/* Ok, one other line algorithm.
 * With subpixel and anti-alias, that's the point !
 * Starting : Tue Sep 21 17:44:20 MET DST 1999
 * Wed Sep 22 00:11:58 MET DST 1999 : almost done.
 * I had some eating time and washing too, so I did not
 * spend all this time on it. I must do a max routine
 * (put the max value between the point and the current
 * point). Not sure about my anti-alias stuff. Gonna watch some
 * demos to see what it give with them.
 * Let's go.
 */

#include "line.h"

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

#include "screen.h"
#include "device.h"
#include "poly.h"

device_line *the_lines;

/* returns a*b/c in int 16:16 */
static inline int muldiv(int a, int b, int c)
{
#ifdef PC_ARCHI
  register int r;
  __asm__ volatile ("imul %2  \n"
		    "idiv %3  \n"
		    : "=a" (r)
		    : "a" (a), "b" (b), "c" (c)
		    : "eax", "ebx", "ecx", "edx"
		    );
  return r;
#else
  return (double)a*(double)b/(double)c;
#endif
}

void drawline8(device *d, line *l)
{
  /* lots of goto for lisibility (if if else if if else else is impossible to debug) */
  register int x1=l->p1.x*65536., x2=l->p2.x*65536.,
               y1=l->p1.y*65536., y2=l->p2.y*65536., dx, dy;
  register int xcur;
  register int xinc;
  register int ycur;
  register int yinc;
  register int i;
  register int y1i=(y1+65535)>>16, y2i=(y2+65535)>>16;
  register int x1i=(x1+65535)>>16, x2i=(x2+65535)>>16;
  register unsigned char *dest;
  register unsigned char col1=l->color;
  register int xf;
  register int yf;

  /* in asm, those two subs are good for us, because we can test if x1<x2 and if y1<y2 at the same time, but it's C */
  dx=x2-x1;
  dy=y2-y1;

  if (dx<0)
    goto x1_sup_x2;
  if (dx)
    goto x1_inf_x2;

  /* here x1==x2 */
  if (dy<0)
    goto x1_eq_x2_and_y1_sup_y2;
  if (dy>0)
    goto x1_eq_x2_and_y1_inf_y2;
  /* here x1==x2 && y1==y2 */
  if (y1i<y2i) {
    dest=d->buffer+y1i*SCREEN_X+(x1>>16);
  } else {
    dest=d->buffer+y2i*SCREEN_X+(x1>>16);
  }
  *dest=col1;
  return;
x1_eq_x2_and_y1_sup_y2:
  dest=d->buffer+y2i*SCREEN_X+(x1>>16);
  for (i=y2i; i<y1i; dest+=SCREEN_X, i++) {
    *dest=col1;
  }
  return;
x1_eq_x2_and_y1_inf_y2:
  dest=d->buffer+y1i*SCREEN_X+(x1>>16);
  for (i=y1i; i<y2i; dest+=SCREEN_X, i++) {
    *dest=col1;
  }
  return;

x1_inf_x2:
  if (dy<0)
    goto x1_inf_x2_and_y1_sup_y2;
  if (dy>0)
    goto x1_inf_x2_and_y1_inf_y2;
  /* x1<x2 && y1==y2 */
  dest=d->buffer+(y1>>16)*SCREEN_X+x1i;
  for (i=x1i; i<x2i; dest++, i++) {
    *dest=col1;
  }
  return;

x1_inf_x2_and_y1_sup_y2:
  /*
                  (x2, y2)


       (x1, y1)
   */
  /* remove those abs in the asm code */
  if (abs(dx)>abs(dy))
    goto x1_inf_x2_and_y1_sup_y2_dx_sup_dy;
  if (abs(dx)<abs(dy))
    goto x1_inf_x2_and_y1_sup_y2_dx_inf_dy;
  /* here dx == dy */
  dest=d->buffer+y2i*SCREEN_X+(x2>>16);
  for (i=y2i; i<y1i; dest+=SCREEN_X-1, i++) {
    *dest=col1;
  }
  return;
x1_inf_x2_and_y1_sup_y2_dx_sup_dy:
  /*
                    xxxx
		xxxx
	    xxxx
	xxxx
  */
  yinc=muldiv(dy, 65536, dx);
  ycur=y2+muldiv(yinc, (x2i<<16)-x2, 65536);
  yf=ycur&65535;
  dest=d->buffer+(ycur>>16)*SCREEN_X+x2i;
  for (i=x2i; i>x1i; i--, dest--) {
    *dest=col1;
    yf-=yinc;
    if (yf>65535) {
      yf-=65536;
      dest+=SCREEN_X;
    }
  }
  return;
x1_inf_x2_and_y1_sup_y2_dx_inf_dy:
  /*
          x
	  x
	  x
	 x
	 x
	 x
	x
	x
	x
  */
  xinc=muldiv(dx, 65536, dy);
  xcur=x2+muldiv(xinc, (y2i<<16)-y2, 65536);
  xf=xcur&65535;
  dest=d->buffer+y2i*SCREEN_X+(xcur>>16);
  for (i=y2i; i<y1i; i++, dest+=SCREEN_X) {
    *dest=col1;
    xf+=xinc;
    if (xf<0) {
      xf+=65536;
      dest--;
    }
  }
  return;

x1_inf_x2_and_y1_inf_y2:
  /*
      (x1, y1)


                  (x2, y2)
  */
  if (abs(dx)>abs(dy))
    goto x1_inf_x2_and_y1_inf_y2_and_dx_sup_dy;
  if (abs(dx)<abs(dy))
    goto x1_inf_x2_and_y1_inf_y2_and_dx_inf_dy;
  /* here dx == dy */
  dest=d->buffer+y1i*SCREEN_X+(x1>>16);
  for (i=y1i; i<y2i; dest+=SCREEN_X+1, i++) {
    *dest=col1;
  }
  return;

x1_inf_x2_and_y1_inf_y2_and_dx_sup_dy:
  /*
        xxxx
	    xxxx
	        xxxx
   */
  yinc=muldiv(dy, 65536, dx);
  ycur=y1+muldiv(yinc, (x1i<<16)-x1, 65536);
  yf=ycur&65535;
  dest=d->buffer+(ycur>>16)*SCREEN_X+x1i;
  for (i=x1i; i<x2i; i++, dest++) {
    *dest=col1;
    yf+=yinc;
    if (yf>65535) {
      yf-=65536;
      dest+=SCREEN_X;
    }
  }
  return;
x1_inf_x2_and_y1_inf_y2_and_dx_inf_dy:
  /*
      x
      x
      x
       x
       x
       x
        x
	x
	x
  */
  xinc=muldiv(dx, 65536, dy);
  xcur=x1+muldiv(xinc, (y1i<<16)-y1, 65536);
  xf=xcur&65535;
  dest=d->buffer+y1i*SCREEN_X+(xcur>>16);
  for (i=y1i; i<y2i; i++, dest+=SCREEN_X) {
    *dest=col1;
    xf+=xinc;
    if (xf>65536) {
      xf-=65536;
      dest++;
    }
  }
  return;

x1_sup_x2:
  if (dy<0)
    goto x1_sup_x2_and_y1_sup_y2;
  if (dy>0)
    goto x1_sup_x2_and_y1_inf_y2;
  /* x1>x2 && y1==y2 */
  dest=d->buffer+(y2>>16)*SCREEN_X+x2i;
  for (i=x2i; i<x1i; dest++, i++) {
    *dest=col1;
  }
  return;
x1_sup_x2_and_y1_sup_y2:
  /*
       (x2, y2)


                 (x1, y1)
   */
  if (abs(dx)>abs(dy))
    goto x1_sup_x2_and_y1_sup_y2_and_dx_sup_dy;
  if (abs(dx)<abs(dy))
    goto x1_sup_x2_and_y1_sup_y2_and_dx_inf_dy;
  /* here dx == dy */
  dest=d->buffer+y2i*SCREEN_X+(x2>>16);
  for (i=y2i; i<y1i; dest+=SCREEN_X+1, i++) {
    *dest=col1;
  }
  return;
x1_sup_x2_and_y1_sup_y2_and_dx_sup_dy:
  yinc=muldiv(dy, 65536, dx);
  ycur=y2+muldiv(yinc, (x2i<<16)-x2, 65536);
  yf=ycur&65535;
  dest=d->buffer+(ycur>>16)*SCREEN_X+x2i;
  for (i=x2i; i<x1i; i++, dest++) {
    *dest=col1;
    yf+=yinc;
    if (yf>65535) {
      yf-=65536;
      dest+=SCREEN_X;
    }
  }
  return;
x1_sup_x2_and_y1_sup_y2_and_dx_inf_dy:
  xinc=muldiv(dx, 65536, dy);
  xcur=x2+muldiv(xinc, (y2i<<16)-y2, 65536);
  xf=xcur&65535;
  dest=d->buffer+y2i*SCREEN_X+(xcur>>16);
  for (i=y2i; i<y1i; i++, dest+=SCREEN_X) {
    *dest=col1;
    xf+=xinc;
    if (xf>65536) {
      xf-=65536;
      dest++;
    }
  }
  return;

x1_sup_x2_and_y1_inf_y2:
  /*
                 (x1, y1)


      (x2, y2)
   */
  if (abs(dx)>abs(dy))
    goto x1_sup_x2_and_y1_inf_y2_and_dx_sup_dy;
  if (abs(dx)<abs(dy))
    goto x1_sup_x2_and_y1_inf_y2_and_dx_inf_dy;
  /* here dx == dy */
  dest=d->buffer+y1i*SCREEN_X+(x1>>16);
  for (i=y1i; i<y2i; dest+=SCREEN_X-1, i++) {
    *dest=col1;
  }
  return;
x1_sup_x2_and_y1_inf_y2_and_dx_sup_dy:
  yinc=muldiv(dy, 65536, dx);
  ycur=y1+muldiv(yinc, (x1i<<16)-x1, 65536);
  yf=ycur&65535;
  dest=d->buffer+(ycur>>16)*SCREEN_X+x1i;
  for (i=x1i; i>x2i; i--, dest--) {
    *dest=col1;
    yf-=yinc;
    if (yf>65535) {
      yf-=65536;
      dest+=SCREEN_X;
    }
  }
  return;
x1_sup_x2_and_y1_inf_y2_and_dx_inf_dy:
  xinc=muldiv(dx, 65536, dy);
  xcur=x1+muldiv(xinc, (y1i<<16)-y1, 65536);
  xf=xcur&65535;
  dest=d->buffer+y1i*SCREEN_X+(xcur>>16);
  for (i=y1i; i<y2i; i++, dest+=SCREEN_X) {
    *dest=col1;
    xf+=xinc;
    if (xf<0) {
      xf+=65536;
      dest--;
    }
  }
  return;
}

void drawline16(device *d, line *l)
{
  /* lots of goto for lisibility (if if else if if else else is impossible to debug) */
  register int x1=l->p1.x*65536., x2=l->p2.x*65536., y1=l->p1.y*65536., y2=l->p2.y*65536., dx, dy;
  register int xcur;
  register int xinc;
  register int ycur;
  register int yinc;
  register int i;
  register int y1i=(y1+65535)>>16, y2i=(y2+65535)>>16;
  register int x1i=(x1+65535)>>16, x2i=(x2+65535)>>16;
  register unsigned short *dest;
  register unsigned short col1=l->color;
  register int xf;
  register int yf;

  /* in asm, those two subs are good for us, because we can test if x1<x2 and if y1<y2 at the same time, but it's C */
  dx=x2-x1;
  dy=y2-y1;

  if (dx<0)
    goto x1_sup_x2;
  if (dx)
    goto x1_inf_x2;

  /* here x1==x2 */
  if (dy<0)
    goto x1_eq_x2_and_y1_sup_y2;
  if (dy>0)
    goto x1_eq_x2_and_y1_inf_y2;
  /* here x1==x2 && y1==y2 */
  if (y1i<y2i) {
    dest=(unsigned short *)d->buffer+y1i*SCREEN_X+(x1>>16);
  } else {
    dest=(unsigned short *)d->buffer+y2i*SCREEN_X+(x1>>16);
  }
  *dest=col1;
  return;
x1_eq_x2_and_y1_sup_y2:
  dest=(unsigned short *)d->buffer+y2i*SCREEN_X+(x1>>16);
  for (i=y2i; i<y1i; dest+=SCREEN_X, i++) {
    *dest=col1;
  }
  return;
x1_eq_x2_and_y1_inf_y2:
  dest=(unsigned short *)d->buffer+y1i*SCREEN_X+(x1>>16);
  for (i=y1i; i<y2i; dest+=SCREEN_X, i++) {
    *dest=col1;
  }
  return;

x1_inf_x2:
  if (dy<0)
    goto x1_inf_x2_and_y1_sup_y2;
  if (dy>0)
    goto x1_inf_x2_and_y1_inf_y2;
  /* x1<x2 && y1==y2 */
  dest=(unsigned short *)d->buffer+(y1>>16)*SCREEN_X+x1i;
  for (i=x1i; i<x2i; dest++, i++) {
    *dest=col1;
  }
  return;

x1_inf_x2_and_y1_sup_y2:
  /*
                  (x2, y2)


       (x1, y1)
   */
  /* remove those abs in the asm code */
  if (abs(dx)>abs(dy))
    goto x1_inf_x2_and_y1_sup_y2_dx_sup_dy;
  if (abs(dx)<abs(dy))
    goto x1_inf_x2_and_y1_sup_y2_dx_inf_dy;
  /* here dx == dy */
  dest=(unsigned short *)d->buffer+y2i*SCREEN_X+(x2>>16);
  for (i=y2i; i<y1i; dest+=SCREEN_X-1, i++) {
    *dest=col1;
  }
  return;
x1_inf_x2_and_y1_sup_y2_dx_sup_dy:
  /*
                    xxxx
		xxxx
	    xxxx
	xxxx
  */
  yinc=muldiv(dy, 65536, dx);
  ycur=y2+muldiv(yinc, (x2i<<16)-x2, 65536);
  yf=ycur&65535;
  dest=(unsigned short *)d->buffer+(ycur>>16)*SCREEN_X+x2i;
  for (i=x2i; i>x1i; i--, dest--) {
    *dest=col1;
    yf-=yinc;
    if (yf>65535) {
      yf-=65536;
      dest+=SCREEN_X;
    }
  }
  return;
x1_inf_x2_and_y1_sup_y2_dx_inf_dy:
  /*
          x
	  x
	  x
	 x
	 x
	 x
	x
	x
	x
  */
  xinc=muldiv(dx, 65536, dy);
  xcur=x2+muldiv(xinc, (y2i<<16)-y2, 65536);
  xf=xcur&65535;
  dest=(unsigned short *)d->buffer+y2i*SCREEN_X+(xcur>>16);
  for (i=y2i; i<y1i; i++, dest+=SCREEN_X) {
    *dest=col1;
    xf+=xinc;
    if (xf<0) {
      xf+=65536;
      dest--;
    }
  }
  return;

x1_inf_x2_and_y1_inf_y2:
  /*
      (x1, y1)


                  (x2, y2)
  */
  if (abs(dx)>abs(dy))
    goto x1_inf_x2_and_y1_inf_y2_and_dx_sup_dy;
  if (abs(dx)<abs(dy))
    goto x1_inf_x2_and_y1_inf_y2_and_dx_inf_dy;
  /* here dx == dy */
  dest=(unsigned short *)d->buffer+y1i*SCREEN_X+(x1>>16);
  for (i=y1i; i<y2i; dest+=SCREEN_X+1, i++) {
    *dest=col1;
  }
  return;

x1_inf_x2_and_y1_inf_y2_and_dx_sup_dy:
  /*
        xxxx
	    xxxx
	        xxxx
   */
  yinc=muldiv(dy, 65536, dx);
  ycur=y1+muldiv(yinc, (x1i<<16)-x1, 65536);
  yf=ycur&65535;
  dest=(unsigned short *)d->buffer+(ycur>>16)*SCREEN_X+x1i;
  for (i=x1i; i<x2i; i++, dest++) {
    *dest=col1;
    yf+=yinc;
    if (yf>65535) {
      yf-=65536;
      dest+=SCREEN_X;
    }
  }
  return;
x1_inf_x2_and_y1_inf_y2_and_dx_inf_dy:
  /*
      x
      x
      x
       x
       x
       x
        x
	x
	x
  */
  xinc=muldiv(dx, 65536, dy);
  xcur=x1+muldiv(xinc, (y1i<<16)-y1, 65536);
  xf=xcur&65535;
  dest=(unsigned short *)d->buffer+y1i*SCREEN_X+(xcur>>16);
  for (i=y1i; i<y2i; i++, dest+=SCREEN_X) {
    *dest=col1;
    xf+=xinc;
    if (xf>65536) {
      xf-=65536;
      dest++;
    }
  }
  return;

x1_sup_x2:
  if (dy<0)
    goto x1_sup_x2_and_y1_sup_y2;
  if (dy>0)
    goto x1_sup_x2_and_y1_inf_y2;
  /* x1>x2 && y1==y2 */
  dest=(unsigned short *)d->buffer+(y2>>16)*SCREEN_X+x2i;
  for (i=x2i; i<x1i; dest++, i++) {
    *dest=col1;
  }
  return;
x1_sup_x2_and_y1_sup_y2:
  /*
       (x2, y2)


                 (x1, y1)
   */
  if (abs(dx)>abs(dy))
    goto x1_sup_x2_and_y1_sup_y2_and_dx_sup_dy;
  if (abs(dx)<abs(dy))
    goto x1_sup_x2_and_y1_sup_y2_and_dx_inf_dy;
  /* here dx == dy */
  dest=(unsigned short *)d->buffer+y2i*SCREEN_X+(x2>>16);
  for (i=y2i; i<y1i; dest+=SCREEN_X+1, i++) {
    *dest=col1;
  }
  return;
x1_sup_x2_and_y1_sup_y2_and_dx_sup_dy:
  yinc=muldiv(dy, 65536, dx);
  ycur=y2+muldiv(yinc, (x2i<<16)-x2, 65536);
  yf=ycur&65535;
  dest=(unsigned short *)d->buffer+(ycur>>16)*SCREEN_X+x2i;
  for (i=x2i; i<x1i; i++, dest++) {
    *dest=col1;
    yf+=yinc;
    if (yf>65535) {
      yf-=65536;
      dest+=SCREEN_X;
    }
  }
  return;
x1_sup_x2_and_y1_sup_y2_and_dx_inf_dy:
  xinc=muldiv(dx, 65536, dy);
  xcur=x2+muldiv(xinc, (y2i<<16)-y2, 65536);
  xf=xcur&65535;
  dest=(unsigned short *)d->buffer+y2i*SCREEN_X+(xcur>>16);
  for (i=y2i; i<y1i; i++, dest+=SCREEN_X) {
    *dest=col1;
    xf+=xinc;
    if (xf>65536) {
      xf-=65536;
      dest++;
    }
  }
  return;

x1_sup_x2_and_y1_inf_y2:
  /*
                 (x1, y1)


      (x2, y2)
   */
  if (abs(dx)>abs(dy))
    goto x1_sup_x2_and_y1_inf_y2_and_dx_sup_dy;
  if (abs(dx)<abs(dy))
    goto x1_sup_x2_and_y1_inf_y2_and_dx_inf_dy;
  /* here dx == dy */
  dest=(unsigned short *)d->buffer+y1i*SCREEN_X+(x1>>16);
  for (i=y1i; i<y2i; dest+=SCREEN_X-1, i++) {
    *dest=col1;
  }
  return;
x1_sup_x2_and_y1_inf_y2_and_dx_sup_dy:
  yinc=muldiv(dy, 65536, dx);
  ycur=y1+muldiv(yinc, (x1i<<16)-x1, 65536);
  yf=ycur&65535;
  dest=(unsigned short *)d->buffer+(ycur>>16)*SCREEN_X+x1i;
  for (i=x1i; i>x2i; i--, dest--) {
    *dest=col1;
    yf-=yinc;
    if (yf>65535) {
      yf-=65536;
      dest+=SCREEN_X;
    }
  }
  return;
x1_sup_x2_and_y1_inf_y2_and_dx_inf_dy:
  xinc=muldiv(dx, 65536, dy);
  xcur=x1+muldiv(xinc, (y1i<<16)-y1, 65536);
  xf=xcur&65535;
  dest=(unsigned short *)d->buffer+y1i*SCREEN_X+(xcur>>16);
  for (i=y1i; i<y2i; i++, dest+=SCREEN_X) {
    *dest=col1;
    xf+=xinc;
    if (xf<0) {
      xf+=65536;
      dest--;
    }
  }
  return;
}

void drawline24(device *d, line *l)
{
  /* lots of goto for lisibility (if if else if if else else is impossible to debug) */
  register int x1=l->p1.x*65536., x2=l->p2.x*65536., y1=l->p1.y*65536., y2=l->p2.y*65536., dx, dy;
  register int xcur;
  register int xinc;
  register int ycur;
  register int yinc;
  register int i;
  register int y1i=(y1+65535)>>16, y2i=(y2+65535)>>16;
  register int x1i=(x1+65535)>>16, x2i=(x2+65535)>>16;
  register unsigned char *dest;
  register unsigned char col1=l->color&255;
  register unsigned char col2=(l->color>>8)&255;
  register unsigned char col3=l->color>>16;
  register int xf;
  register int yf;

  /* in asm, those two subs are good for us, because we can test if x1<x2 and if y1<y2 at the same time, but it's C */
  dx=x2-x1;
  dy=y2-y1;

  if (dx<0)
    goto x1_sup_x2;
  if (dx)
    goto x1_inf_x2;

  /* here x1==x2 */
  if (dy<0)
    goto x1_eq_x2_and_y1_sup_y2;
  if (dy>0)
    goto x1_eq_x2_and_y1_inf_y2;
  /* here x1==x2 && y1==y2 */
  if (y1i<y2i) {
    dest=d->buffer+(y1i*SCREEN_X+(x1>>16))*3;
  } else {
    dest=d->buffer+(y2i*SCREEN_X+(x1>>16))*3;
  }
  *dest=col1;
  *(dest+1)=col2;
  *(dest+2)=col3;
  return;
x1_eq_x2_and_y1_sup_y2:
  dest=d->buffer+(y2i*SCREEN_X+(x1>>16))*3;
  for (i=y2i; i<y1i; dest+=SCREEN_X*3, i++) {
    *dest=col1;
    *(dest+1)=col2;
    *(dest+2)=col3;
  }
  return;
x1_eq_x2_and_y1_inf_y2:
  dest=d->buffer+(y1i*SCREEN_X+(x1>>16))*3;
  for (i=y1i; i<y2i; dest+=SCREEN_X*3, i++) {
    *dest=col1;
    *(dest+1)=col2;
    *(dest+2)=col3;
  }
  return;

x1_inf_x2:
  if (dy<0)
    goto x1_inf_x2_and_y1_sup_y2;
  if (dy>0)
    goto x1_inf_x2_and_y1_inf_y2;
  /* x1<x2 && y1==y2 */
  dest=d->buffer+((y1>>16)*SCREEN_X+x1i)*3;
  for (i=x1i; i<x2i; dest+=3, i++) {
    *dest=col1;
    *(dest+1)=col2;
    *(dest+2)=col3;
  }
  return;

x1_inf_x2_and_y1_sup_y2:
  /*
                  (x2, y2)


       (x1, y1)
   */
  /* remove those abs in the asm code */
  if (abs(dx)>abs(dy))
    goto x1_inf_x2_and_y1_sup_y2_dx_sup_dy;
  if (abs(dx)<abs(dy))
    goto x1_inf_x2_and_y1_sup_y2_dx_inf_dy;
  /* here dx == dy */
  dest=d->buffer+(y2i*SCREEN_X+(x2>>16))*3;
  for (i=y2i; i<y1i; dest+=SCREEN_X-1*3, i++) {
    *dest=col1;
    *(dest+1)=col2;
    *(dest+2)=col3;
  }
  return;
x1_inf_x2_and_y1_sup_y2_dx_sup_dy:
  /*
                    xxxx
		xxxx
	    xxxx
	xxxx
  */
  yinc=muldiv(dy, 65536, dx);
  ycur=y2+muldiv(yinc, (x2i<<16)-x2, 65536);
  yf=ycur&65535;
  dest=d->buffer+((ycur>>16)*SCREEN_X+x2i)*3;
  for (i=x2i; i>x1i; i--, dest-=3) {
    *dest=col1;
    *(dest+1)=col2;
    *(dest+2)=col3;
    yf-=yinc;
    if (yf>65535) {
      yf-=65536;
      dest+=SCREEN_X*3;
    }
  }
  return;
x1_inf_x2_and_y1_sup_y2_dx_inf_dy:
  /*
          x
	  x
	  x
	 x
	 x
	 x
	x
	x
	x
  */
  xinc=muldiv(dx, 65536, dy);
  xcur=x2+muldiv(xinc, (y2i<<16)-y2, 65536);
  xf=xcur&65535;
  dest=d->buffer+(y2i*SCREEN_X+(xcur>>16))*3;
  for (i=y2i; i<y1i; i++, dest+=SCREEN_X*3) {
    *dest=col1;
    *(dest+1)=col2;
    *(dest+2)=col3;
    xf+=xinc;
    if (xf<0) {
      xf+=65536;
      dest-=3;
    }
  }
  return;

x1_inf_x2_and_y1_inf_y2:
  /*
      (x1, y1)


                  (x2, y2)
  */
  if (abs(dx)>abs(dy))
    goto x1_inf_x2_and_y1_inf_y2_and_dx_sup_dy;
  if (abs(dx)<abs(dy))
    goto x1_inf_x2_and_y1_inf_y2_and_dx_inf_dy;
  /* here dx == dy */
  dest=d->buffer+(y1i*SCREEN_X+(x1>>16))*3;
  for (i=y1i; i<y2i; dest+=SCREEN_X+1*3, i++) {
    *dest=col1;
    *(dest+1)=col2;
    *(dest+2)=col3;
  }
  return;

x1_inf_x2_and_y1_inf_y2_and_dx_sup_dy:
  /*
        xxxx
	    xxxx
	        xxxx
   */
  yinc=muldiv(dy, 65536, dx);
  ycur=y1+muldiv(yinc, (x1i<<16)-x1, 65536);
  yf=ycur&65535;
  dest=d->buffer+((ycur>>16)*SCREEN_X+x1i)*3;
  for (i=x1i; i<x2i; i++, dest+=3) {
    *dest=col1;
    *(dest+1)=col2;
    *(dest+2)=col3;
    yf+=yinc;
    if (yf>65535) {
      yf-=65536;
      dest+=SCREEN_X*3;
    }
  }
  return;
x1_inf_x2_and_y1_inf_y2_and_dx_inf_dy:
  /*
      x
      x
      x
       x
       x
       x
        x
	x
	x
  */
  xinc=muldiv(dx, 65536, dy);
  xcur=x1+muldiv(xinc, (y1i<<16)-y1, 65536);
  xf=xcur&65535;
  dest=d->buffer+(y1i*SCREEN_X+(xcur>>16))*3;
  for (i=y1i; i<y2i; i++, dest+=SCREEN_X*3) {
    *dest=col1;
    *(dest+1)=col2;
    *(dest+2)=col3;
    xf+=xinc;
    if (xf>65536) {
      xf-=65536;
      dest+=3;
    }
  }
  return;

x1_sup_x2:
  if (dy<0)
    goto x1_sup_x2_and_y1_sup_y2;
  if (dy>0)
    goto x1_sup_x2_and_y1_inf_y2;
  /* x1>x2 && y1==y2 */
  dest=d->buffer+((y2>>16)*SCREEN_X+x2i)*3;
  for (i=x2i; i<x1i; dest+=3, i++) {
    *dest=col1;
    *(dest+1)=col2;
    *(dest+2)=col3;
  }
  return;
x1_sup_x2_and_y1_sup_y2:
  /*
       (x2, y2)


                 (x1, y1)
   */
  if (abs(dx)>abs(dy))
    goto x1_sup_x2_and_y1_sup_y2_and_dx_sup_dy;
  if (abs(dx)<abs(dy))
    goto x1_sup_x2_and_y1_sup_y2_and_dx_inf_dy;
  /* here dx == dy */
  dest=d->buffer+(y2i*SCREEN_X+(x2>>16))*3;
  for (i=y2i; i<y1i; dest+=SCREEN_X+1*3, i++) {
    *dest=col1;
    *(dest+1)=col2;
    *(dest+2)=col3;
  }
  return;
x1_sup_x2_and_y1_sup_y2_and_dx_sup_dy:
  yinc=muldiv(dy, 65536, dx);
  ycur=y2+muldiv(yinc, (x2i<<16)-x2, 65536);
  yf=ycur&65535;
  dest=d->buffer+((ycur>>16)*SCREEN_X+x2i)*3;
  for (i=x2i; i<x1i; i++, dest+=3) {
    *dest=col1;
    *(dest+1)=col2;
    *(dest+2)=col3;
    yf+=yinc;
    if (yf>65535) {
      yf-=65536;
      dest+=SCREEN_X*3;
    }
  }
  return;
x1_sup_x2_and_y1_sup_y2_and_dx_inf_dy:
  xinc=muldiv(dx, 65536, dy);
  xcur=x2+muldiv(xinc, (y2i<<16)-y2, 65536);
  xf=xcur&65535;
  dest=d->buffer+(y2i*SCREEN_X+(xcur>>16))*3;
  for (i=y2i; i<y1i; i++, dest+=SCREEN_X*3) {
    *dest=col1;
    *(dest+1)=col2;
    *(dest+2)=col3;
    xf+=xinc;
    if (xf>65536) {
      xf-=65536;
      dest+=3;
    }
  }
  return;

x1_sup_x2_and_y1_inf_y2:
  /*
                 (x1, y1)


      (x2, y2)
   */
  if (abs(dx)>abs(dy))
    goto x1_sup_x2_and_y1_inf_y2_and_dx_sup_dy;
  if (abs(dx)<abs(dy))
    goto x1_sup_x2_and_y1_inf_y2_and_dx_inf_dy;
  /* here dx == dy */
  dest=d->buffer+(y1i*SCREEN_X+(x1>>16))*3;
  for (i=y1i; i<y2i; dest+=SCREEN_X-1*3, i++) {
    *dest=col1;
    *(dest+1)=col2;
    *(dest+2)=col3;
  }
  return;
x1_sup_x2_and_y1_inf_y2_and_dx_sup_dy:
  yinc=muldiv(dy, 65536, dx);
  ycur=y1+muldiv(yinc, (x1i<<16)-x1, 65536);
  yf=ycur&65535;
  dest=d->buffer+((ycur>>16)*SCREEN_X+x1i)*3;
  for (i=x1i; i>x2i; i--, dest-=3) {
    *dest=col1;
    *(dest+1)=col2;
    *(dest+2)=col3;
    yf-=yinc;
    if (yf>65535) {
      yf-=65536;
      dest+=SCREEN_X*3;
    }
  }
  return;
x1_sup_x2_and_y1_inf_y2_and_dx_inf_dy:
  xinc=muldiv(dx, 65536, dy);
  xcur=x1+muldiv(xinc, (y1i<<16)-y1, 65536);
  xf=xcur&65535;
  dest=d->buffer+(y1i*SCREEN_X+(xcur>>16))*3;
  for (i=y1i; i<y2i; i++, dest+=SCREEN_X*3) {
    *dest=col1;
    *(dest+1)=col2;
    *(dest+2)=col3;
    xf+=xinc;
    if (xf<0) {
      xf+=65536;
      dest-=3;
    }
  }
  return;
}

void drawline32(device *d, line *l)
{
  /* lots of goto for lisibility (if if else if if else else is impossible to debug) */
  register int x1=l->p1.x*65536., 
               x2=l->p2.x*65536., 
               y1=l->p1.y*65536., 
               y2=l->p2.y*65536.;
  register int dx, dy;
  register int xcur;
  register int xinc;
  register int ycur;
  register int yinc;
  register int i;
  register int y1i=(y1+65535)>>16, y2i=(y2+65535)>>16;
  register int x1i=(x1+65535)>>16, x2i=(x2+65535)>>16;
/*
  register int y1i=y1>>16, y2i=y2>>16;
  register int x1i=x1>>16, x2i=x2>>16;
*/ 
  register unsigned int *dest;
  register unsigned int col1=l->color;
  register int xf;
  register int yf;

  /* in asm, those two subs are good for us, because we can test if x1<x2 and if y1<y2 at the same time, but it's C */
  dx=x2-x1;
  dy=y2-y1;

  if (dx<0)
    goto x1_sup_x2;
  if (dx)
    goto x1_inf_x2;

  /* here x1==x2 */
  if (dy<0)
    goto x1_eq_x2_and_y1_sup_y2;
  if (dy>0)
    goto x1_eq_x2_and_y1_inf_y2;
  /* here x1==x2 && y1==y2 */
  if (y1i<y2i) {
    dest=(unsigned int *)d->buffer+y1i*SCREEN_X+(x1>>16);
  } else {
    dest=(unsigned int *)d->buffer+y2i*SCREEN_X+(x1>>16);
  }
  *dest=col1;
  return;
x1_eq_x2_and_y1_sup_y2:
  dest=(unsigned int *)d->buffer+y2i*SCREEN_X+(x1>>16);
  for (i=y2i; i<y1i; dest+=SCREEN_X, i++) {
    *dest=col1;
  }
  return;
x1_eq_x2_and_y1_inf_y2:
  dest=(unsigned int *)d->buffer+y1i*SCREEN_X+(x1>>16);
  for (i=y1i; i<y2i; dest+=SCREEN_X, i++) {
    *dest=col1;
  }
  return;

x1_inf_x2:
  if (dy<0)
    goto x1_inf_x2_and_y1_sup_y2;
  if (dy>0)
    goto x1_inf_x2_and_y1_inf_y2;
  /* x1<x2 && y1==y2 */
  dest=(unsigned int *)d->buffer+(y1>>16)*SCREEN_X+x1i;
  for (i=x1i; i<x2i; dest++, i++) {
    *dest=col1;
  }
  return;

x1_inf_x2_and_y1_sup_y2:
  /*
                  (x2, y2)


       (x1, y1)
   */
  /* remove those abs in the asm code */
  if (abs(dx)>abs(dy))
    goto x1_inf_x2_and_y1_sup_y2_dx_sup_dy;
  if (abs(dx)<abs(dy))
    goto x1_inf_x2_and_y1_sup_y2_dx_inf_dy;
  /* here dx == dy */
  dest=(unsigned int *)d->buffer+y2i*SCREEN_X+(x2>>16);
  for (i=y2i; i<y1i; dest+=SCREEN_X-1, i++) {
    *dest=col1;
  }
  return;
x1_inf_x2_and_y1_sup_y2_dx_sup_dy:
  /*
                    xxxx
		xxxx
	    xxxx
	xxxx
  */
  yinc=muldiv(dy, 65536, dx);
  ycur=y2+muldiv(yinc, (x2i<<16)-x2, 65536);
  yf=ycur&65535;
  dest=(unsigned int *)d->buffer+(ycur>>16)*SCREEN_X+x2i;
  for (i=x2i; i>x1i; i--, dest--) {
    *dest=col1;
    yf-=yinc;
    if (yf>65535) {
      yf-=65536;
      dest+=SCREEN_X;
    }
  }
  return;
x1_inf_x2_and_y1_sup_y2_dx_inf_dy:
  /*
          x
	  x
	  x
	 x
	 x
	 x
	x
	x
	x
  */
  xinc=muldiv(dx, 65536, dy);
  xcur=x2+muldiv(xinc, (y2i<<16)-y2, 65536);
  xf=xcur&65535;
  dest=(unsigned int *)d->buffer+y2i*SCREEN_X+(xcur>>16);
  for (i=y2i; i<y1i; i++, dest+=SCREEN_X) {
    *dest=col1;
    xf+=xinc;
    if (xf<0) {
      xf+=65536;
      dest--;
    }
  }
  return;

x1_inf_x2_and_y1_inf_y2:
  /*
      (x1, y1)


                  (x2, y2)
  */
  if (abs(dx)>abs(dy))
    goto x1_inf_x2_and_y1_inf_y2_and_dx_sup_dy;
  if (abs(dx)<abs(dy))
    goto x1_inf_x2_and_y1_inf_y2_and_dx_inf_dy;
  /* here dx == dy */
  dest=(unsigned int *)d->buffer+y1i*SCREEN_X+(x1>>16);
  for (i=y1i; i<y2i; dest+=SCREEN_X+1, i++) {
    *dest=col1;
  }
  return;

x1_inf_x2_and_y1_inf_y2_and_dx_sup_dy:
  /*
        xxxx
	    xxxx
	        xxxx
   */
  yinc=muldiv(dy, 65536, dx);
  ycur=y1+muldiv(yinc, (x1i<<16)-x1, 65536);
  yf=ycur&65535;
  dest=(unsigned int *)d->buffer+(ycur>>16)*SCREEN_X+x1i;
  for (i=x1i; i<x2i; i++, dest++) {
    *dest=col1;
    yf+=yinc;
    if (yf>65535) {
      yf-=65536;
      dest+=SCREEN_X;
    }
  }
  return;
x1_inf_x2_and_y1_inf_y2_and_dx_inf_dy:
  /*
      x
      x
      x
       x
       x
       x
        x
	x
	x
  */
  xinc=muldiv(dx, 65536, dy);
  xcur=x1+muldiv(xinc, (y1i<<16)-y1, 65536);
  xf=xcur&65535;
  dest=(unsigned int *)d->buffer+y1i*SCREEN_X+(xcur>>16);
  for (i=y1i; i<y2i; i++, dest+=SCREEN_X) {
    *dest=col1;
    xf+=xinc;
    if (xf>65536) {
      xf-=65536;
      dest++;
    }
  }
  return;

x1_sup_x2:
  if (dy<0)
    goto x1_sup_x2_and_y1_sup_y2;
  if (dy>0)
    goto x1_sup_x2_and_y1_inf_y2;
  /* x1>x2 && y1==y2 */
  dest=(unsigned int *)d->buffer+(y2>>16)*SCREEN_X+x2i;
  for (i=x2i; i<x1i; dest++, i++) {
    *dest=col1;
  }
  return;
x1_sup_x2_and_y1_sup_y2:
  /*
       (x2, y2)


                 (x1, y1)
   */
  if (abs(dx)>abs(dy))
    goto x1_sup_x2_and_y1_sup_y2_and_dx_sup_dy;
  if (abs(dx)<abs(dy))
    goto x1_sup_x2_and_y1_sup_y2_and_dx_inf_dy;
  /* here dx == dy */
  dest=(unsigned int *)d->buffer+y2i*SCREEN_X+(x2>>16);
  for (i=y2i; i<y1i; dest+=SCREEN_X+1, i++) {
    *dest=col1;
  }
  return;
x1_sup_x2_and_y1_sup_y2_and_dx_sup_dy:
  yinc=muldiv(dy, 65536, dx);
  ycur=y2+muldiv(yinc, (x2i<<16)-x2, 65536);
  yf=ycur&65535;
  dest=(unsigned int *)d->buffer+(ycur>>16)*SCREEN_X+x2i;
  for (i=x2i; i<x1i; i++, dest++) {
    *dest=col1;
    yf+=yinc;
    if (yf>65535) {
      yf-=65536;
      dest+=SCREEN_X;
    }
  }
  return;
x1_sup_x2_and_y1_sup_y2_and_dx_inf_dy:
  xinc=muldiv(dx, 65536, dy);
  xcur=x2+muldiv(xinc, (y2i<<16)-y2, 65536);
  xf=xcur&65535;
  dest=(unsigned int *)d->buffer+y2i*SCREEN_X+(xcur>>16);
  for (i=y2i; i<y1i; i++, dest+=SCREEN_X) {
    *dest=col1;
    xf+=xinc;
    if (xf>65536) {
      xf-=65536;
      dest++;
    }
  }
  return;

x1_sup_x2_and_y1_inf_y2:
  /*
                 (x1, y1)


      (x2, y2)
   */
  if (abs(dx)>abs(dy))
    goto x1_sup_x2_and_y1_inf_y2_and_dx_sup_dy;
  if (abs(dx)<abs(dy))
    goto x1_sup_x2_and_y1_inf_y2_and_dx_inf_dy;
  /* here dx == dy */
  dest=(unsigned int *)d->buffer+y1i*SCREEN_X+(x1>>16);
  for (i=y1i; i<y2i; dest+=SCREEN_X-1, i++) {
    *dest=col1;
  }
  return;
x1_sup_x2_and_y1_inf_y2_and_dx_sup_dy:
  yinc=muldiv(dy, 65536, dx);
  ycur=y1+muldiv(yinc, (x1i<<16)-x1, 65536);
  yf=ycur&65535;
  dest=(unsigned int *)d->buffer+(ycur>>16)*SCREEN_X+x1i;
  for (i=x1i; i>x2i; i--, dest--) {
    *dest=col1;
    yf-=yinc;
    if (yf>65535) {
      yf-=65536;
      dest+=SCREEN_X;
    }
  }
  return;
x1_sup_x2_and_y1_inf_y2_and_dx_inf_dy:
  xinc=muldiv(dx, 65536, dy);
  xcur=x1+muldiv(xinc, (y1i<<16)-y1, 65536);
  xf=xcur&65535;
  dest=(unsigned int *)d->buffer+y1i*SCREEN_X+(xcur>>16);
  for (i=y1i; i<y2i; i++, dest+=SCREEN_X) {
    *dest=col1;
    xf+=xinc;
    if (xf<0) {
      xf+=65536;
      dest--;
    }
  }
  return;
}

void drawline(device *d, line *l)
{
#if 0
  void fill_left(device *, point *, point *);
  extern device_line *the_lines;
  int miny, maxy;
  int minx, maxx;

  if (l->p1.y<l->p2.y) {
    miny=((int)(l->p1.y*65536.)+65535)>>16;
    maxy=((int)(l->p2.y*65536.)+65535)>>16;
  } else {
    miny=((int)(l->p2.y*65536.)+65535)>>16;
    maxy=((int)(l->p1.y*65536.)+65535)>>16;
  }

  fill_left(d, &l->p1, &l->p2);

  minx=the_lines[miny].x_left/65536+1;
  maxx=the_lines[maxy-1].x_left/65536+1;

  if (d->depth==8) {
    register int i, j;
    register int col=l->color;
    register unsigned char *p=(unsigned char *)d->buffer, *q;
    register device_line *cur_line=&the_lines[(int)miny];

    p+=miny*SCREEN_X;

    if (minx<maxx) {
      for (i=miny; i<maxy-1; i++, p+=SCREEN_X, cur_line++) {
	register const int xleft=(cur_line->x_left/65536)+1;
	register /*const */int xright=((cur_line+1)->x_left/65536);
	q=p;
	q+=xleft;
	if (xleft>xright)
	  xright++;
	for (j=xleft; j<=xright /*cur_line->x_right*/; j++, q++)
	  *q=col;
      }
    } else {
      for (i=miny; i<maxy-1; i++, p+=SCREEN_X, cur_line++) {
	register /*const */int xright=(cur_line->x_left/65536)+1;
	register const int xleft=(cur_line+1)->x_left/65536;
	q=p;
	q+=xleft;
	if (xleft>xright)
	  xright++;
	for (j=xleft; j<=xright /*cur_line->x_right*/; j++, q++)
	  *q=col;
      }
    }
  } else
    drawline16(d, l);
#endif
  if (d->depth==8)
    drawline8(d, l);
  else if (d->depth==16)
    drawline16(d, l);
  else if (d->depth==24)
    drawline24(d, l);
  else if (d->depth==32)
    drawline32(d, l);   
}

#define LEFT	1
#define RIGHT	2
#define UP	4
#define DOWN	8

static inline void copy_point(point *dest, point *src)
{
  /* this one is useless, take it away !!! */
  memcpy(dest, src, sizeof(point));
}

static inline void do_clip_left(point *res, point *p1, point *p2)
{
  register double factor=p2->x/(p2->x-p1->x);
  res->x=0;
  res->y=p2->y-(p2->y-p1->y)*factor;
}

static inline void do_clip_right(point *res, point *p1, point *p2)
{
  register double factor=(p2->x-SCREEN_X+2)/(p2->x-p1->x);
  res->x=SCREEN_X-2;
  res->y=p2->y-(p2->y-p1->y)*factor;
}

static inline void do_clip_up(point *res, point *p1, point *p2)
{
  register double factor=p2->y/(p2->y-p1->y);
  res->x=p2->x-(p2->x-p1->x)*factor;
  res->y=0;
}

static inline void do_clip_down(point *res, point *p1, point *p2)
{
  register double factor=(p2->y-(SCREEN_Y-2))/(p2->y-p1->y);
  res->x=p2->x-(p2->x-p1->x)*factor;
  res->y=SCREEN_Y-2;
}

void clip_and_draw_line(device *d, line *l)
{
  register int clip=0;
  register double a, b;
  register point *pp, *pp1;
  point next;

  /* step 1 - do we clip ? */
  a=l->p1.x;
  b=l->p1.y;
  if (a<0)
    clip|=LEFT;
  if (a>SCREEN_X-2)
      clip|=RIGHT;
  if (b<0)
    clip|=UP;
  if (b>SCREEN_Y-2)
    clip|=DOWN;
  a=l->p2.x;
  b=l->p2.y;
  if (a<0)
    clip|=LEFT;
  if (a>SCREEN_X-2)
      clip|=RIGHT;
  if (b<0)
    clip|=UP;
  if (b>SCREEN_Y-2)
    clip|=DOWN;

  pp=&l->p1;
  pp1=&l->p2;

  /* step 2 - clip left ? */
  if (!(clip&LEFT))
    goto no_left;
  if (pp->x>0 && pp1->x<0) {
    do_clip_left(&next, pp, pp1);
    copy_point(pp1, &next);
  } else if (pp->x<0 && pp1->x>0) {
    do_clip_left(&next, pp, pp1);
    copy_point(pp, &next);
  }

no_left:
  /* step 3 - right ? */
  if (!(clip&RIGHT))
    goto no_right;
  if (pp->x<SCREEN_X-2 && pp1->x>SCREEN_X-2) {
    do_clip_right(&next, pp, pp1);
    copy_point(pp1, &next);
  } else if (pp->x>SCREEN_X-2 && pp1->x<SCREEN_X-2) {
    do_clip_right(&next, pp, pp1);
    copy_point(pp, &next);
  }

no_right:
  /* step 4 - up ? */
  if (!(clip&UP))
    goto no_up;
  if (pp->y>0 && pp1->y<0) {
    do_clip_up(&next, pp, pp1);
    copy_point(pp1, &next);
  } else if (pp->y<0 && pp1->y>0) {
    do_clip_up(&next, pp, pp1);
    copy_point(pp, &next);
  }

no_up:
  /* step 5 - down ? */
  if (!(clip&DOWN))
    goto no_down;
  if (pp->y<SCREEN_Y-2 && pp1->y>SCREEN_Y-2) {
    do_clip_down(&next, pp, pp1);
    copy_point(pp1, &next);
  } else if (pp->y>SCREEN_Y-2 && pp1->y<SCREEN_Y-2) {
    do_clip_down(&next, pp, pp1);
    copy_point(pp, &next);
  }

no_down:
  /* step 6 - draw the resulting line (clipping may have outed it all, so let's test it) */
  if (pp->x>=0 && pp->x<=SCREEN_X-2 && pp->y>=0 && pp->y<=SCREEN_Y-2
      && pp1->x>=0 && pp1->x<=SCREEN_X-2 && pp1->y>=0 && pp1->y<=SCREEN_Y-2)
    drawline(d, l);
#if 0
  else
    fprintf(stderr, "clip=%d p1=%f %f, p2=%f %f\n", clip, pp->x, pp->y, pp1->x, pp1->y);
#endif
}