Re: [PATCH] Mark object files with "target:" filenames as OBJF_NONLOCAL_FILENAME

[Gary Benson <gbenson@redhat.com>]

Doug Evans wrote:
> On Mon, Apr 13, 2015 at 12:07 PM, Gary Benson <gbenson@redhat.com> wrote:
> > This patch introduces a new objfile flag OBJF_NONLOCAL_FILENAME to
> > denote that objfile->original_name and objfile->obfd->filename are
> > filenames referring to files on filesystems other than GDB's local
> > filesystem.  allocate_objfile is updated to force
> > OBJF_NONLOCAL_FILENAME if the specified name starts with "target:",
> > and to not attempt to expand the name using gdb_abspath if flags has
> > OBJF_NONLOCAL_FILENAME set.  load_auto_scripts_for_objfile is updated
> > to not attempt loading of auto-load scripts for objfiles with
> > OBJF_NONLOCAL_FILENAME set in their flags.
> >
> > A new flag was created rather than reusing OBJF_NOT_FILENAME because
> > setting that flag would stop Python's gdb.lookup_objfile from seeing
> > the file, and it's not clear that that's desirable.
> >
> > Without this patch you *sometimes* get things like:
> >
> >   Reading symbols from /home/gary/target:/lib64/libreadline.so.6...
> >
> > I haven't figured out why this doesn't always happen but it's plainly
> > wrong :)
> >
> > Built and regtested on RHEL6.6 x86_64.
> >
> > Ok to commit?
> >
> > Cheers,
> > Gary
> >
> > gdb/ChangeLog:
> >
> >         * objfiles.h (OBJF_NONLOCAL_FILENAME): New define.
> >         * objfiles.c (allocate_objfile): Force OBJF_NONLOCAL_FILENAME
> >         for BFDs with "target:" filenames.  Do not attempt to expand
> >         name if flags has OBJF_NONLOCAL_FILENAME set.
> >         * auto-load.c (load_auto_scripts_for_objfile): Do not attempt
> >         to auto-load scripts for OBJF_NONLOCAL_FILENAME objfiles.
> 
> My first thought is that we'll be recording something twice, and that
> can lead to problems (e.g., effort has to be expended to keep them in
> sync).
> "is nonlocal" is already specified by the "target:" in the name.
> 
> While I'm all for building on "foo:bar" in path names (target:foo,
> remote:foo, and so on), IWBN to build a library on top of that
> rather than have sideband tables that recorded such extra info.
> [Down the road I can imagine having a class for such things such
> that we could augment what's recorded beyond just a "foo:bar"
> string, but that's later, if ever.]
> 
> IOW, how about having an "is non-local" predicate that is invoked on
> the path whenever needed?
> [it could be the current "is_target_filename" or if you wanted to add
> a layer of abstraction that might be ok, depending on how this might
> evolve]
> 
> I realize this is a bit incongruous with OBJF_NOT_FILENAME, but I'd
> rather head in the above direction than adding more OBJF_ flags.
> 
> Thoughts?

I'm happy to remake this patch using "is_target_filename".  I'll do
that and mail a version 2 tomorrow.

(I've been thinking we might need something more than a prefix at some
point, maybe something more URL-like, but like you say, we don't need
that right now.)

Cheers,
Gary

-- 
http://gbenson.net/