r194095 - Produce direct calls instead of alias to linkonce_odr functions.

[Rafael Espíndola]

> Alternatively, rather than replacing one hack with another hack that may or
> may not break ABI, why not actually implement comdat support?

The current implementation is not a hack. It brings us back to the one
function per comdat we used to have. We should implement comdats (in
my todo list, after pr10367 and mangler), but it is not a strict
superset.

> I feel like this patch doesn't solve any pressing need, and is causing
> gratuitous fallout. It may be that the old pattern was a hack, but it was at
> least a hack that had been well tested in deployment and has caused us no
> problems so far.

On ELF it is an optimization issue. On COFF we need something along
these lines since MSVC will only print one destructor if it knows that
the base and complete ones are identical.

Cheers,
Rafael