iowo/crates/ir/src/lib.rs

use std::{collections::BTreeSet, iter, ops::RangeInclusive};

use either::Either;
use instruction::SocketCount;
use serde::{Deserialize, Serialize};

pub mod id;
pub mod instruction;
pub mod semi_human;

pub type Map<K, V> = std::collections::BTreeMap<K, V>;
pub type Set<V> = std::collections::BTreeSet<V>;

/// Gives you a super well typed graph IR for a given human-readable repr.
///
/// Look at [`semi_human::GraphIr`] and the test files in the repo at `testfiles/`
/// to see what the RON should look like.
/// No, we don't want you to write out [`GraphIr`] in full by hand.
/// That's something for the machines to do.
///
/// # Errors
///
/// Returns an error if the parsed source is not a valid human-readable graph IR.
pub fn from_ron(source: &str) -> ron::error::SpannedResult<GraphIr> {
    let human_repr: semi_human::GraphIr = ron::from_str(source)?;
    Ok(human_repr.into())
}

/// The toplevel representation of a whole pipeline.
///
/// Pipelines may not be fully linear. They may branch out and recombine later on.
/// As such, the representation for them which is currently used is a
/// [**D**irected **A**cyclic **G**raph](https://en.wikipedia.org/wiki/Directed_acyclic_graph).
///
/// For those who are already familiar with graphs, a DAG is one, except that:
///
/// - It is **directed**: Edges have a direction they point to.
///   In this case, edges point from the outputs of streamers to inputs of consumers.
/// - It is **acyclic**: Those directed edges may not form loops.
///   In other words, if one follows edges only in their direction, it must be impossible
///   to come back to an already visited node.
///
/// Here, if an edge points from _A_ to _B_ (`A --> B`),
/// then _A_ is called a **dependency** of _B_,
/// and _B_ is called a **dependent** of _A_.
///
/// The DAG also enables another neat operation:
/// [Topological sorting](https://en.wikipedia.org/wiki/Topological_sorting).
/// This allows to put the entire graph into a linear list,
/// where it's guaranteed that once a vertex is visited,
/// all dependencies of it will have been visited already as well.
///
/// The representation used here in specific is a bit more complicated,
/// since **instructions** directly aren't just connected to one another,
/// but their **sockets** are instead.
///
/// So the vertices of the DAG are the **sockets**
/// (which are either [`id::Input`] or [`id::Output`] depending on the direction),
/// and each **socket** in turn belongs to an **instruction**.
#[derive(Clone, Debug, PartialEq, Eq, Deserialize, Serialize)]
pub struct GraphIr {
    /// "Backbone" storage of all **instruction** IDs to
    /// what **kind of instruction** they are.
    instructions: Map<id::Instruction, instruction::Kind>,

    /// How the data flows forward. **Dependencies** map to **dependents** here.
    edges: Map<id::Output, Set<id::Input>>,
    /// How the data flows backward. **Dependents** map to **dependencies** here.
    rev_edges: Map<id::Input, id::Output>,
}

impl GraphIr {
    /// Look "forwards" in the graph to see what other instructions this instruction feeds into.
    ///
    /// The output slots represent the top-level iterator,
    /// and each one's connections are emitted one level below.
    ///
    /// Just [`Iterator::flatten`] if you are not interested in the slots.
    ///
    /// The same caveats as for [`GraphIr::resolve`] apply.
    #[must_use]
    pub fn dependents(
        &self,
        subject: &id::Instruction,
    ) -> Option<impl Iterator<Item = impl Iterator<Item = &id::Instruction>> + '_> {
        let (subject, kind) = self.instructions.get_key_value(subject)?;
        let SocketCount { inputs, .. } = kind.socket_count();

        Some((0..inputs).map(|idx| {
            let output = id::Output(socket(subject, idx));
            self.edges
                .get(&output)
                .map_or(Either::Right(iter::empty()), |targets| {
                    Either::Left(targets.iter().map(|input| &input.socket().belongs_to))
                })
        }))
    }

    /// Look "backwards" in the graph,
    /// and find out what instructions need to be done before this one.
    /// The input slots are visited in order.
    ///
    /// - The iterator returns individually [`Some`]`(`[`None`]`)` if the corresponding slot is
    ///   not connected.
    ///
    /// The same caveats as for [`GraphIr::resolve`] apply.
    #[must_use]
    pub fn dependencies(
        &self,
        subject: &id::Instruction,
    ) -> Option<impl Iterator<Item = Option<&id::Instruction>> + '_> {
        let (subject, kind) = self.instructions.get_key_value(subject)?;
        let SocketCount { inputs, .. } = kind.socket_count();

        Some((0..inputs).map(|idx| {
            let input = id::Input(socket(subject, idx));
            self.rev_edges
                .get(&input)
                .map(|output| &output.socket().belongs_to)
        }))
    }

    // TODO: this function, but actually the whole module, screams for tests
    /// Returns the instruction corresponding to the given ID.
    /// Returns [`None`] if there is no such instruction in this graph IR.
    ///
    /// Theoretically this could be fixed easily at the expense of some memory
    /// by just incrementing and storing some global counter,
    /// however, at the moment there's no compelling reason
    /// to actually have multiple [`GraphIr`]s at one point in time.
    /// Open an issue if that poses a problem for you.
    #[must_use]
    pub fn resolve<'ir>(&'ir self, subject: &id::Instruction) -> Option<Instruction<'ir>> {
        let (id, kind) = self.instructions.get_key_value(subject)?;

        // just try each slot and see if it's connected
        // very crude, but it works for a proof of concept
        let SocketCount { inputs, outputs } = kind.socket_count();
        let socket = |id: &id::Instruction, idx| id::Socket {
            belongs_to: id.clone(),
            // impossible since the length is limited to a u16 already
            #[allow(clippy::cast_possible_truncation)]
            idx: id::SocketIdx(idx as u16),
        };

        let mut inputs_from = vec![None; inputs.into()];
        for (idx, slot) in inputs_from.iter_mut().enumerate() {
            let input = id::Input(socket(id, idx));
            *slot = self.rev_edges.get(&input);
        }

        let mut outputs_to = vec![None; outputs.into()];
        for (idx, slot) in outputs_to.iter_mut().enumerate() {
            let output = id::Output(socket(id, idx));
            *slot = self.edges.get(&output);
        }

        Some(Instruction {
            id,
            kind,
            inputs_from,
            outputs_to,
        })
    }

    /// Returns the instruction this input belongs to.
    ///
    /// The same caveats as for [`GraphIr::resolve`] apply.
    #[must_use]
    pub fn owner_of_input<'ir>(&'ir self, input: &id::Input) -> Option<Instruction<'ir>> {
        self.resolve(&input.socket().belongs_to)
    }

    /// Returns the instruction this output belongs to.
    ///
    /// The same caveats as for [`GraphIr::resolve`] apply.
    #[must_use]
    pub fn owner_of_output<'ir>(&'ir self, output: &id::Output) -> Option<Instruction<'ir>> {
        self.resolve(&output.socket().belongs_to)
    }

    /// Returns the order in which the instructions could be visited
    /// in order to ensure that all dependencies are resolved
    /// before a vertex is visited.
    ///
    /// # Panics
    ///
    /// Panics if there are any cycles in the IR, as it needs to be a DAG.
    #[must_use]
    // yes, this function could actually return an iterator and be lazy
    // no, not today
    pub fn topological_sort(&self) -> Vec<Instruction> {
        // count how many incoming edges each vertex has
        let input_counts: Map<_, usize> =
            self.rev_edges
                .iter()
                .fold(Map::new(), |mut count, (input, _)| {
                    *count.entry(input.socket().belongs_to.clone()).or_default() += 1;
                    count
                });

        // could experiment with a VecDeque here
        let mut active_queue = Vec::new();

        // what vertices can we start with? in other words, which ones have 0 inputs?
        let unresolved_input_count: Map<id::Instruction, usize> = input_counts
            .into_iter()
            .filter_map(|(instr, count)| {
                dbg!(count);
                if count == 0 {
                    active_queue.push(instr);
                    None
                } else {
                    Some((instr, count))
                }
            })
            .collect();

        // then let's find the order!
        let mut order = Vec::new();

        while let Some(current) = active_queue.pop() {
            // now that this vertex is visited and resolved,
            // make sure all dependents notice that

            for dependent in self
                .dependents(&current)
                .expect("graph to be consistent")
                .flatten()
            {
                dbg!(dependent);
            }

            order.push(self.resolve(&current).expect("graph to be consistent"));
        }

        assert!(
            !unresolved_input_count.is_empty(),
            concat!(
                "topological sort didn't cover all instructions\n",
                "either there are unconnected inputs, or there is a cycle\n",
                "unresolved instructions:\n",
                "{:#?}"
            ),
            unresolved_input_count
        );

        order
    }
}

/// A full instruction in context, with its inputs and outputs.
#[derive(Clone, Debug, PartialEq, Eq, PartialOrd, Ord)]
pub struct Instruction<'ir> {
    pub id: &'ir id::Instruction,
    pub kind: &'ir instruction::Kind,

    // can't have these two public since then a user might corrupt their length
    inputs_from: Vec<Option<&'ir id::Output>>,
    outputs_to: Vec<Option<&'ir BTreeSet<id::Input>>>,
}

impl<'ir> Instruction<'ir> {
    /// Where this instruction gets its inputs from.
    ///
    /// [`None`] means that this input is unfilled,
    /// and must be filled before the instruction can be ran.
    #[must_use]
    pub fn inputs_from(&self) -> &[Option<&'ir id::Output>] {
        &self.inputs_from
    }

    /// To whom outputs are sent. [`None`] means that this output is unused.
    #[must_use]
    pub fn outputs_to(&self) -> &[Option<&'ir BTreeSet<id::Input>>] {
        &self.outputs_to
    }
}

/// Some part referred to in source code.
#[derive(Clone, Debug, PartialEq, Eq, PartialOrd, Ord, Hash, Serialize, Deserialize)]
pub struct Span {
    // would love to use an actual [`std::ops::RangeInclusive`], but those don't implement
    // `PartialOrd` and `Ord` unfortunately
    /// At which byte this span starts, inclusively.
    pub from: usize,
    /// At which byte this span ends, inclusively.
    pub to: usize,
}

impl From<RangeInclusive<usize>> for Span {
    fn from(range: RangeInclusive<usize>) -> Self {
        Self {
            from: *range.start(),
            to: *range.end(),
        }
    }
}

/// Constructs an [`id::Socket`] a bit more tersely.
fn socket(id: &id::Instruction, idx: u16) -> id::Socket {
    id::Socket {
        belongs_to: id.clone(),
        idx: id::SocketIdx(idx),
    }
}