Re: dwarf_aggregate_size() seems to fall over on pointer types

[Josh Stone <jistone at redhat dot com>]

On 10/01/2014 10:15 AM, Conrad Rad wrote:
> Hi all,
> 
> First, I'm not super familiar with DWARF, and perhaps this is a bug in
> the DWARF information emitted by Clang (3.4). It's definitely a
> difference between GCC and Clang DWARF output.
> 
> I've been fiddling with a pahole-alike tool[0] that sits directly on
> elfutils (as opposed to libdwarves, which hasn't been updated for some
> time and doesn't handle some of the tags Clang emits). I've observed
> that dwarf_aggregate_size() doesn't seem to like types that are
> pointers or arrays of pointers[1].
> 
> I've compiled a small example program, repro, and output[2]:
> http://pastie.org/9610702
> 
> dwarf_aggregate_size() has no trouble with most member types, or even
> pointer types on the output from GCC. However, it returns an error on
> Clang's output.
> 
> Perhaps Clang is just eliding pointer size information, assuming
> consumers will consult the Elf machine class and infer?

Indeed, clang doesn't seem to output DW_AT_byte_size on any
DW_TAG_pointer_type, but gcc does.

FWIW, the DWARF4 Appendix A does *not* list DW_AT_byte_size as being
applicable to DW_TAG_pointer_type, nor DW_TAG_reference_type, nor
DW_TAG_rvalue_reference_type.  However, that is just "informative", and
producers are free to omit some of the suggestions as well as add their
own, so it's not an error that gcc produces DW_AT_byte_size.

> Am I missing something? Is this a Clang bug or an elfutils bug?

I don't think it's strictly a bug in either.  Clang is not required to
produce byte_size, and elfutils is just reporting what it knows.

Still, it might be nice if elfutils made that machine-class inference
for you.  (But I worry about things like that mips pointer issue
reported recently.)