cur_data() / cur_data_all() not simplifying list columns

[romainfrancois]

closes #5901

library(dplyr, warn.conflicts = FALSE)

f <- function(x) { print(x, n = nrow(x) )}
bla <- tibble(i = 1:2, x = list(tibble(a = 1), tibble(a = 2))) %>% 
  rowwise() %>% 
  summarise(f(cur_data_all()))
#> # A tibble: 1 x 2
#>       i x               
#>   <int> <list>          
#> 1     1 <tibble [1 × 1]>
#> # A tibble: 1 x 2
#>       i x               
#>   <int> <list>          
#> 1     2 <tibble [1 × 1]>

^{Created on 2021-09-17 by the reprex package (v2.0.0)}

[romainfrancois]

However, the data is still simplified as usual when used directly in the expression, or through across():

library(dplyr, warn.conflicts = FALSE)
library(tibble)

df <- tibble(x = list(tibble(a = 1), tibble(a = 2))) %>%
  rowwise()

summarise(df, is_tibble(x))
#> # A tibble: 2 x 1
#>   `is_tibble(x)`
#>   <lgl>         
#> 1 TRUE          
#> 2 TRUE
summarise(df, across(x, is_tibble))
#> # A tibble: 2 x 1
#>   x    
#>   <lgl>
#> 1 TRUE 
#> 2 TRUE

^{Created on 2021-09-17 by the reprex package (v2.0.0)}

[hadley]

Can you explain why we have to re-wrap in a list? (As opposed to not unwrapping somewhere?)

[romainfrancois]

It happens in this line:

dplyr/src/chop.cpp

Line 20 in bb84360

if (rowwise && vctrs::vec_is_list(column) && Rf_length(column) > 0) {

void dplyr_lazy_vec_chop_grouped(SEXP chops_env, SEXP rows, SEXP data, bool rowwise) {
  SEXP names = PROTECT(Rf_getAttrib(data, R_NamesSymbol));
  R_xlen_t n = XLENGTH(data);

  const SEXP* p_data = VECTOR_PTR_RO(data);
  const SEXP* p_names = STRING_PTR_RO(names);
  for (R_xlen_t i = 0; i < n; i++) {
    SEXP prom = PROTECT(Rf_allocSExp(PROMSXP));
    SET_PRENV(prom, R_EmptyEnv);
    SEXP column = p_data[i];
    if (rowwise && vctrs::vec_is_list(column) && Rf_length(column) > 0) {
      SET_PRCODE(prom, column);
    } else {
      SET_PRCODE(prom, Rf_lang3(dplyr::functions::vec_chop, column, rows));
    }
    SET_PRVALUE(prom, R_UnboundValue);

    Rf_defineVar(rlang::str_as_symbol(p_names[i]), prom, chops_env);
    UNPROTECT(1);
  }

  UNPROTECT(1);
}

When we setup the promise that makes the chops, and this is a rowwise data, we can skip using vec_chop() and simply use the list columns as the chops.

This gives us the rowwise simplification for when we refer to the list column in an expression (not having to [[1]]).

[romainfrancois]

... and for non-original columns, i.e. those that are added, they go through the <DataMask>$add_one() method:

add_one = function(name, chunks, result) {
      if (inherits(private$data, "rowwise_df")){
        is_scalar_list <- function(.x) {
          vec_is_list(.x) && length(.x) == 1L
        }
        if (all(map_lgl(chunks, is_scalar_list))) {
          chunks <- map(chunks, `[[`, 1L)
        }
      }

      .Call(`dplyr_mask_add`, private, name, result, chunks)
    },

[hadley]

Ah makes sense. Thanks!