This is generalized solution for Bob's idea of speed up self::method() calls without ZEND_FETCH_CLASS.
At first, it adds few new opcodes to separate class related behaviour:
FETCH_STATIC_PROP_R
FETCH_STATIC_PROP_W
FETCH_STATIC_PROP_RW
FETCH_STATIC_PROP_FUNC_ARG
FETCH_STATIC_PROP_UNSET
FETCH_STATIC_PROP_IS
UNSET_STATIC_PROP
ISSET_ISEMPTY_STATIC_PROP
FETCH_CLASS_CONSTANT
At seconds, it enables IS_UNUSED operand to fetch (self, parent or static without separate FETCH_CLASS) for new opcodes and the following ones:
INIT_STATIC_METHOD_CALL
NEW
END_INSTANCEOF
Finaly, opcache optimizer had to be fixed to support new opcodes.
This replaces add_new with update for the RW case. This should not
be problematic for performance, as this branch throws a notice.
Alternatively add_new could also be replaced with add. I went with
update, because it makes $a[0] += 1 behavior the same as
$a[0] = $a[0] + 1.
As it turns out, there is actually no reason to prevent this, it even was a bigger BC break than expected...
Also fixes a memory leak (the Closure leaks) when calling internal functions via Closure by moving it out of leave helper onto caller side for TOP_CODE:
$z = new SplStack; $z->push(20);
$x = (new ReflectionMethod("SplStack", "pop"))->getClosure($z);
var_dump($x());
This VM may work only if all tail calls are optimized, otherwaise it will crach because of stack overflow.
Unfortunately, we can't guarantee tail call optimization in C.
It is by far the most used branch; compilers tended here to split the branches via an &2, &1 check, now it only does an &3 check and prefers ZEND_CALL_NESTED_FUNCTION branch
This fix is actually made for array acessing bug fix (#70262) which is
discarded since we have another better fix, anyway now seems this is still useful
This prevented compilation warnings and disclosed few incorrect usages in Zend/zend_vm_def.h and ext/dom/xpath.c.
Now explicit type casting may be required on call site.
This may break some C extension code, but it shoulfn't be a problem to add explicit casting.
Revert "We cannot safely assume that all op array will be refcount 0 after execution"
This reverts commit b6936adb58.
This change turns out to not have been a clever idea and was causing more weirdness than it helped...