feat: add experimental tailcall vm backend

.lazy.lua

@@ -16,6 +16,9 @@ vim.lsp.config("rust_analyzer", {
   settings = {
     ["rust-analyzer"] = {
       cargo = {
+        features = {
+          "tailcall"
+        }
       }
     }
   }

fix-codegen/src/disassembler.rs

@@ -3,7 +3,7 @@ use std::fmt::Write;
 use colored::Colorize;
 use num_enum::TryFromPrimitive;
 
-use crate::{AttrKeyType, InstructionPtr, OperandType, Op};
+use crate::{AttrKeyType, InstructionPtr, Op, OperandType};
 
 pub trait DisassemblerContext {
     fn resolve_string(&self, id: u32) -> &str;
@@ -287,12 +287,13 @@ impl<'a, Ctx: DisassemblerContext> Disassembler<'a, Ctx> {
                 let mut args = format!("size={}", count);
                 for _ in 0..count {
                     let key_tag = self.read_u8();
-                    let key_ty = AttrKeyType::try_from_primitive(key_tag)
+                    let key_ty =
-                        .expect("invalid attr key type");
+                        AttrKeyType::try_from_primitive(key_tag).expect("invalid attr key type");
                     match key_ty {
                         AttrKeyType::Static => {
                             let key_id = self.read_u32();
-                            let _ = write!(args, " [{}={}", self.ctx.resolve_string(key_id), key_id);
+                            let _ =
+                                write!(args, " [{}={}", self.ctx.resolve_string(key_id), key_id);
                         }
                         AttrKeyType::Dynamic => {
                             let _ = write!(args, " [dyn");

fix-vm/Cargo.toml

@@ -3,6 +3,9 @@ name = "fix-vm"
 version = "0.1.0"
 edition = "2024"
 
+[features]
+tailcall = []
+
 [dependencies]
 gc-arena = { workspace = true }
 hashbrown = { workspace = true }

fix-vm/src/bytecode_reader.rs

@@ -12,6 +12,7 @@ pub(crate) struct BytecodeReader<'a> {
 }
 
 impl<'a> BytecodeReader<'a> {
+    #[cfg_attr(feature = "tailcall", allow(dead_code))]
     pub(crate) fn new(bytecode: &'a [u8], pc: usize) -> Self {
         Self {
             bytecode,
@@ -20,6 +21,16 @@ impl<'a> BytecodeReader<'a> {
         }
     }
 
+    #[inline(always)]
+    #[cfg_attr(not(feature = "tailcall"), allow(dead_code))]
+    pub(crate) fn from_after_op(bytecode: &'a [u8], inst_start_pc: usize) -> Self {
+        Self {
+            bytecode,
+            pc: inst_start_pc + 1,
+            inst_start_pc,
+        }
+    }
+
     #[inline(always)]
     #[cfg_attr(debug_assertions, track_caller)]
     fn read_array<const N: usize>(&mut self) -> [u8; N] {
@@ -31,6 +42,7 @@ impl<'a> BytecodeReader<'a> {
     }
 
     #[inline(always)]
+    #[cfg_attr(feature = "tailcall", allow(dead_code))]
     pub(crate) fn read_op(&mut self) -> fix_codegen::Op {
         use fix_codegen::Op;
         self.inst_start_pc = self.pc;
@@ -105,16 +117,17 @@ impl<'a> BytecodeReader<'a> {
     }
 
     #[inline(always)]
-    pub(crate) fn read_attr_key_data<C: crate::VmContext>(&mut self, ctx: &C) -> crate::AttrKeyData {
+    pub(crate) fn read_attr_key_data<C: crate::VmContext>(
+        &mut self,
+        ctx: &C,
+    ) -> crate::AttrKeyData {
         use fix_codegen::AttrKeyType;
         let tag = self.read_u8();
         let ty = AttrKeyType::try_from_primitive(tag)
             .unwrap_or_else(|err| panic!("unknown key tag: {:#04x}", err.number));
         match ty {
             AttrKeyType::Static => crate::AttrKeyData::Static(self.read_string_id()),
-            AttrKeyType::Dynamic => {
+            AttrKeyType::Dynamic => crate::AttrKeyData::Dynamic(self.read_operand_data(ctx)),
-                crate::AttrKeyData::Dynamic(self.read_operand_data(ctx))
-            }
         }
     }
 

fix-vm/src/dispatch_tailcall.rs

@@ -0,0 +1,311 @@
+#![cfg(feature = "tailcall")]
+
+use gc_arena::Mutation;
+
+use crate::{BytecodeReader, ForceInfo, StepResult, Vm, VmContext};
+
+pub(crate) enum TailResult<'gc> {
+    YieldFuel(u32),
+    ForceThunk(ForceInfo<'gc>),
+    Done,
+}
+
+pub(crate) type OpFn<'gc, C> = extern "rust-preserve-none" fn(
+    &mut Vm<'gc>,
+    &Mutation<'gc>,
+    &mut C,
+    &[u8],
+    &DispatchTable<'gc, C>,
+    u32,
+    u32,
+) -> TailResult<'gc>;
+
+pub(crate) struct DispatchTable<'gc, C: VmContext>(pub(crate) [OpFn<'gc, C>; 256]);
+
+extern "rust-preserve-none" fn op_illegal<'gc, C: VmContext>(
+    _vm: &mut Vm<'gc>,
+    _mc: &Mutation<'gc>,
+    _ctx: &mut C,
+    _bc: &[u8],
+    _table: &DispatchTable<'gc, C>,
+    pc: u32,
+    _fuel: u32,
+) -> TailResult<'gc> {
+    panic!("illegal opcode at pc = {pc}");
+}
+
+macro_rules! tail_dispatch_after {
+    ($result:expr, $new_pc:expr, $vm:ident, $mc:ident, $ctx:ident, $bc:ident, $table:ident, $fuel:ident) => {{
+        match $result {
+            StepResult::Continue => {}
+            StepResult::ForceThunk(info) => return TailResult::ForceThunk(info),
+            StepResult::Done => return TailResult::Done,
+        }
+        let new_pc: u32 = $new_pc;
+        if $fuel == 0 {
+            return TailResult::YieldFuel(new_pc);
+        }
+        let next_op = $bc[new_pc as usize] as usize;
+        become $table.0[next_op]($vm, $mc, $ctx, $bc, $table, new_pc, $fuel - 1)
+    }};
+}
+
+macro_rules! tail_fn {
+    ($name:ident, ()) => {
+        extern "rust-preserve-none" fn $name<'gc, C: VmContext>(
+            vm: &mut Vm<'gc>,
+            mc: &Mutation<'gc>,
+            ctx: &mut C,
+            bc: &[u8],
+            table: &DispatchTable<'gc, C>,
+            pc: u32,
+            fuel: u32,
+        ) -> TailResult<'gc> {
+            let result = vm.$name();
+            tail_dispatch_after!(result, pc + 1, vm, mc, ctx, bc, table, fuel)
+        }
+    };
+    ($name:ident, (reader)) => {
+        extern "rust-preserve-none" fn $name<'gc, C: VmContext>(
+            vm: &mut Vm<'gc>,
+            mc: &Mutation<'gc>,
+            ctx: &mut C,
+            bc: &[u8],
+            table: &DispatchTable<'gc, C>,
+            pc: u32,
+            fuel: u32,
+        ) -> TailResult<'gc> {
+            let mut reader = BytecodeReader::from_after_op(bc, pc as usize);
+            let result = vm.$name(&mut reader);
+            tail_dispatch_after!(result, reader.pc() as u32, vm, mc, ctx, bc, table, fuel)
+        }
+    };
+    ($name:ident, (reader, mc)) => {
+        extern "rust-preserve-none" fn $name<'gc, C: VmContext>(
+            vm: &mut Vm<'gc>,
+            mc: &Mutation<'gc>,
+            ctx: &mut C,
+            bc: &[u8],
+            table: &DispatchTable<'gc, C>,
+            pc: u32,
+            fuel: u32,
+        ) -> TailResult<'gc> {
+            let mut reader = BytecodeReader::from_after_op(bc, pc as usize);
+            let result = vm.$name(&mut reader, mc);
+            tail_dispatch_after!(result, reader.pc() as u32, vm, mc, ctx, bc, table, fuel)
+        }
+    };
+    ($name:ident, (ctx, reader, mc)) => {
+        extern "rust-preserve-none" fn $name<'gc, C: VmContext>(
+            vm: &mut Vm<'gc>,
+            mc: &Mutation<'gc>,
+            ctx: &mut C,
+            bc: &[u8],
+            table: &DispatchTable<'gc, C>,
+            pc: u32,
+            fuel: u32,
+        ) -> TailResult<'gc> {
+            let mut reader = BytecodeReader::from_after_op(bc, pc as usize);
+            let result = vm.$name(ctx, &mut reader, mc);
+            tail_dispatch_after!(result, reader.pc() as u32, vm, mc, ctx, bc, table, fuel)
+        }
+    };
+    ($name:ident, (mc, inst_start_pc)) => {
+        extern "rust-preserve-none" fn $name<'gc, C: VmContext>(
+            vm: &mut Vm<'gc>,
+            mc: &Mutation<'gc>,
+            ctx: &mut C,
+            bc: &[u8],
+            table: &DispatchTable<'gc, C>,
+            pc: u32,
+            fuel: u32,
+        ) -> TailResult<'gc> {
+            let result = vm.$name(mc, pc as usize);
+            tail_dispatch_after!(result, pc + 1, vm, mc, ctx, bc, table, fuel)
+        }
+    };
+    ($name:ident, (ctx)) => {
+        extern "rust-preserve-none" fn $name<'gc, C: VmContext>(
+            vm: &mut Vm<'gc>,
+            mc: &Mutation<'gc>,
+            ctx: &mut C,
+            bc: &[u8],
+            table: &DispatchTable<'gc, C>,
+            pc: u32,
+            fuel: u32,
+        ) -> TailResult<'gc> {
+            let result = vm.$name(ctx);
+            tail_dispatch_after!(result, pc + 1, vm, mc, ctx, bc, table, fuel)
+        }
+    };
+}
+
+tail_fn!(op_push_smi, (reader));
+tail_fn!(op_push_bigint, (reader, mc));
+tail_fn!(op_push_float, (reader));
+tail_fn!(op_push_string, (reader));
+tail_fn!(op_push_null, ());
+tail_fn!(op_push_true, ());
+tail_fn!(op_push_false, ());
+
+tail_fn!(op_load_local, (reader));
+tail_fn!(op_load_outer, (reader));
+tail_fn!(op_store_local, (reader, mc));
+tail_fn!(op_alloc_locals, (reader, mc));
+
+tail_fn!(op_make_thunk, (reader, mc));
+tail_fn!(op_make_closure, (reader, mc));
+tail_fn!(op_make_pattern_closure, (reader, mc));
+
+tail_fn!(op_call, (ctx, reader, mc));
+tail_fn!(op_return, (ctx, reader, mc));
+
+tail_fn!(op_make_attrs, (ctx, reader, mc));
+tail_fn!(op_make_empty_attrs, ());
+tail_fn!(op_select_static, (ctx, reader, mc));
+tail_fn!(op_select_dynamic, (ctx, reader, mc));
+tail_fn!(op_jump_if_select_succeeded, (reader));
+tail_fn!(op_has_attr, (reader));
+
+tail_fn!(op_make_list, (ctx, reader, mc));
+tail_fn!(op_make_empty_list, ());
+
+tail_fn!(op_add, (ctx, reader, mc));
+tail_fn!(op_sub, (reader, mc));
+tail_fn!(op_mul, (reader, mc));
+tail_fn!(op_div, (reader, mc));
+tail_fn!(op_eq, (ctx, reader, mc));
+tail_fn!(op_neq, (ctx, reader, mc));
+tail_fn!(op_lt, (ctx, reader, mc));
+tail_fn!(op_gt, (ctx, reader, mc));
+tail_fn!(op_leq, (ctx, reader, mc));
+tail_fn!(op_geq, (ctx, reader, mc));
+tail_fn!(op_concat, (reader, mc));
+tail_fn!(op_update, (mc, inst_start_pc));
+
+tail_fn!(op_neg, ());
+tail_fn!(op_not, ());
+
+tail_fn!(op_jump_if_false, (reader, mc));
+tail_fn!(op_jump_if_true, (reader, mc));
+tail_fn!(op_jump, (reader));
+
+tail_fn!(op_concat_strings, (ctx, reader, mc));
+tail_fn!(op_resolve_path, (ctx));
+
+tail_fn!(op_assert, (reader));
+
+tail_fn!(op_push_with, (ctx, reader, mc));
+tail_fn!(op_pop_with, ());
+tail_fn!(op_lookup_with, (ctx, reader, mc));
+tail_fn!(op_prepare_with, ());
+
+tail_fn!(op_load_builtins, ());
+tail_fn!(op_load_builtin, (reader));
+
+tail_fn!(op_mk_pos, (reader));
+tail_fn!(op_load_repl_binding, (reader));
+tail_fn!(op_load_scoped_binding, (reader));
+
+macro_rules! table {
+    ($($variant:ident => $fn:ident),* $(,)?) => {
+        impl<'gc, C: VmContext> DispatchTable<'gc, C> {
+            pub(crate) const NEW: Self = {
+                let mut arr: [OpFn<'gc, C>; 256] = [op_illegal; 256];
+                $( arr[fix_codegen::Op::$variant as usize] = $fn; )*
+                DispatchTable(arr)
+            };
+        }
+
+        // Exhaustiveness check: fails to compile if `fix_codegen::Op` gains,
+        // loses, or renames a variant that isn't wired up above.
+        #[allow(dead_code)]
+        const _: fn(fix_codegen::Op) = |op| match op {
+            $( fix_codegen::Op::$variant => (), )*
+        };
+    };
+}
+
+table! {
+    PushSmi               => op_push_smi,
+    PushBigInt            => op_push_bigint,
+    PushFloat             => op_push_float,
+    PushString            => op_push_string,
+    PushNull              => op_push_null,
+    PushTrue              => op_push_true,
+    PushFalse             => op_push_false,
+
+    LoadLocal             => op_load_local,
+    LoadOuter             => op_load_outer,
+    StoreLocal            => op_store_local,
+    AllocLocals           => op_alloc_locals,
+
+    MakeThunk             => op_make_thunk,
+    MakeClosure           => op_make_closure,
+    MakePatternClosure    => op_make_pattern_closure,
+
+    Call                  => op_call,
+    Return                => op_return,
+
+    MakeAttrs             => op_make_attrs,
+    MakeEmptyAttrs        => op_make_empty_attrs,
+    SelectStatic          => op_select_static,
+    SelectDynamic         => op_select_dynamic,
+    JumpIfSelectSucceeded => op_jump_if_select_succeeded,
+    HasAttr               => op_has_attr,
+
+    MakeList              => op_make_list,
+    MakeEmptyList         => op_make_empty_list,
+
+    OpAdd                 => op_add,
+    OpSub                 => op_sub,
+    OpMul                 => op_mul,
+    OpDiv                 => op_div,
+    OpEq                  => op_eq,
+    OpNeq                 => op_neq,
+    OpLt                  => op_lt,
+    OpGt                  => op_gt,
+    OpLeq                 => op_leq,
+    OpGeq                 => op_geq,
+    OpConcat              => op_concat,
+    OpUpdate              => op_update,
+
+    OpNeg                 => op_neg,
+    OpNot                 => op_not,
+
+    JumpIfFalse           => op_jump_if_false,
+    JumpIfTrue            => op_jump_if_true,
+    Jump                  => op_jump,
+
+    ConcatStrings         => op_concat_strings,
+    ResolvePath           => op_resolve_path,
+
+    Assert                => op_assert,
+
+    PushWith              => op_push_with,
+    PopWith               => op_pop_with,
+    LookupWith            => op_lookup_with,
+    PrepareWith           => op_prepare_with,
+
+    LoadBuiltins          => op_load_builtins,
+    LoadBuiltin           => op_load_builtin,
+
+    MkPos                 => op_mk_pos,
+    LoadReplBinding       => op_load_repl_binding,
+    LoadScopedBinding     => op_load_scoped_binding,
+
+    Illegal               => op_illegal,
+}
+
+pub(crate) fn run_tailcall<'gc, C: VmContext>(
+    vm: &mut Vm<'gc>,
+    mc: &Mutation<'gc>,
+    ctx: &mut C,
+    bc: &[u8],
+    pc: u32,
+) -> TailResult<'gc> {
+    const FUEL: u32 = 1024;
+    let table = &DispatchTable::<'gc, C>::NEW;
+    let op = bc[pc as usize] as usize;
+    table.0[op](vm, mc, ctx, bc, table, pc, FUEL)
+}

fix-vm/src/instructions/arithmetic.rs

@@ -2,8 +2,7 @@ use std::cmp::Ordering;
 
 use gc_arena::{Gc, Mutation};
 
-use crate::{BytecodeReader, NixNum, StepResult, StrictValue, VmError, Value};
+use crate::{BytecodeReader, NixNum, StepResult, StrictValue, Value, VmContextExt, VmError};
-use crate::VmContextExt;
 
 impl<'gc> crate::Vm<'gc> {
     #[inline(always)]
@@ -21,7 +20,10 @@ impl<'gc> crate::Vm<'gc> {
         }
         let rhs = self.pop_stack_forced();
         let lhs = self.pop_stack_forced();
-        if let (Some(ls), Some(rs)) = (VmContextExt::get_string(ctx, lhs), VmContextExt::get_string(ctx, rhs)) {
+        if let (Some(ls), Some(rs)) = (
+            VmContextExt::get_string(ctx, lhs),
+            VmContextExt::get_string(ctx, rhs),
+        ) {
             let ns = Gc::new(mc, crate::NixString::new(format!("{ls}{rs}")));
             self.push_stack(Value::new_gc(ns));
             return StepResult::Continue;
@@ -300,7 +302,10 @@ impl<'gc> crate::Vm<'gc> {
         if lhs.is::<crate::Null>() && rhs.is::<crate::Null>() {
             return Ok(true);
         }
-        if let (Some(a), Some(b)) = (VmContextExt::get_string(ctx, lhs), VmContextExt::get_string(ctx, rhs)) {
+        if let (Some(a), Some(b)) = (
+            VmContextExt::get_string(ctx, lhs),
+            VmContextExt::get_string(ctx, rhs),
+        ) {
             return Ok(a == b);
         }
         if let (Some(a), Some(b)) = (lhs.as_gc::<crate::List>(), rhs.as_gc::<crate::List>()) {
@@ -358,20 +363,21 @@ impl<'gc> crate::Vm<'gc> {
         if let (Some(a), Some(b)) = (get_num(lhs), get_num(rhs)) {
             let ord = match (a, b) {
                 (NixNum::Int(a), NixNum::Int(b)) => a.cmp(&b),
-                (NixNum::Float(a), NixNum::Float(b)) => {
+                (NixNum::Float(a), NixNum::Float(b)) => a.partial_cmp(&b).unwrap_or(Ordering::Less),
-                    a.partial_cmp(&b).unwrap_or(Ordering::Less)
+                (NixNum::Int(a), NixNum::Float(b)) => {
+                    (a as f64).partial_cmp(&b).unwrap_or(Ordering::Less)
+                }
+                (NixNum::Float(a), NixNum::Int(b)) => {
+                    a.partial_cmp(&(b as f64)).unwrap_or(Ordering::Less)
                 }
-                (NixNum::Int(a), NixNum::Float(b)) => (a as f64)
-                    .partial_cmp(&b)
-                    .unwrap_or(Ordering::Less),
-                (NixNum::Float(a), NixNum::Int(b)) => a
-                    .partial_cmp(&(b as f64))
-                    .unwrap_or(Ordering::Less),
             };
             self.push_stack(Value::new_inline(pred(ord)));
             return Ok(());
         }
-        if let (Some(a), Some(b)) = (VmContextExt::get_string(ctx, lhs), VmContextExt::get_string(ctx, rhs)) {
+        if let (Some(a), Some(b)) = (
+            VmContextExt::get_string(ctx, lhs),
+            VmContextExt::get_string(ctx, rhs),
+        ) {
             self.push_stack(Value::new_inline(pred(a.cmp(b))));
             return Ok(());
         }

fix-vm/src/instructions/builtins_misc.rs

@@ -1,7 +1,8 @@
-use crate::{BytecodeReader, PrimOp, StepResult, Value};
 use fix_builtins::BuiltinId;
 use num_enum::TryFromPrimitive;
 
+use crate::{BytecodeReader, PrimOp, StepResult, Value};
+
 impl<'gc> crate::Vm<'gc> {
     #[inline(always)]
     pub(crate) fn op_load_builtins(&mut self) -> StepResult<'gc> {
@@ -27,13 +28,19 @@ impl<'gc> crate::Vm<'gc> {
     }
 
     #[inline(always)]
-    pub(crate) fn op_load_repl_binding(&mut self, reader: &mut BytecodeReader<'_>) -> StepResult<'gc> {
+    pub(crate) fn op_load_repl_binding(
+        &mut self,
+        reader: &mut BytecodeReader<'_>,
+    ) -> StepResult<'gc> {
         let _name = reader.read_string_id();
         todo!("LoadReplBinding");
     }
 
     #[inline(always)]
-    pub(crate) fn op_load_scoped_binding(&mut self, reader: &mut BytecodeReader<'_>) -> StepResult<'gc> {
+    pub(crate) fn op_load_scoped_binding(
+        &mut self,
+        reader: &mut BytecodeReader<'_>,
+    ) -> StepResult<'gc> {
         let _name = reader.read_string_id();
         todo!("LoadScopedBinding");
     }

fix-vm/src/instructions/calls.rs

@@ -1,11 +1,8 @@
 use fix_error::Error;
-
-use crate::{
-    BytecodeReader, CallFrame, Closure, Env, StepResult, ThunkState,
-};
-use crate::VmContextExt;
 use gc_arena::{Gc, Mutation, RefLock};
 
+use crate::{BytecodeReader, CallFrame, Closure, Env, StepResult, ThunkState, VmContextExt};
+
 impl<'gc> crate::Vm<'gc> {
     #[inline(always)]
     pub(crate) fn op_call(
@@ -18,9 +15,7 @@ impl<'gc> crate::Vm<'gc> {
             return step;
         }
         if self.call_depth > 10000 {
-            return self.finish_err(Error::eval_error(
+            return self.finish_err(Error::eval_error("stack overflow; max-call-depth exceeded"));
-                "stack overflow; max-call-depth exceeded",
-            ));
         }
         self.call_depth += 1;
         let func = self.pop_stack();
@@ -83,11 +78,7 @@ impl<'gc> crate::Vm<'gc> {
             match val.restrict() {
                 Ok(val) => {
                     *outer_thunk.borrow_mut(mc) = ThunkState::Evaluated(val);
-                    if reader
+                    if reader.bytecode().get(ret_pc).copied() == Some(fix_codegen::Op::Return as u8)
-                        .bytecode()
-                        .get(ret_pc)
-                        .copied()
-                        == Some(fix_codegen::Op::Return as u8)
                     {
                         self.push_stack(val.relax());
                     }
@@ -122,10 +113,7 @@ impl<'gc> crate::Vm<'gc> {
                         }
                         ThunkState::Evaluated(val) => {
                             *outer_thunk.borrow_mut(mc) = ThunkState::Evaluated(val);
-                            if reader
+                            if reader.bytecode().get(ret_pc).copied()
-                                .bytecode()
-                                .get(ret_pc)
-                                .copied()
                                 == Some(fix_codegen::Op::Return as u8)
                             {
                                 self.push_stack(val.relax());
@@ -133,9 +121,8 @@ impl<'gc> crate::Vm<'gc> {
                         }
                         ThunkState::Apply { func: _, arg: _ } => todo!("force Apply thunk"),
                         ThunkState::Blackhole => {
-                            return self.finish_err(Error::eval_error(
+                            return self
-                                "infinite recursion encountered",
+                                .finish_err(Error::eval_error("infinite recursion encountered"));
-                            ));
                         }
                     }
                 }

fix-vm/src/instructions/closures.rs

@@ -4,7 +4,11 @@ use crate::{BytecodeReader, StepResult, ThunkState, Value};
 
 impl<'gc> crate::Vm<'gc> {
     #[inline(always)]
-    pub(crate) fn op_make_thunk(&mut self, reader: &mut BytecodeReader<'_>, mc: &Mutation<'gc>) -> StepResult<'gc> {
+    pub(crate) fn op_make_thunk(
+        &mut self,
+        reader: &mut BytecodeReader<'_>,
+        mc: &Mutation<'gc>,
+    ) -> StepResult<'gc> {
         let entry_point = reader.read_u32();
         let thunk = Gc::new(
             mc,
@@ -19,7 +23,11 @@ impl<'gc> crate::Vm<'gc> {
     }
 
     #[inline(always)]
-    pub(crate) fn op_make_closure(&mut self, reader: &mut BytecodeReader<'_>, mc: &Mutation<'gc>) -> StepResult<'gc> {
+    pub(crate) fn op_make_closure(
+        &mut self,
+        reader: &mut BytecodeReader<'_>,
+        mc: &Mutation<'gc>,
+    ) -> StepResult<'gc> {
         let entry_point = reader.read_u32();
         let n_locals = reader.read_u32();
         let closure = Gc::new(
@@ -36,7 +44,11 @@ impl<'gc> crate::Vm<'gc> {
     }
 
     #[inline(always)]
-    pub(crate) fn op_make_pattern_closure(&mut self, reader: &mut BytecodeReader<'_>, mc: &Mutation<'gc>) -> StepResult<'gc> {
+    pub(crate) fn op_make_pattern_closure(
+        &mut self,
+        reader: &mut BytecodeReader<'_>,
+        mc: &Mutation<'gc>,
+    ) -> StepResult<'gc> {
         let entry_point = reader.read_u32();
         let n_locals = reader.read_u32();
         let req_count = reader.read_u16() as usize;

fix-vm/src/instructions/collections.rs

@@ -72,8 +72,7 @@ impl<'gc> crate::Vm<'gc> {
             Some(v) => {
                 self.replace_stack(0, v);
             }
-            None => {
+            None => loop {
-                loop {
                 let byte = reader.bytecode()[reader.pc()];
                 if byte == fix_codegen::Op::SelectStatic as u8 {
                     reader.set_pc(reader.pc() + 1 + 4 + 4);
@@ -85,12 +84,10 @@ impl<'gc> crate::Vm<'gc> {
                     break;
                 } else {
                     let name = ctx.resolve_string(key);
-                        return self.finish_err(Error::eval_error(format!(
+                    return self
-                            "attribute '{name}' missing"
+                        .finish_err(Error::eval_error(format!("attribute '{name}' missing")));
-                        )));
-                    }
-                }
                 }
+            },
         }
         StepResult::Continue
     }
@@ -119,9 +116,7 @@ impl<'gc> crate::Vm<'gc> {
         } else if let Some(ns) = key_val.as_gc::<NixString>() {
             ctx.intern_string(ns.as_str())
         } else {
-            return self.finish_err(Error::eval_error(
+            return self.finish_err(Error::eval_error("dynamic select key must be a string"));
-                "dynamic select key must be a string",
-            ));
         };
 
         let attrset_val = self.stack[self.stack.len() - 2].restrict().expect("forced");
@@ -138,16 +133,17 @@ impl<'gc> crate::Vm<'gc> {
             }
             None => {
                 let name = ctx.resolve_string(key_sid);
-                return self.finish_err(Error::eval_error(format!(
+                return self.finish_err(Error::eval_error(format!("attribute '{name}' missing")));
-                    "attribute '{name}' missing"
-                )));
             }
         }
         StepResult::Continue
     }
 
     #[inline(always)]
-    pub(crate) fn op_jump_if_select_succeeded(&mut self, reader: &mut BytecodeReader<'_>) -> StepResult<'gc> {
+    pub(crate) fn op_jump_if_select_succeeded(
+        &mut self,
+        reader: &mut BytecodeReader<'_>,
+    ) -> StepResult<'gc> {
         let offset = reader.read_i32();
         reader.set_pc(((reader.pc() as isize) + (offset as isize)) as usize);
         StepResult::Continue

fix-vm/src/instructions/literals.rs

@@ -11,7 +11,11 @@ impl<'gc> crate::Vm<'gc> {
     }
 
     #[inline(always)]
-    pub(crate) fn op_push_bigint(&mut self, reader: &mut BytecodeReader<'_>, mc: &Mutation<'gc>) -> StepResult<'gc> {
+    pub(crate) fn op_push_bigint(
+        &mut self,
+        reader: &mut BytecodeReader<'_>,
+        mc: &Mutation<'gc>,
+    ) -> StepResult<'gc> {
         let val = reader.read_i64();
         self.push_stack(Value::new_gc(Gc::new(mc, val)));
         StepResult::Continue

fix-vm/src/instructions/mod.rs

@@ -1,9 +1,9 @@
+pub(crate) mod arithmetic;
+pub(crate) mod builtins_misc;
+pub(crate) mod calls;
+pub(crate) mod closures;
+pub(crate) mod collections;
+pub(crate) mod control;
 pub(crate) mod literals;
 pub(crate) mod variables;
-pub(crate) mod closures;
-pub(crate) mod calls;
-pub(crate) mod collections;
-pub(crate) mod arithmetic;
-pub(crate) mod control;
 pub(crate) mod with_scope;
-pub(crate) mod builtins_misc;

fix-vm/src/instructions/variables.rs

@@ -23,7 +23,11 @@ impl<'gc> crate::Vm<'gc> {
     }
 
     #[inline(always)]
-    pub(crate) fn op_store_local(&mut self, reader: &mut BytecodeReader<'_>, mc: &Mutation<'gc>) -> StepResult<'gc> {
+    pub(crate) fn op_store_local(
+        &mut self,
+        reader: &mut BytecodeReader<'_>,
+        mc: &Mutation<'gc>,
+    ) -> StepResult<'gc> {
         let idx = reader.read_u32() as usize;
         let val = self.pop_stack();
         self.env.borrow_mut(mc).locals[idx] = val;
@@ -31,7 +35,11 @@ impl<'gc> crate::Vm<'gc> {
     }
 
     #[inline(always)]
-    pub(crate) fn op_alloc_locals(&mut self, reader: &mut BytecodeReader<'_>, mc: &Mutation<'gc>) -> StepResult<'gc> {
+    pub(crate) fn op_alloc_locals(
+        &mut self,
+        reader: &mut BytecodeReader<'_>,
+        mc: &Mutation<'gc>,
+    ) -> StepResult<'gc> {
         let count = reader.read_u32() as usize;
         self.env
             .borrow_mut(mc)

fix-vm/src/instructions/with_scope.rs

@@ -1,8 +1,8 @@
-use fix_error::Error;
 use fix_common::Symbol;
+use fix_error::Error;
+use gc_arena::Gc;
 
 use crate::{BytecodeReader, CallFrame, StepResult, WithEnv};
-use gc_arena::Gc;
 
 impl<'gc> crate::Vm<'gc> {
     #[inline(always)]

fix-vm/src/lib.rs

@@ -1,4 +1,9 @@
 #![warn(clippy::unwrap_used)]
+#![cfg_attr(feature = "tailcall", expect(incomplete_features))]
+#![cfg_attr(
+    feature = "tailcall",
+    feature(explicit_tail_calls, rust_preserve_none_cc)
+)]
 
 use std::path::PathBuf;
 
@@ -12,16 +17,17 @@ use hashbrown::HashMap;
 use num_enum::TryFromPrimitive;
 use smallvec::SmallVec;
 
-mod bytecode_reader;
 mod boxing;
+mod bytecode_reader;
+#[cfg(feature = "tailcall")]
+mod dispatch_tailcall;
 mod value;
-use value::*;
 pub use value::StaticValue;
-pub(crate) mod instructions;
+use value::*;
 mod helpers;
-use helpers::*;
+pub(crate) mod instructions;
-
 pub(crate) use bytecode_reader::BytecodeReader;
+use helpers::*;
 
 type VmResult<T> = std::result::Result<T, VmError>;
 
@@ -339,11 +345,7 @@ impl<'gc> Vm<'gc> {
         if let Some(thunk) = val.as_gc::<Thunk>() {
             let mut state = thunk.borrow_mut(mc);
             match *state {
-                ThunkState::Pending {
+                ThunkState::Pending { ip, env, with_env } => {
-                    ip,
-                    env,
-                    with_env,
-                } => {
                     *state = ThunkState::Blackhole;
                     drop(state);
                     StepResult::ForceThunk(ForceInfo {
@@ -381,6 +383,20 @@ impl<'gc> Vm<'gc> {
             other => Some(other),
         }
     }
+
+    #[inline(always)]
+    pub(crate) fn apply_force_thunk(&mut self, info: ForceInfo<'gc>) -> usize {
+        self.call_stack.push(CallFrame {
+            thunk: Some(info.thunk),
+            stack_depth: info.stack_depth,
+            pc: info.inst_start_pc,
+            env: self.env,
+            with_env: self.with_env,
+        });
+        self.env = info.env;
+        self.with_env = info.with_env;
+        info.ip
+    }
 }
 
 #[allow(dead_code)]
@@ -423,7 +439,7 @@ impl Vm<'_> {
         let mut pc = ip.0;
         let bytecode: Vec<u8> = ctx.bytecode().to_vec();
         loop {
-            match arena.mutate_root(|mc, root| root.execute_batch(&bytecode, &mut ctx, pc, mc)) {
+            match arena.mutate_root(|mc, root| root.dispatch_batch(&bytecode, &mut ctx, pc, mc)) {
                 Action::Continue { pc: new_pc } => {
                     pc = new_pc;
                     if arena.metrics().allocation_debt() > COLLECTOR_GRANULARITY {
@@ -442,6 +458,37 @@ impl Vm<'_> {
 
 impl<'gc> Vm<'gc> {
     #[inline(always)]
+    fn dispatch_batch<C: VmContext>(
+        &mut self,
+        bytecode: &[u8],
+        ctx: &mut C,
+        pc: usize,
+        mc: &Mutation<'gc>,
+    ) -> Action {
+        #[cfg(not(feature = "tailcall"))]
+        {
+            self.execute_batch(bytecode, ctx, pc, mc)
+        }
+        #[cfg(feature = "tailcall")]
+        {
+            use crate::dispatch_tailcall::{TailResult, run_tailcall};
+            match run_tailcall(self, mc, ctx, bytecode, pc as u32) {
+                TailResult::YieldFuel(new_pc) => Action::Continue {
+                    pc: new_pc as usize,
+                },
+                TailResult::ForceThunk(info) => {
+                    let new_pc = self.apply_force_thunk(info);
+                    Action::Continue { pc: new_pc }
+                }
+                TailResult::Done => {
+                    Action::Done(self.result.take().expect("TailResult::Done without result"))
+                }
+            }
+        }
+    }
+
+    #[inline(always)]
+    #[cfg(not(feature = "tailcall"))]
     fn execute_batch(
         &mut self,
         bytecode: &[u8],
@@ -537,16 +584,8 @@ impl<'gc> Vm<'gc> {
             match result {
                 StepResult::Continue => {}
                 StepResult::ForceThunk(info) => {
-                    self.call_stack.push(CallFrame {
+                    let new_pc = self.apply_force_thunk(info);
-                        thunk: Some(info.thunk),
+                    reader.set_pc(new_pc);
-                        stack_depth: info.stack_depth,
-                        pc: info.inst_start_pc,
-                        env: self.env,
-                        with_env: self.with_env,
-                    });
-                    reader.set_pc(info.ip);
-                    self.env = info.env;
-                    self.with_env = info.with_env;
                 }
                 StepResult::Done => {
                     return Action::Done(