[wgsl-in] Refactor construction lowering code.

[jimblandy]

(Draft because it depends an a bunch of other stuff.)

• [wgsl-in] Unify ConcreteConstructor and ConcreteConstructorHandle.

  Replace the ConcreteConstructor and ConcreteConstructorHandle
  types in front::wgsl::lower::construction with a single type
  Constructor with a type parameter that determines how it refers to
  Naga types.

  In the single-argument vector construction case, construct Compose
  expressions when the lengths don't match, and leave the problem for
  validation to catch; validate_compose is a better place to invest
  time in improving diagnostics (and arguably producess more specific
  error messages already).

• [wgsl-in] Remove Components enum in favor of slice patterns.

  Remove front::wgsl::lower::construction::Components, and simply use
  slice patterns to decide which Naga code to generate. Supply the type
  of the first argument, if available, as an Option. This is slightly
  redundant, as the Option is Some if and only if the slice is
  non-empty, but in practice it is not too messy, and we can add match
  arms to check for misalignment.

[teoxoy]

• [wgsl-in] Unify ConcreteConstructor and ConcreteConstructorHandle.
  Replace the ConcreteConstructor and ConcreteConstructorHandle
  types in front::wgsl::lower::construction with a single type
  Constructor with a type parameter that determines how it refers to
  Naga types.

Nice refactor!

In the single-argument vector construction case, construct Compose
expressions when the lengths don't match, and leave the problem for
validation to catch; validate_compose is a better place to invest
time in improving diagnostics (and arguably producess more specific
error messages already).

This used to error with:

error: cannot cast a vec3<f32> to a vec2<f32>
   ┌─ wgsl:91:26
   │
91 │     let test = vec2<f32>(vec3<f32>());
   │                          ^^^^^^^^^^ cannot cast a vec3<f32> to a vec2<f32>

It now errors with:

Naga module validation failed on test 'access.wgsl': WithSpan { inner: Function { handle: [2], name: "test_matrix_within_array_within_struct_accesses", source: Expression { handle: [71], source: Compose(ComponentCount { given: 3, expected: 2 }) } }, spans: [(Span { start: 1431, end: 2264 }, "naga::Function [2]"), (Span { start: 1921, end: 1943 }, "naga::Expression [71]")] }

I think we should keep this as it was (at least for now) as there are other cases where you'd run into this same error and it would keep things consistent.

• [wgsl-in] Remove Components enum in favor of slice patterns.
  Remove front::wgsl::lower::construction::Components, and simply use
  slice patterns to decide which Naga code to generate. Supply the type
  of the first argument, if available, as an Option. This is slightly
  redundant, as the Option is Some if and only if the slice is
  non-empty, but in practice it is not too messy, and we can add match
  arms to check for misalignment.

The nice thing about the Components enum was that it worked around rust's limitation on matching on Vecs. With this change, we are matching on slices and then have to copy the components.

I also found it easier to read the code before (even if it was a bit more lengthy) since there are now 3 things we are matching on inside a tuple + the components being a tuple themselves (containing the handle and the span).

[jimblandy]

The nice thing about the Components enum was that it worked around rust's limitation on matching on Vecs. With this change, we are matching on slices and then have to copy the components.

I also found it easier to read the code before (even if it was a bit more lengthy) since there are now 3 things we are matching on inside a tuple + the components being a tuple themselves (containing the handle and the span).

Okay - the intent was to make things clearer. I'll try putting it back. There was a reason I introduced the vector copies, related to my abstract types plan, but I'll see if I can keep them out. I do feel like there was room for improvement in the original Components, but maybe I didn't find the right balance.

[jimblandy]

I ended up removing pretty much any changes to Components, so all that's left here is the stuff that I think you liked.

[teoxoy]

Thanks, looks good!