%1 = "tf.Identity"(%0#3) {device = ""} : (tensor<1x1024xf32>) -> tensor<1x1024xf32>
    func.return %1 : tensor<1x1024xf32>
  }

constexpr absl::string_view kModuleStatic = R"mlir(
  module {
    func.func @main(%arg0: tensor<1x1024xf32>, %arg1: tensor<1024x3xf32>) -> tensor<1x3xf32> attributes {_from_xla_call_module} {
      %0 = stablehlo.dot_general %arg0, %arg1, contracting_dims = [1] x [0], precision = [] : (tensor<1x1024xf32>, tensor<1024x3xf32>) -> tensor<1x3xf32>
      return %0 : tensor<1x3xf32>
    }
  }
)mlir";

    %1 = stablehlo.constant dense<-1.280000e+02> : tensor<1x1024xf32>
    %2 = stablehlo.constant dense<0.003921567> : tensor<1x1024xf32>
    %3 = stablehlo.divide %arg0, %2 : tensor<1x1024xf32>
    %4 = stablehlo.add %3, %1 : tensor<1x1024xf32>
    %5 = "tf.Identity"(%4) {device = ""} : (tensor<1x1024xf32>) -> tensor<1x1024xf32>
    return %5 : tensor<1x1024xf32>
  }

// CHECK: %[[v1:.*]] = "tf.BatchMatMulV2"(%arg0, %0) <{adj_x = false, adj_y = false}> : (tensor<?x?x1024xf32>, tensor<1024x1024xf32>) -> tensor<?x?x1024xf32>
// CHECK: %[[v2:.*]] = "tf.Shape"(%arg0) : (tensor<?x?x1024xf32>) -> tensor<3xi32>
// CHECK: %[[v3:.*]] = "tf.Gather"(%2, %[[cst_1]]) <{validate_indices = true}> : (tensor<3xi32>, tensor<2xi32>) -> tensor<2xi32>

  module {
    func.func private @composite_dot_general_fn_1(%arg0: tensor<1x1024xf32>, %arg1: tensor<1024x3xf32>) -> tensor<1x3xf32> attributes {_from_xla_call_module, tf_quant.composite_function} {
      %0 = stablehlo.dot_general %arg0, %arg1, contracting_dims = [1] x [0], precision = [DEFAULT, DEFAULT] : (tensor<1x1024xf32>, tensor<1024x3xf32>) -> tensor<1x3xf32>
      return %0 : tensor<1x3xf32>
    }
  }
)mlir";

// CHECK:}

func.func @reshape(%arg0 : tensor<1x128xi32>) -> tensor<4x32x1xi32>{
  %0 = "vhlo.reshape_v1"(%arg0) : (tensor<1x128xi32>) -> tensor<4x32x1xi32>
  func.return %0 : tensor<4x32x1xi32>
}

//CHECK:func.func private @reshape(%arg0: tensor<1x128xi32>) -> tensor<4x32x1xi32> {
//CHECK-NEXT: %0 = "vhlo.reshape_v1"(%arg0) : (tensor<1x128xi32>) -> tensor<4x32x1xi32>
//CHECK-NEXT: return %0 : tensor<4x32x1xi32>

  func.return %0 : tensor<2x1024xf32>
  // CHECK-DAG: %[[CST:.*]] = arith.constant dense<1.000000e+00> : tensor<2x512xf32>
  // CHECK-DAG: %[[CST_0:.*]] = arith.constant dense<2.000000e+00> : tensor<1024x512xf32>

}

// Matches %iota generated from the following code (rank 3 example):
//
// %iota_r1 = "mhlo.iota"(){iota_dimension = 0 : i32} : () -> tensor<44xi32>
// %iota = "mhlo.reshape"(%iota_r1): (tensor<44xi32>) -> tensor<1x1x44xi32>
//
// Where $dimensions is of size 1 and $dimensions[0] = 2.
//
// In general matches a 1-D Iota with multiple dimensions of size 1 added
// through a reshape.

// CHECK-NEXT:    %[[RESHAPED_1:.*]] = mhlo.reshape %arg1
// CHECK-NEXT:    %[[BMM_0:.*]] = "tfl.batch_matmul"(%[[RESHAPED_0]], %[[RESHAPED_1]]) <{adj_x = false, adj_y = false, asymmetric_quantize_inputs = false}> : {{.*}} -> tensor<1x1024xf32>
// CHECK-NEXT:    %[[RESHAPED_BMM:.*]] = mhlo.reshape %[[BMM_0]]
// CHECK-NEXT:    return %[[RESHAPED_BMM]] : tensor<1x1x1024xf32>
}