[llvm] [KeyInstr] Add bitcode support (PR #144102)

[Orlando Cazalet-Hyams via llvm-commits]

OCHyams wrote:

> Hmmm -- that's a noteworthy cost (in some scenarios up to 5%) on bitcode size (with -g). While we can rely on the position of "this is delivering a superior debugging experience" because there's strictly more information in the object-file / bitcode, it's worth going the extra mile to see if we can cut the file size increase down a bit.
> 
> IIRC we talked in the past about the numeric distribution of atom-groups and how they're keyed on {Scope,InlinedAt}, thus each independent function can have its own number-sequence, but IIRC we also thought about making them globally unique-ish. Or something. The point being that with VBR-encoded atom-groups we pay a cost for high-numbers that perhaps we could avoid by being smarter?

Doing anything "smart" looks tricky, but I found an easy win in #146497 which seems to cover the cost of the extra fields.

https://github.com/llvm/llvm-project/pull/144102