[llvm] [SME] Stop RA from coalescing COPY instructions that transcend beyond smstart/smstop. (PR #78294)

[David Sherwood via llvm-commits]

================
@@ -189,6 +191,26 @@ def : Pat<(int_aarch64_sme_set_tpidr2 i64:$val),
           (MSR 0xde85, GPR64:$val)>;
 def : Pat<(i64 (int_aarch64_sme_get_tpidr2)),
           (MRS 0xde85)>;
+
+multiclass CoalescerBarrierPseudo<RegisterClass rc, list<ValueType> vts> {
+  def NAME : Pseudo<(outs rc:$dst), (ins rc:$idx), []>, Sched<[]> {
+    let Constraints = "$dst = $idx";
+  }
+  foreach vt = vts in {
+    def : Pat<(vt (AArch64CoalescerBarrier (vt rc:$idx))),
+              (!cast<Instruction>(NAME) rc:$idx)>;
+  }
+}
+
+multiclass CoalescerBarriers {
+  defm _FPR16  : CoalescerBarrierPseudo<FPR16, [f16]>;
+  defm _FPR32  : CoalescerBarrierPseudo<FPR32, [f32]>;
+  defm _FPR64  : CoalescerBarrierPseudo<FPR64, [f64, v8i8, v4i16, v2i32, v1i64, v4f16, v2f32, v1f64, v4bf16]>;
+  defm _FPR128 : CoalescerBarrierPseudo<FPR128, [f128, v16i8, v8i16, v4i32, v2i64, v8f16, v4f32, v2f64, v8bf16]>;
----------------
david-arm wrote:

Do we need to worry about v1i128 - or is that not even legal?

https://github.com/llvm/llvm-project/pull/78294