Re: RFA symtab: Fix for PR c++/1267 ("next" and shared libraries)

[Daniel Jacobowitz <drow at mvista dot com>]

Only the testsuite on i386-linux.  What would you recommend?

On Mon, Jul 21, 2003 at 02:13:55AM -0500, Jim Blandy wrote:
> 
> I think this is a great idea.  How widely have you tested it?
> 
> Daniel Jacobowitz <drow@mvista.com> writes:
> 
> > This patch fixes c++/1267, a bug where stepping over a function call that
> > went through the PLT (as happens when a -fPIC function makes a call to a
> > globally visible symbol) would lose control of the inferior.  I'll spare you
> > the complete debugging session, as it really doesn't make much sense.  But
> > here's the root of the problem:
> > 
> > When we called frame_pc_unwind on the sentinel frame, we got an address in
> > the PLT.  But when we called frame_func_unwind, we got "_init", in ".init",
> > which is generally located right before the PLT.  Then, we'd run the
> > new-and-improved prologue unwinder on _init, and get some completely bogus
> > information, since things weren't actually saved on the stack where it
> > thought they were.  This led to the unwound stack pointer being wrong for
> > the step_resume breakpoint, so when we hit the step_resume breakpoint we
> > kept going.
> > 
> > I fixed this by changing lookup_minimal_symbol_pc_section to be paranoid
> > about returning a minsym in the same section as the PC.  Technically, at
> > least on ELF targets, that doesn't _have_ to be true.  I've never
> > encountered an exception or a good reason for one, though.  Does anyone see
> > any pitfalls for this change?  Symtab maintainers, is this patch OK?
> > 
> > I believe this patch should also fix shlibs/1237, and may also fix
> > shlibs/1280.  Adam, could you check those?
> > 
> > 
> > 
> > 
> > 
> > By the way, I'm convinced that all is not well in step_over_function.  This
> > comment,
> > 
> >   /* NOTE: cagney/2003-04-06:
> > 
> >      The intent of DEPRECATED_SAVED_PC_AFTER_CALL was to:
> > 
> >      - provide a very light weight equivalent to frame_unwind_pc()
> >      (nee FRAME_SAVED_PC) that avoids the prologue analyzer
> > 
> >      - avoid handling the case where the PC hasn't been saved in the
> >      prologue analyzer
> > 
> >      Unfortunatly, not five lines further down, is a call to
> >      get_frame_id() and that is guarenteed to trigger the prologue
> >      analyzer.
> > 
> > is either incorrect or has gotten out of sync with the code:
> > 
> >   if (DEPRECATED_SAVED_PC_AFTER_CALL_P ())
> >     sr_sal.pc = ADDR_BITS_REMOVE (DEPRECATED_SAVED_PC_AFTER_CALL (get_current_frame ()));
> >   else
> >     sr_sal.pc = ADDR_BITS_REMOVE (frame_pc_unwind (get_current_frame ()));
> >   sr_sal.section = find_pc_overlay (sr_sal.pc);
> > 
> >   check_for_old_step_resume_breakpoint ();
> >   step_resume_breakpoint =
> >     set_momentary_breakpoint (sr_sal, get_frame_id (get_current_frame ()),
> >                               bp_step_resume);
> > 
> > 
> > Note that get_frame_id unwinds from the NEXT frame, and
> > frame_pc_unwind/DEPRECATED_SAVED_PC_AFTER_CALL unwind from THIS frame.
> > This throws me a loop every time I have to work in this function.  Also, I
> > have the nagging feeling we're saving the wrong frame.  I have an old MIPS
> > patch where I needed to use get_prev_frame in step_over_function.  As soon
> > as I have time to revisit that patch I'll be back to clean this up some
> > more.
> > 
> > -- 
> > Daniel Jacobowitz
> > MontaVista Software                         Debian GNU/Linux Developer
> > 
> > 2003-07-19  Daniel Jacobowitz  <drow@mvista.com>
> > 
> > 	PR c++/1267
> > 	* minsyms.c (lookup_minimal_symbol_by_pc_section): If SECTION is
> > 	NULL, default to the section containing PC.
> > 
> > Index: minsyms.c
> > ===================================================================
> > RCS file: /cvs/src/src/gdb/minsyms.c,v
> > retrieving revision 1.31
> > diff -u -p -r1.31 minsyms.c
> > --- minsyms.c	15 May 2003 22:23:24 -0000	1.31
> > +++ minsyms.c	19 Jul 2003 18:03:08 -0000
> > @@ -403,12 +403,22 @@ lookup_minimal_symbol_by_pc_section (COR
> >    struct objfile *objfile;
> >    struct minimal_symbol *msymbol;
> >    struct minimal_symbol *best_symbol = NULL;
> > +  struct obj_section *pc_section;
> >  
> >    /* pc has to be in a known section. This ensures that anything beyond
> >       the end of the last segment doesn't appear to be part of the last
> >       function in the last segment.  */
> > -  if (find_pc_section (pc) == NULL)
> > +  pc_section = find_pc_section (pc);
> > +  if (pc_section == NULL)
> >      return NULL;
> > +
> > +  /* If no section was specified, then just make sure that the PC is in
> > +     the same section as the minimal symbol we find.  */
> > +  if (section == NULL)
> > +    section = pc_section->the_bfd_section;
> > +
> > +  /* FIXME drow/2003-07-19: Should we also check that PC is in SECTION
> > +     if we were passed a non-NULL SECTION argument?  */
> >  
> >    for (objfile = object_files;
> >         objfile != NULL;
> 

-- 
Daniel Jacobowitz
MontaVista Software                         Debian GNU/Linux Developer