Re: [RFA] Limited DW_OP_piece support

[Daniel Jacobowitz <drow@mvista.com>]

On Thu, May 22, 2003 at 02:19:31PM -0700, Kevin Buettner wrote:
> On May 22,  2:19pm, Daniel Jacobowitz wrote:
> 
> > On Thu, May 22, 2003 at 10:00:39AM -0700, Kevin Buettner wrote:
> > > The patch below adds limited DW_OP_piece support to dwarf2expr.c.  I
> > > will post a patch to rs6000-tdep.c which illustrates what a
> > > ``dwarf2_compose_register_pieces'' method should look like.
> > > 
> > > Okay?
> > 
> > I would really strongly prefer that we not do it this way.
> > 
> > You'll notice that there are no other gdbarch calls in the expression
> > evaluator.  There might be some implicit ones through macros, for
> > instance there is TARGET_ADDR_BIT.  That needs to be fixed properly
> > some day already.
> > 
> > Instead, IMHO, we should devise a way to represent multiple locations
> > in the evaluator's return value.  This is not suggesting the complete
> > overhaul that we need to support multiple locations in the rest of GDB.
> > Then have the expression evaluator properly return a list of locations,
> > and have the massaging done via gdbarch in the evaluator's client. 
> > Does that sound reasonable?
> 
> I must admit that it sure sounded reasonable when I first read it. 
> I've been looking at the code to see how doable it is, and it's
> looking less reasonable to me now.  It appears to me that there are
> multiple clients and it seems ugly to do the massaging that you speak
> of in multiple places.  (Or perhaps I misunderstand who the client
> is?)

I'm suggesting that the massaging be done in the caller of
dwarf_expr_eval.  There are three of them at present: one which only
cares about whether we need a frame, and two for locations.  One's the
frame base, and the other's via dwarf2_evaluate_loc_desc.

For the moment, I believe everything you need could be done in
dwarf2_evaluate_loc_desc.  The frame base will not (on current
platforms, etc.) use DW_OP_piece, and that call should be going away
anyway.  There will be more calls, as we use the evaluator for more,
but it's not clear how they should react to DW_OP_piece.

Another alternative is to do it in dwarf_expr_eval.  This would
probably want us to separate it into two functions: one for evaluating
an expression as a location, and one otherwise.  i.e. there are times
when DW_OP_piece should be handled, and times when it is not valid. 
They can have different return signatures.

Does that make more sense?

-- 
Daniel Jacobowitz
MontaVista Software                         Debian GNU/Linux Developer