Re: [patch] Performance patch for MIPS conditional move in expr.c

[Andrew Pinski <andrew dot pinski at caviumnetworks dot com>]

On Wed, Nov 14, 2012 at 1:51 PM, Andrew Pinski
<andrew.pinski@caviumnetworks.com> wrote:
> On Wed, Nov 14, 2012 at 1:45 PM, Steve Ellcey <sellcey@mips.com> wrote:
>> On Wed, 2012-11-14 at 12:00 -0800, Andrew Pinski wrote:
>>
>>> I know exactly where this code comes from; I have looked at the
>>> benchmark as one of the reason why I add expand_cond_expr_using_cmove
>>> in the first place.  Anyways you should look into removing
>>> TARGET_PROMOTE_PROTOTYPES because I found that also fixes the problem
>>> mentioned here.
>>>
>>> Thanks,
>>> Andrew Pinski
>>
>> Removing TARGET_PROMOTE_PROTOTYPES looks interesting but I don't know if
>> it is possible for compatibility reasons.  I am still looking at my
>> example though, I see GCC doing:
>>
>> andi    $5,$5,0x1
>> xori    $5,$5,0x1
>> movz    $2,$4,$5
>>
>> When it should just do:
>>
>> andi    $5,$5,0x1
>> movn    $2,$4,$5
>
> Yes I have a few patches for improving this case.  I have not
> submitted them yet though.
> Attached is the assembly I get from a 4.7 toolchain with all of the
> changes I have internally applied; this is for n32 (though o32
> produces the exact same code in this case).  I will try to post some
> more in the next coming weeks.

Removing Richard B. from the CC list since this is now a MIPS specific
change (the original patch is still needed though).
Here is the patch which should improve the above case.  I have not
tested it on the trunk yet but it applies there.

Basically the *mov<GPR:mode>_on_<GPR2:mode>_ne pattern is the one
which is needed in this case.  The other two changes are done for
64bit comparisons.

Thanks,
Andrew Pinski




>
> Thanks,
> Andrew Pinski
>
>
>>
>>
>> Steve Ellcey
>> sellcey@mips.com
>>
>>

Attachment: improvemips_cond_expr.diff.txt
Description: Text document