@fastmath not applying to additions sometimes

[Zentrik]

function test5(x, y, z) 
    @fastmath @inbounds for i in eachindex(x)
        return x[i] + y[i] + z[i]
    end
end

a = [Vec{8, Float32}(0)]
@code_llvm test5(a, a, a)

We see that

; ┌ @ fastmath.jl:265 within `add_fast`
; │┌ @ operators.jl:591 within `+` @ C:\Users\rag\.julia\packages\SIMD\7eukp\src\simdvec.jl:253
; ││┌ @ C:\Users\rag\.julia\packages\SIMD\7eukp\src\LLVM_intrinsics.jl:212 within `fadd` @ C:\Users\rag\.julia\packages\SIMD\7eukp\src\LLVM_intrinsics.jl:212
; │││┌ @ C:\Users\rag\.julia\packages\SIMD\7eukp\src\LLVM_intrinsics.jl:221 within `macro expansion`
      %16 = fadd <8 x float> %.unpack, %.unpack10
      %17 = fadd <8 x float> %16, %.unpack11

For whatever reason doing 2 additions on the same line in a for loop removes the fast flag from fadd.

[Zentrik]

Ok identified what the problem is

julia> @macroexpand test(x, y, z) = @fastmath x[0] + y[0] + z[0]
 :(test(x, y, z) = begin
	#= REPL[45]:1 =#
	Base.FastMath.add_fast(x[0], y[0], z[0])
end) 

The problem is we're calling https://github.com/JuliaLang/julia/blob/e7345b89fd4eb15e8f395395701e19be705d7b06/base/fastmath.jl#L263-L264.

		# fall-back implementation for non-numeric types
        $op_fast(xs...) = $op(xs...)

We probably want something like
https://github.com/JuliaLang/julia/blob/e7345b89fd4eb15e8f395395701e19be705d7b06/base/fastmath.jl#L168-L169.

add_fast(x::T, y::T, zs::T...) where {T<:FloatTypes} = 
     add_fast(add_fast(x, y), zs...) 

[KristofferC]

n-arg parsing strikes again...

[KristofferC]

Fixed by JuliaLang/julia#54513 (on 1.12) at least.