1 //===- llvm/Analysis/ValueTracking.h - Walk computations --------*- C++ -*-===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file contains routines that help analyze properties that chains of 11 // computations have. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #ifndef LLVM_ANALYSIS_VALUETRACKING_H 16 #define LLVM_ANALYSIS_VALUETRACKING_H 17 18 #include "llvm/IR/CallSite.h" 19 #include "llvm/IR/ConstantRange.h" 20 #include "llvm/IR/Instruction.h" 21 #include "llvm/IR/IntrinsicInst.h" 22 #include "llvm/Support/DataTypes.h" 23 24 namespace llvm { 25 template <typename T> class ArrayRef; 26 class APInt; 27 class AddOperator; 28 class AssumptionCache; 29 class DataLayout; 30 class DominatorTree; 31 class GEPOperator; 32 class Instruction; 33 class Loop; 34 class LoopInfo; 35 class MDNode; 36 class StringRef; 37 class TargetLibraryInfo; 38 class Value; 39 40 namespace Intrinsic { 41 enum ID : unsigned; 42 } 43 44 /// Determine which bits of V are known to be either zero or one and return 45 /// them in the KnownZero/KnownOne bit sets. 46 /// 47 /// This function is defined on values with integer type, values with pointer 48 /// type, and vectors of integers. In the case 49 /// where V is a vector, the known zero and known one values are the 50 /// same width as the vector element, and the bit is set only if it is true 51 /// for all of the elements in the vector. 52 void computeKnownBits(Value *V, APInt &KnownZero, APInt &KnownOne, 53 const DataLayout &DL, unsigned Depth = 0, 54 AssumptionCache *AC = nullptr, 55 const Instruction *CxtI = nullptr, 56 const DominatorTree *DT = nullptr); 57 /// Compute known bits from the range metadata. 58 /// \p KnownZero the set of bits that are known to be zero 59 /// \p KnownOne the set of bits that are known to be one 60 void computeKnownBitsFromRangeMetadata(const MDNode &Ranges, 61 APInt &KnownZero, APInt &KnownOne); 62 /// Return true if LHS and RHS have no common bits set. 63 bool haveNoCommonBitsSet(Value *LHS, Value *RHS, const DataLayout &DL, 64 AssumptionCache *AC = nullptr, 65 const Instruction *CxtI = nullptr, 66 const DominatorTree *DT = nullptr); 67 68 /// Determine whether the sign bit is known to be zero or one. Convenience 69 /// wrapper around computeKnownBits. 70 void ComputeSignBit(Value *V, bool &KnownZero, bool &KnownOne, 71 const DataLayout &DL, unsigned Depth = 0, 72 AssumptionCache *AC = nullptr, 73 const Instruction *CxtI = nullptr, 74 const DominatorTree *DT = nullptr); 75 76 /// Return true if the given value is known to have exactly one bit set when 77 /// defined. For vectors return true if every element is known to be a power 78 /// of two when defined. Supports values with integer or pointer type and 79 /// vectors of integers. If 'OrZero' is set, then return true if the given 80 /// value is either a power of two or zero. 81 bool isKnownToBeAPowerOfTwo(Value *V, const DataLayout &DL, 82 bool OrZero = false, unsigned Depth = 0, 83 AssumptionCache *AC = nullptr, 84 const Instruction *CxtI = nullptr, 85 const DominatorTree *DT = nullptr); 86 87 /// Return true if the given value is known to be non-zero when defined. For 88 /// vectors, return true if every element is known to be non-zero when 89 /// defined. Supports values with integer or pointer type and vectors of 90 /// integers. 91 bool isKnownNonZero(Value *V, const DataLayout &DL, unsigned Depth = 0, 92 AssumptionCache *AC = nullptr, 93 const Instruction *CxtI = nullptr, 94 const DominatorTree *DT = nullptr); 95 96 /// Returns true if the give value is known to be non-negative. 97 bool isKnownNonNegative(Value *V, const DataLayout &DL, unsigned Depth = 0, 98 AssumptionCache *AC = nullptr, 99 const Instruction *CxtI = nullptr, 100 const DominatorTree *DT = nullptr); 101 102 /// Returns true if the given value is known be positive (i.e. non-negative 103 /// and non-zero). 104 bool isKnownPositive(Value *V, const DataLayout &DL, unsigned Depth = 0, 105 AssumptionCache *AC = nullptr, 106 const Instruction *CxtI = nullptr, 107 const DominatorTree *DT = nullptr); 108 109 /// Returns true if the given value is known be negative (i.e. non-positive 110 /// and non-zero). 111 bool isKnownNegative(Value *V, const DataLayout &DL, unsigned Depth = 0, 112 AssumptionCache *AC = nullptr, 113 const Instruction *CxtI = nullptr, 114 const DominatorTree *DT = nullptr); 115 116 /// Return true if the given values are known to be non-equal when defined. 117 /// Supports scalar integer types only. 118 bool isKnownNonEqual(Value *V1, Value *V2, const DataLayout &DL, 119 AssumptionCache *AC = nullptr, 120 const Instruction *CxtI = nullptr, 121 const DominatorTree *DT = nullptr); 122 123 /// Return true if 'V & Mask' is known to be zero. We use this predicate to 124 /// simplify operations downstream. Mask is known to be zero for bits that V 125 /// cannot have. 126 /// 127 /// This function is defined on values with integer type, values with pointer 128 /// type, and vectors of integers. In the case 129 /// where V is a vector, the mask, known zero, and known one values are the 130 /// same width as the vector element, and the bit is set only if it is true 131 /// for all of the elements in the vector. 132 bool MaskedValueIsZero(Value *V, const APInt &Mask, const DataLayout &DL, 133 unsigned Depth = 0, AssumptionCache *AC = nullptr, 134 const Instruction *CxtI = nullptr, 135 const DominatorTree *DT = nullptr); 136 137 /// Return the number of times the sign bit of the register is replicated into 138 /// the other bits. We know that at least 1 bit is always equal to the sign 139 /// bit (itself), but other cases can give us information. For example, 140 /// immediately after an "ashr X, 2", we know that the top 3 bits are all 141 /// equal to each other, so we return 3. For vectors, return the number of 142 /// sign bits for the vector element with the mininum number of known sign 143 /// bits. 144 unsigned ComputeNumSignBits(Value *Op, const DataLayout &DL, 145 unsigned Depth = 0, AssumptionCache *AC = nullptr, 146 const Instruction *CxtI = nullptr, 147 const DominatorTree *DT = nullptr); 148 149 /// This function computes the integer multiple of Base that equals V. If 150 /// successful, it returns true and returns the multiple in Multiple. If 151 /// unsuccessful, it returns false. Also, if V can be simplified to an 152 /// integer, then the simplified V is returned in Val. Look through sext only 153 /// if LookThroughSExt=true. 154 bool ComputeMultiple(Value *V, unsigned Base, Value *&Multiple, 155 bool LookThroughSExt = false, 156 unsigned Depth = 0); 157 158 /// Map a call instruction to an intrinsic ID. Libcalls which have equivalent 159 /// intrinsics are treated as-if they were intrinsics. 160 Intrinsic::ID getIntrinsicForCallSite(ImmutableCallSite ICS, 161 const TargetLibraryInfo *TLI); 162 163 /// Return true if we can prove that the specified FP value is never equal to 164 /// -0.0. 165 bool CannotBeNegativeZero(const Value *V, const TargetLibraryInfo *TLI, 166 unsigned Depth = 0); 167 168 /// Return true if we can prove that the specified FP value is either a NaN or 169 /// never less than 0.0. 170 bool CannotBeOrderedLessThanZero(const Value *V, const TargetLibraryInfo *TLI, 171 unsigned Depth = 0); 172 173 /// If the specified value can be set by repeating the same byte in memory, 174 /// return the i8 value that it is represented with. This is true for all i8 175 /// values obviously, but is also true for i32 0, i32 -1, i16 0xF0F0, double 176 /// 0.0 etc. If the value can't be handled with a repeated byte store (e.g. 177 /// i16 0x1234), return null. 178 Value *isBytewiseValue(Value *V); 179 180 /// Given an aggregrate and an sequence of indices, see if the scalar value 181 /// indexed is already around as a register, for example if it were inserted 182 /// directly into the aggregrate. 183 /// 184 /// If InsertBefore is not null, this function will duplicate (modified) 185 /// insertvalues when a part of a nested struct is extracted. 186 Value *FindInsertedValue(Value *V, 187 ArrayRef<unsigned> idx_range, 188 Instruction *InsertBefore = nullptr); 189 190 /// Analyze the specified pointer to see if it can be expressed as a base 191 /// pointer plus a constant offset. Return the base and offset to the caller. 192 Value *GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset, 193 const DataLayout &DL); 194 static inline const Value * GetPointerBaseWithConstantOffset(const Value * Ptr,int64_t & Offset,const DataLayout & DL)195 GetPointerBaseWithConstantOffset(const Value *Ptr, int64_t &Offset, 196 const DataLayout &DL) { 197 return GetPointerBaseWithConstantOffset(const_cast<Value *>(Ptr), Offset, 198 DL); 199 } 200 201 /// Returns true if the GEP is based on a pointer to a string (array of i8), 202 /// and is indexing into this string. 203 bool isGEPBasedOnPointerToString(const GEPOperator *GEP); 204 205 /// This function computes the length of a null-terminated C string pointed to 206 /// by V. If successful, it returns true and returns the string in Str. If 207 /// unsuccessful, it returns false. This does not include the trailing null 208 /// character by default. If TrimAtNul is set to false, then this returns any 209 /// trailing null characters as well as any other characters that come after 210 /// it. 211 bool getConstantStringInfo(const Value *V, StringRef &Str, 212 uint64_t Offset = 0, bool TrimAtNul = true); 213 214 /// If we can compute the length of the string pointed to by the specified 215 /// pointer, return 'len+1'. If we can't, return 0. 216 uint64_t GetStringLength(Value *V); 217 218 /// This method strips off any GEP address adjustments and pointer casts from 219 /// the specified value, returning the original object being addressed. Note 220 /// that the returned value has pointer type if the specified value does. If 221 /// the MaxLookup value is non-zero, it limits the number of instructions to 222 /// be stripped off. 223 Value *GetUnderlyingObject(Value *V, const DataLayout &DL, 224 unsigned MaxLookup = 6); 225 static inline const Value *GetUnderlyingObject(const Value *V, 226 const DataLayout &DL, 227 unsigned MaxLookup = 6) { 228 return GetUnderlyingObject(const_cast<Value *>(V), DL, MaxLookup); 229 } 230 231 /// \brief This method is similar to GetUnderlyingObject except that it can 232 /// look through phi and select instructions and return multiple objects. 233 /// 234 /// If LoopInfo is passed, loop phis are further analyzed. If a pointer 235 /// accesses different objects in each iteration, we don't look through the 236 /// phi node. E.g. consider this loop nest: 237 /// 238 /// int **A; 239 /// for (i) 240 /// for (j) { 241 /// A[i][j] = A[i-1][j] * B[j] 242 /// } 243 /// 244 /// This is transformed by Load-PRE to stash away A[i] for the next iteration 245 /// of the outer loop: 246 /// 247 /// Curr = A[0]; // Prev_0 248 /// for (i: 1..N) { 249 /// Prev = Curr; // Prev = PHI (Prev_0, Curr) 250 /// Curr = A[i]; 251 /// for (j: 0..N) { 252 /// Curr[j] = Prev[j] * B[j] 253 /// } 254 /// } 255 /// 256 /// Since A[i] and A[i-1] are independent pointers, getUnderlyingObjects 257 /// should not assume that Curr and Prev share the same underlying object thus 258 /// it shouldn't look through the phi above. 259 void GetUnderlyingObjects(Value *V, SmallVectorImpl<Value *> &Objects, 260 const DataLayout &DL, LoopInfo *LI = nullptr, 261 unsigned MaxLookup = 6); 262 263 /// Return true if the only users of this pointer are lifetime markers. 264 bool onlyUsedByLifetimeMarkers(const Value *V); 265 266 /// Return true if the instruction does not have any effects besides 267 /// calculating the result and does not have undefined behavior. 268 /// 269 /// This method never returns true for an instruction that returns true for 270 /// mayHaveSideEffects; however, this method also does some other checks in 271 /// addition. It checks for undefined behavior, like dividing by zero or 272 /// loading from an invalid pointer (but not for undefined results, like a 273 /// shift with a shift amount larger than the width of the result). It checks 274 /// for malloc and alloca because speculatively executing them might cause a 275 /// memory leak. It also returns false for instructions related to control 276 /// flow, specifically terminators and PHI nodes. 277 /// 278 /// If the CtxI is specified this method performs context-sensitive analysis 279 /// and returns true if it is safe to execute the instruction immediately 280 /// before the CtxI. 281 /// 282 /// If the CtxI is NOT specified this method only looks at the instruction 283 /// itself and its operands, so if this method returns true, it is safe to 284 /// move the instruction as long as the correct dominance relationships for 285 /// the operands and users hold. 286 /// 287 /// This method can return true for instructions that read memory; 288 /// for such instructions, moving them may change the resulting value. 289 bool isSafeToSpeculativelyExecute(const Value *V, 290 const Instruction *CtxI = nullptr, 291 const DominatorTree *DT = nullptr); 292 293 /// Returns true if the result or effects of the given instructions \p I 294 /// depend on or influence global memory. 295 /// Memory dependence arises for example if the instruction reads from 296 /// memory or may produce effects or undefined behaviour. Memory dependent 297 /// instructions generally cannot be reorderd with respect to other memory 298 /// dependent instructions or moved into non-dominated basic blocks. 299 /// Instructions which just compute a value based on the values of their 300 /// operands are not memory dependent. 301 bool mayBeMemoryDependent(const Instruction &I); 302 303 /// Return true if this pointer couldn't possibly be null by its definition. 304 /// This returns true for allocas, non-extern-weak globals, and byval 305 /// arguments. 306 bool isKnownNonNull(const Value *V); 307 308 /// Return true if this pointer couldn't possibly be null. If the context 309 /// instruction is specified, perform context-sensitive analysis and return 310 /// true if the pointer couldn't possibly be null at the specified 311 /// instruction. 312 bool isKnownNonNullAt(const Value *V, 313 const Instruction *CtxI = nullptr, 314 const DominatorTree *DT = nullptr); 315 316 /// Return true if it is valid to use the assumptions provided by an 317 /// assume intrinsic, I, at the point in the control-flow identified by the 318 /// context instruction, CxtI. 319 bool isValidAssumeForContext(const Instruction *I, const Instruction *CxtI, 320 const DominatorTree *DT = nullptr); 321 322 enum class OverflowResult { AlwaysOverflows, MayOverflow, NeverOverflows }; 323 OverflowResult computeOverflowForUnsignedMul(Value *LHS, Value *RHS, 324 const DataLayout &DL, 325 AssumptionCache *AC, 326 const Instruction *CxtI, 327 const DominatorTree *DT); 328 OverflowResult computeOverflowForUnsignedAdd(Value *LHS, Value *RHS, 329 const DataLayout &DL, 330 AssumptionCache *AC, 331 const Instruction *CxtI, 332 const DominatorTree *DT); 333 OverflowResult computeOverflowForSignedAdd(Value *LHS, Value *RHS, 334 const DataLayout &DL, 335 AssumptionCache *AC = nullptr, 336 const Instruction *CxtI = nullptr, 337 const DominatorTree *DT = nullptr); 338 /// This version also leverages the sign bit of Add if known. 339 OverflowResult computeOverflowForSignedAdd(AddOperator *Add, 340 const DataLayout &DL, 341 AssumptionCache *AC = nullptr, 342 const Instruction *CxtI = nullptr, 343 const DominatorTree *DT = nullptr); 344 345 /// Returns true if the arithmetic part of the \p II 's result is 346 /// used only along the paths control dependent on the computation 347 /// not overflowing, \p II being an <op>.with.overflow intrinsic. 348 bool isOverflowIntrinsicNoWrap(IntrinsicInst *II, DominatorTree &DT); 349 350 /// Return true if this function can prove that the instruction I will 351 /// always transfer execution to one of its successors (including the next 352 /// instruction that follows within a basic block). E.g. this is not 353 /// guaranteed for function calls that could loop infinitely. 354 /// 355 /// In other words, this function returns false for instructions that may 356 /// transfer execution or fail to transfer execution in a way that is not 357 /// captured in the CFG nor in the sequence of instructions within a basic 358 /// block. 359 /// 360 /// Undefined behavior is assumed not to happen, so e.g. division is 361 /// guaranteed to transfer execution to the following instruction even 362 /// though division by zero might cause undefined behavior. 363 bool isGuaranteedToTransferExecutionToSuccessor(const Instruction *I); 364 365 /// Return true if this function can prove that the instruction I 366 /// is executed for every iteration of the loop L. 367 /// 368 /// Note that this currently only considers the loop header. 369 bool isGuaranteedToExecuteForEveryIteration(const Instruction *I, 370 const Loop *L); 371 372 /// Return true if this function can prove that I is guaranteed to yield 373 /// full-poison (all bits poison) if at least one of its operands are 374 /// full-poison (all bits poison). 375 /// 376 /// The exact rules for how poison propagates through instructions have 377 /// not been settled as of 2015-07-10, so this function is conservative 378 /// and only considers poison to be propagated in uncontroversial 379 /// cases. There is no attempt to track values that may be only partially 380 /// poison. 381 bool propagatesFullPoison(const Instruction *I); 382 383 /// Return either nullptr or an operand of I such that I will trigger 384 /// undefined behavior if I is executed and that operand has a full-poison 385 /// value (all bits poison). 386 const Value *getGuaranteedNonFullPoisonOp(const Instruction *I); 387 388 /// Return true if this function can prove that if PoisonI is executed 389 /// and yields a full-poison value (all bits poison), then that will 390 /// trigger undefined behavior. 391 /// 392 /// Note that this currently only considers the basic block that is 393 /// the parent of I. 394 bool isKnownNotFullPoison(const Instruction *PoisonI); 395 396 /// \brief Specific patterns of select instructions we can match. 397 enum SelectPatternFlavor { 398 SPF_UNKNOWN = 0, 399 SPF_SMIN, /// Signed minimum 400 SPF_UMIN, /// Unsigned minimum 401 SPF_SMAX, /// Signed maximum 402 SPF_UMAX, /// Unsigned maximum 403 SPF_FMINNUM, /// Floating point minnum 404 SPF_FMAXNUM, /// Floating point maxnum 405 SPF_ABS, /// Absolute value 406 SPF_NABS /// Negated absolute value 407 }; 408 /// \brief Behavior when a floating point min/max is given one NaN and one 409 /// non-NaN as input. 410 enum SelectPatternNaNBehavior { 411 SPNB_NA = 0, /// NaN behavior not applicable. 412 SPNB_RETURNS_NAN, /// Given one NaN input, returns the NaN. 413 SPNB_RETURNS_OTHER, /// Given one NaN input, returns the non-NaN. 414 SPNB_RETURNS_ANY /// Given one NaN input, can return either (or 415 /// it has been determined that no operands can 416 /// be NaN). 417 }; 418 struct SelectPatternResult { 419 SelectPatternFlavor Flavor; 420 SelectPatternNaNBehavior NaNBehavior; /// Only applicable if Flavor is 421 /// SPF_FMINNUM or SPF_FMAXNUM. 422 bool Ordered; /// When implementing this min/max pattern as 423 /// fcmp; select, does the fcmp have to be 424 /// ordered? 425 426 /// \brief Return true if \p SPF is a min or a max pattern. isMinOrMaxSelectPatternResult427 static bool isMinOrMax(SelectPatternFlavor SPF) { 428 return !(SPF == SPF_UNKNOWN || SPF == SPF_ABS || SPF == SPF_NABS); 429 } 430 }; 431 /// Pattern match integer [SU]MIN, [SU]MAX and ABS idioms, returning the kind 432 /// and providing the out parameter results if we successfully match. 433 /// 434 /// If CastOp is not nullptr, also match MIN/MAX idioms where the type does 435 /// not match that of the original select. If this is the case, the cast 436 /// operation (one of Trunc,SExt,Zext) that must be done to transform the 437 /// type of LHS and RHS into the type of V is returned in CastOp. 438 /// 439 /// For example: 440 /// %1 = icmp slt i32 %a, i32 4 441 /// %2 = sext i32 %a to i64 442 /// %3 = select i1 %1, i64 %2, i64 4 443 /// 444 /// -> LHS = %a, RHS = i32 4, *CastOp = Instruction::SExt 445 /// 446 SelectPatternResult matchSelectPattern(Value *V, Value *&LHS, Value *&RHS, 447 Instruction::CastOps *CastOp = nullptr); 448 449 /// Parse out a conservative ConstantRange from !range metadata. 450 /// 451 /// E.g. if RangeMD is !{i32 0, i32 10, i32 15, i32 20} then return [0, 20). 452 ConstantRange getConstantRangeFromMetadata(MDNode &RangeMD); 453 454 /// Return true if RHS is known to be implied true by LHS. Return false if 455 /// RHS is known to be implied false by LHS. Otherwise, return None if no 456 /// implication can be made. 457 /// A & B must be i1 (boolean) values or a vector of such values. Note that 458 /// the truth table for implication is the same as <=u on i1 values (but not 459 /// <=s!). The truth table for both is: 460 /// | T | F (B) 461 /// T | T | F 462 /// F | T | T 463 /// (A) 464 Optional<bool> isImpliedCondition( 465 Value *LHS, Value *RHS, const DataLayout &DL, bool InvertAPred = false, 466 unsigned Depth = 0, AssumptionCache *AC = nullptr, 467 const Instruction *CxtI = nullptr, const DominatorTree *DT = nullptr); 468 } // end namespace llvm 469 470 #endif 471