Re: binutils 2.17 branch

[Alan Modra <amodra at bigpond dot net dot au>]

On Thu, Mar 16, 2006 at 10:04:06PM -0500, Daniel Jacobowitz wrote:
> The only thing I've got my eye on at the moment is Alan's recent
> --as-needed fixes; experience suggests it will take a little
> while for that to settle down.

Yes, that Alan Modra bloke can be unreliable. :)

Seriously, it's still somewhat of a hack.  The fact that we throw away
syms added for an unneeded lib means I'm relying on nothing
(eg. backend add_symbol_hook) keeping pointers into the symbol hash
table apart from the ones I already handle (undefs, sym_hashes).
I also don't like doubling the symbol table memory usage.

The technically better approach, of scanning as-needed lib syms to
determine whether to load the lib has its own set of problems.  You need
to fairly carefully duplicate existing code behaviour in deciding
whether a definition will satisfy a reference.  Symbol versioning
complicates this process.  Also, you'd need to modify the backend
add_symbol_hook functions to only discard symbols during the scan.
I felt that getting this approach right would take longer than
checkpointing the hash table..

-- 
Alan Modra
IBM OzLabs - Linux Technology Centre