------- Comment #27 from maddox at google dot com 2007-05-25 04:57 -------
(In reply to comment #20)
Ian proposed handling the simple case in which the two constructors ("clones")
are identical by emitting the code only once, but labelling it with the symbols
for both constructors. This scheme fails in the case that the constructors
must go in a comdat group. What do we name the group? If we invent a new name,
this will have ABI impact, as other implementations must not step on it. If we
use the name of one of the constructors, we may find that the section is
discarded in favor of a like-named section produced by another compiler, but
which need not define both symbols, resulting in link-incompatibility. It
appears, then,
that a distinct group name must be used for the combined constructor. As I
understand it, this was the principal issue that kept Stuart's patch
(referenced in comment #23) from being adopted, so I don't see how this more
limited optimization sidesteps the objections to that patch. I spoke with Ian
this afternoon, and he agreed with this analysis. It was suggested in the
discussion of Stuart's patch that it would be sufficient to provide for a
vendor namespace.
This seems like a simple, general facility that could be leveraged in multiple
contexts, e.g., to support a more general subroutinization optimization. How
difficult would it be to get such a provision into the ABI? With a suitably
chosen mangling convention, I don't think that name collisions with other
ABI-conformant implementations would actually be a problem in practice, so it
seems to be more of a matter of agreeing on a convention for the sake of the
ABI standard rather than actually managing the transition. Comments?
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=3187
