zend_closures.c: Call closure directly

[Girgias]

This prevents a bunch of indirection and calling a C function via a function pointer

[iluuu1994]

So, this is avoiding a call to zend_is_callable_ex()? Is this measurably faster?

[iluuu1994]

Nit: This causes some unnecessary initialization or most fields that are set later. It's probably elided if zend_closure_get_closure() is inlined though.

[Girgias]

So, this is avoiding a call to zend_is_callable_ex()? Is this measurably faster?

Basically yes as the current call does the following:

• call_user_function_named (which is just _call_user_function_impl() which sets up a FCI struct to call
• zend_call_function which, due to lack of an FCC, calls
• zend_is_callable_ex (which calls zend_is_callable_at_frame) which calls
• Z_OBJ_HANDLER_P(callable, get_closure) (which corresponds to zend_closure_get_closure) to get the FCC to continue the zend_call_function() call

I did not measure if this is faster or not, as I'm trying to rely less on the function_name field of the FCI to see if this can be removed at one point and just rely on the FCC struct.

Another reason is that this is the only reference to call_user_function_named.

I could also just inline zend_closure_get_closure to prevent the function call together.

(note to self, I should update https://www.phpinternalsbook.com/php7/internal_types/functions/callables.html or move it to the new php-src internal docs as some stuff has changed since I wrote this)

[iluuu1994]

A quick benchmark would be good when touching performance critical code. Sometimes performance changes unexpectedly.

[Girgias]

A quick benchmark would be good when touching performance critical code. Sometimes performance changes unexpectedly.

What sort of benchmark do you want me to run? I also don't see how this could possibly be slower than the existing indirections.

[iluuu1994]

I see this is only called when doing $c->__invoke(), so it should be ok. But generally, it's not too time consuming to test changes that might have performance implications. You can just do a synthetic benchmark, i.e. hit your code in a loop of enough iterations to be measurable.