[PATCH] D18683: Fix bug with duplicate struct types in Linker

[Rafael Espíndola via llvm-commits]

> We could keep track of where each function came from, and if we generate a call to a linked function, grovel around in the linked function signatures, work out what type the linker decided to assign it and then generate a bunch of bitcasts (since the pointer might be the original type if it came from the source module), but I'd rather not resort to doing that.  It seems easier to fix the problem at the source (and also solve it for anyone else who's building jits using LLVM in a similar way).

It is not that much. It is just that instead of doing
"Dst.getTypeByName(Name)" you would do "Dst.getNamedValue(Name)" and
use the type of that. Where is the getTypeByName in the codebase you
pointed to?

> I have no idea what the intended behaviour of the Linker is since there's not much documentation about what the linker is/isn't allow to do, but this is something that has worked historically. It wouldn't expect it to work if the struct types had different names originally, but this is just reversing some renaming that LLVM did.

Merge as many types as possible. The use of names is just an heuristic
to try to merge more types in the presence of cycles and opaque types.
With typeless pointers we will be able to merge *all* isomorphic types
without ever looking at the name.

Cheers,
Rafael