Fully implement structured control flow

[MaikKlein]

No description provided.

[MaikKlein]

#[spirv(fragment)]
fn color_frag(
    frag: Fragment,
    uv: Input<N0, Vec2<f32>>,
    time: Descriptor<N2, N0, f32>,
) -> Output<N0, Vec4<f32>> {
    let uv = uv.data;
    // let mut time = time.data;
    let offset = Vec3::new(0.0f32, 2.0, 4.0);
    let right = uv.x > 0.5;
    let top = uv.y > 0.5;

    let color: Vec3<f32> = match (right, top) {
        (true, true) => Vec3::new(0.0, 0.0, 1.0),
        (true, false) => Vec3::new(0.0, 1.0, 0.0),
        (false, true) => Vec3::new(1.0, 0.0, 0.0),
        (false, false) => Vec3::new(1.0, 1.0, 1.0),
    };

    Output::new(color.extend(1.0))
}

This results in the following CFG

Three branches which all converge into bb7. The dashed lines here are the merge blocks. While 30 is the "correct" merge block, in spir-v all merge blocks have to be unique.

To address this issue we need to generate new blocks for:

bb1, bb2 -> bb8 -> bb7  

bb3, bb4 -> bb9 -> bb7 

bb0 -> bb10 -> bb7

And the new control flow node should be

bb5 => bb8

bb6 => bb9

bb0 => bb10

[MaikKlein]

76c1d30

[MaikKlein]

A more complex example

[MaikKlein]

Interestingly a very simple branch doesn't work.

#[spirv(fragment)]
fn color_frag(
    frag: Fragment,
    uv: Input<N0, Vec2<f32>>,
    time: Descriptor<N2, N0, f32>,
) -> Output<N0, Vec4<f32>> {
    let uv = uv.data;
    let mut color = Vec3::new(1.0, 1.0, 0.0);
    if uv.x > 0.5 {
        color.x = 0.5;
    }

    Output::new(color.extend(1.0))
}

It crashes in Vulkan.
Adding the else branch fixes it.

#[spirv(fragment)]
fn color_frag(
    frag: Fragment,
    uv: Input<N0, Vec2<f32>>,
    time: Descriptor<N2, N0, f32>,
) -> Output<N0, Vec4<f32>> {
    let uv = uv.data;
    let mut color = Vec3::new(1.0, 1.0, 0.0);
    if uv.x > 0.5 {
        color.x = 0.5;
    }
    } else {
        color.x = 0.5;
    }
    Output::new(color.extend(1.0))
}

I am not yet sure which rule I am violating.

[MaikKlein]

After some testing it seems the problem is the OpSwitchinstruction. Switching to OpBranchConditional fixes the problem. This might be a driver problem.

[MaikKlein]

The current implementation seems to break for the ? operator on options. It produces where weird looking cfgs

[MaikKlein]

It seems rlsl doesn't correctly handle unwind and resume.