This is a follow up to #35662. The optimizer regression that resulted in that bug is still not fixed.

That bug created a simplified test case, and fixed it for that case. That's good. However, I have not been able to remove my workaround, so the issue still persists in the original code.

The issue appears for an 8x8 kernel and disappears if the kernel is shrunk to 4x4, so it's somehow related to the sheer size of the function, or the length it goes to in loop unrolling.

Preamble for the code that produces the desired codegen:

let mut ab: [[f32; 8]; 8];
ab = ::std::mem::uninitialized();
loop8!(i, loop8!(j, ab[i][j] = 0.));

What the loop8 macros do is that they expand the expression statically, so it corresponds to 64 assignments.

Initialization part for code which is not optimizing well:

let mut ab: [[f32; 8]; 8];
ab = [[0.; 8]; 8];

Another example which is not optimizing well:

let mut ab: [[f32; 8]; 8];
ab = ::std::mem::uninitialized();
for i in 0..8 {
    for j in 0..8 {
        ab[i][j] = 0.;
    }
}

Full reproducer in the next comment.