[llvm] X86: make VBMI2 funnel shifts use VSHLD/VSHRD for const splats (PR #169401)

[via llvm-commits]

================
@@ -57622,6 +57626,41 @@ static SDValue combineFP_TO_xINT_SAT(SDNode *N, SelectionDAG &DAG,
   return SDValue();
 }
 
+// Combiner: turn uniform-constant splat funnel shifts into VSHLD/VSHRD
+static SDValue combineFunnelShift(SDNode *N, SelectionDAG &DAG,
+                                  TargetLowering::DAGCombinerInfo &DCI,
+                                  const X86Subtarget &Subtarget) {
+  SDLoc DL(N);
+  SDValue Op0 = N->getOperand(0);
+  SDValue Op1 = N->getOperand(1);
+  SDValue Amt = N->getOperand(2);
+  EVT VT = Op0.getValueType();
+
+  if (!VT.isVector())
+    return SDValue();
+
+  // Only combine if the operation is legal for this type.
+  // This ensures we don't try to convert types that need to be
+  // widened/promoted.
+  if (!DAG.getTargetLoweringInfo().isOperationLegal(N->getOpcode(), VT))
----------------
ArnavM3434 wrote:

I tried replacing the isOperationLegal check with the DCI.isBeforeLegalizeDAG() check, but it fails cases like this:
vector-fshl-sub128.ll
define <2 x i32> @splatconstant_funnnel_v2i32(<2 x i32> %x, <2 x i32> %y) {
  %res = call <2 x i32> @llvm.fshl.v2i32(<2 x i32> %x, <2 x i32> %y, <2 x i32> <i32 4, i32 4>)
  ret <2 x i32> %res
}

Without isOperationLegal is tries to combine the illegal type v2i32. isBeforeLegalizeDAG() doesn't catch this

https://github.com/llvm/llvm-project/pull/169401