This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH] i386: Enable AVX512 memory broadcast for FMA
On Thu, Oct 18, 2018 at 1:19 PM H.J. Lu <hjl.tools@gmail.com> wrote:
> >> +(define_insn "*<sd_mask_codefor>fma_fmadd_<mode><sd_maskz_name>_bcst_1"
> >> + [(set (match_operand:VF_AVX512 0 "register_operand" "=v,v")
> >> + (fma:VF_AVX512
> >> + (match_operand:VF_AVX512 1 "nonimmediate_operand" "0,v")
> >> + (match_operand:VF_AVX512 2 "nonimmediate_operand" "v,0")
> >> + (vec_duplicate:VF_AVX512
> >> + (match_operand:<ssescalarmode> 3 "nonimmediate_operand"
> >> "m,m"))))]
> >
> > Please note that having "nonimmediate_operand" predicate with "m"
> > constraint will force scalar value that lives in any register to
> > memory. So, scalar value will be pushed from either integer or SSE
> > register to memory, and will be broadcast to SSE register from here. I
> > guess this is not the optimal way, and we still want (eventual movq
> > from integer reg) + broadcast insn in this case.
> >
> > If this predicate is changed to "memory_operand", then only scalars
> > that live in memory will be considered.
>
> Using "memory_operand" causes:
>
> FAIL: gcc.target/i386/avx512f-fmadd-sf-zmm-7.c scan-assembler-times
> vfmadd...ps[ \\t]+[^\n\r]+\\{1to[1-8]+\\}, %zmm[0-9]+, %zmm0 1
> FAIL: gcc.target/i386/avx512f-fmadd-sf-zmm-7.c scan-assembler-not
> vbroadcastss[^\n]*%zmm[0-9]+
>
> __m512
> foo (__m512 x, __m512 y)
> {
> return _mm512_fmadd_ps (x, y, _mm512_set1_ps (2.f));
> }
>
> Combiner:
>
> Failed to match this instruction:
> (set (reg:V16SF 91)
> (fma:V16SF (reg/v:V16SF 85 [ x ])
> (reg:V16SF 21 xmm1 [ y ])
> (vec_duplicate:V16SF (reg:SF 88))))
This is expected, there is no memory operand there. Can you check what
prevents combiner from propagating memory into the insn?
Uros.