#include <stdint.h>
#include "opmplay.h"
#include "opmfile.h"

// -------------------------
// file I/O procedures
static uint32_t opmplay_mem_read(opmplay_io_t* io, void* dst, uint32_t size) {
    if ((io == NULL) || dst == NULL) return OPMPLAY_ERR_NULLPTR;
    if ((size + io->offset) > io->size) return OPMPLAY_ERR_IO;
    opmplay_memcpy(dst, (uint8_t*)io->buf + io->offset, io->size);
    io->offset += io->size;
    return size;
}
static uint32_t opmplay_mem_seek(opmplay_io_t* io, uint32_t pos) {
    if (io == NULL) return OPMPLAY_ERR_NULLPTR;
    if (pos > io->size) return OPMPLAY_ERR_IO;
    io->offset = pos;
    return 0;
}

#ifdef OPMPLAY_ENABLE_STDIO
static uint32_t opmplay_file_read(opmplay_io_t* io, void* dst, uint32_t size) {
    return fread(dst, 1, size, io->io);
}
static uint32_t opmplay_file_seek(opmplay_io_t* io, uint32_t pos) {
    return fseek(io->io, pos, SEEK_SET);
}
#endif

// -------------------------

int opmplay_init(opmplay_context_t* ctx, opl3_chip* chip) {
    if ((ctx == NULL) || (chip == NULL)) return OPMPLAY_ERR_NULLPTR;

    opmplay_memset(ctx, 0, sizeof(opmplay_context_t));
    ctx->chip = chip;

    return OPMPLAY_ERR_OK;
}

int opmplay_free(opmplay_context_t* ctx)
{
    for (int ch = 0; ch < ctx->header.channels+1; ch++) {
        if (ctx->channels[ch].stream.data != NULL) {
            opmplay_memfree(ctx->channels[ch].stream.data); ctx->channels[ch].stream.data = NULL;
        }
    }
    return OPMPLAY_ERR_OK;
}

// load file header
int opmplay_load_header(opmplay_context_t* ctx, opmplay_io_t* io) {
    if ((ctx == NULL) || (io == NULL)) return OPMPLAY_ERR_NULLPTR;

    // init file I/O handlers
    switch (io->type) {
#ifdef OPMPLAY_ENABLE_STDIO
    case OPMPLAY_IO_FILE:
        io->read = &opmplay_file_read;
        io->seek = &opmplay_file_seek;
        break;
#endif
    case OPMPLAY_IO_MEMORY:
        io->read = &opmplay_mem_read;
        io->seek = &opmplay_mem_seek;
        break;
    case OPMPLAY_IO_USER:
        // use user-provided I/O functions
        if ((io->read == NULL) || (io->seek == NULL)) return OPMPLAY_ERR_BAD_PARAMETER;
        break;
    default:
        return OPMPLAY_ERR_BAD_PARAMETER;
    }

    // read header
    if (io->read(io, &ctx->header, sizeof(ctx->header)) != sizeof(ctx->header)) return OPMPLAY_ERR_IO;

    // and validate it
    if ((opmplay_memcmp(ctx->header.magic, "OPM\x1A", sizeof(ctx->header.magic))) || (ctx->header.version.v != OPM_FORMAT_VERSION))
        return OPMPLAY_ERR_BAD_FILE_STRUCTURE;

    // done for now, waiting for opmplay_load_module :)
    return OPMPLAY_ERR_OK;
}

int opmplay_load_module(opmplay_context_t* ctx, opmplay_io_t* io) {
    if ((ctx == NULL) || (io == NULL)) return OPMPLAY_ERR_NULLPTR;
    uint32_t filepos = sizeof(opm_header_t);

    // allocate and copy stream data
    opm_header_stream_desc_t* streamdesc = (opm_header_stream_desc_t*)opmplay_alloc(sizeof(opm_header_stream_desc_t) * (ctx->header.channels + 1));
    if (streamdesc == NULL) return OPMPLAY_ERR_NULLPTR;
    if (io->seek(io, filepos)) return OPMPLAY_ERR_IO;
    if (io->read(io, streamdesc, sizeof(opm_header_stream_desc_t) * (ctx->header.channels + 1))
        != sizeof(opm_header_stream_desc_t) * (ctx->header.channels + 1))
        return OPMPLAY_ERR_IO;

    // allocate and copy channel streams
    for (int ch = 0; ch < ctx->header.channels + 1; ch++) {
        if (streamdesc[ch].size > 0) {
            ctx->channels[ch].stream.data = (uint8_t*)opmplay_alloc(sizeof(uint8_t*) * (streamdesc[ch].size));
            if (ctx->channels[ch].stream.data == NULL) return OPMPLAY_ERR_MEMALLOC;
            if (io->read(io, ctx->channels[ch].stream.data, streamdesc[ch].size) != streamdesc[ch].size) return OPMPLAY_ERR_IO;
            ctx->channels[ch].stream.delay = 1;
        }
        else ctx->channels[ch].stream.data = NULL;
    }

    // rewind to start
    opmplay_rewind(ctx);

    // done :)
    return OPMPLAY_ERR_OK;
}

void opmplay_pop_stack(opmplay_channel_context_t* chctx) {
    opmplay_channel_stack_t* st = chctx->stack + (--chctx->stack_pos);
    chctx->stream.ptr = st->ptr;
    chctx->stream.samples_to_play = st->frames_to_play;
}

void opmplay_push_stack(opmplay_channel_context_t* chctx) {
    opmplay_channel_stack_t* st = chctx->stack + chctx->stack_pos;
    st->ptr = chctx->stream.ptr;
    st->frames_to_play = chctx->stream.samples_to_play;
    chctx->stack_pos++;
}

int opmplay_loop(opmplay_context_t* ctx) {
    // channel streams
    for (int ch = 0; ch < ctx->header.channels + 1; ch++) {
        // init stack
        ctx->channels[ch].stack_pos = 0;
        ctx->channels[ch].stream.samples_to_play = -1;
        ctx->channels[ch].stream.ptr = ctx->channels[ch].stream.loop;
        ctx->channels[ch].stream.delay = ctx->channels[ch].stream.reload = 1;
    }

    return OPMPLAY_ERR_OK;
}

int opmplay_rewind(opmplay_context_t* ctx) {
    // set chip mode
    switch (ctx->header.flags & OPM_FLAG_CHIP_TYPE) {
    case OPM_FLAG_CHIP_OPL3:
        // set OPL3 mode
        OPL3_WriteRegBuffered(ctx->chip, 0x105, 0x01);
        // disable 4-op
        OPL3_WriteRegBuffered(ctx->chip, 0x104, 0x00);
        // send key off for every channel
        for (int ch = 0; ch < 9; ch++) {
            OPL3_WriteRegBuffered(ctx->chip, 0x0B0 + ch, 0x00);
            OPL3_WriteRegBuffered(ctx->chip, 0x1B0 + ch, 0x00);
        }
        break;
    case OPM_FLAG_CHIP_OPL2:
        // enable waveforms
        OPL3_WriteRegBuffered(ctx->chip, 0x001, 0x20);
        // send key off for every channel
        for (int ch = 0; ch < 9; ch++) {
            OPL3_WriteRegBuffered(ctx->chip, 0x0B0 + ch, 0x00);
        }
        break;

    default: return OPMPLAY_ERR_DEVICE;
    }

    for (int ch = 0; ch < ctx->header.channels + 1; ch++) {
        ctx->channels[ch].stream.loop = ctx->channels[ch].stream.data;
    }
    ctx->_4op = 0;

    opmplay_loop(ctx);

    return OPMPLAY_ERR_OK;
}

// get and parse delay
static uint32_t opmplay_set_delay(uint8_t** data) {
    uint32_t delay = 0;
    if (**data == OPM_STREAM_DELAY_INT32) {
        delay = (
            (*(*data + 1) << 0) |
            (*(*data + 2) << 8) |
            (*(*data + 3) << 16) |
            (*(*data + 4) << 24)
            );
        *data += 5;
    }
    else if (**data == OPM_STREAM_DELAY_INT16) {
        delay = (
            (*(*data + 1) << 0) |
            (*(*data + 2) << 8)
            );
        *data += 3;
    }
    else if ((**data & 0xF0) == OPM_STREAM_DELAY_INT12) {
        delay = ((**data & 0x0F) << 8) | (*(*data + 1));
        *data += 2;
    }
    else if ((**data & 0xF0) == OPM_STERAM_DELAY_SHORT) {
        delay = (**data & 0xF) + 1;
        (*data)++;
    }
    return delay;
}

// translation tables
static int opmplay_opregs_channel_offset[] = {
    0x000, 0x001, 0x002, 0x008, 0x009, 0x00A, 0x010, 0x011, 0x012,
    0x100, 0x101, 0x102, 0x108, 0x109, 0x10A, 0x110, 0x111, 0x112,
};
static int opmplay_regs_channel_offset[] = {
    0x000, 0x001, 0x002, 0x003, 0x004, 0x005, 0x006, 0x007, 0x008,
    0x100, 0x101, 0x102, 0x103, 0x104, 0x105, 0x106, 0x107, 0x108,
};
static int opmplay_4op_bit_offset[] = {
    1<<0, 1<<1, 1<<2, 0, 0, 0, 0, 0, 0,
    1<<3, 1<<4, 1<<5, 0, 0, 0, 0, 0, 0,
};
static int opmplay_ch_parsing_order[] = {
     3,  0,  4,  1,  5,  2,  6,  7,  8,
    12,  9, 13, 10, 14, 11, 15, 16, 17
};

int opmplay_tick(opmplay_context_t* ctx) {

    int ch = 0;
    int rtn = OPMPLAY_ERR_OK;
    uint32_t newdelay = 0;

    // parse channel streams
    for (int i = 0; i < ctx->header.channels; i++) {
        int ch = opmplay_ch_parsing_order[i]; // mangle channel order to make 4op mode work

        opmplay_channel_context_t* chctx = ctx->channels + ch + 1;
        uint8_t* data = chctx->stream.ptr; if (data == NULL) continue;
        bool isRun = true;
        if (--chctx->stream.delay == 0) {
            while (isRun) {
                // get streams
                if ((*(data) & 0x80) == OPM_SET_OPERATOR) {
                    int mask = *data;
                    int op = (mask & OPM_CMD00_SELECT_OPERATOR ? 3 : 0) + opmplay_opregs_channel_offset[ch];
                    data++;
                    if (mask & OPM_CMD00_SET_MULT)      OPL3_WriteRegBuffered(ctx->chip, 0x20 + op, *data++);
                    if (mask & OPM_CMD00_SET_TL)        OPL3_WriteRegBuffered(ctx->chip, 0x40 + op, *data++);
                    if (mask & OPM_CMD00_SET_AD)        OPL3_WriteRegBuffered(ctx->chip, 0x60 + op, *data++);
                    if (mask & OPM_CMD00_SET_SR)        OPL3_WriteRegBuffered(ctx->chip, 0x80 + op, *data++);
                    if (mask & OPM_CMD00_SET_WAVEFORM)  OPL3_WriteRegBuffered(ctx->chip, 0xE0 + op, *data++);
                    if (mask & OPM_CMD00_END_OF_FRAME)  isRun = false;
                    continue;
                }
                if ((*(data) & 0xC0) == OPM_SET_FREQ_FB_VOL) {
                    int mask = *data;
                    data++;
                    if (mask & OPM_CMD80_SET_TL0)       OPL3_WriteRegBuffered(ctx->chip, 0x40 + opmplay_opregs_channel_offset[ch], *data++);
                    if (mask & OPM_CMD80_SET_TL1)       OPL3_WriteRegBuffered(ctx->chip, 0x43 + opmplay_opregs_channel_offset[ch], *data++);
                    if (mask & OPM_CMD80_SET_FEEDBACK)  OPL3_WriteRegBuffered(ctx->chip, 0xC0 + opmplay_regs_channel_offset[ch], *data++);
                    if (mask & OPM_CMD80_SET_FREQ)      OPL3_WriteRegBuffered(ctx->chip, 0xA0 + opmplay_regs_channel_offset[ch], *data++);
                    if (mask & OPM_CMD80_SET_KEYBLOCK) { chctx->block = *data; OPL3_WriteRegBuffered(ctx->chip, 0xB0 + opmplay_regs_channel_offset[ch], *data++); };
                    if (mask & OPM_CMD80_END_OF_FRAME)  isRun = false;
                    continue;
                }
                if ((*data & 0xFC) == OPM_KEY_TRIGGER) {
                    if (*data & OPM_KEY_ON) chctx->block |= (1 << 5); else chctx->block &= ~(1 << 5);
                    OPL3_WriteRegBuffered(ctx->chip, 0xB0 + opmplay_regs_channel_offset[ch], chctx->block);
                    if (*data & OPM_KEY_END_OF_FRAME)  isRun = false;
                    data++;
                    continue;
                }
                if ((*data & 0xFE) == OPM_4OP_TRIGGER) {
                    if (opmplay_4op_bit_offset[ch] != 0) {
                        if (*data & OPM_4OP_ON) ctx->_4op |= opmplay_4op_bit_offset[ch]; else ctx->_4op &= ~opmplay_4op_bit_offset[ch];
                        OPL3_WriteRegBuffered(ctx->chip, 0x104, ctx->_4op);
                    }
                    data++;
                    continue;
                }
                if ((*data & 0xF0) == OPM_STREAM_BACKREF) {
                    // back reference, nested call :)
                    int distance = ((*(data + 0) & 0x0F) << 8) | (*(data + 1));
                    int frames_to_play = *(data + 2);
                    chctx->stream.ptr = data + 3;
                    opmplay_push_stack(chctx);
                    data -= distance;
                    chctx->stream.samples_to_play = frames_to_play; // hack?
                    continue;
                }

                // check for common stuff
                switch (*data) {
                case OPM_STREAM_END:
                    // end of current stream, delay forever
                    chctx->stream.reload = -1;
                    isRun = false;
                    break;
                    // just an NOP, break
                case OPM_STREAM_NEW_ORDER:
                case OPM_STREAM_NOP:
                    data++;
                    break;
                case OPM_STREAM_LOOP:
                    // save loop point
                    chctx->stream.loop = data;
                    data++;
                    break;
                case OPM_STREAM_END_FRAME:
                    // end of frame - special case here
                    data++;
                    isRun = false;
                    break;
                    
                default:
                    // test for delay
                    newdelay = opmplay_set_delay(&data);
                    if (newdelay > 0) {
                        chctx->stream.reload = newdelay;
                    }
                    else
                    {
                        printf("unknonw token %02x!\n", *data);
                        return OPMPLAY_ERR_BAD_FILE_STRUCTURE;
                    }
                }
            }
            chctx->stream.delay = chctx->stream.reload;

            // decrement samples to play counter
            if (--chctx->stream.samples_to_play == 0) {
                // pop context from the stack
                do opmplay_pop_stack(chctx); while (--chctx->stream.samples_to_play == 0);
            }
            else {
                // save data pointer
                chctx->stream.ptr = data;
            }
        }
    }

    // process control stream
    {
        opmplay_channel_context_t* chctx = ctx->channels + 0;
        uint8_t* data = chctx->stream.ptr;
        bool isRun = true;
        if (--chctx->stream.delay == 0) {
            while (isRun) {
                if ((*(data) & 0xE0) == OPM_CTRL_REG_SET) {
                    int mask = *data;
                    data++;
                    if (mask & OPM_CTRL_SET_REG01)  OPL3_WriteRegBuffered(ctx->chip, 0x001, *data++);
                    if (mask & OPM_CTRL_SET_REG08)  OPL3_WriteRegBuffered(ctx->chip, 0x008, *data++);
                    if (mask & OPM_CTRL_SET_REG105) OPL3_WriteRegBuffered(ctx->chip, 0x105, *data++);
                    if (mask & OPM_CTRL_SET_REG104) OPL3_WriteRegBuffered(ctx->chip, 0x104, *data++);
                    if (mask & OPM_CTRL_SET_REGBD)  OPL3_WriteRegBuffered(ctx->chip, 0x0BD, *data++);
                    continue;
                }
                if ((*data & 0xF0) == OPM_STREAM_BACKREF) {
                    // back reference, nested call :)
                    int distance = ((*(data + 0) & 0x0F) << 8) | (*(data + 1));
                    int frames_to_play = *(data + 2);
                    chctx->stream.ptr = data + 3;
                    opmplay_push_stack(chctx);
                    data -= distance;
                    chctx->stream.samples_to_play = frames_to_play; // hack?
                    continue;
                }

                // check for common stuff
                switch (*data) {
                    // end of stream - rewind everything
                case OPM_STREAM_END:
                    opmplay_loop(ctx);
                    isRun = false;
                    return OPMPLAY_ERR_END_OF_STREAM;
                    // just an NOP, break
                case OPM_STREAM_NEW_ORDER:
                case OPM_STREAM_NOP:
                    data++;
                    break;
                case OPM_STREAM_LOOP:
                    // save loop point
                    ctx->pos.looped = ctx->pos.frame;
                    chctx->stream.loop = data;
                    data++;
                    break;
                    // set new frame rate
                case OPM_STREAM_SET_FRAME_RATE:
                    ctx->header.frame_rate = *(uint16_t*)(data + 1); data += 3;
                    break;
                case OPM_STREAM_END_FRAME:
                    // end of frame - special case here
                    data++;
                    isRun = false;
                    break;

                default:
                    // test for delay
                    newdelay = opmplay_set_delay(&data);
                    if (newdelay > 0) {
                        chctx->stream.reload = newdelay;
                    }
                    else
                    {
                        printf("unknonw token %02x!\n", *data);
                        return OPMPLAY_ERR_BAD_FILE_STRUCTURE;
                    }
                }
            }
            chctx->stream.delay = chctx->stream.reload;
            // decrement samples to play counter
            if (--chctx->stream.samples_to_play == 0) {
                // pop context from the stack
                do opmplay_pop_stack(chctx); while (--chctx->stream.samples_to_play == 0);
            }
            else {
                // save data pointer
                chctx->stream.ptr = data;
            }
        }
    }

    ctx->pos.frame++;

    return rtn;
}