auto range = rewriter.create<mhlo::SubtractOp>(op.getLoc(), limit, start);
    auto abs = rewriter.create<mhlo::AbsOp>(op.getLoc(), range);

    // Delta is not necessarily the same type as start and limit.
    auto abs_cast =
        rewriter.create<mhlo::ConvertOp>(op.getLoc(), compute_type, abs);
    auto delta_cast =
        rewriter.create<mhlo::ConvertOp>(op.getLoc(), compute_type, delta);

    Value scalar_zero = CreateI32SplatConst(op.getLoc(), rewriter, {}, 0);
    Type shape_dtype = getElementTypeOrSelf(op.getElementShape().getType());
    Value num_elements = operands[1];
    IntegerAttr attr;
    if (matchPattern(num_elements, m_Constant(&attr))) {
      return CreateI32SplatConst(op.getLoc(), rewriter, {1}, attr.getInt());
    }

      block_arg.getLoc(),
      DenseIntElementsAttr::get(
          RankedTensorType::get({1, num_replicas}, builder.getIntegerType(32)),
          group_assignment_val));

  StringAttr reduce_op = builder.getStringAttr("Add");
  StringAttr mode = builder.getStringAttr("CrossReplica");
  return builder.create<XlaAllReduceOp>(block_arg.getLoc(), block_arg.getType(),

		// reuse.
		x, _ = l.shared.popHead()
		if x == nil {
			x = p.getSlow(pid)
		}
	}
	runtime_procUnpin()
	if race.Enabled {
		race.Enable()
		if x != nil {
			race.Acquire(poolRaceAddr(x))
		}
	}
	if x == nil && p.New != nil {
		x = p.New()
	}
	return x
}

func (p *Pool) getSlow(pid int) any {
	// See the comment in pin regarding ordering of the loads.

    "$_builder.create<TF::ShapeOp>($0.getLoc(), $1, $2)">;

def IsI32 : NativeCodeCall<
    "$_builder.getBoolAttr(getElementTypeOrSelf($0.getType()).isInteger(32))">;

def CreateTFCastOpF32 : NativeCodeCall<
    "CreateTFCastOpF32(&$_builder, $0.getLoc(), $1, $2)">;

def CreateTFCastOpI32 : NativeCodeCall<
    "CreateTFCastOpI32(&$_builder, $0.getLoc(), $1, $2)">;

def CreateTensorScatterNdOp : NativeCodeCall<

              <exclude>org.apache.maven.toolchain.DefaultToolchain#getLog():METHOD_RETURN_TYPE_CHANGED</exclude>
              <exclude>org.apache.maven.toolchain.DefaultToolchain#DefaultToolchain(org.apache.maven.toolchain.model.ToolchainModel,org.codehaus.plexus.logging.Logger):CONSTRUCTOR_REMOVED</exclude>

          tl_element_shape.replaceAllUsesWith(initial_element_shape);
        } else {
          OpBuilder b(use.getOwner());
          Operation* cast_op = InsertCast(
              b, use.getOwner()->getLoc(),
              tl_element_shape.getResult().getType(), initial_element_shape);
          if (!cast_op) return false;
          tl_element_shape.replaceAllUsesWith(cast_op->getResult(0));
        }
        continue;

  std::string graph_name = op_name + std::to_string(subgraph_index);
  auto subgraph = BuildSubGraph(graph_name, region, subgraph_index);
  if (!subgraph.has_value()) {
    mlir::emitError(region->getLoc(), "failed to build subgraph");
    return -1;
  }
  subgraphs_.push_back(subgraph.value());
  subgraph_index_map_[graph_name] = subgraph_index;
  return subgraph_index;
}