[llvm] r209007 - Fix most of PR10367.

[Jordan Rose]

Hi, Rafael. We are actually using the feature that allows GlobalAliases to have constant offsets in an out-of-tree branch. I am surprised Akira's test case in CodeGen/X86/alias-gep.ll didn't break with this change, but that's essentially what we're trying to do in code, and r209007 broke our use case. What's the constant offset code going to look like, and can you bump it up in priority so that we don't regress here? Also, any tips on a better test case that would have caught this breakage?

(I usually stay up in the Clang side of things, so I'm not the one to fix this.)

Jordan