Nodes: Allow JS-like assigns, implement AssignNode

[LeviPesin]

Alternative to #26785

Description

Implements SplitNode.assign() and ArrayElementNode.assign(), also adds a separate class for AssignNode.

Same cases of use are supported:

let vec = ...;

vec = vec.w.assign( 1 );
vec = vec.xy.assign( 1 );
vec = vec.xy.assign( vec2( 1, 2 ) );
vec = vec.z.assign( vec.w );

Following use cases are also supported:

let nodeArr = ...;

nodeArr = nodeArr[ 0 ].assign( 1 );

The following use cases for VarNodes are supported as a bonus:

const variable = temp( ... );

variable.w = 1;
variable.xy = 1;
variable.xy = vec2( 1, 2 );
variable.z = variable.w;

Caching issue

Something is quite wrong with its caching in the case of Background:

nodeVar0 = ( ( NodeUniforms.projectionMatrix * NodeUniforms.modelViewMatrix ) * vec4<f32>( nodeVarying1, 1.0 ) );
nodeVar0.z = ( ( NodeUniforms.projectionMatrix * NodeUniforms.modelViewMatrix ) * vec4<f32>( nodeVarying1, 1.0 ) ).w;

But it also was before:

nodeVar0 = vec4<f32>( ( ( NodeUniforms.projectionMatrix * NodeUniforms.modelViewMatrix ) * vec4<f32>( nodeVarying1, 1.0 ) ).x, ( ( NodeUniforms.projectionMatrix * NodeUniforms.modelViewMatrix ) * vec4<f32>( nodeVarying1, 1.0 ) ).y, ( ( NodeUniforms.projectionMatrix * NodeUniforms.modelViewMatrix ) * vec4<f32>( nodeVarying1, 1.0 ) ).w, ( ( NodeUniforms.projectionMatrix * NodeUniforms.modelViewMatrix ) * vec4<f32>( nodeVarying1, 1.0 ) ).w );

So I just fixed it in Background manually by adding .temp() -- but if using the "bonus" use case the caching issue still appears.

Adds support for JS-like assigns:

const variable = someNode;

variable.w = 1;
variable.xy = 1;
variable.xy = vec2( 1, 2 );
variable.z = variable.w;

[sunag]

The code below cannot return the same vector, see #26493 (comment), #26493 (comment)

vec = vec.xy.assign( 1 );

--

The code below is good!

variable.xy = vec2( 1, 2 );

Is it possible to you simplify this PR to support just this?

[sunag]

Would this be compatible with stack.if?

[sunag]

It can be confusing for the user to try something like this and it doesn't work properly:

const variable = temp( ... );

stack.if( bool, ( stack ) => {

	variable.x = 123;

} );

[LeviPesin]

The code below cannot return the same vector

Yeah, I understand now that it looks quite confusing... Maybe we could say that the "main" approach is to do vec.abc = something and the other one is also supported, but don't use it in examples? Like in the same way we have two approach for nodes, node elements (preferred) and "raw" node classes (not really preferred, but still supported).

Would this be compatible with stack.if?

I think the problem here is in VarNode.assign()... It currently works in "compile-time" (JS code), maybe we should change its behavior to work in "run-time" (shader code) so that it would be more in line with other methods.

[sunag]

I don't think we should support this syntax vec = vec.xy.assign( 1 ), if we could shorten the code it would be better.

[LeviPesin]

I don't really see how to support the "bonus" syntax without supporting this one...

[LeviPesin]

I'm afraid with the latest changes to VarNode.assign() the "bonus" syntax can no longer work... The var = var.z.assign( var.w ) still works, though.
@sunag Is there any way we could enable the bonus syntax to work again? I don't really see such...

[sunag]

There is an approach that could work, although we would have to find a good way to do it. This would be having a global stack reference, like currentStack, if we make this work we can have nodes like assign/if/else/loop without needing to use stack.assign or stack.if for example, so we can do everything without having to use stack property. I've had this in mind for some time but haven't tested it yet.

[LeviPesin]

This wouldn't work normally, because currentStack is null when this is ran -- so I've wrapped it in a some kind of "Node IIFE".

[LeviPesin]

@sunag This PR should work now. It has also became much simpler with the latest changes and also now non-VarNodes should be also supported. (There is a one minor issue with possibility of currentStack being null, but this can be overcame by using "Node IIFEs")

(AssignNode can be moved to its own PR if needed)

Also, not sure whether we need SetNode now? We can now write variable.xy = anything or anything like that and it will work.

[sunag]

Thanks!

[sunag]

I think SetNode will still be useful, for inline code and visual editors.

[sunag]

If Proxy would support operators like +, *, - it would be complete :)

[sunag]

It looks like this PR broke webgpu_compute_particles

[sunag]

It looks like this PR broke webgpu_compute_particles

Don't worry about it, I'll publish a PR to review, Great work!

[LeviPesin]

If Proxy would support operators like +, *, - it would be complete :)

I think JS doesn't allow overriding such operators yet ☹️