Optimize SbTag::eq

[Noratrieb]

The code before generated really bad code with a branch.
This nudges LLVM towards being smarter and simply comparing
the integers.

See #2214 (comment)

[Noratrieb]

This change was suggested by @saethlin somewhere, but I don't recall where.

[Noratrieb]

That's an interesting clippy lint, first time I've seen it

[saethlin]

I suggested this here: rust-lang/rust#49892 (comment)

How does this compare to #2214 ? Obviously we could have both of these (and this change plays nicely with #1935 because it also benefits the cache lookup code path) but it would be good to have benchmarks for both.

Generally I'm okay with the sort of change in this PR though I'd rather solve this codegen problem by improving derive(PartialEq) rather than using a non-obvious implementation like this. The explanation for why we have this is "rustc/LLVM is producing bad code so we're hacking around it" as opposed to something like #2214 where the explanation is "Stacked Borrows has a particular property and we're reflecting that in this implementation".

Also as a sidenote, all this optimization in this and #2214 is very much in the weeds unless you rebase those onto #1935 which I promise I will make progress on soon.

[saethlin]

Also since we're dealing with such small optimizations I wonder if this intersects with #2132

[Noratrieb]

For benches, see my comments in #2214.

[RalfJung]

Thanks, this makes sense!
@bors r+

[bors]

📌 Commit 93db9a6 has been approved by RalfJung

[bors]

⌛ Testing commit 93db9a6 with merge 4d6eca1...

[bors]

☀️ Test successful - checks-actions
Approved by: RalfJung
Pushing 4d6eca1 to master...