// RUN: tf-opt %s -tfl-prepare-quantize-dynamic-range="enable-custom-op-quantization=CustomTestOp=1" -tfl-quantize="enable-dynamic-range-quantization=true enable-custom-op-weight-only=CustomTestOp=false" -tfl-post-quantize="enable-no-side-effect=CustomTestOp=false" | FileCheck --check-prefix=NotPrune %s

  // CHECK-DAG: %[[VAL2:.+]] = "tfl.dequantize"(%[[VAL0]])
  // CHECK-DAG: %[[VAL3:.+]] = "tfl.dequantize"(%[[VAL1]])
  // CHECK-DAG: %[[VAL4:.+]] = "tfl.conv_2d"(%arg0, %[[VAL2]], %[[VAL3]]) <{dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "NONE", padding = "SAME", stride_h = 1 : i32, stride_w = 1 : i32}>

    %0 = "tfl.dequantize"(%arg0) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<100x!quant.uniform<i8:f32, 2.000000e-01:-3>>) -> tensor<100xf32>
    %1 = "tfl.dequantize"(%arg1) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<100x!quant.uniform<i8:f32, 2.000000e-01:-3>>) -> tensor<100xf32>

            op->user_begin()->hasTrait<OpTrait::IsTerminator>())
          return failure();
      }
      // If the quantize op is a requantize op, it is being used in other scale
      // adjustments and should be kept. Instead, moving dequantize op before
      // the requantize op to remove the unnecessary requantize op.
      if (auto qtype = quant::QuantizedType::getQuantizedElementType(
              q.getInput().getType())) {

    const auto float_graph = model_->subgraphs()->Get(subgraph_idx);
    // The output graph should have an extra tensor from the added dequantize
    // op.
    ASSERT_EQ(quantized_graph->tensors()->size(),
              float_graph->tensors()->size() + 1);
    // Check that a dequantize op exists.
    int32_t dequant_input_idx = -1;
    int32_t dequant_output_idx = -1;

// CHECK-DAG:       %[[VAL_8:.*]] = "tfl.pseudo_const"(){{.*}}dense<[384, 128]> : tensor<2xi32>
// CHECK:           %[[VAL_9:.*]] = "tfl.dequantize"(%[[VAL_0]]) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<384x512x!quant.uniform<i8:f32, 1.000000e-01>>) -> tensor<384x512xf32>

        "MLIR dump file name.">,
    Option<"merge_fusion_with_dequantize_",
        "merge-fusion-with-dequantize",
        "bool", /*default=*/"false",
        "Whether to merge quantized conv/dot_general fusion with subsequent dequantize.">,
  ];
  let dependentDialects = [
    "mlir::arith::ArithDialect",
    "mlir::stablehlo::StablehloDialect",
    "mlir::quant::QuantizationDialect",

  // CHECK: %[[DEQUANTIZE:.*]] = mhlo.uniform_dequantize %[[CONVERT_2]] : (tensor<2x!quant.uniform<i8:f32, 1.000000e+00:3>>) -> tensor<2xf32>
  // CHECK: return %[[DEQUANTIZE]] : tensor<2xf32>

  %0 = "tf.UniformQuantize"(%arg0, %scales, %zps) {
    quantization_axis = -1 : i64, quantization_min_val = -128 : i64, quantization_max_val = 127 : i64

        // correct float op should be the user of the last DequantizeOp.
        if (llvm::isa<QuantizeOpT>(user)) {
          user = *user->getResult(0).getUsers().begin();
        }
        if (auto dequantize = llvm::dyn_cast<DequantizeOpT>(user)) {
          // Replace all uses, except not quantizable ops that are being used in
          // the float backbone.

  %6 = "tfl.dequantize"(%5) : (tensor<1x401408x!quant.uniform<i8:f32, 3.906250e-03>>) -> tensor<1x401408xf32>
  func.return %6 : tensor<1x401408xf32>

// CHECK-LABEL: func @modified(%arg0: tensor<1x224x224x3xf32>) -> tensor<1x401408xf32>
// CHECK-NEXT: %[[shape:.*]] = arith.constant dense<[1, 401408]> : tensor<2xi32>