[llvm] [X86] Fold FMSUBADD + FADD into VFCMADDC for fp16 (PR #222896)

[Simon Pilgrim via llvm-commits]

================
@@ -56066,6 +56066,54 @@ static SDValue combineFMulcFCMulc(SDNode *N, SelectionDAG &DAG,
   return Res;
 }
 
+// We try to match the following pattern from FMSUBADD(X, A, M) to lower it
+// into complex conjugate fmadd for fp16 (#216290).
+// for vector of the complex form v <v0r, v0i, v1r, v1i, ...>
+// and 2 complex vectors a, b,
+// X = duplicate real (b) : <b0r, b0r, b1r, b1r, ...>
+// A = a
+// M = FMUL (P, Q)
+//   P = adjacent pair swapped (a) : <a0i, a0r, a1i, a1r, ...>
+//   Q = duplicate imaginary (b) : <b0i, b0i, b1i, b1i, ...>
+static bool isCFMulFromFMSUBADD(SDValue N, SelectionDAG &DAG, SDValue &A,
+                                SDValue &B) {
+  SDValue Op0 = N.getOperand(0);
+  SDValue Op1 = N.getOperand(1);
+  SDValue Op2 = N.getOperand(2);
+  SmallVector<SDValue, 2> Inputs;
----------------
RKSimon wrote:

Passing Inputs by reference like this through multiple lambdas looks like a potential footgun - can you clean it up at all? Maybe have matchShufflePattern return SDValue Inputs[0] on success instead of a bool?

https://github.com/llvm/llvm-project/pull/222896