Steven Bosscher wrote:
On 4/12/07, Vladimir Makarov <[EMAIL PROTECTED]> wrote:
SPECFp2000 compilation time (user time):
machine mainline branch change
---------------------------------
x86_64 104.8s 117.7s +12.3%
ppc64 312.3s 367.8s +17.8%
ia64 377.6s 502.9s +33.2%
Hi Vlad,
Thanks for testing this. Do you also have per benchmark compilation
times, perhaps?
Not really. I don't do that because runtest startup is about 0.4s (on
ppc64) and a few fp tests are compiled for 1.5s. If you are interesting
in analyzing a reason for slowdown, I'd recommend to look at fma3d. For
ppc64 It is compiled for 2m9.6s on the branch and 1m43.7s on the
mainline last merge point. For ia64 it is compiled for 3m20.8s on the
branch and for 2m9.6s on the mainline.
I think that bigger compilation time is partially because of bigger code
generated on the branch. It is very hard to analyze the code size
difference on x86_64 (because the same insns with different registers
e.g. xmm0 and xmm8 or dx and r8 have different length) or on ia64
(because of different number of nops used to fill bundles). PPC64 is
the best target for this because you need to check only number of insns.
Here is the code size difference
2nd column - text segment size on the branch
3rd column - text segment size on mainline on the merge point
PPC64
----------------CFP2000-----------------
-1.553% 29882 29418 168.wupwise
-1.904% 10925 10717 171.swim
-1.934% 17374 17038 172.mgrid
-1.672% 50717 49869 173.applu
-0.994% 587269 581429 177.mesa
-1.539% 230755 227203 178.galgel
-1.128% 18440 18232 179.art
-0.309% 20698 20634 183.equake
-1.094% 68741 67989 187.facerec
-0.780% 135440 134384 188.ammp
-1.024% 45301 44837 189.lucas
-0.425% 1005864 1001592 191.fma3d
-0.316% 870011 867259 200.sixtrack
-1.171% 139383 137751 301.apsi
