Re: Strange behaviour in C++...

[Jason Merrill <jason at cygnus dot com>]

>>>>> Joe Buck <jbuck@synopsys.COM> writes:

 >>   > I assume that this would be an ABI for Intel-like hardware?  (I can't
 >>   > imagine an unqualified ABI for all platforms, I think it is too early).

 >> ia64.  But most of the concepts behind the ABI apply to other
 >> architectures -- name mangling, vtable layouts, empty virtual baseclass
 >> opts, etc etc.

 > Ah, now I understand ... thunks might really cost on that architecture.
 > However, it is far from a typical architecture, and what's optimal for
 > it might be far from best for others.

 > I will be annoyed, however, if only those who have signed the ia64 NDAs
 > get to play in setting this standard, and then we are asked to use it for
 > all the other ports.  Not all the clueful people are in that position.

Participation does not require signing the NDAs; we only discuss non-NDA
stuff at the meetings.  If you'd like to get involved, send email to
dehnert@sgi.com.  We currently have representatives from SGI (who put it
together), Sun, SCO, IBM, and EDG as well as g++ folk.

The rationale for dropping thunks is that the vtable size increase is not
thought to be as important as the performance impact of thunks.  Virtual
inheritance is actually fairly common, I'm told (mixins and whatnot).  The
performance impact of the adjustment is avoided by doing the adjustment on the
callee side (or not, if it isn't needed).

Jason