// CHECK: %[[DQ_1:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>
  // CHECK: %[[DQ_2:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>
  // CHECK: %[[DQ_3:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>
  // CHECK: %[[DQ_4:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>
  // CHECK: %[[DQ_5:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>

    %0 = "tfl.dequantize"(%arg0) : (tensor<1x224x224x3x!quant.uniform<u8:f32, 1.0>>) -> tensor<1x224x224x3xf32>
    %1 = "tfl.dequantize"(%arg1) : (tensor<32x3x3x3x!quant.uniform<u8<1:255>:f32, 1.0>>) -> tensor<32x3x3x3xf32>
    %2 = "tfl.dequantize"(%arg2) : (tensor<32x!quant.uniform<i32:f32, 1.0>>) -> tensor<32xf32>

    // dequantize ops, and insert them between the tf.FakeQuantWithMinMaxVarsOp
    // and its users.
    Value value = tf_op.getOutputs();
    auto quantize = rewriter.create<TFL::QuantizeOp>(
        tf_op.getLoc(), qtype.getValue(), value, qtype);
    auto dequantize = rewriter.create<TFL::DequantizeOp>(
        tf_op.getLoc(), res_type, quantize.getOutput());

  }
};

// Inserts a "tfl.quantize" and "tfl.dequantize" op pair (QDQs) after the
// "tf.FakeQuantWithMinMaxVarsOp" to be constant folded. Since the constant
// folding logic will use a "arith.constant" op to replace the
// "tf.FakeQuantWithMinMaxVarsOp", the "tfl.quantize" op is used to preserve
// the quantization parameters as a TypeAttr and "tfl.dequantize" op used to

// CHECK: %[[CONSTANT:.*]] = "tf.Const"() <{value = dense<0.000000e+00> : tensor<8xf32>}>
// CHECK: %[[QUANTIZE:.*]] = "quantfork.qcast"(%[[CONSTANT]]) : (tensor<8xf32>) -> tensor<8x!quant.uniform<i4:f32, 1.000000e+00:-8>>
// CHECK: %[[DEQUANTIZE:.*]] = "quantfork.dcast"(%[[QUANTIZE]])
// CHECK: return %[[DEQUANTIZE]] : tensor<8xf32>
}

// CHECK-LABEL: fakeQuantNotFolded

// CHECK: %[[CONSTANT:.*]] = "tf.Const"() <{value = dense<0.000000e+00> : tensor<8xf32>}>
// CHECK: %[[QUANTIZE:.*]] = "quantfork.qcast"(%[[CONSTANT]]) : (tensor<8xf32>) -> tensor<8x!quant.uniform<i8:f32, 1.000000e+00:-128>>
// CHECK: %[[DEQUANTIZE:.*]] = "quantfork.dcast"(%[[QUANTIZE]])
// CHECK: return %[[DEQUANTIZE]] : tensor<8xf32>
}

// CHECK-LABEL: fakeQuantNotFolded

  %0 = "tfl.dequantize"(%arg0) : (tensor<1000x1000x!quant.uniform<i8:f32, 7.812500e-03>>) -> tensor<1000x1000xf32>
  %1 = "tfl.gather"(%0, %arg1) {axis = 0 : i32, batch_dims = 0 : i32}: (tensor<1000x1000xf32>, tensor<1x1xi32>) -> tensor<1x1x1000xf32>
  func.return %1 : tensor<1x1x1000xf32>

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

// CHECK-LABEL: dequantize_pushdown_gather_with_reduction

// Dequantize ops will produce 3x larger tensors, so we want to move it after
// some passthrough ops to reduce the memory consumption.
struct PushDownDequantize : public OpRewritePattern<DequantizeOp> {
  explicit PushDownDequantize(MLIRContext* context)
      : OpRewritePattern<DequantizeOp>(context) {}

  LogicalResult matchAndRewrite(DequantizeOp dequantize_op,

  }
  TypeAttr type_attr = TypeAttr::get(new_type);
  auto quantize = builder.create<TFL::QuantizeOp>(value.getLoc(), new_type,
                                                  value, type_attr);
  auto dequantize = builder.create<TFL::DequantizeOp>(
      value.getLoc(), expressed_type, quantize.getOutput());
  value.replaceAllUsesWith(dequantize);

  // `quantize` is using `dequantize` now, so we should set its operand to
  // `value`.

// CHECK:   %{{.*}} = "tfl.dequantize"(%{{.*}}) : (tensor<1x401408x!quant.uniform<u8:f32, 3.906250e-03>>) -> tensor<1x401408xf32>

  %cst = arith.constant dense<[1, 401408]> : tensor<2xi32>