Sources such as
http://preshing.com/20130930/double-checked-locking-is-fixed-in-cpp11/
or https://en.wikipedia.org/wiki/Double-checked_locking suggest that
it wasn't possible to reliably do a portable double-checked initialization
before C++11. It may be true that for all platforms we support those
memory barriers are in fact not needed (which seems to be the assumption
behind OSL_DOUBLE_CHECKED_LOCKING_MEMORY_BARRIER being empty), and
looking at the generated assembly here on x86-64 seems to confirm that, but
in the worst case then this is a more chatty and standard way of writing
a no-op.

I don't want to use threadsafe statics or std::call_once() because
ScGlobal::Clear() does cleanup, which would be non-trivial to do with these,
and also some functions may not necessarily always force
creation of the singleton when touching the pointer, so it can't be easily
hidden behind a single function call.

The need to explicitly use load() with delete (thus preventing DELETEZ)
looks like a Clang bug to me.

Change-Id: Id3b0ef4b273ed25a5c154f90cde090ea1f9674fb
Reviewed-on: https://gerrit.libreoffice.org/55851
Tested-by: Jenkins
Reviewed-by: Michael Meeks <michael.meeks@collabora.com>