p.Optab = uint16(cap(optab) - cap(ops) + i + 1)
			return op
		}
	}

	c.ctxt.Diag("illegal combination %v %v %v %v %v %v", p.As, DRconv(a1), DRconv(a2), DRconv(a3), DRconv(a4), DRconv(a5))
	prasm(p)
	// Turn illegal instruction into an UNDEF, avoid crashing in asmout.
	return &Optab{obj.AUNDEF, C_NONE, C_NONE, C_NONE, C_NONE, C_NONE, 49, 4, 0, 0}
}

		a1 = C_LCON
	}
	if (a3 == C_RCON2A || a3 == C_RCON2S) && p.Scond&C_SBIT != 0 {
		a3 = C_LCON
	}

	if false { /*debug['O']*/
		fmt.Printf("oplook %v %v %v %v\n", p.As, DRconv(a1), DRconv(a2), DRconv(a3))
		fmt.Printf("\t\t%d %d\n", p.From.Type, p.To.Type)
	}

	if (p.As == ASRL || p.As == ASRA) && p.From.Type == obj.TYPE_CONST && p.From.Offset == 0 {

			p.Optab = uint16(cap(optab) - cap(ops) + i + 1)
			return op
		}
	}

	c.ctxt.Diag("illegal combination %v %v %v %v", p.As, DRconv(a1), DRconv(a2), DRconv(a3))
	prasm(p)
	// Turn illegal instruction into an UNDEF, avoid crashing in asmout.
	return &Optab{obj.AUNDEF, C_NONE, C_NONE, C_NONE, 49, 4, 0, 0, 0}
}

func cmp(a int, b int) bool {
	if a == b {

// CHECK: }

// CHECK-LABEL: private @composite_conv_with_bias_dynamic_fn_1
// CHECK: %[[CONV:.*]] = stablehlo.convolution(%arg0, %arg1)
// CHECK: %[[SHAPE_OF:.*]] = shape.shape_of %[[CONV]]
// CHECK: %[[DYNAMIC_BROADCAST_IN_DIM:.*]] = stablehlo.dynamic_broadcast_in_dim %arg2, %[[SHAPE_OF]]
// CHECK: %[[ADD:.*]] = stablehlo.add %[[CONV]], %[[DYNAMIC_BROADCAST_IN_DIM]]
// CHECK: return %[[ADD]] : tensor<?x28x28x16xf32>
// CHECK: }

  %conv = "tfl.conv_2d"(%input, %weight, %bias) {dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32,
    fused_activation_function = "NONE", padding = "SAME", stride_h = 2 : i32, stride_w = 2 : i32}
  : (tensor<1x224x224x3xf32>, tensor<32x3x3x3xf32>, tensor<32xf32>) -> tensor<1x112x112x32xf32>
  func.return %conv : tensor<1x112x112x32xf32>

			conv.getTypeIDs[n.Go[:len(n.Go)-9]] = true
		}
	}
	for i, n := range names {
		if types[i] == nil {
			continue
		}
		pos := f.NamePos[n]
		f, fok := types[i].(*dwarf.FuncType)
		if n.Kind != "type" && fok {
			n.Kind = "func"
			n.FuncType = conv.FuncType(f, pos)
		} else {
			n.Type = conv.Type(types[i], pos)
			switch n.Kind {
			case "iconst":

// CHECK: %[[PAD:.*]] = "tf.PadV2"({{.*}}, %[[CONST]], %[[CONST_1]])
// CHECK: %[[CONV:.*]] = "tf.XlaConvV2"(%[[PAD]], %[[WEIGHT]]
// CHECK-SAME: (tensor<1x4x5x5x3xi8>, tensor<2x3x3x3x2xi8>, tensor<3xi32>, tensor<3x2xi32>, tensor<3xi32>, tensor<3xi32>, tensor<i32>) -> tensor<1x3x2x3x2xi32>
// CHECK: %[[SUB:.*]] = "tf.Sub"(%[[CONV]], %[[CONST_2]])
}

// -----

module attributes {tf_saved_model.semantics} {

      auto pad_output_type = UnrankedTensorType::get(elem_type);
      input = rewriter.create<TF::PadOp>(op->getLoc(), pad_output_type, input,
                                         padding_const);

      // Set Conv padding to `VALID` since padding has been handled by Pad op.
      state.padding = rewriter.getStringAttr("VALID");
    }
    auto conv_op = static_cast<const ConcreteType *>(this)->createTFLOp(

    ConstBoolAttrTrue, $asymmetric_quantize_inputs),
  [(HasRank<2> $input),
   (AreLastTwoDimsTransposed $perm_value),
   (IsBoolAttrEqual<"false"> $adj_y)]>;

// Replace conv-->transpose-->add with conv-->add-->transpose
// The bias needs only reshape (i.e. ReshapeNCHWBiasToNHWC) and not transpose
// because the bias's shape simply changes from NxCx1x1 to Nx1x1xC.
def ReorderNCHWTransposeAdd : Pat <

          {"BN",
           {"FusedBatchNorm", "FusedBatchNormV2", "FusedBatchNormV3",
            "_FusedBatchNormEx", "FusedBatchNormGrad", "FusedBatchNormGradV2",
            "FusedBatchNormGradV3"}},
          {"Conv", {"_FusedConv2D"}},
          {"SORT", {"TopKV2"}},  // XLA version much faster then TF version.
          {"MISC",
           // clang-format off
     {"ApproxTopK", "BroadcastTo", "ExpandDims", "Fill", "NoOp",