ggc_pop_context speedup
Mike Stump
mstump@apple.com
Fri Jan 31 23:13:00 GMT 2003
On Thursday, January 30, 2003, at 11:25 AM, Mike Stump wrote:
> On Thursday, January 30, 2003, at 11:11 AM, Richard Henderson wrote:
>> On Wed, Jan 29, 2003 at 08:40:53PM -0800, Mike Stump wrote:
>>> ggc_pop_context is an utter and complete pig.
>>
>> Hmm, yes. Seems we were working on this simultaneously.
>> Before you check this patch in, will you try my much much
>> simpler patch on your test case?
>>
>> http://gcc.gnu.org/ml/gcc-patches/2003-01/msg02482.html
>
> I'll do that. If ggc_pop_context still comes out in the top 70
> functions, I'm still going to seek the inclusion of my work.
>
> Remember, the goal is a compiler that is 6x (post PCH) faster... We
> ain't there yet.
Nope, no good. In my real benchmark, Finder_FE, your work drops
ggc_pop_context from #3 to about #12, but it leaves most (~50%) of the
cost of the function still on the plate. My patch removes the cost of
the function, not by doing any less work, but by making use of all that
extra processor time during all the memory stalls. :-)
Finder_FE without the popcontext work:
real 620.87
user 855.82
sys 277.09
Finder_FE with popcontext:
real 601.30 3.15% speedup real
user 827.14 3.35% speedup user
sys 267.08 3.61% speedup sys
In absolute terms, 19.57 real, 28.68 user.
Finder_FE with rth's work:
real 610.02
user 833.72
sys 276.15
I'll work on those issues that Geoff pointed out...
More information about the Gcc-patches
mailing list