Re: [PATCH/i386newframe/RFC] DWARF CFI frame unwinder

[Daniel Jacobowitz <drow@mvista.com>]

On Mon, May 05, 2003 at 10:52:54AM -0400, Andrew Cagney wrote:
> 
> >>>The frame's CFA is the basis for identifying the frame and locating
> >>>saved registers in the CFI.  It is always present when you have CFI.
> >>>
> >>>DW_AT_frame_base is the basis for locating saved variables and locals. 
> >>>It is generally present when you have DWARF-2 debug info.
> >
> >>
> >>You and I went through all this not too long ago.  frame-base is for 
> >>this high level thingie, frame-unwind is for the low level register 
> >>information.
> >
> >
> >Then, as Mark said, it shouldn't be providing a frame base at all.  The
> >CFA information is not the right frame base, and the use of
> >DW_AT_frame_base is exactly orthogonal to the use of CFI.
> 
> Daniel, you and I went through all this not too long ago.
> 
> Not providing a dwarf2 specific frame-base (returning DW_AT_frame_base) 
> leads to ``info frame'' printing inconsistent information.  It will 
> report ``id.stack_addr'' as the frame's base and that is definitly not 
> correct - it won't match the high level ``base'' that the user expects 
> to match a disassembler.
> 
> Again, this is why the high level frame-base is separate to the low 
> level frame-unwind.  It is possible to mix 'n' match.
> 
> However, it doesn't need to be implemented right now - at present 
> dwarf2loc short circutes frame-base, implementing DW_AT_frame_base locally.

I think we've managed to end up in violent agreement.  Since right now
the frame base is associated with the unwinder, shall we leave it as
the CFA with a comment saying something like:

  /* When we have DWARF-2 debugging information, this should be
     DW_AT_frame_base.  It should probably be provided by a method
     specific to the function's debug information instead of its
     unwind type.  */

Not sure I got the wording right.

-- 
Daniel Jacobowitz
MontaVista Software                         Debian GNU/Linux Developer