[llvm] [IR] Allow MDString in operand bundles (PR #110805)

[Serge Pavlov via llvm-commits]

spavloff wrote:

> To check my understanding, the reason why we need the bitcode change is that we only encode the operand type if the value ID is larger than the instruction ID, which is never going to be the case for metadata, so we can't use the type to distinguish metadata?

Yes, the type index cannot be used for this purpose, because it is absent.

Also, there is no bitcode change here, everything that worked previously produces the same bit code. It is only the new feature (metadata in operande bundles) that require the new rule.

> And the reason this is not a problem for normal call arguments is that it doesn't encode the types and instead takes them from the function type?

Because previously deserializer always could expect either SSA value, or metadata, but not both.



https://github.com/llvm/llvm-project/pull/110805