064 | const SCEV *FoundRHS); |
| 2065 | |
| 2066 | /// Test whether the condition described by Pred, LHS, and RHS is true |
| 2067 | /// whenever the condition described by Pred, FoundLHS, and FoundRHS is |
| 2068 | /// true. Utility function used by isImpliedCondOperands. Tries to get |
| 2069 | /// cases like "X `sgt` 0 => X - 1 `sgt` -1". |
| 2070 | bool isImpliedCondOperandsViaRanges(CmpPredicate Pred, const SCEV *LHS, |
| 2071 | const SCEV *RHS, CmpPredicate FoundPred, |
| 2072 | const SCEV *FoundLHS, |
| 2073 | const SCEV *FoundRHS); |
| 2074 | |
| 2075 | /// Return true if the condition denoted by \p LHS \p Pred \p RHS is implied |
| 2076 | /// by a call to @llvm.experimental.guard in \p BB. |
| 2077 | bool isImpliedViaGuard(const BasicBlock *BB, CmpPredicate Pred, |
| 2078 | const SCEV *LHS, const SCEV *RHS); |
| 2079 | |
| 2080 | /// Test whether the condition described by Pred, LHS, and RHS is true |
| 2081 | /// whenever the condition described by Pred, FoundLHS, and FoundRHS is |
| 2082 | /// true. |
| 2083 | /// |
| 2084 | /// This routine tries to rule out certain kinds of integer overflow, and |
| 2085 | /// then tries to reason about arithmetic properties of the predicates. |
| 2086 | bool isImpliedCondOperandsViaNoOverflow(CmpPredicate Pred, const SCEV *LHS, |
| 2087 | const SCEV *RHS, const SCEV *FoundLHS, |
| 2088 | const SCEV *FoundRHS); |
| 2089 | |
| 2090 | /// Test whether the condition described by Pred, LHS, and RHS is true |
| 2091 | /// whenever the condition described by Pred, FoundLHS, and FoundRHS is |
| 2092 | /// true. |
| 2093 | /// |
| 2094 | /// This routine tries to weaken the known condition basing on fact that |
| 2095 | /// FoundLHS is an AddRec. |
| 2096 | bool isImpliedCondOperandsViaAddRecStart(CmpPredicate Pred, const SCEV *LHS, |
| 2097 | const SCEV *RHS, |
| 2098 | const SCEV *FoundLHS, |
| 2099 | const SCEV *FoundRHS, |
| 2100 | const Instruction *CtxI); |
| 2101 | |
| 2102 | /// Test whether the condition described by Pred, LHS, and RHS is true |
| 2103 | /// whenever the condition described by Pred, FoundLHS, and FoundRHS is |
| 2104 | /// true. |
| 2105 | /// |
| 2106 | /// This routine tries to figure out predicate for Phis which are SCEVUnknown |
| 2107 | /// if it is true for every possible incoming value from their respective |
| 2108 | /// basic blocks. |
| 2109 | bool isImpliedViaMerge(CmpPredicate Pred, const SCEV *LHS, const SCEV *RHS, |
| 2110 | const SCEV *FoundLHS, const SCEV *FoundRHS, |
| 2111 | unsigned Depth); |
| 2112 | |
| 2113 | /// Test whether the condition described by Pred, LHS, and RHS is true |
| 2114 | /// whenever the condition described by Pred, FoundLHS, and FoundRHS is |
| 2115 | /// true. |
| 2116 | /// |
| 2117 | /// This routine tries to reason about shifts. |
| 2118 | bool isImpliedCondOperandsViaShift(CmpPredicate Pred, const SCEV *LHS, |
| 2119 | const SCEV *RHS, const SCEV *FoundLHS, |
| 2120 | const SCEV *FoundRHS); |
| 2121 | |
| 2122 | /// If we know that the specified Phi is in the header of its containing |
| 2123 | /// loop, we know the loop executes a constant number of times, and the PHI |
| 2124 | /// node is just a recurrence involving constants, fold it. |
| 2125 | Constant *getConstantEvolutionLoopExitValue(PHINode *PN, const APInt &BEs, |
| 2126 | const Loop *L); |
| 2127 | |
| 2128 | /// Test if the given expression is known to satisfy the condition described |
| 2129 | /// by Pred and the known constant ranges of LHS and RHS. |
| 2130 | bool isKnownPredicateViaConstantRanges(CmpPredicate Pred, const SCEV *LHS, |
| 2131 | const SCEV *RHS); |
| 2132 | |
| 2133 | /// Try to prove the condition described by "LHS Pred RHS" by ruling out |
| 2134 | /// integer overflow. |
| 2135 | /// |
| 2136 | /// For instance, this will return true for "A s< (A + C)<nsw>" if C is |
| 2137 | /// positive. |
| 2138 | bool isKnownPredicateViaNoOverflow(CmpPredicate Pred, const SCEV *LHS, |
| 2139 | const SCEV *RHS); |
| 2140 | |
| 2141 | /// Try to split Pred LHS RHS into logical conjunctions (and's) and try to |
| 2142 | /// prove them individually. |
| 2143 | bool isKnownPredicateViaSplitting(CmpPredicate Pred, const SCEV *LHS, |
| 2144 | const SCEV *RHS); |
| 2145 | |
| 2146 | /// Try to match the Expr as "(L + R)<Flags>". |
| 2147 | bool splitBinaryAdd(const SCEV *Expr, const SCEV *&L, const SCEV *&R, |
| 2148 | SCEV::NoWrapFlags &Flags); |
| 2149 | |
| 2150 | /// Forget predicated/non-predicated backedge taken counts for the given loop. |
| 2151 | void forgetBackedgeTakenCounts(const Loop *L, bool Predicated); |
| 2152 | |
| 2153 | /// Drop memoized information for all \p SCEVs. |
| 2154 | void forgetMemoizedResults(ArrayRef<const SCEV *> SCEVs); |
| 2155 | |
| 2156 | /// Helper for forgetMemoizedResults. |
| 2157 | void forgetMemoizedResultsImpl(const SCEV *S); |
| 2158 | |
| 2159 | /// Iterate over instructions in \p Worklist and their users. Erase entries |
| 2160 | /// from ValueExprMap and collect SCEV expressions in \p ToForget |
| 2161 | void visitAndClearUsers(SmallVectorImpl<Instruction *> &Worklist, |
| 2162 | SmallPtrSetImpl<Instruction *> &Visited, |
| 2163 | SmallVectorImpl<const SCEV *> &ToForget); |
| 2164 | |
| 2165 | /// Erase Value from ValueExprMap and ExprValueMap. |
| 2166 | void eraseValueFromMap(Value *V); |
| 2167 | |
| 2168 | /// Insert V to S mapping into ValueExprMap and ExprValueMap. |
| 2169 | void insertValueToMap(Value *V, const SCEV *S); |
| 2170 | |
| 2171 | /// Return false iff given SCEV contains a SCEVUnknown with NULL value- |
| 2172 | /// pointer. |
| 2173 | bool checkValidity(const SCEV *S) const; |
| 2174 | |
| 2175 | /// Return true if `ExtendOpTy`({`Start`,+,`Step`}) can be proved to be |
| 2176 | /// equal to {`ExtendOpTy`(`Start`),+,`ExtendOpTy`(`Step`)}. This is |
| 2177 | /// equivalent to proving no signed (resp. unsigned) wrap in |
| 2178 | /// {`Start`,+,`Step`} if `ExtendOpTy` is `SCEVSignExtendExpr` |
| 2179 | /// (resp. `SCEVZeroExtendExpr`). |
| 2180 | template <typename ExtendOpTy> |
| 2181 | bool proveNoWrapByVaryingStart(const SCEV *Start, const SCEV *Step, |
| 2182 | const Loop *L); |
| 2183 | |
| 2184 | /// Try to prove NSW or NUW on \p AR relying on ConstantRange manipulation. |
| 2185 | SCEV::NoWrapFlags proveNoWrapViaConstantRanges(const SCEVAddRecExpr *AR); |
| 2186 | |
| 2187 | /// Try to prove NSW on \p AR by proving facts about conditions known on |
| 2188 | /// entry and backedge. |
| 2189 | SCEV::NoWrapFlags proveNoSignedWrapViaInduction(const SCEVAddRecExpr *AR); |
| 2190 | |
| 2191 | /// Try to prove NUW on \p AR by proving facts about conditions known on |
| 2192 | /// entry and backedge. |
| 2193 | SCEV::NoWrapFlags proveNoUnsignedWrapViaInduction(const SCEVAddRecExpr *AR); |
| 2194 | |
| 2195 | std::optional<MonotonicPredicateType> |
| 2196 | getMonotonicPredicateTypeImpl(const SCEVAddRecExpr *LHS, |
| 2197 | ICmpInst::Predicate Pred); |
| 2198 | |
| 2199 | /// Return SCEV no-wrap flags that can be proven based on reasoning about |
| 2200 | /// how poison produced from no-wrap flags on this value (e.g. a nuw add) |
| 2201 | /// would trigger undefined behavior on overflow. |
| 2202 | SCEV::NoWrapFlags getNoWrapFlagsFromUB(const Value *V); |
| 2203 | |
| 2204 | /// Return a scope which provides an upper bound on the defining scope of |
| 2205 | /// 'S'. Specifically, return the first instruction in said bounding scope. |
| 2206 | /// Return nullptr if the scope is trivial (function entry). |
| 2207 | /// (See scope definition rules associated with flag discussion above) |
| 2208 | const Instruction *getNonTrivialDefiningScopeBound(const SCEV *S); |
| 2209 | |
| 2210 | /// Return a scope which provides an upper bound on the defining scope for |
| 2211 | /// a SCEV with the operands in Ops. The outparam Precise is set if the |
| 2212 | /// bound found is a precise bound (i.e. must be the defining scope.) |
| 2213 | const Instruction *getDefiningScopeBound(ArrayRef<const SCEV *> Ops, |
| 2214 | bool &Precise); |
| 2215 | |
| 2216 | /// Wrapper around the above for cases which don't care if the bound |
| 2217 | /// is precise. |
| 2218 | const Instruction *getDefiningScopeBound(ArrayRef<const SCEV *> Ops); |
| 2219 | |
| 2220 | /// Given two instructions in the same function, return true if we can |
| 2221 | /// prove B must execute given A executes. |
| 2222 | bool isGuaranteedToTransferExecutionTo(const Instruction *A, |
| 2223 | const Instruction *B); |
| 2224 | |
| 2225 | /// Returns true if \p Op is guaranteed not to cause immediate UB. |
| 2226 | bool isGuaranteedNotToCauseUB(const SCEV *Op); |
| 2227 | |
| 2228 | /// Returns true if \p Op is guaranteed to not be poison. |
| 2229 | static bool isGuaranteedNotToBePoison(const SCEV *Op); |
| 2230 | |
| 2231 | /// Return true if the SCEV corresponding to \p I is never poison. Proving |
| 2232 | /// this is more complex than proving that just \p I is never poison, since |
| 2233 | /// SCEV commons expressions across control flow, and you can have cases |
| 2234 | /// like: |
| 2235 | /// |
| 2236 | /// idx0 = a + b; |
| 2237 | /// ptr[idx0] = 100; |
| 2238 | /// if (<condition>) { |
| 2239 | /// idx1 = a +nsw b; |
| 2240 | /// ptr[idx1] = 200; |
| 2241 | /// } |
| 2242 | /// |
| 2243 | /// where the SCEV expression (+ a b) is guaranteed to not be poison (and |
| 2244 | /// hence not sign-overflow) only if "<condition>" is true. Since both |
| 2245 | /// `idx0` and `idx1` will be mapped to the same SCEV expression, (+ a b), |
| 2246 | /// it is not okay to annotate (+ a b) with <nsw> in the above example. |
| 2247 | bool isSCEVExprNeverPoison(const Instruction *I); |
| 2248 | |
| 2249 | /// This is like \c isSCEVExprNeverPoison but it specifically works for |
| 2250 | /// instructions that will get mapped to SCEV add recurrences. Return true |
| 2251 | /// if \p I will never generate poison under the assumption that \p I is an |
| 2252 | /// add recurrence on the loop \p L. |
| 2253 | bool isAddRecNeverPoison(const Instruction *I, const Loop *L); |
| 2254 | |
| 2255 | /// Similar to createAddRecFromPHI, but with the additional flexibility of |
| 2256 | /// suggesting runtime overflow checks in case casts are encountered. |
| 2257 | /// If successful, the analysis records that for this loop, \p SymbolicPHI, |
| 2258 | /// which is the UnknownSCEV currently representing the PHI, can be rewritten |
| 2259 | /// into an AddRec, assuming some predicates; The function then returns the |
| 2260 | /// AddRec and the predicates as a pair, and caches this pair in |
| 2261 | /// PredicatedSCEVRewrites. |
| 2262 | /// If the analysis is not successful, a mapping from the \p SymbolicPHI to |
| 2263 | /// itself (with no predicates) is recorded, and a nullptr with an empty |
| 2264 | /// predicates vector is returned as a pair. |
| 2265 | std::optional<std::pair<const SCEV *, SmallVector<const SCEVPredicate *, 3>>> |
| 2266 | createAddRecFromPHIWithCastsImpl(const SCEVUnknown *SymbolicPHI); |
| 2267 | |
| 2268 | /// Compute the maximum backedge count based on the range of values |
| 2269 | /// permitted by Start, End, and Stride. This is for loops of the form |
| 2270 | /// {Start, +, Stride} LT End. |
| 2271 | /// |
| 2272 | /// Preconditions: |
| 2273 | /// * the induction variable is known to be positive. |
| 2274 | /// * the induction variable is assumed not to overflow (i.e. either it |
| 2275 | /// actually doesn't, or we'd have to immediately execute UB) |
| 2276 | /// We *don't* assert these preconditions so please be careful. |
| 2277 | const SCEV *computeMaxBECountForLT(const SCEV *Start, const SCEV *Stride, |
| 2278 | const SCEV *End, unsigned BitWidth, |
| 2279 | bool IsSigned); |
| 2280 | |
| 2281 | /// Verify if an linear IV with positive stride can overflow when in a |
| 2282 | /// less-than comparison, knowing the invariant term of the comparison, |
| 2283 | /// the stride. |
| 2284 | bool canIVOverflowOnLT(const SCEV *RHS, const SCEV *Stride, bool IsSigned); |
| 2285 | |
| 2286 | /// Verify if an linear IV with negative stride can overflow when in a |
| 2287 | /// greater-than comparison, knowing the invariant term of the comparison, |
| 2288 | /// the stride. |
| 2289 | bool canIVOverflowOnGT(const SCEV *RHS, const SCEV *Stride, bool IsSigned); |
| 2290 | |
| 2291 | /// Get add expr already created or create a new one. |
| 2292 | const SCEV *getOrCreateAddExpr(ArrayRef<const SCEV *> Ops, |
| 2293 | SCEV::NoWrapFlags Flags); |
| 2294 | |
| 2295 | /// Get mul expr already created or create a new one. |
| 2296 | const SCEV *getOrCreateMulExpr(ArrayRef<const SCEV *> Ops, |
| 2297 | SCEV::NoWrapFlags Flags); |
| 2298 | |
| 2299 | // Get addrec expr already created or create a new one. |
| 2300 | const SCEV *getOrCreateAddRecExpr(ArrayRef<const SCEV *> Ops, |
| 2301 | const Loop *L, SCEV::NoWrapFlags Flags); |
| 2302 | |
| 2303 | /// Return x if \p Val is f(x) where f is a 1-1 function. |
| 2304 | const SCEV *stripInjectiveFunctions(const SCEV *Val) const; |
| 2305 | |
| 2306 | /// Find all of the loops transitively used in \p S, and fill \p LoopsUsed. |
| 2307 | /// A loop is considered "used" by an expression if it contains |
| 2308 | /// an add rec on said loop. |
| 2309 | void getUsedLoops(const SCEV *S, SmallPtrSetImpl<const Loop *> &LoopsUsed); |
| 2310 | |
| 2311 | /// Try to match the pattern generated by getURemExpr(A, B). If successful, |
| 2312 | /// Assign A and B to LHS and RHS, respectively. |
| 2313 | LLVM_ABI bool matchURem(const SCEV *Expr, const SCEV *&LHS, const SCEV *&RHS); |
| 2314 | |
| 2315 | /// Look for a SCEV expression with type `SCEVType` and operands `Ops` in |
| 2316 | /// `UniqueSCEVs`. Return if found, else nullptr. |
| 2317 | SCEV *findExistingSCEVInCache(SCEVTypes SCEVType, ArrayRef<const SCEV *> Ops); |
| 2318 | |
| 2319 | /// Get reachable blocks in this function, making limited use of SCEV |
| 2320 | /// reasoning about conditions. |
| 2321 | void getReachableBlocks(SmallPtrSetImpl<BasicBlock *> &Reachable, |
| 2322 | Function &F); |
| 2323 | |
| 2324 | /// Return the given SCEV expression with a new set of operands. |
| 2325 | /// This preserves the origial nowrap flags. |
| 2326 | const SCEV *getWithOperands(const SCEV *S, |
| 2327 | SmallVectorImpl<const SCEV *> &NewOps); |
| 2328 | |
| 2329 | FoldingSet<SCEV> UniqueSCEVs; |
| 2330 | FoldingSet<SCEVPredicate> UniquePreds; |
| 2331 | BumpPtrAllocator SCEVAllocator; |
| 2332 | |
| 2333 | /// This maps loops to a list of addrecs that directly use said loop. |
| 2334 | DenseMap<const Loop *, SmallVector<const SCEVAddRecExpr *, 4>> LoopUsers; |
| 2335 | |
| 2336 | /// Cache tentative mappings from UnknownSCEVs in a Loop, to a SCEV expression |
| 2337 | /// they can be rewritten into under certain predicates. |
| 2338 | DenseMap<std::pair<const SCEVUnknown *, const Loop *>, |
| 2339 | std::pair<const SCEV *, SmallVector<const SCEVPredicate *, 3>>> |
| 2340 | PredicatedSCEVRewrites; |
| 2341 | |
| 2342 | /// Set of AddRecs for which proving NUW via an induction has already been |
| 2343 | /// tried. |
| 2344 | SmallPtrSet<const SCEVAddRecExpr *, 16> UnsignedWrapViaInductionTried; |
| 2345 | |
| 2346 | /// Set of AddRecs for which proving NSW via an induction has already been |
| 2347 | /// tried. |
| 2348 | SmallPtrSet<const SCEVAddRecExpr *, 16> SignedWrapViaInductionTried; |
| 2349 | |
| 2350 | /// The head of a linked list of all SCEVUnknown values that have been |
| 2351 | /// allocated. This is used by releaseMemory to locate them all and call |
| 2352 | /// their destructors. |
| 2353 | SCEVUnknown *FirstUnknown = nullptr; |
| 2354 | }; |
| 2355 | |
| 2356 | /// Analysis pass that exposes the \c ScalarEvolution for a function. |
| 2357 | class ScalarEvolutionAnalysis |
| 2358 | : public AnalysisInfoMixin<ScalarEvolutionAnalysis> { |
| 2359 | friend AnalysisInfoMixin<ScalarEvolutionAnalysis>; |
| 2360 | |
| 2361 | LLVM_ABI static AnalysisKey Key; |
| 2362 | |
| 2363 | public: |
| 2364 | using Result = ScalarEvolution; |
| 2365 | |
| 2366 | LLVM_ABI ScalarEvolution run(Function &F, FunctionAnalysisManager &AM); |
| 2367 | }; |
| 2368 | |
| 2369 | /// Verifier pass for the \c ScalarEvolutionAnalysis results. |
| 2370 | class ScalarEvolutionVerifierPass |
| 2371 | : public PassInfoMixin<ScalarEvolutionVerifierPass> { |
| 2372 | public: |
| 2373 | LLVM_ABI PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM); |
| 2374 | static bool isRequired() { return true; } |
| 2375 | }; |
| 2376 | |
| 2377 | /// Printer pass for the \c ScalarEvolutionAnalysis results. |
| 2378 | class ScalarEvolutionPrinterPass |
| 2379 | : public PassInfoMixin<ScalarEvolutionPrinterPass> { |
| 2380 | raw_ostream &OS; |
| 2381 | |
| 2382 | public: |
| 2383 | explicit ScalarEvolutionPrinterPass(raw_ostream &OS) : OS(OS) {} |
| 2384 | |
| 2385 | LLVM_ABI PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM); |
| 2386 | |
| 2387 | static bool isRequired() { return true; } |
| 2388 | }; |
| 2389 | |
| 2390 | class LLVM_ABI ScalarEvolutionWrapperPass : public FunctionPass { |
| 2391 | std::unique_ptr<ScalarEvolution> SE; |
| 2392 | |
| 2393 | public: |
| 2394 | static char ID; |
| 2395 | |
| 2396 | ScalarEvolutionWrapperPass(); |
| 2397 | |
| 2398 | ScalarEvolution &getSE() { return *SE; } |
| 2399 | const ScalarEvolution &getSE() const { return *SE; } |
| 2400 | |
| 2401 | bool runOnFunction(Function &F) override; |
| 2402 | void releaseMemory() override; |
| 2403 | void getAnalysisUsage(AnalysisUsage &AU) const override; |
| 2404 | void print(raw_ostream &OS, const Module * = nullptr) const override; |
| 2405 | void verifyAnalysis() const override; |
| 2406 | }; |
| 2407 | |
| 2408 | /// An interface layer with SCEV used to manage how we see SCEV expressions |
| 2409 | /// for values in the context of existing predicates. We can add new |
| 2410 | /// predicates, but we cannot remove them. |
| 2411 | /// |
| 2412 | /// This layer has multiple purposes: |
| 2413 | /// - provides a simple interface for SCEV versioning. |
| 2414 | /// - guarantees that the order of transformations applied on a SCEV |
| 2415 | /// expression for a single Value is consistent across two different |
| 2416 | /// getSCEV calls. This means that, for example, once we've obtained |
| 2417 | /// an AddRec expression for a certain value through expression |
| 2418 | /// rewriting, we will continue to get an AddRec expression for that |
| 2419 | /// Value. |
| 2420 | /// - lowers the number of expression rewrites. |
| 2421 | class PredicatedScalarEvolution { |
| 2422 | public: |
| 2423 | LLVM_ABI PredicatedScalarEvolution(ScalarEvolution &SE, Loop &L); |
| 2424 | |
| 2425 | LLVM_ABI const SCEVPredicate &getPredicate() const; |
| 2426 | |
| 2427 | /// Returns the SCEV expression of V, in the context of the current SCEV |
| 2428 | /// predicate. The order of transformations applied on the expression of V |
| 2429 | /// returned by ScalarEvolution is guaranteed to be preserved, even when |
| 2430 | /// adding new predicates. |
| 2431 | LLVM_ABI const SCEV *getSCEV(Value *V); |
| 2432 | |
| 2433 | /// Get the (predicated) backedge count for the analyzed loop. |
| 2434 | LLVM_ABI const SCEV *getBackedgeTakenCount(); |
| 2435 | |
| 2436 | /// Get the (predicated) symbolic max backedge count for the analyzed loop. |
| 2437 | LLVM_ABI const SCEV *getSymbolicMaxBackedgeTakenCount(); |
| 2438 | |
| 2439 | /// Returns the upper bound of the loop trip count as a normal unsigned |
| 2440 | /// value, or 0 if the trip count is unknown. |
| 2441 | LLVM_ABI unsigned getSmallConstantMaxTripCount(); |
| 2442 | |
| 2443 | /// Adds a new predicate. |
| 2444 | LLVM_ABI void addPredicate(const SCEVPredicate &Pred); |
| 2445 | |
| 2446 | /// Attempts to produce an AddRecExpr for V by adding additional SCEV |
| 2447 | /// predicates. If we can't transform the expression into an AddRecExpr we |
| 2448 | /// return nullptr and not add additional SCEV predicates to the current |
| 2449 | /// context. |
| 2450 | LLVM_ABI const SCEVAddRecExpr *getAsAddRec(Value *V); |
| 2451 | |
| 2452 | /// Proves that V doesn't overflow by adding SCEV predicate. |
| 2453 | LLVM_ABI void setNoOverflow(Value *V, |
| 2454 | SCEVWrapPredicate::IncrementWrapFlags Flags); |
| 2455 | |
| 2456 | /// Returns true if we've proved that V doesn't wrap by means of a SCEV |
| 2457 | /// predicate. |
| 2458 | LLVM_ABI bool hasNoOverflow(Value *V, |
| 2459 | SCEVWrapPredicate::IncrementWrapFlags Flags); |
| 2460 | |
| 2461 | /// Returns the ScalarEvolution analysis used. |
| 2462 | ScalarEvolution *getSE() const { return &SE; } |
| 2463 | |
| 2464 | /// We need to explicitly define the copy constructor because of FlagsMap. |
| 2465 | LLVM_ABI PredicatedScalarEvolution(const PredicatedScalarEvolution &); |
| 2466 | |
| 2467 | /// Print the SCEV mappings done by the Predicated Scalar Evolution. |
| 2468 | /// The printed text is indented by \p Depth. |
| 2469 | LLVM_ABI void print(raw_ostream &OS, unsigned Depth) const; |
| 2470 | |
| 2471 | /// Check if \p AR1 and \p AR2 are equal, while taking into account |
| 2472 | /// Equal predicates in Preds. |
| 2473 | LLVM_ABI bool areAddRecsEqualWithPreds(const SCEVAddRecExpr *AR1, |
| 2474 | const SCEVAddRecExpr *AR2) const; |
| 2475 | |
| 2476 | private: |
| 2477 | /// Increments the version number of the predicate. This needs to be called |
| 2478 | /// every time the SCEV predicate changes. |
| 2479 | void updateGeneration(); |
| 2480 | |
| 2481 | /// Holds a SCEV and the version number of the SCEV predicate used to |
| 2482 | /// perform the rewrite of the expression. |
| 2483 | using RewriteEntry = std::pair<unsigned, const SCEV *>; |
| 2484 | |
| 2485 | /// Maps a SCEV to the rewrite result of that SCEV at a certain version |
| 2486 | /// number. If this number doesn't match the current Generation, we will |
| 2487 | /// need to do a rewrite. To preserve the transformation order of previous |
| 2488 | /// rewrites, we will rewrite the previous result instead of the original |
| 2489 | /// SCEV. |
| 2490 | DenseMap<const SCEV *, RewriteEntry> RewriteMap; |
| 2491 | |
| 2492 | /// Records what NoWrap flags we've added to a Value *. |
| 2493 | ValueMap<Value *, SCEVWrapPredicate::IncrementWrapFlags> FlagsMap; |
| 2494 | |
| 2495 | /// The ScalarEvolution analysis. |
| 2496 | ScalarEvolution &SE; |
| 2497 | |
| 2498 | /// The analyzed Loop. |
| 2499 | const Loop &L; |
| 2500 | |
| 2501 | /// The SCEVPredicate that forms our context. We will rewrite all |
| 2502 | /// expressions assuming that this predicate true. |
| 2503 | std::unique_ptr<SCEVUnionPredicate> Preds; |
| 2504 | |
| 2505 | /// Marks the version of the SCEV predicate used. When rewriting a SCEV |
| 2506 | /// expression we mark it with the version of the predicate. We use this to |
| 2507 | /// figure out if the predicate has changed from the last rewrite of the |
| 2508 | /// SCEV. If so, we need to perform a new rewrite. |
| 2509 | unsigned Generation = 0; |
| 2510 | |
| 2511 | /// The backedge taken count. |
| 2512 | const SCEV *BackedgeCount = nullptr; |
| 2513 | |
| 2514 | /// The symbolic backedge taken count. |
| 2515 | const SCEV *SymbolicMaxBackedgeCount = nullptr; |
| 2516 | |
| 2517 | /// The constant max trip count for the loop. |
| 2518 | std::optional<unsigned> SmallConstantMaxTripCount; |
| 2519 | }; |
| 2520 | |
| 2521 | template <> struct DenseMapInfo<ScalarEvolution::FoldID> { |
| 2522 | static inline ScalarEvolution::FoldID getEmptyKey() { |
| 2523 | ScalarEvolution::FoldID ID(0); |
| 2524 | return ID; |
| 2525 | } |
| 2526 | static inline ScalarEvolution::FoldID getTombstoneKey() { |
| 2527 | ScalarEvolution::FoldID ID(1); |
| 2528 | return ID; |
| 2529 | } |
| 2530 | |
| 2531 | static unsigned getHashValue(const ScalarEvolution::FoldID &Val) { |
| 2532 | return Val.computeHash(); |
| 2533 | } |
| 2534 | |
| 2535 | static bool isEqual(const ScalarEvolution::FoldID &LHS, |
| 2536 | const ScalarEvolution::FoldID &RHS) { |
| 2537 | return LHS == RHS; |
| 2538 | } |
| 2539 | }; |
| 2540 | |
| 2541 | } // end namespace llvm |
| 2542 | |
| 2543 | #endif // LLVM_ANALYSIS_SCALAREVOLUTION_H |
| 2544 | |