Creating a nested MaybeUninit array unnecessarily calls memset

[abgros]

This function calls memset:

fn create_uninit_array() -> [[MaybeUninit<u8>; 4]; 200] {
    [[MaybeUninit::<u8>::uninit(); 4]; 200]
}

and this one doesn't:

fn create_uninit_array_fast() -> [MaybeUninit<u8>; 800] {
    [MaybeUninit::<u8>::uninit(); 800]
}

I would have expected the same assembly to be generated for both cases.

See on Godbolt

[saethlin]

This is pretty subtle. We only get the deoptimization if the inner array length is 2, 4, or 8. Because the deoptimization happens in SROA + EarlyCSE. The lowering we do here involves creating uninit alloca then storing from them repeatedly into a much larger allocation. SROA confuses the story then EarlyCSE for some reason replaces a store of an undef SSA value with a store of 0 instead of a const undef.

[saethlin]

Upon further inspection, the problem is the splat trick in SROA, which generates a zext undef. EarlyCSE does the only reasonable thing with that and turns it into all zeroes. The problem is that we wanted all the bits to be undef and the transformation that SROA is doing adds a guarantee that some bits are zero.

I'm going to see if I can do more with the "this const has all bytes uninit" logic I added in #138634