https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87555
--- Comment #13 from Hongtao.liu <crazylht at gmail dot com> ---
(In reply to Richard Biener from comment #12)
> (In reply to Hongtao.liu from comment #10)
> > > Note I'm not sure that doing fmaddsub as merge of fma and fms will be
> > > optimal since that most definitely will preclude combine from recognizing
> > > fmaddsub from (addsub (mul ..) x) which would be another goal to support
> > > (PR81904)
> >
> > I guess you're talking about
> >
> > #include <x86intrin.h>
> > __m128d f(__m128d x, __m128d y, __m128d z){
> > return _mm_addsub_pd(_mm_mul_pd(x,y),z);
> > }
> >
> > which pass_combine tries
> >
> > Failed to match this instruction:
> > (set (reg:V2DF 88)
> > (vec_merge:V2DF (minus:V2DF (mult:V2DF (reg:V2DF 90)
> > (reg:V2DF 91))
> > (reg:V2DF 92))
> > (plus:V2DF (mult:V2DF (reg:V2DF 90)
> > (reg:V2DF 91))
> > (reg:V2DF 92))
> > (const_int 1 [0x1])))
> >
> > but doesn't realize fisrt merge operand is fms and second is fma.
>
> Yes. This situation will happen when I push the SLP pattern detection
> for addsub - we then no longer detect FMA on the GIMPLE level (we might
> want to improve that as well, of course, exposing standard pattern names
> for fmaddsub and fmsubadd).
if fm{a,s}_optab is supported in the backend, can we always simplify (minus A
(mult B C)) to (fma B C (neg A)) and (plus A (mult B C)) to (fma B C A)?
