harn_vm/

chunk.rs

use std::collections::BTreeMap;
use std::fmt;
use std::sync::atomic::{AtomicU64, Ordering};
use std::sync::Arc;

use harn_parser::TypeExpr;
use parking_lot::Mutex;
use serde::{Deserialize, Serialize};

use crate::harness::HarnessKind;
use crate::runtime_guards::RuntimeParamGuard;

/// Sentinel value stored in [`Chunk::inline_cache_index`] for code offsets
/// that have no inline-cache slot registered. Chosen as `u32::MAX` so the
/// hot dispatch path can treat the side-table as a flat `Vec<u32>` without
/// an `Option` wrapper — the comparison against the sentinel collapses to a
/// single integer compare. The compile-time max useful slot count is bounded
/// by code length (one slot per cacheable opcode), so `u32::MAX` is safely
/// out of the addressable slot range.
pub(crate) const NO_INLINE_CACHE_SLOT: u32 = u32::MAX;
static NEXT_CHUNK_CACHE_ID: AtomicU64 = AtomicU64::new(1);

fn next_chunk_cache_id() -> u64 {
    NEXT_CHUNK_CACHE_ID.fetch_add(1, Ordering::Relaxed)
}

/// Bytecode opcodes for the Harn VM. The enum, the byte-to-variant
/// mapping, the sync and async dispatch tables, the disassembly
/// renderer, and the per-opcode classification helpers are all emitted
/// by `harn_opcode_macros::define_opcodes!` in [`crate::vm::ops`].
/// Re-exported here so callers that import `crate::chunk::Op` need no
/// awareness of the macro layout.
pub use crate::vm::ops::Op;
pub(crate) use crate::vm::ops::{is_adaptive_binary_op, op_reads_outer_name};

/// A constant value in the constant pool.
#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
pub enum Constant {
    Int(i64),
    Float(f64),
    String(String),
    Bool(bool),
    Nil,
    Duration(i64),
}

/// Runtime-only inline-cache state for bytecode instructions that repeatedly
/// see the same dynamic shape. Lookup caches stay monomorphic on a name and
/// receiver shape. Adaptive caches warm on a stable operand or call target,
/// then fall back through the generic opcode and replace or reset state when
/// the observed shape changes.
///
/// This vector is intentionally excluded from [`CachedChunk`]: bytecode cache
/// artifacts keep the slot layout but start with empty runtime feedback in each
/// process.
#[derive(Debug, Clone, PartialEq, Eq)]
pub(crate) enum InlineCacheEntry {
    Empty,
    Property {
        name_idx: u16,
        target: PropertyCacheTarget,
    },
    Method {
        name_idx: u16,
        argc: usize,
        target: MethodCacheTarget,
    },
    AdaptiveBinary {
        op: AdaptiveBinaryOp,
        state: AdaptiveBinaryState,
    },
    DirectCall {
        state: DirectCallState,
    },
}

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub(crate) enum AdaptiveBinaryOp {
    Add,
    Sub,
    Mul,
    Div,
    Mod,
    Equal,
    NotEqual,
    Less,
    Greater,
    LessEqual,
    GreaterEqual,
}

/// Adaptive-binary IC state. All fields are scalar `Copy` (shape is a
/// `Copy` enum, hit/miss counters are integers), so the struct as a whole
/// is `Copy`. This lets `execute_adaptive_binary` extract the cached state
/// by value for the specialization check without cloning the wrapping
/// `InlineCacheEntry` on every dispatch.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub(crate) enum AdaptiveBinaryState {
    Warmup {
        shape: BinaryShape,
        hits: u8,
    },
    Specialized {
        shape: BinaryShape,
        hits: u64,
        misses: u64,
    },
}

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub(crate) enum BinaryShape {
    Int,
    Float,
    Bool,
    String,
}

#[derive(Debug, Clone)]
pub(crate) enum DirectCallState {
    Warmup {
        argc: usize,
        target: DirectCallTarget,
        hits: u8,
    },
    Specialized {
        argc: usize,
        target: DirectCallTarget,
        hits: u64,
        misses: u64,
    },
}

#[derive(Debug, Clone)]
pub(crate) enum DirectCallTarget {
    Closure(Arc<crate::value::VmClosure>),
}

impl PartialEq for DirectCallTarget {
    fn eq(&self, other: &Self) -> bool {
        match (self, other) {
            (Self::Closure(left), Self::Closure(right)) => Arc::ptr_eq(left, right),
        }
    }
}

impl Eq for DirectCallTarget {}

impl PartialEq for DirectCallState {
    fn eq(&self, other: &Self) -> bool {
        match (self, other) {
            (
                Self::Warmup {
                    argc: left_argc,
                    target: left_target,
                    hits: left_hits,
                },
                Self::Warmup {
                    argc: right_argc,
                    target: right_target,
                    hits: right_hits,
                },
            ) => left_argc == right_argc && left_target == right_target && left_hits == right_hits,
            (
                Self::Specialized {
                    argc: left_argc,
                    target: left_target,
                    hits: left_hits,
                    misses: left_misses,
                },
                Self::Specialized {
                    argc: right_argc,
                    target: right_target,
                    hits: right_hits,
                    misses: right_misses,
                },
            ) => {
                left_argc == right_argc
                    && left_target == right_target
                    && left_hits == right_hits
                    && left_misses == right_misses
            }
            _ => false,
        }
    }
}

impl Eq for DirectCallState {}

#[derive(Debug, Clone, PartialEq, Eq)]
pub(crate) enum PropertyCacheTarget {
    DictField(Arc<str>),
    StructField { field_name: Arc<str>, index: usize },
    HarnessSubHandle(HarnessKind),
    ListCount,
    ListEmpty,
    ListFirst,
    ListLast,
    StringCount,
    StringEmpty,
    PairFirst,
    PairSecond,
    EnumVariant,
    EnumFields,
}

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub(crate) enum MethodCacheTarget {
    Harness(HarnessKind),
    ListCount,
    ListEmpty,
    ListContains,
    StringCount,
    StringEmpty,
    StringContains,
    DictCount,
    DictHas,
    RangeCount,
    RangeLen,
    RangeEmpty,
    RangeFirst,
    RangeLast,
    SetCount,
    SetLen,
    SetEmpty,
    SetContains,
}

/// Debug metadata for a slot-indexed local in a compiled chunk.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
pub struct LocalSlotInfo {
    pub name: String,
    pub mutable: bool,
    pub scope_depth: usize,
}

impl fmt::Display for Constant {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        match self {
            Constant::Int(n) => write!(f, "{n}"),
            Constant::Float(n) => write!(f, "{n}"),
            Constant::String(s) => write!(f, "\"{s}\""),
            Constant::Bool(b) => write!(f, "{b}"),
            Constant::Nil => write!(f, "nil"),
            Constant::Duration(ms) => write!(f, "{ms}ms"),
        }
    }
}

/// A compiled chunk of bytecode.
#[derive(Debug)]
pub struct Chunk {
    /// Runtime-only identity for VM-local inline cache storage. It is not
    /// serialized; freshly compiled or loaded chunks get new ids, while clones
    /// keep the same id because they represent the same bytecode object.
    cache_id: u64,
    /// The bytecode instructions.
    pub code: Vec<u8>,
    /// Constant pool.
    pub constants: Vec<Constant>,
    /// Source line numbers for each instruction (for error reporting).
    pub lines: Vec<u32>,
    /// Source column numbers for each instruction (for error reporting).
    /// Parallel to `lines`; 0 means no column info available.
    pub columns: Vec<u32>,
    /// Source file that this chunk was compiled from, when known. Set for
    /// chunks compiled from imported modules so runtime errors can report
    /// the correct file path for each frame instead of always pointing at
    /// the entry-point pipeline.
    pub source_file: Option<String>,
    /// Current column to use when emitting instructions (set by compiler).
    current_col: u32,
    /// Compiled function bodies (for closures).
    pub functions: Vec<CompiledFunctionRef>,
    /// Instruction offset to inline-cache slot. Slots are assigned at emit time
    /// for cacheable instructions while bytecode bytes remain immutable.
    /// Preserved as the serialization-stable representation that round-trips
    /// through [`CachedChunk`]; the runtime hot path reads
    /// [`Chunk::inline_cache_index`] instead.
    inline_cache_slots: BTreeMap<usize, usize>,
    /// Flat side-table indexed by code offset that returns the inline-cache
    /// slot index (or [`NO_INLINE_CACHE_SLOT`] for "no slot at this offset").
    /// Built alongside [`Chunk::inline_cache_slots`] at emit/load time so the
    /// per-dispatch lookup that fires on every adaptive binary op, `Op::Call`,
    /// `Op::MethodCall`, and `Op::GetProperty` is one cache-friendly `Vec`
    /// index instead of a `BTreeMap::get` (O(1) vs O(log n) with the
    /// associated pointer chasing). Derived; intentionally not serialized.
    inline_cache_index: Vec<u32>,
    /// Test/bench scratch entries for validating inline-cache transitions.
    /// Runtime execution keeps live cache entries on each `Vm` isolate so
    /// parallel workers do not contend on shared compiled chunks.
    inline_caches: Arc<Mutex<Vec<InlineCacheEntry>>>,
    /// Lazily-materialized shared string cache for `Constant::String` entries,
    /// parallel to `constants`. String constants are materialized once per
    /// unique constant; subsequent pushes are an `Arc` refcount bump.
    constant_strings: Arc<Mutex<Vec<Option<Arc<str>>>>>,
    /// Source-name metadata for slot-indexed locals in this chunk.
    pub(crate) local_slots: Vec<LocalSlotInfo>,
    /// True when this chunk's bytecode emits an opcode that resolves a
    /// name through the runtime env (`GetVar`, `SetVar`, `CallBuiltin`,
    /// `CallBuiltinSpread`, `CheckType`). The closure-call hot path uses
    /// this as a cheap static guard: if a closure body never reads
    /// outer names by name, the caller-scope late-bind walks in
    /// [`Vm::closure_call_env`] and
    /// [`Vm::closure_call_env_for_current_frame`] are pure overhead and
    /// can be skipped, leaving the closure's captured env as-is.
    ///
    /// Walks exist to inject late-bound closure-typed names — typically
    /// for self/mutually-recursive local fns and for fns whose captured
    /// env predates a sibling definition. Inline arithmetic / comparison
    /// callbacks (the `.map(x -> x * 2)` / `.filter(x -> x % 2 == 0)`
    /// shape) emit none of the flagged opcodes, so the walk is wasted
    /// work on every invocation.
    pub(crate) references_outer_names: bool,
    /// Compile-time operand-stack-depth tracking for the debug-build
    /// balance assertion (issue #2622). `balance_depth` is the running net
    /// effect of every *linearly-modeled* opcode emitted so far;
    /// `balance_nonlinear` counts emits whose effect can't be tracked by a
    /// straight-line sum (jumps, `return`, async/handler ops, variadic ops
    /// whose count isn't an emit argument). A statement is "balance-exact"
    /// only when `balance_nonlinear` is unchanged across its compilation,
    /// at which point `balance_depth`'s delta is its true net stack effect.
    /// Transient compile-time state: reset by [`Chunk::new`], never
    /// serialized into [`CachedChunk`], and read only by debug assertions —
    /// so a wrong absolute value (which a non-exact statement can leave
    /// behind) is harmless; only per-statement *deltas over exact spans*
    /// are ever trusted.
    #[cfg(debug_assertions)]
    balance_depth: i32,
    #[cfg(debug_assertions)]
    balance_nonlinear: u32,
}

pub type ChunkRef = Arc<Chunk>;
pub type CompiledFunctionRef = Arc<CompiledFunction>;

impl Clone for Chunk {
    fn clone(&self) -> Self {
        Self {
            cache_id: self.cache_id,
            code: self.code.clone(),
            constants: self.constants.clone(),
            lines: self.lines.clone(),
            columns: self.columns.clone(),
            source_file: self.source_file.clone(),
            current_col: self.current_col,
            functions: self.functions.clone(),
            inline_cache_slots: self.inline_cache_slots.clone(),
            inline_cache_index: self.inline_cache_index.clone(),
            inline_caches: Arc::new(Mutex::new(vec![
                InlineCacheEntry::Empty;
                self.inline_cache_slot_count()
            ])),
            constant_strings: Arc::new(Mutex::new(vec![None; self.constants.len()])),
            local_slots: self.local_slots.clone(),
            references_outer_names: self.references_outer_names,
            #[cfg(debug_assertions)]
            balance_depth: self.balance_depth,
            #[cfg(debug_assertions)]
            balance_nonlinear: self.balance_nonlinear,
        }
    }
}

/// Serializable snapshot of a [`Chunk`] suitable for the on-disk bytecode
/// cache and for in-memory stdlib artifact caches. Inline-cache state is
/// dropped at freeze time because it warms at runtime per VM isolate; the
/// rest of the chunk round-trips byte-identically.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct CachedChunk {
    pub(crate) code: Vec<u8>,
    pub(crate) constants: Vec<Constant>,
    pub(crate) lines: Vec<u32>,
    pub(crate) columns: Vec<u32>,
    pub(crate) source_file: Option<String>,
    pub(crate) current_col: u32,
    pub(crate) functions: Vec<CachedCompiledFunction>,
    pub(crate) inline_cache_slots: BTreeMap<usize, usize>,
    pub(crate) local_slots: Vec<LocalSlotInfo>,
    #[serde(default)]
    pub(crate) references_outer_names: bool,
}

#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct CachedCompiledFunction {
    pub(crate) name: String,
    pub(crate) type_params: Vec<String>,
    pub(crate) nominal_type_names: Vec<String>,
    pub(crate) params: Vec<CachedParamSlot>,
    pub(crate) default_start: Option<usize>,
    pub(crate) chunk: CachedChunk,
    pub(crate) is_generator: bool,
    pub(crate) is_stream: bool,
    pub(crate) has_rest_param: bool,
    pub(crate) has_runtime_type_checks: bool,
}

#[derive(Debug, Clone, Serialize, Deserialize)]
pub(crate) struct CachedParamSlot {
    pub(crate) name: String,
    pub(crate) type_expr: Option<TypeExpr>,
    pub(crate) has_default: bool,
}

impl CachedParamSlot {
    fn thaw(&self) -> ParamSlot {
        ParamSlot {
            name: self.name.clone(),
            type_expr: self.type_expr.clone(),
            runtime_guard: self
                .type_expr
                .as_ref()
                .map(RuntimeParamGuard::from_type_expr),
            has_default: self.has_default,
        }
    }
}

/// One parameter slot of a compiled user-defined function. Carries the
/// declared name, the (optional) declared type expression, and a flag
/// for whether a default value was provided. The runtime consults the
/// type expression in `bind_param_slots` to enforce declared types
/// against the values supplied at the call site.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct ParamSlot {
    pub name: String,
    /// Declared parameter type. `None` for untyped parameters (gradual
    /// typing); the runtime skips type assertion when absent.
    pub type_expr: Option<TypeExpr>,
    /// Precomputed runtime validation metadata derived from `type_expr`.
    /// Bytecode-cache artifacts omit this field and rebuild it at load time.
    #[serde(skip)]
    pub(crate) runtime_guard: Option<RuntimeParamGuard>,
    /// True when the parameter has a default-value clause. Diagnostic
    /// only — the canonical authority for arity ranges is
    /// [`CompiledFunction::default_start`].
    pub has_default: bool,
}

impl ParamSlot {
    /// Build a [`ParamSlot`] from a parser-side [`harn_parser::TypedParam`].
    /// Centralizes the conversion so every compile path stays in lockstep.
    pub fn from_typed_param(param: &harn_parser::TypedParam) -> Self {
        Self {
            name: param.name.clone(),
            type_expr: param.type_expr.clone(),
            runtime_guard: param
                .type_expr
                .as_ref()
                .map(RuntimeParamGuard::from_type_expr),
            has_default: param.default_value.is_some(),
        }
    }

    fn freeze_for_cache(&self) -> CachedParamSlot {
        CachedParamSlot {
            name: self.name.clone(),
            type_expr: self.type_expr.clone(),
            has_default: self.has_default,
        }
    }

    /// Build a `Vec<ParamSlot>` from a slice of parser-side typed
    /// parameters. Used pervasively at compile sites instead of
    /// `TypedParam::names` (which discarded the type info we now need
    /// at runtime).
    pub fn vec_from_typed(params: &[harn_parser::TypedParam]) -> Vec<Self> {
        params.iter().map(Self::from_typed_param).collect()
    }
}

/// A compiled function (closure body).
#[derive(Debug, Clone)]
pub struct CompiledFunction {
    pub name: String,
    /// Generic type parameters declared by this function. Runtime
    /// validation treats these as static-only constraints because the VM
    /// does not monomorphize function bodies.
    pub type_params: Vec<String>,
    /// User-defined struct and enum names visible when this function was
    /// compiled. These are the only non-primitive named types with runtime
    /// nominal identity; aliases and interfaces remain static-only.
    pub nominal_type_names: Vec<String>,
    pub params: Vec<ParamSlot>,
    /// Index of the first parameter with a default value, or None if all required.
    pub default_start: Option<usize>,
    pub chunk: ChunkRef,
    /// True if the function body contains `yield` expressions (generator function).
    pub is_generator: bool,
    /// True if the function was declared as `gen fn` and should return Stream.
    pub is_stream: bool,
    /// True if the last parameter is a rest parameter (`...name`).
    pub has_rest_param: bool,
    /// True when at least one parameter has a runtime-visible type
    /// assertion. Untyped closures dominate collection callback hot paths,
    /// so this lets the VM skip the per-argument metadata walk after the
    /// arity check.
    pub has_runtime_type_checks: bool,
}

impl CompiledFunction {
    pub(crate) fn has_runtime_type_checks_for_params(params: &[ParamSlot]) -> bool {
        params.iter().any(|param| param.type_expr.is_some())
    }

    /// Returns just the parameter names — convenience for code paths that
    /// don't care about types or defaults.
    pub fn param_names(&self) -> impl Iterator<Item = &str> {
        self.params.iter().map(|p| p.name.as_str())
    }

    /// Number of required parameters (those before `default_start`).
    pub fn required_param_count(&self) -> usize {
        self.default_start.unwrap_or(self.params.len())
    }

    pub fn declares_type_param(&self, name: &str) -> bool {
        self.type_params.iter().any(|param| param == name)
    }

    pub fn has_nominal_type(&self, name: &str) -> bool {
        self.nominal_type_names.iter().any(|ty| ty == name)
    }

    pub(crate) fn freeze_for_cache(&self) -> CachedCompiledFunction {
        CachedCompiledFunction {
            name: self.name.clone(),
            type_params: self.type_params.clone(),
            nominal_type_names: self.nominal_type_names.clone(),
            params: self
                .params
                .iter()
                .map(ParamSlot::freeze_for_cache)
                .collect(),
            default_start: self.default_start,
            chunk: self.chunk.freeze_for_cache(),
            is_generator: self.is_generator,
            is_stream: self.is_stream,
            has_rest_param: self.has_rest_param,
            has_runtime_type_checks: self.has_runtime_type_checks,
        }
    }

    pub(crate) fn from_cached(cached: &CachedCompiledFunction) -> Self {
        Self {
            name: cached.name.clone(),
            type_params: cached.type_params.clone(),
            nominal_type_names: cached.nominal_type_names.clone(),
            params: cached.params.iter().map(CachedParamSlot::thaw).collect(),
            default_start: cached.default_start,
            chunk: Arc::new(Chunk::from_cached(&cached.chunk)),
            is_generator: cached.is_generator,
            is_stream: cached.is_stream,
            has_rest_param: cached.has_rest_param,
            has_runtime_type_checks: cached.has_runtime_type_checks,
        }
    }
}

/// A snapshot of [`Chunk`]'s compile-time balance model, returned by
/// [`Chunk::balance_probe`] and consumed by [`Chunk::balance_delta_since`].
#[cfg(debug_assertions)]
#[derive(Clone, Copy)]
pub(crate) struct BalanceProbe {
    depth: i32,
    nonlinear: u32,
}

/// Net operand-stack effect (`pushes - pops`) of one emitted opcode, for
/// the debug-build balance assertion (issue #2622). `count` is the opcode's
/// variadic arity when that arity is the emit-call argument (`BuildList`
/// length, `Call` argc, …) and `0` otherwise.
///
/// `Some(delta)` means the effect is exactly modeled. `None` marks an
/// opcode a straight-line running sum can't track — control flow that
/// branches or terminates (`Jump*`, `Return`, `Throw`, `TailCall`),
/// async/handler ops, and variadic ops whose arity rides in a raw operand
/// byte rather than the emit argument (`BuildEnum`, `MatchEnum`). Such an
/// opcode taints its enclosing statement as non-exact, so the assertion
/// skips it instead of risking a false trip.
///
/// The `match` is intentionally exhaustive with no `_` arm: adding an
/// opcode forces a classification here (a compile error otherwise), so the
/// balance model can't silently drift out of sync with the instruction set.
#[cfg(debug_assertions)]
fn op_stack_delta(op: Op, count: u16) -> Option<i32> {
    use Op::*;
    let count = count as i32;
    Some(match op {
        // Push one value.
        Constant | Nil | True | False | GetVar | GetArgc | GetLocalSlot | Closure | Dup => 1,
        // Consume one value (into a binding / property / discard). `SetVar`,
        // `SetProperty` and the local-slot stores read their target by name
        // or slot index, so they only pop the value being stored.
        DefLet | DefVar | SetVar | DefLocalSlot | SetLocalSlot | SetProperty | Pop => -1,
        // Value-preserving: unary ops, by-name lookups/checks, and scope /
        // iterator / exception-handler bookkeeping (the last three touch
        // side stacks, not the operand stack).
        Negate | Not | GetProperty | GetPropertyOpt | CheckType | TryUnwrap | TryWrapOk | Swap
        | PushScope | PopScope | PopIterator | PopHandler => 0,
        // Pop two, push one.
        Add | Sub | Mul | Div | Mod | Pow | AddInt | SubInt | MulInt | DivInt | ModInt
        | AddFloat | SubFloat | MulFloat | DivFloat | ModFloat | Equal | NotEqual | Less
        | Greater | LessEqual | GreaterEqual | EqualInt | NotEqualInt | LessInt | GreaterInt
        | LessEqualInt | GreaterEqualInt | EqualFloat | NotEqualFloat | LessFloat
        | GreaterFloat | LessEqualFloat | GreaterEqualFloat | EqualBool | NotEqualBool
        | EqualString | NotEqualString | Contains | Subscript | SubscriptOpt => -1,
        // `IterInit` consumes the iterable and pushes nothing (the iterator
        // lives on a side stack).
        IterInit => -1,
        // Pop three (or two values + a by-name target), push one.
        Slice | SetSubscript => -2,
        // Variadic whose arity is the emit argument: pop `count`, push one.
        BuildList | Concat | CallBuiltin => 1 - count,
        BuildDict => 1 - 2 * count,
        // Calls also pop the callee/receiver beneath the args.
        Call | MethodCall | MethodCallOpt => -count,
        // Non-linear (see doc comment): branches, terminators, async/handler
        // ops, and variadic ops whose arity isn't the emit argument.
        Jump | JumpIfFalse | JumpIfTrue | IterNext | Return | TailCall | Throw | TryCatchSetup
        | Spawn | Pipe | Parallel | ParallelMap | ParallelMapStream | ParallelSettle
        | SyncMutexEnter | Import | SelectiveImport | DeadlineSetup | DeadlineEnd | BuildEnum
        | MatchEnum | Yield | CallSpread | CallBuiltinSpread | MethodCallSpread => return None,
    })
}

impl Chunk {
    pub fn new() -> Self {
        Self {
            cache_id: next_chunk_cache_id(),
            code: Vec::new(),
            constants: Vec::new(),
            lines: Vec::new(),
            columns: Vec::new(),
            source_file: None,
            current_col: 0,
            functions: Vec::new(),
            inline_cache_slots: BTreeMap::new(),
            inline_cache_index: Vec::new(),
            inline_caches: Arc::new(Mutex::new(Vec::new())),
            constant_strings: Arc::new(Mutex::new(Vec::new())),
            local_slots: Vec::new(),
            references_outer_names: false,
            #[cfg(debug_assertions)]
            balance_depth: 0,
            #[cfg(debug_assertions)]
            balance_nonlinear: 0,
        }
    }

    /// Set the current column for subsequent emit calls.
    pub fn set_column(&mut self, col: u32) {
        self.current_col = col;
    }

    /// Add a constant and return its index.
    pub fn add_constant(&mut self, constant: Constant) -> u16 {
        for (i, c) in self.constants.iter().enumerate() {
            if c == &constant {
                return i as u16;
            }
        }
        let idx = self.constants.len();
        self.constants.push(constant);
        idx as u16
    }

    /// Emit a single-byte instruction.
    pub fn emit(&mut self, op: Op, line: u32) {
        #[cfg(debug_assertions)]
        self.note_balance(op, 0);
        let col = self.current_col;
        let op_offset = self.code.len();
        self.code.push(op as u8);
        self.lines.push(line);
        self.columns.push(col);
        if is_adaptive_binary_op(op) {
            self.register_inline_cache(op_offset);
        }
        if op_reads_outer_name(op) {
            self.references_outer_names = true;
        }
    }

    /// Emit an instruction with a u16 argument.
    pub fn emit_u16(&mut self, op: Op, arg: u16, line: u32) {
        #[cfg(debug_assertions)]
        self.note_balance(op, arg);
        let col = self.current_col;
        let op_offset = self.code.len();
        self.code.push(op as u8);
        self.code.push((arg >> 8) as u8);
        self.code.push((arg & 0xFF) as u8);
        self.lines.push(line);
        self.lines.push(line);
        self.lines.push(line);
        self.columns.push(col);
        self.columns.push(col);
        self.columns.push(col);
        if matches!(
            op,
            Op::GetProperty | Op::GetPropertyOpt | Op::MethodCallSpread
        ) {
            self.register_inline_cache(op_offset);
        }
        if op_reads_outer_name(op) {
            self.references_outer_names = true;
        }
    }

    /// Emit an instruction with a u8 argument.
    pub fn emit_u8(&mut self, op: Op, arg: u8, line: u32) {
        #[cfg(debug_assertions)]
        self.note_balance(op, arg as u16);
        let col = self.current_col;
        let op_offset = self.code.len();
        self.code.push(op as u8);
        self.code.push(arg);
        self.lines.push(line);
        self.lines.push(line);
        self.columns.push(col);
        self.columns.push(col);
        if matches!(op, Op::Call) {
            self.register_inline_cache(op_offset);
        }
        if op_reads_outer_name(op) {
            self.references_outer_names = true;
        }
    }

    /// Emit a direct builtin call.
    pub fn emit_call_builtin(
        &mut self,
        id: crate::BuiltinId,
        name_idx: u16,
        arg_count: u8,
        line: u32,
    ) {
        #[cfg(debug_assertions)]
        self.note_balance(Op::CallBuiltin, arg_count as u16);
        let col = self.current_col;
        let op_offset = self.code.len();
        self.code.push(Op::CallBuiltin as u8);
        self.code.extend_from_slice(&id.raw().to_be_bytes());
        self.code.push((name_idx >> 8) as u8);
        self.code.push((name_idx & 0xFF) as u8);
        self.code.push(arg_count);
        for _ in 0..12 {
            self.lines.push(line);
            self.columns.push(col);
        }
        self.register_inline_cache(op_offset);
        self.references_outer_names = true;
    }

    /// Emit a direct builtin spread call.
    pub fn emit_call_builtin_spread(&mut self, id: crate::BuiltinId, name_idx: u16, line: u32) {
        #[cfg(debug_assertions)]
        self.note_balance(Op::CallBuiltinSpread, 0);
        let col = self.current_col;
        self.code.push(Op::CallBuiltinSpread as u8);
        self.code.extend_from_slice(&id.raw().to_be_bytes());
        self.code.push((name_idx >> 8) as u8);
        self.code.push((name_idx & 0xFF) as u8);
        for _ in 0..11 {
            self.lines.push(line);
            self.columns.push(col);
        }
        self.references_outer_names = true;
    }

    /// Emit a method call: op + u16 (method name) + u8 (arg count).
    pub fn emit_method_call(&mut self, name_idx: u16, arg_count: u8, line: u32) {
        self.emit_method_call_inner(Op::MethodCall, name_idx, arg_count, line);
    }

    /// Emit an optional method call (?.) — returns nil if receiver is nil.
    pub fn emit_method_call_opt(&mut self, name_idx: u16, arg_count: u8, line: u32) {
        self.emit_method_call_inner(Op::MethodCallOpt, name_idx, arg_count, line);
    }

    fn emit_method_call_inner(&mut self, op: Op, name_idx: u16, arg_count: u8, line: u32) {
        #[cfg(debug_assertions)]
        self.note_balance(op, arg_count as u16);
        let col = self.current_col;
        let op_offset = self.code.len();
        self.code.push(op as u8);
        self.code.push((name_idx >> 8) as u8);
        self.code.push((name_idx & 0xFF) as u8);
        self.code.push(arg_count);
        self.lines.push(line);
        self.lines.push(line);
        self.lines.push(line);
        self.lines.push(line);
        self.columns.push(col);
        self.columns.push(col);
        self.columns.push(col);
        self.columns.push(col);
        self.register_inline_cache(op_offset);
    }

    /// Current code offset (for jump patching).
    pub fn current_offset(&self) -> usize {
        self.code.len()
    }

    /// Emit a jump instruction with a placeholder offset. Returns the position to patch.
    pub fn emit_jump(&mut self, op: Op, line: u32) -> usize {
        #[cfg(debug_assertions)]
        self.note_balance(op, 0);
        let col = self.current_col;
        self.code.push(op as u8);
        let patch_pos = self.code.len();
        self.code.push(0xFF);
        self.code.push(0xFF);
        self.lines.push(line);
        self.lines.push(line);
        self.lines.push(line);
        self.columns.push(col);
        self.columns.push(col);
        self.columns.push(col);
        patch_pos
    }

    /// Patch a jump instruction at the given position to jump to the current offset.
    pub fn patch_jump(&mut self, patch_pos: usize) {
        let target = self.code.len() as u16;
        self.code[patch_pos] = (target >> 8) as u8;
        self.code[patch_pos + 1] = (target & 0xFF) as u8;
    }

    /// Patch a jump to a specific target position.
    pub fn patch_jump_to(&mut self, patch_pos: usize, target: usize) {
        let target = target as u16;
        self.code[patch_pos] = (target >> 8) as u8;
        self.code[patch_pos + 1] = (target & 0xFF) as u8;
    }

    /// Read a u16 argument at the given position.
    pub fn read_u16(&self, pos: usize) -> u16 {
        ((self.code[pos] as u16) << 8) | (self.code[pos + 1] as u16)
    }

    /// Fold one just-emitted opcode into the compile-time operand-stack
    /// balance model (issue #2622). See [`op_stack_delta`] for the
    /// linear-vs-non-linear classification.
    #[cfg(debug_assertions)]
    fn note_balance(&mut self, op: Op, count: u16) {
        match op_stack_delta(op, count) {
            Some(delta) => self.balance_depth += delta,
            None => self.balance_nonlinear += 1,
        }
    }

    /// Snapshot the balance model before compiling a statement; pair with
    /// [`Chunk::balance_delta_since`].
    #[cfg(debug_assertions)]
    pub(crate) fn balance_probe(&self) -> BalanceProbe {
        BalanceProbe {
            depth: self.balance_depth,
            nonlinear: self.balance_nonlinear,
        }
    }

    /// Net operand-stack effect emitted since `probe`, or `None` when any
    /// non-linearly-modeled opcode was emitted in that span (which makes
    /// the running sum untrustworthy, so callers must not assert on it).
    /// The absolute `balance_depth` may be meaningless after a non-exact
    /// span — only deltas over a fully-exact span are valid.
    #[cfg(debug_assertions)]
    pub(crate) fn balance_delta_since(&self, probe: BalanceProbe) -> Option<i32> {
        if self.balance_nonlinear == probe.nonlinear {
            Some(self.balance_depth - probe.depth)
        } else {
            None
        }
    }

    fn register_inline_cache(&mut self, op_offset: usize) {
        if self.inline_cache_slots.contains_key(&op_offset) {
            return;
        }
        let mut entries = self.inline_caches.lock();
        let slot = entries.len();
        entries.push(InlineCacheEntry::Empty);
        self.inline_cache_slots.insert(op_offset, slot);
        Self::write_inline_cache_index(&mut self.inline_cache_index, op_offset, slot);
    }

    /// Fast-path side-table writer. Pulled out as an associated fn so both
    /// the live emit path and [`Chunk::from_cached`] share the same growth
    /// strategy. Cache slots fit comfortably in `u32` because the slot count
    /// is bounded by the cacheable-opcode count in `code`.
    fn write_inline_cache_index(index: &mut Vec<u32>, op_offset: usize, slot: usize) {
        if op_offset >= index.len() {
            index.resize(op_offset + 1, NO_INLINE_CACHE_SLOT);
        }
        index[op_offset] = slot as u32;
    }

    /// Look up the inline-cache slot for the opcode at `op_offset`. This is
    /// called on every dispatch of an adaptive binary op (Add/Sub/Mul/Div/
    /// Mod/Eq/Neq/Less/Greater/LessEq/GreaterEq), `Op::Call`, `Op::MethodCall`
    /// (and `MethodCallOpt`/`MethodCallSpread`), and `Op::GetProperty`
    /// (`GetPropertyOpt`). Backed by [`Chunk::inline_cache_index`] — a flat
    /// `Vec<u32>` indexed by code offset — so the lookup is a single bounds-
    /// checked array read instead of the prior `BTreeMap::get` which walked
    /// internal nodes for every dispatched op.
    #[inline]
    pub(crate) fn inline_cache_slot(&self, op_offset: usize) -> Option<usize> {
        match self.inline_cache_index.get(op_offset).copied() {
            None | Some(NO_INLINE_CACHE_SLOT) => None,
            Some(slot) => Some(slot as usize),
        }
    }

    pub(crate) fn inline_cache_slot_count(&self) -> usize {
        self.inline_cache_slots.len()
    }

    pub(crate) fn cache_id(&self) -> u64 {
        self.cache_id
    }

    /// Pre-optimization control path: the `BTreeMap`-backed lookup the
    /// dispatcher used before the flat `Vec<u32>` side-table. Exposed
    /// only behind the `vm-bench-internals` feature so the criterion
    /// microbench can A/B the two paths inside one binary on identical
    /// hardware. The production hot path must keep using
    /// [`Chunk::inline_cache_slot`].
    #[cfg(feature = "vm-bench-internals")]
    pub fn inline_cache_slot_via_btreemap_for_bench(&self, op_offset: usize) -> Option<usize> {
        self.inline_cache_slots.get(&op_offset).copied()
    }

    /// Returns a shared string for a `Constant::String` at the given pool
    /// index, materializing it on first access and caching for reuse.
    /// Returns `None` when the constant at `idx` is not a string (the
    /// caller should fall back to the regular `Constant` match).
    pub(crate) fn constant_string_rc(&self, idx: usize) -> Option<Arc<str>> {
        // Borrow the side table mutably so we can lazily extend / fill
        // entries. The borrow is scope-confined to this function; the
        // VM never re-enters constant_string_rc for the same chunk
        // during a single materialization, so no nested-borrow risk.
        let mut entries = self.constant_strings.lock();
        if entries.len() < self.constants.len() {
            entries.resize(self.constants.len(), None);
        }
        if let Some(Some(existing)) = entries.get(idx) {
            return Some(Arc::clone(existing));
        }
        let materialized = match self.constants.get(idx)? {
            Constant::String(s) => Arc::<str>::from(s.as_str()),
            _ => return None,
        };
        entries[idx] = Some(Arc::clone(&materialized));
        Some(materialized)
    }

    #[cfg(feature = "vm-bench-internals")]
    pub(crate) fn inline_cache_entry(&self, slot: usize) -> InlineCacheEntry {
        self.inline_caches
            .lock()
            .get(slot)
            .cloned()
            .unwrap_or(InlineCacheEntry::Empty)
    }

    /// Adaptive-binary fast path read. Returns the cached
    /// `(op, state)` pair by value (both `Copy`) when slot holds an
    /// `AdaptiveBinary` entry, else `None`. Skips the
    /// `InlineCacheEntry::clone` that `inline_cache_entry` performs:
    /// since `AdaptiveBinaryState: Copy`, the read does a single
    /// scalar move out of the cache instead of a 24-32B memcpy of the
    /// wrapping enum (which the variant-checking match destructures
    /// and throws away anyway). Fires on every Add/Sub/Mul/Div/Mod/Eq/
    /// Neq/Less/Greater/LessEq/GreaterEq dispatch, so the per-op
    /// savings compound across the millions of dispatches a typical
    /// loop body issues.
    #[inline]
    #[cfg(any(test, feature = "vm-bench-internals"))]
    pub(crate) fn peek_adaptive_binary_cache(
        &self,
        slot: usize,
    ) -> Option<(AdaptiveBinaryOp, AdaptiveBinaryState)> {
        match self.inline_caches.lock().get(slot)? {
            &InlineCacheEntry::AdaptiveBinary { op, state } => Some((op, state)),
            _ => None,
        }
    }

    /// Method-cache fast path read. Returns the cached `(name_idx, argc,
    /// target)` triple by value (all three are `Copy`) when `slot` holds a
    /// `Method` entry, else `None`. Skips the full `InlineCacheEntry::clone`
    /// that `inline_cache_entry` performs on every `Op::MethodCall`,
    /// `Op::MethodCallOpt`, and `Op::MethodCallSpread` dispatch: the
    /// variant-checking `let-else` in `try_cached_method` destructures and
    /// throws the wrapping enum away anyway, so reading the payload by `Copy`
    /// avoids the 32-48B enum memcpy. Method-call dispatch is the second-
    /// hottest IC-keyed opcode class after the adaptive binary ops, so the
    /// per-dispatch savings compound across the millions of method calls a
    /// typical pipeline (`xs.filter(...).map(...).count()`) issues.
    #[inline]
    #[cfg(any(test, feature = "vm-bench-internals"))]
    pub(crate) fn peek_method_cache(&self, slot: usize) -> Option<(u16, usize, MethodCacheTarget)> {
        match self.inline_caches.lock().get(slot)? {
            &InlineCacheEntry::Method {
                name_idx,
                argc,
                target,
            } => Some((name_idx, argc, target)),
            _ => None,
        }
    }

    /// Property-cache fast path read. Returns the cached `(name_idx, target)`
    /// pair by value when `slot` holds a `Property` entry, else `None`. The
    /// outer `InlineCacheEntry` is the worst-case-sized variant (DirectCall
    /// at ~48 bytes including padding); cloning it just to discard four other
    /// variants in `try_cached_property`'s variant-check is wasted work. The
    /// peek returns just the `Property` payload (`u16` + `PropertyCacheTarget`),
    /// skipping the outer enum tag init and the padding-to-largest-variant
    /// memcpy. Fires on every `Op::GetProperty` / `Op::GetPropertyOpt`
    /// dispatch, which is the dominant opcode for any field-read-heavy code.
    #[inline]
    #[cfg(any(test, feature = "vm-bench-internals"))]
    pub(crate) fn peek_property_cache(&self, slot: usize) -> Option<(u16, PropertyCacheTarget)> {
        match self.inline_caches.lock().get(slot)? {
            InlineCacheEntry::Property { name_idx, target } => Some((*name_idx, target.clone())),
            _ => None,
        }
    }

    /// Direct-call cache state read. Returns just the inner `DirectCallState`
    /// by value when `slot` holds a `DirectCall` entry, else `None`. Used by
    /// both `try_cached_direct_call(_)` (steady-state Specialized hit check)
    /// and `next_direct_call_entry` (Warmup → Specialized state-machine
    /// transition). Peeking the inner state directly skips the outer
    /// `InlineCacheEntry` discriminant check and tag init that the dispatcher
    /// otherwise pays on every `Op::Call` (closure callee) and the named-fn
    /// fast path inside `Op::CallBuiltin`. Single peek per dispatch covers
    /// both the read check and the write-back computation.
    #[inline]
    #[cfg(any(test, feature = "vm-bench-internals"))]
    pub(crate) fn peek_direct_call_state(&self, slot: usize) -> Option<DirectCallState> {
        match self.inline_caches.lock().get(slot)? {
            InlineCacheEntry::DirectCall { state } => Some(state.clone()),
            _ => None,
        }
    }

    #[cfg(any(test, feature = "vm-bench-internals"))]
    pub(crate) fn set_inline_cache_entry(&self, slot: usize, entry: InlineCacheEntry) {
        if let Some(existing) = self.inline_caches.lock().get_mut(slot) {
            *existing = entry;
        }
    }

    pub fn freeze_for_cache(&self) -> CachedChunk {
        CachedChunk {
            code: self.code.clone(),
            constants: self.constants.clone(),
            lines: self.lines.clone(),
            columns: self.columns.clone(),
            source_file: self.source_file.clone(),
            current_col: self.current_col,
            functions: self
                .functions
                .iter()
                .map(|function| function.freeze_for_cache())
                .collect(),
            inline_cache_slots: self.inline_cache_slots.clone(),
            local_slots: self.local_slots.clone(),
            references_outer_names: self.references_outer_names,
        }
    }

    pub fn from_cached(cached: &CachedChunk) -> Self {
        let inline_cache_count = cached.inline_cache_slots.len();
        let constants_count = cached.constants.len();
        // Project the cached `BTreeMap<op_offset, slot>` into the flat
        // dispatch-side lookup table. Sized to `code.len()` so the hottest
        // hot opcodes (binary ops at the end of a long chunk) still hit the
        // fast-path bounds check rather than falling through to the
        // none-found branch. The size is bounded by code length, so the
        // memory footprint is tiny — a few KB for typical chunks.
        let mut inline_cache_index = Vec::new();
        inline_cache_index.resize(cached.code.len(), NO_INLINE_CACHE_SLOT);
        for (&op_offset, &slot) in cached.inline_cache_slots.iter() {
            if op_offset < inline_cache_index.len() {
                inline_cache_index[op_offset] = slot as u32;
            }
        }
        Self {
            cache_id: next_chunk_cache_id(),
            code: cached.code.clone(),
            constants: cached.constants.clone(),
            lines: cached.lines.clone(),
            columns: cached.columns.clone(),
            source_file: cached.source_file.clone(),
            current_col: cached.current_col,
            functions: cached
                .functions
                .iter()
                .map(|function| Arc::new(CompiledFunction::from_cached(function)))
                .collect(),
            inline_cache_slots: cached.inline_cache_slots.clone(),
            inline_cache_index,
            inline_caches: Arc::new(Mutex::new(vec![
                InlineCacheEntry::Empty;
                inline_cache_count
            ])),
            constant_strings: Arc::new(Mutex::new(vec![None; constants_count])),
            local_slots: cached.local_slots.clone(),
            references_outer_names: cached.references_outer_names,
            #[cfg(debug_assertions)]
            balance_depth: 0,
            #[cfg(debug_assertions)]
            balance_nonlinear: 0,
        }
    }

    pub(crate) fn add_local_slot(
        &mut self,
        name: String,
        mutable: bool,
        scope_depth: usize,
    ) -> u16 {
        let idx = self.local_slots.len();
        self.local_slots.push(LocalSlotInfo {
            name,
            mutable,
            scope_depth,
        });
        idx as u16
    }

    /// Read a u64 argument at the given position.
    pub fn read_u64(&self, pos: usize) -> u64 {
        u64::from_be_bytes([
            self.code[pos],
            self.code[pos + 1],
            self.code[pos + 2],
            self.code[pos + 3],
            self.code[pos + 4],
            self.code[pos + 5],
            self.code[pos + 6],
            self.code[pos + 7],
        ])
    }

    /// Disassemble the chunk for debugging. The per-opcode rendering is
    /// macro-generated alongside the dispatch tables in
    /// `crate::vm::ops` — see [`Self::disassemble_op`].
    pub fn disassemble(&self, name: &str) -> String {
        let mut out = format!("== {name} ==\n");
        let mut ip = 0;
        while ip < self.code.len() {
            let op_byte = self.code[ip];
            let line = self.lines.get(ip).copied().unwrap_or(0);
            out.push_str(&format!("{ip:04} [{line:>4}] "));
            ip += 1;

            if let Some(op) = Op::from_byte(op_byte) {
                self.disassemble_op(op, &mut ip, &mut out);
            } else {
                out.push_str(&format!("UNKNOWN(0x{op_byte:02x})\n"));
            }
        }
        out
    }
}

/// Disassembly helpers consumed by the macro-generated
/// [`Chunk::disassemble_op`]. Each helper takes the current code position
/// (already advanced past the opcode byte), advances it over the operand
/// bytes the opcode carries, and renders one human-readable line without
/// a trailing newline (the dispatcher appends it).
///
/// Defining one helper per operand layout — and not one per opcode —
/// keeps adding an opcode a one-line edit in the `define_opcodes!` table
/// rather than a paired edit here. New layouts live with the helpers;
/// new opcodes live with the dispatch.
pub(crate) fn disasm_bare(_chunk: &Chunk, _ip: &mut usize, label: &str) -> String {
    label.to_string()
}

pub(crate) fn disasm_u8(chunk: &Chunk, ip: &mut usize, label: &str) -> String {
    let arg = chunk.code[*ip];
    *ip += 1;
    format!("{label} {arg:>4}")
}

pub(crate) fn disasm_u16(chunk: &Chunk, ip: &mut usize, label: &str) -> String {
    let arg = chunk.read_u16(*ip);
    *ip += 2;
    format!("{label} {arg:>4}")
}

pub(crate) fn disasm_const_pool_u16(chunk: &Chunk, ip: &mut usize, label: &str) -> String {
    let idx = chunk.read_u16(*ip);
    *ip += 2;
    format!("{label} {idx:>4} ({})", chunk.constants[idx as usize])
}

pub(crate) fn disasm_local_slot_u16(chunk: &Chunk, ip: &mut usize, label: &str) -> String {
    let slot = chunk.read_u16(*ip);
    *ip += 2;
    let mut out = format!("{label} {slot:>4}");
    if let Some(info) = chunk.local_slots.get(slot as usize) {
        out.push_str(&format!(" ({})", info.name));
    }
    out
}

pub(crate) fn disasm_method_call(chunk: &Chunk, ip: &mut usize, label: &str) -> String {
    let idx = chunk.read_u16(*ip);
    *ip += 2;
    let argc = chunk.code[*ip];
    *ip += 1;
    format!(
        "{label} {idx:>4} ({}) argc={argc}",
        chunk.constants[idx as usize]
    )
}

pub(crate) fn disasm_match_enum(chunk: &Chunk, ip: &mut usize, label: &str) -> String {
    let enum_idx = chunk.read_u16(*ip);
    *ip += 2;
    let var_idx = chunk.read_u16(*ip);
    *ip += 2;
    format!(
        "{label} {enum_idx:>4} ({}) {var_idx:>4} ({})",
        chunk.constants[enum_idx as usize], chunk.constants[var_idx as usize],
    )
}

pub(crate) fn disasm_build_enum(chunk: &Chunk, ip: &mut usize, label: &str) -> String {
    let enum_idx = chunk.read_u16(*ip);
    *ip += 2;
    let var_idx = chunk.read_u16(*ip);
    *ip += 2;
    let field_count = chunk.read_u16(*ip);
    *ip += 2;
    format!(
        "{label} {enum_idx:>4} ({}) {var_idx:>4} ({}) fields={field_count}",
        chunk.constants[enum_idx as usize], chunk.constants[var_idx as usize],
    )
}

pub(crate) fn disasm_selective_import(chunk: &Chunk, ip: &mut usize, label: &str) -> String {
    let path_idx = chunk.read_u16(*ip);
    *ip += 2;
    let names_idx = chunk.read_u16(*ip);
    *ip += 2;
    format!(
        "{label} {path_idx:>4} ({}) names: {names_idx:>4} ({})",
        chunk.constants[path_idx as usize], chunk.constants[names_idx as usize],
    )
}

pub(crate) fn disasm_check_type(chunk: &Chunk, ip: &mut usize, label: &str) -> String {
    let var_idx = chunk.read_u16(*ip);
    *ip += 2;
    let type_idx = chunk.read_u16(*ip);
    *ip += 2;
    format!(
        "{label} {var_idx:>4} ({}) -> {type_idx:>4} ({})",
        chunk.constants[var_idx as usize], chunk.constants[type_idx as usize],
    )
}

pub(crate) fn disasm_call_builtin(chunk: &Chunk, ip: &mut usize, label: &str) -> String {
    let id = chunk.read_u64(*ip);
    *ip += 8;
    let idx = chunk.read_u16(*ip);
    *ip += 2;
    let argc = chunk.code[*ip];
    *ip += 1;
    format!(
        "{label} {id:#018x} {idx:>4} ({}) argc={argc}",
        chunk.constants[idx as usize],
    )
}

pub(crate) fn disasm_call_builtin_spread(chunk: &Chunk, ip: &mut usize, label: &str) -> String {
    let id = chunk.read_u64(*ip);
    *ip += 8;
    let idx = chunk.read_u16(*ip);
    *ip += 2;
    format!(
        "{label} {id:#018x} {idx:>4} ({})",
        chunk.constants[idx as usize],
    )
}

pub(crate) fn disasm_method_call_spread(chunk: &Chunk, ip: &mut usize, label: &str) -> String {
    // emit_u16(Op::MethodCallSpread, name_idx, ...) writes opcode + 2
    // bytes of u16 name_idx, so the operand is read at *ip with the
    // usual `read_u16`. The previous hand-written disasm read at
    // `ip + 1`, which displayed the wrong constant index — silently
    // corrupting any disassembly that hit a `MethodCallSpread` opcode.
    let idx = chunk.read_u16(*ip);
    *ip += 2;
    format!("{label} {idx:>4} ({})", chunk.constants[idx as usize])
}

impl Default for Chunk {
    fn default() -> Self {
        Self::new()
    }
}

#[cfg(test)]
mod tests {
    use std::sync::Arc;

    use super::{
        Chunk, DirectCallState, DirectCallTarget, InlineCacheEntry, MethodCacheTarget, Op,
        PropertyCacheTarget,
    };
    use crate::BuiltinId;

    #[test]
    fn op_from_byte_matches_repr_order() {
        for (byte, op) in Op::ALL.iter().copied().enumerate() {
            assert_eq!(byte as u8, op as u8);
            assert_eq!(Op::from_byte(byte as u8), Some(op));
        }
        assert_eq!(Op::from_byte(Op::ALL.len() as u8), None);
        assert_eq!(Op::COUNT, Op::ALL.len());
    }

    #[test]
    fn disassemble_covers_every_opcode_variant() {
        // The macro-generated `disassemble_op` match is exhaustive on
        // `Op`, so this is a compile-time guarantee. The runtime check
        // pins that no helper falls through to `UNKNOWN(...)` for a
        // valid opcode byte — catching any future macro refactor that
        // silently drops a helper arm. Each opcode is exercised in
        // isolation against a hand-built chunk so the test logic does
        // not depend on operand sizes (and so a single short opcode
        // does not bleed into reading trailing padding as a follow-on
        // opcode in the chunk-level loop).
        for op in Op::ALL.iter().copied() {
            let mut chunk = Chunk::new();
            chunk.add_constant(super::Constant::String("__probe__".to_string()));
            // Pad to the worst-case operand width (CallBuiltin: u64 +
            // u16 + u8 = 11 bytes) so any helper has well-formed bytes
            // to consume regardless of its layout.
            for _ in 0..16 {
                chunk.code.push(0);
            }
            let mut ip: usize = 0;
            let mut out = String::new();
            chunk.disassemble_op(op, &mut ip, &mut out);
            assert!(
                !out.contains("UNKNOWN"),
                "disasm emitted UNKNOWN for {op:?}: {out}",
            );
            assert!(!out.is_empty(), "disasm produced no output for {op:?}");
        }
    }

    // --- references_outer_names tracking ---
    //
    // Drives the compile-time guard used in `Vm::closure_call_env`
    // and `Vm::closure_call_env_for_current_frame` to skip the
    // per-invocation caller-scope late-bind walks. Coverage parity
    // matters because false negatives would regress recursive /
    // mutually-recursive fns.

    #[test]
    fn empty_chunk_does_not_reference_outer_names() {
        let chunk = Chunk::new();
        assert!(!chunk.references_outer_names);
    }

    #[test]
    fn arithmetic_only_chunk_does_not_reference_outer_names() {
        // The hot `.map(x -> x * 2)` / `.filter(x -> x % 2 == 0)`
        // shape: pure stack/arithmetic ops and slot locals, no env
        // reads. Must NOT flag — that's the whole point of the
        // optimization.
        let mut chunk = Chunk::new();
        chunk.emit_u16(Op::GetLocalSlot, 0, 1);
        chunk.emit_u16(Op::Constant, 0, 1);
        chunk.emit(Op::MulInt, 1);
        chunk.emit(Op::Pop, 1);
        chunk.emit(Op::Return, 1);
        assert!(!chunk.references_outer_names);
    }

    #[test]
    fn slot_only_chunk_does_not_reference_outer_names() {
        // Compiler-resolved locals never need env-based late-bind.
        let mut chunk = Chunk::new();
        chunk.emit_u16(Op::DefLocalSlot, 0, 1);
        chunk.emit_u16(Op::GetLocalSlot, 0, 1);
        chunk.emit_u16(Op::SetLocalSlot, 0, 1);
        assert!(!chunk.references_outer_names);
    }

    #[test]
    fn get_var_flags_outer_name_reference() {
        let mut chunk = Chunk::new();
        chunk.emit_u16(Op::GetVar, 0, 1);
        assert!(chunk.references_outer_names);
    }

    #[test]
    fn set_var_flags_outer_name_reference() {
        let mut chunk = Chunk::new();
        chunk.emit_u16(Op::SetVar, 0, 1);
        assert!(chunk.references_outer_names);
    }

    #[test]
    fn check_type_flags_outer_name_reference() {
        let mut chunk = Chunk::new();
        chunk.emit_u16(Op::CheckType, 0, 1);
        assert!(chunk.references_outer_names);
    }

    #[test]
    fn call_builtin_flags_outer_name_reference() {
        let mut chunk = Chunk::new();
        chunk.emit_call_builtin(BuiltinId::from_name("any_name"), 0, 1, 1);
        assert!(chunk.references_outer_names);
    }

    #[test]
    fn call_builtin_spread_flags_outer_name_reference() {
        let mut chunk = Chunk::new();
        chunk.emit_call_builtin_spread(BuiltinId::from_name("any_name"), 0, 1);
        assert!(chunk.references_outer_names);
    }

    #[test]
    fn tail_call_flags_outer_name_reference() {
        // `return fn_name(...)` compiles to Constant + TailCall —
        // TailCall does a runtime name lookup, so it has to flag.
        let mut chunk = Chunk::new();
        chunk.emit_u8(Op::TailCall, 1, 1);
        assert!(chunk.references_outer_names);
    }

    #[test]
    fn call_flags_outer_name_reference() {
        // Op::Call can receive a String callee from the stack (the
        // by-name dispatch shape), so it has to flag too.
        let mut chunk = Chunk::new();
        chunk.emit_u8(Op::Call, 1, 1);
        assert!(chunk.references_outer_names);
    }

    #[test]
    fn pipe_flags_outer_name_reference() {
        // `x |> name` resolves `name` through env when the value on
        // the stack is a String / BuiltinRef.
        let mut chunk = Chunk::new();
        chunk.emit(Op::Pipe, 1);
        assert!(chunk.references_outer_names);
    }

    #[test]
    fn method_call_does_not_flag_outer_name_reference() {
        // Method receivers come off the operand stack, not the env;
        // emitting MethodCall alone must not force the walk.
        let mut chunk = Chunk::new();
        chunk.emit_method_call(0, 1, 1);
        chunk.emit_method_call_opt(0, 1, 1);
        assert!(!chunk.references_outer_names);
    }

    #[test]
    fn jump_and_control_flow_do_not_flag_outer_name_reference() {
        // Jumps, returns, pops — control flow stays inside the
        // frame and never touches env lookups.
        let mut chunk = Chunk::new();
        chunk.emit_u16(Op::Constant, 0, 1);
        chunk.emit(Op::JumpIfFalse, 1);
        chunk.emit(Op::Jump, 1);
        chunk.emit(Op::Return, 1);
        chunk.emit(Op::Pop, 1);
        assert!(!chunk.references_outer_names);
    }

    #[test]
    fn references_outer_names_is_monotonic() {
        // Once flagged, subsequent non-flagging emits must not
        // clear the bit — flags are sticky.
        let mut chunk = Chunk::new();
        chunk.emit_u16(Op::GetVar, 0, 1);
        assert!(chunk.references_outer_names);
        chunk.emit_u16(Op::GetLocalSlot, 0, 1);
        chunk.emit(Op::MulInt, 1);
        assert!(chunk.references_outer_names);
    }

    #[test]
    fn freeze_thaw_round_trips_references_outer_names() {
        // Bytecode-cache hits must observe the same flag as a
        // fresh compile — otherwise the first call after a cache
        // hit would either over- or under-skip the walk.
        let mut chunk = Chunk::new();
        chunk.emit_u16(Op::GetVar, 0, 1);
        assert!(chunk.references_outer_names);
        let frozen = chunk.freeze_for_cache();
        let thawed = Chunk::from_cached(&frozen);
        assert!(thawed.references_outer_names);

        let plain = Chunk::new();
        assert!(!plain.references_outer_names);
        let frozen_plain = plain.freeze_for_cache();
        let thawed_plain = Chunk::from_cached(&frozen_plain);
        assert!(!thawed_plain.references_outer_names);
    }

    // --- inline_cache_slot flat-index parity ---
    //
    // Slot lookups fire on every dispatch of an adaptive binary op
    // (Add/Sub/Mul/Div/Mod/Eq/Neq/Less/Greater/LessEq/GreaterEq),
    // every `Op::Call`, every `Op::MethodCall(Opt)`, and every
    // `Op::GetProperty(Opt)`. The flat `Vec<u32>` index has to stay
    // perfectly in sync with the serialization-stable BTreeMap or
    // a cached call site would either skip its inline cache (slow
    // path with no learning) or read a stale slot (silently
    // mis-specialized arithmetic). These tests pin the contract.

    #[test]
    fn inline_cache_slot_returns_none_for_non_cacheable_offsets() {
        // GetLocalSlot is a sync-fast-path opcode with no inline
        // cache; the index must report no slot.
        let mut chunk = Chunk::new();
        chunk.emit_u16(Op::GetLocalSlot, 0, 1);
        chunk.emit(Op::Pop, 1);
        chunk.emit(Op::Return, 1);
        assert!(chunk.inline_cache_slot(0).is_none());
        assert!(chunk.inline_cache_slot(3).is_none());
        assert!(chunk.inline_cache_slot(4).is_none());
    }

    #[test]
    fn inline_cache_slot_registered_for_adaptive_binary_op() {
        // Pure-arithmetic ops use the adaptive-binary IC for shape
        // specialization. The slot has to be 0 because the chunk is
        // otherwise empty.
        let mut chunk = Chunk::new();
        chunk.emit(Op::Add, 1);
        assert_eq!(chunk.inline_cache_slot(0), Some(0));
    }

    #[test]
    fn inline_cache_slot_distinct_for_sequential_adaptive_binary_ops() {
        // Three back-to-back Adds must get three distinct slots so
        // each instruction's shape feedback evolves independently
        // (otherwise the same call site would clobber a neighbor's
        // learning every dispatch).
        let mut chunk = Chunk::new();
        chunk.emit(Op::Add, 1);
        chunk.emit(Op::Sub, 1);
        chunk.emit(Op::Mul, 1);
        let s0 = chunk.inline_cache_slot(0).expect("Add slot");
        let s1 = chunk.inline_cache_slot(1).expect("Sub slot");
        let s2 = chunk.inline_cache_slot(2).expect("Mul slot");
        assert_ne!(s0, s1);
        assert_ne!(s1, s2);
        assert_ne!(s0, s2);
    }

    #[test]
    fn inline_cache_slot_returns_none_for_out_of_bounds_offset() {
        // The dispatcher derives `op_offset` from `ip - 1`; an
        // out-of-bounds query must return None rather than panic.
        let mut chunk = Chunk::new();
        chunk.emit(Op::Add, 1);
        assert!(chunk.inline_cache_slot(usize::MAX).is_none());
        assert!(chunk.inline_cache_slot(chunk.code.len()).is_none());
        assert!(chunk.inline_cache_slot(chunk.code.len() + 16).is_none());
    }

    #[test]
    fn inline_cache_slot_for_get_property_and_method_call() {
        // GetProperty(Opt) and MethodCall(Opt) both register an IC
        // slot at emit time — adaptive method-call dispatch and
        // monomorphic property-cache learning depend on it.
        let mut chunk = Chunk::new();
        chunk.emit_u16(Op::GetProperty, 0, 1); // offset 0..3
        chunk.emit_method_call(0, 1, 1); // offset 3..7
        chunk.emit_method_call_opt(0, 1, 1); // offset 7..11
        chunk.emit_u16(Op::GetPropertyOpt, 0, 1); // offset 11..14
        assert!(chunk.inline_cache_slot(0).is_some(), "GetProperty");
        assert!(chunk.inline_cache_slot(3).is_some(), "MethodCall");
        assert!(chunk.inline_cache_slot(7).is_some(), "MethodCallOpt");
        assert!(chunk.inline_cache_slot(11).is_some(), "GetPropertyOpt");
    }

    #[test]
    fn inline_cache_slot_for_call_and_call_builtin() {
        // Both `Op::Call` (closure / by-name callee) and
        // `emit_call_builtin` register IC slots. The latter is the
        // adaptive-call fast path used for every direct user-fn
        // invocation.
        let mut chunk = Chunk::new();
        chunk.emit_u8(Op::Call, 1, 1); // offset 0..2
        let call_builtin_offset = chunk.code.len();
        chunk.emit_call_builtin(BuiltinId::from_name("any"), 0, 1, 1);
        assert!(chunk.inline_cache_slot(0).is_some(), "Op::Call IC slot");
        assert!(
            chunk.inline_cache_slot(call_builtin_offset).is_some(),
            "Op::CallBuiltin IC slot"
        );
    }

    #[test]
    fn inline_cache_slot_register_is_idempotent_for_same_offset() {
        // The compile path uses `BTreeMap::contains_key` to dedup
        // re-registration at the same offset (eg. when a helper
        // re-emits into a still-live position). The flat index has
        // to honor the same semantics — never silently overwriting
        // an existing slot with a fresh one.
        let mut chunk = Chunk::new();
        chunk.emit(Op::Add, 1);
        let slot_before = chunk.inline_cache_slot(0).expect("first registration");
        // Manually re-register the same offset to confirm dedup.
        chunk.register_inline_cache(0);
        let slot_after = chunk.inline_cache_slot(0).expect("re-registration");
        assert_eq!(slot_before, slot_after);
    }

    #[test]
    fn inline_cache_index_round_trips_through_cached_chunk() {
        // The cache freeze drops the flat index (it's derived from
        // the BTreeMap that *is* serialized). On thaw, the flat
        // index must be rebuilt so the first hot dispatch of a
        // cached chunk doesn't fall off the IC-slot cliff (which
        // would silently disable shape specialization until the
        // chunk is recompiled from source).
        let mut chunk = Chunk::new();
        chunk.emit_u16(Op::GetLocalSlot, 0, 1);
        chunk.emit_u16(Op::Constant, 0, 1);
        chunk.emit(Op::Add, 1);
        chunk.emit(Op::Sub, 1);
        chunk.emit_method_call(0, 1, 1);
        chunk.emit_u8(Op::Call, 1, 1);
        let live_slots: Vec<(usize, Option<usize>)> = (0..chunk.code.len())
            .map(|o| (o, chunk.inline_cache_slot(o)))
            .collect();
        let frozen = chunk.freeze_for_cache();
        let thawed = Chunk::from_cached(&frozen);
        let thawed_slots: Vec<(usize, Option<usize>)> = (0..thawed.code.len())
            .map(|o| (o, thawed.inline_cache_slot(o)))
            .collect();
        assert_eq!(live_slots, thawed_slots);
    }

    #[test]
    fn inline_cache_index_agrees_with_btreemap_view() {
        // Authoritative parity check: for every code offset, the
        // flat-index `inline_cache_slot` must return exactly what
        // the underlying BTreeMap would (mod the `Option` boxing).
        // Catches any future emit path that grows `inline_cache_slots`
        // without going through `register_inline_cache`.
        let mut chunk = Chunk::new();
        chunk.emit(Op::Add, 1);
        chunk.emit_u16(Op::GetVar, 0, 1);
        chunk.emit(Op::LessInt, 1);
        chunk.emit_u8(Op::Call, 2, 1);
        chunk.emit(Op::Equal, 1);
        chunk.emit_u16(Op::GetProperty, 0, 1);
        chunk.emit_method_call_opt(0, 0, 1);
        for offset in 0..chunk.code.len() {
            let from_map = chunk.inline_cache_slots.get(&offset).copied();
            let from_index = chunk.inline_cache_slot(offset);
            assert_eq!(from_index, from_map, "parity broken at offset {offset}");
        }
    }

    // --- peek_adaptive_binary_cache contract ---
    //
    // The peek replaces the per-dispatch `InlineCacheEntry::clone` on the
    // hottest opcode class (Add / Sub / Mul / Div / Mod / Eq / Neq /
    // Less / Greater / LessEq / GreaterEq). It must return None for
    // unrelated IC variants — silently mis-extracting a `Property` /
    // `DirectCall` / `Method` slot as `AdaptiveBinary` would feed
    // garbage into `try_specialized_binary` and either spec-mis-fire or
    // crash. These tests pin the variant gate.

    #[test]
    fn peek_adaptive_binary_returns_none_for_empty_slot() {
        let mut chunk = Chunk::new();
        chunk.emit(Op::Add, 1);
        let slot = chunk.inline_cache_slot(0).expect("Add registers a slot");
        // Default state of a freshly-emitted slot is Empty.
        assert!(chunk.peek_adaptive_binary_cache(slot).is_none());
    }

    #[test]
    fn peek_adaptive_binary_returns_op_and_state_after_warmup() {
        use super::{AdaptiveBinaryOp, AdaptiveBinaryState, BinaryShape, InlineCacheEntry};
        let mut chunk = Chunk::new();
        chunk.emit(Op::Add, 1);
        let slot = chunk.inline_cache_slot(0).expect("Add registers a slot");
        chunk.set_inline_cache_entry(
            slot,
            InlineCacheEntry::AdaptiveBinary {
                op: AdaptiveBinaryOp::Add,
                state: AdaptiveBinaryState::Warmup {
                    shape: BinaryShape::Int,
                    hits: 2,
                },
            },
        );
        let (op, state) = chunk
            .peek_adaptive_binary_cache(slot)
            .expect("warmed slot peek");
        assert_eq!(op, AdaptiveBinaryOp::Add);
        assert!(matches!(
            state,
            AdaptiveBinaryState::Warmup {
                shape: BinaryShape::Int,
                hits: 2
            }
        ));
    }

    #[test]
    fn peek_adaptive_binary_returns_none_for_non_binary_variants() {
        // The cache slot may legitimately hold a `Property`, `Method`,
        // or `DirectCall` entry (eg. a Property slot at the offset
        // sequence happens to alias an Add slot during a code rewrite —
        // currently this cannot happen, but the peek must defensively
        // refuse non-AdaptiveBinary variants regardless).
        use super::{InlineCacheEntry, PropertyCacheTarget};
        let mut chunk = Chunk::new();
        chunk.emit(Op::Add, 1);
        let slot = chunk.inline_cache_slot(0).expect("Add registers a slot");
        chunk.set_inline_cache_entry(
            slot,
            InlineCacheEntry::Property {
                name_idx: 0,
                target: PropertyCacheTarget::ListCount,
            },
        );
        assert!(chunk.peek_adaptive_binary_cache(slot).is_none());
    }

    #[test]
    fn peek_adaptive_binary_returns_none_for_out_of_bounds_slot() {
        // Defensive: `execute_adaptive_binary` filters its `slot`
        // through `inline_cache_slot` first, but
        // `peek_adaptive_binary_cache` should still return None for an
        // unmapped slot rather than panicking.
        let chunk = Chunk::new();
        assert!(chunk.peek_adaptive_binary_cache(0).is_none());
        assert!(chunk.peek_adaptive_binary_cache(usize::MAX).is_none());
    }

    #[test]
    fn peek_adaptive_binary_state_is_copy() {
        // Compile-time assertion: `AdaptiveBinaryState: Copy` is the
        // whole point of this optimization — if a future variant adds
        // a non-Copy field, the static check below will fail at compile
        // time before the dispatcher silently regresses to the heavy
        // `InlineCacheEntry::clone` path.
        fn assert_copy<T: Copy>() {}
        assert_copy::<super::AdaptiveBinaryState>();
        assert_copy::<super::AdaptiveBinaryOp>();
        assert_copy::<super::BinaryShape>();
    }

    // --- peek_method_cache contract ---
    //
    // The peek replaces the per-dispatch `InlineCacheEntry::clone` on the
    // method-call dispatch sites (`Op::MethodCall`, `Op::MethodCallOpt`,
    // `Op::MethodCallSpread`). It must return None for unrelated IC variants
    // — silently mis-extracting a `Property` / `DirectCall` / `AdaptiveBinary`
    // slot as `Method` would feed garbage into `try_cached_method` and either
    // spec-mis-fire (wrong target/argc) or skip the cache entirely on a real
    // hit. These tests pin the variant gate.

    #[test]
    fn peek_method_cache_returns_none_for_empty_slot() {
        let mut chunk = Chunk::new();
        chunk.emit_method_call(0, 0, 1);
        let slot = chunk
            .inline_cache_slot(0)
            .expect("MethodCall registers a slot");
        assert!(chunk.peek_method_cache(slot).is_none());
    }

    #[test]
    fn peek_method_cache_returns_triple_after_warmup() {
        let mut chunk = Chunk::new();
        chunk.emit_method_call(7, 2, 1);
        let slot = chunk
            .inline_cache_slot(0)
            .expect("MethodCall registers a slot");
        chunk.set_inline_cache_entry(
            slot,
            InlineCacheEntry::Method {
                name_idx: 7,
                argc: 2,
                target: MethodCacheTarget::ListContains,
            },
        );
        let (name_idx, argc, target) = chunk.peek_method_cache(slot).expect("warmed slot peek");
        assert_eq!(name_idx, 7);
        assert_eq!(argc, 2);
        assert_eq!(target, MethodCacheTarget::ListContains);
    }

    #[test]
    fn peek_method_cache_returns_none_for_non_method_variants() {
        // The cache slot may legitimately hold an `AdaptiveBinary`,
        // `Property`, or `DirectCall` entry. The peek must defensively
        // refuse non-Method variants regardless.
        let mut chunk = Chunk::new();
        chunk.emit_method_call(0, 0, 1);
        let slot = chunk
            .inline_cache_slot(0)
            .expect("MethodCall registers a slot");

        chunk.set_inline_cache_entry(
            slot,
            InlineCacheEntry::Property {
                name_idx: 0,
                target: PropertyCacheTarget::ListCount,
            },
        );
        assert!(chunk.peek_method_cache(slot).is_none());
    }

    #[test]
    fn peek_method_cache_returns_none_for_out_of_bounds_slot() {
        let chunk = Chunk::new();
        assert!(chunk.peek_method_cache(0).is_none());
        assert!(chunk.peek_method_cache(usize::MAX).is_none());
    }

    #[test]
    fn peek_method_cache_target_is_copy() {
        // Compile-time assertion: `MethodCacheTarget: Copy` is the whole
        // point of this peek path — if a future variant adds a non-Copy
        // field (eg. an `Arc<str>` for a dynamic method name), the static
        // check below will fail at compile time before the dispatcher
        // silently regresses to the heavy `InlineCacheEntry::clone` path.
        fn assert_copy<T: Copy>() {}
        assert_copy::<super::MethodCacheTarget>();
    }

    // --- peek_property_cache contract ---
    //
    // The peek replaces the per-dispatch `InlineCacheEntry::clone` on the
    // property-read path (`Op::GetProperty` / `Op::GetPropertyOpt`). It
    // must return None for unrelated IC variants — silently mis-extracting
    // a `Method` / `DirectCall` / `AdaptiveBinary` slot as `Property` would
    // feed garbage into `try_cached_property` (wrong target match, possibly
    // a panic on the field-name lookup). These tests pin the variant gate.

    #[test]
    fn peek_property_cache_returns_none_for_empty_slot() {
        let mut chunk = Chunk::new();
        chunk.emit_u16(Op::GetProperty, 0, 1);
        let slot = chunk
            .inline_cache_slot(0)
            .expect("GetProperty registers a slot");
        assert!(chunk.peek_property_cache(slot).is_none());
    }

    #[test]
    fn peek_property_cache_returns_pair_after_warmup_for_dict_field() {
        let mut chunk = Chunk::new();
        chunk.emit_u16(Op::GetProperty, 0, 1);
        let slot = chunk
            .inline_cache_slot(0)
            .expect("GetProperty registers a slot");
        chunk.set_inline_cache_entry(
            slot,
            InlineCacheEntry::Property {
                name_idx: 11,
                target: PropertyCacheTarget::DictField(Arc::from("count")),
            },
        );
        let (name_idx, target) = chunk
            .peek_property_cache(slot)
            .expect("warmed property slot peek");
        assert_eq!(name_idx, 11);
        match target {
            PropertyCacheTarget::DictField(field) => assert_eq!(field.as_ref(), "count"),
            other => panic!("expected DictField, got {other:?}"),
        }
    }

    #[test]
    fn peek_property_cache_returns_pair_for_unit_target() {
        // Unit targets (eg. ListCount, ListEmpty, PairFirst) carry no Arc,
        // so the cloned PropertyCacheTarget is a pure scalar move at the
        // peek boundary. The hottest case in practice.
        let mut chunk = Chunk::new();
        chunk.emit_u16(Op::GetProperty, 0, 1);
        let slot = chunk
            .inline_cache_slot(0)
            .expect("GetProperty registers a slot");
        chunk.set_inline_cache_entry(
            slot,
            InlineCacheEntry::Property {
                name_idx: 3,
                target: PropertyCacheTarget::ListCount,
            },
        );
        let (name_idx, target) = chunk
            .peek_property_cache(slot)
            .expect("warmed property slot peek");
        assert_eq!(name_idx, 3);
        assert_eq!(target, PropertyCacheTarget::ListCount);
    }

    #[test]
    fn peek_property_cache_returns_none_for_non_property_variants() {
        let mut chunk = Chunk::new();
        chunk.emit_u16(Op::GetProperty, 0, 1);
        let slot = chunk
            .inline_cache_slot(0)
            .expect("GetProperty registers a slot");
        chunk.set_inline_cache_entry(
            slot,
            InlineCacheEntry::Method {
                name_idx: 0,
                argc: 0,
                target: MethodCacheTarget::ListCount,
            },
        );
        assert!(chunk.peek_property_cache(slot).is_none());
    }

    #[test]
    fn peek_property_cache_returns_none_for_out_of_bounds_slot() {
        let chunk = Chunk::new();
        assert!(chunk.peek_property_cache(0).is_none());
        assert!(chunk.peek_property_cache(usize::MAX).is_none());
    }

    // --- peek_direct_call_state contract ---
    //
    // Used on both the hot Specialized-hit check path (`try_cached_direct_call`
    // / `try_cached_named_direct_call`) and the state-machine write-back
    // (`next_direct_call_entry`). Returning None for the non-DirectCall slot
    // shapes is critical: a mis-extracted Method/Property/AdaptiveBinary slot
    // would have the dispatcher attempt a closure call with the wrong argc
    // or Arc::ptr_eq against an unrelated closure.

    #[test]
    fn peek_direct_call_state_returns_none_for_empty_slot() {
        let mut chunk = Chunk::new();
        chunk.emit_u8(Op::Call, 0, 1);
        let slot = chunk
            .inline_cache_slot(0)
            .expect("Op::Call registers a slot");
        assert!(chunk.peek_direct_call_state(slot).is_none());
    }

    #[test]
    fn peek_direct_call_state_returns_warmup_state() {
        let mut chunk = Chunk::new();
        chunk.emit_u8(Op::Call, 0, 1);
        let slot = chunk
            .inline_cache_slot(0)
            .expect("Op::Call registers a slot");
        let target = synthetic_direct_call_target();
        chunk.set_inline_cache_entry(
            slot,
            InlineCacheEntry::DirectCall {
                state: DirectCallState::Warmup {
                    argc: 2,
                    target: target.clone(),
                    hits: 1,
                },
            },
        );
        let state = chunk
            .peek_direct_call_state(slot)
            .expect("warmed direct-call slot peek");
        match state {
            DirectCallState::Warmup {
                argc,
                target: peeked_target,
                hits,
            } => {
                assert_eq!(argc, 2);
                assert_eq!(hits, 1);
                assert_eq!(peeked_target, target);
            }
            other => panic!("expected Warmup, got {other:?}"),
        }
    }

    #[test]
    fn peek_direct_call_state_returns_specialized_state() {
        let mut chunk = Chunk::new();
        chunk.emit_u8(Op::Call, 0, 1);
        let slot = chunk
            .inline_cache_slot(0)
            .expect("Op::Call registers a slot");
        let target = synthetic_direct_call_target();
        chunk.set_inline_cache_entry(
            slot,
            InlineCacheEntry::DirectCall {
                state: DirectCallState::Specialized {
                    argc: 3,
                    target: target.clone(),
                    hits: 100,
                    misses: 0,
                },
            },
        );
        let state = chunk
            .peek_direct_call_state(slot)
            .expect("warmed direct-call slot peek");
        match state {
            DirectCallState::Specialized {
                argc,
                target: peeked_target,
                hits,
                misses,
            } => {
                assert_eq!(argc, 3);
                assert_eq!(hits, 100);
                assert_eq!(misses, 0);
                assert_eq!(peeked_target, target);
            }
            other => panic!("expected Specialized, got {other:?}"),
        }
    }

    #[test]
    fn peek_direct_call_state_returns_none_for_non_direct_call_variants() {
        let mut chunk = Chunk::new();
        chunk.emit_u8(Op::Call, 0, 1);
        let slot = chunk
            .inline_cache_slot(0)
            .expect("Op::Call registers a slot");

        chunk.set_inline_cache_entry(
            slot,
            InlineCacheEntry::Property {
                name_idx: 0,
                target: PropertyCacheTarget::ListCount,
            },
        );
        assert!(chunk.peek_direct_call_state(slot).is_none());
    }

    #[test]
    fn peek_direct_call_state_returns_none_for_out_of_bounds_slot() {
        let chunk = Chunk::new();
        assert!(chunk.peek_direct_call_state(0).is_none());
        assert!(chunk.peek_direct_call_state(usize::MAX).is_none());
    }

    /// Build a synthetic `DirectCallTarget::Closure` for direct-call peek
    /// tests. The closure has an empty body — the IC peek only inspects
    /// the wrapping `Arc`, not the closure internals.
    fn synthetic_direct_call_target() -> DirectCallTarget {
        use crate::value::VmClosure;
        use crate::{CompiledFunction, VmEnv};
        let func = CompiledFunction {
            name: "synthetic".to_string(),
            type_params: Vec::new(),
            nominal_type_names: Vec::new(),
            params: Vec::new(),
            default_start: None,
            chunk: Arc::new(Chunk::new()),
            is_generator: false,
            is_stream: false,
            has_rest_param: false,
            has_runtime_type_checks: false,
        };
        DirectCallTarget::Closure(Arc::new(VmClosure {
            func: Arc::new(func),
            env: VmEnv::new(),
            source_dir: None,
            module_functions: None,
            module_state: None,
        }))
    }
}