------- Additional Comments From pcarlini at suse dot de 2005-01-20 00:34
-------
Hi Gaby,
> Yes. Basically, we need to have tha aligned attribute work correctly.
Agreed, in principle: indeed, we are filing together and taking care of
many PRs in this area.
> | is not present in the original design. I don't think we can implement it
> | now, for 4.0, without changing the ABI. I think we should just document
> | that for our current basic_string memory rerurned by the allocator should
> | be maximally aligned (in some cases less aligned is ok, but details become
> | tricky to spell out).
>
> I rather we fix it. Remember, this is more an optimization issue
> than a semantics issue. An optimization issue that had causes us
> more trouble than benefits I believe. I don't believe it is wise
> for us to go that path down putting more an more restrictions.
> With people playing with fancy allocator around, it is likely that
> we're going to have more and more of this issue popping up.
Again, in principle I agree, but remember that:
1- It's almost impossible (see messages from Mark) that the
__attribute__(aligned) machinery will be satisfactorily fixed for
4.0.
2- We do *not* want to break the ABI for 4.0, in particular, in
the basic_string area. Indeed, as you can see, we are always
monitoring check-abi and very careful about everything (see
constructor / assignment issue of yesterday)
This is way I'm proposing, *for 4.0, only for 4.0*, to document
a bit some weaknesses wrt alignment, which we always had, only
implicitly, and change a bit the ext/array_allocator things to
not trigger problems, currently (see patch posted on the v3 list)
Sounds reasonable?
--
What |Removed |Added
----------------------------------------------------------------------------
CC| |gdr at integrable-solutions
| |dot net
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19495
