[llvm] r263742 - Bitcode: Error out instead of crashing on corrupt metadata

[Rafael Espíndola via llvm-commits]

> I'd need to create another four binaries with manual bit flips for the
> testing of this to be more than just lip service, and I really disagree
> that such tests are at all useful.
>
> If we want to make it a requirement that we can test the bitcode reader
> in a reasonable way before it's worth fixing this kind of crash on bogus
> data, fine. I can revert this change for now and just let it crash. I'm
> not convinced the tests that I am capable of generating at this point
> are worth the cost of maintaining them, so I'm against committing them.

What is the cost of maintaining them? You are just testing what
llvm-dis does on a broken bitcode. The bitcode would never change. The
script would only change to reflect changes in the error message.

Cheers,
Rafael