Core: Reduce assignments in Variant Object constructor

[aurpine]

What problem(s) does this PR solve?

• Removes redundant assignments in the Variant Object* constructor.

Additional information

• In the constructor Variant(Object *), ObjData is empty-initialized because ref_pointer unrefs the old value. This is not needed in a constructor so the new implementation extracts the main body from ref_pointer.

[Ivorforce]

Makes sense, but "Removes redundant assignments" is not a sufficient motivation for a PR.

I assume this is for performance purposes? If yes, we'd need to see some measurements as per our optimization guidelines.
I agree the new code should be faster; the question is if it matters in any practical case (so a profile would be enough for me).

[aurpine]

Yes this was identified through GDScript function call profiling. Setup: calling an empty instance function in a loop.

VS2022 Profiling Before:

After:

However I am unable to get results as high in release templates. Maybe MinGW vs MSVC differences?

---

In a C++ benchmark, I only notice a difference when build with MSVC. N=100,000,000 Variant constructions 0.73s (PR) vs 1.11s (master).

Test snippet

    const int N = 100'000'000;

    std::vector<Object *> objs;

    for (int i = 0; i < 1024; i++) {
        objs.push_back(memnew(CustomObject));
    }

    auto t = std::chrono::high_resolution_clock::now();

    for (int i = 0; i < N; i++) {
        Variant a(objs[i & 1023]);
        if (!a) { // So the construction doesn't get optimized out
            break;
        }
    }

    auto t2 = std::chrono::high_resolution_clock::now();
    auto duration = std::chrono::duration_cast<std::chrono::milliseconds>(t2 - t).count();

    print_line(duration / 1000.);

I am not sure if this is change is desired if GCC already optimizes it. I haven't tested clang/LLVM.

[Ivorforce]

Yea, it's possible compilers will be able to reason away most of the code, but I wouldn't always rely on it. MSVC especially isn't very good at it.

We optimize most constructors for this reason, and I think it's warranted to do it here too, since it's a hot path between C++ and Scripts/GDExtension.

[aurpine]

Since we aren't calling an ObjData function anymore... we can use memnew_placement like the other constructors. I think &_get_obj() is the same as _data._mem...

This makes the profiling results look even better:

Because of the struct initializer list with two arguments, the comma is parsed as part of the macro. The fix I found is to add a comma macro.

Re-requesting review as to not sneak in this substantial change.

[Ivorforce]

Technically more correct than the assignment path.