On Fri, Apr 13, 2012 at 02:24:09PM +0200, Olivier Hainque wrote:
> On typical VxWorks environments, WindRiver integrated tools are used as
> much if not more than gdb for debugging purposes.
>
> These evolve at an industrial pace, traditionally not as fast as GCC
> regarding the support of latest dwarf standards.
>
> As of today, in our experience, the best compromise to let WRS debugging
> tools work fine and have negligible effect on gdb's experience for this
> target is to enforce strict dwarf-2 debugging info by default.
>
> We have been doing this in our production compilers for a while and find
> it really useful, so we're happy to contribute here, in case this is deemed
> more generally appropriate.
>
> The attached patch achieves this in two mini steps:
>
> - Allow to distinguish between a dwarf version picked by default by the
> compiler and one provided on the command line that happens to match the
> default, which is useful to make sure we don't override a request for
> an explicit level on the command line.
>
> This is achieved by initializing dwarf_version to -1 instead of 2 from
> common.opt, then pick 2 as a default from toplev.c:process_options,
> similar to the scheme in place for dwarf_strict already.
>
> - Default to dwarf_strict = 1 && dwarf_version = 2 for VxWorks via
> vxworks_override_options.
>
> Sanity checked on mainline for powerpc-wrs-vxworks and bootstrapped
> on x86_64-pc-linux-gnu.
Looks good to me, but IMHO we should go further at this point, and
default to dwarf_version 4 unless overridden (and override
also on Darwin to too due to its tools not capable of handling it),
probably together with making -fno-debug-types-section the default
(since while DWARF 4 DW_FORM_* etc. support is in gdb/libdw/valgrind/etc.
for quite some time, e.g. .debug_types support is very recent, and
furthermore on lots of programs/DSOs .debug_types actually increases
size instead of decreases. If a post-linking utility is used
undoing the extra duplication caused by .debug_types (yeah, I know
Cary said he'll work on it) is going to be very hard. I've noticed
also lots of intra-CU redundancies with current state of .debug_types,
e.g. lots of DW_TAG_{const,pointer,reference}_type DIEs with the same
DW_FORM_ref_sig8 hash).
Jason, are you ok with Olivier's patch?
If yes, I'll post a follow-up to that.
> 2012-04-13 Olivier Hainque <hain...@adacore.com>
>
> * common.opt (gdwarf-): Initialize dwarf_version to -1 instead of 2.
> * toplev.c (process_options): Default to dwarf_version 2.
> * config/vxworks.c (vxworks_override_options): Default to strict-dwarf
> and dwarf_version 2.
Jakub
