Re: oprofile; Was: [RFA] Kill some linear searches in minsyms.c

[Daniel Jacobowitz <drow at mvista dot com>]

On Tue, Jan 07, 2003 at 07:16:31PM -0500, Andrew Cagney wrote:
> >On Mon, Jan 06, 2003 at 04:02:37PM -0500, Andrew Cagney wrote:
> >
> >>>Future things to examine:
> >
> >>
> >>Now this reminds me.
> >>
> >>Things-to-do-today includes run oprofile on GDB while debugging 
> >>something like mozilla.  That would give a real picture of where GDB is 
> >>spending its time.
> >
> >
> >I have a couple more similar patches for the places we're spending our
> >time :)  After I clear up some more backlog.
> 
> Based on profile or oprofile?
> 
> The problem with profile is that it artifically inflates the call 
> frequency of small functions and that leasily leads to mis-analysis. 
> cf, not so recently where the a finger was pointed at the pid/tid 
> functions as the cause of thread slowness.  The real problem was (and 
> probably still is) too many system calls.

Some of each; this batch is mostly gprof.  For functions with a large
per instance time, this is plenty accurate.  It also gives exact call
counts, which are useful for the accessors even when they're not really
the problem; e.g. the 27M calls to symbol_demangled_name that I could
tell didn't belong there.

-- 
Daniel Jacobowitz
MontaVista Software                         Debian GNU/Linux Developer