Rework vec_ptype() into a generic for S3 types

[DavisVaughan]

Demonstrated benefits of a static vec-ptype:

(It is a bit oversimplified for year-month-day, but the performance improvements remains valid)

It would be up to the developer to ensure that their vec-ptype method returns a "valid" ptype. I think this is a reasonable requirement, since we essentially require that their vec-ptype2 and vec-cast methods are also valid.

library(clock)
library(vctrs)

x <- year_month_day(2019, 1, 1)
xs <- rep(list(x), 1e5)

bench::mark(vec_ptype(x), iterations = 100000)
#> # 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 vec_ptype(x)   14.3µs   15.2µs    62544.    29.8KB     21.9

bench::mark(vec_ptype_common(!!!xs))
#> 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 vec_ptype_common(!!!xs)    7.79s    7.79s     0.128      32KB     24.5

empty_year_month_day <- year_month_day(integer(), integer(), integer())

vec_ptype.clock_year_month_day <- function(x, ..., x_arg = "") {
  empty_year_month_day
}

bench::mark(vec_ptype(x), iterations = 100000)
#> # 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 vec_ptype(x)   1.83µs    2.4µs   388969.        0B     27.2

bench::mark(vec_ptype_common(!!!xs))
#> 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 vec_ptype_common(!!!xs)    1.29s    1.29s     0.773        0B     27.8

^{Created on 2021-02-24 by the reprex package (v1.0.0)}

[DavisVaughan]

Possible benefits for integer64, which could be added in a followup PR

Compare against #1323 (comment)

• Nearly 3x faster for common type determination
• Which translates to roughly 40% increase for vec_c().
• Still much slower than native types, but better than it used to be

library(vctrs, warn.conflicts = FALSE)

vctrs_empty_integer64 <- NULL

vec_ptype.integer64 <- function(x, ...) {
  if (is.null(vctrs_empty_integer64)) {
    vctrs_empty_integer64 <<- bit64::integer64()
  }
  
  out <- vctrs_empty_integer64
  dim <- dim(x)
  
  if (is.null(dim)) {
    return(out)
  }
  
  if (length(dim) == 0L) {
    abort("Corrupt `dim` attribute. `dim` must have length >=1.")
  }
  
  dim[[1]] <- 0L
  dim(out) <- dim
  
  out
}

library(dplyr, warn.conflicts = FALSE)
suppressPackageStartupMessages(library(bit64))
#> Warning: package 'bit64' was built under R version 4.0.2
#> Warning: package 'bit' was built under R version 4.0.2

df <- tibble(x = 1, x64 = integer64(1), xstr = "1", y = sample(10000, 1000000, replace = T)) %>% group_by(y)

first_i64 <- group_map(df, ~.x$x64[1])
first_int <- group_map(df, ~.x$x[1])

type_common <- function(x) vec_ptype_common(!!!x)
vec_c2 <- function(x) vec_c(!!!x)

bench::mark(
  type_common(first_int),
  vec_c2(first_int),
  type_common(first_i64),
  vec_c2(first_i64),
  
  check = FALSE,
  iterations = 20
)
#> Warning: Some expressions had a GC in every iteration; so filtering is disabled.
#> # A tibble: 4 x 6
#>   expression                  min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr>             <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 type_common(first_int)  606.1µs 680.31µs  1402.       4.02KB      0  
#> 2 vec_c2(first_int)        1.88ms   2.27ms   458.     195.45KB      0  
#> 3 type_common(first_i64) 175.62ms 201.31ms     4.88    42.09KB     21.0
#> 4 vec_c2(first_i64)         1.19s    1.26s     0.795  764.06MB     12.2

[lionel-]

I think the main consequence of the "static prototype" argument is that names are no longer part of the prototype. This is something that matches my intuition and longer-term plans: #1020

For the patch release we could merge this to unblock Davis (performance of clock classes). Then at the next non-patch release we'll unname inputs before slicing in the default vec_ptype() method. Does that sound good @hadley?