[LLVMdev] Removing types from metadata

[Chris Lattner]

> On Dec 19, 2014, at 1:58 PM, Rafael Espíndola <rafael.espindola at gmail.com> wrote:
>> 
>> IMO, it's important that the three forms of IR match each other
>> closely -- most of all, assembly and C++.  The metadata/value split is
>> most difficult because of C++ API changes; I think the assembly
>> changes are relatively minor and easy enough to script.  (I
>> particularly don't want to maintain parser support for
>> `metadata !{i32 %val}`, which pretends that `i32 %val` is stored in
>> an `MDNode`.  It was a huge cleanup to remove that from the parser.)
>> 
>> What do others think, though?
> 
> I am also in favor of having a single upgrade path: bitcode.

Me too.

-Chris