Improve path_rel performance

[orgadish]

#519

Improved path_rel performance by:

1. Hoisting path_split(start) out of the loop entirely.
2. Performing path_split(path) in a single vectorized call outside the R loop.
3. Adding a new, internal path_common_parts that takes the components, rather than having to split again, and then passing the vectorized-split parts directly.
4. The internal path_common_parts uses vectorized calls rather than R loops.
5. Deduplicating path input to path_rel.

Performance Improvement:

• ~10x improvement on 5k vector with no duplication (1.77 -> 200ms)
• ~350x improvement on 5k vector with 50x duplication (2.09s -> 6ms)

Tests pass:

[ FAIL 0 | WARN 0 | SKIP 5 | PASS 280 ]

── Skipped tests (5) ────────────────────────────────────────────────────────────────────────────────────────────────────────
• On Windows (5): test-path.R:118:5, test-path.R:129:5, test-path.R:148:5, test-path.R:161:5, test-path.R:716:5

Benchmark

# Load dev path_rel and path_common
devtools::load_all("D:/Github/fs")
#> ℹ Loading fs

# Original path_rel and path_common
original_path_common <- function(path) {
  ...  # Copied from fs
}
original_path_rel <- function(path, start = ".") {
  ...  # Copied from fs, except using original_path_common()
}

all_3_letter_combos <- do.call(paste0, expand.grid(rep(list(LETTERS), 3)))

set.seed(3)
top_path <- "x/y/"

# Case 1: No duplication
paths <- paste0(top_path, sample(all_3_letter_combos, 5000), "/file.csv")
str(paths)
#>  chr [1:5000] "x/y/OHQ/file.csv" "x/y/JXJ/file.csv" "x/y/DDK/file.csv" ...

## path_rel()
microbenchmark::microbenchmark(
  new = path_rel(paths, start=top_path),
  original = original_path_rel(paths, start=top_path),
  times = 10L
)
#> Unit: milliseconds
#>      expr       min        lq      mean    median        uq      max neval cld
#>       new  181.1754  183.2254  256.2644  200.0344  344.2702  383.949    10  a 
#>  original 1723.4413 1757.3969 1920.8330 1769.0271 1786.6630 2693.200    10   b

## path_common()
microbenchmark::microbenchmark(
  new = path_common(paths),
  original = original_path_common(paths),
  times = 10L
)
#> Unit: milliseconds
#>      expr     min      lq     mean  median      uq     max neval cld
#>       new 71.0870 72.2124 79.92248 79.7240 85.2738 97.9682    10   a
#>  original 76.5306 77.0471 82.81926 83.1001 88.1458 89.9067    10   a


# Case 2: With duplication

# Create a list with significant duplication (50x duplication).
dup_paths <- sample(paths, length(paths) / 50) |> 
  rep(50) |> 
  sample()  # Reshuffle to ensure no sort-based benefits.
str(dup_paths)
#>  chr [1:5000] "x/y/EBC/file.csv" "x/y/ZUI/file.csv" "x/y/XSC/file.csv" ...

## path_rel()
microbenchmark::microbenchmark(
  new = path_rel(dup_paths, start=top_path),
  original = original_path_rel(dup_paths, start=top_path),
  deduped_og = deduped::deduped(original_path_rel)(dup_paths, start=top_path),
  deduped_new = deduped::deduped(path_rel)(dup_paths, start=top_path),
  times = 10L
)
#> Unit: milliseconds
#>         expr       min        lq       mean     median        uq       max neval cld
#>          new    4.9530    6.0521    6.29429    6.32245    6.3719    8.5303    10  a 
#>     original 1883.7256 1989.1657 2137.89882 2092.36985 2314.7633 2428.3711    10   b
#>   deduped_og   34.9242   39.0378   44.23877   40.64605   42.8788   68.3704    10  a 
#>  deduped_new    4.7585    5.6040   17.32869    5.80520    6.1606  119.5408    10  a 

## path_common()
microbenchmark::microbenchmark(
  new = path_common(dup_paths),
  original = original_path_common(dup_paths),
  times = 10L
)
#> Unit: milliseconds
#>      expr     min      lq     mean   median      uq     max neval cld
#>       new 72.1070 72.7383 76.19327 74.94015 79.1212 84.9619    10  a 
#>  original 73.5073 76.5788 87.99847 91.37255 97.9526 99.3352    10   b

^{Created on 2026-07-04 with reprex v2.1.1}