wc: improve wc -c perf with unextended pipe input (Linux only)

[oech3]

$ truncate -s 32GB huge
$ time cat huge|wc -c
Executed in    1.31 secs    fish           external
   usr time    0.64 secs    0.53 millis    0.63 secs
   sys time    1.53 secs    2.95 millis    1.52 secs
$ time cat huge|target/release/wc -c
Executed in  388.34 millis    fish           external
   usr time   39.03 millis    0.64 millis   38.39 millis
   sys time  657.96 millis    2.98 millis  654.98 millis

Depends on #11551 for smaller diff.

[oech3]

Is splice() unsupported at CodSpeed?

[sylvestre]

please run your benchmark with hyperfine
time isn't great compared to hyperfine for this

[oech3]

using stdin at hyperfine just fails. I was using -N .

[sylvestre]

are you sure ?
hyperfine --warmup 1 'cat /tmp/huge | wc -c' 'cat /tmp/huge | target/release/wc -c' 'wc -c < /tmp/huge' 'target/release/wc -c < /tmp/huge'
worked for me

[oech3]

Also we cannot use our cat from main since #11551 has same effect for input.

We lost the utility to benchmark this at #11551.

[sylvestre]

and compared with GNU ? :)

[oech3]

I did serious typo at prev comment (missing -c)
(and we actuall should use cat from main)

$ truncate -s 1GB huge
$ hyperfine "cat huge |wc-main -c" "cat huge |wc-fcntl-input -c" "cat huge|/hom
e/a/bin/gnu58d88d243/wc -c"
Benchmark 1: cat huge |wc-main -c
  Time (mean ± σ):      29.5 ms ±   1.8 ms    [User: 11.2 ms, System: 40.7 ms]
  Range (min … max):    27.0 ms …  35.8 ms    98 runs
 
Benchmark 2: cat huge |wc-fcntl-input -c
  Time (mean ± σ):      17.7 ms ±   1.1 ms    [User: 2.3 ms, System: 28.7 ms]
  Range (min … max):    16.0 ms …  22.3 ms    143 runs
 
Benchmark 3: cat huge|/home/a/bin/gnu58d88d243/wc -c
  Time (mean ± σ):      72.8 ms ±   2.2 ms    [User: 5.3 ms, System: 136.6 ms]
  Range (min … max):    69.2 ms …  78.9 ms    40 runs
 
Summary
  cat huge |wc-fcntl-input -c ran
    1.67 ± 0.14 times faster than cat huge |wc-main -c
    4.12 ± 0.28 times faster than cat huge|/home/a/bin/gnu58d88d243/wc -c

GNU does not use splice() even at master branch yet.

[oech3]

I noticed that we don't need to increase pipe size of input fd since

• if sender supports splice(), it probably extended the pipe size of input fd.
• if sender uses write(), it is bottleneck for our splice wc -c.
  Exception is sender utility with splice() support which forgot to increase pipe size or sender is modest.

(this PR's splice()'s length is still meaningful)

I'll remove fcntl at #11636 .