Repository navigation
Relations based on variance markers do not take into account parameters used as key types. #29698
Description
TypeScript Version: 3.3
Code
type Record2<K extends keyof any, T> = {
[P in K]: T;
};
function defaultRecord(x: Record<'a', string>, y: Record<string, string>) {
x = y; // no error, but error expected.
}
function customRecord(x: Record2<'a', string>, y: Record2<string, string>) {
x = y; // no error, but error expected.
}
function mixed1(x: Record2<'a', string>, y: Record<string, string>) {
x = y; // error
}
function mixed2(x: Record<'a', string>, y: Record2<string, string>) {
x = y; // error
}Playground Link: link
Related Issues: #28798
Activity
[-]Relating generic mapped types does not instantiate constraint type.[/-][+]Relating generic mapped types does not instantiate constraint type (in some cases).[/+]ahejlsberg commentedon Feb 6, 2019
The cause of this is #27804. With that PR we started measuring variances for type aliases to object and conditional types. The variance measuring logic determines that
Kis contravariant andTis covariant inRecord<K, T>. This is almost true: An instantiation whereKisstringis unrelated to an instantiation whereKis a literal type (because a string index signature doesn't satisfy an actual property).Wesley Wigham (@weswigham) Perhaps we should exclude aliases to mapped types from the variance measuring shortcut?
jack-williams commentedon Feb 6, 2019
Thanks for looking into this Anders Hejlsberg (@ahejlsberg). I guess the difference I'm seeing when using two different aliases is because variance measurements are not shared across different aliases of identical types?
This seems a similar problem to #28798.
I guess the underlying problem is that the semantics of
stringornumberas a value type is different to the semantics as a key type. The logic to relate keys can reuse the logic to relate values by inverting the relationship, which works in most cases, but not whenstringis involved.Would another---albeit more costly ---solution be to have a key relation that interprets key types into their value type equivalent before using the value relation?
ahejlsberg commentedon Feb 7, 2019
Also see #29393 which appears to be an effect of variance measuring for aliased types.
[-]Relating generic mapped types does not instantiate constraint type (in some cases).[/-][+]Relations based on variance markers do not take into account parameters used as key types.[/+]jack-williams commentedon Feb 7, 2019
I changed the title to something that is hopefully closer to the actual issue. Feel free to change if it is misleading.
My understanding now, after your comments and reading the source, is that a contravariant marker is inferred for the parameter
Kafter:but this does not take into account the fact that
Kis used a key type. Is this correct?Would it be completely unreasonably to collect uses as key types along with variances, and use that to correctly determine when and how two instantiations of
Kshould be related? I feel that there are some wider issues (#29393) that I don't fully understand, so may well be in completely the wrong ballpark.Am I right in saying that this is going to be a problem if #26797 lands? Where you could have: