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

#pragma pack(push, 1)
typedef struct BmHeader{
    uint8_t  ID[2];
    uint32_t file_size, reserved, offset;
} BmHeader;

typedef struct BmInfoHeader{
    uint32_t size, width, height;
    uint16_t  planes, bpp;
    uint32_t compession, image_size, unused[2], colors, important;
} BmInfoHeader;
#pragma pack(pop)

#define debug 0
#if debug
#define log(fnt, ...) fprintf(stderr, fnt "\n", __VA_ARGS__)
#else
#define log(...)
#endif

int main(int args, char const *argv[]) {

    if (args < 3)
    {
        printf("%s <bmp in> <bmp out>", argv[0]);
        exit(1);
    }
    FILE *fr = fopen(argv[1], "rb");
    FILE *fw = fopen(argv[2], "wb");

    BmHeader bh;
    BmInfoHeader bih;

    fread(&bh, sizeof(BmHeader), 1, fr);
    fread(&bih.size, sizeof(unsigned int), 1, fr);
    fread(&bih.width, bih.size - 4, 1, fr);

    if (bih.size != 40 || bih.planes != 1 || bih.compession > 2){
        printf("Bmp file corrupted\n");
        exit(2);
    }
    if (bih.bpp != 24)
    {
        printf("bpp %d not supported\n", bih.bpp);
        exit(3);
    }

    log("fsize %d, offset 0x%x",bh.file_size, bh.offset);
    log("bpp %hd, compress %d, colors used %d",
        bih.bpp, bih.compession, bih.colors);

    unsigned int bufwidth = bih.width, bufheight = bih.height;

    fwrite(&bh, sizeof(BmHeader), 1, fw);
    fwrite(&bih, bih.size, 1, fw);

    char *buffer = malloc(bh.offset - (sizeof(BmHeader) + bih.size));
    fread(buffer, bh.offset - (sizeof(BmHeader) + bih.size), 1, fr);
    fwrite(buffer, bh.offset - (sizeof(BmHeader) + bih.size), 1, fw);
    free(buffer);

    buffer = malloc(bih.image_size);
    fread(buffer, bih.image_size, 1, fr);

    unsigned char bpp = bih.bpp / 8;

    for (int i = 0; i < bufheight / 2; i++) {
        for (int j = 0; j < bufwidth * bpp; j++) {
            char tmp = buffer[bpp * i * bufwidth + j];
            buffer[bpp * i * bufwidth + j] = buffer[bpp * bufwidth * (bufheight - i - 1) + j];
            buffer[bpp *  bufwidth * (bufheight - i - 1) + j] = tmp;
        }
    }

    char exbuf[6];

    for (int i = 0; i < bufheight; i++) {
        for (int j = 0; j < bufwidth / 2; j++) {
            memcpy(exbuf, (&buffer[bpp * i * bufwidth + bpp * j]), bpp);
            memcpy(&buffer[bpp * i * bufwidth + bpp * j], &buffer[bpp * i * bufwidth + bpp * (bufwidth - 1 - j)], bpp);
            memcpy(&buffer[bpp * i * bufwidth + bpp * (bufwidth - 1 - j)], exbuf, bpp);
        }
    }
    fwrite(buffer, bih.image_size, 1, fw);

    free(buffer);

    buffer = malloc(bh.file_size - bh.offset - bih.image_size);

    fread(buffer, 1, bh.file_size - bh.offset - bih.image_size, fr);
    fwrite(buffer, 1, bh.file_size - bh.offset - bih.image_size, fw);

    free(buffer);

    fclose(fr);
    fclose(fw);

    return 0;
}