/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/include/llvm/IR/PatternMatch.h

Bug Summary

File:	llvm/include/llvm/IR/PatternMatch.h
Warning:	line 1460, column 9 Called C++ object pointer is null

Annotated Source Code

Press '?' to see keyboard shortcuts

Show analyzer invocation

clang -cc1 -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name InstCombineAddSub.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -mframe-pointer=none -fmath-errno -fno-rounding-math -mconstructor-aliases -munwind-tables -target-cpu x86-64 -tune-cpu generic -debugger-tuning=gdb -ffunction-sections -fdata-sections -fcoverage-compilation-dir=/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/build-llvm/lib/Transforms/InstCombine -resource-dir /usr/lib/llvm-14/lib/clang/14.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/build-llvm/lib/Transforms/InstCombine -I /build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine -I /build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/build-llvm/include -I /build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/include -D NDEBUG -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/x86_64-linux-gnu/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10/backward -internal-isystem /usr/lib/llvm-14/lib/clang/14.0.0/include -internal-isystem /usr/local/include -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../x86_64-linux-gnu/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-class-memaccess -Wno-redundant-move -Wno-pessimizing-move -Wno-noexcept-type -Wno-comment -std=c++14 -fdeprecated-macro -fdebug-compilation-dir=/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/build-llvm/lib/Transforms/InstCombine -fdebug-prefix-map=/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0=. -ferror-limit 19 -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -faddrsig -D__GCC_HAVE_DWARF2_CFI_ASM=1 -o /tmp/scan-build-2021-08-28-193554-24367-1 -x c++ /build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp

/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp

→

1//===- InstCombineAddSub.cpp ------------------------------------*- C++ -*-===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file implements the visit functions for add, fadd, sub, and fsub.
10//
11//===----------------------------------------------------------------------===//

13#include "InstCombineInternal.h"
14#include "llvm/ADT/APFloat.h"
15#include "llvm/ADT/APInt.h"
16#include "llvm/ADT/STLExtras.h"
17#include "llvm/ADT/SmallVector.h"
18#include "llvm/Analysis/InstructionSimplify.h"
19#include "llvm/Analysis/ValueTracking.h"
20#include "llvm/IR/Constant.h"
21#include "llvm/IR/Constants.h"
22#include "llvm/IR/InstrTypes.h"
23#include "llvm/IR/Instruction.h"
24#include "llvm/IR/Instructions.h"
25#include "llvm/IR/Operator.h"
26#include "llvm/IR/PatternMatch.h"
27#include "llvm/IR/Type.h"
28#include "llvm/IR/Value.h"
29#include "llvm/Support/AlignOf.h"
30#include "llvm/Support/Casting.h"
31#include "llvm/Support/KnownBits.h"
32#include "llvm/Transforms/InstCombine/InstCombiner.h"
33#include <cassert>
34#include <utility>

36using namespace llvm;
37using namespace PatternMatch;

39#define DEBUG_TYPE"instcombine" "instcombine"

41namespace {

/// Class representing coefficient of floating-point addend.
/// This class needs to be highly efficient, which is especially true for
/// the constructor. As of I write this comment, the cost of the default
/// constructor is merely 4-byte-store-zero (Assuming compiler is able to
/// perform write-merging).
///
class FAddendCoef {
public:
  // The constructor has to initialize a APFloat, which is unnecessary for
  // most addends which have coefficient either 1 or -1. So, the constructor
  // is expensive. In order to avoid the cost of the constructor, we should
  // reuse some instances whenever possible. The pre-created instances
  // FAddCombine::Add[0-5] embodies this idea.
  FAddendCoef() = default;
  ~FAddendCoef();

  // If possible, don't define operator+/operator- etc because these
  // operators inevitably call FAddendCoef's constructor which is not cheap.
  void operator=(const FAddendCoef &A);
  void operator+=(const FAddendCoef &A);
  void operator*=(const FAddendCoef &S);

  void set(short C) {
    assert(!insaneIntVal(C) && "Insane coefficient")(static_cast <bool> (!insaneIntVal(C) && "Insane coefficient"
) ? void (0) : __assert_fail ("!insaneIntVal(C) && \"Insane coefficient\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 66, __extension__ __PRETTY_FUNCTION__));
    IsFp = false; IntVal = C;
  }

  void set(const APFloat& C);

  void negate();

  bool isZero() const { return isInt() ? !IntVal : getFpVal().isZero(); }
  Value *getValue(Type *) const;

  bool isOne() const { return isInt() && IntVal == 1; }
  bool isTwo() const { return isInt() && IntVal == 2; }
  bool isMinusOne() const { return isInt() && IntVal == -1; }
  bool isMinusTwo() const { return isInt() && IntVal == -2; }

private:
  bool insaneIntVal(int V) { return V > 4 || V < -4; }

  APFloat *getFpValPtr() { return reinterpret_cast<APFloat *>(&FpValBuf); }

  const APFloat *getFpValPtr() const {
    return reinterpret_cast<const APFloat *>(&FpValBuf);
  }

  const APFloat &getFpVal() const {
    assert(IsFp && BufHasFpVal && "Incorret state")(static_cast <bool> (IsFp && BufHasFpVal &&
 "Incorret state") ? void (0) : __assert_fail ("IsFp && BufHasFpVal && \"Incorret state\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 92, __extension__ __PRETTY_FUNCTION__));
    return *getFpValPtr();
  }

  APFloat &getFpVal() {
    assert(IsFp && BufHasFpVal && "Incorret state")(static_cast <bool> (IsFp && BufHasFpVal &&
 "Incorret state") ? void (0) : __assert_fail ("IsFp && BufHasFpVal && \"Incorret state\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 97, __extension__ __PRETTY_FUNCTION__));
    return *getFpValPtr();
  }

  bool isInt() const { return !IsFp; }

  // If the coefficient is represented by an integer, promote it to a
  // floating point.
  void convertToFpType(const fltSemantics &Sem);

  // Construct an APFloat from a signed integer.
  // TODO: We should get rid of this function when APFloat can be constructed
  //       from an *SIGNED* integer.
  APFloat createAPFloatFromInt(const fltSemantics &Sem, int Val);

  bool IsFp = false;

  // True iff FpValBuf contains an instance of APFloat.
  bool BufHasFpVal = false;

  // The integer coefficient of an individual addend is either 1 or -1,
  // and we try to simplify at most 4 addends from neighboring at most
  // two instructions. So the range of <IntVal> falls in [-4, 4]. APInt
  // is overkill of this end.
  short IntVal = 0;

  AlignedCharArrayUnion<APFloat> FpValBuf;
};

/// FAddend is used to represent floating-point addend. An addend is
/// represented as <C, V>, where the V is a symbolic value, and C is a
/// constant coefficient. A constant addend is represented as <C, 0>.
class FAddend {
public:
  FAddend() = default;

  void operator+=(const FAddend &T) {
    assert((Val == T.Val) && "Symbolic-values disagree")(static_cast <bool> ((Val == T.Val) && "Symbolic-values disagree"
) ? void (0) : __assert_fail ("(Val == T.Val) && \"Symbolic-values disagree\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 134, __extension__ __PRETTY_FUNCTION__));
    Coeff += T.Coeff;
  }

  Value *getSymVal() const { return Val; }
  const FAddendCoef &getCoef() const { return Coeff; }

  bool isConstant() const { return Val == nullptr; }
  bool isZero() const { return Coeff.isZero(); }

  void set(short Coefficient, Value *V) {
    Coeff.set(Coefficient);
    Val = V;
  }
  void set(const APFloat &Coefficient, Value *V) {
    Coeff.set(Coefficient);
    Val = V;
  }
  void set(const ConstantFP *Coefficient, Value *V) {
    Coeff.set(Coefficient->getValueAPF());
    Val = V;
  }

  void negate() { Coeff.negate(); }

  /// Drill down the U-D chain one step to find the definition of V, and
  /// try to break the definition into one or two addends.
  static unsigned drillValueDownOneStep(Value* V, FAddend &A0, FAddend &A1);

  /// Similar to FAddend::drillDownOneStep() except that the value being
  /// splitted is the addend itself.
  unsigned drillAddendDownOneStep(FAddend &Addend0, FAddend &Addend1) const;

private:
  void Scale(const FAddendCoef& ScaleAmt) { Coeff *= ScaleAmt; }

  // This addend has the value of "Coeff * Val".
  Value *Val = nullptr;
  FAddendCoef Coeff;
};

/// FAddCombine is the class for optimizing an unsafe fadd/fsub along
/// with its neighboring at most two instructions.
///
class FAddCombine {
public:
  FAddCombine(InstCombiner::BuilderTy &B) : Builder(B) {}

  Value *simplify(Instruction *FAdd);

private:
  using AddendVect = SmallVector<const FAddend *, 4>;

  Value *simplifyFAdd(AddendVect& V, unsigned InstrQuota);

  /// Convert given addend to a Value
  Value *createAddendVal(const FAddend &A, bool& NeedNeg);

  /// Return the number of instructions needed to emit the N-ary addition.
  unsigned calcInstrNumber(const AddendVect& Vect);

  Value *createFSub(Value *Opnd0, Value *Opnd1);
  Value *createFAdd(Value *Opnd0, Value *Opnd1);
  Value *createFMul(Value *Opnd0, Value *Opnd1);
  Value *createFNeg(Value *V);
  Value *createNaryFAdd(const AddendVect& Opnds, unsigned InstrQuota);
  void createInstPostProc(Instruction *NewInst, bool NoNumber = false);

   // Debugging stuff are clustered here.
  #ifndef NDEBUG
    unsigned CreateInstrNum;
    void initCreateInstNum() { CreateInstrNum = 0; }
    void incCreateInstNum() { CreateInstrNum++; }
  #else
    void initCreateInstNum() {}
    void incCreateInstNum() {}
  #endif

  InstCombiner::BuilderTy &Builder;
  Instruction *Instr = nullptr;
};

216} // end anonymous namespace

218//===----------------------------------------------------------------------===//
219//
220// Implementation of
221//    {FAddendCoef, FAddend, FAddition, FAddCombine}.
222//
223//===----------------------------------------------------------------------===//
224FAddendCoef::~FAddendCoef() {
if (BufHasFpVal)
  getFpValPtr()->~APFloat();
227}

229void FAddendCoef::set(const APFloat& C) {
APFloat *P = getFpValPtr();

if (isInt()) {
  // As the buffer is meanless byte stream, we cannot call
  // APFloat::operator=().
  new(P) APFloat(C);
} else
  *P = C;

IsFp = BufHasFpVal = true;
240}

242void FAddendCoef::convertToFpType(const fltSemantics &Sem) {
if (!isInt())
  return;

APFloat *P = getFpValPtr();
if (IntVal > 0)
  new(P) APFloat(Sem, IntVal);
else {
  new(P) APFloat(Sem, 0 - IntVal);
  P->changeSign();
}
IsFp = BufHasFpVal = true;
254}

256APFloat FAddendCoef::createAPFloatFromInt(const fltSemantics &Sem, int Val) {
if (Val >= 0)
  return APFloat(Sem, Val);

APFloat T(Sem, 0 - Val);
T.changeSign();

return T;
264}

266void FAddendCoef::operator=(const FAddendCoef &That) {
if (That.isInt())
  set(That.IntVal);
else
  set(That.getFpVal());
271}

273void FAddendCoef::operator+=(const FAddendCoef &That) {
RoundingMode RndMode = RoundingMode::NearestTiesToEven;
if (isInt() == That.isInt()) {
  if (isInt())
    IntVal += That.IntVal;
  else
    getFpVal().add(That.getFpVal(), RndMode);
  return;
}

if (isInt()) {
  const APFloat &T = That.getFpVal();
  convertToFpType(T.getSemantics());
  getFpVal().add(T, RndMode);
  return;
}

APFloat &T = getFpVal();
T.add(createAPFloatFromInt(T.getSemantics(), That.IntVal), RndMode);
292}

294void FAddendCoef::operator*=(const FAddendCoef &That) {
if (That.isOne())
  return;

if (That.isMinusOne()) {
  negate();
  return;
}

if (isInt() && That.isInt()) {
  int Res = IntVal * (int)That.IntVal;
  assert(!insaneIntVal(Res) && "Insane int value")(static_cast <bool> (!insaneIntVal(Res) && "Insane int value"
) ? void (0) : __assert_fail ("!insaneIntVal(Res) && \"Insane int value\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 305, __extension__ __PRETTY_FUNCTION__));
  IntVal = Res;
  return;
}

const fltSemantics &Semantic =
  isInt() ? That.getFpVal().getSemantics() : getFpVal().getSemantics();

if (isInt())
  convertToFpType(Semantic);
APFloat &F0 = getFpVal();

if (That.isInt())
  F0.multiply(createAPFloatFromInt(Semantic, That.IntVal),
              APFloat::rmNearestTiesToEven);
else
  F0.multiply(That.getFpVal(), APFloat::rmNearestTiesToEven);
322}

324void FAddendCoef::negate() {
if (isInt())
  IntVal = 0 - IntVal;
else
  getFpVal().changeSign();
329}

331Value *FAddendCoef::getValue(Type *Ty) const {
return isInt() ?
  ConstantFP::get(Ty, float(IntVal)) :
  ConstantFP::get(Ty->getContext(), getFpVal());
335}

337// The definition of <Val>     Addends
338// =========================================
339//  A + B                     <1, A>, <1,B>
340//  A - B                     <1, A>, <1,B>
341//  0 - B                     <-1, B>
342//  C * A,                    <C, A>
343//  A + C                     <1, A> <C, NULL>
344//  0 +/- 0                   <0, NULL> (corner case)
345//
346// Legend: A and B are not constant, C is constant
347unsigned FAddend::drillValueDownOneStep
(Value *Val, FAddend &Addend0, FAddend &Addend1) {
Instruction *I = nullptr;
if (!Val || !(I = dyn_cast<Instruction>(Val)))
  return 0;

unsigned Opcode = I->getOpcode();

if (Opcode == Instruction::FAdd || Opcode == Instruction::FSub) {
  ConstantFP *C0, *C1;
  Value *Opnd0 = I->getOperand(0);
  Value *Opnd1 = I->getOperand(1);
  if ((C0 = dyn_cast<ConstantFP>(Opnd0)) && C0->isZero())
    Opnd0 = nullptr;

  if ((C1 = dyn_cast<ConstantFP>(Opnd1)) && C1->isZero())
    Opnd1 = nullptr;

  if (Opnd0) {
    if (!C0)
      Addend0.set(1, Opnd0);
    else
      Addend0.set(C0, nullptr);
  }

  if (Opnd1) {
    FAddend &Addend = Opnd0 ? Addend1 : Addend0;
    if (!C1)
      Addend.set(1, Opnd1);
    else
      Addend.set(C1, nullptr);
    if (Opcode == Instruction::FSub)
      Addend.negate();
  }

  if (Opnd0 || Opnd1)
    return Opnd0 && Opnd1 ? 2 : 1;

  // Both operands are zero. Weird!
  Addend0.set(APFloat(C0->getValueAPF().getSemantics()), nullptr);
  return 1;
}

if (I->getOpcode() == Instruction::FMul) {
  Value *V0 = I->getOperand(0);
  Value *V1 = I->getOperand(1);
  if (ConstantFP *C = dyn_cast<ConstantFP>(V0)) {
    Addend0.set(C, V1);
    return 1;
  }

  if (ConstantFP *C = dyn_cast<ConstantFP>(V1)) {
    Addend0.set(C, V0);
    return 1;
  }
}

return 0;
405}

407// Try to break *this* addend into two addends. e.g. Suppose this addend is
408// <2.3, V>, and V = X + Y, by calling this function, we obtain two addends,
409// i.e. <2.3, X> and <2.3, Y>.
410unsigned FAddend::drillAddendDownOneStep
(FAddend &Addend0, FAddend &Addend1) const {
if (isConstant())
  return 0;

unsigned BreakNum = FAddend::drillValueDownOneStep(Val, Addend0, Addend1);
if (!BreakNum || Coeff.isOne())
  return BreakNum;

Addend0.Scale(Coeff);

if (BreakNum == 2)
  Addend1.Scale(Coeff);

return BreakNum;
425}

427Value *FAddCombine::simplify(Instruction *I) {
assert(I->hasAllowReassoc() && I->hasNoSignedZeros() &&(static_cast <bool> (I->hasAllowReassoc() &&
 I->hasNoSignedZeros() && "Expected 'reassoc'+'nsz' instruction"
) ? void (0) : __assert_fail ("I->hasAllowReassoc() && I->hasNoSignedZeros() && \"Expected 'reassoc'+'nsz' instruction\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 429, __extension__ __PRETTY_FUNCTION__))
       "Expected 'reassoc'+'nsz' instruction")(static_cast <bool> (I->hasAllowReassoc() &&
 I->hasNoSignedZeros() && "Expected 'reassoc'+'nsz' instruction"
) ? void (0) : __assert_fail ("I->hasAllowReassoc() && I->hasNoSignedZeros() && \"Expected 'reassoc'+'nsz' instruction\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 429, __extension__ __PRETTY_FUNCTION__));

// Currently we are not able to handle vector type.
if (I->getType()->isVectorTy())
  return nullptr;

assert((I->getOpcode() == Instruction::FAdd ||(static_cast <bool> ((I->getOpcode() == Instruction::
FAdd || I->getOpcode() == Instruction::FSub) && "Expect add/sub"
) ? void (0) : __assert_fail ("(I->getOpcode() == Instruction::FAdd || I->getOpcode() == Instruction::FSub) && \"Expect add/sub\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 436, __extension__ __PRETTY_FUNCTION__))
        I->getOpcode() == Instruction::FSub) && "Expect add/sub")(static_cast <bool> ((I->getOpcode() == Instruction::
FAdd || I->getOpcode() == Instruction::FSub) && "Expect add/sub"
) ? void (0) : __assert_fail ("(I->getOpcode() == Instruction::FAdd || I->getOpcode() == Instruction::FSub) && \"Expect add/sub\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 436, __extension__ __PRETTY_FUNCTION__));

// Save the instruction before calling other member-functions.
Instr = I;

FAddend Opnd0, Opnd1, Opnd0_0, Opnd0_1, Opnd1_0, Opnd1_1;

unsigned OpndNum = FAddend::drillValueDownOneStep(I, Opnd0, Opnd1);

// Step 1: Expand the 1st addend into Opnd0_0 and Opnd0_1.
unsigned Opnd0_ExpNum = 0;
unsigned Opnd1_ExpNum = 0;

if (!Opnd0.isConstant())
  Opnd0_ExpNum = Opnd0.drillAddendDownOneStep(Opnd0_0, Opnd0_1);

// Step 2: Expand the 2nd addend into Opnd1_0 and Opnd1_1.
if (OpndNum == 2 && !Opnd1.isConstant())
  Opnd1_ExpNum = Opnd1.drillAddendDownOneStep(Opnd1_0, Opnd1_1);

// Step 3: Try to optimize Opnd0_0 + Opnd0_1 + Opnd1_0 + Opnd1_1
if (Opnd0_ExpNum && Opnd1_ExpNum) {
  AddendVect AllOpnds;
  AllOpnds.push_back(&Opnd0_0);
  AllOpnds.push_back(&Opnd1_0);
  if (Opnd0_ExpNum == 2)
    AllOpnds.push_back(&Opnd0_1);
  if (Opnd1_ExpNum == 2)
    AllOpnds.push_back(&Opnd1_1);

  // Compute instruction quota. We should save at least one instruction.
  unsigned InstQuota = 0;

  Value *V0 = I->getOperand(0);
  Value *V1 = I->getOperand(1);
  InstQuota = ((!isa<Constant>(V0) && V0->hasOneUse()) &&
               (!isa<Constant>(V1) && V1->hasOneUse())) ? 2 : 1;

  if (Value *R = simplifyFAdd(AllOpnds, InstQuota))
    return R;
}

if (OpndNum != 2) {
  // The input instruction is : "I=0.0 +/- V". If the "V" were able to be
  // splitted into two addends, say "V = X - Y", the instruction would have
  // been optimized into "I = Y - X" in the previous steps.
  //
  const FAddendCoef &CE = Opnd0.getCoef();
  return CE.isOne() ? Opnd0.getSymVal() : nullptr;
}

// step 4: Try to optimize Opnd0 + Opnd1_0 [+ Opnd1_1]
if (Opnd1_ExpNum) {
  AddendVect AllOpnds;
  AllOpnds.push_back(&Opnd0);
  AllOpnds.push_back(&Opnd1_0);
  if (Opnd1_ExpNum == 2)
    AllOpnds.push_back(&Opnd1_1);

  if (Value *R = simplifyFAdd(AllOpnds, 1))
    return R;
}

// step 5: Try to optimize Opnd1 + Opnd0_0 [+ Opnd0_1]
if (Opnd0_ExpNum) {
  AddendVect AllOpnds;
  AllOpnds.push_back(&Opnd1);
  AllOpnds.push_back(&Opnd0_0);
  if (Opnd0_ExpNum == 2)
    AllOpnds.push_back(&Opnd0_1);

  if (Value *R = simplifyFAdd(AllOpnds, 1))
    return R;
}

return nullptr;
512}

514Value *FAddCombine::simplifyFAdd(AddendVect& Addends, unsigned InstrQuota) {
unsigned AddendNum = Addends.size();
assert(AddendNum <= 4 && "Too many addends")(static_cast <bool> (AddendNum <= 4 && "Too many addends"
) ? void (0) : __assert_fail ("AddendNum <= 4 && \"Too many addends\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 516, __extension__ __PRETTY_FUNCTION__));

// For saving intermediate results;
unsigned NextTmpIdx = 0;
FAddend TmpResult[3];

// Points to the constant addend of the resulting simplified expression.
// If the resulting expr has constant-addend, this constant-addend is
// desirable to reside at the top of the resulting expression tree. Placing
// constant close to supper-expr(s) will potentially reveal some optimization
// opportunities in super-expr(s).
const FAddend *ConstAdd = nullptr;

// Simplified addends are placed <SimpVect>.
AddendVect SimpVect;

// The outer loop works on one symbolic-value at a time. Suppose the input
// addends are : <a1, x>, <b1, y>, <a2, x>, <c1, z>, <b2, y>, ...
// The symbolic-values will be processed in this order: x, y, z.
for (unsigned SymIdx = 0; SymIdx < AddendNum; SymIdx++) {

  const FAddend *ThisAddend = Addends[SymIdx];
  if (!ThisAddend) {
    // This addend was processed before.
    continue;
  }

  Value *Val = ThisAddend->getSymVal();
  unsigned StartIdx = SimpVect.size();
  SimpVect.push_back(ThisAddend);

  // The inner loop collects addends sharing same symbolic-value, and these
  // addends will be later on folded into a single addend. Following above
  // example, if the symbolic value "y" is being processed, the inner loop
  // will collect two addends "<b1,y>" and "<b2,Y>". These two addends will
  // be later on folded into "<b1+b2, y>".
  for (unsigned SameSymIdx = SymIdx + 1;
       SameSymIdx < AddendNum; SameSymIdx++) {
    const FAddend *T = Addends[SameSymIdx];
    if (T && T->getSymVal() == Val) {
      // Set null such that next iteration of the outer loop will not process
      // this addend again.
      Addends[SameSymIdx] = nullptr;
      SimpVect.push_back(T);
    }
  }

  // If multiple addends share same symbolic value, fold them together.
  if (StartIdx + 1 != SimpVect.size()) {
    FAddend &R = TmpResult[NextTmpIdx ++];
    R = *SimpVect[StartIdx];
    for (unsigned Idx = StartIdx + 1; Idx < SimpVect.size(); Idx++)
      R += *SimpVect[Idx];

    // Pop all addends being folded and push the resulting folded addend.
    SimpVect.resize(StartIdx);
    if (Val) {
      if (!R.isZero()) {
        SimpVect.push_back(&R);
      }
    } else {
      // Don't push constant addend at this time. It will be the last element
      // of <SimpVect>.
      ConstAdd = &R;
    }
  }
}

assert((NextTmpIdx <= array_lengthof(TmpResult) + 1) &&(static_cast <bool> ((NextTmpIdx <= array_lengthof(TmpResult
) + 1) && "out-of-bound access") ? void (0) : __assert_fail
 ("(NextTmpIdx <= array_lengthof(TmpResult) + 1) && \"out-of-bound access\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 585, __extension__ __PRETTY_FUNCTION__))
       "out-of-bound access")(static_cast <bool> ((NextTmpIdx <= array_lengthof(TmpResult
) + 1) && "out-of-bound access") ? void (0) : __assert_fail
 ("(NextTmpIdx <= array_lengthof(TmpResult) + 1) && \"out-of-bound access\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 585, __extension__ __PRETTY_FUNCTION__));

if (ConstAdd)
  SimpVect.push_back(ConstAdd);

Value *Result;
if (!SimpVect.empty())
  Result = createNaryFAdd(SimpVect, InstrQuota);
else {
  // The addition is folded to 0.0.
  Result = ConstantFP::get(Instr->getType(), 0.0);
}

return Result;
599}

601Value *FAddCombine::createNaryFAdd
(const AddendVect &Opnds, unsigned InstrQuota) {
assert(!Opnds.empty() && "Expect at least one addend")(static_cast <bool> (!Opnds.empty() && "Expect at least one addend"
) ? void (0) : __assert_fail ("!Opnds.empty() && \"Expect at least one addend\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 603, __extension__ __PRETTY_FUNCTION__));

// Step 1: Check if the # of instructions needed exceeds the quota.

unsigned InstrNeeded = calcInstrNumber(Opnds);
if (InstrNeeded > InstrQuota)
  return nullptr;

initCreateInstNum();

// step 2: Emit the N-ary addition.
// Note that at most three instructions are involved in Fadd-InstCombine: the
// addition in question, and at most two neighboring instructions.
// The resulting optimized addition should have at least one less instruction
// than the original addition expression tree. This implies that the resulting
// N-ary addition has at most two instructions, and we don't need to worry
// about tree-height when constructing the N-ary addition.

Value *LastVal = nullptr;
bool LastValNeedNeg = false;

// Iterate the addends, creating fadd/fsub using adjacent two addends.
for (const FAddend *Opnd : Opnds) {
  bool NeedNeg;
  Value *V = createAddendVal(*Opnd, NeedNeg);
  if (!LastVal) {
    LastVal = V;
    LastValNeedNeg = NeedNeg;
    continue;
  }

  if (LastValNeedNeg == NeedNeg) {
    LastVal = createFAdd(LastVal, V);
    continue;
  }

  if (LastValNeedNeg)
    LastVal = createFSub(V, LastVal);
  else
    LastVal = createFSub(LastVal, V);

  LastValNeedNeg = false;
}

if (LastValNeedNeg) {
  LastVal = createFNeg(LastVal);
}

651#ifndef NDEBUG
assert(CreateInstrNum == InstrNeeded &&(static_cast <bool> (CreateInstrNum == InstrNeeded &&
 "Inconsistent in instruction numbers") ? void (0) : __assert_fail
 ("CreateInstrNum == InstrNeeded && \"Inconsistent in instruction numbers\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 653, __extension__ __PRETTY_FUNCTION__))
       "Inconsistent in instruction numbers")(static_cast <bool> (CreateInstrNum == InstrNeeded &&
 "Inconsistent in instruction numbers") ? void (0) : __assert_fail
 ("CreateInstrNum == InstrNeeded && \"Inconsistent in instruction numbers\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 653, __extension__ __PRETTY_FUNCTION__));
654#endif

return LastVal;
657}

659Value *FAddCombine::createFSub(Value *Opnd0, Value *Opnd1) {
Value *V = Builder.CreateFSub(Opnd0, Opnd1);
if (Instruction *I = dyn_cast<Instruction>(V))
  createInstPostProc(I);
return V;
664}

666Value *FAddCombine::createFNeg(Value *V) {
Value *NewV = Builder.CreateFNeg(V);
if (Instruction *I = dyn_cast<Instruction>(NewV))
  createInstPostProc(I, true); // fneg's don't receive instruction numbers.
return NewV;
671}

673Value *FAddCombine::createFAdd(Value *Opnd0, Value *Opnd1) {
Value *V = Builder.CreateFAdd(Opnd0, Opnd1);
if (Instruction *I = dyn_cast<Instruction>(V))
  createInstPostProc(I);
return V;
678}

680Value *FAddCombine::createFMul(Value *Opnd0, Value *Opnd1) {
Value *V = Builder.CreateFMul(Opnd0, Opnd1);
if (Instruction *I = dyn_cast<Instruction>(V))
  createInstPostProc(I);
return V;
685}

687void FAddCombine::createInstPostProc(Instruction *NewInstr, bool NoNumber) {
NewInstr->setDebugLoc(Instr->getDebugLoc());

// Keep track of the number of instruction created.
if (!NoNumber)
  incCreateInstNum();

// Propagate fast-math flags
NewInstr->setFastMathFlags(Instr->getFastMathFlags());
696}

698// Return the number of instruction needed to emit the N-ary addition.
699// NOTE: Keep this function in sync with createAddendVal().
700unsigned FAddCombine::calcInstrNumber(const AddendVect &Opnds) {
unsigned OpndNum = Opnds.size();
unsigned InstrNeeded = OpndNum - 1;

// The number of addends in the form of "(-1)*x".
unsigned NegOpndNum = 0;

// Adjust the number of instructions needed to emit the N-ary add.
for (const FAddend *Opnd : Opnds) {
  if (Opnd->isConstant())
    continue;

  // The constant check above is really for a few special constant
  // coefficients.
  if (isa<UndefValue>(Opnd->getSymVal()))
    continue;

  const FAddendCoef &CE = Opnd->getCoef();
  if (CE.isMinusOne() || CE.isMinusTwo())
    NegOpndNum++;

  // Let the addend be "c * x". If "c == +/-1", the value of the addend
  // is immediately available; otherwise, it needs exactly one instruction
  // to evaluate the value.
  if (!CE.isMinusOne() && !CE.isOne())
    InstrNeeded++;
}
return InstrNeeded;
728}

730// Input Addend        Value           NeedNeg(output)
731// ================================================================
732// Constant C          C               false
733// <+/-1, V>           V               coefficient is -1
734// <2/-2, V>          "fadd V, V"      coefficient is -2
735// <C, V>             "fmul V, C"      false
736//
737// NOTE: Keep this function in sync with FAddCombine::calcInstrNumber.
738Value *FAddCombine::createAddendVal(const FAddend &Opnd, bool &NeedNeg) {
const FAddendCoef &Coeff = Opnd.getCoef();

if (Opnd.isConstant()) {
  NeedNeg = false;
  return Coeff.getValue(Instr->getType());
}

Value *OpndVal = Opnd.getSymVal();

if (Coeff.isMinusOne() || Coeff.isOne()) {
  NeedNeg = Coeff.isMinusOne();
  return OpndVal;
}

if (Coeff.isTwo() || Coeff.isMinusTwo()) {
  NeedNeg = Coeff.isMinusTwo();
  return createFAdd(OpndVal, OpndVal);
}

NeedNeg = false;
return createFMul(OpndVal, Coeff.getValue(Instr->getType()));
760}

762// Checks if any operand is negative and we can convert add to sub.
763// This function checks for following negative patterns
764//   ADD(XOR(OR(Z, NOT(C)), C)), 1) == NEG(AND(Z, C))
765//   ADD(XOR(AND(Z, C), C), 1) == NEG(OR(Z, ~C))
766//   XOR(AND(Z, C), (C + 1)) == NEG(OR(Z, ~C)) if C is even
767static Value *checkForNegativeOperand(BinaryOperator &I,
                                    InstCombiner::BuilderTy &Builder) {
Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);

// This function creates 2 instructions to replace ADD, we need at least one
// of LHS or RHS to have one use to ensure benefit in transform.
if (!LHS->hasOneUse() && !RHS->hasOneUse())
  return nullptr;

Value *X = nullptr, *Y = nullptr, *Z = nullptr;
const APInt *C1 = nullptr, *C2 = nullptr;

// if ONE is on other side, swap
if (match(RHS, m_Add(m_Value(X), m_One())))
  std::swap(LHS, RHS);

if (match(LHS, m_Add(m_Value(X), m_One()))) {
  // if XOR on other side, swap
  if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
    std::swap(X, RHS);

  if (match(X, m_Xor(m_Value(Y), m_APInt(C1)))) {
    // X = XOR(Y, C1), Y = OR(Z, C2), C2 = NOT(C1) ==> X == NOT(AND(Z, C1))
    // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, AND(Z, C1))
    if (match(Y, m_Or(m_Value(Z), m_APInt(C2))) && (*C2 == ~(*C1))) {
      Value *NewAnd = Builder.CreateAnd(Z, *C1);
      return Builder.CreateSub(RHS, NewAnd, "sub");
    } else if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && (*C1 == *C2)) {
      // X = XOR(Y, C1), Y = AND(Z, C2), C2 == C1 ==> X == NOT(OR(Z, ~C1))
      // ADD(ADD(X, 1), RHS) == ADD(X, ADD(RHS, 1)) == SUB(RHS, OR(Z, ~C1))
      Value *NewOr = Builder.CreateOr(Z, ~(*C1));
      return Builder.CreateSub(RHS, NewOr, "sub");
    }
  }
}

// Restore LHS and RHS
LHS = I.getOperand(0);
RHS = I.getOperand(1);

// if XOR is on other side, swap
if (match(RHS, m_Xor(m_Value(Y), m_APInt(C1))))
  std::swap(LHS, RHS);

// C2 is ODD
// LHS = XOR(Y, C1), Y = AND(Z, C2), C1 == (C2 + 1) => LHS == NEG(OR(Z, ~C2))
// ADD(LHS, RHS) == SUB(RHS, OR(Z, ~C2))
if (match(LHS, m_Xor(m_Value(Y), m_APInt(C1))))
  if (C1->countTrailingZeros() == 0)
    if (match(Y, m_And(m_Value(Z), m_APInt(C2))) && *C1 == (*C2 + 1)) {
      Value *NewOr = Builder.CreateOr(Z, ~(*C2));
      return Builder.CreateSub(RHS, NewOr, "sub");
    }
return nullptr;
821}

823/// Wrapping flags may allow combining constants separated by an extend.
824static Instruction *foldNoWrapAdd(BinaryOperator &Add,
                                InstCombiner::BuilderTy &Builder) {
Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
Type *Ty = Add.getType();
Constant *Op1C;
if (!match(Op1, m_Constant(Op1C)))
  return nullptr;

// Try this match first because it results in an add in the narrow type.
// (zext (X +nuw C2)) + C1 --> zext (X + (C2 + trunc(C1)))
Value *X;
const APInt *C1, *C2;
if (match(Op1, m_APInt(C1)) &&
    match(Op0, m_OneUse(m_ZExt(m_NUWAdd(m_Value(X), m_APInt(C2))))) &&
    C1->isNegative() && C1->sge(-C2->sext(C1->getBitWidth()))) {
  Constant *NewC =
      ConstantInt::get(X->getType(), *C2 + C1->trunc(C2->getBitWidth()));
  return new ZExtInst(Builder.CreateNUWAdd(X, NewC), Ty);
}

// More general combining of constants in the wide type.
// (sext (X +nsw NarrowC)) + C --> (sext X) + (sext(NarrowC) + C)
Constant *NarrowC;
if (match(Op0, m_OneUse(m_SExt(m_NSWAdd(m_Value(X), m_Constant(NarrowC)))))) {
  Constant *WideC = ConstantExpr::getSExt(NarrowC, Ty);
  Constant *NewC = ConstantExpr::getAdd(WideC, Op1C);
  Value *WideX = Builder.CreateSExt(X, Ty);
  return BinaryOperator::CreateAdd(WideX, NewC);
}
// (zext (X +nuw NarrowC)) + C --> (zext X) + (zext(NarrowC) + C)
if (match(Op0, m_OneUse(m_ZExt(m_NUWAdd(m_Value(X), m_Constant(NarrowC)))))) {
  Constant *WideC = ConstantExpr::getZExt(NarrowC, Ty);
  Constant *NewC = ConstantExpr::getAdd(WideC, Op1C);
  Value *WideX = Builder.CreateZExt(X, Ty);
  return BinaryOperator::CreateAdd(WideX, NewC);
}

return nullptr;
862}

864Instruction *InstCombinerImpl::foldAddWithConstant(BinaryOperator &Add) {
Value *Op0 = Add.getOperand(0), *Op1 = Add.getOperand(1);
Constant *Op1C;
if (!match(Op1, m_ImmConstant(Op1C)))
  return nullptr;

if (Instruction *NV = foldBinOpIntoSelectOrPhi(Add))
  return NV;

Value *X;
Constant *Op00C;

// add (sub C1, X), C2 --> sub (add C1, C2), X
if (match(Op0, m_Sub(m_Constant(Op00C), m_Value(X))))
  return BinaryOperator::CreateSub(ConstantExpr::getAdd(Op00C, Op1C), X);

Value *Y;

// add (sub X, Y), -1 --> add (not Y), X
if (match(Op0, m_OneUse(m_Sub(m_Value(X), m_Value(Y)))) &&
    match(Op1, m_AllOnes()))
  return BinaryOperator::CreateAdd(Builder.CreateNot(Y), X);

// zext(bool) + C -> bool ? C + 1 : C
if (match(Op0, m_ZExt(m_Value(X))) &&
    X->getType()->getScalarSizeInBits() == 1)
  return SelectInst::Create(X, InstCombiner::AddOne(Op1C), Op1);
// sext(bool) + C -> bool ? C - 1 : C
if (match(Op0, m_SExt(m_Value(X))) &&
    X->getType()->getScalarSizeInBits() == 1)
  return SelectInst::Create(X, InstCombiner::SubOne(Op1C), Op1);

// ~X + C --> (C-1) - X
if (match(Op0, m_Not(m_Value(X))))
  return BinaryOperator::CreateSub(InstCombiner::SubOne(Op1C), X);

const APInt *C;
if (!match(Op1, m_APInt(C)))
  return nullptr;

// (X | Op01C) + Op1C --> X + (Op01C + Op1C) iff the `or` is actually an `add`
Constant *Op01C;
if (match(Op0, m_Or(m_Value(X), m_ImmConstant(Op01C))) &&
    haveNoCommonBitsSet(X, Op01C, DL, &AC, &Add, &DT))
  return BinaryOperator::CreateAdd(X, ConstantExpr::getAdd(Op01C, Op1C));

// (X | C2) + C --> (X | C2) ^ C2 iff (C2 == -C)
const APInt *C2;
if (match(Op0, m_Or(m_Value(), m_APInt(C2))) && *C2 == -*C)
  return BinaryOperator::CreateXor(Op0, ConstantInt::get(Add.getType(), *C2));

if (C->isSignMask()) {
  // If wrapping is not allowed, then the addition must set the sign bit:
  // X + (signmask) --> X | signmask
  if (Add.hasNoSignedWrap() || Add.hasNoUnsignedWrap())
    return BinaryOperator::CreateOr(Op0, Op1);

  // If wrapping is allowed, then the addition flips the sign bit of LHS:
  // X + (signmask) --> X ^ signmask
  return BinaryOperator::CreateXor(Op0, Op1);
}

// Is this add the last step in a convoluted sext?
// add(zext(xor i16 X, -32768), -32768) --> sext X
Type *Ty = Add.getType();
if (match(Op0, m_ZExt(m_Xor(m_Value(X), m_APInt(C2)))) &&
    C2->isMinSignedValue() && C2->sext(Ty->getScalarSizeInBits()) == *C)
  return CastInst::Create(Instruction::SExt, X, Ty);

if (match(Op0, m_Xor(m_Value(X), m_APInt(C2)))) {
  // (X ^ signmask) + C --> (X + (signmask ^ C))
  if (C2->isSignMask())
    return BinaryOperator::CreateAdd(X, ConstantInt::get(Ty, *C2 ^ *C));

  // If X has no high-bits set above an xor mask:
  // add (xor X, LowMaskC), C --> sub (LowMaskC + C), X
  if (C2->isMask()) {
    KnownBits LHSKnown = computeKnownBits(X, 0, &Add);
    if ((*C2 | LHSKnown.Zero).isAllOnesValue())
      return BinaryOperator::CreateSub(ConstantInt::get(Ty, *C2 + *C), X);
  }

  // Look for a math+logic pattern that corresponds to sext-in-register of a
  // value with cleared high bits. Convert that into a pair of shifts:
  // add (xor X, 0x80), 0xF..F80 --> (X << ShAmtC) >>s ShAmtC
  // add (xor X, 0xF..F80), 0x80 --> (X << ShAmtC) >>s ShAmtC
  if (Op0->hasOneUse() && *C2 == -(*C)) {
    unsigned BitWidth = Ty->getScalarSizeInBits();
    unsigned ShAmt = 0;
    if (C->isPowerOf2())
      ShAmt = BitWidth - C->logBase2() - 1;
    else if (C2->isPowerOf2())
      ShAmt = BitWidth - C2->logBase2() - 1;
    if (ShAmt && MaskedValueIsZero(X, APInt::getHighBitsSet(BitWidth, ShAmt),
                                   0, &Add)) {
      Constant *ShAmtC = ConstantInt::get(Ty, ShAmt);
      Value *NewShl = Builder.CreateShl(X, ShAmtC, "sext");
      return BinaryOperator::CreateAShr(NewShl, ShAmtC);
    }
  }
}

if (C->isOneValue() && Op0->hasOneUse()) {
  // add (sext i1 X), 1 --> zext (not X)
  // TODO: The smallest IR representation is (select X, 0, 1), and that would
  // not require the one-use check. But we need to remove a transform in
  // visitSelect and make sure that IR value tracking for select is equal or
  // better than for these ops.
  if (match(Op0, m_SExt(m_Value(X))) &&
      X->getType()->getScalarSizeInBits() == 1)
    return new ZExtInst(Builder.CreateNot(X), Ty);

  // Shifts and add used to flip and mask off the low bit:
  // add (ashr (shl i32 X, 31), 31), 1 --> and (not X), 1
  const APInt *C3;
  if (match(Op0, m_AShr(m_Shl(m_Value(X), m_APInt(C2)), m_APInt(C3))) &&
      C2 == C3 && *C2 == Ty->getScalarSizeInBits() - 1) {
    Value *NotX = Builder.CreateNot(X);
    return BinaryOperator::CreateAnd(NotX, ConstantInt::get(Ty, 1));
  }
}

// If all bits affected by the add are included in a high-bit-mask, do the
// add before the mask op:
// (X & 0xFF00) + xx00 --> (X + xx00) & 0xFF00
if (match(Op0, m_OneUse(m_And(m_Value(X), m_APInt(C2)))) &&
    C2->isNegative() && C2->isShiftedMask() && *C == (*C & *C2)) {
  Value *NewAdd = Builder.CreateAdd(X, ConstantInt::get(Ty, *C));
  return BinaryOperator::CreateAnd(NewAdd, ConstantInt::get(Ty, *C2));
}

return nullptr;
996}

998// Matches multiplication expression Op * C where C is a constant. Returns the
999// constant value in C and the other operand in Op. Returns true if such a
1000// match is found.
1001static bool MatchMul(Value *E, Value *&Op, APInt &C) {
const APInt *AI;
if (match(E, m_Mul(m_Value(Op), m_APInt(AI)))) {
  C = *AI;
  return true;
}
if (match(E, m_Shl(m_Value(Op), m_APInt(AI)))) {
  C = APInt(AI->getBitWidth(), 1);
  C <<= *AI;
  return true;
}
return false;
1013}

1015// Matches remainder expression Op % C where C is a constant. Returns the
1016// constant value in C and the other operand in Op. Returns the signedness of
1017// the remainder operation in IsSigned. Returns true if such a match is
1018// found.
1019static bool MatchRem(Value *E, Value *&Op, APInt &C, bool &IsSigned) {
const APInt *AI;
IsSigned = false;
if (match(E, m_SRem(m_Value(Op), m_APInt(AI)))) {
  IsSigned = true;
  C = *AI;
  return true;
}
if (match(E, m_URem(m_Value(Op), m_APInt(AI)))) {
  C = *AI;
  return true;
}
if (match(E, m_And(m_Value(Op), m_APInt(AI))) && (*AI + 1).isPowerOf2()) {
  C = *AI + 1;
  return true;
}
return false;
1036}

1038// Matches division expression Op / C with the given signedness as indicated
1039// by IsSigned, where C is a constant. Returns the constant value in C and the
1040// other operand in Op. Returns true if such a match is found.
1041static bool MatchDiv(Value *E, Value *&Op, APInt &C, bool IsSigned) {
const APInt *AI;
if (IsSigned && match(E, m_SDiv(m_Value(Op), m_APInt(AI)))) {
  C = *AI;
  return true;
}
if (!IsSigned) {
  if (match(E, m_UDiv(m_Value(Op), m_APInt(AI)))) {
    C = *AI;
    return true;
  }
  if (match(E, m_LShr(m_Value(Op), m_APInt(AI)))) {
    C = APInt(AI->getBitWidth(), 1);
    C <<= *AI;
    return true;
  }
}
return false;
1059}

1061// Returns whether C0 * C1 with the given signedness overflows.
1062static bool MulWillOverflow(APInt &C0, APInt &C1, bool IsSigned) {
bool overflow;
if (IsSigned)
  (void)C0.smul_ov(C1, overflow);
else
  (void)C0.umul_ov(C1, overflow);
return overflow;
1069}

1071// Simplifies X % C0 + (( X / C0 ) % C1) * C0 to X % (C0 * C1), where (C0 * C1)
1072// does not overflow.
1073Value *InstCombinerImpl::SimplifyAddWithRemainder(BinaryOperator &I) {
Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
Value *X, *MulOpV;
APInt C0, MulOpC;
bool IsSigned;
// Match I = X % C0 + MulOpV * C0
if (((MatchRem(LHS, X, C0, IsSigned) && MatchMul(RHS, MulOpV, MulOpC)) ||
     (MatchRem(RHS, X, C0, IsSigned) && MatchMul(LHS, MulOpV, MulOpC))) &&
    C0 == MulOpC) {
  Value *RemOpV;
  APInt C1;
  bool Rem2IsSigned;
  // Match MulOpC = RemOpV % C1
  if (MatchRem(MulOpV, RemOpV, C1, Rem2IsSigned) &&
      IsSigned == Rem2IsSigned) {
    Value *DivOpV;
    APInt DivOpC;
    // Match RemOpV = X / C0
    if (MatchDiv(RemOpV, DivOpV, DivOpC, IsSigned) && X == DivOpV &&
        C0 == DivOpC && !MulWillOverflow(C0, C1, IsSigned)) {
      Value *NewDivisor = ConstantInt::get(X->getType(), C0 * C1);
      return IsSigned ? Builder.CreateSRem(X, NewDivisor, "srem")
                      : Builder.CreateURem(X, NewDivisor, "urem");
    }
  }
}

return nullptr;
1101}

1103/// Fold
1104///   (1 << NBits) - 1
1105/// Into:
1106///   ~(-(1 << NBits))
1107/// Because a 'not' is better for bit-tracking analysis and other transforms
1108/// than an 'add'. The new shl is always nsw, and is nuw if old `and` was.
1109static Instruction *canonicalizeLowbitMask(BinaryOperator &I,
                                         InstCombiner::BuilderTy &Builder) {
Value *NBits;
if (!match(&I, m_Add(m_OneUse(m_Shl(m_One(), m_Value(NBits))), m_AllOnes())))
  return nullptr;

Constant *MinusOne = Constant::getAllOnesValue(NBits->getType());
Value *NotMask = Builder.CreateShl(MinusOne, NBits, "notmask");
// Be wary of constant folding.
if (auto *BOp = dyn_cast<BinaryOperator>(NotMask)) {
  // Always NSW. But NUW propagates from `add`.
  BOp->setHasNoSignedWrap();
  BOp->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
}

return BinaryOperator::CreateNot(NotMask, I.getName());
1125}

1127static Instruction *foldToUnsignedSaturatedAdd(BinaryOperator &I) {
assert(I.getOpcode() == Instruction::Add && "Expecting add instruction")(static_cast <bool> (I.getOpcode() == Instruction::Add &&
 "Expecting add instruction") ? void (0) : __assert_fail ("I.getOpcode() == Instruction::Add && \"Expecting add instruction\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 1128, __extension__ __PRETTY_FUNCTION__));
Type *Ty = I.getType();
auto getUAddSat = [&]() {
  return Intrinsic::getDeclaration(I.getModule(), Intrinsic::uadd_sat, Ty);
};

// add (umin X, ~Y), Y --> uaddsat X, Y
Value *X, *Y;
if (match(&I, m_c_Add(m_c_UMin(m_Value(X), m_Not(m_Value(Y))),
                      m_Deferred(Y))))
  return CallInst::Create(getUAddSat(), { X, Y });

// add (umin X, ~C), C --> uaddsat X, C
const APInt *C, *NotC;
if (match(&I, m_Add(m_UMin(m_Value(X), m_APInt(NotC)), m_APInt(C))) &&
    *C == ~*NotC)
  return CallInst::Create(getUAddSat(), { X, ConstantInt::get(Ty, *C) });

return nullptr;
1147}

1149Instruction *InstCombinerImpl::
  canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(
      BinaryOperator &I) {
assert((I.getOpcode() == Instruction::Add ||(static_cast <bool> ((I.getOpcode() == Instruction::Add
 || I.getOpcode() == Instruction::Or || I.getOpcode() == Instruction
::Sub) && "Expecting add/or/sub instruction") ? void (
0) : __assert_fail ("(I.getOpcode() == Instruction::Add || I.getOpcode() == Instruction::Or || I.getOpcode() == Instruction::Sub) && \"Expecting add/or/sub instruction\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 1155, __extension__ __PRETTY_FUNCTION__))
1
Assuming the condition is false→
2
←
Assuming the condition is false→
3
←
Assuming the condition is true→
4
←
'?' condition is true→
        I.getOpcode() == Instruction::Or ||(static_cast <bool> ((I.getOpcode() == Instruction::Add
 || I.getOpcode() == Instruction::Or || I.getOpcode() == Instruction
::Sub) && "Expecting add/or/sub instruction") ? void (
0) : __assert_fail ("(I.getOpcode() == Instruction::Add || I.getOpcode() == Instruction::Or || I.getOpcode() == Instruction::Sub) && \"Expecting add/or/sub instruction\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 1155, __extension__ __PRETTY_FUNCTION__))
        I.getOpcode() == Instruction::Sub) &&(static_cast <bool> ((I.getOpcode() == Instruction::Add
 || I.getOpcode() == Instruction::Or || I.getOpcode() == Instruction
::Sub) && "Expecting add/or/sub instruction") ? void (
0) : __assert_fail ("(I.getOpcode() == Instruction::Add || I.getOpcode() == Instruction::Or || I.getOpcode() == Instruction::Sub) && \"Expecting add/or/sub instruction\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 1155, __extension__ __PRETTY_FUNCTION__))
       "Expecting add/or/sub instruction")(static_cast <bool> ((I.getOpcode() == Instruction::Add
 || I.getOpcode() == Instruction::Or || I.getOpcode() == Instruction
::Sub) && "Expecting add/or/sub instruction") ? void (
0) : __assert_fail ("(I.getOpcode() == Instruction::Add || I.getOpcode() == Instruction::Or || I.getOpcode() == Instruction::Sub) && \"Expecting add/or/sub instruction\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 1155, __extension__ __PRETTY_FUNCTION__));

// We have a subtraction/addition between a (potentially truncated) *logical*
// right-shift of X and a "select".
Value *X, *Select;
Instruction *LowBitsToSkip, *Extract;
if (!match(&I, m_c_BinOp(m_TruncOrSelf(m_CombineAnd(
10
←
Calling 'm_c_BinOp<llvm::PatternMatch::match_combine_or<llvm::PatternMatch::CastClass_match<llvm::PatternMatch::match_combine_and<llvm::PatternMatch::BinaryOp_match<llvm::PatternMatch::bind_ty<llvm::Value>, llvm::PatternMatch::bind_ty<llvm::Instruction>, 26, false>, llvm::PatternMatch::bind_ty<llvm::Instruction>>, 38>, llvm::PatternMatch::match_combine_and<llvm::PatternMatch::BinaryOp_match<llvm::PatternMatch::bind_ty<llvm::Value>, llvm::PatternMatch::bind_ty<llvm::Instruction>, 26, false>, llvm::PatternMatch::bind_ty<llvm::Instruction>>>, llvm::PatternMatch::bind_ty<llvm::Value>>'→
12
←
Returning from 'm_c_BinOp<llvm::PatternMatch::match_combine_or<llvm::PatternMatch::CastClass_match<llvm::PatternMatch::match_combine_and<llvm::PatternMatch::BinaryOp_match<llvm::PatternMatch::bind_ty<llvm::Value>, llvm::PatternMatch::bind_ty<llvm::Instruction>, 26, false>, llvm::PatternMatch::bind_ty<llvm::Instruction>>, 38>, llvm::PatternMatch::match_combine_and<llvm::PatternMatch::BinaryOp_match<llvm::PatternMatch::bind_ty<llvm::Value>, llvm::PatternMatch::bind_ty<llvm::Instruction>, 26, false>, llvm::PatternMatch::bind_ty<llvm::Instruction>>>, llvm::PatternMatch::bind_ty<llvm::Value>>'→
13
←
Calling 'match<llvm::BinaryOperator, llvm::PatternMatch::AnyBinaryOp_match<llvm::PatternMatch::match_combine_or<llvm::PatternMatch::CastClass_match<llvm::PatternMatch::match_combine_and<llvm::PatternMatch::BinaryOp_match<llvm::PatternMatch::bind_ty<llvm::Value>, llvm::PatternMatch::bind_ty<llvm::Instruction>, 26, false>, llvm::PatternMatch::bind_ty<llvm::Instruction>>, 38>, llvm::PatternMatch::match_combine_and<llvm::PatternMatch::BinaryOp_match<llvm::PatternMatch::bind_ty<llvm::Value>, llvm::PatternMatch::bind_ty<llvm::Instruction>, 26, false>, llvm::PatternMatch::bind_ty<llvm::Instruction>>>, llvm::PatternMatch::bind_ty<llvm::Value>, true>>'→
35
←
Returning from 'match<llvm::BinaryOperator, llvm::PatternMatch::AnyBinaryOp_match<llvm::PatternMatch::match_combine_or<llvm::PatternMatch::CastClass_match<llvm::PatternMatch::match_combine_and<llvm::PatternMatch::BinaryOp_match<llvm::PatternMatch::bind_ty<llvm::Value>, llvm::PatternMatch::bind_ty<llvm::Instruction>, 26, false>, llvm::PatternMatch::bind_ty<llvm::Instruction>>, 38>, llvm::PatternMatch::match_combine_and<llvm::PatternMatch::BinaryOp_match<llvm::PatternMatch::bind_ty<llvm::Value>, llvm::PatternMatch::bind_ty<llvm::Instruction>, 26, false>, llvm::PatternMatch::bind_ty<llvm::Instruction>>>, llvm::PatternMatch::bind_ty<llvm::Value>, true>>'→
                             m_LShr(m_Value(X), m_Instruction(LowBitsToSkip)),
                             m_Instruction(Extract))),
                         m_Value(Select))))
5
←
Calling 'm_Value'→
9
←
Returning from 'm_Value'→
  return nullptr;

// `add`/`or` is commutative; but for `sub`, "select" *must* be on RHS.
if (I.getOpcode() == Instruction::Sub && I.getOperand(1) != Select)
36
←
Assuming pointer value is null→
37
←
Taking false branch→
  return nullptr;

Type *XTy = X->getType();
bool HadTrunc = I.getType() != XTy;
38
←
Assuming the condition is false→

// If there was a truncation of extracted value, then we'll need to produce
// one extra instruction, so we need to ensure one instruction will go away.
if (HadTrunc38.1
'HadTrunc' is false
1
'HadTrunc' is false
 && !match(&I, m_c_BinOp(m_OneUse(m_Value()), m_Value())))
39
←
Taking false branch→
  return nullptr;

// Extraction should extract high NBits bits, with shift amount calculated as:
//   low bits to skip = shift bitwidth - high bits to extract
// The shift amount itself may be extended, and we need to look past zero-ext
// when matching NBits, that will matter for matching later.
Constant *C;
Value *NBits;
if (!match(
40
←
Taking false branch→
        LowBitsToSkip,
        m_ZExtOrSelf(m_Sub(m_Constant(C), m_ZExtOrSelf(m_Value(NBits))))) ||
    !match(C, m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_EQ,
                                 APInt(C->getType()->getScalarSizeInBits(),
                                       X->getType()->getScalarSizeInBits()))))
  return nullptr;

// Sign-extending value can be zero-extended if we `sub`tract it,
// or sign-extended otherwise.
auto SkipExtInMagic = [&I](Value *&V) {
  if (I.getOpcode() == Instruction::Sub)
42
←
Taking true branch→
    match(V, m_ZExtOrSelf(m_Value(V)));
43
←
Calling 'm_Value'→
45
←
Returning from 'm_Value'→
46
←
Calling 'm_ZExtOrSelf<llvm::PatternMatch::bind_ty<llvm::Value>>'→
54
←
Returning from 'm_ZExtOrSelf<llvm::PatternMatch::bind_ty<llvm::Value>>'→
55
←
Calling 'match<llvm::Value, llvm::PatternMatch::match_combine_or<llvm::PatternMatch::CastClass_match<llvm::PatternMatch::bind_ty<llvm::Value>, 39>, llvm::PatternMatch::bind_ty<llvm::Value>>>'→
57
←
Returning from 'match<llvm::Value, llvm::PatternMatch::match_combine_or<llvm::PatternMatch::CastClass_match<llvm::PatternMatch::bind_ty<llvm::Value>, 39>, llvm::PatternMatch::bind_ty<llvm::Value>>>'→
  else
    match(V, m_SExtOrSelf(m_Value(V)));
};
58
←
Returning without writing to 'V'→

// Now, finally validate the sign-extending magic.
// `select` itself may be appropriately extended, look past that.
SkipExtInMagic(Select);
41
←
Calling 'operator()'→
59
←
Returning from 'operator()'→

ICmpInst::Predicate Pred;
const APInt *Thr;
Value *SignExtendingValue, *Zero;
bool ShouldSignext;
// It must be a select between two values we will later establish to be a
// sign-extending value and a zero constant. The condition guarding the
// sign-extension must be based on a sign bit of the same X we had in `lshr`.
if (!match(Select, m_Select(m_ICmp(Pred, m_Specific(X), m_APInt(Thr)),
60
←
Passing null pointer value via 1st parameter 'V'→
61
←
Calling 'match<llvm::Value, llvm::PatternMatch::ThreeOps_match<llvm::PatternMatch::CmpClass_match<llvm::PatternMatch::specificval_ty, llvm::PatternMatch::apint_match, llvm::ICmpInst, llvm::CmpInst::Predicate, false>, llvm::PatternMatch::bind_ty<llvm::Value>, llvm::PatternMatch::bind_ty<llvm::Value>, 57>>'→
                            m_Value(SignExtendingValue), m_Value(Zero))) ||
    !isSignBitCheck(Pred, *Thr, ShouldSignext))
  return nullptr;

// icmp-select pair is commutative.
if (!ShouldSignext)
  std::swap(SignExtendingValue, Zero);

// If we should not perform sign-extension then we must add/or/subtract zero.
if (!match(Zero, m_Zero()))
  return nullptr;
// Otherwise, it should be some constant, left-shifted by the same NBits we
// had in `lshr`. Said left-shift can also be appropriately extended.
// Again, we must look past zero-ext when looking for NBits.
SkipExtInMagic(SignExtendingValue);
Constant *SignExtendingValueBaseConstant;
if (!match(SignExtendingValue,
           m_Shl(m_Constant(SignExtendingValueBaseConstant),
                 m_ZExtOrSelf(m_Specific(NBits)))))
  return nullptr;
// If we `sub`, then the constant should be one, else it should be all-ones.
if (I.getOpcode() == Instruction::Sub
        ? !match(SignExtendingValueBaseConstant, m_One())
        : !match(SignExtendingValueBaseConstant, m_AllOnes()))
  return nullptr;

auto *NewAShr = BinaryOperator::CreateAShr(X, LowBitsToSkip,
                                           Extract->getName() + ".sext");
NewAShr->copyIRFlags(Extract); // Preserve `exact`-ness.
if (!HadTrunc)
  return NewAShr;

Builder.Insert(NewAShr);
return TruncInst::CreateTruncOrBitCast(NewAShr, I.getType());
1248}

1250/// This is a specialization of a more general transform from
1251/// SimplifyUsingDistributiveLaws. If that code can be made to work optimally
1252/// for multi-use cases or propagating nsw/nuw, then we would not need this.
1253static Instruction *factorizeMathWithShlOps(BinaryOperator &I,
                                          InstCombiner::BuilderTy &Builder) {
// TODO: Also handle mul by doubling the shift amount?
assert((I.getOpcode() == Instruction::Add ||(static_cast <bool> ((I.getOpcode() == Instruction::Add
 || I.getOpcode() == Instruction::Sub) && "Expected add/sub"
) ? void (0) : __assert_fail ("(I.getOpcode() == Instruction::Add || I.getOpcode() == Instruction::Sub) && \"Expected add/sub\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 1258, __extension__ __PRETTY_FUNCTION__))
        I.getOpcode() == Instruction::Sub) &&(static_cast <bool> ((I.getOpcode() == Instruction::Add
 || I.getOpcode() == Instruction::Sub) && "Expected add/sub"
) ? void (0) : __assert_fail ("(I.getOpcode() == Instruction::Add || I.getOpcode() == Instruction::Sub) && \"Expected add/sub\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 1258, __extension__ __PRETTY_FUNCTION__))
       "Expected add/sub")(static_cast <bool> ((I.getOpcode() == Instruction::Add
 || I.getOpcode() == Instruction::Sub) && "Expected add/sub"
) ? void (0) : __assert_fail ("(I.getOpcode() == Instruction::Add || I.getOpcode() == Instruction::Sub) && \"Expected add/sub\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 1258, __extension__ __PRETTY_FUNCTION__));
auto *Op0 = dyn_cast<BinaryOperator>(I.getOperand(0));
auto *Op1 = dyn_cast<BinaryOperator>(I.getOperand(1));
if (!Op0 || !Op1 || !(Op0->hasOneUse() || Op1->hasOneUse()))
  return nullptr;

Value *X, *Y, *ShAmt;
if (!match(Op0, m_Shl(m_Value(X), m_Value(ShAmt))) ||
    !match(Op1, m_Shl(m_Value(Y), m_Specific(ShAmt))))
  return nullptr;

// No-wrap propagates only when all ops have no-wrap.
bool HasNSW = I.hasNoSignedWrap() && Op0->hasNoSignedWrap() &&
              Op1->hasNoSignedWrap();
bool HasNUW = I.hasNoUnsignedWrap() && Op0->hasNoUnsignedWrap() &&
              Op1->hasNoUnsignedWrap();

// add/sub (X << ShAmt), (Y << ShAmt) --> (add/sub X, Y) << ShAmt
Value *NewMath = Builder.CreateBinOp(I.getOpcode(), X, Y);
if (auto *NewI = dyn_cast<BinaryOperator>(NewMath)) {
  NewI->setHasNoSignedWrap(HasNSW);
  NewI->setHasNoUnsignedWrap(HasNUW);
}
auto *NewShl = BinaryOperator::CreateShl(NewMath, ShAmt);
NewShl->setHasNoSignedWrap(HasNSW);
NewShl->setHasNoUnsignedWrap(HasNUW);
return NewShl;
1285}

1287Instruction *InstCombinerImpl::visitAdd(BinaryOperator &I) {
if (Value *V = SimplifyAddInst(I.getOperand(0), I.getOperand(1),
                               I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
                               SQ.getWithInstruction(&I)))
  return replaceInstUsesWith(I, V);

if (SimplifyAssociativeOrCommutative(I))
  return &I;

if (Instruction *X = foldVectorBinop(I))
  return X;

// (A*B)+(A*C) -> A*(B+C) etc
if (Value *V = SimplifyUsingDistributiveLaws(I))
  return replaceInstUsesWith(I, V);

if (Instruction *R = factorizeMathWithShlOps(I, Builder))
  return R;

if (Instruction *X = foldAddWithConstant(I))
  return X;

if (Instruction *X = foldNoWrapAdd(I, Builder))
  return X;

Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
Type *Ty = I.getType();
if (Ty->isIntOrIntVectorTy(1))
  return BinaryOperator::CreateXor(LHS, RHS);

// X + X --> X << 1
if (LHS == RHS) {
  auto *Shl = BinaryOperator::CreateShl(LHS, ConstantInt::get(Ty, 1));
  Shl->setHasNoSignedWrap(I.hasNoSignedWrap());
  Shl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
  return Shl;
}

Value *A, *B;
if (match(LHS, m_Neg(m_Value(A)))) {
  // -A + -B --> -(A + B)
  if (match(RHS, m_Neg(m_Value(B))))
    return BinaryOperator::CreateNeg(Builder.CreateAdd(A, B));

  // -A + B --> B - A
  return BinaryOperator::CreateSub(RHS, A);
}

// A + -B  -->  A - B
if (match(RHS, m_Neg(m_Value(B))))
  return BinaryOperator::CreateSub(LHS, B);

if (Value *V = checkForNegativeOperand(I, Builder))
  return replaceInstUsesWith(I, V);

// (A + 1) + ~B --> A - B
// ~B + (A + 1) --> A - B
// (~B + A) + 1 --> A - B
// (A + ~B) + 1 --> A - B
if (match(&I, m_c_BinOp(m_Add(m_Value(A), m_One()), m_Not(m_Value(B)))) ||
    match(&I, m_BinOp(m_c_Add(m_Not(m_Value(B)), m_Value(A)), m_One())))
  return BinaryOperator::CreateSub(A, B);

// (A + RHS) + RHS --> A + (RHS << 1)
if (match(LHS, m_OneUse(m_c_Add(m_Value(A), m_Specific(RHS)))))
  return BinaryOperator::CreateAdd(A, Builder.CreateShl(RHS, 1, "reass.add"));

// LHS + (A + LHS) --> A + (LHS << 1)
if (match(RHS, m_OneUse(m_c_Add(m_Value(A), m_Specific(LHS)))))
  return BinaryOperator::CreateAdd(A, Builder.CreateShl(LHS, 1, "reass.add"));

{
  // (A + C1) + (C2 - B) --> (A - B) + (C1 + C2)
  Constant *C1, *C2;
  if (match(&I, m_c_Add(m_Add(m_Value(A), m_ImmConstant(C1)),
                        m_Sub(m_ImmConstant(C2), m_Value(B)))) &&
      (LHS->hasOneUse() || RHS->hasOneUse())) {
    Value *Sub = Builder.CreateSub(A, B);
    return BinaryOperator::CreateAdd(Sub, ConstantExpr::getAdd(C1, C2));
  }
}

// X % C0 + (( X / C0 ) % C1) * C0 => X % (C0 * C1)
if (Value *V = SimplifyAddWithRemainder(I)) return replaceInstUsesWith(I, V);

// ((X s/ C1) << C2) + X => X s% -C1 where -C1 is 1 << C2
const APInt *C1, *C2;
if (match(LHS, m_Shl(m_SDiv(m_Specific(RHS), m_APInt(C1)), m_APInt(C2)))) {
  APInt one(C2->getBitWidth(), 1);
  APInt minusC1 = -(*C1);
  if (minusC1 == (one << *C2)) {
    Constant *NewRHS = ConstantInt::get(RHS->getType(), minusC1);
    return BinaryOperator::CreateSRem(RHS, NewRHS);
  }
}

// A+B --> A|B iff A and B have no bits set in common.
if (haveNoCommonBitsSet(LHS, RHS, DL, &AC, &I, &DT))
  return BinaryOperator::CreateOr(LHS, RHS);

// add (select X 0 (sub n A)) A  -->  select X A n
{
  SelectInst *SI = dyn_cast<SelectInst>(LHS);
  Value *A = RHS;
  if (!SI) {
    SI = dyn_cast<SelectInst>(RHS);
    A = LHS;
  }
  if (SI && SI->hasOneUse()) {
    Value *TV = SI->getTrueValue();
    Value *FV = SI->getFalseValue();
    Value *N;

    // Can we fold the add into the argument of the select?
    // We check both true and false select arguments for a matching subtract.
    if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Specific(A))))
      // Fold the add into the true select value.
      return SelectInst::Create(SI->getCondition(), N, A);

    if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Specific(A))))
      // Fold the add into the false select value.
      return SelectInst::Create(SI->getCondition(), A, N);
  }
}

if (Instruction *Ext = narrowMathIfNoOverflow(I))
  return Ext;

// (add (xor A, B) (and A, B)) --> (or A, B)
// (add (and A, B) (xor A, B)) --> (or A, B)
if (match(&I, m_c_BinOp(m_Xor(m_Value(A), m_Value(B)),
                        m_c_And(m_Deferred(A), m_Deferred(B)))))
  return BinaryOperator::CreateOr(A, B);

// (add (or A, B) (and A, B)) --> (add A, B)
// (add (and A, B) (or A, B)) --> (add A, B)
if (match(&I, m_c_BinOp(m_Or(m_Value(A), m_Value(B)),
                        m_c_And(m_Deferred(A), m_Deferred(B))))) {
  // Replacing operands in-place to preserve nuw/nsw flags.
  replaceOperand(I, 0, A);
  replaceOperand(I, 1, B);
  return &I;
}

// TODO(jingyue): Consider willNotOverflowSignedAdd and
// willNotOverflowUnsignedAdd to reduce the number of invocations of
// computeKnownBits.
bool Changed = false;
if (!I.hasNoSignedWrap() && willNotOverflowSignedAdd(LHS, RHS, I)) {
  Changed = true;
  I.setHasNoSignedWrap(true);
}
if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedAdd(LHS, RHS, I)) {
  Changed = true;
  I.setHasNoUnsignedWrap(true);
}

if (Instruction *V = canonicalizeLowbitMask(I, Builder))
  return V;

if (Instruction *V =
        canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(I))
  return V;

if (Instruction *SatAdd = foldToUnsignedSaturatedAdd(I))
  return SatAdd;

// usub.sat(A, B) + B => umax(A, B)
if (match(&I, m_c_BinOp(
        m_OneUse(m_Intrinsic<Intrinsic::usub_sat>(m_Value(A), m_Value(B))),
        m_Deferred(B)))) {
  return replaceInstUsesWith(I,
      Builder.CreateIntrinsic(Intrinsic::umax, {I.getType()}, {A, B}));
}

// ctpop(A) + ctpop(B) => ctpop(A | B) if A and B have no bits set in common.
if (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(A)))) &&
    match(RHS, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(B)))) &&
    haveNoCommonBitsSet(A, B, DL, &AC, &I, &DT))
  return replaceInstUsesWith(
      I, Builder.CreateIntrinsic(Intrinsic::ctpop, {I.getType()},
                                 {Builder.CreateOr(A, B)}));

return Changed ? &I : nullptr;
1471}

1473/// Eliminate an op from a linear interpolation (lerp) pattern.
1474static Instruction *factorizeLerp(BinaryOperator &I,
                                InstCombiner::BuilderTy &Builder) {
Value *X, *Y, *Z;
if (!match(&I, m_c_FAdd(m_OneUse(m_c_FMul(m_Value(Y),
                                          m_OneUse(m_FSub(m_FPOne(),
                                                          m_Value(Z))))),
                        m_OneUse(m_c_FMul(m_Value(X), m_Deferred(Z))))))
  return nullptr;

// (Y * (1.0 - Z)) + (X * Z) --> Y + Z * (X - Y) [8 commuted variants]
Value *XY = Builder.CreateFSubFMF(X, Y, &I);
Value *MulZ = Builder.CreateFMulFMF(Z, XY, &I);
return BinaryOperator::CreateFAddFMF(Y, MulZ, &I);
1487}

1489/// Factor a common operand out of fadd/fsub of fmul/fdiv.
1490static Instruction *factorizeFAddFSub(BinaryOperator &I,
                                    InstCombiner::BuilderTy &Builder) {
assert((I.getOpcode() == Instruction::FAdd ||(static_cast <bool> ((I.getOpcode() == Instruction::FAdd
 || I.getOpcode() == Instruction::FSub) && "Expecting fadd/fsub"
) ? void (0) : __assert_fail ("(I.getOpcode() == Instruction::FAdd || I.getOpcode() == Instruction::FSub) && \"Expecting fadd/fsub\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 1493, __extension__ __PRETTY_FUNCTION__))
        I.getOpcode() == Instruction::FSub) && "Expecting fadd/fsub")(static_cast <bool> ((I.getOpcode() == Instruction::FAdd
 || I.getOpcode() == Instruction::FSub) && "Expecting fadd/fsub"
) ? void (0) : __assert_fail ("(I.getOpcode() == Instruction::FAdd || I.getOpcode() == Instruction::FSub) && \"Expecting fadd/fsub\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 1493, __extension__ __PRETTY_FUNCTION__));
assert(I.hasAllowReassoc() && I.hasNoSignedZeros() &&(static_cast <bool> (I.hasAllowReassoc() && I.hasNoSignedZeros
() && "FP factorization requires FMF") ? void (0) : __assert_fail
 ("I.hasAllowReassoc() && I.hasNoSignedZeros() && \"FP factorization requires FMF\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 1495, __extension__ __PRETTY_FUNCTION__))
       "FP factorization requires FMF")(static_cast <bool> (I.hasAllowReassoc() && I.hasNoSignedZeros
() && "FP factorization requires FMF") ? void (0) : __assert_fail
 ("I.hasAllowReassoc() && I.hasNoSignedZeros() && \"FP factorization requires FMF\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 1495, __extension__ __PRETTY_FUNCTION__));

if (Instruction *Lerp = factorizeLerp(I, Builder))
  return Lerp;

Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
Value *X, *Y, *Z;
bool IsFMul;
if ((match(Op0, m_OneUse(m_FMul(m_Value(X), m_Value(Z)))) &&
     match(Op1, m_OneUse(m_c_FMul(m_Value(Y), m_Specific(Z))))) ||
    (match(Op0, m_OneUse(m_FMul(m_Value(Z), m_Value(X)))) &&
     match(Op1, m_OneUse(m_c_FMul(m_Value(Y), m_Specific(Z))))))
  IsFMul = true;
else if (match(Op0, m_OneUse(m_FDiv(m_Value(X), m_Value(Z)))) &&
         match(Op1, m_OneUse(m_FDiv(m_Value(Y), m_Specific(Z)))))
  IsFMul = false;
else
  return nullptr;

// (X * Z) + (Y * Z) --> (X + Y) * Z
// (X * Z) - (Y * Z) --> (X - Y) * Z
// (X / Z) + (Y / Z) --> (X + Y) / Z
// (X / Z) - (Y / Z) --> (X - Y) / Z
bool IsFAdd = I.getOpcode() == Instruction::FAdd;
Value *XY = IsFAdd ? Builder.CreateFAddFMF(X, Y, &I)
                   : Builder.CreateFSubFMF(X, Y, &I);

// Bail out if we just created a denormal constant.
// TODO: This is copied from a previous implementation. Is it necessary?
const APFloat *C;
if (match(XY, m_APFloat(C)) && !C->isNormal())
  return nullptr;

return IsFMul ? BinaryOperator::CreateFMulFMF(XY, Z, &I)
              : BinaryOperator::CreateFDivFMF(XY, Z, &I);
1530}

1532Instruction *InstCombinerImpl::visitFAdd(BinaryOperator &I) {
if (Value *V = SimplifyFAddInst(I.getOperand(0), I.getOperand(1),
                                I.getFastMathFlags(),
                                SQ.getWithInstruction(&I)))
  return replaceInstUsesWith(I, V);

if (SimplifyAssociativeOrCommutative(I))
  return &I;

if (Instruction *X = foldVectorBinop(I))
  return X;

if (Instruction *FoldedFAdd = foldBinOpIntoSelectOrPhi(I))
  return FoldedFAdd;

// (-X) + Y --> Y - X
Value *X, *Y;
if (match(&I, m_c_FAdd(m_FNeg(m_Value(X)), m_Value(Y))))
  return BinaryOperator::CreateFSubFMF(Y, X, &I);

// Similar to above, but look through fmul/fdiv for the negated term.
// (-X * Y) + Z --> Z - (X * Y) [4 commuted variants]
Value *Z;
if (match(&I, m_c_FAdd(m_OneUse(m_c_FMul(m_FNeg(m_Value(X)), m_Value(Y))),
                       m_Value(Z)))) {
  Value *XY = Builder.CreateFMulFMF(X, Y, &I);
  return BinaryOperator::CreateFSubFMF(Z, XY, &I);
}
// (-X / Y) + Z --> Z - (X / Y) [2 commuted variants]
// (X / -Y) + Z --> Z - (X / Y) [2 commuted variants]
if (match(&I, m_c_FAdd(m_OneUse(m_FDiv(m_FNeg(m_Value(X)), m_Value(Y))),
                       m_Value(Z))) ||
    match(&I, m_c_FAdd(m_OneUse(m_FDiv(m_Value(X), m_FNeg(m_Value(Y)))),
                       m_Value(Z)))) {
  Value *XY = Builder.CreateFDivFMF(X, Y, &I);
  return BinaryOperator::CreateFSubFMF(Z, XY, &I);
}

// Check for (fadd double (sitofp x), y), see if we can merge this into an
// integer add followed by a promotion.
Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
  Value *LHSIntVal = LHSConv->getOperand(0);
  Type *FPType = LHSConv->getType();

  // TODO: This check is overly conservative. In many cases known bits
  // analysis can tell us that the result of the addition has less significant
  // bits than the integer type can hold.
  auto IsValidPromotion = [](Type *FTy, Type *ITy) {
    Type *FScalarTy = FTy->getScalarType();
    Type *IScalarTy = ITy->getScalarType();

    // Do we have enough bits in the significand to represent the result of
    // the integer addition?
    unsigned MaxRepresentableBits =
        APFloat::semanticsPrecision(FScalarTy->getFltSemantics());
    return IScalarTy->getIntegerBitWidth() <= MaxRepresentableBits;
  };

  // (fadd double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
  // ... if the constant fits in the integer value.  This is useful for things
  // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
  // requires a constant pool load, and generally allows the add to be better
  // instcombined.
  if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
    if (IsValidPromotion(FPType, LHSIntVal->getType())) {
      Constant *CI =
        ConstantExpr::getFPToSI(CFP, LHSIntVal->getType());
      if (LHSConv->hasOneUse() &&
          ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
          willNotOverflowSignedAdd(LHSIntVal, CI, I)) {
        // Insert the new integer add.
        Value *NewAdd = Builder.CreateNSWAdd(LHSIntVal, CI, "addconv");
        return new SIToFPInst(NewAdd, I.getType());
      }
    }

  // (fadd double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
  if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
    Value *RHSIntVal = RHSConv->getOperand(0);
    // It's enough to check LHS types only because we require int types to
    // be the same for this transform.
    if (IsValidPromotion(FPType, LHSIntVal->getType())) {
      // Only do this if x/y have the same type, if at least one of them has a
      // single use (so we don't increase the number of int->fp conversions),
      // and if the integer add will not overflow.
      if (LHSIntVal->getType() == RHSIntVal->getType() &&
          (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
          willNotOverflowSignedAdd(LHSIntVal, RHSIntVal, I)) {
        // Insert the new integer add.
        Value *NewAdd = Builder.CreateNSWAdd(LHSIntVal, RHSIntVal, "addconv");
        return new SIToFPInst(NewAdd, I.getType());
      }
    }
  }
}

// Handle specials cases for FAdd with selects feeding the operation
if (Value *V = SimplifySelectsFeedingBinaryOp(I, LHS, RHS))
  return replaceInstUsesWith(I, V);

if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
  if (Instruction *F = factorizeFAddFSub(I, Builder))
    return F;

  // Try to fold fadd into start value of reduction intrinsic.
  if (match(&I, m_c_FAdd(m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(
                             m_AnyZeroFP(), m_Value(X))),
                         m_Value(Y)))) {
    // fadd (rdx 0.0, X), Y --> rdx Y, X
    return replaceInstUsesWith(
        I, Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
                                   {X->getType()}, {Y, X}, &I));
  }
  const APFloat *StartC, *C;
  if (match(LHS, m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(
                     m_APFloat(StartC), m_Value(X)))) &&
      match(RHS, m_APFloat(C))) {
    // fadd (rdx StartC, X), C --> rdx (C + StartC), X
    Constant *NewStartC = ConstantFP::get(I.getType(), *C + *StartC);
    return replaceInstUsesWith(
        I, Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
                                   {X->getType()}, {NewStartC, X}, &I));
  }

  if (Value *V = FAddCombine(Builder).simplify(&I))
    return replaceInstUsesWith(I, V);
}

return nullptr;
1662}

1664/// Optimize pointer differences into the same array into a size.  Consider:
1665///  &A[10] - &A[0]: we should compile this to "10".  LHS/RHS are the pointer
1666/// operands to the ptrtoint instructions for the LHS/RHS of the subtract.
1667Value *InstCombinerImpl::OptimizePointerDifference(Value *LHS, Value *RHS,
                                                 Type *Ty, bool IsNUW) {
// If LHS is a gep based on RHS or RHS is a gep based on LHS, we can optimize
// this.
bool Swapped = false;
GEPOperator *GEP1 = nullptr, *GEP2 = nullptr;
if (!isa<GEPOperator>(LHS) && isa<GEPOperator>(RHS)) {
  std::swap(LHS, RHS);
  Swapped = true;
}

// Require at least one GEP with a common base pointer on both sides.
if (auto *LHSGEP = dyn_cast<GEPOperator>(LHS)) {
  // (gep X, ...) - X
  if (LHSGEP->getOperand(0) == RHS) {
    GEP1 = LHSGEP;
  } else if (auto *RHSGEP = dyn_cast<GEPOperator>(RHS)) {
    // (gep X, ...) - (gep X, ...)
    if (LHSGEP->getOperand(0)->stripPointerCasts() ==
        RHSGEP->getOperand(0)->stripPointerCasts()) {
      GEP1 = LHSGEP;
      GEP2 = RHSGEP;
    }
  }
}

if (!GEP1)
  return nullptr;

if (GEP2) {
  // (gep X, ...) - (gep X, ...)
  //
  // Avoid duplicating the arithmetic if there are more than one non-constant
  // indices between the two GEPs and either GEP has a non-constant index and
  // multiple users. If zero non-constant index, the result is a constant and
  // there is no duplication. If one non-constant index, the result is an add
  // or sub with a constant, which is no larger than the original code, and
  // there's no duplicated arithmetic, even if either GEP has multiple
  // users. If more than one non-constant indices combined, as long as the GEP
  // with at least one non-constant index doesn't have multiple users, there
  // is no duplication.
  unsigned NumNonConstantIndices1 = GEP1->countNonConstantIndices();
  unsigned NumNonConstantIndices2 = GEP2->countNonConstantIndices();
  if (NumNonConstantIndices1 + NumNonConstantIndices2 > 1 &&
      ((NumNonConstantIndices1 > 0 && !GEP1->hasOneUse()) ||
       (NumNonConstantIndices2 > 0 && !GEP2->hasOneUse()))) {
    return nullptr;
  }
}

// Emit the offset of the GEP and an intptr_t.
Value *Result = EmitGEPOffset(GEP1);

// If this is a single inbounds GEP and the original sub was nuw,
// then the final multiplication is also nuw.
if (auto *I = dyn_cast<Instruction>(Result))
  if (IsNUW && !GEP2 && !Swapped && GEP1->isInBounds() &&
      I->getOpcode() == Instruction::Mul)
    I->setHasNoUnsignedWrap();

// If we have a 2nd GEP of the same base pointer, subtract the offsets.
// If both GEPs are inbounds, then the subtract does not have signed overflow.
if (GEP2) {
  Value *Offset = EmitGEPOffset(GEP2);
  Result = Builder.CreateSub(Result, Offset, "gepdiff", /* NUW */ false,
                             GEP1->isInBounds() && GEP2->isInBounds());
}

// If we have p - gep(p, ...)  then we have to negate the result.
if (Swapped)
  Result = Builder.CreateNeg(Result, "diff.neg");

return Builder.CreateIntCast(Result, Ty, true);
1740}

1742Instruction *InstCombinerImpl::visitSub(BinaryOperator &I) {
if (Value *V = SimplifySubInst(I.getOperand(0), I.getOperand(1),
                               I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
                               SQ.getWithInstruction(&I)))
  return replaceInstUsesWith(I, V);

if (Instruction *X = foldVectorBinop(I))
  return X;

Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);

// If this is a 'B = x-(-A)', change to B = x+A.
// We deal with this without involving Negator to preserve NSW flag.
if (Value *V = dyn_castNegVal(Op1)) {
  BinaryOperator *Res = BinaryOperator::CreateAdd(Op0, V);

  if (const auto *BO = dyn_cast<BinaryOperator>(Op1)) {
    assert(BO->getOpcode() == Instruction::Sub &&(static_cast <bool> (BO->getOpcode() == Instruction::
Sub && "Expected a subtraction operator!") ? void (0)
 : __assert_fail ("BO->getOpcode() == Instruction::Sub && \"Expected a subtraction operator!\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 1760, __extension__ __PRETTY_FUNCTION__))
           "Expected a subtraction operator!")(static_cast <bool> (BO->getOpcode() == Instruction::
Sub && "Expected a subtraction operator!") ? void (0)
 : __assert_fail ("BO->getOpcode() == Instruction::Sub && \"Expected a subtraction operator!\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp"
, 1760, __extension__ __PRETTY_FUNCTION__));
    if (BO->hasNoSignedWrap() && I.hasNoSignedWrap())
      Res->setHasNoSignedWrap(true);
  } else {
    if (cast<Constant>(Op1)->isNotMinSignedValue() && I.hasNoSignedWrap())
      Res->setHasNoSignedWrap(true);
  }

  return Res;
}

// Try this before Negator to preserve NSW flag.
if (Instruction *R = factorizeMathWithShlOps(I, Builder))
  return R;

Constant *C;
if (match(Op0, m_ImmConstant(C))) {
  Value *X;
  Constant *C2;

  // C-(X+C2) --> (C-C2)-X
  if (match(Op1, m_Add(m_Value(X), m_ImmConstant(C2))))
    return BinaryOperator::CreateSub(ConstantExpr::getSub(C, C2), X);
}

auto TryToNarrowDeduceFlags = [this, &I, &Op0, &Op1]() -> Instruction * {
  if (Instruction *Ext = narrowMathIfNoOverflow(I))
    return Ext;

  bool Changed = false;
  if (!I.hasNoSignedWrap() && willNotOverflowSignedSub(Op0, Op1, I)) {
    Changed = true;
    I.setHasNoSignedWrap(true);
  }
  if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedSub(Op0, Op1, I)) {
    Changed = true;
    I.setHasNoUnsignedWrap(true);
  }

  return Changed ? &I : nullptr;
};

// First, let's try to interpret `sub a, b` as `add a, (sub 0, b)`,
// and let's try to sink `(sub 0, b)` into `b` itself. But only if this isn't
// a pure negation used by a select that looks like abs/nabs.
bool IsNegation = match(Op0, m_ZeroInt());
if (!IsNegation || none_of(I.users(), [&I, Op1](const User *U) {
      const Instruction *UI = dyn_cast<Instruction>(U);
      if (!UI)
        return false;
      return match(UI,
                   m_Select(m_Value(), m_Specific(Op1), m_Specific(&I))) ||
             match(UI, m_Select(m_Value(), m_Specific(&I), m_Specific(Op1)));
    })) {
  if (Value *NegOp1 = Negator::Negate(IsNegation, Op1, *this))
    return BinaryOperator::CreateAdd(NegOp1, Op0);
}
if (IsNegation)
  return TryToNarrowDeduceFlags(); // Should have been handled in Negator!

// (A*B)-(A*C) -> A*(B-C) etc
if (Value *V = SimplifyUsingDistributiveLaws(I))
  return replaceInstUsesWith(I, V);

if (I.getType()->isIntOrIntVectorTy(1))
  return BinaryOperator::CreateXor(Op0, Op1);

// Replace (-1 - A) with (~A).
if (match(Op0, m_AllOnes()))
  return BinaryOperator::CreateNot(Op1);

// (X + -1) - Y --> ~Y + X
Value *X, *Y;
if (match(Op0, m_OneUse(m_Add(m_Value(X), m_AllOnes()))))
  return BinaryOperator::CreateAdd(Builder.CreateNot(Op1), X);

// Reassociate sub/add sequences to create more add instructions and
// reduce dependency chains:
// ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
Value *Z;
if (match(Op0, m_OneUse(m_c_Add(m_OneUse(m_Sub(m_Value(X), m_Value(Y))),
                                m_Value(Z))))) {
  Value *XZ = Builder.CreateAdd(X, Z);
  Value *YW = Builder.CreateAdd(Y, Op1);
  return BinaryOperator::CreateSub(XZ, YW);
}

// ((X - Y) - Op1)  -->  X - (Y + Op1)
if (match(Op0, m_OneUse(m_Sub(m_Value(X), m_Value(Y))))) {
  Value *Add = Builder.CreateAdd(Y, Op1);
  return BinaryOperator::CreateSub(X, Add);
}

// (~X) - (~Y) --> Y - X
// This is placed after the other reassociations and explicitly excludes a
// sub-of-sub pattern to avoid infinite looping.
if (isFreeToInvert(Op0, Op0->hasOneUse()) &&
    isFreeToInvert(Op1, Op1->hasOneUse()) &&
    !match(Op0, m_Sub(m_ImmConstant(), m_Value()))) {
  Value *NotOp0 = Builder.CreateNot(Op0);
  Value *NotOp1 = Builder.CreateNot(Op1);
  return BinaryOperator::CreateSub(NotOp1, NotOp0);
}

auto m_AddRdx = [](Value *&Vec) {
  return m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_add>(m_Value(Vec)));
};
Value *V0, *V1;
if (match(Op0, m_AddRdx(V0)) && match(Op1, m_AddRdx(V1)) &&
    V0->getType() == V1->getType()) {
  // Difference of sums is sum of differences:
  // add_rdx(V0) - add_rdx(V1) --> add_rdx(V0 - V1)
  Value *Sub = Builder.CreateSub(V0, V1);
  Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_add,
                                       {Sub->getType()}, {Sub});
  return replaceInstUsesWith(I, Rdx);
}

if (Constant *C = dyn_cast<Constant>(Op0)) {
  Value *X;
  if (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
    // C - (zext bool) --> bool ? C - 1 : C
    return SelectInst::Create(X, InstCombiner::SubOne(C), C);
  if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
    // C - (sext bool) --> bool ? C + 1 : C
    return SelectInst::Create(X, InstCombiner::AddOne(C), C);

  // C - ~X == X + (1+C)
  if (match(Op1, m_Not(m_Value(X))))
    return BinaryOperator::CreateAdd(X, InstCombiner::AddOne(C));

  // Try to fold constant sub into select arguments.
  if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
    if (Instruction *R = FoldOpIntoSelect(I, SI))
      return R;

  // Try to fold constant sub into PHI values.
  if (PHINode *PN = dyn_cast<PHINode>(Op1))
    if (Instruction *R = foldOpIntoPhi(I, PN))
      return R;

  Constant *C2;

  // C-(C2-X) --> X+(C-C2)
  if (match(Op1, m_Sub(m_ImmConstant(C2), m_Value(X))))
    return BinaryOperator::CreateAdd(X, ConstantExpr::getSub(C, C2));
}

const APInt *Op0C;
if (match(Op0, m_APInt(Op0C)) && Op0C->isMask()) {
  // Turn this into a xor if LHS is 2^n-1 and the remaining bits are known
  // zero.
  KnownBits RHSKnown = computeKnownBits(Op1, 0, &I);
  if ((*Op0C | RHSKnown.Zero).isAllOnesValue())
    return BinaryOperator::CreateXor(Op1, Op0);
}

{
  Value *Y;
  // X-(X+Y) == -Y    X-(Y+X) == -Y
  if (match(Op1, m_c_Add(m_Specific(Op0), m_Value(Y))))
    return BinaryOperator::CreateNeg(Y);

  // (X-Y)-X == -Y
  if (match(Op0, m_Sub(m_Specific(Op1), m_Value(Y))))
    return BinaryOperator::CreateNeg(Y);
}

// (sub (or A, B) (and A, B)) --> (xor A, B)
{
  Value *A, *B;
  if (match(Op1, m_And(m_Value(A), m_Value(B))) &&
      match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
    return BinaryOperator::CreateXor(A, B);
}

// (sub (add A, B) (or A, B)) --> (and A, B)
{
  Value *A, *B;
  if (match(Op0, m_Add(m_Value(A), m_Value(B))) &&
      match(Op1, m_c_Or(m_Specific(A), m_Specific(B))))
    return BinaryOperator::CreateAnd(A, B);
}

// (sub (add A, B) (and A, B)) --> (or A, B)
{
  Value *A, *B;
  if (match(Op0, m_Add(m_Value(A), m_Value(B))) &&
      match(Op1, m_c_And(m_Specific(A), m_Specific(B))))
    return BinaryOperator::CreateOr(A, B);
}

// (sub (and A, B) (or A, B)) --> neg (xor A, B)
{
  Value *A, *B;
  if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
      match(Op1, m_c_Or(m_Specific(A), m_Specific(B))) &&
      (Op0->hasOneUse() || Op1->hasOneUse()))
    return BinaryOperator::CreateNeg(Builder.CreateXor(A, B));
}

// (sub (or A, B), (xor A, B)) --> (and A, B)
{
  Value *A, *B;
  if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
      match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
    return BinaryOperator::CreateAnd(A, B);
}

// (sub (xor A, B) (or A, B)) --> neg (and A, B)
{
  Value *A, *B;
  if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
      match(Op1, m_c_Or(m_Specific(A), m_Specific(B))) &&
      (Op0->hasOneUse() || Op1->hasOneUse()))
    return BinaryOperator::CreateNeg(Builder.CreateAnd(A, B));
}

{
  Value *Y;
  // ((X | Y) - X) --> (~X & Y)
  if (match(Op0, m_OneUse(m_c_Or(m_Value(Y), m_Specific(Op1)))))
    return BinaryOperator::CreateAnd(
        Y, Builder.CreateNot(Op1, Op1->getName() + ".not"));
}

{
  // (sub (and Op1, (neg X)), Op1) --> neg (and Op1, (add X, -1))
  Value *X;
  if (match(Op0, m_OneUse(m_c_And(m_Specific(Op1),
                                  m_OneUse(m_Neg(m_Value(X))))))) {
    return BinaryOperator::CreateNeg(Builder.CreateAnd(
        Op1, Builder.CreateAdd(X, Constant::getAllOnesValue(I.getType()))));
  }
}

{
  // (sub (and Op1, C), Op1) --> neg (and Op1, ~C)
  Constant *C;
  if (match(Op0, m_OneUse(m_And(m_Specific(Op1), m_Constant(C))))) {
    return BinaryOperator::CreateNeg(
        Builder.CreateAnd(Op1, Builder.CreateNot(C)));
  }
}

{
  // If we have a subtraction between some value and a select between
  // said value and something else, sink subtraction into select hands, i.e.:
  //   sub (select %Cond, %TrueVal, %FalseVal), %Op1
  //     ->
  //   select %Cond, (sub %TrueVal, %Op1), (sub %FalseVal, %Op1)
  //  or
  //   sub %Op0, (select %Cond, %TrueVal, %FalseVal)
  //     ->
  //   select %Cond, (sub %Op0, %TrueVal), (sub %Op0, %FalseVal)
  // This will result in select between new subtraction and 0.
  auto SinkSubIntoSelect =
      [Ty = I.getType()](Value *Select, Value *OtherHandOfSub,
                         auto SubBuilder) -> Instruction * {
    Value *Cond, *TrueVal, *FalseVal;
    if (!match(Select, m_OneUse(m_Select(m_Value(Cond), m_Value(TrueVal),
                                         m_Value(FalseVal)))))
      return nullptr;
    if (OtherHandOfSub != TrueVal && OtherHandOfSub != FalseVal)
      return nullptr;
    // While it is really tempting to just create two subtractions and let
    // InstCombine fold one of those to 0, it isn't possible to do so
    // because of worklist visitation order. So ugly it is.
    bool OtherHandOfSubIsTrueVal = OtherHandOfSub == TrueVal;
    Value *NewSub = SubBuilder(OtherHandOfSubIsTrueVal ? FalseVal : TrueVal);
    Constant *Zero = Constant::getNullValue(Ty);
    SelectInst *NewSel =
        SelectInst::Create(Cond, OtherHandOfSubIsTrueVal ? Zero : NewSub,
                           OtherHandOfSubIsTrueVal ? NewSub : Zero);
    // Preserve prof metadata if any.
    NewSel->copyMetadata(cast<Instruction>(*Select));
    return NewSel;
  };
  if (Instruction *NewSel = SinkSubIntoSelect(
          /*Select=*/Op0, /*OtherHandOfSub=*/Op1,
          [Builder = &Builder, Op1](Value *OtherHandOfSelect) {
            return Builder->CreateSub(OtherHandOfSelect,
                                      /*OtherHandOfSub=*/Op1);
          }))
    return NewSel;
  if (Instruction *NewSel = SinkSubIntoSelect(
          /*Select=*/Op1, /*OtherHandOfSub=*/Op0,
          [Builder = &Builder, Op0](Value *OtherHandOfSelect) {
            return Builder->CreateSub(/*OtherHandOfSub=*/Op0,
                                      OtherHandOfSelect);
          }))
    return NewSel;
}

// (X - (X & Y))   -->   (X & ~Y)
if (match(Op1, m_c_And(m_Specific(Op0), m_Value(Y))) &&
    (Op1->hasOneUse() || isa<Constant>(Y)))
  return BinaryOperator::CreateAnd(
      Op0, Builder.CreateNot(Y, Y->getName() + ".not"));

{
  // ~A - Min/Max(~A, O) -> Max/Min(A, ~O) - A
  // ~A - Min/Max(O, ~A) -> Max/Min(A, ~O) - A
  // Min/Max(~A, O) - ~A -> A - Max/Min(A, ~O)
  // Min/Max(O, ~A) - ~A -> A - Max/Min(A, ~O)
  // So long as O here is freely invertible, this will be neutral or a win.
  Value *LHS, *RHS, *A;
  Value *NotA = Op0, *MinMax = Op1;
  SelectPatternFlavor SPF = matchSelectPattern(MinMax, LHS, RHS).Flavor;
  if (!SelectPatternResult::isMinOrMax(SPF)) {
    NotA = Op1;
    MinMax = Op0;
    SPF = matchSelectPattern(MinMax, LHS, RHS).Flavor;
  }
  if (SelectPatternResult::isMinOrMax(SPF) &&
      match(NotA, m_Not(m_Value(A))) && (NotA == LHS || NotA == RHS)) {
    if (NotA == LHS)
      std::swap(LHS, RHS);
    // LHS is now O above and expected to have at least 2 uses (the min/max)
    // NotA is epected to have 2 uses from the min/max and 1 from the sub.
    if (isFreeToInvert(LHS, !LHS->hasNUsesOrMore(3)) &&
        !NotA->hasNUsesOrMore(4)) {
      // Note: We don't generate the inverse max/min, just create the not of
      // it and let other folds do the rest.
      Value *Not = Builder.CreateNot(MinMax);
      if (NotA == Op0)
        return BinaryOperator::CreateSub(Not, A);
      else
        return BinaryOperator::CreateSub(A, Not);
    }
  }
}

// Optimize pointer differences into the same array into a size.  Consider:
//  &A[10] - &A[0]: we should compile this to "10".
Value *LHSOp, *RHSOp;
if (match(Op0, m_PtrToInt(m_Value(LHSOp))) &&
    match(Op1, m_PtrToInt(m_Value(RHSOp))))
  if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(),
                                             I.hasNoUnsignedWrap()))
    return replaceInstUsesWith(I, Res);

// trunc(p)-trunc(q) -> trunc(p-q)
if (match(Op0, m_Trunc(m_PtrToInt(m_Value(LHSOp)))) &&
    match(Op1, m_Trunc(m_PtrToInt(m_Value(RHSOp)))))
  if (Value *Res = OptimizePointerDifference(LHSOp, RHSOp, I.getType(),
                                             /* IsNUW */ false))
    return replaceInstUsesWith(I, Res);

// Canonicalize a shifty way to code absolute value to the common pattern.
// There are 2 potential commuted variants.
// We're relying on the fact that we only do this transform when the shift has
// exactly 2 uses and the xor has exactly 1 use (otherwise, we might increase
// instructions).
Value *A;
const APInt *ShAmt;
Type *Ty = I.getType();
if (match(Op1, m_AShr(m_Value(A), m_APInt(ShAmt))) &&
    Op1->hasNUses(2) && *ShAmt == Ty->getScalarSizeInBits() - 1 &&
    match(Op0, m_OneUse(m_c_Xor(m_Specific(A), m_Specific(Op1))))) {
  // B = ashr i32 A, 31 ; smear the sign bit
  // sub (xor A, B), B  ; flip bits if negative and subtract -1 (add 1)
  // --> (A < 0) ? -A : A
  Value *Cmp = Builder.CreateICmpSLT(A, ConstantInt::getNullValue(Ty));
  // Copy the nuw/nsw flags from the sub to the negate.
  Value *Neg = Builder.CreateNeg(A, "", I.hasNoUnsignedWrap(),
                                 I.hasNoSignedWrap());
  return SelectInst::Create(Cmp, Neg, A);
}

// If we are subtracting a low-bit masked subset of some value from an add
// of that same value with no low bits changed, that is clearing some low bits
// of the sum:
// sub (X + AddC), (X & AndC) --> and (X + AddC), ~AndC
const APInt *AddC, *AndC;
if (match(Op0, m_Add(m_Value(X), m_APInt(AddC))) &&
    match(Op1, m_And(m_Specific(X), m_APInt(AndC)))) {
  unsigned BitWidth = Ty->getScalarSizeInBits();
  unsigned Cttz = AddC->countTrailingZeros();
  APInt HighMask(APInt::getHighBitsSet(BitWidth, BitWidth - Cttz));
  if ((HighMask & *AndC).isNullValue())
    return BinaryOperator::CreateAnd(Op0, ConstantInt::get(Ty, ~(*AndC)));
}

if (Instruction *V =
        canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(I))
  return V;

// X - usub.sat(X, Y) => umin(X, Y)
if (match(Op1, m_OneUse(m_Intrinsic<Intrinsic::usub_sat>(m_Specific(Op0),
                                                         m_Value(Y)))))
  return replaceInstUsesWith(
      I, Builder.CreateIntrinsic(Intrinsic::umin, {I.getType()}, {Op0, Y}));

// C - ctpop(X) => ctpop(~X) if C is bitwidth
if (match(Op0, m_SpecificInt(Ty->getScalarSizeInBits())) &&
    match(Op1, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(X)))))
  return replaceInstUsesWith(
      I, Builder.CreateIntrinsic(Intrinsic::ctpop, {I.getType()},
                                 {Builder.CreateNot(X)}));

return TryToNarrowDeduceFlags();
2162}

2164/// This eliminates floating-point negation in either 'fneg(X)' or
2165/// 'fsub(-0.0, X)' form by combining into a constant operand.
2166static Instruction *foldFNegIntoConstant(Instruction &I) {
// This is limited with one-use because fneg is assumed better for
// reassociation and cheaper in codegen than fmul/fdiv.
// TODO: Should the m_OneUse restriction be removed?
Instruction *FNegOp;
if (!match(&I, m_FNeg(m_OneUse(m_Instruction(FNegOp)))))
  return nullptr;

Value *X;
Constant *C;

// Fold negation into constant operand.
// -(X * C) --> X * (-C)
if (match(FNegOp, m_FMul(m_Value(X), m_Constant(C))))
  return BinaryOperator::CreateFMulFMF(X, ConstantExpr::getFNeg(C), &I);
// -(X / C) --> X / (-C)
if (match(FNegOp, m_FDiv(m_Value(X), m_Constant(C))))
  return BinaryOperator::CreateFDivFMF(X, ConstantExpr::getFNeg(C), &I);
// -(C / X) --> (-C) / X
if (match(FNegOp, m_FDiv(m_Constant(C), m_Value(X)))) {
  Instruction *FDiv =
      BinaryOperator::CreateFDivFMF(ConstantExpr::getFNeg(C), X, &I);

  // Intersect 'nsz' and 'ninf' because those special value exceptions may not
  // apply to the fdiv. Everything else propagates from the fneg.
  // TODO: We could propagate nsz/ninf from fdiv alone?
  FastMathFlags FMF = I.getFastMathFlags();
  FastMathFlags OpFMF = FNegOp->getFastMathFlags();
  FDiv->setHasNoSignedZeros(FMF.noSignedZeros() & OpFMF.noSignedZeros());
  FDiv->setHasNoInfs(FMF.noInfs() & OpFMF.noInfs());
  return FDiv;
}
// With NSZ [ counter-example with -0.0: -(-0.0 + 0.0) != 0.0 + -0.0 ]:
// -(X + C) --> -X + -C --> -C - X
if (I.hasNoSignedZeros() && match(FNegOp, m_FAdd(m_Value(X), m_Constant(C))))
  return BinaryOperator::CreateFSubFMF(ConstantExpr::getFNeg(C), X, &I);

return nullptr;
2204}

2206static Instruction *hoistFNegAboveFMulFDiv(Instruction &I,
                                         InstCombiner::BuilderTy &Builder) {
Value *FNeg;
if (!match(&I, m_FNeg(m_Value(FNeg))))
  return nullptr;

Value *X, *Y;
if (match(FNeg, m_OneUse(m_FMul(m_Value(X), m_Value(Y)))))
  return BinaryOperator::CreateFMulFMF(Builder.CreateFNegFMF(X, &I), Y, &I);

if (match(FNeg, m_OneUse(m_FDiv(m_Value(X), m_Value(Y)))))
  return BinaryOperator::CreateFDivFMF(Builder.CreateFNegFMF(X, &I), Y, &I);

return nullptr;
2220}

2222Instruction *InstCombinerImpl::visitFNeg(UnaryOperator &I) {
Value *Op = I.getOperand(0);

if (Value *V = SimplifyFNegInst(Op, I.getFastMathFlags(),
                                getSimplifyQuery().getWithInstruction(&I)))
  return replaceInstUsesWith(I, V);

if (Instruction *X = foldFNegIntoConstant(I))
  return X;

Value *X, *Y;

// If we can ignore the sign of zeros: -(X - Y) --> (Y - X)
if (I.hasNoSignedZeros() &&
    match(Op, m_OneUse(m_FSub(m_Value(X), m_Value(Y)))))
  return BinaryOperator::CreateFSubFMF(Y, X, &I);

if (Instruction *R = hoistFNegAboveFMulFDiv(I, Builder))
  return R;

// Try to eliminate fneg if at least 1 arm of the select is negated.
Value *Cond;
if (match(Op, m_OneUse(m_Select(m_Value(Cond), m_Value(X), m_Value(Y))))) {
  // Unlike most transforms, this one is not safe to propagate nsz unless
  // it is present on the original select. (We are conservatively intersecting
  // the nsz flags from the select and root fneg instruction.)
  auto propagateSelectFMF = [&](SelectInst *S) {
    S->copyFastMathFlags(&I);
    if (auto *OldSel = dyn_cast<SelectInst>(Op))
      if (!OldSel->hasNoSignedZeros())
        S->setHasNoSignedZeros(false);
  };
  // -(Cond ? -P : Y) --> Cond ? P : -Y
  Value *P;
  if (match(X, m_FNeg(m_Value(P)))) {
    Value *NegY = Builder.CreateFNegFMF(Y, &I, Y->getName() + ".neg");
    SelectInst *NewSel = SelectInst::Create(Cond, P, NegY);
    propagateSelectFMF(NewSel);
    return NewSel;
  }
  // -(Cond ? X : -P) --> Cond ? -X : P
  if (match(Y, m_FNeg(m_Value(P)))) {
    Value *NegX = Builder.CreateFNegFMF(X, &I, X->getName() + ".neg");
    SelectInst *NewSel = SelectInst::Create(Cond, NegX, P);
    propagateSelectFMF(NewSel);
    return NewSel;
  }
}

return nullptr;
2272}

2274Instruction *InstCombinerImpl::visitFSub(BinaryOperator &I) {
if (Value *V = SimplifyFSubInst(I.getOperand(0), I.getOperand(1),
                                I.getFastMathFlags(),
                                getSimplifyQuery().getWithInstruction(&I)))
  return replaceInstUsesWith(I, V);

if (Instruction *X = foldVectorBinop(I))
  return X;

// Subtraction from -0.0 is the canonical form of fneg.
// fsub -0.0, X ==> fneg X
// fsub nsz 0.0, X ==> fneg nsz X
//
// FIXME This matcher does not respect FTZ or DAZ yet:
// fsub -0.0, Denorm ==> +-0
// fneg Denorm ==> -Denorm
Value *Op;
if (match(&I, m_FNeg(m_Value(Op))))
  return UnaryOperator::CreateFNegFMF(Op, &I);

if (Instruction *X = foldFNegIntoConstant(I))
  return X;

if (Instruction *R = hoistFNegAboveFMulFDiv(I, Builder))
  return R;

Value *X, *Y;
Constant *C;

Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
// If Op0 is not -0.0 or we can ignore -0.0: Z - (X - Y) --> Z + (Y - X)
// Canonicalize to fadd to make analysis easier.
// This can also help codegen because fadd is commutative.
// Note that if this fsub was really an fneg, the fadd with -0.0 will get
// killed later. We still limit that particular transform with 'hasOneUse'
// because an fneg is assumed better/cheaper than a generic fsub.
if (I.hasNoSignedZeros() || CannotBeNegativeZero(Op0, SQ.TLI)) {
  if (match(Op1, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) {
    Value *NewSub = Builder.CreateFSubFMF(Y, X, &I);
    return BinaryOperator::CreateFAddFMF(Op0, NewSub, &I);
  }
}

// (-X) - Op1 --> -(X + Op1)
if (I.hasNoSignedZeros() && !isa<ConstantExpr>(Op0) &&
    match(Op0, m_OneUse(m_FNeg(m_Value(X))))) {
  Value *FAdd = Builder.CreateFAddFMF(X, Op1, &I);
  return UnaryOperator::CreateFNegFMF(FAdd, &I);
}

if (isa<Constant>(Op0))
  if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
    if (Instruction *NV = FoldOpIntoSelect(I, SI))
      return NV;

// X - C --> X + (-C)
// But don't transform constant expressions because there's an inverse fold
// for X + (-Y) --> X - Y.
if (match(Op1, m_ImmConstant(C)))
  return BinaryOperator::CreateFAddFMF(Op0, ConstantExpr::getFNeg(C), &I);

// X - (-Y) --> X + Y
if (match(Op1, m_FNeg(m_Value(Y))))
  return BinaryOperator::CreateFAddFMF(Op0, Y, &I);

// Similar to above, but look through a cast of the negated value:
// X - (fptrunc(-Y)) --> X + fptrunc(Y)
Type *Ty = I.getType();
if (match(Op1, m_OneUse(m_FPTrunc(m_FNeg(m_Value(Y))))))
  return BinaryOperator::CreateFAddFMF(Op0, Builder.CreateFPTrunc(Y, Ty), &I);

// X - (fpext(-Y)) --> X + fpext(Y)
if (match(Op1, m_OneUse(m_FPExt(m_FNeg(m_Value(Y))))))
  return BinaryOperator::CreateFAddFMF(Op0, Builder.CreateFPExt(Y, Ty), &I);

// Similar to above, but look through fmul/fdiv of the negated value:
// Op0 - (-X * Y) --> Op0 + (X * Y)
// Op0 - (Y * -X) --> Op0 + (X * Y)
if (match(Op1, m_OneUse(m_c_FMul(m_FNeg(m_Value(X)), m_Value(Y))))) {
  Value *FMul = Builder.CreateFMulFMF(X, Y, &I);
  return BinaryOperator::CreateFAddFMF(Op0, FMul, &I);
}
// Op0 - (-X / Y) --> Op0 + (X / Y)
// Op0 - (X / -Y) --> Op0 + (X / Y)
if (match(Op1, m_OneUse(m_FDiv(m_FNeg(m_Value(X)), m_Value(Y)))) ||
    match(Op1, m_OneUse(m_FDiv(m_Value(X), m_FNeg(m_Value(Y)))))) {
  Value *FDiv = Builder.CreateFDivFMF(X, Y, &I);
  return BinaryOperator::CreateFAddFMF(Op0, FDiv, &I);
}

// Handle special cases for FSub with selects feeding the operation
if (Value *V = SimplifySelectsFeedingBinaryOp(I, Op0, Op1))
  return replaceInstUsesWith(I, V);

if (I.hasAllowReassoc() && I.hasNoSignedZeros()) {
  // (Y - X) - Y --> -X
  if (match(Op0, m_FSub(m_Specific(Op1), m_Value(X))))
    return UnaryOperator::CreateFNegFMF(X, &I);

  // Y - (X + Y) --> -X
  // Y - (Y + X) --> -X
  if (match(Op1, m_c_FAdd(m_Specific(Op0), m_Value(X))))
    return UnaryOperator::CreateFNegFMF(X, &I);

  // (X * C) - X --> X * (C - 1.0)
  if (match(Op0, m_FMul(m_Specific(Op1), m_Constant(C)))) {
    Constant *CSubOne = ConstantExpr::getFSub(C, ConstantFP::get(Ty, 1.0));
    return BinaryOperator::CreateFMulFMF(Op1, CSubOne, &I);
  }
  // X - (X * C) --> X * (1.0 - C)
  if (match(Op1, m_FMul(m_Specific(Op0), m_Constant(C)))) {
    Constant *OneSubC = ConstantExpr::getFSub(ConstantFP::get(Ty, 1.0), C);
    return BinaryOperator::CreateFMulFMF(Op0, OneSubC, &I);
  }

  // Reassociate fsub/fadd sequences to create more fadd instructions and
  // reduce dependency chains:
  // ((X - Y) + Z) - Op1 --> (X + Z) - (Y + Op1)
  Value *Z;
  if (match(Op0, m_OneUse(m_c_FAdd(m_OneUse(m_FSub(m_Value(X), m_Value(Y))),
                                   m_Value(Z))))) {
    Value *XZ = Builder.CreateFAddFMF(X, Z, &I);
    Value *YW = Builder.CreateFAddFMF(Y, Op1, &I);
    return BinaryOperator::CreateFSubFMF(XZ, YW, &I);
  }

  auto m_FaddRdx = [](Value *&Sum, Value *&Vec) {
    return m_OneUse(m_Intrinsic<Intrinsic::vector_reduce_fadd>(m_Value(Sum),
                                                               m_Value(Vec)));
  };
  Value *A0, *A1, *V0, *V1;
  if (match(Op0, m_FaddRdx(A0, V0)) && match(Op1, m_FaddRdx(A1, V1)) &&
      V0->getType() == V1->getType()) {
    // Difference of sums is sum of differences:
    // add_rdx(A0, V0) - add_rdx(A1, V1) --> add_rdx(A0, V0 - V1) - A1
    Value *Sub = Builder.CreateFSubFMF(V0, V1, &I);
    Value *Rdx = Builder.CreateIntrinsic(Intrinsic::vector_reduce_fadd,
                                         {Sub->getType()}, {A0, Sub}, &I);
    return BinaryOperator::CreateFSubFMF(Rdx, A1, &I);
  }

  if (Instruction *F = factorizeFAddFSub(I, Builder))
    return F;

  // TODO: This performs reassociative folds for FP ops. Some fraction of the
  // functionality has been subsumed by simple pattern matching here and in
  // InstSimplify. We should let a dedicated reassociation pass handle more
  // complex pattern matching and remove this from InstCombine.
  if (Value *V = FAddCombine(Builder).simplify(&I))
    return replaceInstUsesWith(I, V);

  // (X - Y) - Op1 --> X - (Y + Op1)
  if (match(Op0, m_OneUse(m_FSub(m_Value(X), m_Value(Y))))) {
    Value *FAdd = Builder.CreateFAddFMF(Y, Op1, &I);
    return BinaryOperator::CreateFSubFMF(X, FAdd, &I);
  }
}

return nullptr;
2433}

←

/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/include/llvm/IR/PatternMatch.h

1//===- PatternMatch.h - Match on the LLVM IR --------------------*- C++ -*-===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file provides a simple and efficient mechanism for performing general
10// tree-based pattern matches on the LLVM IR. The power of these routines is
11// that it allows you to write concise patterns that are expressive and easy to
12// understand. The other major advantage of this is that it allows you to
13// trivially capture/bind elements in the pattern to variables. For example,
14// you can do something like this:
15//
16//  Value *Exp = ...
17//  Value *X, *Y;  ConstantInt *C1, *C2;      // (X & C1) | (Y & C2)
18//  if (match(Exp, m_Or(m_And(m_Value(X), m_ConstantInt(C1)),
19//                      m_And(m_Value(Y), m_ConstantInt(C2))))) {
20//    ... Pattern is matched and variables are bound ...
21//  }
22//
23// This is primarily useful to things like the instruction combiner, but can
24// also be useful for static analysis tools or code generators.
25//
26//===----------------------------------------------------------------------===//
27 
28#ifndef LLVM_IR_PATTERNMATCH_H
29#define LLVM_IR_PATTERNMATCH_H
30 
31#include "llvm/ADT/APFloat.h"
32#include "llvm/ADT/APInt.h"
33#include "llvm/IR/Constant.h"
34#include "llvm/IR/Constants.h"
35#include "llvm/IR/DataLayout.h"
36#include "llvm/IR/InstrTypes.h"
37#include "llvm/IR/Instruction.h"
38#include "llvm/IR/Instructions.h"
39#include "llvm/IR/IntrinsicInst.h"
40#include "llvm/IR/Intrinsics.h"
41#include "llvm/IR/Operator.h"
42#include "llvm/IR/Value.h"
43#include "llvm/Support/Casting.h"
44#include <cstdint>
45 
46namespace llvm {
47namespace PatternMatch {
48 
49template <typename Val, typename Pattern> bool match(Val *V, const Pattern &P) {
50  return const_cast<Pattern &>(P).match(V);
14
←
Calling 'AnyBinaryOp_match::match'→
33
←
Returning from 'AnyBinaryOp_match::match'→
34
←
Returning the value 1, which participates in a condition later→
56
←
Returning without writing to 'P.R.VR'→
62
←
Passing null pointer value via 1st parameter 'V'→
63
←
Calling 'ThreeOps_match::match'→
51}
52 
53template <typename Pattern> bool match(ArrayRef<int> Mask, const Pattern &P) {
54  return const_cast<Pattern &>(P).match(Mask);
55}
56 
57template <typename SubPattern_t> struct OneUse_match {
58  SubPattern_t SubPattern;
59 
60  OneUse_match(const SubPattern_t &SP) : SubPattern(SP) {}
61 
62  template <typename OpTy> bool match(OpTy *V) {
63    return V->hasOneUse() && SubPattern.match(V);
64  }
65};
66 
67template <typename T> inline OneUse_match<T> m_OneUse(const T &SubPattern) {
68  return SubPattern;
69}
70 
71template <typename Class> struct class_match {
72  template <typename ITy> bool match(ITy *V) { return isa<Class>(V); }
73};
74 
75/// Match an arbitrary value and ignore it.
76inline class_match<Value> m_Value() { return class_match<Value>(); }
77 
78/// Match an arbitrary unary operation and ignore it.
79inline class_match<UnaryOperator> m_UnOp() {
80  return class_match<UnaryOperator>();
81}
82 
83/// Match an arbitrary binary operation and ignore it.
84inline class_match<BinaryOperator> m_BinOp() {
85  return class_match<BinaryOperator>();
86}
87 
88/// Matches any compare instruction and ignore it.
89inline class_match<CmpInst> m_Cmp() { return class_match<CmpInst>(); }
90 
91struct undef_match {
92  static bool check(const Value *V) {
93    if (isa<UndefValue>(V))
94      return true;
95 
96    const auto *CA = dyn_cast<ConstantAggregate>(V);
97    if (!CA)
98      return false;
99 
100    SmallPtrSet<const ConstantAggregate *, 8> Seen;
101    SmallVector<const ConstantAggregate *, 8> Worklist;
102 
103    // Either UndefValue, PoisonValue, or an aggregate that only contains
104    // these is accepted by matcher.
105    // CheckValue returns false if CA cannot satisfy this constraint.
106    auto CheckValue = [&](const ConstantAggregate *CA) {
107      for (const Value *Op : CA->operand_values()) {
108        if (isa<UndefValue>(Op))
109          continue;
110 
111        const auto *CA = dyn_cast<ConstantAggregate>(Op);
112        if (!CA)
113          return false;
114        if (Seen.insert(CA).second)
115          Worklist.emplace_back(CA);
116      }
117 
118      return true;
119    };
120 
121    if (!CheckValue(CA))
122      return false;
123 
124    while (!Worklist.empty()) {
125      if (!CheckValue(Worklist.pop_back_val()))
126        return false;
127    }
128    return true;
129  }
130  template <typename ITy> bool match(ITy *V) { return check(V); }
131};
132 
133/// Match an arbitrary undef constant. This matches poison as well.
134/// If this is an aggregate and contains a non-aggregate element that is
135/// neither undef nor poison, the aggregate is not matched.
136inline auto m_Undef() { return undef_match(); }
137 
138/// Match an arbitrary poison constant.
139inline class_match<PoisonValue> m_Poison() { return class_match<PoisonValue>(); }
140 
141/// Match an arbitrary Constant and ignore it.
142inline class_match<Constant> m_Constant() { return class_match<Constant>(); }
143 
144/// Match an arbitrary ConstantInt and ignore it.
145inline class_match<ConstantInt> m_ConstantInt() {
146  return class_match<ConstantInt>();
147}
148 
149/// Match an arbitrary ConstantFP and ignore it.
150inline class_match<ConstantFP> m_ConstantFP() {
151  return class_match<ConstantFP>();
152}
153 
154/// Match an arbitrary ConstantExpr and ignore it.
155inline class_match<ConstantExpr> m_ConstantExpr() {
156  return class_match<ConstantExpr>();
157}
158 
159/// Match an arbitrary basic block value and ignore it.
160inline class_match<BasicBlock> m_BasicBlock() {
161  return class_match<BasicBlock>();
162}
163 
164/// Inverting matcher
165template <typename Ty> struct match_unless {
166  Ty M;
167 
168  match_unless(const Ty &Matcher) : M(Matcher) {}
169 
170  template <typename ITy> bool match(ITy *V) { return !M.match(V); }
171};
172 
173/// Match if the inner matcher does *NOT* match.
174template <typename Ty> inline match_unless<Ty> m_Unless(const Ty &M) {
175  return match_unless<Ty>(M);
176}
177 
178/// Matching combinators
179template <typename LTy, typename RTy> struct match_combine_or {
180  LTy L;
181  RTy R;
182 
183  match_combine_or(const LTy &Left, const RTy &Right) : L(Left), R(Right) {}
184 
185  template <typename ITy> bool match(ITy *V) {
186    if (L.match(V))
17
←
Taking false branch→
187      return true;
188    if (R.match(V))
18
←
Calling 'match_combine_and::match'→
29
←
Returning from 'match_combine_and::match'→
30
←
Taking true branch→
189      return true;
190    return false;
191  }
192};
193 
194template <typename LTy, typename RTy> struct match_combine_and {
195  LTy L;
196  RTy R;
197 
198  match_combine_and(const LTy &Left, const RTy &Right) : L(Left), R(Right) {}
199 
200  template <typename ITy> bool match(ITy *V) {
201    if (L.match(V))
19
←
Calling 'BinaryOp_match::match'→
26
←
Returning from 'BinaryOp_match::match'→
27
←
Taking true branch→
202      if (R.match(V))
28
←
Taking true branch→
203        return true;
204    return false;
205  }
206};
207 
208/// Combine two pattern matchers matching L || R
209template <typename LTy, typename RTy>
210inline match_combine_or<LTy, RTy> m_CombineOr(const LTy &L, const RTy &R) {
211  return match_combine_or<LTy, RTy>(L, R);
51
←
Returning without writing to 'L.Op.VR'→
212}
213 
214/// Combine two pattern matchers matching L && R
215template <typename LTy, typename RTy>
216inline match_combine_and<LTy, RTy> m_CombineAnd(const LTy &L, const RTy &R) {
217  return match_combine_and<LTy, RTy>(L, R);
218}
219 
220struct apint_match {
221  const APInt *&Res;
222  bool AllowUndef;
223 
224  apint_match(const APInt *&Res, bool AllowUndef)
225    : Res(Res), AllowUndef(AllowUndef) {}
226 
227  template <typename ITy> bool match(ITy *V) {
228    if (auto *CI = dyn_cast<ConstantInt>(V)) {
229      Res = &CI->getValue();
230      return true;
231    }
232    if (V->getType()->isVectorTy())
233      if (const auto *C = dyn_cast<Constant>(V))
234        if (auto *CI = dyn_cast_or_null<ConstantInt>(
235                C->getSplatValue(AllowUndef))) {
236          Res = &CI->getValue();
237          return true;
238        }
239    return false;
240  }
241};
242// Either constexpr if or renaming ConstantFP::getValueAPF to
243// ConstantFP::getValue is needed to do it via single template
244// function for both apint/apfloat.
245struct apfloat_match {
246  const APFloat *&Res;
247  bool AllowUndef;
248 
249  apfloat_match(const APFloat *&Res, bool AllowUndef)
250      : Res(Res), AllowUndef(AllowUndef) {}
251 
252  template <typename ITy> bool match(ITy *V) {
253    if (auto *CI = dyn_cast<ConstantFP>(V)) {
254      Res = &CI->getValueAPF();
255      return true;
256    }
257    if (V->getType()->isVectorTy())
258      if (const auto *C = dyn_cast<Constant>(V))
259        if (auto *CI = dyn_cast_or_null<ConstantFP>(
260                C->getSplatValue(AllowUndef))) {
261          Res = &CI->getValueAPF();
262          return true;
263        }
264    return false;
265  }
266};
267 
268/// Match a ConstantInt or splatted ConstantVector, binding the
269/// specified pointer to the contained APInt.
270inline apint_match m_APInt(const APInt *&Res) {
271  // Forbid undefs by default to maintain previous behavior.
272  return apint_match(Res, /* AllowUndef */ false);
273}
274 
275/// Match APInt while allowing undefs in splat vector constants.
276inline apint_match m_APIntAllowUndef(const APInt *&Res) {
277  return apint_match(Res, /* AllowUndef */ true);
278}
279 
280/// Match APInt while forbidding undefs in splat vector constants.
281inline apint_match m_APIntForbidUndef(const APInt *&Res) {
282  return apint_match(Res, /* AllowUndef */ false);
283}
284 
285/// Match a ConstantFP or splatted ConstantVector, binding the
286/// specified pointer to the contained APFloat.
287inline apfloat_match m_APFloat(const APFloat *&Res) {
288  // Forbid undefs by default to maintain previous behavior.
289  return apfloat_match(Res, /* AllowUndef */ false);
290}
291 
292/// Match APFloat while allowing undefs in splat vector constants.
293inline apfloat_match m_APFloatAllowUndef(const APFloat *&Res) {
294  return apfloat_match(Res, /* AllowUndef */ true);
295}
296 
297/// Match APFloat while forbidding undefs in splat vector constants.
298inline apfloat_match m_APFloatForbidUndef(const APFloat *&Res) {
299  return apfloat_match(Res, /* AllowUndef */ false);
300}
301 
302template <int64_t Val> struct constantint_match {
303  template <typename ITy> bool match(ITy *V) {
304    if (const auto *CI = dyn_cast<ConstantInt>(V)) {
305      const APInt &CIV = CI->getValue();
306      if (Val >= 0)
307        return CIV == static_cast<uint64_t>(Val);
308      // If Val is negative, and CI is shorter than it, truncate to the right
309      // number of bits.  If it is larger, then we have to sign extend.  Just
310      // compare their negated values.
311      return -CIV == -Val;
312    }
313    return false;
314  }
315};
316 
317/// Match a ConstantInt with a specific value.
318template <int64_t Val> inline constantint_match<Val> m_ConstantInt() {
319  return constantint_match<Val>();
320}
321 
322/// This helper class is used to match constant scalars, vector splats,
323/// and fixed width vectors that satisfy a specified predicate.
324/// For fixed width vector constants, undefined elements are ignored.
325template <typename Predicate, typename ConstantVal>
326struct cstval_pred_ty : public Predicate {
327  template <typename ITy> bool match(ITy *V) {
328    if (const auto *CV = dyn_cast<ConstantVal>(V))
329      return this->isValue(CV->getValue());
330    if (const auto *VTy = dyn_cast<VectorType>(V->getType())) {
331      if (const auto *C = dyn_cast<Constant>(V)) {
332        if (const auto *CV = dyn_cast_or_null<ConstantVal>(C->getSplatValue()))
333          return this->isValue(CV->getValue());
334 
335        // Number of elements of a scalable vector unknown at compile time
336        auto *FVTy = dyn_cast<FixedVectorType>(VTy);
337        if (!FVTy)
338          return false;
339 
340        // Non-splat vector constant: check each element for a match.
341        unsigned NumElts = FVTy->getNumElements();
342        assert(NumElts != 0 && "Constant vector with no elements?")(static_cast <bool> (NumElts != 0 && "Constant vector with no elements?"
) ? void (0) : __assert_fail ("NumElts != 0 && \"Constant vector with no elements?\""
, "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/include/llvm/IR/PatternMatch.h"
, 342, __extension__ __PRETTY_FUNCTION__));
343        bool HasNonUndefElements = false;
344        for (unsigned i = 0; i != NumElts; ++i) {
345          Constant *Elt = C->getAggregateElement(i);
346          if (!Elt)
347            return false;
348          if (isa<UndefValue>(Elt))
349            continue;
350          auto *CV = dyn_cast<ConstantVal>(Elt);
351          if (!CV || !this->isValue(CV->getValue()))
352            return false;
353          HasNonUndefElements = true;
354        }
355        return HasNonUndefElements;
356      }
357    }
358    return false;
359  }
360};
361 
362/// specialization of cstval_pred_ty for ConstantInt
363template <typename Predicate>
364using cst_pred_ty = cstval_pred_ty<Predicate, ConstantInt>;
365 
366/// specialization of cstval_pred_ty for ConstantFP
367template <typename Predicate>
368using cstfp_pred_ty = cstval_pred_ty<Predicate, ConstantFP>;
369 
370/// This helper class is used to match scalar and vector constants that
371/// satisfy a specified predicate, and bind them to an APInt.
372template <typename Predicate> struct api_pred_ty : public Predicate {
373  const APInt *&Res;
374 
375  api_pred_ty(const APInt *&R) : Res(R) {}
376 
377  template <typename ITy> bool match(ITy *V) {
378    if (const auto *CI = dyn_cast<ConstantInt>(V))
379      if (this->isValue(CI->getValue())) {
380        Res = &CI->getValue();
381        return true;
382      }
383    if (V->getType()->isVectorTy())
384      if (const auto *C = dyn_cast<Constant>(V))
385        if (auto *CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue()))
386          if (this->isValue(CI->getValue())) {
387            Res = &CI->getValue();
388            return true;
389          }
390 
391    return false;
392  }
393};
394 
395/// This helper class is used to match scalar and vector constants that
396/// satisfy a specified predicate, and bind them to an APFloat.
397/// Undefs are allowed in splat vector constants.
398template <typename Predicate> struct apf_pred_ty : public Predicate {
399  const APFloat *&Res;
400 
401  apf_pred_ty(const APFloat *&R) : Res(R) {}
402 
403  template <typename ITy> bool match(ITy *V) {
404    if (const auto *CI = dyn_cast<ConstantFP>(V))
405      if (this->isValue(CI->getValue())) {
406        Res = &CI->getValue();
407        return true;
408      }
409    if (V->getType()->isVectorTy())
410      if (const auto *C = dyn_cast<Constant>(V))
411        if (auto *CI = dyn_cast_or_null<ConstantFP>(
412                C->getSplatValue(/* AllowUndef */ true)))
413          if (this->isValue(CI->getValue())) {
414            Res = &CI->getValue();
415            return true;
416          }
417 
418    return false;
419  }
420};
421 
422///////////////////////////////////////////////////////////////////////////////
423//
424// Encapsulate constant value queries for use in templated predicate matchers.
425// This allows checking if constants match using compound predicates and works
426// with vector constants, possibly with relaxed constraints. For example, ignore
427// undef values.
428//
429///////////////////////////////////////////////////////////////////////////////
430 
431struct is_any_apint {
432  bool isValue(const APInt &C) { return true; }
433};
434/// Match an integer or vector with any integral constant.
435/// For vectors, this includes constants with undefined elements.
436inline cst_pred_ty<is_any_apint> m_AnyIntegralConstant() {
437  return cst_pred_ty<is_any_apint>();
438}
439 
440struct is_all_ones {
441  bool isValue(const APInt &C) { return C.isAllOnesValue(); }
442};
443/// Match an integer or vector with all bits set.
444/// For vectors, this includes constants with undefined elements.
445inline cst_pred_ty<is_all_ones> m_AllOnes() {
446  return cst_pred_ty<is_all_ones>();
447}
448 
449struct is_maxsignedvalue {
450  bool isValue(const APInt &C) { return C.isMaxSignedValue(); }
451};
452/// Match an integer or vector with values having all bits except for the high
453/// bit set (0x7f...).
454/// For vectors, this includes constants with undefined elements.
455inline cst_pred_ty<is_maxsignedvalue> m_MaxSignedValue() {
456  return cst_pred_ty<is_maxsignedvalue>();
457}
458inline api_pred_ty<is_maxsignedvalue> m_MaxSignedValue(const APInt *&V) {
459  return V;
460}
461 
462struct is_negative {
463  bool isValue(const APInt &C) { return C.isNegative(); }
464};
465/// Match an integer or vector of negative values.
466/// For vectors, this includes constants with undefined elements.
467inline cst_pred_ty<is_negative> m_Negative() {
468  return cst_pred_ty<is_negative>();
469}
470inline api_pred_ty<is_negative> m_Negative(const APInt *&V) {
471  return V;
472}
473 
474struct is_nonnegative {
475  bool isValue(const APInt &C) { return C.isNonNegative(); }
476};
477/// Match an integer or vector of non-negative values.
478/// For vectors, this includes constants with undefined elements.
479inline cst_pred_ty<is_nonnegative> m_NonNegative() {
480  return cst_pred_ty<is_nonnegative>();
481}
482inline api_pred_ty<is_nonnegative> m_NonNegative(const APInt *&V) {
483  return V;
484}
485 
486struct is_strictlypositive {
487  bool isValue(const APInt &C) { return C.isStrictlyPositive(); }
488};
489/// Match an integer or vector of strictly positive values.
490/// For vectors, this includes constants with undefined elements.
491inline cst_pred_ty<is_strictlypositive> m_StrictlyPositive() {
492  return cst_pred_ty<is_strictlypositive>();
493}
494inline api_pred_ty<is_strictlypositive> m_StrictlyPositive(const APInt *&V) {
495  return V;
496}
497 
498struct is_nonpositive {
499  bool isValue(const APInt &C) { return C.isNonPositive(); }
500};
501/// Match an integer or vector of non-positive values.
502/// For vectors, this includes constants with undefined elements.
503inline cst_pred_ty<is_nonpositive> m_NonPositive() {
504  return cst_pred_ty<is_nonpositive>();
505}
506inline api_pred_ty<is_nonpositive> m_NonPositive(const APInt *&V) { return V; }
507 
508struct is_one {
509  bool isValue(const APInt &C) { return C.isOneValue(); }
510};
511/// Match an integer 1 or a vector with all elements equal to 1.
512/// For vectors, this includes constants with undefined elements.
513inline cst_pred_ty<is_one> m_One() {
514  return cst_pred_ty<is_one>();
515}
516 
517struct is_zero_int {
518  bool isValue(const APInt &C) { return C.isNullValue(); }
519};
520/// Match an integer 0 or a vector with all elements equal to 0.
521/// For vectors, this includes constants with undefined elements.
522inline cst_pred_ty<is_zero_int> m_ZeroInt() {
523  return cst_pred_ty<is_zero_int>();
524}
525 
526struct is_zero {
527  template <typename ITy> bool match(ITy *V) {
528    auto *C = dyn_cast<Constant>(V);
529    // FIXME: this should be able to do something for scalable vectors
530    return C && (C->isNullValue() || cst_pred_ty<is_zero_int>().match(C));
531  }
532};
533/// Match any null constant or a vector with all elements equal to 0.
534/// For vectors, this includes constants with undefined elements.
535inline is_zero m_Zero() {
536  return is_zero();
537}
538 
539struct is_power2 {
540  bool isValue(const APInt &C) { return C.isPowerOf2(); }
541};
542/// Match an integer or vector power-of-2.
543/// For vectors, this includes constants with undefined elements.
544inline cst_pred_ty<is_power2> m_Power2() {
545  return cst_pred_ty<is_power2>();
546}
547inline api_pred_ty<is_power2> m_Power2(const APInt *&V) {
548  return V;
549}
550 
551struct is_negated_power2 {
552  bool isValue(const APInt &C) { return (-C).isPowerOf2(); }
553};
554/// Match a integer or vector negated power-of-2.
555/// For vectors, this includes constants with undefined elements.
556inline cst_pred_ty<is_negated_power2> m_NegatedPower2() {
557  return cst_pred_ty<is_negated_power2>();
558}
559inline api_pred_ty<is_negated_power2> m_NegatedPower2(const APInt *&V) {
560  return V;
561}
562 
563struct is_power2_or_zero {
564  bool isValue(const APInt &C) { return !C || C.isPowerOf2(); }
565};
566/// Match an integer or vector of 0 or power-of-2 values.
567/// For vectors, this includes constants with undefined elements.
568inline cst_pred_ty<is_power2_or_zero> m_Power2OrZero() {
569  return cst_pred_ty<is_power2_or_zero>();
570}
571inline api_pred_ty<is_power2_or_zero> m_Power2OrZero(const APInt *&V) {
572  return V;
573}
574 
575struct is_sign_mask {
576  bool isValue(const APInt &C) { return C.isSignMask(); }
577};
578/// Match an integer or vector with only the sign bit(s) set.
579/// For vectors, this includes constants with undefined elements.
580inline cst_pred_ty<is_sign_mask> m_SignMask() {
581  return cst_pred_ty<is_sign_mask>();
582}
583 
584struct is_lowbit_mask {
585  bool isValue(const APInt &C) { return C.isMask(); }
586};
587/// Match an integer or vector with only the low bit(s) set.
588/// For vectors, this includes constants with undefined elements.
589inline cst_pred_ty<is_lowbit_mask> m_LowBitMask() {
590  return cst_pred_ty<is_lowbit_mask>();
591}
592 
593struct icmp_pred_with_threshold {
594  ICmpInst::Predicate Pred;
595  const APInt *Thr;
596  bool isValue(const APInt &C) {
597    switch (Pred) {
598    case ICmpInst::Predicate::ICMP_EQ:
599      return C.eq(*Thr);
600    case ICmpInst::Predicate::ICMP_NE:
601      return C.ne(*Thr);
602    case ICmpInst::Predicate::ICMP_UGT:
603      return C.ugt(*Thr);
604    case ICmpInst::Predicate::ICMP_UGE:
605      return C.uge(*Thr);
606    case ICmpInst::Predicate::ICMP_ULT:
607      return C.ult(*Thr);
608    case ICmpInst::Predicate::ICMP_ULE:
609      return C.ule(*Thr);
610    case ICmpInst::Predicate::ICMP_SGT:
611      return C.sgt(*Thr);
612    case ICmpInst::Predicate::ICMP_SGE:
613      return C.sge(*Thr);
614    case ICmpInst::Predicate::ICMP_SLT:
615      return C.slt(*Thr);
616    case ICmpInst::Predicate::ICMP_SLE:
617      return C.sle(*Thr);
618    default:
619      llvm_unreachable("Unhandled ICmp predicate")::llvm::llvm_unreachable_internal("Unhandled ICmp predicate",
 "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/include/llvm/IR/PatternMatch.h"
, 619);
620    }
621  }
622};
623/// Match an integer or vector with every element comparing 'pred' (eg/ne/...)
624/// to Threshold. For vectors, this includes constants with undefined elements.
625inline cst_pred_ty<icmp_pred_with_threshold>
626m_SpecificInt_ICMP(ICmpInst::Predicate Predicate, const APInt &Threshold) {
627  cst_pred_ty<icmp_pred_with_threshold> P;
628  P.Pred = Predicate;
629  P.Thr = &Threshold;
630  return P;
631}
632 
633struct is_nan {
634  bool isValue(const APFloat &C) { return C.isNaN(); }
635};
636/// Match an arbitrary NaN constant. This includes quiet and signalling nans.
637/// For vectors, this includes constants with undefined elements.
638inline cstfp_pred_ty<is_nan> m_NaN() {
639  return cstfp_pred_ty<is_nan>();
640}
641 
642struct is_nonnan {
643  bool isValue(const APFloat &C) { return !C.isNaN(); }
644};
645/// Match a non-NaN FP constant.
646/// For vectors, this includes constants with undefined elements.
647inline cstfp_pred_ty<is_nonnan> m_NonNaN() {
648  return cstfp_pred_ty<is_nonnan>();
649}
650 
651struct is_inf {
652  bool isValue(const APFloat &C) { return C.isInfinity(); }
653};
654/// Match a positive or negative infinity FP constant.
655/// For vectors, this includes constants with undefined elements.
656inline cstfp_pred_ty<is_inf> m_Inf() {
657  return cstfp_pred_ty<is_inf>();
658}
659 
660struct is_noninf {
661  bool isValue(const APFloat &C) { return !C.isInfinity(); }
662};
663/// Match a non-infinity FP constant, i.e. finite or NaN.
664/// For vectors, this includes constants with undefined elements.
665inline cstfp_pred_ty<is_noninf> m_NonInf() {
666  return cstfp_pred_ty<is_noninf>();
667}
668 
669struct is_finite {
670  bool isValue(const APFloat &C) { return C.isFinite(); }
671};
672/// Match a finite FP constant, i.e. not infinity or NaN.
673/// For vectors, this includes constants with undefined elements.
674inline cstfp_pred_ty<is_finite> m_Finite() {
675  return cstfp_pred_ty<is_finite>();
676}
677inline apf_pred_ty<is_finite> m_Finite(const APFloat *&V) { return V; }
678 
679struct is_finitenonzero {
680  bool isValue(const APFloat &C) { return C.isFiniteNonZero(); }
681};
682/// Match a finite non-zero FP constant.
683/// For vectors, this includes constants with undefined elements.
684inline cstfp_pred_ty<is_finitenonzero> m_FiniteNonZero() {
685  return cstfp_pred_ty<is_finitenonzero>();
686}
687inline apf_pred_ty<is_finitenonzero> m_FiniteNonZero(const APFloat *&V) {
688  return V;
689}
690 
691struct is_any_zero_fp {
692  bool isValue(const APFloat &C) { return C.isZero(); }
693};
694/// Match a floating-point negative zero or positive zero.
695/// For vectors, this includes constants with undefined elements.
696inline cstfp_pred_ty<is_any_zero_fp> m_AnyZeroFP() {
697  return cstfp_pred_ty<is_any_zero_fp>();
698}
699 
700struct is_pos_zero_fp {
701  bool isValue(const APFloat &C) { return C.isPosZero(); }
702};
703/// Match a floating-point positive zero.
704/// For vectors, this includes constants with undefined elements.
705inline cstfp_pred_ty<is_pos_zero_fp> m_PosZeroFP() {
706  return cstfp_pred_ty<is_pos_zero_fp>();
707}
708 
709struct is_neg_zero_fp {
710  bool isValue(const APFloat &C) { return C.isNegZero(); }
711};
712/// Match a floating-point negative zero.
713/// For vectors, this includes constants with undefined elements.
714inline cstfp_pred_ty<is_neg_zero_fp> m_NegZeroFP() {
715  return cstfp_pred_ty<is_neg_zero_fp>();
716}
717 
718struct is_non_zero_fp {
719  bool isValue(const APFloat &C) { return C.isNonZero(); }
720};
721/// Match a floating-point non-zero.
722/// For vectors, this includes constants with undefined elements.
723inline cstfp_pred_ty<is_non_zero_fp> m_NonZeroFP() {
724  return cstfp_pred_ty<is_non_zero_fp>();
725}
726 
727///////////////////////////////////////////////////////////////////////////////
728 
729template <typename Class> struct bind_ty {
730  Class *&VR;
731 
732  bind_ty(Class *&V) : VR(V) {}
733 
734  template <typename ITy> bool match(ITy *V) {
735    if (auto *CV = dyn_cast<Class>(V)) {
736      VR = CV;
737      return true;
738    }
739    return false;
740  }
741};
742 
743/// Match a value, capturing it if we match.
744inline bind_ty<Value> m_Value(Value *&V) { return V; }
6
←
Calling constructor for 'bind_ty<llvm::Value>'→
7
←
Returning from constructor for 'bind_ty<llvm::Value>'→
8
←
Returning without writing to 'V'→
44
←
Returning without writing to 'V'→
745inline bind_ty<const Value> m_Value(const Value *&V) { return V; }
746 
747/// Match an instruction, capturing it if we match.
748inline bind_ty<Instruction> m_Instruction(Instruction *&I) { return I; }
749/// Match a unary operator, capturing it if we match.
750inline bind_ty<UnaryOperator> m_UnOp(UnaryOperator *&I) { return I; }
751/// Match a binary operator, capturing it if we match.
752inline bind_ty<BinaryOperator> m_BinOp(BinaryOperator *&I) { return I; }
753/// Match a with overflow intrinsic, capturing it if we match.
754inline bind_ty<WithOverflowInst> m_WithOverflowInst(WithOverflowInst *&I) { return I; }
755inline bind_ty<const WithOverflowInst>
756m_WithOverflowInst(const WithOverflowInst *&I) {
757  return I;
758}
759 
760/// Match a Constant, capturing the value if we match.
761inline bind_ty<Constant> m_Constant(Constant *&C) { return C; }
762 
763/// Match a ConstantInt, capturing the value if we match.
764inline bind_ty<ConstantInt> m_ConstantInt(ConstantInt *&CI) { return CI; }
765 
766/// Match a ConstantFP, capturing the value if we match.
767inline bind_ty<ConstantFP> m_ConstantFP(ConstantFP *&C) { return C; }
768 
769/// Match a ConstantExpr, capturing the value if we match.
770inline bind_ty<ConstantExpr> m_ConstantExpr(ConstantExpr *&C) { return C; }
771 
772/// Match a basic block value, capturing it if we match.
773inline bind_ty<BasicBlock> m_BasicBlock(BasicBlock *&V) { return V; }
774inline bind_ty<const BasicBlock> m_BasicBlock(const BasicBlock *&V) {
775  return V;
776}
777 
778/// Match an arbitrary immediate Constant and ignore it.
779inline match_combine_and<class_match<Constant>,
780                         match_unless<class_match<ConstantExpr>>>
781m_ImmConstant() {
782  return m_CombineAnd(m_Constant(), m_Unless(m_ConstantExpr()));
783}
784 
785/// Match an immediate Constant, capturing the value if we match.
786inline match_combine_and<bind_ty<Constant>,
787                         match_unless<class_match<ConstantExpr>>>
788m_ImmConstant(Constant *&C) {
789  return m_CombineAnd(m_Constant(C), m_Unless(m_ConstantExpr()));
790}
791 
792/// Match a specified Value*.
793struct specificval_ty {
794  const Value *Val;
795 
796  specificval_ty(const Value *V) : Val(V) {}
797 
798  template <typename ITy> bool match(ITy *V) { return V == Val; }
799};
800 
801/// Match if we have a specific specified value.
802inline specificval_ty m_Specific(const Value *V) { return V; }
803 
804/// Stores a reference to the Value *, not the Value * itself,
805/// thus can be used in commutative matchers.
806template <typename Class> struct deferredval_ty {
807  Class *const &Val;
808 
809  deferredval_ty(Class *const &V) : Val(V) {}
810 
811  template <typename ITy> bool match(ITy *const V) { return V == Val; }
812};
813 
814/// Like m_Specific(), but works if the specific value to match is determined
815/// as part of the same match() expression. For example:
816/// m_Add(m_Value(X), m_Specific(X)) is incorrect, because m_Specific() will
817/// bind X before the pattern match starts.
818/// m_Add(m_Value(X), m_Deferred(X)) is correct, and will check against
819/// whichever value m_Value(X) populated.
820inline deferredval_ty<Value> m_Deferred(Value *const &V) { return V; }
821inline deferredval_ty<const Value> m_Deferred(const Value *const &V) {
822  return V;
823}
824 
825/// Match a specified floating point value or vector of all elements of
826/// that value.
827struct specific_fpval {
828  double Val;
829 
830  specific_fpval(double V) : Val(V) {}
831 
832  template <typename ITy> bool match(ITy *V) {
833    if (const auto *CFP = dyn_cast<ConstantFP>(V))
834      return CFP->isExactlyValue(Val);
835    if (V->getType()->isVectorTy())
836      if (const auto *C = dyn_cast<Constant>(V))
837        if (auto *CFP = dyn_cast_or_null<ConstantFP>(C->getSplatValue()))
838          return CFP->isExactlyValue(Val);
839    return false;
840  }
841};
842 
843/// Match a specific floating point value or vector with all elements
844/// equal to the value.
845inline specific_fpval m_SpecificFP(double V) { return specific_fpval(V); }
846 
847/// Match a float 1.0 or vector with all elements equal to 1.0.
848inline specific_fpval m_FPOne() { return m_SpecificFP(1.0); }
849 
850struct bind_const_intval_ty {
851  uint64_t &VR;
852 
853  bind_const_intval_ty(uint64_t &V) : VR(V) {}
854 
855  template <typename ITy> bool match(ITy *V) {
856    if (const auto *CV = dyn_cast<ConstantInt>(V))
857      if (CV->getValue().ule(UINT64_MAX(18446744073709551615UL))) {
858        VR = CV->getZExtValue();
859        return true;
860      }
861    return false;
862  }
863};
864 
865/// Match a specified integer value or vector of all elements of that
866/// value.
867template <bool AllowUndefs>
868struct specific_intval {
869  APInt Val;
870 
871  specific_intval(APInt V) : Val(std::move(V)) {}
872 
873  template <typename ITy> bool match(ITy *V) {
874    const auto *CI = dyn_cast<ConstantInt>(V);
875    if (!CI && V->getType()->isVectorTy())
876      if (const auto *C = dyn_cast<Constant>(V))
877        CI = dyn_cast_or_null<ConstantInt>(C->getSplatValue(AllowUndefs));
878 
879    return CI && APInt::isSameValue(CI->getValue(), Val);
880  }
881};
882 
883/// Match a specific integer value or vector with all elements equal to
884/// the value.
885inline specific_intval<false> m_SpecificInt(APInt V) {
886  return specific_intval<false>(std::move(V));
887}
888 
889inline specific_intval<false> m_SpecificInt(uint64_t V) {
890  return m_SpecificInt(APInt(64, V));
891}
892 
893inline specific_intval<true> m_SpecificIntAllowUndef(APInt V) {
894  return specific_intval<true>(std::move(V));
895}
896 
897inline specific_intval<true> m_SpecificIntAllowUndef(uint64_t V) {
898  return m_SpecificIntAllowUndef(APInt(64, V));
899}
900 
901/// Match a ConstantInt and bind to its value.  This does not match
902/// ConstantInts wider than 64-bits.
903inline bind_const_intval_ty m_ConstantInt(uint64_t &V) { return V; }
904 
905/// Match a specified basic block value.
906struct specific_bbval {
907  BasicBlock *Val;
908 
909  specific_bbval(BasicBlock *Val) : Val(Val) {}
910 
911  template <typename ITy> bool match(ITy *V) {
912    const auto *BB = dyn_cast<BasicBlock>(V);
913    return BB && BB == Val;
914  }
915};
916 
917/// Match a specific basic block value.
918inline specific_bbval m_SpecificBB(BasicBlock *BB) {
919  return specific_bbval(BB);
920}
921 
922/// A commutative-friendly version of m_Specific().
923inline deferredval_ty<BasicBlock> m_Deferred(BasicBlock *const &BB) {
924  return BB;
925}
926inline deferredval_ty<const BasicBlock>
927m_Deferred(const BasicBlock *const &BB) {
928  return BB;
929}
930 
931//===----------------------------------------------------------------------===//
932// Matcher for any binary operator.
933//
934template <typename LHS_t, typename RHS_t, bool Commutable = false>
935struct AnyBinaryOp_match {
936  LHS_t L;
937  RHS_t R;
938 
939  // The evaluation order is always stable, regardless of Commutability.
940  // The LHS is always matched first.
941  AnyBinaryOp_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {}
942 
943  template <typename OpTy> bool match(OpTy *V) {
944    if (auto *I15.1
'I' is non-null
15.1
'I' is non-null
 = dyn_cast<BinaryOperator>(V))
15
←
Assuming 'V' is a 'BinaryOperator'→
945      return (L.match(I->getOperand(0)) && R.match(I->getOperand(1))) ||
16
←
Calling 'match_combine_or::match'→
31
←
Returning from 'match_combine_or::match'→
32
←
Returning the value 1, which participates in a condition later→
946             (Commutable && L.match(I->getOperand(1)) &&
947              R.match(I->getOperand(0)));
948    return false;
949  }
950};
951 
952template <typename LHS, typename RHS>
953inline AnyBinaryOp_match<LHS, RHS> m_BinOp(const LHS &L, const RHS &R) {
954  return AnyBinaryOp_match<LHS, RHS>(L, R);
955}
956 
957//===----------------------------------------------------------------------===//
958// Matcher for any unary operator.
959// TODO fuse unary, binary matcher into n-ary matcher
960//
961template <typename OP_t> struct AnyUnaryOp_match {
962  OP_t X;
963 
964  AnyUnaryOp_match(const OP_t &X) : X(X) {}
965 
966  template <typename OpTy> bool match(OpTy *V) {
967    if (auto *I = dyn_cast<UnaryOperator>(V))
968      return X.match(I->getOperand(0));
969    return false;
970  }
971};
972 
973template <typename OP_t> inline AnyUnaryOp_match<OP_t> m_UnOp(const OP_t &X) {
974  return AnyUnaryOp_match<OP_t>(X);
975}
976 
977//===----------------------------------------------------------------------===//
978// Matchers for specific binary operators.
979//
980 
981template <typename LHS_t, typename RHS_t, unsigned Opcode,
982          bool Commutable = false>
983struct BinaryOp_match {
984  LHS_t L;
985  RHS_t R;
986 
987  // The evaluation order is always stable, regardless of Commutability.
988  // The LHS is always matched first.
989  BinaryOp_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {}
990 
991  template <typename OpTy> bool match(OpTy *V) {
992    if (V->getValueID() == Value::InstructionVal + Opcode) {
20
←
Assuming the condition is true→
21
←
Taking true branch→
993      auto *I = cast<BinaryOperator>(V);
22
←
'V' is a 'BinaryOperator'→
994      return (L.match(I->getOperand(0)) && R.match(I->getOperand(1))) ||
23
←
Value assigned to 'Select'→
24
←
Assuming the condition is true→
25
←
Assuming the condition is true→
995             (Commutable && L.match(I->getOperand(1)) &&
996              R.match(I->getOperand(0)));
997    }
998    if (auto *CE = dyn_cast<ConstantExpr>(V))
999      return CE->getOpcode() == Opcode &&
1000             ((L.match(CE->getOperand(0)) && R.match(CE->getOperand(1))) ||
1001              (Commutable && L.match(CE->getOperand(1)) &&
1002               R.match(CE->getOperand(0))));
1003    return false;
1004  }
1005};
1006 
1007template <typename LHS, typename RHS>
1008inline BinaryOp_match<LHS, RHS, Instruction::Add> m_Add(const LHS &L,
1009                                                        const RHS &R) {
1010  return BinaryOp_match<LHS, RHS, Instruction::Add>(L, R);
1011}
1012 
1013template <typename LHS, typename RHS>
1014inline BinaryOp_match<LHS, RHS, Instruction::FAdd> m_FAdd(const LHS &L,
1015                                                          const RHS &R) {
1016  return BinaryOp_match<LHS, RHS, Instruction::FAdd>(L, R);
1017}
1018 
1019template <typename LHS, typename RHS>
1020inline BinaryOp_match<LHS, RHS, Instruction::Sub> m_Sub(const LHS &L,
1021                                                        const RHS &R) {
1022  return BinaryOp_match<LHS, RHS, Instruction::Sub>(L, R);
1023}
1024 
1025template <typename LHS, typename RHS>
1026inline BinaryOp_match<LHS, RHS, Instruction::FSub> m_FSub(const LHS &L,
1027                                                          const RHS &R) {
1028  return BinaryOp_match<LHS, RHS, Instruction::FSub>(L, R);
1029}
1030 
1031template <typename Op_t> struct FNeg_match {
1032  Op_t X;
1033 
1034  FNeg_match(const Op_t &Op) : X(Op) {}
1035  template <typename OpTy> bool match(OpTy *V) {
1036    auto *FPMO = dyn_cast<FPMathOperator>(V);
1037    if (!FPMO) return false;
1038 
1039    if (FPMO->getOpcode() == Instruction::FNeg)
1040      return X.match(FPMO->getOperand(0));
1041 
1042    if (FPMO->getOpcode() == Instruction::FSub) {
1043      if (FPMO->hasNoSignedZeros()) {
1044        // With 'nsz', any zero goes.
1045        if (!cstfp_pred_ty<is_any_zero_fp>().match(FPMO->getOperand(0)))
1046          return false;
1047      } else {
1048        // Without 'nsz', we need fsub -0.0, X exactly.
1049        if (!cstfp_pred_ty<is_neg_zero_fp>().match(FPMO->getOperand(0)))
1050          return false;
1051      }
1052 
1053      return X.match(FPMO->getOperand(1));
1054    }
1055 
1056    return false;
1057  }
1058};
1059 
1060/// Match 'fneg X' as 'fsub -0.0, X'.
1061template <typename OpTy>
1062inline FNeg_match<OpTy>
1063m_FNeg(const OpTy &X) {
1064  return FNeg_match<OpTy>(X);
1065}
1066 
1067/// Match 'fneg X' as 'fsub +-0.0, X'.
1068template <typename RHS>
1069inline BinaryOp_match<cstfp_pred_ty<is_any_zero_fp>, RHS, Instruction::FSub>
1070m_FNegNSZ(const RHS &X) {
1071  return m_FSub(m_AnyZeroFP(), X);
1072}
1073 
1074template <typename LHS, typename RHS>
1075inline BinaryOp_match<LHS, RHS, Instruction::Mul> m_Mul(const LHS &L,
1076                                                        const RHS &R) {
1077  return BinaryOp_match<LHS, RHS, Instruction::Mul>(L, R);
1078}
1079 
1080template <typename LHS, typename RHS>
1081inline BinaryOp_match<LHS, RHS, Instruction::FMul> m_FMul(const LHS &L,
1082                                                          const RHS &R) {
1083  return BinaryOp_match<LHS, RHS, Instruction::FMul>(L, R);
1084}
1085 
1086template <typename LHS, typename RHS>
1087inline BinaryOp_match<LHS, RHS, Instruction::UDiv> m_UDiv(const LHS &L,
1088                                                          const RHS &R) {
1089  return BinaryOp_match<LHS, RHS, Instruction::UDiv>(L, R);
1090}
1091 
1092template <typename LHS, typename RHS>
1093inline BinaryOp_match<LHS, RHS, Instruction::SDiv> m_SDiv(const LHS &L,
1094                                                          const RHS &R) {
1095  return BinaryOp_match<LHS, RHS, Instruction::SDiv>(L, R);
1096}
1097 
1098template <typename LHS, typename RHS>
1099inline BinaryOp_match<LHS, RHS, Instruction::FDiv> m_FDiv(const LHS &L,
1100                                                          const RHS &R) {
1101  return BinaryOp_match<LHS, RHS, Instruction::FDiv>(L, R);
1102}
1103 
1104template <typename LHS, typename RHS>
1105inline BinaryOp_match<LHS, RHS, Instruction::URem> m_URem(const LHS &L,
1106                                                          const RHS &R) {
1107  return BinaryOp_match<LHS, RHS, Instruction::URem>(L, R);
1108}
1109 
1110template <typename LHS, typename RHS>
1111inline BinaryOp_match<LHS, RHS, Instruction::SRem> m_SRem(const LHS &L,
1112                                                          const RHS &R) {
1113  return BinaryOp_match<LHS, RHS, Instruction::SRem>(L, R);
1114}
1115 
1116template <typename LHS, typename RHS>
1117inline BinaryOp_match<LHS, RHS, Instruction::FRem> m_FRem(const LHS &L,
1118                                                          const RHS &R) {
1119  return BinaryOp_match<LHS, RHS, Instruction::FRem>(L, R);
1120}
1121 
1122template <typename LHS, typename RHS>
1123inline BinaryOp_match<LHS, RHS, Instruction::And> m_And(const LHS &L,
1124                                                        const RHS &R) {
1125  return BinaryOp_match<LHS, RHS, Instruction::And>(L, R);
1126}
1127 
1128template <typename LHS, typename RHS>
1129inline BinaryOp_match<LHS, RHS, Instruction::Or> m_Or(const LHS &L,
1130                                                      const RHS &R) {
1131  return BinaryOp_match<LHS, RHS, Instruction::Or>(L, R);
1132}
1133 
1134template <typename LHS, typename RHS>
1135inline BinaryOp_match<LHS, RHS, Instruction::Xor> m_Xor(const LHS &L,
1136                                                        const RHS &R) {
1137  return BinaryOp_match<LHS, RHS, Instruction::Xor>(L, R);
1138}
1139 
1140template <typename LHS, typename RHS>
1141inline BinaryOp_match<LHS, RHS, Instruction::Shl> m_Shl(const LHS &L,
1142                                                        const RHS &R) {
1143  return BinaryOp_match<LHS, RHS, Instruction::Shl>(L, R);
1144}
1145 
1146template <typename LHS, typename RHS>
1147inline BinaryOp_match<LHS, RHS, Instruction::LShr> m_LShr(const LHS &L,
1148                                                          const RHS &R) {
1149  return BinaryOp_match<LHS, RHS, Instruction::LShr>(L, R);
1150}
1151 
1152template <typename LHS, typename RHS>
1153inline BinaryOp_match<LHS, RHS, Instruction::AShr> m_AShr(const LHS &L,
1154                                                          const RHS &R) {
1155  return BinaryOp_match<LHS, RHS, Instruction::AShr>(L, R);
1156}
1157 
1158template <typename LHS_t, typename RHS_t, unsigned Opcode,
1159          unsigned WrapFlags = 0>
1160struct OverflowingBinaryOp_match {
1161  LHS_t L;
1162  RHS_t R;
1163 
1164  OverflowingBinaryOp_match(const LHS_t &LHS, const RHS_t &RHS)
1165      : L(LHS), R(RHS) {}
1166 
1167  template <typename OpTy> bool match(OpTy *V) {
1168    if (auto *Op = dyn_cast<OverflowingBinaryOperator>(V)) {
1169      if (Op->getOpcode() != Opcode)
1170        return false;
1171      if ((WrapFlags & OverflowingBinaryOperator::NoUnsignedWrap) &&
1172          !Op->hasNoUnsignedWrap())
1173        return false;
1174      if ((WrapFlags & OverflowingBinaryOperator::NoSignedWrap) &&
1175          !Op->hasNoSignedWrap())
1176        return false;
1177      return L.match(Op->getOperand(0)) && R.match(Op->getOperand(1));
1178    }
1179    return false;
1180  }
1181};
1182 
1183template <typename LHS, typename RHS>
1184inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Add,
1185                                 OverflowingBinaryOperator::NoSignedWrap>
1186m_NSWAdd(const LHS &L, const RHS &R) {
1187  return OverflowingBinaryOp_match<LHS, RHS, Instruction::Add,
1188                                   OverflowingBinaryOperator::NoSignedWrap>(
1189      L, R);
1190}
1191template <typename LHS, typename RHS>
1192inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub,
1193                                 OverflowingBinaryOperator::NoSignedWrap>
1194m_NSWSub(const LHS &L, const RHS &R) {
1195  return OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub,
1196                                   OverflowingBinaryOperator::NoSignedWrap>(
1197      L, R);
1198}
1199template <typename LHS, typename RHS>
1200inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul,
1201                                 OverflowingBinaryOperator::NoSignedWrap>
1202m_NSWMul(const LHS &L, const RHS &R) {
1203  return OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul,
1204                                   OverflowingBinaryOperator::NoSignedWrap>(
1205      L, R);
1206}
1207template <typename LHS, typename RHS>
1208inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl,
1209                                 OverflowingBinaryOperator::NoSignedWrap>
1210m_NSWShl(const LHS &L, const RHS &R) {
1211  return OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl,
1212                                   OverflowingBinaryOperator::NoSignedWrap>(
1213      L, R);
1214}
1215 
1216template <typename LHS, typename RHS>
1217inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Add,
1218                                 OverflowingBinaryOperator::NoUnsignedWrap>
1219m_NUWAdd(const LHS &L, const RHS &R) {
1220  return OverflowingBinaryOp_match<LHS, RHS, Instruction::Add,
1221                                   OverflowingBinaryOperator::NoUnsignedWrap>(
1222      L, R);
1223}
1224template <typename LHS, typename RHS>
1225inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub,
1226                                 OverflowingBinaryOperator::NoUnsignedWrap>
1227m_NUWSub(const LHS &L, const RHS &R) {
1228  return OverflowingBinaryOp_match<LHS, RHS, Instruction::Sub,
1229                                   OverflowingBinaryOperator::NoUnsignedWrap>(
1230      L, R);
1231}
1232template <typename LHS, typename RHS>
1233inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul,
1234                                 OverflowingBinaryOperator::NoUnsignedWrap>
1235m_NUWMul(const LHS &L, const RHS &R) {
1236  return OverflowingBinaryOp_match<LHS, RHS, Instruction::Mul,
1237                                   OverflowingBinaryOperator::NoUnsignedWrap>(
1238      L, R);
1239}
1240template <typename LHS, typename RHS>
1241inline OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl,
1242                                 OverflowingBinaryOperator::NoUnsignedWrap>
1243m_NUWShl(const LHS &L, const RHS &R) {
1244  return OverflowingBinaryOp_match<LHS, RHS, Instruction::Shl,
1245                                   OverflowingBinaryOperator::NoUnsignedWrap>(
1246      L, R);
1247}
1248 
1249//===----------------------------------------------------------------------===//
1250// Class that matches a group of binary opcodes.
1251//
1252template <typename LHS_t, typename RHS_t, typename Predicate>
1253struct BinOpPred_match : Predicate {
1254  LHS_t L;
1255  RHS_t R;
1256 
1257  BinOpPred_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {}
1258 
1259  template <typename OpTy> bool match(OpTy *V) {
1260    if (auto *I = dyn_cast<Instruction>(V))
1261      return this->isOpType(I->getOpcode()) && L.match(I->getOperand(0)) &&
1262             R.match(I->getOperand(1));
1263    if (auto *CE = dyn_cast<ConstantExpr>(V))
1264      return this->isOpType(CE->getOpcode()) && L.match(CE->getOperand(0)) &&
1265             R.match(CE->getOperand(1));
1266    return false;
1267  }
1268};
1269 
1270struct is_shift_op {
1271  bool isOpType(unsigned Opcode) { return Instruction::isShift(Opcode); }
1272};
1273 
1274struct is_right_shift_op {
1275  bool isOpType(unsigned Opcode) {
1276    return Opcode == Instruction::LShr || Opcode == Instruction::AShr;
1277  }
1278};
1279 
1280struct is_logical_shift_op {
1281  bool isOpType(unsigned Opcode) {
1282    return Opcode == Instruction::LShr || Opcode == Instruction::Shl;
1283  }
1284};
1285 
1286struct is_bitwiselogic_op {
1287  bool isOpType(unsigned Opcode) {
1288    return Instruction::isBitwiseLogicOp(Opcode);
1289  }
1290};
1291 
1292struct is_idiv_op {
1293  bool isOpType(unsigned Opcode) {
1294    return Opcode == Instruction::SDiv || Opcode == Instruction::UDiv;
1295  }
1296};
1297 
1298struct is_irem_op {
1299  bool isOpType(unsigned Opcode) {
1300    return Opcode == Instruction::SRem || Opcode == Instruction::URem;
1301  }
1302};
1303 
1304/// Matches shift operations.
1305template <typename LHS, typename RHS>
1306inline BinOpPred_match<LHS, RHS, is_shift_op> m_Shift(const LHS &L,
1307                                                      const RHS &R) {
1308  return BinOpPred_match<LHS, RHS, is_shift_op>(L, R);
1309}
1310 
1311/// Matches logical shift operations.
1312template <typename LHS, typename RHS>
1313inline BinOpPred_match<LHS, RHS, is_right_shift_op> m_Shr(const LHS &L,
1314                                                          const RHS &R) {
1315  return BinOpPred_match<LHS, RHS, is_right_shift_op>(L, R);
1316}
1317 
1318/// Matches logical shift operations.
1319template <typename LHS, typename RHS>
1320inline BinOpPred_match<LHS, RHS, is_logical_shift_op>
1321m_LogicalShift(const LHS &L, const RHS &R) {
1322  return BinOpPred_match<LHS, RHS, is_logical_shift_op>(L, R);
1323}
1324 
1325/// Matches bitwise logic operations.
1326template <typename LHS, typename RHS>
1327inline BinOpPred_match<LHS, RHS, is_bitwiselogic_op>
1328m_BitwiseLogic(const LHS &L, const RHS &R) {
1329  return BinOpPred_match<LHS, RHS, is_bitwiselogic_op>(L, R);
1330}
1331 
1332/// Matches integer division operations.
1333template <typename LHS, typename RHS>
1334inline BinOpPred_match<LHS, RHS, is_idiv_op> m_IDiv(const LHS &L,
1335                                                    const RHS &R) {
1336  return BinOpPred_match<LHS, RHS, is_idiv_op>(L, R);
1337}
1338 
1339/// Matches integer remainder operations.
1340template <typename LHS, typename RHS>
1341inline BinOpPred_match<LHS, RHS, is_irem_op> m_IRem(const LHS &L,
1342                                                    const RHS &R) {
1343  return BinOpPred_match<LHS, RHS, is_irem_op>(L, R);
1344}
1345 
1346//===----------------------------------------------------------------------===//
1347// Class that matches exact binary ops.
1348//
1349template <typename SubPattern_t> struct Exact_match {
1350  SubPattern_t SubPattern;
1351 
1352  Exact_match(const SubPattern_t &SP) : SubPattern(SP) {}
1353 
1354  template <typename OpTy> bool match(OpTy *V) {
1355    if (auto *PEO = dyn_cast<PossiblyExactOperator>(V))
1356      return PEO->isExact() && SubPattern.match(V);
1357    return false;
1358  }
1359};
1360 
1361template <typename T> inline Exact_match<T> m_Exact(const T &SubPattern) {
1362  return SubPattern;
1363}
1364 
1365//===----------------------------------------------------------------------===//
1366// Matchers for CmpInst classes
1367//
1368 
1369template <typename LHS_t, typename RHS_t, typename Class, typename PredicateTy,
1370          bool Commutable = false>
1371struct CmpClass_match {
1372  PredicateTy &Predicate;
1373  LHS_t L;
1374  RHS_t R;
1375 
1376  // The evaluation order is always stable, regardless of Commutability.
1377  // The LHS is always matched first.
1378  CmpClass_match(PredicateTy &Pred, const LHS_t &LHS, const RHS_t &RHS)
1379      : Predicate(Pred), L(LHS), R(RHS) {}
1380 
1381  template <typename OpTy> bool match(OpTy *V) {
1382    if (auto *I = dyn_cast<Class>(V)) {
1383      if (L.match(I->getOperand(0)) && R.match(I->getOperand(1))) {
1384        Predicate = I->getPredicate();
1385        return true;
1386      } else if (Commutable && L.match(I->getOperand(1)) &&
1387           R.match(I->getOperand(0))) {
1388        Predicate = I->getSwappedPredicate();
1389        return true;
1390      }
1391    }
1392    return false;
1393  }
1394};
1395 
1396template <typename LHS, typename RHS>
1397inline CmpClass_match<LHS, RHS, CmpInst, CmpInst::Predicate>
1398m_Cmp(CmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
1399  return CmpClass_match<LHS, RHS, CmpInst, CmpInst::Predicate>(Pred, L, R);
1400}
1401 
1402template <typename LHS, typename RHS>
1403inline CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate>
1404m_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
1405  return CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate>(Pred, L, R);
1406}
1407 
1408template <typename LHS, typename RHS>
1409inline CmpClass_match<LHS, RHS, FCmpInst, FCmpInst::Predicate>
1410m_FCmp(FCmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
1411  return CmpClass_match<LHS, RHS, FCmpInst, FCmpInst::Predicate>(Pred, L, R);
1412}
1413 
1414//===----------------------------------------------------------------------===//
1415// Matchers for instructions with a given opcode and number of operands.
1416//
1417 
1418/// Matches instructions with Opcode and three operands.
1419template <typename T0, unsigned Opcode> struct OneOps_match {
1420  T0 Op1;
1421 
1422  OneOps_match(const T0 &Op1) : Op1(Op1) {}
1423 
1424  template <typename OpTy> bool match(OpTy *V) {
1425    if (V->getValueID() == Value::InstructionVal + Opcode) {
1426      auto *I = cast<Instruction>(V);
1427      return Op1.match(I->getOperand(0));
1428    }
1429    return false;
1430  }
1431};
1432 
1433/// Matches instructions with Opcode and three operands.
1434template <typename T0, typename T1, unsigned Opcode> struct TwoOps_match {
1435  T0 Op1;
1436  T1 Op2;
1437 
1438  TwoOps_match(const T0 &Op1, const T1 &Op2) : Op1(Op1), Op2(Op2) {}
1439 
1440  template <typename OpTy> bool match(OpTy *V) {
1441    if (V->getValueID() == Value::InstructionVal + Opcode) {
1442      auto *I = cast<Instruction>(V);
1443      return Op1.match(I->getOperand(0)) && Op2.match(I->getOperand(1));
1444    }
1445    return false;
1446  }
1447};
1448 
1449/// Matches instructions with Opcode and three operands.
1450template <typename T0, typename T1, typename T2, unsigned Opcode>
1451struct ThreeOps_match {
1452  T0 Op1;
1453  T1 Op2;
1454  T2 Op3;
1455 
1456  ThreeOps_match(const T0 &Op1, const T1 &Op2, const T2 &Op3)
1457      : Op1(Op1), Op2(Op2), Op3(Op3) {}
1458 
1459  template <typename OpTy> bool match(OpTy *V) {
1460    if (V->getValueID() == Value::InstructionVal + Opcode) {
64
←
Called C++ object pointer is null
1461      auto *I = cast<Instruction>(V);
1462      return Op1.match(I->getOperand(0)) && Op2.match(I->getOperand(1)) &&
1463             Op3.match(I->getOperand(2));
1464    }
1465    return false;
1466  }
1467};
1468 
1469/// Matches SelectInst.
1470template <typename Cond, typename LHS, typename RHS>
1471inline ThreeOps_match<Cond, LHS, RHS, Instruction::Select>
1472m_Select(const Cond &C, const LHS &L, const RHS &R) {
1473  return ThreeOps_match<Cond, LHS, RHS, Instruction::Select>(C, L, R);
1474}
1475 
1476/// This matches a select of two constants, e.g.:
1477/// m_SelectCst<-1, 0>(m_Value(V))
1478template <int64_t L, int64_t R, typename Cond>
1479inline ThreeOps_match<Cond, constantint_match<L>, constantint_match<R>,
1480                      Instruction::Select>
1481m_SelectCst(const Cond &C) {
1482  return m_Select(C, m_ConstantInt<L>(), m_ConstantInt<R>());
1483}
1484 
1485/// Matches FreezeInst.
1486template <typename OpTy>
1487inline OneOps_match<OpTy, Instruction::Freeze> m_Freeze(const OpTy &Op) {
1488  return OneOps_match<OpTy, Instruction::Freeze>(Op);
1489}
1490 
1491/// Matches InsertElementInst.
1492template <typename Val_t, typename Elt_t, typename Idx_t>
1493inline ThreeOps_match<Val_t, Elt_t, Idx_t, Instruction::InsertElement>
1494m_InsertElt(const Val_t &Val, const Elt_t &Elt, const Idx_t &Idx) {
1495  return ThreeOps_match<Val_t, Elt_t, Idx_t, Instruction::InsertElement>(
1496      Val, Elt, Idx);
1497}
1498 
1499/// Matches ExtractElementInst.
1500template <typename Val_t, typename Idx_t>
1501inline TwoOps_match<Val_t, Idx_t, Instruction::ExtractElement>
1502m_ExtractElt(const Val_t &Val, const Idx_t &Idx) {
1503  return TwoOps_match<Val_t, Idx_t, Instruction::ExtractElement>(Val, Idx);
1504}
1505 
1506/// Matches shuffle.
1507template <typename T0, typename T1, typename T2> struct Shuffle_match {
1508  T0 Op1;
1509  T1 Op2;
1510  T2 Mask;
1511 
1512  Shuffle_match(const T0 &Op1, const T1 &Op2, const T2 &Mask)
1513      : Op1(Op1), Op2(Op2), Mask(Mask) {}
1514 
1515  template <typename OpTy> bool match(OpTy *V) {
1516    if (auto *I = dyn_cast<ShuffleVectorInst>(V)) {
1517      return Op1.match(I->getOperand(0)) && Op2.match(I->getOperand(1)) &&
1518             Mask.match(I->getShuffleMask());
1519    }
1520    return false;
1521  }
1522};
1523 
1524struct m_Mask {
1525  ArrayRef<int> &MaskRef;
1526  m_Mask(ArrayRef<int> &MaskRef) : MaskRef(MaskRef) {}
1527  bool match(ArrayRef<int> Mask) {
1528    MaskRef = Mask;
1529    return true;
1530  }
1531};
1532 
1533struct m_ZeroMask {
1534  bool match(ArrayRef<int> Mask) {
1535    return all_of(Mask, [](int Elem) { return Elem == 0 || Elem == -1; });
1536  }
1537};
1538 
1539struct m_SpecificMask {
1540  ArrayRef<int> &MaskRef;
1541  m_SpecificMask(ArrayRef<int> &MaskRef) : MaskRef(MaskRef) {}
1542  bool match(ArrayRef<int> Mask) { return MaskRef == Mask; }
1543};
1544 
1545struct m_SplatOrUndefMask {
1546  int &SplatIndex;
1547  m_SplatOrUndefMask(int &SplatIndex) : SplatIndex(SplatIndex) {}
1548  bool match(ArrayRef<int> Mask) {
1549    auto First = find_if(Mask, [](int Elem) { return Elem != -1; });
1550    if (First == Mask.end())
1551      return false;
1552    SplatIndex = *First;
1553    return all_of(Mask,
1554                  [First](int Elem) { return Elem == *First || Elem == -1; });
1555  }
1556};
1557 
1558/// Matches ShuffleVectorInst independently of mask value.
1559template <typename V1_t, typename V2_t>
1560inline TwoOps_match<V1_t, V2_t, Instruction::ShuffleVector>
1561m_Shuffle(const V1_t &v1, const V2_t &v2) {
1562  return TwoOps_match<V1_t, V2_t, Instruction::ShuffleVector>(v1, v2);
1563}
1564 
1565template <typename V1_t, typename V2_t, typename Mask_t>
1566inline Shuffle_match<V1_t, V2_t, Mask_t>
1567m_Shuffle(const V1_t &v1, const V2_t &v2, const Mask_t &mask) {
1568  return Shuffle_match<V1_t, V2_t, Mask_t>(v1, v2, mask);
1569}
1570 
1571/// Matches LoadInst.
1572template <typename OpTy>
1573inline OneOps_match<OpTy, Instruction::Load> m_Load(const OpTy &Op) {
1574  return OneOps_match<OpTy, Instruction::Load>(Op);
1575}
1576 
1577/// Matches StoreInst.
1578template <typename ValueOpTy, typename PointerOpTy>
1579inline TwoOps_match<ValueOpTy, PointerOpTy, Instruction::Store>
1580m_Store(const ValueOpTy &ValueOp, const PointerOpTy &PointerOp) {
1581  return TwoOps_match<ValueOpTy, PointerOpTy, Instruction::Store>(ValueOp,
1582                                                                  PointerOp);
1583}
1584 
1585//===----------------------------------------------------------------------===//
1586// Matchers for CastInst classes
1587//
1588 
1589template <typename Op_t, unsigned Opcode> struct CastClass_match {
1590  Op_t Op;
1591 
1592  CastClass_match(const Op_t &OpMatch) : Op(OpMatch) {}
1593 
1594  template <typename OpTy> bool match(OpTy *V) {
1595    if (auto *O = dyn_cast<Operator>(V))
1596      return O->getOpcode() == Opcode && Op.match(O->getOperand(0));
1597    return false;
1598  }
1599};
1600 
1601/// Matches BitCast.
1602template <typename OpTy>
1603inline CastClass_match<OpTy, Instruction::BitCast> m_BitCast(const OpTy &Op) {
1604  return CastClass_match<OpTy, Instruction::BitCast>(Op);
1605}
1606 
1607/// Matches PtrToInt.
1608template <typename OpTy>
1609inline CastClass_match<OpTy, Instruction::PtrToInt> m_PtrToInt(const OpTy &Op) {
1610  return CastClass_match<OpTy, Instruction::PtrToInt>(Op);
1611}
1612 
1613/// Matches IntToPtr.
1614template <typename OpTy>
1615inline CastClass_match<OpTy, Instruction::IntToPtr> m_IntToPtr(const OpTy &Op) {
1616  return CastClass_match<OpTy, Instruction::IntToPtr>(Op);
1617}
1618 
1619/// Matches Trunc.
1620template <typename OpTy>
1621inline CastClass_match<OpTy, Instruction::Trunc> m_Trunc(const OpTy &Op) {
1622  return CastClass_match<OpTy, Instruction::Trunc>(Op);
1623}
1624 
1625template <typename OpTy>
1626inline match_combine_or<CastClass_match<OpTy, Instruction::Trunc>, OpTy>
1627m_TruncOrSelf(const OpTy &Op) {
1628  return m_CombineOr(m_Trunc(Op), Op);
1629}
1630 
1631/// Matches SExt.
1632template <typename OpTy>
1633inline CastClass_match<OpTy, Instruction::SExt> m_SExt(const OpTy &Op) {
1634  return CastClass_match<OpTy, Instruction::SExt>(Op);
1635}
1636 
1637/// Matches ZExt.
1638template <typename OpTy>
1639inline CastClass_match<OpTy, Instruction::ZExt> m_ZExt(const OpTy &Op) {
1640  return CastClass_match<OpTy, Instruction::ZExt>(Op);
48
←
Returning without writing to 'Op.VR'→
1641}
1642 
1643template <typename OpTy>
1644inline match_combine_or<CastClass_match<OpTy, Instruction::ZExt>, OpTy>
1645m_ZExtOrSelf(const OpTy &Op) {
1646  return m_CombineOr(m_ZExt(Op), Op);
47
←
Calling 'm_ZExt<llvm::PatternMatch::bind_ty<llvm::Value>>'→
49
←
Returning from 'm_ZExt<llvm::PatternMatch::bind_ty<llvm::Value>>'→
50
←
Calling 'm_CombineOr<llvm::PatternMatch::CastClass_match<llvm::PatternMatch::bind_ty<llvm::Value>, 39>, llvm::PatternMatch::bind_ty<llvm::Value>>'→
52
←
Returning from 'm_CombineOr<llvm::PatternMatch::CastClass_match<llvm::PatternMatch::bind_ty<llvm::Value>, 39>, llvm::PatternMatch::bind_ty<llvm::Value>>'→
53
←
Returning without writing to 'Op.VR'→
1647}
1648 
1649template <typename OpTy>
1650inline match_combine_or<CastClass_match<OpTy, Instruction::SExt>, OpTy>
1651m_SExtOrSelf(const OpTy &Op) {
1652  return m_CombineOr(m_SExt(Op), Op);
1653}
1654 
1655template <typename OpTy>
1656inline match_combine_or<CastClass_match<OpTy, Instruction::ZExt>,
1657                        CastClass_match<OpTy, Instruction::SExt>>
1658m_ZExtOrSExt(const OpTy &Op) {
1659  return m_CombineOr(m_ZExt(Op), m_SExt(Op));
1660}
1661 
1662template <typename OpTy>
1663inline match_combine_or<
1664    match_combine_or<CastClass_match<OpTy, Instruction::ZExt>,
1665                     CastClass_match<OpTy, Instruction::SExt>>,
1666    OpTy>
1667m_ZExtOrSExtOrSelf(const OpTy &Op) {
1668  return m_CombineOr(m_ZExtOrSExt(Op), Op);
1669}
1670 
1671template <typename OpTy>
1672inline CastClass_match<OpTy, Instruction::UIToFP> m_UIToFP(const OpTy &Op) {
1673  return CastClass_match<OpTy, Instruction::UIToFP>(Op);
1674}
1675 
1676template <typename OpTy>
1677inline CastClass_match<OpTy, Instruction::SIToFP> m_SIToFP(const OpTy &Op) {
1678  return CastClass_match<OpTy, Instruction::SIToFP>(Op);
1679}
1680 
1681template <typename OpTy>
1682inline CastClass_match<OpTy, Instruction::FPToUI> m_FPToUI(const OpTy &Op) {
1683  return CastClass_match<OpTy, Instruction::FPToUI>(Op);
1684}
1685 
1686template <typename OpTy>
1687inline CastClass_match<OpTy, Instruction::FPToSI> m_FPToSI(const OpTy &Op) {
1688  return CastClass_match<OpTy, Instruction::FPToSI>(Op);
1689}
1690 
1691template <typename OpTy>
1692inline CastClass_match<OpTy, Instruction::FPTrunc> m_FPTrunc(const OpTy &Op) {
1693  return CastClass_match<OpTy, Instruction::FPTrunc>(Op);
1694}
1695 
1696template <typename OpTy>
1697inline CastClass_match<OpTy, Instruction::FPExt> m_FPExt(const OpTy &Op) {
1698  return CastClass_match<OpTy, Instruction::FPExt>(Op);
1699}
1700 
1701//===----------------------------------------------------------------------===//
1702// Matchers for control flow.
1703//
1704 
1705struct br_match {
1706  BasicBlock *&Succ;
1707 
1708  br_match(BasicBlock *&Succ) : Succ(Succ) {}
1709 
1710  template <typename OpTy> bool match(OpTy *V) {
1711    if (auto *BI = dyn_cast<BranchInst>(V))
1712      if (BI->isUnconditional()) {
1713        Succ = BI->getSuccessor(0);
1714        return true;
1715      }
1716    return false;
1717  }
1718};
1719 
1720inline br_match m_UnconditionalBr(BasicBlock *&Succ) { return br_match(Succ); }
1721 
1722template <typename Cond_t, typename TrueBlock_t, typename FalseBlock_t>
1723struct brc_match {
1724  Cond_t Cond;
1725  TrueBlock_t T;
1726  FalseBlock_t F;
1727 
1728  brc_match(const Cond_t &C, const TrueBlock_t &t, const FalseBlock_t &f)
1729      : Cond(C), T(t), F(f) {}
1730 
1731  template <typename OpTy> bool match(OpTy *V) {
1732    if (auto *BI = dyn_cast<BranchInst>(V))
1733      if (BI->isConditional() && Cond.match(BI->getCondition()))
1734        return T.match(BI->getSuccessor(0)) && F.match(BI->getSuccessor(1));
1735    return false;
1736  }
1737};
1738 
1739template <typename Cond_t>
1740inline brc_match<Cond_t, bind_ty<BasicBlock>, bind_ty<BasicBlock>>
1741m_Br(const Cond_t &C, BasicBlock *&T, BasicBlock *&F) {
1742  return brc_match<Cond_t, bind_ty<BasicBlock>, bind_ty<BasicBlock>>(
1743      C, m_BasicBlock(T), m_BasicBlock(F));
1744}
1745 
1746template <typename Cond_t, typename TrueBlock_t, typename FalseBlock_t>
1747inline brc_match<Cond_t, TrueBlock_t, FalseBlock_t>
1748m_Br(const Cond_t &C, const TrueBlock_t &T, const FalseBlock_t &F) {
1749  return brc_match<Cond_t, TrueBlock_t, FalseBlock_t>(C, T, F);
1750}
1751 
1752//===----------------------------------------------------------------------===//
1753// Matchers for max/min idioms, eg: "select (sgt x, y), x, y" -> smax(x,y).
1754//
1755 
1756template <typename CmpInst_t, typename LHS_t, typename RHS_t, typename Pred_t,
1757          bool Commutable = false>
1758struct MaxMin_match {
1759  using PredType = Pred_t;
1760  LHS_t L;
1761  RHS_t R;
1762 
1763  // The evaluation order is always stable, regardless of Commutability.
1764  // The LHS is always matched first.
1765  MaxMin_match(const LHS_t &LHS, const RHS_t &RHS) : L(LHS), R(RHS) {}
1766 
1767  template <typename OpTy> bool match(OpTy *V) {
1768    if (auto *II = dyn_cast<IntrinsicInst>(V)) {
1769      Intrinsic::ID IID = II->getIntrinsicID();
1770      if ((IID == Intrinsic::smax && Pred_t::match(ICmpInst::ICMP_SGT)) ||
1771          (IID == Intrinsic::smin && Pred_t::match(ICmpInst::ICMP_SLT)) ||
1772          (IID == Intrinsic::umax && Pred_t::match(ICmpInst::ICMP_UGT)) ||
1773          (IID == Intrinsic::umin && Pred_t::match(ICmpInst::ICMP_ULT))) {
1774        Value *LHS = II->getOperand(0), *RHS = II->getOperand(1);
1775        return (L.match(LHS) && R.match(RHS)) ||
1776               (Commutable && L.match(RHS) && R.match(LHS));
1777      }
1778    }
1779    // Look for "(x pred y) ? x : y" or "(x pred y) ? y : x".
1780    auto *SI = dyn_cast<SelectInst>(V);
1781    if (!SI)
1782      return false;
1783    auto *Cmp = dyn_cast<CmpInst_t>(SI->getCondition());
1784    if (!Cmp)
1785      return false;
1786    // At this point we have a select conditioned on a comparison.  Check that
1787    // it is the values returned by the select that are being compared.
1788    auto *TrueVal = SI->getTrueValue();
1789    auto *FalseVal = SI->getFalseValue();
1790    auto *LHS = Cmp->getOperand(0);
1791    auto *RHS = Cmp->getOperand(1);
1792    if ((TrueVal != LHS || FalseVal != RHS) &&
1793        (TrueVal != RHS || FalseVal != LHS))
1794      return false;
1795    typename CmpInst_t::Predicate Pred =
1796        LHS == TrueVal ? Cmp->getPredicate() : Cmp->getInversePredicate();
1797    // Does "(x pred y) ? x : y" represent the desired max/min operation?
1798    if (!Pred_t::match(Pred))
1799      return false;
1800    // It does!  Bind the operands.
1801    return (L.match(LHS) && R.match(RHS)) ||
1802           (Commutable && L.match(RHS) && R.match(LHS));
1803  }
1804};
1805 
1806/// Helper class for identifying signed max predicates.
1807struct smax_pred_ty {
1808  static bool match(ICmpInst::Predicate Pred) {
1809    return Pred == CmpInst::ICMP_SGT || Pred == CmpInst::ICMP_SGE;
1810  }
1811};
1812 
1813/// Helper class for identifying signed min predicates.
1814struct smin_pred_ty {
1815  static bool match(ICmpInst::Predicate Pred) {
1816    return Pred == CmpInst::ICMP_SLT || Pred == CmpInst::ICMP_SLE;
1817  }
1818};
1819 
1820/// Helper class for identifying unsigned max predicates.
1821struct umax_pred_ty {
1822  static bool match(ICmpInst::Predicate Pred) {
1823    return Pred == CmpInst::ICMP_UGT || Pred == CmpInst::ICMP_UGE;
1824  }
1825};
1826 
1827/// Helper class for identifying unsigned min predicates.
1828struct umin_pred_ty {
1829  static bool match(ICmpInst::Predicate Pred) {
1830    return Pred == CmpInst::ICMP_ULT || Pred == CmpInst::ICMP_ULE;
1831  }
1832};
1833 
1834/// Helper class for identifying ordered max predicates.
1835struct ofmax_pred_ty {
1836  static bool match(FCmpInst::Predicate Pred) {
1837    return Pred == CmpInst::FCMP_OGT || Pred == CmpInst::FCMP_OGE;
1838  }
1839};
1840 
1841/// Helper class for identifying ordered min predicates.
1842struct ofmin_pred_ty {
1843  static bool match(FCmpInst::Predicate Pred) {
1844    return Pred == CmpInst::FCMP_OLT || Pred == CmpInst::FCMP_OLE;
1845  }
1846};
1847 
1848/// Helper class for identifying unordered max predicates.
1849struct ufmax_pred_ty {
1850  static bool match(FCmpInst::Predicate Pred) {
1851    return Pred == CmpInst::FCMP_UGT || Pred == CmpInst::FCMP_UGE;
1852  }
1853};
1854 
1855/// Helper class for identifying unordered min predicates.
1856struct ufmin_pred_ty {
1857  static bool match(FCmpInst::Predicate Pred) {
1858    return Pred == CmpInst::FCMP_ULT || Pred == CmpInst::FCMP_ULE;
1859  }
1860};
1861 
1862template <typename LHS, typename RHS>
1863inline MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty> m_SMax(const LHS &L,
1864                                                             const RHS &R) {
1865  return MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty>(L, R);
1866}
1867 
1868template <typename LHS, typename RHS>
1869inline MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty> m_SMin(const LHS &L,
1870                                                             const RHS &R) {
1871  return MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty>(L, R);
1872}
1873 
1874template <typename LHS, typename RHS>
1875inline MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty> m_UMax(const LHS &L,
1876                                                             const RHS &R) {
1877  return MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty>(L, R);
1878}
1879 
1880template <typename LHS, typename RHS>
1881inline MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty> m_UMin(const LHS &L,
1882                                                             const RHS &R) {
1883  return MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty>(L, R);
1884}
1885 
1886template <typename LHS, typename RHS>
1887inline match_combine_or<
1888    match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty>,
1889                     MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty>>,
1890    match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty>,
1891                     MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty>>>
1892m_MaxOrMin(const LHS &L, const RHS &R) {
1893  return m_CombineOr(m_CombineOr(m_SMax(L, R), m_SMin(L, R)),
1894                     m_CombineOr(m_UMax(L, R), m_UMin(L, R)));
1895}
1896 
1897/// Match an 'ordered' floating point maximum function.
1898/// Floating point has one special value 'NaN'. Therefore, there is no total
1899/// order. However, if we can ignore the 'NaN' value (for example, because of a
1900/// 'no-nans-float-math' flag) a combination of a fcmp and select has 'maximum'
1901/// semantics. In the presence of 'NaN' we have to preserve the original
1902/// select(fcmp(ogt/ge, L, R), L, R) semantics matched by this predicate.
1903///
1904///                         max(L, R)  iff L and R are not NaN
1905///  m_OrdFMax(L, R) =      R          iff L or R are NaN
1906template <typename LHS, typename RHS>
1907inline MaxMin_match<FCmpInst, LHS, RHS, ofmax_pred_ty> m_OrdFMax(const LHS &L,
1908                                                                 const RHS &R) {
1909  return MaxMin_match<FCmpInst, LHS, RHS, ofmax_pred_ty>(L, R);
1910}
1911 
1912/// Match an 'ordered' floating point minimum function.
1913/// Floating point has one special value 'NaN'. Therefore, there is no total
1914/// order. However, if we can ignore the 'NaN' value (for example, because of a
1915/// 'no-nans-float-math' flag) a combination of a fcmp and select has 'minimum'
1916/// semantics. In the presence of 'NaN' we have to preserve the original
1917/// select(fcmp(olt/le, L, R), L, R) semantics matched by this predicate.
1918///
1919///                         min(L, R)  iff L and R are not NaN
1920///  m_OrdFMin(L, R) =      R          iff L or R are NaN
1921template <typename LHS, typename RHS>
1922inline MaxMin_match<FCmpInst, LHS, RHS, ofmin_pred_ty> m_OrdFMin(const LHS &L,
1923                                                                 const RHS &R) {
1924  return MaxMin_match<FCmpInst, LHS, RHS, ofmin_pred_ty>(L, R);
1925}
1926 
1927/// Match an 'unordered' floating point maximum function.
1928/// Floating point has one special value 'NaN'. Therefore, there is no total
1929/// order. However, if we can ignore the 'NaN' value (for example, because of a
1930/// 'no-nans-float-math' flag) a combination of a fcmp and select has 'maximum'
1931/// semantics. In the presence of 'NaN' we have to preserve the original
1932/// select(fcmp(ugt/ge, L, R), L, R) semantics matched by this predicate.
1933///
1934///                         max(L, R)  iff L and R are not NaN
1935///  m_UnordFMax(L, R) =    L          iff L or R are NaN
1936template <typename LHS, typename RHS>
1937inline MaxMin_match<FCmpInst, LHS, RHS, ufmax_pred_ty>
1938m_UnordFMax(const LHS &L, const RHS &R) {
1939  return MaxMin_match<FCmpInst, LHS, RHS, ufmax_pred_ty>(L, R);
1940}
1941 
1942/// Match an 'unordered' floating point minimum function.
1943/// Floating point has one special value 'NaN'. Therefore, there is no total
1944/// order. However, if we can ignore the 'NaN' value (for example, because of a
1945/// 'no-nans-float-math' flag) a combination of a fcmp and select has 'minimum'
1946/// semantics. In the presence of 'NaN' we have to preserve the original
1947/// select(fcmp(ult/le, L, R), L, R) semantics matched by this predicate.
1948///
1949///                          min(L, R)  iff L and R are not NaN
1950///  m_UnordFMin(L, R) =     L          iff L or R are NaN
1951template <typename LHS, typename RHS>
1952inline MaxMin_match<FCmpInst, LHS, RHS, ufmin_pred_ty>
1953m_UnordFMin(const LHS &L, const RHS &R) {
1954  return MaxMin_match<FCmpInst, LHS, RHS, ufmin_pred_ty>(L, R);
1955}
1956 
1957//===----------------------------------------------------------------------===//
1958// Matchers for overflow check patterns: e.g. (a + b) u< a, (a ^ -1) <u b
1959// Note that S might be matched to other instructions than AddInst.
1960//
1961 
1962template <typename LHS_t, typename RHS_t, typename Sum_t>
1963struct UAddWithOverflow_match {
1964  LHS_t L;
1965  RHS_t R;
1966  Sum_t S;
1967 
1968  UAddWithOverflow_match(const LHS_t &L, const RHS_t &R, const Sum_t &S)
1969      : L(L), R(R), S(S) {}
1970 
1971  template <typename OpTy> bool match(OpTy *V) {
1972    Value *ICmpLHS, *ICmpRHS;
1973    ICmpInst::Predicate Pred;
1974    if (!m_ICmp(Pred, m_Value(ICmpLHS), m_Value(ICmpRHS)).match(V))
1975      return false;
1976 
1977    Value *AddLHS, *AddRHS;
1978    auto AddExpr = m_Add(m_Value(AddLHS), m_Value(AddRHS));
1979 
1980    // (a + b) u< a, (a + b) u< b
1981    if (Pred == ICmpInst::ICMP_ULT)
1982      if (AddExpr.match(ICmpLHS) && (ICmpRHS == AddLHS || ICmpRHS == AddRHS))
1983        return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpLHS);
1984 
1985    // a >u (a + b), b >u (a + b)
1986    if (Pred == ICmpInst::ICMP_UGT)
1987      if (AddExpr.match(ICmpRHS) && (ICmpLHS == AddLHS || ICmpLHS == AddRHS))
1988        return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpRHS);
1989 
1990    Value *Op1;
1991    auto XorExpr = m_OneUse(m_Xor(m_Value(Op1), m_AllOnes()));
1992    // (a ^ -1) <u b
1993    if (Pred == ICmpInst::ICMP_ULT) {
1994      if (XorExpr.match(ICmpLHS))
1995        return L.match(Op1) && R.match(ICmpRHS) && S.match(ICmpLHS);
1996    }
1997    //  b > u (a ^ -1)
1998    if (Pred == ICmpInst::ICMP_UGT) {
1999      if (XorExpr.match(ICmpRHS))
2000        return L.match(Op1) && R.match(ICmpLHS) && S.match(ICmpRHS);
2001    }
2002 
2003    // Match special-case for increment-by-1.
2004    if (Pred == ICmpInst::ICMP_EQ) {
2005      // (a + 1) == 0
2006      // (1 + a) == 0
2007      if (AddExpr.match(ICmpLHS) && m_ZeroInt().match(ICmpRHS) &&
2008          (m_One().match(AddLHS) || m_One().match(AddRHS)))
2009        return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpLHS);
2010      // 0 == (a + 1)
2011      // 0 == (1 + a)
2012      if (m_ZeroInt().match(ICmpLHS) && AddExpr.match(ICmpRHS) &&
2013          (m_One().match(AddLHS) || m_One().match(AddRHS)))
2014        return L.match(AddLHS) && R.match(AddRHS) && S.match(ICmpRHS);
2015    }
2016 
2017    return false;
2018  }
2019};
2020 
2021/// Match an icmp instruction checking for unsigned overflow on addition.
2022///
2023/// S is matched to the addition whose result is being checked for overflow, and
2024/// L and R are matched to the LHS and RHS of S.
2025template <typename LHS_t, typename RHS_t, typename Sum_t>
2026UAddWithOverflow_match<LHS_t, RHS_t, Sum_t>
2027m_UAddWithOverflow(const LHS_t &L, const RHS_t &R, const Sum_t &S) {
2028  return UAddWithOverflow_match<LHS_t, RHS_t, Sum_t>(L, R, S);
2029}
2030 
2031template <typename Opnd_t> struct Argument_match {
2032  unsigned OpI;
2033  Opnd_t Val;
2034 
2035  Argument_match(unsigned OpIdx, const Opnd_t &V) : OpI(OpIdx), Val(V) {}
2036 
2037  template <typename OpTy> bool match(OpTy *V) {
2038    // FIXME: Should likely be switched to use `CallBase`.
2039    if (const auto *CI = dyn_cast<CallInst>(V))
2040      return Val.match(CI->getArgOperand(OpI));
2041    return false;
2042  }
2043};
2044 
2045/// Match an argument.
2046template <unsigned OpI, typename Opnd_t>
2047inline Argument_match<Opnd_t> m_Argument(const Opnd_t &Op) {
2048  return Argument_match<Opnd_t>(OpI, Op);
2049}
2050 
2051/// Intrinsic matchers.
2052struct IntrinsicID_match {
2053  unsigned ID;
2054 
2055  IntrinsicID_match(Intrinsic::ID IntrID) : ID(IntrID) {}
2056 
2057  template <typename OpTy> bool match(OpTy *V) {
2058    if (const auto *CI = dyn_cast<CallInst>(V))
2059      if (const auto *F = CI->getCalledFunction())
2060        return F->getIntrinsicID() == ID;
2061    return false;
2062  }
2063};
2064 
2065/// Intrinsic matches are combinations of ID matchers, and argument
2066/// matchers. Higher arity matcher are defined recursively in terms of and-ing
2067/// them with lower arity matchers. Here's some convenient typedefs for up to
2068/// several arguments, and more can be added as needed
2069template <typename T0 = void, typename T1 = void, typename T2 = void,
2070          typename T3 = void, typename T4 = void, typename T5 = void,
2071          typename T6 = void, typename T7 = void, typename T8 = void,
2072          typename T9 = void, typename T10 = void>
2073struct m_Intrinsic_Ty;
2074template <typename T0> struct m_Intrinsic_Ty<T0> {
2075  using Ty = match_combine_and<IntrinsicID_match, Argument_match<T0>>;
2076};
2077template <typename T0, typename T1> struct m_Intrinsic_Ty<T0, T1> {
2078  using Ty =
2079      match_combine_and<typename m_Intrinsic_Ty<T0>::Ty, Argument_match<T1>>;
2080};
2081template <typename T0, typename T1, typename T2>
2082struct m_Intrinsic_Ty<T0, T1, T2> {
2083  using Ty =
2084      match_combine_and<typename m_Intrinsic_Ty<T0, T1>::Ty,
2085                        Argument_match<T2>>;
2086};
2087template <typename T0, typename T1, typename T2, typename T3>
2088struct m_Intrinsic_Ty<T0, T1, T2, T3> {
2089  using Ty =
2090      match_combine_and<typename m_Intrinsic_Ty<T0, T1, T2>::Ty,
2091                        Argument_match<T3>>;
2092};
2093 
2094template <typename T0, typename T1, typename T2, typename T3, typename T4>
2095struct m_Intrinsic_Ty<T0, T1, T2, T3, T4> {
2096  using Ty = match_combine_and<typename m_Intrinsic_Ty<T0, T1, T2, T3>::Ty,
2097                               Argument_match<T4>>;
2098};
2099 
2100template <typename T0, typename T1, typename T2, typename T3, typename T4, typename T5>
2101struct m_Intrinsic_Ty<T0, T1, T2, T3, T4, T5> {
2102  using Ty = match_combine_and<typename m_Intrinsic_Ty<T0, T1, T2, T3, T4>::Ty,
2103                               Argument_match<T5>>;
2104};
2105 
2106/// Match intrinsic calls like this:
2107/// m_Intrinsic<Intrinsic::fabs>(m_Value(X))
2108template <Intrinsic::ID IntrID> inline IntrinsicID_match m_Intrinsic() {
2109  return IntrinsicID_match(IntrID);
2110}
2111 
2112/// Matches MaskedLoad Intrinsic.
2113template <typename Opnd0, typename Opnd1, typename Opnd2, typename Opnd3>
2114inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2, Opnd3>::Ty
2115m_MaskedLoad(const Opnd0 &Op0, const Opnd1 &Op1, const Opnd2 &Op2,
2116             const Opnd3 &Op3) {
2117  return m_Intrinsic<Intrinsic::masked_load>(Op0, Op1, Op2, Op3);
2118}
2119 
2120template <Intrinsic::ID IntrID, typename T0>
2121inline typename m_Intrinsic_Ty<T0>::Ty m_Intrinsic(const T0 &Op0) {
2122  return m_CombineAnd(m_Intrinsic<IntrID>(), m_Argument<0>(Op0));
2123}
2124 
2125template <Intrinsic::ID IntrID, typename T0, typename T1>
2126inline typename m_Intrinsic_Ty<T0, T1>::Ty m_Intrinsic(const T0 &Op0,
2127                                                       const T1 &Op1) {
2128  return m_CombineAnd(m_Intrinsic<IntrID>(Op0), m_Argument<1>(Op1));
2129}
2130 
2131template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2>
2132inline typename m_Intrinsic_Ty<T0, T1, T2>::Ty
2133m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2) {
2134  return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1), m_Argument<2>(Op2));
2135}
2136 
2137template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2,
2138          typename T3>
2139inline typename m_Intrinsic_Ty<T0, T1, T2, T3>::Ty
2140m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2, const T3 &Op3) {
2141  return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1, Op2), m_Argument<3>(Op3));
2142}
2143 
2144template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2,
2145          typename T3, typename T4>
2146inline typename m_Intrinsic_Ty<T0, T1, T2, T3, T4>::Ty
2147m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2, const T3 &Op3,
2148            const T4 &Op4) {
2149  return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1, Op2, Op3),
2150                      m_Argument<4>(Op4));
2151}
2152 
2153template <Intrinsic::ID IntrID, typename T0, typename T1, typename T2,
2154          typename T3, typename T4, typename T5>
2155inline typename m_Intrinsic_Ty<T0, T1, T2, T3, T4, T5>::Ty
2156m_Intrinsic(const T0 &Op0, const T1 &Op1, const T2 &Op2, const T3 &Op3,
2157            const T4 &Op4, const T5 &Op5) {
2158  return m_CombineAnd(m_Intrinsic<IntrID>(Op0, Op1, Op2, Op3, Op4),
2159                      m_Argument<5>(Op5));
2160}
2161 
2162// Helper intrinsic matching specializations.
2163template <typename Opnd0>
2164inline typename m_Intrinsic_Ty<Opnd0>::Ty m_BitReverse(const Opnd0 &Op0) {
2165  return m_Intrinsic<Intrinsic::bitreverse>(Op0);
2166}
2167 
2168template <typename Opnd0>
2169inline typename m_Intrinsic_Ty<Opnd0>::Ty m_BSwap(const Opnd0 &Op0) {
2170  return m_Intrinsic<Intrinsic::bswap>(Op0);
2171}
2172 
2173template <typename Opnd0>
2174inline typename m_Intrinsic_Ty<Opnd0>::Ty m_FAbs(const Opnd0 &Op0) {
2175  return m_Intrinsic<Intrinsic::fabs>(Op0);
2176}
2177 
2178template <typename Opnd0>
2179inline typename m_Intrinsic_Ty<Opnd0>::Ty m_FCanonicalize(const Opnd0 &Op0) {
2180  return m_Intrinsic<Intrinsic::canonicalize>(Op0);
2181}
2182 
2183template <typename Opnd0, typename Opnd1>
2184inline typename m_Intrinsic_Ty<Opnd0, Opnd1>::Ty m_FMin(const Opnd0 &Op0,
2185                                                        const Opnd1 &Op1) {
2186  return m_Intrinsic<Intrinsic::minnum>(Op0, Op1);
2187}
2188 
2189template <typename Opnd0, typename Opnd1>
2190inline typename m_Intrinsic_Ty<Opnd0, Opnd1>::Ty m_FMax(const Opnd0 &Op0,
2191                                                        const Opnd1 &Op1) {
2192  return m_Intrinsic<Intrinsic::maxnum>(Op0, Op1);
2193}
2194 
2195template <typename Opnd0, typename Opnd1, typename Opnd2>
2196inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2>::Ty
2197m_FShl(const Opnd0 &Op0, const Opnd1 &Op1, const Opnd2 &Op2) {
2198  return m_Intrinsic<Intrinsic::fshl>(Op0, Op1, Op2);
2199}
2200 
2201template <typename Opnd0, typename Opnd1, typename Opnd2>
2202inline typename m_Intrinsic_Ty<Opnd0, Opnd1, Opnd2>::Ty
2203m_FShr(const Opnd0 &Op0, const Opnd1 &Op1, const Opnd2 &Op2) {
2204  return m_Intrinsic<Intrinsic::fshr>(Op0, Op1, Op2);
2205}
2206 
2207//===----------------------------------------------------------------------===//
2208// Matchers for two-operands operators with the operators in either order
2209//
2210 
2211/// Matches a BinaryOperator with LHS and RHS in either order.
2212template <typename LHS, typename RHS>
2213inline AnyBinaryOp_match<LHS, RHS, true> m_c_BinOp(const LHS &L, const RHS &R) {
2214  return AnyBinaryOp_match<LHS, RHS, true>(L, R);
11
←
Returning without writing to 'R.VR'→
2215}
2216 
2217/// Matches an ICmp with a predicate over LHS and RHS in either order.
2218/// Swaps the predicate if operands are commuted.
2219template <typename LHS, typename RHS>
2220inline CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate, true>
2221m_c_ICmp(ICmpInst::Predicate &Pred, const LHS &L, const RHS &R) {
2222  return CmpClass_match<LHS, RHS, ICmpInst, ICmpInst::Predicate, true>(Pred, L,
2223                                                                       R);
2224}
2225 
2226/// Matches a Add with LHS and RHS in either order.
2227template <typename LHS, typename RHS>
2228inline BinaryOp_match<LHS, RHS, Instruction::Add, true> m_c_Add(const LHS &L,
2229                                                                const RHS &R) {
2230  return BinaryOp_match<LHS, RHS, Instruction::Add, true>(L, R);
2231}
2232 
2233/// Matches a Mul with LHS and RHS in either order.
2234template <typename LHS, typename RHS>
2235inline BinaryOp_match<LHS, RHS, Instruction::Mul, true> m_c_Mul(const LHS &L,
2236                                                                const RHS &R) {
2237  return BinaryOp_match<LHS, RHS, Instruction::Mul, true>(L, R);
2238}
2239 
2240/// Matches an And with LHS and RHS in either order.
2241template <typename LHS, typename RHS>
2242inline BinaryOp_match<LHS, RHS, Instruction::And, true> m_c_And(const LHS &L,
2243                                                                const RHS &R) {
2244  return BinaryOp_match<LHS, RHS, Instruction::And, true>(L, R);
2245}
2246 
2247/// Matches an Or with LHS and RHS in either order.
2248template <typename LHS, typename RHS>
2249inline BinaryOp_match<LHS, RHS, Instruction::Or, true> m_c_Or(const LHS &L,
2250                                                              const RHS &R) {
2251  return BinaryOp_match<LHS, RHS, Instruction::Or, true>(L, R);
2252}
2253 
2254/// Matches an Xor with LHS and RHS in either order.
2255template <typename LHS, typename RHS>
2256inline BinaryOp_match<LHS, RHS, Instruction::Xor, true> m_c_Xor(const LHS &L,
2257                                                                const RHS &R) {
2258  return BinaryOp_match<LHS, RHS, Instruction::Xor, true>(L, R);
2259}
2260 
2261/// Matches a 'Neg' as 'sub 0, V'.
2262template <typename ValTy>
2263inline BinaryOp_match<cst_pred_ty<is_zero_int>, ValTy, Instruction::Sub>
2264m_Neg(const ValTy &V) {
2265  return m_Sub(m_ZeroInt(), V);
2266}
2267 
2268/// Matches a 'Neg' as 'sub nsw 0, V'.
2269template <typename ValTy>
2270inline OverflowingBinaryOp_match<cst_pred_ty<is_zero_int>, ValTy,
2271                                 Instruction::Sub,
2272                                 OverflowingBinaryOperator::NoSignedWrap>
2273m_NSWNeg(const ValTy &V) {
2274  return m_NSWSub(m_ZeroInt(), V);
2275}
2276 
2277/// Matches a 'Not' as 'xor V, -1' or 'xor -1, V'.
2278template <typename ValTy>
2279inline BinaryOp_match<ValTy, cst_pred_ty<is_all_ones>, Instruction::Xor, true>
2280m_Not(const ValTy &V) {
2281  return m_c_Xor(V, m_AllOnes());
2282}
2283 
2284/// Matches an SMin with LHS and RHS in either order.
2285template <typename LHS, typename RHS>
2286inline MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true>
2287m_c_SMin(const LHS &L, const RHS &R) {
2288  return MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true>(L, R);
2289}
2290/// Matches an SMax with LHS and RHS in either order.
2291template <typename LHS, typename RHS>
2292inline MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true>
2293m_c_SMax(const LHS &L, const RHS &R) {
2294  return MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true>(L, R);
2295}
2296/// Matches a UMin with LHS and RHS in either order.
2297template <typename LHS, typename RHS>
2298inline MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true>
2299m_c_UMin(const LHS &L, const RHS &R) {
2300  return MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true>(L, R);
2301}
2302/// Matches a UMax with LHS and RHS in either order.
2303template <typename LHS, typename RHS>
2304inline MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true>
2305m_c_UMax(const LHS &L, const RHS &R) {
2306  return MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true>(L, R);
2307}
2308 
2309template <typename LHS, typename RHS>
2310inline match_combine_or<
2311    match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, smax_pred_ty, true>,
2312                     MaxMin_match<ICmpInst, LHS, RHS, smin_pred_ty, true>>,
2313    match_combine_or<MaxMin_match<ICmpInst, LHS, RHS, umax_pred_ty, true>,
2314                     MaxMin_match<ICmpInst, LHS, RHS, umin_pred_ty, true>>>
2315m_c_MaxOrMin(const LHS &L, const RHS &R) {
2316  return m_CombineOr(m_CombineOr(m_c_SMax(L, R), m_c_SMin(L, R)),
2317                     m_CombineOr(m_c_UMax(L, R), m_c_UMin(L, R)));
2318}
2319 
2320/// Matches FAdd with LHS and RHS in either order.
2321template <typename LHS, typename RHS>
2322inline BinaryOp_match<LHS, RHS, Instruction::FAdd, true>
2323m_c_FAdd(const LHS &L, const RHS &R) {
2324  return BinaryOp_match<LHS, RHS, Instruction::FAdd, true>(L, R);
2325}
2326 
2327/// Matches FMul with LHS and RHS in either order.
2328template <typename LHS, typename RHS>
2329inline BinaryOp_match<LHS, RHS, Instruction::FMul, true>
2330m_c_FMul(const LHS &L, const RHS &R) {
2331  return BinaryOp_match<LHS, RHS, Instruction::FMul, true>(L, R);
2332}
2333 
2334template <typename Opnd_t> struct Signum_match {
2335  Opnd_t Val;
2336  Signum_match(const Opnd_t &V) : Val(V) {}
2337 
2338  template <typename OpTy> bool match(OpTy *V) {
2339    unsigned TypeSize = V->getType()->getScalarSizeInBits();
2340    if (TypeSize == 0)
2341      return false;
2342 
2343    unsigned ShiftWidth = TypeSize - 1;
2344    Value *OpL = nullptr, *OpR = nullptr;
2345 
2346    // This is the representation of signum we match:
2347    //
2348    //  signum(x) == (x >> 63) | (-x >>u 63)
2349    //
2350    // An i1 value is its own signum, so it's correct to match
2351    //
2352    //  signum(x) == (x >> 0)  | (-x >>u 0)
2353    //
2354    // for i1 values.
2355 
2356    auto LHS = m_AShr(m_Value(OpL), m_SpecificInt(ShiftWidth));
2357    auto RHS = m_LShr(m_Neg(m_Value(OpR)), m_SpecificInt(ShiftWidth));
2358    auto Signum = m_Or(LHS, RHS);
2359 
2360    return Signum.match(V) && OpL == OpR && Val.match(OpL);
2361  }
2362};
2363 
2364/// Matches a signum pattern.
2365///
2366/// signum(x) =
2367///      x >  0  ->  1
2368///      x == 0  ->  0
2369///      x <  0  -> -1
2370template <typename Val_t> inline Signum_match<Val_t> m_Signum(const Val_t &V) {
2371  return Signum_match<Val_t>(V);
2372}
2373 
2374template <int Ind, typename Opnd_t> struct ExtractValue_match {
2375  Opnd_t Val;
2376  ExtractValue_match(const Opnd_t &V) : Val(V) {}
2377 
2378  template <typename OpTy> bool match(OpTy *V) {
2379    if (auto *I = dyn_cast<ExtractValueInst>(V)) {
2380      // If Ind is -1, don't inspect indices
2381      if (Ind != -1 &&
2382          !(I->getNumIndices() == 1 && I->getIndices()[0] == (unsigned)Ind))
2383        return false;
2384      return Val.match(I->getAggregateOperand());
2385    }
2386    return false;
2387  }
2388};
2389 
2390/// Match a single index ExtractValue instruction.
2391/// For example m_ExtractValue<1>(...)
2392template <int Ind, typename Val_t>
2393inline ExtractValue_match<Ind, Val_t> m_ExtractValue(const Val_t &V) {
2394  return ExtractValue_match<Ind, Val_t>(V);
2395}
2396 
2397/// Match an ExtractValue instruction with any index.
2398/// For example m_ExtractValue(...)
2399template <typename Val_t>
2400inline ExtractValue_match<-1, Val_t> m_ExtractValue(const Val_t &V) {
2401  return ExtractValue_match<-1, Val_t>(V);
2402}
2403 
2404/// Matcher for a single index InsertValue instruction.
2405template <int Ind, typename T0, typename T1> struct InsertValue_match {
2406  T0 Op0;
2407  T1 Op1;
2408 
2409  InsertValue_match(const T0 &Op0, const T1 &Op1) : Op0(Op0), Op1(Op1) {}
2410 
2411  template <typename OpTy> bool match(OpTy *V) {
2412    if (auto *I = dyn_cast<InsertValueInst>(V)) {
2413      return Op0.match(I->getOperand(0)) && Op1.match(I->getOperand(1)) &&
2414             I->getNumIndices() == 1 && Ind == I->getIndices()[0];
2415    }
2416    return false;
2417  }
2418};
2419 
2420/// Matches a single index InsertValue instruction.
2421template <int Ind, typename Val_t, typename Elt_t>
2422inline InsertValue_match<Ind, Val_t, Elt_t> m_InsertValue(const Val_t &Val,
2423                                                          const Elt_t &Elt) {
2424  return InsertValue_match<Ind, Val_t, Elt_t>(Val, Elt);
2425}
2426 
2427/// Matches patterns for `vscale`. This can either be a call to `llvm.vscale` or
2428/// the constant expression
2429///  `ptrtoint(gep <vscale x 1 x i8>, <vscale x 1 x i8>* null, i32 1>`
2430/// under the right conditions determined by DataLayout.
2431struct VScaleVal_match {
2432  const DataLayout &DL;
2433  VScaleVal_match(const DataLayout &DL) : DL(DL) {}
2434 
2435  template <typename ITy> bool match(ITy *V) {
2436    if (m_Intrinsic<Intrinsic::vscale>().match(V))
2437      return true;
2438 
2439    Value *Ptr;
2440    if (m_PtrToInt(m_Value(Ptr)).match(V)) {
2441      if (auto *GEP = dyn_cast<GEPOperator>(Ptr)) {
2442        auto *DerefTy = GEP->getSourceElementType();
2443        if (GEP->getNumIndices() == 1 && isa<ScalableVectorType>(DerefTy) &&
2444            m_Zero().match(GEP->getPointerOperand()) &&
2445            m_SpecificInt(1).match(GEP->idx_begin()->get()) &&
2446            DL.getTypeAllocSizeInBits(DerefTy).getKnownMinSize() == 8)
2447          return true;
2448      }
2449    }
2450 
2451    return false;
2452  }
2453};
2454 
2455inline VScaleVal_match m_VScale(const DataLayout &DL) {
2456  return VScaleVal_match(DL);
2457}
2458 
2459template <typename LHS, typename RHS, unsigned Opcode>
2460struct LogicalOp_match {
2461  LHS L;
2462  RHS R;
2463 
2464  LogicalOp_match(const LHS &L, const RHS &R) : L(L), R(R) {}
2465 
2466  template <typename T> bool match(T *V) {
2467    if (auto *I = dyn_cast<Instruction>(V)) {
2468      if (!I->getType()->isIntOrIntVectorTy(1))
2469        return false;
2470 
2471      if (I->getOpcode() == Opcode && L.match(I->getOperand(0)) &&
2472          R.match(I->getOperand(1)))
2473        return true;
2474 
2475      if (auto *SI = dyn_cast<SelectInst>(I)) {
2476        if (Opcode == Instruction::And) {
2477          if (const auto *C = dyn_cast<Constant>(SI->getFalseValue()))
2478            if (C->isNullValue() && L.match(SI->getCondition()) &&
2479                R.match(SI->getTrueValue()))
2480              return true;
2481        } else {
2482          assert(Opcode == Instruction::Or)(static_cast <bool> (Opcode == Instruction::Or) ? void (
0) : __assert_fail ("Opcode == Instruction::Or", "/build/llvm-toolchain-snapshot-14~++20210828111110+16086d47c0d0/llvm/include/llvm/IR/PatternMatch.h"
, 2482, __extension__ __PRETTY_FUNCTION__));
2483          if (const auto *C = dyn_cast<Constant>(SI->getTrueValue()))
2484            if (C->isOneValue() && L.match(SI->getCondition()) &&
2485                R.match(SI->getFalseValue()))
2486              return true;
2487        }
2488      }
2489    }
2490 
2491    return false;
2492  }
2493};
2494 
2495/// Matches L && R either in the form of L & R or L ? R : false.
2496/// Note that the latter form is poison-blocking.
2497template <typename LHS, typename RHS>
2498inline LogicalOp_match<LHS, RHS, Instruction::And>
2499m_LogicalAnd(const LHS &L, const RHS &R) {
2500  return LogicalOp_match<LHS, RHS, Instruction::And>(L, R);
2501}
2502 
2503/// Matches L && R where L and R are arbitrary values.
2504inline auto m_LogicalAnd() { return m_LogicalAnd(m_Value(), m_Value()); }
2505 
2506/// Matches L || R either in the form of L | R or L ? true : R.
2507/// Note that the latter form is poison-blocking.
2508template <typename LHS, typename RHS>
2509inline LogicalOp_match<LHS, RHS, Instruction::Or>
2510m_LogicalOr(const LHS &L, const RHS &R) {
2511  return LogicalOp_match<LHS, RHS, Instruction::Or>(L, R);
2512}
2513 
2514/// Matches L || R where L and R are arbitrary values.
2515inline auto m_LogicalOr() {
2516  return m_LogicalOr(m_Value(), m_Value());
2517}
2518 
2519} // end namespace PatternMatch
2520} // end namespace llvm
2521 
2522#endif // LLVM_IR_PATTERNMATCH_H