// CHECK: %[[LHS_2:.*]] = "tf.Reshape"(%[[LHS_SPLIT]]#1, %[[MATMUL_LHS_SHAPE]]) : (tensor<1x4x5xf32>, tensor<2xi64>) -> tensor<4x5xf32>
  // CHECK: %[[LHS_3:.*]] = "tf.Reshape"(%[[LHS_SPLIT]]#2, %[[MATMUL_LHS_SHAPE]]) : (tensor<1x4x5xf32>, tensor<2xi64>) -> tensor<4x5xf32>
  // CHECK: %[[LHS_4:.*]] = "tf.Reshape"(%[[LHS_SPLIT]]#3, %[[MATMUL_LHS_SHAPE]]) : (tensor<1x4x5xf32>, tensor<2xi64>) -> tensor<4x5xf32>

namespace mlir {
namespace odml {
// Converts mhlo.dot_general to tfl.BatchMatMul. Reshape and Transpose ops will
// be inserted to convert to well-formed matrix multiply; i.e., mhlo.dot_general
// -> tfl.batch_matmul(mhlo.transpose(mhlo.reshape(operand)), ...).
// Note:
// 1) Reshape/transpose are inserted because tfl.BatchMatMul requires
// size(contracting_dimensions) = 1 and size(output_dim) = 1, whereas

    %0 = stablehlo.reshape %arg0 : (tensor<10x1x3xf32>) -> tensor<3x10xf32>
    return %0 : tensor<3x10xf32>
  }
  // CHECK: %[[RESHAPE:.*]] = stablehlo.reshape
  // CHECK-NEXT: return %[[RESHAPE]]

  // CHECK: @main_1
  func.func private @main_1(%arg0: tensor<3x10xf32>) -> tensor<6x5xf32> {
    %0 = stablehlo.reshape %arg0 : (tensor<3x10xf32>) -> tensor<6x5xf32>
    return %0 : tensor<6x5xf32>

    }
    shapes->push_back(x_shape);

    TensorShape y_shape;
    if (is_y_batch) {
      // y.shape = [b, k, n]
      y_shape = ty ? TensorShape({b, n, k}) : TensorShape({b, k, n});
    } else {
      // y.shape = [k, n]
      y_shape = ty ? TensorShape({n, k}) : TensorShape({k, n});
    }
    shapes->push_back(y_shape);

    TensorShape z_shape;
    if (is_x_batch || is_y_batch) {

// CHECK:           %[[VAL_6:.*]] = "tfl.reshape"(%[[VAL_0]], %[[VAL_4]]) : (tensor<384x384xf32>, tensor<4xi32>) -> tensor<1x1x384x384xf32>
// CHECK:           %[[VAL_7:.*]] = "tfl.reshape"(%[[VAL_1]], %[[VAL_4]]) : (tensor<384x384xf32>, tensor<4xi32>) -> tensor<1x1x384x384xf32>
// CHECK:           %[[VAL_8:.*]] = "tfl.reshape"(%[[VAL_2]], %[[VAL_4]]) : (tensor<384x384xf32>, tensor<4xi32>) -> tensor<1x1x384x384xf32>

// CHECK-SAME: quant.uniform<i8:f32, 5.000000e-02:-10>
// CHECK: %[[dq1:.*]] = "quantfork.dcast"(%[[q1]])
// CHECK-SAME: quant.uniform<i8:f32, 5.000000e-02:-10>
// CHECK: %[[reshape:.*]] = "tf.Reshape"(%[[dq1]]
// CHECK: %[[q2:.*]] = "quantfork.qcast"(%[[reshape]])
// CHECK-SAME: quant.uniform<i8:f32, 5.000000e-02:-10>
// CHECK: %[[dq2:.*]] = "quantfork.dcast"(%[[q2]])
// CHECK-SAME: quant.uniform<i8:f32, 5.000000e-02:-10>

     // `input`. In other words, the shape of the `Reshape` op are not
     // changed after the transformation.
     (IsTailOfShape $rhs, $input),
     (HasRankAtMost<4> $input),
     (HasRankAtMost<4> $lhs),
     (HasRankAtMost<4> $rhs),
     (SameElementType $input, $rhs)]>;

  // Move binary op before reshape:
  // binary(reshape(lhs), reshape(rhs)) => reshape(binary(lhs, rhs))

// CHECK-NEXT: %[[reshape:.*]] = "tfl.reshape"(%[[conv]], %[[shape]]) : (tensor<1x112x112x32x!quant.uniform<i8:f32, 0.023528476789885875>>, tensor<2xi32>) -> tensor<1x401408x!quant.uniform<i8:f32, 0.023528476789885875>>

  func.return %1 : tensor<64xf32>

// CHECK-LABEL: func @reshape_removeAdjacent
// CHECK:  %[[CST:.*]] = arith.constant dense<64> : tensor<1xi32>
// CHECK:  %[[RESHAPE:.*]] = "tfl.reshape"(%arg0, %[[CST]]) : (tensor<4x4x4xf32>, tensor<1xi32>) -> tensor<64xf32>
// CHECK:  return %[[RESHAPE]]
}

// Checks that tfl.reshape should be removed if its output has more than one

  func.return %3 : tensor<2x1xf32>
}

// CHECK: %[[CST:.*]] = arith.constant dense<1> : tensor<4xi32>
// CHECK:  [[VAL_0:%.*]] = "tfl.reshape"(%1, %[[CST]]) {tac.device = "GPU",  tac.inference_type = "FLOAT"} : (tensor<1xf32>, tensor<4xi32>) -> tensor<1x1x1x1xf32>
// CHECK:  [[VAL_1:%.*]] = "tfl.reshape"(%2, %[[CST]]) {tac.device = "GPU",  tac.inference_type = "FLOAT"} : (tensor<1xf32>, tensor<4xi32>) -> tensor<1x1x1x1xf32>