/* AILang reference-counting runtime.
 *
 * Allocator + counter primitives for `ail build --alloc=rc`, plus the
 * worklist allocator for `(drop-iterative)` data types. The runtime
 * shape is set by Decision 10 (RC + uniqueness inference) and grew
 * across Iter 18b–18e:
 *
 *   - 18b shipped the 8-byte-refcount-header layout and
 *     `ailang_rc_alloc`. Programs leaked everything because codegen
 *     did not yet emit inc/dec.
 *   - 18c.3 added `ailang_rc_inc` / `ailang_rc_dec` emission at
 *     `Term::Clone` and at `Term::Let` scope close (when the binder is
 *     a unique RC-allocated value).
 *   - 18c.4 added per-type `drop_<m>_<T>(ptr)` functions emitted by
 *     codegen; on dec-to-zero the cascade walks pointer-typed fields
 *     before freeing the outer cell.
 *   - 18d.1–18d.4 added explicit `(reuse-as)` rewrites and move-aware
 *     pattern-binder + Own-param dec at scope close, all built on top
 *     of these runtime primitives without changing the ABI.
 *   - 18e added `ailang_drop_worklist_*` for ADTs annotated
 *     `(drop-iterative)`. Codegen swaps the recursive cascade for a
 *     worklist loop; deep ADT chains free without stack growth.
 *
 * The ABI defined here is stable; adding behaviour to the codegen
 * (further uniqueness elision, atomic refcounts under threading, etc.)
 * will not require runtime changes unless the layout itself shifts.
 *
 * Layout:
 *
 *   high address ┐
 *                │  payload (size bytes, 8-byte aligned)
 *                ┤  ← returned pointer (`p`)
 *                │  uint64_t refcount    ← header (8 bytes)
 *   low address  ┘  ← ailang_rc_alloc's internal allocation
 *
 * The returned pointer points to the *payload*. The header is at
 * `p - 8`. Codegen treats the returned pointer exactly like a
 * `GC_malloc`-returned pointer; it stores the ADT tag at offset 0,
 * fields from offset 8, env-cells from offset 0 in lambda envs, etc.
 *
 * FROZEN ABI for the embedding boundary — see DESIGN.md §"Embedding
 * ABI" > "Frozen value layout". A boundary-crossing single-ctor
 * scalar record's box offsets MUST NOT move.
 *
 * Single-threaded: counter ops are non-atomic. AILang has no
 * concurrency primitives yet; when it acquires them, atomic-vs-non-
 * atomic becomes a separate decision per allocation kind (see
 * Decision 10's "Does not commit to atomic refcounts" clause).
 */

#include <stddef.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>

/* Header lives in the 8 bytes preceding every payload. */
typedef uint64_t ailang_rc_header_t;

#define HEADER_SIZE ((size_t)sizeof(ailang_rc_header_t))

static inline ailang_rc_header_t *header_of(void *payload) {
    return (ailang_rc_header_t *)((uint8_t *)payload - HEADER_SIZE);
}

/* ---------------------------------------------------------------------------
 * Iter 18g.0: opt-in alloc/free stats counter.
 *
 * Two non-atomic 64-bit counters incremented from `ailang_rc_alloc` and
 * the to-zero branch of `ailang_rc_dec`. Their difference is the live
 * cell count at any point in execution; at program exit a non-zero
 * difference is a leak.
 *
 * Output is gated by the `AILANG_RC_STATS` env var: when set to a
 * non-empty value, an `atexit` handler prints
 *
 *     ailang_rc_stats: allocs=N frees=M live=K
 *
 * to stderr. The default-disabled path keeps production binaries quiet
 * and adds only two unconditional `++` operations to the hot path —
 * negligible relative to the libc malloc/free already on each path,
 * and the bench numbers in JOURNAL 18f.2 were taken with the counters
 * compiled in but disabled, so the figures are still valid.
 *
 * Single-threaded; non-atomic. Same scope as the rest of `runtime/rc.c`.
 * The counters are intentionally not exposed via FFI symbols — the
 * env-var-gated atexit print is the supported readback path, and that
 * is sufficient for the e2e leak tests that consume the diagnostic.
 * --------------------------------------------------------------------------- */
static uint64_t g_rc_alloc_count = 0;
static uint64_t g_rc_free_count = 0;

/* M2: per-thread embedding context. A scalar kernel allocates nothing,
 * so this carries only the de-globalised RC accounting. Set by the
 * generated C `@<sym>` forwarder into __ail_tls_ctx for the synchronous
 * duration of one call (never held across a suspension point). The
 * g_rc_* statics + atexit below are RETAINED as the null-ctx
 * (single-threaded executable) fallback. */
typedef struct ailang_ctx {
    uint64_t alloc_count;
    uint64_t free_count;
} ailang_ctx_t;

__thread ailang_ctx_t *__ail_tls_ctx = NULL;

ailang_ctx_t *ailang_ctx_new(void) {
    return (ailang_ctx_t *)calloc(1, sizeof(ailang_ctx_t));
}

void ailang_ctx_free(ailang_ctx_t *ctx) {
    if (ctx != NULL) {
        const char *flag = getenv("AILANG_RC_STATS");
        if (flag != NULL && flag[0] != '\0') {
            fprintf(stderr,
                    "ailang_rc_stats: allocs=%llu frees=%llu live=%lld\n",
                    (unsigned long long)ctx->alloc_count,
                    (unsigned long long)ctx->free_count,
                    (long long)(ctx->alloc_count - ctx->free_count));
        }
    }
    free(ctx);
}

static void ailang_rc_stats_atexit(void) {
    fprintf(stderr,
            "ailang_rc_stats: allocs=%llu frees=%llu live=%lld\n",
            (unsigned long long)g_rc_alloc_count,
            (unsigned long long)g_rc_free_count,
            (long long)(g_rc_alloc_count - g_rc_free_count));
}

__attribute__((constructor))
static void ailang_rc_stats_install(void) {
    const char *flag = getenv("AILANG_RC_STATS");
    if (flag != NULL && flag[0] != '\0') {
        atexit(ailang_rc_stats_atexit);
    }
}

/* Allocate `size` bytes of payload, prefixed by an 8-byte refcount
 * header initialised to 1. Returns a pointer to the payload.
 *
 * Aborts on out-of-memory; AILang has no exception machinery yet, and
 * Boehm's behaviour on OOM is also "abort", so this matches.
 *
 * Zero-initialises the payload to match `GC_malloc`'s contract — codegen
 * may rely on uninitialised fields reading as zero in some paths. */
void *ailang_rc_alloc(size_t size) {
    void *block = malloc(HEADER_SIZE + size);
    if (block == NULL) {
        fprintf(stderr,
                "ailang_rc_alloc: out of memory (requested payload %zu bytes)\n",
                size);
        abort();
    }
    ailang_rc_header_t *hdr = (ailang_rc_header_t *)block;
    *hdr = 1;
    void *payload = (uint8_t *)block + HEADER_SIZE;
    memset(payload, 0, size);
    ailang_ctx_t *_ctx = __ail_tls_ctx;
    if (_ctx != NULL) _ctx->alloc_count++; else g_rc_alloc_count++;
    return payload;
}

/* Refcount += 1. No-op on null (codegen never asks for inc on a known-
 * null pointer, but defensive — top-level fn-value pointers may be
 * null-env closure pairs in static memory which must not be incremented). */
void ailang_rc_inc(void *payload) {
    if (payload == NULL) {
        return;
    }
    /* Static closure-pair env pointers (Iter 8b) live in the LLVM data
     * segment, not in heap memory we allocated. Codegen elides inc/dec
     * for known-static pointers (the `@`-prefix gate added in 18c.3),
     * so this path is not reached for them in practice. The runtime
     * itself has no header-bit flag distinguishing static from heap;
     * if codegen ever loses the elision, inc on a static pointer is
     * undefined behaviour. */
    ailang_rc_header_t *hdr = header_of(payload);
    *hdr += 1;
}

/* Refcount -= 1. If it reaches zero, frees the underlying block.
 *
 * This function performs only the *outer* free. The per-type
 * `drop_<m>_<T>(ptr)` functions emitted by codegen (Iter 18c.4) are
 * what walk pointer-typed children before calling `ailang_rc_dec` on
 * the outer cell. For ADTs annotated `(drop-iterative)`, codegen
 * emits a worklist loop using `ailang_drop_worklist_*` (Iter 18e)
 * instead of recursive cascade, allowing arbitrarily deep ADT chains
 * to free without stack growth.
 *
 * Calling `ailang_rc_dec` directly on a pointer whose type has boxed
 * children will leak those children. Codegen routes through
 * `drop_<m>_<T>` when it knows the type; this entry point is the
 * shared bottom they all converge on. */
void ailang_rc_dec(void *payload) {
    if (payload == NULL) {
        return;
    }
    ailang_rc_header_t *hdr = header_of(payload);
    if (*hdr == 0) {
        fprintf(stderr,
                "ailang_rc_dec: refcount underflow at %p (already zero)\n",
                payload);
        abort();
    }
    *hdr -= 1;
    if (*hdr == 0) {
        free(hdr);
        ailang_ctx_t *_ctx = __ail_tls_ctx;
        if (_ctx != NULL) _ctx->free_count++; else g_rc_free_count++;
    }
}

/* ---------------------------------------------------------------------------
 * Iter 18e: drop worklist.
 *
 * Backs the `(drop-iterative)` data attribute. When a type is annotated
 * `(drop-iterative)`, codegen emits `drop_<m>_<T>` with an iterative-with-
 * worklist body in place of the recursive cascade. The worklist is a
 * heap-allocated stretchy buffer of `void*` pointers — one entry per
 * not-yet-processed cell. Each entry is mono-typed to T (the annotated
 * ADT being dropped); fields of T whose type is `T` itself are pushed,
 * fields whose type is a different ADT call that ADT's drop fn directly.
 * (See `Emitter::emit_iterative_drop_fn_for_type` in the codegen for the
 * IR shape and the same-type / different-type dispatch.)
 *
 * Strategy: heap-allocated buffer, doubled on overflow. We chose this
 * over a stack-allocated small buffer (overcomplicates the IR seam — the
 * codegen body would need to track "which buffer is live") and over Lean
 * 4's "thread the worklist through one of the cell's own pointer slots"
 * technique (requires the codegen to know which slot of each ctor is
 * "free to repurpose" — non-trivial since AILang ctors are heterogeneous
 * and slot 0 is always the tag). The runtime-helper approach keeps the
 * IR-level body of `drop_<m>_<T>` small: three calls (new / push / pop /
 * free) drive the loop.
 *
 * Precedent: Lean 4's `lean_dec_ref_cold` and Roc's iterative-free path
 * both use a worklist to break tail recursion in their drop cascades.
 * Lean threads the worklist through field slots (the "in-place" variant);
 * we use a separate heap buffer because AILang's ctor layout makes slot
 * repurposing fragile. The semantic invariant matches: every cell whose
 * refcount reaches zero is dec'd exactly once, regardless of cascade
 * depth, without consuming proportional C stack space.
 *
 * Single-threaded; non-atomic. Same scope as the rest of `runtime/rc.c`.
 * --------------------------------------------------------------------------- */

typedef struct {
    void **data;     /* heap buffer of `cap` pointers; null once freed */
    size_t len;      /* number of live entries (always <= cap) */
    size_t cap;      /* current capacity in slots */
} ailang_drop_worklist_t;

/* Initial capacity. 16 slots * 8 bytes = 128 bytes — small enough that
 * very-shallow drops don't waste memory, large enough that 16-deep
 * cascades (very common) never realloc. Doubled on overflow. */
#define DROP_WORKLIST_INIT_CAP ((size_t)16)

void *ailang_drop_worklist_new(void) {
    ailang_drop_worklist_t *wl = malloc(sizeof(ailang_drop_worklist_t));
    if (wl == NULL) {
        fprintf(stderr,
                "ailang_drop_worklist_new: out of memory (header)\n");
        abort();
    }
    wl->data = malloc(DROP_WORKLIST_INIT_CAP * sizeof(void *));
    if (wl->data == NULL) {
        fprintf(stderr,
                "ailang_drop_worklist_new: out of memory (initial buffer)\n");
        abort();
    }
    wl->len = 0;
    wl->cap = DROP_WORKLIST_INIT_CAP;
    return (void *)wl;
}

/* Push `payload` onto the worklist. Skips null payloads — pushed nulls
 * would dispatch on `load i64, ptr null` at pop time and segfault, so
 * we filter here. The check is symmetric with `ailang_rc_dec`'s null
 * guard (a null payload is a no-op everywhere in the rc runtime). */
void ailang_drop_worklist_push(void *wl_opaque, void *payload) {
    if (payload == NULL) {
        return;
    }
    ailang_drop_worklist_t *wl = (ailang_drop_worklist_t *)wl_opaque;
    if (wl->len == wl->cap) {
        size_t new_cap = wl->cap * 2;
        void **new_data = realloc(wl->data, new_cap * sizeof(void *));
        if (new_data == NULL) {
            fprintf(stderr,
                    "ailang_drop_worklist_push: out of memory (grow to %zu slots)\n",
                    new_cap);
            abort();
        }
        wl->data = new_data;
        wl->cap = new_cap;
    }
    wl->data[wl->len] = payload;
    wl->len += 1;
}

/* Pop one payload from the worklist. Returns NULL when the worklist is
 * empty. Since `push` filters nulls, a returned null is unambiguous and
 * can be used by the IR body as the loop-exit sentinel. */
void *ailang_drop_worklist_pop(void *wl_opaque) {
    ailang_drop_worklist_t *wl = (ailang_drop_worklist_t *)wl_opaque;
    if (wl->len == 0) {
        return NULL;
    }
    wl->len -= 1;
    return wl->data[wl->len];
}

/* Free the worklist itself. Called once at the end of the iterative
 * drop loop. Does NOT free any payloads still in the buffer — the IR
 * body must drain the buffer first via repeated `pop` calls before
 * calling free. */
void ailang_drop_worklist_free(void *wl_opaque) {
    if (wl_opaque == NULL) {
        return;
    }
    ailang_drop_worklist_t *wl = (ailang_drop_worklist_t *)wl_opaque;
    free(wl->data);
    free(wl);
}