we generate correct code for both strong and weak atomic compare-exchanges, even
though we don't propagate to the IR enough information to generate optimal weak
compare-exchanges on architectures which support them.

llvm-svn: 155161