Consistent by-value and by-ref parameters for geometric types

[Ughuuu]

I am trying to make a physics server 2d and physics server 3d, and I want to use:

Rect2
Aabb

If I look at the docs, it says Aabb is the 3d counterpart of Rect2.

The method for merge of Aabb is:

pub fn merge(&self, b: &Aabb) -> Self {

The method for merge for Rect2 is:

pub fn merge(&self, b: Rect2) -> Self {

Could the 2 change so either both have reference or neither have reference for the parameter b?

[Bromeon]

This is a wider issue than merge, we need to establish a clear policy when to use by-value and by-ref for geometric types.

So far, the idea was that Rect2 is by-value, because it has the same size as Vector4, and vectors are always by-value for convenience. However, this is not reflected in Rect2's current &self receivers.

In general, things could be more consistent. I compiled an overview of the parameter passing in our builtin types:

Type	Words (à 4 bytes)	Parameter
Vector2, Vector2i	2	by-value
Vector3, Vector3i	3	by-value
Vector4, Vector4i	4	by-value
Quaternion	4	by-value
Color	4	by-value
Rect2	4	by-ref (mostly)
Rect2i	4	by-ref (mostly)
Aabb	6	by-ref
Plane	4	by-ref
Transform2D	6	mixed
Basis	9	mixed
Transform3D	12	mixed
Projection	16	by-ref

"mixed" means by-value is used if the return type is Self, by-ref otherwise.

[Ughuuu]

Impressive investigation, thanks.

I believe we should do by value those that are up to 4 words, and also include Aabb (6 words) for symmetry reasons. All the others I propose to be by reference, for performance reasons(eg. with worst case being double builds, where a word is 8 bytes).

[Bromeon]

Some more input on this, in terms of performance?

I checked glam, and it seems:

• All vectors use by-val, even DVec4 -- 8 words
• All matrices use by-ref, even Mat2 -- 4 words
• Quat uses by-val -- 4 words

So after consistency changes, we do the same for the above types, and additionally treat Aabb, Plane, Rect2, Rect2i as by-val.