Thanks for the suggestion, here is the updated patch that survived
bootstrap/regtest.

> Please note reg_mentioned_p in the above condition. This function
> returns nonzero if register op0 appears somewhere within op1 and is
> critical for the correct operation of your patch.
I added reg_mentioned_p for all insns except fp16 complex mult, since
they have constraint & to the dest so it must be allocated different
register from src.

Uros Bizjak <ubizjak@gmail.com> 于2022年1月14日周五 23:49写道：


>
> On Fri, Jan 14, 2022 at 2:44 PM Hongyu Wang <wwwhhhyyy333@gmail.com> wrote:
> >
> > > Are there any technical obstacles to introduce subst to
> > > define_{,insn_and_}split?
> >
> > gccint says: define_subst can be used only in define_insn and
> > define_expand, it cannot be used in other expressions (e.g. in
> > define_insn_and_split).
>
> Hm, hm ... annoying ...
>
> > I have no idea how to implement it in current infrastructure.
> >
> > > In the proposed patch, if the output register is also mentioned in the
> > > input, then clearing before insn will clear the value in the input
> > > register. The solution in the i386.md also takes care of this issue.
> > >
> >
> > For this, I think we can add REGNO checks for operands in condition
> > (which means there is true dependency).
>
> Let's go in your direction, considering the limitations of current
> infrastructure.
> +{
> +  if (TARGET_DEST_FALSE_DEPENDENCY
> +      && get_attr_dest_false_dep (insn) ==
> +        DEST_FALSE_DEP_TRUE)
> +    output_asm_insn ("vxorps\t{%x0, %x0, %x0}", operands);
> +  return "vgetmant<ssemodesuffix>\t{%2, <round_saeonly_mask_op3>%1,
> %0<mask_operand3>|%0<mask_operand3>, %1<round_saeonly_mask_op3>, %2}";
> +}
>
> There is no need to pass the information via attributes. IMO, you
> shoud use subst attribute directly in the condition:
>
> {
>   if (TARGET_DEST_FALSE_DEPENDENCY
>       && <mask3_dest_false_dep_attr>
>       && !reg_mentioned_p (operands[0], operands[1]))
>    output_asm_insn ("vxorps\t{%x0, %x0, %x0}", operands);
>   return ".....";
> }
>
> Assume the above works, so please:
>
> - rename TARGET_DEST_FALSE_DEPENDENCY to something less generic, maybe
> following existing BMI example with TARGET_AVOID_FALSE_DEP_FOR_AVX512F
> - rename "mask3_dest_false_dep_attr" to "mask3_false_dep_for_avx512f_cond"
>
> Please note reg_mentioned_p in the above condition. This function
> returns nonzero if register op0 appears somewhere within op1 and is
> critical for the correct operation of your patch.
>
> Uros.