Re: Profiling Octave

[Paul Thomas]

Clinton Chee wrote:

> Hi Paul, Ben, others,
>
> From Paul's recent email see below, I get the impression that octave 
> can be compiled with gprof (-pg). However, when I build octave, it 
> compiles successfully (I think) with the -pg but when I run it, I get:
>
> /home/chee/programs/octave-2.1.57$>src/octave
> panic: Profiling timer expired -- stopping myself...
> Profiling timer expired
Clinton,

I have to confess to having blundered around somewhat to arrive at a 
build that worked and depended on helpful remarks from others to get 
there.  Also, although it should not make any difference, Ben and I were 
working under Cygwin.  I and, I rather think, Ben do not claim any 
expertise in setting up profiling under gcc.

That not withstanding, I will make one remark:  We could not get the 
build with -pg to complete with shared enabled.  The linker complained 
of an unrequited reference to  __monstartup which is to be found in 
gcrt1.o. Looking at the specs file for gcc-3.2.2 on RH9, there appears 
to be a similar exclusion:

*startfile:
%{!shared:      %{pg:gcrt1.o%s} %{!pg:%{p:gcrt1.o%s}                
%{!p:%{profile:gcrt1.o%s}              %{!profile:crt1.o%s}}}}    
crti.o%s %{static:crtbeginT.o%s}   %{!static:%{!shared:crtbegin.o%s} 
%{shared:crtbeginS.o%s}}

This prevents gcrt1.o from being loaded unless share is disabled.  We 
took this at face value and satisfied the reference by doing just that. 
The other approach would be to link gcrt1.o explicitly.  I have no idea 
if this will work or not.

Note also, it is necessary to set -pg in the compile and the link steps.

A final remark concerns the profiling output; whilst there appears to be 
reasonable proportionality between the reported times and the elapsed 
cputime, in absolute terms it is out by a factor of a few.  I have no 
idea, therefore, of how much trust can be placed in the profiling.  I am 
exploring this with a smaller, simpler code before returning to octave.

Paul T