On Sun, 2010-05-16 at 14:04 +0200, Jonas Maebe wrote:
> On 15 May 2010, at 23:44, Jan Kratochvil wrote:
> 
> > But if the "allocation" is just an internal
> > compiler issue which should be hidden by the same compiler at the DWARF level.
> > Therefore I would guess to use some:
> > 	drop DW_TAG_array_type -> DW_AT_allocated
> > 	DW_TAG_subrange_type -> DW_AT_upper_bound:
> > 		DW_OP_push_object_address
> > 		DW_OP_deref
> > 		DW_OP_dup
> > 		DW_OP_bra allocated
> > 			DW_OP_lit0
> > 			DW_OP_skip end
> > 		allocated:
> > 			DW_OP_lit8
> > 			DW_OP_minus
> > 			DW_OP_deref
> > 		end:
> 
> I agree. Joost, I've attached a patch for FPC's DWARF writer to fix it. I can't test whether it works (the dumped DWARF info looks ok though), because I can't get gdb/F-13 to build:

It doesn't work.

That's because type_length_get() (gdbtypes.c) does not return 0 when the
high and low-bound are equal to each other. Instead it returns the
element-size. Then allocate_value_lazy tries to read element-size bytes
from the base-address (being 0x0) -> av.

I think that type_length_get() not returning 0 is a bug, (also because
there is a comment in the function that some part of the code doesn't
handle count=0 right. But I think that should be count=1) but it could
be that it's on purpose for Fortran. Jan, can you comment on this?

There's also a second problem. The lower bound is 1 for strings. With
Jonas' patch for fpc the upper bound of strings is 0 when not allocated.
Leading to a size of -1. An easy fix, new patch is attached.

Joost.