Performance optimization ideas for PHP 7.3

Hi Nicolas,

Don’t you think the biggest performance upgrade that PHP could bring is a
real application server that could keep the app bootsrapped between
requests, much like php-pm/reactphp/aerys are trying to do?

Best,

Albert

On Thu, 19 Apr 2018 at 16:11, Nicolas Grekas nicolas.grekas+php@gmail.com
wrote:

Hello internals,

During the last months, I've been working on optimizing Symfony/Twig for
PHP7.
I found a few situations where the PHP engine could likely be optimized,
and where it would benefit real-world apps.

In each case, these might look like micro-optimizations, but they are not
in hot loops.
Here are the links to issues I created in the php bug tracker:

• Class constants are slow: they should be inlined at runtime -
  https://bugs.php.net/76178
• Add array_key_exists() to the list of specialy compiled functions -
  https://bugs.php.net/76148
• Don't trigger copy-on-write when assigning same value -
  https://bugs.php.net/76150

As a bonus, I also created this one today:

• Add hrtime() to zend_try_compile_special_func() -
  https://bugs.php.net/76241

It would be awesome if those ideas could be implemented in a future PHP
version.
(I'm sorry I don't have the knowledge to do it myself.)

Keep up the good work!

Nicolas

On Thu, Apr 19, 2018 at 4:10 PM, Nicolas Grekas <
nicolas.grekas+php@gmail.com> wrote:

Hello internals,

During the last months, I've been working on optimizing Symfony/Twig for
PHP7.
I found a few situations where the PHP engine could likely be optimized,
and where it would benefit real-world apps.

In each case, these might look like micro-optimizations, but they are not
in hot loops.
Here are the links to issues I created in the php bug tracker:

• Class constants are slow: they should be inlined at runtime -
  https://bugs.php.net/76178

This should in principle be straightforward. We just need to add a runtime
cache for the constant lookup (or more generally, the constexpr AST
evaluation). Possible complications:

• The responsible opcode ZEND_RECV_INIT already uses a runtime cache slot
  for the class typehint, so there might be a bit of juggling involved in the
  runtime cache allocation. It might be easiest to always allocate the class
  name cache slot if the default value cache slot is used, even if it's not
  necessary, to avoid making this too complicated.
• Some thought needs to be put into whether this is really valid in all
  cases. Apart from the things we already ignore in other runtime caches
  (value of a constant might change due to NS fallback), there are other
  cases to consider, such as function($foo = "" . M_PI). The output could
  change at runtime based on the LC_NUMERIC locale and the precision ini
  setting, so the default value might change between calls. However, we
  already ignore this issue during compile-time evaluation, so I wouldn't
  consider this to be particularly problematic.

• Add array_key_exists() to the list of specialy compiled functions -
  https://bugs.php.net/76148

This is also straightforward. This will have to be a new opcode, because
the key behavior is subtly different (e.g. for floating point keys) than in
isset(). (And isset has all kinds of special handling not relevant to this.)

• Don't trigger copy-on-write when assigning same value -
  https://bugs.php.net/76150

This one on the other hand would be a major change, and I'm not sure it's a
beneficial one (on average). It might give a large improvement for the
specific, rare case discussed in the bug report, but will be a minor
regression (on a very common operation) for everything else.

As a bonus, I also created this one today:

• Add hrtime() to zend_try_compile_special_func() -
  https://bugs.php.net/76241

It would be awesome if those ideas could be implemented in a future PHP
version.
(I'm sorry I don't have the knowledge to do it myself.)

Keep up the good work!

If someone wants to try their hand implementing the 1st or 2nd
optimization, feel free to ping me for some pointers.

Nikita