The original support for embedded bitcode used appending linkage for the
magic internal variables that hold the bitcode and command line options,
but that private linkage is a better fit. The only real reason for
appending linkage was to prevent those variables from being optimized away.
r269706 limits the use of appending linkage so that it cannot be used for
those variables, so this switches to use private linkage and keep the
variables alive with llvm.compiler.used. This is basically copied from
clang r269679. rdar://problem/28685198.
As of the swift-3.1-branch versions of Clang/LLVM, embedded bitcode is now
working well enough that the tests can be reenabled. rdar://problem/26247134
Together with -wmo it enables multi-threaded compilation.
I didn't want to reuse the -j option for this, because -num-threads (even if n == 1) does change the generated code.
For details see commit message of r25930.
Swift SVN r26258
