Improve performance of in_array

[jorgsowa]

Added checks for packed array inside function _php_search_array that is used by in_array() and array_search().

Function in_array() is used frequently with lists, and this modification improves the performance.

For array_search() in all cases the diference is negligible.

+-----------+------------------+--------+--------------------------------------+
| arraySize | randomKeysSize   | strict | Result                               |
+-----------+------------------+--------+--------------------------------------+
| 5000      | 5000             | false  | master 1.00 ± 0.00× faster than new  |
| 5000      | 5000             | true   | master 1.00 ± 0.00× faster than new  |
| 5         | 5                | true   | master 1.00 ± 0.09× faster than new  |
| 5         | 5                | false  | master 1.00 ± 0.09× faster than new  |
+-----------+------------------+--------+--------------------------------------+

For in_array() bigger the array, faster the execution.

| arraySize | randomKeysSize | packed | valueType | strict | Result (new vs master)    
|-----------|----------------|--------|-----------|--------|----------------------------------
| 500,000   | 100            | yes    | int       | false  | 1.32 ± 0.01 × faster     
| 500,000   | 100            | yes    | int       | true   | 1.19 ± 0.02 × faster        
| 50,000    | 100            | yes    | mixed     | false  | 1.07 ± 0.16 × faster         
| 500,000   | 100            | yes    | mixed     | true   | 1.19 ± 0.01 × faster       
| 500,000   | 100            | yes    | string    | false  | 1.11 ± 0.01 × faster     
| 500,000   | 100            | yes    | string    | true   | 1.08 ± 0.01 × faster   
| 50,000    | 100            | no     | string    | false  | 1.01 ± 0.01 × faster            
| 50,000    | 100            | no     | string    | true   | 1.00 ± 0.01 × faster       
| 500       | 100            | yes    | int       | false  | 1.05 ± 0.12 x faster          
| 500       | 100            | yes    | int       | true   | 1.05 ± 0.06 x faster   
| 500       | 100            | yes    | mixed     | false  | 1.03 ± 0.01 x faster   
| 500       | 100            | yes    | mixed     | true   | 1.08 ± 0.30 x faster            
| 500       | 100            | yes    | string    | false  | 1.08 ± 0.15 x faster        
| 500       | 100            | yes    | string    | true   | 1.04 ± 0.03 x faster      
| 500       | 100            | no     | string    | false  | 1.01 ± 0.18 x faster           
| 500       | 100            | no     | string    | true   | 1.01 ± 0.14 x slower       
| 5         | 5              | yes    | mixed     | true   | 1.01 ± 0.23 x faster (10000 loops)
| 5         | 5              | no     | mixed     | true   | 1.01 ± 0.05 x faster (10000 loops)

Benchmark code

<?php

$arraySize = 100;
$randomKeysSize = 100;
$isPacked = true;
$valueType = 0; // 0 - only strings, 1 - only integers, 2 - string/int 50/50
$strict = false;
$in_array = true;

$array = [];
for ($i = 0; $i < $arraySize; $i++) {
    if($isPacked) {
        $array[] = ($valueType === 0 || $i % $valueType == 1) ? "String $i" : $i;
    } else {
        $array["key_$i"] = "String $i";
    }
}

mt_srand(1234);
$randomKeys = array_rand($array, $randomKeysSize);
$values = array_map(fn($key) => $array[$key], $randomKeys);

for ($i = 0; $i < 100; $i++) {
    foreach ($values as $value) {
        if($in_array) {
            in_array($value, $array, $strict);
        } else {
            array_search($value, $array, $strict);
        }
    }
}

Command I used for comparison:
hyperfine --warmup 3 "sapi/cli/php_master test_in_array.php" "sapi/cli/php test_in_array.php"

Must be merged after merging the extended tests:

• test: improve the tests for in_array #21087

[github-actions]

AWS x86_64 (c6id.metal)

Attribute	Value
Environment	aws
Instance type	c6id.metal
Architecture	x86_64
CPU	Intel(R) Xeon(R) Platinum 8375C CPU @ 2.90GHz, 64 cores @ 2900 MHz
CPU settings	disabled deeper C-states, disabled turbo boost, disabled hyper-threading
RAM	251 GB
Kernel	6.12.66-88.122.amzn2023.x86_64
OS	Amazon Linux 2023.10.20260202
GCC	14.2.1
Time	2026-02-24 18:49:07 UTC
Job details	https://github.com/php/php-src/actions/runs/22365239226 (Artifacts)
Changeset	https://github.com/php/php-src/compare/7158e6376e..c9c025d489

Laravel 12.11.0 demo app - 50 iterations, 50 warmups, 100 requests (sec)

PHP	Min	Max	Std dev	Rel std dev %	Mean	Mean diff %	Median	Median diff %	Skewness	Z-stat	P-value	Memory
PHP - baseline@7158e63	0.38896	0.39196	0.00052	0.13%	0.38951	0.00%	0.38939	0.00%	3.203	0.000	1.000	25.26 MB
PHP - perf/in_array	0.38876	0.39009	0.00035	0.09%	0.38923	-0.07%	0.38914	-0.06%	0.841	3.657	0.000	25.25 MB

Symfony 2.8.0 demo app - 50 iterations, 50 warmups, 100 requests (sec)

PHP	Min	Max	Std dev	Rel std dev %	Mean	Mean diff %	Median	Median diff %	Skewness	Z-stat	P-value	Memory
PHP - baseline@7158e63	0.65851	0.66081	0.00046	0.07%	0.65928	0.00%	0.65921	0.00%	1.574	0.000	1.000	25.30 MB
PHP - perf/in_array	0.66310	0.66701	0.00074	0.11%	0.66384	0.69%	0.66365	0.67%	2.524	-8.614	0.000	25.30 MB

Wordpress 6.9 main page - 50 iterations, 20 warmups, 20 requests (sec)

PHP	Min	Max	Std dev	Rel std dev %	Mean	Mean diff %	Median	Median diff %	Skewness	Z-stat	P-value	Memory
PHP - baseline@7158e63	0.58594	0.58994	0.00069	0.12%	0.58670	0.00%	0.58652	0.00%	3.112	0.000	1.000	25.39 MB
PHP - perf/in_array	0.58365	0.59109	0.00125	0.21%	0.58444	-0.39%	0.58406	-0.42%	3.764	7.263	0.000	25.34 MB

bench.php - 50 iterations, 20 warmups, 2 requests (sec)

PHP	Min	Max	Std dev	Rel std dev %	Mean	Mean diff %	Median	Median diff %	Skewness	Z-stat	P-value	Memory
PHP - baseline@7158e63	0.44446	0.44655	0.00050	0.11%	0.44551	0.00%	0.44548	0.00%	0.050	0.000	1.000	25.39 MB
PHP - perf/in_array	0.44322	0.44613	0.00068	0.15%	0.44494	-0.13%	0.44487	-0.14%	-0.284	4.147	0.000	25.34 MB

[github-actions]

AWS x86_64 (c6id.metal)

Attribute	Value
Environment	aws
Instance type	c6id.metal
Architecture	x86_64
CPU	Intel(R) Xeon(R) Platinum 8375C CPU @ 2.90GHz, 64 cores @ 2900 MHz
CPU settings	disabled deeper C-states, disabled turbo boost, disabled hyper-threading
RAM	251 GB
Kernel	6.12.66-88.122.amzn2023.x86_64
OS	Amazon Linux 2023.10.20260202
GCC	14.2.1
Time	2026-02-24 22:34:32 UTC
Job details	https://github.com/php/php-src/actions/runs/22372998923 (Artifacts)
Changeset	https://github.com/php/php-src/compare/7158e6376e..c9c025d489

Laravel 12.11.0 demo app - 50 iterations, 50 warmups, 100 requests (sec)

PHP	Min	Max	Std dev	Rel std dev %	Mean	Mean diff %	Median	Median diff %	Skewness	Z-stat	P-value	Memory
PHP - baseline@7158e63	0.38922	0.39076	0.00034	0.09%	0.38964	0.00%	0.38954	0.00%	1.683	0.000	1.000	25.27 MB
PHP - perf/in_array	0.38890	0.39144	0.00051	0.13%	0.38962	-0.00%	0.38958	0.01%	1.263	0.596	0.551	25.26 MB

Symfony 2.8.0 demo app - 50 iterations, 50 warmups, 100 requests (sec)

PHP	Min	Max	Std dev	Rel std dev %	Mean	Mean diff %	Median	Median diff %	Skewness	Z-stat	P-value	Memory
PHP - baseline@7158e63	0.65959	0.66443	0.00074	0.11%	0.66027	0.00%	0.66011	0.00%	3.932	0.000	1.000	25.32 MB
PHP - perf/in_array	0.66380	0.66812	0.00075	0.11%	0.66462	0.66%	0.66441	0.65%	2.880	-8.428	0.000	25.32 MB

Wordpress 6.9 main page - 50 iterations, 20 warmups, 20 requests (sec)

PHP	Min	Max	Std dev	Rel std dev %	Mean	Mean diff %	Median	Median diff %	Skewness	Z-stat	P-value	Memory
PHP - baseline@7158e63	0.58585	0.58895	0.00066	0.11%	0.58650	0.00%	0.58635	0.00%	2.776	0.000	1.000	25.39 MB
PHP - perf/in_array	0.58356	0.58692	0.00080	0.14%	0.58419	-0.39%	0.58393	-0.41%	2.552	7.483	0.000	25.33 MB

bench.php - 50 iterations, 20 warmups, 2 requests (sec)

PHP	Min	Max	Std dev	Rel std dev %	Mean	Mean diff %	Median	Median diff %	Skewness	Z-stat	P-value	Memory
PHP - baseline@7158e63	0.44437	0.44702	0.00054	0.12%	0.44570	0.00%	0.44571	0.00%	0.095	0.000	1.000	25.39 MB
PHP - perf/in_array	0.44305	0.44656	0.00085	0.19%	0.44496	-0.17%	0.44505	-0.15%	-0.342	4.581	0.000	25.33 MB

[iluuu1994]

Hmm, the two benchmark runs do seem to indicate a significant slowdown in Symfony. Could be overspecialization. Hard to say whether the trade-off is worth it. I'll try benchmarking locally tomorrow to verify.

[Girgias]

I think at this point it just makes more sense to split _php_search_array into two C functions, one for in_array and the other for array_search.

[iluuu1994]

This is inlined (or should be at least, should be zend_always_inline), so not sure if there's a point to splitting, unless the duplication can be reduced.

[github-actions]

AWS x86_64 (c6id.metal)

Attribute	Value
Environment	aws
Instance type	c6id.metal
Architecture	x86_64
CPU	Intel(R) Xeon(R) Platinum 8375C CPU @ 2.90GHz, 64 cores @ 2900 MHz
CPU settings	disabled deeper C-states, disabled turbo boost, disabled hyper-threading
RAM	251 GB
Kernel	6.12.66-88.122.amzn2023.x86_64
OS	Amazon Linux 2023.10.20260202
GCC	14.2.1
Time	2026-03-04 09:49:20 UTC
Job details	https://github.com/php/php-src/actions/runs/22663940877 (Artifacts)
Changeset	https://github.com/php/php-src/compare/7a1c2612c0..b107170a01

Laravel 12.11.0 demo app - 50 iterations, 50 warmups, 100 requests (sec)

PHP	Min	Max	Std dev	Rel std dev %	Mean	Mean diff %	Median	Median diff %	Skewness	Z-stat	P-value	Memory
PHP - baseline@7a1c261	0.38828	0.38962	0.00020	0.05%	0.38873	0.00%	0.38869	0.00%	1.719	0.000	1.000	25.24 MB
PHP - perf/in_array	0.38870	0.39072	0.00033	0.08%	0.38923	0.13%	0.38919	0.13%	2.027	-7.649	0.000	25.18 MB

Symfony 2.8.0 demo app - 50 iterations, 50 warmups, 100 requests (sec)

PHP	Min	Max	Std dev	Rel std dev %	Mean	Mean diff %	Median	Median diff %	Skewness	Z-stat	P-value	Memory
PHP - baseline@7a1c261	0.66326	0.66598	0.00046	0.07%	0.66382	0.00%	0.66372	0.00%	2.379	0.000	1.000	25.21 MB
PHP - perf/in_array	0.66389	0.67015	0.00086	0.13%	0.66448	0.10%	0.66432	0.09%	6.010	-6.866	0.000	25.28 MB

Wordpress 6.9 main page - 50 iterations, 20 warmups, 20 requests (sec)

PHP	Min	Max	Std dev	Rel std dev %	Mean	Mean diff %	Median	Median diff %	Skewness	Z-stat	P-value	Memory
PHP - baseline@7a1c261	0.58552	0.58889	0.00085	0.15%	0.58641	0.00%	0.58611	0.00%	1.762	0.000	1.000	25.23 MB
PHP - perf/in_array	0.58446	0.59139	0.00114	0.19%	0.58521	-0.20%	0.58492	-0.20%	3.924	7.097	0.000	25.30 MB

bench.php - 50 iterations, 20 warmups, 2 requests (sec)

PHP	Min	Max	Std dev	Rel std dev %	Mean	Mean diff %	Median	Median diff %	Skewness	Z-stat	P-value	Memory
PHP - baseline@7a1c261	0.44448	0.44661	0.00045	0.10%	0.44578	0.00%	0.44585	0.00%	-0.697	0.000	1.000	25.23 MB
PHP - perf/in_array	0.44422	0.44691	0.00056	0.13%	0.44566	-0.03%	0.44576	-0.02%	-0.355	1.299	0.194	25.30 MB

[jorgsowa]

Hey @iluuu1994 Is there anything more I can regarding this PR? Should I measure something else?

[github-actions]

AWS x86_64 (c6id.metal)

Attribute	Value
Environment	aws
Instance type	c6id.metal
Architecture	x86_64
CPU	Intel(R) Xeon(R) Platinum 8375C CPU @ 2.90GHz, 64 cores @ 2900 MHz
CPU settings	disabled deeper C-states, disabled turbo boost, disabled hyper-threading
RAM	251 GB
Kernel	6.18.38-76.139.amzn2023.x86_64
OS	Amazon Linux 2023.12.20260727
GCC	14.2.1
Binary layout strategy	bolt-align
Time	2026-09-23 21:38:23 UTC
Job details	https://github.com/php/php-src/actions/runs/35923641336 (Artifacts)
Changeset	https://github.com/php/php-src/compare/2e20074310..0dea8bf816

Laravel 12.11.0 demo app - 50 iterations, 50 warmups, 100 requests (sec)

PHP	Min	Max	Std dev	Rel std dev %	Mean	Mean diff %	Median	Median diff %	Skewness	Z-stat	P-value	Memory
PHP - baseline@2e20074	0.39754	0.40040	0.00045	0.11%	0.39799	0.00%	0.39789	0.00%	3.626	0.000	1.000	30.82 MB
PHP - perf/in_array	0.39701	0.40072	0.00054	0.14%	0.39743	-0.14%	0.39733	-0.14%	4.902	7.228	0.000	30.80 MB

Symfony 2.8.0 demo app - 50 iterations, 50 warmups, 100 requests (sec)

PHP	Min	Max	Std dev	Rel std dev %	Mean	Mean diff %	Median	Median diff %	Skewness	Z-stat	P-value	Memory
PHP - baseline@2e20074	0.68609	0.68936	0.00056	0.08%	0.68688	0.00%	0.68676	0.00%	2.486	0.000	1.000	30.75 MB
PHP - perf/in_array	0.68650	0.68818	0.00040	0.06%	0.68707	0.03%	0.68695	0.03%	1.022	-3.264	0.001	30.74 MB

Wordpress 6.9 main page - 50 iterations, 20 warmups, 20 requests (sec)

PHP	Min	Max	Std dev	Rel std dev %	Mean	Mean diff %	Median	Median diff %	Skewness	Z-stat	P-value	Memory
PHP - baseline@2e20074	0.59549	0.59820	0.00055	0.09%	0.59610	0.00%	0.59597	0.00%	2.131	0.000	1.000	30.76 MB
PHP - perf/in_array	0.59421	0.59599	0.00040	0.07%	0.59483	-0.21%	0.59475	-0.21%	1.089	8.186	0.000	30.75 MB

bench.php - 50 iterations, 20 warmups, 2 requests (sec)

PHP	Min	Max	Std dev	Rel std dev %	Mean	Mean diff %	Median	Median diff %	Skewness	Z-stat	P-value	Memory
PHP - baseline@2e20074	0.45024	0.45215	0.00042	0.09%	0.45102	0.00%	0.45093	0.00%	0.783	0.000	1.000	30.76 MB
PHP - perf/in_array	0.45164	0.46682	0.00203	0.45%	0.45310	0.46%	0.45275	0.40%	6.531	-8.572	0.000	30.75 MB

micro_bench.php - 50 iterations, 5 warmups, 1 request (sec)

PHP	Min	Max	Std dev	Rel std dev %	Mean	Mean diff %	Median	Median diff %	Skewness	Z-stat	P-value	Memory
PHP - baseline@2e20074	1.29728	1.31153	0.00324	0.25%	1.30563	0.00%	1.30539	0.00%	-0.129	0.000	1.000	30.76 MB
PHP - perf/in_array	1.29738	1.31420	0.00353	0.27%	1.30581	0.01%	1.30587	0.04%	-0.008	-0.238	0.812	30.75 MB

[laruence]

I don't think this optimization is worth it.

The benefit is too small: array_search() gains nothing at all (1.00× in every case), and in_array() only shows a real gain at 500k elements. At 500 elements it's 1.05 ± 0.12× and at 5 it's 1.01 ± 0.23× — both within noise. Arrays with hundreds of thousands of packed elements aren't the common shape for in_array(), so the extreme case isn't very informative.

What I do care about is code size. I compiled ext/standard/array.c at -O2 with and without this change (identical trees otherwise) and the affected functions grow by ~2.4 KB — _php_search_array itself goes from 1484 to 3016 bytes on arm64, and the whole family from 3372 to 5884. Same on x86_64: 3348 to 5713.

The reason is that _php_search_array is outlined (both zif_in_array and zif_array_search tail-call into it), so behavior stays a runtime value and each of the six branches now instantiates three loops instead of one. More code in the hot path means more icache misses, and every caller pays that — including all array_search() users who get no benefit.

I think this is a net loss.