Emit vmovdqa instructions on x86

[vbgl]

Description

The compiler emits VMOVDQA instructions instead of VMOVDQU when the underlying memory access is known to be aligned.

Checklist

• Add a changelog entry in changes if the PR is a user-visible change

[vbgl]

I’ve doubts regarding this PR: I think that we are too optimistic about the alignment of arrays that are arguments to export functions.

[eponier]

Hm, the compiler should report the expected alignment for the arguments. You think users don't care and don't check that they align their arguments properly, or that it is anyway difficult to ensure for the users (a bit like when we align the stack of export functions ourselves instead of asking the external world to call the export function with the right alignment)?

[vbgl]

I think that even if user care, it is difficult to achieve the desired alignment constraint (from the Jasmin side). For instance, the copy primitive will use aligned array accesses. Here is a typical snippet:

export
fn in_place_computation(reg ptr u8[192] state) -> reg ptr u8[192] {
  reg u256[6] w;

  w = #copy_256(state);
  w = do_stuff(w);
  state = #copy_256(w);
  return state;
}

If such a case, if we want to relax the alignment constraint, we have to give up on the copy operator and write the loops manually.