/* test_exec.c -- every opcode of DECOMPOSITION.md section 1.3, and the slot
 * sequencing of section 1.2.
 *
 * Each expected value below is worked out from the table in section 1.3, not
 * from the executor.  Words are packed by hand with v4_iword_assemble so that
 * this file does not depend on asm.c, which has its own test.
 */
#include "v4/exec.h"
#include <stdio.h>

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

#define NOP  V4_OP_NOP
#define MINC ((v4_cell)V4_MSB)              /* most negative cell */
#define MAXC ((v4_cell)(V4_MSB - 1u))       /* most positive cell */

static v4_node       n;
static v4_exec_state es;
static v4_heat       h;

static void fresh(void)
{
    v4_node_reset(&n);
    v4_exec_reset(&es);
    v4_heat_reset(&h);
}

static void dpush(v4_cell x) { v4_dstack_push(&n.ds, x); }
static void rpush(v4_cell x) { v4_rstack_push(&n.rs, x); }

/* A word of six plain opcodes. */
static v4_cell word6(unsigned a, unsigned b, unsigned c,
                     unsigned d, unsigned e, unsigned f)
{
    unsigned op[V4_SLOT_COUNT];
    op[0] = a; op[1] = b; op[2] = c; op[3] = d; op[4] = e; op[5] = f;
    return (v4_cell)v4_iword_assemble(op, 0);
}

/* A word with `npre` nops, then a branch in the next slot, then its target. */
static v4_cell bword(unsigned npre, unsigned op, v4_iword target)
{
    v4_iword w = 0;
    for (unsigned k = 0; k < npre; k++) w |= (v4_iword)NOP << V4_SLOT_LOW_BIT(k);
    w |= (v4_iword)op << V4_SLOT_LOW_BIT(npre);
    w |= target & v4_iword_slot_mask(npre);
    return (v4_cell)w;
}

/* Run one opcode, padded with nops, on the current stacks.  Returns the count
 * of instructions retired. */
static unsigned run1(unsigned op)
{
    n.mem[0] = word6(op, NOP, NOP, NOP, NOP, NOP);
    n.p = 0;
    return v4_exec_step_word(&n, &es, &h);
}

static void test_stack_ops(void)
{
    fresh(); dpush(1); dpush(2); run1(V4_OP_DUP);
    CHECK(n.ds.t == 2 && n.ds.s == 2, "dup");
    (void)v4_dstack_pop(&n.ds);
    CHECK(n.ds.t == 2 && n.ds.s == 1, "dup keeps what was below");

    fresh(); dpush(1); dpush(2); run1(V4_OP_OVER);
    CHECK(n.ds.t == 1 && n.ds.s == 2, "over");
    (void)v4_dstack_pop(&n.ds);
    CHECK(n.ds.t == 2 && n.ds.s == 1, "over keeps what was below");

    fresh(); dpush(1); dpush(2); dpush(3); run1(V4_OP_DROP);
    CHECK(n.ds.t == 2 && n.ds.s == 1, "drop");

    fresh(); dpush(1); dpush(2); CHECK(run1(NOP) == 6, "nop word retires six");
    CHECK(n.ds.t == 2 && n.ds.s == 1 && n.p == 1, "nop changes nothing but P");
}

static void test_alu(void)
{
    fresh(); dpush(9); dpush(3); dpush(4); run1(V4_OP_ADD);
    CHECK(n.ds.t == 7 && n.ds.s == 9, "3 + 4, and 9 moves up to S");
    fresh(); dpush(MAXC); dpush(1); run1(V4_OP_ADD);
    CHECK(n.ds.t == MINC, "+ wraps");

    fresh(); dpush(9); dpush(12); dpush(10); run1(V4_OP_AND);
    CHECK(n.ds.t == 8 && n.ds.s == 9, "and");
    fresh(); dpush(9); dpush(12); dpush(10); run1(V4_OP_XOR);
    CHECK(n.ds.t == 6 && n.ds.s == 9, "xor");

    fresh(); dpush(5); dpush(0); run1(V4_OP_INV);
    CHECK(n.ds.t == -1 && n.ds.s == 5, "inv 0");
    fresh(); dpush(MAXC); run1(V4_OP_INV);
    CHECK(n.ds.t == MINC, "inv max");

    fresh(); dpush(3); run1(V4_OP_TWO_STAR);     CHECK(n.ds.t == 6, "2* 3");
    fresh(); dpush(-1); run1(V4_OP_TWO_STAR);    CHECK(n.ds.t == -2, "2* -1");
    fresh(); dpush(MINC); run1(V4_OP_TWO_STAR);  CHECK(n.ds.t == 0, "2* drops the top bit");

    fresh(); dpush(6); run1(V4_OP_TWO_SLASH);    CHECK(n.ds.t == 3, "2/ 6");
    fresh(); dpush(-1); run1(V4_OP_TWO_SLASH);   CHECK(n.ds.t == -1, "2/ -1 is -1, not 0");
    fresh(); dpush(-7); run1(V4_OP_TWO_SLASH);   CHECK(n.ds.t == -4, "2/ -7 floors");
    fresh(); dpush(MINC); run1(V4_OP_TWO_SLASH);
    CHECK((v4_ucell)n.ds.t == (V4_MSB | (V4_MSB >> 1)), "2/ keeps the sign bit");
}

static void test_mul_step(void)
{
    /* A even: no add. */
    fresh(); dpush(5); dpush(6); n.a = 4; run1(V4_OP_MUL_STEP);
    CHECK(n.ds.t == 3 && n.ds.s == 5 && n.a == 2, "+* A even");

    /* A odd: T+S = 11, shifted to 5, the 1 shifted out enters the top of A. */
    fresh(); dpush(5); dpush(6); n.a = 5; run1(V4_OP_MUL_STEP);
    CHECK(n.ds.t == 5 && n.ds.s == 5, "+* A odd, T");
    CHECK((v4_ucell)n.a == (V4_MSB | 2u), "+* A odd, A");

    fresh(); dpush(0); dpush(3); n.a = 0; run1(V4_OP_MUL_STEP);
    CHECK(n.ds.t == 1 && (v4_ucell)n.a == V4_MSB, "+* T bit 0 enters A");

    fresh(); dpush(0); dpush(-2); n.a = 0; run1(V4_OP_MUL_STEP);
    CHECK(n.ds.t == -1 && n.a == 0, "+* keeps T's sign bit");

    /* D-3: all-ones + 1 carries out of T and the carry is gone. */
    fresh(); dpush(1); dpush(-1); n.a = 1; run1(V4_OP_MUL_STEP);
    CHECK(n.ds.t == 0 && n.a == 0, "+* does not keep the carry (D-3)");
}

static void test_registers_and_rstack(void)
{
    fresh(); dpush(4); dpush(5); run1(V4_OP_PUSH);
    CHECK(n.rs.r == 5 && n.ds.t == 4, "push");
    run1(V4_OP_RPOP);
    CHECK(n.ds.t == 5 && n.ds.s == 4, "pop");

    fresh(); dpush(4); dpush(77); run1(V4_OP_BANG_A);
    CHECK(n.a == 77 && n.ds.t == 4, "a!");
    run1(V4_OP_PUSH_A);
    CHECK(n.ds.t == 77 && n.ds.s == 4 && n.a == 77, "a");

    fresh(); dpush(4); dpush(88); run1(V4_OP_BANG_B);
    CHECK(n.b == 88 && n.ds.t == 4 && n.a == 0, "b!");
}

static void test_memory(void)
{
    fresh(); n.a = 100; n.b = 101; n.mem[100] = 11; n.mem[101] = 22;
    run1(V4_OP_FETCH_A);    CHECK(n.ds.t == 11 && n.a == 100, "@");
    run1(V4_OP_FETCH_B);    CHECK(n.ds.t == 22 && n.ds.s == 11 && n.b == 101, "@b");
    run1(V4_OP_FETCH_INC);  CHECK(n.ds.t == 11 && n.a == 101, "@+ first");
    run1(V4_OP_FETCH_INC);  CHECK(n.ds.t == 22 && n.a == 102, "@+ second");

    fresh(); n.a = 100; n.b = 200; dpush(1); dpush(5);
    run1(V4_OP_STORE_A);
    CHECK(n.mem[100] == 5 && n.ds.t == 1 && n.a == 100, "!");
    dpush(6); run1(V4_OP_STORE_INC);
    CHECK(n.mem[100] == 6 && n.ds.t == 1 && n.a == 101, "!+");
    dpush(7); run1(V4_OP_STORE_B);
    CHECK(n.mem[200] == 7 && n.ds.t == 1 && n.b == 200, "!b");

    /* @p takes the word after the instruction word and steps over it. */
    fresh(); n.mem[1] = (v4_cell)0x1122334455667788LL;
    run1(V4_OP_FETCH_P);
    CHECK(n.ds.t == n.mem[1] && n.p == 2, "@p is a full cell and advances P");

    fresh(); dpush(1); dpush(9); run1(V4_OP_STORE_P);
    CHECK(n.mem[1] == 9 && n.ds.t == 1 && n.p == 2, "!p");

    /* Two literals in one word are the two words after it, in order. */
    fresh(); n.mem[0] = word6(V4_OP_FETCH_P, V4_OP_FETCH_P, V4_OP_ADD, NOP, NOP, NOP);
    n.mem[1] = 30; n.mem[2] = 12; n.p = 0;
    (void)v4_exec_step_word(&n, &es, &h);
    CHECK(n.ds.t == 42 && n.p == 3, "two literals");
}

static void test_return_and_ex(void)
{
    fresh(); rpush(7); rpush(50);
    n.mem[0] = word6(V4_OP_SEMI, V4_OP_DUP, V4_OP_DUP, V4_OP_DUP, V4_OP_DUP, V4_OP_DUP);
    n.p = 0;
    CHECK(v4_exec_step_word(&n, &es, &h) == 1, "; ends the word");
    CHECK(n.p == 50 && n.rs.r == 7, "; returns and pops R");
    CHECK(h.op[V4_OP_DUP] == 0 && es.anticlock == 1, "nothing after ; retires");

    fresh(); rpush(7); rpush(50);
    n.mem[20] = word6(V4_OP_EX, V4_OP_DUP, NOP, NOP, NOP, NOP);
    n.p = 20;
    CHECK(v4_exec_step_word(&n, &es, &h) == 1, "ex ends the word");
    CHECK(n.p == 50 && n.rs.r == 21, "ex swaps P and R");
    (void)v4_rstack_pop(&n.rs);
    CHECK(n.rs.r == 7, "ex does not change the return stack depth");
}

static void test_branches(void)
{
    /* jump in each legal slot; the slots before it execute. */
    for (unsigned slot = 0; slot < 4; slot++) {
        fresh(); n.mem[10] = bword(slot, V4_OP_JUMP, 300); n.p = 10;
        CHECK(v4_exec_step_word(&n, &es, &h) == slot + 1, "jump slot %u count", slot);
        CHECK(n.p == 300, "jump slot %u", slot);
        CHECK(h.op[NOP] == slot && h.op[V4_OP_JUMP] == 1, "jump slot %u heat", slot);
    }

    fresh(); rpush(7); n.mem[10] = bword(0, V4_OP_CALL, 200); n.p = 10;
    (void)v4_exec_step_word(&n, &es, &h);
    CHECK(n.p == 200 && n.rs.r == 11, "call pushes the return address");
    CHECK(v4_heat_call_get(&h, 200) == 1 && h.op[V4_OP_CALL] == 1, "call heats its target");
    (void)v4_rstack_pop(&n.rs);
    CHECK(n.rs.r == 7, "call pushes exactly one");

    fresh(); rpush(99); rpush(2); n.mem[5] = bword(0, V4_OP_NEXT, 40); n.p = 5;
    (void)v4_exec_step_word(&n, &es, &h);
    CHECK(n.p == 40 && n.rs.r == 1, "next, R nonzero");
    fresh(); rpush(99); rpush(0); n.mem[5] = bword(0, V4_OP_NEXT, 40); n.p = 5;
    (void)v4_exec_step_word(&n, &es, &h);
    CHECK(n.p == 6 && n.rs.r == 99, "next, R zero");

    fresh(); dpush(8); dpush(0); n.mem[5] = bword(0, V4_OP_IF, 40); n.p = 5;
    (void)v4_exec_step_word(&n, &es, &h);
    CHECK(n.p == 40 && n.ds.t == 0 && n.ds.s == 8, "if, T zero: taken, T kept");
    /* The target bits 0x0FFFF would decode as opcodes if they were executed. */
    fresh(); dpush(8); dpush(5); n.mem[5] = bword(0, V4_OP_IF, 0x0FFFF); n.p = 5;
    CHECK(v4_exec_step_word(&n, &es, &h) == 1, "if not taken still ends the word");
    CHECK(n.p == 6 && n.ds.t == 5 && n.ds.s == 8, "if, T nonzero: not taken, T kept");

    fresh(); dpush(8); dpush(5); n.mem[5] = bword(0, V4_OP_MINUS_IF, 40); n.p = 5;
    (void)v4_exec_step_word(&n, &es, &h);
    CHECK(n.p == 40 && n.ds.t == 5 && n.ds.s == 8, "-if, T positive: taken");
    fresh(); dpush(0); n.mem[5] = bword(0, V4_OP_MINUS_IF, 40); n.p = 5;
    (void)v4_exec_step_word(&n, &es, &h);
    CHECK(n.p == 40, "-if, T zero: taken");
    fresh(); dpush(8); dpush(-1); n.mem[5] = bword(0, V4_OP_MINUS_IF, 40); n.p = 5;
    (void)v4_exec_step_word(&n, &es, &h);
    CHECK(n.p == 6 && n.ds.t == -1 && n.ds.s == 8, "-if, T negative: not taken");
}

static void test_unext(void)
{
    /* R = 3: the body runs four times, then the rest of the word runs once. */
    fresh(); dpush(1); rpush(99); rpush(3);
    n.mem[0] = word6(V4_OP_TWO_STAR, V4_OP_UNEXT, NOP, NOP, NOP, NOP); n.p = 0;
    CHECK(v4_exec_step_word(&n, &es, &h) == 12, "unext retire count");
    CHECK(n.ds.t == 16, "unext body ran R+1 times");
    CHECK(n.rs.r == 99, "unext pops R when done");
    CHECK(n.p == 1, "unext leaves P at the next word");
    CHECK(h.op[V4_OP_TWO_STAR] == 4 && h.op[V4_OP_UNEXT] == 4 && h.op[NOP] == 4,
          "unext heat");
    CHECK(es.anticlock == 12, "unext anticlock");

    fresh(); dpush(1); rpush(99); rpush(0);
    n.mem[0] = word6(V4_OP_TWO_STAR, V4_OP_UNEXT, NOP, NOP, NOP, NOP); n.p = 0;
    CHECK(v4_exec_step_word(&n, &es, &h) == 6, "unext with R zero runs once");
    CHECK(n.ds.t == 2 && n.rs.r == 99, "unext with R zero");

    /* The word is not refetched and P is not rewound, so @p walks forward. */
    fresh(); rpush(99); rpush(2); n.a = 100;
    n.mem[0] = word6(V4_OP_FETCH_P, V4_OP_STORE_INC, V4_OP_UNEXT, NOP, NOP, NOP);
    n.mem[1] = 11; n.mem[2] = 22; n.mem[3] = 33; n.p = 0;
    (void)v4_exec_step_word(&n, &es, &h);
    CHECK(n.mem[100] == 11 && n.mem[101] == 22 && n.mem[102] == 33,
          "@p in a unext loop copies successive words");
    CHECK(n.a == 103 && n.p == 4 && n.rs.r == 99, "@p in a unext loop, registers");
}

static void test_retirement(void)
{
    fresh();
    for (unsigned op = 0; op < V4_OPCODE_COUNT; op++) v4_exec_op(&n, &es, &h, op, 0, 0);
    for (unsigned op = 0; op < V4_OPCODE_COUNT; op++)
        CHECK(h.op[op] == 1, "opcode %u retires once", op);
    CHECK(es.anticlock == V4_OPCODE_COUNT, "one anticlock tick per opcode");

    v4_exec_op(&n, &es, &h, V4_OPCODE_COUNT, 0, 0);
    CHECK(es.anticlock == V4_OPCODE_COUNT, "a non-opcode retires nothing");
}

#if V4_CELL_BITS == 64
static void test_iword_in_wide_cell(void)
{
    /* D-9: the instruction word is the low 32 bits of the cell.  The high half
     * is not decoded and does not reach a branch target. */
    v4_cell junk = (v4_cell)0xA5A5A5A500000000ULL;

    fresh(); dpush(3);
    n.mem[0] = junk | word6(V4_OP_DUP, V4_OP_ADD, NOP, NOP, NOP, NOP); n.p = 0;
    CHECK(v4_exec_step_word(&n, &es, &h) == 6, "wide cell: six slots");
    CHECK(n.ds.t == 6, "wide cell: decode ignores the high half");

    fresh(); n.mem[10] = junk | bword(0, V4_OP_JUMP, 300); n.p = 10;
    (void)v4_exec_step_word(&n, &es, &h);
    CHECK(n.p == 300, "wide cell: branch ignores the high half");
}
#endif

/* Many opcodes on arbitrary values.  The point is the sanitizer build: no
 * opcode may overflow a signed cell or shift out of range, whatever is on the
 * stacks.  Memory opcodes are left out because A, B and P would be arbitrary. */
static void test_soak(void)
{
    static const unsigned safe[] = {
        V4_OP_MUL_STEP, V4_OP_TWO_STAR, V4_OP_TWO_SLASH, V4_OP_INV, V4_OP_ADD,
        V4_OP_AND, V4_OP_XOR, V4_OP_DROP, V4_OP_DUP, V4_OP_RPOP, V4_OP_OVER,
        V4_OP_PUSH_A, V4_OP_NOP, V4_OP_PUSH, V4_OP_BANG_B, V4_OP_BANG_A
    };
    uint64_t seed = 0x9E3779B97F4A7C15ULL;
    const unsigned count = 200000;

    fresh();
    dpush(MINC); dpush(MAXC); dpush(-1); dpush(1);
    for (unsigned i = 0; i < count; i++) {
        seed = seed * 6364136223846793005ULL + 1442695040888963407ULL;
        if ((seed >> 60) == 0) dpush((v4_cell)(seed >> 7));
        v4_exec_op(&n, &es, &h, safe[(seed >> 33) % (sizeof safe / sizeof safe[0])], 0, 0);
    }
    CHECK(es.anticlock == count, "soak retired every opcode");
    CHECK(v4_node_guards_intact(&n), "soak left the guards intact");
}

int main(void)
{
    printf("v4 exec tests: V4_CELL_BITS=%d\n", V4_CELL_BITS);

    test_stack_ops();
    test_alu();
    test_mul_step();
    test_registers_and_rstack();
    test_memory();
    test_return_and_ex();
    test_branches();
    test_unext();
    test_retirement();
#if V4_CELL_BITS == 64
    test_iword_in_wide_cell();
#endif
    test_soak();

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