Make T[never] = never instead of erroring or being any

[kpdonn]

This change makes it so that when indexing a type with never(e.g. T[never]), the result is never. Previously it was an error to explicitly index a type with never and it would non-intuitively become any if it was inferred somewhere inside generics.

There is more discussion about the reasoning for the change and the implications of it from @sandersn and @yortus in #22042.

Fixes #22042

[yortus]

@kpdonn maybe clarify that this change only affects types with no index signature. So for example:

• {[x: string]: boolean}[never] is boolean (unchanged behaviour)
• {foo: boolean}[never] is never (new behaviour, previously was an error)

With that qualification, this is a non-breaking change because it only affects cases that are currently errors.

[mhegazy]

@ahejlsberg and @sandersn any comments?

[sandersn]

After discussing it in person with @ahejlsberg, we didn't like the fact that ({ [n: number]: boolean})[never] is inconsistent: it's not never like everything else, just to make sure that lodash doesn't break. To fix this inconsistency, we are thinking about (finally) introducing the numeric-string type, which would what keyof { [n: number]: boolean } would be:

type Ni = { [n: number]: boolean }
type Kni = keyof Ni // numeric-string
type Nikni = Ni[Kni] // boolean

lodash would no longer break because T[keyof T] would never end up indexing a numeric-index-only type with never, but with numeric-string. Then this PR could change so that never-indexing applies to all types, not just those with properties.

[sandersn]

We should wait on this until we have decided what to do about the numeric-string type.

[ahejlsberg]

I'm actually okay with taking this PR since it moves us closer to where we eventually want to be, i.e. that T[never] is always never regardless of T. Longer term, if and when we fix keyof { [x: number]: T } to produce numeric-string (or whatever we call it) instead of never, we can then also change { [x: number]: T }[never] to be never.

That said, we need a small change to the PR. I will review below.

[ahejlsberg]

Move these three lines up and insert then before line 8122. That way an obj[key] expression where key is an expression of type never will produce never (instead of the any it produces now).

[kpdonn]

@ahejlsberg Updated with the requested change and also a test for the obj[key] case where key is never.

Not sure why the CI builds failed though, I ran gulp runtests locally before committing and did not have failures.

Edit: Nevermind, looks like rebasing off of latest master resolved the CI build failures.

[sandersn]

All right, it’s in! Thanks @kpdonn!