Re: some profiling numbers

[Andrew Pinski <pinskia at physics dot uc dot edu>]

On Wednesday, June 25, 2003, at 1:39 PM, graydon hoare wrote:

> hi,
>
> I've recently done some profiling and analysis of gcc and g++, using 
> the oprofile hardware-assisted profiling module in linux 2.5. I 
> collected stats on a variety of versions of gcc (some only of 
> historical interest) and some different branches and workloads, and 
> characterized some aspects of the cache and branch-mispredict 
> penalties in addition to cycle hotspots.
>
> the results of the work are posted at
>      http://people.redhat.com/graydon/gcc-optimizing/
>  and http://people.redhat.com/graydon/g++-report/
>
> any comments or questions are welcome (including instructions and 
> hand-holding on how to set up the profiler for your own runs, if you 
> want to do more of this stuff).

In your c++ report, you site side_effects_p to have a large number of 
branch mispredicts but
there is another problem you did not see, there are invariant loads in 
the 'E' loop as each
time through the loop you have to load XVECLEN (x, i) and XVEC (x, i) 
for XVECEXP (x, i, j).
There are large number of places this happens too.
In comp_template_args there is an invariant load of TREE_VEC_LENGTH 
(oldargs).
These fixes are simple and by making sure there is no extra loads in 
the loop
(but this fix might hurt on x86 but x86 is the old case since there are 
not enough
registers to do anything useful to speed up by removing invariant load).

I used Shark from the CHUD tools on Mac OS X to see that these are 
important fixes and should be fixed.
Also I will look at mispredicted branches on my G4 laptop to see if I 
get anything different, I think
I will but who knows.

Thanks,
Andrew Pinski