GitHub - karans4/sysp: A systems Lisp that compiles to C. Homoiconic macros, refcounted memory, Hindley-Milner type inference.

sysp

A systems Lisp that compiles to readable C. Designed for numerical computing and GPU work, where you want to generate specialized kernels at compile time using homoiconic macros. Manual memory management, Hindley-Milner type inference with monomorphization, and direct C generation with no runtime.

Prior Art

sysp draws inspiration from Carp, Zig, and other systems Lisps.

Why

Most numerical computing is either written in C/CUDA (fast but verbose) or Python/JAX (expressive but slow). sysp targets the middle: write Lisp macros that generate specialized C or CUDA code at compile time. You get homoiconic code generation for kernel specialization, gradual typing without annotation burden, and readable generated code you can inspect and debug. The type system infers precise types where possible and falls back to a boxed Value type (tagged union) where it can't.

Features

• S-expression syntax with [] array literals
• Compile-time macros - defmacro over s-expressions, quasiquote with ~ and ~@
• Compile-time evaluation - defn-ct for real homoiconic metaprogramming with gensym
• Tail-call optimization - recur compiles to goto
• Cons cells with automatic refcounting (val_retain/val_release)
• Symbols as interned integers with compile-time symbol table
• Structs, enums, tuples, vectors with generated C type definitions
• Foreign function interface - extern declarations, pointer types, casts
• Type inference - Hindley-Milner unification engine (WIP: constraint generation, monomorphization)

Example

(defn factorial [n :int, acc :int] :int
  (if (== n 0) (return acc)
  else (recur (- n 1) (* acc n)))
  0)

(defn main []
  (println (factorial 12 1)))

Compiles to:

int factorial(int n, int acc) {
_recur_top:
    if ((n == 0)) { return acc; }
    else { int _t0 = (n - 1); int _t1 = (acc * n); n = _t0; acc = _t1; goto _recur_top; }
    return 0;
}
int main(void) {
    printf("%d\n", factorial(12, 1));
    return 0;
}

Building

Requires SBCL (Steel Bank Common Lisp) and a C compiler.

# Compile a .sysp file to C
sbcl --script sysp.lisp input.sysp output.c

# Then compile the C
cc output.c -o program

For the raylib demo (tic-tac-toe):

sbcl --script sysp.lisp tictactoe.sysp tictactoe.c
cc tictactoe.c -lraylib -lm -o tictactoe

Architecture

The compiler is a single Common Lisp file (sysp.lisp, ~1900 lines):

1. Read - custom readtable for [] arrays and backquote syntax
2. Macro-expand - recursive expansion of defmacro + built-in macros (->, ->>, when-let, dotimes, etc.)
3. Infer (WIP) - Hindley-Milner type inference with unification, generalize/instantiate for let-polymorphism
4. Compile - AST → C code generation with type tracking
5. Emit - headers, type declarations, forward decls, function bodies

Type System (in progress)

The goal is gradual typing with aggressive monomorphization:

• Unresolved types box into Value (tagged union: int, float, string, symbol, cons)
• Resolved types generate specialized C: Vector_int, Fn_int_int, Cons_int_symbol
• The inference engine supports type variables, structural unification, occurs check, and let-polymorphism
• Value acts as the "any" type - unifies with everything, zero annotation burden

Roadmap

• Core language (functions, control flow, operators)
• Macro system (defmacro + compile-time eval)
• Cons cells with refcounting
• Symbols, quasiquote, splice
• Type inference (constraint generation from AST)
• Monomorphization (specialized C output per concrete type)
• Pattern matching / destructuring
• Tagged unions (sum types)
• Borrow checker (linear types)
• CUDA backend (GPU kernels from Lisp)
• HPC primitives (SIMD, parallel loops)
• Self-hosting

License

MIT