[llvm] r189954 - Revert "Add r159136 back now that pr13124 has been fixed."

[Rafael Espíndola]

> We could just specify this the other way: if we have two definitions and
> either is unnamed_addr, the merged definition is also unnamed_addr.  Is
> there some reason we shouldn't?

Disallows so merging opportunities and enables other optimizations,
but I don't think it would enable this one.  In the above example

TU1:
linkonce_odr unnamed_addr @foo....

TU2:
linkonce_odr @foo.....

I was assuming that each TU produces an ELF file. The ELF linker knows
nothing about unnamed_addr and will just produce a hidden symbol for
foo. For this optimization to be legal we would have to require the
frontends to produce a "linkonce_odr unnamed_addr" in all copies of a
symbol or none.

Cheers,
Rafael