// Try to reuse result transposes.
    TransposeOp transpose = ReuseExistingTranspose(
        &operand, permutation, op, permutation_op, &transpose_ops);
    // If no transpose available for using, create new one.
    if (!transpose)
      transpose =
          builder.create<TransposeOp>(loc, operand.get(), permutation_op);

    operand.set(transpose);
  }

  // Remove unused transpose operations.

    return false;
  }

  return true;
}

// In some cases, transposes may commute with elementwise operations. In order
// to make as many tranposes redudant as possible, we can "push" transposes
// back so that they fuse later on. These patterns handles 2 such cases in
// a conservative fashion; on-net it will never add to the number of transposes
// in the graph.

// ewise(tpose(x), tpose(y)) -> tpose(ewise(x, y))

      "== mlir::BoolAttr::get($_builder.getContext(), true)">>;

def IsNhwcLayoutOp: Constraint<Neg<IsNchwLayoutOp.predicate>>;

// TODO(b/343278954): Move the creation of transposes to a separate prepare pass
// to avoid creating multiple pattern-rewrite rules for the same composite op.
def LegalizeTorchUpsampleBlinear2dComposite: Pat<
    (MHLO_CompositeOp:$old_val
    (variadic $input),

    // `TransposeOp`'s permutation attribute.
    const DenseElementsTransposer transposer(original_shape,
                                             op.getPermutation());
    SmallVector<float> transposed_values =
        transposer.TransposeValues(original_values);

    // Create a new constant op with the transposed values.
    const Location combined_loc =

      tf_executor::CreateTFExecutorIslandCoarseningPass());
  pm.addPass(CreateTFShapeInferencePass());

  // Assign optimal data layout to layout sensitive operations and delete
  // redundant transposes from the IR.
  LayoutOptimizationPipelineOptions layout_optimization_options;
  CreateLayoutOptimizationPipeline(pm.nest<func::FuncOp>(),
                                   layout_optimization_options);

  // CHECK: %[[TRANSPOSE_1:.*]] = "tf.Transpose"(%[[EXPAND_DIMS]], %[[CST_0]]) : (tensor<8x64x1xf32>, tensor<3xi32>) -> tensor<1x8x64xf32>
  // CHECK: %[[TRANSPOSE_2:.*]] = "tf.Transpose"(%arg1, %[[CST_0]]) : (tensor<8x64x64xf32>, tensor<3xi32>) -> tensor<64x8x64xf32>
  // CHECK: %[[ADD:.*]] = "tf.AddV2"(%[[TRANSPOSE_1]], %[[TRANSPOSE_2]]) {device = ""} : (tensor<1x8x64xf32>, tensor<64x8x64xf32>) -> tensor<64x8x64xf32>

  // CHECK: %[[RES_TRANSPOSE_0:[0-9]*]] = "tf.Transpose"(%[[ADD]], %[[RES_PERM]])
  // CHECK: %[[RES_TRANSPOSE_1:[0-9]*]] = "tf.Transpose"(%[[RES_TRANSPOSE_0]], %[[RES_PERM]])
  // CHECK: return %[[RES_TRANSPOSE_1]]

  %0 = "tf.Const"() {value = dense<[0, 3, 1, 2]> : tensor<4xi32>} : () -> tensor<4xi32>
  %1 = "tf.Transpose"(%arg0, %0) : (tensor<1x4x4x8xf32>, tensor<4xi32>) -> tensor<1x8x4x4xf32>

  func.return %1 : tensor<4x7xcomplex<f64>>
}

// BatchMatMulV2 has native as well as fallback lowering patterns available.
// The fallback pattern uses dot_general without broadcast on operands and then
// transposes the output which is faster. However, the fallback pattern doesn't
// support dynamic shaped operands like the native lowering. Verify that
// fallback lowering is preferred for static shaped operands when available.

}

static Value CreateTranspose(OpBuilder& builder, Value source,
                             ArrayRef<int64_t> perm) {
  return builder.create<stablehlo::TransposeOp>(source.getLoc(), source, perm)
      ->getResult(0);
}

// Transform pad(transpose(x)) to transpose(pad(x))
struct TransposeCommuteWithPad : public OpRewritePattern<stablehlo::PadOp> {
  using OpRewritePattern<stablehlo::PadOp>::OpRewritePattern;

// quantized typed tensors and converts them to equivalent ops in the TFLite
// dialect.
std::unique_ptr<OperationPass<func::FuncOp>>
CreateUniformQuantizedStableHloToTflPass();

// Create a pass that commute transposes through specific ops
std::unique_ptr<OperationPass<ModuleOp>> CreateTransposeCommuteOpsPass();

// Create a pass that legalizes MHLO to TF dialect.
std::unique_ptr<OperationPass<ModuleOp>> CreateLegalizeHloToTfPass();