NFA parser for mbe matcher

[edwin0cheng]

Almost straight porting from rustc one, but a little bit slow :(

rust-analyzer analysis-stats -q . 

From:

Database loaded:     636.11ms, 277minstr
  crates: 36, mods: 594, decls: 11527, fns: 9017
Item Collection:     10.99s, 60ginstr
  exprs: 249618, ??ty: 2699 (1%), ?ty: 2101 (0%), !ty: 932
Inference:           28.94s, 123ginstr
Total:               39.93s, 184ginstr

To:

Database loaded:     630.90ms, 277minstr
  crates: 36, mods: 594, decls: 11528, fns: 9018
Item Collection:     13.70s, 77ginstr
  exprs: 249482, ??ty: 2699 (1%), ?ty: 2101 (0%), !ty: 932
Inference:           30.27s, 133ginstr
Total:               43.97s, 211ginstr

Fixes #4777

[matklad]

Yeah, the regression is sizable, but it's not like we can not do it. I think some of it can be clawed back by just profiling and microoptimiaton.

However, to do that, we really need a benchmark here. I ... now realize we don't have a great benchmarking story at all, but it's better to start somewhere.

@edwin0cheng, could you do a separate PR which includes some bencmarks? I think they should be just usual #[test] tests (with run_slow_test check) which just use StopWatch to observe timings. The bench should live in the mbe crate. The bench should show the same impact for this PR, but it should be much easier to run.

With the benchmark in place, and with a good-first-issuer "optimize this", I'll be more than comfortable to merge this.

[edwin0cheng]

Rebased from #7566 and here is the benchmark result :

before:
mbe expand macro rules: 1.21s, 4966minstr

after
mbe expand macro rules: 2.24s, 8464minstr

2x slower :(

[matklad]

bors r+

Could you create an issue with mentoring instructions on how to run the benchmark? I wonder if we can nerd-snipe someone to look at the code with a profiler...

[bors]

Build succeeded:

• Rust (macos-latest)
• Rust (ubuntu-latest)
• Rust (windows-latest)
• TypeScript (ubuntu-latest)
• TypeScript (windows-latest)

[edwin0cheng]

Could you create an issue with mentoring instructions on how to run the benchmark? I wonder if we can nerd-snipe someone to look at the code with a profiler...

#7857