[Gcl-devel] Re: 10%, for what it's worth

[Camm Maguire]

Greetings! Please excuse my silence of late as I've tried to catch up
on work that had accumulated during my vacation.


Robert Boyer <address@hidden> writes:

> It looks like an icc-9 version of GCL is about 9% faster than a gcc version
> on the Nqthm tests under GCL 2.7.0.
> 
>    ICC
> 
>    run-gbc time    :   2362.520 secs
>    child run time  :    115.530 secs
>    gbc time        :     97.970 secs
> 
>    GCC
> 
>    run-gbc time    :   2447.650 secs
>    child run time  :    156.270 secs
>    gbc time        :    215.780 secs
> 
>    (/ (+ 2447.6 156.2 215.8) (+ 2362.5 115.5 98)) =   1.094


Thank you for these very interesting results!

1) I'd very much be interested in the icc compiler warning output that
   bounced in the mail -- perhaps you can tell me where to retrieve!

2) Of course the lion's chare of the improvement is in gc time, and I
   bet I know exactly where -- sweeping the relocatable area.  I'd
   wager icc uses sse and prefetch instructions to do the copy much
   faster.  Thankfully, we can get the same using pure open source
   tools.  gcc-4.0 now spits out sse, and even without it we can make
   use of atlas tuned blas copy routines.  You can confirm/refute in
   part by doing objjdump -d gbc.o | grep xmm

3) I'd love to see a breakdown preferably by routine of the run-gbc
   time improvements.  I don't suppose icc speaks gprof?

Take care,

> 
> Bob
> 
> 
> 

-- 
Camm Maguire                                            address@hidden
==========================================================================
"The earth is but one country, and mankind its citizens."  --  Baha'u'llah