[llvm] [X86] added computeKnownBits for X86ISD::GF2P8AFFINEQB (PR #194197)

[via llvm-commits]

================
@@ -39022,6 +39026,51 @@ void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
   Known.resetAll();
   switch (Opc) {
   default: break;
+  case X86ISD::GF2P8AFFINEQB: {
+    SDValue Input = Op.getOperand(0);
+    SDValue Matrix = Op.getOperand(1);
+    SDValue Imm = Op.getOperand(2);
+
+    auto *ImmN = dyn_cast<ConstantSDNode>(Imm);
+    if (!ImmN)
+      break;
+
+    uint8_t Imm8 = ImmN->getZExtValue();
+    APInt UndefElts;
+    SmallVector<APInt, 64> MatEltBits;
+    if (!getTargetConstantBitsFromNode(Matrix, 8, UndefElts, MatEltBits,
+                                       /*AllowWholeUndefs=*/false,
+                                       /*AllowPartialUndefs=*/false))
+      break;
+
+    KnownBits InputKnown = DAG.computeKnownBits(Input, DemandedElts, Depth + 1);
+
+    APInt KnownMask = InputKnown.Zero | InputKnown.One;
+    KnownBits Res(BitWidth);
+    Res.resetAll();
+
+    for (unsigned OutBit = 0; OutBit != 8; ++OutBit) {
+      unsigned RowIdx = 7 - OutBit;
+
+      uint8_t Row = MatEltBits[RowIdx].getZExtValue();
----------------
WalterKruger wrote:

Or you can guard the entire analysis by checking if the matrix is splat. Although it shouldn't be too expensive since it doesn't need additional `computeKnownBits` calls.

https://github.com/llvm/llvm-project/pull/194197