[llvm] [LV] Handle FSub Partial Reductions (PR #197134)

[Benjamin Maxwell via llvm-commits]

================
@@ -412,5 +412,7 @@ for.exit:
 !8 = !{!"llvm.loop.mustprogress"}
 !9 = !{!"llvm.loop.vectorize.predicate.enable", i1 true}
 !10 = !{!"llvm.loop.vectorize.enable", i1 true}
+
 attributes #0 = { vscale_range(1,16) "target-features"="+sve" }
 attributes #1 = { "target-cpu"="apple-m1" }
+attributes #2 = { vscale_range(1,16) "target-features"="+sve2" }
----------------
MacDue wrote:

I guess requiring SVE2 was a consequence of https://github.com/llvm/llvm-project/pull/194594 removing the lower cost for sub expansions? So we now need `fmlalb/t` from SVE2?

https://github.com/llvm/llvm-project/pull/197134