Re: [RFA]: LOC_COMPUTED + abstracted dwarf2 evaluator, again

[Jim Blandy <jimb at redhat dot com>]

Daniel Berlin <dberlin@dberlin.org> writes:
> > > +   if (ctx->in_reg)
> > > +     {
> > > +       store_typed_address (VALUE_CONTENTS_RAW (retval), 
> > > + 			   SYMBOL_TYPE (var), dwarf_expr_fetch (ctx, 0));
> > > +       VALUE_LVAL (retval) = not_lval;
> > > +     }
> > > +   else
> > > +     {
> > > +       result = dwarf_expr_fetch (ctx, 0);
> > > +       VALUE_LAZY (retval) = 1;
> > > +       VALUE_ADDRESS (retval) = result;
> > > +     }
> > 
> > This looks wrong.  If evaluating the Dwarf location expression yields
> > an `in_reg' result, shouldn't the value be an lval_register, with its
> > the address set to the register number?
> 
> This doesn't work, in reality.
> 
> IIRC, you run into problems dereferencing the value, because it 
> thinks it *only* lives in a register (IE it's not really a memory 
> location), regardless of what you tell it.
> 
> I may be misremembering the exact reason why, but I know things don't work 
> if you do this, even though it would seem to make sense (I had originally 
> written it to do what you suggest).

The `in_reg' case needs to call value_from_register, not explicitly
set the value's `lval' and `address' fields, as I implied.  Calling
value_from_register will take care of finding registers that have been
spilled to the stack.