Design Meeting Notes, 1/22/2021

[DanielRosenwasser]

Template String Types for Template Expressions

(continued)

• Strong feelings on this?

  • Seems better, and more consistent, with templates with no holes.
  • As soon as you added holes in positions, then you lose it.
  • So part of the idea is we're trying to faithfully reflect what's happening in the type system.

• + concatenations don't do this though - that's very inconsistent.

• If we were to introduce + for any T and U.

• We don't support algebraic operators in general.

  const a: 42 = 42;
  const b: 3  = 3;
  const c = a + b

• In DefinitelyTyped, we recently had a PR with a huge number of strings - one could argue that's strictly better.

  • This was a template type - not a template literal.
  • feat(terra-spacer): Add type defs for terra-spacer DefinitelyTyped/DefinitelyTyped#50748

• Should a new programmer have to see this behavior?

• New programmer will usually see a union type.

  • Will sometimes see a template with an interpolated string though!

• There's a widening/non-widening form, right?

  • Yes

  • But we don't always do literal widening for functions!

  • https://github.com/ionic-team/ionicons/blob/1543daa0103c9065805f7075b5dc6abd1c2dd9c1/scripts/build.ts#L262

  • Simplified

    declare var srcSvgData: SvgData[];
    interface SvgData {
        fileName: string;
        iconName: string;
    }
    
    const c = srcSvgData.map(
        (svgData) => `<a href="./svg/${svgData.fileName}"><svg><use href="#${svgData.iconName}" xlink:href="#${svgData.iconName}"/></svg></a>`,
    );

  • Could argue the problem is that we don't look at the whole program.

• Consistency arguments don't exactly feel useful because you want the literal type so you can compare for literal unions.

  • When you have an infinite set...that feels somewhat less useful.
  • Consistency vs. usability argument, usability feels more compelling to me.

• Could imagine "only create a literal type if it's finite"

  • Still hard to explain "if you fill your hole with a string, you won't get a literal-ish type."

• Maybe the most minimal change is to make return statements do different widening.

• Can always opt in with as const with old behavior.

• There is some sentiment that the behavior is good; but should just make concessions to widening on function returns.

  // This should probably keep working like it does post-PR:
  let a = 12;
  const lit = `${a}px`;
  usePxValue(lit);
  const obj = {
      lit
  };

• But feels strange, that if you wrote this as makePx, and we widened for you, that'd be bad.

• unique symbols widen to symbols.

• Seems like these should work the same.

  let a = 12;
  const c = `${x}px`;
  
  function makePxValue1(x: number) {
      return `${x}px`;
  }
  
  function makePxValue2(x: number) {
      const c = `${x}px`;
      return c;
  }

• If we revert, we can avoid all these contextual rules.

  • Only time we infer literal types is if they're finite.
  • Also, can maintain the string slots when contextually typed.

• Seems like reverting is a reasonable option, so you'd just end up with string unless you write as const - same way with arrays/object literals.

• What kind of inferences would we need to make this better? Seems like these are desirable.

  • Well you have to speculate with these super-specific types - that's why we have widening. So you don't really know.
  • We need to see where/when/how types are mutated.

[awerlogus]

It would be nice if everything will work without features that are specific for .ts files and are not currently supported by checkjs, like as const, for example.

[AlCalzone]

I'd really like a way to opt in with JS which seems impossible since JS has no "as const".

[jcalz]

@DanielRosenwasser 1/22/2020 was more than a year ago; how did you remember all these meeting notes? 😜

[pedrodurek]

It'd be nice having the Template Literal Inference back. I was really expecting this feature, I even mentioned it on our doc 😢 . In my view, as long as the expression from the template string is a literal value, we should appropriately infer that template string (return the merged literal value rather than just string). I hope you consider bringing it back 🙏.