Performance regression in 4.0 with semantics-only createWatchProgram?

[dfreeman]

The wiki here has a page that outlines ways to programmatically create a watched build, and suggests that createSemanticDiagnosticsBuilderProgram is likely preferable to createEmitAndSemanticDiagnosticsBuilderProgram for use cases that aren't interested in emit.

The ember-cli-typescript package uses the APIs outlined in that wiki page to integrate typechecking results into Ember CLI builds, and there've been several reports of slower re-checks after a file change from folks that have updated to typescript@4.

TypeScript Version:

• 3.9.7 (works as expected)
• 4.0.3 (exhibits slowdown)
• 4.1.0-dev.20200928 (exhibits slowdown)

Search Terms:

• createWatchProgram
• createSemanticDiagnosticsBuilderProgram
• 4.0 performance

Code

const ts = require('typescript');
const host = ts.createWatchCompilerHost(
  'tsconfig.json',
  { noEmit: true },
  ts.sys,
  ts.createSemanticDiagnosticsBuilderProgram
);

ts.createWatchProgram(host);

Expected behavior:
Re-checks are at least as fast with createSemanticDiagnosticsBuilderProgram as with its counterpart that tracks emit.

Actual behavior:
In a project that takes on the order of 15s to typecheck at startup, re-checking takes less than 1s with the above invocation in 3.9, as well as in 4.0 and the most recent 4.1 nightly when using createEmitAndSemanticDiagnosticsBuilderProgram.

When using createSemanticDiagnosticsBuilderProgram in 4.0 and 4.1, though, re-checking consistently takes around 5s.

[sheetalkamat]

@dfreeman Can you please provide the repro where this happens.. i ran this on sample and it seems to be using cached diagnostics correctly.

[dfreeman]

@sheetalkamat Apologies, the project I mentioned above is private, but I'd (probably foolishly 😄) assumed since it had been reported by several others that it would reproduce easily.

Here's a reproduction repo I put together: https://github.com/dfreeman/ts-40808-perf-demo
(the use of Babel there is incidental; it's just the first thing that popped to mind to give some mass to the typecheck without generating a bunch of code myself)

[sheetalkamat]

Note that the issue is coming up because of changes in #39122 where we emit build info.
Your code is doing the right thing because you want to emit tsbuildinfo so it can be reused.

[andreyfel]

@sheetalkamat will the fix be released as a patch fix for 4.0.x series?

[sheetalkamat]

@DanielRosenwasser is it feasible to take #40880 for 4.0 patch

[DanielRosenwasser]

Yeah, I think that'd be reasonable.