[clang] Patch for 6037

[Richard Smith]

On Wed, Dec 24, 2014 at 8:46 AM, Nathan Sidwell <nathan at acm.org> wrote:
> [*] As an aside, I wonder if completing the std:set with the direct bases
> and keeping it with the RecordDecl would speed up base conversion checks --
> the base conversion machinery could use it for a quick 'is this even worth
> figuring out' check.

It shouldn't be a std::set; node-based containers are particularly bad
for malloc traffic. Also, for a deep hierarchy, we'd end up storing
O(N^2) nodes, which seems unfortunate (we already get that as a time
cost with this patch; it'd be nice to avoid that too). I wonder if
there's some relatively compact way of representing this information
that supports fast query and fast updates...