Re: New AIX libstdc++ testsuite regression 20_util/allocator_members

[Benjamin Kosnik <bkoz at redhat dot com>]

I agree with Jason here: I think being able to replace the new/delete
global signatures is a very important bit of functionality (see 18.4.1).
To me, not being able to do this on AIX is a bug, and should be fixed.

(Actually, I think tests should be added to
testsuite/18_support/[new.cc, delete.cc] to make sure this is indeed
possible, and that class-scope new/delete works as well. That can be
done later, I guess. I'm surprised to be finding out about this now.)

> 	My concern is how much to do things automatically versus how
> 	much to allow the user to have control.  Most people using GCC
> 	probably assume SVR4-like semantics, but some people do develop
> 	code for the AIX model. Automatically exporting all symbols and
> 	invoking runtime linking behavior can cause problems in the
> 	opposite direction.

Certainly, all the 18.4.1 symbols should be allowed. 

Is there a way you can re-use the config/linker-map.gnu information for
AIX? That has a list of symbols that shared libstdc++-v3 intends to
export: perhaps that would be helpful? (It does export all local
symbols, however.) You are welcome to make a new file, linker-map.aix,
if it would be of use for you.

> 	Unfortunately, there is no "right" answer.

I'm not sure this really should be catagorized as a "high" priority
regression then, since this underlying behavior has been an issue for
3.0.x compilers on AIX as well. (We just didn't know about it.)

-benjamin