[PATCH] Adjust memory_move_cost for MASK_REGS when MODE_SIZE > 8.

[PATCH] Adjust memory_move_cost for MASK_REGS when MODE_SIZE > 8.

[liuhongt]

RA sometimes will use lowest the cost of the mode with all different regclasses
w/o check if it's hard_regno_mode_ok.
It's impossible to put modes whose size > 8 into MASK_REGS, ajdust the cost to
avoid potential performance issue.

Bootstrapped and regtested on x86_64-pc-linux-gnu{-m32,}.
Ok for trunk?

gcc/ChangeLog:

	* config/i386/i386.cc (inline_memory_move_cost): Return 100
	for MASK_REGS when MODE_SIZE > 8.
---
 gcc/config/i386/i386.cc | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/gcc/config/i386/i386.cc b/gcc/config/i386/i386.cc
index 2cc8e9548a9..2581b800a06 100644
--- a/gcc/config/i386/i386.cc
+++ b/gcc/config/i386/i386.cc
@@ -19847,9 +19847,12 @@ inline_memory_move_cost (machine_mode mode, enum reg_class regclass, int in)
 	  index = 1;
 	  break;
 	/* DImode loads and stores assumed to cost the same as SImode.  */
-	default:
+	case 4:
+	case 8:
 	  index = 2;
 	  break;
+	default:
+	  return 100;
 	}
 
       if (in == 2)
-- 
2.39.1.388.g2fc9e9ca3c

Re: [PATCH] Adjust memory_move_cost for MASK_REGS when MODE_SIZE > 8.

[Uros Bizjak]

On Fri, Mar 31, 2023 at 7:11 AM liuhongt <hongtao.liu@intel.com> wrote:
>
> RA sometimes will use lowest the cost of the mode with all different regclasses
> w/o check if it's hard_regno_mode_ok.
> It's impossible to put modes whose size > 8 into MASK_REGS, ajdust the cost to
> avoid potential performance issue.

I was going to ask to open a PR in order to fix RA instead of the
proposed workaround, but the patch just prevents a theoretical issue
and follows the established practice with other regclasses. Perhaps
these workarounds are not needed with the current RA, but this is a
tangential issue which should be fixed for all regclasses.

> Bootstrapped and regtested on x86_64-pc-linux-gnu{-m32,}.
> Ok for trunk?

OK.

Thanks,
Uros.

>
> gcc/ChangeLog:
>
>         * config/i386/i386.cc (inline_memory_move_cost): Return 100
>         for MASK_REGS when MODE_SIZE > 8.
> ---
>  gcc/config/i386/i386.cc | 5 ++++-
>  1 file changed, 4 insertions(+), 1 deletion(-)
>
> diff --git a/gcc/config/i386/i386.cc b/gcc/config/i386/i386.cc
> index 2cc8e9548a9..2581b800a06 100644
> --- a/gcc/config/i386/i386.cc
> +++ b/gcc/config/i386/i386.cc
> @@ -19847,9 +19847,12 @@ inline_memory_move_cost (machine_mode mode, enum reg_class regclass, int in)
>           index = 1;
>           break;
>         /* DImode loads and stores assumed to cost the same as SImode.  */
> -       default:
> +       case 4:
> +       case 8:
>           index = 2;
>           break;
> +       default:
> +         return 100;
>         }
>
>        if (in == 2)
> --
> 2.39.1.388.g2fc9e9ca3c
>

Re: [PATCH] Adjust memory_move_cost for MASK_REGS when MODE_SIZE > 8.

[Hongtao Liu]

On Fri, Mar 31, 2023 at 1:57 PM Uros Bizjak <ubizjak@gmail.com> wrote:
>
> On Fri, Mar 31, 2023 at 7:11 AM liuhongt <hongtao.liu@intel.com> wrote:
> >
> > RA sometimes will use lowest the cost of the mode with all different regclasses
> > w/o check if it's hard_regno_mode_ok.
> > It's impossible to put modes whose size > 8 into MASK_REGS, ajdust the cost to
> > avoid potential performance issue.
>
> I was going to ask to open a PR in order to fix RA instead of the
> proposed workaround, but the patch just prevents a theoretical issue
> and follows the established practice with other regclasses. Perhaps
> these workarounds are not needed with the current RA, but this is a
> tangential issue which should be fixed for all regclasses.
Opened https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109351
>
> > Bootstrapped and regtested on x86_64-pc-linux-gnu{-m32,}.
> > Ok for trunk?
>
> OK.
>
> Thanks,
> Uros.
>
> >
> > gcc/ChangeLog:
> >
> >         * config/i386/i386.cc (inline_memory_move_cost): Return 100
> >         for MASK_REGS when MODE_SIZE > 8.
> > ---
> >  gcc/config/i386/i386.cc | 5 ++++-
> >  1 file changed, 4 insertions(+), 1 deletion(-)
> >
> > diff --git a/gcc/config/i386/i386.cc b/gcc/config/i386/i386.cc
> > index 2cc8e9548a9..2581b800a06 100644
> > --- a/gcc/config/i386/i386.cc
> > +++ b/gcc/config/i386/i386.cc
> > @@ -19847,9 +19847,12 @@ inline_memory_move_cost (machine_mode mode, enum reg_class regclass, int in)
> >           index = 1;
> >           break;
> >         /* DImode loads and stores assumed to cost the same as SImode.  */
> > -       default:
> > +       case 4:
> > +       case 8:
> >           index = 2;
> >           break;
> > +       default:
> > +         return 100;
> >         }
> >
> >        if (in == 2)
> > --
> > 2.39.1.388.g2fc9e9ca3c
> >



-- 
BR,
Hongtao