[llvm] [AArch64] Make use of byte FPR stores for bytes extracted from vectors (PR #134117)

[Paul Walker via llvm-commits]

================
@@ -24066,11 +24086,44 @@ static SDValue performSTORECombine(SDNode *N,
     SDValue ExtIdx = Value.getOperand(1);
     EVT VectorVT = Vector.getValueType();
     EVT ElemVT = VectorVT.getVectorElementType();
-    if (!ValueVT.isInteger() || ElemVT == MVT::i8 || MemVT == MVT::i8)
+    if (!ValueVT.isInteger())
       return SDValue();
     if (ValueVT != MemVT && !ST->isTruncatingStore())
       return SDValue();
 
+    if (MemVT == MVT::i8) {
+      auto *ExtCst = dyn_cast<ConstantSDNode>(ExtIdx);
+      if (Subtarget->isNeonAvailable() &&
+          (VectorVT == MVT::v8i8 || VectorVT == MVT::v16i8) && ExtCst &&
+          !ExtCst->isZero() && ST->getBasePtr().getOpcode() != ISD::ADD) {
+        // These can lower to st1.b, which is preferable if we're unlikely to
+        // fold the addressing into the store.
+        return SDValue();
+      }
+
+      // Lower as truncstore of v1i64 -> v1i8 (which can lower to a bsub store).
+      SDValue Zero = DAG.getConstant(0, DL, MVT::i64);
+      SDValue ExtVector;
+      EVT VecVT64 = get64BitVector(ElemVT);
+      if (ExtCst && ExtCst->isZero()) {
+        ExtVector =
+            DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VecVT64, Vector, Zero);
+      } else {
+        SDValue Ext = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
+                                  Value.getValueType(), Vector, ExtIdx);
+        ExtVector = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VecVT64,
+                                DAG.getUNDEF(VecVT64), Ext, Zero);
+      }
+
----------------
paulwalker-arm wrote:

Rather than the extracts and inserts would it be possible to first `nvcast` the input to an `i64` based vector of the same bit length and then optionally extract a `v1i64` subvector to truncstore?

Perhaps it's also worth revisiting the previous patch that added the i16-i64 support to see if it's possible to send all the types down the same new path?

https://github.com/llvm/llvm-project/pull/134117