[ndossche]

This PHP code:

<?php
final class Foo {
    public static $prop = 1;
    function foo() {
        $test = 0;
        for ($i = 0; $i < 100; $i++)
            $test += static::$prop;
        return $test;
    }
}

(new Foo)->foo();

previously generated these loop checks:

.L3:
	cmpq $0x64, %rbx
	jge jit$$trace_exit_1
	movq 0x20(%r14), %rax ; dependency for scope check
	movq 0x40(%r14), %rcx ; load runtime cache ptr
	movq 8(%rcx), %rcx ; scope check
	cmpq 0x10(%rax), %rcx ; scope check
	jne .L5 ; scope check
	movq 0x40(%r14), %rax ; load runtime cache ptr
	movq 0x10(%rax), %rax
	testq %rax, %rax
	je .L5
	cmpb $0, 8(%rax)
	je .L6

Notice how the commented lines are unnecessary: the runtime cache ptr is loaded twice, and the dependency isn't necessary for this final CE case.

Now it generates this:

.L3:
	cmpq $0x64, %rbx
	jge jit$$trace_exit_1
	movq 0x40(%r14), %rcx
	movq 0x10(%rcx), %rax
	testq %rax, %rax
	je .L5
	cmpb $0, 8(%rax)
	je .L6

This is almost optimal, the last check isn't necessary but that's out of scope here.

[iliaal]

Nit: this check is redundant. final trait T {} is a parse error, so a trait scope never has ZEND_ACC_FINAL, and the check above already excludes it.

[ndossche]

Yeah you're correct, I added this after checking for TRAIT_CLONE below, but didn't realize that the combination with final already excludes a trait