Re: [PATCH v2] gdb/dwarf: disable per-BFD resource sharing for -readnow objfiles

[Simon Marchi <simon.marchi@polymtl.ca>]

On 2021-03-30 12:13 p.m., Tom Tromey wrote:> Simon> I think the most sensible fix would be to not share a dwarf2_per_bfd
> Simon> between two objfiles loaded with different methods.  That means that two
> Simon> objfiles sharing the same BFD and loaded the same way would share a
> Simon> dwarf2_per_bfd.  Two objfiles sharing the same BFD but loaded with
> Simon> different methods would use two different dwarf2_per_bfd structures.
> 
> It seems to me that -readnow is a funny situation, because it is
> intended to do something unusual.
> 
> For all other kinds of possible sharing, I think that deferring to
> whatever was already done and recorded in the per-BFD object should work fine.
> The idea here is just to bypass the reading step when possible.
> So IMO it's fine to ignore the index cache if the psymtabs have already
> been scanned -- reading from the cache won't give any benefits relative
> to reusing the per-BFD object.
> 
> This seems like it would mainly be an issue for the ordering of checks
> in dwarf2_initialize_objfile.  When I look there, it seems correct
> already though -- the existing per-BFD is checked before trying to do
> any scan or reading from the index cache.

Yes, that case works fine (because it was broken at first and a PR was
reported :)).  I would be curious what you think about the (still
theoritical) index -> psymtabs case mentioned here:

    https://sourceware.org/pipermail/gdb-patches/2021-March/177366.html

> 
> Simon> 	* dwarf2/read.c (dwarf2_has_info): Don't share dwarf2_per_bfd
> Simon> 	with objfiles using READNOW.
> 
> Thank you.  This looks good to me.  Thanks especially for the test case.

> I think the ChangeLog entry should probably mention the PR though.

Thanks, I will add that and push to both branches.

Simon