When a type is #[repr(simd)], #[repr(align(N))] annotations are ignored.

[fu5ha]

See this example playground

#[repr(simd, align(32))]
struct Test([f32; 4]);

fn main() {
    println!("{}", core::mem::align_of::<Test>());
}

I expected to see this happen: an alignment of at least 32 to be printed.

Instead, this happened: the alignment is only 16, the align of the underlying platform 128bit simd vector type.

Meta

Seen on playground nightly with version

(2024-09-14 9b72238eb813e9d06e9e)

[workingjubilee]

@fu5ha Just curious: For the use-case in question (the Vec3A type), do you expect unbounded align-up?

[fu5ha]

@workingjubilee It mostly matters for me up to align(16) and I guess there could be similar reasons for align(32) or align(64)

[workingjubilee]

I will be able to fix this after #137256