Insufficiently aligned x86_64 SIMD stores

[finnbear]

bvh/src/ray/intersect_x86_64.rs

Lines 68 to 69 in e61924c

	let mut data = std::mem::MaybeUninit::<[f32; 4]>::uninit();
	_mm_store_ps(data.as_mut_ptr() as *mut f32, res);

The docs of _mm_store_ps say:

If the pointer is not aligned to a 128-bit boundary (16 bytes) a general protection fault will be triggered (fatal program crash).

[f32; 4] has the same alignment as f32, or only 4 bytes.

This code appears to be relying on Rust function stack memory layout properties that are not guaranteed.

Possible solutions:

• Create a wrapper type with #[repr(C, align(16))]
• Switch to unaligned SIMD stores, e.g. _mm_storeu_ps

This applies to other x86_64 SIMD stores, not just on f32x4.

[svenstaro]

Good you bring this up. Frankly, we probably shouldn't be using explicit SIMD intrinsics in the first place. There are plenty of libraries that would make this a lot easier and incidentally even allow us to automatically targets other architectures implicitly. In addition, they all run on stable:

• pulp
• wide
• simdeez
• probably more

I don't think we should fix the current SIMD implementation at all.