statistical profiling of gcc

statistical profiling of gcc

[David Morrison]

Hi all,

I'm sure something like this must have been tried before, but on the 
off-change it hasn't, I hacked in a line or two to gcc/main.c to set up 
"jprof" to do statistical profiling on gcc runs.  Here's a link to info 
on jprof:

http://lxr.mozilla.org/mozilla/source/tools/jprof/README.html

It provides fairly detailed information that I've found useful on other 
projects for diagnosing CPU usage hotspots.  I used the resulting 
compiler to compile gcc/expr.o - for no particularly good reason, just a 
nice big file - with and without optimization.  The full results can be 
found here:

http://www.phenix.bnl.gov/~dave/gcc/expr_opt.html

and here:

http://www.phenix.bnl.gov/~dave/gcc/expr_noopt.html

The first part of the "flat" profile is shown below.  The interrupt was 
set for 1 ms and the "Count" should be roughly proportional to CPU time 
spent in that routine.  The full report has a lot more information.

Count   Function Name
750     seen_objects
497     htab_traverse
279     ggc_set_mark
190     count_or_remove_death_notes
148     clear_table
139     compute_transp
117     verify_flow_info
  95     ggc_alloc
  95     side_effects_p
  90     for_each_rtx
  90     gt_ggc_mx_lang_tree_node
  81     varray_clear
  77     constrain_operands
  77     sbitmap_vector_alloc
  72     subst_reloads
  69     gt_ggc_mx_rtx_def

Useful?

Dave

Re: statistical profiling of gcc

[Dale Johannesen]

On Wednesday, February 19, 2003, at 11:43  AM, David Morrison wrote:
> The first part of the "flat" profile is shown below.  The interrupt 
> was set for 1 ms and the "Count" should be roughly proportional to CPU 
> time spent in that routine.  The full report has a lot more 
> information.
>
> Count   Function Name
> 750     seen_objects
> 497     htab_traverse
> 279     ggc_set_mark
> 190     count_or_remove_death_notes

The only "seen_objects" I can find is a data object, so I think jprof 
needs a little work.
Thanks for trying.