[llvm] [CodeGen][AArch64] Use vector parts for internal non-power-of-two vectors (PR #213982)

[Hari Limaye via llvm-commits]

================
@@ -336,14 +336,19 @@ define <3 x i8> @tv3i8(i1 %c, ptr %p, <3 x i8> %a, <3 x i8> %b) {
 ; CHECK-SD:       // %bb.0: // %entry
 ; CHECK-SD-NEXT:    tbz w0, #0, .LBB11_2
 ; CHECK-SD-NEXT:  // %bb.1: // %t
-; CHECK-SD-NEXT:    mov w5, w2
-; CHECK-SD-NEXT:    mov w6, w3
-; CHECK-SD-NEXT:    mov w7, w4
+; CHECK-SD-NEXT:    fmov s0, w2
----------------
hazzlim wrote:

Yeah I think this is just because we are shuffling to and from the scalarized representation for the calling-convention - I'm not sure how much we expect these weird types to come up as actual parameters?

https://github.com/llvm/llvm-project/pull/213982