access to fused multiply-subtract etc.

[stevengj]

Currently, f(x,y,z) = @fastmath muladd(x, y, -z) generates an explicit negation operation for the -z in the case of floating-point arguments, as can be seen by @code_native f(3.0, 4.0, 5.0). Similarly for muladd(-x, y, z) etcetera.

In #9881, @yuyichao indicated that this was due to the FastSel optimization in LLVM.

As I mentioned in #6330, it would be good to have access to fused multiply-subtract and similar operations. LLVM doesn't provide intrinsics for these, but it would be good to have a way to write code that LLVM can transform into such instructions when they are available, and that will avoid explicit negations when fused operations aren't available.

cc: @eschnett, @simonbyrne

[yuyichao]

I had a hard time following the isel logic in LLVM but it seems that the logic of replacing fmadd(x, y, -z) with fmsub(x, y, z) only appears at one place and looks pretty general (eliminate the neg if it is only used by one fma-like instruction) so I don't think there's much we can do to help llvm recognizing the code pattern here.

Can we make #15984 close this and make this an upstream issue?

[stevengj]

Sounds good @yuyichao.

[yuyichao]

Might be related https://llvm.org/bugs/show_bug.cgi?id=13118