Should `parent.frame()` in S7 methods match S3?

[t-kalinowski]

Spun out from #592.

parent.frame() inside an S7 method currently differs from S3:

library(S7)

foo <- new_generic("foo", "x")
method(foo, class_any) <- function(x) parent.frame()

bar <- function(x) UseMethod("bar")
bar.default <- function(x) parent.frame()

foo(1)
#> <environment: 0x117954580>
bar(1)
#> <environment: R_GlobalEnv>

It would be nice if parent.frame() in an S7 method matched S3 here.

[hadley]

We should discuss this. I think there's some disagreement about how much we want to elide the details of method dispatch. I think I'd prefer to keep it explicit (as long as it's consistent, so that parent.frame(2) would always get you the caller of the generic).

[lawremi]

To clarify, it's consistency with both S3 and S4.

I agree that it's confusing when the return value of sys.parents() repeats values (and parent.frame() skips frames):

Browse[1]> sys.parents()
[1] 0 0
Browse[1]> where
where 1 at #1: bar.default(1)
where 2: bar(1)

The at #1 is not intuitive and is easy to miss.

Of course with the current S7 implementation, the same complexity exists, and there isn't even an indicator in the trace:

Browse[1]> sys.parents()
[1] 0 1 1
Browse[1]> where
where 1: `method(foo, class_any)`(x = 1, ...)
where 2: S7::S7_dispatch()
where 3: foo(1)

One direction would be to improve the clarity of the traces, so that the user understands what is happening.

[hadley]

I wonder if we should expose two helpers for what you'd typically use parent.frame() — getting the frame of the generic (for error messages) and getting the user frame (for various NSE shenanigans). If we provided these helpers we could ensure that they always worked even when super() was in the mix.

That would also free us to consider alternative implementation in the future.

[lawremi]

Yes, those helpers be.. helpful.. regardless of what we do with this issue. Abstraction is a good idea here.

[hadley]

Maybe S7_generic_frame() and S7_user_frame()?

[lawremi]

Seems reasonable.