// CHECK: {
// CHECK-NEXT:   version: 3,
// CHECK-NEXT:   operator_codes: [ {
// CHECK-NEXT:     deprecated_builtin_code: 44,
// CHECK-NEXT:     version: 1,
// CHECK-NEXT:     builtin_code: UNIDIRECTIONAL_SEQUENCE_LSTM
// CHECK-NEXT:   } ],
// CHECK-NEXT:   subgraphs: [ {
// CHECK-NEXT:     tensors: [ {
// CHECK-NEXT:       shape: [ 4, 4, 4 ],
// CHECK-NEXT:       buffer: 1,
// CHECK-NEXT:       name: "arg0",

  // CHECK-DAG:  %[[CST_1:.*]] = "tfl.pseudo_const"() <{value = dense<0.000000e+00> : tensor<4x4xf32>}> : () -> tensor<4x4xf32>

// CHECK-DAG: %[[input_19:.*]] = "quantfork.stats"({{.*}}) <{layerStats = dense<[-2.000000e+00, 4.000000e+00]> : tensor<2xf32>}> : (tensor<1x2xf32>) -> tensor<1x2xf32>

// CHECK: "tfl.unidirectional_sequence_lstm"({{.*}}, %[[input_18]], %[[input_19]], %{{[0-9]+}}, %{{[0-9]+}}, %{{[0-9]+}}, %{{[0-9]+}})
// CHECK-SAME: effective_hidden_scale_intermediate = tensor<*x!quant.calibrated<f32<-5.000000e-01:5.000000e-01>>>

  %cell_input = "tfl.pseudo_const"() {value = dense<1.000000e+00> : tensor<1x3xf32>} : () -> tensor<1x3xf32>
  %16 = "tfl.unidirectional_sequence_lstm"(
    %arg0,
    %1, %2, %3, %4,
    %5, %6, %7, %8,
    %9, %9, %9,
    %10, %11,
    %10, %10,
    %9, %9,
    %recurrent_input, %cell_input,
    %9, %9, %9, %9) {

  // CHECK: "tfl.unidirectional_sequence_lstm"(%arg0, %arg1, %arg2, %arg3, %arg4, %arg5, %arg6, %arg7, %arg8, %arg9, %arg10, %arg11, %arg12, %arg13, %arg14, %arg15, %arg16, %arg17, %arg18, %arg19, %arg20, %arg21, %arg22, %arg23) <{fused_activation_function = "NONE", time_major = false}>...

// TFLite has these variable tensors semantics. So the variable mapping from TF
// to TFLite is actually broken here, we sort of hard-code the variable tensors
// based on the actual ops using them, such as unidirectional_sequence_lstm.
//
// MLIRConverter also benefits from lots of typical compiler optimization like
// merging same input values if they're identical. These optimizations are

  %cell_stats = "quantfork.stats"(%cell_input) {layerStats = dense<[0.0, 1.0]> : tensor<2xf32>} : (tensor<1x3xf32>) -> tensor<1x3xf32>
  %16 = "tfl.unidirectional_sequence_lstm"(
    %input,
    %1, %2, %3, %4,
    %5, %6, %7, %8,
    %9, %9, %9,
    %10, %11,
    %10, %10,
    %9, %9,
    %recurrent_stats, %cell_stats,
    %9, %9, %9, %9) {

    %cell_stats = "quantfork.stats"(%cell_input) {layerStats = dense<[-2.73090601, 7.94872093]> : tensor<2xf32>} : (tensor<1x20xf32>) -> tensor<1x20xf32>
    %0 = "tfl.unidirectional_sequence_lstm"(%arg0,
      %cst_11, %cst_11, %cst_11, %cst_11,
      %cst_3, %cst_3, %cst_3, %cst_3,
      %cst_2, %cst_2, %cst_2,
      %cst_7, %cst_7, %cst_7, %cst_7,
      %cst_2, %cst_2,

    %cell_stats = "quantfork.stats"(%cell_input) {layerStats = dense<[-2.73090601, 7.94872093]> : tensor<2xf32>} : (tensor<1x20xf32>) -> tensor<1x20xf32>
    %0 = "tfl.unidirectional_sequence_lstm"(%arg0,
      %cst_3, %cst_3, %cst_3, %cst_3,
      %cst_3, %cst_3, %cst_3, %cst_3,
      %cst_7, %cst_7, %cst_7,
      %cst_7, %cst_7, %cst_7, %cst_7,
      %cst_3, %cst_2,

  // CHECK: "tfl.unidirectional_sequence_lstm"(%arg0, %arg1, %arg2, %arg3, %arg4, %arg5, %arg6, %arg7, %arg8, %arg9, %arg10, %arg11, %arg12, %arg13, %arg14, %arg15, %arg16, %arg17, %arg18, %arg19, %arg20, %arg21, %arg22, %arg23) <{asymmetric_quantize_inputs = false, cell_clip = 1.000000e+01...