xref: /netbsd-src/external/apache2/llvm/dist/llvm/include/llvm/Analysis/ValueTracking.h (revision 82d56013d7b633d116a93943de88e08335357a7c)
1 //===- llvm/Analysis/ValueTracking.h - Walk computations --------*- 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 // This file contains routines that help analyze properties that chains of
10 // computations have.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #ifndef LLVM_ANALYSIS_VALUETRACKING_H
15 #define LLVM_ANALYSIS_VALUETRACKING_H
16 
17 #include "llvm/ADT/ArrayRef.h"
18 #include "llvm/ADT/Optional.h"
19 #include "llvm/ADT/SmallSet.h"
20 #include "llvm/IR/Constants.h"
21 #include "llvm/IR/DataLayout.h"
22 #include "llvm/IR/InstrTypes.h"
23 #include "llvm/IR/Intrinsics.h"
24 #include "llvm/IR/Operator.h"
25 #include <cassert>
26 #include <cstdint>
27 
28 namespace llvm {
29 
30 class AddOperator;
31 class AllocaInst;
32 class APInt;
33 class AssumptionCache;
34 class DominatorTree;
35 class GEPOperator;
36 class IntrinsicInst;
37 class LoadInst;
38 class WithOverflowInst;
39 struct KnownBits;
40 class Loop;
41 class LoopInfo;
42 class MDNode;
43 class OptimizationRemarkEmitter;
44 class StringRef;
45 class TargetLibraryInfo;
46 class Value;
47 
48 constexpr unsigned MaxAnalysisRecursionDepth = 6;
49 
50   /// Determine which bits of V are known to be either zero or one and return
51   /// them in the KnownZero/KnownOne bit sets.
52   ///
53   /// This function is defined on values with integer type, values with pointer
54   /// type, and vectors of integers.  In the case
55   /// where V is a vector, the known zero and known one values are the
56   /// same width as the vector element, and the bit is set only if it is true
57   /// for all of the elements in the vector.
58   void computeKnownBits(const Value *V, KnownBits &Known,
59                         const DataLayout &DL, unsigned Depth = 0,
60                         AssumptionCache *AC = nullptr,
61                         const Instruction *CxtI = nullptr,
62                         const DominatorTree *DT = nullptr,
63                         OptimizationRemarkEmitter *ORE = nullptr,
64                         bool UseInstrInfo = true);
65 
66   /// Determine which bits of V are known to be either zero or one and return
67   /// them in the KnownZero/KnownOne bit sets.
68   ///
69   /// This function is defined on values with integer type, values with pointer
70   /// type, and vectors of integers.  In the case
71   /// where V is a vector, the known zero and known one values are the
72   /// same width as the vector element, and the bit is set only if it is true
73   /// for all of the demanded elements in the vector.
74   void computeKnownBits(const Value *V, const APInt &DemandedElts,
75                         KnownBits &Known, const DataLayout &DL,
76                         unsigned Depth = 0, AssumptionCache *AC = nullptr,
77                         const Instruction *CxtI = nullptr,
78                         const DominatorTree *DT = nullptr,
79                         OptimizationRemarkEmitter *ORE = nullptr,
80                         bool UseInstrInfo = true);
81 
82   /// Returns the known bits rather than passing by reference.
83   KnownBits computeKnownBits(const Value *V, const DataLayout &DL,
84                              unsigned Depth = 0, AssumptionCache *AC = nullptr,
85                              const Instruction *CxtI = nullptr,
86                              const DominatorTree *DT = nullptr,
87                              OptimizationRemarkEmitter *ORE = nullptr,
88                              bool UseInstrInfo = true);
89 
90   /// Returns the known bits rather than passing by reference.
91   KnownBits computeKnownBits(const Value *V, const APInt &DemandedElts,
92                              const DataLayout &DL, unsigned Depth = 0,
93                              AssumptionCache *AC = nullptr,
94                              const Instruction *CxtI = nullptr,
95                              const DominatorTree *DT = nullptr,
96                              OptimizationRemarkEmitter *ORE = nullptr,
97                              bool UseInstrInfo = true);
98 
99   /// Compute known bits from the range metadata.
100   /// \p KnownZero the set of bits that are known to be zero
101   /// \p KnownOne the set of bits that are known to be one
102   void computeKnownBitsFromRangeMetadata(const MDNode &Ranges,
103                                          KnownBits &Known);
104 
105   /// Return true if LHS and RHS have no common bits set.
106   bool haveNoCommonBitsSet(const Value *LHS, const Value *RHS,
107                            const DataLayout &DL,
108                            AssumptionCache *AC = nullptr,
109                            const Instruction *CxtI = nullptr,
110                            const DominatorTree *DT = nullptr,
111                            bool UseInstrInfo = true);
112 
113   /// Return true if the given value is known to have exactly one bit set when
114   /// defined. For vectors return true if every element is known to be a power
115   /// of two when defined. Supports values with integer or pointer type and
116   /// vectors of integers. If 'OrZero' is set, then return true if the given
117   /// value is either a power of two or zero.
118   bool isKnownToBeAPowerOfTwo(const Value *V, const DataLayout &DL,
119                               bool OrZero = false, unsigned Depth = 0,
120                               AssumptionCache *AC = nullptr,
121                               const Instruction *CxtI = nullptr,
122                               const DominatorTree *DT = nullptr,
123                               bool UseInstrInfo = true);
124 
125   bool isOnlyUsedInZeroEqualityComparison(const Instruction *CxtI);
126 
127   /// Return true if the given value is known to be non-zero when defined. For
128   /// vectors, return true if every element is known to be non-zero when
129   /// defined. For pointers, if the context instruction and dominator tree are
130   /// specified, perform context-sensitive analysis and return true if the
131   /// pointer couldn't possibly be null at the specified instruction.
132   /// Supports values with integer or pointer type and vectors of integers.
133   bool isKnownNonZero(const Value *V, const DataLayout &DL, unsigned Depth = 0,
134                       AssumptionCache *AC = nullptr,
135                       const Instruction *CxtI = nullptr,
136                       const DominatorTree *DT = nullptr,
137                       bool UseInstrInfo = true);
138 
139   /// Return true if the two given values are negation.
140   /// Currently can recoginze Value pair:
141   /// 1: <X, Y> if X = sub (0, Y) or Y = sub (0, X)
142   /// 2: <X, Y> if X = sub (A, B) and Y = sub (B, A)
143   bool isKnownNegation(const Value *X, const Value *Y, bool NeedNSW = false);
144 
145   /// Returns true if the give value is known to be non-negative.
146   bool isKnownNonNegative(const Value *V, const DataLayout &DL,
147                           unsigned Depth = 0,
148                           AssumptionCache *AC = nullptr,
149                           const Instruction *CxtI = nullptr,
150                           const DominatorTree *DT = nullptr,
151                           bool UseInstrInfo = true);
152 
153   /// Returns true if the given value is known be positive (i.e. non-negative
154   /// and non-zero).
155   bool isKnownPositive(const Value *V, const DataLayout &DL, unsigned Depth = 0,
156                        AssumptionCache *AC = nullptr,
157                        const Instruction *CxtI = nullptr,
158                        const DominatorTree *DT = nullptr,
159                        bool UseInstrInfo = true);
160 
161   /// Returns true if the given value is known be negative (i.e. non-positive
162   /// and non-zero).
163   bool isKnownNegative(const Value *V, const DataLayout &DL, unsigned Depth = 0,
164                        AssumptionCache *AC = nullptr,
165                        const Instruction *CxtI = nullptr,
166                        const DominatorTree *DT = nullptr,
167                        bool UseInstrInfo = true);
168 
169   /// Return true if the given values are known to be non-equal when defined.
170   /// Supports scalar integer types only.
171   bool isKnownNonEqual(const Value *V1, const Value *V2, const DataLayout &DL,
172                        AssumptionCache *AC = nullptr,
173                        const Instruction *CxtI = nullptr,
174                        const DominatorTree *DT = nullptr,
175                        bool UseInstrInfo = true);
176 
177   /// Return true if 'V & Mask' is known to be zero. We use this predicate to
178   /// simplify operations downstream. Mask is known to be zero for bits that V
179   /// cannot have.
180   ///
181   /// This function is defined on values with integer type, values with pointer
182   /// type, and vectors of integers.  In the case
183   /// where V is a vector, the mask, known zero, and known one values are the
184   /// same width as the vector element, and the bit is set only if it is true
185   /// for all of the elements in the vector.
186   bool MaskedValueIsZero(const Value *V, const APInt &Mask,
187                          const DataLayout &DL,
188                          unsigned Depth = 0, AssumptionCache *AC = nullptr,
189                          const Instruction *CxtI = nullptr,
190                          const DominatorTree *DT = nullptr,
191                          bool UseInstrInfo = true);
192 
193   /// Return the number of times the sign bit of the register is replicated into
194   /// the other bits. We know that at least 1 bit is always equal to the sign
195   /// bit (itself), but other cases can give us information. For example,
196   /// immediately after an "ashr X, 2", we know that the top 3 bits are all
197   /// equal to each other, so we return 3. For vectors, return the number of
198   /// sign bits for the vector element with the mininum number of known sign
199   /// bits.
200   unsigned ComputeNumSignBits(const Value *Op, const DataLayout &DL,
201                               unsigned Depth = 0, AssumptionCache *AC = nullptr,
202                               const Instruction *CxtI = nullptr,
203                               const DominatorTree *DT = nullptr,
204                               bool UseInstrInfo = true);
205 
206   /// This function computes the integer multiple of Base that equals V. If
207   /// successful, it returns true and returns the multiple in Multiple. If
208   /// unsuccessful, it returns false. Also, if V can be simplified to an
209   /// integer, then the simplified V is returned in Val. Look through sext only
210   /// if LookThroughSExt=true.
211   bool ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
212                        bool LookThroughSExt = false,
213                        unsigned Depth = 0);
214 
215   /// Map a call instruction to an intrinsic ID.  Libcalls which have equivalent
216   /// intrinsics are treated as-if they were intrinsics.
217   Intrinsic::ID getIntrinsicForCallSite(const CallBase &CB,
218                                         const TargetLibraryInfo *TLI);
219 
220   /// Return true if we can prove that the specified FP value is never equal to
221   /// -0.0.
222   bool CannotBeNegativeZero(const Value *V, const TargetLibraryInfo *TLI,
223                             unsigned Depth = 0);
224 
225   /// Return true if we can prove that the specified FP value is either NaN or
226   /// never less than -0.0.
227   ///
228   ///      NaN --> true
229   ///       +0 --> true
230   ///       -0 --> true
231   ///   x > +0 --> true
232   ///   x < -0 --> false
233   bool CannotBeOrderedLessThanZero(const Value *V, const TargetLibraryInfo *TLI);
234 
235   /// Return true if the floating-point scalar value is not an infinity or if
236   /// the floating-point vector value has no infinities. Return false if a value
237   /// could ever be infinity.
238   bool isKnownNeverInfinity(const Value *V, const TargetLibraryInfo *TLI,
239                             unsigned Depth = 0);
240 
241   /// Return true if the floating-point scalar value is not a NaN or if the
242   /// floating-point vector value has no NaN elements. Return false if a value
243   /// could ever be NaN.
244   bool isKnownNeverNaN(const Value *V, const TargetLibraryInfo *TLI,
245                        unsigned Depth = 0);
246 
247   /// Return true if we can prove that the specified FP value's sign bit is 0.
248   ///
249   ///      NaN --> true/false (depending on the NaN's sign bit)
250   ///       +0 --> true
251   ///       -0 --> false
252   ///   x > +0 --> true
253   ///   x < -0 --> false
254   bool SignBitMustBeZero(const Value *V, const TargetLibraryInfo *TLI);
255 
256   /// If the specified value can be set by repeating the same byte in memory,
257   /// return the i8 value that it is represented with. This is true for all i8
258   /// values obviously, but is also true for i32 0, i32 -1, i16 0xF0F0, double
259   /// 0.0 etc. If the value can't be handled with a repeated byte store (e.g.
260   /// i16 0x1234), return null. If the value is entirely undef and padding,
261   /// return undef.
262   Value *isBytewiseValue(Value *V, const DataLayout &DL);
263 
264   /// Given an aggregate and an sequence of indices, see if the scalar value
265   /// indexed is already around as a register, for example if it were inserted
266   /// directly into the aggregate.
267   ///
268   /// If InsertBefore is not null, this function will duplicate (modified)
269   /// insertvalues when a part of a nested struct is extracted.
270   Value *FindInsertedValue(Value *V,
271                            ArrayRef<unsigned> idx_range,
272                            Instruction *InsertBefore = nullptr);
273 
274   /// Analyze the specified pointer to see if it can be expressed as a base
275   /// pointer plus a constant offset. Return the base and offset to the caller.
276   ///
277   /// This is a wrapper around Value::stripAndAccumulateConstantOffsets that
278   /// creates and later unpacks the required APInt.
279   inline Value *GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
280                                                  const DataLayout &DL,
281                                                  bool AllowNonInbounds = true) {
282     APInt OffsetAPInt(DL.getIndexTypeSizeInBits(Ptr->getType()), 0);
283     Value *Base =
284         Ptr->stripAndAccumulateConstantOffsets(DL, OffsetAPInt, AllowNonInbounds);
285 
286     Offset = OffsetAPInt.getSExtValue();
287     return Base;
288   }
289   inline const Value *
290   GetPointerBaseWithConstantOffset(const Value *Ptr, int64_t &Offset,
291                                    const DataLayout &DL,
292                                    bool AllowNonInbounds = true) {
293     return GetPointerBaseWithConstantOffset(const_cast<Value *>(Ptr), Offset, DL,
294                                             AllowNonInbounds);
295   }
296 
297   /// Returns true if the GEP is based on a pointer to a string (array of
298   // \p CharSize integers) and is indexing into this string.
299   bool isGEPBasedOnPointerToString(const GEPOperator *GEP,
300                                    unsigned CharSize = 8);
301 
302   /// Represents offset+length into a ConstantDataArray.
303   struct ConstantDataArraySlice {
304     /// ConstantDataArray pointer. nullptr indicates a zeroinitializer (a valid
305     /// initializer, it just doesn't fit the ConstantDataArray interface).
306     const ConstantDataArray *Array;
307 
308     /// Slice starts at this Offset.
309     uint64_t Offset;
310 
311     /// Length of the slice.
312     uint64_t Length;
313 
314     /// Moves the Offset and adjusts Length accordingly.
moveConstantDataArraySlice315     void move(uint64_t Delta) {
316       assert(Delta < Length);
317       Offset += Delta;
318       Length -= Delta;
319     }
320 
321     /// Convenience accessor for elements in the slice.
322     uint64_t operator[](unsigned I) const {
323       return Array==nullptr ? 0 : Array->getElementAsInteger(I + Offset);
324     }
325   };
326 
327   /// Returns true if the value \p V is a pointer into a ConstantDataArray.
328   /// If successful \p Slice will point to a ConstantDataArray info object
329   /// with an appropriate offset.
330   bool getConstantDataArrayInfo(const Value *V, ConstantDataArraySlice &Slice,
331                                 unsigned ElementSize, uint64_t Offset = 0);
332 
333   /// This function computes the length of a null-terminated C string pointed to
334   /// by V. If successful, it returns true and returns the string in Str. If
335   /// unsuccessful, it returns false. This does not include the trailing null
336   /// character by default. If TrimAtNul is set to false, then this returns any
337   /// trailing null characters as well as any other characters that come after
338   /// it.
339   bool getConstantStringInfo(const Value *V, StringRef &Str,
340                              uint64_t Offset = 0, bool TrimAtNul = true);
341 
342   /// If we can compute the length of the string pointed to by the specified
343   /// pointer, return 'len+1'.  If we can't, return 0.
344   uint64_t GetStringLength(const Value *V, unsigned CharSize = 8);
345 
346   /// This function returns call pointer argument that is considered the same by
347   /// aliasing rules. You CAN'T use it to replace one value with another. If
348   /// \p MustPreserveNullness is true, the call must preserve the nullness of
349   /// the pointer.
350   const Value *getArgumentAliasingToReturnedPointer(const CallBase *Call,
351                                                     bool MustPreserveNullness);
352   inline Value *
getArgumentAliasingToReturnedPointer(CallBase * Call,bool MustPreserveNullness)353   getArgumentAliasingToReturnedPointer(CallBase *Call,
354                                        bool MustPreserveNullness) {
355     return const_cast<Value *>(getArgumentAliasingToReturnedPointer(
356         const_cast<const CallBase *>(Call), MustPreserveNullness));
357   }
358 
359   /// {launder,strip}.invariant.group returns pointer that aliases its argument,
360   /// and it only captures pointer by returning it.
361   /// These intrinsics are not marked as nocapture, because returning is
362   /// considered as capture. The arguments are not marked as returned neither,
363   /// because it would make it useless. If \p MustPreserveNullness is true,
364   /// the intrinsic must preserve the nullness of the pointer.
365   bool isIntrinsicReturningPointerAliasingArgumentWithoutCapturing(
366       const CallBase *Call, bool MustPreserveNullness);
367 
368   /// This method strips off any GEP address adjustments and pointer casts from
369   /// the specified value, returning the original object being addressed. Note
370   /// that the returned value has pointer type if the specified value does. If
371   /// the MaxLookup value is non-zero, it limits the number of instructions to
372   /// be stripped off.
373   const Value *getUnderlyingObject(const Value *V, unsigned MaxLookup = 6);
374   inline Value *getUnderlyingObject(Value *V, unsigned MaxLookup = 6) {
375     // Force const to avoid infinite recursion.
376     const Value *VConst = V;
377     return const_cast<Value *>(getUnderlyingObject(VConst, MaxLookup));
378   }
379 
380   /// This method is similar to getUnderlyingObject except that it can
381   /// look through phi and select instructions and return multiple objects.
382   ///
383   /// If LoopInfo is passed, loop phis are further analyzed.  If a pointer
384   /// accesses different objects in each iteration, we don't look through the
385   /// phi node. E.g. consider this loop nest:
386   ///
387   ///   int **A;
388   ///   for (i)
389   ///     for (j) {
390   ///        A[i][j] = A[i-1][j] * B[j]
391   ///     }
392   ///
393   /// This is transformed by Load-PRE to stash away A[i] for the next iteration
394   /// of the outer loop:
395   ///
396   ///   Curr = A[0];          // Prev_0
397   ///   for (i: 1..N) {
398   ///     Prev = Curr;        // Prev = PHI (Prev_0, Curr)
399   ///     Curr = A[i];
400   ///     for (j: 0..N) {
401   ///        Curr[j] = Prev[j] * B[j]
402   ///     }
403   ///   }
404   ///
405   /// Since A[i] and A[i-1] are independent pointers, getUnderlyingObjects
406   /// should not assume that Curr and Prev share the same underlying object thus
407   /// it shouldn't look through the phi above.
408   void getUnderlyingObjects(const Value *V,
409                             SmallVectorImpl<const Value *> &Objects,
410                             LoopInfo *LI = nullptr, unsigned MaxLookup = 6);
411 
412   /// This is a wrapper around getUnderlyingObjects and adds support for basic
413   /// ptrtoint+arithmetic+inttoptr sequences.
414   bool getUnderlyingObjectsForCodeGen(const Value *V,
415                                       SmallVectorImpl<Value *> &Objects);
416 
417   /// Returns unique alloca where the value comes from, or nullptr.
418   /// If OffsetZero is true check that V points to the begining of the alloca.
419   AllocaInst *findAllocaForValue(Value *V, bool OffsetZero = false);
420   inline const AllocaInst *findAllocaForValue(const Value *V,
421                                               bool OffsetZero = false) {
422     return findAllocaForValue(const_cast<Value *>(V), OffsetZero);
423   }
424 
425   /// Return true if the only users of this pointer are lifetime markers.
426   bool onlyUsedByLifetimeMarkers(const Value *V);
427 
428   /// Return true if the only users of this pointer are lifetime markers or
429   /// droppable instructions.
430   bool onlyUsedByLifetimeMarkersOrDroppableInsts(const Value *V);
431 
432   /// Return true if speculation of the given load must be suppressed to avoid
433   /// ordering or interfering with an active sanitizer.  If not suppressed,
434   /// dereferenceability and alignment must be proven separately.  Note: This
435   /// is only needed for raw reasoning; if you use the interface below
436   /// (isSafeToSpeculativelyExecute), this is handled internally.
437   bool mustSuppressSpeculation(const LoadInst &LI);
438 
439   /// Return true if the instruction does not have any effects besides
440   /// calculating the result and does not have undefined behavior.
441   ///
442   /// This method never returns true for an instruction that returns true for
443   /// mayHaveSideEffects; however, this method also does some other checks in
444   /// addition. It checks for undefined behavior, like dividing by zero or
445   /// loading from an invalid pointer (but not for undefined results, like a
446   /// shift with a shift amount larger than the width of the result). It checks
447   /// for malloc and alloca because speculatively executing them might cause a
448   /// memory leak. It also returns false for instructions related to control
449   /// flow, specifically terminators and PHI nodes.
450   ///
451   /// If the CtxI is specified this method performs context-sensitive analysis
452   /// and returns true if it is safe to execute the instruction immediately
453   /// before the CtxI.
454   ///
455   /// If the CtxI is NOT specified this method only looks at the instruction
456   /// itself and its operands, so if this method returns true, it is safe to
457   /// move the instruction as long as the correct dominance relationships for
458   /// the operands and users hold.
459   ///
460   /// This method can return true for instructions that read memory;
461   /// for such instructions, moving them may change the resulting value.
462   bool isSafeToSpeculativelyExecute(const Value *V,
463                                     const Instruction *CtxI = nullptr,
464                                     const DominatorTree *DT = nullptr,
465                                     const TargetLibraryInfo *TLI = nullptr);
466 
467   /// Returns true if the result or effects of the given instructions \p I
468   /// depend on or influence global memory.
469   /// Memory dependence arises for example if the instruction reads from
470   /// memory or may produce effects or undefined behaviour. Memory dependent
471   /// instructions generally cannot be reorderd with respect to other memory
472   /// dependent instructions or moved into non-dominated basic blocks.
473   /// Instructions which just compute a value based on the values of their
474   /// operands are not memory dependent.
475   bool mayBeMemoryDependent(const Instruction &I);
476 
477   /// Return true if it is an intrinsic that cannot be speculated but also
478   /// cannot trap.
479   bool isAssumeLikeIntrinsic(const Instruction *I);
480 
481   /// Return true if it is valid to use the assumptions provided by an
482   /// assume intrinsic, I, at the point in the control-flow identified by the
483   /// context instruction, CxtI.
484   bool isValidAssumeForContext(const Instruction *I, const Instruction *CxtI,
485                                const DominatorTree *DT = nullptr);
486 
487   enum class OverflowResult {
488     /// Always overflows in the direction of signed/unsigned min value.
489     AlwaysOverflowsLow,
490     /// Always overflows in the direction of signed/unsigned max value.
491     AlwaysOverflowsHigh,
492     /// May or may not overflow.
493     MayOverflow,
494     /// Never overflows.
495     NeverOverflows,
496   };
497 
498   OverflowResult computeOverflowForUnsignedMul(const Value *LHS,
499                                                const Value *RHS,
500                                                const DataLayout &DL,
501                                                AssumptionCache *AC,
502                                                const Instruction *CxtI,
503                                                const DominatorTree *DT,
504                                                bool UseInstrInfo = true);
505   OverflowResult computeOverflowForSignedMul(const Value *LHS, const Value *RHS,
506                                              const DataLayout &DL,
507                                              AssumptionCache *AC,
508                                              const Instruction *CxtI,
509                                              const DominatorTree *DT,
510                                              bool UseInstrInfo = true);
511   OverflowResult computeOverflowForUnsignedAdd(const Value *LHS,
512                                                const Value *RHS,
513                                                const DataLayout &DL,
514                                                AssumptionCache *AC,
515                                                const Instruction *CxtI,
516                                                const DominatorTree *DT,
517                                                bool UseInstrInfo = true);
518   OverflowResult computeOverflowForSignedAdd(const Value *LHS, const Value *RHS,
519                                              const DataLayout &DL,
520                                              AssumptionCache *AC = nullptr,
521                                              const Instruction *CxtI = nullptr,
522                                              const DominatorTree *DT = nullptr);
523   /// This version also leverages the sign bit of Add if known.
524   OverflowResult computeOverflowForSignedAdd(const AddOperator *Add,
525                                              const DataLayout &DL,
526                                              AssumptionCache *AC = nullptr,
527                                              const Instruction *CxtI = nullptr,
528                                              const DominatorTree *DT = nullptr);
529   OverflowResult computeOverflowForUnsignedSub(const Value *LHS, const Value *RHS,
530                                                const DataLayout &DL,
531                                                AssumptionCache *AC,
532                                                const Instruction *CxtI,
533                                                const DominatorTree *DT);
534   OverflowResult computeOverflowForSignedSub(const Value *LHS, const Value *RHS,
535                                              const DataLayout &DL,
536                                              AssumptionCache *AC,
537                                              const Instruction *CxtI,
538                                              const DominatorTree *DT);
539 
540   /// Returns true if the arithmetic part of the \p WO 's result is
541   /// used only along the paths control dependent on the computation
542   /// not overflowing, \p WO being an <op>.with.overflow intrinsic.
543   bool isOverflowIntrinsicNoWrap(const WithOverflowInst *WO,
544                                  const DominatorTree &DT);
545 
546 
547   /// Determine the possible constant range of an integer or vector of integer
548   /// value. This is intended as a cheap, non-recursive check.
549   ConstantRange computeConstantRange(const Value *V, bool UseInstrInfo = true,
550                                      AssumptionCache *AC = nullptr,
551                                      const Instruction *CtxI = nullptr,
552                                      unsigned Depth = 0);
553 
554   /// Return true if this function can prove that the instruction I will
555   /// always transfer execution to one of its successors (including the next
556   /// instruction that follows within a basic block). E.g. this is not
557   /// guaranteed for function calls that could loop infinitely.
558   ///
559   /// In other words, this function returns false for instructions that may
560   /// transfer execution or fail to transfer execution in a way that is not
561   /// captured in the CFG nor in the sequence of instructions within a basic
562   /// block.
563   ///
564   /// Undefined behavior is assumed not to happen, so e.g. division is
565   /// guaranteed to transfer execution to the following instruction even
566   /// though division by zero might cause undefined behavior.
567   bool isGuaranteedToTransferExecutionToSuccessor(const Instruction *I);
568 
569   /// Returns true if this block does not contain a potential implicit exit.
570   /// This is equivelent to saying that all instructions within the basic block
571   /// are guaranteed to transfer execution to their successor within the basic
572   /// block. This has the same assumptions w.r.t. undefined behavior as the
573   /// instruction variant of this function.
574   bool isGuaranteedToTransferExecutionToSuccessor(const BasicBlock *BB);
575 
576   /// Return true if this function can prove that the instruction I
577   /// is executed for every iteration of the loop L.
578   ///
579   /// Note that this currently only considers the loop header.
580   bool isGuaranteedToExecuteForEveryIteration(const Instruction *I,
581                                               const Loop *L);
582 
583   /// Return true if I yields poison or raises UB if any of its operands is
584   /// poison.
585   /// Formally, given I = `r = op v1 v2 .. vN`, propagatesPoison returns true
586   /// if, for all i, r is evaluated to poison or op raises UB if vi = poison.
587   /// If vi is a vector or an aggregate and r is a single value, any poison
588   /// element in vi should make r poison or raise UB.
589   /// To filter out operands that raise UB on poison, you can use
590   /// getGuaranteedNonPoisonOp.
591   bool propagatesPoison(const Operator *I);
592 
593   /// Insert operands of I into Ops such that I will trigger undefined behavior
594   /// if I is executed and that operand has a poison value.
595   void getGuaranteedNonPoisonOps(const Instruction *I,
596                                  SmallPtrSetImpl<const Value *> &Ops);
597   /// Insert operands of I into Ops such that I will trigger undefined behavior
598   /// if I is executed and that operand is not a well-defined value
599   /// (i.e. has undef bits or poison).
600   void getGuaranteedWellDefinedOps(const Instruction *I,
601                                    SmallPtrSetImpl<const Value *> &Ops);
602 
603   /// Return true if the given instruction must trigger undefined behavior
604   /// when I is executed with any operands which appear in KnownPoison holding
605   /// a poison value at the point of execution.
606   bool mustTriggerUB(const Instruction *I,
607                      const SmallSet<const Value *, 16>& KnownPoison);
608 
609   /// Return true if this function can prove that if Inst is executed
610   /// and yields a poison value or undef bits, then that will trigger
611   /// undefined behavior.
612   ///
613   /// Note that this currently only considers the basic block that is
614   /// the parent of Inst.
615   bool programUndefinedIfUndefOrPoison(const Instruction *Inst);
616   bool programUndefinedIfPoison(const Instruction *Inst);
617 
618   /// canCreateUndefOrPoison returns true if Op can create undef or poison from
619   /// non-undef & non-poison operands.
620   /// For vectors, canCreateUndefOrPoison returns true if there is potential
621   /// poison or undef in any element of the result when vectors without
622   /// undef/poison poison are given as operands.
623   /// For example, given `Op = shl <2 x i32> %x, <0, 32>`, this function returns
624   /// true. If Op raises immediate UB but never creates poison or undef
625   /// (e.g. sdiv I, 0), canCreatePoison returns false.
626   ///
627   /// canCreatePoison returns true if Op can create poison from non-poison
628   /// operands.
629   bool canCreateUndefOrPoison(const Operator *Op);
630   bool canCreatePoison(const Operator *Op);
631 
632   /// Return true if V is poison given that ValAssumedPoison is already poison.
633   /// For example, if ValAssumedPoison is `icmp X, 10` and V is `icmp X, 5`,
634   /// impliesPoison returns true.
635   bool impliesPoison(const Value *ValAssumedPoison, const Value *V);
636 
637   /// Return true if this function can prove that V does not have undef bits
638   /// and is never poison. If V is an aggregate value or vector, check whether
639   /// all elements (except padding) are not undef or poison.
640   /// Note that this is different from canCreateUndefOrPoison because the
641   /// function assumes Op's operands are not poison/undef.
642   ///
643   /// If CtxI and DT are specified this method performs flow-sensitive analysis
644   /// and returns true if it is guaranteed to be never undef or poison
645   /// immediately before the CtxI.
646   bool isGuaranteedNotToBeUndefOrPoison(const Value *V,
647                                         AssumptionCache *AC = nullptr,
648                                         const Instruction *CtxI = nullptr,
649                                         const DominatorTree *DT = nullptr,
650                                         unsigned Depth = 0);
651   bool isGuaranteedNotToBePoison(const Value *V, AssumptionCache *AC = nullptr,
652                                  const Instruction *CtxI = nullptr,
653                                  const DominatorTree *DT = nullptr,
654                                  unsigned Depth = 0);
655 
656   /// Specific patterns of select instructions we can match.
657   enum SelectPatternFlavor {
658     SPF_UNKNOWN = 0,
659     SPF_SMIN,                   /// Signed minimum
660     SPF_UMIN,                   /// Unsigned minimum
661     SPF_SMAX,                   /// Signed maximum
662     SPF_UMAX,                   /// Unsigned maximum
663     SPF_FMINNUM,                /// Floating point minnum
664     SPF_FMAXNUM,                /// Floating point maxnum
665     SPF_ABS,                    /// Absolute value
666     SPF_NABS                    /// Negated absolute value
667   };
668 
669   /// Behavior when a floating point min/max is given one NaN and one
670   /// non-NaN as input.
671   enum SelectPatternNaNBehavior {
672     SPNB_NA = 0,                /// NaN behavior not applicable.
673     SPNB_RETURNS_NAN,           /// Given one NaN input, returns the NaN.
674     SPNB_RETURNS_OTHER,         /// Given one NaN input, returns the non-NaN.
675     SPNB_RETURNS_ANY            /// Given one NaN input, can return either (or
676                                 /// it has been determined that no operands can
677                                 /// be NaN).
678   };
679 
680   struct SelectPatternResult {
681     SelectPatternFlavor Flavor;
682     SelectPatternNaNBehavior NaNBehavior; /// Only applicable if Flavor is
683                                           /// SPF_FMINNUM or SPF_FMAXNUM.
684     bool Ordered;               /// When implementing this min/max pattern as
685                                 /// fcmp; select, does the fcmp have to be
686                                 /// ordered?
687 
688     /// Return true if \p SPF is a min or a max pattern.
isMinOrMaxSelectPatternResult689     static bool isMinOrMax(SelectPatternFlavor SPF) {
690       return SPF != SPF_UNKNOWN && SPF != SPF_ABS && SPF != SPF_NABS;
691     }
692   };
693 
694   /// Pattern match integer [SU]MIN, [SU]MAX and ABS idioms, returning the kind
695   /// and providing the out parameter results if we successfully match.
696   ///
697   /// For ABS/NABS, LHS will be set to the input to the abs idiom. RHS will be
698   /// the negation instruction from the idiom.
699   ///
700   /// If CastOp is not nullptr, also match MIN/MAX idioms where the type does
701   /// not match that of the original select. If this is the case, the cast
702   /// operation (one of Trunc,SExt,Zext) that must be done to transform the
703   /// type of LHS and RHS into the type of V is returned in CastOp.
704   ///
705   /// For example:
706   ///   %1 = icmp slt i32 %a, i32 4
707   ///   %2 = sext i32 %a to i64
708   ///   %3 = select i1 %1, i64 %2, i64 4
709   ///
710   /// -> LHS = %a, RHS = i32 4, *CastOp = Instruction::SExt
711   ///
712   SelectPatternResult matchSelectPattern(Value *V, Value *&LHS, Value *&RHS,
713                                          Instruction::CastOps *CastOp = nullptr,
714                                          unsigned Depth = 0);
715 
716   inline SelectPatternResult
matchSelectPattern(const Value * V,const Value * & LHS,const Value * & RHS)717   matchSelectPattern(const Value *V, const Value *&LHS, const Value *&RHS) {
718     Value *L = const_cast<Value *>(LHS);
719     Value *R = const_cast<Value *>(RHS);
720     auto Result = matchSelectPattern(const_cast<Value *>(V), L, R);
721     LHS = L;
722     RHS = R;
723     return Result;
724   }
725 
726   /// Determine the pattern that a select with the given compare as its
727   /// predicate and given values as its true/false operands would match.
728   SelectPatternResult matchDecomposedSelectPattern(
729       CmpInst *CmpI, Value *TrueVal, Value *FalseVal, Value *&LHS, Value *&RHS,
730       Instruction::CastOps *CastOp = nullptr, unsigned Depth = 0);
731 
732   /// Return the canonical comparison predicate for the specified
733   /// minimum/maximum flavor.
734   CmpInst::Predicate getMinMaxPred(SelectPatternFlavor SPF,
735                                    bool Ordered = false);
736 
737   /// Return the inverse minimum/maximum flavor of the specified flavor.
738   /// For example, signed minimum is the inverse of signed maximum.
739   SelectPatternFlavor getInverseMinMaxFlavor(SelectPatternFlavor SPF);
740 
741   Intrinsic::ID getInverseMinMaxIntrinsic(Intrinsic::ID MinMaxID);
742 
743   /// Return the canonical inverse comparison predicate for the specified
744   /// minimum/maximum flavor.
745   CmpInst::Predicate getInverseMinMaxPred(SelectPatternFlavor SPF);
746 
747   /// Check if the values in \p VL are select instructions that can be converted
748   /// to a min or max (vector) intrinsic. Returns the intrinsic ID, if such a
749   /// conversion is possible, together with a bool indicating whether all select
750   /// conditions are only used by the selects. Otherwise return
751   /// Intrinsic::not_intrinsic.
752   std::pair<Intrinsic::ID, bool>
753   canConvertToMinOrMaxIntrinsic(ArrayRef<Value *> VL);
754 
755   /// Attempt to match a simple first order recurrence cycle of the form:
756   ///   %iv = phi Ty [%Start, %Entry], [%Inc, %backedge]
757   ///   %inc = binop %iv, %step
758   /// OR
759   ///   %iv = phi Ty [%Start, %Entry], [%Inc, %backedge]
760   ///   %inc = binop %step, %iv
761   ///
762   /// A first order recurrence is a formula with the form: X_n = f(X_(n-1))
763   ///
764   /// A couple of notes on subtleties in that definition:
765   /// * The Step does not have to be loop invariant.  In math terms, it can
766   ///   be a free variable.  We allow recurrences with both constant and
767   ///   variable coefficients. Callers may wish to filter cases where Step
768   ///   does not dominate P.
769   /// * For non-commutative operators, we will match both forms.  This
770   ///   results in some odd recurrence structures.  Callers may wish to filter
771   ///   out recurrences where the phi is not the LHS of the returned operator.
772   /// * Because of the structure matched, the caller can assume as a post
773   ///   condition of the match the presence of a Loop with P's parent as it's
774   ///   header *except* in unreachable code.  (Dominance decays in unreachable
775   ///   code.)
776   ///
777   /// NOTE: This is intentional simple.  If you want the ability to analyze
778   /// non-trivial loop conditons, see ScalarEvolution instead.
779   bool matchSimpleRecurrence(const PHINode *P, BinaryOperator *&BO,
780                              Value *&Start, Value *&Step);
781 
782   /// Analogous to the above, but starting from the binary operator
783   bool matchSimpleRecurrence(const BinaryOperator *I, PHINode *&P,
784                                     Value *&Start, Value *&Step);
785 
786   /// Return true if RHS is known to be implied true by LHS.  Return false if
787   /// RHS is known to be implied false by LHS.  Otherwise, return None if no
788   /// implication can be made.
789   /// A & B must be i1 (boolean) values or a vector of such values. Note that
790   /// the truth table for implication is the same as <=u on i1 values (but not
791   /// <=s!).  The truth table for both is:
792   ///    | T | F (B)
793   ///  T | T | F
794   ///  F | T | T
795   /// (A)
796   Optional<bool> isImpliedCondition(const Value *LHS, const Value *RHS,
797                                     const DataLayout &DL, bool LHSIsTrue = true,
798                                     unsigned Depth = 0);
799   Optional<bool> isImpliedCondition(const Value *LHS,
800                                     CmpInst::Predicate RHSPred,
801                                     const Value *RHSOp0, const Value *RHSOp1,
802                                     const DataLayout &DL, bool LHSIsTrue = true,
803                                     unsigned Depth = 0);
804 
805   /// Return the boolean condition value in the context of the given instruction
806   /// if it is known based on dominating conditions.
807   Optional<bool> isImpliedByDomCondition(const Value *Cond,
808                                          const Instruction *ContextI,
809                                          const DataLayout &DL);
810   Optional<bool> isImpliedByDomCondition(CmpInst::Predicate Pred,
811                                          const Value *LHS, const Value *RHS,
812                                          const Instruction *ContextI,
813                                          const DataLayout &DL);
814 
815   /// If Ptr1 is provably equal to Ptr2 plus a constant offset, return that
816   /// offset. For example, Ptr1 might be &A[42], and Ptr2 might be &A[40]. In
817   /// this case offset would be -8.
818   Optional<int64_t> isPointerOffset(const Value *Ptr1, const Value *Ptr2,
819                                     const DataLayout &DL);
820 } // end namespace llvm
821 
822 #endif // LLVM_ANALYSIS_VALUETRACKING_H
823