graphix_compiler/

lib.rs

#![doc(
    html_logo_url = "https://graphix-lang.github.io/graphix/graphix-icon.svg",
    html_favicon_url = "https://graphix-lang.github.io/graphix/graphix-icon.svg"
)]
#[macro_use]
extern crate netidx_core;
#[macro_use]
extern crate combine;
#[macro_use]
extern crate serde_derive;

pub mod env;
pub mod expr;
pub mod node;
pub mod typ;

use crate::{
    env::Env,
    expr::{ExprId, ModPath},
    node::lambda::LambdaDef,
    typ::{FnType, Type},
};
use anyhow::{bail, Result};
use arcstr::ArcStr;
use enumflags2::{bitflags, BitFlags};
use expr::Expr;
use futures::channel::mpsc;
use fxhash::{FxHashMap, FxHashSet};
use log::info;
use netidx::{
    path::Path,
    publisher::{Id, Val, WriteRequest},
    subscriber::{self, Dval, SubId, UpdatesFlags, Value},
};
use netidx_protocols::rpc::server::{ArgSpec, RpcCall};
use netidx_value::{abstract_type::AbstractWrapper, Abstract};
use node::compiler;
use parking_lot::{Mutex, RwLock};
use poolshark::{
    global::{GPooled, Pool},
    local::LPooled,
};
use std::{
    any::{Any, TypeId},
    cell::Cell,
    collections::{
        hash_map::{self, Entry},
        HashMap,
    },
    fmt::Debug,
    mem,
    sync::{
        self,
        atomic::{AtomicBool, Ordering},
        LazyLock,
    },
    time::Duration,
};
use tokio::{task, time::Instant};
use triomphe::Arc;
use uuid::Uuid;

#[derive(Debug, Clone, Copy)]
#[bitflags]
#[repr(u64)]
pub enum CFlag {
    WarnUnhandled,
    WarnUnused,
    WarningsAreErrors,
}

#[allow(dead_code)]
static TRACE: AtomicBool = AtomicBool::new(false);

#[allow(dead_code)]
pub fn set_trace(b: bool) {
    TRACE.store(b, Ordering::Relaxed)
}

#[allow(dead_code)]
pub fn with_trace<F: FnOnce() -> Result<R>, R>(
    enable: bool,
    spec: &Expr,
    f: F,
) -> Result<R> {
    let prev = trace();
    set_trace(enable);
    if !prev && enable {
        eprintln!("trace enabled at {}, spec: {}", spec.pos, spec);
    } else if prev && !enable {
        eprintln!("trace disabled at {}, spec: {}", spec.pos, spec);
    }
    let r = match f() {
        Err(e) => {
            eprintln!("traced at {} failed with {e:?}", spec.pos);
            Err(e)
        }
        r => r,
    };
    if prev && !enable {
        eprintln!("trace reenabled")
    }
    set_trace(prev);
    r
}

#[allow(dead_code)]
pub fn trace() -> bool {
    TRACE.load(Ordering::Relaxed)
}

#[macro_export]
macro_rules! tdbg {
    ($e:expr) => {
        if $crate::trace() {
            dbg!($e)
        } else {
            $e
        }
    };
}

#[macro_export]
macro_rules! err {
    ($tag:expr, $err:literal) => {{
        let e: Value = ($tag.clone(), ::arcstr::literal!($err)).into();
        Value::Error(::triomphe::Arc::new(e))
    }};
}

#[macro_export]
macro_rules! errf {
    ($tag:expr, $fmt:expr, $($args:expr),*) => {{
        let msg: ArcStr = ::compact_str::format_compact!($fmt, $($args),*).as_str().into();
        let e: Value = ($tag.clone(), msg).into();
        Value::Error(::triomphe::Arc::new(e))
    }};
    ($tag:expr, $fmt:expr) => {{
        let msg: ArcStr = ::compact_str::format_compact!($fmt).as_str().into();
        let e: Value = ($tag.clone(), msg).into();
        Value::Error(::triomphe::Arc::new(e))
    }};
}

#[macro_export]
macro_rules! defetyp {
    ($name:ident, $tag_name:ident, $tag:literal, $typ:expr) => {
        static $tag_name: ArcStr = ::arcstr::literal!($tag);
        static $name: ::std::sync::LazyLock<$crate::typ::Type> =
            ::std::sync::LazyLock::new(|| {
                let scope = $crate::expr::ModPath::root();
                $crate::expr::parser::parse_type(&format!($typ, $tag))
                    .expect("failed to parse type")
                    .scope_refs(&scope)
            });
    };
}

defetyp!(CAST_ERR, CAST_ERR_TAG, "InvalidCast", "Error<`{}(string)>");

atomic_id!(LambdaId);

impl From<u64> for LambdaId {
    fn from(v: u64) -> Self {
        LambdaId(v)
    }
}

atomic_id!(BindId);

impl From<u64> for BindId {
    fn from(v: u64) -> Self {
        BindId(v)
    }
}

impl TryFrom<Value> for BindId {
    type Error = anyhow::Error;

    fn try_from(value: Value) -> Result<Self> {
        match value {
            Value::U64(id) => Ok(BindId(id)),
            v => bail!("invalid bind id {v}"),
        }
    }
}

pub trait UserEvent: Clone + Debug + Any {
    fn clear(&mut self);
}

pub trait CustomBuiltinType: Debug + Any + Send + Sync {}

impl CustomBuiltinType for Value {}
impl CustomBuiltinType for Option<Value> {}

#[derive(Debug, Clone)]
pub struct NoUserEvent;

impl UserEvent for NoUserEvent {
    fn clear(&mut self) {}
}

#[derive(Debug, Clone, Copy)]
#[bitflags]
#[repr(u64)]
pub enum PrintFlag {
    /// Dereference type variables and print both the tvar name and the bound
    /// type or "unbound".
    DerefTVars,
    /// Replace common primitives with shorter type names as defined
    /// in core. e.g. Any, instead of the set of every primitive type.
    ReplacePrims,
    /// When formatting an Origin don't print the source, just the location
    NoSource,
    /// When formatting an Origin don't print the origin's parents
    NoParents,
}

thread_local! {
    static PRINT_FLAGS: Cell<BitFlags<PrintFlag>> = Cell::new(PrintFlag::ReplacePrims.into());
}

/// global pool of channel watch batches
pub static CBATCH_POOL: LazyLock<Pool<Vec<(BindId, Box<dyn CustomBuiltinType>)>>> =
    LazyLock::new(|| Pool::new(10000, 1000));

/// For the duration of the closure F change the way type variables
/// are formatted (on this thread only) according to the specified
/// flags.
pub fn format_with_flags<G: Into<BitFlags<PrintFlag>>, R, F: FnOnce() -> R>(
    flags: G,
    f: F,
) -> R {
    let prev = PRINT_FLAGS.replace(flags.into());
    let res = f();
    PRINT_FLAGS.set(prev);
    res
}

/// Event represents all the things that happened simultaneously in a
/// given execution cycle. Event may contain only one update for each
/// variable and netidx subscription in a given cycle, if more updates
/// happen simultaneously they must be queued and deferred to later
/// cycles.
#[derive(Debug)]
pub struct Event<E: UserEvent> {
    pub init: bool,
    pub variables: FxHashMap<BindId, Value>,
    pub netidx: FxHashMap<SubId, subscriber::Event>,
    pub writes: FxHashMap<Id, WriteRequest>,
    pub rpc_calls: FxHashMap<BindId, RpcCall>,
    pub custom: FxHashMap<BindId, Box<dyn CustomBuiltinType>>,
    pub user: E,
}

impl<E: UserEvent> Event<E> {
    pub fn new(user: E) -> Self {
        Event {
            init: false,
            variables: HashMap::default(),
            netidx: HashMap::default(),
            writes: HashMap::default(),
            rpc_calls: HashMap::default(),
            custom: HashMap::default(),
            user,
        }
    }

    pub fn clear(&mut self) {
        let Self { init, variables, netidx, rpc_calls, writes, custom, user } = self;
        *init = false;
        variables.clear();
        netidx.clear();
        rpc_calls.clear();
        custom.clear();
        writes.clear();
        user.clear();
    }
}

#[derive(Debug, Clone, Default)]
pub struct Refs {
    refed: LPooled<FxHashSet<BindId>>,
    bound: LPooled<FxHashSet<BindId>>,
}

pub use combine::stream::position::SourcePosition;

/// A textual occurrence of a name at a specific source position that
/// the compiler resolved to a particular `BindId`. Populated as a side
/// effect of compilation so IDE tooling can answer
/// `textDocument/references` and `textDocument/definition` without
/// re-implementing name resolution.
///
/// `def_pos` and `def_ori` mirror the bind's declaration site at
/// resolution time. They're captured here because some bindings
/// (notably lambda parameters) are unbound from the env when the
/// callsite that created them is dropped — but their declaration
/// site is still meaningful to the user.
#[derive(Debug, Clone)]
pub struct ReferenceSite {
    pub pos: SourcePosition,
    pub ori: Arc<expr::Origin>,
    pub name: expr::ModPath,
    pub bind_id: BindId,
    pub def_pos: SourcePosition,
    pub def_ori: Arc<expr::Origin>,
}

/// A textual occurrence of a module reference (either `use foo;` or
/// `mod foo;`). For the `mod foo;` case `def_ori` points at the file
/// the module's body was loaded from — that's the natural target for
/// go-to-definition on a module name.
#[derive(Debug, Clone)]
pub struct ModuleRefSite {
    pub pos: SourcePosition,
    pub ori: Arc<expr::Origin>,
    /// Module name as the user wrote it (might be relative).
    pub name: expr::ModPath,
    /// Absolute module path the compiler resolved this reference to.
    pub canonical: expr::ModPath,
    /// Origin of the module's body (the file it was loaded from)
    /// when this site is itself a declaration that pulled the
    /// module in. `None` for plain `use` sites.
    pub def_ori: Option<Arc<expr::Origin>>,
}

/// One entry in the per-compile scope map: the compiler descended
/// into an `Expr` at this `(pos, ori)` while in this `scope`. IDE
/// tooling answers `cursor → scope` by finding the entry with the
/// greatest `pos` ≤ the cursor in the same file.
#[derive(Debug, Clone)]
pub struct ScopeMapEntry {
    pub pos: SourcePosition,
    pub ori: Arc<expr::Origin>,
    pub scope: Scope,
}

/// A textual occurrence of a type reference (e.g. `Foo` in `let x: Foo`).
/// Captured by the compiler when a `Type::Ref` carrying parse-time
/// position info gets dereferenced. `def_pos`/`def_ori` point at the
/// `type Foo = …` declaration site so go-to-def on a type name lands
/// on the typedef.
#[derive(Debug, Clone)]
pub struct TypeRefSite {
    pub pos: SourcePosition,
    pub ori: Arc<expr::Origin>,
    /// The name as written in source (e.g. `Result`, `array::Foo`).
    pub name: expr::ModPath,
    /// Canonical scope of the typedef the reference resolved to.
    pub canonical_scope: expr::ModPath,
    pub def_pos: SourcePosition,
    pub def_ori: Arc<expr::Origin>,
}

/// Maps a `val foo: T` declaration in a `.gxi` interface to its
/// `let foo = …` implementation site in the paired `.gx`. Populated by
/// `check_sig` whenever it matches a sig proxy bind to its impl bind.
/// Used by IDE tooling to (a) goto-def from a sig val site to the impl,
/// and (b) union find-references results across both `BindId`s.
/// Only populated when `env.lsp_mode` is set.
#[derive(Debug, Clone)]
pub struct SigImplLink {
    pub scope: expr::ModPath,
    pub name: compact_str::CompactString,
    pub sig_id: BindId,
    pub impl_id: BindId,
}

/// Per-module snapshot of the *internal* env (the impl's view, where
/// implementation bindings shadow sig proxies). The CheckResult's
/// top-level `env` is the *external* view across the project; this
/// per-module entry lets IDE queries on names inside a module reach
/// the impl bind metadata. Only populated when `env.lsp_mode` is set.
#[derive(Debug, Clone)]
pub struct ModuleInternalView {
    pub scope: expr::ModPath,
    pub env: env::Env,
}

/// Pools backing the IDE side-channel collections. `GPooled` so the
/// buffers can return to the same pool after crossing the
/// runtime-task → LSP-thread boundary as part of `CheckResult`. Sized
/// generously since the LSP recompiles on every keystroke and these
/// can grow into the tens of thousands of entries on large modules.
pub static REFERENCE_SITE_POOL: LazyLock<Pool<Vec<ReferenceSite>>> =
    LazyLock::new(|| Pool::new(64, 65536));
pub static MODULE_REF_SITE_POOL: LazyLock<Pool<Vec<ModuleRefSite>>> =
    LazyLock::new(|| Pool::new(64, 65536));
pub static SCOPE_MAP_ENTRY_POOL: LazyLock<Pool<Vec<ScopeMapEntry>>> =
    LazyLock::new(|| Pool::new(64, 65536));
// `TYPE_REF_SITE_POOL`, `SIG_LINK_POOL`, and `MODULE_INTERNAL_VIEW_POOL`
// back the per-check `Lsp` sinks; they live in `env` next to the
// `Lsp` struct that consumes them.

impl Refs {
    pub fn clear(&mut self) {
        self.refed.clear();
        self.bound.clear();
    }

    pub fn with_external_refs(&self, mut f: impl FnMut(BindId)) {
        for id in &*self.refed {
            if !self.bound.contains(id) {
                f(*id);
            }
        }
    }
}

pub type Node<R, E> = Box<dyn Update<R, E>>;

/// Phase indicator for Apply::typecheck
#[derive(Debug)]
pub enum TypecheckPhase<'a> {
    /// During Lambda::typecheck — faux args, building FnType
    Lambda,
    /// During deferred check or bind — resolved FnType available
    CallSite(&'a FnType),
}

pub type InitFn<R, E> = sync::Arc<
    dyn for<'a, 'b, 'c, 'd> Fn(
            &'a Scope,
            &'b mut ExecCtx<R, E>,
            &'c mut [Node<R, E>],
            Option<&'d FnType>,
            ExprId,
        ) -> Result<Box<dyn Apply<R, E>>>
        + Send
        + Sync
        + 'static,
>;

/// Apply is a kind of node that represents a function application. It
/// does not hold ownership of it's arguments, instead those are held
/// by a CallSite node. This allows us to change the function called
/// at runtime without recompiling the arguments.
pub trait Apply<R: Rt, E: UserEvent>: Debug + Send + Sync + Any {
    fn update(
        &mut self,
        ctx: &mut ExecCtx<R, E>,
        from: &mut [Node<R, E>],
        event: &mut Event<E>,
    ) -> Option<Value>;

    /// delete any internally generated nodes, only needed for
    /// builtins that dynamically generate code at runtime
    fn delete(&mut self, _ctx: &mut ExecCtx<R, E>) {
        ()
    }

    /// apply custom typechecking. Phase indicates context:
    /// - Lambda: during lambda body checking (faux args). Return NeedsCallSite
    ///   to opt in to deferred call-site type checking.
    /// - CallSite: during deferred check or bind (resolved FnType available)
    fn typecheck(
        &mut self,
        _ctx: &mut ExecCtx<R, E>,
        _from: &mut [Node<R, E>],
        _phase: TypecheckPhase<'_>,
    ) -> Result<()> {
        Ok(())
    }

    /// return the lambdas type, builtins do not need to implement
    /// this, it is implemented by the BuiltIn wrapper
    fn typ(&self) -> Arc<FnType> {
        static EMPTY: LazyLock<Arc<FnType>> = LazyLock::new(|| {
            Arc::new(FnType {
                args: Arc::from_iter([]),
                constraints: Arc::new(RwLock::new(LPooled::take())),
                rtype: Type::Bottom,
                throws: Type::Bottom,
                vargs: None,
                explicit_throws: false,
                ..Default::default()
            })
        });
        Arc::clone(&*EMPTY)
    }

    /// Populate the Refs structure with all the ids bound and refed by this
    /// node. It is only necessary for builtins to implement this if they create
    /// nodes, such as call sites.
    fn refs<'a>(&self, _refs: &mut Refs) {}

    /// put the node to sleep, used in conditions like select for branches that
    /// are not selected. Any cached values should be cleared on sleep.
    fn sleep(&mut self, _ctx: &mut ExecCtx<R, E>);
}

/// Update represents a regular graph node, as opposed to a function
/// application represented by Apply. Regular graph nodes are used for
/// every built in node except for builtin functions.
pub trait Update<R: Rt, E: UserEvent>: Debug + Send + Sync + Any + 'static {
    /// update the node with the specified event and return any output
    /// it might generate
    fn update(&mut self, ctx: &mut ExecCtx<R, E>, event: &mut Event<E>) -> Option<Value>;

    /// delete the node and it's children from the specified context
    fn delete(&mut self, ctx: &mut ExecCtx<R, E>);

    /// type check the node and it's children
    fn typecheck(&mut self, ctx: &mut ExecCtx<R, E>) -> Result<()>;

    /// return the node type
    fn typ(&self) -> &Type;

    /// Populate the Refs structure with all the bind ids either refed or bound
    /// by the node and it's children
    fn refs(&self, refs: &mut Refs);

    /// return the original expression used to compile this node
    fn spec(&self) -> &Expr;

    /// put the node to sleep, called on unselected branches
    fn sleep(&mut self, ctx: &mut ExecCtx<R, E>);
}

pub type BuiltInInitFn<R, E> = for<'a, 'b, 'c, 'd> fn(
    &'a mut ExecCtx<R, E>,
    &'a FnType,
    Option<&'d FnType>,
    &'b Scope,
    &'c [Node<R, E>],
    ExprId,
) -> Result<Box<dyn Apply<R, E>>>;

/// Trait implemented by graphix built-in functions implemented in rust. This
/// trait isn't meant to be implemented manually, use derive(BuiltIn) from the
/// graphix-derive crate
pub trait BuiltIn<R: Rt, E: UserEvent> {
    const NAME: &str;
    const NEEDS_CALLSITE: bool;

    fn init<'a, 'b, 'c, 'd>(
        ctx: &'a mut ExecCtx<R, E>,
        typ: &'a FnType,
        resolved_type: Option<&'d FnType>,
        scope: &'b Scope,
        from: &'c [Node<R, E>],
        top_id: ExprId,
    ) -> Result<Box<dyn Apply<R, E>>>;
}

pub trait Abortable {
    fn abort(&self);
}

impl Abortable for task::AbortHandle {
    fn abort(&self) {
        task::AbortHandle::abort(self)
    }
}

pub trait Rt: Debug + Any {
    type AbortHandle: Abortable;

    fn clear(&mut self);

    /// Subscribe to the specified netidx path
    ///
    /// When the subscription updates you are expected to deliver
    /// Netidx events to the expression specified by ref_by.
    fn subscribe(&mut self, flags: UpdatesFlags, path: Path, ref_by: ExprId) -> Dval;

    /// Called when a subscription is no longer needed
    fn unsubscribe(&mut self, path: Path, dv: Dval, ref_by: ExprId);

    /// List the netidx path, return Value::Null if the path did not
    /// change. When the path did update you should send the output
    /// back as a properly formatted struct with two fields, rows and
    /// columns both containing string arrays.
    fn list(&mut self, id: BindId, path: Path);

    /// List the table at path, return Value::Null if the path did not
    /// change
    fn list_table(&mut self, id: BindId, path: Path);

    /// list or table will no longer be called on this BindId, and
    /// related resources can be cleaned up.
    fn stop_list(&mut self, id: BindId);

    /// Publish the specified value, returning it's Id, which must be
    /// used to update the value and unpublish it. If the path is
    /// already published, return an error.
    fn publish(&mut self, path: Path, value: Value, ref_by: ExprId) -> Result<Val>;

    /// Update the specified value
    fn update(&mut self, id: &Val, value: Value);

    /// Stop publishing the specified id
    fn unpublish(&mut self, id: Val, ref_by: ExprId);

    /// This will be called by the compiler whenever a bound variable
    /// is referenced. The ref_by is the toplevel expression that
    /// contains the variable reference. When a variable event
    /// happens, you should update all the toplevel expressions that
    /// ref that variable.
    ///
    /// ref_var will also be called when a bound lambda expression is
    /// referenced, in that case the ref_by id will be the toplevel
    /// expression containing the call site.
    fn ref_var(&mut self, id: BindId, ref_by: ExprId);
    fn unref_var(&mut self, id: BindId, ref_by: ExprId);

    /// Called by the ExecCtx when set_var is called on it.
    ///
    /// All expressions that ref the id should be updated when this happens. The
    /// runtime must deliver all set_vars in a single event except that set_vars
    /// for the same variable in the same cycle must be queued and deferred to
    /// the next cycle.
    ///
    /// The runtime MUST NOT change event while a cycle is in
    /// progress. set_var must be queued until the cycle ends and then
    /// presented as a new batch.
    fn set_var(&mut self, id: BindId, value: Value);

    /// Notify the RT that a top level variable has been set internally
    ///
    /// This is called when the compiler has determined that it's safe to set a
    /// variable without waiting a cycle. When the updated variable is a
    /// toplevel node this method is called to notify the runtime that needs to
    /// update any dependent toplevel nodes.
    fn notify_set(&mut self, id: BindId);

    /// This must return results from the same path in the call order.
    ///
    /// when the rpc returns you are expected to deliver a Variable
    /// event with the specified id to the expression specified by
    /// ref_by.
    fn call_rpc(&mut self, name: Path, args: Vec<(ArcStr, Value)>, id: BindId);

    /// Publish an rpc at the specified path with the specified
    /// procedure level doc and arg spec.
    ///
    /// When the RPC is called the rpc table in event will be
    /// populated under the specified bind id.
    ///
    /// If the procedure is already published an error will be
    /// returned
    fn publish_rpc(
        &mut self,
        name: Path,
        doc: Value,
        spec: Vec<ArgSpec>,
        id: BindId,
    ) -> Result<()>;

    /// unpublish the rpc identified by the bind id.
    fn unpublish_rpc(&mut self, name: Path);

    /// arrange to have a Timer event delivered after timeout. When
    /// the timer expires you are expected to deliver a Variable event
    /// for the id, containing the current time.
    fn set_timer(&mut self, id: BindId, timeout: Duration);

    /// Spawn a task
    ///
    /// When the task completes it's output must be delivered as a
    /// custom event using the returned `BindId`
    ///
    /// Calling `abort` must guarantee that if it is called before the
    /// task completes then no update will be delivered.
    fn spawn<F: Future<Output = (BindId, Box<dyn CustomBuiltinType>)> + Send + 'static>(
        &mut self,
        f: F,
    ) -> Self::AbortHandle;

    /// Spawn a task
    ///
    /// When the task completes it's output must be delivered as a
    /// variable event using the returned `BindId`
    ///
    /// Calling `abort` must guarantee that if it is called before the
    /// task completes then no update will be delivered.
    fn spawn_var<F: Future<Output = (BindId, Value)> + Send + 'static>(
        &mut self,
        f: F,
    ) -> Self::AbortHandle;

    /// Ask the runtime to watch a channel
    ///
    /// When event batches arrive via the channel the runtime must
    /// deliver the events as custom updates.
    fn watch(
        &mut self,
        s: mpsc::Receiver<GPooled<Vec<(BindId, Box<dyn CustomBuiltinType>)>>>,
    );

    /// Ask the runtime to watch a channel
    ///
    /// When event batches arrive via the channel the runtime must
    /// deliver the events variable updates.
    fn watch_var(&mut self, s: mpsc::Receiver<GPooled<Vec<(BindId, Value)>>>);
}

#[derive(Default)]
pub struct LibState(FxHashMap<TypeId, Box<dyn Any + Send + Sync>>);

impl LibState {
    /// Look up and return the context global library state of type
    /// `T`.
    ///
    /// If none is registered in this context for `T` then create one
    /// using `T::default`
    pub fn get_or_default<T>(&mut self) -> &mut T
    where
        T: Default + Any + Send + Sync,
    {
        self.0
            .entry(TypeId::of::<T>())
            .or_insert_with(|| Box::new(T::default()) as Box<dyn Any + Send + Sync>)
            .downcast_mut::<T>()
            .unwrap()
    }

    /// Look up and return the context global library state of type
    /// `T`.
    ///
    /// If none is registered in this context for `T` then create one
    /// using the provided function.
    pub fn get_or_else<T, F>(&mut self, f: F) -> &mut T
    where
        T: Any + Send + Sync,
        F: FnOnce() -> T,
    {
        self.0
            .entry(TypeId::of::<T>())
            .or_insert_with(|| Box::new(f()) as Box<dyn Any + Send + Sync>)
            .downcast_mut::<T>()
            .unwrap()
    }

    pub fn entry<'a, T>(
        &'a mut self,
    ) -> hash_map::Entry<'a, TypeId, Box<dyn Any + Send + Sync>>
    where
        T: Any + Send + Sync,
    {
        self.0.entry(TypeId::of::<T>())
    }

    /// return true if `T` is present
    pub fn contains<T>(&self) -> bool
    where
        T: Any + Send + Sync,
    {
        self.0.contains_key(&TypeId::of::<T>())
    }

    /// Look up and return a reference to the context global library
    /// state of type `T`.
    ///
    /// If none is registered in this context for `T` return `None`
    pub fn get<T>(&mut self) -> Option<&T>
    where
        T: Any + Send + Sync,
    {
        self.0.get(&TypeId::of::<T>()).map(|t| t.downcast_ref::<T>().unwrap())
    }

    /// Look up and return a mutable reference to the context global
    /// library state of type `T`.
    ///
    /// If none is registered return `None`
    pub fn get_mut<T>(&mut self) -> Option<&mut T>
    where
        T: Any + Send + Sync,
    {
        self.0.get_mut(&TypeId::of::<T>()).map(|t| t.downcast_mut::<T>().unwrap())
    }

    /// Set the context global library state of type `T`
    ///
    /// Any existing state will be returned
    pub fn set<T>(&mut self, t: T) -> Option<Box<T>>
    where
        T: Any + Send + Sync,
    {
        self.0
            .insert(TypeId::of::<T>(), Box::new(t) as Box<dyn Any + Send + Sync>)
            .map(|t| t.downcast::<T>().unwrap())
    }

    /// Remove and refurn the context global state library state of type `T`
    pub fn remove<T>(&mut self) -> Option<Box<T>>
    where
        T: Any + Send + Sync,
    {
        self.0.remove(&TypeId::of::<T>()).map(|t| t.downcast::<T>().unwrap())
    }
}

/// A registry of abstract type UUIDs used by graphix and graphix libraries,
/// along with a string tag describing what the type is. We must do this because
/// you can't register different type ids with the same uuid in netidx's
/// abstract type system, and because abstract types often need to be
/// parameterized by the Rt and UserEvent they will have different a different
/// type id for each monomorphization, and thus they must have a different uuid.
///
/// The tag is necessary because non parameterized functions often end up with
/// an abstract netidx type and want to know generally what it is, for example
/// printing functions.
#[derive(Default)]
pub struct AbstractTypeRegistry {
    by_tid: FxHashMap<TypeId, Uuid>,
    by_uuid: FxHashMap<Uuid, &'static str>,
}

impl AbstractTypeRegistry {
    fn with<V, F: FnMut(&mut AbstractTypeRegistry) -> V>(mut f: F) -> V {
        static REG: LazyLock<Mutex<AbstractTypeRegistry>> =
            LazyLock::new(|| Mutex::new(AbstractTypeRegistry::default()));
        let mut g = REG.lock();
        f(&mut *g)
    }

    /// Get the UUID of abstract type T
    pub(crate) fn uuid<T: Any>(tag: &'static str) -> Uuid {
        Self::with(|rg| {
            *rg.by_tid.entry(TypeId::of::<T>()).or_insert_with(|| {
                let id = Uuid::new_v4();
                rg.by_uuid.insert(id, tag);
                id
            })
        })
    }

    /// return the tag of this abstract type, or None if it isn't registered
    pub fn tag(a: &Abstract) -> Option<&'static str> {
        Self::with(|rg| rg.by_uuid.get(&a.id()).map(|r| *r))
    }

    /// return true if the abstract type has tag
    pub fn is_a(a: &Abstract, tag: &str) -> bool {
        match Self::tag(a) {
            Some(t) => t == tag,
            None => false,
        }
    }
}

pub struct ExecCtx<R: Rt, E: UserEvent> {
    // used to wrap lambdas into an abstract netidx value type
    lambdawrap: AbstractWrapper<LambdaDef<R, E>>,
    // all registered built-in functions
    builtins: FxHashMap<&'static str, (BuiltInInitFn<R, E>, bool)>,
    // whether calling built-in functions is allowed in this context, used for
    // sandboxing
    builtins_allowed: bool,
    // hash consed variant tags
    tags: FxHashSet<ArcStr>,
    /// context global library state for built-in functions
    pub libstate: LibState,
    /// the language environment, typdefs, binds, lambdas, etc
    pub env: Env,
    /// the last value of every bound variable
    pub cached: FxHashMap<BindId, Value>,
    /// the runtime
    pub rt: R,
    /// LambdaDefs indexed by LambdaId, for deferred type checking
    pub lambda_defs: FxHashMap<LambdaId, Value>,
    /// deferred type check closures, evaluated after all primary type checking
    pub deferred_checks:
        Vec<Box<dyn FnOnce(&mut ExecCtx<R, E>) -> Result<()> + Send + Sync>>,
    /// Reference sites accumulated during compilation. Each is a
    /// textual occurrence of a name and the `BindId` it resolved to.
    /// At compile boundaries, swap with a fresh `REFERENCE_SITE_POOL`
    /// container via `mem::replace` (not `mem::take` — `Default`
    /// routes to the unsized thread-local registry, not our named
    /// pool). Only populated when `env.lsp_mode` is set.
    pub references: GPooled<Vec<ReferenceSite>>,
    /// Module reference sites — `use foo;` and `mod foo;` mentions.
    /// Same scoping rules as `references`.
    pub module_references: GPooled<Vec<ModuleRefSite>>,
    /// Per-compile scope map. `compile()` pushes one entry every
    /// time it's invoked, recording the scope it was called with.
    /// IDE tooling reads this to answer `cursor → scope` queries.
    pub scope_map: GPooled<Vec<ScopeMapEntry>>,
}

impl<R: Rt, E: UserEvent> ExecCtx<R, E> {
    pub fn clear(&mut self) {
        self.env.clear();
        self.rt.clear();
    }

    /// Build a new execution context.
    ///
    /// This is a very low level interface that you can use to build a
    /// custom runtime with deep integration to your code. It is very
    /// difficult to use, and if you don't implement everything
    /// correctly the semantics of the language can be wrong.
    ///
    /// Most likely you want to use the `rt` module instead.
    pub fn new(user: R) -> Result<Self> {
        let id = AbstractTypeRegistry::uuid::<LambdaDef<R, E>>("lambda");
        Ok(Self {
            lambdawrap: Abstract::register(id)?,
            env: Env::default(),
            builtins: FxHashMap::default(),
            builtins_allowed: true,
            libstate: LibState::default(),
            tags: FxHashSet::default(),
            cached: HashMap::default(),
            rt: user,
            lambda_defs: FxHashMap::default(),
            deferred_checks: Vec::new(),
            references: REFERENCE_SITE_POOL.take(),
            module_references: MODULE_REF_SITE_POOL.take(),
            scope_map: SCOPE_MAP_ENTRY_POOL.take(),
        })
    }

    pub fn register_builtin<T: BuiltIn<R, E>>(&mut self) -> Result<()> {
        match self.builtins.entry(T::NAME) {
            Entry::Vacant(e) => {
                e.insert((T::init, T::NEEDS_CALLSITE));
            }
            Entry::Occupied(_) => bail!("builtin {} is already registered", T::NAME),
        }
        Ok(())
    }

    /// Wrap a `LambdaDef` into a `Value` that can be returned from a builtin
    /// as a first-class function value. The runtime handles the resulting
    /// Value as a callable lambda — call sites resolve against the LambdaDef's
    /// `init` to construct an Apply impl. Also registers the LambdaDef in
    /// `lambda_defs` so deferred typechecking can find it.
    pub fn wrap_lambda(&mut self, def: LambdaDef<R, E>) -> Value {
        let id = def.id;
        let v = self.lambdawrap.wrap(def);
        self.lambda_defs.insert(id, v.clone());
        v
    }

    /// Built in functions should call this when variables are set
    /// unless they are sure the variable does not need to be
    /// cached. This will also call the user ctx set_var.
    pub fn set_var(&mut self, id: BindId, v: Value) {
        self.cached.insert(id, v.clone());
        self.rt.set_var(id, v)
    }

    fn tag(&mut self, s: &ArcStr) -> ArcStr {
        match self.tags.get(s) {
            Some(s) => s.clone(),
            None => {
                self.tags.insert(s.clone());
                s.clone()
            }
        }
    }

    /// Restore the lexical environment to the snapshot `env` for the duration
    /// of `f` restoring it to it's original value afterwords. `by_id` and
    /// `lambdas` defined by the closure will be retained.
    pub fn with_restored<T, F: FnOnce(&mut Self) -> T>(&mut self, env: Env, f: F) -> T {
        let snap = self.env.restore_lexical_env(env);
        let orig = mem::replace(&mut self.env, snap);
        let r = f(self);
        self.env = self.env.restore_lexical_env(orig);
        r
    }

    /// Restore the lexical environment to the snapshot `env` for the duration
    /// of `f` restoring it to it's original value afterwords. `by_id` and
    /// `lambdas` defined by the closure will be retained. `env` will be mutated
    /// instead of requiring a clone, this allows maintaining continuity in two
    /// different envs across multiple invocations
    pub fn with_restored_mut<T, F: FnOnce(&mut Self) -> T>(
        &mut self,
        env: &mut Env,
        f: F,
    ) -> T {
        let snap = self.env.restore_lexical_env_mut(env);
        let orig = mem::replace(&mut self.env, snap);
        let r = f(self);
        *env = self.env.clone();
        self.env = self.env.restore_lexical_env(orig);
        r
    }
}

#[derive(Debug, Clone)]
pub struct Scope {
    pub lexical: ModPath,
    pub dynamic: ModPath,
}

impl Scope {
    pub fn append<S: AsRef<str> + ?Sized>(&self, s: &S) -> Self {
        Self {
            lexical: ModPath(self.lexical.append(s)),
            dynamic: ModPath(self.dynamic.append(s)),
        }
    }

    pub fn root() -> Self {
        Self { lexical: ModPath::root(), dynamic: ModPath::root() }
    }
}

/// compile the expression into a node graph in the specified context
/// and scope, return the root node or an error if compilation failed.
pub fn compile<R: Rt, E: UserEvent>(
    ctx: &mut ExecCtx<R, E>,
    flags: BitFlags<CFlag>,
    scope: &Scope,
    spec: Expr,
) -> Result<Node<R, E>> {
    let top_id = spec.id;
    let env = ctx.env.clone();
    let st = Instant::now();
    let mut node = match compiler::compile(ctx, flags, spec, scope, top_id) {
        Ok(n) => n,
        Err(e) => {
            ctx.env = env;
            return Err(e);
        }
    };
    info!("compile time {:?}", st.elapsed());
    let st = Instant::now();
    if let Err(e) = node.typecheck(ctx) {
        ctx.env = env;
        return Err(e);
    }
    // run deferred builtin type checks after all primary type checking completes
    while let Some(check) = ctx.deferred_checks.pop() {
        if let Err(e) = check(ctx) {
            ctx.env = env;
            return Err(e);
        }
    }
    info!("typecheck time {:?}", st.elapsed());
    Ok(node)
}