Re: [Patch] Use resume instead of target_resume when stepping over watchpoint.

[Pedro Alves <pedro@codesourcery.com>]

On Thursday 30 October 2008 03:08:05, Joel Brobecker wrote:
> Pedro, Others,
> 
> What do you think of this patch? Personally, I have pretty much
> convinced myself that it shouldn't do any harm, but I really
> wished that "resume" would take a ptid as an argument. Except
> that this is not trivial to do, and I think that the current
> "resume" would need to be split a bit, to remove the code that
> determines what to resume.

Won't this slightly change the behaviour on hardware single-step
archs?  Before, we'd always tell the target to resume a single-thread
(keeping the others stopped, on target that support scheduler locking);
with this change, I think you'll tell the target to resume all
threads.

Urglhs, infrun could use a facelift.  The natural function
to call would be keep_going instead of resume/prepare_to_wait,
but keep_going doesn't know a think about watchpoints...
Would setting ecs->event_thread->trap_expected = 1 in addition
to switching to resume so we trigger this:

resume:
     if ((step || singlestep_breakpoints_inserted_p)
	  && tp->trap_expected)
	{
...
	  resume_ptid = inferior_ptid;
        }

be too ugly?  Hmmm, maybe not OK, it can have other side
effects, like tripping on this...

 if (ecs->event_thread->stop_signal == TARGET_SIGNAL_TRAP
      && ecs->event_thread->trap_expected
      && gdbarch_single_step_through_delay_p (current_gdbarch)
      && currently_stepping (ecs->event_thread))
    {

Can you confirm what I think I'm seeing?

> 
> Anyway, I don't see anything wrong with this patch, but I'd love
> for someone to take a look as well. This is a pretty delicate
> part of the debugger.  Do we really need the gdb_assert thought?
> 
> On Tue, Sep 09, 2008 at 05:31:33PM -0700, David Daney wrote:
> > In handle_inferior_event() when stepping over a watch point currently we
> > issue target_resume().  This only works on architectures that have
> > hardware single step support.  For gdbarch_software_single_step_p()
> > systems (like MIPS), we need to insert a single step breakpoint instead.
> > 
> > The fix is to call resume() as it does the right thing already.  I also
> > added an assert that inferior_ptid == ecs->ptid to be sure that resume()
> > was stepping the proper thread.
> > 
> > This is essentially the change requested by Daniel in:
> > http://sourceware.org/ml/gdb-patches/2008-04/msg00443.html
> > 
> > This change is a prerequisite for my forthcoming MIPS hardware watch patch.
> > 
> > Tested on x86_64-pc-linux-gnu as well as mipsel-linux (in conjunction
> > with the MIPS hardware watch patch).
> > 
> > OK to commit?
> > 
> > 2008-09-09  David Daney  <ddaney@avtrex.com>
> > 
> > 	* infrun.c (handle_inferior_event): Call resume instead of
> > 	target_resume when stepping over watchpoint.
> > 
> > 
> 
> > Index: infrun.c
> > ===================================================================
> > RCS file: /cvs/src/src/gdb/infrun.c,v
> > retrieving revision 1.316
> > diff -u -p -r1.316 infrun.c
> > --- infrun.c	8 Sep 2008 22:10:20 -0000	1.316
> > +++ infrun.c	9 Sep 2008 23:37:09 -0000
> > @@ -2472,7 +2472,8 @@ targets should add new threads to the th
> >        if (!HAVE_STEPPABLE_WATCHPOINT)
> >  	remove_breakpoints ();
> >        registers_changed ();
> > -      target_resume (ecs->ptid, 1, TARGET_SIGNAL_0);	/* Single step */
> > +      gdb_assert (ptid_equal (inferior_ptid, ecs->ptid));
> > +      resume (1, TARGET_SIGNAL_0);	/* Single step */
> >        waiton_ptid = ecs->ptid;
> >        if (HAVE_STEPPABLE_WATCHPOINT)
> >  	infwait_state = infwait_step_watch_state;
> 
> 



-- 
Pedro Alves