[RFC] [VOTE] Constant scalar expressions

Hi Tjerk,

On Fri, Nov 22, 2013 at 1:34 PM, Tjerk Meesters tjerk.meesters@gmail.comwrote:

Yes, of course that's all possible, because define() is just a function
and therefore the normal assignment rules apply. However, you didn't
mention the biggest drawback to such an approach, which is that it creates
global constants; there's no way to namespace them.

What was the reason why we don't have name space for define()?

Sorry for the confusion; of course it's possible to supply a fully
qualified name to define(), thus making the constant appear in a namespace.

With namespacing I also meant the ability to "attach" constants to class
definitions, which isn't possible with define().

I prefer 'const' to be a static for better performance, since
we have dynamic constant by define() already.

Perhaps we'd have to consider how much performance is impacted on
existing code, e.g.

class A
 {
    const B = 123;
}

How much slower would that become if this RFC is accepted, cached or
otherwise?

I made a script to experiment. I have background job running on this
machine,
so I don't get stable result now, but embedding value is constantly faster.

[yohgaki@dev tmp]$ php bench.php
Time: 6.4675929546356
Time: 7.4527719020844 15.232544075653% slower

That's the time for a define()'ed constant.

Time: 7.4178550243378 14.69266968975% slower

That's the time for a const'ed constant. Still seems faster.

Time: 8.1120491027832 25.426092205894% slower

This is probably due to zend_fetch_class I assume?

Time: 7.4948840141296 15.883669035753% slower

I can't really explain that :)

[yohgaki@dev tmp]$ php bench.php
Time: 6.4104981422424
Time: 7.7053151130676 20.198383060014% slower
Time: 8.022579908371 25.14752723358% slower
Time: 7.5032601356506 17.046444272519% slower
Time: 7.713919878006 20.332612331241% slower

This is way too much variation to be reliable.

[yohgaki@dev tmp]$ cat bench.php
<?php
define('a', 1);
const b = 1;

class foo {
const c = 1;
}

class bar extends foo {};

$start = microtime(true);
for ($i=0 ; $i < 99999999; $i++) {
$v = 1 + 1;
}
$test0 = microtime(true) - $start;
echo 'Time: '. $test0 . PHP_EOL;

$start = microtime(true);
for ($i=0 ; $i < 99999999; $i++) {
$v = 1 + a;
}
$test = microtime(true) - $start;
echo 'Time: '. $test . ' '. (($test - $test0)/$test0)*100 . '% slower' .
PHP_EOL;

$start = microtime(true);
for ($i=0 ; $i < 99999999; $i++) {
$v = 1 + b;
}
$test = microtime(true) - $start;
echo 'Time: '. $test . ' '. (($test - $test0)/$test0)*100 . '% slower' .
PHP_EOL;

$start = microtime(true);
for ($i=0 ; $i < 99999999; $i++) {
$v = 1 + foo::c;
}
$test = microtime(true) - $start;
echo 'Time: '. $test . ' '. (($test - $test0)/$test0)*100 . '% slower' .
PHP_EOL;

$start = microtime(true);
for ($i=0 ; $i < 99999999; $i++) {
$v = 1 + bar::c;
}
$test = microtime(true) - $start;
echo 'Time: '. $test . ' '. (($test - $test0)/$test0)*100 . '% slower' .
PHP_EOL;
?>

Actually, the test should be done against two separate builds:

1. build without patch
2. build with patch

:)

--
Yasuo Ohgaki
yohgaki@ohgaki.net

--

Tjerk