[llvm] [ISel] Introduce llvm.clmul intrinsic (PR #168731)

[Oscar Smith via llvm-commits]

================
@@ -8302,6 +8302,57 @@ SDValue TargetLowering::expandFunnelShift(SDNode *Node,
   return DAG.getNode(ISD::OR, DL, VT, ShX, ShY);
 }
 
+SDValue TargetLowering::expandCLMUL(SDNode *Node, SelectionDAG &DAG) const {
+  SDLoc DL(Node);
+  EVT VT = Node->getValueType(0);
+  SDValue X = Node->getOperand(0);
+  SDValue Y = Node->getOperand(1);
+  unsigned BW = VT.getScalarSizeInBits();
+  unsigned Opcode = Node->getOpcode();
+
+  if (VT.isVector() &&
+      isOperationLegalOrCustomOrPromote(Opcode, VT.getVectorElementType()))
+    return DAG.UnrollVectorOp(Node);
----------------
oscardssmith wrote:

Unfortunately, this is a bit of damned if you do, damned if you don't. Given a clmul(8 x i64, 8x i64), scalarizing will be ~8 ops, while lowering to vector bitmath will be ~120, so the scalar version will be a lot faster, (unless the scalar clmul is also missing).

https://github.com/llvm/llvm-project/pull/168731