[PATCH] i386: Clear REG_UNUSED and REG_DEAD notes from the IL at the end of vzeroupper pass [PR113059]

[PATCH] i386: Clear REG_UNUSED and REG_DEAD notes from the IL at the end of vzeroupper pass [PR113059]

[Jakub Jelinek]

Hi!

The move of the vzeroupper pass from after reload pass to after
postreload_cse helped only partially, CSE-like passes can still invalidate
those notes (especially REG_UNUSED) if they use some earlier register
holding some value later on in the IL.

So, either we could try to move it one pass further after gcse2 and hope
no later pass invalidates the notes, or the following patch attempts to
restore the REG_DEAD/REG_UNUSED state from GCC 13 and earlier, where
the LRA or reload passes remove all REG_DEAD/REG_UNUSED notes and the notes
reappear only at the start of dse2 pass when it calls
  df_note_add_problem ();
  df_analyze ();
So, effectively
          NEXT_PASS (pass_postreload_cse);
          NEXT_PASS (pass_gcse2);
          NEXT_PASS (pass_split_after_reload);
          NEXT_PASS (pass_ree);
          NEXT_PASS (pass_compare_elim_after_reload);
          NEXT_PASS (pass_thread_prologue_and_epilogue);
passes operate without those notes in the IL.
While in GCC 14 mode switching computes the notes problem at the start of
vzeroupper, the patch below removes them at the end of the pass again, so
that the above passes continue to operate without them.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

2024-01-31  Jakub Jelinek  <jakub@redhat.com>

	PR target/113059
	* config/i386/i386-features.cc (rest_of_handle_insert_vzeroupper):
	Remove REG_DEAD/REG_UNUSED notes at the end of the pass before
	df_analyze call.

--- gcc/config/i386/i386-features.cc.jj	2024-01-08 12:15:13.611477047 +0100
+++ gcc/config/i386/i386-features.cc	2024-01-30 12:36:27.834515803 +0100
@@ -2664,6 +2664,32 @@ rest_of_handle_insert_vzeroupper (void)
   /* Call optimize_mode_switching.  */
   g->get_passes ()->execute_pass_mode_switching ();
 
+  /* LRA removes all REG_DEAD/REG_UNUSED notes and normally they
+     reappear in the IL only at the start of pass_rtl_dse2, which does
+     df_note_add_problem (); df_analyze ();
+     The vzeroupper is scheduled after postreload_cse pass and mode
+     switching computes the notes as well, the problem is that e.g.
+     pass_gcse2 doesn't maintain the notes, see PR113059 and
+     PR112760.  Remove the notes now to restore status quo ante
+     until we figure out how to maintain the notes or what else
+     to do.  */
+  basic_block bb;
+  rtx_insn *insn;
+  FOR_EACH_BB_FN (bb, cfun)
+    FOR_BB_INSNS (bb, insn)
+      if (NONDEBUG_INSN_P (insn))
+	{
+	  rtx *pnote = &REG_NOTES (insn);
+	  while (*pnote != 0)
+	    {
+	      if (REG_NOTE_KIND (*pnote) == REG_DEAD
+		  || REG_NOTE_KIND (*pnote) == REG_UNUSED)
+		*pnote = XEXP (*pnote, 1);
+	      else
+		pnote = &XEXP (*pnote, 1);
+	    }
+	}
+
   df_analyze ();
   return 0;
 }

	Jakub

Re: [PATCH] i386: Clear REG_UNUSED and REG_DEAD notes from the IL at the end of vzeroupper pass [PR113059]

[Uros Bizjak]

On Wed, Jan 31, 2024 at 9:23 AM Jakub Jelinek <jakub@redhat.com> wrote:
>
> Hi!
>
> The move of the vzeroupper pass from after reload pass to after
> postreload_cse helped only partially, CSE-like passes can still invalidate
> those notes (especially REG_UNUSED) if they use some earlier register
> holding some value later on in the IL.
>
> So, either we could try to move it one pass further after gcse2 and hope
> no later pass invalidates the notes, or the following patch attempts to
> restore the REG_DEAD/REG_UNUSED state from GCC 13 and earlier, where
> the LRA or reload passes remove all REG_DEAD/REG_UNUSED notes and the notes
> reappear only at the start of dse2 pass when it calls
>   df_note_add_problem ();
>   df_analyze ();
> So, effectively
>           NEXT_PASS (pass_postreload_cse);
>           NEXT_PASS (pass_gcse2);
>           NEXT_PASS (pass_split_after_reload);
>           NEXT_PASS (pass_ree);
>           NEXT_PASS (pass_compare_elim_after_reload);
>           NEXT_PASS (pass_thread_prologue_and_epilogue);
> passes operate without those notes in the IL.
> While in GCC 14 mode switching computes the notes problem at the start of
> vzeroupper, the patch below removes them at the end of the pass again, so
> that the above passes continue to operate without them.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
>
> 2024-01-31  Jakub Jelinek  <jakub@redhat.com>
>
>         PR target/113059
>         * config/i386/i386-features.cc (rest_of_handle_insert_vzeroupper):
>         Remove REG_DEAD/REG_UNUSED notes at the end of the pass before
>         df_analyze call.

Not really a review, but let's rubber stamp this workaround OK.

Thanks,
Uros.

>
> --- gcc/config/i386/i386-features.cc.jj 2024-01-08 12:15:13.611477047 +0100
> +++ gcc/config/i386/i386-features.cc    2024-01-30 12:36:27.834515803 +0100
> @@ -2664,6 +2664,32 @@ rest_of_handle_insert_vzeroupper (void)
>    /* Call optimize_mode_switching.  */
>    g->get_passes ()->execute_pass_mode_switching ();
>
> +  /* LRA removes all REG_DEAD/REG_UNUSED notes and normally they
> +     reappear in the IL only at the start of pass_rtl_dse2, which does
> +     df_note_add_problem (); df_analyze ();
> +     The vzeroupper is scheduled after postreload_cse pass and mode
> +     switching computes the notes as well, the problem is that e.g.
> +     pass_gcse2 doesn't maintain the notes, see PR113059 and
> +     PR112760.  Remove the notes now to restore status quo ante
> +     until we figure out how to maintain the notes or what else
> +     to do.  */
> +  basic_block bb;
> +  rtx_insn *insn;
> +  FOR_EACH_BB_FN (bb, cfun)
> +    FOR_BB_INSNS (bb, insn)
> +      if (NONDEBUG_INSN_P (insn))
> +       {
> +         rtx *pnote = &REG_NOTES (insn);
> +         while (*pnote != 0)
> +           {
> +             if (REG_NOTE_KIND (*pnote) == REG_DEAD
> +                 || REG_NOTE_KIND (*pnote) == REG_UNUSED)
> +               *pnote = XEXP (*pnote, 1);
> +             else
> +               pnote = &XEXP (*pnote, 1);
> +           }
> +       }
> +
>    df_analyze ();
>    return 0;
>  }
>
>         Jakub
>