[LLVM][AArch64][SVE] Mark DUP immediate instructions with isAsCheapAsAMove.

[paulwalker-arm]

Doing this means we'll regenerate an immediate rather than copy the result of an existing one, reducing instruction dependency chains.

Looking at the output difference of #133929 I noticed this omission and so I figured it would be better to land this PR first so the effect of #133929 becomes more obvious. In general it seems like a common sense change which I think might be applicable to some other instructions (e.g. pfalse) but for now I've limited to regular vector splat immediate operations.

[MacDue]

This looks reasonable to me 👍

[david-arm]

This does seem like a sensible patch, but can it also have unwanted side-effects? For example, MachineLICM is often incredibly aggressive and likes to hoist invariants as far away from the original loop as it possibly can. When you throw intervening function calls in to the mix, could it lead to more fills and spills? I guess it depends upon the ABI and whether or not it's a level playing field, i.e. if SVE and NEON are treated equally then at least it's no different to marking NEON dups as cheap as a move.

[paulwalker-arm]

When you throw intervening function calls in to the mix, could it lead to more fills and spills?

The affected instructions are also marked with isReMaterializable so we should not need to spill them. Whilst this change might trigger more LICM I figure if that does trigger the need to spill, the end result will be to spill one of these rematerialisable instructions first and so the affect would be the same as if the COPY/DUP had not been hoisted in the first place?

[david-arm]

LGTM!