Relax vector, matrix and array OOB validation

[teoxoy]

Relax vector, matrix and array OOB validation by handling OOB accesses at runtime.

The following example should work; however, validation currently errors because the index is out of bounds.

@compute @workgroup_size(1)
fn main() {
    let idx = 9;
    let v = vec2<f32>();
    let m = mat2x2<f32>();
    let a = array<f32,1>();
    let vi = v[idx];
    let mi = m[idx];
    let ai = a[idx];
}

[edit] Note that WGSL says that let-bound identifiers cannot be constant expressions, so the shader translation time bounds checking rules do not apply to this example.

Related

• error X3504: literal loop terminated early due to out of bounds array access #4381
• Handle vector, matrix and array OOB accesses for HLSL #4389
• Naga SPIR-V backend assumes that let-bound indexes are bounds checked in validation  #6396

[jimblandy]

Note that this issue only covers the excessive validation. There may be missing bounds checks, and those should be filed as separate issues.

[nical]

See also: #4337

[bavalpey]

Curious, why should this work?
Do you just mean it should pass validation? And then the bounds-check policy is used to determine what should happen with vi, mi, etc.?

I think this would be a regression. If you can detect OOB accesses, as these clearly are, at shader translation time, why not report them?

[ErichDonGubler]

@bavalpey: It's for compliance with the WebGPU spec., which forces safe (but implementation-defined) behavior: https://www.w3.org/TR/WGSL/#out-of-bounds-access-sec