[llvm] [X86][BF16] Improve vectorization of BF16 (PR #88486)

[Simon Pilgrim via llvm-commits]

================
@@ -56517,17 +56501,40 @@ static SDValue combineFP16_TO_FP(SDNode *N, SelectionDAG &DAG,
 
 static SDValue combineFP_EXTEND(SDNode *N, SelectionDAG &DAG,
                                 const X86Subtarget &Subtarget) {
+  EVT VT = N->getValueType(0);
+  bool IsStrict = N->isStrictFPOpcode();
+  SDValue Src = N->getOperand(IsStrict ? 1 : 0);
+  EVT SrcVT = Src.getValueType();
+
+  SDLoc dl(N);
+  if (SrcVT.getScalarType() == MVT::bf16) {
+    assert(!IsStrict && "Strict FP doesn't support BF16");
+    if (Src.getOpcode() == ISD::FP_ROUND &&
+        Src.getOperand(0).getValueType() == VT)
+      return Src.getOperand(0);
+
+    if (!SrcVT.isVector())
+      return SDValue();
+
+    if (VT.getVectorElementType() == MVT::f64) {
+      MVT TmpVT = VT.getSimpleVT().changeVectorElementType(MVT::f32);
+      return DAG.getNode(ISD::FP_EXTEND, dl, VT,
+                         DAG.getNode(ISD::FP_EXTEND, dl, TmpVT, Src));
+    }
----------------
RKSimon wrote:

Would this be OK to perform on scalars too?

https://github.com/llvm/llvm-project/pull/88486