1 |
Branko Badrljica <brankob@××××××××××.com> posted |
2 |
49FCA81D.4040104@××××××××××.com, excerpted below, on Sat, 02 May 2009 |
3 |
22:07:57 +0200: |
4 |
|
5 |
> BTW: How much of a performance gain would be if one could compile |
6 |
> everything with -profile-generate, run resulting program on some |
7 |
> representative test data and then use profile files to generate final |
8 |
> binary with -fprofile-use ? |
9 |
|
10 |
In theory, that's more likely to give you a decent gain than most of |
11 |
those fancy optimizations you listed. However, few people do it for one- |
12 |
off or only a few users, as the time to compile, profile, and recompile, |
13 |
is way more than the time you'd save by the better optimizations, for |
14 |
just a few users over the life of most packages. |
15 |
|
16 |
OTOH, for something like folding@home or the like that you'll be running |
17 |
24/7 at full CPU untilization, it's likely to make sense, as it would for |
18 |
say Red Hat for x86 especially for packages like glibc, xorg, etc, that |
19 |
most folks use, since they likely have tens of thousands of people |
20 |
running the same binaries, often for a much longer time between releases, |
21 |
than Gentoo has. That's an entirely different scenario than every Gentoo |
22 |
user compiling his own, meanwhile, pointing out one of the weaknesses in |
23 |
the Gentoo setup, since that profile data really doesn't make sense to |
24 |
collect and use on an individual basis for the vast majority of usage |
25 |
scenarios. |
26 |
|
27 |
-- |
28 |
Duncan - List replies preferred. No HTML msgs. |
29 |
"Every nonfree program has a lord, a master -- |
30 |
and if you use the program, he is your master." Richard Stallman |