[adapik]

In 107ae86 specialized comparisons were added to in_array() for integer and string needles, to avoid repeating the type checks on every iteration. That covered loose mode only. Strict mode kept going through the generic identity check for every element.

Later @nikic noticed the gap for integers: it doesn't make sense that strict=false is much faster than strict just for lack of a specialized codepath, and added an integer loop for strict mode (ea86a92). Strings were not touched, so strict mode with a string needle still calls zend_is_identical() for every element.

The result is that the same gap is still there for strings. On master in_array($str, $list, true) is about 2x slower than the loose version when the needle isn't found (1.7x in instructions, 2.0x in time).

This PR adds the missing loop, next to the integer one. After it, strict and loose are on par for string needles (about 21 vs 23 instructions per element).

Instructions per call, array of N strings (callgrind, release build):

• N=5: -15% if found, -27% if not found
• N=50: -23% / -45%
• N=500: -23% / -48%
• N=5000: -23% / -49%

At N=500 the time goes from 95 to 62 ms (found) and from 79 to 40 ms (not found). Other needle types are unchanged, except null, which costs 3 more instructions per call.

[ndossche]

Callgrind measures the instructions as if it operates as an in-order CPU, and treats all instructions as equally expensive.
Please share your benchmark script and test the benchmark using hyperfine.
Code itself looks good but we need better evidence this is worth it

[adapik]

@ndossche Here are wall-clock numbers with hyperfine instead.

The main motivation: a strict check should be at least as fast as a loose one, and right now it's slower.

Benchmark script

Release build (./configure --disable-all --enable-cli), run with -n, 30 runs each, mean times in ms. Time includes process startup. The needle is never found and is as long as the last element, so each comparison has to look at the string contents.

bench.php:

<?php
[$_, $mode, $size] = $argv;
$size = (int)$size;

$list = [];
for ($i = 0; $i < $size; $i++) {
    $list[] = "item_{$i}_xxxxxxxx";
}
$needle = 'item_' . ($size - 1) . '_xxxxxxxY';

$calls = intdiv(20_000_000, $size);
$hits = 0;
if ($mode === 'strict') {
    for ($i = 0; $i < $calls; $i++) {
        if (in_array($needle, $list, true)) $hits++;
    }
} else {
    for ($i = 0; $i < $calls; $i++) {
        if (in_array($needle, $list)) $hits++;
    }
}
echo $hits, "\n";

1. master only, loose vs strict:

hyperfine -N --warmup 3 --runs 30 -L mode strict,loose -L size 10,100,1000 \
  "./php_master -n bench.php {mode} {size}"

array size	loose (ms)	strict (ms)	slowdown
10	55.7	64.7	1.16x
100	42.8	52.2	1.22x
1000	42.0	51.1	1.22x

2. strict only, master vs patched:

hyperfine -N --warmup 3 --runs 30 -L bin master,patched -L size 10,100,1000 \
  "./php_{bin} -n bench.php strict {size}"

array size	master (ms)	patched (ms)	speedup
10	65.2	48.0	1.36x
100	51.8	35.9	1.44x
1000	51.2	36.5	1.40x

Note: I updated the branch but only to resolve the conflict in NEWS, all the rest is still the same.

[ndossche]

Thanks, please add an entry to UPGRADING under "performance improvements"

[adapik]

UPGRADING entry added, thanks for the review!

[ndossche]

Merged,thanks!