Separate internal and user function extension handles

[iluuu1994]

This allows us to skip zend_init_internal_run_time_cache() when opcache is enabled. This function can be quite expensive.

This is backwards in compatible with other extensions that use extension handles for internal functions. If they rely on zend_get_op_array_extension_handle() to register cache for internal functions, they might run into segfaults due to NULL pointers, since the cache won't actually be allocated. Even worse, the offset they rely on may clash with that of other, correctly installed internal function handles, and/or overflow the internal function completely, leading to unpredictable behavior. Hence, it might be better to delay this to master.

Benchmark

Tested with Symfony Demo. The time win is constant across applications, so the relative time win will look more favorable for shorter requests.

Before:
3.145369 + 3.146658 + 3.146848 + 3.148367 + 3.148696 + 3.148788 + 3.149109 + 3.149189 + 3.149542 + 3.149608 + 3.150335 + 3.151215 + 3.151918 + 3.151977 + 3.153058 + 3.153311 + 3.158170 + 3.161079 + 3.185014 + 3.249802
Avg: 3.15740265

After:
3.141876 + 3.142185 + 3.142216 + 3.144321 + 3.144610 + 3.145507 + 3.146388 + 3.146561 + 3.146796 + 3.147445 + 3.147642 + 3.147791 + 3.148390 + 3.148433 + 3.149037 + 3.150181 + 3.150525 + 3.151157 + 3.151939 + 3.153810
Avg: 3.1473405

Diff: -0.32%

[iluuu1994]

Actually, I'm just seeing that zend_observer_fcall_op_array_extension is exported and used in a few other places. I'll have to take a closer look at those. Edit: Done.

[dstogov]

I think it's OK.

@bwoebi why you don't use indirect map_ptrs?
they should be already used for internal classes.
I noticed this function profiling Symfony Demo.
It took ~2% with disabled observer and enabled JIT.

[iluuu1994]

@dstogov I think the downside is that then we have to do the run time check for internal functions too, as well as allocate more arena space. This might still be faster overall, I don't know. We'd have to try.

[iluuu1994]

@dstogov I started working on your idea in master...iluuu1994:php-src:alloc-internal-rtc-at-runtime. I'm very skeptical it will be faster than this solution, because (compared to this one) it only avoids a fbc->type == ZEND_USER_FUNCTION check for user functions. For internal functions, the !RUN_TIME_CACHE(&fbc->op_array) check is likely more expensive.

Anyway, I haven't measured this yet. I will do so tomorrow.

[bwoebi]

@iluuu1994 The current form of the PR looks pretty reasonable and about zero extra cost to observers for me, so I like it :-)