#include <stdlib.h>
#include <stdio.h>
#include <stdarg.h>
#include <unistd.h>
#include <string.h>
#include <errno.h>
#include <assert.h>
#include <sysexits.h>

#include "dcpu16.h"

/*
 *  emulates the DCPU16 system from http://0x10c.com/doc/dcpu-16.txt
 *  
 *  I couldn't remember ever implementing an emulator before, so this
 *  happened.  As such, consider this a toy in progress.
 *  There are likely many improvable aspects.
 *  
 *  Justin Wind <justin.wind@gmail.com>
 *    2012 04 05 - implementation started
 *    2012 04 06 - first functionality achieved
 *    2012 04 09 - minor cleanups
 *    2012 04 10 - moved cli to separate module
 *
 *  TODO
 *    drop checks for assigning to literals -- it won't affect anything anyhow
 *    debug short literal decoding
 */

static const char * const src_id_ = "$Id$";

#define WORD DCPU16_WORD

static const char regnames_[] = "ABCXYZIJ";


static unsigned int trace_mode_ = 0; /* spew overly verbose internals */

#define WARN(...) do { if (warn_cb_) warn_cb_(__VA_ARGS__); } while (0)
static inline void warn_(char *fmt, ...) __attribute__((format(printf, 1, 2)));
static inline
void warn_(char *fmt, ...) {
    va_list ap;

    fprintf(stderr, "[warning] ");
    va_start(ap, fmt);
    vfprintf(stderr, fmt, ap);
    va_end(ap);
    fprintf(stderr, "\n");
    fflush(stderr);
}
static void (*warn_cb_)(char *fmt, ...) = warn_;
void dcpu16_warn_cb_set(void (*fn)(char *fmt, ...)) {
    warn_cb_ = fn;
}

#define TRACE(...) do { if (trace_cb_ && trace_mode_) trace_cb_(__VA_ARGS__); } while (0)
static inline void trace_(char *fmt, ...) __attribute__((format(printf, 1, 2)));
static inline
void trace_(char *fmt, ...) {
    va_list ap;

    fprintf(stdout, "[debug] ");
    va_start(ap, fmt);
    vfprintf(stdout, fmt, ap);
    va_end(ap);
    fprintf(stdout, "\n");
    fflush(stdout);
}
static void (*trace_cb_)(char *fmt, ...) = trace_;
void dcpu16_trace_cb_set(void (*fn)(char *fmt, ...)) {
    trace_cb_ = fn;
}


/*  value_decode_
 * sets *v to be the destination of the value
 * workv is buffer to use to accumulate literal value before use
 * returns true if destination points to literal (id est *v should ignore writes)
 */
static unsigned int value_decode(struct dcpu16 *d, WORD value, WORD *work_v, WORD **v) {
    WORD nextword;
    unsigned int retval = 0;

    assert(value <= 0x3f);

    /* does this value indicate a literal */
    if (value >= 0x1f)
        retval = 1;

    /* if we're skipping this instruction, just advance the pc if needed */ 
    if (d->skip_) {
        TRACE(">>      SKIP decode");
        if (value == 0x1e || value == 0x1f)
            d->pc++;
        return retval;
    }

    if (value <= 0x07) { /* register */
        *v = d->reg + value;
        TRACE(">>     %c (0x%04x)",
              regnames_[value],
              **v);

    } else if (value <= 0x0f) { /* [register] */
        *v = &(d->ram[ d->reg[(value & 0x07)] ]);
        TRACE(">>     [%c] [0x%04x] (0x%04x)",
              regnames_[value&0x07],
              d->reg[value&0x07],
              **v);

    } else if (value <= 0x17) { /* [next word + register] */
        nextword = d->ram[ d->pc++ ];
        d->cycle++;
        *v = &(d->ram[ nextword + d->reg[(value & 0x07)] ]);
        TRACE(">>     [nextword + %c] [0x%04x + 0x%04x] (0x%04x)",
              regnames_[(value & 0x07)],
              nextword,
              d->reg[(value & 0x07)],
              **v);

    } else switch (value) {
        case 0x18: /* POP / [sp++] */
        *v = &(d->ram[ d->sp++ ]);
        TRACE(">>     POP [0x%04x] (0x%04x)",
              d->sp - 1,
              **v);
        break;

        case 0x19: /* PEEK / [sp] */
        *v = &(d->ram[ d->sp ]);
        TRACE(">>     PEEK [0x%04x] (0x%04x)",
              d->sp,
              **v);
        break;

        case 0x1a: /* PUSH / [--sp] */
        *v = &(d->ram[ --d->sp ]);
        TRACE(">>     PUSH [0x%04x] (0x%04x)",
              d->sp + 1,
              **v);
        break;

        case 0x1b: /* SP */
        *v = &(d->sp);
        TRACE(">>     SP (0x%04x)",
              **v);
        break;

        case 0x1c: /* PC */
        *v = &(d->pc);
        TRACE(">>     PC (0x%04x)", **v);
        break;

        case 0x1d: /* O */
        *v = &(d->o);
        TRACE(">>     O (0x%04x)", **v);
        break;

        case 0x1e: /* [next word] / [[pc++]] */
        nextword = d->ram[ d->pc++ ];
        d->cycle++;
        *v = &(d->ram[ nextword ]);
        TRACE(">>     [nextword] [0x%04x] (0x%04x)",
              nextword,
              **v);
        break;

        case 0x1f: /* next word (literal) / [pc++] */
        nextword = d->ram[ d->pc++ ];
        d->cycle++;
        *work_v = nextword;
        *v = work_v;
        TRACE(">>     nextword (0x%04x)", **v);
        break;

        default: /* 0x20-0x3f: literal values 0x00-0x1f */
        *work_v = value & 0x1f;
        *v = work_v;
        TRACE(">>     literal (0x%04x)", **v);
    }

    return retval;
}

#define OPCODE_BASIC_BITS (4)
#define OPCODE_BASIC_SHIFT (0)

#define OPCODE_NBI_BITS (6)
#define OPCODE_NBI_SHIFT (4)

#define OPCODE_FUTURE_BITS (16)
#define OPCODE_FUTURE_SHIFT (10)

#define OPCODE_NAME_LEN 16
struct opcode_entry {
    unsigned short value;
    char name[OPCODE_NAME_LEN];
    void (*impl)(struct dcpu16 *, WORD, WORD);
};

/* messy boilerplate for opcode handlers */

#define OP_IMPL(x) static void op_##x(struct dcpu16 *d, WORD val_a, WORD val_b)

#define OP_NBI_ (void)val_b, (void)b
#define OP_BASIC_ (void)value_decode(d, val_b, &d->reg_work_[0], &b)
#define OP_TYPE(op_type) WORD *a, *b;\
    unsigned int lit_a;\
    do {\
        lit_a = value_decode(d, val_a, &d->reg_work_[0], &a);\
        op_type;\
        if (d->skip_) {\
            TRACE("++ SKIPPED");\
            d->skip_ = 0;\
            return;\
        }\
    } while (0)
#define OP_BASIC(x) OP_TYPE(OP_BASIC_)
#define OP_NBI(x) OP_TYPE(OP_NBI_)


/* extended opcodes */

/*
    N.B. this next function currently decodes values -- id est, it is
    an opcode processing terminus; however, if 'future instruction set
    expansion' happens, this will probably need to behave more like
    the OP_IMPL(_nbi_) function which invoked it, if those instructions
    have zero or differently styled operands.
*/
OP_IMPL(nbi__reserved_) {
    OP_NBI(nbi__reserved_);
    /* reserved for future expansion */

    WORD future_opcode = (d->ram[d->pc] >> OPCODE_FUTURE_SHIFT);
    WARN("reserved future opcode 0x%04x invoked", future_opcode);
}

OP_IMPL(nbi_jsr) {
    OP_NBI(nbi_jsr);
    /* pushes the address of the next instruction to the stack, then sets PC to a */

    d->ram[ --d->sp ] = d->pc;
    d->pc = *a;

    d->cycle += 2;
}

OP_IMPL(nbi__reserved2_) {
    OP_NBI(nbi__reserved2_);
    /* reserved */

    WARN("reserved nbi opcode invoked");
}

static const struct opcode_entry opcode_nbi_entries[] = {
    {0x0, "(reserved)", op_nbi__reserved_},
    {0x1, "JSR", op_nbi_jsr},
    {0x2, "(reserved)", op_nbi__reserved2_},
    {0x0, "", NULL}
};
#define OPCODE_NBI_MAX (((sizeof(opcode_nbi_entries)) / (sizeof(struct opcode_entry))) - 1)


/* basic opcodes */

/*
    N.B. the following function does not decode values, as the nbi
    instructions only have one operand.
*/
OP_IMPL(_nbi_) {
    /* non-basic instruction */

    /* don't do normal value decoding here */

    WORD nbi_opcode = val_a;
    const struct opcode_entry *e = opcode_nbi_entries;

    e = opcode_nbi_entries + ( (nbi_opcode < OPCODE_NBI_MAX) ? nbi_opcode : (OPCODE_NBI_MAX - 1) );

    assert(e->impl != NULL);

    TRACE(">> %s 0x%04x", e->name, val_b);
    e->impl(d, val_b, 0);
}

OP_IMPL(set) {
    OP_BASIC(set);
    /* sets a to b */

    /* only set non-literal target */
    if (val_a < 0x1f) {
        *a = *b;
    }

    d->cycle += 1;
}

OP_IMPL(add) {
    OP_BASIC(add);
    /* sets a to a+b, sets O to 0x0001 if there's an overflow, 0x0 otherwise */
    unsigned int acc = *a + *b;

    if (val_a < 0x1f) {
        *a = acc;
    }
    d->o = acc >> 16;

    d->cycle += 2;
}

OP_IMPL(sub) {
    OP_BASIC(sub);
    /* sets a to a-b, sets O to 0xffff if there's an underflow, 0x0 otherwise */
    unsigned int acc = *a - *b;

    if (val_a < 0x1f) {
        *a = acc;
    }
    d->o = acc >> 16;

    d->cycle += 2;
}

OP_IMPL(mul) {
    OP_BASIC(mul);
    /* sets a to a*b, sets O to ((a*b)>>16)&0xffff */
    unsigned int acc = *a * *b;

    if (val_a < 0x1f) {
        *a = acc;
    }
    d->o = acc >> 16;
    d->cycle += 2;
}

OP_IMPL(div) {
    OP_BASIC(div);
    /* sets a to a/b, sets O to ((a<<16)/b)&0xffff. if b==0, sets a and O to 0 instead. */

    if (*b == 0) {
        if (val_a < 0x1f) {
            *a = 0;
        }
        d->o = 0;
    } else {
        unsigned int acc = *a / *b;

        if (val_a < 0x1f) {
            *a = acc;
        }

        acc = (*a << 16) / *b;
        d->o = acc;
    }

    d->cycle += 3;
}

OP_IMPL(mod) {
    OP_BASIC(mod);
    /* sets a to a%b. if b==0, sets a to 0 instead. */

    if (*b == 0) {
        if (val_a < 0x1f) {
            *a = 0;
        }
    } else {
        if (val_a < 0x1f) {
            *a = *a % *b;
        }
    }

    d->cycle += 3;
}

OP_IMPL(shl) {
    OP_BASIC(shl);
    /* sets a to a<<b, sets O to ((a<<b)>>16)&0xffff */
    unsigned int acc = *a << *b;

    if (val_a < 0x1f) {
        *a = acc;
    }
    d->o = acc >> 16;

    d->cycle += 2;
}

OP_IMPL(shr) {
    OP_BASIC(shr);
    /* sets a to a>>b, sets O to ((a<<16)>>b)&0xffff */
    unsigned int acc = *a >> *b;

    if (val_a < 0x1f) {
        *a = acc;
    }
    d->o = (*a << 16) >> *b;

    d->cycle += 2;
}

OP_IMPL(and) {
    OP_BASIC(and);
    /* sets a to a&b */

    if (val_a < 0x1f) {
        *a = *a & *b;
    }

    d->cycle += 1;
}

OP_IMPL(bor) {
    OP_BASIC(bor);
    /* sets a to a|b */

    if (val_a < 0x1f) {
        *a = *a | *b;
    }

    d->cycle += 1;
}

OP_IMPL(xor) {
    OP_BASIC(xor);
    /* sets a to a^b */

    if (val_a < 0x1f) {
        *a = *a ^ *b;
    }

    d->cycle += 1;
}

OP_IMPL(ife) {
    OP_BASIC(ife);
    /* performs next instruction only if a==b */

    if (*a == *b) {
        /* */
    } else {
        d->skip_ = 1;
        d->cycle++;
    }

    d->cycle += 2;
}

OP_IMPL(ifn) {
    OP_BASIC(ifn);
    /* performs next instruction only if a!=b */

    if (*a != *b) {
        /* */
    } else {
        d->skip_ = 1;
        d->cycle++;
    }

    d->cycle += 2;
}

OP_IMPL(ifg) {
    OP_BASIC(ifg);
    /* performs next instruction only if a>b */

    if (*a > *b) {
        /* */
    } else {
        d->skip_ = 1;
        d->cycle++;
    }

    d->cycle += 2;
}

OP_IMPL(ifb) {
    OP_BASIC(ifb);
    /* performs next instruction only if (a&b)!=0 */

    if ((*a & *b) != 0) {
        /* */
    } else {
        d->skip_ = 1;
        d->cycle++;
    }

    d->cycle += 2;
}

static const struct opcode_entry opcode_basic_entries[] = {
    {0x0, "(nbi)", op__nbi_},
    {0x1, "SET", op_set },
    {0x2, "ADD", op_add },
    {0x3, "SUB", op_sub },
    {0x4, "MUL", op_mul },
    {0x5, "DIV", op_div },
    {0x6, "MOD", op_mod },
    {0x7, "SHL", op_shl },
    {0x8, "SHR", op_shr },
    {0x9, "AND", op_and },
    {0xa, "BOR", op_bor },
    {0xb, "XOR", op_xor },
    {0xc, "IFE", op_ife },
    {0xd, "IFN", op_ifn },
    {0xe, "IFG", op_ifg },
    {0xf, "IFB", op_ifb },
    {0x0, "", NULL }
};

void dump_value(WORD value, WORD nextword) {
    if (value < 0x07) {
        printf(" %c", regnames_[value]);
    } else if (value < 0x0f) {
        printf(" [%c]", regnames_[value & 0x07]);
    } else if (value < 0x17) {
        printf(" [0x%04x + %c]", nextword, regnames_[value & 0x07]);
    } else switch (value) {
        case 0x18: printf(" POP"); break;
        case 0x19: printf(" PEEK"); break;
        case 0x1a: printf(" PUSH"); break;
        case 0x1b: printf(" SP"); break;
        case 0x1c: printf(" PC"); break;
        case 0x1d: printf(" O"); break;
        case 0x1e: printf(" [0x%04x]", nextword); break;
        case 0x1f: printf(" 0x%04x", nextword); break;
        default:   printf(" 0x%02x", value - 0x20);
    }
}

void dcpu16_disassemble_print(struct dcpu16 *d, WORD addr) {
    WORD opcode, a, b;
    unsigned int instr_len = 1;
    const struct opcode_entry *e;

    opcode = (d->ram[addr] >> OPCODE_BASIC_SHIFT) & ((1 << OPCODE_BASIC_BITS) - 1);
    a = (d->ram[addr] >> (OPCODE_BASIC_SHIFT + OPCODE_BASIC_BITS)) & ((1 << 6) - 1);
    b = (d->ram[addr] >> (OPCODE_BASIC_SHIFT + OPCODE_BASIC_BITS + 6)) & ((1 << 6) - 1); 

    assert(opcode <= 0x0f);
    assert(a <= 0x3f);
    assert(b <= 0x3f);

    printf("   next instr 0x%04x: %04x", addr, d->ram[addr]);

    if (opcode != 0)
    {
        if (a == 0x1e || a == 0x1f) {
            printf(" %04x", d->ram[addr + instr_len]);
            instr_len++;
        }
    }
    if (b == 0x1e || b == 0x1f) {
        printf(" %04x", d->ram[addr + instr_len]);
        instr_len++;
    }

    if (opcode)
        e = opcode_basic_entries + opcode;
    else
        e = opcode_nbi_entries + ( (a < OPCODE_NBI_MAX) ? a : (OPCODE_NBI_MAX - 1) );

    printf("%s%s ; %s",
           instr_len < 3 ? "     " : "",
           instr_len < 2 ? "     " : "",
           e->name);
    if (opcode != 0) {
        dump_value(a, d->ram[addr + 1]);
        if (a == 0x1e || a == 0x1f)
            addr++;
        printf(",");
    }

    dump_value(b, d->ram[addr + 1]);

    printf("\n");
}

void dcpu16_step(struct dcpu16 *d) {
    WORD opcode;
    WORD val_a, val_b;
    const struct opcode_entry *e;

    /* decode opcode and invoke */

    opcode = (d->ram[ d->pc ] >> OPCODE_BASIC_SHIFT) & ((1 << OPCODE_BASIC_BITS) - 1);
    val_a = (d->ram[d->pc] >> (OPCODE_BASIC_SHIFT + OPCODE_BASIC_BITS)) & ((1 << 6) - 1);
    val_b = (d->ram[d->pc] >> (OPCODE_BASIC_SHIFT + OPCODE_BASIC_BITS + 6)) & ((1 << 6) - 1);

    d->pc++;

    for (e = opcode_basic_entries; e->impl; e++) {
        if (e->value == opcode) {
            TRACE(">> %s 0x%04x, 0x%04x", e->name, val_a, val_b);
            e->impl(d, val_a, val_b);
            break;
        }
    }
}

void dcpu16_state_print(struct dcpu16 *d) {
    unsigned int i;

    printf("---- cycle:0x%08llx %2s:0x%04x %2s:0x%04x %2s:0x%04x\n",
           d->cycle,
           "PC", d->pc,
           "SP", d->sp,
           "O", d->o);
    printf(" ");
    for (i = 0; i < 8; i++)
        printf("  %c:0x%04x", regnames_[i], d->reg[i]);
    printf("\n");
}

void dcpu16_dump_ram(struct dcpu16 *d, WORD start, WORD stop) {
    unsigned int i, j;
    const unsigned int n = 8; /* words per line */

    for (i = start, j = 0; i <= stop; i++, j++) {
        if (j % n == 0)
            printf("0x%04x:\t", i);
        printf(" %04x%s", d->ram[i], (j % n) == (n - 1) ? "\n" : "");
    }
}

/*  dcpu16_new
 *  allocate a new dcpu16 instance
 */
struct dcpu16 *dcpu16_new(void) {
    struct dcpu16 *vm;

    vm = calloc(1, sizeof *vm);
    if (vm == NULL)
        WARN("%s: %s(%zu): %s", __func__, "calloc", strerror(errno));

    return vm;
}

/*  dcpu16_delete
 *  release a dcpu16 instance
 */
void dcpu16_delete(struct dcpu16 **vm) {
    free(*vm);
    *vm = NULL;
}