[llvm] [LV][AArch64]: Utilise SVE ld4/st4 instructions via auto-vectorisation (PR #89018)

[Paul Walker via llvm-commits]

paulwalker-arm wrote:

> Hi - Is there a plan for how to handle ld3? We have seen a lot of issues recently with the canonical shuffle representation for fixed-vector ld2/ld3/ld4, and I was wondering if it made sense to move away from shuffles for fixed-length too.

There is but that will require a new intrinsic.  My hope is that rather than having an intrinsic per interleave factor we could model them all using interleave2 and interleave3 (once it's created).  This is why we've started with ld4/st4 support to see if there are any pitfalls to this approach.

Personally I'd love us to move to using these intrinsics for all vector types because it will streamline several code paths.

https://github.com/llvm/llvm-project/pull/89018