cmd/api: API checker doesn't seem to handle Alias uses correctly

[griesemer]

Background: API check failure in https://go.dev/cl/559435.

[randall77]

I think this might not be a bug in the API checker. I think it might actually be a problem with whether we can move types.

Suppose package a has a type T.
Then if we do fmt.Printf("%v\n", reflect.ValueOf(a.T{}).Type()), it prints a.T.
Then if we move T to package b and replace the declaration of T in a with type T = b.T, then that same print statement will now print b.T.
So that move isn't undetectable by callers.

[randall77]

@qiulaidongfeng

[qiulaidongfeng]

I think it might actually be a problem with whether we can move types.

Then if we do fmt.Printf("%v\n", reflect.ValueOf(a.T{}).Type()), it prints a.T.
Then if we move T to package b and replace the declaration of T in a with type T = b.T, then that same print statement will now print b.T.

See https://go.dev/doc/go1compat , Compatibility is at the source level.
See https://go.dev/blog/compat , Perhaps the clearest fact about compatibility is that we can’t take away API, or else programs using it will break.

From the above blog about compatibility, I can't tell if fmt.Printf("%v\n", reflect.ValueOf(a.T{}).Type()) output a.T backward compatibility is guaranteed to not change it?

[griesemer]

@randall77 I seems to me that explicit Alias type nodes should allow us to fix this: that is, if package a; type T = b.T, printing an a.T value could still print a.T with reflect. But I haven't thought it through in detail.

That said, per the original alias proposal, we already have issue with reflect (link).
Also, in the document it seems that we don't mention the reflect name of alias types (not just anonymous field names) - possibly an oversight. In any case, I don't think we must require unchanged reflect type names with the backward-compatibility guarantee.

[griesemer]

Moving to 1.24. Not urgent.

[griesemer]

Didn't get to this. For 1.25.