Design Meeting Notes, 3/26/2024

[DanielRosenwasser]

Format Types Consistently in Baselines and Underlining Type Node Generation

• Apply same single-line style to copied nodes as manufactured ones #57890

• Underline type nodes without position mappings in .types baselines weswigham/TypeScript#69

• As we try to maximize node reuse, we realized we have no indication as to whether that is taking place apart from whitespace as a heuristic.

• Even then, preserving whitespace isn't always desirable.

• This PR adjusts type baselines to always apply the same formatting consistently.

• A separate PR adds an underline to indicate when a node was synthesized.

• What are we trying to catch here?

  • Lack of node reuse.
  • Could exhaustively check things, but better off using our current infrastructure.

• Kind of a pain to read these.

• Is it arguable that the underlines actually indicate that you should pay attention even if you only care about the types?

  • Maybe.

• Aside: do we need symbol baselines?

tsc --noCheck

#57934

• For TypeScript? Isn't that weird??
• Idea: allows people to perform a check and a build separately.
• Declaration emit actually surprisingly works here. Most of the information can be calculated lazily, and so it avoids a full check.
• Unfortunately, JavaScript emit actually calculates some information in the checker.
• This is very surprisingly fast.
• The only downside of using --noCheck for declaration emit is that it may differ in output from stuff like union ordering (which depends on types being assigned IDs, which is sensitive to check order).
• Our main goal is to be a type-checker.
• How does this work with syntax errors?
  • You won't get them.
• Unclear if it's a good idea to bring this in without making this possible for the JS side. We need to try decoupling JS emit to see what if anything regresses.
• Mixture of support for the concept of --noCheck, but if we can unlock this for transpileModule that would be a big win.
• Don't want to add a command line flag just yet.
• Bike-shed a name.
  • skipCheck?
  • noCheck to parallel noEmit?

Respect type in package.json in more module modes

#57896

• Some notable changes:
  • A .cjs module under --module esnext got emitted with ES syntax; no longer does.
  • People using .cts files under --moduleResolution bundler --module esnext (aside: nowadays --module preserve is what most of these people want), we would resolve with the import condition, now resolves with the require condition.
    • Kind of weird under --preserve, but 🤷‍♂️.
  • Under --module preserve, we issue something like verbatimModuleSyntax errors because it won't be interpreted the same way from a declaration file.
• Did not add logic to disallow CommonJS input files. Presumably a program-level error if we wanted to add it.