How is SizeClass1D::X4 initializing row_coeffs?

[kornelski]

In forward_transform_avx2 columns skip by 8:

rav1e/src/asm/x86/transform/forward.rs

Line 361 in 620d541

for cg in (0..txfm_size_col).step_by(8) {

but buf can initialize either 4 or 8:

rav1e/src/asm/x86/transform/forward.rs

Line 427 in 620d541

let buf = cast_mut::<4, _>(buf);

which makes me wonder if the buffer is fully initialized, and whether

rav1e/src/asm/x86/transform/forward.rs

Line 439 in 620d541

row_coeffs.reverse();

reverses the right thing (isn't this buf a 4 coeffs, 4 uninit, 4 coeffs, 4 uninit…)?

[kornelski]

Oh wait,

rav1e/src/asm/x86/transform/forward.rs

Line 427 in 620d541

let buf = cast_mut::<4, _>(buf);

this doesn't even do let buf = &mut buf[cg..]; like the previous case. So it's overwriting start of the buffer over and over? Is that a bug?