execute_launch.erase();

  // Move all regions from old parallel_execute to new parallel_execute.
  for (auto region : llvm::zip(new_parallel_execute_op->getRegions(),
                               parallel_execute_op->getRegions()))
    std::get<0>(region).takeBody(std::get<1>(region));

  // Remove the original parallel_execute.
  parallel_execute_op->dropAllUses();
  parallel_execute.erase();
}

        rewriter.setInsertionPointAfter(value.getDefiningOp());
        ConvertAndWrapUsesInUnrealizedCast(value, converter, rewriter);
      }

      // Convert block arguments
      for (auto &region : op->getRegions()) {
        for (auto &block : region.getBlocks()) {
          rewriter.setInsertionPointToStart(&block);
          for (auto arg : block.getArguments()) {

                     << "  No device for output #" << output_index << "\n");
          continue;
        }
        // Transfer the annotation to both region arguments
        for (Region* region : while_region.getRegions()) {
          BlockArgument arg = region->getArgument(output_index);
          LLVM_DEBUG(llvm::dbgs()
                     << "  Propagating device = '" << *device << "' to arg #"

    OperationState state(op->getLoc(), op->getName().getStringRef(), operands,
                         new_results, op->getAttrs(), op->getSuccessors());
    for (Region &region : op->getRegions()) {
      Region &new_region = *state.addRegion();
      rewriter.inlineRegionBefore(region, new_region, new_region.begin());
      if (failed(rewriter.convertRegionTypes(&new_region, *getTypeConverter())))

    builder.setInsertionPoint(&item);
    OperationState state(item.getLoc(), item.getName());
    types.assign(item.result_type_begin(), item.result_type_end());
    state.addTypes(types);
    for (Region& region : item.getRegions()) {
      state.addRegion()->takeBody(region);
    }
    // Assign existing operands for item.
    operands.assign(item.operand_begin(), item.operand_end());

    // Collect defining ops for existing operands.

      if (quantizing_op->getNumRegions() != 0) {
        for (const auto& indexed_regions :
             llvm::enumerate(quantizing_op->getRegions())) {
          IRMapping mapping;
          indexed_regions.value().cloneInto(
              &quantized_op->getRegion(indexed_regions.index()), mapping);
        }
      }
      for (const auto& output_index_pair : outputs_replaced) {

             << result.getResultNumber()
             << ", resource type results are expected to have been "
                "canonicalized away for region based control flow ops";
    }
  }
  for (Region& region : op.getRegions()) {
    if (failed(DecomposeStackOpsInternal(&region.front(), module,
                                         data_var_to_size_var,
                                         decomposed_partitioned_call_callees)))

//===----------------------------------------------------------------------===//

LogicalResult ParallelExecuteOp::verify() {
  ParallelExecuteOp op = *this;
  const auto& regions = op.getOperation()->getRegions();
  if (regions.empty()) {
    return op.emitOpError() << "must have at least one region.";
  }

  int output_index = 0;
  for (const auto& region_and_index : llvm::enumerate(regions)) {

      if (quantizing_op->getNumRegions() != 0) {
        for (const auto& indexed_regions :
             llvm::enumerate(quantizing_op->getRegions())) {
          Region& target_region =
              quantized_op->getRegion(indexed_regions.index());
          IRMapping mapping;
          indexed_regions.value().cloneInto(&target_region, mapping);
        }
      }

  // If there is no TPUReplicateMetadata op in this block, process blocks in
  // regions attached to the op's in the block.
  if (metadata_map.empty()) {
    for (Operation& op : *block) {
      for (Region& region : op.getRegions()) {
        if (!llvm::hasSingleElement(region))
          return op.emitOpError("Expected single block region");
        if (failed(FormClustersInBlock(&region.front(), side_effect_analysis,