/* test_stack.c -- the F18 circular stacks, checked against an independent
 * model of the same semantics.
 *
 * The implementation in stack.c is specified in DECOMPOSITION.md as
 * "T, S + 8 circular" / "R + 8 circular".  This test does not take that
 * decomposition on trust.  It builds a second, deliberately dumb model of
 * what D-2 describes -- a flat fixed-depth circular buffer with no bounds
 * check at all -- and drives both with identical operation sequences,
 * requiring them to agree after every single operation.  If the register/ring
 * bookkeeping in stack.c has a wrong pointer direction or an off-by-one in
 * the wrap, this catches it; an inspection-only check would not.
 *
 * Depth, ordering, and the "silently overwrites the oldest entry" behaviour
 * are additionally pinned with explicit cases, because those are the three
 * properties the ISA's push-heavy words actually depend on.
 *
 * Built and run at both V4_CELL_BITS=32 and 64; see v4/Makefile.
 */
#include "v4/stack.h"

#include <stdio.h>
#include <stdlib.h>

static int failures = 0;
static int checks   = 0;

#define CHECK(cond, ...)                                                   \
    do {                                                                   \
        checks++;                                                          \
        if (!(cond)) {                                                     \
            failures++;                                                    \
            printf("  FAIL %s:%d: ", __FILE__, __LINE__);                  \
            printf(__VA_ARGS__);                                           \
            printf("\n");                                                  \
        }                                                                  \
    } while (0)

/* ---- the independent reference model ---------------------------------------
 * A flat circular buffer of depth N with no overflow or underflow detection,
 * which is exactly what D-2 describes.  top indexes the newest element. */

typedef struct {
    v4_cell buf[64];
    unsigned top;
    unsigned depth;
} flat_t;

static void flat_reset(flat_t *f, unsigned depth)
{
    f->depth = depth;
    f->top = 0;
    for (unsigned i = 0; i < 64; i++) f->buf[i] = 0;
}

static void flat_push(flat_t *f, v4_cell x)
{
    f->top = (f->top + 1u) % f->depth;
    f->buf[f->top] = x;
}

static v4_cell flat_pop(flat_t *f)
{
    v4_cell x = f->buf[f->top];
    f->top = (f->top + f->depth - 1u) % f->depth;
    return x;
}

static v4_cell flat_peek(const flat_t *f)
{
    return f->buf[f->top];
}

/* ---- deterministic PRNG ---------------------------------------------------
 * xorshift64, so a failure is reproducible from the seed alone.  No rand(),
 * whose sequence is implementation-defined and would make a failure on one
 * host unreproducible on another. */

static uint64_t rng_state = 0x9E3779B97F4A7C15ull;

static uint64_t rng_next(void)
{
    uint64_t x = rng_state;
    x ^= x << 13;
    x ^= x >> 7;
    x ^= x << 17;
    rng_state = x;
    return x;
}

/* ---- tests ---------------------------------------------------------------- */

static void test_reset_is_zero(void)
{
    v4_dstack d;
    v4_rstack r;
    v4_dstack_reset(&d);
    v4_rstack_reset(&r);
    CHECK(v4_dstack_peek(&d) == 0, "data stack peek after reset != 0");
    CHECK(v4_dstack_peek2(&d) == 0, "data stack peek2 after reset != 0");
    CHECK(v4_rstack_peek(&r) == 0, "return stack peek after reset != 0");
}

static void test_exact_depth_data(void)
{
    /* Fill to exactly V4_DATA_DEPTH, then drain and confirm the order. */
    v4_dstack d;
    flat_t f;
    v4_dstack_reset(&d);
    flat_reset(&f, V4_DATA_DEPTH);

    for (unsigned i = 0; i < V4_DATA_DEPTH; i++) {
        v4_cell v = (v4_cell)(i + 1);
        v4_dstack_push(&d, v);
        flat_push(&f, v);
    }
    CHECK(v4_dstack_peek(&d) == (v4_cell)V4_DATA_DEPTH,
          "peek at full depth should be the last pushed value (%d)",
          (int)V4_DATA_DEPTH);

    for (unsigned i = V4_DATA_DEPTH; i > 0; i--) {
        v4_cell got = v4_dstack_pop(&d);
        v4_cell exp = flat_pop(&f);
        CHECK(got == exp, "drain at depth %u: got %lld want %lld", i,
              (long long)got, (long long)exp);
        CHECK(got == (v4_cell)i, "drain at depth %u: got %lld want %d", i,
              (long long)got, (int)i);
    }
}

static void test_overflow_overwrites_oldest(void)
{
    /* D-2: "pushing past the bottom silently overwrites the oldest entry."
     * Push one past depth: the very first value pushed must be gone, and the
     * remaining V4_DATA_DEPTH-1 must come back newest-first. */
    v4_dstack d;
    flat_t f;
    v4_dstack_reset(&d);
    flat_reset(&f, V4_DATA_DEPTH);

    for (unsigned i = 0; i < V4_DATA_DEPTH + 1u; i++) {
        v4_cell v = (v4_cell)(i + 1);
        v4_dstack_push(&d, v);
        flat_push(&f, v);
    }
    CHECK(v4_dstack_peek(&d) == (v4_cell)(V4_DATA_DEPTH + 1u),
          "peek after overflow should be the newest value");

    for (unsigned i = 0; i < V4_DATA_DEPTH; i++) {
        v4_cell got = v4_dstack_pop(&d);
        v4_cell exp = flat_pop(&f);
        CHECK(got == exp, "post-overflow drain %u: got %lld want %lld", i,
              (long long)got, (long long)exp);
        CHECK(got != 1,
              "post-overflow drain %u: value 1 should have been overwritten",
              i);
    }
}

static void test_exact_depth_return(void)
{
    v4_rstack r;
    flat_t f;
    v4_rstack_reset(&r);
    flat_reset(&f, V4_RET_DEPTH);

    for (unsigned i = 0; i < V4_RET_DEPTH + 3u; i++) {
        v4_cell v = (v4_cell)(i + 1);
        v4_rstack_push(&r, v);
        flat_push(&f, v);
    }
    for (unsigned i = 0; i < V4_RET_DEPTH; i++) {
        v4_cell got = v4_rstack_pop(&r);
        v4_cell exp = flat_pop(&f);
        CHECK(got == exp, "return drain %u: got %lld want %lld", i,
              (long long)got, (long long)exp);
    }
}

static void test_underflow_wraps_rather_than_trapping(void)
{
    /* D-2 says there is no underflow detection.  Popping an "empty" stack must
     * therefore return a defined (stale) value, not fault and not abort.  The
     * golden model must diverge from hardware in neither direction, so this is
     * asserted rather than left to chance. */
    v4_dstack d;
    flat_t f;
    v4_dstack_reset(&d);
    flat_reset(&f, V4_DATA_DEPTH);

    for (unsigned i = 0; i < 5; i++) {
        v4_dstack_pop(&d);
        flat_pop(&f);
    }
    CHECK(1, "popping past empty must not trap");
}

/* ---- live-depth-aware comparison ------------------------------------------
 * D-2 specifies what these stacks do while they hold live entries: LIFO order
 * within the depth, and oldest-entry-overwritten past it.  It explicitly does
 * NOT specify the contents once the stack has been popped empty, because
 * "no overflow or underflow" means the residue is whatever the physical
 * register file happened to hold.  Two independent models of a circular buffer
 * will legitimately disagree down there -- it is not a semantic difference.
 *
 * So the differential test tracks how many live entries each stack holds and
 * asserts agreement only where the spec makes a claim: the value a pop returns
 * and the top peek while depth > 0, and the second element while depth > 1.
 * The stale region is still exercised (the sequence runs right through it) and
 * is still required not to trap, it is simply not required to agree. */

typedef struct {
    v4_dstack hw;
    flat_t    model;
    int       live;
} pair_t;

static void pair_reset(pair_t *p)
{
    v4_dstack_reset(&p->hw);
    flat_reset(&p->model, V4_DATA_DEPTH);
    p->live = 0;
}

static void pair_step(pair_t *p, int push, v4_cell v, int step, const char *tag)
{
    if (push) {
        v4_dstack_push(&p->hw, v);
        flat_push(&p->model, v);
        if (p->live < V4_DATA_DEPTH) p->live++;
    } else {
        v4_cell got = v4_dstack_pop(&p->hw);
        v4_cell exp = flat_pop(&p->model);
        if (p->live > 0) {
            CHECK(got == exp, "%s op %d: pop got %lld want %lld (live=%d)",
                  tag, step, (long long)got, (long long)exp, p->live);
            p->live--;
        }
    }
    if (p->live > 0) {
        CHECK(v4_dstack_peek(&p->hw) == flat_peek(&p->model),
              "%s op %d: peek mismatch (live=%d)", tag, step, p->live);
    }
    if (p->live > 1) {
        /* Second element of a depth-N circular buffer whose top is at `top` is
         * N-1 further along the fill direction, i.e. (top-1) mod N. */
        v4_cell want = p->model.buf[(p->model.top + p->model.depth - 1u)
                                    % p->model.depth];
        CHECK(v4_dstack_peek2(&p->hw) == want,
              "%s op %d: peek2 got %lld want %lld (live=%d)", tag, step,
              (long long)v4_dstack_peek2(&p->hw), (long long)want, p->live);
    }
}

static void test_exhaustive_sequences(void)
{
    /* Every push/pop sequence up to length 10 -- 2046 of them -- driven through
     * both models.  Exhaustive over short sequences rather than random, because
     * a wrap-direction bug shows up in a handful of specific short patterns
     * (notably push x N+1 then pop x N, which is the only sequence that ever
     * overwrites the oldest entry) and a random driver finds those only by
     * luck.  This finds all of them, every run, deterministically. */
    enum { MAXLEN = 10 };
    int total = 0;
    for (int len = 1; len <= MAXLEN; len++) total += 1 << len;

    for (int len = 1; len <= MAXLEN; len++) {
        for (int bits = 0; bits < (1 << len); bits++) {
            pair_t p;
            pair_reset(&p);
            for (int i = 0; i < len; i++) {
                int push = (bits >> i) & 1;
                /* Distinct, non-zero values so a slot mix-up is visible. */
                v4_cell v = (v4_cell)(100 + i * 7);
                pair_step(&p, push, v, i, "exhaustive");
            }
        }
    }
    printf("  exhaustive: %d sequences of length <= %d\n", total, MAXLEN);
}

static void test_differential_random(void)
{
    /* Long random walk across the wrap points, both widths, both stacks. */
    enum { OPS = 200000 };
    pair_t dp;
    v4_rstack r;
    flat_t rf;
    int rlive;

    pair_reset(&dp);
    v4_rstack_reset(&r);
    flat_reset(&rf, V4_RET_DEPTH);
    rlive = 0;

    for (int i = 0; i < OPS; i++) {
        uint64_t bits = rng_next();

        pair_step(&dp, (int)(bits & 1u), (v4_cell)bits, i, "differential data");

        if (bits & 2u) {
            v4_cell v = (v4_cell)(bits >> 8);
            v4_rstack_push(&r, v);
            flat_push(&rf, v);
            if (rlive < V4_RET_DEPTH) rlive++;
        } else {
            v4_cell got = v4_rstack_pop(&r);
            v4_cell exp = flat_pop(&rf);
            if (rlive > 0) {
                CHECK(got == exp,
                      "differential ret op %d: pop got %lld want %lld (live=%d)",
                      i, (long long)got, (long long)exp, rlive);
                rlive--;
            }
        }
        if (rlive > 0) {
            CHECK(v4_rstack_peek(&r) == flat_peek(&rf),
                  "differential ret op %d: peek mismatch (live=%d)", i, rlive);
        }
    }
    printf("  differential: %d ops\n", OPS);
}

/* D-16: each stack counts what it holds.  The count stops at the stack's size
 * and at zero; push and pop go on doing what they always did there, and it is
 * the executor that turns those two cases into faults (test_exec.c). */
static void test_depth_count(void)
{
    v4_dstack d;
    v4_rstack r;
    unsigned i;

    v4_dstack_reset(&d);
    v4_rstack_reset(&r);
    CHECK(d.depth == 0 && r.depth == 0, "reset stacks hold nothing");
    for (i = 1; i <= V4_DATA_DEPTH + 3u; i++) {
        v4_dstack_push(&d, (v4_cell)i);
        CHECK(d.depth == (i < V4_DATA_DEPTH ? i : (unsigned)V4_DATA_DEPTH), "data depth after %u pushes is %u", i, d.depth);
    }
    for (i = V4_DATA_DEPTH; i-- > 0; ) {
        (void)v4_dstack_pop(&d);
        CHECK(d.depth == i, "data depth counts down to %u", i);
    }
    (void)v4_dstack_pop(&d);
    CHECK(d.depth == 0, "and stays at zero");
    for (i = 1; i <= V4_RET_DEPTH + 3u; i++) {
        v4_rstack_push(&r, (v4_cell)i);
        CHECK(r.depth == (i < V4_RET_DEPTH ? i : (unsigned)V4_RET_DEPTH), "return depth after %u pushes is %u", i, r.depth);
    }
    for (i = V4_RET_DEPTH; i-- > 0; ) {
        (void)v4_rstack_pop(&r);
        CHECK(r.depth == i, "return depth counts down to %u", i);
    }
    (void)v4_rstack_pop(&r);
    CHECK(r.depth == 0, "and stays at zero");

    v4_dstack_push(&d, 5); v4_dstack_push(&d, 6); v4_rstack_push(&r, 7);
    v4_dstack_clear(&d); v4_rstack_clear(&r);
    CHECK(d.depth == 0 && r.depth == 0, "clear empties a stack");
    CHECK(d.t == 6 && d.s == 5 && r.r == 7, "and changes nothing else");
    CHECK(v4_dstack_guards_intact(&d) && v4_rstack_guards_intact(&r), "guards intact");
}

int main(void)
{
    printf("v4 stack tests: V4_CELL_BITS=%d, data depth %d, return depth %d\n",
           V4_CELL_BITS, V4_DATA_DEPTH, V4_RET_DEPTH);

    test_reset_is_zero();
    test_exact_depth_data();
    test_overflow_overwrites_oldest();
    test_exact_depth_return();
    test_underflow_wraps_rather_than_trapping();
    test_exhaustive_sequences();
    test_differential_random();
    test_depth_count();

    printf("  %d checks, %d failures\n", checks, failures);
    return failures ? 1 : 0;
}