ludic/selfhost/emit_core.ludic

# emit_core.ludic — emitter state, type mapping, struct/slice helpers, and the
# module header. Mirrors the pieces of compiler/back/ that this subset needs.
# structs and slices are references, so every non-scalar type lowers to `ptr`.

property Val { code: pointer = null, ty: pointer = null }
function val(code: pointer, ty: pointer) -> Val { let v = new Val;  v.code = code;  v.ty = ty;  return v }

var head: Buf                 # module-level: types, globals, string constants
var code: Buf                 # function bodies
var falloc: Buf              # entry-block allocas for the current function
var ll_t: int = 0             # temp register counter (reset per function)
var ll_lbl: int = 0           # label counter
var ll_str: int = 0           # string-constant counter

# local environment (parallel slices), reset per function
var loc_name: []pointer
var loc_reg:  []pointer
var loc_ty:   []pointer
var loc_mut:  []int          # 1 = mutable (var / param / loop-var), 0 = immutable (let)
var nloc: int = 0
var g_uses_str: bool = false  # a `str + str` / `str == str` was emitted -> emit the prelude
var g_uses_intstr: bool = false  # `string(int)` was emitted -> emit the int->string prelude
var g_uses_strslice: bool = false  # `s[a..b]` was emitted -> emit the substring prelude
var g_uses_mathrt: bool = false  # Math.sqrt/sin/cos/tan was emitted -> emit the math runtime prelude
var g_uses_textrt: bool = false  # Text.upper/lower/trim/repeat/pad was emitted -> emit the text builders
var g_uses_textrt2: bool = false  # Text.split/join/replace was emitted -> emit the string/slice builders
var g_uses_hashrt: bool = false  # Hash.of/fnv1a/crc32 was emitted -> emit the byte-stream hashers
var g_uses_cryptort: bool = false  # Crypto.* was emitted -> emit the SHA-256 / HMAC runtime
var g_uses_datert: bool = false  # Date.*/DateTime.* was emitted -> emit the civil<->epoch conversions
var g_uses_longstr: bool = false  # string(long) / interpolating a long was emitted -> emit fn_long_str

# loop targets for break/continue (innermost last)
var brk_lbl: []pointer
var cnt_lbl: []pointer
var nloop: int = 0

var ret_ty: pointer              # current function's return type
var g_term: bool = false     # did the current block end in a terminator?
var self_stk: []pointer          # entity-index slot (ip) per enclosing query, for self()
var nself: int = 0
var mach_stk: []Node         # enclosing `machine` nodes, so `become` finds its register
var nmach: int = 0

# scenes: one implicit active-scene register (@L_scene). A scene lowers to a
# machine the compiler writes for you — `become <Scene>` runs the source scene's
# on-exit, stores the target id, and runs its on-enter.
var g_scenes: []Node         # every `scene` declaration, in source order (ival = id)
var g_scene_count: int = 0   # parse-time id counter
var g_start_scene: int = 0   # id of the scene marked `start` (else the first)
var g_cur_scene: Node = null # scene owning the handler being emitted, for `become`

function find_scene(name: pointer) -> Node {
  var i = 0
  while i < len(g_scenes) { if (g_scenes[i].s == name) { return g_scenes[i] };  i = i + 1 }
  return null
}

function emit(s: pointer) -> void { buf_puts(code, s) }
function emith(s: pointer) -> void { buf_puts(head, s) }

# stack slots MUST live in the entry block (an alloca in a loop walks the stack
# off its end), so they go into a per-function buffer spliced in at entry.
function emit_alloca(llt: pointer) -> pointer {
  let r = `%t{itoa(ll_t)}`;  ll_t = ll_t + 1
  buf_puts(falloc, "  ");  buf_puts(falloc, r);  buf_puts(falloc, " = alloca ");  buf_puts(falloc, llt);  buf_puts(falloc, "\n")
  return r
}

# "%t<n>" fresh register
function nreg() -> pointer { let r = `%t{itoa(ll_t)}`;  ll_t = ll_t + 1;  return r }
function lbl(pfx: pointer) -> pointer { let r = (pfx + itoa(ll_lbl));  ll_lbl = ll_lbl + 1;  return r }

# Ludic type -> LLVM type. int/bool are i32; everything else (ptr/str/struct/
# slice) is a pointer; void is void.
function llty(t: pointer) -> pointer {
  if (t == "int") or (t == "bool") or (t == "fixed") or (t == "entity") { return "i32" }   # entity = an i32 handle (self())
  if (t == "long") { return "i64" }               # a 64-bit signed integer
  if (t == "Vector") { return "i64" }             # a 2D vector: (x, y) fixeds packed into one i64
  if (t == "byte") { return "i8" }                # a single byte (p[i] on a raw ptr)
  if (t == "words") or (t == "fixeds") or (t == "pointers") { return "ptr" }   # typed buffers
  if (t == "void") { return "void" }
  return "ptr"
}

function is_slice_ty(t: pointer) -> bool { return t[0] == 91 and t[1] == 93 }   # "[]"
function slice_elem(t: pointer) -> pointer { return t[2..len(t)] }

function find_arch(name: pointer) -> Node {
  var i = 0
  while i < len(prog) { let d = prog[i];  if d.kind == N_ARCH and (d.s == name) { return d };  i = i + 1 }
  return null
}
function find_comp(name: pointer) -> Node {
  var i = 0
  while i < len(prog) { let d = prog[i];  if d.kind == N_COMP and (d.s == name) { return d };  i = i + 1 }
  return null
}
# every record is a `property` with a %Cmp_ layout of named fields — whether it
# is stored per-entity by the ECS or heap-allocated by `new` is a matter of use.
function layout_node(name: pointer) -> Node { return find_comp(name) }
function layout_ty(name: pointer) -> pointer { return (("%Cmp_") + name) }

function field_index(s: Node, fname: pointer) -> int {
  var i = 0
  while i < len(s.kids) { if (s.kids[i].s == fname) { return i }; i = i + 1 }
  return 0 - 1
}
function field_type(s: Node, fname: pointer) -> pointer {
  var i = 0
  while i < len(s.kids) { if (s.kids[i].s == fname) { return s.kids[i].ty }; i = i + 1 }
  return "int"
}

# find a global var/const by name
function find_global(name: pointer) -> Node {
  var i = 0
  while i < len(prog) {
    let d = prog[i]
    if d.kind == N_VAR and (d.s == name) { return d }
    if d.kind == N_CONST and (d.s == name) { return d }
    i = i + 1
  }
  return null
}

# `Enum.Variant` -> the variant's ordinal (its index), or -1 if `ename` names no
# enum with that variant. Enum names live in `prog` like any other declaration.
function enum_ordinal(ename: pointer, vname: pointer) -> int {
  var i = 0
  while i < len(prog) {
    let d = prog[i]
    if d.kind == N_ENUM and (d.s == ename) {
      var j = 0
      while j < len(d.kids) { if (d.kids[j].s == vname) { return j };  j = j + 1 }
    }
    i = i + 1
  }
  # LC1: the compiler owns `EndReason` — the reason bound by a reason-carrying
  # teardown (`@OnDespawn(M, reason: r)`). Each despawn site passes one of these.
  if (ename == "EndReason") {
    if (vname == "Despawned") { return 0 }      # explicit `despawn e`
    if (vname == "SceneExit") { return 1 }      # a scene tearing down its owned entities
    if (vname == "Quit")      { return 2 }      # program shutdown
  }
  return 0 - 1
}
function find_fn(name: pointer) -> Node {
  var i = 0
  while i < len(prog) { let d = prog[i];  if d.kind == N_FN and (d.s == name) { return d };  i = i + 1 }
  return null
}

# `extern function name(params) -> T = "sym"` binds a Ludic name to a link symbol. A
# call to `name` lowers to a direct `@<sym>` call (no @fn_ prefix — the string is
# the exact linked symbol), and emit_extern_decls emits a matching `declare`. This
# is the transport seam (net_send/net_poll), the windowing/socket FFI, and any
# C/Rust/Zig library binding — the same seam NETWORKING-DESIGN §5 names.
function find_extern(name: pointer) -> Node {
  var i = 0
  while i < len(prog) { let d = prog[i];  if d.kind == N_EXTERN and (d.s == name) { return d };  i = i + 1 }
  return null
}

# @Computed derived fields: a per-property (Prop.field -> expression) registry.
# These are NOT stored in the component layout; `x.field` expands inline to the
# expression with its bare names read as fields of `x`. Populated at parse time.
var g_computed: []Node          # each: s = "Prop.field", ty = result type, a = expr

function register_computed(prop: pointer, field: pointer, ty: pointer, e: Node) -> void {
  let cf = node(N_FIELD); cf.s = `{prop}.{field}`; cf.ty = ty; cf.a = e
  push(g_computed, cf)
}
function computed_expr(prop: pointer, field: pointer) -> Node {
  if (prop == null) { return null }
  let key = `{prop}.{field}`
  var i = 0
  while i < len(g_computed) { if (g_computed[i].s == key) { return g_computed[i].a };  i = i + 1 }
  return null
}
# best-effort static type of an expression (for computed-field lookup; emits nothing)
function static_type(e: Node) -> pointer {
  if e.kind == E_ID { let li = loc_find(e.s);  if li >= 0 { return loc_ty[li] } }
  return null
}

# @OnSpawn(Model) hooks: a Model -> hook-body registry. Populated at parse time;
# `spawn Model { … }` runs the body with the model's properties bound (like a
# constructor). Spawn statically knows the model, so no runtime dispatch is needed.
var g_onspawn: []Node           # each: s = Model name, a = hook body block

function register_onspawn(model: pointer, body: Node) -> void {
  let n = node(N_BLOCK); n.s = model; n.a = body; push(g_onspawn, n)
}
function onspawn_body(model: pointer) -> Node {
  var i = 0
  while i < len(g_onspawn) { if (g_onspawn[i].s == model) { return g_onspawn[i].a };  i = i + 1 }
  return null
}

# @OnDespawn(Model): a Model -> hook-body registry. Despawn does not statically
# know an entity's model, so these are emitted as functions and dispatched on the
# entity's kind at each `despawn`. @OnAttach(Property) fires per property-attach.
var g_ondespawn: []Node         # each: s = Model name, a = hook body block
var g_onattach: []Node          # each: s = Property name, a = hook body block

# LC1: `.ty` carries the optional `reason:` binding name (null if the hook took
# no reason). The despawn hook function gains an `i32 %reason` parameter and each
# teardown site passes a constant EndReason (see emit_despawn_hooks / emit_despawn).
function register_ondespawn(model: pointer, body: Node, reason: pointer) -> void {
  let n = node(N_BLOCK); n.s = model; n.a = body; n.ty = reason; push(g_ondespawn, n)
}
function ondespawn_body(model: pointer) -> Node {
  var i = 0
  while i < len(g_ondespawn) { if (g_ondespawn[i].s == model) { return g_ondespawn[i].a };  i = i + 1 }
  return null
}
function register_onattach(prop: pointer, body: Node) -> void {
  let n = node(N_BLOCK); n.s = prop; n.a = body; push(g_onattach, n)
}
function onattach_body(prop: pointer) -> Node {
  var i = 0
  while i < len(g_onattach) { if (g_onattach[i].s == prop) { return g_onattach[i].a };  i = i + 1 }
  return null
}

# @OnDetach(Property): the teardown paired with @OnAttach — fires when a property
# is removed from a live entity (`detach P on e`), with the property bound by name
# so the body can read its outgoing value before it is cleared.
var g_ondetach: []Node          # each: s = Property name, a = hook body block
function register_ondetach(prop: pointer, body: Node) -> void {
  let n = node(N_BLOCK); n.s = prop; n.a = body; push(g_ondetach, n)
}
function ondetach_body(prop: pointer) -> Node {
  var i = 0
  while i < len(g_ondetach) { if (g_ondetach[i].s == prop) { return g_ondetach[i].a };  i = i + 1 }
  return null
}

# @OnEnable(Property) / @OnDisable(Property): run when a property is toggled on an
# entity, with the property bound by name.
var g_onenable: []Node
var g_ondisable: []Node
function register_onenable(prop: pointer, body: Node) -> void { let n = node(N_BLOCK); n.s = prop; n.a = body; push(g_onenable, n) }
function register_ondisable(prop: pointer, body: Node) -> void { let n = node(N_BLOCK); n.s = prop; n.a = body; push(g_ondisable, n) }
function onenable_body(prop: pointer) -> Node {
  var i = 0
  while i < len(g_onenable) { if (g_onenable[i].s == prop) { return g_onenable[i].a };  i = i + 1 }
  return null
}
function ondisable_body(prop: pointer) -> Node {
  var i = 0
  while i < len(g_ondisable) { if (g_ondisable[i].s == prop) { return g_ondisable[i].a };  i = i + 1 }
  return null
}

# EV0 — the event bus core. `event E { fields }` declares a POD payload; `@On(E)
# handler …` registers a compile-time listener; `emit E(…)` fires it. An event
# lowers to a `@ev_<E>(payload)` function whose body is the concatenation of its
# listeners in declaration order — a direct call at each `emit` site, no runtime.
# Gated on `len(g_events) > 0`, so a program with no events is byte-identical.
var g_events: []Node            # each: an N_EVENT node (s = name, kids = payload fields, ival=1 if cancellable)
var g_onlisten: []Node          # each: N_BLOCK, s = event name, a = listener body block
var g_cancel_addr: pointer = null   # EV3: address of the current cancellable dispatch's flag (null outside one)

function register_event(n: Node) -> void { push(g_events, n) }
function find_event(name: pointer) -> Node {
  var i = 0
  while i < len(g_events) { if (g_events[i].s == name) { return g_events[i] };  i = i + 1 }
  return null
}
function register_onlisten(evt: pointer, body: Node) -> void {
  let n = node(N_BLOCK);  n.s = evt;  n.a = body;  push(g_onlisten, n)
}

# EV1 — `@Public` promotes a lifecycle hook to a public event. A promoted event
# is synthesized here (payload = the entity, plus a reason for despawn); its
# presence in g_events is what makes each lifecycle fire site also `emit` it, so
# `find_event(name) != null` doubles as the "is this hook public?" test. Names are
# the stable ABI contract: `model_<M>_spawn`, `model_<M>_despawn`, etc.
function ensure_event(name: pointer, with_reason: bool) -> void {
  if (find_event(name) != null) { return }
  let n = node(N_EVENT);  n.s = name
  let ent = node(N_FIELD);  ent.s = "entity";  ent.ty = "int";  push(n.kids, ent)
  if with_reason { let r = node(N_FIELD);  r.s = "reason";  r.ty = "int";  push(n.kids, r) }
  register_event(n)
}
# a promoted scene/program event has no per-entity payload
function ensure_event_empty(name: pointer) -> void {
  if (find_event(name) != null) { return }
  let n = node(N_EVENT);  n.s = name;  register_event(n)
}

# EV1/SCENES-E2 — layer toggle. A layer named in an `enable layer L`/`disable
# layer L` statement becomes "managed": it gets an @LE_<L> enabled flag and its
# handlers gate on it. Only managed layers pay for this, so a scene program that
# never toggles a layer is byte-identical.
var g_toggled_layers: []pointer
function note_toggled_layer(name: pointer) -> void {
  var i = 0
  while i < len(g_toggled_layers) { if (g_toggled_layers[i] == name) { return };  i = i + 1 }
  push(g_toggled_layers, name)
}
function is_toggled_layer(name: pointer) -> bool {
  var i = 0
  while i < len(g_toggled_layers) { if (g_toggled_layers[i] == name) { return true };  i = i + 1 }
  return false
}

# is `name` a model (archetype)? — chooses model-vs-handler for a bare enable/disable
function is_model(name: pointer) -> bool { return find_arch_id(name) > 0 }

# local variable environment
function loc_reset() -> void { nloc = 0 }
# push a local. Defaults to mutable (params, loop and query bindings are all
# reassignable/rebindable); a `let` binding marks its slot immutable afterward
# via loc_set_mut, so a later `name = …` can be rejected.
function loc_push(name: pointer, r: pointer, ty: pointer) -> void {
  if nloc < len(loc_name) { loc_name[nloc] = name;  loc_reg[nloc] = r;  loc_ty[nloc] = ty;  loc_mut[nloc] = 1 }
  else { push(loc_name, name);  push(loc_reg, r);  push(loc_ty, ty);  push(loc_mut, 1) }
  nloc = nloc + 1
}
function loc_set_mut(m: int) -> void { if nloc > 0 { loc_mut[nloc - 1] = m } }
function loc_find(name: pointer) -> int {
  var i = nloc - 1
  while i >= 0 { if (loc_name[i] == name) { return i }; i = i - 1 }
  return 0 - 1
}