Re: GDB 13 Release 2022-09-11 update

[Tom de Vries <tdevries@suse.de>]

On 9/16/22 00:41, Tom Tromey wrote:
> Tom> Perhaps it would be worth clarifying the status of support for .gdb_index.
> 
> Tom> It is the defacto standard for the index cache, but ada support is
> Tom> broken ( https://sourceware.org/bugzilla/show_bug.cgi?id=29179 ).
> 
> Tom> Do we want to fix this?
> 
> Tom> Do we want to deprecate .gdb_index and start using .debug_names for
> Tom> the index cache?
> 
> Tom> Or are we fine with the status quo (which then might need documenting)?
> 
> My impression is that .gdb_index never really worked correctly for Ada.

Hmm, I committed 7ab96794115 ("[gdb/symtab] Enable ada .gdb_index").

So my understanding was that it was working, starting gdb 10, at least 
to the extent that running cc-with-gdb-index didn't show any regressions 
vs native for ada test-cases.

> So, while that is a regression of sorts, since the feature isn't really
> usable for Ada in general,

Um, can you elaborate a bit on why you think that?

> it doesn't seem worthwhile to fix it.
> 
> I did make .debug_names work for Ada.  See commit 3b00ef10a2a ("Add Ada
> support for .debug_names").  I'm not at my work machine right now but
> IIRC this is tested internally.
> 

Yeah, I also try to run cc-with-debug-names target board once in a while 
and report/fix regressions, and I don't see any ada-specific problems.

> As for the overall plan, my view is:
> 
> * .gdb_index is (or should be) deprecated.  It was good but now we can
>    do better.  However...
> 
> * In order to switch over fully, we have to fix the bugs in .debug_names.
>    Fixing the writing side is easy (I have a patch somewhere) but I
>    haven't worked on the reader yet.
> 
> * Once that is done we could switch over the index cache, but this also
>    requires some more work to handle the situation where the index needs
>    a string that isn't already in .debug_str.
> 
> I don't mind if this situation is documented.  We could even make
> .gdb_index writing fail for Ada, though it seems to me that for the
> index cache it is better to just fail silently.  Based on the Ada
> programs I've seen, the new DWARF reader seems plenty fast enough
> anyway.

The current failure mode is a disfunctional index, and consequently 
things are not being found, so there are lots of test-suite FAILs.

We could refuse to write an index (which would be non-silent), or we 
could even write an empty index (which could be silent, I think).

Thanks,
- Tom