Retitling. The right fix is to kill off ghost linkage, which has always been a hack.

Is the fix to replace this with a function attribute denoting its rematerializability? Or a different redesign where you can query the Function-ModuleProvider pair? (I'd prefer the materialize method be on Function, not the MP.)

I'd suggest adding a ModuleProvider* to the Module (reversing the ownership relation), and moving ModuleProvider::materializeFunction() to GlobalValue::Materialize(). The function currently named GlobalValue::hasNotBeenLoadedFromBitcode() could be renamed GlovalValue::materializable() and could either query getParent()->getMP()->some_dense_map or cache that result in an attribute.

Similarly, ModuleProvider::dematerializeFunction() would move to GlobalValue::Dematerialize(). The current BitcodeReader implementation assert-fails when the function wasn't originally loaded from bitcode, but it would be nice to either return false when the function wasn't dematerializable or have a GlobalValue::dematerializable() query method. ("dematerializable" or "rematerializable"?)

Thoughts?

I have a preliminary fix at http://codereview.appspot.com/193064. I still need to proofread it and check whether it immediately fixes #6107 , but this is the overall direction I'm proposing. Comments are welcome. I'll ping when it's ready for final review.

Fixed in r94686.