[lld] [lld] Report temporal BP profile resolution (PR #212127)

[Karim Alweheshy via llvm-commits]

karim-alweheshy wrote:

Agreed — this is a profile-consumption diagnostic, not a monotonic quality or staleness score. Expected cross-image names, or other intentionally extra trace names, can lower the resolved ratio without indicating a problem.

I think both raw counts are useful because they answer different questions:

- reference count measures how much of the temporal trace stream BP actually consumes, including repetition;
- unique-root count measures naming breadth, so frequently repeated hot names cannot mask unresolved-name coverage.

The production framework case illustrates the distinction: 44.32% of references resolved, but only 31.51% of unique roots did. That divergence is the reason to retain both, while interpreting them as consumption signals rather than a quality score.

https://github.com/llvm/llvm-project/pull/212127