[wgsl-in] Implement complete validation for size and align attributes

[teoxoy]

From the WGSL spec:

align

Must be a power of 2, and must satisfy the required-alignment for the member type:

If align(n) is applied to a member of S with type T, and S is the store type or contained in the store type for a variable in address space C, then n must satisfy: n = k × RequiredAlignOf(T,C) for some positive integer k.

size

This number must be at least the byte-size of the type of the member:

If size(n) is applied to a member with type T, then SizeOf(T) ≤ n.

Note

• tests have been modified accordingly
• commits are self-contained

[jimblandy]

This all looks fantastic. Much cleaner. Just a few points to address.

[jimblandy]

It seems to me that we need to keep this an error. Naga should never panic on bad input, and this function is used by front ends, so we can't assume the types have been validated yet.

[teoxoy]

I was planning to change the width in the IR from a u8 to either a PowerOfTwoU8 (concrete type or create some generic POT type used by widths and the Alignment type) or at least a NonZeroU8. Thoughts on this?

[jimblandy]

Hmm... In src/lib.rs, at least, Bytes is used exclusively for 2, 4, and 8. Maybe an enum like VectorSize would suffice, with implementations of Mul with Alignment. Stuff like this would need to be changed to use a more relaxed type:

src/back/glsl/mod.rs

    /// A scalar with an unsupported width was requested.
    #[error("A scalar with an unsupported width was requested: {0:?} {1:?}")]
    UnsupportedScalar(crate::ScalarKind, crate::Bytes),

But that seems like an independent change. I think that would be best as a separate PR.

[teoxoy]

@jimblandy I'm a bit confused, I thought we'll revert the removal of the error. The commits have been force-pushed but nothing seems to have changed; just a rebase?