dfir_lang/graph/

meta_graph.rs

#![warn(missing_docs)]

extern crate proc_macro;

use std::collections::{BTreeMap, BTreeSet};
use std::fmt::Debug;
use std::iter::FusedIterator;

use itertools::Itertools;
use proc_macro2::{Ident, Literal, Span, TokenStream};
use quote::{ToTokens, format_ident, quote, quote_spanned};
use serde::{Deserialize, Serialize};
use slotmap::{Key, SecondaryMap, SlotMap, SparseSecondaryMap};
use syn::spanned::Spanned;

use super::graph_write::{Dot, GraphWrite, Mermaid};
use super::ops::{
    DelayType, FloType, OPERATORS, OperatorWriteOutput, WriteContextArgs, find_op_op_constraints,
    null_write_iterator_fn,
};
use super::{
    CONTEXT, Color, DiMulGraph, GRAPH, GraphEdgeId, GraphLoopId, GraphNode, GraphNodeId,
    GraphSubgraphId, HANDOFF_NODE_STR, HandoffKind, MODULE_BOUNDARY_NODE_STR, OperatorInstance,
    PortIndexValue, SINGLETON_SLOT_NODE_STR, Varname, change_spans, get_operator_generics,
};
use crate::diagnostic::{Diagnostic, Diagnostics, Level};
use crate::pretty_span::{PrettyRowCol, PrettySpan};
use crate::process_singletons;

/// A resolved handoff reference: the target node ID plus mutability and access group info.
#[derive(Clone, Debug, Serialize, Deserialize)]
pub struct ResolvedHandoffRef {
    /// The resolved target node ID (`None` if unresolved/error).
    pub node_id: Option<GraphNodeId>,
    /// Whether this is a mutable reference (`#mut var`).
    pub is_mut: bool,
    /// Optional access group for ordering (`#{N} var`).
    pub access_group: Option<u32>,
}

/// An abstract "meta graph" representation of a DFIR graph.
///
/// Can be with or without subgraph partitioning, stratification, and handoff insertion. This is
/// the meta graph used for generating Rust source code in macros from DFIR sytnax.
///
/// This struct has a lot of methods for manipulating the graph, vaguely grouped together in
/// separate `impl` blocks. You might notice a few particularly specific arbitray-seeming methods
/// in here--those are just what was needed for the compilation algorithms. If you need another
/// method then add it.
#[derive(Default, Debug, Serialize, Deserialize)]
pub struct DfirGraph {
    /// Each node type (operator or handoff).
    nodes: SlotMap<GraphNodeId, GraphNode>,

    /// Instance data corresponding to each operator node.
    /// This field will be empty after deserialization.
    #[serde(skip)]
    operator_instances: SecondaryMap<GraphNodeId, OperatorInstance>,
    /// Debugging/tracing tag for each operator node.
    operator_tag: SecondaryMap<GraphNodeId, String>,
    /// Graph data structure (two-way adjacency list).
    graph: DiMulGraph<GraphNodeId, GraphEdgeId>,
    /// Input and output port for each edge.
    ports: SecondaryMap<GraphEdgeId, (PortIndexValue, PortIndexValue)>,

    /// Which loop a node belongs to (or none for top-level).
    node_loops: SecondaryMap<GraphNodeId, GraphLoopId>,
    /// Which nodes belong to each loop.
    loop_nodes: SlotMap<GraphLoopId, Vec<GraphNodeId>>,
    /// For the loop, what is its parent (`None` for top-level).
    loop_parent: SparseSecondaryMap<GraphLoopId, GraphLoopId>,
    /// What loops are at the root.
    root_loops: Vec<GraphLoopId>,
    /// For the loop, what are its child loops.
    loop_children: SecondaryMap<GraphLoopId, Vec<GraphLoopId>>,

    /// Which subgraph each node belongs to.
    node_subgraph: SecondaryMap<GraphNodeId, GraphSubgraphId>,

    /// Which nodes belong to each subgraph.
    subgraph_nodes: SlotMap<GraphSubgraphId, Vec<GraphNodeId>>,
    /// Subgraph IDs in topological sort order (set during partitioning).
    subgraph_toposort: Vec<GraphSubgraphId>,

    /// Resolved handoff varnames references, per node.
    node_handoff_references: SparseSecondaryMap<GraphNodeId, Vec<ResolvedHandoffRef>>,
    /// What variable name each graph node belongs to (if any). For debugging (graph writing) purposes only.
    node_varnames: SparseSecondaryMap<GraphNodeId, Varname>,

    /// Delay type for handoff nodes that represent tick-boundary back-edges.
    /// Set by `order_subgraphs` for `defer_tick` / `defer_tick_lazy`, either on handoff nodes
    /// it injects or on existing handoff nodes that it marks as tick-boundary back-edges.
    handoff_delay_type: SparseSecondaryMap<GraphNodeId, DelayType>,
}

/// Basic methods.
impl DfirGraph {
    /// Create a new empty graph.
    pub fn new() -> Self {
        Default::default()
    }
}

/// Node methods.
impl DfirGraph {
    /// Get a node with its operator instance (if applicable).
    pub fn node(&self, node_id: GraphNodeId) -> &GraphNode {
        self.nodes.get(node_id).expect("Node not found.")
    }

    /// Get the `OperatorInstance` for a given node. Node must be an operator and have an
    /// `OperatorInstance` present, otherwise will return `None`.
    ///
    /// Note that no operator instances will be persent after deserialization.
    pub fn node_op_inst(&self, node_id: GraphNodeId) -> Option<&OperatorInstance> {
        self.operator_instances.get(node_id)
    }

    /// Get the debug variable name attached to a graph node.
    pub fn node_varname(&self, node_id: GraphNodeId) -> Option<&Varname> {
        self.node_varnames.get(node_id)
    }

    /// Get subgraph for node.
    pub fn node_subgraph(&self, node_id: GraphNodeId) -> Option<GraphSubgraphId> {
        self.node_subgraph.get(node_id).copied()
    }

    /// Degree into a node, i.e. the number of predecessors.
    pub fn node_degree_in(&self, node_id: GraphNodeId) -> usize {
        self.graph.degree_in(node_id)
    }

    /// Degree out of a node, i.e. the number of successors.
    pub fn node_degree_out(&self, node_id: GraphNodeId) -> usize {
        self.graph.degree_out(node_id)
    }

    /// Successors, iterator of `(GraphEdgeId, GraphNodeId)` of outgoing edges.
    pub fn node_successors(
        &self,
        src: GraphNodeId,
    ) -> impl '_
    + DoubleEndedIterator<Item = (GraphEdgeId, GraphNodeId)>
    + ExactSizeIterator
    + FusedIterator
    + Clone
    + Debug {
        self.graph.successors(src)
    }

    /// Predecessors, iterator of `(GraphEdgeId, GraphNodeId)` of incoming edges.
    pub fn node_predecessors(
        &self,
        dst: GraphNodeId,
    ) -> impl '_
    + DoubleEndedIterator<Item = (GraphEdgeId, GraphNodeId)>
    + ExactSizeIterator
    + FusedIterator
    + Clone
    + Debug {
        self.graph.predecessors(dst)
    }

    /// Successor edges, iterator of `GraphEdgeId` of outgoing edges.
    pub fn node_successor_edges(
        &self,
        src: GraphNodeId,
    ) -> impl '_
    + DoubleEndedIterator<Item = GraphEdgeId>
    + ExactSizeIterator
    + FusedIterator
    + Clone
    + Debug {
        self.graph.successor_edges(src)
    }

    /// Predecessor edges, iterator of `GraphEdgeId` of incoming edges.
    pub fn node_predecessor_edges(
        &self,
        dst: GraphNodeId,
    ) -> impl '_
    + DoubleEndedIterator<Item = GraphEdgeId>
    + ExactSizeIterator
    + FusedIterator
    + Clone
    + Debug {
        self.graph.predecessor_edges(dst)
    }

    /// Successor nodes, iterator of `GraphNodeId`.
    pub fn node_successor_nodes(
        &self,
        src: GraphNodeId,
    ) -> impl '_
    + DoubleEndedIterator<Item = GraphNodeId>
    + ExactSizeIterator
    + FusedIterator
    + Clone
    + Debug {
        self.graph.successor_vertices(src)
    }

    /// Predecessor nodes, iterator of `GraphNodeId`.
    pub fn node_predecessor_nodes(
        &self,
        dst: GraphNodeId,
    ) -> impl '_
    + DoubleEndedIterator<Item = GraphNodeId>
    + ExactSizeIterator
    + FusedIterator
    + Clone
    + Debug {
        self.graph.predecessor_vertices(dst)
    }

    /// Iterator of node IDs `GraphNodeId`.
    pub fn node_ids(&self) -> slotmap::basic::Keys<'_, GraphNodeId, GraphNode> {
        self.nodes.keys()
    }

    /// Iterator over `(GraphNodeId, &Node)` pairs.
    pub fn nodes(&self) -> slotmap::basic::Iter<'_, GraphNodeId, GraphNode> {
        self.nodes.iter()
    }

    /// Insert a node, assigning the given varname.
    pub fn insert_node(
        &mut self,
        node: GraphNode,
        varname_opt: Option<Ident>,
        loop_opt: Option<GraphLoopId>,
    ) -> GraphNodeId {
        let node_id = self.nodes.insert(node);
        if let Some(varname) = varname_opt {
            self.node_varnames.insert(node_id, Varname(varname));
        }
        if let Some(loop_id) = loop_opt {
            self.node_loops.insert(node_id, loop_id);
            self.loop_nodes[loop_id].push(node_id);
        }
        node_id
    }

    /// Insert an operator instance for the given node. Panics if already set.
    pub fn insert_node_op_inst(&mut self, node_id: GraphNodeId, op_inst: OperatorInstance) {
        assert!(matches!(
            self.nodes.get(node_id),
            Some(GraphNode::Operator(_))
        ));
        let old_inst = self.operator_instances.insert(node_id, op_inst);
        assert!(old_inst.is_none());
    }

    /// Assign all operator instances if not set. Write diagnostic messages/errors into `diagnostics`.
    pub fn insert_node_op_insts_all(&mut self, diagnostics: &mut Diagnostics) {
        // Handle all nodes in two phases, since the helper methods take total ownership of `&self`.
        // Possible to do in one phase, but would require accessing fields directly for partial mutable ownership.

        // Collect operator instances, then assign.
        let mut op_insts = Vec::new();
        // Collect nodes that should be lowered to handoffs (the `handoff()`/`singleton()` pseudo-operators).
        let mut handoff_nodes: Vec<(GraphNodeId, HandoffKind, Span)> = Vec::new();

        for (node_id, node) in self.nodes() {
            let GraphNode::Operator(operator) = node else {
                continue;
            };
            if self.node_op_inst(node_id).is_some() {
                continue;
            };

            // Recognize `handoff()`/`singleton()` pseudo-operators and lower to GraphNode::Handoff.
            let handoff_kind = match &*operator.name_string() {
                "handoff" => Some(HandoffKind::Vec),
                "singleton" => Some(HandoffKind::Singleton),
                "optional" => Some(HandoffKind::Optional),
                _ => None,
            };
            if let Some(kind) = handoff_kind {
                if !operator.args.is_empty() {
                    diagnostics.push(Diagnostic::spanned(
                        operator.path.span(),
                        Level::Error,
                        format!("`{}` takes no arguments.", operator.name_string()),
                    ));
                }
                if operator.type_arguments().is_some() {
                    diagnostics.push(Diagnostic::spanned(
                        operator.path.span(),
                        Level::Error,
                        format!("`{}` takes no generic arguments.", operator.name_string()),
                    ));
                }
                handoff_nodes.push((node_id, kind, operator.path.span()));
                continue;
            }

            // Op constraints.
            let Some(op_constraints) = find_op_op_constraints(operator) else {
                diagnostics.push(Diagnostic::spanned(
                    operator.path.span(),
                    Level::Error,
                    format!("Unknown operator `{}`", operator.name_string()),
                ));
                continue;
            };

            // Input and output ports.
            let (input_ports, output_ports) = {
                let mut input_edges: Vec<(&PortIndexValue, GraphNodeId)> = self
                    .node_predecessors(node_id)
                    .map(|(edge_id, pred_id)| (self.edge_ports(edge_id).1, pred_id))
                    .collect();
                // Ensure sorted by port index.
                input_edges.sort();
                let input_ports: Vec<PortIndexValue> = input_edges
                    .into_iter()
                    .map(|(port, _pred)| port)
                    .cloned()
                    .collect();

                // Collect output arguments (successors).
                let mut output_edges: Vec<(&PortIndexValue, GraphNodeId)> = self
                    .node_successors(node_id)
                    .map(|(edge_id, succ)| (self.edge_ports(edge_id).0, succ))
                    .collect();
                // Ensure sorted by port index.
                output_edges.sort();
                let output_ports: Vec<PortIndexValue> = output_edges
                    .into_iter()
                    .map(|(port, _succ)| port)
                    .cloned()
                    .collect();

                (input_ports, output_ports)
            };

            // Generic arguments.
            let generics = get_operator_generics(diagnostics, operator);
            // Generic argument errors.
            {
                // Span of `generic_args` (if it exists), otherwise span of the operator name.
                let generics_span = generics
                    .generic_args
                    .as_ref()
                    .map(Spanned::span)
                    .unwrap_or_else(|| operator.path.span());

                if !op_constraints
                    .persistence_args
                    .contains(&generics.persistence_args.len())
                {
                    diagnostics.push(Diagnostic::spanned(
                        generics.persistence_args_span().unwrap_or(generics_span),
                        Level::Error,
                        format!(
                            "`{}` should have {} persistence lifetime arguments, actually has {}.",
                            op_constraints.name,
                            op_constraints.persistence_args.human_string(),
                            generics.persistence_args.len()
                        ),
                    ));
                }
                if !op_constraints.type_args.contains(&generics.type_args.len()) {
                    diagnostics.push(Diagnostic::spanned(
                        generics.type_args_span().unwrap_or(generics_span),
                        Level::Error,
                        format!(
                            "`{}` should have {} generic type arguments, actually has {}.",
                            op_constraints.name,
                            op_constraints.type_args.human_string(),
                            generics.type_args.len()
                        ),
                    ));
                }
            }

            op_insts.push((
                node_id,
                OperatorInstance {
                    op_constraints,
                    input_ports,
                    output_ports,
                    singletons_referenced: operator.singletons_referenced.clone(),
                    generics,
                    arguments_pre: operator.args.clone(),
                    arguments_raw: operator.args_raw.clone(),
                },
            ));
        }

        for (node_id, op_inst) in op_insts {
            self.insert_node_op_inst(node_id, op_inst);
        }

        // Replace pseudo-operator nodes with GraphNode::Handoff.
        for (node_id, kind, span) in handoff_nodes {
            self.nodes[node_id] = GraphNode::Handoff {
                kind,
                src_span: span,
                dst_span: span,
            };
        }
    }

    /// Inserts a node between two existing nodes connected by the given `edge_id`.
    ///
    /// `edge`: (src, dst, dst_idx)
    ///
    /// Before: A (src) ------------> B (dst)
    /// After:  A (src) -> X (new) -> B (dst)
    ///
    /// Returns the ID of X & ID of edge OUT of X.
    ///
    /// Note that both the edges will be new and `edge_id` will be removed. Both new edges will
    /// get the edge type of the original edge.
    pub fn insert_intermediate_node(
        &mut self,
        edge_id: GraphEdgeId,
        new_node: GraphNode,
    ) -> (GraphNodeId, GraphEdgeId) {
        let span = Some(new_node.span());

        // Make corresponding operator instance (if `node` is an operator).
        let op_inst_opt = 'oc: {
            let GraphNode::Operator(operator) = &new_node else {
                break 'oc None;
            };
            let Some(op_constraints) = find_op_op_constraints(operator) else {
                break 'oc None;
            };
            let (input_port, output_port) = self.ports.get(edge_id).cloned().unwrap();

            let mut dummy_diagnostics = Diagnostics::new();
            let generics = get_operator_generics(&mut dummy_diagnostics, operator);
            assert!(dummy_diagnostics.is_empty());

            Some(OperatorInstance {
                op_constraints,
                input_ports: vec![input_port],
                output_ports: vec![output_port],
                singletons_referenced: operator.singletons_referenced.clone(),
                generics,
                arguments_pre: operator.args.clone(),
                arguments_raw: operator.args_raw.clone(),
            })
        };

        // Insert new `node`.
        let node_id = self.nodes.insert(new_node);
        // Insert corresponding `OperatorInstance` if applicable.
        if let Some(op_inst) = op_inst_opt {
            self.operator_instances.insert(node_id, op_inst);
        }
        // Update edges to insert node within `edge_id`.
        let (e0, e1) = self
            .graph
            .insert_intermediate_vertex(node_id, edge_id)
            .unwrap();

        // Update corresponding ports.
        let (src_idx, dst_idx) = self.ports.remove(edge_id).unwrap();
        self.ports
            .insert(e0, (src_idx, PortIndexValue::Elided(span)));
        self.ports
            .insert(e1, (PortIndexValue::Elided(span), dst_idx));

        (node_id, e1)
    }

    /// Remove the node `node_id` but preserves and connects the single predecessor and single successor.
    /// Panics if the node does not have exactly one predecessor and one successor, or is not in the graph.
    pub fn remove_intermediate_node(&mut self, node_id: GraphNodeId) {
        assert_eq!(
            1,
            self.node_degree_in(node_id),
            "Removed intermediate node must have one predecessor"
        );
        assert_eq!(
            1,
            self.node_degree_out(node_id),
            "Removed intermediate node must have one successor"
        );
        assert!(
            self.node_subgraph.is_empty() && self.subgraph_nodes.is_empty(),
            "Should not remove intermediate node after subgraph partitioning"
        );

        assert!(self.nodes.remove(node_id).is_some());
        let (new_edge_id, (pred_edge_id, succ_edge_id)) =
            self.graph.remove_intermediate_vertex(node_id).unwrap();
        self.operator_instances.remove(node_id);
        self.node_varnames.remove(node_id);

        let (src_port, _) = self.ports.remove(pred_edge_id).unwrap();
        let (_, dst_port) = self.ports.remove(succ_edge_id).unwrap();
        self.ports.insert(new_edge_id, (src_port, dst_port));
    }

    /// Helper method: determine the "color" (pull vs push) of a node based on its in and out degree,
    /// excluding reference edges. If linear (1 in, 1 out), color is `None`, indicating it can be
    /// either push or pull.
    ///
    /// Note that this does NOT consider `DelayType` barriers (which generally implies `Pull`).
    pub(crate) fn node_color(&self, node_id: GraphNodeId) -> Option<Color> {
        if matches!(self.node(node_id), GraphNode::Handoff { .. }) {
            return Some(Color::Hoff);
        }

        // TODO(shadaj): this is a horrible hack
        if let GraphNode::Operator(op) = self.node(node_id)
            && (op.name_string() == "resolve_futures_blocking"
                || op.name_string() == "resolve_futures_blocking_ordered")
        {
            return Some(Color::Push);
        }

        // In-degree, excluding ref-edges.
        let inn_degree = self.node_predecessor_nodes(node_id).len();
        // Out-degree excluding ref-edges.
        let out_degree = self.node_successor_nodes(node_id).len();

        match (inn_degree, out_degree) {
            (0, 0) => None, // Generally should not happen, "Degenerate subgraph detected".
            (0, 1) => Some(Color::Pull),
            (1, 0) => Some(Color::Push),
            (1, 1) => None, // Linear, can be either push or pull.
            (_many, 0 | 1) => Some(Color::Pull),
            (0 | 1, _many) => Some(Color::Push),
            (_many, _to_many) => Some(Color::Comp),
        }
    }

    /// Set the operator tag (for debugging/tracing).
    pub fn set_operator_tag(&mut self, node_id: GraphNodeId, tag: String) {
        self.operator_tag.insert(node_id, tag);
    }
}

/// Handoff references.
impl DfirGraph {
    /// Set the handoff references for the `node_id` operator. Each reference corresponds to the
    /// same index in the [`crate::parse::Operator::singletons_referenced`] vec.
    pub fn set_node_handoff_references(
        &mut self,
        node_id: GraphNodeId,
        singletons_referenced: Vec<ResolvedHandoffRef>,
    ) -> Option<Vec<ResolvedHandoffRef>> {
        self.node_handoff_references
            .insert(node_id, singletons_referenced)
    }

    /// Gets the handoff references for a node. Returns an empty slice for non-operators and
    /// operators that do not reference handoffs.
    pub fn node_handoff_references(&self, node_id: GraphNodeId) -> &[ResolvedHandoffRef] {
        self.node_handoff_references
            .get(node_id)
            .map(std::ops::Deref::deref)
            .unwrap_or_default()
    }

    /// Collect all refs, grouped by the handoff they're pointing at, then by the access group idx `Option<u32>`.
    pub fn node_handoff_reference_groups(&self) -> NodeHandoffReferenceGroups<'_> {
        let mut handoff_references = NodeHandoffReferenceGroups::new();
        for node_id in self.node_ids() {
            if let GraphNode::Operator(operator) = self.node(node_id) {
                let resolved = self.node_handoff_references(node_id);
                for (resolved_ref, ref_token) in
                    resolved.iter().zip(operator.singletons_referenced.iter())
                {
                    if let Some(target_nid) = resolved_ref.node_id {
                        handoff_references
                            .entry(target_nid)
                            .or_default()
                            .entry(resolved_ref.access_group)
                            .or_default()
                            .push((node_id, resolved_ref, ref_token.span()));
                    }
                }
            }
        }
        handoff_references
    }
}

/// Per-node handoff references, in turn grouped by access group.
/// Map: handoff_node_id -> access_group -> (source `GraphNodeId`, `ResolvedHandoffRef`, `#ref` span)
pub type NodeHandoffReferenceGroups<'a> =
    BTreeMap<GraphNodeId, BTreeMap<Option<u32>, Vec<(GraphNodeId, &'a ResolvedHandoffRef, Span)>>>;

/// Module methods.
impl DfirGraph {
    /// When modules are imported into a flat graph, they come with an input and output ModuleBoundary node.
    /// The partitioner doesn't understand these nodes and will panic if it encounters them.
    /// merge_modules removes them from the graph, stitching the input and ouput sides of the ModuleBondaries based on their ports
    /// For example:
    ///     source_iter([]) -> \[myport\]ModuleBoundary(input)\[my_port\] -> map(|x| x) -> ModuleBoundary(output) -> null();
    /// in the above eaxmple, the \[myport\] port will be used to connect the source_iter with the map that is inside of the module.
    /// The output module boundary has elided ports, this is also used to match up the input/output across the module boundary.
    pub fn merge_modules(&mut self) -> Result<(), Diagnostic> {
        let mod_bound_nodes = self
            .nodes()
            .filter(|(_nid, node)| matches!(node, GraphNode::ModuleBoundary { .. }))
            .map(|(nid, _node)| nid)
            .collect::<Vec<_>>();

        for mod_bound_node in mod_bound_nodes {
            self.remove_module_boundary(mod_bound_node)?;
        }

        Ok(())
    }

    /// see `merge_modules`
    /// This function removes a singular module boundary from the graph and performs the necessary stitching to fix the graph afterward.
    /// `merge_modules` calls this function for each module boundary in the graph.
    fn remove_module_boundary(&mut self, mod_bound_node: GraphNodeId) -> Result<(), Diagnostic> {
        assert!(
            self.node_subgraph.is_empty() && self.subgraph_nodes.is_empty(),
            "Should not remove intermediate node after subgraph partitioning"
        );

        let mut mod_pred_ports = BTreeMap::new();
        let mut mod_succ_ports = BTreeMap::new();

        for mod_out_edge in self.node_predecessor_edges(mod_bound_node) {
            let (pred_port, succ_port) = self.edge_ports(mod_out_edge);
            mod_pred_ports.insert(succ_port.clone(), (mod_out_edge, pred_port.clone()));
        }

        for mod_inn_edge in self.node_successor_edges(mod_bound_node) {
            let (pred_port, succ_port) = self.edge_ports(mod_inn_edge);
            mod_succ_ports.insert(pred_port.clone(), (mod_inn_edge, succ_port.clone()));
        }

        if mod_pred_ports.keys().collect::<BTreeSet<_>>()
            != mod_succ_ports.keys().collect::<BTreeSet<_>>()
        {
            // get module boundary node
            let GraphNode::ModuleBoundary { input, import_expr } = self.node(mod_bound_node) else {
                panic!();
            };

            if *input {
                return Err(Diagnostic {
                    span: *import_expr,
                    level: Level::Error,
                    message: format!(
                        "The ports into the module did not match. input: {:?}, expected: {:?}",
                        mod_pred_ports.keys().map(|x| x.to_string()).join(", "),
                        mod_succ_ports.keys().map(|x| x.to_string()).join(", ")
                    ),
                });
            } else {
                return Err(Diagnostic {
                    span: *import_expr,
                    level: Level::Error,
                    message: format!(
                        "The ports out of the module did not match. output: {:?}, expected: {:?}",
                        mod_succ_ports.keys().map(|x| x.to_string()).join(", "),
                        mod_pred_ports.keys().map(|x| x.to_string()).join(", "),
                    ),
                });
            }
        }

        for (port, (pred_edge, pred_port)) in mod_pred_ports {
            let (succ_edge, succ_port) = mod_succ_ports.remove(&port).unwrap();

            let (src, _) = self.edge(pred_edge);
            let (_, dst) = self.edge(succ_edge);
            self.remove_edge(pred_edge);
            self.remove_edge(succ_edge);

            let new_edge_id = self.graph.insert_edge(src, dst);
            self.ports.insert(new_edge_id, (pred_port, succ_port));
        }

        self.graph.remove_vertex(mod_bound_node);
        self.nodes.remove(mod_bound_node);

        Ok(())
    }
}

/// Edge methods.
impl DfirGraph {
    /// Get the `src` and `dst` for an edge: `(src GraphNodeId, dst GraphNodeId)`.
    pub fn edge(&self, edge_id: GraphEdgeId) -> (GraphNodeId, GraphNodeId) {
        let (src, dst) = self.graph.edge(edge_id).expect("Edge not found.");
        (src, dst)
    }

    /// Get the source and destination ports for an edge: `(src &PortIndexValue, dst &PortIndexValue)`.
    pub fn edge_ports(&self, edge_id: GraphEdgeId) -> (&PortIndexValue, &PortIndexValue) {
        let (src_port, dst_port) = self.ports.get(edge_id).expect("Edge not found.");
        (src_port, dst_port)
    }

    /// Iterator of all edge IDs `GraphEdgeId`.
    pub fn edge_ids(&self) -> slotmap::basic::Keys<'_, GraphEdgeId, (GraphNodeId, GraphNodeId)> {
        self.graph.edge_ids()
    }

    /// Iterator over all edges: `(GraphEdgeId, (src GraphNodeId, dst GraphNodeId))`.
    pub fn edges(
        &self,
    ) -> impl '_
    + ExactSizeIterator<Item = (GraphEdgeId, (GraphNodeId, GraphNodeId))>
    + FusedIterator
    + Clone
    + Debug {
        self.graph.edges()
    }

    /// Insert an edge between nodes thru the given ports.
    pub fn insert_edge(
        &mut self,
        src: GraphNodeId,
        src_port: PortIndexValue,
        dst: GraphNodeId,
        dst_port: PortIndexValue,
    ) -> GraphEdgeId {
        let edge_id = self.graph.insert_edge(src, dst);
        self.ports.insert(edge_id, (src_port, dst_port));
        edge_id
    }

    /// Removes an edge and its corresponding ports and edge type info.
    pub fn remove_edge(&mut self, edge: GraphEdgeId) {
        let (_src, _dst) = self.graph.remove_edge(edge).unwrap();
        let (_src_port, _dst_port) = self.ports.remove(edge).unwrap();
    }
}

/// Subgraph methods.
impl DfirGraph {
    /// Nodes belonging to the given subgraph.
    pub fn subgraph(&self, subgraph_id: GraphSubgraphId) -> &Vec<GraphNodeId> {
        self.subgraph_nodes
            .get(subgraph_id)
            .expect("Subgraph not found.")
    }

    /// Iterator over all subgraph IDs.
    pub fn subgraph_ids(&self) -> slotmap::basic::Keys<'_, GraphSubgraphId, Vec<GraphNodeId>> {
        self.subgraph_nodes.keys()
    }

    /// Subgraph IDs in topological sort order.
    pub fn subgraph_toposort(&self) -> &[GraphSubgraphId] {
        &self.subgraph_toposort
    }

    /// Set the topological sort order for subgraphs.
    pub fn set_subgraph_toposort(&mut self, order: Vec<GraphSubgraphId>) {
        self.subgraph_toposort = order;
    }

    /// Iterator over all subgraphs, ID and members: `(GraphSubgraphId, Vec<GraphNodeId>)`.
    pub fn subgraphs(&self) -> slotmap::basic::Iter<'_, GraphSubgraphId, Vec<GraphNodeId>> {
        self.subgraph_nodes.iter()
    }

    /// Create a subgraph consisting of `node_ids`. Returns an error if any of the nodes are already in a subgraph.
    pub fn insert_subgraph(
        &mut self,
        node_ids: Vec<GraphNodeId>,
    ) -> Result<GraphSubgraphId, (GraphNodeId, GraphSubgraphId)> {
        // Check none are already in subgraphs
        for &node_id in node_ids.iter() {
            if let Some(&old_sg_id) = self.node_subgraph.get(node_id) {
                return Err((node_id, old_sg_id));
            }
        }
        let subgraph_id = self.subgraph_nodes.insert_with_key(|sg_id| {
            for &node_id in node_ids.iter() {
                self.node_subgraph.insert(node_id, sg_id);
            }
            node_ids
        });

        Ok(subgraph_id)
    }

    /// Removes a node from its subgraph. Returns true if the node was in a subgraph.
    pub fn remove_from_subgraph(&mut self, node_id: GraphNodeId) -> bool {
        if let Some(old_sg_id) = self.node_subgraph.remove(node_id) {
            self.subgraph_nodes[old_sg_id].retain(|&other_node_id| other_node_id != node_id);
            true
        } else {
            false
        }
    }

    /// Gets the delay type for a handoff node, if set.
    pub fn handoff_delay_type(&self, node_id: GraphNodeId) -> Option<DelayType> {
        self.handoff_delay_type.get(node_id).copied()
    }

    /// Sets the delay type for a handoff node.
    pub fn set_handoff_delay_type(&mut self, node_id: GraphNodeId, delay_type: DelayType) {
        self.handoff_delay_type.insert(node_id, delay_type);
    }

    /// Helper: finds the first index in `subgraph_nodes` where it transitions from pull to push.
    fn find_pull_to_push_idx(&self, subgraph_nodes: &[GraphNodeId]) -> usize {
        subgraph_nodes
            .iter()
            .position(|&node_id| {
                self.node_color(node_id)
                    .is_some_and(|color| Color::Pull != color)
            })
            .unwrap_or(subgraph_nodes.len())
    }
}

/// Display/output methods.
impl DfirGraph {
    /// Helper to generate a deterministic `Ident` for the given node.
    fn node_as_ident(&self, node_id: GraphNodeId, is_pred: bool) -> Ident {
        let name = match &self.nodes[node_id] {
            GraphNode::Operator(_) => format!("op_{:?}", node_id.data()),
            GraphNode::Handoff {
                kind: HandoffKind::Vec,
                ..
            } => format!(
                "hoff_{:?}_{}",
                node_id.data(),
                if is_pred { "recv" } else { "send" }
            ),
            GraphNode::Handoff {
                kind: HandoffKind::Singleton | HandoffKind::Optional,
                ..
            } => format!(
                "singleton_{:?}_{}",
                node_id.data(),
                if is_pred { "recv" } else { "send" }
            ),
            GraphNode::ModuleBoundary { .. } => panic!(),
        };
        let span = match (is_pred, &self.nodes[node_id]) {
            (_, GraphNode::Operator(operator)) => operator.span(),
            (true, &GraphNode::Handoff { src_span, .. }) => src_span,
            (false, &GraphNode::Handoff { dst_span, .. }) => dst_span,
            (_, GraphNode::ModuleBoundary { .. }) => panic!(),
        };
        Ident::new(&name, span)
    }

    /// Helper to generate the main buffer `Ident` for a handoff node.
    fn hoff_buf_ident(&self, hoff_id: GraphNodeId, span: Span) -> Ident {
        Ident::new(&format!("hoff_{:?}_buf", hoff_id.data()), span)
    }

    /// Helper to generate the back (double-buffer) `Ident` for a handoff node.
    fn hoff_back_ident(&self, hoff_id: GraphNodeId, span: Span) -> Ident {
        Ident::new(&format!("hoff_{:?}_back", hoff_id.data()), span)
    }

    /// Resolve the handoff references via [`Self::node_handoff_references`] for the given `node_id`.
    /// Returns token streams for each reference:
    /// - For HandoffKind::Singleton: `buf.as_ref().unwrap()` (shared, `&T`) or
    ///   `buf.as_mut().unwrap()` (mutable, `&mut T`)
    /// - For HandoffKind::Optional: `&buf` (shared, `&Option<T>`) or
    ///   `&mut buf` (mutable, `&mut Option<T>`)
    /// - For HandoffKind::Vec: `&buf` (shared, `&Vec<T>`) or
    ///   `&mut buf` (mutable, `&mut Vec<T>`)
    fn helper_resolve_singletons(&self, node_id: GraphNodeId, span: Span) -> Vec<TokenStream> {
        self.node_handoff_references(node_id)
            .iter()
            .map(|resolved_ref| {
                // TODO(mingwei): this `expect` should be caught in error checking
                let ref_node_id = resolved_ref
                    .node_id
                    .expect("Expected singleton to be resolved but was not, this is a bug.");
                let is_mut = resolved_ref.is_mut;
                match self.node(ref_node_id) {
                    GraphNode::Handoff {
                        kind: HandoffKind::Singleton,
                        ..
                    } => {
                        let buf_ident = self.hoff_buf_ident(ref_node_id, span);
                        if is_mut {
                            quote_spanned! {span=> #buf_ident.as_mut().unwrap() }
                        } else {
                            quote_spanned! {span=> #buf_ident.as_ref().unwrap() }
                        }
                    }
                    GraphNode::Handoff {
                        kind: HandoffKind::Optional | HandoffKind::Vec,
                        ..
                    } => {
                        let buf_ident = self.hoff_buf_ident(ref_node_id, span);
                        if is_mut {
                            quote_spanned! {span=> &mut #buf_ident }
                        } else {
                            quote_spanned! {span=> &#buf_ident }
                        }
                    }
                    _ => {
                        unreachable!("Only handoff nodes should be reachable as handoff references")
                    }
                }
            })
            .collect::<Vec<_>>()
    }

    /// Returns each subgraph's receive and send handoffs.
    /// `Map<GraphSubgraphId, (recv handoffs, send handoffs)>`
    fn helper_collect_subgraph_handoffs(
        &self,
    ) -> SecondaryMap<GraphSubgraphId, (Vec<GraphNodeId>, Vec<GraphNodeId>)> {
        // Get data on handoff src and dst subgraphs.
        let mut subgraph_handoffs: SecondaryMap<
            GraphSubgraphId,
            (Vec<GraphNodeId>, Vec<GraphNodeId>),
        > = self
            .subgraph_nodes
            .keys()
            .map(|k| (k, Default::default()))
            .collect();

        // For each handoff/singleton node, add it to the `send`/`recv` lists for the corresponding subgraphs.
        for (hoff_id, hoff) in self.nodes() {
            if !matches!(hoff, GraphNode::Handoff { .. }) {
                continue;
            }
            // Receivers from the handoff. (Should really only be one).
            for (_edge, succ_id) in self.node_successors(hoff_id) {
                let succ_sg = self
                    .node_subgraph(succ_id)
                    .expect("bug: successor not in subgraph, may be a doubled/adjacent handoff");
                subgraph_handoffs[succ_sg].0.push(hoff_id);
            }
            // Senders into the handoff. (Should really only be one).
            for (_edge, pred_id) in self.node_predecessors(hoff_id) {
                let pred_sg = self
                    .node_subgraph(pred_id)
                    .expect("bug: predecessor not in subgraph, may be a doubled/adjacent handoff");
                subgraph_handoffs[pred_sg].1.push(hoff_id);
            }
        }

        subgraph_handoffs
    }

    /// Compute the output handoffs exiting each loop (sender inside, receiver outside).
    /// Returns a map from loop ID to the list of handoff node IDs that exit that loop.
    fn helper_loop_output_handoffs(&self) -> SecondaryMap<GraphLoopId, Vec<GraphNodeId>> {
        let mut loop_hoffs_out = SecondaryMap::<GraphLoopId, Vec<GraphNodeId>>::new();

        for (hoff_id, hoff) in self.nodes() {
            if !matches!(hoff, GraphNode::Handoff { .. }) {
                continue;
            }

            let loop_pred = self
                .node_predecessors(hoff_id)
                .next()
                .and_then(|(_, pred)| self.node_loop(pred));
            let loop_succ = self
                .node_successors(hoff_id)
                .next()
                .and_then(|(_, succ)| self.node_loop(succ));

            if let Some(loop_pred) = loop_pred
                && loop_succ == self.loop_parent(loop_pred)
            {
                // Pred is inside a child loop, succ is in the parent/outer.
                loop_hoffs_out
                    .entry(loop_pred)
                    .expect("loop removed")
                    .or_default()
                    .push(hoff_id);
            }
        }

        loop_hoffs_out
    }

    /// Returns true if `node_loop` is `loop_id` or a (transitive) child of `loop_id`.
    fn is_inside_loop(&self, node_loop: Option<GraphLoopId>, loop_id: GraphLoopId) -> bool {
        let mut current = node_loop;
        while let Some(l) = current {
            if l == loop_id {
                return true;
            }
            current = self.loop_parent(l);
        }
        false
    }

    /// Emit a loop gate: wraps `child_body` in the appropriate control structure
    /// and appends the result + swap code to `output`.
    ///
    /// - **Root-level loops** (no parent) are fused with the tick: they emit an `if`
    ///   so the body runs at most once per tick when their entry condition is met.
    /// - **Nested loops** (have a parent loop) emit a `while` so they can iterate
    ///   until fixpoint (driven by `defer_tick` back-edges).
    /// - If there are no gate checks, the body is emitted unconditionally.
    fn emit_loop_gate(
        &self,
        loop_id: GraphLoopId,
        child_body: TokenStream,
        loop_input_handoffs: &SecondaryMap<GraphLoopId, Vec<GraphNodeId>>,
        back_edge_hoffs_and_lazyness: &SparseSecondaryMap<GraphNodeId, bool>,
        loop_swap_code: &std::collections::HashMap<GraphLoopId, Vec<TokenStream>>,
        output: &mut TokenStream,
    ) {
        // Get swap code for this loop's defer_tick handoffs.
        let swap_code = loop_swap_code
            .get(&loop_id)
            .map(|v| v.as_slice())
            .unwrap_or(&[]);

        // Root-level loops are fused with the tick: emit `if` instead of `while`.
        let is_root_loop = self.loop_parent(loop_id).is_none();

        // Build the gate condition from entry handoffs (excluding lazy windowing operators).
        let entry_handoffs = loop_input_handoffs.get(loop_id).expect("loop missing");
        let mut gate_checks: Vec<TokenStream> = entry_handoffs
            .iter()
            .filter(|&&hoff_id| {
                // Check if the successor (windowing operator) is lazy.
                // If so, exclude from the gate — it doesn't trigger the loop.
                let is_lazy = self
                    .node_successors(hoff_id)
                    .next()
                    .and_then(|(_, succ)| self.node_op_inst(succ))
                    .is_some_and(|op_inst| {
                        op_inst.op_constraints.flo_type == Some(FloType::WindowingLazy)
                    });
                !is_lazy
            })
            .map(|&hoff_id| {
                let span = self.node(hoff_id).span();
                let buf_ident = self.hoff_buf_ident(hoff_id, span);
                if back_edge_hoffs_and_lazyness.contains_key(hoff_id) {
                    let back_ident = self.hoff_back_ident(hoff_id, span);
                    quote_spanned! {span=> !#back_ident.is_empty() }
                } else {
                    quote_spanned! {span=> !#buf_ident.is_empty() }
                }
            })
            .collect();

        // Non-lazy defer_tick back-buffers also contribute to the gate (nested loops only).
        if !is_root_loop {
            for (hoff_id, hoff) in self.nodes() {
                if !matches!(hoff, GraphNode::Handoff { .. }) {
                    continue;
                }
                let Some(delay_type) = self.handoff_delay_type(hoff_id) else {
                    continue;
                };
                if delay_type != DelayType::Loop {
                    continue;
                }
                // Check this handoff belongs to loop_id.
                let hoff_loop = self
                    .node_successors(hoff_id)
                    .next()
                    .and_then(|(_, succ)| self.node_subgraph(succ))
                    .and_then(|sg| self.subgraph_loop(sg));
                if hoff_loop != Some(loop_id) {
                    continue;
                }
                let span = self.node(hoff_id).span();
                let back_ident = self.hoff_back_ident(hoff_id, span);
                gate_checks.push(quote_spanned! {span=> !#back_ident.is_empty() });
            }
        }

        // For root-level loops: non-lazy defer_tick back-buffers also contribute to the gate.
        // This ensures the loop fires on the next tick when data was deferred via defer_tick.
        if is_root_loop {
            for (hoff_id, hoff) in self.nodes() {
                if !matches!(hoff, GraphNode::Handoff { .. }) {
                    continue;
                }
                let Some(delay_type) = self.handoff_delay_type(hoff_id) else {
                    continue;
                };
                if delay_type != DelayType::Tick {
                    continue;
                }
                // Check this handoff's consumer is inside this root-level loop.
                let hoff_loop = self
                    .node_successors(hoff_id)
                    .next()
                    .and_then(|(_, succ)| self.node_subgraph(succ))
                    .and_then(|sg| self.subgraph_loop(sg));
                if hoff_loop != Some(loop_id) {
                    continue;
                }
                let span = self.node(hoff_id).span();
                let back_ident = self.hoff_back_ident(hoff_id, span);
                gate_checks.push(quote_spanned! {span=> !#back_ident.is_empty() });
            }
        }

        // An eager windowing operator (`batch_eager()`) forces the loop to fire unconditionally,
        // even when its windowed input is empty. It is only valid at the entry of a root-level
        // loop (disallowed in nested loops during validation, since forcing a nested loop to
        // always fire would prevent its fixpoint iteration from terminating).
        let has_eager = entry_handoffs.iter().any(|&hoff_id| {
            self.node_successors(hoff_id)
                .next()
                .and_then(|(_, succ)| self.node_op_inst(succ))
                .is_some_and(|op_inst| {
                    op_inst.op_constraints.flo_type == Some(FloType::WindowingEager)
                })
        });

        if has_eager && is_root_loop {
            // Eager entry: always run the loop body (gate forced true).
            output.extend(child_body);
            output.extend(quote! { #( #swap_code )* });
        } else if gate_checks.is_empty() {
            // No entry handoffs — always run.
            output.extend(child_body);
            output.extend(quote! { #( #swap_code )* });
        } else if is_root_loop {
            // Root-level loop: fused with tick, fire at most once.
            output.extend(quote! {
                #[allow(clippy::nonminimal_bool, reason = "codegen")]
                if false #( || #gate_checks )* {
                    #child_body
                    #( #swap_code )*
                }
            });
        } else {
            // Nested loop: iterate until fixpoint.
            output.extend(quote! {
                #[allow(clippy::nonminimal_bool, reason = "codegen")]
                while false #( || #gate_checks )* {
                    #child_body
                    #( #swap_code )*
                }
            });
        }
    }

    /// Compute the input handoffs into each loop (predecessor outside, successor inside).
    fn helper_loop_input_handoffs(&self) -> SecondaryMap<GraphLoopId, Vec<GraphNodeId>> {
        let mut loop_hoffs_inn = SecondaryMap::<GraphLoopId, Vec<GraphNodeId>>::new();

        // Check each handoff node.
        for (hoff_id, hoff) in self.nodes() {
            if !matches!(hoff, GraphNode::Handoff { .. }) {
                continue;
            }

            // Get the loop context of the predecessor and successor.
            let loop_pred = self
                .node_predecessors(hoff_id)
                .next()
                .and_then(|(_, pred)| self.node_loop(pred));
            let loop_succ = self
                .node_successors(hoff_id)
                .next()
                .and_then(|(_, succ)| self.node_loop(succ));

            if let Some(loop_succ) = loop_succ
                && loop_pred == self.loop_parent(loop_succ)
            {
                // Pred is parent/outer loop of succ.
                loop_hoffs_inn
                    .entry(loop_succ)
                    .expect("loop removed")
                    .or_default()
                    .push(hoff_id);
            }
        }

        loop_hoffs_inn
    }

    /// Emit this graph as runnable Rust source code tokens that execute inline.
    /// Generates a flat `async move |df: &mut Context|` closure where subgraph
    /// blocks are inlined in topological order, using local `Vec<T>` buffers
    /// instead of runtime handoffs. Each call to the closure runs one tick.
    ///
    /// The generated code block evaluates to a `Dfir` instance wrapping the
    /// closure. Operator prologues run at construction time on the `Context`
    /// before it is moved into `Dfir::new`. `Dfir` provides the `Context`
    /// to the closure on each tick run.
    ///
    /// # Errors
    ///
    /// Returns all diagnostics as `Err(diagnostics)` if any are errors
    /// (leaving `&mut diagnostics` empty).
    pub fn as_code(
        &self,
        root: &TokenStream,
        include_type_guards: bool,
        prefix: TokenStream,
        diagnostics: &mut Diagnostics,
    ) -> Result<TokenStream, Diagnostics> {
        self.as_code_with_options(root, include_type_guards, true, prefix, diagnostics)
    }

    /// Like [`Self::as_code`], but with `include_meta` controlling whether
    /// the runtime meta graph + diagnostics JSON blobs are baked into the
    /// generated `Dfir::new(...)` call.
    ///
    /// The simulator calls Dfir::new() on each iteration, and as a part of that
    /// it does parsing of the metagraph and diganostics blob. One of them causes spans to get allocated,
    /// each time a span is allocated, some threadlocal u32 is being incremented, and, on a long simulator run,
    /// the u32 overflows and panics.
    pub fn as_code_with_options(
        &self,
        root: &TokenStream,
        include_type_guards: bool,
        include_meta: bool,
        prefix: TokenStream,
        diagnostics: &mut Diagnostics,
    ) -> Result<TokenStream, Diagnostics> {
        let df = Ident::new(GRAPH, Span::call_site());
        let context = Ident::new(CONTEXT, Span::call_site());
        // Tick-local bump-allocated Vec handoff declarations (inside the tick closure).
        let bump_ident = Ident::new("__dfir_bump", Span::call_site());

        // 1. Collect all handoff nodes.
        let handoff_nodes = self
            .nodes
            .iter()
            .filter_map(|(node_id, node)| match node {
                &GraphNode::Handoff {
                    kind,
                    src_span,
                    dst_span,
                } => Some((node_id, kind, (src_span, dst_span))),
                GraphNode::Operator(_) => None,
                GraphNode::ModuleBoundary { .. } => panic!(),
            })
            .collect::<Vec<_>>();

        // Determine which handoff nodes are tick-boundary (defer_tick) back-edges.
        // These must remain as captured Vec<T> since they persist across ticks.
        // All other Vec handoffs will be bump-allocated (tick-local).
        let back_edge_hoffs_and_lazyness = handoff_nodes
            .iter()
            .map(|&(node_id, _, _)| node_id)
            .filter_map(|node_id| {
                let delay_type = self.handoff_delay_type(node_id)?;
                Some((
                    node_id,
                    matches!(delay_type, DelayType::TickLazy | DelayType::LoopLazy),
                ))
            })
            .collect::<SparseSecondaryMap<_, _>>();

        // Back buffer idents, buf idents, and if they are lazy.
        let back_buffer_idents_laziness = handoff_nodes
            .iter()
            .filter_map(|&(hoff_id, _kind, (src_span, dst_span))| {
                back_edge_hoffs_and_lazyness.get(hoff_id).map(|&is_lazy| {
                    let span = src_span.join(dst_span).unwrap_or(src_span);
                    let back_ident = self.hoff_back_ident(hoff_id, span);
                    let buf_ident = self.hoff_buf_ident(hoff_id, span);
                    (back_ident, buf_ident, is_lazy)
                })
            })
            .collect::<Vec<_>>();

        // Generate swap code for tick-boundary (defer_tick / defer_tick_lazy) handoffs.
        // At the end of each tick, swap the regular buffer and back buffer so the
        // consumer reads last tick's data from the back buffer.
        // Only tick-level swaps go here; loop-level swaps are emitted inside the loop gate.
        // IMPORTANT: For defer_tick handoffs whose consumer is inside a root-level loop,
        // the swap is emitted inside the `if` gate (via loop_swap_code), not at tick level.
        let back_edge_swap_code = handoff_nodes
            .iter()
            .filter(|&&(node_id, _kind, _)| {
                self.handoff_delay_type(node_id)
                    .is_some_and(|dt| matches!(dt, DelayType::Tick | DelayType::TickLazy))
            })
            .filter(|&&(hoff_id, _kind, _)| {
                // Exclude handoffs whose consumer is inside a root-level loop.
                // Those get their swap emitted inside the loop gate.
                let consumer_loop = self
                    .node_successors(hoff_id)
                    .next()
                    .and_then(|(_, succ)| self.node_subgraph(succ))
                    .and_then(|sg| self.subgraph_loop(sg));
                if let Some(loop_id) = consumer_loop {
                    // If it's a root-level loop, don't include in tick-level swap.
                    self.loop_parent(loop_id).is_some()
                } else {
                    // No loop context: emit at tick level (original behavior).
                    true
                }
            })
            .map(|&(hoff_id, _kind, _)| {
                let span = self.nodes[hoff_id].span();
                let buf_ident = self.hoff_buf_ident(hoff_id, span);
                let back_ident = self.hoff_back_ident(hoff_id, span);
                quote_spanned! {span=>
                    ::std::mem::swap(&mut #buf_ident, &mut #back_ident);
                }
            })
            .collect::<Vec<_>>();

        // Collect per-loop swap code for defer_tick / defer_tick_lazy handoffs.
        // AND defer_tick / defer_tick_lazy handoffs inside root-level loops.
        // Keyed by the loop ID of the consumer (successor) of the handoff.
        let mut loop_swap_code: std::collections::HashMap<GraphLoopId, Vec<TokenStream>> =
            std::collections::HashMap::new();
        for &(hoff_id, _kind, _) in handoff_nodes.iter() {
            let Some(delay_type) = self.handoff_delay_type(hoff_id) else {
                continue;
            };
            // Find the loop this handoff belongs to (from its consumer's loop context).
            let loop_id = self
                .node_successors(hoff_id)
                .next()
                .and_then(|(_, succ)| self.node_subgraph(succ))
                .and_then(|sg| self.subgraph_loop(sg));
            let Some(loop_id) = loop_id else {
                continue;
            };
            let include = match delay_type {
                DelayType::Loop | DelayType::LoopLazy => true,
                DelayType::Tick | DelayType::TickLazy => {
                    // Only include in loop swap if this is a root-level loop.
                    self.loop_parent(loop_id).is_none()
                }
            };
            if !include {
                continue;
            }
            let span = self.nodes[hoff_id].span();
            let buf_ident = self.hoff_buf_ident(hoff_id, span);
            let back_ident = self.hoff_back_ident(hoff_id, span);
            loop_swap_code
                .entry(loop_id)
                .or_default()
                .push(quote_spanned! {span=>
                    ::std::mem::swap(&mut #buf_ident, &mut #back_ident);
                });
        }

        // 2. Collect per-subgraph recv & send handoffs.
        let subgraph_handoffs = self.helper_collect_subgraph_handoffs();

        // 3. Use pre-computed subgraph topological order.
        let all_subgraphs: Vec<_> = self
            .subgraph_toposort()
            .iter()
            .map(|&sg_id| (sg_id, self.subgraph(sg_id)))
            .collect();

        // TODO(mingwei): If a handoff has no pipe consumers we should drop it as soon as possible, after all reference
        // consumers. Right now we just let these handoffs die at the end of the tick.

        let mut op_prologue_code = Vec::new();
        let mut op_tick_end_code = Vec::new();

        // Stack-based hierarchical code generation.
        // Each entry is (loop_id, body_tokens) for an open loop context.
        // The "current output" is always the innermost open context (or root).
        let mut loop_stack: Vec<(GraphLoopId, TokenStream)> = Vec::new();
        let mut current_output = TokenStream::new();

        // Pre-compute loop gate data.
        let loop_input_handoffs = self.helper_loop_input_handoffs();
        let loop_output_handoffs = self.helper_loop_output_handoffs();

        {
            for &(subgraph_id, subgraph_nodes) in all_subgraphs.iter() {
                let sg_loop = self.subgraph_loop(subgraph_id);

                // Transition loop contexts: close loops we've exited, open loops we've entered.
                // Close loops until we're at the right level.
                while let Some(&(top_loop, _)) = loop_stack.last() {
                    if sg_loop == Some(top_loop) || self.is_inside_loop(sg_loop, top_loop) {
                        break;
                    }
                    // Pop: wrap the body in a loop gate and append to parent.
                    let (closed_loop, child_body) = loop_stack.pop().unwrap();
                    let target = if let Some((_, parent_body)) = loop_stack.last_mut() {
                        parent_body
                    } else {
                        &mut current_output
                    };
                    self.emit_loop_gate(
                        closed_loop,
                        child_body,
                        &loop_input_handoffs,
                        &back_edge_hoffs_and_lazyness,
                        &loop_swap_code,
                        target,
                    );
                }

                // Open new loops if we've descended.
                if let Some(target_loop) = sg_loop
                    && loop_stack.last().map(|&(l, _)| l) != Some(target_loop)
                {
                    // Find the path of loops to open (from outermost to target).
                    let mut path = Vec::new();
                    let mut cur = Some(target_loop);
                    while let Some(l) = cur {
                        if loop_stack.last().map(|&(top, _)| top) == Some(l) {
                            break;
                        }
                        path.push(l);
                        cur = self.loop_parent(l);
                    }
                    // Push in outermost-first order, emitting exit-handoff declarations
                    // to the parent level before the while loop.
                    for &loop_id in path.iter().rev() {
                        // Declare exit-handoff buffers at the current (parent) level.
                        if let Some(exit_hoffs) = loop_output_handoffs.get(loop_id) {
                            let exit_hoff_decls = exit_hoffs.iter().map(|&hoff_id| {
                                let span = self.nodes[hoff_id].span();
                                let buf_ident = self.hoff_buf_ident(hoff_id, span);
                                let GraphNode::Handoff { kind, .. } = self.node(hoff_id) else {
                                    panic!()
                                };
                                match kind {
                                    HandoffKind::Vec => quote_spanned! {span=>
                                        let mut #buf_ident = #root::bumpalo::collections::Vec::new_in(&#bump_ident);
                                    },
                                    HandoffKind::Singleton | HandoffKind::Optional => quote_spanned! {span=>
                                        let mut #buf_ident = ::std::option::Option::None;
                                    },
                                }
                            });
                            let target = if let Some((_, body)) = loop_stack.last_mut() {
                                body
                            } else {
                                &mut current_output
                            };
                            target.extend(quote! { #( #exit_hoff_decls )* });
                        }
                        loop_stack.push((loop_id, TokenStream::new()));
                    }
                }
                let sg_metrics_ffi = subgraph_id.data().as_ffi();
                let (recv_hoffs, send_hoffs) = &subgraph_handoffs[subgraph_id];

                // Generate buffer ident helpers for this subgraph's handoffs.
                let recv_port_idents: Vec<Ident> = recv_hoffs
                    .iter()
                    .map(|&hoff_id| self.node_as_ident(hoff_id, true))
                    .collect();
                let send_port_idents: Vec<Ident> = send_hoffs
                    .iter()
                    .map(|&hoff_id| self.node_as_ident(hoff_id, false))
                    .collect();

                // Map handoff node IDs to buffer idents.
                let recv_buf_idents: Vec<Ident> = recv_hoffs
                    .iter()
                    .map(|&hoff_id| self.hoff_buf_ident(hoff_id, self.nodes[hoff_id].span()))
                    .collect();
                let send_buf_idents: Vec<Ident> = send_hoffs
                    .iter()
                    .map(|&hoff_id| self.hoff_buf_ident(hoff_id, self.nodes[hoff_id].span()))
                    .collect();

                // Handoff kinds
                let recv_kinds = recv_hoffs
                    .iter()
                    .map(|&hoff_id| {
                        let GraphNode::Handoff { kind, .. } = self.node(hoff_id) else {
                            panic!()
                        };
                        *kind
                    })
                    .collect::<Vec<_>>();
                let send_kinds = send_hoffs
                    .iter()
                    .map(|&hoff_id| {
                        let GraphNode::Handoff { kind, .. } = self.node(hoff_id) else {
                            panic!()
                        };
                        *kind
                    })
                    .collect::<Vec<_>>();

                // Recv port code: drain from buffer into iterator, tracking if non-empty.
                // For back-edge (defer_tick) handoffs, drain from the back buffer instead.
                // Also update handoff metrics (measured at recv, not send — see graph.rs).
                let recv_port_code: Vec<TokenStream> = recv_port_idents
                    .iter()
                    .zip(recv_buf_idents.iter())
                    .zip(recv_kinds.iter())
                    .zip(recv_hoffs.iter())
                    .map(|(((port_ident, buf_ident), &kind), &hoff_id)| {
                        let hoff_ffi = hoff_id.data().as_ffi();
                        // Use call_site span for internal identifiers to avoid
                        // hygiene issues when invoked through declarative macros
                        // (e.g. dfir_expect_warnings!). TODO(#2781): define these once.
                        let work_done = Ident::new("__dfir_work_done", Span::call_site());
                        let metrics = Ident::new("__dfir_metrics", Span::call_site());

                        // Compute len and drain expressions based on handoff kind.
                        let (len_expr, drain_expr) = match kind {
                            HandoffKind::Singleton | HandoffKind::Optional => (
                                quote! { if #buf_ident.is_some() { 1usize } else { 0usize } },
                                quote! { #root::dfir_pipes::pull::iter(#buf_ident.take().into_iter()) },
                            ),
                            HandoffKind::Vec => {
                                // Special asymmetric handling for defer tick handoffs, which are double-buffered.
                                // The producer writes to the regular buffer; at end-of-tick the buffers are swapped,
                                // so the consumer drains from the back buffer (here).
                                let drain_ident = if back_edge_hoffs_and_lazyness.contains_key(hoff_id) {
                                    &self.hoff_back_ident(hoff_id, buf_ident.span())
                                } else {
                                    buf_ident
                                };
                                (
                                    quote! { #drain_ident.len() },
                                    quote! { #root::dfir_pipes::pull::iter(#drain_ident.drain(..)) },
                                )
                            }
                        };

                        quote_spanned! {port_ident.span()=>
                            {
                                let hoff_len = #len_expr;
                                if hoff_len > 0 {
                                    #work_done = true;
                                }
                                let hoff_metrics = &#metrics.handoffs[
                                    #root::slotmap::KeyData::from_ffi(#hoff_ffi).into()
                                ];
                                hoff_metrics.total_items_count.update(|x| x + hoff_len);
                                hoff_metrics.curr_items_count.set(hoff_len);
                            }
                            let #port_ident = #drain_expr;
                        }
                    })
                    .collect();

                // Send port code: push into buffer.
                let send_port_code: Vec<TokenStream> = send_port_idents
                    .iter()
                    .zip(send_buf_idents.iter())
                    .zip(send_kinds.iter())
                    .map(|((port_ident, buf_ident), &kind)| {
                        match kind {
                            HandoffKind::Singleton => {
                                // Singleton slot: store exactly one item, panic on duplicate.
                                quote_spanned! {port_ident.span()=>
                                    let #port_ident = #root::dfir_pipes::push::for_each(|__item| {
                                        if #buf_ident.replace(__item).is_some() {
                                            panic!("singleton() received more than one item");
                                        }
                                    });
                                }
                            }
                            HandoffKind::Optional => {
                                // Optional slot: store at most one item, panic on duplicate.
                                quote_spanned! {port_ident.span()=>
                                    let #port_ident = #root::dfir_pipes::push::for_each(|__item| {
                                        if #buf_ident.replace(__item).is_some() {
                                            panic!("optional() received more than one item");
                                        }
                                    });
                                }
                            }
                            HandoffKind::Vec => {
                                quote_spanned! {port_ident.span()=>
                                    // TODO(mingwei): use `#root::dfir_pipes::push::vec_push`?
                                    let #port_ident = #root::dfir_pipes::push::for_each(|item| { #buf_ident.push(item); });
                                }
                            }
                        }
                    })
                    .collect();

                // All nodes in a subgraph should be in the same loop.
                let loop_id = self.node_loop(subgraph_nodes[0]);

                let mut subgraph_op_iter_code = Vec::new();
                let mut subgraph_op_iter_after_code = Vec::new();
                {
                    let pull_to_push_idx = self.find_pull_to_push_idx(subgraph_nodes);

                    let (pull_half, push_half) = subgraph_nodes.split_at(pull_to_push_idx);
                    let nodes_iter = pull_half.iter().chain(push_half.iter().rev());

                    for (idx, &node_id) in nodes_iter.enumerate() {
                        let node = &self.nodes[node_id];
                        assert!(
                            matches!(node, GraphNode::Operator(_)),
                            "Handoffs are not part of subgraphs."
                        );
                        let op_inst = &self.operator_instances[node_id];

                        let op_span = node.span();
                        let op_name = op_inst.op_constraints.name;
                        // Use op's span for root. #root is expected to be correct, any errors should span back to the op gen.
                        let root = change_spans(root.clone(), op_span);
                        let op_constraints = OPERATORS
                            .iter()
                            .find(|op| op_name == op.name)
                            .unwrap_or_else(|| panic!("Failed to find op: {}", op_name));

                        let ident = self.node_as_ident(node_id, false);

                        {
                            // TODO clean this up.
                            // Collect input arguments (predecessors).
                            let mut input_edges = self
                                .graph
                                .predecessor_edges(node_id)
                                .map(|edge_id| (self.edge_ports(edge_id).1, edge_id))
                                .collect::<Vec<_>>();
                            // Ensure sorted by port index.
                            input_edges.sort();

                            let inputs = input_edges
                                .iter()
                                .map(|&(_port, edge_id)| {
                                    let (pred, _) = self.edge(edge_id);
                                    self.node_as_ident(pred, true)
                                })
                                .collect::<Vec<_>>();

                            // Collect output arguments (successors).
                            let mut output_edges = self
                                .graph
                                .successor_edges(node_id)
                                .map(|edge_id| (&self.ports[edge_id].0, edge_id))
                                .collect::<Vec<_>>();
                            // Ensure sorted by port index.
                            output_edges.sort();

                            let outputs = output_edges
                                .iter()
                                .map(|&(_port, edge_id)| {
                                    let (_, succ) = self.edge(edge_id);
                                    self.node_as_ident(succ, false)
                                })
                                .collect::<Vec<_>>();

                            let is_pull = idx < pull_to_push_idx;

                            // There's a bit of dark magic hidden in `Span`s... you'd think it's just a `file:line:column`,
                            // but it has one extra bit of info for _name resolution_, used for `Ident`s. `Span::call_site()`
                            // has the (unhygienic) resolution we want, an ident is just solely determined by its string name,
                            // which is what you'd expect out of unhygienic proc macros like this. Meanwhile, declarative macros
                            // use `Span::mixed_site()` which is weird and I don't understand it. It turns out that if you call
                            // the dfir syntax proc macro from _within_ a declarative macro then `op_span` will have the
                            // bad `Span::mixed_site()` name resolution and cause "Cannot find value `df/context`" errors. So
                            // we call `.resolved_at()` to fix resolution back to `Span::call_site()`. -Mingwei
                            let df_local = &Ident::new(GRAPH, op_span.resolved_at(df.span()));
                            let context = &Ident::new(CONTEXT, op_span.resolved_at(context.span()));

                            let singletons_resolved =
                                self.helper_resolve_singletons(node_id, op_span);

                            let arguments = &process_singletons::postprocess_singletons(
                                op_inst.arguments_raw.clone(),
                                singletons_resolved,
                            );

                            let source_tag = 'a: {
                                if let Some(tag) = self.operator_tag.get(node_id).cloned() {
                                    break 'a tag;
                                }

                                if proc_macro::is_available() {
                                    let op_span = op_span.unwrap();
                                    break 'a format!(
                                        "loc_{}_{}_{}_{}_{}",
                                        crate::pretty_span::make_source_path_relative(
                                            &op_span.file()
                                        )
                                        .display()
                                        .to_string()
                                        .replace(|x: char| !x.is_ascii_alphanumeric(), "_"),
                                        op_span.start().line(),
                                        op_span.start().column(),
                                        op_span.end().line(),
                                        op_span.end().column(),
                                    );
                                }

                                format!(
                                    "loc_nopath_{}_{}_{}_{}",
                                    op_span.start().line,
                                    op_span.start().column,
                                    op_span.end().line,
                                    op_span.end().column
                                )
                            };

                            let work_fn = format_ident!(
                                "{}__{}__{}",
                                ident,
                                op_name,
                                source_tag,
                                span = op_span
                            );
                            let work_fn_async = format_ident!("{}__async", work_fn, span = op_span);

                            let context_args = WriteContextArgs {
                                root: &root,
                                df_ident: df_local,
                                context,
                                subgraph_id,
                                node_id,
                                loop_id,
                                op_span,
                                op_tag: self.operator_tag.get(node_id).cloned(),
                                work_fn: &work_fn,
                                work_fn_async: &work_fn_async,
                                ident: &ident,
                                is_pull,
                                inputs: &inputs,
                                outputs: &outputs,
                                op_name,
                                op_inst,
                                arguments,
                            };

                            let write_result =
                                (op_constraints.write_fn)(&context_args, diagnostics);
                            let OperatorWriteOutput {
                                write_prologue,
                                write_iterator,
                                write_iterator_after,
                                write_tick_end,
                            } = write_result.unwrap_or_else(|()| {
                                assert!(
                                    diagnostics.has_error(),
                                    "Operator `{}` returned `Err` but emitted no diagnostics, this is a bug.",
                                    op_name,
                                );
                                OperatorWriteOutput {
                                    write_iterator: null_write_iterator_fn(&context_args),
                                    ..Default::default()
                                }
                            });

                            op_prologue_code.push(syn::parse_quote! {
                                #[allow(dead_code, non_snake_case, reason = "codegen")]
                                #[inline(always)]
                                fn #work_fn<T>(thunk: impl ::std::ops::FnOnce() -> T) -> T {
                                    thunk()
                                }

                                #[allow(dead_code, non_snake_case, reason = "codegen")]
                                #[inline(always)]
                                async fn #work_fn_async<T>(
                                    thunk: impl ::std::future::Future<Output = T>,
                                ) -> T {
                                    thunk.await
                                }
                            });
                            op_prologue_code.push(write_prologue);
                            op_tick_end_code.push(write_tick_end);
                            subgraph_op_iter_code.push(write_iterator);

                            if include_type_guards {
                                let type_guard = if is_pull {
                                    quote_spanned! {op_span=>
                                        let #ident = {
                                            #[allow(non_snake_case)]
                                            #[inline(always)]
                                            pub fn #work_fn<Item, Input>(input: Input)
                                                -> impl #root::dfir_pipes::pull::Pull<Item = Item, Meta = (), CanPend = Input::CanPend, CanEnd = Input::CanEnd>
                                            where
                                                Input: #root::dfir_pipes::pull::Pull<Item = Item, Meta = ()>,
                                            {
                                                #root::pin_project_lite::pin_project! {
                                                    #[repr(transparent)]
                                                    struct Pull<Item, Input: #root::dfir_pipes::pull::Pull<Item = Item>> {
                                                        #[pin]
                                                        inner: Input
                                                    }
                                                }

                                                impl<Item, Input> #root::dfir_pipes::pull::Pull for Pull<Item, Input>
                                                where
                                                    Input: #root::dfir_pipes::pull::Pull<Item = Item>,
                                                {
                                                    type Ctx<'ctx> = Input::Ctx<'ctx>;

                                                    type Item = Item;
                                                    type Meta = Input::Meta;
                                                    type CanPend = Input::CanPend;
                                                    type CanEnd = Input::CanEnd;

                                                    #[inline(always)]
                                                    fn pull(
                                                        self: ::std::pin::Pin<&mut Self>,
                                                        ctx: &mut Self::Ctx<'_>,
                                                    ) -> #root::dfir_pipes::pull::PullStep<Self::Item, Self::Meta, Self::CanPend, Self::CanEnd> {
                                                        #root::dfir_pipes::pull::Pull::pull(self.project().inner, ctx)
                                                    }

                                                    #[inline(always)]
                                                    fn size_hint(&self) -> (usize, Option<usize>) {
                                                        #root::dfir_pipes::pull::Pull::size_hint(&self.inner)
                                                    }
                                                }

                                                Pull {
                                                    inner: input
                                                }
                                            }
                                            #work_fn::<_, _>( #ident )
                                        };
                                    }
                                } else {
                                    quote_spanned! {op_span=>
                                        let #ident = {
                                            #[allow(non_snake_case)]
                                            #[inline(always)]
                                            pub fn #work_fn<Item, Psh>(psh: Psh) -> impl #root::dfir_pipes::push::Push<Item, (), CanPend = Psh::CanPend>
                                            where
                                                Psh: #root::dfir_pipes::push::Push<Item, ()>
                                            {
                                                #root::pin_project_lite::pin_project! {
                                                    #[repr(transparent)]
                                                    struct PushGuard<Psh> {
                                                        #[pin]
                                                        inner: Psh,
                                                    }
                                                }

                                                impl<Item, Psh> #root::dfir_pipes::push::Push<Item, ()> for PushGuard<Psh>
                                                where
                                                    Psh: #root::dfir_pipes::push::Push<Item, ()>,
                                                {
                                                    type Ctx<'ctx> = Psh::Ctx<'ctx>;

                                                    type CanPend = Psh::CanPend;

                                                    #[inline(always)]
                                                    fn poll_ready(
                                                        self: ::std::pin::Pin<&mut Self>,
                                                        ctx: &mut Self::Ctx<'_>,
                                                    ) -> #root::dfir_pipes::push::PushStep<Self::CanPend> {
                                                        #root::dfir_pipes::push::Push::poll_ready(self.project().inner, ctx)
                                                    }

                                                    #[inline(always)]
                                                    fn start_send(
                                                        self: ::std::pin::Pin<&mut Self>,
                                                        item: Item,
                                                        meta: (),
                                                    ) {
                                                        #root::dfir_pipes::push::Push::start_send(self.project().inner, item, meta)
                                                    }

                                                    #[inline(always)]
                                                    fn poll_finalize(
                                                        self: ::std::pin::Pin<&mut Self>,
                                                        ctx: &mut Self::Ctx<'_>,
                                                    ) -> #root::dfir_pipes::push::PushStep<Self::CanPend> {
                                                        #root::dfir_pipes::push::Push::poll_finalize(self.project().inner, ctx)
                                                    }

                                                    #[inline(always)]
                                                    fn size_hint(
                                                        self: ::std::pin::Pin<&mut Self>,
                                                        hint: (usize, Option<usize>),
                                                    ) {
                                                        #root::dfir_pipes::push::Push::size_hint(self.project().inner, hint)
                                                    }
                                                }

                                                PushGuard {
                                                    inner: psh
                                                }
                                            }
                                            #work_fn( #ident )
                                        };
                                    }
                                };
                                subgraph_op_iter_code.push(type_guard);
                            }
                            subgraph_op_iter_after_code.push(write_iterator_after);
                        }
                    }

                    {
                        // Determine pull and push halves of the `Pivot`.
                        let pull_ident = if 0 < pull_to_push_idx {
                            self.node_as_ident(subgraph_nodes[pull_to_push_idx - 1], false)
                        } else {
                            // Entire subgraph is push (with a single recv/pull handoff input).
                            recv_port_idents[0].clone()
                        };

                        #[rustfmt::skip]
                        let push_ident = if let Some(&node_id) =
                            subgraph_nodes.get(pull_to_push_idx)
                        {
                            self.node_as_ident(node_id, false)
                        } else if 1 == send_port_idents.len() {
                            // Entire subgraph is pull (with a single send/push handoff output).
                            send_port_idents[0].clone()
                        } else {
                            diagnostics.push(Diagnostic::spanned(
                                pull_ident.span(),
                                Level::Error,
                                "Degenerate subgraph detected, is there a disconnected `null()` or other degenerate pipeline somewhere?",
                            ));
                            continue;
                        };

                        // Pivot span is combination of pull and push spans (or if not possible, just take the push).
                        let pivot_span = pull_ident
                            .span()
                            .join(push_ident.span())
                            .unwrap_or_else(|| push_ident.span());
                        let pivot_fn_ident = Ident::new(
                            &format!("pivot_run_sg_{:?}", subgraph_id.data()),
                            pivot_span,
                        );
                        let root = change_spans(root.clone(), pivot_span);
                        subgraph_op_iter_code.push(quote_spanned! {pivot_span=>
                            #[inline(always)]
                            fn #pivot_fn_ident<Pul, Psh, Item>(pull: Pul, push: Psh)
                                -> impl ::std::future::Future<Output = ()>
                            where
                                Pul: #root::dfir_pipes::pull::Pull<Item = Item>,
                                Psh: #root::dfir_pipes::push::Push<Item, Pul::Meta>,
                            {
                                #root::dfir_pipes::pull::Pull::send_push(pull, push)
                            }
                            (#pivot_fn_ident)(#pull_ident, #push_ident).await;
                        });
                    }
                };

                // Each subgraph block is an async block so it can be individually instrumented.
                // Note: this ident is for the subgraph future, not a runtime SubgraphId binding
                // (unlike the scheduled path's `sg_ident`).
                let sg_fut_ident = subgraph_id.as_ident(Span::call_site());

                // Generate send-side curr_items_count updates (after subgraph runs).
                let send_metrics_code = send_hoffs
                    .iter()
                    .zip(send_buf_idents.iter())
                    .zip(send_kinds.iter())
                    .map(|((&hoff_id, buf_ident), &kind)| {
                        let hoff_ffi = hoff_id.data().as_ffi();
                        let len_expr = match kind {
                            HandoffKind::Singleton | HandoffKind::Optional => {
                                quote! { if #buf_ident.is_some() { 1 } else { 0 } }
                            }
                            HandoffKind::Vec => {
                                quote! { #buf_ident.len() }
                            }
                        };
                        quote! {
                            __dfir_metrics.handoffs[
                                #root::slotmap::KeyData::from_ffi(#hoff_ffi).into()
                            ].curr_items_count.set(#len_expr);
                        }
                    })
                    .collect::<Vec<_>>();

                // Create the handoffs we are about to push to (send).
                // Exit handoffs (sender inside a loop, receiver in parent) are already declared
                // before the while loop, so skip them here.
                let send_hoff_make_code = send_buf_idents.iter()
                    .zip(send_kinds.iter())
                    .zip(send_hoffs.iter())
                    .filter_map(|((buf_ident, &kind), &hoff_id)| {
                        let span = buf_ident.span();
                        if back_edge_hoffs_and_lazyness.contains_key(hoff_id) {
                            // Defer_tick send buffers are declared outside the tick closure
                            // as std::vec::Vec for O(1) swap. Just clear here.
                            Some(quote_spanned! {span=>
                                #buf_ident.clear();
                            })
                        } else {
                            // Check if this is a loop-exit handoff: sender is in a loop,
                            // receiver is in the parent (already declared outside the while loop).
                            let receiver_loop = self
                                .node_successors(hoff_id)
                                .next()
                                .and_then(|(_, succ)| self.node_loop(succ));
                            let is_exit = if let Some(sender_loop) = sg_loop {
                                receiver_loop == self.loop_parent(sender_loop)
                            } else {
                                false
                            };
                            if is_exit {
                                // Exit handoff: buffer already declared at parent level.
                                None
                            } else {
                                Some(match kind {
                                    HandoffKind::Vec => quote_spanned! {span=>
                                        let mut #buf_ident = #root::bumpalo::collections::Vec::new_in(&#bump_ident);
                                    },
                                    HandoffKind::Singleton | HandoffKind::Optional => quote_spanned! {span=>
                                        let mut #buf_ident = ::std::option::Option::None;
                                    },
                                })
                            }
                        }
                    })
                    .collect::<Vec<_>>();
                // Drop the handoffs we just drained (recv).
                // TODO(mingwei): we could use `.into_iter()` instead of `.drain(..)` to consume the handoffs directly.
                // This only works for handoffs within the tick, though, not `defer_tick` handoffs.
                let recv_hoff_drop_code = recv_buf_idents
                    .iter()
                    .zip(recv_hoffs.iter())
                    .filter(|&(_, &hoff_id)| !back_edge_hoffs_and_lazyness.contains_key(hoff_id))
                    .map(|(buf_ident, _)| {
                        let span = buf_ident.span();
                        quote_spanned! {span=>
                            let _ = #buf_ident;
                        }
                    });

                // Emit subgraph block to the current loop level (top of stack or root).
                let sg_block = quote! {
                    // Create the handoffs we are about to push to (send).
                    #( #send_hoff_make_code )*

                    let #sg_fut_ident = async {
                        let #context = &#df;
                        #( #recv_port_code )*
                        #( #send_port_code )*
                        #( #subgraph_op_iter_code )*
                        #( #subgraph_op_iter_after_code )*
                    };
                    {
                        // Instrument w/ the subgraph metrics.
                        let sg_metrics = &__dfir_metrics.subgraphs[
                            #root::slotmap::KeyData::from_ffi(#sg_metrics_ffi).into()
                        ];
                        #root::scheduled::metrics::InstrumentSubgraph::new(
                            #sg_fut_ident, sg_metrics
                        ).await;
                        sg_metrics.total_run_count.update(|x| x + 1);

                        // Update send (output) handoff metrics.
                        #( #send_metrics_code )*

                        // Drop the handoffs we just drained (recv).
                        #( #recv_hoff_drop_code )*
                    }
                };
                if let Some((_, body)) = loop_stack.last_mut() {
                    body.extend(sg_block);
                } else {
                    current_output.extend(sg_block);
                }
            }
        }

        // Close any remaining open loops.
        let gated_subgraph_code = {
            while let Some((closed_loop, child_body)) = loop_stack.pop() {
                let target = if let Some((_, parent_body)) = loop_stack.last_mut() {
                    parent_body
                } else {
                    &mut current_output
                };
                self.emit_loop_gate(
                    closed_loop,
                    child_body,
                    &loop_input_handoffs,
                    &back_edge_hoffs_and_lazyness,
                    &loop_swap_code,
                    target,
                );
            }
            current_output
        };

        if diagnostics.has_error() {
            return Err(std::mem::take(diagnostics));
        }
        let _ = diagnostics; // Ensure no more diagnostics may be added after checking for errors.

        let (meta_graph_arg, diagnostics_arg) = if include_meta {
            let meta_graph_json = serde_json::to_string(&self).unwrap();
            let meta_graph_json = Literal::string(&meta_graph_json);

            let serde_diagnostics: Vec<_> = diagnostics.iter().map(Diagnostic::to_serde).collect();
            let diagnostics_json = serde_json::to_string(&*serde_diagnostics).unwrap();
            let diagnostics_json = Literal::string(&diagnostics_json);

            (
                quote! { Some(#meta_graph_json) },
                quote! { Some(#diagnostics_json) },
            )
        } else {
            (quote! { None }, quote! { None })
        };

        // Generate metrics initialization: one entry per handoff and per subgraph.
        let metrics_init_code = {
            let handoff_inits = handoff_nodes.iter().map(|&(node_id, _, _)| {
                let ffi = node_id.data().as_ffi();
                quote! {
                    dfir_metrics.handoffs.insert(
                        #root::slotmap::KeyData::from_ffi(#ffi).into(),
                        ::std::default::Default::default(),
                    );
                }
            });
            let subgraph_inits = all_subgraphs.iter().map(|&(sg_id, _)| {
                let ffi = sg_id.data().as_ffi();
                quote! {
                    dfir_metrics.subgraphs.insert(
                        #root::slotmap::KeyData::from_ffi(#ffi).into(),
                        ::std::default::Default::default(),
                    );
                }
            });
            handoff_inits.chain(subgraph_inits).collect::<Vec<_>>()
        };

        // For creating back-buffer handoff vecs.
        let back_buffer_idents = back_buffer_idents_laziness
            .iter()
            .map(|(back_ident, _, _)| back_ident);
        // For creating the send-side buffer for defer_tick handoffs (also outside the closure).
        let defer_tick_buf_idents = back_buffer_idents_laziness
            .iter()
            .map(|(_, buf_ident, _)| buf_ident);
        // For checking if we should start the next tick (`schedule_subgraph`):
        // Collect the ident to check for each non-lazy back-edge handoff.
        // - For defer_tick handoffs in a root-level loop: check `back` (swap happened inside `if`)
        // - For all others: check `buf` (original behavior; tick-level swap hasn't happened yet)
        let non_lazy_schedule_idents: Vec<&Ident> = handoff_nodes
            .iter()
            .filter_map(|&(hoff_id, _, _)| {
                let delay_type = self.handoff_delay_type(hoff_id)?;
                // Only non-lazy.
                if matches!(delay_type, DelayType::TickLazy | DelayType::LoopLazy) {
                    return None;
                }
                let span = self.nodes[hoff_id].span();
                let expected_back_ident = self.hoff_back_ident(hoff_id, span);
                let entry = back_buffer_idents_laziness
                    .iter()
                    .find(|(back_ident, _, _)| *back_ident == expected_back_ident)?;

                // For defer_tick inside a root-level loop, check `back`.
                if delay_type == DelayType::Tick {
                    let consumer_loop = self
                        .node_successors(hoff_id)
                        .next()
                        .and_then(|(_, succ)| self.node_subgraph(succ))
                        .and_then(|sg| self.subgraph_loop(sg));
                    if consumer_loop.is_some_and(|lid| self.loop_parent(lid).is_none()) {
                        return Some(&entry.0); // back ident
                    }
                }
                Some(&entry.1) // buf ident
            })
            .collect();

        // Prologues and buffer declarations persist across ticks (outside the closure).
        // Subgraph blocks run each tick (inside the closure).
        Ok(quote! {
            {
                #prefix

                use #root::{var_expr, var_args};

                let __dfir_wake_state = ::std::sync::Arc::new(
                    #root::scheduled::context::WakeState::default()
                );

                let __dfir_metrics = {
                    let mut dfir_metrics = #root::scheduled::metrics::DfirMetrics::default();
                    #( #metrics_init_code )*
                    ::std::rc::Rc::new(dfir_metrics)
                };

                #[allow(unused_mut)]
                let mut #df = #root::scheduled::context::Context::new(
                    ::std::clone::Clone::clone(&__dfir_wake_state),
                    __dfir_metrics,
                );

                #( #op_prologue_code )*

                // For tick-boundary handoffs (`defer_tick` / `defer_tick_lazy`), declare both the
                // send buffer and the "back" buffer as std::vec::Vec outside the tick closure.
                // This enables O(1) mem::swap at end of tick for double-buffering.
                #( let mut #back_buffer_idents = ::std::vec::Vec::new(); )*
                #( let mut #defer_tick_buf_idents = ::std::vec::Vec::new(); )*

                // Bump allocator for handoffs (except for back-edge handoffs, above).
                let mut #bump_ident = #root::bumpalo::Bump::new();

                // Pre-set to true so the first tick always returns true
                // (matching Dfir pre-scheduling behavior). Subsequent ticks
                // start false (from take()) and are set true by recv port code
                // if any handoff buffer has data.
                let mut __dfir_work_done = true;
                #[allow(unused_qualifications, unused_mut, unused_variables, clippy::await_holding_refcell_ref, clippy::deref_addrof)]
                let __dfir_inline_tick = async move |#df: &mut #root::scheduled::context::Context| {
                    // Reset arena between ticks (start-of-tick)
                    #bump_ident.reset();

                    {
                        let __dfir_metrics = #df.metrics();

                        #gated_subgraph_code

                        // For non-lazy defer_tick: if any deferred buffer has data,
                        // signal that another tick should run.
                        #[allow(clippy::nonminimal_bool, reason = "codegen")]
                        if false #( || !#non_lazy_schedule_idents.is_empty() )* {
                            #df.schedule_subgraph(true);
                        }

                        // Double-buffer swap for defer_tick handoffs: move last tick's producer output (regular buffer)
                        // into the back buffer for the consumer to drain.
                        #( #back_edge_swap_code )*
                    }

                    // End-of-tick per-operator state handling (i.e. 'tick persistence).
                    #( #op_tick_end_code )*

                    #df.__end_tick();

                    ::std::mem::take(&mut __dfir_work_done)
                };
                #root::scheduled::context::Dfir::new(
                    __dfir_inline_tick,
                    #df,
                    #meta_graph_arg,
                    #diagnostics_arg,
                )
            }
        })
    }

    /// Color mode (pull vs. push, handoff vs. comp) for nodes. Some nodes can be push *OR* pull;
    /// those nodes will not be set in the returned map.
    pub fn node_color_map(&self) -> SparseSecondaryMap<GraphNodeId, Color> {
        let mut node_color_map: SparseSecondaryMap<GraphNodeId, Color> = self
            .node_ids()
            .filter_map(|node_id| {
                let op_color = self.node_color(node_id)?;
                Some((node_id, op_color))
            })
            .collect();

        // Fill in rest via subgraphs.
        for sg_nodes in self.subgraph_nodes.values() {
            let pull_to_push_idx = self.find_pull_to_push_idx(sg_nodes);

            for (idx, node_id) in sg_nodes.iter().copied().enumerate() {
                let is_pull = idx < pull_to_push_idx;
                node_color_map.insert(node_id, if is_pull { Color::Pull } else { Color::Push });
            }
        }

        node_color_map
    }

    /// Writes this graph as mermaid into a string.
    pub fn to_mermaid(&self, write_config: &WriteConfig) -> String {
        let mut output = String::new();
        self.write_mermaid(&mut output, write_config).unwrap();
        output
    }

    /// Writes this graph as mermaid into the given `Write`.
    pub fn write_mermaid(
        &self,
        output: impl std::fmt::Write,
        write_config: &WriteConfig,
    ) -> std::fmt::Result {
        let mut graph_write = Mermaid::new(output);
        self.write_graph(&mut graph_write, write_config)
    }

    /// Writes this graph as DOT (graphviz) into a string.
    pub fn to_dot(&self, write_config: &WriteConfig) -> String {
        let mut output = String::new();
        let mut graph_write = Dot::new(&mut output);
        self.write_graph(&mut graph_write, write_config).unwrap();
        output
    }

    /// Writes this graph as DOT (graphviz) into the given `Write`.
    pub fn write_dot(
        &self,
        output: impl std::fmt::Write,
        write_config: &WriteConfig,
    ) -> std::fmt::Result {
        let mut graph_write = Dot::new(output);
        self.write_graph(&mut graph_write, write_config)
    }

    /// Write out this graph using the given `GraphWrite`. E.g. `Mermaid` or `Dot.
    pub(crate) fn write_graph<W>(
        &self,
        mut graph_write: W,
        write_config: &WriteConfig,
    ) -> Result<(), W::Err>
    where
        W: GraphWrite,
    {
        fn helper_edge_label(
            src_port: &PortIndexValue,
            dst_port: &PortIndexValue,
        ) -> Option<String> {
            let src_label = match src_port {
                PortIndexValue::Path(path) => Some(path.to_token_stream().to_string()),
                PortIndexValue::Int(index) => Some(index.value.to_string()),
                _ => None,
            };
            let dst_label = match dst_port {
                PortIndexValue::Path(path) => Some(path.to_token_stream().to_string()),
                PortIndexValue::Int(index) => Some(index.value.to_string()),
                _ => None,
            };
            let label = match (src_label, dst_label) {
                (Some(l1), Some(l2)) => Some(format!("{}\n{}", l1, l2)),
                (Some(l1), None) => Some(l1),
                (None, Some(l2)) => Some(l2),
                (None, None) => None,
            };
            label
        }

        // Make node color map one time.
        let node_color_map = self.node_color_map();

        // Write prologue.
        graph_write.write_prologue()?;

        // Define nodes.
        let mut skipped_handoffs = BTreeSet::new();
        for (node_id, node) in self.nodes() {
            if matches!(node, GraphNode::Handoff { .. }) && write_config.no_handoffs {
                skipped_handoffs.insert(node_id);
                continue;
            }
            graph_write.write_node_definition(
                node_id,
                &if write_config.op_short_text {
                    node.to_name_string()
                } else if write_config.op_text_no_imports {
                    // Remove any lines that start with "use" (imports)
                    let full_text = node.to_pretty_string();
                    let mut output = String::new();
                    for sentence in full_text.split('\n') {
                        if sentence.trim().starts_with("use") {
                            continue;
                        }
                        output.push('\n');
                        output.push_str(sentence);
                    }
                    output.into()
                } else {
                    node.to_pretty_string()
                },
                if write_config.no_pull_push {
                    None
                } else {
                    node_color_map.get(node_id).copied()
                },
            )?;
        }

        // Write edges.
        for (edge_id, (src_id, mut dst_id)) in self.edges() {
            // Handling for if `write_config.no_handoffs` true.
            if skipped_handoffs.contains(&src_id) {
                continue;
            }

            let (src_port, mut dst_port) = self.edge_ports(edge_id);
            if skipped_handoffs.contains(&dst_id) {
                // The destination is a hidden handoff. If it has a successor, skip through.
                // If it has 0 successors (ref-only singleton), drop this edge entirely —
                // the data dependency is captured via the reference edge instead.
                let mut handoff_succs = self.node_successors(dst_id);
                if handoff_succs.len() == 0 {
                    continue;
                }
                let (succ_edge, succ_node) = handoff_succs.next().unwrap();
                dst_id = succ_node;
                dst_port = self.edge_ports(succ_edge).1;
            }

            let label = helper_edge_label(src_port, dst_port);
            let delay_type = self
                .node_op_inst(dst_id)
                .and_then(|op_inst| (op_inst.op_constraints.input_delaytype_fn)(dst_port));
            graph_write.write_edge(src_id, dst_id, delay_type, label.as_deref(), false)?;
        }

        // Write reference edges.
        if !write_config.no_references {
            for dst_id in self.node_ids() {
                for src_ref_id in self
                    .node_handoff_references(dst_id)
                    .iter()
                    .filter_map(|r| r.node_id)
                {
                    // When handoffs are hidden, resolve through to the predecessor of
                    // the singleton handoff so the edge points from the actual writer.
                    let resolved_src = if skipped_handoffs.contains(&src_ref_id) {
                        self.node_predecessor_nodes(src_ref_id).next()
                    } else {
                        Some(src_ref_id)
                    };
                    let Some(resolved_src) = resolved_src else {
                        continue;
                    };
                    let label = None;
                    graph_write.write_edge(resolved_src, dst_id, None, label, true)?;
                }
            }
        }

        // The following code is a little bit tricky. Generally, the graph has the hierarchy:
        // `loop -> subgraph -> varname -> node`. However, each of these can be disabled via the `write_config`. To
        // handle both the enabled and disabled case, this code is structured as a series of nested loops. If the layer
        // is disabled, then the HashMap<Option<KEY>, Vec<VALUE>> will only have a single key (`None`) with a
        // corresponding `Vec` value containing everything. This way no special handling is needed for the next layer.

        // Loop -> Subgraphs
        let loop_subgraphs = self.subgraph_ids().map(|sg_id| {
            let loop_id = if write_config.no_loops {
                None
            } else {
                self.subgraph_loop(sg_id)
            };
            (loop_id, sg_id)
        });
        let loop_subgraphs = into_group_map(loop_subgraphs);
        for (loop_id, subgraph_ids) in loop_subgraphs {
            if let Some(loop_id) = loop_id {
                graph_write.write_loop_start(loop_id)?;
            }

            // Subgraph -> Varnames.
            let subgraph_varnames_nodes = subgraph_ids.into_iter().flat_map(|sg_id| {
                self.subgraph(sg_id).iter().copied().map(move |node_id| {
                    let opt_sg_id = if write_config.no_subgraphs {
                        None
                    } else {
                        Some(sg_id)
                    };
                    (opt_sg_id, (self.node_varname(node_id), node_id))
                })
            });
            let subgraph_varnames_nodes = into_group_map(subgraph_varnames_nodes);
            for (sg_id, varnames) in subgraph_varnames_nodes {
                if let Some(sg_id) = sg_id {
                    graph_write.write_subgraph_start(sg_id)?;
                }

                // Varnames -> Nodes.
                let varname_nodes = varnames.into_iter().map(|(varname, node)| {
                    let varname = if write_config.no_varnames {
                        None
                    } else {
                        varname
                    };
                    (varname, node)
                });
                let varname_nodes = into_group_map(varname_nodes);
                for (varname, node_ids) in varname_nodes {
                    if let Some(varname) = varname {
                        graph_write.write_varname_start(&varname.0.to_string(), sg_id)?;
                    }

                    // Write all nodes.
                    for node_id in node_ids {
                        graph_write.write_node(node_id)?;
                    }

                    if varname.is_some() {
                        graph_write.write_varname_end()?;
                    }
                }

                if sg_id.is_some() {
                    graph_write.write_subgraph_end()?;
                }
            }

            if loop_id.is_some() {
                graph_write.write_loop_end()?;
            }
        }

        // Write epilogue.
        graph_write.write_epilogue()?;

        Ok(())
    }

    /// Convert back into surface syntax.
    pub fn surface_syntax_string(&self) -> String {
        let mut string = String::new();
        self.write_surface_syntax(&mut string).unwrap();
        string
    }

    /// Convert back into surface syntax.
    pub fn write_surface_syntax(&self, write: &mut impl std::fmt::Write) -> std::fmt::Result {
        for (key, node) in self.nodes.iter() {
            match node {
                GraphNode::Operator(op) => {
                    writeln!(write, "_{:?} = {};", key.data(), op.to_token_stream())?;
                }
                GraphNode::Handoff {
                    kind: HandoffKind::Vec,
                    ..
                } => {
                    writeln!(write, "_{:?} = handoff();", key.data())?;
                }
                GraphNode::Handoff {
                    kind: HandoffKind::Singleton,
                    ..
                } => {
                    writeln!(write, "_{:?} = singleton();", key.data())?;
                }
                GraphNode::Handoff {
                    kind: HandoffKind::Optional,
                    ..
                } => {
                    writeln!(write, "_{:?} = optional();", key.data())?;
                }
                GraphNode::ModuleBoundary { .. } => panic!(),
            }
        }
        writeln!(write)?;
        for (e, (src_key, dst_key)) in self.graph.edges() {
            let (src_port, dst_port) = self.edge_ports(e);
            let src_port_str = if src_port.is_specified() {
                format!("[{}]", src_port)
            } else {
                String::new()
            };
            let dst_port_str = if dst_port.is_specified() {
                format!("[{}]", dst_port)
            } else {
                String::new()
            };
            writeln!(
                write,
                "_{:?}{} -> {}_{:?};",
                src_key.data(),
                src_port_str,
                dst_port_str,
                dst_key.data()
            )?;
        }
        Ok(())
    }

    /// Convert into a [mermaid](https://mermaid-js.github.io/) graph. Ignores subgraphs.
    pub fn mermaid_string_flat(&self) -> String {
        let mut string = String::new();
        self.write_mermaid_flat(&mut string).unwrap();
        string
    }

    /// Convert into a [mermaid](https://mermaid-js.github.io/) graph. Ignores subgraphs.
    pub fn write_mermaid_flat(&self, write: &mut impl std::fmt::Write) -> std::fmt::Result {
        writeln!(write, "flowchart TB")?;
        for (key, node) in self.nodes.iter() {
            match node {
                GraphNode::Operator(operator) => writeln!(
                    write,
                    "    %% {span}\n    {id:?}[\"{row_col} <tt>{code}</tt>\"]",
                    span = PrettySpan(node.span()),
                    id = key.data(),
                    row_col = PrettyRowCol(node.span()),
                    code = operator
                        .to_token_stream()
                        .to_string()
                        .replace('&', "&amp;")
                        .replace('<', "&lt;")
                        .replace('>', "&gt;")
                        .replace('"', "&quot;")
                        .replace('\n', "<br>"),
                ),
                GraphNode::Handoff {
                    kind: HandoffKind::Vec,
                    ..
                } => {
                    writeln!(write, r#"    {:?}{{"{}"}}"#, key.data(), HANDOFF_NODE_STR)
                }
                GraphNode::Handoff {
                    kind: HandoffKind::Singleton | HandoffKind::Optional,
                    ..
                } => {
                    writeln!(
                        write,
                        r#"    {:?}{{"{}"}}"#,
                        key.data(),
                        SINGLETON_SLOT_NODE_STR
                    )
                }
                GraphNode::ModuleBoundary { .. } => {
                    writeln!(
                        write,
                        r#"    {:?}{{"{}"}}"#,
                        key.data(),
                        MODULE_BOUNDARY_NODE_STR
                    )
                }
            }?;
        }
        writeln!(write)?;
        for (_e, (src_key, dst_key)) in self.graph.edges() {
            writeln!(write, "    {:?}-->{:?}", src_key.data(), dst_key.data())?;
        }
        Ok(())
    }
}

/// Loops
impl DfirGraph {
    /// Iterator over all loop IDs.
    pub fn loop_ids(&self) -> slotmap::basic::Keys<'_, GraphLoopId, Vec<GraphNodeId>> {
        self.loop_nodes.keys()
    }

    /// Iterator over all loops, ID and members: `(GraphLoopId, Vec<GraphNodeId>)`.
    pub fn loops(&self) -> slotmap::basic::Iter<'_, GraphLoopId, Vec<GraphNodeId>> {
        self.loop_nodes.iter()
    }

    /// Get a loop's member nodes.
    pub fn loop_nodes(&self, loop_id: GraphLoopId) -> &[GraphNodeId] {
        self.loop_nodes.get(loop_id).unwrap()
    }

    /// Create a new loop context, with the given parent loop (or `None`).
    pub fn insert_loop(&mut self, parent_loop: Option<GraphLoopId>) -> GraphLoopId {
        let loop_id = self.loop_nodes.insert(Vec::new());
        self.loop_children.insert(loop_id, Vec::new());
        if let Some(parent_loop) = parent_loop {
            self.loop_parent.insert(loop_id, parent_loop);
            self.loop_children
                .get_mut(parent_loop)
                .unwrap()
                .push(loop_id);
        } else {
            self.root_loops.push(loop_id);
        }
        loop_id
    }

    /// Get a node's loop context (or `None` for root).
    pub fn node_loop(&self, node_id: GraphNodeId) -> Option<GraphLoopId> {
        self.node_loops.get(node_id).copied()
    }

    /// Get a subgraph's loop context (or `None` for root).
    pub fn subgraph_loop(&self, subgraph_id: GraphSubgraphId) -> Option<GraphLoopId> {
        let &node_id = self.subgraph(subgraph_id).first().unwrap();
        let out = self.node_loop(node_id);
        debug_assert!(
            self.subgraph(subgraph_id)
                .iter()
                .all(|&node_id| self.node_loop(node_id) == out),
            "Subgraph nodes should all have the same loop context."
        );
        out
    }

    /// Get a loop context's parent loop context (or `None` for root).
    pub fn loop_parent(&self, loop_id: GraphLoopId) -> Option<GraphLoopId> {
        self.loop_parent.get(loop_id).copied()
    }

    /// Get a loop context's child loops.
    pub fn loop_children(&self, loop_id: GraphLoopId) -> &Vec<GraphLoopId> {
        self.loop_children.get(loop_id).unwrap()
    }

    /// Get root-level loops (those with no parent loop).
    pub fn root_loops(&self) -> &[GraphLoopId] {
        &self.root_loops
    }
}

/// Configuration for writing graphs.
#[derive(Clone, Debug, Default)]
#[cfg_attr(feature = "clap-derive", derive(clap::Args))]
pub struct WriteConfig {
    /// Subgraphs will not be rendered if set.
    #[cfg_attr(feature = "clap-derive", arg(long))]
    pub no_subgraphs: bool,
    /// Variable names will not be rendered if set.
    #[cfg_attr(feature = "clap-derive", arg(long))]
    pub no_varnames: bool,
    /// Will not render pull/push shapes if set.
    #[cfg_attr(feature = "clap-derive", arg(long))]
    pub no_pull_push: bool,
    /// Will not render handoffs if set.
    #[cfg_attr(feature = "clap-derive", arg(long))]
    pub no_handoffs: bool,
    /// Will not render singleton references if set.
    #[cfg_attr(feature = "clap-derive", arg(long))]
    pub no_references: bool,
    /// Will not render loops if set.
    #[cfg_attr(feature = "clap-derive", arg(long))]
    pub no_loops: bool,

    /// Op text will only be their name instead of the whole source.
    #[cfg_attr(feature = "clap-derive", arg(long))]
    pub op_short_text: bool,
    /// Op text will exclude any line that starts with "use".
    #[cfg_attr(feature = "clap-derive", arg(long))]
    pub op_text_no_imports: bool,
}

/// Enum for choosing between mermaid and dot graph writing.
#[derive(Copy, Clone, Debug)]
#[cfg_attr(feature = "clap-derive", derive(clap::Parser, clap::ValueEnum))]
pub enum WriteGraphType {
    /// Mermaid graphs.
    Mermaid,
    /// Dot (Graphviz) graphs.
    Dot,
}

/// [`itertools::Itertools::into_group_map`], but for `BTreeMap`.
fn into_group_map<K, V>(iter: impl IntoIterator<Item = (K, V)>) -> BTreeMap<K, Vec<V>>
where
    K: Ord,
{
    let mut out: BTreeMap<_, Vec<_>> = BTreeMap::new();
    for (k, v) in iter {
        out.entry(k).or_default().push(v);
    }
    out
}