[llvm] [AArch64] Consider runtime mode when deciding to use SVE for fixed-length vectors. (PR #96081)

[Paul Walker via llvm-commits]

================
@@ -1528,14 +1528,24 @@ AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM,
       }
     }
 
-    // NEON doesn't support masked loads/stores/gathers/scatters, but SVE does
-    for (auto VT : {MVT::v4f16, MVT::v8f16, MVT::v2f32, MVT::v4f32, MVT::v1f64,
-                    MVT::v2f64, MVT::v8i8, MVT::v16i8, MVT::v4i16, MVT::v8i16,
-                    MVT::v2i32, MVT::v4i32, MVT::v1i64, MVT::v2i64}) {
+    // NEON doesn't support masked loads/stores, but SME and SVE do.
+    for (auto VT :
+         {MVT::v4f16, MVT::v8f16, MVT::v2f32, MVT::v4f32, MVT::v1f64,
+          MVT::v2f64, MVT::v8i8, MVT::v16i8, MVT::v4i16, MVT::v8i16,
+          MVT::v2i32, MVT::v4i32, MVT::v1i64, MVT::v2i64}) {
       setOperationAction(ISD::MLOAD, VT, Custom);
       setOperationAction(ISD::MSTORE, VT, Custom);
-      setOperationAction(ISD::MGATHER, VT, Custom);
-      setOperationAction(ISD::MSCATTER, VT, Custom);
+    }
+
+    // NEON doesn't support masked gathers/scatters, but SVE does.
+    if (Subtarget->isSVEAvailable()) {
+      for (auto VT :
+           {MVT::v4f16, MVT::v8f16, MVT::v2f32, MVT::v4f32, MVT::v1f64,
+            MVT::v2f64, MVT::v8i8, MVT::v16i8, MVT::v4i16, MVT::v8i16,
+            MVT::v2i32, MVT::v4i32, MVT::v1i64, MVT::v2i64}) {
+        setOperationAction(ISD::MGATHER, VT, Custom);
+        setOperationAction(ISD::MSCATTER, VT, Custom);
+      }
----------------
paulwalker-arm wrote:

I'd rather it be done within this PR rather than introduce new code only to change it straight after.  If you want to use this PR to fix the same issue with VECREDUCE_SEQ_FADD  then I'm fine with that as well.

https://github.com/llvm/llvm-project/pull/96081