• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===- ValueTracking.cpp - Walk computations to compute properties --------===//
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 #include "llvm/Analysis/ValueTracking.h"
16 #include "llvm/ADT/APFloat.h"
17 #include "llvm/ADT/APInt.h"
18 #include "llvm/ADT/ArrayRef.h"
19 #include "llvm/ADT/None.h"
20 #include "llvm/ADT/Optional.h"
21 #include "llvm/ADT/STLExtras.h"
22 #include "llvm/ADT/SmallPtrSet.h"
23 #include "llvm/ADT/SmallSet.h"
24 #include "llvm/ADT/SmallVector.h"
25 #include "llvm/ADT/StringRef.h"
26 #include "llvm/ADT/iterator_range.h"
27 #include "llvm/Analysis/AliasAnalysis.h"
28 #include "llvm/Analysis/AssumptionCache.h"
29 #include "llvm/Analysis/InstructionSimplify.h"
30 #include "llvm/Analysis/Loads.h"
31 #include "llvm/Analysis/LoopInfo.h"
32 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
33 #include "llvm/Analysis/TargetLibraryInfo.h"
34 #include "llvm/IR/Argument.h"
35 #include "llvm/IR/Attributes.h"
36 #include "llvm/IR/BasicBlock.h"
37 #include "llvm/IR/CallSite.h"
38 #include "llvm/IR/Constant.h"
39 #include "llvm/IR/ConstantRange.h"
40 #include "llvm/IR/Constants.h"
41 #include "llvm/IR/DataLayout.h"
42 #include "llvm/IR/DerivedTypes.h"
43 #include "llvm/IR/DiagnosticInfo.h"
44 #include "llvm/IR/Dominators.h"
45 #include "llvm/IR/Function.h"
46 #include "llvm/IR/GetElementPtrTypeIterator.h"
47 #include "llvm/IR/GlobalAlias.h"
48 #include "llvm/IR/GlobalValue.h"
49 #include "llvm/IR/GlobalVariable.h"
50 #include "llvm/IR/InstrTypes.h"
51 #include "llvm/IR/Instruction.h"
52 #include "llvm/IR/Instructions.h"
53 #include "llvm/IR/IntrinsicInst.h"
54 #include "llvm/IR/Intrinsics.h"
55 #include "llvm/IR/LLVMContext.h"
56 #include "llvm/IR/Metadata.h"
57 #include "llvm/IR/Module.h"
58 #include "llvm/IR/Operator.h"
59 #include "llvm/IR/PatternMatch.h"
60 #include "llvm/IR/Type.h"
61 #include "llvm/IR/User.h"
62 #include "llvm/IR/Value.h"
63 #include "llvm/Support/Casting.h"
64 #include "llvm/Support/CommandLine.h"
65 #include "llvm/Support/Compiler.h"
66 #include "llvm/Support/ErrorHandling.h"
67 #include "llvm/Support/KnownBits.h"
68 #include "llvm/Support/MathExtras.h"
69 #include <algorithm>
70 #include <array>
71 #include <cassert>
72 #include <cstdint>
73 #include <iterator>
74 #include <utility>
75 
76 using namespace llvm;
77 using namespace llvm::PatternMatch;
78 
79 const unsigned MaxDepth = 6;
80 
81 // Controls the number of uses of the value searched for possible
82 // dominating comparisons.
83 static cl::opt<unsigned> DomConditionsMaxUses("dom-conditions-max-uses",
84                                               cl::Hidden, cl::init(20));
85 
86 /// Returns the bitwidth of the given scalar or pointer type. For vector types,
87 /// returns the element type's bitwidth.
getBitWidth(Type * Ty,const DataLayout & DL)88 static unsigned getBitWidth(Type *Ty, const DataLayout &DL) {
89   if (unsigned BitWidth = Ty->getScalarSizeInBits())
90     return BitWidth;
91 
92   return DL.getIndexTypeSizeInBits(Ty);
93 }
94 
95 namespace {
96 
97 // Simplifying using an assume can only be done in a particular control-flow
98 // context (the context instruction provides that context). If an assume and
99 // the context instruction are not in the same block then the DT helps in
100 // figuring out if we can use it.
101 struct Query {
102   const DataLayout &DL;
103   AssumptionCache *AC;
104   const Instruction *CxtI;
105   const DominatorTree *DT;
106 
107   // Unlike the other analyses, this may be a nullptr because not all clients
108   // provide it currently.
109   OptimizationRemarkEmitter *ORE;
110 
111   /// Set of assumptions that should be excluded from further queries.
112   /// This is because of the potential for mutual recursion to cause
113   /// computeKnownBits to repeatedly visit the same assume intrinsic. The
114   /// classic case of this is assume(x = y), which will attempt to determine
115   /// bits in x from bits in y, which will attempt to determine bits in y from
116   /// bits in x, etc. Regarding the mutual recursion, computeKnownBits can call
117   /// isKnownNonZero, which calls computeKnownBits and isKnownToBeAPowerOfTwo
118   /// (all of which can call computeKnownBits), and so on.
119   std::array<const Value *, MaxDepth> Excluded;
120 
121   unsigned NumExcluded = 0;
122 
Query__anonfca0c9b60111::Query123   Query(const DataLayout &DL, AssumptionCache *AC, const Instruction *CxtI,
124         const DominatorTree *DT, OptimizationRemarkEmitter *ORE = nullptr)
125       : DL(DL), AC(AC), CxtI(CxtI), DT(DT), ORE(ORE) {}
126 
Query__anonfca0c9b60111::Query127   Query(const Query &Q, const Value *NewExcl)
128       : DL(Q.DL), AC(Q.AC), CxtI(Q.CxtI), DT(Q.DT), ORE(Q.ORE),
129         NumExcluded(Q.NumExcluded) {
130     Excluded = Q.Excluded;
131     Excluded[NumExcluded++] = NewExcl;
132     assert(NumExcluded <= Excluded.size());
133   }
134 
isExcluded__anonfca0c9b60111::Query135   bool isExcluded(const Value *Value) const {
136     if (NumExcluded == 0)
137       return false;
138     auto End = Excluded.begin() + NumExcluded;
139     return std::find(Excluded.begin(), End, Value) != End;
140   }
141 };
142 
143 } // end anonymous namespace
144 
145 // Given the provided Value and, potentially, a context instruction, return
146 // the preferred context instruction (if any).
safeCxtI(const Value * V,const Instruction * CxtI)147 static const Instruction *safeCxtI(const Value *V, const Instruction *CxtI) {
148   // If we've been provided with a context instruction, then use that (provided
149   // it has been inserted).
150   if (CxtI && CxtI->getParent())
151     return CxtI;
152 
153   // If the value is really an already-inserted instruction, then use that.
154   CxtI = dyn_cast<Instruction>(V);
155   if (CxtI && CxtI->getParent())
156     return CxtI;
157 
158   return nullptr;
159 }
160 
161 static void computeKnownBits(const Value *V, KnownBits &Known,
162                              unsigned Depth, const Query &Q);
163 
computeKnownBits(const Value * V,KnownBits & Known,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT,OptimizationRemarkEmitter * ORE)164 void llvm::computeKnownBits(const Value *V, KnownBits &Known,
165                             const DataLayout &DL, unsigned Depth,
166                             AssumptionCache *AC, const Instruction *CxtI,
167                             const DominatorTree *DT,
168                             OptimizationRemarkEmitter *ORE) {
169   ::computeKnownBits(V, Known, Depth,
170                      Query(DL, AC, safeCxtI(V, CxtI), DT, ORE));
171 }
172 
173 static KnownBits computeKnownBits(const Value *V, unsigned Depth,
174                                   const Query &Q);
175 
computeKnownBits(const Value * V,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT,OptimizationRemarkEmitter * ORE)176 KnownBits llvm::computeKnownBits(const Value *V, const DataLayout &DL,
177                                  unsigned Depth, AssumptionCache *AC,
178                                  const Instruction *CxtI,
179                                  const DominatorTree *DT,
180                                  OptimizationRemarkEmitter *ORE) {
181   return ::computeKnownBits(V, Depth,
182                             Query(DL, AC, safeCxtI(V, CxtI), DT, ORE));
183 }
184 
haveNoCommonBitsSet(const Value * LHS,const Value * RHS,const DataLayout & DL,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)185 bool llvm::haveNoCommonBitsSet(const Value *LHS, const Value *RHS,
186                                const DataLayout &DL,
187                                AssumptionCache *AC, const Instruction *CxtI,
188                                const DominatorTree *DT) {
189   assert(LHS->getType() == RHS->getType() &&
190          "LHS and RHS should have the same type");
191   assert(LHS->getType()->isIntOrIntVectorTy() &&
192          "LHS and RHS should be integers");
193   // Look for an inverted mask: (X & ~M) op (Y & M).
194   Value *M;
195   if (match(LHS, m_c_And(m_Not(m_Value(M)), m_Value())) &&
196       match(RHS, m_c_And(m_Specific(M), m_Value())))
197     return true;
198   if (match(RHS, m_c_And(m_Not(m_Value(M)), m_Value())) &&
199       match(LHS, m_c_And(m_Specific(M), m_Value())))
200     return true;
201   IntegerType *IT = cast<IntegerType>(LHS->getType()->getScalarType());
202   KnownBits LHSKnown(IT->getBitWidth());
203   KnownBits RHSKnown(IT->getBitWidth());
204   computeKnownBits(LHS, LHSKnown, DL, 0, AC, CxtI, DT);
205   computeKnownBits(RHS, RHSKnown, DL, 0, AC, CxtI, DT);
206   return (LHSKnown.Zero | RHSKnown.Zero).isAllOnesValue();
207 }
208 
isOnlyUsedInZeroEqualityComparison(const Instruction * CxtI)209 bool llvm::isOnlyUsedInZeroEqualityComparison(const Instruction *CxtI) {
210   for (const User *U : CxtI->users()) {
211     if (const ICmpInst *IC = dyn_cast<ICmpInst>(U))
212       if (IC->isEquality())
213         if (Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
214           if (C->isNullValue())
215             continue;
216     return false;
217   }
218   return true;
219 }
220 
221 static bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth,
222                                    const Query &Q);
223 
isKnownToBeAPowerOfTwo(const Value * V,const DataLayout & DL,bool OrZero,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)224 bool llvm::isKnownToBeAPowerOfTwo(const Value *V, const DataLayout &DL,
225                                   bool OrZero,
226                                   unsigned Depth, AssumptionCache *AC,
227                                   const Instruction *CxtI,
228                                   const DominatorTree *DT) {
229   return ::isKnownToBeAPowerOfTwo(V, OrZero, Depth,
230                                   Query(DL, AC, safeCxtI(V, CxtI), DT));
231 }
232 
233 static bool isKnownNonZero(const Value *V, unsigned Depth, const Query &Q);
234 
isKnownNonZero(const Value * V,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)235 bool llvm::isKnownNonZero(const Value *V, const DataLayout &DL, unsigned Depth,
236                           AssumptionCache *AC, const Instruction *CxtI,
237                           const DominatorTree *DT) {
238   return ::isKnownNonZero(V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT));
239 }
240 
isKnownNonNegative(const Value * V,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)241 bool llvm::isKnownNonNegative(const Value *V, const DataLayout &DL,
242                               unsigned Depth,
243                               AssumptionCache *AC, const Instruction *CxtI,
244                               const DominatorTree *DT) {
245   KnownBits Known = computeKnownBits(V, DL, Depth, AC, CxtI, DT);
246   return Known.isNonNegative();
247 }
248 
isKnownPositive(const Value * V,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)249 bool llvm::isKnownPositive(const Value *V, const DataLayout &DL, unsigned Depth,
250                            AssumptionCache *AC, const Instruction *CxtI,
251                            const DominatorTree *DT) {
252   if (auto *CI = dyn_cast<ConstantInt>(V))
253     return CI->getValue().isStrictlyPositive();
254 
255   // TODO: We'd doing two recursive queries here.  We should factor this such
256   // that only a single query is needed.
257   return isKnownNonNegative(V, DL, Depth, AC, CxtI, DT) &&
258     isKnownNonZero(V, DL, Depth, AC, CxtI, DT);
259 }
260 
isKnownNegative(const Value * V,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)261 bool llvm::isKnownNegative(const Value *V, const DataLayout &DL, unsigned Depth,
262                            AssumptionCache *AC, const Instruction *CxtI,
263                            const DominatorTree *DT) {
264   KnownBits Known = computeKnownBits(V, DL, Depth, AC, CxtI, DT);
265   return Known.isNegative();
266 }
267 
268 static bool isKnownNonEqual(const Value *V1, const Value *V2, const Query &Q);
269 
isKnownNonEqual(const Value * V1,const Value * V2,const DataLayout & DL,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)270 bool llvm::isKnownNonEqual(const Value *V1, const Value *V2,
271                            const DataLayout &DL,
272                            AssumptionCache *AC, const Instruction *CxtI,
273                            const DominatorTree *DT) {
274   return ::isKnownNonEqual(V1, V2, Query(DL, AC,
275                                          safeCxtI(V1, safeCxtI(V2, CxtI)),
276                                          DT));
277 }
278 
279 static bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth,
280                               const Query &Q);
281 
MaskedValueIsZero(const Value * V,const APInt & Mask,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)282 bool llvm::MaskedValueIsZero(const Value *V, const APInt &Mask,
283                              const DataLayout &DL,
284                              unsigned Depth, AssumptionCache *AC,
285                              const Instruction *CxtI, const DominatorTree *DT) {
286   return ::MaskedValueIsZero(V, Mask, Depth,
287                              Query(DL, AC, safeCxtI(V, CxtI), DT));
288 }
289 
290 static unsigned ComputeNumSignBits(const Value *V, unsigned Depth,
291                                    const Query &Q);
292 
ComputeNumSignBits(const Value * V,const DataLayout & DL,unsigned Depth,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)293 unsigned llvm::ComputeNumSignBits(const Value *V, const DataLayout &DL,
294                                   unsigned Depth, AssumptionCache *AC,
295                                   const Instruction *CxtI,
296                                   const DominatorTree *DT) {
297   return ::ComputeNumSignBits(V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT));
298 }
299 
computeKnownBitsAddSub(bool Add,const Value * Op0,const Value * Op1,bool NSW,KnownBits & KnownOut,KnownBits & Known2,unsigned Depth,const Query & Q)300 static void computeKnownBitsAddSub(bool Add, const Value *Op0, const Value *Op1,
301                                    bool NSW,
302                                    KnownBits &KnownOut, KnownBits &Known2,
303                                    unsigned Depth, const Query &Q) {
304   unsigned BitWidth = KnownOut.getBitWidth();
305 
306   // If an initial sequence of bits in the result is not needed, the
307   // corresponding bits in the operands are not needed.
308   KnownBits LHSKnown(BitWidth);
309   computeKnownBits(Op0, LHSKnown, Depth + 1, Q);
310   computeKnownBits(Op1, Known2, Depth + 1, Q);
311 
312   KnownOut = KnownBits::computeForAddSub(Add, NSW, LHSKnown, Known2);
313 }
314 
computeKnownBitsMul(const Value * Op0,const Value * Op1,bool NSW,KnownBits & Known,KnownBits & Known2,unsigned Depth,const Query & Q)315 static void computeKnownBitsMul(const Value *Op0, const Value *Op1, bool NSW,
316                                 KnownBits &Known, KnownBits &Known2,
317                                 unsigned Depth, const Query &Q) {
318   unsigned BitWidth = Known.getBitWidth();
319   computeKnownBits(Op1, Known, Depth + 1, Q);
320   computeKnownBits(Op0, Known2, Depth + 1, Q);
321 
322   bool isKnownNegative = false;
323   bool isKnownNonNegative = false;
324   // If the multiplication is known not to overflow, compute the sign bit.
325   if (NSW) {
326     if (Op0 == Op1) {
327       // The product of a number with itself is non-negative.
328       isKnownNonNegative = true;
329     } else {
330       bool isKnownNonNegativeOp1 = Known.isNonNegative();
331       bool isKnownNonNegativeOp0 = Known2.isNonNegative();
332       bool isKnownNegativeOp1 = Known.isNegative();
333       bool isKnownNegativeOp0 = Known2.isNegative();
334       // The product of two numbers with the same sign is non-negative.
335       isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) ||
336         (isKnownNonNegativeOp1 && isKnownNonNegativeOp0);
337       // The product of a negative number and a non-negative number is either
338       // negative or zero.
339       if (!isKnownNonNegative)
340         isKnownNegative = (isKnownNegativeOp1 && isKnownNonNegativeOp0 &&
341                            isKnownNonZero(Op0, Depth, Q)) ||
342                           (isKnownNegativeOp0 && isKnownNonNegativeOp1 &&
343                            isKnownNonZero(Op1, Depth, Q));
344     }
345   }
346 
347   assert(!Known.hasConflict() && !Known2.hasConflict());
348   // Compute a conservative estimate for high known-0 bits.
349   unsigned LeadZ =  std::max(Known.countMinLeadingZeros() +
350                              Known2.countMinLeadingZeros(),
351                              BitWidth) - BitWidth;
352   LeadZ = std::min(LeadZ, BitWidth);
353 
354   // The result of the bottom bits of an integer multiply can be
355   // inferred by looking at the bottom bits of both operands and
356   // multiplying them together.
357   // We can infer at least the minimum number of known trailing bits
358   // of both operands. Depending on number of trailing zeros, we can
359   // infer more bits, because (a*b) <=> ((a/m) * (b/n)) * (m*n) assuming
360   // a and b are divisible by m and n respectively.
361   // We then calculate how many of those bits are inferrable and set
362   // the output. For example, the i8 mul:
363   //  a = XXXX1100 (12)
364   //  b = XXXX1110 (14)
365   // We know the bottom 3 bits are zero since the first can be divided by
366   // 4 and the second by 2, thus having ((12/4) * (14/2)) * (2*4).
367   // Applying the multiplication to the trimmed arguments gets:
368   //    XX11 (3)
369   //    X111 (7)
370   // -------
371   //    XX11
372   //   XX11
373   //  XX11
374   // XX11
375   // -------
376   // XXXXX01
377   // Which allows us to infer the 2 LSBs. Since we're multiplying the result
378   // by 8, the bottom 3 bits will be 0, so we can infer a total of 5 bits.
379   // The proof for this can be described as:
380   // Pre: (C1 >= 0) && (C1 < (1 << C5)) && (C2 >= 0) && (C2 < (1 << C6)) &&
381   //      (C7 == (1 << (umin(countTrailingZeros(C1), C5) +
382   //                    umin(countTrailingZeros(C2), C6) +
383   //                    umin(C5 - umin(countTrailingZeros(C1), C5),
384   //                         C6 - umin(countTrailingZeros(C2), C6)))) - 1)
385   // %aa = shl i8 %a, C5
386   // %bb = shl i8 %b, C6
387   // %aaa = or i8 %aa, C1
388   // %bbb = or i8 %bb, C2
389   // %mul = mul i8 %aaa, %bbb
390   // %mask = and i8 %mul, C7
391   //   =>
392   // %mask = i8 ((C1*C2)&C7)
393   // Where C5, C6 describe the known bits of %a, %b
394   // C1, C2 describe the known bottom bits of %a, %b.
395   // C7 describes the mask of the known bits of the result.
396   APInt Bottom0 = Known.One;
397   APInt Bottom1 = Known2.One;
398 
399   // How many times we'd be able to divide each argument by 2 (shr by 1).
400   // This gives us the number of trailing zeros on the multiplication result.
401   unsigned TrailBitsKnown0 = (Known.Zero | Known.One).countTrailingOnes();
402   unsigned TrailBitsKnown1 = (Known2.Zero | Known2.One).countTrailingOnes();
403   unsigned TrailZero0 = Known.countMinTrailingZeros();
404   unsigned TrailZero1 = Known2.countMinTrailingZeros();
405   unsigned TrailZ = TrailZero0 + TrailZero1;
406 
407   // Figure out the fewest known-bits operand.
408   unsigned SmallestOperand = std::min(TrailBitsKnown0 - TrailZero0,
409                                       TrailBitsKnown1 - TrailZero1);
410   unsigned ResultBitsKnown = std::min(SmallestOperand + TrailZ, BitWidth);
411 
412   APInt BottomKnown = Bottom0.getLoBits(TrailBitsKnown0) *
413                       Bottom1.getLoBits(TrailBitsKnown1);
414 
415   Known.resetAll();
416   Known.Zero.setHighBits(LeadZ);
417   Known.Zero |= (~BottomKnown).getLoBits(ResultBitsKnown);
418   Known.One |= BottomKnown.getLoBits(ResultBitsKnown);
419 
420   // Only make use of no-wrap flags if we failed to compute the sign bit
421   // directly.  This matters if the multiplication always overflows, in
422   // which case we prefer to follow the result of the direct computation,
423   // though as the program is invoking undefined behaviour we can choose
424   // whatever we like here.
425   if (isKnownNonNegative && !Known.isNegative())
426     Known.makeNonNegative();
427   else if (isKnownNegative && !Known.isNonNegative())
428     Known.makeNegative();
429 }
430 
computeKnownBitsFromRangeMetadata(const MDNode & Ranges,KnownBits & Known)431 void llvm::computeKnownBitsFromRangeMetadata(const MDNode &Ranges,
432                                              KnownBits &Known) {
433   unsigned BitWidth = Known.getBitWidth();
434   unsigned NumRanges = Ranges.getNumOperands() / 2;
435   assert(NumRanges >= 1);
436 
437   Known.Zero.setAllBits();
438   Known.One.setAllBits();
439 
440   for (unsigned i = 0; i < NumRanges; ++i) {
441     ConstantInt *Lower =
442         mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0));
443     ConstantInt *Upper =
444         mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1));
445     ConstantRange Range(Lower->getValue(), Upper->getValue());
446 
447     // The first CommonPrefixBits of all values in Range are equal.
448     unsigned CommonPrefixBits =
449         (Range.getUnsignedMax() ^ Range.getUnsignedMin()).countLeadingZeros();
450 
451     APInt Mask = APInt::getHighBitsSet(BitWidth, CommonPrefixBits);
452     Known.One &= Range.getUnsignedMax() & Mask;
453     Known.Zero &= ~Range.getUnsignedMax() & Mask;
454   }
455 }
456 
isEphemeralValueOf(const Instruction * I,const Value * E)457 static bool isEphemeralValueOf(const Instruction *I, const Value *E) {
458   SmallVector<const Value *, 16> WorkSet(1, I);
459   SmallPtrSet<const Value *, 32> Visited;
460   SmallPtrSet<const Value *, 16> EphValues;
461 
462   // The instruction defining an assumption's condition itself is always
463   // considered ephemeral to that assumption (even if it has other
464   // non-ephemeral users). See r246696's test case for an example.
465   if (is_contained(I->operands(), E))
466     return true;
467 
468   while (!WorkSet.empty()) {
469     const Value *V = WorkSet.pop_back_val();
470     if (!Visited.insert(V).second)
471       continue;
472 
473     // If all uses of this value are ephemeral, then so is this value.
474     if (llvm::all_of(V->users(), [&](const User *U) {
475                                    return EphValues.count(U);
476                                  })) {
477       if (V == E)
478         return true;
479 
480       if (V == I || isSafeToSpeculativelyExecute(V)) {
481        EphValues.insert(V);
482        if (const User *U = dyn_cast<User>(V))
483          for (User::const_op_iterator J = U->op_begin(), JE = U->op_end();
484               J != JE; ++J)
485            WorkSet.push_back(*J);
486       }
487     }
488   }
489 
490   return false;
491 }
492 
493 // Is this an intrinsic that cannot be speculated but also cannot trap?
isAssumeLikeIntrinsic(const Instruction * I)494 bool llvm::isAssumeLikeIntrinsic(const Instruction *I) {
495   if (const CallInst *CI = dyn_cast<CallInst>(I))
496     if (Function *F = CI->getCalledFunction())
497       switch (F->getIntrinsicID()) {
498       default: break;
499       // FIXME: This list is repeated from NoTTI::getIntrinsicCost.
500       case Intrinsic::assume:
501       case Intrinsic::sideeffect:
502       case Intrinsic::dbg_declare:
503       case Intrinsic::dbg_value:
504       case Intrinsic::dbg_label:
505       case Intrinsic::invariant_start:
506       case Intrinsic::invariant_end:
507       case Intrinsic::lifetime_start:
508       case Intrinsic::lifetime_end:
509       case Intrinsic::objectsize:
510       case Intrinsic::ptr_annotation:
511       case Intrinsic::var_annotation:
512         return true;
513       }
514 
515   return false;
516 }
517 
isValidAssumeForContext(const Instruction * Inv,const Instruction * CxtI,const DominatorTree * DT)518 bool llvm::isValidAssumeForContext(const Instruction *Inv,
519                                    const Instruction *CxtI,
520                                    const DominatorTree *DT) {
521   // There are two restrictions on the use of an assume:
522   //  1. The assume must dominate the context (or the control flow must
523   //     reach the assume whenever it reaches the context).
524   //  2. The context must not be in the assume's set of ephemeral values
525   //     (otherwise we will use the assume to prove that the condition
526   //     feeding the assume is trivially true, thus causing the removal of
527   //     the assume).
528 
529   if (DT) {
530     if (DT->dominates(Inv, CxtI))
531       return true;
532   } else if (Inv->getParent() == CxtI->getParent()->getSinglePredecessor()) {
533     // We don't have a DT, but this trivially dominates.
534     return true;
535   }
536 
537   // With or without a DT, the only remaining case we will check is if the
538   // instructions are in the same BB.  Give up if that is not the case.
539   if (Inv->getParent() != CxtI->getParent())
540     return false;
541 
542   // If we have a dom tree, then we now know that the assume doesn't dominate
543   // the other instruction.  If we don't have a dom tree then we can check if
544   // the assume is first in the BB.
545   if (!DT) {
546     // Search forward from the assume until we reach the context (or the end
547     // of the block); the common case is that the assume will come first.
548     for (auto I = std::next(BasicBlock::const_iterator(Inv)),
549          IE = Inv->getParent()->end(); I != IE; ++I)
550       if (&*I == CxtI)
551         return true;
552   }
553 
554   // The context comes first, but they're both in the same block. Make sure
555   // there is nothing in between that might interrupt the control flow.
556   for (BasicBlock::const_iterator I =
557          std::next(BasicBlock::const_iterator(CxtI)), IE(Inv);
558        I != IE; ++I)
559     if (!isSafeToSpeculativelyExecute(&*I) && !isAssumeLikeIntrinsic(&*I))
560       return false;
561 
562   return !isEphemeralValueOf(Inv, CxtI);
563 }
564 
computeKnownBitsFromAssume(const Value * V,KnownBits & Known,unsigned Depth,const Query & Q)565 static void computeKnownBitsFromAssume(const Value *V, KnownBits &Known,
566                                        unsigned Depth, const Query &Q) {
567   // Use of assumptions is context-sensitive. If we don't have a context, we
568   // cannot use them!
569   if (!Q.AC || !Q.CxtI)
570     return;
571 
572   unsigned BitWidth = Known.getBitWidth();
573 
574   // Note that the patterns below need to be kept in sync with the code
575   // in AssumptionCache::updateAffectedValues.
576 
577   for (auto &AssumeVH : Q.AC->assumptionsFor(V)) {
578     if (!AssumeVH)
579       continue;
580     CallInst *I = cast<CallInst>(AssumeVH);
581     assert(I->getParent()->getParent() == Q.CxtI->getParent()->getParent() &&
582            "Got assumption for the wrong function!");
583     if (Q.isExcluded(I))
584       continue;
585 
586     // Warning: This loop can end up being somewhat performance sensitive.
587     // We're running this loop for once for each value queried resulting in a
588     // runtime of ~O(#assumes * #values).
589 
590     assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&
591            "must be an assume intrinsic");
592 
593     Value *Arg = I->getArgOperand(0);
594 
595     if (Arg == V && isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
596       assert(BitWidth == 1 && "assume operand is not i1?");
597       Known.setAllOnes();
598       return;
599     }
600     if (match(Arg, m_Not(m_Specific(V))) &&
601         isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
602       assert(BitWidth == 1 && "assume operand is not i1?");
603       Known.setAllZero();
604       return;
605     }
606 
607     // The remaining tests are all recursive, so bail out if we hit the limit.
608     if (Depth == MaxDepth)
609       continue;
610 
611     Value *A, *B;
612     auto m_V = m_CombineOr(m_Specific(V),
613                            m_CombineOr(m_PtrToInt(m_Specific(V)),
614                            m_BitCast(m_Specific(V))));
615 
616     CmpInst::Predicate Pred;
617     uint64_t C;
618     // assume(v = a)
619     if (match(Arg, m_c_ICmp(Pred, m_V, m_Value(A))) &&
620         Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
621       KnownBits RHSKnown(BitWidth);
622       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
623       Known.Zero |= RHSKnown.Zero;
624       Known.One  |= RHSKnown.One;
625     // assume(v & b = a)
626     } else if (match(Arg,
627                      m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)), m_Value(A))) &&
628                Pred == ICmpInst::ICMP_EQ &&
629                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
630       KnownBits RHSKnown(BitWidth);
631       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
632       KnownBits MaskKnown(BitWidth);
633       computeKnownBits(B, MaskKnown, Depth+1, Query(Q, I));
634 
635       // For those bits in the mask that are known to be one, we can propagate
636       // known bits from the RHS to V.
637       Known.Zero |= RHSKnown.Zero & MaskKnown.One;
638       Known.One  |= RHSKnown.One  & MaskKnown.One;
639     // assume(~(v & b) = a)
640     } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))),
641                                    m_Value(A))) &&
642                Pred == ICmpInst::ICMP_EQ &&
643                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
644       KnownBits RHSKnown(BitWidth);
645       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
646       KnownBits MaskKnown(BitWidth);
647       computeKnownBits(B, MaskKnown, Depth+1, Query(Q, I));
648 
649       // For those bits in the mask that are known to be one, we can propagate
650       // inverted known bits from the RHS to V.
651       Known.Zero |= RHSKnown.One  & MaskKnown.One;
652       Known.One  |= RHSKnown.Zero & MaskKnown.One;
653     // assume(v | b = a)
654     } else if (match(Arg,
655                      m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)), m_Value(A))) &&
656                Pred == ICmpInst::ICMP_EQ &&
657                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
658       KnownBits RHSKnown(BitWidth);
659       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
660       KnownBits BKnown(BitWidth);
661       computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
662 
663       // For those bits in B that are known to be zero, we can propagate known
664       // bits from the RHS to V.
665       Known.Zero |= RHSKnown.Zero & BKnown.Zero;
666       Known.One  |= RHSKnown.One  & BKnown.Zero;
667     // assume(~(v | b) = a)
668     } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))),
669                                    m_Value(A))) &&
670                Pred == ICmpInst::ICMP_EQ &&
671                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
672       KnownBits RHSKnown(BitWidth);
673       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
674       KnownBits BKnown(BitWidth);
675       computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
676 
677       // For those bits in B that are known to be zero, we can propagate
678       // inverted known bits from the RHS to V.
679       Known.Zero |= RHSKnown.One  & BKnown.Zero;
680       Known.One  |= RHSKnown.Zero & BKnown.Zero;
681     // assume(v ^ b = a)
682     } else if (match(Arg,
683                      m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)), m_Value(A))) &&
684                Pred == ICmpInst::ICMP_EQ &&
685                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
686       KnownBits RHSKnown(BitWidth);
687       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
688       KnownBits BKnown(BitWidth);
689       computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
690 
691       // For those bits in B that are known to be zero, we can propagate known
692       // bits from the RHS to V. For those bits in B that are known to be one,
693       // we can propagate inverted known bits from the RHS to V.
694       Known.Zero |= RHSKnown.Zero & BKnown.Zero;
695       Known.One  |= RHSKnown.One  & BKnown.Zero;
696       Known.Zero |= RHSKnown.One  & BKnown.One;
697       Known.One  |= RHSKnown.Zero & BKnown.One;
698     // assume(~(v ^ b) = a)
699     } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))),
700                                    m_Value(A))) &&
701                Pred == ICmpInst::ICMP_EQ &&
702                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
703       KnownBits RHSKnown(BitWidth);
704       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
705       KnownBits BKnown(BitWidth);
706       computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
707 
708       // For those bits in B that are known to be zero, we can propagate
709       // inverted known bits from the RHS to V. For those bits in B that are
710       // known to be one, we can propagate known bits from the RHS to V.
711       Known.Zero |= RHSKnown.One  & BKnown.Zero;
712       Known.One  |= RHSKnown.Zero & BKnown.Zero;
713       Known.Zero |= RHSKnown.Zero & BKnown.One;
714       Known.One  |= RHSKnown.One  & BKnown.One;
715     // assume(v << c = a)
716     } else if (match(Arg, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)),
717                                    m_Value(A))) &&
718                Pred == ICmpInst::ICMP_EQ &&
719                isValidAssumeForContext(I, Q.CxtI, Q.DT) &&
720                C < BitWidth) {
721       KnownBits RHSKnown(BitWidth);
722       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
723       // For those bits in RHS that are known, we can propagate them to known
724       // bits in V shifted to the right by C.
725       RHSKnown.Zero.lshrInPlace(C);
726       Known.Zero |= RHSKnown.Zero;
727       RHSKnown.One.lshrInPlace(C);
728       Known.One  |= RHSKnown.One;
729     // assume(~(v << c) = a)
730     } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))),
731                                    m_Value(A))) &&
732                Pred == ICmpInst::ICMP_EQ &&
733                isValidAssumeForContext(I, Q.CxtI, Q.DT) &&
734                C < BitWidth) {
735       KnownBits RHSKnown(BitWidth);
736       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
737       // For those bits in RHS that are known, we can propagate them inverted
738       // to known bits in V shifted to the right by C.
739       RHSKnown.One.lshrInPlace(C);
740       Known.Zero |= RHSKnown.One;
741       RHSKnown.Zero.lshrInPlace(C);
742       Known.One  |= RHSKnown.Zero;
743     // assume(v >> c = a)
744     } else if (match(Arg,
745                      m_c_ICmp(Pred, m_Shr(m_V, m_ConstantInt(C)),
746                               m_Value(A))) &&
747                Pred == ICmpInst::ICMP_EQ &&
748                isValidAssumeForContext(I, Q.CxtI, Q.DT) &&
749                C < BitWidth) {
750       KnownBits RHSKnown(BitWidth);
751       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
752       // For those bits in RHS that are known, we can propagate them to known
753       // bits in V shifted to the right by C.
754       Known.Zero |= RHSKnown.Zero << C;
755       Known.One  |= RHSKnown.One  << C;
756     // assume(~(v >> c) = a)
757     } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_Shr(m_V, m_ConstantInt(C))),
758                                    m_Value(A))) &&
759                Pred == ICmpInst::ICMP_EQ &&
760                isValidAssumeForContext(I, Q.CxtI, Q.DT) &&
761                C < BitWidth) {
762       KnownBits RHSKnown(BitWidth);
763       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
764       // For those bits in RHS that are known, we can propagate them inverted
765       // to known bits in V shifted to the right by C.
766       Known.Zero |= RHSKnown.One  << C;
767       Known.One  |= RHSKnown.Zero << C;
768     // assume(v >=_s c) where c is non-negative
769     } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
770                Pred == ICmpInst::ICMP_SGE &&
771                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
772       KnownBits RHSKnown(BitWidth);
773       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
774 
775       if (RHSKnown.isNonNegative()) {
776         // We know that the sign bit is zero.
777         Known.makeNonNegative();
778       }
779     // assume(v >_s c) where c is at least -1.
780     } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
781                Pred == ICmpInst::ICMP_SGT &&
782                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
783       KnownBits RHSKnown(BitWidth);
784       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
785 
786       if (RHSKnown.isAllOnes() || RHSKnown.isNonNegative()) {
787         // We know that the sign bit is zero.
788         Known.makeNonNegative();
789       }
790     // assume(v <=_s c) where c is negative
791     } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
792                Pred == ICmpInst::ICMP_SLE &&
793                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
794       KnownBits RHSKnown(BitWidth);
795       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
796 
797       if (RHSKnown.isNegative()) {
798         // We know that the sign bit is one.
799         Known.makeNegative();
800       }
801     // assume(v <_s c) where c is non-positive
802     } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
803                Pred == ICmpInst::ICMP_SLT &&
804                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
805       KnownBits RHSKnown(BitWidth);
806       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
807 
808       if (RHSKnown.isZero() || RHSKnown.isNegative()) {
809         // We know that the sign bit is one.
810         Known.makeNegative();
811       }
812     // assume(v <=_u c)
813     } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
814                Pred == ICmpInst::ICMP_ULE &&
815                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
816       KnownBits RHSKnown(BitWidth);
817       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
818 
819       // Whatever high bits in c are zero are known to be zero.
820       Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros());
821       // assume(v <_u c)
822     } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
823                Pred == ICmpInst::ICMP_ULT &&
824                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
825       KnownBits RHSKnown(BitWidth);
826       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
827 
828       // If the RHS is known zero, then this assumption must be wrong (nothing
829       // is unsigned less than zero). Signal a conflict and get out of here.
830       if (RHSKnown.isZero()) {
831         Known.Zero.setAllBits();
832         Known.One.setAllBits();
833         break;
834       }
835 
836       // Whatever high bits in c are zero are known to be zero (if c is a power
837       // of 2, then one more).
838       if (isKnownToBeAPowerOfTwo(A, false, Depth + 1, Query(Q, I)))
839         Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros() + 1);
840       else
841         Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros());
842     }
843   }
844 
845   // If assumptions conflict with each other or previous known bits, then we
846   // have a logical fallacy. It's possible that the assumption is not reachable,
847   // so this isn't a real bug. On the other hand, the program may have undefined
848   // behavior, or we might have a bug in the compiler. We can't assert/crash, so
849   // clear out the known bits, try to warn the user, and hope for the best.
850   if (Known.Zero.intersects(Known.One)) {
851     Known.resetAll();
852 
853     if (Q.ORE)
854       Q.ORE->emit([&]() {
855         auto *CxtI = const_cast<Instruction *>(Q.CxtI);
856         return OptimizationRemarkAnalysis("value-tracking", "BadAssumption",
857                                           CxtI)
858                << "Detected conflicting code assumptions. Program may "
859                   "have undefined behavior, or compiler may have "
860                   "internal error.";
861       });
862   }
863 }
864 
865 /// Compute known bits from a shift operator, including those with a
866 /// non-constant shift amount. Known is the output of this function. Known2 is a
867 /// pre-allocated temporary with the same bit width as Known. KZF and KOF are
868 /// operator-specific functions that, given the known-zero or known-one bits
869 /// respectively, and a shift amount, compute the implied known-zero or
870 /// known-one bits of the shift operator's result respectively for that shift
871 /// amount. The results from calling KZF and KOF are conservatively combined for
872 /// all permitted shift amounts.
computeKnownBitsFromShiftOperator(const Operator * I,KnownBits & Known,KnownBits & Known2,unsigned Depth,const Query & Q,function_ref<APInt (const APInt &,unsigned)> KZF,function_ref<APInt (const APInt &,unsigned)> KOF)873 static void computeKnownBitsFromShiftOperator(
874     const Operator *I, KnownBits &Known, KnownBits &Known2,
875     unsigned Depth, const Query &Q,
876     function_ref<APInt(const APInt &, unsigned)> KZF,
877     function_ref<APInt(const APInt &, unsigned)> KOF) {
878   unsigned BitWidth = Known.getBitWidth();
879 
880   if (auto *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
881     unsigned ShiftAmt = SA->getLimitedValue(BitWidth-1);
882 
883     computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
884     Known.Zero = KZF(Known.Zero, ShiftAmt);
885     Known.One  = KOF(Known.One, ShiftAmt);
886     // If the known bits conflict, this must be an overflowing left shift, so
887     // the shift result is poison. We can return anything we want. Choose 0 for
888     // the best folding opportunity.
889     if (Known.hasConflict())
890       Known.setAllZero();
891 
892     return;
893   }
894 
895   computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
896 
897   // If the shift amount could be greater than or equal to the bit-width of the
898   // LHS, the value could be poison, but bail out because the check below is
899   // expensive. TODO: Should we just carry on?
900   if ((~Known.Zero).uge(BitWidth)) {
901     Known.resetAll();
902     return;
903   }
904 
905   // Note: We cannot use Known.Zero.getLimitedValue() here, because if
906   // BitWidth > 64 and any upper bits are known, we'll end up returning the
907   // limit value (which implies all bits are known).
908   uint64_t ShiftAmtKZ = Known.Zero.zextOrTrunc(64).getZExtValue();
909   uint64_t ShiftAmtKO = Known.One.zextOrTrunc(64).getZExtValue();
910 
911   // It would be more-clearly correct to use the two temporaries for this
912   // calculation. Reusing the APInts here to prevent unnecessary allocations.
913   Known.resetAll();
914 
915   // If we know the shifter operand is nonzero, we can sometimes infer more
916   // known bits. However this is expensive to compute, so be lazy about it and
917   // only compute it when absolutely necessary.
918   Optional<bool> ShifterOperandIsNonZero;
919 
920   // Early exit if we can't constrain any well-defined shift amount.
921   if (!(ShiftAmtKZ & (PowerOf2Ceil(BitWidth) - 1)) &&
922       !(ShiftAmtKO & (PowerOf2Ceil(BitWidth) - 1))) {
923     ShifterOperandIsNonZero = isKnownNonZero(I->getOperand(1), Depth + 1, Q);
924     if (!*ShifterOperandIsNonZero)
925       return;
926   }
927 
928   computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
929 
930   Known.Zero.setAllBits();
931   Known.One.setAllBits();
932   for (unsigned ShiftAmt = 0; ShiftAmt < BitWidth; ++ShiftAmt) {
933     // Combine the shifted known input bits only for those shift amounts
934     // compatible with its known constraints.
935     if ((ShiftAmt & ~ShiftAmtKZ) != ShiftAmt)
936       continue;
937     if ((ShiftAmt | ShiftAmtKO) != ShiftAmt)
938       continue;
939     // If we know the shifter is nonzero, we may be able to infer more known
940     // bits. This check is sunk down as far as possible to avoid the expensive
941     // call to isKnownNonZero if the cheaper checks above fail.
942     if (ShiftAmt == 0) {
943       if (!ShifterOperandIsNonZero.hasValue())
944         ShifterOperandIsNonZero =
945             isKnownNonZero(I->getOperand(1), Depth + 1, Q);
946       if (*ShifterOperandIsNonZero)
947         continue;
948     }
949 
950     Known.Zero &= KZF(Known2.Zero, ShiftAmt);
951     Known.One  &= KOF(Known2.One, ShiftAmt);
952   }
953 
954   // If the known bits conflict, the result is poison. Return a 0 and hope the
955   // caller can further optimize that.
956   if (Known.hasConflict())
957     Known.setAllZero();
958 }
959 
computeKnownBitsFromOperator(const Operator * I,KnownBits & Known,unsigned Depth,const Query & Q)960 static void computeKnownBitsFromOperator(const Operator *I, KnownBits &Known,
961                                          unsigned Depth, const Query &Q) {
962   unsigned BitWidth = Known.getBitWidth();
963 
964   KnownBits Known2(Known);
965   switch (I->getOpcode()) {
966   default: break;
967   case Instruction::Load:
968     if (MDNode *MD = cast<LoadInst>(I)->getMetadata(LLVMContext::MD_range))
969       computeKnownBitsFromRangeMetadata(*MD, Known);
970     break;
971   case Instruction::And: {
972     // If either the LHS or the RHS are Zero, the result is zero.
973     computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
974     computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
975 
976     // Output known-1 bits are only known if set in both the LHS & RHS.
977     Known.One &= Known2.One;
978     // Output known-0 are known to be clear if zero in either the LHS | RHS.
979     Known.Zero |= Known2.Zero;
980 
981     // and(x, add (x, -1)) is a common idiom that always clears the low bit;
982     // here we handle the more general case of adding any odd number by
983     // matching the form add(x, add(x, y)) where y is odd.
984     // TODO: This could be generalized to clearing any bit set in y where the
985     // following bit is known to be unset in y.
986     Value *X = nullptr, *Y = nullptr;
987     if (!Known.Zero[0] && !Known.One[0] &&
988         match(I, m_c_BinOp(m_Value(X), m_Add(m_Deferred(X), m_Value(Y))))) {
989       Known2.resetAll();
990       computeKnownBits(Y, Known2, Depth + 1, Q);
991       if (Known2.countMinTrailingOnes() > 0)
992         Known.Zero.setBit(0);
993     }
994     break;
995   }
996   case Instruction::Or:
997     computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
998     computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
999 
1000     // Output known-0 bits are only known if clear in both the LHS & RHS.
1001     Known.Zero &= Known2.Zero;
1002     // Output known-1 are known to be set if set in either the LHS | RHS.
1003     Known.One |= Known2.One;
1004     break;
1005   case Instruction::Xor: {
1006     computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
1007     computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1008 
1009     // Output known-0 bits are known if clear or set in both the LHS & RHS.
1010     APInt KnownZeroOut = (Known.Zero & Known2.Zero) | (Known.One & Known2.One);
1011     // Output known-1 are known to be set if set in only one of the LHS, RHS.
1012     Known.One = (Known.Zero & Known2.One) | (Known.One & Known2.Zero);
1013     Known.Zero = std::move(KnownZeroOut);
1014     break;
1015   }
1016   case Instruction::Mul: {
1017     bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1018     computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW, Known,
1019                         Known2, Depth, Q);
1020     break;
1021   }
1022   case Instruction::UDiv: {
1023     // For the purposes of computing leading zeros we can conservatively
1024     // treat a udiv as a logical right shift by the power of 2 known to
1025     // be less than the denominator.
1026     computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1027     unsigned LeadZ = Known2.countMinLeadingZeros();
1028 
1029     Known2.resetAll();
1030     computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1031     unsigned RHSMaxLeadingZeros = Known2.countMaxLeadingZeros();
1032     if (RHSMaxLeadingZeros != BitWidth)
1033       LeadZ = std::min(BitWidth, LeadZ + BitWidth - RHSMaxLeadingZeros - 1);
1034 
1035     Known.Zero.setHighBits(LeadZ);
1036     break;
1037   }
1038   case Instruction::Select: {
1039     const Value *LHS, *RHS;
1040     SelectPatternFlavor SPF = matchSelectPattern(I, LHS, RHS).Flavor;
1041     if (SelectPatternResult::isMinOrMax(SPF)) {
1042       computeKnownBits(RHS, Known, Depth + 1, Q);
1043       computeKnownBits(LHS, Known2, Depth + 1, Q);
1044     } else {
1045       computeKnownBits(I->getOperand(2), Known, Depth + 1, Q);
1046       computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1047     }
1048 
1049     unsigned MaxHighOnes = 0;
1050     unsigned MaxHighZeros = 0;
1051     if (SPF == SPF_SMAX) {
1052       // If both sides are negative, the result is negative.
1053       if (Known.isNegative() && Known2.isNegative())
1054         // We can derive a lower bound on the result by taking the max of the
1055         // leading one bits.
1056         MaxHighOnes =
1057             std::max(Known.countMinLeadingOnes(), Known2.countMinLeadingOnes());
1058       // If either side is non-negative, the result is non-negative.
1059       else if (Known.isNonNegative() || Known2.isNonNegative())
1060         MaxHighZeros = 1;
1061     } else if (SPF == SPF_SMIN) {
1062       // If both sides are non-negative, the result is non-negative.
1063       if (Known.isNonNegative() && Known2.isNonNegative())
1064         // We can derive an upper bound on the result by taking the max of the
1065         // leading zero bits.
1066         MaxHighZeros = std::max(Known.countMinLeadingZeros(),
1067                                 Known2.countMinLeadingZeros());
1068       // If either side is negative, the result is negative.
1069       else if (Known.isNegative() || Known2.isNegative())
1070         MaxHighOnes = 1;
1071     } else if (SPF == SPF_UMAX) {
1072       // We can derive a lower bound on the result by taking the max of the
1073       // leading one bits.
1074       MaxHighOnes =
1075           std::max(Known.countMinLeadingOnes(), Known2.countMinLeadingOnes());
1076     } else if (SPF == SPF_UMIN) {
1077       // We can derive an upper bound on the result by taking the max of the
1078       // leading zero bits.
1079       MaxHighZeros =
1080           std::max(Known.countMinLeadingZeros(), Known2.countMinLeadingZeros());
1081     } else if (SPF == SPF_ABS) {
1082       // RHS from matchSelectPattern returns the negation part of abs pattern.
1083       // If the negate has an NSW flag we can assume the sign bit of the result
1084       // will be 0 because that makes abs(INT_MIN) undefined.
1085       if (cast<Instruction>(RHS)->hasNoSignedWrap())
1086         MaxHighZeros = 1;
1087     }
1088 
1089     // Only known if known in both the LHS and RHS.
1090     Known.One &= Known2.One;
1091     Known.Zero &= Known2.Zero;
1092     if (MaxHighOnes > 0)
1093       Known.One.setHighBits(MaxHighOnes);
1094     if (MaxHighZeros > 0)
1095       Known.Zero.setHighBits(MaxHighZeros);
1096     break;
1097   }
1098   case Instruction::FPTrunc:
1099   case Instruction::FPExt:
1100   case Instruction::FPToUI:
1101   case Instruction::FPToSI:
1102   case Instruction::SIToFP:
1103   case Instruction::UIToFP:
1104     break; // Can't work with floating point.
1105   case Instruction::PtrToInt:
1106   case Instruction::IntToPtr:
1107     // Fall through and handle them the same as zext/trunc.
1108     LLVM_FALLTHROUGH;
1109   case Instruction::ZExt:
1110   case Instruction::Trunc: {
1111     Type *SrcTy = I->getOperand(0)->getType();
1112 
1113     unsigned SrcBitWidth;
1114     // Note that we handle pointer operands here because of inttoptr/ptrtoint
1115     // which fall through here.
1116     Type *ScalarTy = SrcTy->getScalarType();
1117     SrcBitWidth = ScalarTy->isPointerTy() ?
1118       Q.DL.getIndexTypeSizeInBits(ScalarTy) :
1119       Q.DL.getTypeSizeInBits(ScalarTy);
1120 
1121     assert(SrcBitWidth && "SrcBitWidth can't be zero");
1122     Known = Known.zextOrTrunc(SrcBitWidth);
1123     computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1124     Known = Known.zextOrTrunc(BitWidth);
1125     // Any top bits are known to be zero.
1126     if (BitWidth > SrcBitWidth)
1127       Known.Zero.setBitsFrom(SrcBitWidth);
1128     break;
1129   }
1130   case Instruction::BitCast: {
1131     Type *SrcTy = I->getOperand(0)->getType();
1132     if (SrcTy->isIntOrPtrTy() &&
1133         // TODO: For now, not handling conversions like:
1134         // (bitcast i64 %x to <2 x i32>)
1135         !I->getType()->isVectorTy()) {
1136       computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1137       break;
1138     }
1139     break;
1140   }
1141   case Instruction::SExt: {
1142     // Compute the bits in the result that are not present in the input.
1143     unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
1144 
1145     Known = Known.trunc(SrcBitWidth);
1146     computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1147     // If the sign bit of the input is known set or clear, then we know the
1148     // top bits of the result.
1149     Known = Known.sext(BitWidth);
1150     break;
1151   }
1152   case Instruction::Shl: {
1153     // (shl X, C1) & C2 == 0   iff   (X & C2 >>u C1) == 0
1154     bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1155     auto KZF = [NSW](const APInt &KnownZero, unsigned ShiftAmt) {
1156       APInt KZResult = KnownZero << ShiftAmt;
1157       KZResult.setLowBits(ShiftAmt); // Low bits known 0.
1158       // If this shift has "nsw" keyword, then the result is either a poison
1159       // value or has the same sign bit as the first operand.
1160       if (NSW && KnownZero.isSignBitSet())
1161         KZResult.setSignBit();
1162       return KZResult;
1163     };
1164 
1165     auto KOF = [NSW](const APInt &KnownOne, unsigned ShiftAmt) {
1166       APInt KOResult = KnownOne << ShiftAmt;
1167       if (NSW && KnownOne.isSignBitSet())
1168         KOResult.setSignBit();
1169       return KOResult;
1170     };
1171 
1172     computeKnownBitsFromShiftOperator(I, Known, Known2, Depth, Q, KZF, KOF);
1173     break;
1174   }
1175   case Instruction::LShr: {
1176     // (lshr X, C1) & C2 == 0   iff  (-1 >> C1) & C2 == 0
1177     auto KZF = [](const APInt &KnownZero, unsigned ShiftAmt) {
1178       APInt KZResult = KnownZero.lshr(ShiftAmt);
1179       // High bits known zero.
1180       KZResult.setHighBits(ShiftAmt);
1181       return KZResult;
1182     };
1183 
1184     auto KOF = [](const APInt &KnownOne, unsigned ShiftAmt) {
1185       return KnownOne.lshr(ShiftAmt);
1186     };
1187 
1188     computeKnownBitsFromShiftOperator(I, Known, Known2, Depth, Q, KZF, KOF);
1189     break;
1190   }
1191   case Instruction::AShr: {
1192     // (ashr X, C1) & C2 == 0   iff  (-1 >> C1) & C2 == 0
1193     auto KZF = [](const APInt &KnownZero, unsigned ShiftAmt) {
1194       return KnownZero.ashr(ShiftAmt);
1195     };
1196 
1197     auto KOF = [](const APInt &KnownOne, unsigned ShiftAmt) {
1198       return KnownOne.ashr(ShiftAmt);
1199     };
1200 
1201     computeKnownBitsFromShiftOperator(I, Known, Known2, Depth, Q, KZF, KOF);
1202     break;
1203   }
1204   case Instruction::Sub: {
1205     bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1206     computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
1207                            Known, Known2, Depth, Q);
1208     break;
1209   }
1210   case Instruction::Add: {
1211     bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
1212     computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
1213                            Known, Known2, Depth, Q);
1214     break;
1215   }
1216   case Instruction::SRem:
1217     if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1218       APInt RA = Rem->getValue().abs();
1219       if (RA.isPowerOf2()) {
1220         APInt LowBits = RA - 1;
1221         computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1222 
1223         // The low bits of the first operand are unchanged by the srem.
1224         Known.Zero = Known2.Zero & LowBits;
1225         Known.One = Known2.One & LowBits;
1226 
1227         // If the first operand is non-negative or has all low bits zero, then
1228         // the upper bits are all zero.
1229         if (Known2.isNonNegative() || LowBits.isSubsetOf(Known2.Zero))
1230           Known.Zero |= ~LowBits;
1231 
1232         // If the first operand is negative and not all low bits are zero, then
1233         // the upper bits are all one.
1234         if (Known2.isNegative() && LowBits.intersects(Known2.One))
1235           Known.One |= ~LowBits;
1236 
1237         assert((Known.Zero & Known.One) == 0 && "Bits known to be one AND zero?");
1238         break;
1239       }
1240     }
1241 
1242     // The sign bit is the LHS's sign bit, except when the result of the
1243     // remainder is zero.
1244     computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1245     // If it's known zero, our sign bit is also zero.
1246     if (Known2.isNonNegative())
1247       Known.makeNonNegative();
1248 
1249     break;
1250   case Instruction::URem: {
1251     if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1252       const APInt &RA = Rem->getValue();
1253       if (RA.isPowerOf2()) {
1254         APInt LowBits = (RA - 1);
1255         computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1256         Known.Zero |= ~LowBits;
1257         Known.One &= LowBits;
1258         break;
1259       }
1260     }
1261 
1262     // Since the result is less than or equal to either operand, any leading
1263     // zero bits in either operand must also exist in the result.
1264     computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1265     computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
1266 
1267     unsigned Leaders =
1268         std::max(Known.countMinLeadingZeros(), Known2.countMinLeadingZeros());
1269     Known.resetAll();
1270     Known.Zero.setHighBits(Leaders);
1271     break;
1272   }
1273 
1274   case Instruction::Alloca: {
1275     const AllocaInst *AI = cast<AllocaInst>(I);
1276     unsigned Align = AI->getAlignment();
1277     if (Align == 0)
1278       Align = Q.DL.getABITypeAlignment(AI->getAllocatedType());
1279 
1280     if (Align > 0)
1281       Known.Zero.setLowBits(countTrailingZeros(Align));
1282     break;
1283   }
1284   case Instruction::GetElementPtr: {
1285     // Analyze all of the subscripts of this getelementptr instruction
1286     // to determine if we can prove known low zero bits.
1287     KnownBits LocalKnown(BitWidth);
1288     computeKnownBits(I->getOperand(0), LocalKnown, Depth + 1, Q);
1289     unsigned TrailZ = LocalKnown.countMinTrailingZeros();
1290 
1291     gep_type_iterator GTI = gep_type_begin(I);
1292     for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
1293       Value *Index = I->getOperand(i);
1294       if (StructType *STy = GTI.getStructTypeOrNull()) {
1295         // Handle struct member offset arithmetic.
1296 
1297         // Handle case when index is vector zeroinitializer
1298         Constant *CIndex = cast<Constant>(Index);
1299         if (CIndex->isZeroValue())
1300           continue;
1301 
1302         if (CIndex->getType()->isVectorTy())
1303           Index = CIndex->getSplatValue();
1304 
1305         unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
1306         const StructLayout *SL = Q.DL.getStructLayout(STy);
1307         uint64_t Offset = SL->getElementOffset(Idx);
1308         TrailZ = std::min<unsigned>(TrailZ,
1309                                     countTrailingZeros(Offset));
1310       } else {
1311         // Handle array index arithmetic.
1312         Type *IndexedTy = GTI.getIndexedType();
1313         if (!IndexedTy->isSized()) {
1314           TrailZ = 0;
1315           break;
1316         }
1317         unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
1318         uint64_t TypeSize = Q.DL.getTypeAllocSize(IndexedTy);
1319         LocalKnown.Zero = LocalKnown.One = APInt(GEPOpiBits, 0);
1320         computeKnownBits(Index, LocalKnown, Depth + 1, Q);
1321         TrailZ = std::min(TrailZ,
1322                           unsigned(countTrailingZeros(TypeSize) +
1323                                    LocalKnown.countMinTrailingZeros()));
1324       }
1325     }
1326 
1327     Known.Zero.setLowBits(TrailZ);
1328     break;
1329   }
1330   case Instruction::PHI: {
1331     const PHINode *P = cast<PHINode>(I);
1332     // Handle the case of a simple two-predecessor recurrence PHI.
1333     // There's a lot more that could theoretically be done here, but
1334     // this is sufficient to catch some interesting cases.
1335     if (P->getNumIncomingValues() == 2) {
1336       for (unsigned i = 0; i != 2; ++i) {
1337         Value *L = P->getIncomingValue(i);
1338         Value *R = P->getIncomingValue(!i);
1339         Operator *LU = dyn_cast<Operator>(L);
1340         if (!LU)
1341           continue;
1342         unsigned Opcode = LU->getOpcode();
1343         // Check for operations that have the property that if
1344         // both their operands have low zero bits, the result
1345         // will have low zero bits.
1346         if (Opcode == Instruction::Add ||
1347             Opcode == Instruction::Sub ||
1348             Opcode == Instruction::And ||
1349             Opcode == Instruction::Or ||
1350             Opcode == Instruction::Mul) {
1351           Value *LL = LU->getOperand(0);
1352           Value *LR = LU->getOperand(1);
1353           // Find a recurrence.
1354           if (LL == I)
1355             L = LR;
1356           else if (LR == I)
1357             L = LL;
1358           else
1359             break;
1360           // Ok, we have a PHI of the form L op= R. Check for low
1361           // zero bits.
1362           computeKnownBits(R, Known2, Depth + 1, Q);
1363 
1364           // We need to take the minimum number of known bits
1365           KnownBits Known3(Known);
1366           computeKnownBits(L, Known3, Depth + 1, Q);
1367 
1368           Known.Zero.setLowBits(std::min(Known2.countMinTrailingZeros(),
1369                                          Known3.countMinTrailingZeros()));
1370 
1371           auto *OverflowOp = dyn_cast<OverflowingBinaryOperator>(LU);
1372           if (OverflowOp && OverflowOp->hasNoSignedWrap()) {
1373             // If initial value of recurrence is nonnegative, and we are adding
1374             // a nonnegative number with nsw, the result can only be nonnegative
1375             // or poison value regardless of the number of times we execute the
1376             // add in phi recurrence. If initial value is negative and we are
1377             // adding a negative number with nsw, the result can only be
1378             // negative or poison value. Similar arguments apply to sub and mul.
1379             //
1380             // (add non-negative, non-negative) --> non-negative
1381             // (add negative, negative) --> negative
1382             if (Opcode == Instruction::Add) {
1383               if (Known2.isNonNegative() && Known3.isNonNegative())
1384                 Known.makeNonNegative();
1385               else if (Known2.isNegative() && Known3.isNegative())
1386                 Known.makeNegative();
1387             }
1388 
1389             // (sub nsw non-negative, negative) --> non-negative
1390             // (sub nsw negative, non-negative) --> negative
1391             else if (Opcode == Instruction::Sub && LL == I) {
1392               if (Known2.isNonNegative() && Known3.isNegative())
1393                 Known.makeNonNegative();
1394               else if (Known2.isNegative() && Known3.isNonNegative())
1395                 Known.makeNegative();
1396             }
1397 
1398             // (mul nsw non-negative, non-negative) --> non-negative
1399             else if (Opcode == Instruction::Mul && Known2.isNonNegative() &&
1400                      Known3.isNonNegative())
1401               Known.makeNonNegative();
1402           }
1403 
1404           break;
1405         }
1406       }
1407     }
1408 
1409     // Unreachable blocks may have zero-operand PHI nodes.
1410     if (P->getNumIncomingValues() == 0)
1411       break;
1412 
1413     // Otherwise take the unions of the known bit sets of the operands,
1414     // taking conservative care to avoid excessive recursion.
1415     if (Depth < MaxDepth - 1 && !Known.Zero && !Known.One) {
1416       // Skip if every incoming value references to ourself.
1417       if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
1418         break;
1419 
1420       Known.Zero.setAllBits();
1421       Known.One.setAllBits();
1422       for (Value *IncValue : P->incoming_values()) {
1423         // Skip direct self references.
1424         if (IncValue == P) continue;
1425 
1426         Known2 = KnownBits(BitWidth);
1427         // Recurse, but cap the recursion to one level, because we don't
1428         // want to waste time spinning around in loops.
1429         computeKnownBits(IncValue, Known2, MaxDepth - 1, Q);
1430         Known.Zero &= Known2.Zero;
1431         Known.One &= Known2.One;
1432         // If all bits have been ruled out, there's no need to check
1433         // more operands.
1434         if (!Known.Zero && !Known.One)
1435           break;
1436       }
1437     }
1438     break;
1439   }
1440   case Instruction::Call:
1441   case Instruction::Invoke:
1442     // If range metadata is attached to this call, set known bits from that,
1443     // and then intersect with known bits based on other properties of the
1444     // function.
1445     if (MDNode *MD = cast<Instruction>(I)->getMetadata(LLVMContext::MD_range))
1446       computeKnownBitsFromRangeMetadata(*MD, Known);
1447     if (const Value *RV = ImmutableCallSite(I).getReturnedArgOperand()) {
1448       computeKnownBits(RV, Known2, Depth + 1, Q);
1449       Known.Zero |= Known2.Zero;
1450       Known.One |= Known2.One;
1451     }
1452     if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1453       switch (II->getIntrinsicID()) {
1454       default: break;
1455       case Intrinsic::bitreverse:
1456         computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1457         Known.Zero |= Known2.Zero.reverseBits();
1458         Known.One |= Known2.One.reverseBits();
1459         break;
1460       case Intrinsic::bswap:
1461         computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1462         Known.Zero |= Known2.Zero.byteSwap();
1463         Known.One |= Known2.One.byteSwap();
1464         break;
1465       case Intrinsic::ctlz: {
1466         computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1467         // If we have a known 1, its position is our upper bound.
1468         unsigned PossibleLZ = Known2.One.countLeadingZeros();
1469         // If this call is undefined for 0, the result will be less than 2^n.
1470         if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1471           PossibleLZ = std::min(PossibleLZ, BitWidth - 1);
1472         unsigned LowBits = Log2_32(PossibleLZ)+1;
1473         Known.Zero.setBitsFrom(LowBits);
1474         break;
1475       }
1476       case Intrinsic::cttz: {
1477         computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1478         // If we have a known 1, its position is our upper bound.
1479         unsigned PossibleTZ = Known2.One.countTrailingZeros();
1480         // If this call is undefined for 0, the result will be less than 2^n.
1481         if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
1482           PossibleTZ = std::min(PossibleTZ, BitWidth - 1);
1483         unsigned LowBits = Log2_32(PossibleTZ)+1;
1484         Known.Zero.setBitsFrom(LowBits);
1485         break;
1486       }
1487       case Intrinsic::ctpop: {
1488         computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
1489         // We can bound the space the count needs.  Also, bits known to be zero
1490         // can't contribute to the population.
1491         unsigned BitsPossiblySet = Known2.countMaxPopulation();
1492         unsigned LowBits = Log2_32(BitsPossiblySet)+1;
1493         Known.Zero.setBitsFrom(LowBits);
1494         // TODO: we could bound KnownOne using the lower bound on the number
1495         // of bits which might be set provided by popcnt KnownOne2.
1496         break;
1497       }
1498       case Intrinsic::x86_sse42_crc32_64_64:
1499         Known.Zero.setBitsFrom(32);
1500         break;
1501       }
1502     }
1503     break;
1504   case Instruction::ExtractElement:
1505     // Look through extract element. At the moment we keep this simple and skip
1506     // tracking the specific element. But at least we might find information
1507     // valid for all elements of the vector (for example if vector is sign
1508     // extended, shifted, etc).
1509     computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
1510     break;
1511   case Instruction::ExtractValue:
1512     if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
1513       const ExtractValueInst *EVI = cast<ExtractValueInst>(I);
1514       if (EVI->getNumIndices() != 1) break;
1515       if (EVI->getIndices()[0] == 0) {
1516         switch (II->getIntrinsicID()) {
1517         default: break;
1518         case Intrinsic::uadd_with_overflow:
1519         case Intrinsic::sadd_with_overflow:
1520           computeKnownBitsAddSub(true, II->getArgOperand(0),
1521                                  II->getArgOperand(1), false, Known, Known2,
1522                                  Depth, Q);
1523           break;
1524         case Intrinsic::usub_with_overflow:
1525         case Intrinsic::ssub_with_overflow:
1526           computeKnownBitsAddSub(false, II->getArgOperand(0),
1527                                  II->getArgOperand(1), false, Known, Known2,
1528                                  Depth, Q);
1529           break;
1530         case Intrinsic::umul_with_overflow:
1531         case Intrinsic::smul_with_overflow:
1532           computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1), false,
1533                               Known, Known2, Depth, Q);
1534           break;
1535         }
1536       }
1537     }
1538   }
1539 }
1540 
1541 /// Determine which bits of V are known to be either zero or one and return
1542 /// them.
computeKnownBits(const Value * V,unsigned Depth,const Query & Q)1543 KnownBits computeKnownBits(const Value *V, unsigned Depth, const Query &Q) {
1544   KnownBits Known(getBitWidth(V->getType(), Q.DL));
1545   computeKnownBits(V, Known, Depth, Q);
1546   return Known;
1547 }
1548 
1549 /// Determine which bits of V are known to be either zero or one and return
1550 /// them in the Known bit set.
1551 ///
1552 /// NOTE: we cannot consider 'undef' to be "IsZero" here.  The problem is that
1553 /// we cannot optimize based on the assumption that it is zero without changing
1554 /// it to be an explicit zero.  If we don't change it to zero, other code could
1555 /// optimized based on the contradictory assumption that it is non-zero.
1556 /// Because instcombine aggressively folds operations with undef args anyway,
1557 /// this won't lose us code quality.
1558 ///
1559 /// This function is defined on values with integer type, values with pointer
1560 /// type, and vectors of integers.  In the case
1561 /// where V is a vector, known zero, and known one values are the
1562 /// same width as the vector element, and the bit is set only if it is true
1563 /// for all of the elements in the vector.
computeKnownBits(const Value * V,KnownBits & Known,unsigned Depth,const Query & Q)1564 void computeKnownBits(const Value *V, KnownBits &Known, unsigned Depth,
1565                       const Query &Q) {
1566   assert(V && "No Value?");
1567   assert(Depth <= MaxDepth && "Limit Search Depth");
1568   unsigned BitWidth = Known.getBitWidth();
1569 
1570   assert((V->getType()->isIntOrIntVectorTy(BitWidth) ||
1571           V->getType()->isPtrOrPtrVectorTy()) &&
1572          "Not integer or pointer type!");
1573 
1574   Type *ScalarTy = V->getType()->getScalarType();
1575   unsigned ExpectedWidth = ScalarTy->isPointerTy() ?
1576     Q.DL.getIndexTypeSizeInBits(ScalarTy) : Q.DL.getTypeSizeInBits(ScalarTy);
1577   assert(ExpectedWidth == BitWidth && "V and Known should have same BitWidth");
1578   (void)BitWidth;
1579   (void)ExpectedWidth;
1580 
1581   const APInt *C;
1582   if (match(V, m_APInt(C))) {
1583     // We know all of the bits for a scalar constant or a splat vector constant!
1584     Known.One = *C;
1585     Known.Zero = ~Known.One;
1586     return;
1587   }
1588   // Null and aggregate-zero are all-zeros.
1589   if (isa<ConstantPointerNull>(V) || isa<ConstantAggregateZero>(V)) {
1590     Known.setAllZero();
1591     return;
1592   }
1593   // Handle a constant vector by taking the intersection of the known bits of
1594   // each element.
1595   if (const ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
1596     // We know that CDS must be a vector of integers. Take the intersection of
1597     // each element.
1598     Known.Zero.setAllBits(); Known.One.setAllBits();
1599     for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
1600       APInt Elt = CDS->getElementAsAPInt(i);
1601       Known.Zero &= ~Elt;
1602       Known.One &= Elt;
1603     }
1604     return;
1605   }
1606 
1607   if (const auto *CV = dyn_cast<ConstantVector>(V)) {
1608     // We know that CV must be a vector of integers. Take the intersection of
1609     // each element.
1610     Known.Zero.setAllBits(); Known.One.setAllBits();
1611     for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
1612       Constant *Element = CV->getAggregateElement(i);
1613       auto *ElementCI = dyn_cast_or_null<ConstantInt>(Element);
1614       if (!ElementCI) {
1615         Known.resetAll();
1616         return;
1617       }
1618       const APInt &Elt = ElementCI->getValue();
1619       Known.Zero &= ~Elt;
1620       Known.One &= Elt;
1621     }
1622     return;
1623   }
1624 
1625   // Start out not knowing anything.
1626   Known.resetAll();
1627 
1628   // We can't imply anything about undefs.
1629   if (isa<UndefValue>(V))
1630     return;
1631 
1632   // There's no point in looking through other users of ConstantData for
1633   // assumptions.  Confirm that we've handled them all.
1634   assert(!isa<ConstantData>(V) && "Unhandled constant data!");
1635 
1636   // Limit search depth.
1637   // All recursive calls that increase depth must come after this.
1638   if (Depth == MaxDepth)
1639     return;
1640 
1641   // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
1642   // the bits of its aliasee.
1643   if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
1644     if (!GA->isInterposable())
1645       computeKnownBits(GA->getAliasee(), Known, Depth + 1, Q);
1646     return;
1647   }
1648 
1649   if (const Operator *I = dyn_cast<Operator>(V))
1650     computeKnownBitsFromOperator(I, Known, Depth, Q);
1651 
1652   // Aligned pointers have trailing zeros - refine Known.Zero set
1653   if (V->getType()->isPointerTy()) {
1654     unsigned Align = V->getPointerAlignment(Q.DL);
1655     if (Align)
1656       Known.Zero.setLowBits(countTrailingZeros(Align));
1657   }
1658 
1659   // computeKnownBitsFromAssume strictly refines Known.
1660   // Therefore, we run them after computeKnownBitsFromOperator.
1661 
1662   // Check whether a nearby assume intrinsic can determine some known bits.
1663   computeKnownBitsFromAssume(V, Known, Depth, Q);
1664 
1665   assert((Known.Zero & Known.One) == 0 && "Bits known to be one AND zero?");
1666 }
1667 
1668 /// Return true if the given value is known to have exactly one
1669 /// bit set when defined. For vectors return true if every element is known to
1670 /// be a power of two when defined. Supports values with integer or pointer
1671 /// types and vectors of integers.
isKnownToBeAPowerOfTwo(const Value * V,bool OrZero,unsigned Depth,const Query & Q)1672 bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth,
1673                             const Query &Q) {
1674   assert(Depth <= MaxDepth && "Limit Search Depth");
1675 
1676   // Attempt to match against constants.
1677   if (OrZero && match(V, m_Power2OrZero()))
1678       return true;
1679   if (match(V, m_Power2()))
1680       return true;
1681 
1682   // 1 << X is clearly a power of two if the one is not shifted off the end.  If
1683   // it is shifted off the end then the result is undefined.
1684   if (match(V, m_Shl(m_One(), m_Value())))
1685     return true;
1686 
1687   // (signmask) >>l X is clearly a power of two if the one is not shifted off
1688   // the bottom.  If it is shifted off the bottom then the result is undefined.
1689   if (match(V, m_LShr(m_SignMask(), m_Value())))
1690     return true;
1691 
1692   // The remaining tests are all recursive, so bail out if we hit the limit.
1693   if (Depth++ == MaxDepth)
1694     return false;
1695 
1696   Value *X = nullptr, *Y = nullptr;
1697   // A shift left or a logical shift right of a power of two is a power of two
1698   // or zero.
1699   if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
1700                  match(V, m_LShr(m_Value(X), m_Value()))))
1701     return isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q);
1702 
1703   if (const ZExtInst *ZI = dyn_cast<ZExtInst>(V))
1704     return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q);
1705 
1706   if (const SelectInst *SI = dyn_cast<SelectInst>(V))
1707     return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q) &&
1708            isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q);
1709 
1710   if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
1711     // A power of two and'd with anything is a power of two or zero.
1712     if (isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q) ||
1713         isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, Depth, Q))
1714       return true;
1715     // X & (-X) is always a power of two or zero.
1716     if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
1717       return true;
1718     return false;
1719   }
1720 
1721   // Adding a power-of-two or zero to the same power-of-two or zero yields
1722   // either the original power-of-two, a larger power-of-two or zero.
1723   if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
1724     const OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
1725     if (OrZero || VOBO->hasNoUnsignedWrap() || VOBO->hasNoSignedWrap()) {
1726       if (match(X, m_And(m_Specific(Y), m_Value())) ||
1727           match(X, m_And(m_Value(), m_Specific(Y))))
1728         if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q))
1729           return true;
1730       if (match(Y, m_And(m_Specific(X), m_Value())) ||
1731           match(Y, m_And(m_Value(), m_Specific(X))))
1732         if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q))
1733           return true;
1734 
1735       unsigned BitWidth = V->getType()->getScalarSizeInBits();
1736       KnownBits LHSBits(BitWidth);
1737       computeKnownBits(X, LHSBits, Depth, Q);
1738 
1739       KnownBits RHSBits(BitWidth);
1740       computeKnownBits(Y, RHSBits, Depth, Q);
1741       // If i8 V is a power of two or zero:
1742       //  ZeroBits: 1 1 1 0 1 1 1 1
1743       // ~ZeroBits: 0 0 0 1 0 0 0 0
1744       if ((~(LHSBits.Zero & RHSBits.Zero)).isPowerOf2())
1745         // If OrZero isn't set, we cannot give back a zero result.
1746         // Make sure either the LHS or RHS has a bit set.
1747         if (OrZero || RHSBits.One.getBoolValue() || LHSBits.One.getBoolValue())
1748           return true;
1749     }
1750   }
1751 
1752   // An exact divide or right shift can only shift off zero bits, so the result
1753   // is a power of two only if the first operand is a power of two and not
1754   // copying a sign bit (sdiv int_min, 2).
1755   if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
1756       match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
1757     return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero,
1758                                   Depth, Q);
1759   }
1760 
1761   return false;
1762 }
1763 
1764 /// Test whether a GEP's result is known to be non-null.
1765 ///
1766 /// Uses properties inherent in a GEP to try to determine whether it is known
1767 /// to be non-null.
1768 ///
1769 /// Currently this routine does not support vector GEPs.
isGEPKnownNonNull(const GEPOperator * GEP,unsigned Depth,const Query & Q)1770 static bool isGEPKnownNonNull(const GEPOperator *GEP, unsigned Depth,
1771                               const Query &Q) {
1772   const Function *F = nullptr;
1773   if (const Instruction *I = dyn_cast<Instruction>(GEP))
1774     F = I->getFunction();
1775 
1776   if (!GEP->isInBounds() ||
1777       NullPointerIsDefined(F, GEP->getPointerAddressSpace()))
1778     return false;
1779 
1780   // FIXME: Support vector-GEPs.
1781   assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
1782 
1783   // If the base pointer is non-null, we cannot walk to a null address with an
1784   // inbounds GEP in address space zero.
1785   if (isKnownNonZero(GEP->getPointerOperand(), Depth, Q))
1786     return true;
1787 
1788   // Walk the GEP operands and see if any operand introduces a non-zero offset.
1789   // If so, then the GEP cannot produce a null pointer, as doing so would
1790   // inherently violate the inbounds contract within address space zero.
1791   for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1792        GTI != GTE; ++GTI) {
1793     // Struct types are easy -- they must always be indexed by a constant.
1794     if (StructType *STy = GTI.getStructTypeOrNull()) {
1795       ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
1796       unsigned ElementIdx = OpC->getZExtValue();
1797       const StructLayout *SL = Q.DL.getStructLayout(STy);
1798       uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
1799       if (ElementOffset > 0)
1800         return true;
1801       continue;
1802     }
1803 
1804     // If we have a zero-sized type, the index doesn't matter. Keep looping.
1805     if (Q.DL.getTypeAllocSize(GTI.getIndexedType()) == 0)
1806       continue;
1807 
1808     // Fast path the constant operand case both for efficiency and so we don't
1809     // increment Depth when just zipping down an all-constant GEP.
1810     if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
1811       if (!OpC->isZero())
1812         return true;
1813       continue;
1814     }
1815 
1816     // We post-increment Depth here because while isKnownNonZero increments it
1817     // as well, when we pop back up that increment won't persist. We don't want
1818     // to recurse 10k times just because we have 10k GEP operands. We don't
1819     // bail completely out because we want to handle constant GEPs regardless
1820     // of depth.
1821     if (Depth++ >= MaxDepth)
1822       continue;
1823 
1824     if (isKnownNonZero(GTI.getOperand(), Depth, Q))
1825       return true;
1826   }
1827 
1828   return false;
1829 }
1830 
isKnownNonNullFromDominatingCondition(const Value * V,const Instruction * CtxI,const DominatorTree * DT)1831 static bool isKnownNonNullFromDominatingCondition(const Value *V,
1832                                                   const Instruction *CtxI,
1833                                                   const DominatorTree *DT) {
1834   assert(V->getType()->isPointerTy() && "V must be pointer type");
1835   assert(!isa<ConstantData>(V) && "Did not expect ConstantPointerNull");
1836 
1837   if (!CtxI || !DT)
1838     return false;
1839 
1840   unsigned NumUsesExplored = 0;
1841   for (auto *U : V->users()) {
1842     // Avoid massive lists
1843     if (NumUsesExplored >= DomConditionsMaxUses)
1844       break;
1845     NumUsesExplored++;
1846 
1847     // If the value is used as an argument to a call or invoke, then argument
1848     // attributes may provide an answer about null-ness.
1849     if (auto CS = ImmutableCallSite(U))
1850       if (auto *CalledFunc = CS.getCalledFunction())
1851         for (const Argument &Arg : CalledFunc->args())
1852           if (CS.getArgOperand(Arg.getArgNo()) == V &&
1853               Arg.hasNonNullAttr() && DT->dominates(CS.getInstruction(), CtxI))
1854             return true;
1855 
1856     // Consider only compare instructions uniquely controlling a branch
1857     CmpInst::Predicate Pred;
1858     if (!match(const_cast<User *>(U),
1859                m_c_ICmp(Pred, m_Specific(V), m_Zero())) ||
1860         (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE))
1861       continue;
1862 
1863     for (auto *CmpU : U->users()) {
1864       if (const BranchInst *BI = dyn_cast<BranchInst>(CmpU)) {
1865         assert(BI->isConditional() && "uses a comparison!");
1866 
1867         BasicBlock *NonNullSuccessor =
1868             BI->getSuccessor(Pred == ICmpInst::ICMP_EQ ? 1 : 0);
1869         BasicBlockEdge Edge(BI->getParent(), NonNullSuccessor);
1870         if (Edge.isSingleEdge() && DT->dominates(Edge, CtxI->getParent()))
1871           return true;
1872       } else if (Pred == ICmpInst::ICMP_NE &&
1873                  match(CmpU, m_Intrinsic<Intrinsic::experimental_guard>()) &&
1874                  DT->dominates(cast<Instruction>(CmpU), CtxI)) {
1875         return true;
1876       }
1877     }
1878   }
1879 
1880   return false;
1881 }
1882 
1883 /// Does the 'Range' metadata (which must be a valid MD_range operand list)
1884 /// ensure that the value it's attached to is never Value?  'RangeType' is
1885 /// is the type of the value described by the range.
rangeMetadataExcludesValue(const MDNode * Ranges,const APInt & Value)1886 static bool rangeMetadataExcludesValue(const MDNode* Ranges, const APInt& Value) {
1887   const unsigned NumRanges = Ranges->getNumOperands() / 2;
1888   assert(NumRanges >= 1);
1889   for (unsigned i = 0; i < NumRanges; ++i) {
1890     ConstantInt *Lower =
1891         mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0));
1892     ConstantInt *Upper =
1893         mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1));
1894     ConstantRange Range(Lower->getValue(), Upper->getValue());
1895     if (Range.contains(Value))
1896       return false;
1897   }
1898   return true;
1899 }
1900 
1901 /// Return true if the given value is known to be non-zero when defined. For
1902 /// vectors, return true if every element is known to be non-zero when
1903 /// defined. For pointers, if the context instruction and dominator tree are
1904 /// specified, perform context-sensitive analysis and return true if the
1905 /// pointer couldn't possibly be null at the specified instruction.
1906 /// Supports values with integer or pointer type and vectors of integers.
isKnownNonZero(const Value * V,unsigned Depth,const Query & Q)1907 bool isKnownNonZero(const Value *V, unsigned Depth, const Query &Q) {
1908   if (auto *C = dyn_cast<Constant>(V)) {
1909     if (C->isNullValue())
1910       return false;
1911     if (isa<ConstantInt>(C))
1912       // Must be non-zero due to null test above.
1913       return true;
1914 
1915     // For constant vectors, check that all elements are undefined or known
1916     // non-zero to determine that the whole vector is known non-zero.
1917     if (auto *VecTy = dyn_cast<VectorType>(C->getType())) {
1918       for (unsigned i = 0, e = VecTy->getNumElements(); i != e; ++i) {
1919         Constant *Elt = C->getAggregateElement(i);
1920         if (!Elt || Elt->isNullValue())
1921           return false;
1922         if (!isa<UndefValue>(Elt) && !isa<ConstantInt>(Elt))
1923           return false;
1924       }
1925       return true;
1926     }
1927 
1928     // A global variable in address space 0 is non null unless extern weak
1929     // or an absolute symbol reference. Other address spaces may have null as a
1930     // valid address for a global, so we can't assume anything.
1931     if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
1932       if (!GV->isAbsoluteSymbolRef() && !GV->hasExternalWeakLinkage() &&
1933           GV->getType()->getAddressSpace() == 0)
1934         return true;
1935     } else
1936       return false;
1937   }
1938 
1939   if (auto *I = dyn_cast<Instruction>(V)) {
1940     if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) {
1941       // If the possible ranges don't contain zero, then the value is
1942       // definitely non-zero.
1943       if (auto *Ty = dyn_cast<IntegerType>(V->getType())) {
1944         const APInt ZeroValue(Ty->getBitWidth(), 0);
1945         if (rangeMetadataExcludesValue(Ranges, ZeroValue))
1946           return true;
1947       }
1948     }
1949   }
1950 
1951   // Some of the tests below are recursive, so bail out if we hit the limit.
1952   if (Depth++ >= MaxDepth)
1953     return false;
1954 
1955   // Check for pointer simplifications.
1956   if (V->getType()->isPointerTy()) {
1957     // Alloca never returns null, malloc might.
1958     if (isa<AllocaInst>(V) && Q.DL.getAllocaAddrSpace() == 0)
1959       return true;
1960 
1961     // A byval, inalloca, or nonnull argument is never null.
1962     if (const Argument *A = dyn_cast<Argument>(V))
1963       if (A->hasByValOrInAllocaAttr() || A->hasNonNullAttr())
1964         return true;
1965 
1966     // A Load tagged with nonnull metadata is never null.
1967     if (const LoadInst *LI = dyn_cast<LoadInst>(V))
1968       if (LI->getMetadata(LLVMContext::MD_nonnull))
1969         return true;
1970 
1971     if (auto CS = ImmutableCallSite(V)) {
1972       if (CS.isReturnNonNull())
1973         return true;
1974       if (const auto *RP = getArgumentAliasingToReturnedPointer(CS))
1975         return isKnownNonZero(RP, Depth, Q);
1976     }
1977   }
1978 
1979 
1980   // Check for recursive pointer simplifications.
1981   if (V->getType()->isPointerTy()) {
1982     if (isKnownNonNullFromDominatingCondition(V, Q.CxtI, Q.DT))
1983       return true;
1984 
1985     if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V))
1986       if (isGEPKnownNonNull(GEP, Depth, Q))
1987         return true;
1988   }
1989 
1990   unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), Q.DL);
1991 
1992   // X | Y != 0 if X != 0 or Y != 0.
1993   Value *X = nullptr, *Y = nullptr;
1994   if (match(V, m_Or(m_Value(X), m_Value(Y))))
1995     return isKnownNonZero(X, Depth, Q) || isKnownNonZero(Y, Depth, Q);
1996 
1997   // ext X != 0 if X != 0.
1998   if (isa<SExtInst>(V) || isa<ZExtInst>(V))
1999     return isKnownNonZero(cast<Instruction>(V)->getOperand(0), Depth, Q);
2000 
2001   // shl X, Y != 0 if X is odd.  Note that the value of the shift is undefined
2002   // if the lowest bit is shifted off the end.
2003   if (match(V, m_Shl(m_Value(X), m_Value(Y)))) {
2004     // shl nuw can't remove any non-zero bits.
2005     const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
2006     if (BO->hasNoUnsignedWrap())
2007       return isKnownNonZero(X, Depth, Q);
2008 
2009     KnownBits Known(BitWidth);
2010     computeKnownBits(X, Known, Depth, Q);
2011     if (Known.One[0])
2012       return true;
2013   }
2014   // shr X, Y != 0 if X is negative.  Note that the value of the shift is not
2015   // defined if the sign bit is shifted off the end.
2016   else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
2017     // shr exact can only shift out zero bits.
2018     const PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
2019     if (BO->isExact())
2020       return isKnownNonZero(X, Depth, Q);
2021 
2022     KnownBits Known = computeKnownBits(X, Depth, Q);
2023     if (Known.isNegative())
2024       return true;
2025 
2026     // If the shifter operand is a constant, and all of the bits shifted
2027     // out are known to be zero, and X is known non-zero then at least one
2028     // non-zero bit must remain.
2029     if (ConstantInt *Shift = dyn_cast<ConstantInt>(Y)) {
2030       auto ShiftVal = Shift->getLimitedValue(BitWidth - 1);
2031       // Is there a known one in the portion not shifted out?
2032       if (Known.countMaxLeadingZeros() < BitWidth - ShiftVal)
2033         return true;
2034       // Are all the bits to be shifted out known zero?
2035       if (Known.countMinTrailingZeros() >= ShiftVal)
2036         return isKnownNonZero(X, Depth, Q);
2037     }
2038   }
2039   // div exact can only produce a zero if the dividend is zero.
2040   else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
2041     return isKnownNonZero(X, Depth, Q);
2042   }
2043   // X + Y.
2044   else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
2045     KnownBits XKnown = computeKnownBits(X, Depth, Q);
2046     KnownBits YKnown = computeKnownBits(Y, Depth, Q);
2047 
2048     // If X and Y are both non-negative (as signed values) then their sum is not
2049     // zero unless both X and Y are zero.
2050     if (XKnown.isNonNegative() && YKnown.isNonNegative())
2051       if (isKnownNonZero(X, Depth, Q) || isKnownNonZero(Y, Depth, Q))
2052         return true;
2053 
2054     // If X and Y are both negative (as signed values) then their sum is not
2055     // zero unless both X and Y equal INT_MIN.
2056     if (XKnown.isNegative() && YKnown.isNegative()) {
2057       APInt Mask = APInt::getSignedMaxValue(BitWidth);
2058       // The sign bit of X is set.  If some other bit is set then X is not equal
2059       // to INT_MIN.
2060       if (XKnown.One.intersects(Mask))
2061         return true;
2062       // The sign bit of Y is set.  If some other bit is set then Y is not equal
2063       // to INT_MIN.
2064       if (YKnown.One.intersects(Mask))
2065         return true;
2066     }
2067 
2068     // The sum of a non-negative number and a power of two is not zero.
2069     if (XKnown.isNonNegative() &&
2070         isKnownToBeAPowerOfTwo(Y, /*OrZero*/ false, Depth, Q))
2071       return true;
2072     if (YKnown.isNonNegative() &&
2073         isKnownToBeAPowerOfTwo(X, /*OrZero*/ false, Depth, Q))
2074       return true;
2075   }
2076   // X * Y.
2077   else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
2078     const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
2079     // If X and Y are non-zero then so is X * Y as long as the multiplication
2080     // does not overflow.
2081     if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
2082         isKnownNonZero(X, Depth, Q) && isKnownNonZero(Y, Depth, Q))
2083       return true;
2084   }
2085   // (C ? X : Y) != 0 if X != 0 and Y != 0.
2086   else if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
2087     if (isKnownNonZero(SI->getTrueValue(), Depth, Q) &&
2088         isKnownNonZero(SI->getFalseValue(), Depth, Q))
2089       return true;
2090   }
2091   // PHI
2092   else if (const PHINode *PN = dyn_cast<PHINode>(V)) {
2093     // Try and detect a recurrence that monotonically increases from a
2094     // starting value, as these are common as induction variables.
2095     if (PN->getNumIncomingValues() == 2) {
2096       Value *Start = PN->getIncomingValue(0);
2097       Value *Induction = PN->getIncomingValue(1);
2098       if (isa<ConstantInt>(Induction) && !isa<ConstantInt>(Start))
2099         std::swap(Start, Induction);
2100       if (ConstantInt *C = dyn_cast<ConstantInt>(Start)) {
2101         if (!C->isZero() && !C->isNegative()) {
2102           ConstantInt *X;
2103           if ((match(Induction, m_NSWAdd(m_Specific(PN), m_ConstantInt(X))) ||
2104                match(Induction, m_NUWAdd(m_Specific(PN), m_ConstantInt(X)))) &&
2105               !X->isNegative())
2106             return true;
2107         }
2108       }
2109     }
2110     // Check if all incoming values are non-zero constant.
2111     bool AllNonZeroConstants = llvm::all_of(PN->operands(), [](Value *V) {
2112       return isa<ConstantInt>(V) && !cast<ConstantInt>(V)->isZero();
2113     });
2114     if (AllNonZeroConstants)
2115       return true;
2116   }
2117 
2118   KnownBits Known(BitWidth);
2119   computeKnownBits(V, Known, Depth, Q);
2120   return Known.One != 0;
2121 }
2122 
2123 /// Return true if V2 == V1 + X, where X is known non-zero.
isAddOfNonZero(const Value * V1,const Value * V2,const Query & Q)2124 static bool isAddOfNonZero(const Value *V1, const Value *V2, const Query &Q) {
2125   const BinaryOperator *BO = dyn_cast<BinaryOperator>(V1);
2126   if (!BO || BO->getOpcode() != Instruction::Add)
2127     return false;
2128   Value *Op = nullptr;
2129   if (V2 == BO->getOperand(0))
2130     Op = BO->getOperand(1);
2131   else if (V2 == BO->getOperand(1))
2132     Op = BO->getOperand(0);
2133   else
2134     return false;
2135   return isKnownNonZero(Op, 0, Q);
2136 }
2137 
2138 /// Return true if it is known that V1 != V2.
isKnownNonEqual(const Value * V1,const Value * V2,const Query & Q)2139 static bool isKnownNonEqual(const Value *V1, const Value *V2, const Query &Q) {
2140   if (V1 == V2)
2141     return false;
2142   if (V1->getType() != V2->getType())
2143     // We can't look through casts yet.
2144     return false;
2145   if (isAddOfNonZero(V1, V2, Q) || isAddOfNonZero(V2, V1, Q))
2146     return true;
2147 
2148   if (V1->getType()->isIntOrIntVectorTy()) {
2149     // Are any known bits in V1 contradictory to known bits in V2? If V1
2150     // has a known zero where V2 has a known one, they must not be equal.
2151     KnownBits Known1 = computeKnownBits(V1, 0, Q);
2152     KnownBits Known2 = computeKnownBits(V2, 0, Q);
2153 
2154     if (Known1.Zero.intersects(Known2.One) ||
2155         Known2.Zero.intersects(Known1.One))
2156       return true;
2157   }
2158   return false;
2159 }
2160 
2161 /// Return true if 'V & Mask' is known to be zero.  We use this predicate to
2162 /// simplify operations downstream. Mask is known to be zero for bits that V
2163 /// cannot have.
2164 ///
2165 /// This function is defined on values with integer type, values with pointer
2166 /// type, and vectors of integers.  In the case
2167 /// where V is a vector, the mask, known zero, and known one values are the
2168 /// same width as the vector element, and the bit is set only if it is true
2169 /// for all of the elements in the vector.
MaskedValueIsZero(const Value * V,const APInt & Mask,unsigned Depth,const Query & Q)2170 bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth,
2171                        const Query &Q) {
2172   KnownBits Known(Mask.getBitWidth());
2173   computeKnownBits(V, Known, Depth, Q);
2174   return Mask.isSubsetOf(Known.Zero);
2175 }
2176 
2177 /// For vector constants, loop over the elements and find the constant with the
2178 /// minimum number of sign bits. Return 0 if the value is not a vector constant
2179 /// or if any element was not analyzed; otherwise, return the count for the
2180 /// element with the minimum number of sign bits.
computeNumSignBitsVectorConstant(const Value * V,unsigned TyBits)2181 static unsigned computeNumSignBitsVectorConstant(const Value *V,
2182                                                  unsigned TyBits) {
2183   const auto *CV = dyn_cast<Constant>(V);
2184   if (!CV || !CV->getType()->isVectorTy())
2185     return 0;
2186 
2187   unsigned MinSignBits = TyBits;
2188   unsigned NumElts = CV->getType()->getVectorNumElements();
2189   for (unsigned i = 0; i != NumElts; ++i) {
2190     // If we find a non-ConstantInt, bail out.
2191     auto *Elt = dyn_cast_or_null<ConstantInt>(CV->getAggregateElement(i));
2192     if (!Elt)
2193       return 0;
2194 
2195     MinSignBits = std::min(MinSignBits, Elt->getValue().getNumSignBits());
2196   }
2197 
2198   return MinSignBits;
2199 }
2200 
2201 static unsigned ComputeNumSignBitsImpl(const Value *V, unsigned Depth,
2202                                        const Query &Q);
2203 
ComputeNumSignBits(const Value * V,unsigned Depth,const Query & Q)2204 static unsigned ComputeNumSignBits(const Value *V, unsigned Depth,
2205                                    const Query &Q) {
2206   unsigned Result = ComputeNumSignBitsImpl(V, Depth, Q);
2207   assert(Result > 0 && "At least one sign bit needs to be present!");
2208   return Result;
2209 }
2210 
2211 /// Return the number of times the sign bit of the register is replicated into
2212 /// the other bits. We know that at least 1 bit is always equal to the sign bit
2213 /// (itself), but other cases can give us information. For example, immediately
2214 /// after an "ashr X, 2", we know that the top 3 bits are all equal to each
2215 /// other, so we return 3. For vectors, return the number of sign bits for the
2216 /// vector element with the minimum number of known sign bits.
ComputeNumSignBitsImpl(const Value * V,unsigned Depth,const Query & Q)2217 static unsigned ComputeNumSignBitsImpl(const Value *V, unsigned Depth,
2218                                        const Query &Q) {
2219   assert(Depth <= MaxDepth && "Limit Search Depth");
2220 
2221   // We return the minimum number of sign bits that are guaranteed to be present
2222   // in V, so for undef we have to conservatively return 1.  We don't have the
2223   // same behavior for poison though -- that's a FIXME today.
2224 
2225   Type *ScalarTy = V->getType()->getScalarType();
2226   unsigned TyBits = ScalarTy->isPointerTy() ?
2227     Q.DL.getIndexTypeSizeInBits(ScalarTy) :
2228     Q.DL.getTypeSizeInBits(ScalarTy);
2229 
2230   unsigned Tmp, Tmp2;
2231   unsigned FirstAnswer = 1;
2232 
2233   // Note that ConstantInt is handled by the general computeKnownBits case
2234   // below.
2235 
2236   if (Depth == MaxDepth)
2237     return 1;  // Limit search depth.
2238 
2239   const Operator *U = dyn_cast<Operator>(V);
2240   switch (Operator::getOpcode(V)) {
2241   default: break;
2242   case Instruction::SExt:
2243     Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
2244     return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q) + Tmp;
2245 
2246   case Instruction::SDiv: {
2247     const APInt *Denominator;
2248     // sdiv X, C -> adds log(C) sign bits.
2249     if (match(U->getOperand(1), m_APInt(Denominator))) {
2250 
2251       // Ignore non-positive denominator.
2252       if (!Denominator->isStrictlyPositive())
2253         break;
2254 
2255       // Calculate the incoming numerator bits.
2256       unsigned NumBits = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2257 
2258       // Add floor(log(C)) bits to the numerator bits.
2259       return std::min(TyBits, NumBits + Denominator->logBase2());
2260     }
2261     break;
2262   }
2263 
2264   case Instruction::SRem: {
2265     const APInt *Denominator;
2266     // srem X, C -> we know that the result is within [-C+1,C) when C is a
2267     // positive constant.  This let us put a lower bound on the number of sign
2268     // bits.
2269     if (match(U->getOperand(1), m_APInt(Denominator))) {
2270 
2271       // Ignore non-positive denominator.
2272       if (!Denominator->isStrictlyPositive())
2273         break;
2274 
2275       // Calculate the incoming numerator bits. SRem by a positive constant
2276       // can't lower the number of sign bits.
2277       unsigned NumrBits =
2278           ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2279 
2280       // Calculate the leading sign bit constraints by examining the
2281       // denominator.  Given that the denominator is positive, there are two
2282       // cases:
2283       //
2284       //  1. the numerator is positive.  The result range is [0,C) and [0,C) u<
2285       //     (1 << ceilLogBase2(C)).
2286       //
2287       //  2. the numerator is negative.  Then the result range is (-C,0] and
2288       //     integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)).
2289       //
2290       // Thus a lower bound on the number of sign bits is `TyBits -
2291       // ceilLogBase2(C)`.
2292 
2293       unsigned ResBits = TyBits - Denominator->ceilLogBase2();
2294       return std::max(NumrBits, ResBits);
2295     }
2296     break;
2297   }
2298 
2299   case Instruction::AShr: {
2300     Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2301     // ashr X, C   -> adds C sign bits.  Vectors too.
2302     const APInt *ShAmt;
2303     if (match(U->getOperand(1), m_APInt(ShAmt))) {
2304       if (ShAmt->uge(TyBits))
2305         break;  // Bad shift.
2306       unsigned ShAmtLimited = ShAmt->getZExtValue();
2307       Tmp += ShAmtLimited;
2308       if (Tmp > TyBits) Tmp = TyBits;
2309     }
2310     return Tmp;
2311   }
2312   case Instruction::Shl: {
2313     const APInt *ShAmt;
2314     if (match(U->getOperand(1), m_APInt(ShAmt))) {
2315       // shl destroys sign bits.
2316       Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2317       if (ShAmt->uge(TyBits) ||      // Bad shift.
2318           ShAmt->uge(Tmp)) break;    // Shifted all sign bits out.
2319       Tmp2 = ShAmt->getZExtValue();
2320       return Tmp - Tmp2;
2321     }
2322     break;
2323   }
2324   case Instruction::And:
2325   case Instruction::Or:
2326   case Instruction::Xor:    // NOT is handled here.
2327     // Logical binary ops preserve the number of sign bits at the worst.
2328     Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2329     if (Tmp != 1) {
2330       Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2331       FirstAnswer = std::min(Tmp, Tmp2);
2332       // We computed what we know about the sign bits as our first
2333       // answer. Now proceed to the generic code that uses
2334       // computeKnownBits, and pick whichever answer is better.
2335     }
2336     break;
2337 
2338   case Instruction::Select:
2339     Tmp = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2340     if (Tmp == 1) break;
2341     Tmp2 = ComputeNumSignBits(U->getOperand(2), Depth + 1, Q);
2342     return std::min(Tmp, Tmp2);
2343 
2344   case Instruction::Add:
2345     // Add can have at most one carry bit.  Thus we know that the output
2346     // is, at worst, one more bit than the inputs.
2347     Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2348     if (Tmp == 1) break;
2349 
2350     // Special case decrementing a value (ADD X, -1):
2351     if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1)))
2352       if (CRHS->isAllOnesValue()) {
2353         KnownBits Known(TyBits);
2354         computeKnownBits(U->getOperand(0), Known, Depth + 1, Q);
2355 
2356         // If the input is known to be 0 or 1, the output is 0/-1, which is all
2357         // sign bits set.
2358         if ((Known.Zero | 1).isAllOnesValue())
2359           return TyBits;
2360 
2361         // If we are subtracting one from a positive number, there is no carry
2362         // out of the result.
2363         if (Known.isNonNegative())
2364           return Tmp;
2365       }
2366 
2367     Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2368     if (Tmp2 == 1) break;
2369     return std::min(Tmp, Tmp2)-1;
2370 
2371   case Instruction::Sub:
2372     Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2373     if (Tmp2 == 1) break;
2374 
2375     // Handle NEG.
2376     if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0)))
2377       if (CLHS->isNullValue()) {
2378         KnownBits Known(TyBits);
2379         computeKnownBits(U->getOperand(1), Known, Depth + 1, Q);
2380         // If the input is known to be 0 or 1, the output is 0/-1, which is all
2381         // sign bits set.
2382         if ((Known.Zero | 1).isAllOnesValue())
2383           return TyBits;
2384 
2385         // If the input is known to be positive (the sign bit is known clear),
2386         // the output of the NEG has the same number of sign bits as the input.
2387         if (Known.isNonNegative())
2388           return Tmp2;
2389 
2390         // Otherwise, we treat this like a SUB.
2391       }
2392 
2393     // Sub can have at most one carry bit.  Thus we know that the output
2394     // is, at worst, one more bit than the inputs.
2395     Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2396     if (Tmp == 1) break;
2397     return std::min(Tmp, Tmp2)-1;
2398 
2399   case Instruction::Mul: {
2400     // The output of the Mul can be at most twice the valid bits in the inputs.
2401     unsigned SignBitsOp0 = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2402     if (SignBitsOp0 == 1) break;
2403     unsigned SignBitsOp1 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
2404     if (SignBitsOp1 == 1) break;
2405     unsigned OutValidBits =
2406         (TyBits - SignBitsOp0 + 1) + (TyBits - SignBitsOp1 + 1);
2407     return OutValidBits > TyBits ? 1 : TyBits - OutValidBits + 1;
2408   }
2409 
2410   case Instruction::PHI: {
2411     const PHINode *PN = cast<PHINode>(U);
2412     unsigned NumIncomingValues = PN->getNumIncomingValues();
2413     // Don't analyze large in-degree PHIs.
2414     if (NumIncomingValues > 4) break;
2415     // Unreachable blocks may have zero-operand PHI nodes.
2416     if (NumIncomingValues == 0) break;
2417 
2418     // Take the minimum of all incoming values.  This can't infinitely loop
2419     // because of our depth threshold.
2420     Tmp = ComputeNumSignBits(PN->getIncomingValue(0), Depth + 1, Q);
2421     for (unsigned i = 1, e = NumIncomingValues; i != e; ++i) {
2422       if (Tmp == 1) return Tmp;
2423       Tmp = std::min(
2424           Tmp, ComputeNumSignBits(PN->getIncomingValue(i), Depth + 1, Q));
2425     }
2426     return Tmp;
2427   }
2428 
2429   case Instruction::Trunc:
2430     // FIXME: it's tricky to do anything useful for this, but it is an important
2431     // case for targets like X86.
2432     break;
2433 
2434   case Instruction::ExtractElement:
2435     // Look through extract element. At the moment we keep this simple and skip
2436     // tracking the specific element. But at least we might find information
2437     // valid for all elements of the vector (for example if vector is sign
2438     // extended, shifted, etc).
2439     return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
2440   }
2441 
2442   // Finally, if we can prove that the top bits of the result are 0's or 1's,
2443   // use this information.
2444 
2445   // If we can examine all elements of a vector constant successfully, we're
2446   // done (we can't do any better than that). If not, keep trying.
2447   if (unsigned VecSignBits = computeNumSignBitsVectorConstant(V, TyBits))
2448     return VecSignBits;
2449 
2450   KnownBits Known(TyBits);
2451   computeKnownBits(V, Known, Depth, Q);
2452 
2453   // If we know that the sign bit is either zero or one, determine the number of
2454   // identical bits in the top of the input value.
2455   return std::max(FirstAnswer, Known.countMinSignBits());
2456 }
2457 
2458 /// This function computes the integer multiple of Base that equals V.
2459 /// If successful, it returns true and returns the multiple in
2460 /// Multiple. If unsuccessful, it returns false. It looks
2461 /// through SExt instructions only if LookThroughSExt is true.
ComputeMultiple(Value * V,unsigned Base,Value * & Multiple,bool LookThroughSExt,unsigned Depth)2462 bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
2463                            bool LookThroughSExt, unsigned Depth) {
2464   const unsigned MaxDepth = 6;
2465 
2466   assert(V && "No Value?");
2467   assert(Depth <= MaxDepth && "Limit Search Depth");
2468   assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
2469 
2470   Type *T = V->getType();
2471 
2472   ConstantInt *CI = dyn_cast<ConstantInt>(V);
2473 
2474   if (Base == 0)
2475     return false;
2476 
2477   if (Base == 1) {
2478     Multiple = V;
2479     return true;
2480   }
2481 
2482   ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
2483   Constant *BaseVal = ConstantInt::get(T, Base);
2484   if (CO && CO == BaseVal) {
2485     // Multiple is 1.
2486     Multiple = ConstantInt::get(T, 1);
2487     return true;
2488   }
2489 
2490   if (CI && CI->getZExtValue() % Base == 0) {
2491     Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
2492     return true;
2493   }
2494 
2495   if (Depth == MaxDepth) return false;  // Limit search depth.
2496 
2497   Operator *I = dyn_cast<Operator>(V);
2498   if (!I) return false;
2499 
2500   switch (I->getOpcode()) {
2501   default: break;
2502   case Instruction::SExt:
2503     if (!LookThroughSExt) return false;
2504     // otherwise fall through to ZExt
2505     LLVM_FALLTHROUGH;
2506   case Instruction::ZExt:
2507     return ComputeMultiple(I->getOperand(0), Base, Multiple,
2508                            LookThroughSExt, Depth+1);
2509   case Instruction::Shl:
2510   case Instruction::Mul: {
2511     Value *Op0 = I->getOperand(0);
2512     Value *Op1 = I->getOperand(1);
2513 
2514     if (I->getOpcode() == Instruction::Shl) {
2515       ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
2516       if (!Op1CI) return false;
2517       // Turn Op0 << Op1 into Op0 * 2^Op1
2518       APInt Op1Int = Op1CI->getValue();
2519       uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
2520       APInt API(Op1Int.getBitWidth(), 0);
2521       API.setBit(BitToSet);
2522       Op1 = ConstantInt::get(V->getContext(), API);
2523     }
2524 
2525     Value *Mul0 = nullptr;
2526     if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
2527       if (Constant *Op1C = dyn_cast<Constant>(Op1))
2528         if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
2529           if (Op1C->getType()->getPrimitiveSizeInBits() <
2530               MulC->getType()->getPrimitiveSizeInBits())
2531             Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
2532           if (Op1C->getType()->getPrimitiveSizeInBits() >
2533               MulC->getType()->getPrimitiveSizeInBits())
2534             MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
2535 
2536           // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
2537           Multiple = ConstantExpr::getMul(MulC, Op1C);
2538           return true;
2539         }
2540 
2541       if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
2542         if (Mul0CI->getValue() == 1) {
2543           // V == Base * Op1, so return Op1
2544           Multiple = Op1;
2545           return true;
2546         }
2547     }
2548 
2549     Value *Mul1 = nullptr;
2550     if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
2551       if (Constant *Op0C = dyn_cast<Constant>(Op0))
2552         if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
2553           if (Op0C->getType()->getPrimitiveSizeInBits() <
2554               MulC->getType()->getPrimitiveSizeInBits())
2555             Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
2556           if (Op0C->getType()->getPrimitiveSizeInBits() >
2557               MulC->getType()->getPrimitiveSizeInBits())
2558             MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
2559 
2560           // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
2561           Multiple = ConstantExpr::getMul(MulC, Op0C);
2562           return true;
2563         }
2564 
2565       if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
2566         if (Mul1CI->getValue() == 1) {
2567           // V == Base * Op0, so return Op0
2568           Multiple = Op0;
2569           return true;
2570         }
2571     }
2572   }
2573   }
2574 
2575   // We could not determine if V is a multiple of Base.
2576   return false;
2577 }
2578 
getIntrinsicForCallSite(ImmutableCallSite ICS,const TargetLibraryInfo * TLI)2579 Intrinsic::ID llvm::getIntrinsicForCallSite(ImmutableCallSite ICS,
2580                                             const TargetLibraryInfo *TLI) {
2581   const Function *F = ICS.getCalledFunction();
2582   if (!F)
2583     return Intrinsic::not_intrinsic;
2584 
2585   if (F->isIntrinsic())
2586     return F->getIntrinsicID();
2587 
2588   if (!TLI)
2589     return Intrinsic::not_intrinsic;
2590 
2591   LibFunc Func;
2592   // We're going to make assumptions on the semantics of the functions, check
2593   // that the target knows that it's available in this environment and it does
2594   // not have local linkage.
2595   if (!F || F->hasLocalLinkage() || !TLI->getLibFunc(*F, Func))
2596     return Intrinsic::not_intrinsic;
2597 
2598   if (!ICS.onlyReadsMemory())
2599     return Intrinsic::not_intrinsic;
2600 
2601   // Otherwise check if we have a call to a function that can be turned into a
2602   // vector intrinsic.
2603   switch (Func) {
2604   default:
2605     break;
2606   case LibFunc_sin:
2607   case LibFunc_sinf:
2608   case LibFunc_sinl:
2609     return Intrinsic::sin;
2610   case LibFunc_cos:
2611   case LibFunc_cosf:
2612   case LibFunc_cosl:
2613     return Intrinsic::cos;
2614   case LibFunc_exp:
2615   case LibFunc_expf:
2616   case LibFunc_expl:
2617     return Intrinsic::exp;
2618   case LibFunc_exp2:
2619   case LibFunc_exp2f:
2620   case LibFunc_exp2l:
2621     return Intrinsic::exp2;
2622   case LibFunc_log:
2623   case LibFunc_logf:
2624   case LibFunc_logl:
2625     return Intrinsic::log;
2626   case LibFunc_log10:
2627   case LibFunc_log10f:
2628   case LibFunc_log10l:
2629     return Intrinsic::log10;
2630   case LibFunc_log2:
2631   case LibFunc_log2f:
2632   case LibFunc_log2l:
2633     return Intrinsic::log2;
2634   case LibFunc_fabs:
2635   case LibFunc_fabsf:
2636   case LibFunc_fabsl:
2637     return Intrinsic::fabs;
2638   case LibFunc_fmin:
2639   case LibFunc_fminf:
2640   case LibFunc_fminl:
2641     return Intrinsic::minnum;
2642   case LibFunc_fmax:
2643   case LibFunc_fmaxf:
2644   case LibFunc_fmaxl:
2645     return Intrinsic::maxnum;
2646   case LibFunc_copysign:
2647   case LibFunc_copysignf:
2648   case LibFunc_copysignl:
2649     return Intrinsic::copysign;
2650   case LibFunc_floor:
2651   case LibFunc_floorf:
2652   case LibFunc_floorl:
2653     return Intrinsic::floor;
2654   case LibFunc_ceil:
2655   case LibFunc_ceilf:
2656   case LibFunc_ceill:
2657     return Intrinsic::ceil;
2658   case LibFunc_trunc:
2659   case LibFunc_truncf:
2660   case LibFunc_truncl:
2661     return Intrinsic::trunc;
2662   case LibFunc_rint:
2663   case LibFunc_rintf:
2664   case LibFunc_rintl:
2665     return Intrinsic::rint;
2666   case LibFunc_nearbyint:
2667   case LibFunc_nearbyintf:
2668   case LibFunc_nearbyintl:
2669     return Intrinsic::nearbyint;
2670   case LibFunc_round:
2671   case LibFunc_roundf:
2672   case LibFunc_roundl:
2673     return Intrinsic::round;
2674   case LibFunc_pow:
2675   case LibFunc_powf:
2676   case LibFunc_powl:
2677     return Intrinsic::pow;
2678   case LibFunc_sqrt:
2679   case LibFunc_sqrtf:
2680   case LibFunc_sqrtl:
2681     return Intrinsic::sqrt;
2682   }
2683 
2684   return Intrinsic::not_intrinsic;
2685 }
2686 
2687 /// Return true if we can prove that the specified FP value is never equal to
2688 /// -0.0.
2689 ///
2690 /// NOTE: this function will need to be revisited when we support non-default
2691 /// rounding modes!
CannotBeNegativeZero(const Value * V,const TargetLibraryInfo * TLI,unsigned Depth)2692 bool llvm::CannotBeNegativeZero(const Value *V, const TargetLibraryInfo *TLI,
2693                                 unsigned Depth) {
2694   if (auto *CFP = dyn_cast<ConstantFP>(V))
2695     return !CFP->getValueAPF().isNegZero();
2696 
2697   // Limit search depth.
2698   if (Depth == MaxDepth)
2699     return false;
2700 
2701   auto *Op = dyn_cast<Operator>(V);
2702   if (!Op)
2703     return false;
2704 
2705   // Check if the nsz fast-math flag is set.
2706   if (auto *FPO = dyn_cast<FPMathOperator>(Op))
2707     if (FPO->hasNoSignedZeros())
2708       return true;
2709 
2710   // (fadd x, 0.0) is guaranteed to return +0.0, not -0.0.
2711   if (match(Op, m_FAdd(m_Value(), m_PosZeroFP())))
2712     return true;
2713 
2714   // sitofp and uitofp turn into +0.0 for zero.
2715   if (isa<SIToFPInst>(Op) || isa<UIToFPInst>(Op))
2716     return true;
2717 
2718   if (auto *Call = dyn_cast<CallInst>(Op)) {
2719     Intrinsic::ID IID = getIntrinsicForCallSite(Call, TLI);
2720     switch (IID) {
2721     default:
2722       break;
2723     // sqrt(-0.0) = -0.0, no other negative results are possible.
2724     case Intrinsic::sqrt:
2725       return CannotBeNegativeZero(Call->getArgOperand(0), TLI, Depth + 1);
2726     // fabs(x) != -0.0
2727     case Intrinsic::fabs:
2728       return true;
2729     }
2730   }
2731 
2732   return false;
2733 }
2734 
2735 /// If \p SignBitOnly is true, test for a known 0 sign bit rather than a
2736 /// standard ordered compare. e.g. make -0.0 olt 0.0 be true because of the sign
2737 /// bit despite comparing equal.
cannotBeOrderedLessThanZeroImpl(const Value * V,const TargetLibraryInfo * TLI,bool SignBitOnly,unsigned Depth)2738 static bool cannotBeOrderedLessThanZeroImpl(const Value *V,
2739                                             const TargetLibraryInfo *TLI,
2740                                             bool SignBitOnly,
2741                                             unsigned Depth) {
2742   // TODO: This function does not do the right thing when SignBitOnly is true
2743   // and we're lowering to a hypothetical IEEE 754-compliant-but-evil platform
2744   // which flips the sign bits of NaNs.  See
2745   // https://llvm.org/bugs/show_bug.cgi?id=31702.
2746 
2747   if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
2748     return !CFP->getValueAPF().isNegative() ||
2749            (!SignBitOnly && CFP->getValueAPF().isZero());
2750   }
2751 
2752   // Handle vector of constants.
2753   if (auto *CV = dyn_cast<Constant>(V)) {
2754     if (CV->getType()->isVectorTy()) {
2755       unsigned NumElts = CV->getType()->getVectorNumElements();
2756       for (unsigned i = 0; i != NumElts; ++i) {
2757         auto *CFP = dyn_cast_or_null<ConstantFP>(CV->getAggregateElement(i));
2758         if (!CFP)
2759           return false;
2760         if (CFP->getValueAPF().isNegative() &&
2761             (SignBitOnly || !CFP->getValueAPF().isZero()))
2762           return false;
2763       }
2764 
2765       // All non-negative ConstantFPs.
2766       return true;
2767     }
2768   }
2769 
2770   if (Depth == MaxDepth)
2771     return false; // Limit search depth.
2772 
2773   const Operator *I = dyn_cast<Operator>(V);
2774   if (!I)
2775     return false;
2776 
2777   switch (I->getOpcode()) {
2778   default:
2779     break;
2780   // Unsigned integers are always nonnegative.
2781   case Instruction::UIToFP:
2782     return true;
2783   case Instruction::FMul:
2784     // x*x is always non-negative or a NaN.
2785     if (I->getOperand(0) == I->getOperand(1) &&
2786         (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()))
2787       return true;
2788 
2789     LLVM_FALLTHROUGH;
2790   case Instruction::FAdd:
2791   case Instruction::FDiv:
2792   case Instruction::FRem:
2793     return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2794                                            Depth + 1) &&
2795            cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
2796                                            Depth + 1);
2797   case Instruction::Select:
2798     return cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
2799                                            Depth + 1) &&
2800            cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly,
2801                                            Depth + 1);
2802   case Instruction::FPExt:
2803   case Instruction::FPTrunc:
2804     // Widening/narrowing never change sign.
2805     return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2806                                            Depth + 1);
2807   case Instruction::ExtractElement:
2808     // Look through extract element. At the moment we keep this simple and skip
2809     // tracking the specific element. But at least we might find information
2810     // valid for all elements of the vector.
2811     return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2812                                            Depth + 1);
2813   case Instruction::Call:
2814     const auto *CI = cast<CallInst>(I);
2815     Intrinsic::ID IID = getIntrinsicForCallSite(CI, TLI);
2816     switch (IID) {
2817     default:
2818       break;
2819     case Intrinsic::maxnum:
2820       return (isKnownNeverNaN(I->getOperand(0)) &&
2821               cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI,
2822                                               SignBitOnly, Depth + 1)) ||
2823              (isKnownNeverNaN(I->getOperand(1)) &&
2824               cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI,
2825                                               SignBitOnly, Depth + 1));
2826 
2827     case Intrinsic::minnum:
2828       return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2829                                              Depth + 1) &&
2830              cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
2831                                              Depth + 1);
2832     case Intrinsic::exp:
2833     case Intrinsic::exp2:
2834     case Intrinsic::fabs:
2835       return true;
2836 
2837     case Intrinsic::sqrt:
2838       // sqrt(x) is always >= -0 or NaN.  Moreover, sqrt(x) == -0 iff x == -0.
2839       if (!SignBitOnly)
2840         return true;
2841       return CI->hasNoNaNs() && (CI->hasNoSignedZeros() ||
2842                                  CannotBeNegativeZero(CI->getOperand(0), TLI));
2843 
2844     case Intrinsic::powi:
2845       if (ConstantInt *Exponent = dyn_cast<ConstantInt>(I->getOperand(1))) {
2846         // powi(x,n) is non-negative if n is even.
2847         if (Exponent->getBitWidth() <= 64 && Exponent->getSExtValue() % 2u == 0)
2848           return true;
2849       }
2850       // TODO: This is not correct.  Given that exp is an integer, here are the
2851       // ways that pow can return a negative value:
2852       //
2853       //   pow(x, exp)    --> negative if exp is odd and x is negative.
2854       //   pow(-0, exp)   --> -inf if exp is negative odd.
2855       //   pow(-0, exp)   --> -0 if exp is positive odd.
2856       //   pow(-inf, exp) --> -0 if exp is negative odd.
2857       //   pow(-inf, exp) --> -inf if exp is positive odd.
2858       //
2859       // Therefore, if !SignBitOnly, we can return true if x >= +0 or x is NaN,
2860       // but we must return false if x == -0.  Unfortunately we do not currently
2861       // have a way of expressing this constraint.  See details in
2862       // https://llvm.org/bugs/show_bug.cgi?id=31702.
2863       return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
2864                                              Depth + 1);
2865 
2866     case Intrinsic::fma:
2867     case Intrinsic::fmuladd:
2868       // x*x+y is non-negative if y is non-negative.
2869       return I->getOperand(0) == I->getOperand(1) &&
2870              (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()) &&
2871              cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly,
2872                                              Depth + 1);
2873     }
2874     break;
2875   }
2876   return false;
2877 }
2878 
CannotBeOrderedLessThanZero(const Value * V,const TargetLibraryInfo * TLI)2879 bool llvm::CannotBeOrderedLessThanZero(const Value *V,
2880                                        const TargetLibraryInfo *TLI) {
2881   return cannotBeOrderedLessThanZeroImpl(V, TLI, false, 0);
2882 }
2883 
SignBitMustBeZero(const Value * V,const TargetLibraryInfo * TLI)2884 bool llvm::SignBitMustBeZero(const Value *V, const TargetLibraryInfo *TLI) {
2885   return cannotBeOrderedLessThanZeroImpl(V, TLI, true, 0);
2886 }
2887 
isKnownNeverNaN(const Value * V)2888 bool llvm::isKnownNeverNaN(const Value *V) {
2889   assert(V->getType()->isFPOrFPVectorTy() && "Querying for NaN on non-FP type");
2890 
2891   // If we're told that NaNs won't happen, assume they won't.
2892   if (auto *FPMathOp = dyn_cast<FPMathOperator>(V))
2893     if (FPMathOp->hasNoNaNs())
2894       return true;
2895 
2896   // TODO: Handle instructions and potentially recurse like other 'isKnown'
2897   // functions. For example, the result of sitofp is never NaN.
2898 
2899   // Handle scalar constants.
2900   if (auto *CFP = dyn_cast<ConstantFP>(V))
2901     return !CFP->isNaN();
2902 
2903   // Bail out for constant expressions, but try to handle vector constants.
2904   if (!V->getType()->isVectorTy() || !isa<Constant>(V))
2905     return false;
2906 
2907   // For vectors, verify that each element is not NaN.
2908   unsigned NumElts = V->getType()->getVectorNumElements();
2909   for (unsigned i = 0; i != NumElts; ++i) {
2910     Constant *Elt = cast<Constant>(V)->getAggregateElement(i);
2911     if (!Elt)
2912       return false;
2913     if (isa<UndefValue>(Elt))
2914       continue;
2915     auto *CElt = dyn_cast<ConstantFP>(Elt);
2916     if (!CElt || CElt->isNaN())
2917       return false;
2918   }
2919   // All elements were confirmed not-NaN or undefined.
2920   return true;
2921 }
2922 
2923 /// If the specified value can be set by repeating the same byte in memory,
2924 /// return the i8 value that it is represented with.  This is
2925 /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
2926 /// i16 0xF0F0, double 0.0 etc.  If the value can't be handled with a repeated
2927 /// byte store (e.g. i16 0x1234), return null.
isBytewiseValue(Value * V)2928 Value *llvm::isBytewiseValue(Value *V) {
2929   // All byte-wide stores are splatable, even of arbitrary variables.
2930   if (V->getType()->isIntegerTy(8)) return V;
2931 
2932   // Handle 'null' ConstantArrayZero etc.
2933   if (Constant *C = dyn_cast<Constant>(V))
2934     if (C->isNullValue())
2935       return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
2936 
2937   // Constant float and double values can be handled as integer values if the
2938   // corresponding integer value is "byteable".  An important case is 0.0.
2939   if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
2940     if (CFP->getType()->isFloatTy())
2941       V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
2942     if (CFP->getType()->isDoubleTy())
2943       V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
2944     // Don't handle long double formats, which have strange constraints.
2945   }
2946 
2947   // We can handle constant integers that are multiple of 8 bits.
2948   if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
2949     if (CI->getBitWidth() % 8 == 0) {
2950       assert(CI->getBitWidth() > 8 && "8 bits should be handled above!");
2951 
2952       if (!CI->getValue().isSplat(8))
2953         return nullptr;
2954       return ConstantInt::get(V->getContext(), CI->getValue().trunc(8));
2955     }
2956   }
2957 
2958   // A ConstantDataArray/Vector is splatable if all its members are equal and
2959   // also splatable.
2960   if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
2961     Value *Elt = CA->getElementAsConstant(0);
2962     Value *Val = isBytewiseValue(Elt);
2963     if (!Val)
2964       return nullptr;
2965 
2966     for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
2967       if (CA->getElementAsConstant(I) != Elt)
2968         return nullptr;
2969 
2970     return Val;
2971   }
2972 
2973   // Conceptually, we could handle things like:
2974   //   %a = zext i8 %X to i16
2975   //   %b = shl i16 %a, 8
2976   //   %c = or i16 %a, %b
2977   // but until there is an example that actually needs this, it doesn't seem
2978   // worth worrying about.
2979   return nullptr;
2980 }
2981 
2982 // This is the recursive version of BuildSubAggregate. It takes a few different
2983 // arguments. Idxs is the index within the nested struct From that we are
2984 // looking at now (which is of type IndexedType). IdxSkip is the number of
2985 // indices from Idxs that should be left out when inserting into the resulting
2986 // struct. To is the result struct built so far, new insertvalue instructions
2987 // build on that.
BuildSubAggregate(Value * From,Value * To,Type * IndexedType,SmallVectorImpl<unsigned> & Idxs,unsigned IdxSkip,Instruction * InsertBefore)2988 static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
2989                                 SmallVectorImpl<unsigned> &Idxs,
2990                                 unsigned IdxSkip,
2991                                 Instruction *InsertBefore) {
2992   StructType *STy = dyn_cast<StructType>(IndexedType);
2993   if (STy) {
2994     // Save the original To argument so we can modify it
2995     Value *OrigTo = To;
2996     // General case, the type indexed by Idxs is a struct
2997     for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2998       // Process each struct element recursively
2999       Idxs.push_back(i);
3000       Value *PrevTo = To;
3001       To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
3002                              InsertBefore);
3003       Idxs.pop_back();
3004       if (!To) {
3005         // Couldn't find any inserted value for this index? Cleanup
3006         while (PrevTo != OrigTo) {
3007           InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
3008           PrevTo = Del->getAggregateOperand();
3009           Del->eraseFromParent();
3010         }
3011         // Stop processing elements
3012         break;
3013       }
3014     }
3015     // If we successfully found a value for each of our subaggregates
3016     if (To)
3017       return To;
3018   }
3019   // Base case, the type indexed by SourceIdxs is not a struct, or not all of
3020   // the struct's elements had a value that was inserted directly. In the latter
3021   // case, perhaps we can't determine each of the subelements individually, but
3022   // we might be able to find the complete struct somewhere.
3023 
3024   // Find the value that is at that particular spot
3025   Value *V = FindInsertedValue(From, Idxs);
3026 
3027   if (!V)
3028     return nullptr;
3029 
3030   // Insert the value in the new (sub) aggregate
3031   return InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
3032                                  "tmp", InsertBefore);
3033 }
3034 
3035 // This helper takes a nested struct and extracts a part of it (which is again a
3036 // struct) into a new value. For example, given the struct:
3037 // { a, { b, { c, d }, e } }
3038 // and the indices "1, 1" this returns
3039 // { c, d }.
3040 //
3041 // It does this by inserting an insertvalue for each element in the resulting
3042 // struct, as opposed to just inserting a single struct. This will only work if
3043 // each of the elements of the substruct are known (ie, inserted into From by an
3044 // insertvalue instruction somewhere).
3045 //
3046 // All inserted insertvalue instructions are inserted before InsertBefore
BuildSubAggregate(Value * From,ArrayRef<unsigned> idx_range,Instruction * InsertBefore)3047 static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
3048                                 Instruction *InsertBefore) {
3049   assert(InsertBefore && "Must have someplace to insert!");
3050   Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
3051                                                              idx_range);
3052   Value *To = UndefValue::get(IndexedType);
3053   SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
3054   unsigned IdxSkip = Idxs.size();
3055 
3056   return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
3057 }
3058 
3059 /// Given an aggregate and a sequence of indices, see if the scalar value
3060 /// indexed is already around as a register, for example if it was inserted
3061 /// directly into the aggregate.
3062 ///
3063 /// If InsertBefore is not null, this function will duplicate (modified)
3064 /// insertvalues when a part of a nested struct is extracted.
FindInsertedValue(Value * V,ArrayRef<unsigned> idx_range,Instruction * InsertBefore)3065 Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
3066                                Instruction *InsertBefore) {
3067   // Nothing to index? Just return V then (this is useful at the end of our
3068   // recursion).
3069   if (idx_range.empty())
3070     return V;
3071   // We have indices, so V should have an indexable type.
3072   assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
3073          "Not looking at a struct or array?");
3074   assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
3075          "Invalid indices for type?");
3076 
3077   if (Constant *C = dyn_cast<Constant>(V)) {
3078     C = C->getAggregateElement(idx_range[0]);
3079     if (!C) return nullptr;
3080     return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
3081   }
3082 
3083   if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
3084     // Loop the indices for the insertvalue instruction in parallel with the
3085     // requested indices
3086     const unsigned *req_idx = idx_range.begin();
3087     for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
3088          i != e; ++i, ++req_idx) {
3089       if (req_idx == idx_range.end()) {
3090         // We can't handle this without inserting insertvalues
3091         if (!InsertBefore)
3092           return nullptr;
3093 
3094         // The requested index identifies a part of a nested aggregate. Handle
3095         // this specially. For example,
3096         // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
3097         // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
3098         // %C = extractvalue {i32, { i32, i32 } } %B, 1
3099         // This can be changed into
3100         // %A = insertvalue {i32, i32 } undef, i32 10, 0
3101         // %C = insertvalue {i32, i32 } %A, i32 11, 1
3102         // which allows the unused 0,0 element from the nested struct to be
3103         // removed.
3104         return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
3105                                  InsertBefore);
3106       }
3107 
3108       // This insert value inserts something else than what we are looking for.
3109       // See if the (aggregate) value inserted into has the value we are
3110       // looking for, then.
3111       if (*req_idx != *i)
3112         return FindInsertedValue(I->getAggregateOperand(), idx_range,
3113                                  InsertBefore);
3114     }
3115     // If we end up here, the indices of the insertvalue match with those
3116     // requested (though possibly only partially). Now we recursively look at
3117     // the inserted value, passing any remaining indices.
3118     return FindInsertedValue(I->getInsertedValueOperand(),
3119                              makeArrayRef(req_idx, idx_range.end()),
3120                              InsertBefore);
3121   }
3122 
3123   if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
3124     // If we're extracting a value from an aggregate that was extracted from
3125     // something else, we can extract from that something else directly instead.
3126     // However, we will need to chain I's indices with the requested indices.
3127 
3128     // Calculate the number of indices required
3129     unsigned size = I->getNumIndices() + idx_range.size();
3130     // Allocate some space to put the new indices in
3131     SmallVector<unsigned, 5> Idxs;
3132     Idxs.reserve(size);
3133     // Add indices from the extract value instruction
3134     Idxs.append(I->idx_begin(), I->idx_end());
3135 
3136     // Add requested indices
3137     Idxs.append(idx_range.begin(), idx_range.end());
3138 
3139     assert(Idxs.size() == size
3140            && "Number of indices added not correct?");
3141 
3142     return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
3143   }
3144   // Otherwise, we don't know (such as, extracting from a function return value
3145   // or load instruction)
3146   return nullptr;
3147 }
3148 
3149 /// Analyze the specified pointer to see if it can be expressed as a base
3150 /// pointer plus a constant offset. Return the base and offset to the caller.
GetPointerBaseWithConstantOffset(Value * Ptr,int64_t & Offset,const DataLayout & DL)3151 Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
3152                                               const DataLayout &DL) {
3153   unsigned BitWidth = DL.getIndexTypeSizeInBits(Ptr->getType());
3154   APInt ByteOffset(BitWidth, 0);
3155 
3156   // We walk up the defs but use a visited set to handle unreachable code. In
3157   // that case, we stop after accumulating the cycle once (not that it
3158   // matters).
3159   SmallPtrSet<Value *, 16> Visited;
3160   while (Visited.insert(Ptr).second) {
3161     if (Ptr->getType()->isVectorTy())
3162       break;
3163 
3164     if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
3165       // If one of the values we have visited is an addrspacecast, then
3166       // the pointer type of this GEP may be different from the type
3167       // of the Ptr parameter which was passed to this function.  This
3168       // means when we construct GEPOffset, we need to use the size
3169       // of GEP's pointer type rather than the size of the original
3170       // pointer type.
3171       APInt GEPOffset(DL.getIndexTypeSizeInBits(Ptr->getType()), 0);
3172       if (!GEP->accumulateConstantOffset(DL, GEPOffset))
3173         break;
3174 
3175       ByteOffset += GEPOffset.getSExtValue();
3176 
3177       Ptr = GEP->getPointerOperand();
3178     } else if (Operator::getOpcode(Ptr) == Instruction::BitCast ||
3179                Operator::getOpcode(Ptr) == Instruction::AddrSpaceCast) {
3180       Ptr = cast<Operator>(Ptr)->getOperand(0);
3181     } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
3182       if (GA->isInterposable())
3183         break;
3184       Ptr = GA->getAliasee();
3185     } else {
3186       break;
3187     }
3188   }
3189   Offset = ByteOffset.getSExtValue();
3190   return Ptr;
3191 }
3192 
isGEPBasedOnPointerToString(const GEPOperator * GEP,unsigned CharSize)3193 bool llvm::isGEPBasedOnPointerToString(const GEPOperator *GEP,
3194                                        unsigned CharSize) {
3195   // Make sure the GEP has exactly three arguments.
3196   if (GEP->getNumOperands() != 3)
3197     return false;
3198 
3199   // Make sure the index-ee is a pointer to array of \p CharSize integers.
3200   // CharSize.
3201   ArrayType *AT = dyn_cast<ArrayType>(GEP->getSourceElementType());
3202   if (!AT || !AT->getElementType()->isIntegerTy(CharSize))
3203     return false;
3204 
3205   // Check to make sure that the first operand of the GEP is an integer and
3206   // has value 0 so that we are sure we're indexing into the initializer.
3207   const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
3208   if (!FirstIdx || !FirstIdx->isZero())
3209     return false;
3210 
3211   return true;
3212 }
3213 
getConstantDataArrayInfo(const Value * V,ConstantDataArraySlice & Slice,unsigned ElementSize,uint64_t Offset)3214 bool llvm::getConstantDataArrayInfo(const Value *V,
3215                                     ConstantDataArraySlice &Slice,
3216                                     unsigned ElementSize, uint64_t Offset) {
3217   assert(V);
3218 
3219   // Look through bitcast instructions and geps.
3220   V = V->stripPointerCasts();
3221 
3222   // If the value is a GEP instruction or constant expression, treat it as an
3223   // offset.
3224   if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
3225     // The GEP operator should be based on a pointer to string constant, and is
3226     // indexing into the string constant.
3227     if (!isGEPBasedOnPointerToString(GEP, ElementSize))
3228       return false;
3229 
3230     // If the second index isn't a ConstantInt, then this is a variable index
3231     // into the array.  If this occurs, we can't say anything meaningful about
3232     // the string.
3233     uint64_t StartIdx = 0;
3234     if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
3235       StartIdx = CI->getZExtValue();
3236     else
3237       return false;
3238     return getConstantDataArrayInfo(GEP->getOperand(0), Slice, ElementSize,
3239                                     StartIdx + Offset);
3240   }
3241 
3242   // The GEP instruction, constant or instruction, must reference a global
3243   // variable that is a constant and is initialized. The referenced constant
3244   // initializer is the array that we'll use for optimization.
3245   const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
3246   if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
3247     return false;
3248 
3249   const ConstantDataArray *Array;
3250   ArrayType *ArrayTy;
3251   if (GV->getInitializer()->isNullValue()) {
3252     Type *GVTy = GV->getValueType();
3253     if ( (ArrayTy = dyn_cast<ArrayType>(GVTy)) ) {
3254       // A zeroinitializer for the array; there is no ConstantDataArray.
3255       Array = nullptr;
3256     } else {
3257       const DataLayout &DL = GV->getParent()->getDataLayout();
3258       uint64_t SizeInBytes = DL.getTypeStoreSize(GVTy);
3259       uint64_t Length = SizeInBytes / (ElementSize / 8);
3260       if (Length <= Offset)
3261         return false;
3262 
3263       Slice.Array = nullptr;
3264       Slice.Offset = 0;
3265       Slice.Length = Length - Offset;
3266       return true;
3267     }
3268   } else {
3269     // This must be a ConstantDataArray.
3270     Array = dyn_cast<ConstantDataArray>(GV->getInitializer());
3271     if (!Array)
3272       return false;
3273     ArrayTy = Array->getType();
3274   }
3275   if (!ArrayTy->getElementType()->isIntegerTy(ElementSize))
3276     return false;
3277 
3278   uint64_t NumElts = ArrayTy->getArrayNumElements();
3279   if (Offset > NumElts)
3280     return false;
3281 
3282   Slice.Array = Array;
3283   Slice.Offset = Offset;
3284   Slice.Length = NumElts - Offset;
3285   return true;
3286 }
3287 
3288 /// This function computes the length of a null-terminated C string pointed to
3289 /// by V. If successful, it returns true and returns the string in Str.
3290 /// If unsuccessful, it returns false.
getConstantStringInfo(const Value * V,StringRef & Str,uint64_t Offset,bool TrimAtNul)3291 bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
3292                                  uint64_t Offset, bool TrimAtNul) {
3293   ConstantDataArraySlice Slice;
3294   if (!getConstantDataArrayInfo(V, Slice, 8, Offset))
3295     return false;
3296 
3297   if (Slice.Array == nullptr) {
3298     if (TrimAtNul) {
3299       Str = StringRef();
3300       return true;
3301     }
3302     if (Slice.Length == 1) {
3303       Str = StringRef("", 1);
3304       return true;
3305     }
3306     // We cannot instantiate a StringRef as we do not have an appropriate string
3307     // of 0s at hand.
3308     return false;
3309   }
3310 
3311   // Start out with the entire array in the StringRef.
3312   Str = Slice.Array->getAsString();
3313   // Skip over 'offset' bytes.
3314   Str = Str.substr(Slice.Offset);
3315 
3316   if (TrimAtNul) {
3317     // Trim off the \0 and anything after it.  If the array is not nul
3318     // terminated, we just return the whole end of string.  The client may know
3319     // some other way that the string is length-bound.
3320     Str = Str.substr(0, Str.find('\0'));
3321   }
3322   return true;
3323 }
3324 
3325 // These next two are very similar to the above, but also look through PHI
3326 // nodes.
3327 // TODO: See if we can integrate these two together.
3328 
3329 /// If we can compute the length of the string pointed to by
3330 /// the specified pointer, return 'len+1'.  If we can't, return 0.
GetStringLengthH(const Value * V,SmallPtrSetImpl<const PHINode * > & PHIs,unsigned CharSize)3331 static uint64_t GetStringLengthH(const Value *V,
3332                                  SmallPtrSetImpl<const PHINode*> &PHIs,
3333                                  unsigned CharSize) {
3334   // Look through noop bitcast instructions.
3335   V = V->stripPointerCasts();
3336 
3337   // If this is a PHI node, there are two cases: either we have already seen it
3338   // or we haven't.
3339   if (const PHINode *PN = dyn_cast<PHINode>(V)) {
3340     if (!PHIs.insert(PN).second)
3341       return ~0ULL;  // already in the set.
3342 
3343     // If it was new, see if all the input strings are the same length.
3344     uint64_t LenSoFar = ~0ULL;
3345     for (Value *IncValue : PN->incoming_values()) {
3346       uint64_t Len = GetStringLengthH(IncValue, PHIs, CharSize);
3347       if (Len == 0) return 0; // Unknown length -> unknown.
3348 
3349       if (Len == ~0ULL) continue;
3350 
3351       if (Len != LenSoFar && LenSoFar != ~0ULL)
3352         return 0;    // Disagree -> unknown.
3353       LenSoFar = Len;
3354     }
3355 
3356     // Success, all agree.
3357     return LenSoFar;
3358   }
3359 
3360   // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
3361   if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
3362     uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs, CharSize);
3363     if (Len1 == 0) return 0;
3364     uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs, CharSize);
3365     if (Len2 == 0) return 0;
3366     if (Len1 == ~0ULL) return Len2;
3367     if (Len2 == ~0ULL) return Len1;
3368     if (Len1 != Len2) return 0;
3369     return Len1;
3370   }
3371 
3372   // Otherwise, see if we can read the string.
3373   ConstantDataArraySlice Slice;
3374   if (!getConstantDataArrayInfo(V, Slice, CharSize))
3375     return 0;
3376 
3377   if (Slice.Array == nullptr)
3378     return 1;
3379 
3380   // Search for nul characters
3381   unsigned NullIndex = 0;
3382   for (unsigned E = Slice.Length; NullIndex < E; ++NullIndex) {
3383     if (Slice.Array->getElementAsInteger(Slice.Offset + NullIndex) == 0)
3384       break;
3385   }
3386 
3387   return NullIndex + 1;
3388 }
3389 
3390 /// If we can compute the length of the string pointed to by
3391 /// the specified pointer, return 'len+1'.  If we can't, return 0.
GetStringLength(const Value * V,unsigned CharSize)3392 uint64_t llvm::GetStringLength(const Value *V, unsigned CharSize) {
3393   if (!V->getType()->isPointerTy())
3394     return 0;
3395 
3396   SmallPtrSet<const PHINode*, 32> PHIs;
3397   uint64_t Len = GetStringLengthH(V, PHIs, CharSize);
3398   // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
3399   // an empty string as a length.
3400   return Len == ~0ULL ? 1 : Len;
3401 }
3402 
getArgumentAliasingToReturnedPointer(ImmutableCallSite CS)3403 const Value *llvm::getArgumentAliasingToReturnedPointer(ImmutableCallSite CS) {
3404   assert(CS &&
3405          "getArgumentAliasingToReturnedPointer only works on nonnull CallSite");
3406   if (const Value *RV = CS.getReturnedArgOperand())
3407     return RV;
3408   // This can be used only as a aliasing property.
3409   if (isIntrinsicReturningPointerAliasingArgumentWithoutCapturing(CS))
3410     return CS.getArgOperand(0);
3411   return nullptr;
3412 }
3413 
isIntrinsicReturningPointerAliasingArgumentWithoutCapturing(ImmutableCallSite CS)3414 bool llvm::isIntrinsicReturningPointerAliasingArgumentWithoutCapturing(
3415     ImmutableCallSite CS) {
3416   return CS.getIntrinsicID() == Intrinsic::launder_invariant_group ||
3417          CS.getIntrinsicID() == Intrinsic::strip_invariant_group;
3418 }
3419 
3420 /// \p PN defines a loop-variant pointer to an object.  Check if the
3421 /// previous iteration of the loop was referring to the same object as \p PN.
isSameUnderlyingObjectInLoop(const PHINode * PN,const LoopInfo * LI)3422 static bool isSameUnderlyingObjectInLoop(const PHINode *PN,
3423                                          const LoopInfo *LI) {
3424   // Find the loop-defined value.
3425   Loop *L = LI->getLoopFor(PN->getParent());
3426   if (PN->getNumIncomingValues() != 2)
3427     return true;
3428 
3429   // Find the value from previous iteration.
3430   auto *PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(0));
3431   if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
3432     PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(1));
3433   if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
3434     return true;
3435 
3436   // If a new pointer is loaded in the loop, the pointer references a different
3437   // object in every iteration.  E.g.:
3438   //    for (i)
3439   //       int *p = a[i];
3440   //       ...
3441   if (auto *Load = dyn_cast<LoadInst>(PrevValue))
3442     if (!L->isLoopInvariant(Load->getPointerOperand()))
3443       return false;
3444   return true;
3445 }
3446 
GetUnderlyingObject(Value * V,const DataLayout & DL,unsigned MaxLookup)3447 Value *llvm::GetUnderlyingObject(Value *V, const DataLayout &DL,
3448                                  unsigned MaxLookup) {
3449   if (!V->getType()->isPointerTy())
3450     return V;
3451   for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
3452     if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
3453       V = GEP->getPointerOperand();
3454     } else if (Operator::getOpcode(V) == Instruction::BitCast ||
3455                Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
3456       V = cast<Operator>(V)->getOperand(0);
3457     } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
3458       if (GA->isInterposable())
3459         return V;
3460       V = GA->getAliasee();
3461     } else if (isa<AllocaInst>(V)) {
3462       // An alloca can't be further simplified.
3463       return V;
3464     } else {
3465       if (auto CS = CallSite(V)) {
3466         // CaptureTracking can know about special capturing properties of some
3467         // intrinsics like launder.invariant.group, that can't be expressed with
3468         // the attributes, but have properties like returning aliasing pointer.
3469         // Because some analysis may assume that nocaptured pointer is not
3470         // returned from some special intrinsic (because function would have to
3471         // be marked with returns attribute), it is crucial to use this function
3472         // because it should be in sync with CaptureTracking. Not using it may
3473         // cause weird miscompilations where 2 aliasing pointers are assumed to
3474         // noalias.
3475         if (auto *RP = getArgumentAliasingToReturnedPointer(CS)) {
3476           V = RP;
3477           continue;
3478         }
3479       }
3480 
3481       // See if InstructionSimplify knows any relevant tricks.
3482       if (Instruction *I = dyn_cast<Instruction>(V))
3483         // TODO: Acquire a DominatorTree and AssumptionCache and use them.
3484         if (Value *Simplified = SimplifyInstruction(I, {DL, I})) {
3485           V = Simplified;
3486           continue;
3487         }
3488 
3489       return V;
3490     }
3491     assert(V->getType()->isPointerTy() && "Unexpected operand type!");
3492   }
3493   return V;
3494 }
3495 
GetUnderlyingObjects(Value * V,SmallVectorImpl<Value * > & Objects,const DataLayout & DL,LoopInfo * LI,unsigned MaxLookup)3496 void llvm::GetUnderlyingObjects(Value *V, SmallVectorImpl<Value *> &Objects,
3497                                 const DataLayout &DL, LoopInfo *LI,
3498                                 unsigned MaxLookup) {
3499   SmallPtrSet<Value *, 4> Visited;
3500   SmallVector<Value *, 4> Worklist;
3501   Worklist.push_back(V);
3502   do {
3503     Value *P = Worklist.pop_back_val();
3504     P = GetUnderlyingObject(P, DL, MaxLookup);
3505 
3506     if (!Visited.insert(P).second)
3507       continue;
3508 
3509     if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
3510       Worklist.push_back(SI->getTrueValue());
3511       Worklist.push_back(SI->getFalseValue());
3512       continue;
3513     }
3514 
3515     if (PHINode *PN = dyn_cast<PHINode>(P)) {
3516       // If this PHI changes the underlying object in every iteration of the
3517       // loop, don't look through it.  Consider:
3518       //   int **A;
3519       //   for (i) {
3520       //     Prev = Curr;     // Prev = PHI (Prev_0, Curr)
3521       //     Curr = A[i];
3522       //     *Prev, *Curr;
3523       //
3524       // Prev is tracking Curr one iteration behind so they refer to different
3525       // underlying objects.
3526       if (!LI || !LI->isLoopHeader(PN->getParent()) ||
3527           isSameUnderlyingObjectInLoop(PN, LI))
3528         for (Value *IncValue : PN->incoming_values())
3529           Worklist.push_back(IncValue);
3530       continue;
3531     }
3532 
3533     Objects.push_back(P);
3534   } while (!Worklist.empty());
3535 }
3536 
3537 /// This is the function that does the work of looking through basic
3538 /// ptrtoint+arithmetic+inttoptr sequences.
getUnderlyingObjectFromInt(const Value * V)3539 static const Value *getUnderlyingObjectFromInt(const Value *V) {
3540   do {
3541     if (const Operator *U = dyn_cast<Operator>(V)) {
3542       // If we find a ptrtoint, we can transfer control back to the
3543       // regular getUnderlyingObjectFromInt.
3544       if (U->getOpcode() == Instruction::PtrToInt)
3545         return U->getOperand(0);
3546       // If we find an add of a constant, a multiplied value, or a phi, it's
3547       // likely that the other operand will lead us to the base
3548       // object. We don't have to worry about the case where the
3549       // object address is somehow being computed by the multiply,
3550       // because our callers only care when the result is an
3551       // identifiable object.
3552       if (U->getOpcode() != Instruction::Add ||
3553           (!isa<ConstantInt>(U->getOperand(1)) &&
3554            Operator::getOpcode(U->getOperand(1)) != Instruction::Mul &&
3555            !isa<PHINode>(U->getOperand(1))))
3556         return V;
3557       V = U->getOperand(0);
3558     } else {
3559       return V;
3560     }
3561     assert(V->getType()->isIntegerTy() && "Unexpected operand type!");
3562   } while (true);
3563 }
3564 
3565 /// This is a wrapper around GetUnderlyingObjects and adds support for basic
3566 /// ptrtoint+arithmetic+inttoptr sequences.
3567 /// It returns false if unidentified object is found in GetUnderlyingObjects.
getUnderlyingObjectsForCodeGen(const Value * V,SmallVectorImpl<Value * > & Objects,const DataLayout & DL)3568 bool llvm::getUnderlyingObjectsForCodeGen(const Value *V,
3569                           SmallVectorImpl<Value *> &Objects,
3570                           const DataLayout &DL) {
3571   SmallPtrSet<const Value *, 16> Visited;
3572   SmallVector<const Value *, 4> Working(1, V);
3573   do {
3574     V = Working.pop_back_val();
3575 
3576     SmallVector<Value *, 4> Objs;
3577     GetUnderlyingObjects(const_cast<Value *>(V), Objs, DL);
3578 
3579     for (Value *V : Objs) {
3580       if (!Visited.insert(V).second)
3581         continue;
3582       if (Operator::getOpcode(V) == Instruction::IntToPtr) {
3583         const Value *O =
3584           getUnderlyingObjectFromInt(cast<User>(V)->getOperand(0));
3585         if (O->getType()->isPointerTy()) {
3586           Working.push_back(O);
3587           continue;
3588         }
3589       }
3590       // If GetUnderlyingObjects fails to find an identifiable object,
3591       // getUnderlyingObjectsForCodeGen also fails for safety.
3592       if (!isIdentifiedObject(V)) {
3593         Objects.clear();
3594         return false;
3595       }
3596       Objects.push_back(const_cast<Value *>(V));
3597     }
3598   } while (!Working.empty());
3599   return true;
3600 }
3601 
3602 /// Return true if the only users of this pointer are lifetime markers.
onlyUsedByLifetimeMarkers(const Value * V)3603 bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
3604   for (const User *U : V->users()) {
3605     const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
3606     if (!II) return false;
3607 
3608     if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
3609         II->getIntrinsicID() != Intrinsic::lifetime_end)
3610       return false;
3611   }
3612   return true;
3613 }
3614 
isSafeToSpeculativelyExecute(const Value * V,const Instruction * CtxI,const DominatorTree * DT)3615 bool llvm::isSafeToSpeculativelyExecute(const Value *V,
3616                                         const Instruction *CtxI,
3617                                         const DominatorTree *DT) {
3618   const Operator *Inst = dyn_cast<Operator>(V);
3619   if (!Inst)
3620     return false;
3621 
3622   for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
3623     if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
3624       if (C->canTrap())
3625         return false;
3626 
3627   switch (Inst->getOpcode()) {
3628   default:
3629     return true;
3630   case Instruction::UDiv:
3631   case Instruction::URem: {
3632     // x / y is undefined if y == 0.
3633     const APInt *V;
3634     if (match(Inst->getOperand(1), m_APInt(V)))
3635       return *V != 0;
3636     return false;
3637   }
3638   case Instruction::SDiv:
3639   case Instruction::SRem: {
3640     // x / y is undefined if y == 0 or x == INT_MIN and y == -1
3641     const APInt *Numerator, *Denominator;
3642     if (!match(Inst->getOperand(1), m_APInt(Denominator)))
3643       return false;
3644     // We cannot hoist this division if the denominator is 0.
3645     if (*Denominator == 0)
3646       return false;
3647     // It's safe to hoist if the denominator is not 0 or -1.
3648     if (*Denominator != -1)
3649       return true;
3650     // At this point we know that the denominator is -1.  It is safe to hoist as
3651     // long we know that the numerator is not INT_MIN.
3652     if (match(Inst->getOperand(0), m_APInt(Numerator)))
3653       return !Numerator->isMinSignedValue();
3654     // The numerator *might* be MinSignedValue.
3655     return false;
3656   }
3657   case Instruction::Load: {
3658     const LoadInst *LI = cast<LoadInst>(Inst);
3659     if (!LI->isUnordered() ||
3660         // Speculative load may create a race that did not exist in the source.
3661         LI->getFunction()->hasFnAttribute(Attribute::SanitizeThread) ||
3662         // Speculative load may load data from dirty regions.
3663         LI->getFunction()->hasFnAttribute(Attribute::SanitizeAddress) ||
3664         LI->getFunction()->hasFnAttribute(Attribute::SanitizeHWAddress))
3665       return false;
3666     const DataLayout &DL = LI->getModule()->getDataLayout();
3667     return isDereferenceableAndAlignedPointer(LI->getPointerOperand(),
3668                                               LI->getAlignment(), DL, CtxI, DT);
3669   }
3670   case Instruction::Call: {
3671     auto *CI = cast<const CallInst>(Inst);
3672     const Function *Callee = CI->getCalledFunction();
3673 
3674     // The called function could have undefined behavior or side-effects, even
3675     // if marked readnone nounwind.
3676     return Callee && Callee->isSpeculatable();
3677   }
3678   case Instruction::VAArg:
3679   case Instruction::Alloca:
3680   case Instruction::Invoke:
3681   case Instruction::PHI:
3682   case Instruction::Store:
3683   case Instruction::Ret:
3684   case Instruction::Br:
3685   case Instruction::IndirectBr:
3686   case Instruction::Switch:
3687   case Instruction::Unreachable:
3688   case Instruction::Fence:
3689   case Instruction::AtomicRMW:
3690   case Instruction::AtomicCmpXchg:
3691   case Instruction::LandingPad:
3692   case Instruction::Resume:
3693   case Instruction::CatchSwitch:
3694   case Instruction::CatchPad:
3695   case Instruction::CatchRet:
3696   case Instruction::CleanupPad:
3697   case Instruction::CleanupRet:
3698     return false; // Misc instructions which have effects
3699   }
3700 }
3701 
mayBeMemoryDependent(const Instruction & I)3702 bool llvm::mayBeMemoryDependent(const Instruction &I) {
3703   return I.mayReadOrWriteMemory() || !isSafeToSpeculativelyExecute(&I);
3704 }
3705 
computeOverflowForUnsignedMul(const Value * LHS,const Value * RHS,const DataLayout & DL,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)3706 OverflowResult llvm::computeOverflowForUnsignedMul(const Value *LHS,
3707                                                    const Value *RHS,
3708                                                    const DataLayout &DL,
3709                                                    AssumptionCache *AC,
3710                                                    const Instruction *CxtI,
3711                                                    const DominatorTree *DT) {
3712   // Multiplying n * m significant bits yields a result of n + m significant
3713   // bits. If the total number of significant bits does not exceed the
3714   // result bit width (minus 1), there is no overflow.
3715   // This means if we have enough leading zero bits in the operands
3716   // we can guarantee that the result does not overflow.
3717   // Ref: "Hacker's Delight" by Henry Warren
3718   unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
3719   KnownBits LHSKnown(BitWidth);
3720   KnownBits RHSKnown(BitWidth);
3721   computeKnownBits(LHS, LHSKnown, DL, /*Depth=*/0, AC, CxtI, DT);
3722   computeKnownBits(RHS, RHSKnown, DL, /*Depth=*/0, AC, CxtI, DT);
3723   // Note that underestimating the number of zero bits gives a more
3724   // conservative answer.
3725   unsigned ZeroBits = LHSKnown.countMinLeadingZeros() +
3726                       RHSKnown.countMinLeadingZeros();
3727   // First handle the easy case: if we have enough zero bits there's
3728   // definitely no overflow.
3729   if (ZeroBits >= BitWidth)
3730     return OverflowResult::NeverOverflows;
3731 
3732   // Get the largest possible values for each operand.
3733   APInt LHSMax = ~LHSKnown.Zero;
3734   APInt RHSMax = ~RHSKnown.Zero;
3735 
3736   // We know the multiply operation doesn't overflow if the maximum values for
3737   // each operand will not overflow after we multiply them together.
3738   bool MaxOverflow;
3739   (void)LHSMax.umul_ov(RHSMax, MaxOverflow);
3740   if (!MaxOverflow)
3741     return OverflowResult::NeverOverflows;
3742 
3743   // We know it always overflows if multiplying the smallest possible values for
3744   // the operands also results in overflow.
3745   bool MinOverflow;
3746   (void)LHSKnown.One.umul_ov(RHSKnown.One, MinOverflow);
3747   if (MinOverflow)
3748     return OverflowResult::AlwaysOverflows;
3749 
3750   return OverflowResult::MayOverflow;
3751 }
3752 
computeOverflowForSignedMul(const Value * LHS,const Value * RHS,const DataLayout & DL,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)3753 OverflowResult llvm::computeOverflowForSignedMul(const Value *LHS,
3754                                                  const Value *RHS,
3755                                                  const DataLayout &DL,
3756                                                  AssumptionCache *AC,
3757                                                  const Instruction *CxtI,
3758                                                  const DominatorTree *DT) {
3759   // Multiplying n * m significant bits yields a result of n + m significant
3760   // bits. If the total number of significant bits does not exceed the
3761   // result bit width (minus 1), there is no overflow.
3762   // This means if we have enough leading sign bits in the operands
3763   // we can guarantee that the result does not overflow.
3764   // Ref: "Hacker's Delight" by Henry Warren
3765   unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
3766 
3767   // Note that underestimating the number of sign bits gives a more
3768   // conservative answer.
3769   unsigned SignBits = ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) +
3770                       ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT);
3771 
3772   // First handle the easy case: if we have enough sign bits there's
3773   // definitely no overflow.
3774   if (SignBits > BitWidth + 1)
3775     return OverflowResult::NeverOverflows;
3776 
3777   // There are two ambiguous cases where there can be no overflow:
3778   //   SignBits == BitWidth + 1    and
3779   //   SignBits == BitWidth
3780   // The second case is difficult to check, therefore we only handle the
3781   // first case.
3782   if (SignBits == BitWidth + 1) {
3783     // It overflows only when both arguments are negative and the true
3784     // product is exactly the minimum negative number.
3785     // E.g. mul i16 with 17 sign bits: 0xff00 * 0xff80 = 0x8000
3786     // For simplicity we just check if at least one side is not negative.
3787     KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT);
3788     KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT);
3789     if (LHSKnown.isNonNegative() || RHSKnown.isNonNegative())
3790       return OverflowResult::NeverOverflows;
3791   }
3792   return OverflowResult::MayOverflow;
3793 }
3794 
computeOverflowForUnsignedAdd(const Value * LHS,const Value * RHS,const DataLayout & DL,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)3795 OverflowResult llvm::computeOverflowForUnsignedAdd(const Value *LHS,
3796                                                    const Value *RHS,
3797                                                    const DataLayout &DL,
3798                                                    AssumptionCache *AC,
3799                                                    const Instruction *CxtI,
3800                                                    const DominatorTree *DT) {
3801   KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT);
3802   if (LHSKnown.isNonNegative() || LHSKnown.isNegative()) {
3803     KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT);
3804 
3805     if (LHSKnown.isNegative() && RHSKnown.isNegative()) {
3806       // The sign bit is set in both cases: this MUST overflow.
3807       // Create a simple add instruction, and insert it into the struct.
3808       return OverflowResult::AlwaysOverflows;
3809     }
3810 
3811     if (LHSKnown.isNonNegative() && RHSKnown.isNonNegative()) {
3812       // The sign bit is clear in both cases: this CANNOT overflow.
3813       // Create a simple add instruction, and insert it into the struct.
3814       return OverflowResult::NeverOverflows;
3815     }
3816   }
3817 
3818   return OverflowResult::MayOverflow;
3819 }
3820 
3821 /// Return true if we can prove that adding the two values of the
3822 /// knownbits will not overflow.
3823 /// Otherwise return false.
checkRippleForSignedAdd(const KnownBits & LHSKnown,const KnownBits & RHSKnown)3824 static bool checkRippleForSignedAdd(const KnownBits &LHSKnown,
3825                                     const KnownBits &RHSKnown) {
3826   // Addition of two 2's complement numbers having opposite signs will never
3827   // overflow.
3828   if ((LHSKnown.isNegative() && RHSKnown.isNonNegative()) ||
3829       (LHSKnown.isNonNegative() && RHSKnown.isNegative()))
3830     return true;
3831 
3832   // If either of the values is known to be non-negative, adding them can only
3833   // overflow if the second is also non-negative, so we can assume that.
3834   // Two non-negative numbers will only overflow if there is a carry to the
3835   // sign bit, so we can check if even when the values are as big as possible
3836   // there is no overflow to the sign bit.
3837   if (LHSKnown.isNonNegative() || RHSKnown.isNonNegative()) {
3838     APInt MaxLHS = ~LHSKnown.Zero;
3839     MaxLHS.clearSignBit();
3840     APInt MaxRHS = ~RHSKnown.Zero;
3841     MaxRHS.clearSignBit();
3842     APInt Result = std::move(MaxLHS) + std::move(MaxRHS);
3843     return Result.isSignBitClear();
3844   }
3845 
3846   // If either of the values is known to be negative, adding them can only
3847   // overflow if the second is also negative, so we can assume that.
3848   // Two negative number will only overflow if there is no carry to the sign
3849   // bit, so we can check if even when the values are as small as possible
3850   // there is overflow to the sign bit.
3851   if (LHSKnown.isNegative() || RHSKnown.isNegative()) {
3852     APInt MinLHS = LHSKnown.One;
3853     MinLHS.clearSignBit();
3854     APInt MinRHS = RHSKnown.One;
3855     MinRHS.clearSignBit();
3856     APInt Result = std::move(MinLHS) + std::move(MinRHS);
3857     return Result.isSignBitSet();
3858   }
3859 
3860   // If we reached here it means that we know nothing about the sign bits.
3861   // In this case we can't know if there will be an overflow, since by
3862   // changing the sign bits any two values can be made to overflow.
3863   return false;
3864 }
3865 
computeOverflowForSignedAdd(const Value * LHS,const Value * RHS,const AddOperator * Add,const DataLayout & DL,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)3866 static OverflowResult computeOverflowForSignedAdd(const Value *LHS,
3867                                                   const Value *RHS,
3868                                                   const AddOperator *Add,
3869                                                   const DataLayout &DL,
3870                                                   AssumptionCache *AC,
3871                                                   const Instruction *CxtI,
3872                                                   const DominatorTree *DT) {
3873   if (Add && Add->hasNoSignedWrap()) {
3874     return OverflowResult::NeverOverflows;
3875   }
3876 
3877   // If LHS and RHS each have at least two sign bits, the addition will look
3878   // like
3879   //
3880   // XX..... +
3881   // YY.....
3882   //
3883   // If the carry into the most significant position is 0, X and Y can't both
3884   // be 1 and therefore the carry out of the addition is also 0.
3885   //
3886   // If the carry into the most significant position is 1, X and Y can't both
3887   // be 0 and therefore the carry out of the addition is also 1.
3888   //
3889   // Since the carry into the most significant position is always equal to
3890   // the carry out of the addition, there is no signed overflow.
3891   if (ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) > 1 &&
3892       ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT) > 1)
3893     return OverflowResult::NeverOverflows;
3894 
3895   KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT);
3896   KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT);
3897 
3898   if (checkRippleForSignedAdd(LHSKnown, RHSKnown))
3899     return OverflowResult::NeverOverflows;
3900 
3901   // The remaining code needs Add to be available. Early returns if not so.
3902   if (!Add)
3903     return OverflowResult::MayOverflow;
3904 
3905   // If the sign of Add is the same as at least one of the operands, this add
3906   // CANNOT overflow. This is particularly useful when the sum is
3907   // @llvm.assume'ed non-negative rather than proved so from analyzing its
3908   // operands.
3909   bool LHSOrRHSKnownNonNegative =
3910       (LHSKnown.isNonNegative() || RHSKnown.isNonNegative());
3911   bool LHSOrRHSKnownNegative =
3912       (LHSKnown.isNegative() || RHSKnown.isNegative());
3913   if (LHSOrRHSKnownNonNegative || LHSOrRHSKnownNegative) {
3914     KnownBits AddKnown = computeKnownBits(Add, DL, /*Depth=*/0, AC, CxtI, DT);
3915     if ((AddKnown.isNonNegative() && LHSOrRHSKnownNonNegative) ||
3916         (AddKnown.isNegative() && LHSOrRHSKnownNegative)) {
3917       return OverflowResult::NeverOverflows;
3918     }
3919   }
3920 
3921   return OverflowResult::MayOverflow;
3922 }
3923 
computeOverflowForUnsignedSub(const Value * LHS,const Value * RHS,const DataLayout & DL,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)3924 OverflowResult llvm::computeOverflowForUnsignedSub(const Value *LHS,
3925                                                    const Value *RHS,
3926                                                    const DataLayout &DL,
3927                                                    AssumptionCache *AC,
3928                                                    const Instruction *CxtI,
3929                                                    const DominatorTree *DT) {
3930   // If the LHS is negative and the RHS is non-negative, no unsigned wrap.
3931   KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT);
3932   KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT);
3933   if (LHSKnown.isNegative() && RHSKnown.isNonNegative())
3934     return OverflowResult::NeverOverflows;
3935 
3936   return OverflowResult::MayOverflow;
3937 }
3938 
computeOverflowForSignedSub(const Value * LHS,const Value * RHS,const DataLayout & DL,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)3939 OverflowResult llvm::computeOverflowForSignedSub(const Value *LHS,
3940                                                  const Value *RHS,
3941                                                  const DataLayout &DL,
3942                                                  AssumptionCache *AC,
3943                                                  const Instruction *CxtI,
3944                                                  const DominatorTree *DT) {
3945   // If LHS and RHS each have at least two sign bits, the subtraction
3946   // cannot overflow.
3947   if (ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) > 1 &&
3948       ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT) > 1)
3949     return OverflowResult::NeverOverflows;
3950 
3951   KnownBits LHSKnown = computeKnownBits(LHS, DL, 0, AC, CxtI, DT);
3952 
3953   KnownBits RHSKnown = computeKnownBits(RHS, DL, 0, AC, CxtI, DT);
3954 
3955   // Subtraction of two 2's complement numbers having identical signs will
3956   // never overflow.
3957   if ((LHSKnown.isNegative() && RHSKnown.isNegative()) ||
3958       (LHSKnown.isNonNegative() && RHSKnown.isNonNegative()))
3959     return OverflowResult::NeverOverflows;
3960 
3961   // TODO: implement logic similar to checkRippleForAdd
3962   return OverflowResult::MayOverflow;
3963 }
3964 
isOverflowIntrinsicNoWrap(const IntrinsicInst * II,const DominatorTree & DT)3965 bool llvm::isOverflowIntrinsicNoWrap(const IntrinsicInst *II,
3966                                      const DominatorTree &DT) {
3967 #ifndef NDEBUG
3968   auto IID = II->getIntrinsicID();
3969   assert((IID == Intrinsic::sadd_with_overflow ||
3970           IID == Intrinsic::uadd_with_overflow ||
3971           IID == Intrinsic::ssub_with_overflow ||
3972           IID == Intrinsic::usub_with_overflow ||
3973           IID == Intrinsic::smul_with_overflow ||
3974           IID == Intrinsic::umul_with_overflow) &&
3975          "Not an overflow intrinsic!");
3976 #endif
3977 
3978   SmallVector<const BranchInst *, 2> GuardingBranches;
3979   SmallVector<const ExtractValueInst *, 2> Results;
3980 
3981   for (const User *U : II->users()) {
3982     if (const auto *EVI = dyn_cast<ExtractValueInst>(U)) {
3983       assert(EVI->getNumIndices() == 1 && "Obvious from CI's type");
3984 
3985       if (EVI->getIndices()[0] == 0)
3986         Results.push_back(EVI);
3987       else {
3988         assert(EVI->getIndices()[0] == 1 && "Obvious from CI's type");
3989 
3990         for (const auto *U : EVI->users())
3991           if (const auto *B = dyn_cast<BranchInst>(U)) {
3992             assert(B->isConditional() && "How else is it using an i1?");
3993             GuardingBranches.push_back(B);
3994           }
3995       }
3996     } else {
3997       // We are using the aggregate directly in a way we don't want to analyze
3998       // here (storing it to a global, say).
3999       return false;
4000     }
4001   }
4002 
4003   auto AllUsesGuardedByBranch = [&](const BranchInst *BI) {
4004     BasicBlockEdge NoWrapEdge(BI->getParent(), BI->getSuccessor(1));
4005     if (!NoWrapEdge.isSingleEdge())
4006       return false;
4007 
4008     // Check if all users of the add are provably no-wrap.
4009     for (const auto *Result : Results) {
4010       // If the extractvalue itself is not executed on overflow, the we don't
4011       // need to check each use separately, since domination is transitive.
4012       if (DT.dominates(NoWrapEdge, Result->getParent()))
4013         continue;
4014 
4015       for (auto &RU : Result->uses())
4016         if (!DT.dominates(NoWrapEdge, RU))
4017           return false;
4018     }
4019 
4020     return true;
4021   };
4022 
4023   return llvm::any_of(GuardingBranches, AllUsesGuardedByBranch);
4024 }
4025 
4026 
computeOverflowForSignedAdd(const AddOperator * Add,const DataLayout & DL,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)4027 OverflowResult llvm::computeOverflowForSignedAdd(const AddOperator *Add,
4028                                                  const DataLayout &DL,
4029                                                  AssumptionCache *AC,
4030                                                  const Instruction *CxtI,
4031                                                  const DominatorTree *DT) {
4032   return ::computeOverflowForSignedAdd(Add->getOperand(0), Add->getOperand(1),
4033                                        Add, DL, AC, CxtI, DT);
4034 }
4035 
computeOverflowForSignedAdd(const Value * LHS,const Value * RHS,const DataLayout & DL,AssumptionCache * AC,const Instruction * CxtI,const DominatorTree * DT)4036 OverflowResult llvm::computeOverflowForSignedAdd(const Value *LHS,
4037                                                  const Value *RHS,
4038                                                  const DataLayout &DL,
4039                                                  AssumptionCache *AC,
4040                                                  const Instruction *CxtI,
4041                                                  const DominatorTree *DT) {
4042   return ::computeOverflowForSignedAdd(LHS, RHS, nullptr, DL, AC, CxtI, DT);
4043 }
4044 
isGuaranteedToTransferExecutionToSuccessor(const Instruction * I)4045 bool llvm::isGuaranteedToTransferExecutionToSuccessor(const Instruction *I) {
4046   // A memory operation returns normally if it isn't volatile. A volatile
4047   // operation is allowed to trap.
4048   //
4049   // An atomic operation isn't guaranteed to return in a reasonable amount of
4050   // time because it's possible for another thread to interfere with it for an
4051   // arbitrary length of time, but programs aren't allowed to rely on that.
4052   if (const LoadInst *LI = dyn_cast<LoadInst>(I))
4053     return !LI->isVolatile();
4054   if (const StoreInst *SI = dyn_cast<StoreInst>(I))
4055     return !SI->isVolatile();
4056   if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I))
4057     return !CXI->isVolatile();
4058   if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I))
4059     return !RMWI->isVolatile();
4060   if (const MemIntrinsic *MII = dyn_cast<MemIntrinsic>(I))
4061     return !MII->isVolatile();
4062 
4063   // If there is no successor, then execution can't transfer to it.
4064   if (const auto *CRI = dyn_cast<CleanupReturnInst>(I))
4065     return !CRI->unwindsToCaller();
4066   if (const auto *CatchSwitch = dyn_cast<CatchSwitchInst>(I))
4067     return !CatchSwitch->unwindsToCaller();
4068   if (isa<ResumeInst>(I))
4069     return false;
4070   if (isa<ReturnInst>(I))
4071     return false;
4072   if (isa<UnreachableInst>(I))
4073     return false;
4074 
4075   // Calls can throw, or contain an infinite loop, or kill the process.
4076   if (auto CS = ImmutableCallSite(I)) {
4077     // Call sites that throw have implicit non-local control flow.
4078     if (!CS.doesNotThrow())
4079       return false;
4080 
4081     // Non-throwing call sites can loop infinitely, call exit/pthread_exit
4082     // etc. and thus not return.  However, LLVM already assumes that
4083     //
4084     //  - Thread exiting actions are modeled as writes to memory invisible to
4085     //    the program.
4086     //
4087     //  - Loops that don't have side effects (side effects are volatile/atomic
4088     //    stores and IO) always terminate (see http://llvm.org/PR965).
4089     //    Furthermore IO itself is also modeled as writes to memory invisible to
4090     //    the program.
4091     //
4092     // We rely on those assumptions here, and use the memory effects of the call
4093     // target as a proxy for checking that it always returns.
4094 
4095     // FIXME: This isn't aggressive enough; a call which only writes to a global
4096     // is guaranteed to return.
4097     return CS.onlyReadsMemory() || CS.onlyAccessesArgMemory() ||
4098            match(I, m_Intrinsic<Intrinsic::assume>()) ||
4099            match(I, m_Intrinsic<Intrinsic::sideeffect>());
4100   }
4101 
4102   // Other instructions return normally.
4103   return true;
4104 }
4105 
isGuaranteedToTransferExecutionToSuccessor(const BasicBlock * BB)4106 bool llvm::isGuaranteedToTransferExecutionToSuccessor(const BasicBlock *BB) {
4107   // TODO: This is slightly consdervative for invoke instruction since exiting
4108   // via an exception *is* normal control for them.
4109   for (auto I = BB->begin(), E = BB->end(); I != E; ++I)
4110     if (!isGuaranteedToTransferExecutionToSuccessor(&*I))
4111       return false;
4112   return true;
4113 }
4114 
isGuaranteedToExecuteForEveryIteration(const Instruction * I,const Loop * L)4115 bool llvm::isGuaranteedToExecuteForEveryIteration(const Instruction *I,
4116                                                   const Loop *L) {
4117   // The loop header is guaranteed to be executed for every iteration.
4118   //
4119   // FIXME: Relax this constraint to cover all basic blocks that are
4120   // guaranteed to be executed at every iteration.
4121   if (I->getParent() != L->getHeader()) return false;
4122 
4123   for (const Instruction &LI : *L->getHeader()) {
4124     if (&LI == I) return true;
4125     if (!isGuaranteedToTransferExecutionToSuccessor(&LI)) return false;
4126   }
4127   llvm_unreachable("Instruction not contained in its own parent basic block.");
4128 }
4129 
propagatesFullPoison(const Instruction * I)4130 bool llvm::propagatesFullPoison(const Instruction *I) {
4131   switch (I->getOpcode()) {
4132   case Instruction::Add:
4133   case Instruction::Sub:
4134   case Instruction::Xor:
4135   case Instruction::Trunc:
4136   case Instruction::BitCast:
4137   case Instruction::AddrSpaceCast:
4138   case Instruction::Mul:
4139   case Instruction::Shl:
4140   case Instruction::GetElementPtr:
4141     // These operations all propagate poison unconditionally. Note that poison
4142     // is not any particular value, so xor or subtraction of poison with
4143     // itself still yields poison, not zero.
4144     return true;
4145 
4146   case Instruction::AShr:
4147   case Instruction::SExt:
4148     // For these operations, one bit of the input is replicated across
4149     // multiple output bits. A replicated poison bit is still poison.
4150     return true;
4151 
4152   case Instruction::ICmp:
4153     // Comparing poison with any value yields poison.  This is why, for
4154     // instance, x s< (x +nsw 1) can be folded to true.
4155     return true;
4156 
4157   default:
4158     return false;
4159   }
4160 }
4161 
getGuaranteedNonFullPoisonOp(const Instruction * I)4162 const Value *llvm::getGuaranteedNonFullPoisonOp(const Instruction *I) {
4163   switch (I->getOpcode()) {
4164     case Instruction::Store:
4165       return cast<StoreInst>(I)->getPointerOperand();
4166 
4167     case Instruction::Load:
4168       return cast<LoadInst>(I)->getPointerOperand();
4169 
4170     case Instruction::AtomicCmpXchg:
4171       return cast<AtomicCmpXchgInst>(I)->getPointerOperand();
4172 
4173     case Instruction::AtomicRMW:
4174       return cast<AtomicRMWInst>(I)->getPointerOperand();
4175 
4176     case Instruction::UDiv:
4177     case Instruction::SDiv:
4178     case Instruction::URem:
4179     case Instruction::SRem:
4180       return I->getOperand(1);
4181 
4182     default:
4183       return nullptr;
4184   }
4185 }
4186 
programUndefinedIfFullPoison(const Instruction * PoisonI)4187 bool llvm::programUndefinedIfFullPoison(const Instruction *PoisonI) {
4188   // We currently only look for uses of poison values within the same basic
4189   // block, as that makes it easier to guarantee that the uses will be
4190   // executed given that PoisonI is executed.
4191   //
4192   // FIXME: Expand this to consider uses beyond the same basic block. To do
4193   // this, look out for the distinction between post-dominance and strong
4194   // post-dominance.
4195   const BasicBlock *BB = PoisonI->getParent();
4196 
4197   // Set of instructions that we have proved will yield poison if PoisonI
4198   // does.
4199   SmallSet<const Value *, 16> YieldsPoison;
4200   SmallSet<const BasicBlock *, 4> Visited;
4201   YieldsPoison.insert(PoisonI);
4202   Visited.insert(PoisonI->getParent());
4203 
4204   BasicBlock::const_iterator Begin = PoisonI->getIterator(), End = BB->end();
4205 
4206   unsigned Iter = 0;
4207   while (Iter++ < MaxDepth) {
4208     for (auto &I : make_range(Begin, End)) {
4209       if (&I != PoisonI) {
4210         const Value *NotPoison = getGuaranteedNonFullPoisonOp(&I);
4211         if (NotPoison != nullptr && YieldsPoison.count(NotPoison))
4212           return true;
4213         if (!isGuaranteedToTransferExecutionToSuccessor(&I))
4214           return false;
4215       }
4216 
4217       // Mark poison that propagates from I through uses of I.
4218       if (YieldsPoison.count(&I)) {
4219         for (const User *User : I.users()) {
4220           const Instruction *UserI = cast<Instruction>(User);
4221           if (propagatesFullPoison(UserI))
4222             YieldsPoison.insert(User);
4223         }
4224       }
4225     }
4226 
4227     if (auto *NextBB = BB->getSingleSuccessor()) {
4228       if (Visited.insert(NextBB).second) {
4229         BB = NextBB;
4230         Begin = BB->getFirstNonPHI()->getIterator();
4231         End = BB->end();
4232         continue;
4233       }
4234     }
4235 
4236     break;
4237   }
4238   return false;
4239 }
4240 
isKnownNonNaN(const Value * V,FastMathFlags FMF)4241 static bool isKnownNonNaN(const Value *V, FastMathFlags FMF) {
4242   if (FMF.noNaNs())
4243     return true;
4244 
4245   if (auto *C = dyn_cast<ConstantFP>(V))
4246     return !C->isNaN();
4247   return false;
4248 }
4249 
isKnownNonZero(const Value * V)4250 static bool isKnownNonZero(const Value *V) {
4251   if (auto *C = dyn_cast<ConstantFP>(V))
4252     return !C->isZero();
4253   return false;
4254 }
4255 
4256 /// Match clamp pattern for float types without care about NaNs or signed zeros.
4257 /// Given non-min/max outer cmp/select from the clamp pattern this
4258 /// function recognizes if it can be substitued by a "canonical" min/max
4259 /// pattern.
matchFastFloatClamp(CmpInst::Predicate Pred,Value * CmpLHS,Value * CmpRHS,Value * TrueVal,Value * FalseVal,Value * & LHS,Value * & RHS)4260 static SelectPatternResult matchFastFloatClamp(CmpInst::Predicate Pred,
4261                                                Value *CmpLHS, Value *CmpRHS,
4262                                                Value *TrueVal, Value *FalseVal,
4263                                                Value *&LHS, Value *&RHS) {
4264   // Try to match
4265   //   X < C1 ? C1 : Min(X, C2) --> Max(C1, Min(X, C2))
4266   //   X > C1 ? C1 : Max(X, C2) --> Min(C1, Max(X, C2))
4267   // and return description of the outer Max/Min.
4268 
4269   // First, check if select has inverse order:
4270   if (CmpRHS == FalseVal) {
4271     std::swap(TrueVal, FalseVal);
4272     Pred = CmpInst::getInversePredicate(Pred);
4273   }
4274 
4275   // Assume success now. If there's no match, callers should not use these anyway.
4276   LHS = TrueVal;
4277   RHS = FalseVal;
4278 
4279   const APFloat *FC1;
4280   if (CmpRHS != TrueVal || !match(CmpRHS, m_APFloat(FC1)) || !FC1->isFinite())
4281     return {SPF_UNKNOWN, SPNB_NA, false};
4282 
4283   const APFloat *FC2;
4284   switch (Pred) {
4285   case CmpInst::FCMP_OLT:
4286   case CmpInst::FCMP_OLE:
4287   case CmpInst::FCMP_ULT:
4288   case CmpInst::FCMP_ULE:
4289     if (match(FalseVal,
4290               m_CombineOr(m_OrdFMin(m_Specific(CmpLHS), m_APFloat(FC2)),
4291                           m_UnordFMin(m_Specific(CmpLHS), m_APFloat(FC2)))) &&
4292         FC1->compare(*FC2) == APFloat::cmpResult::cmpLessThan)
4293       return {SPF_FMAXNUM, SPNB_RETURNS_ANY, false};
4294     break;
4295   case CmpInst::FCMP_OGT:
4296   case CmpInst::FCMP_OGE:
4297   case CmpInst::FCMP_UGT:
4298   case CmpInst::FCMP_UGE:
4299     if (match(FalseVal,
4300               m_CombineOr(m_OrdFMax(m_Specific(CmpLHS), m_APFloat(FC2)),
4301                           m_UnordFMax(m_Specific(CmpLHS), m_APFloat(FC2)))) &&
4302         FC1->compare(*FC2) == APFloat::cmpResult::cmpGreaterThan)
4303       return {SPF_FMINNUM, SPNB_RETURNS_ANY, false};
4304     break;
4305   default:
4306     break;
4307   }
4308 
4309   return {SPF_UNKNOWN, SPNB_NA, false};
4310 }
4311 
4312 /// Recognize variations of:
4313 ///   CLAMP(v,l,h) ==> ((v) < (l) ? (l) : ((v) > (h) ? (h) : (v)))
matchClamp(CmpInst::Predicate Pred,Value * CmpLHS,Value * CmpRHS,Value * TrueVal,Value * FalseVal)4314 static SelectPatternResult matchClamp(CmpInst::Predicate Pred,
4315                                       Value *CmpLHS, Value *CmpRHS,
4316                                       Value *TrueVal, Value *FalseVal) {
4317   // Swap the select operands and predicate to match the patterns below.
4318   if (CmpRHS != TrueVal) {
4319     Pred = ICmpInst::getSwappedPredicate(Pred);
4320     std::swap(TrueVal, FalseVal);
4321   }
4322   const APInt *C1;
4323   if (CmpRHS == TrueVal && match(CmpRHS, m_APInt(C1))) {
4324     const APInt *C2;
4325     // (X <s C1) ? C1 : SMIN(X, C2) ==> SMAX(SMIN(X, C2), C1)
4326     if (match(FalseVal, m_SMin(m_Specific(CmpLHS), m_APInt(C2))) &&
4327         C1->slt(*C2) && Pred == CmpInst::ICMP_SLT)
4328       return {SPF_SMAX, SPNB_NA, false};
4329 
4330     // (X >s C1) ? C1 : SMAX(X, C2) ==> SMIN(SMAX(X, C2), C1)
4331     if (match(FalseVal, m_SMax(m_Specific(CmpLHS), m_APInt(C2))) &&
4332         C1->sgt(*C2) && Pred == CmpInst::ICMP_SGT)
4333       return {SPF_SMIN, SPNB_NA, false};
4334 
4335     // (X <u C1) ? C1 : UMIN(X, C2) ==> UMAX(UMIN(X, C2), C1)
4336     if (match(FalseVal, m_UMin(m_Specific(CmpLHS), m_APInt(C2))) &&
4337         C1->ult(*C2) && Pred == CmpInst::ICMP_ULT)
4338       return {SPF_UMAX, SPNB_NA, false};
4339 
4340     // (X >u C1) ? C1 : UMAX(X, C2) ==> UMIN(UMAX(X, C2), C1)
4341     if (match(FalseVal, m_UMax(m_Specific(CmpLHS), m_APInt(C2))) &&
4342         C1->ugt(*C2) && Pred == CmpInst::ICMP_UGT)
4343       return {SPF_UMIN, SPNB_NA, false};
4344   }
4345   return {SPF_UNKNOWN, SPNB_NA, false};
4346 }
4347 
4348 /// Recognize variations of:
4349 ///   a < c ? min(a,b) : min(b,c) ==> min(min(a,b),min(b,c))
matchMinMaxOfMinMax(CmpInst::Predicate Pred,Value * CmpLHS,Value * CmpRHS,Value * TVal,Value * FVal,unsigned Depth)4350 static SelectPatternResult matchMinMaxOfMinMax(CmpInst::Predicate Pred,
4351                                                Value *CmpLHS, Value *CmpRHS,
4352                                                Value *TVal, Value *FVal,
4353                                                unsigned Depth) {
4354   // TODO: Allow FP min/max with nnan/nsz.
4355   assert(CmpInst::isIntPredicate(Pred) && "Expected integer comparison");
4356 
4357   Value *A, *B;
4358   SelectPatternResult L = matchSelectPattern(TVal, A, B, nullptr, Depth + 1);
4359   if (!SelectPatternResult::isMinOrMax(L.Flavor))
4360     return {SPF_UNKNOWN, SPNB_NA, false};
4361 
4362   Value *C, *D;
4363   SelectPatternResult R = matchSelectPattern(FVal, C, D, nullptr, Depth + 1);
4364   if (L.Flavor != R.Flavor)
4365     return {SPF_UNKNOWN, SPNB_NA, false};
4366 
4367   // We have something like: x Pred y ? min(a, b) : min(c, d).
4368   // Try to match the compare to the min/max operations of the select operands.
4369   // First, make sure we have the right compare predicate.
4370   switch (L.Flavor) {
4371   case SPF_SMIN:
4372     if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE) {
4373       Pred = ICmpInst::getSwappedPredicate(Pred);
4374       std::swap(CmpLHS, CmpRHS);
4375     }
4376     if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
4377       break;
4378     return {SPF_UNKNOWN, SPNB_NA, false};
4379   case SPF_SMAX:
4380     if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) {
4381       Pred = ICmpInst::getSwappedPredicate(Pred);
4382       std::swap(CmpLHS, CmpRHS);
4383     }
4384     if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
4385       break;
4386     return {SPF_UNKNOWN, SPNB_NA, false};
4387   case SPF_UMIN:
4388     if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) {
4389       Pred = ICmpInst::getSwappedPredicate(Pred);
4390       std::swap(CmpLHS, CmpRHS);
4391     }
4392     if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE)
4393       break;
4394     return {SPF_UNKNOWN, SPNB_NA, false};
4395   case SPF_UMAX:
4396     if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
4397       Pred = ICmpInst::getSwappedPredicate(Pred);
4398       std::swap(CmpLHS, CmpRHS);
4399     }
4400     if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
4401       break;
4402     return {SPF_UNKNOWN, SPNB_NA, false};
4403   default:
4404     return {SPF_UNKNOWN, SPNB_NA, false};
4405   }
4406 
4407   // If there is a common operand in the already matched min/max and the other
4408   // min/max operands match the compare operands (either directly or inverted),
4409   // then this is min/max of the same flavor.
4410 
4411   // a pred c ? m(a, b) : m(c, b) --> m(m(a, b), m(c, b))
4412   // ~c pred ~a ? m(a, b) : m(c, b) --> m(m(a, b), m(c, b))
4413   if (D == B) {
4414     if ((CmpLHS == A && CmpRHS == C) || (match(C, m_Not(m_Specific(CmpLHS))) &&
4415                                          match(A, m_Not(m_Specific(CmpRHS)))))
4416       return {L.Flavor, SPNB_NA, false};
4417   }
4418   // a pred d ? m(a, b) : m(b, d) --> m(m(a, b), m(b, d))
4419   // ~d pred ~a ? m(a, b) : m(b, d) --> m(m(a, b), m(b, d))
4420   if (C == B) {
4421     if ((CmpLHS == A && CmpRHS == D) || (match(D, m_Not(m_Specific(CmpLHS))) &&
4422                                          match(A, m_Not(m_Specific(CmpRHS)))))
4423       return {L.Flavor, SPNB_NA, false};
4424   }
4425   // b pred c ? m(a, b) : m(c, a) --> m(m(a, b), m(c, a))
4426   // ~c pred ~b ? m(a, b) : m(c, a) --> m(m(a, b), m(c, a))
4427   if (D == A) {
4428     if ((CmpLHS == B && CmpRHS == C) || (match(C, m_Not(m_Specific(CmpLHS))) &&
4429                                          match(B, m_Not(m_Specific(CmpRHS)))))
4430       return {L.Flavor, SPNB_NA, false};
4431   }
4432   // b pred d ? m(a, b) : m(a, d) --> m(m(a, b), m(a, d))
4433   // ~d pred ~b ? m(a, b) : m(a, d) --> m(m(a, b), m(a, d))
4434   if (C == A) {
4435     if ((CmpLHS == B && CmpRHS == D) || (match(D, m_Not(m_Specific(CmpLHS))) &&
4436                                          match(B, m_Not(m_Specific(CmpRHS)))))
4437       return {L.Flavor, SPNB_NA, false};
4438   }
4439 
4440   return {SPF_UNKNOWN, SPNB_NA, false};
4441 }
4442 
4443 /// Match non-obvious integer minimum and maximum sequences.
matchMinMax(CmpInst::Predicate Pred,Value * CmpLHS,Value * CmpRHS,Value * TrueVal,Value * FalseVal,Value * & LHS,Value * & RHS,unsigned Depth)4444 static SelectPatternResult matchMinMax(CmpInst::Predicate Pred,
4445                                        Value *CmpLHS, Value *CmpRHS,
4446                                        Value *TrueVal, Value *FalseVal,
4447                                        Value *&LHS, Value *&RHS,
4448                                        unsigned Depth) {
4449   // Assume success. If there's no match, callers should not use these anyway.
4450   LHS = TrueVal;
4451   RHS = FalseVal;
4452 
4453   SelectPatternResult SPR = matchClamp(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal);
4454   if (SPR.Flavor != SelectPatternFlavor::SPF_UNKNOWN)
4455     return SPR;
4456 
4457   SPR = matchMinMaxOfMinMax(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, Depth);
4458   if (SPR.Flavor != SelectPatternFlavor::SPF_UNKNOWN)
4459     return SPR;
4460 
4461   if (Pred != CmpInst::ICMP_SGT && Pred != CmpInst::ICMP_SLT)
4462     return {SPF_UNKNOWN, SPNB_NA, false};
4463 
4464   // Z = X -nsw Y
4465   // (X >s Y) ? 0 : Z ==> (Z >s 0) ? 0 : Z ==> SMIN(Z, 0)
4466   // (X <s Y) ? 0 : Z ==> (Z <s 0) ? 0 : Z ==> SMAX(Z, 0)
4467   if (match(TrueVal, m_Zero()) &&
4468       match(FalseVal, m_NSWSub(m_Specific(CmpLHS), m_Specific(CmpRHS))))
4469     return {Pred == CmpInst::ICMP_SGT ? SPF_SMIN : SPF_SMAX, SPNB_NA, false};
4470 
4471   // Z = X -nsw Y
4472   // (X >s Y) ? Z : 0 ==> (Z >s 0) ? Z : 0 ==> SMAX(Z, 0)
4473   // (X <s Y) ? Z : 0 ==> (Z <s 0) ? Z : 0 ==> SMIN(Z, 0)
4474   if (match(FalseVal, m_Zero()) &&
4475       match(TrueVal, m_NSWSub(m_Specific(CmpLHS), m_Specific(CmpRHS))))
4476     return {Pred == CmpInst::ICMP_SGT ? SPF_SMAX : SPF_SMIN, SPNB_NA, false};
4477 
4478   const APInt *C1;
4479   if (!match(CmpRHS, m_APInt(C1)))
4480     return {SPF_UNKNOWN, SPNB_NA, false};
4481 
4482   // An unsigned min/max can be written with a signed compare.
4483   const APInt *C2;
4484   if ((CmpLHS == TrueVal && match(FalseVal, m_APInt(C2))) ||
4485       (CmpLHS == FalseVal && match(TrueVal, m_APInt(C2)))) {
4486     // Is the sign bit set?
4487     // (X <s 0) ? X : MAXVAL ==> (X >u MAXVAL) ? X : MAXVAL ==> UMAX
4488     // (X <s 0) ? MAXVAL : X ==> (X >u MAXVAL) ? MAXVAL : X ==> UMIN
4489     if (Pred == CmpInst::ICMP_SLT && C1->isNullValue() &&
4490         C2->isMaxSignedValue())
4491       return {CmpLHS == TrueVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false};
4492 
4493     // Is the sign bit clear?
4494     // (X >s -1) ? MINVAL : X ==> (X <u MINVAL) ? MINVAL : X ==> UMAX
4495     // (X >s -1) ? X : MINVAL ==> (X <u MINVAL) ? X : MINVAL ==> UMIN
4496     if (Pred == CmpInst::ICMP_SGT && C1->isAllOnesValue() &&
4497         C2->isMinSignedValue())
4498       return {CmpLHS == FalseVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false};
4499   }
4500 
4501   // Look through 'not' ops to find disguised signed min/max.
4502   // (X >s C) ? ~X : ~C ==> (~X <s ~C) ? ~X : ~C ==> SMIN(~X, ~C)
4503   // (X <s C) ? ~X : ~C ==> (~X >s ~C) ? ~X : ~C ==> SMAX(~X, ~C)
4504   if (match(TrueVal, m_Not(m_Specific(CmpLHS))) &&
4505       match(FalseVal, m_APInt(C2)) && ~(*C1) == *C2)
4506     return {Pred == CmpInst::ICMP_SGT ? SPF_SMIN : SPF_SMAX, SPNB_NA, false};
4507 
4508   // (X >s C) ? ~C : ~X ==> (~X <s ~C) ? ~C : ~X ==> SMAX(~C, ~X)
4509   // (X <s C) ? ~C : ~X ==> (~X >s ~C) ? ~C : ~X ==> SMIN(~C, ~X)
4510   if (match(FalseVal, m_Not(m_Specific(CmpLHS))) &&
4511       match(TrueVal, m_APInt(C2)) && ~(*C1) == *C2)
4512     return {Pred == CmpInst::ICMP_SGT ? SPF_SMAX : SPF_SMIN, SPNB_NA, false};
4513 
4514   return {SPF_UNKNOWN, SPNB_NA, false};
4515 }
4516 
isKnownNegation(const Value * X,const Value * Y,bool NeedNSW)4517 bool llvm::isKnownNegation(const Value *X, const Value *Y, bool NeedNSW) {
4518   assert(X && Y && "Invalid operand");
4519 
4520   // X = sub (0, Y) || X = sub nsw (0, Y)
4521   if ((!NeedNSW && match(X, m_Sub(m_ZeroInt(), m_Specific(Y)))) ||
4522       (NeedNSW && match(X, m_NSWSub(m_ZeroInt(), m_Specific(Y)))))
4523     return true;
4524 
4525   // Y = sub (0, X) || Y = sub nsw (0, X)
4526   if ((!NeedNSW && match(Y, m_Sub(m_ZeroInt(), m_Specific(X)))) ||
4527       (NeedNSW && match(Y, m_NSWSub(m_ZeroInt(), m_Specific(X)))))
4528     return true;
4529 
4530   // X = sub (A, B), Y = sub (B, A) || X = sub nsw (A, B), Y = sub nsw (B, A)
4531   Value *A, *B;
4532   return (!NeedNSW && (match(X, m_Sub(m_Value(A), m_Value(B))) &&
4533                         match(Y, m_Sub(m_Specific(B), m_Specific(A))))) ||
4534          (NeedNSW && (match(X, m_NSWSub(m_Value(A), m_Value(B))) &&
4535                        match(Y, m_NSWSub(m_Specific(B), m_Specific(A)))));
4536 }
4537 
matchSelectPattern(CmpInst::Predicate Pred,FastMathFlags FMF,Value * CmpLHS,Value * CmpRHS,Value * TrueVal,Value * FalseVal,Value * & LHS,Value * & RHS,unsigned Depth)4538 static SelectPatternResult matchSelectPattern(CmpInst::Predicate Pred,
4539                                               FastMathFlags FMF,
4540                                               Value *CmpLHS, Value *CmpRHS,
4541                                               Value *TrueVal, Value *FalseVal,
4542                                               Value *&LHS, Value *&RHS,
4543                                               unsigned Depth) {
4544   LHS = CmpLHS;
4545   RHS = CmpRHS;
4546 
4547   // Signed zero may return inconsistent results between implementations.
4548   //  (0.0 <= -0.0) ? 0.0 : -0.0 // Returns 0.0
4549   //  minNum(0.0, -0.0)          // May return -0.0 or 0.0 (IEEE 754-2008 5.3.1)
4550   // Therefore, we behave conservatively and only proceed if at least one of the
4551   // operands is known to not be zero or if we don't care about signed zero.
4552   switch (Pred) {
4553   default: break;
4554   // FIXME: Include OGT/OLT/UGT/ULT.
4555   case CmpInst::FCMP_OGE: case CmpInst::FCMP_OLE:
4556   case CmpInst::FCMP_UGE: case CmpInst::FCMP_ULE:
4557     if (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) &&
4558         !isKnownNonZero(CmpRHS))
4559       return {SPF_UNKNOWN, SPNB_NA, false};
4560   }
4561 
4562   SelectPatternNaNBehavior NaNBehavior = SPNB_NA;
4563   bool Ordered = false;
4564 
4565   // When given one NaN and one non-NaN input:
4566   //   - maxnum/minnum (C99 fmaxf()/fminf()) return the non-NaN input.
4567   //   - A simple C99 (a < b ? a : b) construction will return 'b' (as the
4568   //     ordered comparison fails), which could be NaN or non-NaN.
4569   // so here we discover exactly what NaN behavior is required/accepted.
4570   if (CmpInst::isFPPredicate(Pred)) {
4571     bool LHSSafe = isKnownNonNaN(CmpLHS, FMF);
4572     bool RHSSafe = isKnownNonNaN(CmpRHS, FMF);
4573 
4574     if (LHSSafe && RHSSafe) {
4575       // Both operands are known non-NaN.
4576       NaNBehavior = SPNB_RETURNS_ANY;
4577     } else if (CmpInst::isOrdered(Pred)) {
4578       // An ordered comparison will return false when given a NaN, so it
4579       // returns the RHS.
4580       Ordered = true;
4581       if (LHSSafe)
4582         // LHS is non-NaN, so if RHS is NaN then NaN will be returned.
4583         NaNBehavior = SPNB_RETURNS_NAN;
4584       else if (RHSSafe)
4585         NaNBehavior = SPNB_RETURNS_OTHER;
4586       else
4587         // Completely unsafe.
4588         return {SPF_UNKNOWN, SPNB_NA, false};
4589     } else {
4590       Ordered = false;
4591       // An unordered comparison will return true when given a NaN, so it
4592       // returns the LHS.
4593       if (LHSSafe)
4594         // LHS is non-NaN, so if RHS is NaN then non-NaN will be returned.
4595         NaNBehavior = SPNB_RETURNS_OTHER;
4596       else if (RHSSafe)
4597         NaNBehavior = SPNB_RETURNS_NAN;
4598       else
4599         // Completely unsafe.
4600         return {SPF_UNKNOWN, SPNB_NA, false};
4601     }
4602   }
4603 
4604   if (TrueVal == CmpRHS && FalseVal == CmpLHS) {
4605     std::swap(CmpLHS, CmpRHS);
4606     Pred = CmpInst::getSwappedPredicate(Pred);
4607     if (NaNBehavior == SPNB_RETURNS_NAN)
4608       NaNBehavior = SPNB_RETURNS_OTHER;
4609     else if (NaNBehavior == SPNB_RETURNS_OTHER)
4610       NaNBehavior = SPNB_RETURNS_NAN;
4611     Ordered = !Ordered;
4612   }
4613 
4614   // ([if]cmp X, Y) ? X : Y
4615   if (TrueVal == CmpLHS && FalseVal == CmpRHS) {
4616     switch (Pred) {
4617     default: return {SPF_UNKNOWN, SPNB_NA, false}; // Equality.
4618     case ICmpInst::ICMP_UGT:
4619     case ICmpInst::ICMP_UGE: return {SPF_UMAX, SPNB_NA, false};
4620     case ICmpInst::ICMP_SGT:
4621     case ICmpInst::ICMP_SGE: return {SPF_SMAX, SPNB_NA, false};
4622     case ICmpInst::ICMP_ULT:
4623     case ICmpInst::ICMP_ULE: return {SPF_UMIN, SPNB_NA, false};
4624     case ICmpInst::ICMP_SLT:
4625     case ICmpInst::ICMP_SLE: return {SPF_SMIN, SPNB_NA, false};
4626     case FCmpInst::FCMP_UGT:
4627     case FCmpInst::FCMP_UGE:
4628     case FCmpInst::FCMP_OGT:
4629     case FCmpInst::FCMP_OGE: return {SPF_FMAXNUM, NaNBehavior, Ordered};
4630     case FCmpInst::FCMP_ULT:
4631     case FCmpInst::FCMP_ULE:
4632     case FCmpInst::FCMP_OLT:
4633     case FCmpInst::FCMP_OLE: return {SPF_FMINNUM, NaNBehavior, Ordered};
4634     }
4635   }
4636 
4637   if (isKnownNegation(TrueVal, FalseVal)) {
4638     // Sign-extending LHS does not change its sign, so TrueVal/FalseVal can
4639     // match against either LHS or sext(LHS).
4640     auto MaybeSExtCmpLHS =
4641         m_CombineOr(m_Specific(CmpLHS), m_SExt(m_Specific(CmpLHS)));
4642     auto ZeroOrAllOnes = m_CombineOr(m_ZeroInt(), m_AllOnes());
4643     auto ZeroOrOne = m_CombineOr(m_ZeroInt(), m_One());
4644     if (match(TrueVal, MaybeSExtCmpLHS)) {
4645       // Set the return values. If the compare uses the negated value (-X >s 0),
4646       // swap the return values because the negated value is always 'RHS'.
4647       LHS = TrueVal;
4648       RHS = FalseVal;
4649       if (match(CmpLHS, m_Neg(m_Specific(FalseVal))))
4650         std::swap(LHS, RHS);
4651 
4652       // (X >s 0) ? X : -X or (X >s -1) ? X : -X --> ABS(X)
4653       // (-X >s 0) ? -X : X or (-X >s -1) ? -X : X --> ABS(X)
4654       if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, ZeroOrAllOnes))
4655         return {SPF_ABS, SPNB_NA, false};
4656 
4657       // (X <s 0) ? X : -X or (X <s 1) ? X : -X --> NABS(X)
4658       // (-X <s 0) ? -X : X or (-X <s 1) ? -X : X --> NABS(X)
4659       if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, ZeroOrOne))
4660         return {SPF_NABS, SPNB_NA, false};
4661     }
4662     else if (match(FalseVal, MaybeSExtCmpLHS)) {
4663       // Set the return values. If the compare uses the negated value (-X >s 0),
4664       // swap the return values because the negated value is always 'RHS'.
4665       LHS = FalseVal;
4666       RHS = TrueVal;
4667       if (match(CmpLHS, m_Neg(m_Specific(TrueVal))))
4668         std::swap(LHS, RHS);
4669 
4670       // (X >s 0) ? -X : X or (X >s -1) ? -X : X --> NABS(X)
4671       // (-X >s 0) ? X : -X or (-X >s -1) ? X : -X --> NABS(X)
4672       if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, ZeroOrAllOnes))
4673         return {SPF_NABS, SPNB_NA, false};
4674 
4675       // (X <s 0) ? -X : X or (X <s 1) ? -X : X --> ABS(X)
4676       // (-X <s 0) ? X : -X or (-X <s 1) ? X : -X --> ABS(X)
4677       if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, ZeroOrOne))
4678         return {SPF_ABS, SPNB_NA, false};
4679     }
4680   }
4681 
4682   if (CmpInst::isIntPredicate(Pred))
4683     return matchMinMax(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS, Depth);
4684 
4685   // According to (IEEE 754-2008 5.3.1), minNum(0.0, -0.0) and similar
4686   // may return either -0.0 or 0.0, so fcmp/select pair has stricter
4687   // semantics than minNum. Be conservative in such case.
4688   if (NaNBehavior != SPNB_RETURNS_ANY ||
4689       (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) &&
4690        !isKnownNonZero(CmpRHS)))
4691     return {SPF_UNKNOWN, SPNB_NA, false};
4692 
4693   return matchFastFloatClamp(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS);
4694 }
4695 
4696 /// Helps to match a select pattern in case of a type mismatch.
4697 ///
4698 /// The function processes the case when type of true and false values of a
4699 /// select instruction differs from type of the cmp instruction operands because
4700 /// of a cast instruction. The function checks if it is legal to move the cast
4701 /// operation after "select". If yes, it returns the new second value of
4702 /// "select" (with the assumption that cast is moved):
4703 /// 1. As operand of cast instruction when both values of "select" are same cast
4704 /// instructions.
4705 /// 2. As restored constant (by applying reverse cast operation) when the first
4706 /// value of the "select" is a cast operation and the second value is a
4707 /// constant.
4708 /// NOTE: We return only the new second value because the first value could be
4709 /// accessed as operand of cast instruction.
lookThroughCast(CmpInst * CmpI,Value * V1,Value * V2,Instruction::CastOps * CastOp)4710 static Value *lookThroughCast(CmpInst *CmpI, Value *V1, Value *V2,
4711                               Instruction::CastOps *CastOp) {
4712   auto *Cast1 = dyn_cast<CastInst>(V1);
4713   if (!Cast1)
4714     return nullptr;
4715 
4716   *CastOp = Cast1->getOpcode();
4717   Type *SrcTy = Cast1->getSrcTy();
4718   if (auto *Cast2 = dyn_cast<CastInst>(V2)) {
4719     // If V1 and V2 are both the same cast from the same type, look through V1.
4720     if (*CastOp == Cast2->getOpcode() && SrcTy == Cast2->getSrcTy())
4721       return Cast2->getOperand(0);
4722     return nullptr;
4723   }
4724 
4725   auto *C = dyn_cast<Constant>(V2);
4726   if (!C)
4727     return nullptr;
4728 
4729   Constant *CastedTo = nullptr;
4730   switch (*CastOp) {
4731   case Instruction::ZExt:
4732     if (CmpI->isUnsigned())
4733       CastedTo = ConstantExpr::getTrunc(C, SrcTy);
4734     break;
4735   case Instruction::SExt:
4736     if (CmpI->isSigned())
4737       CastedTo = ConstantExpr::getTrunc(C, SrcTy, true);
4738     break;
4739   case Instruction::Trunc:
4740     Constant *CmpConst;
4741     if (match(CmpI->getOperand(1), m_Constant(CmpConst)) &&
4742         CmpConst->getType() == SrcTy) {
4743       // Here we have the following case:
4744       //
4745       //   %cond = cmp iN %x, CmpConst
4746       //   %tr = trunc iN %x to iK
4747       //   %narrowsel = select i1 %cond, iK %t, iK C
4748       //
4749       // We can always move trunc after select operation:
4750       //
4751       //   %cond = cmp iN %x, CmpConst
4752       //   %widesel = select i1 %cond, iN %x, iN CmpConst
4753       //   %tr = trunc iN %widesel to iK
4754       //
4755       // Note that C could be extended in any way because we don't care about
4756       // upper bits after truncation. It can't be abs pattern, because it would
4757       // look like:
4758       //
4759       //   select i1 %cond, x, -x.
4760       //
4761       // So only min/max pattern could be matched. Such match requires widened C
4762       // == CmpConst. That is why set widened C = CmpConst, condition trunc
4763       // CmpConst == C is checked below.
4764       CastedTo = CmpConst;
4765     } else {
4766       CastedTo = ConstantExpr::getIntegerCast(C, SrcTy, CmpI->isSigned());
4767     }
4768     break;
4769   case Instruction::FPTrunc:
4770     CastedTo = ConstantExpr::getFPExtend(C, SrcTy, true);
4771     break;
4772   case Instruction::FPExt:
4773     CastedTo = ConstantExpr::getFPTrunc(C, SrcTy, true);
4774     break;
4775   case Instruction::FPToUI:
4776     CastedTo = ConstantExpr::getUIToFP(C, SrcTy, true);
4777     break;
4778   case Instruction::FPToSI:
4779     CastedTo = ConstantExpr::getSIToFP(C, SrcTy, true);
4780     break;
4781   case Instruction::UIToFP:
4782     CastedTo = ConstantExpr::getFPToUI(C, SrcTy, true);
4783     break;
4784   case Instruction::SIToFP:
4785     CastedTo = ConstantExpr::getFPToSI(C, SrcTy, true);
4786     break;
4787   default:
4788     break;
4789   }
4790 
4791   if (!CastedTo)
4792     return nullptr;
4793 
4794   // Make sure the cast doesn't lose any information.
4795   Constant *CastedBack =
4796       ConstantExpr::getCast(*CastOp, CastedTo, C->getType(), true);
4797   if (CastedBack != C)
4798     return nullptr;
4799 
4800   return CastedTo;
4801 }
4802 
matchSelectPattern(Value * V,Value * & LHS,Value * & RHS,Instruction::CastOps * CastOp,unsigned Depth)4803 SelectPatternResult llvm::matchSelectPattern(Value *V, Value *&LHS, Value *&RHS,
4804                                              Instruction::CastOps *CastOp,
4805                                              unsigned Depth) {
4806   if (Depth >= MaxDepth)
4807     return {SPF_UNKNOWN, SPNB_NA, false};
4808 
4809   SelectInst *SI = dyn_cast<SelectInst>(V);
4810   if (!SI) return {SPF_UNKNOWN, SPNB_NA, false};
4811 
4812   CmpInst *CmpI = dyn_cast<CmpInst>(SI->getCondition());
4813   if (!CmpI) return {SPF_UNKNOWN, SPNB_NA, false};
4814 
4815   CmpInst::Predicate Pred = CmpI->getPredicate();
4816   Value *CmpLHS = CmpI->getOperand(0);
4817   Value *CmpRHS = CmpI->getOperand(1);
4818   Value *TrueVal = SI->getTrueValue();
4819   Value *FalseVal = SI->getFalseValue();
4820   FastMathFlags FMF;
4821   if (isa<FPMathOperator>(CmpI))
4822     FMF = CmpI->getFastMathFlags();
4823 
4824   // Bail out early.
4825   if (CmpI->isEquality())
4826     return {SPF_UNKNOWN, SPNB_NA, false};
4827 
4828   // Deal with type mismatches.
4829   if (CastOp && CmpLHS->getType() != TrueVal->getType()) {
4830     if (Value *C = lookThroughCast(CmpI, TrueVal, FalseVal, CastOp)) {
4831       // If this is a potential fmin/fmax with a cast to integer, then ignore
4832       // -0.0 because there is no corresponding integer value.
4833       if (*CastOp == Instruction::FPToSI || *CastOp == Instruction::FPToUI)
4834         FMF.setNoSignedZeros();
4835       return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
4836                                   cast<CastInst>(TrueVal)->getOperand(0), C,
4837                                   LHS, RHS, Depth);
4838     }
4839     if (Value *C = lookThroughCast(CmpI, FalseVal, TrueVal, CastOp)) {
4840       // If this is a potential fmin/fmax with a cast to integer, then ignore
4841       // -0.0 because there is no corresponding integer value.
4842       if (*CastOp == Instruction::FPToSI || *CastOp == Instruction::FPToUI)
4843         FMF.setNoSignedZeros();
4844       return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
4845                                   C, cast<CastInst>(FalseVal)->getOperand(0),
4846                                   LHS, RHS, Depth);
4847     }
4848   }
4849   return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, TrueVal, FalseVal,
4850                               LHS, RHS, Depth);
4851 }
4852 
getMinMaxPred(SelectPatternFlavor SPF,bool Ordered)4853 CmpInst::Predicate llvm::getMinMaxPred(SelectPatternFlavor SPF, bool Ordered) {
4854   if (SPF == SPF_SMIN) return ICmpInst::ICMP_SLT;
4855   if (SPF == SPF_UMIN) return ICmpInst::ICMP_ULT;
4856   if (SPF == SPF_SMAX) return ICmpInst::ICMP_SGT;
4857   if (SPF == SPF_UMAX) return ICmpInst::ICMP_UGT;
4858   if (SPF == SPF_FMINNUM)
4859     return Ordered ? FCmpInst::FCMP_OLT : FCmpInst::FCMP_ULT;
4860   if (SPF == SPF_FMAXNUM)
4861     return Ordered ? FCmpInst::FCMP_OGT : FCmpInst::FCMP_UGT;
4862   llvm_unreachable("unhandled!");
4863 }
4864 
getInverseMinMaxFlavor(SelectPatternFlavor SPF)4865 SelectPatternFlavor llvm::getInverseMinMaxFlavor(SelectPatternFlavor SPF) {
4866   if (SPF == SPF_SMIN) return SPF_SMAX;
4867   if (SPF == SPF_UMIN) return SPF_UMAX;
4868   if (SPF == SPF_SMAX) return SPF_SMIN;
4869   if (SPF == SPF_UMAX) return SPF_UMIN;
4870   llvm_unreachable("unhandled!");
4871 }
4872 
getInverseMinMaxPred(SelectPatternFlavor SPF)4873 CmpInst::Predicate llvm::getInverseMinMaxPred(SelectPatternFlavor SPF) {
4874   return getMinMaxPred(getInverseMinMaxFlavor(SPF));
4875 }
4876 
4877 /// Return true if "icmp Pred LHS RHS" is always true.
isTruePredicate(CmpInst::Predicate Pred,const Value * LHS,const Value * RHS,const DataLayout & DL,unsigned Depth)4878 static bool isTruePredicate(CmpInst::Predicate Pred, const Value *LHS,
4879                             const Value *RHS, const DataLayout &DL,
4880                             unsigned Depth) {
4881   assert(!LHS->getType()->isVectorTy() && "TODO: extend to handle vectors!");
4882   if (ICmpInst::isTrueWhenEqual(Pred) && LHS == RHS)
4883     return true;
4884 
4885   switch (Pred) {
4886   default:
4887     return false;
4888 
4889   case CmpInst::ICMP_SLE: {
4890     const APInt *C;
4891 
4892     // LHS s<= LHS +_{nsw} C   if C >= 0
4893     if (match(RHS, m_NSWAdd(m_Specific(LHS), m_APInt(C))))
4894       return !C->isNegative();
4895     return false;
4896   }
4897 
4898   case CmpInst::ICMP_ULE: {
4899     const APInt *C;
4900 
4901     // LHS u<= LHS +_{nuw} C   for any C
4902     if (match(RHS, m_NUWAdd(m_Specific(LHS), m_APInt(C))))
4903       return true;
4904 
4905     // Match A to (X +_{nuw} CA) and B to (X +_{nuw} CB)
4906     auto MatchNUWAddsToSameValue = [&](const Value *A, const Value *B,
4907                                        const Value *&X,
4908                                        const APInt *&CA, const APInt *&CB) {
4909       if (match(A, m_NUWAdd(m_Value(X), m_APInt(CA))) &&
4910           match(B, m_NUWAdd(m_Specific(X), m_APInt(CB))))
4911         return true;
4912 
4913       // If X & C == 0 then (X | C) == X +_{nuw} C
4914       if (match(A, m_Or(m_Value(X), m_APInt(CA))) &&
4915           match(B, m_Or(m_Specific(X), m_APInt(CB)))) {
4916         KnownBits Known(CA->getBitWidth());
4917         computeKnownBits(X, Known, DL, Depth + 1, /*AC*/ nullptr,
4918                          /*CxtI*/ nullptr, /*DT*/ nullptr);
4919         if (CA->isSubsetOf(Known.Zero) && CB->isSubsetOf(Known.Zero))
4920           return true;
4921       }
4922 
4923       return false;
4924     };
4925 
4926     const Value *X;
4927     const APInt *CLHS, *CRHS;
4928     if (MatchNUWAddsToSameValue(LHS, RHS, X, CLHS, CRHS))
4929       return CLHS->ule(*CRHS);
4930 
4931     return false;
4932   }
4933   }
4934 }
4935 
4936 /// Return true if "icmp Pred BLHS BRHS" is true whenever "icmp Pred
4937 /// ALHS ARHS" is true.  Otherwise, return None.
4938 static Optional<bool>
isImpliedCondOperands(CmpInst::Predicate Pred,const Value * ALHS,const Value * ARHS,const Value * BLHS,const Value * BRHS,const DataLayout & DL,unsigned Depth)4939 isImpliedCondOperands(CmpInst::Predicate Pred, const Value *ALHS,
4940                       const Value *ARHS, const Value *BLHS, const Value *BRHS,
4941                       const DataLayout &DL, unsigned Depth) {
4942   switch (Pred) {
4943   default:
4944     return None;
4945 
4946   case CmpInst::ICMP_SLT:
4947   case CmpInst::ICMP_SLE:
4948     if (isTruePredicate(CmpInst::ICMP_SLE, BLHS, ALHS, DL, Depth) &&
4949         isTruePredicate(CmpInst::ICMP_SLE, ARHS, BRHS, DL, Depth))
4950       return true;
4951     return None;
4952 
4953   case CmpInst::ICMP_ULT:
4954   case CmpInst::ICMP_ULE:
4955     if (isTruePredicate(CmpInst::ICMP_ULE, BLHS, ALHS, DL, Depth) &&
4956         isTruePredicate(CmpInst::ICMP_ULE, ARHS, BRHS, DL, Depth))
4957       return true;
4958     return None;
4959   }
4960 }
4961 
4962 /// Return true if the operands of the two compares match.  IsSwappedOps is true
4963 /// when the operands match, but are swapped.
isMatchingOps(const Value * ALHS,const Value * ARHS,const Value * BLHS,const Value * BRHS,bool & IsSwappedOps)4964 static bool isMatchingOps(const Value *ALHS, const Value *ARHS,
4965                           const Value *BLHS, const Value *BRHS,
4966                           bool &IsSwappedOps) {
4967 
4968   bool IsMatchingOps = (ALHS == BLHS && ARHS == BRHS);
4969   IsSwappedOps = (ALHS == BRHS && ARHS == BLHS);
4970   return IsMatchingOps || IsSwappedOps;
4971 }
4972 
4973 /// Return true if "icmp1 APred ALHS ARHS" implies "icmp2 BPred BLHS BRHS" is
4974 /// true.  Return false if "icmp1 APred ALHS ARHS" implies "icmp2 BPred BLHS
4975 /// BRHS" is false.  Otherwise, return None if we can't infer anything.
isImpliedCondMatchingOperands(CmpInst::Predicate APred,const Value * ALHS,const Value * ARHS,CmpInst::Predicate BPred,const Value * BLHS,const Value * BRHS,bool IsSwappedOps)4976 static Optional<bool> isImpliedCondMatchingOperands(CmpInst::Predicate APred,
4977                                                     const Value *ALHS,
4978                                                     const Value *ARHS,
4979                                                     CmpInst::Predicate BPred,
4980                                                     const Value *BLHS,
4981                                                     const Value *BRHS,
4982                                                     bool IsSwappedOps) {
4983   // Canonicalize the operands so they're matching.
4984   if (IsSwappedOps) {
4985     std::swap(BLHS, BRHS);
4986     BPred = ICmpInst::getSwappedPredicate(BPred);
4987   }
4988   if (CmpInst::isImpliedTrueByMatchingCmp(APred, BPred))
4989     return true;
4990   if (CmpInst::isImpliedFalseByMatchingCmp(APred, BPred))
4991     return false;
4992 
4993   return None;
4994 }
4995 
4996 /// Return true if "icmp1 APred ALHS C1" implies "icmp2 BPred BLHS C2" is
4997 /// true.  Return false if "icmp1 APred ALHS C1" implies "icmp2 BPred BLHS
4998 /// C2" is false.  Otherwise, return None if we can't infer anything.
4999 static Optional<bool>
isImpliedCondMatchingImmOperands(CmpInst::Predicate APred,const Value * ALHS,const ConstantInt * C1,CmpInst::Predicate BPred,const Value * BLHS,const ConstantInt * C2)5000 isImpliedCondMatchingImmOperands(CmpInst::Predicate APred, const Value *ALHS,
5001                                  const ConstantInt *C1,
5002                                  CmpInst::Predicate BPred,
5003                                  const Value *BLHS, const ConstantInt *C2) {
5004   assert(ALHS == BLHS && "LHS operands must match.");
5005   ConstantRange DomCR =
5006       ConstantRange::makeExactICmpRegion(APred, C1->getValue());
5007   ConstantRange CR =
5008       ConstantRange::makeAllowedICmpRegion(BPred, C2->getValue());
5009   ConstantRange Intersection = DomCR.intersectWith(CR);
5010   ConstantRange Difference = DomCR.difference(CR);
5011   if (Intersection.isEmptySet())
5012     return false;
5013   if (Difference.isEmptySet())
5014     return true;
5015   return None;
5016 }
5017 
5018 /// Return true if LHS implies RHS is true.  Return false if LHS implies RHS is
5019 /// false.  Otherwise, return None if we can't infer anything.
isImpliedCondICmps(const ICmpInst * LHS,const ICmpInst * RHS,const DataLayout & DL,bool LHSIsTrue,unsigned Depth)5020 static Optional<bool> isImpliedCondICmps(const ICmpInst *LHS,
5021                                          const ICmpInst *RHS,
5022                                          const DataLayout &DL, bool LHSIsTrue,
5023                                          unsigned Depth) {
5024   Value *ALHS = LHS->getOperand(0);
5025   Value *ARHS = LHS->getOperand(1);
5026   // The rest of the logic assumes the LHS condition is true.  If that's not the
5027   // case, invert the predicate to make it so.
5028   ICmpInst::Predicate APred =
5029       LHSIsTrue ? LHS->getPredicate() : LHS->getInversePredicate();
5030 
5031   Value *BLHS = RHS->getOperand(0);
5032   Value *BRHS = RHS->getOperand(1);
5033   ICmpInst::Predicate BPred = RHS->getPredicate();
5034 
5035   // Can we infer anything when the two compares have matching operands?
5036   bool IsSwappedOps;
5037   if (isMatchingOps(ALHS, ARHS, BLHS, BRHS, IsSwappedOps)) {
5038     if (Optional<bool> Implication = isImpliedCondMatchingOperands(
5039             APred, ALHS, ARHS, BPred, BLHS, BRHS, IsSwappedOps))
5040       return Implication;
5041     // No amount of additional analysis will infer the second condition, so
5042     // early exit.
5043     return None;
5044   }
5045 
5046   // Can we infer anything when the LHS operands match and the RHS operands are
5047   // constants (not necessarily matching)?
5048   if (ALHS == BLHS && isa<ConstantInt>(ARHS) && isa<ConstantInt>(BRHS)) {
5049     if (Optional<bool> Implication = isImpliedCondMatchingImmOperands(
5050             APred, ALHS, cast<ConstantInt>(ARHS), BPred, BLHS,
5051             cast<ConstantInt>(BRHS)))
5052       return Implication;
5053     // No amount of additional analysis will infer the second condition, so
5054     // early exit.
5055     return None;
5056   }
5057 
5058   if (APred == BPred)
5059     return isImpliedCondOperands(APred, ALHS, ARHS, BLHS, BRHS, DL, Depth);
5060   return None;
5061 }
5062 
5063 /// Return true if LHS implies RHS is true.  Return false if LHS implies RHS is
5064 /// false.  Otherwise, return None if we can't infer anything.  We expect the
5065 /// RHS to be an icmp and the LHS to be an 'and' or an 'or' instruction.
isImpliedCondAndOr(const BinaryOperator * LHS,const ICmpInst * RHS,const DataLayout & DL,bool LHSIsTrue,unsigned Depth)5066 static Optional<bool> isImpliedCondAndOr(const BinaryOperator *LHS,
5067                                          const ICmpInst *RHS,
5068                                          const DataLayout &DL, bool LHSIsTrue,
5069                                          unsigned Depth) {
5070   // The LHS must be an 'or' or an 'and' instruction.
5071   assert((LHS->getOpcode() == Instruction::And ||
5072           LHS->getOpcode() == Instruction::Or) &&
5073          "Expected LHS to be 'and' or 'or'.");
5074 
5075   assert(Depth <= MaxDepth && "Hit recursion limit");
5076 
5077   // If the result of an 'or' is false, then we know both legs of the 'or' are
5078   // false.  Similarly, if the result of an 'and' is true, then we know both
5079   // legs of the 'and' are true.
5080   Value *ALHS, *ARHS;
5081   if ((!LHSIsTrue && match(LHS, m_Or(m_Value(ALHS), m_Value(ARHS)))) ||
5082       (LHSIsTrue && match(LHS, m_And(m_Value(ALHS), m_Value(ARHS))))) {
5083     // FIXME: Make this non-recursion.
5084     if (Optional<bool> Implication =
5085             isImpliedCondition(ALHS, RHS, DL, LHSIsTrue, Depth + 1))
5086       return Implication;
5087     if (Optional<bool> Implication =
5088             isImpliedCondition(ARHS, RHS, DL, LHSIsTrue, Depth + 1))
5089       return Implication;
5090     return None;
5091   }
5092   return None;
5093 }
5094 
isImpliedCondition(const Value * LHS,const Value * RHS,const DataLayout & DL,bool LHSIsTrue,unsigned Depth)5095 Optional<bool> llvm::isImpliedCondition(const Value *LHS, const Value *RHS,
5096                                         const DataLayout &DL, bool LHSIsTrue,
5097                                         unsigned Depth) {
5098   // Bail out when we hit the limit.
5099   if (Depth == MaxDepth)
5100     return None;
5101 
5102   // A mismatch occurs when we compare a scalar cmp to a vector cmp, for
5103   // example.
5104   if (LHS->getType() != RHS->getType())
5105     return None;
5106 
5107   Type *OpTy = LHS->getType();
5108   assert(OpTy->isIntOrIntVectorTy(1) && "Expected integer type only!");
5109 
5110   // LHS ==> RHS by definition
5111   if (LHS == RHS)
5112     return LHSIsTrue;
5113 
5114   // FIXME: Extending the code below to handle vectors.
5115   if (OpTy->isVectorTy())
5116     return None;
5117 
5118   assert(OpTy->isIntegerTy(1) && "implied by above");
5119 
5120   // Both LHS and RHS are icmps.
5121   const ICmpInst *LHSCmp = dyn_cast<ICmpInst>(LHS);
5122   const ICmpInst *RHSCmp = dyn_cast<ICmpInst>(RHS);
5123   if (LHSCmp && RHSCmp)
5124     return isImpliedCondICmps(LHSCmp, RHSCmp, DL, LHSIsTrue, Depth);
5125 
5126   // The LHS should be an 'or' or an 'and' instruction.  We expect the RHS to be
5127   // an icmp. FIXME: Add support for and/or on the RHS.
5128   const BinaryOperator *LHSBO = dyn_cast<BinaryOperator>(LHS);
5129   if (LHSBO && RHSCmp) {
5130     if ((LHSBO->getOpcode() == Instruction::And ||
5131          LHSBO->getOpcode() == Instruction::Or))
5132       return isImpliedCondAndOr(LHSBO, RHSCmp, DL, LHSIsTrue, Depth);
5133   }
5134   return None;
5135 }
5136