Skip to content

Conversation

alexandre-daubois
Copy link
Member

@alexandre-daubois alexandre-daubois commented Jul 28, 2025

I think it's worth adding a 1 char fast path for implode(). A quick search on Github reveals the following stats:

  • Implode usage with one char glue: /implode\(\s*["'][^"']{1}["']\s*,\s*/ language:PHP reveals 1.4 million occurrences
  • Implode usage with 2+ char glue: /implode\(\s*["'][^"']{2,}["']\s*,\s*/ language:PHP reveals 1.1 million occurrences

Here is the benchmark code:

<?php

$iterations = 1000;
$array = array_fill(0, 10000, 'item');

for ($i = 0; $i < $iterations; $i++) {
    implode(',', $array);
}

And the results:

Benchmark 1: ./sapi/cli/php.branch bench.php
  Time (mean ± σ):      33.7 ms ±   1.2 ms    [User: 30.8 ms, System: 2.1 ms]
  Range (min … max):    32.7 ms …  41.4 ms    82 runs
 
Benchmark 2: ./sapi/cli/php.master bench.php
  Time (mean ± σ):      45.1 ms ±   6.7 ms    [User: 41.4 ms, System: 2.2 ms]
  Range (min … max):    42.9 ms …  97.5 ms    65 runs

Summary
  ./sapi/cli/php.branch bench.php ran
    1.34 ± 0.21 times faster than ./sapi/cli/php.master bench.php

No change for multi char glue, but a nice improved performance for single char glue especially one big arrays.

@nielsdos
Copy link
Member

1-char bench on an i7-4790:

Benchmark 1: ./sapi/cli/php x.php 
  Time (mean ± σ):      47.9 ms ±   0.7 ms    [User: 45.5 ms, System: 2.0 ms]
  Range (min … max):    46.5 ms …  49.3 ms    60 runs
 
Benchmark 2: ./sapi/cli/php_old x.php
  Time (mean ± σ):      69.8 ms ±   1.1 ms    [User: 67.7 ms, System: 1.8 ms]
  Range (min … max):    67.4 ms …  73.0 ms    42 runs
 
Summary
  ./sapi/cli/php x.php  ran
    1.46 ± 0.03 times faster than ./sapi/cli/php_old x.php

Curiously a large difference with your measurement, did you benchmark a release build?

2-char glue is slightly slower:

Benchmark 1: ./sapi/cli/php x.php 
  Time (mean ± σ):      72.5 ms ±   1.9 ms    [User: 70.2 ms, System: 1.9 ms]
  Range (min … max):    70.2 ms …  82.4 ms    40 runs
 
Benchmark 2: ./sapi/cli/php_old x.php
  Time (mean ± σ):      66.4 ms ±   0.8 ms    [User: 63.9 ms, System: 2.0 ms]
  Range (min … max):    64.5 ms …  68.9 ms    45 runs
 
Summary
  ./sapi/cli/php_old x.php ran
    1.09 ± 0.03 times faster than ./sapi/cli/php x.php 

@alexandre-daubois
Copy link
Member Author

alexandre-daubois commented Aug 10, 2025

I checked again and --enable-debug was used in configure. Removing it reveals results very close to yours. 2-char glue results on about the same perf on Apple M4:

Summary
  ./sapi/cli/php.branch bench.php ran
    1.01 ± 0.03 times faster than ./sapi/cli/php.master bench.php

Given the number of occurrences of 1 char glue as written in the description, I think this fast path would still worth it?

Description updated with the non-debug numbers for 1-char glue.

Copy link
Member

@nielsdos nielsdos left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ok for me

@alexandre-daubois alexandre-daubois merged commit 29c7ee4 into php:master Sep 10, 2025
9 checks passed
@alexandre-daubois alexandre-daubois deleted the implode-opt branch September 10, 2025 14:52

cptr -= ZSTR_LEN(glue);
memcpy(cptr, ZSTR_VAL(glue), ZSTR_LEN(glue));
if (ZSTR_LEN(glue) == 1) {
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is in loop, is ZSTR_LEN(glue) guaranteed to be evaluated only once?

What about moving the condition outside the loop, ie. "duplicating" the loop body?

Also what about the case when the glue is empty string? (we use that strongly in https://github.com/atk4/ui/blob/6.0.0/src/HtmlTemplate.php#L570 - with 10k calls per webpage request easily)

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is in loop, is ZSTR_LEN(glue) guaranteed to be evaluated only once?

No. But chances are that it doesn't matter if this is in a variable or not because you'll need a load from a spill slot in the variable case most likely.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants