* [tree-ssa]: darwin SPEC numbers
@ 2004-01-20 20:08 Dale Johannesen
2004-01-20 20:14 ` Andreas Jaeger
0 siblings, 1 reply; 5+ messages in thread
From: Dale Johannesen @ 2004-01-20 20:08 UTC (permalink / raw)
To: gcc; +Cc: Dale Johannesen
This compares results of tree-ssa branch to (FSF) mainline on Darwin
SPEC,
in case anyone's interested. They're in line with what people are
seeing on Intel.
(gcc and vortex contain code that uses nonstandard aliasing.)
We're more interested in comparing to Apple's 3.3-based branch, but it
looks
like the absence of tree-ssa-based loop optimizations is the main thing
hurting
floating point. It looks like the RTL-based ones do not do as well as
they did in 3.4,
because the RTL is different. Some degradation may actually have
occurred between
FSF 3.3 and 3.4 (I know the rewritten loop unroller was much inferior
on ppc at one point)
but I don't have measurements on that.
We did have to work through several bugs to get tree-ssa to work.
mainline tree-ssa
164.gzip 740 731
175.vpr 515 531
176.gcc ----- -----
181.mcf 526 513
186.crafty 1028 1020
197.parser 625 625
252.eon 949 749
253.perlbmk 804 820
254.gap 748 785
255.vortex ----- -----
256.bzip2 621 618
300.twolf 608 604
GEO MEAN 699 685 (-2%)
168.wupwise 883 899
171.swim 210 245
172.mgrid 600 533
173.applu 663 588
177.mesa 918 921
178.galgel 1010 937
179.art 655 647
183.equake 997 882
187.facerec 1151 1059
188.ammp 440 454
189.lucas 891 858
191.fma3d 984 979
200.sixtrack 546 520
301.apsi 571 569
GEO MEAN 695 675 (-3%)
^ permalink raw reply [flat|nested] 5+ messages in thread
* Re: [tree-ssa]: darwin SPEC numbers
2004-01-20 20:08 [tree-ssa]: darwin SPEC numbers Dale Johannesen
@ 2004-01-20 20:14 ` Andreas Jaeger
2004-01-20 20:21 ` Dale Johannesen
0 siblings, 1 reply; 5+ messages in thread
From: Andreas Jaeger @ 2004-01-20 20:14 UTC (permalink / raw)
To: Dale Johannesen; +Cc: gcc
[-- Attachment #1: Type: text/plain, Size: 1922 bytes --]
Dale Johannesen <dalej@apple.com> writes:
> [...]
> mainline tree-ssa
Are these runtimes or SPEC numbers (1/s) ?
> 164.gzip 740 731
> 175.vpr 515 531
> 176.gcc ----- -----
> 181.mcf 526 513
> 186.crafty 1028 1020
> 197.parser 625 625
> 252.eon 949 749
> 253.perlbmk 804 820
> 254.gap 748 785
> 255.vortex ----- -----
> 256.bzip2 621 618
> 300.twolf 608 604
> GEO MEAN 699 685 (-2%)
>
> 168.wupwise 883 899
> 171.swim 210 245
> 172.mgrid 600 533
> 173.applu 663 588
> 177.mesa 918 921
> 178.galgel 1010 937
> 179.art 655 647
> 183.equake 997 882
> 187.facerec 1151 1059
> 188.ammp 440 454
> 189.lucas 891 858
> 191.fma3d 984 979
> 200.sixtrack 546 520
> 301.apsi 571 569
> GEO MEAN 695 675 (-3%)
Andreas
--
Andreas Jaeger, aj@suse.de, http://www.suse.de/~aj
SuSE Linux AG, Maxfeldstr. 5, 90409 Nürnberg, Germany
GPG fingerprint = 93A3 365E CE47 B889 DF7F FED1 389A 563C C272 A126
[-- Attachment #2: Type: application/pgp-signature, Size: 188 bytes --]
^ permalink raw reply [flat|nested] 5+ messages in thread
* Re: [tree-ssa]: darwin SPEC numbers
2004-01-20 20:14 ` Andreas Jaeger
@ 2004-01-20 20:21 ` Dale Johannesen
2004-01-20 20:56 ` Andreas Jaeger
0 siblings, 1 reply; 5+ messages in thread
From: Dale Johannesen @ 2004-01-20 20:21 UTC (permalink / raw)
To: Andreas Jaeger; +Cc: gcc, Dale Johannesen
On Jan 20, 2004, at 12:14 PM, Andreas Jaeger wrote:
> Dale Johannesen <dalej@apple.com> writes:
>
>> [...]
>> mainline tree-ssa
>
> Are these runtimes or SPEC numbers (1/s) ?
SPEC numbers. 2 GHz G5.
-O3 -funroll-loops -mcpu=G5 -mpowerpc64 -mdynamic-no-pic -ffast-math
-fprofile-arcs/-freorder-blocks -Wa,-force_cpusubtype_ALL
>> 164.gzip 740 731
>> 175.vpr 515 531
>> 176.gcc ----- -----
>> 181.mcf 526 513
>> 186.crafty 1028 1020
>> 197.parser 625 625
>> 252.eon 949 749
>> 253.perlbmk 804 820
>> 254.gap 748 785
>> 255.vortex ----- -----
>> 256.bzip2 621 618
>> 300.twolf 608 604
>> GEO MEAN 699 685 (-2%)
>>
>> 168.wupwise 883 899
>> 171.swim 210 245
>> 172.mgrid 600 533
>> 173.applu 663 588
>> 177.mesa 918 921
>> 178.galgel 1010 937
>> 179.art 655 647
>> 183.equake 997 882
>> 187.facerec 1151 1059
>> 188.ammp 440 454
>> 189.lucas 891 858
>> 191.fma3d 984 979
>> 200.sixtrack 546 520
>> 301.apsi 571 569
>> GEO MEAN 695 675 (-3%)
>
>
> Andreas
> --
> Andreas Jaeger, aj@suse.de, http://www.suse.de/~aj
> SuSE Linux AG, Maxfeldstr. 5, 90409 Nürnberg, Germany
> GPG fingerprint = 93A3 365E CE47 B889 DF7F FED1 389A 563C C272 A126
^ permalink raw reply [flat|nested] 5+ messages in thread
* Re: [tree-ssa]: darwin SPEC numbers
2004-01-20 20:21 ` Dale Johannesen
@ 2004-01-20 20:56 ` Andreas Jaeger
2004-01-20 21:05 ` Dale Johannesen
0 siblings, 1 reply; 5+ messages in thread
From: Andreas Jaeger @ 2004-01-20 20:56 UTC (permalink / raw)
To: Dale Johannesen; +Cc: gcc
[-- Attachment #1: Type: text/plain, Size: 2726 bytes --]
Dale Johannesen <dalej@apple.com> writes:
> On Jan 20, 2004, at 12:14 PM, Andreas Jaeger wrote:
>> Dale Johannesen <dalej@apple.com> writes:
>>
>>> [...]
>>> mainline tree-ssa
>>
>> Are these runtimes or SPEC numbers (1/s) ?
>
> SPEC numbers. 2 GHz G5.
Ok.
> -O3 -funroll-loops -mcpu=G5 -mpowerpc64 -mdynamic-no-pic -ffast-math
> -fprofile-arcs/-freorder-blocks -Wa,-force_cpusubtype_ALL
So, let's look at some differences:
>>> 164.gzip 740 731
>>> 175.vpr 515 531
>>> 176.gcc ----- -----
>>> 181.mcf 526 513
>>> 186.crafty 1028 1020
>>> 197.parser 625 625
>>> 252.eon 949 749
This is really the main difference and fixing it would mean that
tree-ssa looks much better....
>>> 253.perlbmk 804 820
>>> 254.gap 748 785
>>> 255.vortex ----- -----
>>> 256.bzip2 621 618
>>> 300.twolf 608 604
>>> GEO MEAN 699 685 (-2%)
>>>
>>> 168.wupwise 883 899
>>> 171.swim 210 245
Those two above are fine.
>>> 172.mgrid 600 533
This one has to be considered but it's Fortran.
>>> 173.applu 663 588
>>> 177.mesa 918 921
>>> 178.galgel 1010 937
galgel and applu are also Fortran, I would not count any Fortran
program right now as significant.
>>> 179.art 655 647
>>> 183.equake 997 882
equake is C, so this is significant.
>>> 187.facerec 1151 1059
>>> 188.ammp 440 454
>>> 189.lucas 891 858
>>> 191.fma3d 984 979
>>> 200.sixtrack 546 520
>>> 301.apsi 571 569
How did you get e.g 301.apsi working with tree-ssa? This needs a
fortran compiler and gfortran is not able to build it (at least not
for me)....
Andreas
--
Andreas Jaeger, aj@suse.de, http://www.suse.de/~aj
SuSE Linux AG, Maxfeldstr. 5, 90409 Nürnberg, Germany
GPG fingerprint = 93A3 365E CE47 B889 DF7F FED1 389A 563C C272 A126
[-- Attachment #2: Type: application/pgp-signature, Size: 188 bytes --]
^ permalink raw reply [flat|nested] 5+ messages in thread
* Re: [tree-ssa]: darwin SPEC numbers
2004-01-20 20:56 ` Andreas Jaeger
@ 2004-01-20 21:05 ` Dale Johannesen
0 siblings, 0 replies; 5+ messages in thread
From: Dale Johannesen @ 2004-01-20 21:05 UTC (permalink / raw)
To: Andreas Jaeger; +Cc: gcc, Dale Johannesen
On Jan 20, 2004, at 12:56 PM, Andreas Jaeger wrote:
> Dale Johannesen <dalej@apple.com> writes:
>
> So, let's look at some differences:
>
>>>> 164.gzip 740 731
>>>> 175.vpr 515 531
>>>> 176.gcc ----- -----
>>>> 181.mcf 526 513
>>>> 186.crafty 1028 1020
>>>> 197.parser 625 625
>>>> 252.eon 949 749
>
> This is really the main difference and fixing it would mean that
> tree-ssa looks much better....
Yes. It's very sensitive to inlining, has anybody been making
any x86-oriented changes to the defaults?
>>>> 253.perlbmk 804 820
>>>> 254.gap 748 785
>>>> 255.vortex ----- -----
>>>> 256.bzip2 621 618
>>>> 300.twolf 608 604
>>>> GEO MEAN 699 685 (-2%)
>>>>
>>>> 168.wupwise 883 899
>>>> 171.swim 210 245
>
> Those two above are fine.
>
>>>> 172.mgrid 600 533
>
> This one has to be considered but it's Fortran.
>
>>>> 173.applu 663 588
>>>> 177.mesa 918 921
>>>> 178.galgel 1010 937
>
> galgel and applu are also Fortran, I would not count any Fortran
> program right now as significant.
>
>>>> 179.art 655 647
>>>> 183.equake 997 882
>
> equake is C, so this is significant.
>
>>>> 187.facerec 1151 1059
>>>> 188.ammp 440 454
>>>> 189.lucas 891 858
>>>> 191.fma3d 984 979
>>>> 200.sixtrack 546 520
>>>> 301.apsi 571 569
>
> How did you get e.g 301.apsi working with tree-ssa? This needs a
> fortran compiler and gfortran is not able to build it (at least not
> for me)....
We use the NAG non-free compiler (actually F-to-C translator) as we
do for published numbers. So the Fortran ones are actually using
gcc and probably are significant.
^ permalink raw reply [flat|nested] 5+ messages in thread
end of thread, other threads:[~2004-01-20 21:05 UTC | newest]
Thread overview: 5+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2004-01-20 20:08 [tree-ssa]: darwin SPEC numbers Dale Johannesen
2004-01-20 20:14 ` Andreas Jaeger
2004-01-20 20:21 ` Dale Johannesen
2004-01-20 20:56 ` Andreas Jaeger
2004-01-20 21:05 ` Dale Johannesen
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).