vec_rbind() performance regression on R 4.0 with df-cols

[DavisVaughan]

Performance regression when rbinding lots of data frames that have df-cols. I'm sure this has to do with making extra copies, but I'm not sure where yet. I'll take a look.

This is with dev vctrs.

R 3.6

library(vctrs)

df_col <- new_data_frame(list(x = 1:2))
df <- new_data_frame(list(y = df_col))

x <- rep_len(list(df), 10000)
y <- rep_len(list(df_col), 10000)

lst_rbind <- function(x) {
  vec_rbind(!!!x)
}

bench::mark(lst_rbind(x))
#> # A tibble: 1 x 6
#>   expression        min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr>   <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 lst_rbind(x)   37.5ms   37.9ms      26.2     277KB     52.4

bench::mark(lst_rbind(y))
#> # A tibble: 1 x 6
#>   expression        min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr>   <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 lst_rbind(y)   16.5ms   18.2ms      55.1     274KB     16.5

R 4.0

library(vctrs)

df_col <- new_data_frame(list(x = 1:2))
df <- new_data_frame(list(y = df_col))

x <- rep_len(list(df), 10000)
y <- rep_len(list(df_col), 10000)

lst_rbind <- function(x) {
  vec_rbind(!!!x)
}

bench::mark(lst_rbind(x))
#> Warning: Some expressions had a GC in every iteration; so filtering is disabled.
#> # A tibble: 1 x 6
#>   expression        min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr>   <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 lst_rbind(x)    316ms    352ms      2.84     764MB     48.2

bench::mark(lst_rbind(y))
#> # A tibble: 1 x 6
#>   expression        min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr>   <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 lst_rbind(y)   15.8ms   17.6ms      55.6     274KB     23.4

[DavisVaughan]

It seems like when the output df-col is restored with vec_restore() before assigning it into out, that somehow increments the refcnt on each individual column of the df-col from 1 to 2, so then the columns of the output df-col are needlessly copied at the next assignment iteration

vctrs/src/slice-assign.c

Line 395 in 843c03c

assigned = vec_restore(assigned, out_elt, R_NilValue);

[DavisVaughan]

Hmm okay, so vec_restore() on a bare data frame / tibble passes through to bare_df_restore_impl() which calls r_clone_referenced() here

vctrs/src/proxy-restore.c

Lines 116 to 118 in 843c03c

	static SEXP bare_df_restore_impl(SEXP x, SEXP to, R_len_t size) {
	x = PROTECT(r_clone_referenced(x));
	x = PROTECT(vec_restore_default(x, to));

This is a problem, because this is a df-col that we are restoring, meaning that it has already been set inside a data frame and the df-col itself has a refcnt of 1 already. So it is referenced, and a shallow duplication does happen here. This triggers a refcnt bump of all of the columns of that df-col, which bumps them from 1 up to 2.

It is possible we need to pass the ownership parameter down to vec_restore()