Type alignment

[freesig]

Type alignment is set to #[repr(C, align(4))] which seems to be an optimization for risk0.
We should review if this is the correct way to align.
The main issue is this is harmful to performance on the host side so this should be at least under a feature flag.
Something like:

#[derive(Clone, Debug, Hash)]
#[cfg_attr(feature = "zk_alignment", repr(C, align(4)))]
pub struct Address {
    pub words: [u32; 5],
}

Thoughts @pixelcircuits ?

[pixelcircuits]

AI says we definitely don't need align(4) since everything is already made up of u32s. the repr(C) can be removed for the types, but should stay in place for the trie nodes because of how they are being serialized. However, AI does suggest that we push all large fields in the structs to the top.

[freesig]

Yeh unless you are using repr(C) the ordering in the struct doesn't matter. If you do use it then you have to manually align things which I'm not sure we want to do as rust already does that without it. I'm not sure why repr(C) would help with serialization unless we are doing something really unsafe when we serialize

[pixelcircuits]

This issue should really be an investigation as to why the use of [u32] is necessary instead of using #[repr(C, align(4))] on [u8]. The reason it's not like this, is because I previously saw a huge performance difference in proving speed on RiscZero. [u32] forces everything to be word aligned.

[freesig]

Yeh once we have the MVP we should write some experiments to figure out what's going on here. [u8; 4] should be able to be aligned to a single word