Mail Archives: cygwin-developers/2001/05/03/17:39:23
Chris! Can I please have DJ's profiler. Pretty please!
If I get this going I'll happily document it.
Rob
----- Original Message -----
From: "DJ Delorie" <dj AT delorie DOT com>
To: <robert DOT collins AT itdomain DOT com DOT au>
Cc: <cygwin-developers AT cygwin DOT com>
Sent: Friday, May 04, 2001 7:30 AM
Subject: Re: DJ Please read - profiling
>
> > I recall you have done serious work on profiling cygwin? I don't
> > have dejagnu running here, nor _any idea_ on how to profile cygwin.
>
> To use gprof, you need to build cygwin as a static library with -pg.
> That in itself is tricky, because you first have to replace malloc
> with some other implementation so that it doesn't recurse (cygwin
> calls the user's malloc, which *is* cygwin's malloc if you link
> statically). You also have to rewrite all the startup code to work in
> a non-dll situation. It's not easy to get cygwin to run this way, and
> expect to end up with a useless mess when you're done profiling.
>
> The next step is the profiler itself - you can use gprof, but the
> granularity is pretty bad for quick-lived things like cygwin. I wrote
> something that single-stepped the program using the debug API and
> built a profile from that; you'd have to ask Chris if I can publish
> it, assuming it still works. It (my profiler) should work with cygwin
> as a DLL, so that makes things easier, but programs run about 1000
> times slower :-( and you have to manually specify the range of
> addresses you want to profile. You can, however, profile individual
> functions via OutputDebugString controls to the profiler.
>
> I don't remember if my profiler infers call graph counts or if you
> have to build cygwin1.dll with -pg. I think it infers the graph.
>
> I will gladly impart my wisdom and experience in detail if someone is
> willing to validate what still currently works and write up a page for
> cygwin.com.
>
> You don't need dejagnu at all, unless that's what you're profiling.
>
- Raw text -