Make if_any() and if_all() consistent in all contexts

[DavisVaughan]

Closes #7746
Closes #7077

It turns out that our application of if_any() and if_all() was fairly inconsistent. This is due to the fact that they are tricky to get right, we have 2 different implementations of them. One for the expansion case and one for the evaluation case. I've now tried to unify these to use the same underlying implementation, dplyr_list_pany() or dplyr_list_pall() depending on the scenario (which just use vec_pany() and vec_pall() under the hood)

In addition to greater consistency across the board, you'll also note that in errors the In argument: label also now reports the original expression pre-expansion in the filter() cases, which is a much better error

In the examples below, for filter(), note that adding () around the if_any() or if_all() calls triggers the evaluation case rather than the expansion case.

library(dplyr)

# With zero inputs, if_any

# Before
df <- tibble(x = 1:2)
filter(df, if_any(c(), identity))
#> # A tibble: 0 × 1
#> # ℹ 1 variable: x <int>
filter(df, (if_any(c(), identity)))
#> # A tibble: 2 × 1
#>       x
#>   <int>
#> 1     1
#> 2     2
filter(df, any())
#> # A tibble: 0 × 1
#> # ℹ 1 variable: x <int>

# After
df <- tibble(x = 1:2)
filter(df, if_any(c(), identity))
#> # A tibble: 0 × 1
#> # ℹ 1 variable: x <int>
filter(df, (if_any(c(), identity)))
#> # A tibble: 0 × 1
#> # ℹ 1 variable: x <int>
filter(df, any())
#> # A tibble: 0 × 1
#> # ℹ 1 variable: x <int>

# With one non-logical input

# Before
df <- tibble(x = 1:2)
filter(df, if_any(x, identity))
#> Error in `filter()`:
#> ℹ In argument: `(function (x) ...`.
#> Caused by error:
#> ! `..1` must be a logical vector, not an integer vector.
filter(df, (if_any(x, identity)))
#> Error in `filter()`:
#> ℹ In argument: `(if_any(x, identity))`.
#> Caused by error:
#> ! `..1` must be a logical vector, not an integer vector.
mutate(df, a = if_any(x, identity))
#> # A tibble: 2 × 2
#>       x     a
#>   <int> <int>
#> 1     1     1
#> 2     2     2

# After
df <- tibble(x = 1:2)
filter(df, if_any(x, identity))
#> Error in `filter()`:
#> ℹ In argument: `if_any(x, identity)`.
#> Caused by error in `if_any()`:
#> ! `x` must be a logical vector, not an integer vector.
filter(df, (if_any(x, identity)))
#> Error in `filter()`:
#> ℹ In argument: `(if_any(x, identity))`.
#> Caused by error in `if_any()`:
#> ! `x` must be a logical vector, not an integer vector.
mutate(df, a = if_any(x, identity))
#> Error in `mutate()`:
#> ℹ In argument: `a = if_any(x, identity)`.
#> Caused by error in `if_any()`:
#> ! `x` must be a logical vector, not an integer vector.

# In general, with non-logical types resulting from applying `.fns` we now error
# more appropriately

# Before
df <- tibble(x = c(TRUE, FALSE), y = c("a", "b"))
filter(df, if_any(c(x, y), identity))
#> Error in `filter()`:
#> ℹ In argument: `|...`.
#> Caused by error in `<function(x) x>(x) | <function(x) x>(y)`:
#> ! operations are possible only for numeric, logical or complex types
filter(df, (if_any(c(x, y), identity)))
#> Error in `filter()`:
#> ℹ In argument: `(if_any(c(x, y), identity))`.
#> Caused by error in `op()`:
#> ! operations are possible only for numeric, logical or complex types

# After
df <- tibble(x = c(TRUE, FALSE), y = c("a", "b"))
filter(df, if_any(c(x, y), identity))
#> Error in `filter()`:
#> ℹ In argument: `if_any(c(x, y), identity)`.
#> Caused by error in `if_any()`:
#> ! `y` must be a logical vector, not a character vector.
filter(df, (if_any(c(x, y), identity)))
#> Error in `filter()`:
#> ℹ In argument: `(if_any(c(x, y), identity))`.
#> Caused by error in `if_any()`:
#> ! `y` must be a logical vector, not a character vector.

[DavisVaughan]

Plucked from ellmer, with modifications. I'm just trying it out.

[DavisVaughan]

I'm hoping to just remove this in favor of the vctrs versions soon, but I needed to get the semantics of it right enough to be able to add all the tests here

[DavisVaughan]

Having an initial value is an important reason that the 0 and 1 input cases now "just work" correctly. This was NULL before.

[DavisVaughan]

This is the same kind of trick we use in as_pick_expansion() for the expansion path. I've carefully tried to match it to the evaluation path.

Basically we replace if_any(c(x, y), fn) with something like

x <- list(x = x, y = y)
ns <- asNamespace("dplyr")
size <- ns[["dplyr_list_size_common"]](x, absent = 1L, call = call(if_any()))
x <- ns[["dplyr_list_recycle_common"]](x, size = size, call = call(if_any()))
ns[["dplyr_list_pany"]](x, size = size, error_call = call(if_any()))

[DavisVaughan]

Along the way I nearly changed across() to recycle inputs to the group size rather than recycling them to their common size. I think that would have been a mistake so I've added a test to prevent us from ever thinking of doing this.

[DavisVaughan]

This is a mega test to make sure things are consistent everywhere.

It's obviously a lot of tests, but I think we really do need them all to be sure we aren't missing an edge case. These are all very hard to reason about since there are so many dimensions that intersect (filter vs mutate, expansion vs evaluation, groups vs no groups, etc)

Let's put it this way, I feel way more confident about this now that we have this test that hits every edge case

[lionel-]

This previous expansion was written with dbplyr and other backends in mind. We thought the dplyr expansions could be used as is in these other settings, removing some work for them as they would already know how to deal with the bare expansion.

Unfortunately, as the linked issues reveal, these bare expansions don't work that well for the dplyr backend because they are missing things like input validation. If we add these in the expansion, then the original purpose of generic translation is defeated.

We've never actually pushed towards using these expansions in other packages, so although this feels like a step backward, we don't currently lose anything by making the expansion untranslatable.

@DavisVaughan Maybe add a comment about why we still need the expansion at all (to avoid tidyselect getting evaluated on every group).