Re: GCC gprof statistics

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

On Sunday, Jun 29, 2003, at 17:07 US/Eastern, Ishikawa wrote:

> Dear Andrew Pinski,
>
> Care to share information about Shark tool?

It is part of the CHUD tools 3.0.0 (only in beta right now) 
<http://developer.apple.com/tools/performance/>
from Apple for Mac OS X, it used to be called Shikari in CHUD 2.0.
It provides what vTunes does but I like it better (I help to develop 
some of Shark).
It provides assembly and source level profiling, it is a statical one.
It gives you a backtrace for each function.
It also can use the performance monitors counters (PMCs) on the CPU, so 
you can look at miss-predicting of branches or
how many unaligned access happened and where.  It also can use the PMCs 
on the memory controller or inside the OS.
Other information about Shark can be found here 
<http://developer.apple.com/technotes/tn/tn2086.html>, it is about
Apple's new computer the G5 (IBM's 970) but it gives a good overview of 
Shark.


> I used google, but I hit so many movie
> (Jaws?) pages and got lost.
>
> > Using Shark from the CHUD tools, sampling for 30 seconds, here are the
> > top 3 functions which call expr_equiv_p:
> >         8.2%      ldst_entry    cc1
> >         3.1%      expr_equiv_p  cc1             <--- itself (already 
> > lowered)
> >         2.1%      trim_ld_motion_mems   cc1
>
> Also, it would be interesting to look at the mods
> to for_each_rtx.

I think the only one that will help the most is the other most common 
case of having one 'e'
so it will be optimized to loop for that case, too I will look into to 
see if it is a mod to do.

Thanks,
Andrew Pinski