Re: PATCH: new linker option for shared library COMDAT optimization

[Sandra Loosemore <sandra at codesourcery dot com>]

On 02/20/2012 04:40 PM, Alan Modra wrote:
> On Fri, Feb 17, 2012 at 11:41:03AM -0700, Sandra Loosemore wrote:
> > The slight wrinkle in this scheme is that the ELF specification
> > defines COMDAT groups as an attribute of relocatable files only, not
> > shared objects.  So we cannot say definitely that a shared object
> > provides a named COMDAT group.  The expedient hack I've adopted is
> > to consider that a shared library provides a COMDAT group if it
> > provides weak definitions for all symbols defined in sections that
> > make up the group.
>
> I wondered how robust this will prove to be.
> [snip]
> That's rather a lot of missing symbols.  I think for this to work you
> need another linker option to export all global symbols defined in
> comdat sections.

Hmmmm, it seems to me that this is a separate issue.  Maybe it even 
makes more sense to implement it as a compiler option instead, 
consistently with -fvisibility, -fvisibility-inlines-hidden, and #pragma 
GCC visibility.  Between those options and linker EXPORT lists, it seems 
to me that it is certainly already possible to build shared libraries 
that export the necessary symbols, although it may take some extra work 
to do so.

Would it suffice for the purposes of the current patch to add a bit to 
the documentation for --shared-comdat explaining that it requires that 
the comdat symbols have default (exported) visibility in the shared 
library, and that this may require changes to how the library is built?

> Then, given that you will need to build shared
> libraries differently, it would be a much better idea to mark them as
> supporting shared-comdat, perhaps by exporting each comdat signature
> symbol.  If you did that, you could probably dispense with all the
> code you've written to match symbols..

I previously considered keying off the comdat signature symbol only, but 
thought that it might be too easily fooled.  Perhaps as a less-strict 
version of what I have now, I could require that the shared library 
provide definitions for only the comdat symbols that are referenced, 
instead of all of them?  I think that is consistent with what building 
with -frepo would do.

-Sandra