Implement pick()

[markfairbanks]

Closes #341

Overview: Translate pick() and across() calls to a data.table() call.

devtools::load_all(".")
#> ℹ Loading dtplyr

df <- lazy_dt(tibble(x = 1:3, y = 1:3, z = c("a", "a", "b")))

df %>%
  mutate(row_sum = rowSums(pick(x, y)))
#> Source: local data table [3 x 4]
#> Call:   copy(`_DT1`)[, `:=`(row_sum = rowSums(data.table(x = x, y = y)))]
#> 
#>       x     y z     row_sum
#>   <int> <int> <chr>   <dbl>
#> 1     1     1 a           2
#> 2     2     2 a           4
#> 3     3     3 b           6
#> 
#> # Use as.data.table()/as.data.frame()/as_tibble() to access results

df %>%
  mutate(new = rowSums(across(c(x, y), ~ .x + 1)))
#> Source: local data table [3 x 4]
#> Call:   copy(`_DT1`)[, `:=`(new = rowSums(data.table(x = x + 1, y = y + 
#>     1)))]
#> 
#>       x     y z       new
#>   <int> <int> <chr> <dbl>
#> 1     1     1 a         4
#> 2     2     2 a         6
#> 3     3     3 b         8
#> 
#> # Use as.data.table()/as.data.frame()/as_tibble() to access results

[eutwt]

Top level pick calls aren't expanded in this PR's current state. For example this fails

test_that("top level across() and pick() are expanded", {
  df <- lazy_dt(tibble(x = 1:3, y = 1:3, z = c("a", "a", "b")))
  expect_identical(group_by(df, x, y), group_by(df, pick(x, y)))
  expect_identical(group_by(df, x, y), group_by(df, across(c(x, y))))
})

My first thought was to just change the top_across definition in capture_dots() so that it's TRUE for pick() calls too.

dtplyr/R/tidyeval.R

Line 33 in 5c80dc3

top_across <- map(dots, quo_is_call, "across")

But I think the info we need is just whether it's a top-level expression; we could avoid checking for across/pick calls at that stage and just set is_top = TRUE. More concretely, this seems to work:

• Rename is_top_across to just is_top everywhere it appears in tidyeval.R
• In capture_dots() and capture_new_vars(), set dt_squash(..., is_top = TRUE) for all dots, rather than checking dots for across calls
• Change the dt_squash call below to use is_top = FALSE (arguments inside calls being squashed aren't top level)
  

  dtplyr/R/tidyeval.R

  Lines 250 to 253 in 5c80dc3

  	} else {
  	x[-1] <- lapply(x[-1], dt_squash, env, data, j = j, is_top_across)
  	x
  	}

I'd lean towards the second option, but I don't think there are real issues with either, just a style difference.

[markfairbanks]

Ready for review again. Simplified per the suggestion - the logic seems much better this way.

[eutwt]

Looks good. I'm pretty happy to have this feature