Re: PATCH: Add STB_SECONDARY support

[Alan Modra <amodra at gmail dot com>]

On Sat, Jun 30, 2012 at 02:18:41PM -0700, H.J. Lu wrote:
> This patch adds STB_SECONDARY support.  I am enclosing the STB_SECONDARY 
> proposal posted on the gABI group. This patch is backward compatible
> with existing ld.so.  STB_SECONDARY symbols are generated only with
> "ld -shared -z secondary" switch.  OK to install?

I think there are a few problems with this proposal.  One is that weak
symbol behaviour is not well defined by the gABI.  In fact the spec
says

"The behavior of weak symbols in areas not specified by this document
is implementation defined. Weak symbols are intended primarily for use
in system software. Applications using weak symbols are unreliable
since changes in the runtime environment might cause the execution to
fail."

Left unspecified is the behaviour of weak symbols in dynamic objects
at runtime.  So I think your specification is standing on shaky ground
when saying anything about runtime behaviour of secondary symbols.
You really ought to first specify runtime behaviour of weak symbols
(and trying to do that will likely result in a conflict).

After reading some of the history of your proposal on
https://groups.google.com/forum/?fromgroups#!forum/generic-abi I see
that you originally only wanted secondary symbols for relocatable
objects, but were pushed to specify them more widely by other well-
meaning but perhaps inexperienced people.  (I'm not intending any
insult in that comment, I simply don't know the experience or
otherwise of people who responded.)  I also have not seen positive
comments from anyone I know is experienced in GNU toolchain work.
Maybe that is simply because I haven't looked far enough, but right
now I can't see any compelling reason to accept this proposal.  You
need to convince me or another global binutils maintainer that
STB_SECONDARY is a good idea.

If the reason for your proposal still only remains a way of providing
lower precedence symbols at link time, why can you not get the
behaviour you want just with weak symbols?  For example, place all
your lower precedence weak symbol functions/data in an archive library
that is linked last?  That might not be convenient.  I'm just asking
whether it would be possible.  (*)  If it is possible then no
extension to the gABI is necessary.


*) Convenience might be improved by implementing a linker switch that
says "link this object file as if it were inside an archive library
placed last on the linker command line".

-- 
Alan Modra
Australia Development Lab, IBM