binary_ops.jl

function binary_op(op, W, @nospecialize(T))
  ty = LLVM_TYPES[T]
  if isone(W)
    V = T
  else
    ty = "<$W x $ty>"
    V = NTuple{W,VecElement{T}}
  end
  instrs = "%res = $op $ty %0, %1\nret $ty %res"
  call = :($LLVMCALL($instrs, $V, Tuple{$V,$V}, data(v1), data(v2)))
  W > 1 && (call = Expr(:call, :Vec, call))
  Expr(:block, Expr(:meta, :inline), call)
end

# Integer
for (op, f) ∈ [("add", :+), ("sub", :-), ("mul", :*), ("shl", :<<)]
  ff = Symbol('v', op)
  fnsw = Symbol(ff, "_nsw")
  fnuw = Symbol(ff, "_nuw")
  fnw = Symbol(ff, "_nw")
  ff_fast = Symbol(ff, :_fast)
  @eval begin
    # @inline $ff(a,b) = $ff_fast(a,b)
    @inline $ff(
      a::T,
      b::T
    ) where {T<:Union{FloatingTypes,IntegerTypesHW,AbstractSIMD}} =
      $ff_fast(a, b)
    # @generated $ff_fast(v1::Vec{W,T}, v2::Vec{W,T}) where {W,T<:IntegerTypesHW} = binary_op($op * (T <: Signed ? " nsw" : " nuw"), W, T)
    @generated $ff_fast(
      v1::Vec{W,T},
      v2::Vec{W,T}
    ) where {W,T<:IntegerTypesHW} = binary_op($op, W, T)
    @generated $fnsw(v1::Vec{W,T}, v2::Vec{W,T}) where {W,T<:IntegerTypesHW} =
      binary_op($(op * " nsw"), W, T)
    @generated $fnuw(v1::Vec{W,T}, v2::Vec{W,T}) where {W,T<:IntegerTypesHW} =
      binary_op($(op * " nuw"), W, T)
    @generated $fnw(v1::Vec{W,T}, v2::Vec{W,T}) where {W,T<:IntegerTypesHW} =
      binary_op($(op * " nsw nuw"), W, T)
    # @generated Base.$f(v1::Vec{W,T}, v2::Vec{W,T}) where {W,T<:IntegerTypesHW} = binary_op($op, W, T)
    @inline Base.$f(v1::Vec{W,T}, v2::Vec{W,T}) where {W,T<:IntegerTypesHW} =
      $ff_fast(v1, v2)
    # @generated $ff(v1::Vec{W,T}, v2::Vec{W,T}) where {W,T<:IntegerTypesHW} = binary_op($op, W, T)
    @inline $ff_fast(x, y) = $f(x, y)

    # @generated $ff_fast(v1::T, v2::T) where {T<:IntegerTypesHW} = binary_op($op * (T <: Signed ? " nsw" : " nuw"), 1, T)
    # @generated $ff_fast(v1::T, v2::T) where {T<:IntegerTypesHW} = binary_op($op, 1, T)
    @generated $fnsw(v1::T, v2::T) where {T<:IntegerTypesHW} =
      binary_op($(op * " nsw"), 1, T)
    @generated $fnuw(v1::T, v2::T) where {T<:IntegerTypesHW} =
      binary_op($(op * " nuw"), 1, T)
    @generated $fnw(v1::T, v2::T) where {T<:IntegerTypesHW} =
      binary_op($(op * " nsw nuw"), 1, T)
  end
end
@inline vadd_fast(v1::T, v2::T) where {T<:IntegerTypesHW} = Base.add_int(v1, v2)
@inline vsub_fast(v1::T, v2::T) where {T<:IntegerTypesHW} = Base.sub_int(v1, v2)
@inline vmul_fast(v1::T, v2::T) where {T<:IntegerTypesHW} = Base.mul_int(v1, v2)
@inline vshl_fast(v1::T, v2::T) where {T<:IntegerTypesHW} = Base.shl_int(v1, v2)
for (op, f) ∈ [("div", :÷), ("rem", :%)]
  ff = Symbol('v', op) #_ff = Symbol(:_, ff)
  sbf = Symbol('s', op, :_int)
  ubf = Symbol('u', op, :_int)
  sbf128 = Sys.WORD_SIZE == 32 ? f : sbf
  ubf128 = Sys.WORD_SIZE == 32 ? f : ubf
  @eval begin
    @generated $ff(v1::Vec{W,T}, v2::Vec{W,T}) where {W,T<:Integer} =
      binary_op((T <: Signed ? 's' : 'u') * $op, W, T)
    @inline $ff(a::I, b::I) where {I<:SignedHW} = Base.$sbf(a, b)
    @inline $ff(a::I, b::I) where {I<:UnsignedHW} = Base.$ubf(a, b)
    @inline $ff(a::Int128, b::Int128) = Base.$sbf128(a, b)
    @inline $ff(a::UInt128, b::UInt128) = Base.$ubf128(a, b)
    # @generated $_ff(v1::T, v2::T) where {T<:Integer} = binary_op((T <: Signed ? 's' : 'u') * $op, 1, T)
    # @inline $ff(v1::T, v2::T) where {T<:IntegerTypesHW} = $_ff(v1, v2)
  end
end
# for (op,f) ∈ [("div",:÷),("rem",:%)]
#   @eval begin
#     @generated Base.$f(v1::Vec{W,T}, v2::Vec{W,T}) where {W,T<:Integer} = binary_op((T <: Signed ? 's' : 'u') * $op, W, T)
#     @generated Base.$f(v1::T, v2::T) where {T<:IntegerTypesHW} = binary_op((T <: Signed ? 's' : 'u') * $op, 1, T)
#   end
# end
@inline vcld(x, y) = vadd(vdiv(vsub(x, one(x)), y), one(x))
@inline function vdivrem(x, y)
  d = vdiv(x, y)
  r = vsub(x, vmul(d, y))
  d, r
end
for (op, sub) ∈ [
  ("ashr", :SignedHW),
  ("lshr", :UnsignedHW),
  ("lshr", :IntegerTypesHW),
  ("and", :IntegerTypesHW),
  ("or", :IntegerTypesHW),
  ("xor", :IntegerTypesHW)
]
  ff = sub === :UnsignedHW ? :vashr : Symbol('v', op)
  @eval begin
    @generated $ff(v1::Vec{W,T}, v2::Vec{W,T}) where {W,T<:$sub} =
      binary_op($op, W, T)
    @generated $ff(v1::T, v2::T) where {T<:$sub} = binary_op($op, 1, T)
  end
end

for (op, f) ∈
    [("fadd", :vadd), ("fsub", :vsub), ("fmul", :vmul), ("fdiv", :vfdiv)]#,("frem",:vrem)]
  ff = Symbol(f, :_fast)
  fop_fast = f === :vfdiv ? "fdiv fast" : op * ' ' * fast_flags(true)
  fop_contract = op * ' ' * fast_flags(false)
  @eval begin
    @generated $f(
      v1::Vec{W,T},
      v2::Vec{W,T}
    ) where {W,T<:Union{Float32,Float64}} = binary_op($fop_contract, W, T)
    @generated $ff(
      v1::Vec{W,T},
      v2::Vec{W,T}
    ) where {W,T<:Union{Float32,Float64}} = binary_op($fop_fast, W, T)
    @inline $f(v1::Vec{W,Float16}, v2::Vec{W,Float16}) where {W} =
      $f(convert(Vec{W,Float32}, v1), convert(Vec{W,Float32}, v2))
    @inline $ff(v1::Vec{W,Float16}, v2::Vec{W,Float16}) where {W} =
      $ff(convert(Vec{W,Float32}, v1), convert(Vec{W,Float32}, v2))
  end
end
@inline vsub(a::T, b::T) where {T<:Union{Float32,Float64}} =
  Base.sub_float(a, b)
@inline vadd(a::T, b::T) where {T<:Union{Float32,Float64}} =
  Base.add_float(a, b)
@inline vmul(a::T, b::T) where {T<:Union{Float32,Float64}} =
  Base.mul_float(a, b)
@inline vsub_fast(a::T, b::T) where {T<:Union{Float32,Float64}} =
  Base.sub_float_fast(a, b)
@inline vadd_fast(a::T, b::T) where {T<:Union{Float32,Float64}} =
  Base.add_float_fast(a, b)
@inline vmul_fast(a::T, b::T) where {T<:Union{Float32,Float64}} =
  Base.mul_float_fast(a, b)

@inline vdiv(
  v1::AbstractSIMD{W,T},
  v2::AbstractSIMD{W,T}
) where {W,T<:FloatingTypes} = trunc(vfdiv_fast(v1, v2))
@inline vdiv_fast(
  v1::AbstractSIMD{W,T},
  v2::AbstractSIMD{W,T}
) where {W,T<:FloatingTypes} = trunc(vfdiv_fast(v1, v2))
@inline vdiv_fast(v1::T, v2::T) where {T<:FloatingTypes} =
  trunc(Base.FastMath.div_float_fast(v1, v2))
@inline vdiv_fast(v1::T, v2::T) where {T<:Number} = v1 ÷ v2
@inline vdiv(v1::T, v2::T) where {T<:Number} = v1 ÷ v2
@inline vdiv(v1::T, v2::T) where {T<:FloatingTypes} = vdiv_fast(v1, v2)
@inline vrem(a, b) = vfnmadd(vdiv_fast(a, b), b, a)
@inline vrem_fast(a, b) = vfnmadd(vdiv_fast(a, b), b, a)
# @inline vdiv_fast(v1::AbstractSIMD{W,T}, v2::AbstractSIMD{W,T}) where {W,T<:IntegerTypesHW} = trunc(T, vfloat_fast(v1) / vfloat_fast(v2))
@inline vdiv_fast(
  v1::AbstractSIMD{W,T},
  v2::AbstractSIMD{W,T}
) where {W,T<:IntegerTypesHW} = trunc(T, vfloat(v1) / vfloat(v2))
@inline function vdiv_fast(v1, v2)
  v3, v4 = promote_div(v1, v2)
  vdiv_fast(v3, v4)
end
@inline vdiv_fast(v1::VecUnroll{N,1,T,T}, s::T) where {N,T<:SignedHW} =
  VecUnroll(fmap(Base.sdiv_int, data(v1), s))
@inline vdiv_fast(v1::VecUnroll{N,1,T,T}, s::T) where {N,T<:UnsignedHW} =
  VecUnroll(fmap(Base.udiv_int, data(v1), s))
@inline vdiv_fast(
  v1::VecUnroll{N,1,T,T},
  v2::VecUnroll{N,1,T,T}
) where {N,T<:SignedHW} = VecUnroll(fmap(Base.sdiv_int, data(v1), data(v2)))
@inline vdiv_fast(
  v1::VecUnroll{N,1,T,T},
  v2::VecUnroll{N,1,T,T}
) where {N,T<:UnsignedHW} = VecUnroll(fmap(Base.udiv_int, data(v1), data(v2)))

@inline vfdiv(a::AbstractSIMDVector{W}, b::AbstractSIMDVector{W}) where {W} =
  vfdiv(vfloat(a), vfloat(b))
# @inline vfdiv_fast(a::AbstractSIMDVector{W}, b::AbstractSIMDVector{W}) where {W} = vfdiv_fast(vfloat_fast(a), vfloat_fast(b))
@inline vfdiv_fast(
  a::AbstractSIMDVector{W},
  b::AbstractSIMDVector{W}
) where {W} = vfdiv_fast(vfloat(a), vfloat(b))
@inline vfdiv(a, b) = a / b
@inline vfdiv_fast(a, b) = Base.FastMath.div_fast(a, b)

for (f, op) ∈ ((:vand, :and_int), (:vor, :or_int), (:vxor, :xor_int))
  @eval @inline $f(b1::Bool, b2::Bool) = Base.$op(b1, b2)
end

for f ∈ [:vadd, :vsub, :vmul]
  for s ∈ [Symbol(""), :_fast, :_nsw, :_nuw, :_nw]
    fs = Symbol(f, s)
    @eval begin
      @inline function $fs(
        a::Union{FloatingTypes,IntegerTypesHW,AbstractSIMD},
        b::Union{FloatingTypes,IntegerTypesHW,AbstractSIMD}
      )
        c, d = promote(a, b)
        $fs(c, d)
      end
    end
  end
end
# @inline vsub(a::T, b::T) where {T<:Base.BitInteger} = Base.sub_int(a, b)
for (vf, bf) ∈ [
  (:vadd, :add_int),
  (:vsub, :sub_int),
  (:vmul, :mul_int),
  (:vadd_fast, :add_int),
  (:vsub_fast, :sub_int),
  (:vmul_fast, :mul_int),
  (:vadd_nsw, :add_int),
  (:vsub_nsw, :sub_int),
  (:vmul_nsw, :mul_int),
  (:vadd_nuw, :add_int),
  (:vsub_nuw, :sub_int),
  (:vmul_nuw, :mul_int),
  (:vadd_nw, :add_int),
  (:vsub_nw, :sub_int),
  (:vmul_nw, :mul_int)
]
  @eval begin
    @inline $vf(a::Int128, b::Int128) = Base.$bf(a, b)
    @inline $vf(a::UInt128, b::UInt128) = Base.$bf(a, b)
  end
end
# @inline vrem(a::Float32, b::Float32) = Base.rem_float_fast(a, b)
# @inline vrem(a::Float64, b::Float64) = Base.rem_float_fast(a, b)

@inline function Base.FastMath.add_fast(
  a::AbstractSIMD,
  b::AbstractSIMD,
  c::AbstractSIMD
)
  Base.FastMath.add_fast(Base.FastMath.add_fast(a, b), c)
end
@inline function Base.FastMath.add_fast(
  a::T,
  b::T,
  c::T
) where {T<:AbstractSIMD}
  Base.FastMath.add_fast(Base.FastMath.add_fast(a, b), c)
end

@inline function Base.FastMath.add_fast(
  a::AbstractSIMD,
  b::AbstractSIMD,
  c::AbstractSIMD,
  d::AbstractSIMD
)
  x = Base.FastMath.add_fast(a, b)
  y = Base.FastMath.add_fast(c, d)
  Base.FastMath.add_fast(x, y)
end
@inline function Base.FastMath.add_fast(
  a::T,
  b::T,
  c::T,
  d::T
) where {T<:AbstractSIMD}
  x = Base.FastMath.add_fast(a, b)
  y = Base.FastMath.add_fast(c, d)
  Base.FastMath.add_fast(x, y)
end

@inline function Base.FastMath.add_fast(
  a::AbstractSIMD,
  b::AbstractSIMD,
  c::AbstractSIMD,
  d::AbstractSIMD,
  e::AbstractSIMD,
  f::Vararg{Number,K}
) where {K}
  x = Base.FastMath.add_fast(a, b)
  y = Base.FastMath.add_fast(c, d)
  Base.FastMath.add_fast(Base.FastMath.add_fast(x, y), e, f...)
end
@inline function Base.FastMath.add_fast(
  a::T,
  b::T,
  c::T,
  d::T,
  e::T,
  f::Vararg{T,K}
) where {T<:AbstractSIMD,K}
  x = Base.FastMath.add_fast(a, b)
  y = Base.FastMath.add_fast(c, d)
  Base.FastMath.add_fast(Base.FastMath.add_fast(x, y), e, f...)
end