%5 = "tfl.log"(%4) : (tensor<10x1xf32>) -> tensor<10x1xf32>
  %6 = tfl.sub(%2, %5) {fused_activation_function = "NONE"} : (tensor<10x10xf32>, tensor<10x1xf32>) -> tensor<10x10xf32>
  return %6 : tensor<10x10xf32>
 // CHECK: "tfl.log_softmax"(%arg0) : (tensor<10x10xf32>) -> tensor<10x10xf32>
}

// CHECK-LABEL: fuseLogSoftmaxAxisNegativeOne

// CHECK-DAG:       [[VAL_21:%.*]] = arith.constant...

tensor<1024xf32>, tensor<32x1024xf32>, tensor<1024x1xf32>, tensor<1024xf32>, tensor<1024x1xf32>, tensor<i64>, tensor<i32>, tensor<32x1024xf32>, tensor<i64>, tensor<i64>, tensor<1024xf32>, tensor<32x1024xf32>, tensor<1024x1xf32>, tensor<1024xf32>, tensor<1024x1xf32>, tensor<i64>, tensor<i32>, tensor<32x1024xf32>, tensor<i64>, tensor<i64>, tensor<1024xf32>, tensor<32x1024xf32>, tensor<1024x1xf32>, tensor<1024xf32>, tensor<1024x1xf32>, tensor<i64>, tensor<i32>, tensor<32x1024xf32>, tensor<i64>, tensor<i64>, tensor<1024xf32>,...

    %12 = stablehlo.subtract %10, %11 : tensor<1x3xf32>  // q1 * q2 - z1 * q2
    %13 = stablehlo.multiply %12, %4 : tensor<1x3xf32>  // s1 * s2
    %14 = call @uniform_quantize_1(%13, %5, %6) : (tensor<1x3xf32>, tensor<1x1xf32>, tensor<1x1xi8>) -> tensor<1x3xi8>
    %15 = call @uniform_dequantize_0(%14, %5, %6) : (tensor<1x3xi8>, tensor<1x1xf32>, tensor<1x1xi8>) -> tensor<1x3xf32>
    return %15 : tensor<1x3xf32>
  }
// CHECK-NOT: stablehlo.uniform_quantize

func.func @no_pushdown_multiple_outputs(%arg0: tensor<1000x2x!quant.uniform<i8:f32, 7.812500e-03>>) -> tensor<1000xf32> {
  %0 = "tfl.dequantize"(%arg0) : (tensor<1000x2x!quant.uniform<i8:f32, 7.812500e-03>>) -> tensor<1000x2xf32>
  %1:2 = "tfl.unpack"(%0) {axis = 1 : i32, num = 2 : i32} : (tensor<1000x2xf32>) -> (tensor<1000xf32>, tensor<1000xf32>)
  func.return %1#0 : tensor<1000xf32>

// CHECK-NEXT: tfl.dequantize
// CHECK-NEXT: tfl.unpack
}

  constexpr absl::string_view kProgram = R"mlir(
func.func @main(
    %arg0: tensor<10x10xf32>, %scale: tensor<10xf32>, %zp: tensor<10xi32>
  ) -> tensor<10x10xi8> {
  %0 = "tf.UniformQuantize"(%arg0, %scale, %zp) {
    quantization_axis = 1 : i64,
    quantization_min_val = -128 : i64,
    quantization_max_val = 127 : i64
  } : (tensor<10x10xf32>, tensor<10xf32>, tensor<10xi32>) -> tensor<10x10x!tf_type.qint8>

  %1 = "tf.MatMul"(%arg0, %arg0) {device = "", transpose_a = false, transpose_b = false} : (tensor<10x10xi32>, tensor<10x10xi32>) -> tensor<10x10xi32>
  %2 = "tf.PreventGradient"(%0) : (tensor<10x10xi32>) -> tensor<10x10xi32>
  %3 = "tf.PreventGradient"(%1) : (tensor<10x10xi32>) -> tensor<10x10xi32>
  %4 = "tf.AddV2"(%2, %3) {device = ""} : (tensor<10x10xi32>, tensor<10x10xi32>) -> tensor<10x10xi32>

  %9 = "mhlo.compare"(%0, %8) {comparison_direction = #mhlo<comparison_direction NE>} : (tensor<10x8xf32>, tensor<10x8xf32>) -> tensor<10x8xi1>
  %10 = mhlo.and %7, %9 : tensor<10x8xi1>
  %11 = mhlo.subtract %arg0, %6 : tensor<10x8xf32>
  %12 = mhlo.divide %11, %5 : tensor<10x8xf32>
  %13 = mhlo.add %12, %3 : tensor<10x8xf32>

  func.return %0 : tensor<10x10xf32>
}

func.func @testTanh(%arg0: tensor<10x10xf32>) -> tensor<10x10xf32> {
  // CHECK: _arithmetic_count = 6400 : i64
  %0 = "tfl.tanh"(%arg0) : (tensor<10x10xf32>) -> tensor<10x10xf32>
  func.return %0 : tensor<10x10xf32>
}

func.func @testAddN(%arg0: tensor<10x10xf32>, %arg1: tensor<10x10xf32>, %arg2: tensor<10x10xf32>) -> tensor<10x10xf32> {

func.func @no_change_to_dag(%arg0: tensor<!tf_type.resource<tensor<10x3xf32>>>, %arg1: tensor<!tf_type.resource<tensor<10x3xf32>>>, %arg2: tensor<!tf_type.resource<tensor<10x3xf32>>>, %arg3: tensor<!tf_type.resource<tensor<10x3xf32>>>) -> (tensor<!tf_type.resource<tensor<10x3xf32>>>, tensor<!tf_type.resource<tensor<10x3xf32>>>, tensor<!tf_type.resource...