AlisdairM wrote:
John Maddock wrote:
However the change should be restricted to VC7.1 just to be sure.
Triple-checking, I believe you are correct. (and I have been through this dance more than once over the last 18 months)
It does mean that array would never be a POD, and not be admissible in unions. Those are lesser restrictions though, so I am less worried by them.
Actually if array ever became a non-POD that would be a serious breaking change: POD's are initialised in the static initialisation phase and we really do want: boost::array<int 3> a = { 1, 2, 3 }; To be initialised before everything else. However, section 12.8p9 says: "A user-declared copy assignment operator X::operator= is a non-static non-template member function of class X with exactly one parameter of type X, X&, const X&, volatile X& or const volatile X&.109)" So I think we are OK with a *template* assignement operator, as it doesn't count for considering whether a type is a POD or not.
In principle the proposed workaround should be doing exactly the same as the implicitly declared copy-assignment operator, although it is potentially less efficient - especially in the POD case.
The question is - should we make this the default behaviour - banning POD-behaviour and unions; accept these restrictions on the MS compiler only and deem such use non-portable; or specialize on MS only, purely for the pointer-to-member case?
Personally I would add the template-assignment op for BOOST_MSVC < 1400 only, would we gain anything from adding it for other compilers? It opens up an assignment that wasn't there before, so I guess it means array<double, N> becomes assignable to array<int, N> which may or may not be a good idea depending on your POV. John.