dwarf2_get_pc_bounds problem

dwarf2_get_pc_bounds problem

[Martin M. Hunt]

I'm investigating several errors in recent versions of gdb and they all
seem to be caused by bogus values for lowpc and highpc returned from 
dwarf2_get_pc_bounds(). I'm not a DWARF expert so maybe the problem is
bad debug info, but the code in dwarf2_get_pc_bounds seems suspicious.

What I'm seeing is that with a program linked at 0x80000000, all the
highpc and lowpc look fine, except those derived from DW_AT_ranges
information.  Those are all very small, like 0x100.  Looking at the code
in dwarf2_get_pc_bounds(), most of it seems to be trying to calculate a
variable "base" which is then never used. Perhaps a simple addition was
left out?

Martin

Re: dwarf2_get_pc_bounds problem

[Daniel Jacobowitz]

On Fri, Feb 14, 2003 at 02:59:40AM -0800, Martin M. Hunt wrote:
> I'm investigating several errors in recent versions of gdb and they all
> seem to be caused by bogus values for lowpc and highpc returned from 
> dwarf2_get_pc_bounds(). I'm not a DWARF expert so maybe the problem is
> bad debug info, but the code in dwarf2_get_pc_bounds seems suspicious.
> 
> What I'm seeing is that with a program linked at 0x80000000, all the
> highpc and lowpc look fine, except those derived from DW_AT_ranges
> information.  Those are all very small, like 0x100.  Looking at the code
> in dwarf2_get_pc_bounds(), most of it seems to be trying to calculate a
> variable "base" which is then never used. Perhaps a simple addition was
> left out?

Ah, er, um, er.... I tested this, how the heck did it work?  Aha, my
test case involved multiple sections, so GCC used a base of 0.  That's
how.

Could you try this obvious fix?  Still doesn't fix Jakub's test (we
really do need discontiguous ranges for that to work) but r_type
searches one local block instead of going straight to the function.

Oddly, I remember this happening before.  I may have lost the addition
in a merge somewhere.

-- 
Daniel Jacobowitz
MontaVista Software                         Debian GNU/Linux Developer

2003-02-14  Daniel Jacobowitz  <drow@mvista.com>

	* dwarf2read.c (dwarf2_get_pc_bounds): Offset addresses by base.

Index: dwarf2read.c
===================================================================
RCS file: /cvs/src/src/gdb/dwarf2read.c,v
retrieving revision 1.85
diff -u -p -r1.85 dwarf2read.c
--- dwarf2read.c	4 Feb 2003 20:17:02 -0000	1.85
+++ dwarf2read.c	14 Feb 2003 15:10:24 -0000
@@ -2195,6 +2195,9 @@ dwarf2_get_pc_bounds (struct die_info *d
 		  return 0;
 		}
 
+	      range_beginning += base;
+	      range_end += base;
+
 	      /* FIXME: This is recording everything as a low-high
 		 segment of consecutive addresses.  We should have a
 		 data structure for discontiguous block ranges

Re: dwarf2_get_pc_bounds problem

[Martin M. Hunt]

Your patch fixes the problems I was seeing.

Martin

On Fri, 2003-02-14 at 07:12, Daniel Jacobowitz wrote:
> On Fri, Feb 14, 2003 at 02:59:40AM -0800, Martin M. Hunt wrote:
> > I'm investigating several errors in recent versions of gdb and they all
> > seem to be caused by bogus values for lowpc and highpc returned from 
> > dwarf2_get_pc_bounds(). I'm not a DWARF expert so maybe the problem is
> > bad debug info, but the code in dwarf2_get_pc_bounds seems suspicious.
> > 
> > What I'm seeing is that with a program linked at 0x80000000, all the
> > highpc and lowpc look fine, except those derived from DW_AT_ranges
> > information.  Those are all very small, like 0x100.  Looking at the code
> > in dwarf2_get_pc_bounds(), most of it seems to be trying to calculate a
> > variable "base" which is then never used. Perhaps a simple addition was
> > left out?
> 
> Ah, er, um, er.... I tested this, how the heck did it work?  Aha, my
> test case involved multiple sections, so GCC used a base of 0.  That's
> how.
> 
> Could you try this obvious fix?  Still doesn't fix Jakub's test (we
> really do need discontiguous ranges for that to work) but r_type
> searches one local block instead of going straight to the function.
> 
> Oddly, I remember this happening before.  I may have lost the addition
> in a merge somewhere.
> 
> -- 
> Daniel Jacobowitz
> MontaVista Software                         Debian GNU/Linux Developer
> 
> 2003-02-14  Daniel Jacobowitz  <drow@mvista.com>
> 
> 	* dwarf2read.c (dwarf2_get_pc_bounds): Offset addresses by base.
> 
> Index: dwarf2read.c
> ===================================================================
> RCS file: /cvs/src/src/gdb/dwarf2read.c,v
> retrieving revision 1.85
> diff -u -p -r1.85 dwarf2read.c
> --- dwarf2read.c	4 Feb 2003 20:17:02 -0000	1.85
> +++ dwarf2read.c	14 Feb 2003 15:10:24 -0000
> @@ -2195,6 +2195,9 @@ dwarf2_get_pc_bounds (struct die_info *d
>  		  return 0;
>  		}
>  
> +	      range_beginning += base;
> +	      range_end += base;
> +
>  	      /* FIXME: This is recording everything as a low-high
>  		 segment of consecutive addresses.  We should have a
>  		 data structure for discontiguous block ranges