[AArch64] Use SVE for fixed-length bf16 operations with +sve-b16b16 (#169329)

This can avoid the promotion bf16 -> f32 -> bf16 round trip (or costly
expansions).

Author: MacDue

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

@@ -1783,9 +1783,13 @@ AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM,
       setOperationAction(ISD::VECTOR_DEINTERLEAVE, VT, Custom);
       setOperationAction(ISD::VECTOR_INTERLEAVE, VT, Custom);
       setOperationAction(ISD::VECTOR_SPLICE, VT, Custom);
+    }
 
-      if (Subtarget->hasSVEB16B16() &&
-          Subtarget->isNonStreamingSVEorSME2Available()) {
+    if (Subtarget->hasSVEB16B16() &&
+        Subtarget->isNonStreamingSVEorSME2Available()) {
+      // Note: Use SVE for bfloat16 operations when +sve-b16b16 is available.
+      for (auto VT : {MVT::v4bf16, MVT::v8bf16, MVT::nxv2bf16, MVT::nxv4bf16,
+                      MVT::nxv8bf16}) {
         setOperationAction(ISD::FADD, VT, Custom);
         setOperationAction(ISD::FMA, VT, Custom);
         setOperationAction(ISD::FMAXIMUM, VT, Custom);