Optimize performances of `str_pad()` #19272

alexandre-daubois · 2025-07-28T08:57:39Z

I would like to propose this optimization for str_pad(). Here is the benchmark code:

<?php

$iterations = 1000000;

for ($i = 0; $i < $iterations; $i++) {
    str_pad('hello', 2000, ' ', STR_PAD_RIGHT);
}

And the results with right padding:

Benchmark 1: ./sapi/cli/php.branch validation_benchmark.php
  Time (mean ± σ):      44.6 ms ±   0.8 ms    [User: 41.3 ms, System: 2.3 ms]
  Range (min … max):    43.2 ms …  47.4 ms    64 runs
 
Benchmark 2: ./sapi/cli/php.master validation_benchmark.php
  Time (mean ± σ):      2.977 s ±  0.076 s    [User: 2.962 s, System: 0.009 s]
  Range (min … max):    2.863 s …  3.061 s    10 runs
 
Summary
  ./sapi/cli/php.branch validation_benchmark.php ran
   66.75 ± 2.06 times faster than ./sapi/cli/php.master validation_benchmark.php

Left padding results:

alex@alex-macos php-src % hyperfine './sapi/cli/php.branch validation_benchmark.php' './sapi/cli/php.master validation_benchmark.php' --warmup 10
Benchmark 1: ./sapi/cli/php.branch validation_benchmark.php
  Time (mean ± σ):      43.8 ms ±   1.5 ms    [User: 40.5 ms, System: 2.3 ms]
  Range (min … max):    42.2 ms …  49.8 ms    58 runs
 
Benchmark 2: ./sapi/cli/php.master validation_benchmark.php
  Time (mean ± σ):      1.150 s ±  0.020 s    [User: 1.130 s, System: 0.009 s]
  Range (min … max):    1.125 s …  1.174 s    10 runs
 
Summary
  ./sapi/cli/php.branch validation_benchmark.php ran
   26.23 ± 1.01 times faster than ./sapi/cli/php.master validation_benchmark.php

The idea is to avoid modulo operation in loops and copying char by char. Instead, this PR prefers the bulk copy approach.

SakiTakamachi · 2025-07-28T12:10:59Z

ext/standard/string.c

@@ -5737,6 +5737,27 @@ PHP_FUNCTION(substr_count)
 }
 /* }}} */

+static inline void php_str_pad_fill(char *dest, size_t pad_chars, const char *pad_str, size_t pad_str_len) {


Suggested change

static inline void php_str_pad_fill(char *dest, size_t pad_chars, const char *pad_str, size_t pad_str_len) {

static zend_always_inline void php_str_pad_fill(char *dest, size_t pad_chars, const char *pad_str, size_t pad_str_len) {

Using zend_always_inline increases the likelihood that the function will actually be inlined.

Also, instead of passing char *dest and size_t pad_chars as separate arguments, how about accepting a zend_string and extracting the values within this inline function?

Good idea! PR updated with both suggestions

ext/standard/string.c

TimWolla

I feel that this makes the API of php_str_pad_fill() a little safer to use, because it makes less assumptions about the target pointer.

With regard to the inlining: I generally trust the compiler to make better than decisions than I can do myself. I would make the function just static void without any inlining hints and let the compiler decide.

ext/standard/string.c

TimWolla · 2025-07-28T13:19:46Z

ext/standard/string.c

-	for (i = 0; i < left_pad; i++)
-		ZSTR_VAL(result)[ZSTR_LEN(result)++] = pad_str[i % pad_str_len];
+	if (left_pad > 0) {
+		php_str_pad_fill(result, left_pad, pad_str, pad_str_len);


Suggested change

php_str_pad_fill(result, left_pad, pad_str, pad_str_len);

php_str_pad_fill(ZSTR_VAL(result) + ZSTR_LEN(result), left_pad, pad_str, pad_str_len);

@SakiTakamachi advised otherwise if I get it right?

It appears so, but the suggestion doesn't really make sense, since pad_chars cannot be piggy-backed on the zend_string. Perhaps Saki meant to pass pad_str and pad_str_len as a zend_string* (which would make sense to me).

It makes the code a bit slower, going from ~48ms per run to ~58ms. Do we still want this?

I think it’s fine to go with the faster one! :)

TimWolla · 2025-07-28T13:24:51Z

Please also see #19220 (comment) and #19220 (comment) for benchmarking advice.

alexandre-daubois · 2025-07-28T13:30:02Z

Thanks for the link, I'm having a look to hyperfine for this PR and #19276

alexandre-daubois · 2025-07-28T13:47:20Z

PR description updated to use hyperfine

TimWolla · 2025-07-29T09:29:00Z

PR description updated to use hyperfine

The php binary appears to be the system binary, it is not unlikely that it was compiled with different flags or a different compiler (version). You should make sure to test two binaries compiled with identical flags and compilers.

alexandre-daubois · 2025-07-29T09:43:25Z

I think that my system-wide install right now is PHP 8.5 nightly. Just to be sure, I rerun against master and the branch:

Benchmark 1: ./sapi/cli/php.branch validation_benchmark.php
  Time (mean ± σ):      44.6 ms ±   0.8 ms    [User: 41.3 ms, System: 2.3 ms]
  Range (min … max):    43.2 ms …  47.4 ms    64 runs
 
Benchmark 2: ./sapi/cli/php.master validation_benchmark.php
  Time (mean ± σ):      2.977 s ±  0.076 s    [User: 2.962 s, System: 0.009 s]
  Range (min … max):    2.863 s …  3.061 s    10 runs
 
Summary
  ./sapi/cli/php.branch validation_benchmark.php ran
   66.75 ± 2.06 times faster than ./sapi/cli/php.master validation_benchmark.php

github-actions · 2025-08-01T08:47:05Z

AWS x86_64 (c7i.24xl)

Attribute	Value
Environment	aws
Runner	host
Instance type	c7i.metal-24xl (dedicated)
Architecture	x86_64
CPU	48 cores
CPU settings	disabled deeper C-states, disabled turbo boost, disabled hyper-threading
RAM	188 GB
Kernel	6.1.144-170.251.amzn2023.x86_64
OS	Amazon Linux 2023.8.20250721
GCC	11.5.0
Time	2025-08-01 08:20:50 UTC

Laravel 11.1.2 demo app - 30 consecutive runs, 100 requests (sec)

PHP	Min	Max	Std dev	Average	Average diff %	Median	Median diff %	Memory
PHP - baseline@eaf2	0.46107	0.46896	0.00123	0.46674	0.00%	0.46677	0.00%	43.54 MB
PHP - str-pad-opt	0.46154	0.46938	0.00127	0.46765	0.20%	0.46767	0.19%	43.46 MB

Symfony 2.6.0 demo app - 30 consecutive runs, 100 requests (sec)

PHP	Min	Max	Std dev	Average	Average diff %	Median	Median diff %	Memory
PHP - baseline@eaf2	0.72773	0.73390	0.00138	0.72954	0.00%	0.72921	0.00%	39.69 MB
PHP - str-pad-opt	0.73090	0.73512	0.00087	0.73262	0.42%	0.73241	0.44%	39.77 MB

Wordpress 6.2 main page - 30 consecutive runs, 20 requests (sec)

PHP	Min	Max	Std dev	Average	Average diff %	Median	Median diff %	Memory
PHP - baseline@eaf2	0.57803	0.58039	0.00059	0.57886	0.00%	0.57869	0.00%	43.54 MB
PHP - str-pad-opt	0.58101	0.65067	0.02497	0.63512	9.72%	0.64595	11.62%	43.54 MB

bench.php - 25 consecutive runs (sec)

PHP	Min	Max	Std dev	Average	Average diff %	Median	Median diff %	Memory
PHP - baseline@eaf2	0.21721	0.22099	0.00093	0.21825	0.00%	0.21808	0.00%	26.53 MB
PHP - str-pad-opt	0.21425	0.21712	0.00077	0.21580	-1.12%	0.21585	-1.02%	26.53 MB

micro_bench.php - 25 consecutive runs (sec)

PHP	Min	Max	Std dev	Average	Average diff %	Median	Median diff %	Memory
PHP - baseline@eaf2	1.26166	1.28426	0.00498	1.26842	0.00%	1.26635	0.00%	20.82 MB
PHP - str-pad-opt	1.31617	1.33954	0.00573	1.32695	4.61%	1.32688	4.78%	20.82 MB

kocsismate · 2025-08-01T09:04:27Z

The above comment was due to test the real time benchmark :) Unfortunately, I guess most results should show zero difference, which is not currently the case :( The wordpress results are especially weird. I'll have to dig into it...

cc. @iluuu1994 @arnaud-lb

alexandre-daubois · 2025-08-01T09:54:37Z

If I can be of any help, don't hesitate to tell me! It is indeed surprising

github-actions · 2025-08-01T10:10:56Z

AWS x86_64 (c7i.24xl)

Attribute	Value
Environment	aws
Runner	host
Instance type	c7i.metal-24xl (dedicated)
Architecture	x86_64
CPU	48 cores
CPU settings	disabled deeper C-states, disabled turbo boost, disabled hyper-threading
RAM	188 GB
Kernel	6.1.144-170.251.amzn2023.x86_64
OS	Amazon Linux 2023.8.20250721
GCC	11.5.0
Time	2025-08-01 09:41:35 UTC

Laravel 11.1.2 demo app - 30 consecutive runs, 100 requests (sec)

PHP	Min	Max	Std dev	Average	Average diff %	Median	Median diff %	Instr count	Memory
PHP - baseline@eaf2	0.45780	0.46629	0.00139	0.46472	0.00%	0.46488	0.00%	177247194	43.48 MB
PHP - str-pad-opt	0.46514	0.46736	0.00050	0.46594	0.26%	0.46581	0.20%	177251261	43.48 MB

Symfony 2.6.0 demo app - 30 consecutive runs, 100 requests (sec)

PHP	Min	Max	Std dev	Average	Average diff %	Median	Median diff %	Instr count	Memory
PHP - baseline@eaf2	0.72645	0.72932	0.00071	0.72785	0.00%	0.72779	0.00%	288207116	39.78 MB
PHP - str-pad-opt	0.72983	0.73390	0.00103	0.73178	0.54%	0.73152	0.51%	288197741	39.78 MB

Wordpress 6.2 main page - 30 consecutive runs, 20 requests (sec)

PHP	Min	Max	Std dev	Average	Average diff %	Median	Median diff %	Instr count	Memory
PHP - baseline@eaf2	0.57872	0.58180	0.00064	0.57969	0.00%	0.57960	0.00%	1129301005	43.56 MB
PHP - str-pad-opt	0.57772	0.58038	0.00056	0.57900	-0.12%	0.57888	-0.12%	1129302558	43.56 MB

bench.php - 25 consecutive runs (sec)

PHP	Min	Max	Std dev	Average	Average diff %	Median	Median diff %	Instr count	Memory
PHP - baseline@eaf2	0.21719	0.22107	0.00105	0.21863	0.00%	0.21866	0.00%	1791734911	26.55 MB
PHP - str-pad-opt	0.21282	0.21641	0.00090	0.21439	-1.94%	0.21441	-1.94%	1794186569	26.55 MB

micro_bench.php - 25 consecutive runs (sec)

PHP	Min	Max	Std dev	Average	Average diff %	Median	Median diff %	Instr count	Memory
PHP - baseline@eaf2	1.25910	1.27670	0.00515	1.26772	0.00%	1.26679	0.00%	1269754366	20.84 MB
PHP - str-pad-opt	1.31593	1.33956	0.00516	1.32448	4.48%	1.32418	4.53%	1271643119	20.83 MB

iluuu1994 · 2025-08-05T13:09:50Z

@kocsismate FYI, I've written a small benchmarking helper for myself.

https://github.com/iluuu1994/php-benchmark-diff

It differs slightly from things like hyperfine in a few ways:

It performs alternating tests to minimize noise from throttling. In other words, rather than running A 50 times and B 50 times, it runs A then B 50 times.
It allows using perf events rather than just wall time. It also has a "cgi" mode (should be renamed) that can measure sections of an application by using hrtime() and then printing the time of the section to stderr.
It looks at the fastest 25% of runs, i.e. filtering out the slow statistical outliers. I'm not sure yet if this is effective. This window is configurable.
It calculates the statistical p-value, i.e. how likely it is for the result to be random.

Not sure if this is useful to you, but wanted to mention it in case it is. It doesn't support JSON output yet.

kocsismate · 2025-08-05T13:28:11Z

@iluuu1994 Nice thank you very much for the input! I've seen Tim's comment somewhere about hyperfine (and read a blog post about it by Volker), but didn't know how it exactly differs from what I do, but based on your description it looks very promising! Specifically, I'm pretty much interested about the p-value. I think my benchmark also discards some of the results, but I should look into it how I exactly implemented it. Alternating tests is also a very clever idea.

I'll try to implement it at some point, because I'm still pretty much disappointed by my results :( Most tests used to be stable a few months ago, but they stopped being reliable a while ago. And curiously, microbenchmarks perform the worst.

iluuu1994 · 2025-08-05T13:37:07Z

I see I kept the branch with the JSON output too. iluuu1994/php-benchmark-diff@master...json I planned to replace the Valgrind benchmark, but unfortunately collecting CPU cycles with perf doesn't work on GH runners.

alexandre-daubois requested a review from bukka as a code owner July 28, 2025 08:57

github-actions bot added the Extension: standard label Jul 28, 2025

SakiTakamachi reviewed Jul 28, 2025

View reviewed changes

alexandre-daubois force-pushed the str-pad-opt branch from fbfa654 to 24d2823 Compare July 28, 2025 12:24

TimWolla reviewed Jul 28, 2025

View reviewed changes

ext/standard/string.c Outdated Show resolved Hide resolved

alexandre-daubois force-pushed the str-pad-opt branch from 24d2823 to ba1eb4c Compare July 28, 2025 13:12

Girgias reviewed Jul 28, 2025

View reviewed changes

ext/standard/string.c Outdated Show resolved Hide resolved

TimWolla reviewed Jul 28, 2025

View reviewed changes

Optimize performances of str_pad()

78dd005

alexandre-daubois force-pushed the str-pad-opt branch from ba1eb4c to 78dd005 Compare July 28, 2025 13:46

	static inline void php_str_pad_fill(char dest, size_t pad_chars, const char pad_str, size_t pad_str_len) {
	static zend_always_inline void php_str_pad_fill(char dest, size_t pad_chars, const char pad_str, size_t pad_str_len) {

	php_str_pad_fill(result, left_pad, pad_str, pad_str_len);
	php_str_pad_fill(ZSTR_VAL(result) + ZSTR_LEN(result), left_pad, pad_str, pad_str_len);

Optimize performances of str_pad() #19272

Are you sure you want to change the base?

Optimize performances of str_pad() #19272

Conversation

alexandre-daubois commented Jul 28, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

SakiTakamachi Jul 28, 2025

Choose a reason for hiding this comment

Uh oh!

alexandre-daubois Jul 28, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

TimWolla left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

TimWolla Jul 28, 2025

Choose a reason for hiding this comment

Uh oh!

alexandre-daubois Jul 28, 2025

Choose a reason for hiding this comment

Uh oh!

TimWolla Jul 29, 2025

Choose a reason for hiding this comment

Uh oh!

alexandre-daubois Jul 29, 2025

Choose a reason for hiding this comment

Uh oh!

SakiTakamachi Aug 5, 2025

Choose a reason for hiding this comment

Uh oh!

TimWolla commented Jul 28, 2025

Uh oh!

alexandre-daubois commented Jul 28, 2025

Uh oh!

alexandre-daubois commented Jul 28, 2025

Uh oh!

TimWolla commented Jul 29, 2025

Uh oh!

alexandre-daubois commented Jul 29, 2025

Uh oh!

github-actions bot commented Aug 1, 2025

AWS x86_64 (c7i.24xl)

Laravel 11.1.2 demo app - 30 consecutive runs, 100 requests (sec)

Symfony 2.6.0 demo app - 30 consecutive runs, 100 requests (sec)

Wordpress 6.2 main page - 30 consecutive runs, 20 requests (sec)

bench.php - 25 consecutive runs (sec)

micro_bench.php - 25 consecutive runs (sec)

Uh oh!

kocsismate commented Aug 1, 2025

Uh oh!

alexandre-daubois commented Aug 1, 2025

Uh oh!

github-actions bot commented Aug 1, 2025

AWS x86_64 (c7i.24xl)

Laravel 11.1.2 demo app - 30 consecutive runs, 100 requests (sec)

Symfony 2.6.0 demo app - 30 consecutive runs, 100 requests (sec)

Wordpress 6.2 main page - 30 consecutive runs, 20 requests (sec)

bench.php - 25 consecutive runs (sec)

micro_bench.php - 25 consecutive runs (sec)

Uh oh!

iluuu1994 commented Aug 5, 2025

Uh oh!

kocsismate commented Aug 5, 2025

Uh oh!

iluuu1994 commented Aug 5, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Optimize performances of `str_pad()` #19272

Optimize performances of `str_pad()` #19272

alexandre-daubois commented Jul 28, 2025 •

edited

Loading

iluuu1994 commented Aug 5, 2025 •

edited

Loading