This is the mail archive of the
gdb-patches@sourceware.org
mailing list for the GDB project.
Re: [rfc] Fix PowerPC displaced stepping regression
- From: Pedro Alves <pedro at codesourcery dot com>
- To: gdb-patches at sourceware dot org
- Cc: "Ulrich Weigand" <uweigand at de dot ibm dot com>, Julian Brown <julian at codesourcery dot com>, Daniel Jacobowitz <drow at false dot org>
- Date: Mon, 28 Sep 2009 20:07:22 +0100
- Subject: Re: [rfc] Fix PowerPC displaced stepping regression
- References: <200909281745.n8SHjIkG030649@d12av02.megacenter.de.ibm.com>
On Monday 28 September 2009 18:45:18, Ulrich Weigand wrote:
> Pedro Alves wrote:
>
> > On Monday 28 September 2009 18:27:03, Ulrich Weigand wrote:
> > > Maybe I misunderstood your point here, but I don't think we can
> > > actually do SW single-step on the displaced copy (using the normal
> > > SW single-step mechanism). The way SW single-step ususally works
> > > is to place breakpoints at all potential branch targets. But if
> > > we have a displaced PC-relative branch, for example, the branch
> > > target may not even point to addressable memory, so we cannot put
> > > breakpoints there.
> >
> > If you get yourself such an instruction in the buffer, usually you'd
> > want the branch offset had to be adjusted at displaced copy time,
> > otherwise it seems to be you're already broken.
>
> If that's possible. In general, the real branch target may be out of
> range relative to the address of the copied instruction for a branch in
> the original instruction format ... (You could redirect to some temporary
> target in the copy buffer, but at this point you're probably better off
> just emulating the whole thing in the first place.)
Yes, of course. But, the point is that whatever ends up in the
displaced step scratch pad after displaced_step_copy time, be it simply
a copy of the original insn, an adjusted pc-relative instruction, or
a sequence of insns emulating the original insn, _could_ be single-stepped
using software-single stepping. It's the latter case of single instruction
emulation with more than one insn that is generatly more efficient to
execute in one go with a break+continue, irrespective or HW or software
single stepping being supported.
Anyway, we're both clearly aware of these issues, and getting off topic. :-)
--
Pedro Alves