Core2 and Corei7 tuning patches for review

[Maxim Kuvyrkov]

On 10/19/10 4:17 PM, H.J. Lu wrote:
> On Tue, Oct 19, 2010 at 5:02 AM, Maxim Kuvyrkov<maxim@codesourcery.com>  wrote:
>> On 10/19/10 3:57 PM, H.J. Lu wrote:
>>>
>>> On Tue, Oct 19, 2010 at 4:21 AM, Maxim Kuvyrkov<maxim@codesourcery.com>
>>>   wrote:
>>>>
>>>> I'll be posting Core2 and Corei7 tuning patches in this thread for
>>>> review.
>>>>
>>>> Last week I posted benchmark results and analysis which show impact of
>>>> individual changes on the performance [*].  In this thread I'll be asking
>>>> for review of some of those patches that positively affect performance.
>>>>
>>>> [*] http://gcc.gnu.org/ml/gcc-patches/2010-10/msg01318.html
>>>>
>>>
>>> Did you run SPEC CPU 2006 on 64bit Core i7? It failed to compile
>>> 400.perlbench:
>>>
>>> gcc -c -o pp_hot.o -DSPEC_CPU -DNDEBUG -DPERL_CORE  -O3 -funroll-loops
>>> -ffast-math -mtune=corei7       -DSPEC_CPU_LP64  -DSPEC_CPU_LINUX_X64
>>> -fno-strict-aliasing       perly.c
>>> perly.c: In function 'Perl_yyparse':
>>> perly.c:2531:1: internal compiler error: in
>>> core2_first_cycle_multipass_issue, at config/i386/i386.c:21775
>>
>> Which patch set are you testing (the initial, or the one I sent later)?  I
>> fixed this or a very similar ICE in the later set of patches.
>
> I tested:
>
> http://gcc.gnu.org/ml/gcc-patches/2010-10/msg01318.html
>
> This is the only set sent to gcc-patches. Please send me the updated
> set of changes. The last set you sent me directly has patches which
> just cancel each other:
>
> 0023-issue_rate-4.patch
> 0019-Set-issue_rate-to-3.patch

For the record, here is the patch set I have so far, it it is the same 
as I sent H.J. earlier.  The patch 0016 in this set (add modeling of 
Core 2/i7 decoder bottleneck) has the above ICE fixed.  Other patches 
may have minor changes relative to the initial patch set posted to 
gcc-patches@.

> I'd like to test the whole set first since it takes several days to run
> SPEC CPU 2K/2006.

H.J.,

I'm not sure there is much value in retesting the patches I posted 
together with tuning analysis.  It may be better use of machine time to 
benchmark the patches I'm posting now to double check the performance 
improvements I claim.  I expect that every patch posted in this thread 
will improve performance.

Does this seem reasonable?

Thank you,

-- 
Maxim Kuvyrkov
CodeSourcery
maxim@codesourcery.com
(650) 331-3385 x724
-------------- next part --------------
A non-text attachment was scrubbed...
Name: core2i7-3.tar.bz2
Type: application/x-bzip2
Size: 21610 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20101019/fa4ece83/attachment.bz2>