Re: frame->unwind->this_base()

[Daniel Jacobowitz <drow@mvista.com>]

On Wed, Mar 19, 2003 at 10:24:37AM -0500, Andrew Cagney wrote:
> 
> >>True dwarf2 debug info or that .eh_frame stuff (i'm curious)?
> >
> >
> >Hmm, I thought it would write out .debug_frame without DWARF-2 but
> >peering at the GCC source I seem to be wrong again.  So just .eh_frame.
> 
> So using .eh_frame is along the same lines as using ia64's libunwind. 
> .eh_frame just happens to be implemented using something very like 
> dwarf2's unwind code.
> 
> >In any case, we'll parse both, so I stand by my statement.  We'll have
> >.eh_frame even without normal debug info.
> 
> I think there needs to be separate eh-frame and cfi-frame but with a 
> common implementation.  That way the subtle, but important, differences 
> are clear.

What's your perception of the difference?  Just based on the different
on-disk format?  The information contained is essentially the same.  I
certainly don't object, though.  This will be painless.

> >>For stabs to work, it needs FRAME_LOCALS_ADDRESS(); and 
> >>FRAME_LOCALS_ADDRESS() relies on the prologue analyzer (since frame ID 
> >>won't correspond to `frame-base') for the computation of the correct 
> >>value; and that means unwinding the same frame two ways.  Outch.
> >
> >
> >Yeah...
> > - if we have CFI use it to find the frame address.  Does this
> >become the frame ID?
> > - if we have dwarf2 debug and CFI, then we don't need to do prologue
> >analysis; CFI should give us everything we need
> > - if we have stabs debug and CFI, then we do need to do prologue
> >analysis to get FRAME_LOCALS_ADDRESS
> > - if we have either kind of debug info and no CFI then we need to do
> >prologue analysis; for dwarf2 we'll also need to calculate the frame
> >base from DW_AT_frame_base in order to use it to find locals
> >
> >Is that about right?
> 
> Yes.  Try the following higher-level view of the problem:
> 
> On the left is the unwinder.  It exports methods to obtain the frame's 
> ID and the registers.  It can be implemented using CFA, EH, libunwind, 
> prologue analysis, ...; and the implementation is selected based on the 
> low-level unwind information, or lack there of.
> 
> On the right is the local variable code and that needs a frame-base / 
> frame-locals-address / ....  It uses high-level debug info and unwound 
> register values to compute that base.  It can be implemented using 
> dwarf2's frame_base, or prologue analysis (stabs), or ...; and the 
> implementation is selected based on the frame's high-level debug info.
> 
> The mess occures when the high-level RHS frame-locals-address starts 
> assuming the flavour of the low-level LHS unwinder and, consequently, 
> tries to directly exploit that knowledge.  For instance, a RHS prologue 
> based frame-locals-address assuming that the LHS is also prologue based, 
> and hence, can directly access the LHS's prologue analysis cache.
> 
> It can be `fixed' two ways:
> 
> - refusing to allow that sharing of data, forcing the RHS 
> frame-locals-address to re-analyze the prologue.
> 
> - make it possible to tease out the prologue analysis object so that 
> both the LHS and RHS can share it.
> 
> I guess the second is it.

I agree.  Personally, what I would like would be to generate something
structurally/conceptually equivalent to the unwind data from the
prologue reader.  This requires a bit of rewriting and a great deal of
testing/patience.

-- 
Daniel Jacobowitz
MontaVista Software                         Debian GNU/Linux Developer