Interpreters

Interpreters and Bytecode VMs

Purpose

Guide agents through implementing efficient bytecode interpreters and simple JITs in C/C++: dispatch strategies, VM architecture choices, and performance patterns.

Triggers

• "How do I implement a fast bytecode dispatch loop?"
• "What is the difference between switch dispatch and computed goto?"
• "How do I implement a register-based vs stack-based VM?"
• "How do I add basic JIT compilation to my interpreter?"
• "Why is my interpreter slow?"

Workflow

1. VM architecture choice

Stack-based: easier to implement, compile to; code generation is simpler. More instructions per expression. Register-based: fewer dispatch iterations; needs register allocation in the compiler; better cache behaviour for complex expressions.

2. Dispatch loop strategies

Switch dispatch (simplest, baseline)

while (1) {
    uint8_t op = *ip++;
    switch (op) {
        case OP_LOAD:  push(constants[*ip++]); break;
        case OP_ADD:   { Value b = pop(); Value a = pop(); push(a + b); } break;
        case OP_HALT:  return;
        // ...
    }
}

Problem: switch compiles to a single indirect branch from a jump table. Modern CPUs can mispredict it heavily because the same indirect branch is used for all opcodes.

Computed goto (GCC/Clang extension — fastest portable approach)

// Table of label addresses
static const void *dispatch_table[] = {
    [OP_LOAD]  = &&op_load,
    [OP_ADD]   = &&op_add,
    [OP_HALT]  = &&op_halt,
    // ...
};

#define DISPATCH() goto *dispatch_table[*ip++]

DISPATCH();  // start

op_load:
    push(constants[*ip++]);
    DISPATCH();

op_add: {
    Value b = pop(); Value a = pop(); push(a + b);
    DISPATCH();
}

op_halt:
    return;

Each opcode ends with its own indirect branch. The CPU can train the branch predictor per-opcode, dramatically improving prediction rates.

Note: &&label is a GCC/Clang extension, not standard C. Use #ifdef __GNUC__ to guard and fall back to switch for other compilers.

Direct threaded code (most aggressive)

Each bytecode word is a function pointer or label address; the VM is the fetch-decode-execute loop itself.

typedef void (*Handler)(VM *vm);

// Bytecode is an array of handlers
Handler bytecode[] = { op_load, op_push_1, op_add, op_halt };

for (int i = 0; ; i++) {
    bytecode[i](vm);
}

3. Value representation

Tagged pointer: Store type tag in low bits of pointer (pointer alignment guarantees ≥ 2 bits free):

typedef uintptr_t Value;
#define TAG_INT    0x0
#define TAG_FLOAT  0x1
#define TAG_PTR    0x2
#define TAG_MASK   0x3

#define INT_VAL(v)   ((int64_t)(v) >> 2)
#define FLOAT_VAL(v) (*(float*)((v) & ~TAG_MASK))
#define IS_INT(v)    (((v) & TAG_MASK) == TAG_INT)

NaN boxing (64-bit): Store non-double values in NaN bit patterns:

// IEEE 754 quiet NaN: exponent all 1s, mantissa != 0
// Use high mantissa bits as type tag, low 48 bits as payload
// Allows pointer/int/bool/nil to fit in a double-sized slot

Used by V8 (formerly), LuaJIT, JavaScriptCore.

4. Stack management

#define STACK_SIZE 4096
Value stack[STACK_SIZE];
Value *sp = stack;  // stack pointer

#define PUSH(v) (*sp++ = (v))
#define POP()   (*--sp)
#define TOP()   (sp[-1])
#define PEEK(n) (sp[-(n)-1])

// Check for overflow
#define PUSH_SAFE(v) do { \
    if (sp >= stack + STACK_SIZE) { vm_error("stack overflow"); } \
    PUSH(v); \
} while(0)

5. Inline caching (IC)

Inline caching speeds up property lookups and method dispatch by caching the last observed type at each call site.

struct CallSite {
    Type   cached_type;      // Last observed receiver type
    void  *cached_method;    // Cached function pointer
    int    miss_count;       // Number of misses
};

void invoke_method(VM *vm, CallSite *cs, Value receiver, ...) {
    Type t = GET_TYPE(receiver);
    if (t == cs->cached_type) {
        // Cache hit: direct call, no lookup
        cs->cached_method(vm, receiver, ...);
    } else {
        // Cache miss: look up, update cache
        void *method = lookup_method(t, name);
        cs->cached_type = t;
        cs->cached_method = method;
        cs->miss_count++;
        method(vm, receiver, ...);
    }
}

Polymorphic IC (PIC): cache up to N (typ. 4) type-method pairs.

6. Simple JIT (mmap + machine code)

For x86-64: allocate executable memory, write machine code bytes, call it.

#include <sys/mman.h>
#include <string.h>

typedef int (*JitFn)(int a, int b);

JitFn jit_compile_add(void) {
    // x86-64: add rdi, rsi; mov rax, rdi; ret
    static const uint8_t code[] = {
        0x48, 0x01, 0xF7,   // add rdi, rsi
        0x48, 0x89, 0xF8,   // mov rax, rdi
        0xC3                 // ret
    };

    void *mem = mmap(NULL, sizeof(code),
                     PROT_READ | PROT_WRITE | PROT_EXEC,
                     MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
    if (mem == MAP_FAILED) return NULL;

    memcpy(mem, code, sizeof(code));

    // On some systems (macOS Apple Silicon): must use mprotect
    // mprotect(mem, sizeof(code), PROT_READ | PROT_EXEC);

    return (JitFn)mem;
}

On macOS Apple Silicon (M-series): use pthread_jit_write_protect_np() or MAP_JIT flag.

7. Performance tips

1. Dispatch: Use computed goto over switch on GCC/Clang 2. Values: Use NaN boxing or tagged pointers; avoid boxing/unboxing in hot paths 3. Stack: Keep stack pointer in a callee-saved register (register Value *sp asm("r15") — GCC global register variable) 4. Locals access: Keep frequently accessed locals in VM registers (struct fields), not stack 5. Profiling: Use perf or sampling to find dispatch overhead vs actual work 6. Specialisation: Generate specialised handler variants for common type combinations (int+int add vs generic add) 7. Trace recording: Trace JITs (LuaJIT approach) compile hot traces instead of full functions

For a benchmark of dispatch strategies, see references/benchmarks.md.

Related skills

• Use skills/profilers/linux-perf to profile the interpreter dispatch loop
• Use skills/low-level-programming/assembly-x86 to understand JIT output
• Use skills/runtimes/fuzzing to fuzz the bytecode parser/loader
• Use skills/compilers/llvm for LLVM IR-based JIT (MCJIT / ORC JIT)

Interpreter Dispatch Benchmark Notes

Dispatch strategy relative performance

Results vary by workload and CPU, but the general ordering is consistent:

Source: various academic papers and LuaJIT blog posts by Mike Pall.

Measuring dispatch overhead

# Profile the interpreter loop
perf record -g -F 9999 ./my_interpreter benchmark.bc
perf report --no-children   # look at self time

# Expected: high self-time on the dispatch instruction itself
# means CPU is mispredicting the indirect branch

Symptom of dispatch overhead: perf report shows high time at the jmp [r14] (or equivalent indirect jump) instruction, not inside the opcode handlers.

Fix: switch to computed goto or add a tracing JIT.

JIT threshold heuristics

Most production runtimes use multi-tier JIT:

LuaJIT uses trace recording: compile a hot loop's trace after 56 iterations. V8 uses function-level tiering: Ignition (bytecode) → Maglev (mid-tier) → Turbofan (optimising).

Stack frame layout recommendations

// Keep hot fields first in the VM struct (better cache line usage)
typedef struct VM {
    Value  *sp;        // stack pointer — accessed every instruction
    uint8_t *ip;       // instruction pointer — accessed every instruction
    Value  *stack;     // stack base
    Value  *locals;    // current frame locals
    // ... less-hot fields ...
    int     call_depth;
    GC     *gc;
    Table  *globals;
} VM;

Pack hot VM state into 1–2 cache lines (64 bytes each). Accessing sp and ip should never cause cache misses in a tight loop.