LCOV - code coverage report
Current view: top level - lib/Analysis - ValueTracking.cpp (source / functions) Hit Total Coverage
Test: llvm-toolchain.info Lines: 1530 1676 91.3 %
Date: 2018-02-23 15:42:53 Functions: 94 97 96.9 %
Legend: Lines: hit not hit

          Line data    Source code
       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       81686 : static cl::opt<unsigned> DomConditionsMaxUses("dom-conditions-max-uses",
      84       81686 :                                               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.
      88    25844694 : static unsigned getBitWidth(Type *Ty, const DataLayout &DL) {
      89    25844694 :   if (unsigned BitWidth = Ty->getScalarSizeInBits())
      90             :     return BitWidth;
      91             : 
      92     9523539 :   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             : 
     123             :   Query(const DataLayout &DL, AssumptionCache *AC, const Instruction *CxtI,
     124             :         const DominatorTree *DT, OptimizationRemarkEmitter *ORE = nullptr)
     125    40015874 :       : DL(DL), AC(AC), CxtI(CxtI), DT(DT), ORE(ORE) {}
     126             : 
     127             :   Query(const Query &Q, const Value *NewExcl)
     128        1000 :       : DL(Q.DL), AC(Q.AC), CxtI(Q.CxtI), DT(Q.DT), ORE(Q.ORE),
     129        1000 :         NumExcluded(Q.NumExcluded) {
     130         200 :     Excluded = Q.Excluded;
     131         200 :     Excluded[NumExcluded++] = NewExcl;
     132             :     assert(NumExcluded <= Excluded.size());
     133             :   }
     134             : 
     135        1483 :   bool isExcluded(const Value *Value) const {
     136        1483 :     if (NumExcluded == 0)
     137             :       return false;
     138          26 :     auto End = Excluded.begin() + NumExcluded;
     139          26 :     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).
     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    40504394 :   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     2102954 :   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             : 
     164     9432400 : 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     9432400 :   ::computeKnownBits(V, Known, Depth,
     170     9432400 :                      Query(DL, AC, safeCxtI(V, CxtI), DT, ORE));
     171     9432400 : }
     172             : 
     173             : static KnownBits computeKnownBits(const Value *V, unsigned Depth,
     174             :                                   const Query &Q);
     175             : 
     176    23301812 : 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    23301812 :                             Query(DL, AC, safeCxtI(V, CxtI), DT, ORE));
     183             : }
     184             : 
     185     3279395 : 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     3279395 :   IntegerType *IT = cast<IntegerType>(LHS->getType()->getScalarType());
     194     6558790 :   KnownBits LHSKnown(IT->getBitWidth());
     195     6558790 :   KnownBits RHSKnown(IT->getBitWidth());
     196     3279395 :   computeKnownBits(LHS, LHSKnown, DL, 0, AC, CxtI, DT);
     197     3279395 :   computeKnownBits(RHS, RHSKnown, DL, 0, AC, CxtI, DT);
     198     9838185 :   return (LHSKnown.Zero | RHSKnown.Zero).isAllOnesValue();
     199             : }
     200             : 
     201        2916 : bool llvm::isOnlyUsedInZeroEqualityComparison(const Instruction *CxtI) {
     202        3281 :   for (const User *U : CxtI->users()) {
     203             :     if (const ICmpInst *IC = dyn_cast<ICmpInst>(U))
     204         504 :       if (IC->isEquality())
     205             :         if (Constant *C = dyn_cast<Constant>(IC->getOperand(1)))
     206         365 :           if (C->isNullValue())
     207             :             continue;
     208             :     return false;
     209             :   }
     210             :   return true;
     211             : }
     212             : 
     213             : static bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth,
     214             :                                    const Query &Q);
     215             : 
     216        3538 : bool llvm::isKnownToBeAPowerOfTwo(const Value *V, const DataLayout &DL,
     217             :                                   bool OrZero,
     218             :                                   unsigned Depth, AssumptionCache *AC,
     219             :                                   const Instruction *CxtI,
     220             :                                   const DominatorTree *DT) {
     221        3538 :   return ::isKnownToBeAPowerOfTwo(V, OrZero, Depth,
     222        3538 :                                   Query(DL, AC, safeCxtI(V, CxtI), DT));
     223             : }
     224             : 
     225             : static bool isKnownNonZero(const Value *V, unsigned Depth, const Query &Q);
     226             : 
     227     2965565 : bool llvm::isKnownNonZero(const Value *V, const DataLayout &DL, unsigned Depth,
     228             :                           AssumptionCache *AC, const Instruction *CxtI,
     229             :                           const DominatorTree *DT) {
     230     2965565 :   return ::isKnownNonZero(V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT));
     231             : }
     232             : 
     233        1218 : bool llvm::isKnownNonNegative(const Value *V, const DataLayout &DL,
     234             :                               unsigned Depth,
     235             :                               AssumptionCache *AC, const Instruction *CxtI,
     236             :                               const DominatorTree *DT) {
     237        2436 :   KnownBits Known = computeKnownBits(V, DL, Depth, AC, CxtI, DT);
     238        1218 :   return Known.isNonNegative();
     239             : }
     240             : 
     241          10 : bool llvm::isKnownPositive(const Value *V, const DataLayout &DL, unsigned Depth,
     242             :                            AssumptionCache *AC, const Instruction *CxtI,
     243             :                            const DominatorTree *DT) {
     244             :   if (auto *CI = dyn_cast<ConstantInt>(V))
     245           1 :     return CI->getValue().isStrictlyPositive();
     246             : 
     247             :   // TODO: We'd doing two recursive queries here.  We should factor this such
     248             :   // that only a single query is needed.
     249          15 :   return isKnownNonNegative(V, DL, Depth, AC, CxtI, DT) &&
     250           6 :     isKnownNonZero(V, DL, Depth, AC, CxtI, DT);
     251             : }
     252             : 
     253           0 : bool llvm::isKnownNegative(const Value *V, const DataLayout &DL, unsigned Depth,
     254             :                            AssumptionCache *AC, const Instruction *CxtI,
     255             :                            const DominatorTree *DT) {
     256           0 :   KnownBits Known = computeKnownBits(V, DL, Depth, AC, CxtI, DT);
     257           0 :   return Known.isNegative();
     258             : }
     259             : 
     260             : static bool isKnownNonEqual(const Value *V1, const Value *V2, const Query &Q);
     261             : 
     262      813245 : bool llvm::isKnownNonEqual(const Value *V1, const Value *V2,
     263             :                            const DataLayout &DL,
     264             :                            AssumptionCache *AC, const Instruction *CxtI,
     265             :                            const DominatorTree *DT) {
     266      813245 :   return ::isKnownNonEqual(V1, V2, Query(DL, AC,
     267             :                                          safeCxtI(V1, safeCxtI(V2, CxtI)),
     268     1626490 :                                          DT));
     269             : }
     270             : 
     271             : static bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth,
     272             :                               const Query &Q);
     273             : 
     274       58176 : bool llvm::MaskedValueIsZero(const Value *V, const APInt &Mask,
     275             :                              const DataLayout &DL,
     276             :                              unsigned Depth, AssumptionCache *AC,
     277             :                              const Instruction *CxtI, const DominatorTree *DT) {
     278             :   return ::MaskedValueIsZero(V, Mask, Depth,
     279       58176 :                              Query(DL, AC, safeCxtI(V, CxtI), DT));
     280             : }
     281             : 
     282             : static unsigned ComputeNumSignBits(const Value *V, unsigned Depth,
     283             :                                    const Query &Q);
     284             : 
     285     3441138 : unsigned llvm::ComputeNumSignBits(const Value *V, const DataLayout &DL,
     286             :                                   unsigned Depth, AssumptionCache *AC,
     287             :                                   const Instruction *CxtI,
     288             :                                   const DominatorTree *DT) {
     289     3441138 :   return ::ComputeNumSignBits(V, Depth, Query(DL, AC, safeCxtI(V, CxtI), DT));
     290             : }
     291             : 
     292     2314970 : static void computeKnownBitsAddSub(bool Add, const Value *Op0, const Value *Op1,
     293             :                                    bool NSW,
     294             :                                    KnownBits &KnownOut, KnownBits &Known2,
     295             :                                    unsigned Depth, const Query &Q) {
     296             :   unsigned BitWidth = KnownOut.getBitWidth();
     297             : 
     298             :   // If an initial sequence of bits in the result is not needed, the
     299             :   // corresponding bits in the operands are not needed.
     300     4629940 :   KnownBits LHSKnown(BitWidth);
     301     2314970 :   computeKnownBits(Op0, LHSKnown, Depth + 1, Q);
     302     2314970 :   computeKnownBits(Op1, Known2, Depth + 1, Q);
     303             : 
     304     2314970 :   KnownOut = KnownBits::computeForAddSub(Add, NSW, LHSKnown, Known2);
     305     2314970 : }
     306             : 
     307       70352 : static void computeKnownBitsMul(const Value *Op0, const Value *Op1, bool NSW,
     308             :                                 KnownBits &Known, KnownBits &Known2,
     309             :                                 unsigned Depth, const Query &Q) {
     310       70352 :   unsigned BitWidth = Known.getBitWidth();
     311       70352 :   computeKnownBits(Op1, Known, Depth + 1, Q);
     312       70352 :   computeKnownBits(Op0, Known2, Depth + 1, Q);
     313             : 
     314             :   bool isKnownNegative = false;
     315             :   bool isKnownNonNegative = false;
     316             :   // If the multiplication is known not to overflow, compute the sign bit.
     317       70352 :   if (NSW) {
     318       16988 :     if (Op0 == Op1) {
     319             :       // The product of a number with itself is non-negative.
     320             :       isKnownNonNegative = true;
     321             :     } else {
     322             :       bool isKnownNonNegativeOp1 = Known.isNonNegative();
     323             :       bool isKnownNonNegativeOp0 = Known2.isNonNegative();
     324             :       bool isKnownNegativeOp1 = Known.isNegative();
     325             :       bool isKnownNegativeOp0 = Known2.isNegative();
     326             :       // The product of two numbers with the same sign is non-negative.
     327       28655 :       isKnownNonNegative = (isKnownNegativeOp1 && isKnownNegativeOp0) ||
     328       14323 :         (isKnownNonNegativeOp1 && isKnownNonNegativeOp0);
     329             :       // The product of a negative number and a non-negative number is either
     330             :       // negative or zero.
     331             :       if (!isKnownNonNegative)
     332        9305 :         isKnownNegative = (isKnownNegativeOp1 && isKnownNonNegativeOp0 &&
     333       18412 :                            isKnownNonZero(Op0, Depth, Q)) ||
     334        9206 :                           (isKnownNegativeOp0 && isKnownNonNegativeOp1 &&
     335           0 :                            isKnownNonZero(Op1, Depth, Q));
     336             :     }
     337             :   }
     338             : 
     339             :   assert(!Known.hasConflict() && !Known2.hasConflict());
     340             :   // Compute a conservative estimate for high known-0 bits.
     341      211056 :   unsigned LeadZ =  std::max(Known.countMinLeadingZeros() +
     342             :                              Known2.countMinLeadingZeros(),
     343       70352 :                              BitWidth) - BitWidth;
     344       70352 :   LeadZ = std::min(LeadZ, BitWidth);
     345             : 
     346             :   // The result of the bottom bits of an integer multiply can be
     347             :   // inferred by looking at the bottom bits of both operands and
     348             :   // multiplying them together.
     349             :   // We can infer at least the minimum number of known trailing bits
     350             :   // of both operands. Depending on number of trailing zeros, we can
     351             :   // infer more bits, because (a*b) <=> ((a/m) * (b/n)) * (m*n) assuming
     352             :   // a and b are divisible by m and n respectively.
     353             :   // We then calculate how many of those bits are inferrable and set
     354             :   // the output. For example, the i8 mul:
     355             :   //  a = XXXX1100 (12)
     356             :   //  b = XXXX1110 (14)
     357             :   // We know the bottom 3 bits are zero since the first can be divided by
     358             :   // 4 and the second by 2, thus having ((12/4) * (14/2)) * (2*4).
     359             :   // Applying the multiplication to the trimmed arguments gets:
     360             :   //    XX11 (3)
     361             :   //    X111 (7)
     362             :   // -------
     363             :   //    XX11
     364             :   //   XX11
     365             :   //  XX11
     366             :   // XX11
     367             :   // -------
     368             :   // XXXXX01
     369             :   // Which allows us to infer the 2 LSBs. Since we're multiplying the result
     370             :   // by 8, the bottom 3 bits will be 0, so we can infer a total of 5 bits.
     371             :   // The proof for this can be described as:
     372             :   // Pre: (C1 >= 0) && (C1 < (1 << C5)) && (C2 >= 0) && (C2 < (1 << C6)) &&
     373             :   //      (C7 == (1 << (umin(countTrailingZeros(C1), C5) +
     374             :   //                    umin(countTrailingZeros(C2), C6) +
     375             :   //                    umin(C5 - umin(countTrailingZeros(C1), C5),
     376             :   //                         C6 - umin(countTrailingZeros(C2), C6)))) - 1)
     377             :   // %aa = shl i8 %a, C5
     378             :   // %bb = shl i8 %b, C6
     379             :   // %aaa = or i8 %aa, C1
     380             :   // %bbb = or i8 %bb, C2
     381             :   // %mul = mul i8 %aaa, %bbb
     382             :   // %mask = and i8 %mul, C7
     383             :   //   =>
     384             :   // %mask = i8 ((C1*C2)&C7)
     385             :   // Where C5, C6 describe the known bits of %a, %b
     386             :   // C1, C2 describe the known bottom bits of %a, %b.
     387             :   // C7 describes the mask of the known bits of the result.
     388       70352 :   APInt Bottom0 = Known.One;
     389       70352 :   APInt Bottom1 = Known2.One;
     390             : 
     391             :   // How many times we'd be able to divide each argument by 2 (shr by 1).
     392             :   // This gives us the number of trailing zeros on the multiplication result.
     393      140704 :   unsigned TrailBitsKnown0 = (Known.Zero | Known.One).countTrailingOnes();
     394      140704 :   unsigned TrailBitsKnown1 = (Known2.Zero | Known2.One).countTrailingOnes();
     395             :   unsigned TrailZero0 = Known.countMinTrailingZeros();
     396             :   unsigned TrailZero1 = Known2.countMinTrailingZeros();
     397       70352 :   unsigned TrailZ = TrailZero0 + TrailZero1;
     398             : 
     399             :   // Figure out the fewest known-bits operand.
     400      140704 :   unsigned SmallestOperand = std::min(TrailBitsKnown0 - TrailZero0,
     401      211056 :                                       TrailBitsKnown1 - TrailZero1);
     402      140704 :   unsigned ResultBitsKnown = std::min(SmallestOperand + TrailZ, BitWidth);
     403             : 
     404      140704 :   APInt BottomKnown = Bottom0.getLoBits(TrailBitsKnown0) *
     405      211056 :                       Bottom1.getLoBits(TrailBitsKnown1);
     406             : 
     407             :   Known.resetAll();
     408             :   Known.Zero.setHighBits(LeadZ);
     409      281408 :   Known.Zero |= (~BottomKnown).getLoBits(ResultBitsKnown);
     410      140704 :   Known.One |= BottomKnown.getLoBits(ResultBitsKnown);
     411             : 
     412             :   // Only make use of no-wrap flags if we failed to compute the sign bit
     413             :   // directly.  This matters if the multiplication always overflows, in
     414             :   // which case we prefer to follow the result of the direct computation,
     415             :   // though as the program is invoking undefined behaviour we can choose
     416             :   // whatever we like here.
     417       78134 :   if (isKnownNonNegative && !Known.isNegative())
     418             :     Known.makeNonNegative();
     419       62570 :   else if (isKnownNegative && !Known.isNonNegative())
     420             :     Known.makeNegative();
     421       70352 : }
     422             : 
     423      211578 : void llvm::computeKnownBitsFromRangeMetadata(const MDNode &Ranges,
     424             :                                              KnownBits &Known) {
     425             :   unsigned BitWidth = Known.getBitWidth();
     426      211578 :   unsigned NumRanges = Ranges.getNumOperands() / 2;
     427             :   assert(NumRanges >= 1);
     428             : 
     429      211578 :   Known.Zero.setAllBits();
     430      211578 :   Known.One.setAllBits();
     431             : 
     432      634776 :   for (unsigned i = 0; i < NumRanges; ++i) {
     433             :     ConstantInt *Lower =
     434      211599 :         mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 0));
     435             :     ConstantInt *Upper =
     436      211599 :         mdconst::extract<ConstantInt>(Ranges.getOperand(2 * i + 1));
     437      846396 :     ConstantRange Range(Lower->getValue(), Upper->getValue());
     438             : 
     439             :     // The first CommonPrefixBits of all values in Range are equal.
     440             :     unsigned CommonPrefixBits =
     441     1057995 :         (Range.getUnsignedMax() ^ Range.getUnsignedMin()).countLeadingZeros();
     442             : 
     443      211599 :     APInt Mask = APInt::getHighBitsSet(BitWidth, CommonPrefixBits);
     444      634797 :     Known.One &= Range.getUnsignedMax() & Mask;
     445      846396 :     Known.Zero &= ~Range.getUnsignedMax() & Mask;
     446             :   }
     447      211578 : }
     448             : 
     449        1018 : static bool isEphemeralValueOf(const Instruction *I, const Value *E) {
     450        2036 :   SmallVector<const Value *, 16> WorkSet(1, I);
     451             :   SmallPtrSet<const Value *, 32> Visited;
     452             :   SmallPtrSet<const Value *, 16> EphValues;
     453             : 
     454             :   // The instruction defining an assumption's condition itself is always
     455             :   // considered ephemeral to that assumption (even if it has other
     456             :   // non-ephemeral users). See r246696's test case for an example.
     457        2036 :   if (is_contained(I->operands(), E))
     458             :     return true;
     459             : 
     460        3527 :   while (!WorkSet.empty()) {
     461             :     const Value *V = WorkSet.pop_back_val();
     462        3478 :     if (!Visited.insert(V).second)
     463           2 :       continue;
     464             : 
     465             :     // If all uses of this value are ephemeral, then so is this value.
     466        6952 :     if (llvm::all_of(V->users(), [&](const User *U) {
     467        3081 :                                    return EphValues.count(U);
     468        3081 :                                  })) {
     469        2135 :       if (V == E)
     470             :         return true;
     471             : 
     472        1613 :       if (V == I || isSafeToSpeculativelyExecute(V)) {
     473        1454 :        EphValues.insert(V);
     474             :        if (const User *U = dyn_cast<User>(V))
     475        2907 :          for (User::const_op_iterator J = U->op_begin(), JE = U->op_end();
     476        4361 :               J != JE; ++J)
     477        2907 :            WorkSet.push_back(*J);
     478             :       }
     479             :     }
     480             :   }
     481             : 
     482             :   return false;
     483             : }
     484             : 
     485             : // Is this an intrinsic that cannot be speculated but also cannot trap?
     486      208168 : bool llvm::isAssumeLikeIntrinsic(const Instruction *I) {
     487             :   if (const CallInst *CI = dyn_cast<CallInst>(I))
     488             :     if (Function *F = CI->getCalledFunction())
     489      208020 :       switch (F->getIntrinsicID()) {
     490             :       default: break;
     491             :       // FIXME: This list is repeated from NoTTI::getIntrinsicCost.
     492             :       case Intrinsic::assume:
     493             :       case Intrinsic::sideeffect:
     494             :       case Intrinsic::dbg_declare:
     495             :       case Intrinsic::dbg_value:
     496             :       case Intrinsic::invariant_start:
     497             :       case Intrinsic::invariant_end:
     498             :       case Intrinsic::lifetime_start:
     499             :       case Intrinsic::lifetime_end:
     500             :       case Intrinsic::objectsize:
     501             :       case Intrinsic::ptr_annotation:
     502             :       case Intrinsic::var_annotation:
     503             :         return true;
     504             :       }
     505             : 
     506             :   return false;
     507             : }
     508             : 
     509        1699 : bool llvm::isValidAssumeForContext(const Instruction *Inv,
     510             :                                    const Instruction *CxtI,
     511             :                                    const DominatorTree *DT) {
     512             :   // There are two restrictions on the use of an assume:
     513             :   //  1. The assume must dominate the context (or the control flow must
     514             :   //     reach the assume whenever it reaches the context).
     515             :   //  2. The context must not be in the assume's set of ephemeral values
     516             :   //     (otherwise we will use the assume to prove that the condition
     517             :   //     feeding the assume is trivially true, thus causing the removal of
     518             :   //     the assume).
     519             : 
     520        1699 :   if (DT) {
     521        1666 :     if (DT->dominates(Inv, CxtI))
     522             :       return true;
     523          33 :   } else if (Inv->getParent() == CxtI->getParent()->getSinglePredecessor()) {
     524             :     // We don't have a DT, but this trivially dominates.
     525             :     return true;
     526             :   }
     527             : 
     528             :   // With or without a DT, the only remaining case we will check is if the
     529             :   // instructions are in the same BB.  Give up if that is not the case.
     530        1427 :   if (Inv->getParent() != CxtI->getParent())
     531             :     return false;
     532             : 
     533             :   // If we have a dom tree, then we now know that the assume doens't dominate
     534             :   // the other instruction.  If we don't have a dom tree then we can check if
     535             :   // the assume is first in the BB.
     536        1196 :   if (!DT) {
     537             :     // Search forward from the assume until we reach the context (or the end
     538             :     // of the block); the common case is that the assume will come first.
     539             :     for (auto I = std::next(BasicBlock::const_iterator(Inv)),
     540          12 :          IE = Inv->getParent()->end(); I != IE; ++I)
     541          12 :       if (&*I == CxtI)
     542             :         return true;
     543             :   }
     544             : 
     545             :   // The context comes first, but they're both in the same block. Make sure
     546             :   // there is nothing in between that might interrupt the control flow.
     547             :   for (BasicBlock::const_iterator I =
     548             :          std::next(BasicBlock::const_iterator(CxtI)), IE(Inv);
     549        2737 :        I != IE; ++I)
     550        1719 :     if (!isSafeToSpeculativelyExecute(&*I) && !isAssumeLikeIntrinsic(&*I))
     551             :       return false;
     552             : 
     553        1018 :   return !isEphemeralValueOf(Inv, CxtI);
     554             : }
     555             : 
     556    44716325 : static void computeKnownBitsFromAssume(const Value *V, KnownBits &Known,
     557             :                                        unsigned Depth, const Query &Q) {
     558             :   // Use of assumptions is context-sensitive. If we don't have a context, we
     559             :   // cannot use them!
     560    44716325 :   if (!Q.AC || !Q.CxtI)
     561             :     return;
     562             : 
     563             :   unsigned BitWidth = Known.getBitWidth();
     564             : 
     565             :   // Note that the patterns below need to be kept in sync with the code
     566             :   // in AssumptionCache::updateAffectedValues.
     567             : 
     568    80429382 :   for (auto &AssumeVH : Q.AC->assumptionsFor(V)) {
     569        1671 :     if (!AssumeVH)
     570         397 :       continue;
     571             :     CallInst *I = cast<CallInst>(AssumeVH);
     572             :     assert(I->getParent()->getParent() == Q.CxtI->getParent()->getParent() &&
     573             :            "Got assumption for the wrong function!");
     574        1483 :     if (Q.isExcluded(I))
     575          21 :       continue;
     576             : 
     577             :     // Warning: This loop can end up being somewhat performance sensetive.
     578             :     // We're running this loop for once for each value queried resulting in a
     579             :     // runtime of ~O(#assumes * #values).
     580             : 
     581             :     assert(I->getCalledFunction()->getIntrinsicID() == Intrinsic::assume &&
     582             :            "must be an assume intrinsic");
     583             : 
     584        1462 :     Value *Arg = I->getArgOperand(0);
     585             : 
     586        1462 :     if (Arg == V && isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
     587             :       assert(BitWidth == 1 && "assume operand is not i1?");
     588             :       Known.setAllOnes();
     589           8 :       return;
     590             :     }
     591        2943 :     if (match(Arg, m_Not(m_Specific(V))) &&
     592          29 :         isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
     593             :       assert(BitWidth == 1 && "assume operand is not i1?");
     594             :       Known.setAllZero();
     595             :       return;
     596             :     }
     597             : 
     598             :     // The remaining tests are all recursive, so bail out if we hit the limit.
     599        1454 :     if (Depth == MaxDepth)
     600           0 :       continue;
     601             : 
     602             :     Value *A, *B;
     603             :     auto m_V = m_CombineOr(m_Specific(V),
     604             :                            m_CombineOr(m_PtrToInt(m_Specific(V)),
     605             :                            m_BitCast(m_Specific(V))));
     606             : 
     607             :     CmpInst::Predicate Pred;
     608             :     uint64_t C;
     609             :     // assume(v = a)
     610        1835 :     if (match(Arg, m_c_ICmp(Pred, m_V, m_Value(A))) &&
     611        1664 :         Pred == ICmpInst::ICMP_EQ && isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
     612          60 :       KnownBits RHSKnown(BitWidth);
     613          30 :       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
     614          30 :       Known.Zero |= RHSKnown.Zero;
     615          30 :       Known.One  |= RHSKnown.One;
     616             :     // assume(v & b = a)
     617             :     } else if (match(Arg,
     618        1985 :                      m_c_ICmp(Pred, m_c_And(m_V, m_Value(B)), m_Value(A))) &&
     619        2546 :                Pred == ICmpInst::ICMP_EQ &&
     620         561 :                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
     621         118 :       KnownBits RHSKnown(BitWidth);
     622          59 :       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
     623         118 :       KnownBits MaskKnown(BitWidth);
     624          59 :       computeKnownBits(B, MaskKnown, Depth+1, Query(Q, I));
     625             : 
     626             :       // For those bits in the mask that are known to be one, we can propagate
     627             :       // known bits from the RHS to V.
     628         177 :       Known.Zero |= RHSKnown.Zero & MaskKnown.One;
     629         177 :       Known.One  |= RHSKnown.One  & MaskKnown.One;
     630             :     // assume(~(v & b) = a)
     631        1365 :     } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_And(m_V, m_Value(B))),
     632           6 :                                    m_Value(A))) &&
     633        1377 :                Pred == ICmpInst::ICMP_EQ &&
     634           6 :                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
     635           0 :       KnownBits RHSKnown(BitWidth);
     636           0 :       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
     637           0 :       KnownBits MaskKnown(BitWidth);
     638           0 :       computeKnownBits(B, MaskKnown, Depth+1, Query(Q, I));
     639             : 
     640             :       // For those bits in the mask that are known to be one, we can propagate
     641             :       // inverted known bits from the RHS to V.
     642           0 :       Known.Zero |= RHSKnown.One  & MaskKnown.One;
     643           0 :       Known.One  |= RHSKnown.Zero & MaskKnown.One;
     644             :     // assume(v | b = a)
     645             :     } else if (match(Arg,
     646        1383 :                      m_c_ICmp(Pred, m_c_Or(m_V, m_Value(B)), m_Value(A))) &&
     647        1401 :                Pred == ICmpInst::ICMP_EQ &&
     648          18 :                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
     649           4 :       KnownBits RHSKnown(BitWidth);
     650           2 :       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
     651           4 :       KnownBits BKnown(BitWidth);
     652           2 :       computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
     653             : 
     654             :       // For those bits in B that are known to be zero, we can propagate known
     655             :       // bits from the RHS to V.
     656           6 :       Known.Zero |= RHSKnown.Zero & BKnown.Zero;
     657           6 :       Known.One  |= RHSKnown.One  & BKnown.Zero;
     658             :     // assume(~(v | b) = a)
     659        1363 :     } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Or(m_V, m_Value(B))),
     660           7 :                                    m_Value(A))) &&
     661        1377 :                Pred == ICmpInst::ICMP_EQ &&
     662           7 :                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
     663           0 :       KnownBits RHSKnown(BitWidth);
     664           0 :       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
     665           0 :       KnownBits BKnown(BitWidth);
     666           0 :       computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
     667             : 
     668             :       // For those bits in B that are known to be zero, we can propagate
     669             :       // inverted known bits from the RHS to V.
     670           0 :       Known.Zero |= RHSKnown.One  & BKnown.Zero;
     671           0 :       Known.One  |= RHSKnown.Zero & BKnown.Zero;
     672             :     // assume(v ^ b = a)
     673             :     } else if (match(Arg,
     674        1373 :                      m_c_ICmp(Pred, m_c_Xor(m_V, m_Value(B)), m_Value(A))) &&
     675        1383 :                Pred == ICmpInst::ICMP_EQ &&
     676          10 :                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
     677           0 :       KnownBits RHSKnown(BitWidth);
     678           0 :       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
     679           0 :       KnownBits BKnown(BitWidth);
     680           0 :       computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
     681             : 
     682             :       // For those bits in B that are known to be zero, we can propagate known
     683             :       // bits from the RHS to V. For those bits in B that are known to be one,
     684             :       // we can propagate inverted known bits from the RHS to V.
     685           0 :       Known.Zero |= RHSKnown.Zero & BKnown.Zero;
     686           0 :       Known.One  |= RHSKnown.One  & BKnown.Zero;
     687           0 :       Known.Zero |= RHSKnown.One  & BKnown.One;
     688           0 :       Known.One  |= RHSKnown.Zero & BKnown.One;
     689             :     // assume(~(v ^ b) = a)
     690        1363 :     } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_c_Xor(m_V, m_Value(B))),
     691           0 :                                    m_Value(A))) &&
     692        1363 :                Pred == ICmpInst::ICMP_EQ &&
     693           0 :                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
     694           0 :       KnownBits RHSKnown(BitWidth);
     695           0 :       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
     696           0 :       KnownBits BKnown(BitWidth);
     697           0 :       computeKnownBits(B, BKnown, Depth+1, Query(Q, I));
     698             : 
     699             :       // For those bits in B that are known to be zero, we can propagate
     700             :       // inverted known bits from the RHS to V. For those bits in B that are
     701             :       // known to be one, we can propagate known bits from the RHS to V.
     702           0 :       Known.Zero |= RHSKnown.One  & BKnown.Zero;
     703           0 :       Known.One  |= RHSKnown.Zero & BKnown.Zero;
     704           0 :       Known.Zero |= RHSKnown.Zero & BKnown.One;
     705           0 :       Known.One  |= RHSKnown.One  & BKnown.One;
     706             :     // assume(v << c = a)
     707        1363 :     } else if (match(Arg, m_c_ICmp(Pred, m_Shl(m_V, m_ConstantInt(C)),
     708          12 :                                    m_Value(A))) &&
     709          24 :                Pred == ICmpInst::ICMP_EQ &&
     710        1381 :                isValidAssumeForContext(I, Q.CxtI, Q.DT) &&
     711           6 :                C < BitWidth) {
     712          12 :       KnownBits RHSKnown(BitWidth);
     713           6 :       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
     714             :       // For those bits in RHS that are known, we can propagate them to known
     715             :       // bits in V shifted to the right by C.
     716           6 :       RHSKnown.Zero.lshrInPlace(C);
     717           6 :       Known.Zero |= RHSKnown.Zero;
     718           6 :       RHSKnown.One.lshrInPlace(C);
     719           6 :       Known.One  |= RHSKnown.One;
     720             :     // assume(~(v << c) = a)
     721        1357 :     } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_Shl(m_V, m_ConstantInt(C))),
     722           0 :                                    m_Value(A))) &&
     723           0 :                Pred == ICmpInst::ICMP_EQ &&
     724        1357 :                isValidAssumeForContext(I, Q.CxtI, Q.DT) &&
     725           0 :                C < BitWidth) {
     726           0 :       KnownBits RHSKnown(BitWidth);
     727           0 :       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
     728             :       // For those bits in RHS that are known, we can propagate them inverted
     729             :       // to known bits in V shifted to the right by C.
     730           0 :       RHSKnown.One.lshrInPlace(C);
     731           0 :       Known.Zero |= RHSKnown.One;
     732           0 :       RHSKnown.Zero.lshrInPlace(C);
     733           0 :       Known.One  |= RHSKnown.Zero;
     734             :     // assume(v >> c = a)
     735             :     } else if (match(Arg,
     736        1357 :                      m_c_ICmp(Pred, m_Shr(m_V, m_ConstantInt(C)),
     737          12 :                               m_Value(A))) &&
     738          24 :                Pred == ICmpInst::ICMP_EQ &&
     739        1370 :                isValidAssumeForContext(I, Q.CxtI, Q.DT) &&
     740           1 :                C < BitWidth) {
     741           0 :       KnownBits RHSKnown(BitWidth);
     742           0 :       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
     743             :       // For those bits in RHS that are known, we can propagate them to known
     744             :       // bits in V shifted to the right by C.
     745           0 :       Known.Zero |= RHSKnown.Zero << C;
     746           0 :       Known.One  |= RHSKnown.One  << C;
     747             :     // assume(~(v >> c) = a)
     748        1357 :     } else if (match(Arg, m_c_ICmp(Pred, m_Not(m_Shr(m_V, m_ConstantInt(C))),
     749           0 :                                    m_Value(A))) &&
     750           0 :                Pred == ICmpInst::ICMP_EQ &&
     751        1357 :                isValidAssumeForContext(I, Q.CxtI, Q.DT) &&
     752           0 :                C < BitWidth) {
     753           0 :       KnownBits RHSKnown(BitWidth);
     754           0 :       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
     755             :       // For those bits in RHS that are known, we can propagate them inverted
     756             :       // to known bits in V shifted to the right by C.
     757           0 :       Known.Zero |= RHSKnown.One  << C;
     758           0 :       Known.One  |= RHSKnown.Zero << C;
     759             :     // assume(v >=_s c) where c is non-negative
     760        1687 :     } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
     761        1689 :                Pred == ICmpInst::ICMP_SGE &&
     762           2 :                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
     763           2 :       KnownBits RHSKnown(BitWidth);
     764           1 :       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
     765             : 
     766           1 :       if (RHSKnown.isNonNegative()) {
     767             :         // We know that the sign bit is zero.
     768             :         Known.makeNonNegative();
     769             :       }
     770             :     // assume(v >_s c) where c is at least -1.
     771        1685 :     } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
     772        1819 :                Pred == ICmpInst::ICMP_SGT &&
     773         134 :                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
     774          62 :       KnownBits RHSKnown(BitWidth);
     775          31 :       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
     776             : 
     777          35 :       if (RHSKnown.isAllOnes() || RHSKnown.isNonNegative()) {
     778             :         // We know that the sign bit is zero.
     779             :         Known.makeNonNegative();
     780             :       }
     781             :     // assume(v <=_s c) where c is negative
     782        1623 :     } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
     783        1623 :                Pred == ICmpInst::ICMP_SLE &&
     784           0 :                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
     785           0 :       KnownBits RHSKnown(BitWidth);
     786           0 :       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
     787             : 
     788           0 :       if (RHSKnown.isNegative()) {
     789             :         // We know that the sign bit is one.
     790             :         Known.makeNegative();
     791             :       }
     792             :     // assume(v <_s c) where c is non-positive
     793        1623 :     } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
     794        1626 :                Pred == ICmpInst::ICMP_SLT &&
     795           3 :                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
     796           2 :       KnownBits RHSKnown(BitWidth);
     797           1 :       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
     798             : 
     799           2 :       if (RHSKnown.isZero() || RHSKnown.isNegative()) {
     800             :         // We know that the sign bit is one.
     801             :         Known.makeNegative();
     802             :       }
     803             :     // assume(v <=_u c)
     804        1621 :     } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
     805        1621 :                Pred == ICmpInst::ICMP_ULE &&
     806           0 :                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
     807           0 :       KnownBits RHSKnown(BitWidth);
     808           0 :       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
     809             : 
     810             :       // Whatever high bits in c are zero are known to be zero.
     811           0 :       Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros());
     812             :       // assume(v <_u c)
     813        1615 :     } else if (match(Arg, m_ICmp(Pred, m_V, m_Value(A))) &&
     814        1630 :                Pred == ICmpInst::ICMP_ULT &&
     815           9 :                isValidAssumeForContext(I, Q.CxtI, Q.DT)) {
     816           9 :       KnownBits RHSKnown(BitWidth);
     817           6 :       computeKnownBits(A, RHSKnown, Depth+1, Query(Q, I));
     818             : 
     819             :       // If the RHS is known zero, then this assumption must be wrong (nothing
     820             :       // is unsigned less than zero). Signal a conflict and get out of here.
     821           6 :       if (RHSKnown.isZero()) {
     822           3 :         Known.Zero.setAllBits();
     823           3 :         Known.One.setAllBits();
     824           3 :         break;
     825             :       }
     826             : 
     827             :       // Whatever high bits in c are zero are known to be zero (if c is a power
     828             :       // of 2, then one more).
     829           3 :       if (isKnownToBeAPowerOfTwo(A, false, Depth + 1, Query(Q, I)))
     830           4 :         Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros() + 1);
     831             :       else
     832           1 :         Known.Zero.setHighBits(RHSKnown.countMinLeadingZeros());
     833             :     }
     834             :   }
     835             : 
     836             :   // If assumptions conflict with each other or previous known bits, then we
     837             :   // have a logical fallacy. It's possible that the assumption is not reachable,
     838             :   // so this isn't a real bug. On the other hand, the program may have undefined
     839             :   // behavior, or we might have a bug in the compiler. We can't assert/crash, so
     840             :   // clear out the known bits, try to warn the user, and hope for the best.
     841    80426046 :   if (Known.Zero.intersects(Known.One)) {
     842             :     Known.resetAll();
     843             : 
     844           9 :     if (Q.ORE)
     845           8 :       Q.ORE->emit([&]() {
     846           4 :         auto *CxtI = const_cast<Instruction *>(Q.CxtI);
     847           8 :         return OptimizationRemarkAnalysis("value-tracking", "BadAssumption",
     848             :                                           CxtI)
     849             :                << "Detected conflicting code assumptions. Program may "
     850             :                   "have undefined behavior, or compiler may have "
     851           4 :                   "internal error.";
     852             :       });
     853             :   }
     854             : }
     855             : 
     856             : /// Compute known bits from a shift operator, including those with a
     857             : /// non-constant shift amount. Known is the output of this function. Known2 is a
     858             : /// pre-allocated temporary with the same bit width as Known. KZF and KOF are
     859             : /// operator-specific functors that, given the known-zero or known-one bits
     860             : /// respectively, and a shift amount, compute the implied known-zero or
     861             : /// known-one bits of the shift operator's result respectively for that shift
     862             : /// amount. The results from calling KZF and KOF are conservatively combined for
     863             : /// all permitted shift amounts.
     864      368147 : static void computeKnownBitsFromShiftOperator(
     865             :     const Operator *I, KnownBits &Known, KnownBits &Known2,
     866             :     unsigned Depth, const Query &Q,
     867             :     function_ref<APInt(const APInt &, unsigned)> KZF,
     868             :     function_ref<APInt(const APInt &, unsigned)> KOF) {
     869             :   unsigned BitWidth = Known.getBitWidth();
     870             : 
     871      368147 :   if (auto *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
     872      603236 :     unsigned ShiftAmt = SA->getLimitedValue(BitWidth-1);
     873             : 
     874      603236 :     computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
     875      603236 :     Known.Zero = KZF(Known.Zero, ShiftAmt);
     876      603236 :     Known.One  = KOF(Known.One, ShiftAmt);
     877             :     // If the known bits conflict, this must be an overflowing left shift, so
     878             :     // the shift result is poison. We can return anything we want. Choose 0 for
     879             :     // the best folding opportunity.
     880      301618 :     if (Known.hasConflict())
     881             :       Known.setAllZero();
     882             : 
     883             :     return;
     884             :   }
     885             : 
     886       66529 :   computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
     887             : 
     888             :   // If the shift amount could be greater than or equal to the bit-width of the
     889             :   // LHS, the value could be poison, but bail out because the check below is
     890             :   // expensive. TODO: Should we just carry on?
     891      266116 :   if ((~Known.Zero).uge(BitWidth)) {
     892             :     Known.resetAll();
     893             :     return;
     894             :   }
     895             : 
     896             :   // Note: We cannot use Known.Zero.getLimitedValue() here, because if
     897             :   // BitWidth > 64 and any upper bits are known, we'll end up returning the
     898             :   // limit value (which implies all bits are known).
     899       43656 :   uint64_t ShiftAmtKZ = Known.Zero.zextOrTrunc(64).getZExtValue();
     900       43656 :   uint64_t ShiftAmtKO = Known.One.zextOrTrunc(64).getZExtValue();
     901             : 
     902             :   // It would be more-clearly correct to use the two temporaries for this
     903             :   // calculation. Reusing the APInts here to prevent unnecessary allocations.
     904             :   Known.resetAll();
     905             : 
     906             :   // If we know the shifter operand is nonzero, we can sometimes infer more
     907             :   // known bits. However this is expensive to compute, so be lazy about it and
     908             :   // only compute it when absolutely necessary.
     909             :   Optional<bool> ShifterOperandIsNonZero;
     910             : 
     911             :   // Early exit if we can't constrain any well-defined shift amount.
     912       36213 :   if (!(ShiftAmtKZ & (PowerOf2Ceil(BitWidth) - 1)) &&
     913       14385 :       !(ShiftAmtKO & (PowerOf2Ceil(BitWidth) - 1))) {
     914       13354 :     ShifterOperandIsNonZero = isKnownNonZero(I->getOperand(1), Depth + 1, Q);
     915       13354 :     if (!*ShifterOperandIsNonZero)
     916             :       return;
     917             :   }
     918             : 
     919        8492 :   computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
     920             : 
     921        8492 :   Known.Zero.setAllBits();
     922        8492 :   Known.One.setAllBits();
     923      598718 :   for (unsigned ShiftAmt = 0; ShiftAmt < BitWidth; ++ShiftAmt) {
     924             :     // Combine the shifted known input bits only for those shift amounts
     925             :     // compatible with its known constraints.
     926      295113 :     if ((ShiftAmt & ~ShiftAmtKZ) != ShiftAmt)
     927      210365 :       continue;
     928       84748 :     if ((ShiftAmt | ShiftAmtKO) != ShiftAmt)
     929       37424 :       continue;
     930             :     // If we know the shifter is nonzero, we may be able to infer more known
     931             :     // bits. This check is sunk down as far as possible to avoid the expensive
     932             :     // call to isKnownNonZero if the cheaper checks above fail.
     933       47324 :     if (ShiftAmt == 0) {
     934        4283 :       if (!ShifterOperandIsNonZero.hasValue())
     935             :         ShifterOperandIsNonZero =
     936        4265 :             isKnownNonZero(I->getOperand(1), Depth + 1, Q);
     937        4283 :       if (*ShifterOperandIsNonZero)
     938          70 :         continue;
     939             :     }
     940             : 
     941       94508 :     Known.Zero &= KZF(Known2.Zero, ShiftAmt);
     942       94508 :     Known.One  &= KOF(Known2.One, ShiftAmt);
     943             :   }
     944             : 
     945             :   // If the known bits conflict, the result is poison. Return a 0 and hope the
     946             :   // caller can further optimize that.
     947        8492 :   if (Known.hasConflict())
     948             :     Known.setAllZero();
     949             : }
     950             : 
     951    34111575 : static void computeKnownBitsFromOperator(const Operator *I, KnownBits &Known,
     952             :                                          unsigned Depth, const Query &Q) {
     953    34111575 :   unsigned BitWidth = Known.getBitWidth();
     954             : 
     955    68223150 :   KnownBits Known2(Known);
     956    34111575 :   switch (I->getOpcode()) {
     957             :   default: break;
     958             :   case Instruction::Load:
     959    35102467 :     if (MDNode *MD = cast<LoadInst>(I)->getMetadata(LLVMContext::MD_range))
     960      132849 :       computeKnownBitsFromRangeMetadata(*MD, Known);
     961             :     break;
     962      181826 :   case Instruction::And: {
     963             :     // If either the LHS or the RHS are Zero, the result is zero.
     964      363652 :     computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
     965      181826 :     computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
     966             : 
     967             :     // Output known-1 bits are only known if set in both the LHS & RHS.
     968      181826 :     Known.One &= Known2.One;
     969             :     // Output known-0 are known to be clear if zero in either the LHS | RHS.
     970      181826 :     Known.Zero |= Known2.Zero;
     971             : 
     972             :     // and(x, add (x, -1)) is a common idiom that always clears the low bit;
     973             :     // here we handle the more general case of adding any odd number by
     974             :     // matching the form add(x, add(x, y)) where y is odd.
     975             :     // TODO: This could be generalized to clearing any bit set in y where the
     976             :     // following bit is known to be unset in y.
     977      181826 :     Value *Y = nullptr;
     978      464565 :     if (!Known.Zero[0] && !Known.One[0] &&
     979      323182 :         (match(I->getOperand(0), m_Add(m_Specific(I->getOperand(1)),
     980      137554 :                                        m_Value(Y))) ||
     981      319380 :          match(I->getOperand(1), m_Add(m_Specific(I->getOperand(0)),
     982             :                                        m_Value(Y))))) {
     983             :       Known2.resetAll();
     984        3910 :       computeKnownBits(Y, Known2, Depth + 1, Q);
     985        3910 :       if (Known2.countMinTrailingOnes() > 0)
     986             :         Known.Zero.setBit(0);
     987             :     }
     988             :     break;
     989             :   }
     990       56717 :   case Instruction::Or:
     991      113434 :     computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
     992       56717 :     computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
     993             : 
     994             :     // Output known-0 bits are only known if clear in both the LHS & RHS.
     995       56717 :     Known.Zero &= Known2.Zero;
     996             :     // Output known-1 are known to be set if set in either the LHS | RHS.
     997       56717 :     Known.One |= Known2.One;
     998             :     break;
     999       67615 :   case Instruction::Xor: {
    1000      135230 :     computeKnownBits(I->getOperand(1), Known, Depth + 1, Q);
    1001       67615 :     computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
    1002             : 
    1003             :     // Output known-0 bits are known if clear or set in both the LHS & RHS.
    1004      405690 :     APInt KnownZeroOut = (Known.Zero & Known2.Zero) | (Known.One & Known2.One);
    1005             :     // Output known-1 are known to be set if set in only one of the LHS, RHS.
    1006      270460 :     Known.One = (Known.Zero & Known2.One) | (Known.One & Known2.Zero);
    1007             :     Known.Zero = std::move(KnownZeroOut);
    1008             :     break;
    1009             :   }
    1010             :   case Instruction::Mul: {
    1011             :     bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
    1012      140576 :     computeKnownBitsMul(I->getOperand(0), I->getOperand(1), NSW, Known,
    1013             :                         Known2, Depth, Q);
    1014       70288 :     break;
    1015             :   }
    1016       26248 :   case Instruction::UDiv: {
    1017             :     // For the purposes of computing leading zeros we can conservatively
    1018             :     // treat a udiv as a logical right shift by the power of 2 known to
    1019             :     // be less than the denominator.
    1020       52496 :     computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
    1021             :     unsigned LeadZ = Known2.countMinLeadingZeros();
    1022             : 
    1023             :     Known2.resetAll();
    1024       26248 :     computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
    1025             :     unsigned RHSMaxLeadingZeros = Known2.countMaxLeadingZeros();
    1026       26248 :     if (RHSMaxLeadingZeros != BitWidth)
    1027       10380 :       LeadZ = std::min(BitWidth, LeadZ + BitWidth - RHSMaxLeadingZeros - 1);
    1028             : 
    1029       26248 :     Known.Zero.setHighBits(LeadZ);
    1030             :     break;
    1031             :   }
    1032     1318438 :   case Instruction::Select: {
    1033             :     const Value *LHS, *RHS;
    1034     1318438 :     SelectPatternFlavor SPF = matchSelectPattern(I, LHS, RHS).Flavor;
    1035             :     if (SelectPatternResult::isMinOrMax(SPF)) {
    1036       21840 :       computeKnownBits(RHS, Known, Depth + 1, Q);
    1037       21840 :       computeKnownBits(LHS, Known2, Depth + 1, Q);
    1038             :     } else {
    1039     2593196 :       computeKnownBits(I->getOperand(2), Known, Depth + 1, Q);
    1040     1296598 :       computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
    1041             :     }
    1042             : 
    1043             :     unsigned MaxHighOnes = 0;
    1044             :     unsigned MaxHighZeros = 0;
    1045     1318438 :     if (SPF == SPF_SMAX) {
    1046             :       // If both sides are negative, the result is negative.
    1047        4668 :       if (Known.isNegative() && Known2.isNegative())
    1048             :         // We can derive a lower bound on the result by taking the max of the
    1049             :         // leading one bits.
    1050           0 :         MaxHighOnes =
    1051           0 :             std::max(Known.countMinLeadingOnes(), Known2.countMinLeadingOnes());
    1052             :       // If either side is non-negative, the result is non-negative.
    1053        6849 :       else if (Known.isNonNegative() || Known2.isNonNegative())
    1054             :         MaxHighZeros = 1;
    1055     1313812 :     } else if (SPF == SPF_SMIN) {
    1056             :       // If both sides are non-negative, the result is non-negative.
    1057        3710 :       if (Known.isNonNegative() && Known2.isNonNegative())
    1058             :         // We can derive an upper bound on the result by taking the max of the
    1059             :         // leading zero bits.
    1060         116 :         MaxHighZeros = std::max(Known.countMinLeadingZeros(),
    1061         116 :                                 Known2.countMinLeadingZeros());
    1062             :       // If either side is negative, the result is negative.
    1063        5935 :       else if (Known.isNegative() || Known2.isNegative())
    1064             :         MaxHighOnes = 1;
    1065     1310761 :     } else if (SPF == SPF_UMAX) {
    1066             :       // We can derive a lower bound on the result by taking the max of the
    1067             :       // leading one bits.
    1068        6349 :       MaxHighOnes =
    1069       19047 :           std::max(Known.countMinLeadingOnes(), Known2.countMinLeadingOnes());
    1070     1304412 :     } else if (SPF == SPF_UMIN) {
    1071             :       // We can derive an upper bound on the result by taking the max of the
    1072             :       // leading zero bits.
    1073        7814 :       MaxHighZeros =
    1074       23442 :           std::max(Known.countMinLeadingZeros(), Known2.countMinLeadingZeros());
    1075             :     }
    1076             : 
    1077             :     // Only known if known in both the LHS and RHS.
    1078     1318438 :     Known.One &= Known2.One;
    1079     1318438 :     Known.Zero &= Known2.Zero;
    1080     1318438 :     if (MaxHighOnes > 0)
    1081             :       Known.One.setHighBits(MaxHighOnes);
    1082     1318438 :     if (MaxHighZeros > 0)
    1083             :       Known.Zero.setHighBits(MaxHighZeros);
    1084             :     break;
    1085             :   }
    1086             :   case Instruction::FPTrunc:
    1087             :   case Instruction::FPExt:
    1088             :   case Instruction::FPToUI:
    1089             :   case Instruction::FPToSI:
    1090             :   case Instruction::SIToFP:
    1091             :   case Instruction::UIToFP:
    1092             :     break; // Can't work with floating point.
    1093      480098 :   case Instruction::PtrToInt:
    1094             :   case Instruction::IntToPtr:
    1095             :     // Fall through and handle them the same as zext/trunc.
    1096             :     LLVM_FALLTHROUGH;
    1097             :   case Instruction::ZExt:
    1098             :   case Instruction::Trunc: {
    1099      960196 :     Type *SrcTy = I->getOperand(0)->getType();
    1100             : 
    1101             :     unsigned SrcBitWidth;
    1102             :     // Note that we handle pointer operands here because of inttoptr/ptrtoint
    1103             :     // which fall through here.
    1104             :     Type *ScalarTy = SrcTy->getScalarType();
    1105      960196 :     SrcBitWidth = ScalarTy->isPointerTy() ?
    1106      255922 :       Q.DL.getIndexTypeSizeInBits(ScalarTy) :
    1107      224176 :       Q.DL.getTypeSizeInBits(ScalarTy);
    1108             : 
    1109             :     assert(SrcBitWidth && "SrcBitWidth can't be zero");
    1110      480098 :     Known = Known.zextOrTrunc(SrcBitWidth);
    1111      960196 :     computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
    1112      480098 :     Known = Known.zextOrTrunc(BitWidth);
    1113             :     // Any top bits are known to be zero.
    1114      480098 :     if (BitWidth > SrcBitWidth)
    1115      129884 :       Known.Zero.setBitsFrom(SrcBitWidth);
    1116             :     break;
    1117             :   }
    1118      815558 :   case Instruction::BitCast: {
    1119     1631116 :     Type *SrcTy = I->getOperand(0)->getType();
    1120     1590266 :     if ((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) &&
    1121             :         // TODO: For now, not handling conversions like:
    1122             :         // (bitcast i64 %x to <2 x i32>)
    1123      774708 :         !I->getType()->isVectorTy()) {
    1124      771701 :       computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
    1125      771701 :       break;
    1126             :     }
    1127             :     break;
    1128             :   }
    1129       36759 :   case Instruction::SExt: {
    1130             :     // Compute the bits in the result that are not present in the input.
    1131       73518 :     unsigned SrcBitWidth = I->getOperand(0)->getType()->getScalarSizeInBits();
    1132             : 
    1133       36759 :     Known = Known.trunc(SrcBitWidth);
    1134       73518 :     computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
    1135             :     // If the sign bit of the input is known set or clear, then we know the
    1136             :     // top bits of the result.
    1137       36759 :     Known = Known.sext(BitWidth);
    1138       36759 :     break;
    1139             :   }
    1140             :   case Instruction::Shl: {
    1141             :     // (shl X, C1) & C2 == 0   iff   (X & C2 >>u C1) == 0
    1142             :     bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
    1143      324308 :     auto KZF = [NSW](const APInt &KnownZero, unsigned ShiftAmt) {
    1144             :       APInt KZResult = KnownZero << ShiftAmt;
    1145             :       KZResult.setLowBits(ShiftAmt); // Low bits known 0.
    1146             :       // If this shift has "nsw" keyword, then the result is either a poison
    1147             :       // value or has the same sign bit as the first operand.
    1148      173713 :       if (NSW && KnownZero.isSignBitSet())
    1149        6353 :         KZResult.setSignBit();
    1150      162154 :       return KZResult;
    1151      167038 :     };
    1152             : 
    1153      324308 :     auto KOF = [NSW](const APInt &KnownOne, unsigned ShiftAmt) {
    1154             :       APInt KOResult = KnownOne << ShiftAmt;
    1155      173713 :       if (NSW && KnownOne.isSignBitSet())
    1156          20 :         KOResult.setSignBit();
    1157      162154 :       return KOResult;
    1158      167038 :     };
    1159             : 
    1160      167038 :     computeKnownBitsFromShiftOperator(I, Known, Known2, Depth, Q, KZF, KOF);
    1161             :     break;
    1162             :   }
    1163             :   case Instruction::LShr: {
    1164             :     // (lshr X, C1) & C2 == 0   iff  (-1 >> C1) & C2 == 0
    1165      143665 :     auto KZF = [](const APInt &KnownZero, unsigned ShiftAmt) {
    1166      143665 :       APInt KZResult = KnownZero.lshr(ShiftAmt);
    1167             :       // High bits known zero.
    1168             :       KZResult.setHighBits(ShiftAmt);
    1169      143665 :       return KZResult;
    1170             :     };
    1171             : 
    1172             :     auto KOF = [](const APInt &KnownOne, unsigned ShiftAmt) {
    1173             :       return KnownOne.lshr(ShiftAmt);
    1174      143665 :     };
    1175             : 
    1176      158009 :     computeKnownBitsFromShiftOperator(I, Known, Known2, Depth, Q, KZF, KOF);
    1177             :     break;
    1178             :   }
    1179             :   case Instruction::AShr: {
    1180             :     // (ashr X, C1) & C2 == 0   iff  (-1 >> C1) & C2 == 0
    1181             :     auto KZF = [](const APInt &KnownZero, unsigned ShiftAmt) {
    1182             :       return KnownZero.ashr(ShiftAmt);
    1183       43053 :     };
    1184             : 
    1185             :     auto KOF = [](const APInt &KnownOne, unsigned ShiftAmt) {
    1186             :       return KnownOne.ashr(ShiftAmt);
    1187       43053 :     };
    1188             : 
    1189       43100 :     computeKnownBitsFromShiftOperator(I, Known, Known2, Depth, Q, KZF, KOF);
    1190             :     break;
    1191             :   }
    1192             :   case Instruction::Sub: {
    1193             :     bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
    1194      357528 :     computeKnownBitsAddSub(false, I->getOperand(0), I->getOperand(1), NSW,
    1195             :                            Known, Known2, Depth, Q);
    1196      178764 :     break;
    1197             :   }
    1198             :   case Instruction::Add: {
    1199             :     bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
    1200     4268878 :     computeKnownBitsAddSub(true, I->getOperand(0), I->getOperand(1), NSW,
    1201             :                            Known, Known2, Depth, Q);
    1202     2134439 :     break;
    1203             :   }
    1204        1315 :   case Instruction::SRem:
    1205        1315 :     if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
    1206         680 :       APInt RA = Rem->getValue().abs();
    1207         680 :       if (RA.isPowerOf2()) {
    1208         426 :         APInt LowBits = RA - 1;
    1209         852 :         computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
    1210             : 
    1211             :         // The low bits of the first operand are unchanged by the srem.
    1212         852 :         Known.Zero = Known2.Zero & LowBits;
    1213         852 :         Known.One = Known2.One & LowBits;
    1214             : 
    1215             :         // If the first operand is non-negative or has all low bits zero, then
    1216             :         // the upper bits are all zero.
    1217         769 :         if (Known2.isNonNegative() || LowBits.isSubsetOf(Known2.Zero))
    1218         182 :           Known.Zero |= ~LowBits;
    1219             : 
    1220             :         // If the first operand is negative and not all low bits are zero, then
    1221             :         // the upper bits are all one.
    1222         442 :         if (Known2.isNegative() && LowBits.intersects(Known2.One))
    1223          28 :           Known.One |= ~LowBits;
    1224             : 
    1225             :         assert((Known.Zero & Known.One) == 0 && "Bits known to be one AND zero?");
    1226             :         break;
    1227             :       }
    1228             :     }
    1229             : 
    1230             :     // The sign bit is the LHS's sign bit, except when the result of the
    1231             :     // remainder is zero.
    1232        1778 :     computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
    1233             :     // If it's known zero, our sign bit is also zero.
    1234         889 :     if (Known2.isNonNegative())
    1235             :       Known.makeNonNegative();
    1236             : 
    1237             :     break;
    1238       12864 :   case Instruction::URem: {
    1239       12864 :     if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
    1240             :       const APInt &RA = Rem->getValue();
    1241        5547 :       if (RA.isPowerOf2()) {
    1242        1142 :         APInt LowBits = (RA - 1);
    1243        2284 :         computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
    1244        3426 :         Known.Zero |= ~LowBits;
    1245        1142 :         Known.One &= LowBits;
    1246             :         break;
    1247             :       }
    1248             :     }
    1249             : 
    1250             :     // Since the result is less than or equal to either operand, any leading
    1251             :     // zero bits in either operand must also exist in the result.
    1252       23444 :     computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
    1253       11722 :     computeKnownBits(I->getOperand(1), Known2, Depth + 1, Q);
    1254             : 
    1255             :     unsigned Leaders =
    1256       35166 :         std::max(Known.countMinLeadingZeros(), Known2.countMinLeadingZeros());
    1257             :     Known.resetAll();
    1258             :     Known.Zero.setHighBits(Leaders);
    1259             :     break;
    1260             :   }
    1261             : 
    1262             :   case Instruction::Alloca: {
    1263             :     const AllocaInst *AI = cast<AllocaInst>(I);
    1264             :     unsigned Align = AI->getAlignment();
    1265      416137 :     if (Align == 0)
    1266        1327 :       Align = Q.DL.getABITypeAlignment(AI->getAllocatedType());
    1267             : 
    1268      416137 :     if (Align > 0)
    1269      832274 :       Known.Zero.setLowBits(countTrailingZeros(Align));
    1270             :     break;
    1271             :   }
    1272     7255049 :   case Instruction::GetElementPtr: {
    1273             :     // Analyze all of the subscripts of this getelementptr instruction
    1274             :     // to determine if we can prove known low zero bits.
    1275    14510098 :     KnownBits LocalKnown(BitWidth);
    1276    14510098 :     computeKnownBits(I->getOperand(0), LocalKnown, Depth + 1, Q);
    1277     7255049 :     unsigned TrailZ = LocalKnown.countMinTrailingZeros();
    1278             : 
    1279     7255049 :     gep_type_iterator GTI = gep_type_begin(I);
    1280    37120939 :     for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
    1281             :       Value *Index = I->getOperand(i);
    1282     1223586 :       if (StructType *STy = GTI.getStructTypeOrNull()) {
    1283             :         // Handle struct member offset arithmetic.
    1284             : 
    1285             :         // Handle case when index is vector zeroinitializer
    1286             :         Constant *CIndex = cast<Constant>(Index);
    1287     1223586 :         if (CIndex->isZeroValue())
    1288      835435 :           continue;
    1289             : 
    1290      776302 :         if (CIndex->getType()->isVectorTy())
    1291           3 :           Index = CIndex->getSplatValue();
    1292             : 
    1293      388151 :         unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
    1294      388151 :         const StructLayout *SL = Q.DL.getStructLayout(STy);
    1295             :         uint64_t Offset = SL->getElementOffset(Idx);
    1296      388151 :         TrailZ = std::min<unsigned>(TrailZ,
    1297      776302 :                                     countTrailingZeros(Offset));
    1298             :       } else {
    1299             :         // Handle array index arithmetic.
    1300    13709359 :         Type *IndexedTy = GTI.getIndexedType();
    1301    13709359 :         if (!IndexedTy->isSized()) {
    1302           0 :           TrailZ = 0;
    1303           0 :           break;
    1304             :         }
    1305    13709359 :         unsigned GEPOpiBits = Index->getType()->getScalarSizeInBits();
    1306    13709359 :         uint64_t TypeSize = Q.DL.getTypeAllocSize(IndexedTy);
    1307    27418718 :         LocalKnown.Zero = LocalKnown.One = APInt(GEPOpiBits, 0);
    1308    13709359 :         computeKnownBits(Index, LocalKnown, Depth + 1, Q);
    1309    13709359 :         TrailZ = std::min(TrailZ,
    1310    41128077 :                           unsigned(countTrailingZeros(TypeSize) +
    1311             :                                    LocalKnown.countMinTrailingZeros()));
    1312             :       }
    1313             :     }
    1314             : 
    1315     7255049 :     Known.Zero.setLowBits(TrailZ);
    1316             :     break;
    1317             :   }
    1318             :   case Instruction::PHI: {
    1319             :     const PHINode *P = cast<PHINode>(I);
    1320             :     // Handle the case of a simple two-predecessor recurrence PHI.
    1321             :     // There's a lot more that could theoretically be done here, but
    1322             :     // this is sufficient to catch some interesting cases.
    1323     1590673 :     if (P->getNumIncomingValues() == 2) {
    1324     5823230 :       for (unsigned i = 0; i != 2; ++i) {
    1325             :         Value *L = P->getIncomingValue(i);
    1326     2748088 :         Value *R = P->getIncomingValue(!i);
    1327             :         Operator *LU = dyn_cast<Operator>(L);
    1328      566313 :         if (!LU)
    1329      566313 :           continue;
    1330             :         unsigned Opcode = LU->getOpcode();
    1331             :         // Check for operations that have the property that if
    1332             :         // both their operands have low zero bits, the result
    1333             :         // will have low zero bits.
    1334     4363550 :         if (Opcode == Instruction::Add ||
    1335     2181775 :             Opcode == Instruction::Sub ||
    1336     1623972 :             Opcode == Instruction::And ||
    1337     1612258 :             Opcode == Instruction::Or ||
    1338             :             Opcode == Instruction::Mul) {
    1339      573822 :           Value *LL = LU->getOperand(0);
    1340             :           Value *LR = LU->getOperand(1);
    1341             :           // Find a recurrence.
    1342      573822 :           if (LL == I)
    1343             :             L = LR;
    1344      177596 :           else if (LR == I)
    1345             :             L = LL;
    1346             :           else
    1347             :             break;
    1348             :           // Ok, we have a PHI of the form L op= R. Check for low
    1349             :           // zero bits.
    1350      404277 :           computeKnownBits(R, Known2, Depth + 1, Q);
    1351             : 
    1352             :           // We need to take the minimum number of known bits
    1353      808554 :           KnownBits Known3(Known);
    1354      404277 :           computeKnownBits(L, Known3, Depth + 1, Q);
    1355             : 
    1356     1617108 :           Known.Zero.setLowBits(std::min(Known2.countMinTrailingZeros(),
    1357      808554 :                                          Known3.countMinTrailingZeros()));
    1358             : 
    1359             :           auto *OverflowOp = dyn_cast<OverflowingBinaryOperator>(LU);
    1360      402539 :           if (OverflowOp && OverflowOp->hasNoSignedWrap()) {
    1361             :             // If initial value of recurrence is nonnegative, and we are adding
    1362             :             // a nonnegative number with nsw, the result can only be nonnegative
    1363             :             // or poison value regardless of the number of times we execute the
    1364             :             // add in phi recurrence. If initial value is negative and we are
    1365             :             // adding a negative number with nsw, the result can only be
    1366             :             // negative or poison value. Similar arguments apply to sub and mul.
    1367             :             //
    1368             :             // (add non-negative, non-negative) --> non-negative
    1369             :             // (add negative, negative) --> negative
    1370      154593 :             if (Opcode == Instruction::Add) {
    1371      294244 :               if (Known2.isNonNegative() && Known3.isNonNegative())
    1372             :                 Known.makeNonNegative();
    1373       24944 :               else if (Known2.isNegative() && Known3.isNegative())
    1374             :                 Known.makeNegative();
    1375             :             }
    1376             : 
    1377             :             // (sub nsw non-negative, negative) --> non-negative
    1378             :             // (sub nsw negative, non-negative) --> negative
    1379         691 :             else if (Opcode == Instruction::Sub && LL == I) {
    1380         610 :               if (Known2.isNonNegative() && Known3.isNegative())
    1381             :                 Known.makeNonNegative();
    1382         576 :               else if (Known2.isNegative() && Known3.isNonNegative())
    1383             :                 Known.makeNegative();
    1384             :             }
    1385             : 
    1386             :             // (mul nsw non-negative, non-negative) --> non-negative
    1387         249 :             else if (Opcode == Instruction::Mul && Known2.isNonNegative() &&
    1388             :                      Known3.isNonNegative())
    1389             :               Known.makeNonNegative();
    1390             :           }
    1391             : 
    1392             :           break;
    1393             :         }
    1394             :       }
    1395             :     }
    1396             : 
    1397             :     // Unreachable blocks may have zero-operand PHI nodes.
    1398     1590673 :     if (P->getNumIncomingValues() == 0)
    1399             :       break;
    1400             : 
    1401             :     // Otherwise take the unions of the known bit sets of the operands,
    1402             :     // taking conservative care to avoid excessive recursion.
    1403     4009189 :     if (Depth < MaxDepth - 1 && !Known.Zero && !Known.One) {
    1404             :       // Skip if every incoming value references to ourself.
    1405     1131611 :       if (dyn_cast_or_null<UndefValue>(P->hasConstantValue()))
    1406             :         break;
    1407             : 
    1408     1131610 :       Known.Zero.setAllBits();
    1409     1131610 :       Known.One.setAllBits();
    1410     1878848 :       for (Value *IncValue : P->incoming_values()) {
    1411             :         // Skip direct self references.
    1412     1472348 :         if (IncValue == P) continue;
    1413             : 
    1414     1472319 :         Known2 = KnownBits(BitWidth);
    1415             :         // Recurse, but cap the recursion to one level, because we don't
    1416             :         // want to waste time spinning around in loops.
    1417     1472319 :         computeKnownBits(IncValue, Known2, MaxDepth - 1, Q);
    1418             :         Known.Zero &= Known2.Zero;
    1419             :         Known.One &= Known2.One;
    1420             :         // If all bits have been ruled out, there's no need to check
    1421             :         // more operands.
    1422     2589004 :         if (!Known.Zero && !Known.One)
    1423             :           break;
    1424             :       }
    1425             :     }
    1426             :     break;
    1427             :   }
    1428             :   case Instruction::Call:
    1429             :   case Instruction::Invoke:
    1430             :     // If range metadata is attached to this call, set known bits from that,
    1431             :     // and then intersect with known bits based on other properties of the
    1432             :     // function.
    1433      446933 :     if (MDNode *MD = cast<Instruction>(I)->getMetadata(LLVMContext::MD_range))
    1434       76884 :       computeKnownBitsFromRangeMetadata(*MD, Known);
    1435      505324 :     if (const Value *RV = ImmutableCallSite(I).getReturnedArgOperand()) {
    1436         193 :       computeKnownBits(RV, Known2, Depth + 1, Q);
    1437         193 :       Known.Zero |= Known2.Zero;
    1438         193 :       Known.One |= Known2.One;
    1439             :     }
    1440             :     if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
    1441      128047 :       switch (II->getIntrinsicID()) {
    1442             :       default: break;
    1443         415 :       case Intrinsic::bitreverse:
    1444         830 :         computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
    1445         830 :         Known.Zero |= Known2.Zero.reverseBits();
    1446         830 :         Known.One |= Known2.One.reverseBits();
    1447         415 :         break;
    1448        2846 :       case Intrinsic::bswap:
    1449        5692 :         computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
    1450        5692 :         Known.Zero |= Known2.Zero.byteSwap();
    1451        5692 :         Known.One |= Known2.One.byteSwap();
    1452        2846 :         break;
    1453       13345 :       case Intrinsic::ctlz: {
    1454       26690 :         computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
    1455             :         // If we have a known 1, its position is our upper bound.
    1456       13345 :         unsigned PossibleLZ = Known2.One.countLeadingZeros();
    1457             :         // If this call is undefined for 0, the result will be less than 2^n.
    1458       26690 :         if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
    1459       25230 :           PossibleLZ = std::min(PossibleLZ, BitWidth - 1);
    1460       26690 :         unsigned LowBits = Log2_32(PossibleLZ)+1;
    1461       13345 :         Known.Zero.setBitsFrom(LowBits);
    1462             :         break;
    1463             :       }
    1464       21000 :       case Intrinsic::cttz: {
    1465       42000 :         computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
    1466             :         // If we have a known 1, its position is our upper bound.
    1467       21000 :         unsigned PossibleTZ = Known2.One.countTrailingZeros();
    1468             :         // If this call is undefined for 0, the result will be less than 2^n.
    1469       42000 :         if (II->getArgOperand(1) == ConstantInt::getTrue(II->getContext()))
    1470       41260 :           PossibleTZ = std::min(PossibleTZ, BitWidth - 1);
    1471       42000 :         unsigned LowBits = Log2_32(PossibleTZ)+1;
    1472       21000 :         Known.Zero.setBitsFrom(LowBits);
    1473             :         break;
    1474             :       }
    1475        2767 :       case Intrinsic::ctpop: {
    1476        5534 :         computeKnownBits(I->getOperand(0), Known2, Depth + 1, Q);
    1477             :         // We can bound the space the count needs.  Also, bits known to be zero
    1478             :         // can't contribute to the population.
    1479             :         unsigned BitsPossiblySet = Known2.countMaxPopulation();
    1480        2767 :         unsigned LowBits = Log2_32(BitsPossiblySet)+1;
    1481        2767 :         Known.Zero.setBitsFrom(LowBits);
    1482             :         // TODO: we could bound KnownOne using the lower bound on the number
    1483             :         // of bits which might be set provided by popcnt KnownOne2.
    1484             :         break;
    1485             :       }
    1486           8 :       case Intrinsic::x86_sse42_crc32_64_64:
    1487           8 :         Known.Zero.setBitsFrom(32);
    1488             :         break;
    1489             :       }
    1490             :     }
    1491             :     break;
    1492       16112 :   case Instruction::ExtractElement:
    1493             :     // Look through extract element. At the moment we keep this simple and skip
    1494             :     // tracking the specific element. But at least we might find information
    1495             :     // valid for all elements of the vector (for example if vector is sign
    1496             :     // extended, shifted, etc).
    1497       32224 :     computeKnownBits(I->getOperand(0), Known, Depth + 1, Q);
    1498       16112 :     break;
    1499      329620 :   case Instruction::ExtractValue:
    1500      329620 :     if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->getOperand(0))) {
    1501             :       const ExtractValueInst *EVI = cast<ExtractValueInst>(I);
    1502       31303 :       if (EVI->getNumIndices() != 1) break;
    1503       31303 :       if (EVI->getIndices()[0] == 0) {
    1504             :         switch (II->getIntrinsicID()) {
    1505             :         default: break;
    1506         899 :         case Intrinsic::uadd_with_overflow:
    1507             :         case Intrinsic::sadd_with_overflow:
    1508         899 :           computeKnownBitsAddSub(true, II->getArgOperand(0),
    1509         899 :                                  II->getArgOperand(1), false, Known, Known2,
    1510             :                                  Depth, Q);
    1511         899 :           break;
    1512         868 :         case Intrinsic::usub_with_overflow:
    1513             :         case Intrinsic::ssub_with_overflow:
    1514         868 :           computeKnownBitsAddSub(false, II->getArgOperand(0),
    1515         868 :                                  II->getArgOperand(1), false, Known, Known2,
    1516             :                                  Depth, Q);
    1517         868 :           break;
    1518          64 :         case Intrinsic::umul_with_overflow:
    1519             :         case Intrinsic::smul_with_overflow:
    1520         128 :           computeKnownBitsMul(II->getArgOperand(0), II->getArgOperand(1), false,
    1521             :                               Known, Known2, Depth, Q);
    1522          64 :           break;
    1523             :         }
    1524             :       }
    1525             :     }
    1526             :   }
    1527    34111575 : }
    1528             : 
    1529             : /// Determine which bits of V are known to be either zero or one and return
    1530             : /// them.
    1531    24505594 : KnownBits computeKnownBits(const Value *V, unsigned Depth, const Query &Q) {
    1532    24505594 :   KnownBits Known(getBitWidth(V->getType(), Q.DL));
    1533    24505594 :   computeKnownBits(V, Known, Depth, Q);
    1534    24505594 :   return Known;
    1535             : }
    1536             : 
    1537             : /// Determine which bits of V are known to be either zero or one and return
    1538             : /// them in the Known bit set.
    1539             : ///
    1540             : /// NOTE: we cannot consider 'undef' to be "IsZero" here.  The problem is that
    1541             : /// we cannot optimize based on the assumption that it is zero without changing
    1542             : /// it to be an explicit zero.  If we don't change it to zero, other code could
    1543             : /// optimized based on the contradictory assumption that it is non-zero.
    1544             : /// Because instcombine aggressively folds operations with undef args anyway,
    1545             : /// this won't lose us code quality.
    1546             : ///
    1547             : /// This function is defined on values with integer type, values with pointer
    1548             : /// type, and vectors of integers.  In the case
    1549             : /// where V is a vector, known zero, and known one values are the
    1550             : /// same width as the vector element, and the bit is set only if it is true
    1551             : /// for all of the elements in the vector.
    1552    71899748 : void computeKnownBits(const Value *V, KnownBits &Known, unsigned Depth,
    1553             :                       const Query &Q) {
    1554             :   assert(V && "No Value?");
    1555             :   assert(Depth <= MaxDepth && "Limit Search Depth");
    1556             :   unsigned BitWidth = Known.getBitWidth();
    1557             : 
    1558             :   assert((V->getType()->isIntOrIntVectorTy(BitWidth) ||
    1559             :           V->getType()->isPtrOrPtrVectorTy()) &&
    1560             :          "Not integer or pointer type!");
    1561             : 
    1562    71899748 :   Type *ScalarTy = V->getType()->getScalarType();
    1563             :   unsigned ExpectedWidth = ScalarTy->isPointerTy() ?
    1564    71899748 :     Q.DL.getIndexTypeSizeInBits(ScalarTy) : Q.DL.getTypeSizeInBits(ScalarTy);
    1565             :   assert(ExpectedWidth == BitWidth && "V and Known should have same BitWidth");
    1566             :   (void)BitWidth;
    1567             :   (void)ExpectedWidth;
    1568             : 
    1569             :   const APInt *C;
    1570   143799496 :   if (match(V, m_APInt(C))) {
    1571             :     // We know all of the bits for a scalar constant or a splat vector constant!
    1572    26313149 :     Known.One = *C;
    1573    26313149 :     Known.Zero = ~Known.One;
    1574    53496572 :     return;
    1575             :   }
    1576             :   // Null and aggregate-zero are all-zeros.
    1577    45586599 :   if (isa<ConstantPointerNull>(V) || isa<ConstantAggregateZero>(V)) {
    1578             :     Known.setAllZero();
    1579             :     return;
    1580             :   }
    1581             :   // Handle a constant vector by taking the intersection of the known bits of
    1582             :   // each element.
    1583             :   if (const ConstantDataSequential *CDS = dyn_cast<ConstantDataSequential>(V)) {
    1584             :     // We know that CDS must be a vector of integers. Take the intersection of
    1585             :     // each element.
    1586       12304 :     Known.Zero.setAllBits(); Known.One.setAllBits();
    1587       52925 :     for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
    1588       40621 :       APInt Elt = CDS->getElementAsAPInt(i);
    1589       81242 :       Known.Zero &= ~Elt;
    1590             :       Known.One &= Elt;
    1591             :     }
    1592             :     return;
    1593             :   }
    1594             : 
    1595             :   if (const auto *CV = dyn_cast<ConstantVector>(V)) {
    1596             :     // We know that CV must be a vector of integers. Take the intersection of
    1597             :     // each element.
    1598         534 :     Known.Zero.setAllBits(); Known.One.setAllBits();
    1599        2146 :     for (unsigned i = 0, e = CV->getNumOperands(); i != e; ++i) {
    1600        1235 :       Constant *Element = CV->getAggregateElement(i);
    1601             :       auto *ElementCI = dyn_cast_or_null<ConstantInt>(Element);
    1602             :       if (!ElementCI) {
    1603             :         Known.resetAll();
    1604             :         return;
    1605             :       }
    1606             :       const APInt &Elt = ElementCI->getValue();
    1607        1612 :       Known.Zero &= ~Elt;
    1608             :       Known.One &= Elt;
    1609             :     }
    1610             :     return;
    1611             :   }
    1612             : 
    1613             :   // Start out not knowing anything.
    1614             :   Known.resetAll();
    1615             : 
    1616             :   // We can't imply anything about undefs.
    1617    45455810 :   if (isa<UndefValue>(V))
    1618             :     return;
    1619             : 
    1620             :   // There's no point in looking through other users of ConstantData for
    1621             :   // assumptions.  Confirm that we've handled them all.
    1622             :   assert(!isa<ConstantData>(V) && "Unhandled constant data!");
    1623             : 
    1624             :   // Limit search depth.
    1625             :   // All recursive calls that increase depth must come after this.
    1626    45450825 :   if (Depth == MaxDepth)
    1627             :     return;
    1628             : 
    1629             :   // A weak GlobalAlias is totally unknown. A non-weak GlobalAlias has
    1630             :   // the bits of its aliasee.
    1631             :   if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
    1632             :     if (!GA->isInterposable())
    1633          32 :       computeKnownBits(GA->getAliasee(), Known, Depth + 1, Q);
    1634             :     return;
    1635             :   }
    1636             : 
    1637             :   if (const Operator *I = dyn_cast<Operator>(V))
    1638    34111575 :     computeKnownBitsFromOperator(I, Known, Depth, Q);
    1639             : 
    1640             :   // Aligned pointers have trailing zeros - refine Known.Zero set
    1641    89432650 :   if (V->getType()->isPointerTy()) {
    1642    19875725 :     unsigned Align = V->getPointerAlignment(Q.DL);
    1643    19875725 :     if (Align)
    1644     9729070 :       Known.Zero.setLowBits(countTrailingZeros(Align));
    1645             :   }
    1646             : 
    1647             :   // computeKnownBitsFromAssume strictly refines Known.
    1648             :   // Therefore, we run them after computeKnownBitsFromOperator.
    1649             : 
    1650             :   // Check whether a nearby assume intrinsic can determine some known bits.
    1651    44716325 :   computeKnownBitsFromAssume(V, Known, Depth, Q);
    1652             : 
    1653             :   assert((Known.Zero & Known.One) == 0 && "Bits known to be one AND zero?");
    1654             : }
    1655             : 
    1656             : /// Return true if the given value is known to have exactly one
    1657             : /// bit set when defined. For vectors return true if every element is known to
    1658             : /// be a power of two when defined. Supports values with integer or pointer
    1659             : /// types and vectors of integers.
    1660        5980 : bool isKnownToBeAPowerOfTwo(const Value *V, bool OrZero, unsigned Depth,
    1661             :                             const Query &Q) {
    1662             :   assert(Depth <= MaxDepth && "Limit Search Depth");
    1663             : 
    1664             :   // Attempt to match against constants.
    1665        9502 :   if (OrZero && match(V, m_Power2OrZero()))
    1666             :       return true;
    1667        5472 :   if (match(V, m_Power2()))
    1668             :       return true;
    1669             : 
    1670             :   // 1 << X is clearly a power of two if the one is not shifted off the end.  If
    1671             :   // it is shifted off the end then the result is undefined.
    1672        5428 :   if (match(V, m_Shl(m_One(), m_Value())))
    1673             :     return true;
    1674             : 
    1675             :   // (signmask) >>l X is clearly a power of two if the one is not shifted off
    1676             :   // the bottom.  If it is shifted off the bottom then the result is undefined.
    1677        5416 :   if (match(V, m_LShr(m_SignMask(), m_Value())))
    1678             :     return true;
    1679             : 
    1680             :   // The remaining tests are all recursive, so bail out if we hit the limit.
    1681        5412 :   if (Depth++ == MaxDepth)
    1682             :     return false;
    1683             : 
    1684        5404 :   Value *X = nullptr, *Y = nullptr;
    1685             :   // A shift left or a logical shift right of a power of two is a power of two
    1686             :   // or zero.
    1687       16822 :   if (OrZero && (match(V, m_Shl(m_Value(X), m_Value())) ||
    1688        8409 :                  match(V, m_LShr(m_Value(X), m_Value()))))
    1689          46 :     return isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q);
    1690             : 
    1691             :   if (const ZExtInst *ZI = dyn_cast<ZExtInst>(V))
    1692          16 :     return isKnownToBeAPowerOfTwo(ZI->getOperand(0), OrZero, Depth, Q);
    1693             : 
    1694             :   if (const SelectInst *SI = dyn_cast<SelectInst>(V))
    1695          69 :     return isKnownToBeAPowerOfTwo(SI->getTrueValue(), OrZero, Depth, Q) &&
    1696           3 :            isKnownToBeAPowerOfTwo(SI->getFalseValue(), OrZero, Depth, Q);
    1697             : 
    1698        8242 :   if (OrZero && match(V, m_And(m_Value(X), m_Value(Y)))) {
    1699             :     // A power of two and'd with anything is a power of two or zero.
    1700           2 :     if (isKnownToBeAPowerOfTwo(X, /*OrZero*/ true, Depth, Q) ||
    1701           1 :         isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, Depth, Q))
    1702             :       return true;
    1703             :     // X & (-X) is always a power of two or zero.
    1704           3 :     if (match(X, m_Neg(m_Specific(Y))) || match(Y, m_Neg(m_Specific(X))))
    1705             :       return true;
    1706             :     return false;
    1707             :   }
    1708             : 
    1709             :   // Adding a power-of-two or zero to the same power-of-two or zero yields
    1710             :   // either the original power-of-two, a larger power-of-two or zero.
    1711       10632 :   if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
    1712             :     const OverflowingBinaryOperator *VOBO = cast<OverflowingBinaryOperator>(V);
    1713         234 :     if (OrZero || VOBO->hasNoUnsignedWrap() || VOBO->hasNoSignedWrap()) {
    1714         879 :       if (match(X, m_And(m_Specific(Y), m_Value())) ||
    1715         439 :           match(X, m_And(m_Value(), m_Specific(Y))))
    1716           2 :         if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth, Q))
    1717           3 :           return true;
    1718         872 :       if (match(Y, m_And(m_Specific(X), m_Value())) ||
    1719         436 :           match(Y, m_And(m_Value(), m_Specific(X))))
    1720           0 :         if (isKnownToBeAPowerOfTwo(X, OrZero, Depth, Q))
    1721             :           return true;
    1722             : 
    1723         218 :       unsigned BitWidth = V->getType()->getScalarSizeInBits();
    1724         435 :       KnownBits LHSBits(BitWidth);
    1725         218 :       computeKnownBits(X, LHSBits, Depth, Q);
    1726             : 
    1727         435 :       KnownBits RHSBits(BitWidth);
    1728         218 :       computeKnownBits(Y, RHSBits, Depth, Q);
    1729             :       // If i8 V is a power of two or zero:
    1730             :       //  ZeroBits: 1 1 1 0 1 1 1 1
    1731             :       // ~ZeroBits: 0 0 0 1 0 0 0 0
    1732         872 :       if ((~(LHSBits.Zero & RHSBits.Zero)).isPowerOf2())
    1733             :         // If OrZero isn't set, we cannot give back a zero result.
    1734             :         // Make sure either the LHS or RHS has a bit set.
    1735           1 :         if (OrZero || RHSBits.One.getBoolValue() || LHSBits.One.getBoolValue())
    1736           1 :           return true;
    1737             :     }
    1738             :   }
    1739             : 
    1740             :   // An exact divide or right shift can only shift off zero bits, so the result
    1741             :   // is a power of two only if the first operand is a power of two and not
    1742             :   // copying a sign bit (sdiv int_min, 2).
    1743       10626 :   if (match(V, m_Exact(m_LShr(m_Value(), m_Value()))) ||
    1744             :       match(V, m_Exact(m_UDiv(m_Value(), m_Value())))) {
    1745           0 :     return isKnownToBeAPowerOfTwo(cast<Operator>(V)->getOperand(0), OrZero,
    1746           0 :                                   Depth, Q);
    1747             :   }
    1748             : 
    1749             :   return false;
    1750             : }
    1751             : 
    1752             : /// \brief Test whether a GEP's result is known to be non-null.
    1753             : ///
    1754             : /// Uses properties inherent in a GEP to try to determine whether it is known
    1755             : /// to be non-null.
    1756             : ///
    1757             : /// Currently this routine does not support vector GEPs.
    1758      212764 : static bool isGEPKnownNonNull(const GEPOperator *GEP, unsigned Depth,
    1759             :                               const Query &Q) {
    1760      423782 :   if (!GEP->isInBounds() || GEP->getPointerAddressSpace() != 0)
    1761             :     return false;
    1762             : 
    1763             :   // FIXME: Support vector-GEPs.
    1764             :   assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
    1765             : 
    1766             :   // If the base pointer is non-null, we cannot walk to a null address with an
    1767             :   // inbounds GEP in address space zero.
    1768      210646 :   if (isKnownNonZero(GEP->getPointerOperand(), Depth, Q))
    1769             :     return true;
    1770             : 
    1771             :   // Walk the GEP operands and see if any operand introduces a non-zero offset.
    1772             :   // If so, then the GEP cannot produce a null pointer, as doing so would
    1773             :   // inherently violate the inbounds contract within address space zero.
    1774      273891 :   for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
    1775      273891 :        GTI != GTE; ++GTI) {
    1776             :     // Struct types are easy -- they must always be indexed by a constant.
    1777      106133 :     if (StructType *STy = GTI.getStructTypeOrNull()) {
    1778             :       ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
    1779      106133 :       unsigned ElementIdx = OpC->getZExtValue();
    1780      106133 :       const StructLayout *SL = Q.DL.getStructLayout(STy);
    1781             :       uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
    1782      106133 :       if (ElementOffset > 0)
    1783       44421 :         return true;
    1784       78481 :       continue;
    1785             :     }
    1786             : 
    1787             :     // If we have a zero-sized type, the index doesn't matter. Keep looping.
    1788      108323 :     if (Q.DL.getTypeAllocSize(GTI.getIndexedType()) == 0)
    1789           0 :       continue;
    1790             : 
    1791             :     // Fast path the constant operand case both for efficiency and so we don't
    1792             :     // increment Depth when just zipping down an all-constant GEP.
    1793      186378 :     if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
    1794       94514 :       if (!OpC->isZero())
    1795             :         return true;
    1796       78055 :       continue;
    1797             :     }
    1798             : 
    1799             :     // We post-increment Depth here because while isKnownNonZero increments it
    1800             :     // as well, when we pop back up that increment won't persist. We don't want
    1801             :     // to recurse 10k times just because we have 10k GEP operands. We don't
    1802             :     // bail completely out because we want to handle constant GEPs regardless
    1803             :     // of depth.
    1804       13809 :     if (Depth++ >= MaxDepth)
    1805           0 :       continue;
    1806             : 
    1807       13809 :     if (isKnownNonZero(GTI.getOperand(), Depth, Q))
    1808             :       return true;
    1809             :   }
    1810             : 
    1811       59435 :   return false;
    1812             : }
    1813             : 
    1814     1293818 : static bool isKnownNonNullFromDominatingCondition(const Value *V,
    1815             :                                                   const Instruction *CtxI,
    1816             :                                                   const DominatorTree *DT) {
    1817             :   assert(V->getType()->isPointerTy() && "V must be pointer type");
    1818             :   assert(!isa<ConstantData>(V) && "Did not expect ConstantPointerNull");
    1819             : 
    1820     1293818 :   if (!CtxI || !DT)
    1821             :     return false;
    1822             : 
    1823             :   unsigned NumUsesExplored = 0;
    1824     2066129 :   for (auto *U : V->users()) {
    1825             :     // Avoid massive lists
    1826     1454193 :     if (NumUsesExplored >= DomConditionsMaxUses)
    1827             :       break;
    1828     1446430 :     NumUsesExplored++;
    1829             : 
    1830             :     // If the value is used as an argument to a call or invoke, then argument
    1831             :     // attributes may provide an answer about null-ness.
    1832     1446430 :     if (auto CS = ImmutableCallSite(U))
    1833             :       if (auto *CalledFunc = CS.getCalledFunction())
    1834     3286056 :         for (const Argument &Arg : CalledFunc->args())
    1835     3308826 :           if (CS.getArgOperand(Arg.getArgNo()) == V &&
    1836     2337590 :               Arg.hasNonNullAttr() && DT->dominates(CS.getInstruction(), CtxI))
    1837         426 :             return true;
    1838             : 
    1839             :     // Consider only compare instructions uniquely controlling a branch
    1840             :     CmpInst::Predicate Pred;
    1841     1293971 :     if (!match(const_cast<User *>(U),
    1842     1598037 :                m_c_ICmp(Pred, m_Specific(V), m_Zero())) ||
    1843      152033 :         (Pred != ICmpInst::ICMP_EQ && Pred != ICmpInst::ICMP_NE))
    1844     1293971 :       continue;
    1845             : 
    1846      437158 :     for (auto *CmpU : U->users()) {
    1847             :       if (const BranchInst *BI = dyn_cast<BranchInst>(CmpU)) {
    1848             :         assert(BI->isConditional() && "uses a comparison!");
    1849             : 
    1850             :         BasicBlock *NonNullSuccessor =
    1851      147958 :             BI->getSuccessor(Pred == ICmpInst::ICMP_EQ ? 1 : 0);
    1852      147958 :         BasicBlockEdge Edge(BI->getParent(), NonNullSuccessor);
    1853      147958 :         if (Edge.isSingleEdge() && DT->dominates(Edge, CtxI->getParent()))
    1854        1324 :           return true;
    1855      138544 :       } else if (Pred == ICmpInst::ICMP_NE &&
    1856      138497 :                  match(CmpU, m_Intrinsic<Intrinsic::experimental_guard>()) &&
    1857           4 :                  DT->dominates(cast<Instruction>(CmpU), CtxI)) {
    1858             :         return true;
    1859             :       }
    1860             :     }
    1861             :   }
    1862             : 
    1863             :   return false;
    1864             : }
    1865             : 
    1866             : /// Does the 'Range' metadata (which must be a valid MD_range operand list)
    1867             : /// ensure that the value it's attached to is never Value?  'RangeType' is
    1868             : /// is the type of the value described by the range.
    1869       48825 : static bool rangeMetadataExcludesValue(const MDNode* Ranges, const APInt& Value) {
    1870       48825 :   const unsigned NumRanges = Ranges->getNumOperands() / 2;
    1871             :   assert(NumRanges >= 1);
    1872       48851 :   for (unsigned i = 0; i < NumRanges; ++i) {
    1873             :     ConstantInt *Lower =
    1874       48826 :         mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 0));
    1875             :     ConstantInt *Upper =
    1876       48826 :         mdconst::extract<ConstantInt>(Ranges->getOperand(2 * i + 1));
    1877      146491 :     ConstantRange Range(Lower->getValue(), Upper->getValue());
    1878       48826 :     if (Range.contains(Value))
    1879       48813 :       return false;
    1880             :   }
    1881             :   return true;
    1882             : }
    1883             : 
    1884             : /// Return true if the given value is known to be non-zero when defined. For
    1885             : /// vectors, return true if every element is known to be non-zero when
    1886             : /// defined. For pointers, if the context instruction and dominator tree are
    1887             : /// specified, perform context-sensitive analysis and return true if the
    1888             : /// pointer couldn't possibly be null at the specified instruction.
    1889             : /// Supports values with integer or pointer type and vectors of integers.
    1890     3221855 : bool isKnownNonZero(const Value *V, unsigned Depth, const Query &Q) {
    1891             :   if (auto *C = dyn_cast<Constant>(V)) {
    1892     1550206 :     if (C->isNullValue())
    1893             :       return false;
    1894     1499260 :     if (isa<ConstantInt>(C))
    1895             :       // Must be non-zero due to null test above.
    1896             :       return true;
    1897             : 
    1898             :     // For constant vectors, check that all elements are undefined or known
    1899             :     // non-zero to determine that the whole vector is known non-zero.
    1900     1496389 :     if (auto *VecTy = dyn_cast<VectorType>(C->getType())) {
    1901         275 :       for (unsigned i = 0, e = VecTy->getNumElements(); i != e; ++i) {
    1902         217 :         Constant *Elt = C->getAggregateElement(i);
    1903         217 :         if (!Elt || Elt->isNullValue())
    1904             :           return false;
    1905         152 :         if (!isa<UndefValue>(Elt) && !isa<ConstantInt>(Elt))
    1906             :           return false;
    1907             :       }
    1908             :       return true;
    1909             :     }
    1910             : 
    1911             :     // A global variable in address space 0 is non null unless extern weak
    1912             :     // or an absolute symbol reference. Other address spaces may have null as a
    1913             :     // valid address for a global, so we can't assume anything.
    1914             :     if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
    1915      320503 :       if (!GV->isAbsoluteSymbolRef() && !GV->hasExternalWeakLinkage() &&
    1916             :           GV->getType()->getAddressSpace() == 0)
    1917             :         return true;
    1918             :     } else
    1919             :       return false;
    1920             :   }
    1921             : 
    1922             :   if (auto *I = dyn_cast<Instruction>(V)) {
    1923     1151396 :     if (MDNode *Ranges = I->getMetadata(LLVMContext::MD_range)) {
    1924             :       // If the possible ranges don't contain zero, then the value is
    1925             :       // definitely non-zero.
    1926       48825 :       if (auto *Ty = dyn_cast<IntegerType>(V->getType())) {
    1927             :         const APInt ZeroValue(Ty->getBitWidth(), 0);
    1928       48825 :         if (rangeMetadataExcludesValue(Ranges, ZeroValue))
    1929             :           return true;
    1930             :       }
    1931             :     }
    1932             :   }
    1933             : 
    1934             :   // Check for pointer simplifications.
    1935     3343312 :   if (V->getType()->isPointerTy()) {
    1936             :     // Alloca never returns null, malloc might.
    1937      143299 :     if (isa<AllocaInst>(V) && Q.DL.getAllocaAddrSpace() == 0)
    1938             :       return true;
    1939             : 
    1940             :     // A byval, inalloca, or nonnull argument is never null.
    1941             :     if (const Argument *A = dyn_cast<Argument>(V))
    1942      275773 :       if (A->hasByValOrInAllocaAttr() || A->hasNonNullAttr())
    1943             :         return true;
    1944             : 
    1945             :     // A Load tagged with nonnull metadata is never null.
    1946             :     if (const LoadInst *LI = dyn_cast<LoadInst>(V))
    1947      902079 :       if (LI->getMetadata(LLVMContext::MD_nonnull))
    1948             :         return true;
    1949             : 
    1950     1302973 :     if (auto CS = ImmutableCallSite(V))
    1951       61616 :       if (CS.isReturnNonNull())
    1952        9045 :         return true;
    1953             :   }
    1954             : 
    1955             :   // The remaining tests are all recursive, so bail out if we hit the limit.
    1956     1492175 :   if (Depth++ >= MaxDepth)
    1957             :     return false;
    1958             : 
    1959             :   // Check for recursive pointer simplifications.
    1960     2984126 :   if (V->getType()->isPointerTy()) {
    1961     1293818 :     if (isKnownNonNullFromDominatingCondition(V, Q.CxtI, Q.DT))
    1962             :       return true;
    1963             : 
    1964             :     if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V))
    1965      212764 :       if (isGEPKnownNonNull(GEP, Depth, Q))
    1966             :         return true;
    1967             :   }
    1968             : 
    1969     2678200 :   unsigned BitWidth = getBitWidth(V->getType()->getScalarType(), Q.DL);
    1970             : 
    1971             :   // X | Y != 0 if X != 0 or Y != 0.
    1972     1339100 :   Value *X = nullptr, *Y = nullptr;
    1973     2678200 :   if (match(V, m_Or(m_Value(X), m_Value(Y))))
    1974         408 :     return isKnownNonZero(X, Depth, Q) || isKnownNonZero(Y, Depth, Q);
    1975             : 
    1976             :   // ext X != 0 if X != 0.
    1977             :   if (isa<SExtInst>(V) || isa<ZExtInst>(V))
    1978       12422 :     return isKnownNonZero(cast<Instruction>(V)->getOperand(0), Depth, Q);
    1979             : 
    1980             :   // shl X, Y != 0 if X is odd.  Note that the value of the shift is undefined
    1981             :   // if the lowest bit is shifted off the end.
    1982     2664962 :   if (match(V, m_Shl(m_Value(X), m_Value(Y)))) {
    1983             :     // shl nuw can't remove any non-zero bits.
    1984             :     const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
    1985        1199 :     if (BO->hasNoUnsignedWrap())
    1986         139 :       return isKnownNonZero(X, Depth, Q);
    1987             : 
    1988        2259 :     KnownBits Known(BitWidth);
    1989        1130 :     computeKnownBits(X, Known, Depth, Q);
    1990        1130 :     if (Known.One[0])
    1991           1 :       return true;
    1992             :   }
    1993             :   // shr X, Y != 0 if X is negative.  Note that the value of the shift is not
    1994             :   // defined if the sign bit is shifted off the end.
    1995     2662564 :   else if (match(V, m_Shr(m_Value(X), m_Value(Y)))) {
    1996             :     // shr exact can only shift out zero bits.
    1997             :     const PossiblyExactOperator *BO = cast<PossiblyExactOperator>(V);
    1998        3659 :     if (BO->isExact())
    1999        3216 :       return isKnownNonZero(X, Depth, Q);
    2000             : 
    2001        4102 :     KnownBits Known = computeKnownBits(X, Depth, Q);
    2002        2054 :     if (Known.isNegative())
    2003           6 :       return true;
    2004             : 
    2005             :     // If the shifter operand is a constant, and all of the bits shifted
    2006             :     // out are known to be zero, and X is known non-zero then at least one
    2007             :     // non-zero bit must remain.
    2008        2049 :     if (ConstantInt *Shift = dyn_cast<ConstantInt>(Y)) {
    2009        1757 :       auto ShiftVal = Shift->getLimitedValue(BitWidth - 1);
    2010             :       // Is there a known one in the portion not shifted out?
    2011        1757 :       if (Known.countMaxLeadingZeros() < BitWidth - ShiftVal)
    2012             :         return true;
    2013             :       // Are all the bits to be shifted out known zero?
    2014        1756 :       if (Known.countMinTrailingZeros() >= ShiftVal)
    2015           0 :         return isKnownNonZero(X, Depth, Q);
    2016             :     }
    2017             :   }
    2018             :   // div exact can only produce a zero if the dividend is zero.
    2019     2655246 :   else if (match(V, m_Exact(m_IDiv(m_Value(X), m_Value())))) {
    2020         492 :     return isKnownNonZero(X, Depth, Q);
    2021             :   }
    2022             :   // X + Y.
    2023     2654262 :   else if (match(V, m_Add(m_Value(X), m_Value(Y)))) {
    2024       12888 :     KnownBits XKnown = computeKnownBits(X, Depth, Q);
    2025       12888 :     KnownBits YKnown = computeKnownBits(Y, Depth, Q);
    2026             : 
    2027             :     // If X and Y are both non-negative (as signed values) then their sum is not
    2028             :     // zero unless both X and Y are zero.
    2029        7421 :     if (XKnown.isNonNegative() && YKnown.isNonNegative())
    2030         561 :       if (isKnownNonZero(X, Depth, Q) || isKnownNonZero(Y, Depth, Q))
    2031         562 :         return true;
    2032             : 
    2033             :     // If X and Y are both negative (as signed values) then their sum is not
    2034             :     // zero unless both X and Y equal INT_MIN.
    2035        6166 :     if (XKnown.isNegative() && YKnown.isNegative()) {
    2036           1 :       APInt Mask = APInt::getSignedMaxValue(BitWidth);
    2037             :       // The sign bit of X is set.  If some other bit is set then X is not equal
    2038             :       // to INT_MIN.
    2039           1 :       if (XKnown.One.intersects(Mask))
    2040             :         return true;
    2041             :       // The sign bit of Y is set.  If some other bit is set then Y is not equal
    2042             :       // to INT_MIN.
    2043           1 :       if (YKnown.One.intersects(Mask))
    2044             :         return true;
    2045             :     }
    2046             : 
    2047             :     // The sum of a non-negative number and a power of two is not zero.
    2048        6300 :     if (XKnown.isNonNegative() &&
    2049         136 :         isKnownToBeAPowerOfTwo(Y, /*OrZero*/ false, Depth, Q))
    2050             :       return true;
    2051        8372 :     if (YKnown.isNonNegative() &&
    2052        2209 :         isKnownToBeAPowerOfTwo(X, /*OrZero*/ false, Depth, Q))
    2053             :       return true;
    2054             :   }
    2055             :   // X * Y.
    2056     2640812 :   else if (match(V, m_Mul(m_Value(X), m_Value(Y)))) {
    2057             :     const OverflowingBinaryOperator *BO = cast<OverflowingBinaryOperator>(V);
    2058             :     // If X and Y are non-zero then so is X * Y as long as the multiplication
    2059             :     // does not overflow.
    2060          55 :     if ((BO->hasNoSignedWrap() || BO->hasNoUnsignedWrap()) &&
    2061          70 :         isKnownNonZero(X, Depth, Q) && isKnownNonZero(Y, Depth, Q))
    2062             :       return true;
    2063             :   }
    2064             :   // (C ? X : Y) != 0 if X != 0 and Y != 0.
    2065             :   else if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
    2066        3621 :     if (isKnownNonZero(SI->getTrueValue(), Depth, Q) &&
    2067         706 :         isKnownNonZero(SI->getFalseValue(), Depth, Q))
    2068             :       return true;
    2069             :   }
    2070             :   // PHI
    2071             :   else if (const PHINode *PN = dyn_cast<PHINode>(V)) {
    2072             :     // Try and detect a recurrence that monotonically increases from a
    2073             :     // starting value, as these are common as induction variables.
    2074       86237 :     if (PN->getNumIncomingValues() == 2) {
    2075             :       Value *Start = PN->getIncomingValue(0);
    2076             :       Value *Induction = PN->getIncomingValue(1);
    2077       85998 :       if (isa<ConstantInt>(Induction) && !isa<ConstantInt>(Start))
    2078             :         std::swap(Start, Induction);
    2079             :       if (ConstantInt *C = dyn_cast<ConstantInt>(Start)) {
    2080       12252 :         if (!C->isZero() && !C->isNegative()) {
    2081             :           ConstantInt *X;
    2082        5039 :           if ((match(Induction, m_NSWAdd(m_Specific(PN), m_ConstantInt(X))) ||
    2083        2567 :                match(Induction, m_NUWAdd(m_Specific(PN), m_ConstantInt(X)))) &&
    2084         487 :               !X->isNegative())
    2085         242 :             return true;
    2086             :         }
    2087             :       }
    2088             :     }
    2089             :     // Check if all incoming values are non-zero constant.
    2090      174595 :     bool AllNonZeroConstants = llvm::all_of(PN->operands(), [](Value *V) {
    2091      100399 :       return isa<ConstantInt>(V) && !cast<ConstantInt>(V)->isZero();
    2092       88600 :     });
    2093       85995 :     if (AllNonZeroConstants)
    2094             :       return true;
    2095             :   }
    2096             : 
    2097     2658748 :   KnownBits Known(BitWidth);
    2098     1329374 :   computeKnownBits(V, Known, Depth, Q);
    2099             :   return Known.One != 0;
    2100             : }
    2101             : 
    2102             : /// Return true if V2 == V1 + X, where X is known non-zero.
    2103     1626476 : static bool isAddOfNonZero(const Value *V1, const Value *V2, const Query &Q) {
    2104             :   const BinaryOperator *BO = dyn_cast<BinaryOperator>(V1);
    2105       61796 :   if (!BO || BO->getOpcode() != Instruction::Add)
    2106             :     return false;
    2107             :   Value *Op = nullptr;
    2108       20982 :   if (V2 == BO->getOperand(0))
    2109             :     Op = BO->getOperand(1);
    2110       20970 :   else if (V2 == BO->getOperand(1))
    2111             :     Op = BO->getOperand(0);
    2112             :   else
    2113             :     return false;
    2114         187 :   return isKnownNonZero(Op, 0, Q);
    2115             : }
    2116             : 
    2117             : /// Return true if it is known that V1 != V2.
    2118      813245 : static bool isKnownNonEqual(const Value *V1, const Value *V2, const Query &Q) {
    2119      813245 :   if (V1 == V2)
    2120             :     return false;
    2121      813240 :   if (V1->getType() != V2->getType())
    2122             :     // We can't look through casts yet.
    2123             :     return false;
    2124      813240 :   if (isAddOfNonZero(V1, V2, Q) || isAddOfNonZero(V2, V1, Q))
    2125             :     return true;
    2126             : 
    2127     1626462 :   if (V1->getType()->isIntOrIntVectorTy()) {
    2128             :     // Are any known bits in V1 contradictory to known bits in V2? If V1
    2129             :     // has a known zero where V2 has a known one, they must not be equal.
    2130      894356 :     KnownBits Known1 = computeKnownBits(V1, 0, Q);
    2131      894356 :     KnownBits Known2 = computeKnownBits(V2, 0, Q);
    2132             : 
    2133      909244 :     if (Known1.Zero.intersects(Known2.One) ||
    2134             :         Known2.Zero.intersects(Known1.One))
    2135      293922 :       return true;
    2136             :   }
    2137             :   return false;
    2138             : }
    2139             : 
    2140             : /// Return true if 'V & Mask' is known to be zero.  We use this predicate to
    2141             : /// simplify operations downstream. Mask is known to be zero for bits that V
    2142             : /// cannot have.
    2143             : ///
    2144             : /// This function is defined on values with integer type, values with pointer
    2145             : /// type, and vectors of integers.  In the case
    2146             : /// where V is a vector, the mask, known zero, and known one values are the
    2147             : /// same width as the vector element, and the bit is set only if it is true
    2148             : /// for all of the elements in the vector.
    2149       58176 : bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth,
    2150             :                        const Query &Q) {
    2151      116352 :   KnownBits Known(Mask.getBitWidth());
    2152       58176 :   computeKnownBits(V, Known, Depth, Q);
    2153       58176 :   return Mask.isSubsetOf(Known.Zero);
    2154             : }
    2155             : 
    2156             : /// For vector constants, loop over the elements and find the constant with the
    2157             : /// minimum number of sign bits. Return 0 if the value is not a vector constant
    2158             : /// or if any element was not analyzed; otherwise, return the count for the
    2159             : /// element with the minimum number of sign bits.
    2160     3504759 : static unsigned computeNumSignBitsVectorConstant(const Value *V,
    2161             :                                                  unsigned TyBits) {
    2162             :   const auto *CV = dyn_cast<Constant>(V);
    2163      189172 :   if (!CV || !CV->getType()->isVectorTy())
    2164             :     return 0;
    2165             : 
    2166        1722 :   unsigned MinSignBits = TyBits;
    2167             :   unsigned NumElts = CV->getType()->getVectorNumElements();
    2168       14590 :   for (unsigned i = 0; i != NumElts; ++i) {
    2169             :     // If we find a non-ConstantInt, bail out.
    2170        6475 :     auto *Elt = dyn_cast_or_null<ConstantInt>(CV->getAggregateElement(i));
    2171             :     if (!Elt)
    2172             :       return 0;
    2173             : 
    2174       12868 :     MinSignBits = std::min(MinSignBits, Elt->getValue().getNumSignBits());
    2175             :   }
    2176             : 
    2177        1681 :   return MinSignBits;
    2178             : }
    2179             : 
    2180             : static unsigned ComputeNumSignBitsImpl(const Value *V, unsigned Depth,
    2181             :                                        const Query &Q);
    2182             : 
    2183      408462 : static unsigned ComputeNumSignBits(const Value *V, unsigned Depth,
    2184             :                                    const Query &Q) {
    2185     3849600 :   unsigned Result = ComputeNumSignBitsImpl(V, Depth, Q);
    2186             :   assert(Result > 0 && "At least one sign bit needs to be present!");
    2187      408462 :   return Result;
    2188             : }
    2189             : 
    2190             : /// Return the number of times the sign bit of the register is replicated into
    2191             : /// the other bits. We know that at least 1 bit is always equal to the sign bit
    2192             : /// (itself), but other cases can give us information. For example, immediately
    2193             : /// after an "ashr X, 2", we know that the top 3 bits are all equal to each
    2194             : /// other, so we return 3. For vectors, return the number of sign bits for the
    2195             : /// vector element with the mininum number of known sign bits.
    2196     3849600 : static unsigned ComputeNumSignBitsImpl(const Value *V, unsigned Depth,
    2197             :                                        const Query &Q) {
    2198             :   assert(Depth <= MaxDepth && "Limit Search Depth");
    2199             : 
    2200             :   // We return the minimum number of sign bits that are guaranteed to be present
    2201             :   // in V, so for undef we have to conservatively return 1.  We don't have the
    2202             :   // same behavior for poison though -- that's a FIXME today.
    2203             : 
    2204     3849600 :   Type *ScalarTy = V->getType()->getScalarType();
    2205             :   unsigned TyBits = ScalarTy->isPointerTy() ?
    2206       50360 :     Q.DL.getIndexTypeSizeInBits(ScalarTy) :
    2207     3899960 :     Q.DL.getTypeSizeInBits(ScalarTy);
    2208             : 
    2209             :   unsigned Tmp, Tmp2;
    2210     3849600 :   unsigned FirstAnswer = 1;
    2211             : 
    2212             :   // Note that ConstantInt is handled by the general computeKnownBits case
    2213             :   // below.
    2214             : 
    2215     3849600 :   if (Depth == MaxDepth)
    2216             :     return 1;  // Limit search depth.
    2217             : 
    2218             :   const Operator *U = dyn_cast<Operator>(V);
    2219     3672291 :   switch (Operator::getOpcode(V)) {
    2220             :   default: break;
    2221         718 :   case Instruction::SExt:
    2222        1436 :     Tmp = TyBits - U->getOperand(0)->getType()->getScalarSizeInBits();
    2223         718 :     return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q) + Tmp;
    2224             : 
    2225             :   case Instruction::SDiv: {
    2226             :     const APInt *Denominator;
    2227             :     // sdiv X, C -> adds log(C) sign bits.
    2228        2780 :     if (match(U->getOperand(1), m_APInt(Denominator))) {
    2229             : 
    2230             :       // Ignore non-positive denominator.
    2231        1372 :       if (!Denominator->isStrictlyPositive())
    2232             :         break;
    2233             : 
    2234             :       // Calculate the incoming numerator bits.
    2235        2600 :       unsigned NumBits = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
    2236             : 
    2237             :       // Add floor(log(C)) bits to the numerator bits.
    2238        3900 :       return std::min(TyBits, NumBits + Denominator->logBase2());
    2239             :     }
    2240             :     break;
    2241             :   }
    2242             : 
    2243             :   case Instruction::SRem: {
    2244             :     const APInt *Denominator;
    2245             :     // srem X, C -> we know that the result is within [-C+1,C) when C is a
    2246             :     // positive constant.  This let us put a lower bound on the number of sign
    2247             :     // bits.
    2248         280 :     if (match(U->getOperand(1), m_APInt(Denominator))) {
    2249             : 
    2250             :       // Ignore non-positive denominator.
    2251          92 :       if (!Denominator->isStrictlyPositive())
    2252             :         break;
    2253             : 
    2254             :       // Calculate the incoming numerator bits. SRem by a positive constant
    2255             :       // can't lower the number of sign bits.
    2256             :       unsigned NumrBits =
    2257         164 :           ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
    2258             : 
    2259             :       // Calculate the leading sign bit constraints by examining the
    2260             :       // denominator.  Given that the denominator is positive, there are two
    2261             :       // cases:
    2262             :       //
    2263             :       //  1. the numerator is positive.  The result range is [0,C) and [0,C) u<
    2264             :       //     (1 << ceilLogBase2(C)).
    2265             :       //
    2266             :       //  2. the numerator is negative.  Then the result range is (-C,0] and
    2267             :       //     integers in (-C,0] are either 0 or >u (-1 << ceilLogBase2(C)).
    2268             :       //
    2269             :       // Thus a lower bound on the number of sign bits is `TyBits -
    2270             :       // ceilLogBase2(C)`.
    2271             : 
    2272          82 :       unsigned ResBits = TyBits - Denominator->ceilLogBase2();
    2273          82 :       return std::max(NumrBits, ResBits);
    2274             :     }
    2275             :     break;
    2276             :   }
    2277             : 
    2278        2405 :   case Instruction::AShr: {
    2279        4810 :     Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
    2280             :     // ashr X, C   -> adds C sign bits.  Vectors too.
    2281             :     const APInt *ShAmt;
    2282        4810 :     if (match(U->getOperand(1), m_APInt(ShAmt))) {
    2283        4794 :       if (ShAmt->uge(TyBits))
    2284             :         break;  // Bad shift.
    2285        2385 :       unsigned ShAmtLimited = ShAmt->getZExtValue();
    2286        2385 :       Tmp += ShAmtLimited;
    2287        2385 :       if (Tmp > TyBits) Tmp = TyBits;
    2288             :     }
    2289        2393 :     return Tmp;
    2290             :   }
    2291             :   case Instruction::Shl: {
    2292             :     const APInt *ShAmt;
    2293       23540 :     if (match(U->getOperand(1), m_APInt(ShAmt))) {
    2294             :       // shl destroys sign bits.
    2295       21740 :       Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
    2296       32604 :       if (ShAmt->uge(TyBits) ||      // Bad shift.
    2297       10864 :           ShAmt->uge(Tmp)) break;    // Shifted all sign bits out.
    2298         585 :       Tmp2 = ShAmt->getZExtValue();
    2299         585 :       return Tmp - Tmp2;
    2300             :     }
    2301             :     break;
    2302             :   }
    2303       11035 :   case Instruction::And:
    2304             :   case Instruction::Or:
    2305             :   case Instruction::Xor:    // NOT is handled here.
    2306             :     // Logical binary ops preserve the number of sign bits at the worst.
    2307       22070 :     Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
    2308       11035 :     if (Tmp != 1) {
    2309        3579 :       Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
    2310        3579 :       FirstAnswer = std::min(Tmp, Tmp2);
    2311             :       // We computed what we know about the sign bits as our first
    2312             :       // answer. Now proceed to the generic code that uses
    2313             :       // computeKnownBits, and pick whichever answer is better.
    2314             :     }
    2315             :     break;
    2316             : 
    2317       14674 :   case Instruction::Select:
    2318       29348 :     Tmp = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
    2319       14674 :     if (Tmp == 1) return 1;  // Early out.
    2320       12503 :     Tmp2 = ComputeNumSignBits(U->getOperand(2), Depth + 1, Q);
    2321       12503 :     return std::min(Tmp, Tmp2);
    2322             : 
    2323      108705 :   case Instruction::Add:
    2324             :     // Add can have at most one carry bit.  Thus we know that the output
    2325             :     // is, at worst, one more bit than the inputs.
    2326      217410 :     Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
    2327      108705 :     if (Tmp == 1) return 1;  // Early out.
    2328             : 
    2329             :     // Special case decrementing a value (ADD X, -1):
    2330             :     if (const auto *CRHS = dyn_cast<Constant>(U->getOperand(1)))
    2331        3107 :       if (CRHS->isAllOnesValue()) {
    2332          53 :         KnownBits Known(TyBits);
    2333          43 :         computeKnownBits(U->getOperand(0), Known, Depth + 1, Q);
    2334             : 
    2335             :         // If the input is known to be 0 or 1, the output is 0/-1, which is all
    2336             :         // sign bits set.
    2337         129 :         if ((Known.Zero | 1).isAllOnesValue())
    2338          33 :           return TyBits;
    2339             : 
    2340             :         // If we are subtracting one from a positive number, there is no carry
    2341             :         // out of the result.
    2342          41 :         if (Known.isNonNegative())
    2343             :           return Tmp;
    2344             :       }
    2345             : 
    2346        3698 :     Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
    2347        3698 :     if (Tmp2 == 1) return 1;
    2348        3385 :     return std::min(Tmp, Tmp2)-1;
    2349             : 
    2350       44855 :   case Instruction::Sub:
    2351       89710 :     Tmp2 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
    2352       44855 :     if (Tmp2 == 1) return 1;
    2353             : 
    2354             :     // Handle NEG.
    2355             :     if (const auto *CLHS = dyn_cast<Constant>(U->getOperand(0)))
    2356          11 :       if (CLHS->isNullValue()) {
    2357           5 :         KnownBits Known(TyBits);
    2358           5 :         computeKnownBits(U->getOperand(1), Known, Depth + 1, Q);
    2359             :         // If the input is known to be 0 or 1, the output is 0/-1, which is all
    2360             :         // sign bits set.
    2361          15 :         if ((Known.Zero | 1).isAllOnesValue())
    2362           5 :           return TyBits;
    2363             : 
    2364             :         // If the input is known to be positive (the sign bit is known clear),
    2365             :         // the output of the NEG has the same number of sign bits as the input.
    2366           5 :         if (Known.isNonNegative())
    2367             :           return Tmp2;
    2368             : 
    2369             :         // Otherwise, we treat this like a SUB.
    2370             :       }
    2371             : 
    2372             :     // Sub can have at most one carry bit.  Thus we know that the output
    2373             :     // is, at worst, one more bit than the inputs.
    2374        1816 :     Tmp = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
    2375        1816 :     if (Tmp == 1) return 1;  // Early out.
    2376         732 :     return std::min(Tmp, Tmp2)-1;
    2377             : 
    2378        8541 :   case Instruction::Mul: {
    2379             :     // The output of the Mul can be at most twice the valid bits in the inputs.
    2380       17082 :     unsigned SignBitsOp0 = ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
    2381        8541 :     if (SignBitsOp0 == 1) return 1;  // Early out.
    2382        1654 :     unsigned SignBitsOp1 = ComputeNumSignBits(U->getOperand(1), Depth + 1, Q);
    2383        1654 :     if (SignBitsOp1 == 1) return 1;
    2384        1423 :     unsigned OutValidBits =
    2385        1423 :         (TyBits - SignBitsOp0 + 1) + (TyBits - SignBitsOp1 + 1);
    2386        1423 :     return OutValidBits > TyBits ? 1 : TyBits - OutValidBits + 1;
    2387             :   }
    2388             : 
    2389             :   case Instruction::PHI: {
    2390             :     const PHINode *PN = cast<PHINode>(U);
    2391             :     unsigned NumIncomingValues = PN->getNumIncomingValues();
    2392             :     // Don't analyze large in-degree PHIs.
    2393      129140 :     if (NumIncomingValues > 4) break;
    2394             :     // Unreachable blocks may have zero-operand PHI nodes.
    2395      128173 :     if (NumIncomingValues == 0) break;
    2396             : 
    2397             :     // Take the minimum of all incoming values.  This can't infinitely loop
    2398             :     // because of our depth threshold.
    2399      256346 :     Tmp = ComputeNumSignBits(PN->getIncomingValue(0), Depth + 1, Q);
    2400      234097 :     for (unsigned i = 1, e = NumIncomingValues; i != e; ++i) {
    2401      129711 :       if (Tmp == 1) return Tmp;
    2402       52962 :       Tmp = std::min(
    2403      105924 :           Tmp, ComputeNumSignBits(PN->getIncomingValue(i), Depth + 1, Q));
    2404             :     }
    2405       51424 :     return Tmp;
    2406             :   }
    2407             : 
    2408             :   case Instruction::Trunc:
    2409             :     // FIXME: it's tricky to do anything useful for this, but it is an important
    2410             :     // case for targets like X86.
    2411             :     break;
    2412             : 
    2413         892 :   case Instruction::ExtractElement:
    2414             :     // Look through extract element. At the moment we keep this simple and skip
    2415             :     // tracking the specific element. But at least we might find information
    2416             :     // valid for all elements of the vector (for example if vector is sign
    2417             :     // extended, shifted, etc).
    2418        1784 :     return ComputeNumSignBits(U->getOperand(0), Depth + 1, Q);
    2419             :   }
    2420             : 
    2421             :   // Finally, if we can prove that the top bits of the result are 0's or 1's,
    2422             :   // use this information.
    2423             : 
    2424             :   // If we can examine all elements of a vector constant successfully, we're
    2425             :   // done (we can't do any better than that). If not, keep trying.
    2426     3504759 :   if (unsigned VecSignBits = computeNumSignBitsVectorConstant(V, TyBits))
    2427             :     return VecSignBits;
    2428             : 
    2429     7006156 :   KnownBits Known(TyBits);
    2430     3503078 :   computeKnownBits(V, Known, Depth, Q);
    2431             : 
    2432             :   // If we know that the sign bit is either zero or one, determine the number of
    2433             :   // identical bits in the top of the input value.
    2434     7006156 :   return std::max(FirstAnswer, Known.countMinSignBits());
    2435             : }
    2436             : 
    2437             : /// This function computes the integer multiple of Base that equals V.
    2438             : /// If successful, it returns true and returns the multiple in
    2439             : /// Multiple. If unsuccessful, it returns false. It looks
    2440             : /// through SExt instructions only if LookThroughSExt is true.
    2441          50 : bool llvm::ComputeMultiple(Value *V, unsigned Base, Value *&Multiple,
    2442             :                            bool LookThroughSExt, unsigned Depth) {
    2443             :   const unsigned MaxDepth = 6;
    2444             : 
    2445             :   assert(V && "No Value?");
    2446             :   assert(Depth <= MaxDepth && "Limit Search Depth");
    2447             :   assert(V->getType()->isIntegerTy() && "Not integer or pointer type!");
    2448             : 
    2449          50 :   Type *T = V->getType();
    2450             : 
    2451             :   ConstantInt *CI = dyn_cast<ConstantInt>(V);
    2452             : 
    2453          50 :   if (Base == 0)
    2454             :     return false;
    2455             : 
    2456          50 :   if (Base == 1) {
    2457           0 :     Multiple = V;
    2458           0 :     return true;
    2459             :   }
    2460             : 
    2461             :   ConstantExpr *CO = dyn_cast<ConstantExpr>(V);
    2462          50 :   Constant *BaseVal = ConstantInt::get(T, Base);
    2463          50 :   if (CO && CO == BaseVal) {
    2464             :     // Multiple is 1.
    2465           0 :     Multiple = ConstantInt::get(T, 1);
    2466           0 :     return true;
    2467             :   }
    2468             : 
    2469          70 :   if (CI && CI->getZExtValue() % Base == 0) {
    2470          18 :     Multiple = ConstantInt::get(T, CI->getZExtValue() / Base);
    2471          18 :     return true;
    2472             :   }
    2473             : 
    2474          32 :   if (Depth == MaxDepth) return false;  // Limit search depth.
    2475             : 
    2476             :   Operator *I = dyn_cast<Operator>(V);
    2477             :   if (!I) return false;
    2478             : 
    2479          24 :   switch (I->getOpcode()) {
    2480             :   default: break;
    2481           1 :   case Instruction::SExt:
    2482           1 :     if (!LookThroughSExt) return false;
    2483             :     // otherwise fall through to ZExt
    2484             :     LLVM_FALLTHROUGH;
    2485             :   case Instruction::ZExt:
    2486           1 :     return ComputeMultiple(I->getOperand(0), Base, Multiple,
    2487           1 :                            LookThroughSExt, Depth+1);
    2488          13 :   case Instruction::Shl:
    2489             :   case Instruction::Mul: {
    2490          13 :     Value *Op0 = I->getOperand(0);
    2491             :     Value *Op1 = I->getOperand(1);
    2492             : 
    2493          13 :     if (I->getOpcode() == Instruction::Shl) {
    2494             :       ConstantInt *Op1CI = dyn_cast<ConstantInt>(Op1);
    2495           0 :       if (!Op1CI) return false;
    2496             :       // Turn Op0 << Op1 into Op0 * 2^Op1
    2497             :       APInt Op1Int = Op1CI->getValue();
    2498           6 :       uint64_t BitToSet = Op1Int.getLimitedValue(Op1Int.getBitWidth() - 1);
    2499             :       APInt API(Op1Int.getBitWidth(), 0);
    2500           6 :       API.setBit(BitToSet);
    2501           6 :       Op1 = ConstantInt::get(V->getContext(), API);
    2502             :     }
    2503             : 
    2504          13 :     Value *Mul0 = nullptr;
    2505          13 :     if (ComputeMultiple(Op0, Base, Mul0, LookThroughSExt, Depth+1)) {
    2506             :       if (Constant *Op1C = dyn_cast<Constant>(Op1))
    2507           1 :         if (Constant *MulC = dyn_cast<Constant>(Mul0)) {
    2508           2 :           if (Op1C->getType()->getPrimitiveSizeInBits() <
    2509           1 :               MulC->getType()->getPrimitiveSizeInBits())
    2510           0 :             Op1C = ConstantExpr::getZExt(Op1C, MulC->getType());
    2511           2 :           if (Op1C->getType()->getPrimitiveSizeInBits() >
    2512           1 :               MulC->getType()->getPrimitiveSizeInBits())
    2513           1 :             MulC = ConstantExpr::getZExt(MulC, Op1C->getType());
    2514             : 
    2515             :           // V == Base * (Mul0 * Op1), so return (Mul0 * Op1)
    2516           1 :           Multiple = ConstantExpr::getMul(MulC, Op1C);
    2517           1 :           return true;
    2518             :         }
    2519             : 
    2520           2 :       if (ConstantInt *Mul0CI = dyn_cast<ConstantInt>(Mul0))
    2521           2 :         if (Mul0CI->getValue() == 1) {
    2522             :           // V == Base * Op1, so return Op1
    2523           2 :           Multiple = Op1;
    2524           2 :           return true;
    2525             :         }
    2526             :     }
    2527             : 
    2528          10 :     Value *Mul1 = nullptr;
    2529          10 :     if (ComputeMultiple(Op1, Base, Mul1, LookThroughSExt, Depth+1)) {
    2530             :       if (Constant *Op0C = dyn_cast<Constant>(Op0))
    2531           0 :         if (Constant *MulC = dyn_cast<Constant>(Mul1)) {
    2532           0 :           if (Op0C->getType()->getPrimitiveSizeInBits() <
    2533           0 :               MulC->getType()->getPrimitiveSizeInBits())
    2534           0 :             Op0C = ConstantExpr::getZExt(Op0C, MulC->getType());
    2535           0 :           if (Op0C->getType()->getPrimitiveSizeInBits() >
    2536           0 :               MulC->getType()->getPrimitiveSizeInBits())
    2537           0 :             MulC = ConstantExpr::getZExt(MulC, Op0C->getType());
    2538             : 
    2539             :           // V == Base * (Mul1 * Op0), so return (Mul1 * Op0)
    2540           0 :           Multiple = ConstantExpr::getMul(MulC, Op0C);
    2541           0 :           return true;
    2542             :         }
    2543             : 
    2544           8 :       if (ConstantInt *Mul1CI = dyn_cast<ConstantInt>(Mul1))
    2545           8 :         if (Mul1CI->getValue() == 1) {
    2546             :           // V == Base * Op0, so return Op0
    2547           8 :           Multiple = Op0;
    2548           8 :           return true;
    2549             :         }
    2550           2 :     }
    2551             :   }
    2552             :   }
    2553             : 
    2554             :   // We could not determine if V is a multiple of Base.
    2555             :   return false;
    2556             : }
    2557             : 
    2558       11524 : Intrinsic::ID llvm::getIntrinsicForCallSite(ImmutableCallSite ICS,
    2559             :                                             const TargetLibraryInfo *TLI) {
    2560             :   const Function *F = ICS.getCalledFunction();
    2561             :   if (!F)
    2562             :     return Intrinsic::not_intrinsic;
    2563             : 
    2564       11483 :   if (F->isIntrinsic())
    2565       10626 :     return F->getIntrinsicID();
    2566             : 
    2567         857 :   if (!TLI)
    2568             :     return Intrinsic::not_intrinsic;
    2569             : 
    2570             :   LibFunc Func;
    2571             :   // We're going to make assumptions on the semantics of the functions, check
    2572             :   // that the target knows that it's available in this environment and it does
    2573             :   // not have local linkage.
    2574         848 :   if (!F || F->hasLocalLinkage() || !TLI->getLibFunc(*F, Func))
    2575             :     return Intrinsic::not_intrinsic;
    2576             : 
    2577         572 :   if (!ICS.onlyReadsMemory())
    2578             :     return Intrinsic::not_intrinsic;
    2579             : 
    2580             :   // Otherwise check if we have a call to a function that can be turned into a
    2581             :   // vector intrinsic.
    2582         490 :   switch (Func) {
    2583             :   default:
    2584             :     break;
    2585             :   case LibFunc_sin:
    2586             :   case LibFunc_sinf:
    2587             :   case LibFunc_sinl:
    2588             :     return Intrinsic::sin;
    2589          37 :   case LibFunc_cos:
    2590             :   case LibFunc_cosf:
    2591             :   case LibFunc_cosl:
    2592          37 :     return Intrinsic::cos;
    2593          32 :   case LibFunc_exp:
    2594             :   case LibFunc_expf:
    2595             :   case LibFunc_expl:
    2596          32 :     return Intrinsic::exp;
    2597           5 :   case LibFunc_exp2:
    2598             :   case LibFunc_exp2f:
    2599             :   case LibFunc_exp2l:
    2600           5 :     return Intrinsic::exp2;
    2601          32 :   case LibFunc_log:
    2602             :   case LibFunc_logf:
    2603             :   case LibFunc_logl:
    2604          32 :     return Intrinsic::log;
    2605          16 :   case LibFunc_log10:
    2606             :   case LibFunc_log10f:
    2607             :   case LibFunc_log10l:
    2608          16 :     return Intrinsic::log10;
    2609           0 :   case LibFunc_log2:
    2610             :   case LibFunc_log2f:
    2611             :   case LibFunc_log2l:
    2612           0 :     return Intrinsic::log2;
    2613          25 :   case LibFunc_fabs:
    2614             :   case LibFunc_fabsf:
    2615             :   case LibFunc_fabsl:
    2616          25 :     return Intrinsic::fabs;
    2617           0 :   case LibFunc_fmin:
    2618             :   case LibFunc_fminf:
    2619             :   case LibFunc_fminl:
    2620           0 :     return Intrinsic::minnum;
    2621           0 :   case LibFunc_fmax:
    2622             :   case LibFunc_fmaxf:
    2623             :   case LibFunc_fmaxl:
    2624           0 :     return Intrinsic::maxnum;
    2625           0 :   case LibFunc_copysign:
    2626             :   case LibFunc_copysignf:
    2627             :   case LibFunc_copysignl:
    2628           0 :     return Intrinsic::copysign;
    2629          47 :   case LibFunc_floor:
    2630             :   case LibFunc_floorf:
    2631             :   case LibFunc_floorl:
    2632          47 :     return Intrinsic::floor;
    2633          16 :   case LibFunc_ceil:
    2634             :   case LibFunc_ceilf:
    2635             :   case LibFunc_ceill:
    2636          16 :     return Intrinsic::ceil;
    2637           0 :   case LibFunc_trunc:
    2638             :   case LibFunc_truncf:
    2639             :   case LibFunc_truncl:
    2640           0 :     return Intrinsic::trunc;
    2641           0 :   case LibFunc_rint:
    2642             :   case LibFunc_rintf:
    2643             :   case LibFunc_rintl:
    2644           0 :     return Intrinsic::rint;
    2645           0 :   case LibFunc_nearbyint:
    2646             :   case LibFunc_nearbyintf:
    2647             :   case LibFunc_nearbyintl:
    2648           0 :     return Intrinsic::nearbyint;
    2649           0 :   case LibFunc_round:
    2650             :   case LibFunc_roundf:
    2651             :   case LibFunc_roundl:
    2652           0 :     return Intrinsic::round;
    2653          22 :   case LibFunc_pow:
    2654             :   case LibFunc_powf:
    2655             :   case LibFunc_powl:
    2656          22 :     return Intrinsic::pow;
    2657          63 :   case LibFunc_sqrt:
    2658             :   case LibFunc_sqrtf:
    2659             :   case LibFunc_sqrtl:
    2660          63 :     return Intrinsic::sqrt;
    2661             :   }
    2662             : 
    2663         158 :   return Intrinsic::not_intrinsic;
    2664             : }
    2665             : 
    2666             : /// Return true if we can prove that the specified FP value is never equal to
    2667             : /// -0.0.
    2668             : ///
    2669             : /// NOTE: this function will need to be revisited when we support non-default
    2670             : /// rounding modes!
    2671         271 : bool llvm::CannotBeNegativeZero(const Value *V, const TargetLibraryInfo *TLI,
    2672             :                                 unsigned Depth) {
    2673             :   if (auto *CFP = dyn_cast<ConstantFP>(V))
    2674           0 :     return !CFP->getValueAPF().isNegZero();
    2675             : 
    2676             :   // Limit search depth.
    2677         271 :   if (Depth == MaxDepth)
    2678             :     return false;
    2679             : 
    2680             :   auto *Op = dyn_cast<Operator>(V);
    2681             :   if (!Op)
    2682             :     return false;
    2683             : 
    2684             :   // Check if the nsz fast-math flag is set.
    2685             :   if (auto *FPO = dyn_cast<FPMathOperator>(Op))
    2686         225 :     if (FPO->hasNoSignedZeros())
    2687             :       return true;
    2688             : 
    2689             :   // (fadd x, 0.0) is guaranteed to return +0.0, not -0.0.
    2690         223 :   if (match(Op, m_FAdd(m_Value(), m_Zero())))
    2691             :     return true;
    2692             : 
    2693             :   // sitofp and uitofp turn into +0.0 for zero.
    2694             :   if (isa<SIToFPInst>(Op) || isa<UIToFPInst>(Op))
    2695             :     return true;
    2696             : 
    2697             :   if (auto *Call = dyn_cast<CallInst>(Op)) {
    2698           8 :     Intrinsic::ID IID = getIntrinsicForCallSite(Call, TLI);
    2699           8 :     switch (IID) {
    2700             :     default:
    2701             :       break;
    2702             :     // sqrt(-0.0) = -0.0, no other negative results are possible.
    2703           1 :     case Intrinsic::sqrt:
    2704           2 :       return CannotBeNegativeZero(Call->getArgOperand(0), TLI, Depth + 1);
    2705             :     // fabs(x) != -0.0
    2706             :     case Intrinsic::fabs:
    2707             :       return true;
    2708             :     }
    2709             :   }
    2710             : 
    2711             :   return false;
    2712             : }
    2713             : 
    2714             : /// If \p SignBitOnly is true, test for a known 0 sign bit rather than a
    2715             : /// standard ordered compare. e.g. make -0.0 olt 0.0 be true because of the sign
    2716             : /// bit despite comparing equal.
    2717        2860 : static bool cannotBeOrderedLessThanZeroImpl(const Value *V,
    2718             :                                             const TargetLibraryInfo *TLI,
    2719             :                                             bool SignBitOnly,
    2720             :                                             unsigned Depth) {
    2721             :   // TODO: This function does not do the right thing when SignBitOnly is true
    2722             :   // and we're lowering to a hypothetical IEEE 754-compliant-but-evil platform
    2723             :   // which flips the sign bits of NaNs.  See
    2724             :   // https://llvm.org/bugs/show_bug.cgi?id=31702.
    2725             : 
    2726             :   if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
    2727          31 :     return !CFP->getValueAPF().isNegative() ||
    2728           3 :            (!SignBitOnly && CFP->getValueAPF().isZero());
    2729             :   }
    2730             : 
    2731        2829 :   if (Depth == MaxDepth)
    2732             :     return false; // Limit search depth.
    2733             : 
    2734             :   const Operator *I = dyn_cast<Operator>(V);
    2735             :   if (!I)
    2736             :     return false;
    2737             : 
    2738        1675 :   switch (I->getOpcode()) {
    2739             :   default:
    2740             :     break;
    2741             :   // Unsigned integers are always nonnegative.
    2742             :   case Instruction::UIToFP:
    2743             :     return true;
    2744          80 :   case Instruction::FMul:
    2745             :     // x*x is always non-negative or a NaN.
    2746         160 :     if (I->getOperand(0) == I->getOperand(1) &&
    2747          13 :         (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()))
    2748             :       return true;
    2749             : 
    2750             :     LLVM_FALLTHROUGH;
    2751             :   case Instruction::FAdd:
    2752             :   case Instruction::FDiv:
    2753             :   case Instruction::FRem:
    2754         420 :     return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
    2755         215 :                                            Depth + 1) &&
    2756           5 :            cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
    2757             :                                            Depth + 1);
    2758          24 :   case Instruction::Select:
    2759          48 :     return cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
    2760          32 :                                            Depth + 1) &&
    2761           8 :            cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly,
    2762             :                                            Depth + 1);
    2763         144 :   case Instruction::FPExt:
    2764             :   case Instruction::FPTrunc:
    2765             :     // Widening/narrowing never change sign.
    2766         144 :     return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
    2767         144 :                                            Depth + 1);
    2768             :   case Instruction::Call:
    2769             :     const auto *CI = cast<CallInst>(I);
    2770         176 :     Intrinsic::ID IID = getIntrinsicForCallSite(CI, TLI);
    2771         176 :     switch (IID) {
    2772             :     default:
    2773             :       break;
    2774           1 :     case Intrinsic::maxnum:
    2775           2 :       return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
    2776           1 :                                              Depth + 1) ||
    2777           0 :              cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
    2778             :                                              Depth + 1);
    2779           1 :     case Intrinsic::minnum:
    2780           2 :       return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
    2781           2 :                                              Depth + 1) &&
    2782           1 :              cannotBeOrderedLessThanZeroImpl(I->getOperand(1), TLI, SignBitOnly,
    2783             :                                              Depth + 1);
    2784             :     case Intrinsic::exp:
    2785             :     case Intrinsic::exp2:
    2786             :     case Intrinsic::fabs:
    2787             :       return true;
    2788             : 
    2789          31 :     case Intrinsic::sqrt:
    2790             :       // sqrt(x) is always >= -0 or NaN.  Moreover, sqrt(x) == -0 iff x == -0.
    2791          31 :       if (!SignBitOnly)
    2792             :         return true;
    2793          31 :       return CI->hasNoNaNs() && (CI->hasNoSignedZeros() ||
    2794           4 :                                  CannotBeNegativeZero(CI->getOperand(0), TLI));
    2795             : 
    2796           1 :     case Intrinsic::powi:
    2797           1 :       if (ConstantInt *Exponent = dyn_cast<ConstantInt>(I->getOperand(1))) {
    2798             :         // powi(x,n) is non-negative if n is even.
    2799           2 :         if (Exponent->getBitWidth() <= 64 && Exponent->getSExtValue() % 2u == 0)
    2800             :           return true;
    2801             :       }
    2802             :       // TODO: This is not correct.  Given that exp is an integer, here are the
    2803             :       // ways that pow can return a negative value:
    2804             :       //
    2805             :       //   pow(x, exp)    --> negative if exp is odd and x is negative.
    2806             :       //   pow(-0, exp)   --> -inf if exp is negative odd.
    2807             :       //   pow(-0, exp)   --> -0 if exp is positive odd.
    2808             :       //   pow(-inf, exp) --> -0 if exp is negative odd.
    2809             :       //   pow(-inf, exp) --> -inf if exp is positive odd.
    2810             :       //
    2811             :       // Therefore, if !SignBitOnly, we can return true if x >= +0 or x is NaN,
    2812             :       // but we must return false if x == -0.  Unfortunately we do not currently
    2813             :       // have a way of expressing this constraint.  See details in
    2814             :       // https://llvm.org/bugs/show_bug.cgi?id=31702.
    2815           1 :       return cannotBeOrderedLessThanZeroImpl(I->getOperand(0), TLI, SignBitOnly,
    2816           1 :                                              Depth + 1);
    2817             : 
    2818          11 :     case Intrinsic::fma:
    2819             :     case Intrinsic::fmuladd:
    2820             :       // x*x+y is non-negative if y is non-negative.
    2821          16 :       return I->getOperand(0) == I->getOperand(1) &&
    2822          18 :              (!SignBitOnly || cast<FPMathOperator>(I)->hasNoNaNs()) &&
    2823           6 :              cannotBeOrderedLessThanZeroImpl(I->getOperand(2), TLI, SignBitOnly,
    2824             :                                              Depth + 1);
    2825             :     }
    2826             :     break;
    2827             :   }
    2828             :   return false;
    2829             : }
    2830             : 
    2831         290 : bool llvm::CannotBeOrderedLessThanZero(const Value *V,
    2832             :                                        const TargetLibraryInfo *TLI) {
    2833         290 :   return cannotBeOrderedLessThanZeroImpl(V, TLI, false, 0);
    2834             : }
    2835             : 
    2836        2172 : bool llvm::SignBitMustBeZero(const Value *V, const TargetLibraryInfo *TLI) {
    2837        2172 :   return cannotBeOrderedLessThanZeroImpl(V, TLI, true, 0);
    2838             : }
    2839             : 
    2840         962 : bool llvm::isKnownNeverNaN(const Value *V) {
    2841             :   assert(V->getType()->isFPOrFPVectorTy() && "Querying for NaN on non-FP type");
    2842             : 
    2843             :   // If we're told that NaNs won't happen, assume they won't.
    2844         464 :   if (auto *FPMathOp = dyn_cast<FPMathOperator>(V))
    2845         464 :     if (FPMathOp->hasNoNaNs())
    2846             :       return true;
    2847             : 
    2848             :   // TODO: Handle instructions and potentially recurse like other 'isKnown'
    2849             :   // functions. For example, the result of sitofp is never NaN.
    2850             : 
    2851             :   // Handle scalar constants.
    2852             :   if (auto *CFP = dyn_cast<ConstantFP>(V))
    2853          21 :     return !CFP->isNaN();
    2854             : 
    2855             :   // Bail out for constant expressions, but try to handle vector constants.
    2856        1870 :   if (!V->getType()->isVectorTy() || !isa<Constant>(V))
    2857             :     return false;
    2858             : 
    2859             :   // For vectors, verify that each element is not NaN.
    2860             :   unsigned NumElts = V->getType()->getVectorNumElements();
    2861         176 :   for (unsigned i = 0; i != NumElts; ++i) {
    2862          76 :     Constant *Elt = cast<Constant>(V)->getAggregateElement(i);
    2863          76 :     if (!Elt)
    2864             :       return false;
    2865          76 :     if (isa<UndefValue>(Elt))
    2866             :       continue;
    2867             :     auto *CElt = dyn_cast<ConstantFP>(Elt);
    2868          76 :     if (!CElt || CElt->isNaN())
    2869             :       return false;
    2870             :   }
    2871             :   // All elements were confirmed not-NaN or undefined.
    2872             :   return true;
    2873             : }
    2874             : 
    2875             : /// If the specified value can be set by repeating the same byte in memory,
    2876             : /// return the i8 value that it is represented with.  This is
    2877             : /// true for all i8 values obviously, but is also true for i32 0, i32 -1,
    2878             : /// i16 0xF0F0, double 0.0 etc.  If the value can't be handled with a repeated
    2879             : /// byte store (e.g. i16 0x1234), return null.
    2880      660421 : Value *llvm::isBytewiseValue(Value *V) {
    2881             :   // All byte-wide stores are splatable, even of arbitrary variables.
    2882      660421 :   if (V->getType()->isIntegerTy(8)) return V;
    2883             : 
    2884             :   // Handle 'null' ConstantArrayZero etc.
    2885             :   if (Constant *C = dyn_cast<Constant>(V))
    2886      252051 :     if (C->isNullValue())
    2887      222629 :       return Constant::getNullValue(Type::getInt8Ty(V->getContext()));
    2888             : 
    2889             :   // Constant float and double values can be handled as integer values if the
    2890             :   // corresponding integer value is "byteable".  An important case is 0.0.
    2891             :   if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
    2892          86 :     if (CFP->getType()->isFloatTy())
    2893          29 :       V = ConstantExpr::getBitCast(CFP, Type::getInt32Ty(V->getContext()));
    2894          86 :     if (CFP->getType()->isDoubleTy())
    2895          12 :       V = ConstantExpr::getBitCast(CFP, Type::getInt64Ty(V->getContext()));
    2896             :     // Don't handle long double formats, which have strange constraints.
    2897             :   }
    2898             : 
    2899             :   // We can handle constant integers that are multiple of 8 bits.
    2900             :   if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
    2901       24769 :     if (CI->getBitWidth() % 8 == 0) {
    2902             :       assert(CI->getBitWidth() > 8 && "8 bits should be handled above!");
    2903             : 
    2904       24756 :       if (!CI->getValue().isSplat(8))
    2905             :         return nullptr;
    2906         688 :       return ConstantInt::get(V->getContext(), CI->getValue().trunc(8));
    2907             :     }
    2908             :   }
    2909             : 
    2910             :   // A ConstantDataArray/Vector is splatable if all its members are equal and
    2911             :   // also splatable.
    2912             :   if (ConstantDataSequential *CA = dyn_cast<ConstantDataSequential>(V)) {
    2913         103 :     Value *Elt = CA->getElementAsConstant(0);
    2914         103 :     Value *Val = isBytewiseValue(Elt);
    2915         103 :     if (!Val)
    2916             :       return nullptr;
    2917             : 
    2918         118 :     for (unsigned I = 1, E = CA->getNumElements(); I != E; ++I)
    2919         106 :       if (CA->getElementAsConstant(I) != Elt)
    2920             :         return nullptr;
    2921             : 
    2922             :     return Val;
    2923             :   }
    2924             : 
    2925             :   // Conceptually, we could handle things like:
    2926             :   //   %a = zext i8 %X to i16
    2927             :   //   %b = shl i16 %a, 8
    2928             :   //   %c = or i16 %a, %b
    2929             :   // but until there is an example that actually needs this, it doesn't seem
    2930             :   // worth worrying about.
    2931             :   return nullptr;
    2932             : }
    2933             : 
    2934             : // This is the recursive version of BuildSubAggregate. It takes a few different
    2935             : // arguments. Idxs is the index within the nested struct From that we are
    2936             : // looking at now (which is of type IndexedType). IdxSkip is the number of
    2937             : // indices from Idxs that should be left out when inserting into the resulting
    2938             : // struct. To is the result struct built so far, new insertvalue instructions
    2939             : // build on that.
    2940           0 : static Value *BuildSubAggregate(Value *From, Value* To, Type *IndexedType,
    2941             :                                 SmallVectorImpl<unsigned> &Idxs,
    2942             :                                 unsigned IdxSkip,
    2943             :                                 Instruction *InsertBefore) {
    2944             :   StructType *STy = dyn_cast<StructType>(IndexedType);
    2945             :   if (STy) {
    2946             :     // Save the original To argument so we can modify it
    2947             :     Value *OrigTo = To;
    2948             :     // General case, the type indexed by Idxs is a struct
    2949           0 :     for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
    2950             :       // Process each struct element recursively
    2951           0 :       Idxs.push_back(i);
    2952             :       Value *PrevTo = To;
    2953           0 :       To = BuildSubAggregate(From, To, STy->getElementType(i), Idxs, IdxSkip,
    2954             :                              InsertBefore);
    2955             :       Idxs.pop_back();
    2956           0 :       if (!To) {
    2957             :         // Couldn't find any inserted value for this index? Cleanup
    2958           0 :         while (PrevTo != OrigTo) {
    2959             :           InsertValueInst* Del = cast<InsertValueInst>(PrevTo);
    2960             :           PrevTo = Del->getAggregateOperand();
    2961           0 :           Del->eraseFromParent();
    2962             :         }
    2963             :         // Stop processing elements
    2964             :         break;
    2965             :       }
    2966             :     }
    2967             :     // If we successfully found a value for each of our subaggregates
    2968           0 :     if (To)
    2969             :       return To;
    2970             :   }
    2971             :   // Base case, the type indexed by SourceIdxs is not a struct, or not all of
    2972             :   // the struct's elements had a value that was inserted directly. In the latter
    2973             :   // case, perhaps we can't determine each of the subelements individually, but
    2974             :   // we might be able to find the complete struct somewhere.
    2975             : 
    2976             :   // Find the value that is at that particular spot
    2977           0 :   Value *V = FindInsertedValue(From, Idxs);
    2978             : 
    2979           0 :   if (!V)
    2980             :     return nullptr;
    2981             : 
    2982             :   // Insert the value in the new (sub) aggregrate
    2983           0 :   return InsertValueInst::Create(To, V, makeArrayRef(Idxs).slice(IdxSkip),
    2984           0 :                                  "tmp", InsertBefore);
    2985             : }
    2986             : 
    2987             : // This helper takes a nested struct and extracts a part of it (which is again a
    2988             : // struct) into a new value. For example, given the struct:
    2989             : // { a, { b, { c, d }, e } }
    2990             : // and the indices "1, 1" this returns
    2991             : // { c, d }.
    2992             : //
    2993             : // It does this by inserting an insertvalue for each element in the resulting
    2994             : // struct, as opposed to just inserting a single struct. This will only work if
    2995             : // each of the elements of the substruct are known (ie, inserted into From by an
    2996             : // insertvalue instruction somewhere).
    2997             : //
    2998             : // All inserted insertvalue instructions are inserted before InsertBefore
    2999           0 : static Value *BuildSubAggregate(Value *From, ArrayRef<unsigned> idx_range,
    3000             :                                 Instruction *InsertBefore) {
    3001             :   assert(InsertBefore && "Must have someplace to insert!");
    3002           0 :   Type *IndexedType = ExtractValueInst::getIndexedType(From->getType(),
    3003           0 :                                                              idx_range);
    3004           0 :   Value *To = UndefValue::get(IndexedType);
    3005             :   SmallVector<unsigned, 10> Idxs(idx_range.begin(), idx_range.end());
    3006           0 :   unsigned IdxSkip = Idxs.size();
    3007             : 
    3008           0 :   return BuildSubAggregate(From, To, IndexedType, Idxs, IdxSkip, InsertBefore);
    3009             : }
    3010             : 
    3011             : /// Given an aggregrate and an sequence of indices, see if
    3012             : /// the scalar value indexed is already around as a register, for example if it
    3013             : /// were inserted directly into the aggregrate.
    3014             : ///
    3015             : /// If InsertBefore is not null, this function will duplicate (modified)
    3016             : /// insertvalues when a part of a nested struct is extracted.
    3017          27 : Value *llvm::FindInsertedValue(Value *V, ArrayRef<unsigned> idx_range,
    3018             :                                Instruction *InsertBefore) {
    3019             :   // Nothing to index? Just return V then (this is useful at the end of our
    3020             :   // recursion).
    3021          27 :   if (idx_range.empty())
    3022             :     return V;
    3023             :   // We have indices, so V should have an indexable type.
    3024             :   assert((V->getType()->isStructTy() || V->getType()->isArrayTy()) &&
    3025             :          "Not looking at a struct or array?");
    3026             :   assert(ExtractValueInst::getIndexedType(V->getType(), idx_range) &&
    3027             :          "Invalid indices for type?");
    3028             : 
    3029             :   if (Constant *C = dyn_cast<Constant>(V)) {
    3030           0 :     C = C->getAggregateElement(idx_range[0]);
    3031           0 :     if (!C) return nullptr;
    3032           0 :     return FindInsertedValue(C, idx_range.slice(1), InsertBefore);
    3033             :   }
    3034             : 
    3035             :   if (InsertValueInst *I = dyn_cast<InsertValueInst>(V)) {
    3036             :     // Loop the indices for the insertvalue instruction in parallel with the
    3037             :     // requested indices
    3038             :     const unsigned *req_idx = idx_range.begin();
    3039          10 :     for (const unsigned *i = I->idx_begin(), *e = I->idx_end();
    3040          25 :          i != e; ++i, ++req_idx) {
    3041          15 :       if (req_idx == idx_range.end()) {
    3042             :         // We can't handle this without inserting insertvalues
    3043           0 :         if (!InsertBefore)
    3044             :           return nullptr;
    3045             : 
    3046             :         // The requested index identifies a part of a nested aggregate. Handle
    3047             :         // this specially. For example,
    3048             :         // %A = insertvalue { i32, {i32, i32 } } undef, i32 10, 1, 0
    3049             :         // %B = insertvalue { i32, {i32, i32 } } %A, i32 11, 1, 1
    3050             :         // %C = extractvalue {i32, { i32, i32 } } %B, 1
    3051             :         // This can be changed into
    3052             :         // %A = insertvalue {i32, i32 } undef, i32 10, 0
    3053             :         // %C = insertvalue {i32, i32 } %A, i32 11, 1
    3054             :         // which allows the unused 0,0 element from the nested struct to be
    3055             :         // removed.
    3056             :         return BuildSubAggregate(V, makeArrayRef(idx_range.begin(), req_idx),
    3057           0 :                                  InsertBefore);
    3058             :       }
    3059             : 
    3060             :       // This insert value inserts something else than what we are looking for.
    3061             :       // See if the (aggregate) value inserted into has the value we are
    3062             :       // looking for, then.
    3063          15 :       if (*req_idx != *i)
    3064             :         return FindInsertedValue(I->getAggregateOperand(), idx_range,
    3065           5 :                                  InsertBefore);
    3066             :     }
    3067             :     // If we end up here, the indices of the insertvalue match with those
    3068             :     // requested (though possibly only partially). Now we recursively look at
    3069             :     // the inserted value, passing any remaining indices.
    3070             :     return FindInsertedValue(I->getInsertedValueOperand(),
    3071             :                              makeArrayRef(req_idx, idx_range.end()),
    3072          10 :                              InsertBefore);
    3073             :   }
    3074             : 
    3075             :   if (ExtractValueInst *I = dyn_cast<ExtractValueInst>(V)) {
    3076             :     // If we're extracting a value from an aggregate that was extracted from
    3077             :     // something else, we can extract from that something else directly instead.
    3078             :     // However, we will need to chain I's indices with the requested indices.
    3079             : 
    3080             :     // Calculate the number of indices required
    3081           0 :     unsigned size = I->getNumIndices() + idx_range.size();
    3082             :     // Allocate some space to put the new indices in
    3083             :     SmallVector<unsigned, 5> Idxs;
    3084           0 :     Idxs.reserve(size);
    3085             :     // Add indices from the extract value instruction
    3086           0 :     Idxs.append(I->idx_begin(), I->idx_end());
    3087             : 
    3088             :     // Add requested indices
    3089           0 :     Idxs.append(idx_range.begin(), idx_range.end());
    3090             : 
    3091             :     assert(Idxs.size() == size
    3092             :            && "Number of indices added not correct?");
    3093             : 
    3094           0 :     return FindInsertedValue(I->getAggregateOperand(), Idxs, InsertBefore);
    3095             :   }
    3096             :   // Otherwise, we don't know (such as, extracting from a function return value
    3097             :   // or load instruction)
    3098             :   return nullptr;
    3099             : }
    3100             : 
    3101             : /// Analyze the specified pointer to see if it can be expressed as a base
    3102             : /// pointer plus a constant offset. Return the base and offset to the caller.
    3103      219740 : Value *llvm::GetPointerBaseWithConstantOffset(Value *Ptr, int64_t &Offset,
    3104             :                                               const DataLayout &DL) {
    3105      219740 :   unsigned BitWidth = DL.getIndexTypeSizeInBits(Ptr->getType());
    3106             :   APInt ByteOffset(BitWidth, 0);
    3107             : 
    3108             :   // We walk up the defs but use a visited set to handle unreachable code. In
    3109             :   // that case, we stop after accumulating the cycle once (not that it
    3110             :   // matters).
    3111             :   SmallPtrSet<Value *, 16> Visited;
    3112      377249 :   while (Visited.insert(Ptr).second) {
    3113      754498 :     if (Ptr->getType()->isVectorTy())
    3114             :       break;
    3115             : 
    3116             :     if (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
    3117             :       // If one of the values we have visited is an addrspacecast, then
    3118             :       // the pointer type of this GEP may be different from the type
    3119             :       // of the Ptr parameter which was passed to this function.  This
    3120             :       // means when we construct GEPOffset, we need to use the size
    3121             :       // of GEP's pointer type rather than the size of the original
    3122             :       // pointer type.
    3123      202494 :       APInt GEPOffset(DL.getIndexTypeSizeInBits(Ptr->getType()), 0);
    3124      202494 :       if (!GEP->accumulateConstantOffset(DL, GEPOffset))
    3125             :         break;
    3126             : 
    3127      138437 :       ByteOffset += GEPOffset.getSExtValue();
    3128             : 
    3129             :       Ptr = GEP->getPointerOperand();
    3130       58373 :     } else if (Operator::getOpcode(Ptr) == Instruction::BitCast ||
    3131             :                Operator::getOpcode(Ptr) == Instruction::AddrSpaceCast) {
    3132       19072 :       Ptr = cast<Operator>(Ptr)->getOperand(0);
    3133             :     } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
    3134             :       if (GA->isInterposable())
    3135             :         break;
    3136             :       Ptr = GA->getAliasee();
    3137             :     } else {
    3138             :       break;
    3139             :     }
    3140             :   }
    3141      219740 :   Offset = ByteOffset.getSExtValue();
    3142      219740 :   return Ptr;
    3143             : }
    3144             : 
    3145     4362963 : bool llvm::isGEPBasedOnPointerToString(const GEPOperator *GEP,
    3146             :                                        unsigned CharSize) {
    3147             :   // Make sure the GEP has exactly three arguments.
    3148     4362963 :   if (GEP->getNumOperands() != 3)
    3149             :     return false;
    3150             : 
    3151             :   // Make sure the index-ee is a pointer to array of \p CharSize integers.
    3152             :   // CharSize.
    3153     4361528 :   ArrayType *AT = dyn_cast<ArrayType>(GEP->getSourceElementType());
    3154     4342103 :   if (!AT || !AT->getElementType()->isIntegerTy(CharSize))
    3155             :     return false;
    3156             : 
    3157             :   // Check to make sure that the first operand of the GEP is an integer and
    3158             :   // has value 0 so that we are sure we're indexing into the initializer.
    3159        2105 :   const ConstantInt *FirstIdx = dyn_cast<ConstantInt>(GEP->getOperand(1));
    3160        2105 :   if (!FirstIdx || !FirstIdx->isZero())
    3161             :     return false;
    3162             : 
    3163             :   return true;
    3164             : }
    3165             : 
    3166     7364278 : bool llvm::getConstantDataArrayInfo(const Value *V,
    3167             :                                     ConstantDataArraySlice &Slice,
    3168             :                                     unsigned ElementSize, uint64_t Offset) {
    3169             :   assert(V);
    3170             : 
    3171             :   // Look through bitcast instructions and geps.
    3172     7364388 :   V = V->stripPointerCasts();
    3173             : 
    3174             :   // If the value is a GEP instruction or constant expression, treat it as an
    3175             :   // offset.
    3176             :   if (const GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
    3177             :     // The GEP operator should be based on a pointer to string constant, and is
    3178             :     // indexing into the string constant.
    3179     4360953 :     if (!isGEPBasedOnPointerToString(GEP, ElementSize))
    3180             :       return false;
    3181             : 
    3182             :     // If the second index isn't a ConstantInt, then this is a variable index
    3183             :     // into the array.  If this occurs, we can't say anything meaningful about
    3184             :     // the string.
    3185             :     uint64_t StartIdx = 0;
    3186         134 :     if (const ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(2)))
    3187             :       StartIdx = CI->getZExtValue();
    3188             :     else
    3189             :       return false;
    3190         110 :     return getConstantDataArrayInfo(GEP->getOperand(0), Slice, ElementSize,
    3191         110 :                                     StartIdx + Offset);
    3192             :   }
    3193             : 
    3194             :   // The GEP instruction, constant or instruction, must reference a global
    3195             :   // variable that is a constant and is initialized. The referenced constant
    3196             :   // initializer is the array that we'll use for optimization.
    3197             :   const GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
    3198     2995357 :   if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer())
    3199             :     return false;
    3200             : 
    3201             :   const ConstantDataArray *Array;
    3202             :   ArrayType *ArrayTy;
    3203        3945 :   if (GV->getInitializer()->isNullValue()) {
    3204          68 :     Type *GVTy = GV->getValueType();
    3205             :     if ( (ArrayTy = dyn_cast<ArrayType>(GVTy)) ) {
    3206             :       // A zeroinitializer for the array; there is no ConstantDataArray.
    3207             :       Array = nullptr;
    3208             :     } else {
    3209           6 :       const DataLayout &DL = GV->getParent()->getDataLayout();
    3210             :       uint64_t SizeInBytes = DL.getTypeStoreSize(GVTy);
    3211           6 :       uint64_t Length = SizeInBytes / (ElementSize / 8);
    3212           6 :       if (Length <= Offset)
    3213             :         return false;
    3214             : 
    3215           5 :       Slice.Array = nullptr;
    3216           5 :       Slice.Offset = 0;
    3217           5 :       Slice.Length = Length - Offset;
    3218           5 :       return true;
    3219             :     }
    3220             :   } else {
    3221             :     // This must be a ConstantDataArray.
    3222             :     Array = dyn_cast<ConstantDataArray>(GV->getInitializer());
    3223             :     if (!Array)
    3224             :       return false;
    3225             :     ArrayTy = Array->getType();
    3226             :   }
    3227        3852 :   if (!ArrayTy->getElementType()->isIntegerTy(ElementSize))
    3228             :     return false;
    3229             : 
    3230             :   uint64_t NumElts = ArrayTy->getArrayNumElements();
    3231        3811 :   if (Offset > NumElts)
    3232             :     return false;
    3233             : 
    3234        3811 :   Slice.Array = Array;
    3235        3811 :   Slice.Offset = Offset;
    3236        3811 :   Slice.Length = NumElts - Offset;
    3237        3811 :   return true;
    3238             : }
    3239             : 
    3240             : /// This function computes the length of a null-terminated C string pointed to
    3241             : /// by V. If successful, it returns true and returns the string in Str.
    3242             : /// If unsuccessful, it returns false.
    3243     7355710 : bool llvm::getConstantStringInfo(const Value *V, StringRef &Str,
    3244             :                                  uint64_t Offset, bool TrimAtNul) {
    3245             :   ConstantDataArraySlice Slice;
    3246     7355710 :   if (!getConstantDataArrayInfo(V, Slice, 8, Offset))
    3247             :     return false;
    3248             : 
    3249        2274 :   if (Slice.Array == nullptr) {
    3250          30 :     if (TrimAtNul) {
    3251          28 :       Str = StringRef();
    3252          28 :       return true;
    3253             :     }
    3254           2 :     if (Slice.Length == 1) {
    3255           2 :       Str = StringRef("", 1);
    3256           2 :       return true;
    3257             :     }
    3258             :     // We cannot instantiate a StringRef as we do not have an appropriate string
    3259             :     // of 0s at hand.
    3260             :     return false;
    3261             :   }
    3262             : 
    3263             :   // Start out with the entire array in the StringRef.
    3264        4488 :   Str = Slice.Array->getAsString();
    3265             :   // Skip over 'offset' bytes.
    3266        4488 :   Str = Str.substr(Slice.Offset);
    3267             : 
    3268        2244 :   if (TrimAtNul) {
    3269             :     // Trim off the \0 and anything after it.  If the array is not nul
    3270             :     // terminated, we just return the whole end of string.  The client may know
    3271             :     // some other way that the string is length-bound.
    3272        2228 :     Str = Str.substr(0, Str.find('\0'));
    3273             :   }
    3274             :   return true;
    3275             : }
    3276             : 
    3277             : // These next two are very similar to the above, but also look through PHI
    3278             : // nodes.
    3279             : // TODO: See if we can integrate these two together.
    3280             : 
    3281             : /// If we can compute the length of the string pointed to by
    3282             : /// the specified pointer, return 'len+1'.  If we can't, return 0.
    3283        6366 : static uint64_t GetStringLengthH(const Value *V,
    3284             :                                  SmallPtrSetImpl<const PHINode*> &PHIs,
    3285             :                                  unsigned CharSize) {
    3286             :   // Look through noop bitcast instructions.
    3287        6366 :   V = V->stripPointerCasts();
    3288             : 
    3289             :   // If this is a PHI node, there are two cases: either we have already seen it
    3290             :   // or we haven't.
    3291             :   if (const PHINode *PN = dyn_cast<PHINode>(V)) {
    3292          62 :     if (!PHIs.insert(PN).second)
    3293             :       return ~0ULL;  // already in the set.
    3294             : 
    3295             :     // If it was new, see if all the input strings are the same length.
    3296             :     uint64_t LenSoFar = ~0ULL;
    3297          74 :     for (Value *IncValue : PN->incoming_values()) {
    3298          66 :       uint64_t Len = GetStringLengthH(IncValue, PHIs, CharSize);
    3299          66 :       if (Len == 0) return 0; // Unknown length -> unknown.
    3300             : 
    3301           6 :       if (Len == ~0ULL) continue;
    3302             : 
    3303           6 :       if (Len != LenSoFar && LenSoFar != ~0ULL)
    3304             :         return 0;    // Disagree -> unknown.
    3305             :       LenSoFar = Len;
    3306             :     }
    3307             : 
    3308             :     // Success, all agree.
    3309             :     return LenSoFar;
    3310             :   }
    3311             : 
    3312             :   // strlen(select(c,x,y)) -> strlen(x) ^ strlen(y)
    3313             :   if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {
    3314          38 :     uint64_t Len1 = GetStringLengthH(SI->getTrueValue(), PHIs, CharSize);
    3315          38 :     if (Len1 == 0) return 0;
    3316          38 :     uint64_t Len2 = GetStringLengthH(SI->getFalseValue(), PHIs, CharSize);
    3317          38 :     if (Len2 == 0) return 0;
    3318          38 :     if (Len1 == ~0ULL) return Len2;
    3319          38 :     if (Len2 == ~0ULL) return Len1;
    3320          38 :     if (Len1 != Len2) return 0;
    3321           5 :     return Len1;
    3322             :   }
    3323             : 
    3324             :   // Otherwise, see if we can read the string.
    3325             :   ConstantDataArraySlice Slice;
    3326        6266 :   if (!getConstantDataArrayInfo(V, Slice, CharSize))
    3327             :     return 0;
    3328             : 
    3329        1394 :   if (Slice.Array == nullptr)
    3330             :     return 1;
    3331             : 
    3332             :   // Search for nul characters
    3333             :   unsigned NullIndex = 0;
    3334       17218 :   for (unsigned E = Slice.Length; NullIndex < E; ++NullIndex) {
    3335       17218 :     if (Slice.Array->getElementAsInteger(Slice.Offset + NullIndex) == 0)
    3336             :       break;
    3337             :   }
    3338             : 
    3339        1362 :   return NullIndex + 1;
    3340             : }
    3341             : 
    3342             : /// If we can compute the length of the string pointed to by
    3343             : /// the specified pointer, return 'len+1'.  If we can't, return 0.
    3344        6224 : uint64_t llvm::GetStringLength(const Value *V, unsigned CharSize) {
    3345       12448 :   if (!V->getType()->isPointerTy()) return 0;
    3346             : 
    3347             :   SmallPtrSet<const PHINode*, 32> PHIs;
    3348        6224 :   uint64_t Len = GetStringLengthH(V, PHIs, CharSize);
    3349             :   // If Len is ~0ULL, we had an infinite phi cycle: this is dead code, so return
    3350             :   // an empty string as a length.
    3351        6224 :   return Len == ~0ULL ? 1 : Len;
    3352             : }
    3353             : 
    3354             : /// \brief \p PN defines a loop-variant pointer to an object.  Check if the
    3355             : /// previous iteration of the loop was referring to the same object as \p PN.
    3356         270 : static bool isSameUnderlyingObjectInLoop(const PHINode *PN,
    3357             :                                          const LoopInfo *LI) {
    3358             :   // Find the loop-defined value.
    3359         270 :   Loop *L = LI->getLoopFor(PN->getParent());
    3360         270 :   if (PN->getNumIncomingValues() != 2)
    3361             :     return true;
    3362             : 
    3363             :   // Find the value from previous iteration.
    3364             :   auto *PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(0));
    3365         472 :   if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
    3366             :     PrevValue = dyn_cast<Instruction>(PN->getIncomingValue(1));
    3367         540 :   if (!PrevValue || LI->getLoopFor(PrevValue->getParent()) != L)
    3368             :     return true;
    3369             : 
    3370             :   // If a new pointer is loaded in the loop, the pointer references a different
    3371             :   // object in every iteration.  E.g.:
    3372             :   //    for (i)
    3373             :   //       int *p = a[i];
    3374             :   //       ...
    3375             :   if (auto *Load = dyn_cast<LoadInst>(PrevValue))
    3376           2 :     if (!L->isLoopInvariant(Load->getPointerOperand()))
    3377             :       return false;
    3378             :   return true;
    3379             : }
    3380             : 
    3381    63431393 : Value *llvm::GetUnderlyingObject(Value *V, const DataLayout &DL,
    3382             :                                  unsigned MaxLookup) {
    3383   126862786 :   if (!V->getType()->isPointerTy())
    3384             :     return V;
    3385   142430320 :   for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) {
    3386             :     if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) {
    3387             :       V = GEP->getPointerOperand();
    3388    22662293 :     } else if (Operator::getOpcode(V) == Instruction::BitCast ||
    3389             :                Operator::getOpcode(V) == Instruction::AddrSpaceCast) {
    3390     6252102 :       V = cast<Operator>(V)->getOperand(0);
    3391             :     } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) {
    3392             :       if (GA->isInterposable())
    3393             :         return V;
    3394             :       V = GA->getAliasee();
    3395             :     } else if (isa<AllocaInst>(V)) {
    3396             :       // An alloca can't be further simplified.
    3397             :       return V;
    3398             :     } else {
    3399    57779704 :       if (auto CS = CallSite(V))
    3400      220739 :         if (Value *RV = CS.getReturnedArgOperand()) {
    3401             :           V = RV;
    3402          81 :           continue;
    3403             :         }
    3404             : 
    3405             :       // See if InstructionSimplify knows any relevant tricks.
    3406             :       if (Instruction *I = dyn_cast<Instruction>(V))
    3407             :         // TODO: Acquire a DominatorTree and AssumptionCache and use them.
    3408     2530718 :         if (Value *Simplified = SimplifyInstruction(I, {DL, I})) {
    3409             :           V = Simplified;
    3410        8717 :           continue;
    3411             :         }
    3412             : 
    3413             :       return V;
    3414             :     }
    3415             :     assert(V->getType()->isPointerTy() && "Unexpected operand type!");
    3416             :   }
    3417             :   return V;
    3418             : }
    3419             : 
    3420     1810705 : void llvm::GetUnderlyingObjects(Value *V, SmallVectorImpl<Value *> &Objects,
    3421             :                                 const DataLayout &DL, LoopInfo *LI,
    3422             :                                 unsigned MaxLookup) {
    3423             :   SmallPtrSet<Value *, 4> Visited;
    3424             :   SmallVector<Value *, 4> Worklist;
    3425     1810705 :   Worklist.push_back(V);
    3426             :   do {
    3427     1931786 :     Value *P = Worklist.pop_back_val();
    3428     1931786 :     P = GetUnderlyingObject(P, DL, MaxLookup);
    3429             : 
    3430     1931786 :     if (!Visited.insert(P).second)
    3431      139388 :       continue;
    3432             : 
    3433     1891919 :     if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
    3434        1636 :       Worklist.push_back(SI->getTrueValue());
    3435        1636 :       Worklist.push_back(SI->getFalseValue());
    3436        1636 :       continue;
    3437             :     }
    3438             : 
    3439             :     if (PHINode *PN = dyn_cast<PHINode>(P)) {
    3440             :       // If this PHI changes the underlying object in every iteration of the
    3441             :       // loop, don't look through it.  Consider:
    3442             :       //   int **A;
    3443             :       //   for (i) {
    3444             :       //     Prev = Curr;     // Prev = PHI (Prev_0, Curr)
    3445             :       //     Curr = A[i];
    3446             :       //     *Prev, *Curr;
    3447             :       //
    3448             :       // Prev is tracking Curr one iteration behind so they refer to different
    3449             :       // underlying objects.
    3450       55016 :       if (!LI || !LI->isLoopHeader(PN->getParent()) ||
    3451         270 :           isSameUnderlyingObjectInLoop(PN, LI))
    3452      290362 :         for (Value *IncValue : PN->incoming_values())
    3453      117809 :           Worklist.push_back(IncValue);
    3454       54746 :       continue;
    3455             :     }
    3456             : 
    3457     1833901 :     Objects.push_back(P);
    3458     1931786 :   } while (!Worklist.empty());
    3459     1810705 : }
    3460             : 
    3461             : /// This is the function that does the work of looking through basic
    3462             : /// ptrtoint+arithmetic+inttoptr sequences.
    3463        2931 : static const Value *getUnderlyingObjectFromInt(const Value *V) {
    3464             :   do {
    3465             :     if (const Operator *U = dyn_cast<Operator>(V)) {
    3466             :       // If we find a ptrtoint, we can transfer control back to the
    3467             :       // regular getUnderlyingObjectFromInt.
    3468        3247 :       if (U->getOpcode() == Instruction::PtrToInt)
    3469         348 :         return U->getOperand(0);
    3470             :       // If we find an add of a constant, a multiplied value, or a phi, it's
    3471             :       // likely that the other operand will lead us to the base
    3472             :       // object. We don't have to worry about the case where the
    3473             :       // object address is somehow being computed by the multiply,
    3474             :       // because our callers only care when the result is an
    3475             :       // identifiable object.
    3476        4301 :       if (U->getOpcode() != Instruction::Add ||
    3477        1517 :           (!isa<ConstantInt>(U->getOperand(1)) &&
    3478             :            Operator::getOpcode(U->getOperand(1)) != Instruction::Mul &&
    3479             :            !isa<PHINode>(U->getOperand(1))))
    3480             :         return V;
    3481             :       V = U->getOperand(0);
    3482             :     } else {
    3483             :       return V;
    3484             :     }
    3485         789 :     assert(V->getType()->isIntegerTy() && "Unexpected operand type!");
    3486             :   } while (true);
    3487             : }
    3488             : 
    3489             : /// This is a wrapper around GetUnderlyingObjects and adds support for basic
    3490             : /// ptrtoint+arithmetic+inttoptr sequences.
    3491             : /// It returns false if unidentified object is found in GetUnderlyingObjects.
    3492     1226823 : bool llvm::getUnderlyingObjectsForCodeGen(const Value *V,
    3493             :                           SmallVectorImpl<Value *> &Objects,
    3494             :                           const DataLayout &DL) {
    3495             :   SmallPtrSet<const Value *, 16> Visited;
    3496             :   SmallVector<const Value *, 4> Working(1, V);
    3497             :   do {
    3498     1226997 :     V = Working.pop_back_val();
    3499             : 
    3500             :     SmallVector<Value *, 4> Objs;
    3501     1226997 :     GetUnderlyingObjects(const_cast<Value *>(V), Objs, DL);
    3502             : 
    3503     3361415 :     for (Value *V : Objs) {
    3504     1227623 :       if (!Visited.insert(V).second)
    3505           0 :         continue;
    3506      109508 :       if (Operator::getOpcode(V) == Instruction::IntToPtr) {
    3507             :         const Value *O =
    3508        2931 :           getUnderlyingObjectFromInt(cast<User>(V)->getOperand(0));
    3509        6036 :         if (O->getType()->isPointerTy()) {
    3510         174 :           Working.push_back(O);
    3511         174 :           continue;
    3512             :         }
    3513             :       }
    3514             :       // If GetUnderlyingObjects fails to find an identifiable object,
    3515             :       // getUnderlyingObjectsForCodeGen also fails for safety.
    3516     1227449 :       if (!isIdentifiedObject(V)) {
    3517             :         Objects.clear();
    3518             :         return false;
    3519             :       }
    3520     1067035 :       Objects.push_back(const_cast<Value *>(V));
    3521             :     }
    3522     1066583 :   } while (!Working.empty());
    3523             :   return true;
    3524             : }
    3525             : 
    3526             : /// Return true if the only users of this pointer are lifetime markers.
    3527        3309 : bool llvm::onlyUsedByLifetimeMarkers(const Value *V) {
    3528        3487 :   for (const User *U : V->users()) {
    3529             :     const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
    3530             :     if (!II) return false;
    3531             : 
    3532        3270 :     if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
    3533             :         II->getIntrinsicID() != Intrinsic::lifetime_end)
    3534             :       return false;
    3535             :   }
    3536             :   return true;
    3537             : }
    3538             : 
    3539     1533876 : bool llvm::isSafeToSpeculativelyExecute(const Value *V,
    3540             :                                         const Instruction *CtxI,
    3541             :                                         const DominatorTree *DT) {
    3542             :   const Operator *Inst = dyn_cast<Operator>(V);
    3543             :   if (!Inst)
    3544             :     return false;
    3545             : 
    3546     7673179 :   for (unsigned i = 0, e = Inst->getNumOperands(); i != e; ++i)
    3547     3069766 :     if (Constant *C = dyn_cast<Constant>(Inst->getOperand(i)))
    3548     1337096 :       if (C->canTrap())
    3549             :         return false;
    3550             : 
    3551     1533654 :   switch (Inst->getOpcode()) {
    3552             :   default:
    3553             :     return true;
    3554             :   case Instruction::UDiv:
    3555             :   case Instruction::URem: {
    3556             :     // x / y is undefined if y == 0.
    3557             :     const APInt *V;
    3558        2436 :     if (match(Inst->getOperand(1), m_APInt(V)))
    3559         794 :       return *V != 0;
    3560             :     return false;
    3561             :   }
    3562             :   case Instruction::SDiv:
    3563             :   case Instruction::SRem: {
    3564             :     // x / y is undefined if y == 0 or x == INT_MIN and y == -1
    3565             :     const APInt *Numerator, *Denominator;
    3566        1498 :     if (!match(Inst->getOperand(1), m_APInt(Denominator)))
    3567             :       return false;
    3568             :     // We cannot hoist this division if the denominator is 0.
    3569         591 :     if (*Denominator == 0)
    3570             :       return false;
    3571             :     // It's safe to hoist if the denominator is not 0 or -1.
    3572         583 :     if (*Denominator != -1)
    3573             :       return true;
    3574             :     // At this point we know that the denominator is -1.  It is safe to hoist as
    3575             :     // long we know that the numerator is not INT_MIN.
    3576           2 :     if (match(Inst->getOperand(0), m_APInt(Numerator)))
    3577           0 :       return !Numerator->isMinSignedValue();
    3578             :     // The numerator *might* be MinSignedValue.
    3579             :     return false;
    3580             :   }
    3581             :   case Instruction::Load: {
    3582             :     const LoadInst *LI = cast<LoadInst>(Inst);
    3583       94029 :     if (!LI->isUnordered() ||
    3584             :         // Speculative load may create a race that did not exist in the source.
    3585      188041 :         LI->getFunction()->hasFnAttribute(Attribute::SanitizeThread) ||
    3586             :         // Speculative load may load data from dirty regions.
    3587      188006 :         LI->getFunction()->hasFnAttribute(Attribute::SanitizeAddress) ||
    3588       93994 :         LI->getFunction()->hasFnAttribute(Attribute::SanitizeHWAddress))
    3589             :       return false;
    3590       93976 :     const DataLayout &DL = LI->getModule()->getDataLayout();
    3591             :     return isDereferenceableAndAlignedPointer(LI->getPointerOperand(),
    3592       93976 :                                               LI->getAlignment(), DL, CtxI, DT);
    3593             :   }
    3594             :   case Instruction::Call: {
    3595             :     auto *CI = cast<const CallInst>(Inst);
    3596             :     const Function *Callee = CI->getCalledFunction();
    3597             : 
    3598             :     // The called function could have undefined behavior or side-effects, even
    3599             :     // if marked readnone nounwind.
    3600       96525 :     return Callee && Callee->isSpeculatable();
    3601             :   }
    3602      232602 :   case Instruction::VAArg:
    3603             :   case Instruction::Alloca:
    3604             :   case Instruction::Invoke:
    3605             :   case Instruction::PHI:
    3606             :   case Instruction::Store:
    3607             :   case Instruction::Ret:
    3608             :   case Instruction::Br:
    3609             :   case Instruction::IndirectBr:
    3610             :   case Instruction::Switch:
    3611             :   case Instruction::Unreachable:
    3612             :   case Instruction::Fence:
    3613             :   case Instruction::AtomicRMW:
    3614             :   case Instruction::AtomicCmpXchg:
    3615             :   case Instruction::LandingPad:
    3616             :   case Instruction::Resume:
    3617             :   case Instruction::CatchSwitch:
    3618             :   case Instruction::CatchPad:
    3619             :   case Instruction::CatchRet:
    3620             :   case Instruction::CleanupPad:
    3621             :   case Instruction::CleanupRet:
    3622      232602 :     return false; // Misc instructions which have effects
    3623             :   }
    3624             : }
    3625             : 
    3626     2231545 : bool llvm::mayBeMemoryDependent(const Instruction &I) {
    3627     2231545 :   return I.mayReadOrWriteMemory() || !isSafeToSpeculativelyExecute(&I);
    3628             : }
    3629             : 
    3630        6792 : OverflowResult llvm::computeOverflowForUnsignedMul(const Value *LHS,
    3631             :                                                    const Value *RHS,
    3632             :                                                    const DataLayout &DL,
    3633             :                                                    AssumptionCache *AC,
    3634             :                                                    const Instruction *CxtI,
    3635             :                                                    const DominatorTree *DT) {
    3636             :   // Multiplying n * m significant bits yields a result of n + m significant
    3637             :   // bits. If the total number of significant bits does not exceed the
    3638             :   // result bit width (minus 1), there is no overflow.
    3639             :   // This means if we have enough leading zero bits in the operands
    3640             :   // we can guarantee that the result does not overflow.
    3641             :   // Ref: "Hacker's Delight" by Henry Warren
    3642        6792 :   unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
    3643       13584 :   KnownBits LHSKnown(BitWidth);
    3644       13584 :   KnownBits RHSKnown(BitWidth);
    3645        6792 :   computeKnownBits(LHS, LHSKnown, DL, /*Depth=*/0, AC, CxtI, DT);
    3646        6792 :   computeKnownBits(RHS, RHSKnown, DL, /*Depth=*/0, AC, CxtI, DT);
    3647             :   // Note that underestimating the number of zero bits gives a more
    3648             :   // conservative answer.
    3649             :   unsigned ZeroBits = LHSKnown.countMinLeadingZeros() +
    3650        6792 :                       RHSKnown.countMinLeadingZeros();
    3651             :   // First handle the easy case: if we have enough zero bits there's
    3652             :   // definitely no overflow.
    3653        6792 :   if (ZeroBits >= BitWidth)
    3654             :     return OverflowResult::NeverOverflows;
    3655             : 
    3656             :   // Get the largest possible values for each operand.
    3657        6745 :   APInt LHSMax = ~LHSKnown.Zero;
    3658        6745 :   APInt RHSMax = ~RHSKnown.Zero;
    3659             : 
    3660             :   // We know the multiply operation doesn't overflow if the maximum values for
    3661             :   // each operand will not overflow after we multiply them together.
    3662             :   bool MaxOverflow;
    3663       13490 :   (void)LHSMax.umul_ov(RHSMax, MaxOverflow);
    3664        6745 :   if (!MaxOverflow)
    3665             :     return OverflowResult::NeverOverflows;
    3666             : 
    3667             :   // We know it always overflows if multiplying the smallest possible values for
    3668             :   // the operands also results in overflow.
    3669             :   bool MinOverflow;
    3670       13458 :   (void)LHSKnown.One.umul_ov(RHSKnown.One, MinOverflow);
    3671        6729 :   if (MinOverflow)
    3672             :     return OverflowResult::AlwaysOverflows;
    3673             : 
    3674        6726 :   return OverflowResult::MayOverflow;
    3675             : }
    3676             : 
    3677     3263248 : OverflowResult llvm::computeOverflowForUnsignedAdd(const Value *LHS,
    3678             :                                                    const Value *RHS,
    3679             :                                                    const DataLayout &DL,
    3680             :                                                    AssumptionCache *AC,
    3681             :                                                    const Instruction *CxtI,
    3682             :                                                    const DominatorTree *DT) {
    3683     6526496 :   KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT);
    3684     6522101 :   if (LHSKnown.isNonNegative() || LHSKnown.isNegative()) {
    3685        7884 :     KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT);
    3686             : 
    3687        4509 :     if (LHSKnown.isNegative() && RHSKnown.isNegative()) {
    3688             :       // The sign bit is set in both cases: this MUST overflow.
    3689             :       // Create a simple add instruction, and insert it into the struct.
    3690        1020 :       return OverflowResult::AlwaysOverflows;
    3691             :     }
    3692             : 
    3693        8829 :     if (LHSKnown.isNonNegative() && RHSKnown.isNonNegative()) {
    3694             :       // The sign bit is clear in both cases: this CANNOT overflow.
    3695             :       // Create a simple add instruction, and insert it into the struct.
    3696             :       return OverflowResult::NeverOverflows;
    3697             :     }
    3698             :   }
    3699             : 
    3700             :   return OverflowResult::MayOverflow;
    3701             : }
    3702             : 
    3703             : /// \brief Return true if we can prove that adding the two values of the
    3704             : /// knownbits will not overflow.
    3705             : /// Otherwise return false.
    3706     3248757 : static bool checkRippleForSignedAdd(const KnownBits &LHSKnown,
    3707             :                                     const KnownBits &RHSKnown) {
    3708             :   // Addition of two 2's complement numbers having opposite signs will never
    3709             :   // overflow.
    3710     6497553 :   if ((LHSKnown.isNegative() && RHSKnown.isNonNegative()) ||
    3711        1021 :       (LHSKnown.isNonNegative() && RHSKnown.isNegative()))
    3712             :     return true;
    3713             : 
    3714             :   // If either of the values is known to be non-negative, adding them can only
    3715             :   // overflow if the second is also non-negative, so we can assume that.
    3716             :   // Two non-negative numbers will only overflow if there is a carry to the 
    3717             :   // sign bit, so we can check if even when the values are as big as possible
    3718             :   // there is no overflow to the sign bit.
    3719     6496466 :   if (LHSKnown.isNonNegative() || RHSKnown.isNonNegative()) {
    3720     6446912 :     APInt MaxLHS = ~LHSKnown.Zero;
    3721             :     MaxLHS.clearSignBit();
    3722     6446912 :     APInt MaxRHS = ~RHSKnown.Zero;
    3723             :     MaxRHS.clearSignBit();
    3724             :     APInt Result = std::move(MaxLHS) + std::move(MaxRHS);
    3725             :     return Result.isSignBitClear();
    3726             :   }
    3727             : 
    3728             :   // If either of the values is known to be negative, adding them can only
    3729             :   // overflow if the second is also negative, so we can assume that.
    3730             :   // Two negative number will only overflow if there is no carry to the sign
    3731             :   // bit, so we can check if even when the values are as small as possible
    3732             :   // there is overflow to the sign bit.
    3733       50511 :   if (LHSKnown.isNegative() || RHSKnown.isNegative()) {
    3734       10511 :     APInt MinLHS = LHSKnown.One;
    3735       10511 :     MinLHS.clearSignBit();
    3736       10511 :     APInt MinRHS = RHSKnown.One;
    3737       10511 :     MinRHS.clearSignBit();
    3738             :     APInt Result = std::move(MinLHS) + std::move(MinRHS);
    3739             :     return Result.isSignBitSet();
    3740             :   }
    3741             : 
    3742             :   // If we reached here it means that we know nothing about the sign bits.
    3743             :   // In this case we can't know if there will be an overflow, since by 
    3744             :   // changing the sign bits any two values can be made to overflow.
    3745             :   return false;
    3746             : }
    3747             : 
    3748     3249671 : static OverflowResult computeOverflowForSignedAdd(const Value *LHS,
    3749             :                                                   const Value *RHS,
    3750             :                                                   const AddOperator *Add,
    3751             :                                                   const DataLayout &DL,
    3752             :                                                   AssumptionCache *AC,
    3753             :                                                   const Instruction *CxtI,
    3754             :                                                   const DominatorTree *DT) {
    3755     3249709 :   if (Add && Add->hasNoSignedWrap()) {
    3756             :     return OverflowResult::NeverOverflows;
    3757             :   }
    3758             : 
    3759             :   // If LHS and RHS each have at least two sign bits, the addition will look
    3760             :   // like
    3761             :   //
    3762             :   // XX..... +
    3763             :   // YY.....
    3764             :   //
    3765             :   // If the carry into the most significant position is 0, X and Y can't both
    3766             :   // be 1 and therefore the carry out of the addition is also 0.
    3767             :   //
    3768             :   // If the carry into the most significant position is 1, X and Y can't both
    3769             :   // be 0 and therefore the carry out of the addition is also 1.
    3770             :   //
    3771             :   // Since the carry into the most significant position is always equal to
    3772             :   // the carry out of the addition, there is no signed overflow.
    3773     3251110 :   if (ComputeNumSignBits(LHS, DL, 0, AC, CxtI, DT) > 1 &&
    3774        1463 :       ComputeNumSignBits(RHS, DL, 0, AC, CxtI, DT) > 1)
    3775             :     return OverflowResult::NeverOverflows;
    3776             : 
    3777     6497514 :   KnownBits LHSKnown = computeKnownBits(LHS, DL, /*Depth=*/0, AC, CxtI, DT);
    3778     6497514 :   KnownBits RHSKnown = computeKnownBits(RHS, DL, /*Depth=*/0, AC, CxtI, DT);
    3779             : 
    3780     3248757 :   if (checkRippleForSignedAdd(LHSKnown, RHSKnown))
    3781             :     return OverflowResult::NeverOverflows;
    3782             : 
    3783             :   // The remaining code needs Add to be available. Early returns if not so.
    3784     3248713 :   if (!Add)
    3785             :     return OverflowResult::MayOverflow;
    3786             : 
    3787             :   // If the sign of Add is the same as at least one of the operands, this add
    3788             :   // CANNOT overflow. This is particularly useful when the sum is
    3789             :   // @llvm.assume'ed non-negative rather than proved so from analyzing its
    3790             :   // operands.
    3791             :   bool LHSOrRHSKnownNonNegative =
    3792          14 :       (LHSKnown.isNonNegative() || RHSKnown.isNonNegative());
    3793             :   bool LHSOrRHSKnownNegative = 
    3794          16 :       (LHSKnown.isNegative() || RHSKnown.isNegative());
    3795           8 :   if (LHSOrRHSKnownNonNegative || LHSOrRHSKnownNegative) {
    3796          10 :     KnownBits AddKnown = computeKnownBits(Add, DL, /*Depth=*/0, AC, CxtI, DT);
    3797          10 :     if ((AddKnown.isNonNegative() && LHSOrRHSKnownNonNegative) ||
    3798           0 :         (AddKnown.isNegative() && LHSOrRHSKnownNegative)) {
    3799           2 :       return OverflowResult::NeverOverflows;
    3800             :     }
    3801             :   }
    3802             : 
    3803             :   return OverflowResult::MayOverflow;
    3804             : }
    3805             : 
    3806         138 : bool llvm::isOverflowIntrinsicNoWrap(const IntrinsicInst *II,
    3807             :                                      const DominatorTree &DT) {
    3808             : #ifndef NDEBUG
    3809             :   auto IID = II->getIntrinsicID();
    3810             :   assert((IID == Intrinsic::sadd_with_overflow ||
    3811             :           IID == Intrinsic::uadd_with_overflow ||
    3812             :           IID == Intrinsic::ssub_with_overflow ||
    3813             :           IID == Intrinsic::usub_with_overflow ||
    3814             :           IID == Intrinsic::smul_with_overflow ||
    3815             :           IID == Intrinsic::umul_with_overflow) &&
    3816             :          "Not an overflow intrinsic!");
    3817             : #endif
    3818             : 
    3819             :   SmallVector<const BranchInst *, 2> GuardingBranches;
    3820             :   SmallVector<const ExtractValueInst *, 2> Results;
    3821             : 
    3822         414 :   for (const User *U : II->users()) {
    3823         276 :     if (const auto *EVI = dyn_cast<ExtractValueInst>(U)) {
    3824             :       assert(EVI->getNumIndices() == 1 && "Obvious from CI's type");
    3825             : 
    3826         276 :       if (EVI->getIndices()[0] == 0)
    3827         138 :         Results.push_back(EVI);
    3828             :       else {
    3829             :         assert(EVI->getIndices()[0] == 1 && "Obvious from CI's type");
    3830             : 
    3831         276 :         for (const auto *U : EVI->users())
    3832         138 :           if (const auto *B = dyn_cast<BranchInst>(U)) {
    3833             :             assert(B->isConditional() && "How else is it using an i1?");
    3834         106 :             GuardingBranches.push_back(B);
    3835             :           }
    3836             :       }
    3837             :     } else {
    3838             :       // We are using the aggregate directly in a way we don't want to analyze
    3839             :       // here (storing it to a global, say).
    3840           0 :       return false;
    3841             :     }
    3842             :   }
    3843             : 
    3844         104 :   auto AllUsesGuardedByBranch = [&](const BranchInst *BI) {
    3845         104 :     BasicBlockEdge NoWrapEdge(BI->getParent(), BI->getSuccessor(1));
    3846         104 :     if (!NoWrapEdge.isSingleEdge())
    3847             :       return false;
    3848             : 
    3849             :     // Check if all users of the add are provably no-wrap.
    3850         384 :     for (const auto *Result : Results) {
    3851             :       // If the extractvalue itself is not executed on overflow, the we don't
    3852             :       // need to check each use separately, since domination is transitive.
    3853         129 :       if (DT.dominates(NoWrapEdge, Result->getParent()))
    3854          79 :         continue;
    3855             : 
    3856          34 :       for (auto &RU : Result->uses())
    3857          25 :         if (!DT.dominates(NoWrapEdge, RU))
    3858             :           return false;
    3859             :     }
    3860             : 
    3861             :     return true;
    3862             :   };
    3863             : 
    3864         138 :   return llvm::any_of(GuardingBranches, AllUsesGuardedByBranch);
    3865             : }
    3866             : 
    3867             : 
    3868          38 : OverflowResult llvm::computeOverflowForSignedAdd(const AddOperator *Add,
    3869             :                                                  const DataLayout &DL,
    3870             :                                                  AssumptionCache *AC,
    3871             :                                                  const Instruction *CxtI,
    3872             :                                                  const DominatorTree *DT) {
    3873          38 :   return ::computeOverflowForSignedAdd(Add->getOperand(0), Add->getOperand(1),
    3874          38 :                                        Add, DL, AC, CxtI, DT);
    3875             : }
    3876             : 
    3877     3249633 : OverflowResult llvm::computeOverflowForSignedAdd(const Value *LHS,
    3878             :                                                  const Value *RHS,
    3879             :                                                  const DataLayout &DL,
    3880             :                                                  AssumptionCache *AC,
    3881             :                                                  const Instruction *CxtI,
    3882             :                                                  const DominatorTree *DT) {
    3883     3249633 :   return ::computeOverflowForSignedAdd(LHS, RHS, nullptr, DL, AC, CxtI, DT);
    3884             : }
    3885             : 
    3886     3678958 : bool llvm::isGuaranteedToTransferExecutionToSuccessor(const Instruction *I) {
    3887             :   // A memory operation returns normally if it isn't volatile. A volatile
    3888             :   // operation is allowed to trap.
    3889             :   //
    3890             :   // An atomic operation isn't guaranteed to return in a reasonable amount of
    3891             :   // time because it's possible for another thread to interfere with it for an
    3892             :   // arbitrary length of time, but programs aren't allowed to rely on that.
    3893             :   if (const LoadInst *LI = dyn_cast<LoadInst>(I))
    3894      648445 :     return !LI->isVolatile();
    3895             :   if (const StoreInst *SI = dyn_cast<StoreInst>(I))
    3896      658380 :     return !SI->isVolatile();
    3897             :   if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(I))
    3898          67 :     return !CXI->isVolatile();
    3899             :   if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(I))
    3900         312 :     return !RMWI->isVolatile();
    3901             :   if (const MemIntrinsic *MII = dyn_cast<MemIntrinsic>(I))
    3902        3797 :     return !MII->isVolatile();
    3903             : 
    3904             :   // If there is no successor, then execution can't transfer to it.
    3905             :   if (const auto *CRI = dyn_cast<CleanupReturnInst>(I))
    3906           1 :     return !CRI->unwindsToCaller();
    3907             :   if (const auto *CatchSwitch = dyn_cast<CatchSwitchInst>(I))
    3908           2 :     return !CatchSwitch->unwindsToCaller();
    3909     2367954 :   if (isa<ResumeInst>(I))
    3910             :     return false;
    3911     2366143 :   if (isa<ReturnInst>(I))
    3912             :     return false;
    3913     2341629 :   if (isa<UnreachableInst>(I))
    3914             :     return false;
    3915             : 
    3916             :   // Calls can throw, or contain an infinite loop, or kill the process.
    3917     2341096 :   if (auto CS = ImmutableCallSite(I)) {
    3918             :     // Call sites that throw have implicit non-local control flow.
    3919      465235 :     if (!CS.doesNotThrow())
    3920      465235 :       return false;
    3921             : 
    3922             :     // Non-throwing call sites can loop infinitely, call exit/pthread_exit
    3923             :     // etc. and thus not return.  However, LLVM already assumes that
    3924             :     //
    3925             :     //  - Thread exiting actions are modeled as writes to memory invisible to
    3926             :     //    the program.
    3927             :     //
    3928             :     //  - Loops that don't have side effects (side effects are volatile/atomic
    3929             :     //    stores and IO) always terminate (see http://llvm.org/PR965).
    3930             :     //    Furthermore IO itself is also modeled as writes to memory invisible to
    3931             :     //    the program.
    3932             :     //
    3933             :     // We rely on those assumptions here, and use the memory effects of the call
    3934             :     // target as a proxy for checking that it always returns.
    3935             : 
    3936             :     // FIXME: This isn't aggressive enough; a call which only writes to a global
    3937             :     // is guaranteed to return.
    3938      726914 :     return CS.onlyReadsMemory() || CS.onlyAccessesArgMemory() ||
    3939      546205 :            match(I, m_Intrinsic<Intrinsic::assume>()) ||
    3940             :            match(I, m_Intrinsic<Intrinsic::sideeffect>());
    3941             :   }
    3942             : 
    3943             :   // Other instructions return normally.
    3944     1875861 :   return true;
    3945             : }
    3946             : 
    3947        4684 : bool llvm::isGuaranteedToExecuteForEveryIteration(const Instruction *I,
    3948             :                                                   const Loop *L) {
    3949             :   // The loop header is guaranteed to be executed for every iteration.
    3950             :   //
    3951             :   // FIXME: Relax this constraint to cover all basic blocks that are
    3952             :   // guaranteed to be executed at every iteration.
    3953        9368 :   if (I->getParent() != L->getHeader()) return false;
    3954             : 
    3955      522757 :   for (const Instruction &LI : *L->getHeader()) {
    3956      522757 :     if (&LI == I) return true;
    3957      518387 :     if (!isGuaranteedToTransferExecutionToSuccessor(&LI)) return false;
    3958             :   }
    3959           0 :   llvm_unreachable("Instruction not contained in its own parent basic block.");
    3960             : }
    3961             : 
    3962      306318 : bool llvm::propagatesFullPoison(const Instruction *I) {
    3963             :   switch (I->getOpcode()) {
    3964             :   case Instruction::Add:
    3965             :   case Instruction::Sub:
    3966             :   case Instruction::Xor:
    3967             :   case Instruction::Trunc:
    3968             :   case Instruction::BitCast:
    3969             :   case Instruction::AddrSpaceCast:
    3970             :   case Instruction::Mul:
    3971             :   case Instruction::Shl:
    3972             :   case Instruction::GetElementPtr:
    3973             :     // These operations all propagate poison unconditionally. Note that poison
    3974             :     // is not any particular value, so xor or subtraction of poison with
    3975             :     // itself still yields poison, not zero.
    3976             :     return true;
    3977             : 
    3978             :   case Instruction::AShr:
    3979             :   case Instruction::SExt:
    3980             :     // For these operations, one bit of the input is replicated across
    3981             :     // multiple output bits. A replicated poison bit is still poison.
    3982             :     return true;
    3983             : 
    3984             :   case Instruction::ICmp:
    3985             :     // Comparing poison with any value yields poison.  This is why, for
    3986             :     // instance, x s< (x +nsw 1) can be folded to true.
    3987             :     return true;
    3988             : 
    3989      123485 :   default:
    3990      123485 :     return false;
    3991             :   }
    3992             : }
    3993             : 
    3994      307112 : const Value *llvm::getGuaranteedNonFullPoisonOp(const Instruction *I) {
    3995      307112 :   switch (I->getOpcode()) {
    3996             :     case Instruction::Store:
    3997             :       return cast<StoreInst>(I)->getPointerOperand();
    3998             : 
    3999             :     case Instruction::Load:
    4000             :       return cast<LoadInst>(I)->getPointerOperand();
    4001             : 
    4002             :     case Instruction::AtomicCmpXchg:
    4003             :       return cast<AtomicCmpXchgInst>(I)->getPointerOperand();
    4004             : 
    4005             :     case Instruction::AtomicRMW:
    4006             :       return cast<AtomicRMWInst>(I)->getPointerOperand();
    4007             : 
    4008         856 :     case Instruction::UDiv:
    4009             :     case Instruction::SDiv:
    4010             :     case Instruction::URem:
    4011             :     case Instruction::SRem:
    4012        1712 :       return I->getOperand(1);
    4013             : 
    4014             :     default:
    4015             :       return nullptr;
    4016             :   }
    4017             : }
    4018             : 
    4019       40367 : bool llvm::programUndefinedIfFullPoison(const Instruction *PoisonI) {
    4020             :   // We currently only look for uses of poison values within the same basic
    4021             :   // block, as that makes it easier to guarantee that the uses will be
    4022             :   // executed given that PoisonI is executed.
    4023             :   //
    4024             :   // FIXME: Expand this to consider uses beyond the same basic block. To do
    4025             :   // this, look out for the distinction between post-dominance and strong
    4026             :   // post-dominance.
    4027       40367 :   const BasicBlock *BB = PoisonI->getParent();
    4028             : 
    4029             :   // Set of instructions that we have proved will yield poison if PoisonI
    4030             :   // does.
    4031             :   SmallSet<const Value *, 16> YieldsPoison;
    4032             :   SmallSet<const BasicBlock *, 4> Visited;
    4033       40367 :   YieldsPoison.insert(PoisonI);
    4034       40367 :   Visited.insert(PoisonI->getParent());
    4035             : 
    4036       40367 :   BasicBlock::const_iterator Begin = PoisonI->getIterator(), End = BB->end();
    4037             : 
    4038             :   unsigned Iter = 0;
    4039       41591 :   while (Iter++ < MaxDepth) {
    4040      379951 :     for (auto &I : make_range(Begin, End)) {
    4041      347479 :       if (&I != PoisonI) {
    4042      307112 :         const Value *NotPoison = getGuaranteedNonFullPoisonOp(&I);
    4043      307112 :         if (NotPoison != nullptr && YieldsPoison.count(NotPoison))
    4044             :           return true;
    4045      300455 :         if (!isGuaranteedToTransferExecutionToSuccessor(&I))
    4046             :           return false;
    4047             :       }
    4048             : 
    4049             :       // Mark poison that propagates from I through uses of I.
    4050      338972 :       if (YieldsPoison.count(&I)) {
    4051      428772 :         for (const User *User : I.users()) {
    4052             :           const Instruction *UserI = cast<Instruction>(User);
    4053      237483 :           if (propagatesFullPoison(UserI))
    4054      160868 :             YieldsPoison.insert(User);
    4055             :         }
    4056             :       }
    4057             :     }
    4058             : 
    4059       32472 :     if (auto *NextBB = BB->getSingleSuccessor()) {
    4060        1351 :       if (Visited.insert(NextBB).second) {
    4061             :         BB = NextBB;
    4062         612 :         Begin = BB->getFirstNonPHI()->getIterator();
    4063             :         End = BB->end();
    4064             :         continue;
    4065             :       }
    4066             :     }
    4067             : 
    4068             :     break;
    4069             :   }
    4070             :   return false;
    4071             : }
    4072             : 
    4073       14288 : static bool isKnownNonNaN(const Value *V, FastMathFlags FMF) {
    4074       14288 :   if (FMF.noNaNs())
    4075             :     return true;
    4076             : 
    4077             :   if (auto *C = dyn_cast<ConstantFP>(V))
    4078        2608 :     return !C->isNaN();
    4079             :   return false;
    4080             : }
    4081             : 
    4082             : static bool isKnownNonZero(const Value *V) {
    4083             :   if (auto *C = dyn_cast<ConstantFP>(V))
    4084        1111 :     return !C->isZero();
    4085             :   return false;
    4086             : }
    4087             : 
    4088             : /// Match clamp pattern for float types without care about NaNs or signed zeros.
    4089             : /// Given non-min/max outer cmp/select from the clamp pattern this
    4090             : /// function recognizes if it can be substitued by a "canonical" min/max
    4091             : /// pattern.
    4092         100 : static SelectPatternResult matchFastFloatClamp(CmpInst::Predicate Pred,
    4093             :                                                Value *CmpLHS, Value *CmpRHS,
    4094             :                                                Value *TrueVal, Value *FalseVal,
    4095             :                                                Value *&LHS, Value *&RHS) {
    4096             :   // Try to match
    4097             :   //   X < C1 ? C1 : Min(X, C2) --> Max(C1, Min(X, C2))
    4098             :   //   X > C1 ? C1 : Max(X, C2) --> Min(C1, Max(X, C2))
    4099             :   // and return description of the outer Max/Min.
    4100             : 
    4101             :   // First, check if select has inverse order:
    4102         100 :   if (CmpRHS == FalseVal) {
    4103             :     std::swap(TrueVal, FalseVal);
    4104          10 :     Pred = CmpInst::getInversePredicate(Pred);
    4105             :   }
    4106             : 
    4107             :   // Assume success now. If there's no match, callers should not use these anyway.
    4108         100 :   LHS = TrueVal;
    4109         100 :   RHS = FalseVal;
    4110             : 
    4111             :   const APFloat *FC1;
    4112         152 :   if (CmpRHS != TrueVal || !match(CmpRHS, m_APFloat(FC1)) || !FC1->isFinite())
    4113          74 :     return {SPF_UNKNOWN, SPNB_NA, false};
    4114             : 
    4115             :   const APFloat *FC2;
    4116             :   switch (Pred) {
    4117             :   case CmpInst::FCMP_OLT:
    4118             :   case CmpInst::FCMP_OLE:
    4119             :   case CmpInst::FCMP_ULT:
    4120             :   case CmpInst::FCMP_ULE:
    4121             :     if (match(FalseVal,
    4122           4 :               m_CombineOr(m_OrdFMin(m_Specific(CmpLHS), m_APFloat(FC2)),
    4123          28 :                           m_UnordFMin(m_Specific(CmpLHS), m_APFloat(FC2)))) &&
    4124          14 :         FC1->compare(*FC2) == APFloat::cmpResult::cmpLessThan)
    4125          10 :       return {SPF_FMAXNUM, SPNB_RETURNS_ANY, false};
    4126             :     break;
    4127             :   case CmpInst::FCMP_OGT:
    4128             :   case CmpInst::FCMP_OGE:
    4129             :   case CmpInst::FCMP_UGT:
    4130             :   case CmpInst::FCMP_UGE:
    4131             :     if (match(FalseVal,
    4132           4 :               m_CombineOr(m_OrdFMax(m_Specific(CmpLHS), m_APFloat(FC2)),
    4133          20 :                           m_UnordFMax(m_Specific(CmpLHS), m_APFloat(FC2)))) &&
    4134           8 :         FC1->compare(*FC2) == APFloat::cmpResult::cmpGreaterThan)
    4135           8 :       return {SPF_FMINNUM, SPNB_RETURNS_ANY, false};
    4136             :     break;
    4137             :   default:
    4138             :     break;
    4139             :   }
    4140             : 
    4141           8 :   return {SPF_UNKNOWN, SPNB_NA, false};
    4142             : }
    4143             : 
    4144             : /// Recognize variations of:
    4145             : ///   CLAMP(v,l,h) ==> ((v) < (l) ? (l) : ((v) > (h) ? (h) : (v)))
    4146      362488 : static SelectPatternResult matchClamp(CmpInst::Predicate Pred,
    4147             :                                       Value *CmpLHS, Value *CmpRHS,
    4148             :                                       Value *TrueVal, Value *FalseVal) {
    4149             :   // Swap the select operands and predicate to match the patterns below.
    4150      362488 :   if (CmpRHS != TrueVal) {
    4151      360729 :     Pred = ICmpInst::getSwappedPredicate(Pred);
    4152             :     std::swap(TrueVal, FalseVal);
    4153             :   }
    4154             :   const APInt *C1;
    4155      725104 :   if (CmpRHS == TrueVal && match(CmpRHS, m_APInt(C1))) {
    4156             :     const APInt *C2;
    4157             :     // (X <s C1) ? C1 : SMIN(X, C2) ==> SMAX(SMIN(X, C2), C1)
    4158       12852 :     if (match(FalseVal, m_SMin(m_Specific(CmpLHS), m_APInt(C2))) &&
    4159        6459 :         C1->slt(*C2) && Pred == CmpInst::ICMP_SLT)
    4160          68 :       return {SPF_SMAX, SPNB_NA, false};
    4161             : 
    4162             :     // (X >s C1) ? C1 : SMAX(X, C2) ==> SMIN(SMAX(X, C2), C1)
    4163       12798 :     if (match(FalseVal, m_SMax(m_Specific(CmpLHS), m_APInt(C2))) &&
    4164        6415 :         C1->sgt(*C2) && Pred == CmpInst::ICMP_SGT)
    4165          10 :       return {SPF_SMIN, SPNB_NA, false};
    4166             : 
    4167             :     // (X <u C1) ? C1 : UMIN(X, C2) ==> UMAX(UMIN(X, C2), C1)
    4168       12773 :     if (match(FalseVal, m_UMin(m_Specific(CmpLHS), m_APInt(C2))) &&
    4169        6397 :         C1->ult(*C2) && Pred == CmpInst::ICMP_ULT)
    4170           7 :       return {SPF_UMAX, SPNB_NA, false};
    4171             : 
    4172             :     // (X >u C1) ? C1 : UMAX(X, C2) ==> UMIN(UMAX(X, C2), C1)
    4173       12752 :     if (match(FalseVal, m_UMax(m_Specific(CmpLHS), m_APInt(C2))) &&
    4174        6390 :         C1->ugt(*C2) && Pred == CmpInst::ICMP_UGT)
    4175           7 :       return {SPF_UMIN, SPNB_NA, false};
    4176             :   }
    4177      362442 :   return {SPF_UNKNOWN, SPNB_NA, false};
    4178             : }
    4179             : 
    4180             : /// Recognize variations of:
    4181             : ///   a < c ? min(a,b) : min(b,c) ==> min(min(a,b),min(b,c))
    4182      362442 : static SelectPatternResult matchMinMaxOfMinMax(CmpInst::Predicate Pred,
    4183             :                                                Value *CmpLHS, Value *CmpRHS,
    4184             :                                                Value *TVal, Value *FVal,
    4185             :                                                unsigned Depth) {
    4186             :   // TODO: Allow FP min/max with nnan/nsz.
    4187             :   assert(CmpInst::isIntPredicate(Pred) && "Expected integer comparison");
    4188             : 
    4189             :   Value *A, *B;
    4190      362442 :   SelectPatternResult L = matchSelectPattern(TVal, A, B, nullptr, Depth + 1);
    4191             :   if (!SelectPatternResult::isMinOrMax(L.Flavor))
    4192      362286 :     return {SPF_UNKNOWN, SPNB_NA, false};
    4193             : 
    4194             :   Value *C, *D;
    4195         156 :   SelectPatternResult R = matchSelectPattern(FVal, C, D, nullptr, Depth + 1);
    4196         156 :   if (L.Flavor != R.Flavor)
    4197          13 :     return {SPF_UNKNOWN, SPNB_NA, false};
    4198             : 
    4199             :   // We have something like: x Pred y ? min(a, b) : min(c, d).
    4200             :   // Try to match the compare to the min/max operations of the select operands.
    4201             :   // First, make sure we have the right compare predicate.
    4202         143 :   switch (L.Flavor) {
    4203          32 :   case SPF_SMIN:
    4204          32 :     if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE) {
    4205          16 :       Pred = ICmpInst::getSwappedPredicate(Pred);
    4206             :       std::swap(CmpLHS, CmpRHS);
    4207             :     }
    4208          32 :     if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE)
    4209             :       break;
    4210           0 :     return {SPF_UNKNOWN, SPNB_NA, false};
    4211          32 :   case SPF_SMAX:
    4212          32 :     if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) {
    4213          16 :       Pred = ICmpInst::getSwappedPredicate(Pred);
    4214             :       std::swap(CmpLHS, CmpRHS);
    4215             :     }
    4216          32 :     if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE)
    4217             :       break;
    4218           0 :     return {SPF_UNKNOWN, SPNB_NA, false};
    4219          36 :   case SPF_UMIN:
    4220          36 :     if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) {
    4221          16 :       Pred = ICmpInst::getSwappedPredicate(Pred);
    4222             :       std::swap(CmpLHS, CmpRHS);
    4223             :     }
    4224          36 :     if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE)
    4225             :       break;
    4226           0 :     return {SPF_UNKNOWN, SPNB_NA, false};
    4227          37 :   case SPF_UMAX:
    4228          37 :     if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) {
    4229          21 :       Pred = ICmpInst::getSwappedPredicate(Pred);
    4230             :       std::swap(CmpLHS, CmpRHS);
    4231             :     }
    4232          37 :     if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)
    4233             :       break;
    4234           0 :     return {SPF_UNKNOWN, SPNB_NA, false};
    4235           6 :   default:
    4236           6 :     return {SPF_UNKNOWN, SPNB_NA, false};
    4237             :   }
    4238             : 
    4239             :   // If there is a common operand in the already matched min/max and the other
    4240             :   // min/max operands match the compare operands (either directly or inverted),
    4241             :   // then this is min/max of the same flavor.
    4242             : 
    4243             :   // a pred c ? m(a, b) : m(c, b) --> m(m(a, b), m(c, b))
    4244             :   // ~c pred ~a ? m(a, b) : m(c, b) --> m(m(a, b), m(c, b))
    4245         137 :   if (D == B) {
    4246         122 :     if ((CmpLHS == A && CmpRHS == C) || (match(C, m_Not(m_Specific(CmpLHS))) &&
    4247          61 :                                          match(A, m_Not(m_Specific(CmpRHS)))))
    4248          41 :       return {L.Flavor, SPNB_NA, false};
    4249             :   }
    4250             :   // a pred d ? m(a, b) : m(b, d) --> m(m(a, b), m(b, d))
    4251             :   // ~d pred ~a ? m(a, b) : m(b, d) --> m(m(a, b), m(b, d))
    4252          96 :   if (C == B) {
    4253          96 :     if ((CmpLHS == A && CmpRHS == D) || (match(D, m_Not(m_Specific(CmpLHS))) &&
    4254          48 :                                          match(A, m_Not(m_Specific(CmpRHS)))))
    4255          32 :       return {L.Flavor, SPNB_NA, false};
    4256             :   }
    4257             :   // b pred c ? m(a, b) : m(c, a) --> m(m(a, b), m(c, a))
    4258             :   // ~c pred ~b ? m(a, b) : m(c, a) --> m(m(a, b), m(c, a))
    4259          64 :   if (D == A) {
    4260          96 :     if ((CmpLHS == B && CmpRHS == C) || (match(C, m_Not(m_Specific(CmpLHS))) &&
    4261          48 :                                          match(B, m_Not(m_Specific(CmpRHS)))))
    4262          32 :       return {L.Flavor, SPNB_NA, false};
    4263             :   }
    4264             :   // b pred d ? m(a, b) : m(a, d) --> m(m(a, b), m(a, d))
    4265             :   // ~d pred ~b ? m(a, b) : m(a, d) --> m(m(a, b), m(a, d))
    4266          32 :   if (C == A) {
    4267          96 :     if ((CmpLHS == B && CmpRHS == D) || (match(D, m_Not(m_Specific(CmpLHS))) &&
    4268          48 :                                          match(B, m_Not(m_Specific(CmpRHS)))))
    4269          32 :       return {L.Flavor, SPNB_NA, false};
    4270             :   }
    4271             : 
    4272           0 :   return {SPF_UNKNOWN, SPNB_NA, false};
    4273             : }
    4274             : 
    4275             : /// Match non-obvious integer minimum and maximum sequences.
    4276      362488 : static SelectPatternResult matchMinMax(CmpInst::Predicate Pred,
    4277             :                                        Value *CmpLHS, Value *CmpRHS,
    4278             :                                        Value *TrueVal, Value *FalseVal,
    4279             :                                        Value *&LHS, Value *&RHS,
    4280             :                                        unsigned Depth) {
    4281             :   // Assume success. If there's no match, callers should not use these anyway.
    4282      362488 :   LHS = TrueVal;
    4283      362488 :   RHS = FalseVal;
    4284             : 
    4285      362488 :   SelectPatternResult SPR = matchClamp(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal);
    4286      362488 :   if (SPR.Flavor != SelectPatternFlavor::SPF_UNKNOWN)
    4287          46 :     return SPR;
    4288             : 
    4289      362442 :   SPR = matchMinMaxOfMinMax(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, Depth);
    4290      362442 :   if (SPR.Flavor != SelectPatternFlavor::SPF_UNKNOWN)
    4291         137 :     return SPR;
    4292             :   
    4293      362305 :   if (Pred != CmpInst::ICMP_SGT && Pred != CmpInst::ICMP_SLT)
    4294      341727 :     return {SPF_UNKNOWN, SPNB_NA, false};
    4295             : 
    4296             :   // Z = X -nsw Y
    4297             :   // (X >s Y) ? 0 : Z ==> (Z >s 0) ? 0 : Z ==> SMIN(Z, 0)
    4298             :   // (X <s Y) ? 0 : Z ==> (Z <s 0) ? 0 : Z ==> SMAX(Z, 0)
    4299       18845 :   if (match(TrueVal, m_Zero()) &&
    4300       20713 :       match(FalseVal, m_NSWSub(m_Specific(CmpLHS), m_Specific(CmpRHS))))
    4301          24 :     return {Pred == CmpInst::ICMP_SGT ? SPF_SMIN : SPF_SMAX, SPNB_NA, false};
    4302             : 
    4303             :   // Z = X -nsw Y
    4304             :   // (X >s Y) ? Z : 0 ==> (Z >s 0) ? Z : 0 ==> SMAX(Z, 0)
    4305             :   // (X <s Y) ? Z : 0 ==> (Z <s 0) ? Z : 0 ==> SMIN(Z, 0)
    4306       19076 :   if (match(FalseVal, m_Zero()) &&
    4307       20957 :       match(TrueVal, m_NSWSub(m_Specific(CmpLHS), m_Specific(CmpRHS))))
    4308           7 :     return {Pred == CmpInst::ICMP_SGT ? SPF_SMAX : SPF_SMIN, SPNB_NA, false};
    4309             : 
    4310             :   const APInt *C1;
    4311       41094 :   if (!match(CmpRHS, m_APInt(C1)))
    4312        8105 :     return {SPF_UNKNOWN, SPNB_NA, false};
    4313             : 
    4314             :   // An unsigned min/max can be written with a signed compare.
    4315             :   const APInt *C2;
    4316       24977 :   if ((CmpLHS == TrueVal && match(FalseVal, m_APInt(C2))) ||
    4317       12919 :       (CmpLHS == FalseVal && match(TrueVal, m_APInt(C2)))) {
    4318             :     // Is the sign bit set?
    4319             :     // (X <s 0) ? X : MAXVAL ==> (X >u MAXVAL) ? X : MAXVAL ==> UMAX
    4320             :     // (X <s 0) ? MAXVAL : X ==> (X >u MAXVAL) ? MAXVAL : X ==> UMIN
    4321         213 :     if (Pred == CmpInst::ICMP_SLT && C1->isNullValue() &&
    4322          24 :         C2->isMaxSignedValue())
    4323          11 :       return {CmpLHS == TrueVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false};
    4324             : 
    4325             :     // Is the sign bit clear?
    4326             :     // (X >s -1) ? MINVAL : X ==> (X <u MINVAL) ? MINVAL : X ==> UMAX
    4327             :     // (X >s -1) ? X : MINVAL ==> (X <u MINVAL) ? X : MINVAL ==> UMIN
    4328         201 :     if (Pred == CmpInst::ICMP_SGT && C1->isAllOnesValue() &&
    4329          29 :         C2->isMinSignedValue())
    4330          11 :       return {CmpLHS == FalseVal ? SPF_UMAX : SPF_UMIN, SPNB_NA, false};
    4331             :   }
    4332             : 
    4333             :   // Look through 'not' ops to find disguised signed min/max.
    4334             :   // (X >s C) ? ~X : ~C ==> (~X <s ~C) ? ~X : ~C ==> SMIN(~X, ~C)
    4335             :   // (X <s C) ? ~X : ~C ==> (~X >s ~C) ? ~X : ~C ==> SMAX(~X, ~C)
    4336       24869 :   if (match(TrueVal, m_Not(m_Specific(CmpLHS))) &&
    4337       37318 :       match(FalseVal, m_APInt(C2)) && ~(*C1) == *C2)
    4338          29 :     return {Pred == CmpInst::ICMP_SGT ? SPF_SMIN : SPF_SMAX, SPNB_NA, false};
    4339             : 
    4340             :   // (X >s C) ? ~C : ~X ==> (~X <s ~C) ? ~C : ~X ==> SMAX(~C, ~X)
    4341             :   // (X <s C) ? ~C : ~X ==> (~X >s ~C) ? ~C : ~X ==> SMIN(~C, ~X)
    4342       24796 :   if (match(FalseVal, m_Not(m_Specific(CmpLHS))) &&
    4343       37201 :       match(TrueVal, m_APInt(C2)) && ~(*C1) == *C2)
    4344          14 :     return {Pred == CmpInst::ICMP_SGT ? SPF_SMAX : SPF_SMIN, SPNB_NA, false};
    4345             : 
    4346       12377 :   return {SPF_UNKNOWN, SPNB_NA, false};
    4347             : }
    4348             : 
    4349      424782 : static SelectPatternResult matchSelectPattern(CmpInst::Predicate Pred,
    4350             :                                               FastMathFlags FMF,
    4351             :                                               Value *CmpLHS, Value *CmpRHS,
    4352             :                                               Value *TrueVal, Value *FalseVal,
    4353             :                                               Value *&LHS, Value *&RHS,
    4354             :                                               unsigned Depth) {
    4355      424782 :   LHS = CmpLHS;
    4356      424782 :   RHS = CmpRHS;
    4357             : 
    4358             :   // Signed zero may return inconsistent results between implementations.
    4359             :   //  (0.0 <= -0.0) ? 0.0 : -0.0 // Returns 0.0
    4360             :   //  minNum(0.0, -0.0)          // May return -0.0 or 0.0 (IEEE 754-2008 5.3.1)
    4361             :   // Therefore, we behave conservatively and only proceed if at least one of the
    4362             :   // operands is known to not be zero or if we don't care about signed zero.
    4363             :   switch (Pred) {
    4364             :   default: break;
    4365             :   // FIXME: Include OGT/OLT/UGT/ULT.
    4366             :   case CmpInst::FCMP_OGE: case CmpInst::FCMP_OLE:
    4367             :   case CmpInst::FCMP_UGE: case CmpInst::FCMP_ULE:
    4368        3096 :     if (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) &&
    4369             :         !isKnownNonZero(CmpRHS))
    4370         960 :       return {SPF_UNKNOWN, SPNB_NA, false};
    4371             :   }
    4372             : 
    4373             :   SelectPatternNaNBehavior NaNBehavior = SPNB_NA;
    4374             :   bool Ordered = false;
    4375             : 
    4376             :   // When given one NaN and one non-NaN input:
    4377             :   //   - maxnum/minnum (C99 fmaxf()/fminf()) return the non-NaN input.
    4378             :   //   - A simple C99 (a < b ? a : b) construction will return 'b' (as the
    4379             :   //     ordered comparison fails), which could be NaN or non-NaN.
    4380             :   // so here we discover exactly what NaN behavior is required/accepted.
    4381      423822 :   if (CmpInst::isFPPredicate(Pred)) {
    4382        7144 :     bool LHSSafe = isKnownNonNaN(CmpLHS, FMF);
    4383        7144 :     bool RHSSafe = isKnownNonNaN(CmpRHS, FMF);
    4384             : 
    4385        7144 :     if (LHSSafe && RHSSafe) {
    4386             :       // Both operands are known non-NaN.
    4387             :       NaNBehavior = SPNB_RETURNS_ANY;
    4388        6595 :     } else if (CmpInst::isOrdered(Pred)) {
    4389             :       // An ordered comparison will return false when given a NaN, so it
    4390             :       // returns the RHS.
    4391             :       Ordered = true;
    4392        5014 :       if (LHSSafe)
    4393             :         // LHS is non-NaN, so if RHS is NaN then NaN will be returned.
    4394             :         NaNBehavior = SPNB_RETURNS_NAN;
    4395        4846 :       else if (RHSSafe)
    4396             :         NaNBehavior = SPNB_RETURNS_OTHER;
    4397             :       else
    4398             :         // Completely unsafe.
    4399        3754 :         return {SPF_UNKNOWN, SPNB_NA, false};
    4400             :     } else {
    4401             :       Ordered = false;
    4402             :       // An unordered comparison will return true when given a NaN, so it
    4403             :       // returns the LHS.
    4404        1581 :       if (LHSSafe)
    4405             :         // LHS is non-NaN, so if RHS is NaN then non-NaN will be returned.
    4406             :         NaNBehavior = SPNB_RETURNS_OTHER;
    4407        1237 :       else if (RHSSafe)
    4408             :         NaNBehavior = SPNB_RETURNS_NAN;
    4409             :       else
    4410             :         // Completely unsafe.
    4411         309 :         return {SPF_UNKNOWN, SPNB_NA, false};
    4412             :     }
    4413             :   }
    4414             : 
    4415      419759 :   if (TrueVal == CmpRHS && FalseVal == CmpLHS) {
    4416             :     std::swap(CmpLHS, CmpRHS);
    4417        9636 :     Pred = CmpInst::getSwappedPredicate(Pred);
    4418        9636 :     if (NaNBehavior == SPNB_RETURNS_NAN)
    4419             :       NaNBehavior = SPNB_RETURNS_OTHER;
    4420        9335 :     else if (NaNBehavior == SPNB_RETURNS_OTHER)
    4421             :       NaNBehavior = SPNB_RETURNS_NAN;
    4422        9636 :     Ordered = !Ordered;
    4423             :   }
    4424             : 
    4425             :   // ([if]cmp X, Y) ? X : Y
    4426      419759 :   if (TrueVal == CmpLHS && FalseVal == CmpRHS) {
    4427       52473 :     switch (Pred) {
    4428           0 :     default: return {SPF_UNKNOWN, SPNB_NA, false}; // Equality.
    4429       12890 :     case ICmpInst::ICMP_UGT:
    4430       12890 :     case ICmpInst::ICMP_UGE: return {SPF_UMAX, SPNB_NA, false};
    4431       11134 :     case ICmpInst::ICMP_SGT:
    4432       11134 :     case ICmpInst::ICMP_SGE: return {SPF_SMAX, SPNB_NA, false};
    4433       17801 :     case ICmpInst::ICMP_ULT:
    4434       17801 :     case ICmpInst::ICMP_ULE: return {SPF_UMIN, SPNB_NA, false};
    4435        9272 :     case ICmpInst::ICMP_SLT:
    4436        9272 :     case ICmpInst::ICMP_SLE: return {SPF_SMIN, SPNB_NA, false};
    4437         701 :     case FCmpInst::FCMP_UGT:
    4438             :     case FCmpInst::FCMP_UGE:
    4439             :     case FCmpInst::FCMP_OGT:
    4440         701 :     case FCmpInst::FCMP_OGE: return {SPF_FMAXNUM, NaNBehavior, Ordered};
    4441         675 :     case FCmpInst::FCMP_ULT:
    4442             :     case FCmpInst::FCMP_ULE:
    4443             :     case FCmpInst::FCMP_OLT:
    4444         675 :     case FCmpInst::FCMP_OLE: return {SPF_FMINNUM, NaNBehavior, Ordered};
    4445             :     }
    4446             :   }
    4447             : 
    4448             :   const APInt *C1;
    4449      734572 :   if (match(CmpRHS, m_APInt(C1))) {
    4450      508955 :     if ((CmpLHS == TrueVal && match(FalseVal, m_Neg(m_Specific(CmpLHS)))) ||
    4451      253142 :         (CmpLHS == FalseVal && match(TrueVal, m_Neg(m_Specific(CmpLHS))))) {
    4452             : 
    4453             :       // ABS(X) ==> (X >s 0) ? X : -X and (X >s -1) ? X : -X
    4454             :       // NABS(X) ==> (X >s 0) ? -X : X and (X >s -1) ? -X : X
    4455        5275 :       if (Pred == ICmpInst::ICMP_SGT &&
    4456        3088 :           (C1->isNullValue() || C1->isAllOnesValue())) {
    4457        3916 :         return {(CmpLHS == TrueVal) ? SPF_ABS : SPF_NABS, SPNB_NA, false};
    4458             :       }
    4459             : 
    4460             :       // ABS(X) ==> (X <s 0) ? -X : X and (X <s 1) ? -X : X
    4461             :       // NABS(X) ==> (X <s 0) ? X : -X and (X <s 1) ? X : -X
    4462        2494 :       if (Pred == ICmpInst::ICMP_SLT &&
    4463        1619 :           (C1->isNullValue() || C1->isOneValue())) {
    4464        1135 :         return {(CmpLHS == FalseVal) ? SPF_ABS : SPF_NABS, SPNB_NA, false};
    4465             :       }
    4466             :     }
    4467             :   }
    4468             : 
    4469      364193 :   if (CmpInst::isIntPredicate(Pred))
    4470      362488 :     return matchMinMax(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS, Depth);
    4471             : 
    4472             :   // According to (IEEE 754-2008 5.3.1), minNum(0.0, -0.0) and similar
    4473             :   // may return either -0.0 or 0.0, so fcmp/select pair has stricter
    4474             :   // semantics than minNum. Be conservative in such case.
    4475        1807 :   if (NaNBehavior != SPNB_RETURNS_ANY ||
    4476          43 :       (!FMF.noSignedZeros() && !isKnownNonZero(CmpLHS) &&
    4477             :        !isKnownNonZero(CmpRHS)))
    4478        1605 :     return {SPF_UNKNOWN, SPNB_NA, false};
    4479             : 
    4480         100 :   return matchFastFloatClamp(Pred, CmpLHS, CmpRHS, TrueVal, FalseVal, LHS, RHS);
    4481             : }
    4482             : 
    4483             : /// Helps to match a select pattern in case of a type mismatch.
    4484             : ///
    4485             : /// The function processes the case when type of true and false values of a
    4486             : /// select instruction differs from type of the cmp instruction operands because
    4487             : /// of a cast instructon. The function checks if it is legal to move the cast
    4488             : /// operation after "select". If yes, it returns the new second value of
    4489             : /// "select" (with the assumption that cast is moved):
    4490             : /// 1. As operand of cast instruction when both values of "select" are same cast
    4491             : /// instructions.
    4492             : /// 2. As restored constant (by applying reverse cast operation) when the first
    4493             : /// value of the "select" is a cast operation and the second value is a
    4494             : /// constant.
    4495             : /// NOTE: We return only the new second value because the first value could be
    4496             : /// accessed as operand of cast instruction.
    4497       21252 : static Value *lookThroughCast(CmpInst *CmpI, Value *V1, Value *V2,
    4498             :                               Instruction::CastOps *CastOp) {
    4499             :   auto *Cast1 = dyn_cast<CastInst>(V1);
    4500             :   if (!Cast1)
    4501             :     return nullptr;
    4502             : 
    4503          74 :   *CastOp = Cast1->getOpcode();
    4504             :   Type *SrcTy = Cast1->getSrcTy();
    4505             :   if (auto *Cast2 = dyn_cast<CastInst>(V2)) {
    4506             :     // If V1 and V2 are both the same cast from the same type, look through V1.
    4507          18 :     if (*CastOp == Cast2->getOpcode() && SrcTy == Cast2->getSrcTy())
    4508             :       return Cast2->getOperand(0);
    4509             :     return nullptr;
    4510             :   }
    4511             : 
    4512             :   auto *C = dyn_cast<Constant>(V2);
    4513             :   if (!C)
    4514             :     return nullptr;
    4515             : 
    4516             :   Constant *CastedTo = nullptr;
    4517          40 :   switch (*CastOp) {
    4518             :   case Instruction::ZExt:
    4519           2 :     if (CmpI->isUnsigned())
    4520           1 :       CastedTo = ConstantExpr::getTrunc(C, SrcTy);
    4521             :     break;
    4522             :   case Instruction::SExt:
    4523           2 :     if (CmpI->isSigned())
    4524           2 :       CastedTo = ConstantExpr::getTrunc(C, SrcTy, true);
    4525             :     break;
    4526             :   case Instruction::Trunc:
    4527             :     Constant *CmpConst;
    4528           7 :     if (match(CmpI->getOperand(1), m_Constant(CmpConst)) &&
    4529           7 :         CmpConst->getType() == SrcTy) {
    4530             :       // Here we have the following case:
    4531             :       //
    4532             :       //   %cond = cmp iN %x, CmpConst
    4533             :       //   %tr = trunc iN %x to iK
    4534             :       //   %narrowsel = select i1 %cond, iK %t, iK C
    4535             :       //
    4536             :       // We can always move trunc after select operation:
    4537             :       //
    4538             :       //   %cond = cmp iN %x, CmpConst
    4539             :       //   %widesel = select i1 %cond, iN %x, iN CmpConst
    4540             :       //   %tr = trunc iN %widesel to iK
    4541             :       //
    4542             :       // Note that C could be extended in any way because we don't care about
    4543             :       // upper bits after truncation. It can't be abs pattern, because it would
    4544             :       // look like:
    4545             :       //
    4546             :       //   select i1 %cond, x, -x.
    4547             :       //
    4548             :       // So only min/max pattern could be matched. Such match requires widened C
    4549             :       // == CmpConst. That is why set widened C = CmpConst, condition trunc
    4550             :       // CmpConst == C is checked below.
    4551             :       CastedTo = CmpConst;
    4552             :     } else {
    4553           0 :       CastedTo = ConstantExpr::getIntegerCast(C, SrcTy, CmpI->isSigned());
    4554             :     }
    4555             :     break;
    4556           3 :   case Instruction::FPTrunc:
    4557           3 :     CastedTo = ConstantExpr::getFPExtend(C, SrcTy, true);
    4558           3 :     break;
    4559           4 :   case Instruction::FPExt:
    4560           4 :     CastedTo = ConstantExpr::getFPTrunc(C, SrcTy, true);
    4561           4 :     break;
    4562           1 :   case Instruction::FPToUI:
    4563           1 :     CastedTo = ConstantExpr::getUIToFP(C, SrcTy, true);
    4564           1 :     break;
    4565           6 :   case Instruction::FPToSI:
    4566           6 :     CastedTo = ConstantExpr::getSIToFP(C, SrcTy, true);
    4567           6 :     break;
    4568           2 :   case Instruction::UIToFP:
    4569           2 :     CastedTo = ConstantExpr::getFPToUI(C, SrcTy, true);
    4570           2 :     break;
    4571           7 :   case Instruction::SIToFP:
    4572           7 :     CastedTo = ConstantExpr::getFPToSI(C, SrcTy, true);
    4573           7 :     break;
    4574             :   default:
    4575             :     break;
    4576             :   }
    4577             : 
    4578          33 :   if (!CastedTo)
    4579             :     return nullptr;
    4580             : 
    4581             :   // Make sure the cast doesn't lose any information.
    4582             :   Constant *CastedBack =
    4583          33 :       ConstantExpr::getCast(*CastOp, CastedTo, C->getType(), true);
    4584          33 :   if (CastedBack != C)
    4585             :     return nullptr;
    4586             : 
    4587          29 :   return CastedTo;
    4588             : }
    4589             : 
    4590     3886689 : SelectPatternResult llvm::matchSelectPattern(Value *V, Value *&LHS, Value *&RHS,
    4591             :                                              Instruction::CastOps *CastOp,
    4592             :                                              unsigned Depth) {
    4593     3886689 :   if (Depth >= MaxDepth)
    4594          11 :     return {SPF_UNKNOWN, SPNB_NA, false};
    4595             : 
    4596             :   SelectInst *SI = dyn_cast<SelectInst>(V);
    4597     2046984 :   if (!SI) return {SPF_UNKNOWN, SPNB_NA, false};
    4598             : 
    4599             :   CmpInst *CmpI = dyn_cast<CmpInst>(SI->getCondition());
    4600      242932 :   if (!CmpI) return {SPF_UNKNOWN, SPNB_NA, false};
    4601             : 
    4602             :   CmpInst::Predicate Pred = CmpI->getPredicate();
    4603             :   Value *CmpLHS = CmpI->getOperand(0);
    4604             :   Value *CmpRHS = CmpI->getOperand(1);
    4605             :   Value *TrueVal = SI->getTrueValue();
    4606             :   Value *FalseVal = SI->getFalseValue();
    4607             :   FastMathFlags FMF;
    4608     1596762 :   if (isa<FPMathOperator>(CmpI))
    4609       10170 :     FMF = CmpI->getFastMathFlags();
    4610             : 
    4611             :   // Bail out early.
    4612     1596762 :   if (CmpI->isEquality())
    4613     1171980 :     return {SPF_UNKNOWN, SPNB_NA, false};
    4614             : 
    4615             :   // Deal with type mismatches.
    4616      424782 :   if (CastOp && CmpLHS->getType() != TrueVal->getType()) {
    4617       10641 :     if (Value *C = lookThroughCast(CmpI, TrueVal, FalseVal, CastOp)) {
    4618             :       // If this is a potential fmin/fmax with a cast to integer, then ignore
    4619             :       // -0.0 because there is no corresponding integer value.
    4620          30 :       if (*CastOp == Instruction::FPToSI || *CastOp == Instruction::FPToUI)
    4621             :         FMF.setNoSignedZeros();
    4622             :       return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
    4623             :                                   cast<CastInst>(TrueVal)->getOperand(0), C,
    4624          30 :                                   LHS, RHS, Depth);
    4625             :     }
    4626       10611 :     if (Value *C = lookThroughCast(CmpI, FalseVal, TrueVal, CastOp)) {
    4627             :       // If this is a potential fmin/fmax with a cast to integer, then ignore
    4628             :       // -0.0 because there is no corresponding integer value.
    4629           7 :       if (*CastOp == Instruction::FPToSI || *CastOp == Instruction::FPToUI)
    4630             :         FMF.setNoSignedZeros();
    4631             :       return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS,
    4632             :                                   C, cast<CastInst>(FalseVal)->getOperand(0),
    4633           7 :                                   LHS, RHS, Depth);
    4634             :     }
    4635             :   }
    4636             :   return ::matchSelectPattern(Pred, FMF, CmpLHS, CmpRHS, TrueVal, FalseVal,
    4637      424745 :                               LHS, RHS, Depth);
    4638             : }
    4639             : 
    4640             : /// Return true if "icmp Pred LHS RHS" is always true.
    4641        1341 : static bool isTruePredicate(CmpInst::Predicate Pred, const Value *LHS,
    4642             :                             const Value *RHS, const DataLayout &DL,
    4643             :                             unsigned Depth) {
    4644             :   assert(!LHS->getType()->isVectorTy() && "TODO: extend to handle vectors!");
    4645        1341 :   if (ICmpInst::isTrueWhenEqual(Pred) && LHS == RHS)
    4646             :     return true;
    4647             : 
    4648        1310 :   switch (Pred) {
    4649             :   default:
    4650             :     return false;
    4651             : 
    4652             :   case CmpInst::ICMP_SLE: {
    4653             :     const APInt *C;
    4654             : 
    4655             :     // LHS s<= LHS +_{nsw} C   if C >= 0
    4656         404 :     if (match(RHS, m_NSWAdd(m_Specific(LHS), m_APInt(C))))
    4657          16 :       return !C->isNegative();
    4658             :     return false;
    4659             :   }
    4660             : 
    4661             :   case CmpInst::ICMP_ULE: {
    4662             :     const APInt *C;
    4663             : 
    4664             :     // LHS u<= LHS +_{nuw} C   for any C
    4665        2216 :     if (match(RHS, m_NUWAdd(m_Specific(LHS), m_APInt(C))))
    4666             :       return true;
    4667             : 
    4668             :     // Match A to (X +_{nuw} CA) and B to (X +_{nuw} CB)
    4669             :     auto MatchNUWAddsToSameValue = [&](const Value *A, const Value *B,
    4670             :                                        const Value *&X,
    4671        1102 :                                        const APInt *&CA, const APInt *&CB) {
    4672        3326 :       if (match(A, m_NUWAdd(m_Value(X), m_APInt(CA))) &&
    4673        1122 :           match(B, m_NUWAdd(m_Specific(X), m_APInt(CB))))
    4674             :         return true;
    4675             : 
    4676             :       // If X & C == 0 then (X | C) == X +_{nuw} C
    4677        3342 :       if (match(A, m_Or(m_Value(X), m_APInt(CA))) &&
    4678        1142 :           match(B, m_Or(m_Specific(X), m_APInt(CB)))) {
    4679          24 :         KnownBits Known(CA->getBitWidth());
    4680          13 :         computeKnownBits(X, Known, DL, Depth + 1, /*AC*/ nullptr,
    4681             :                          /*CxtI*/ nullptr, /*DT*/ nullptr);
    4682          29 :         if (CA->isSubsetOf(Known.Zero) && CB->isSubsetOf(Known.Zero))
    4683           2 :           return true;
    4684             :       }
    4685             : 
    4686             :       return false;
    4687        1102 :     };
    4688             : 
    4689             :     const Value *X;
    4690             :     const APInt *CLHS, *CRHS;
    4691        1102 :     if (MatchNUWAddsToSameValue(LHS, RHS, X, CLHS, CRHS))
    4692           8 :       return CLHS->ule(*CRHS);
    4693             : 
    4694             :     return false;
    4695             :   }
    4696             :   }
    4697             : }
    4698             : 
    4699             : /// Return true if "icmp Pred BLHS BRHS" is true whenever "icmp Pred
    4700             : /// ALHS ARHS" is true.  Otherwise, return None.
    4701             : static Optional<bool>
    4702       25324 : isImpliedCondOperands(CmpInst::Predicate Pred, const Value *ALHS,
    4703             :                       const Value *ARHS, const Value *BLHS, const Value *BRHS,
    4704             :                       const DataLayout &DL, unsigned Depth) {
    4705       25324 :   switch (Pred) {
    4706             :   default:
    4707             :     return None;
    4708             : 
    4709         201 :   case CmpInst::ICMP_SLT:
    4710             :   case CmpInst::ICMP_SLE:
    4711         208 :     if (isTruePredicate(CmpInst::ICMP_SLE, BLHS, ALHS, DL, Depth) &&
    4712           7 :         isTruePredicate(CmpInst::ICMP_SLE, ARHS, BRHS, DL, Depth))
    4713             :       return true;
    4714             :     return None;
    4715             : 
    4716        1107 :   case CmpInst::ICMP_ULT:
    4717             :   case CmpInst::ICMP_ULE:
    4718        1133 :     if (isTruePredicate(CmpInst::ICMP_ULE, BLHS, ALHS, DL, Depth) &&
    4719          26 :         isTruePredicate(CmpInst::ICMP_ULE, ARHS, BRHS, DL, Depth))
    4720             :       return true;
    4721             :     return None;
    4722             :   }
    4723             : }
    4724             : 
    4725             : /// Return true if the operands of the two compares match.  IsSwappedOps is true
    4726             : /// when the operands match, but are swapped.
    4727             : static bool isMatchingOps(const Value *ALHS, const Value *ARHS,
    4728             :                           const Value *BLHS, const Value *BRHS,
    4729             :                           bool &IsSwappedOps) {
    4730             : 
    4731      143221 :   bool IsMatchingOps = (ALHS == BLHS && ARHS == BRHS);
    4732      143221 :   IsSwappedOps = (ALHS == BRHS && ARHS == BLHS);
    4733      143221 :   return IsMatchingOps || IsSwappedOps;
    4734             : }
    4735             : 
    4736             : /// Return true if "icmp1 APred ALHS ARHS" implies "icmp2 BPred BLHS BRHS" is
    4737             : /// true.  Return false if "icmp1 APred ALHS ARHS" implies "icmp2 BPred BLHS
    4738             : /// BRHS" is false.  Otherwise, return None if we can't infer anything.
    4739        1448 : static Optional<bool> isImpliedCondMatchingOperands(CmpInst::Predicate APred,
    4740             :                                                     const Value *ALHS,
    4741             :                                                     const Value *ARHS,
    4742             :                                                     CmpInst::Predicate BPred,
    4743             :                                                     const Value *BLHS,
    4744             :                                                     const Value *BRHS,
    4745             :                                                     bool IsSwappedOps) {
    4746             :   // Canonicalize the operands so they're matching.
    4747        1448 :   if (IsSwappedOps) {
    4748             :     std::swap(BLHS, BRHS);
    4749          35 :     BPred = ICmpInst::getSwappedPredicate(BPred);
    4750             :   }
    4751        1448 :   if (CmpInst::isImpliedTrueByMatchingCmp(APred, BPred))
    4752             :     return true;
    4753        1390 :   if (CmpInst::isImpliedFalseByMatchingCmp(APred, BPred))
    4754             :     return false;
    4755             : 
    4756             :   return None;
    4757             : }
    4758             : 
    4759             : /// Return true if "icmp1 APred ALHS C1" implies "icmp2 BPred BLHS C2" is
    4760             : /// true.  Return false if "icmp1 APred ALHS C1" implies "icmp2 BPred BLHS
    4761             : /// C2" is false.  Otherwise, return None if we can't infer anything.
    4762             : static Optional<bool>
    4763       12591 : isImpliedCondMatchingImmOperands(CmpInst::Predicate APred, const Value *ALHS,
    4764             :                                  const ConstantInt *C1,
    4765             :                                  CmpInst::Predicate BPred,
    4766             :                                  const Value *BLHS, const ConstantInt *C2) {
    4767             :   assert(ALHS == BLHS && "LHS operands must match.");
    4768             :   ConstantRange DomCR =
    4769       25182 :       ConstantRange::makeExactICmpRegion(APred, C1->getValue());
    4770             :   ConstantRange CR =
    4771       37773 :       ConstantRange::makeAllowedICmpRegion(BPred, C2->getValue());
    4772       25182 :   ConstantRange Intersection = DomCR.intersectWith(CR);
    4773       25182 :   ConstantRange Difference = DomCR.difference(CR);
    4774       12591 :   if (Intersection.isEmptySet())
    4775             :     return false;
    4776       12585 :   if (Difference.isEmptySet())
    4777             :     return true;
    4778             :   return None;
    4779             : }
    4780             : 
    4781             : /// Return true if LHS implies RHS is true.  Return false if LHS implies RHS is
    4782             : /// false.  Otherwise, return None if we can't infer anything.
    4783      143221 : static Optional<bool> isImpliedCondICmps(const ICmpInst *LHS,
    4784             :                                          const ICmpInst *RHS,
    4785             :                                          const DataLayout &DL, bool LHSIsTrue,
    4786             :                                          unsigned Depth) {
    4787             :   Value *ALHS = LHS->getOperand(0);
    4788             :   Value *ARHS = LHS->getOperand(1);
    4789             :   // The rest of the logic assumes the LHS condition is true.  If that's not the
    4790             :   // case, invert the predicate to make it so.
    4791             :   ICmpInst::Predicate APred =
    4792      143221 :       LHSIsTrue ? LHS->getPredicate() : LHS->getInversePredicate();
    4793             : 
    4794             :   Value *BLHS = RHS->getOperand(0);
    4795             :   Value *BRHS = RHS->getOperand(1);
    4796             :   ICmpInst::Predicate BPred = RHS->getPredicate();
    4797             : 
    4798             :   // Can we infer anything when the two compares have matching operands?
    4799             :   bool IsSwappedOps;
    4800             :   if (isMatchingOps(ALHS, ARHS, BLHS, BRHS, IsSwappedOps)) {
    4801        1448 :     if (Optional<bool> Implication = isImpliedCondMatchingOperands(
    4802        1448 :             APred, ALHS, ARHS, BPred, BLHS, BRHS, IsSwappedOps))
    4803             :       return Implication;
    4804             :     // No amount of additional analysis will infer the second condition, so
    4805             :     // early exit.
    4806             :     return None;
    4807             :   }
    4808             : 
    4809             :   // Can we infer anything when the LHS operands match and the RHS operands are
    4810             :   // constants (not necessarily matching)?
    4811      175154 :   if (ALHS == BLHS && isa<ConstantInt>(ARHS) && isa<ConstantInt>(BRHS)) {
    4812       12591 :     if (Optional<bool> Implication = isImpliedCondMatchingImmOperands(
    4813             :             APred, ALHS, cast<ConstantInt>(ARHS), BPred, BLHS,
    4814       12591 :             cast<ConstantInt>(BRHS)))
    4815             :       return Implication;
    4816             :     // No amount of additional analysis will infer the second condition, so
    4817             :     // early exit.
    4818             :     return None;
    4819             :   }
    4820             : 
    4821      129182 :   if (APred == BPred)
    4822       25324 :     return isImpliedCondOperands(APred, ALHS, ARHS, BLHS, BRHS, DL, Depth);
    4823             :   return None;
    4824             : }
    4825             : 
    4826             : /// Return true if LHS implies RHS is true.  Return false if LHS implies RHS is
    4827             : /// false.  Otherwise, return None if we can't infer anything.  We expect the
    4828             : /// RHS to be an icmp and the LHS to be an 'and' or an 'or' instruction.
    4829        2238 : static Optional<bool> isImpliedCondAndOr(const BinaryOperator *LHS,
    4830             :                                          const ICmpInst *RHS,
    4831             :                                          const DataLayout &DL, bool LHSIsTrue,
    4832             :                                          unsigned Depth) {
    4833             :   // The LHS must be an 'or' or an 'and' instruction.
    4834             :   assert((LHS->getOpcode() == Instruction::And ||
    4835             :           LHS->getOpcode() == Instruction::Or) &&
    4836             :          "Expected LHS to be 'and' or 'or'.");
    4837             : 
    4838             :   assert(Depth <= MaxDepth && "Hit recursion limit");
    4839             : 
    4840             :   // If the result of an 'or' is false, then we know both legs of the 'or' are
    4841             :   // false.  Similarly, if the result of an 'and' is true, then we know both
    4842             :   // legs of the 'and' are true.
    4843             :   Value *ALHS, *ARHS;
    4844        5048 :   if ((!LHSIsTrue && match(LHS, m_Or(m_Value(ALHS), m_Value(ARHS)))) ||
    4845        4947 :       (LHSIsTrue && match(LHS, m_And(m_Value(ALHS), m_Value(ARHS))))) {
    4846             :     // FIXME: Make this non-recursion.
    4847         623 :     if (Optional<bool> Implication =
    4848         623 :             isImpliedCondition(ALHS, RHS, DL, LHSIsTrue, Depth + 1))
    4849             :       return Implication;
    4850         615 :     if (Optional<bool> Implication =
    4851         615 :             isImpliedCondition(ARHS, RHS, DL, LHSIsTrue, Depth + 1))
    4852             :       return Implication;
    4853             :     return None;
    4854             :   }
    4855             :   return None;
    4856             : }
    4857             : 
    4858      228924 : Optional<bool> llvm::isImpliedCondition(const Value *LHS, const Value *RHS,
    4859             :                                         const DataLayout &DL, bool LHSIsTrue,
    4860             :                                         unsigned Depth) {
    4861             :   // Bail out when we hit the limit.
    4862      228924 :   if (Depth == MaxDepth)
    4863             :     return None;
    4864             : 
    4865             :   // A mismatch occurs when we compare a scalar cmp to a vector cmp, for
    4866             :   // example.
    4867      228908 :   if (LHS->getType() != RHS->getType())
    4868             :     return None;
    4869             : 
    4870             :   Type *OpTy = LHS->getType();
    4871             :   assert(OpTy->isIntOrIntVectorTy(1) && "Expected integer type only!");
    4872             : 
    4873             :   // LHS ==> RHS by definition
    4874      228838 :   if (LHS == RHS)
    4875             :     return LHSIsTrue;
    4876             : 
    4877             :   // FIXME: Extending the code below to handle vectors.
    4878      228639 :   if (OpTy->isVectorTy())
    4879             :     return None;
    4880             : 
    4881             :   assert(OpTy->isIntegerTy(1) && "implied by above");
    4882             : 
    4883             :   // Both LHS and RHS are icmps.
    4884             :   const ICmpInst *LHSCmp = dyn_cast<ICmpInst>(LHS);
    4885             :   const ICmpInst *RHSCmp = dyn_cast<ICmpInst>(RHS);
    4886      228635 :   if (LHSCmp && RHSCmp)
    4887      143221 :     return isImpliedCondICmps(LHSCmp, RHSCmp, DL, LHSIsTrue, Depth);
    4888             : 
    4889             :   // The LHS should be an 'or' or an 'and' instruction.  We expect the RHS to be
    4890             :   // an icmp. FIXME: Add support for and/or on the RHS.
    4891             :   const BinaryOperator *LHSBO = dyn_cast<BinaryOperator>(LHS);
    4892       85414 :   if (LHSBO && RHSCmp) {
    4893        2298 :     if ((LHSBO->getOpcode() == Instruction::And ||
    4894             :          LHSBO->getOpcode() == Instruction::Or))
    4895        2238 :       return isImpliedCondAndOr(LHSBO, RHSCmp, DL, LHSIsTrue, Depth);
    4896             :   }
    4897             :   return None;
    4898      245058 : }

Generated by: LCOV version 1.13