Design Meeting Notes, 4/9/2021

[DanielRosenwasser]

Generalized Index Signatures

#26797

• People ask for an index signature for symbol.
• But also now that we have template literal types, people want stuff like `data-${string}`.
  • But if you allow things like that, you have to allow generics.
• You can get the same effect today with mapped types.
• Today we do relation checking for string and number index signatures in a pretty loose way.
  • Seems like we need some loose rules.
• Do we? Seems like you can create a slight loosening.
• You need to be sensitive to the signatures with which an overload overlaps.
• With noUncheckedIndexedAccess, we have another layer of relating differently.
• Would it be possible to have generic index signatures?
  • Maybe, but we don't exactly know how that would work.
• The original idea was to unify mapped types and index signatures.

importsNotUsedAsValues: preserve Doesn't Preserve Names

#43393

Perf Regressions

#43422

• Caused by better aliasing of union types.
• 1.5s to 5s regression
  • The actual repro was 30s to 50s.
• But this is expected - like a see-saw, we did better on other codebases, but we'll do worse on this one.
• It's expected because we're manufacturing new type IDs.
  • But we can potentially do some cleverness with the non-aliased union type.
• Experiment.

#43437

• Swapping a declaration/statement actually slows down the type-checking process.
• We cache a whole bunch of type relationships before hitting one.
  • Doing the assignment first forces a ton of work to happen against an anonymous type.
  • While we do this, we stick in the keys of the things we're relating into a stack.
  • That stack gets about 17,500 entries inserted.
  • But as you relate, it's not caching.
  • So you have an O(n * m) check.
  • Fix is to use a map or set.
  • Seems to avoid the odd regression.