[llvm-dev] FMA canonicalization in IR

[Nicolai Hähnle via llvm-dev]

On 19.11.2016 17:26, Sanjay Patel via llvm-dev wrote:
> If I have my FMA intrinsics story straight now (thanks for the
> explanation, Hal!), I think it raises another question about IR
> canonicalization (and may affect the proposed revision to IR FMF):
>
> define float @foo(float %a, float %b, float %c) {
>   %mul = fmul fast float %a, %b     ; using 'fast' because there is no
> 'fma' flag
>   %add = fadd fast float %mul, %c
>   ret float %add
> }
>
> Should this be:
>
> define float @goo(float %a, float %b, float %c) {
>   %maybe.fma = call fast float @llvm.fmuladd.f32(float %a, float %b,
> float %c)
>   ret float %maybe.fma
> }
> declare float @llvm.fmuladd.f32(float %a, float %b, float %c)

Doing this would raise another issue. What about:

   %t = fmul fast float %x, %y
   %u = fmul fast float %t, %x
   %v = fadd fast float %u, %z

If you _first_ canonicalize this to an fmuladd, then you might miss the 
associative transform to

   %t1 = fmul fast float %x, %x
   %u = fmul fast float %t1, %y
   %v = fadd fast flat %u, %z

So _if_ you add that canonicalization, you somehow have to teach all the 
visitFMul logic to visitFMuladd as well. Maybe that's actually desirable 
if the front-end emits fmuladd for C even when fast-math is enabled.

Cheers,
Nicolai