Re: [PATCH Atom] Fix for PR target/50962 (bad AGU stall avoidance)

[Ilya Enkovich <enkovich dot gnu at gmail dot com>]

2011/11/3 Uros Bizjak <ubizjak@gmail.com>:
> Hello!
>
>> Here is a fix for PR 50962. Fix improves AGU stall avoidance
>> optimization by adding opportunity to emit lea instead of mov when it
>> is profitable.
>>
>> 2011-11-03 ?Enkovich Ilya ?<ilya.enkovich@intel.com>
>>
>> ? ? ? PR target/50962
>> ? ? ? * config/i386/i386-protos.h (ix86_use_lea_for_mov): New.
>> ? ? ? * config/i386/i386.c (ix86_use_lea_for_mov): Likewise.
>> ? ? ? * config/i386/i386.md (movsi_internal): Emit lea if profitable.
>> ? ? ? (movdi_internal_rex64): Likewise.
>
> OK for mainline SVN with a small change, see below.
>
> +/* Return true if we should emit lea instruction instead of mov
> + ? instruction. ?*/
> +
> +bool
> +ix86_use_lea_for_mov (rtx insn, rtx operands[])
> +{
> + ?unsigned int regno0;
> + ?unsigned int regno1;
> +
> + ?/* Check we can transform mov to lea. ?*/
> + ?if (!REG_P (operands[0]) || !REG_P (operands[1]))
> + ? ?return false;
>
> Please fix the comment and ...
>
> + ?/* Check if we need to optimize. ?*/
> + ?if (!TARGET_OPT_AGU || optimize_function_for_size_p (cfun))
> + ? ?return false;
>
> ... switch this check with the one above, it is faster to exit early this way.
>
> + ?regno0 = true_regnum (operands[0]);
> + ?regno1 = true_regnum (operands[1]);
> +
> + ?return ix86_lea_outperforms (insn, regno0, regno1, -1, 0);
> +}
>
> Thanks,
> Uros.
>

Hello!

Thanks for review!

Here is a patch with requested fixes. Could please someone commit it?

Thanks,
Ilya
---
gcc/

2011-11-04  Enkovich Ilya  <ilya.enkovich@intel.com>

	PR target/50962
	* config/i386/i386-protos.h (ix86_use_lea_for_mov): New.
	* config/i386/i386.c (ix86_use_lea_for_mov): Likewise.
	* config/i386/i386.md (movsi_internal): Emit lea if profitable.
	(movdi_internal_rex64): Likewise.

Attachment: PR50962.diff
Description: Binary data