Figure out why Brc::new is slower than Arc::new

[Techcable]

Currently, Brc::new consistently benchmarks as 20-100% slower than Arc::new. While I consider 20% an acceptable slowdown, being 2x as slow as Arc::new is unacceptable.

This has been a problem for a real-world user. A post on reddit by @nynjawitay said

I clone them a bunch, but I call new a LOT in this app. I switched out all the Arcs in this app for Brc and startup went from taking 4 seconds to taking 13 seconds. It's not an open source project or I'd share more.

This is even worse than microbenchmarks would suggest.

I need to figure out why there is a slowdown at all. The only thing we do differently than Arc::new is load a TLS and execute a couple of very rarely taken branches.

[Techcable]

The thread-local storage access appears to be the problem. I did a hacky test passing an explicit short_id context parameter. This fixes the performnace issue, although there could be other contributing factors like eliminating some rarely taken branches.

It is strange because this TLS access doesn't slow down Brc::clone or Brc::drop relative to Arc. I suspect it has something to do with storing the ShortThreadId in memory rather than doing a comparison against it and branching.