Simply create a new file when calling updateFile

[DanielRosenwasser]

Previously the code used TypeScript's updateFile, which performs an incremental parse; however, if we don't expect the contents to be similar enough, this wastes a lot of work. Reuse should still be done for untouched files from old versions of the program.

If I measured correctly (big if!), the before is something like this:

Duration: 358.75s, Total samples = 330527.55ms (92.13%)
Showing nodes accounting for 234432.24ms, 70.93% of 330527.55ms total
Dropped 12123 nodes (cum <= 1652.64ms)
      flat  flat%   sum%        cum   cum%
43480.84ms 13.15% 13.15% 43480.84ms 13.15%  (garbage collector)
22431.38ms  6.79% 19.94% 29808.65ms  9.02%  scan
14276.95ms  4.32% 24.26% 16485.47ms  4.99%  explainThemeScope
8470.93ms  2.56% 26.82% 17812.28ms  5.39%  doJSDocScan
7503.03ms  2.27% 29.09% 50711.16ms 15.34%  bind
6731.53ms  2.04% 31.13% 14063.59ms  4.25%  declareSymbol

And the after (with this PR) looks like this:

Duration: 136.13s, Total samples = 104413.98ms (76.70%)
Showing nodes accounting for 66596.18ms, 63.78% of 104413.98ms total
Dropped 10304 nodes (cum <= 522.07ms)
      flat  flat%   sum%        cum   cum%
12840.82ms 12.30% 12.30% 14840.34ms 14.21%  explainThemeScope
10529.17ms 10.08% 22.38% 10529.17ms 10.08%  (garbage collector)
3089.92ms  2.96% 25.34% 24318.65ms 23.29%  tokenizeWithTheme
2659.21ms  2.55% 27.89%  2659.21ms  2.55%  readFileUtf8
2524.28ms  2.42% 30.31% 39548.81ms 37.88%  visit
2190.34ms  2.10% 32.40%  2713.73ms  2.60%  stat

I did run a clean in between steps, so hopefully I did everything correctly - but this is a pretty nice drop!

[DanielRosenwasser]

I don't see the same results in CI, and apparently it was right for me to call out that "big if!" - switching back to the v2 branch, running clean, and re-running, I see similar results. So some stuff must be getting cached. That said, this is quite a bit of work being done.

[jakebailey]

Yeah, locally on rebuild, things are cached. If you're profiling, you kinda want to clean everything and rebuild from scratch to get the "true" time spent in CI.

[jakebailey]

Is this required to make the fix?

[andrewbranch]

What are your performance results from? This is a public package that (I guess) we encourage people to use for all kinds of stuff. My guess is that most instances of updateFile in the wild would be small changes that benefit from an incremental parse. Do we have any evidence that our usage, which benefits from this change, is representative of community usage?

[jakebailey]

This was actually from pprof-it gatsby build in this repo, on the thread that does the TS stuff, IIRC.

[andrewbranch]

Are we calling updateFile when we switch between compiling different twoslash snippets or something? Seems like we should just have/use a different API for that, and let updateFile be for, conceptually, updating a file.