This is the mail archive of the
gdb-patches@sourceware.org
mailing list for the GDB project.
Re: [commit] Improve performance with lots of shared libraries
- From: Pedro Alves <pedro at codesourcery dot com>
- To: Gary Benson <gbenson at redhat dot com>
- Cc: gdb-patches at sourceware dot org
- Date: Wed, 12 Oct 2011 17:15:41 +0100
- Subject: Re: [commit] Improve performance with lots of shared libraries
- References: <20110922171206.GB5874@redhat.com> <201110111753.27321.pedro@codesourcery.com> <20111012155918.GA4216@redhat.com>
On Wednesday 12 October 2011 16:59:18, Gary Benson wrote:
> Pedro Alves wrote:
> > To make this generic for all breakpoints/stops, what I have in mind
> > would be:
> >
> > - at breakpoint creation or re-set time, check if the locations
> > we've created point at inlined code, and set a flag in the
> > breakpoint's locations. We know the location is inlined or not
> > from the debug info. Breakpoint creation is the slow path, so
> > that's okay.
> >
> > - given that we need to single-step over those breakpoints, we also
> > need to know whether the PC after stepping over those breakpoints
> > points at inlined code. I think we can still do that at
> > breakpoint creation or re-set time. We'd need to reuse the
> > software single-step machinery to know where the single-step
> > would take us, and record somewhere that those locations point to
> > inline code or not. We'd also check this list in
> > stopped_at_non_inline_function. The software single-step
> > machinery would need some cleaning up to make this possible.
> > It's interface, gdbarch_software_single_step, isn't fit for this.
> > The gdbarch hook should return a list of locations where to put
> > the breakpoint, instead of implementations planting the
> > breakpoints themselves, which would be a nice cleanup for other
> > things too. We'd also need to implement this hook for x86. It's
> > not implemented currently because x86 can do hardware
> > single-stepping.
>
> Ah, nice! Would it be appropriate to file a bug containing this
> information? So it doesn't get lost before I have a chance to work
> on it?
Sure! What I haven't thought about much is whether this
optimization would be indeed a general win. :-) It'd make a
difference if you tend to have planted breakpoints that
don't cause a stop often (e.g., some python breakpoint),
and maybe it'd make a difference on software single-step
targets, and a tiny bit on handling step-resume
breakpoints on hardware step targets? I don't have a clear
picture where time is being spent (other than roundtripping
to the target). Thread event breakpoints sound like low
hang fruit though.
--
Pedro Alves