output_types, quantizing_op->getAttrs());
      for (int i = 0; i < quantizing_op->getNumRegions(); ++i) {
        new_state.addRegion();
      }
      Operation* quantized_op = rewriter.create(new_state);
      if (quantizing_op->getNumRegions() != 0) {
        for (const auto& indexed_regions :
             llvm::enumerate(quantizing_op->getRegions())) {

        for (const auto& index_and_output :
             llvm::enumerate(partitioned_output.getOutput())) {
          auto idx = (cluster_idx + index_and_output.index()) %
                     new_parallel_execute->getNumRegions();
          const auto output_from_logical_device =
              new_parallel_execute.GetRegionOutputs(
                  idx)[tpu_cluster_output_index];
          index_and_output.value().replaceAllUsesWith(

                               op_with_region->getName().getStringRef(), inputs,
                               output_types, op_with_region->getAttrs());
      for (int i = 0; i < op_with_region->getNumRegions(); ++i) {
        new_state.addRegion();
      }
      Operation* quantized_op = rewriter.create(new_state);
      for (const auto& [index, region] :
           llvm::enumerate(op_with_region->getRegions())) {

  auto new_op = OpBuilder(case_op).create<TF::CaseRegionOp>(
      case_op.getLoc(),
      first_branch->front().getTerminator()->getOperandTypes(),
      case_op.getOperand(), case_op->getAttrs(), case_op.getNumRegions());
  for (const auto& entry : output_buffer_to_size) {
    (*buffer_to_size)[new_op.getResult(std::get<0>(entry))] = {
        new_op.getResult(std::get<1>(entry)), std::get<2>(entry)};
  }

    auto new_op = CreateTfOp<CaseOrIfRegionOp>(rewriter, op, new_result_types,
                                               op.getOperand(), op->getAttrs(),
                                               op.getNumRegions());

    int next_index = 0;
    for (auto result : op.getResults()) {
      if (!result_to_extern_value.count(result)) {
        result.replaceAllUsesWith(new_op.getResult(next_index++));
        continue;

  LogicalResult matchAndRewrite(
      mhlo::WhileOp while_op, OpAdaptor adaptor,
      ConversionPatternRewriter& rewriter) const final {
    // HLO WhileOp should have two regions: cond and body.
    if (while_op->getNumRegions() != 2) return failure();

    // This rule doesn't support mhlo::WhileOp with tuple inputs.
    for (auto type : while_op->getOperandTypes()) {
      if (mlir::isa<TupleType>(type)) return failure();
    }

    } else {
      mhlo_op = rewriter.create<DstOpT>(loc, op.getResultTypes(),
                                        adaptor.getOperands());
    }

    int64_t num_regions = op.getNumRegions();
    for (int64_t idx = 0; idx < num_regions; ++idx) {
      Region &region = mhlo_op.getBodyRegion(idx);
      rewriter.inlineRegionBefore(op.getBodyRegion(idx), region, region.end());