}) {is_stateless = true} : (tensor<i32>, tensor<i32>, tensor<4x4x2xf32>, tensor<4x2xf32>, tensor<4x2xf32>, tensor<4x4x3xf32>) -> (tensor<i32>, tensor<i32>, tensor<*xf32>, tensor<4x2xf32>, tensor<4x2xf32>, tensor<*xf32>)
  %2 = "tfl.shape"(%1#2) : (tensor<*xf32>) -> tensor<?xi32>
  %3 = "tfl.reshape"(%1#2, %2) : (tensor<*xf32>, tensor<?xi32>) -> tensor<?x4x2xf32>

    %13 = stablehlo.subtract %10, %12 : tensor<1x4x3xf32>  // Precalculated zp_neg.
    %14 = stablehlo.broadcast_in_dim %4, dims = [0, 1, 2] : (tensor<1x1x3xf32>) -> tensor<1x4x3xf32>  // Optional
    %15 = stablehlo.multiply %13, %14 : tensor<1x4x3xf32>  // s1 * s2
    %16 = call @uniform_quantize_1(%15, %5, %6) : (tensor<1x4x3xf32>, tensor<1x1x1xf32>, tensor<1x1x1xi8>) -> tensor<1x4x3xi8>

  // CHECK: %[[LHS_SPLIT:.*]]:6 = "tf.Split"(%[[SPLITTING_AXIS]], %[[LHS_RESHAPED]]) : (tensor<i32>, tensor<6x4x5xf32>) -> (tensor<1x4x5xf32>, tensor<1x4x5xf32>, tensor<1x4x5xf32>, tensor<1x4x5xf32>, tensor<1x4x5xf32>, tensor<1x4x5xf32>)
  // CHECK: %[[LHS_1:.*]] = "tf.Reshape"(%[[LHS_SPLIT]]#0, %[[MATMUL_LHS_SHAPE]]) : (tensor<1x4x5xf32>, tensor<2xi64>) -> tensor<4x5xf32>

      return %0 : tensor<1x1x4xf32>
    }
  )mlir";

  OwningOpRef<ModuleOp> module_op =
      ParseModuleOpString(kModuleXlaCallModuleOpWithDefaultQuantizationMethod);
  ASSERT_TRUE(module_op);

  func.return %0, %1, %2 : tensor<1x4x4xf32>, tensor<1x4x4xf32>, tensor<14x4x4xf32>
}

// -----

func.func @testSplitVOpWithBadSizeSplitsConstantSum(%arg0: tensor<16x4x4xf32>) -> (tensor<0x4x4xf32>, tensor<16x4x4xf32>) {

  // CHECK-LABEL: test_batch_matmul_broadcast_to_einsum
  // CHECK: "tf.Einsum"(%arg0, %arg1) <{equation = "...mk,...kn->...mn"}> : (tensor<2x2x4xf32>, tensor<2x4x2xf32>) -> tensor<2x2x2xf32>
  %0 = "tf.BatchMatMul"(%arg0, %arg1) {adj_x = false, adj_y = false} : (tensor<2x2x4xf32>, tensor<2x4x2xf32>) -> tensor<2x2x2xf32>
  func.return %0: tensor<2x2x2xf32>
}

  %2 = "tfl.fully_connected" (%arg0, %arg1, %cst) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<1x40x37xf32>, tensor<4x37xf32>, none) -> (tensor<1x40x4xf32>)
  %3 = "tfl.add"(%2, %cst1) {fused_activation_function = "NONE"} : (tensor<1x40x4xf32>, tensor<1x1x4xf32>) -> tensor<1x40x4xf32>
  func.return %1, %3 : tensor<1x40x4xf32>, tensor<1x40x4xf32>

func.func @lift_float_batch_matmul(%arg0: tensor<4x4x3xf32>) -> (tensor<4x4x3xf32>) {
  %cst = "tf.Const"() {device = "", value = dense<1.0> : tensor<4x3x3xf32>} : () -> tensor<4x3x3xf32>
  %0 = "tf.BatchMatMulV2"(%arg0, %cst) {adj_x = false, adj_y = false, device = ""} : (tensor<4x4x3xf32>, tensor<4x3x3xf32>) -> tensor<4x4x3xf32>
  return %0 : tensor<4x4x3xf32>

      return %0 : tensor<1x1x4xf32>
    }
  )mlir";

  const OwningOpRef<ModuleOp> module_op =
      ParseModuleOpString(kXlaCallModuleOpWithQuantizationMethodAttr);
  ASSERT_TRUE(module_op);

  func::FuncOp main_fn = FindMainFuncOp(*module_op);

  %43 = "quantfork.stats"(%42) {layerStats = dense<[-1.0, 1.0]> : tensor<2xf32>} : (tensor<1x4x3xf32>) -> tensor<1x4x3xf32>
  %44 = "tfl.strided_slice"(%43, %1, %2, %3) {begin_mask = 7 : i32, ellipsis_mask = 0 : i32, end_mask = 5 : i32, new_axis_mask = 0 : i32, shrink_axis_mask = 0 : i32, offset = false} : (tensor<1x4x3xf32>, tensor<3xi32>, tensor<3xi32>, tensor<3xi32>) -> tensor<1x2x3xf32>