More efficient vec_group_loc() vector access

[DavisVaughan]

This PR further optimizes vec_group_loc(), which should directly result in performance improvements for group_by().

In Instruments I noticed that the biggest killer in vec_group_loc() is the double deference in a tight loop of INTEGER(VECTOR_ELT()). That can be replaced with an array of integer pointers that index directly into those integer vectors. That really helped.

Side note: It is insane how fast this is if you order the data ahead of time. As pointed out by @brodieG

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

set.seed(42)

g <- sample(1e6, 1e7, replace = TRUE)
tbl <- tibble(g = g)

o <- order(g)
ordered_tbl <- tbl[o,]

bench::mark(
  vec_group_loc(tbl),
  vec_group_loc(ordered_tbl),
  check = FALSE,
  iterations = 20
)

# Master
#> Warning: Some expressions had a GC in every iteration; so filtering is disabled.
#> # A tibble: 2 x 6
#>   expression                      min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr>                 <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 vec_group_loc(tbl)            2.86s    2.99s     0.333     163MB    0.416
#> 2 vec_group_loc(ordered_tbl) 277.95ms 304.13ms     3.18      163MB    3.33

# This PR
#> Warning: Some expressions had a GC in every iteration; so filtering is disabled.
#> # A tibble: 2 x 6
#>   expression                      min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr>                 <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 vec_group_loc(tbl)            1.63s    1.71s     0.578     171MB    0.722
#> 2 vec_group_loc(ordered_tbl) 258.03ms 276.63ms     3.46      171MB    3.63

[brodieG]

Awesome! Obviously the ratios are not quite as good once you include the ordering time, but at least for this 1e7/1e6 case it seems to be worth it.

If you do go ahead and end up using the order thing, please considering giving a nod to the data table folk that contributed the radix sort to R. This trick is out of the question without that.

Also, as is probably obvious to you already, ordering might allow simplifying the group computation logic as well (maybe, I haven't looked at how you do it).

Thanks for keeping me in the loop, appreciate it.

[lionel-]

Cool improvement.