RE: [PATCH] i386: Extend cvtps2pd to memory

[Liu, Hongtao via Gcc-patches]

> -----Original Message-----
> From: Uros Bizjak <ubiz...@gmail.com>
> Sent: Thursday, June 30, 2022 4:53 PM
> To: Jiang, Haochen <haochen.ji...@intel.com>
> Cc: gcc-patches@gcc.gnu.org; Liu, Hongtao <hongtao....@intel.com>
> Subject: Re: [PATCH] i386: Extend cvtps2pd to memory
> 
> On Thu, Jun 30, 2022 at 10:45 AM Uros Bizjak <ubiz...@gmail.com> wrote:
> >
> > On Thu, Jun 30, 2022 at 9:41 AM Uros Bizjak <ubiz...@gmail.com> wrote:
> > >
> > > On Thu, Jun 30, 2022 at 9:24 AM Jiang, Haochen <haochen.ji...@intel.com>
> wrote:
> > > >
> > > > > -----Original Message-----
> > > > > From: Uros Bizjak <ubiz...@gmail.com>
> > > > > Sent: Thursday, June 30, 2022 2:20 PM
> > > > > To: Jiang, Haochen <haochen.ji...@intel.com>
> > > > > Cc: gcc-patches@gcc.gnu.org; Liu, Hongtao
> > > > > <hongtao....@intel.com>
> > > > > Subject: Re: [PATCH] i386: Extend cvtps2pd to memory
> > > > >
> > > > > On Thu, Jun 30, 2022 at 7:59 AM Haochen Jiang
> > > > > <haochen.ji...@intel.com>
> > > > > wrote:
> > > > > >
> > > > > > Hi all,
> > > > > >
> > > > > > This patch aims to fix the cvtps2pd insn, which should also
> > > > > > work on memory operand but currently does not. After this fix,
> > > > > > when loop == 2, it will eliminate movq instruction.
> > > > > >
> > > > > > Regtested on x86_64-pc-linux-gnu. Ok for trunk?
> > > > > >
> > > > > > BRs,
> > > > > > Haochen
> > > > > >
> > > > > > gcc/ChangeLog:
> > > > > >
> > > > > >         PR target/43618
> > > > > >         * config/i386/sse.md (extendv2sfv2df2): New define_expand.
> > > > > >         (sse2_cvtps2pd_load<mask_name>): Rename extendvsdfv2df2.
> >
> > Rename FROM ...
> >
> > Please also mention change to sse2_cvtps2pd<mask_name>.
> >
> > > > > >
> > > > > > gcc/testsuite/ChangeLog:
> > > > > >
> > > > > >         PR target/43618
> > > > > >         * gcc.target/i386/pr43618-1.c: New test.
> > > > >
> > > > > This patch could be as simple as:
> > > > >
> > > > > diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md
> > > > > index 8cd0f617bf3..c331445cb2d 100644
> > > > > --- a/gcc/config/i386/sse.md
> > > > > +++ b/gcc/config/i386/sse.md
> > > > > @@ -9195,7 +9195,7 @@
> > > > > (define_insn "extendv2sfv2df2"
> > > > >   [(set (match_operand:V2DF 0 "register_operand" "=v")
> > > > >        (float_extend:V2DF
> > > > > -         (match_operand:V2SF 1 "register_operand" "v")))]
> > > > > +         (match_operand:V2SF 1 "nonimmediate_operand" "vm")))]
> > > > >   "TARGET_MMX_WITH_SSE"
> > > > >   "%vcvtps2pd\t{%1, %0|%0, %1}"
> > > > >   [(set_attr "type" "ssecvt")
> > > >
> > > > We also tested on this version, it is ok.
> > > >
> > > > The reason why the patch looks like this is because in the
> > > > previous insn sse2_cvtps2pd<mask_name>, the constraint vm and
> > > > vector_operand actually does not match the actual instruction.
> > > > Memory operand is V2SF, not V4SF.
> > > >
> > > > Therefore, we changed the constraint in that insn. Then it caused 
> > > > another
> issue.
> > > > For memory operand, it seems that we cannot generate those mask
> instructions.
> > > > So I change the pattern to how extendv2hfv2df2 works.
> > >
> > > If you want to change the memory access in sse2_cvtps2pd<mask_name>,
> > > then please see how e.g. <insn>v2hiv2di is handled in sse.md. In
> > > addition to two instructions, you will need one
> > > define_insn_and_split with a pre-reload splitter.
> >
> > Oh, nowadays combine does vec_select from a paradoxical subreg on its own.
> >
> > +(define_expand "extendv2sfv2df2"
> > +  [(set (match_operand:V2DF 0 "register_operand")
> > +        (float_extend:V2DF
> > +          (match_operand:V2SF 1 "nonimmediate_operand")))]
> > +  "TARGET_MMX_WITH_SSE"
> > +{
> > +  if (!MEM_P (operands[1]))
> > +    {
> >
> > You will need force reg here:
> >
> >         rtx op1 = force_reg (V2SFmode, operands[1]);
> > +      operands[1] = lowpart_subreg (V4SFmode, op1, V2SFmode);
> > +      emit_insn (gen_sse2_cvtps2pd (operands[0], operands[1]));
> > +      DONE;
> > +    }
> > +})
> >
> >
> > -(define_insn "extendv2sfv2df2"
> > +(define_insn "sse2_cvtps2pd_load<mask_name>"
> >
> > Please name this insn "*sse2_cvtps2pd<mask_name>_1". Please note the
> > star at the beginning, You don't have to make the name public.
> >
> > OK with the above changes.
> 
> Forgot to mention:
> 
> 
> -         (match_operand:V2SF 1 "register_operand" "v")))]
> -  "TARGET_MMX_WITH_SSE"
> -  "%vcvtps2pd\t{%1, %0|%0, %1}"
> +         (match_operand:V2SF 1 "memory_operand" "m")))]
> + "TARGET_MMX_WITH_SSE && <mask_avx512vl_condition>"
> +  "%vcvtps2pd\t{%1, %0<mask_operand2>|%0<mask_oper
> and2>, %q1}"
>    [(set_attr "type" "ssecvt")
> 
> The new insn does not need to be limited to TARGET_MMX_WITH_SSE, so we
> can use TARGET_SSE2 here.
> 
> Which opens the question if the expander could also be TARGET_SSE2 only.
> There are no MMX registers involved in any of the patterns anymore.
Yes.
> 
> Uros.
> >
> > Thanks,
> > Uros,