/**************************************************************************
 *                                                                        *
 *               Copyright (C) 1994, Silicon Graphics, Inc.               *
 *                                                                        *
 *  These coded instructions, statements, and computer programs  contain  *
 *  unpublished  proprietary  information of Silicon Graphics, Inc., and  *
 *  are protected by Federal copyright law.  They  may  not be disclosed  *
 *  to  third  parties  or copied or duplicated in any form, in whole or  *
 *  in part, without the prior written consent of Silicon Graphics, Inc.  *
 *                                                                        *
 **************************************************************************/

/*
 *  Edge Walker Unit
 *
 *
 *  RJM 8/13/94
 */

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

#include "ints.h"
#include "ew.h"


#define POSEDGE         (save_clk && !save_clk_old)
#define NEGEDGE         (!save_clk && save_clk_old)


#ifndef MAX
#define MAX(x, y)       (((x) > (y)) ? (x) : (y))
#endif

#ifndef MIN
#define MIN(x, y)       (((x) < (y)) ? (x) : (y))
#endif

#define SIGN_EXTEND_27(x)	(((x) & 0x4000000) ? (x) | ~0x3ffffff : (x))
#define SIGN_EXTEND_12(x)	(((x) & 0x800) ? (x) | ~0xfff : (x))
#define SIGN_EXTEND_14(x)	(((x) & 0x2000) ? (x) | ~0x3fff : (x))


/*
 *    G l o b a l s
 */

/* Debugging/checking */
static int ew_dump;

/* edge walker memory */
static unsigned int EwMemoryA[9]; /* attribute */
static unsigned int EwMemoryB[9]; /* da/de */

static int Trigger = 0;

/* =========================================================================
 *  An 18-word by 32-bit register file.  Two writes at the same address
 *  at the same time is not allowed.  Read and write to the same address at 
 *  the same time is ?not allowed?  The da/de memory can write data directly
 *  to the read register.
 * =========================================================================
 */
static void
  ew_reg_file( ew_t *p0, ew_t *p1, 
	       int ld_a, int attr_adrs, int attr_we, 
               int de_adrs, int de_we,
               int64 cs_ew_d)
{
  if(attr_we)
    EwMemoryA[attr_adrs] = ld_a ? cs_ew_d.word1 : p1->sum_att;

  if(de_we)
    EwMemoryB[de_adrs] = cs_ew_d.word1;

  if((p1->cnt_span_attr == de_adrs) && de_we)
    p0->rb_data = p1->ew_stall_attr ? p1->rb_data : cs_ew_d.word1;
  else
    p0->rb_data = p1->ew_stall_attr ? p1->rb_data : EwMemoryB[p1->cnt_span_attr];

  p0->ra_data = p1->ew_stall_attr ? p1->ra_data : EwMemoryA[p1->cnt_span_attr];
}

/* =========================================================================
 *  ew_attr_adder - two 16-bit adders that step attributes along attribute
 *  edge.  This function includes the sum register.
 *  
 * =========================================================================
 */
static void
  ew_attr_adder(ew_t *p0, ew_t *p1)
{
  unsigned int in_attr, in_dade;
  unsigned int sum_ls, sum_ms;
  unsigned int ms_ci, ls_ci;
  unsigned int ms_co, ls_co;

  in_attr = p1->ra_data;
  in_dade = p1->rb_data;

  /* do LS adder */
  ls_ci  = p1->ls_co && !p1->add_clear;
  sum_ls = (in_attr & 0xffff) + (in_dade & 0xffff) + ls_ci;
  ls_co  = (sum_ls & 0x10000) >> 16;
  p0->ls_co = p1->ew_stall_attr ? p1->ls_co : ls_co;
  
  /* do MS adder */
  ms_ci = p1->add32b ? ls_co : p1->ms_co && !p1->add_clear;
  sum_ms = ((in_attr >> 16) & 0xffff) + ((in_dade >> 16) & 0xffff) + ms_ci; 
  ms_co = (sum_ms & 0x10000) >> 16;
  p0->ms_co = p1->ew_stall_attr ? p1->ms_co : ms_co;

  /* save sum */
  p0->sum_att = p1->ew_stall_attr ? p1->sum_att : ((sum_ms & 0xffff) << 16) | (sum_ls & 0xffff);
}


/* =========================================================================
 *  ew_shuffle - swaps integer and fraction to make complete word (16.16)
 *   
 * =========================================================================
 */
static void
  ew_shuffle(ew_t *p0, ew_t *p1)
{
  int int_shuffle, frac_shuffle;
  int result;

  p0->inta_1delay = (p1->ra_data >> 16) & 0xffff;
  p0->fraca_1delay = p1->ra_data & 0xffff;
  p0->fraca_2delay = p1->fraca_1delay;

  p0->intd_1delay = (p1->rb_data >> 16) & 0xffff;
  p0->fracd_1delay = p1->rb_data & 0xffff;
  p0->fracd_2delay = p1->fracd_1delay;

  int_shuffle = p1->shuffle ? p1->fraca_1delay : (p1->ra_data >> 16) & 0xffff;
  frac_shuffle = p1->shuffle ? p1->fraca_2delay : p1->inta_1delay;
  result = (p1->noshuffle ? p1->inta_1delay : int_shuffle) << 16;
  result |= p1->noshuffle ? p1->fraca_1delay : frac_shuffle;
  p0->att_out_s = result;
  p0->att_out   = p1->att_out_s >> 9; /* reduce to s,15.7 */

  int_shuffle = p1->shuffle ? p1->fracd_1delay : (p1->rb_data >> 16) & 0xffff;
  frac_shuffle = p1->shuffle ? p1->fracd_2delay : p1->intd_1delay;
  result = (p1->noshuffle ? p1->intd_1delay : int_shuffle) << 16;
  result |= p1->noshuffle ? p1->fracd_1delay : frac_shuffle;
  p0->dade_out_s = result;
  p0->dade_out   = p1->dade_out_s >> 9; /* reduce to s,15.7 */
}


/* ======================================================================
 *  ew_offset():
 *
 *  dade_out, att_out, dx_att, dy_att - s,15.7
 *  xfrac - 0.8
 *   
 *  s,15.7 * .8 = s,15.15
 *  
 *  Offset attribute as follows: A = A + 3/4(da/de - da/dy) - dx * xfrac;
 *  
 *
 * ======================================================================
 */

#define SIGN_EXTEND_22(x)	(((x) & 0x400000) ? ((x) | ~0x7fffff) : (x))

static void
    ew_offset(ew_t *p0, ew_t *p1, int dx_att, int dy_att)
{
  int x_frac_ss, do_offset, de_offset, dy_offset;
  int att;

  p0->x_frac_s = p1->ld_x_frac ? p1->x_frac : p1->x_frac_s;

  x_frac_ss = p1->x_frac_s & (!(p1->cycle_type == 2 || p1->load_cmd_offset) ? 0xff : 0x00);
  
  do_offset = !(p1->sign_dxhdy_offset ^ p1->left_offset);
  de_offset =  do_offset ? p1->dade_out : 0x0;
  dy_offset =  do_offset ? dy_att : 0x0;

  de_offset = SIGN_EXTEND_22(de_offset);
  dy_offset = SIGN_EXTEND_22(dy_offset);
  att       = SIGN_EXTEND_22(p1->att_out);
  dx_att    = SIGN_EXTEND_22(dx_att);

  /* result is s,15.15 */
  p0->att_d_out = (att << 8) + 
	 (de_offset << 7) + (de_offset << 6) -
	 (dy_offset << 7) - (dy_offset << 6) -
	 (dx_att * x_frac_ss);
#ifdef DUMB
  if(p1->cnt_span_attr == 5 || p1->cnt_span_attr == 6 || p1->cnt_span_attr == 7
       || p1->cnt_span_attr == 8)
  {
    printf("cnt_attr: %d  ew_ep_d = %08x  att_out = %08x att = %08x, dx_att = %08x, de_off = %08x, de_y = %08x, x_frac = %02x\n", p1->cnt_span_attr, p0->ew_ep_d, p0->att_d_out, att, dx_att, de_offset, dy_offset, x_frac_ss);
  }
#endif
  p0->ew_ep_d = (p1->att_d_out >> 9) & 0x3fffff; /* s,15.6 */

}


/* ======================================================================
 *  ew_x_calc():
 *
 *  Compute xh0, xh1, xh2, xh3, xm0, xm1, xm2, xm3
 *  for each span in a triangle or rectangle.
 *
 * ======================================================================
 */

static void
    ew_x_calc(ew_t *p0, ew_t *p1, int ld_xmh, int switch_xl, int ld_dxmdy, 
           int ld_dxldy, int ld_dxhdy, int64 cs_ew_d)
{
  int dxmdy, dxhdy, xm, xh, sum;

  dxmdy = ld_xmh ? ((cs_ew_d.word0 >> 3) &  0x7ffffff) : p1->dxhdy;
  dxhdy = switch_xl ? p1->dxldy : p1->dxmdy;

  sum = SIGN_EXTEND_27(p1->dxhdy) + SIGN_EXTEND_27(p1->xh); 

  xm = ld_xmh ? (cs_ew_d.word1 >> 1) & 0x7ffffff : sum & 0x7ffffff;
  xh = switch_xl ? p1->xl : p1->xm;

  p0->dxmdy = ld_dxmdy ? dxmdy : p1->dxmdy;
  p0->dxldy = ld_dxldy ? (cs_ew_d.word0 >> 3) & 0x7ffffff : p1->dxldy;
  p0->dxhdy = ld_dxhdy ? dxhdy : p1->dxhdy;
  p0->xm = ld_dxmdy ? xm : p1->xm;
  p0->xl = ld_dxldy ? (cs_ew_d.word1 >> 1) & 0x7ffffff : p1->xl;
  p0->xh = ld_dxhdy ? xh : p1->xh;
  p0->x_sticky = (p0->xh & 0x1fff) > 0;

}


/* ======================================================================
 *  ew_scissor_y(): counts and scissors y coordinate
 *
 * ======================================================================
 */
static void
    ew_scissor_y(ew_t *p0, ew_t *p1, int ld_y, int count_y, int *end_prim_y, int64 cs_ew_d)
{
  int yl_ismin_scymax;
  int min_yl_scymax;
  int yh_ismax;
  int max_yh_scmin;
  int xval0, xval1, xval2, xval3;
  int x_invalid;

  p0->ym = ld_y ? (cs_ew_d.word0 >> 16) & 0x3fff : p1->ym;
  p0->yl = ld_y ? cs_ew_d.word1 & 0x3fff : p1->yl;
  p0->yh = ld_y ? cs_ew_d.word0 & 0x3fff : p1->yh;

  /* increment y counter */
  p0->y_cur = ld_y ? cs_ew_d.word0 & 0x3ffc : p1->y_cur + count_y;

  /* min(yl, scbox_ymax) to determine end of primitive */
  if(p1->yl & 0x2000)
    yl_ismin_scymax = 1;
  else if(p1->yl & 0x1000)
    yl_ismin_scymax = 0;
  else
    yl_ismin_scymax = (p1->yl & 0xfff) < p1->scbox_ymax;

  /* Terminate primitive if current y >= min(yl, scbox_ymax) */
  min_yl_scymax = (yl_ismin_scymax || p1->load_cmd_scissor) ? p1->yl : p1->scbox_ymax & 0xfff;
  *end_prim_y = (SIGN_EXTEND_14(p1->y_cur) >= SIGN_EXTEND_14(min_yl_scymax)) || p1->flush;

  /* cancel the span if current y < max(scbox_ymin, yh) */
  if(p1->yh & 0x2000)
    yh_ismax = 0;
  else if (p1->yh & 0x1000)
    yh_ismax = 1;
  else
    yh_ismax = p1->yh >= p1->scbox_ymin;

  max_yh_scmin = (yh_ismax || p1->load_cmd_scissor) ? p1->yh : p1->scbox_ymin & 0xfff;
  x_invalid = SIGN_EXTEND_14(p1->y_cur) < SIGN_EXTEND_14(max_yh_scmin);
  p0->y_invalid = x_invalid || (*end_prim_y); /* used by ew controlto clear x_minor */

  /* get xval */
  xval0 = p1->xval & 1;
  xval1 = (p1->xval >> 1) & 1;
  xval2 = (p1->xval >> 2) & 1;
  xval3 = (p1->xval >> 3) & 1;
 
  p0->xval = (p1->cnt_span_x & 1) ? !((*end_prim_y) || x_invalid) : xval0;
  p0->xval |= ((p1->cnt_span_x & 1) ? xval0 : xval1) << 1;
  p0->xval |= ((p1->cnt_span_x & 1) ? xval1 : xval2) << 2;
  p0->xval |= ((p1->cnt_span_x & 1) ? xval2 : xval3) << 3;
}


/* ======================================================================
 *  ew_scissor_x():
 *
 *  Scissor X coordinates
 * 
 *  Scissored x, x_sc, should only use 13 LSBs outside this function.
 * ======================================================================
 */

static void
    ew_scissor_x(ew_t *p0, ew_t *p1, int ld_xmajor, int clear_allxgemax)
{
  int x_unsc;			/* s,11.8 */
  int less_xmin; 
  int xsc_scmin; 		/*   11.2,sticky */
  int allxlmin;
  int ge_xmax; 
  int xsc_scmax;
  int allxgemax;
  int x_sc_maxnew; 
  int x_sc_max;
  int x_sc_minnew; 
  int x_sc_min;
  int xsc;
  int xsc_m;

  /*
   *  scissor with scbox_xmin, max(scbox_xmin, x_in)
   *  p1->xh is s,11.15
   */
  x_unsc = (p1->xh >> 7) & 0xfffff;
 
  if(x_unsc & 0x80000)  /* x negative, out of scissor box */
    less_xmin = 1;
  else if(x_unsc & 0x40000) /* x out of maximum positive scissor box */
    less_xmin = 0;
  else  /* check against scissor box min (10.2) */
    less_xmin = (((x_unsc & 0x3ffc0) >> 5) | p1->x_sticky) < (p1->scbox_xmin << 1);

  xsc_scmin = less_xmin ? (p1->scbox_xmin << 1) & 0x1ffe : 
	(((x_unsc & 0x7ffc0) >> 5) | p1->x_sticky);
  allxlmin = (p1->allxlmin || clear_allxgemax) && less_xmin && !p1->load_cmd_scissor;

  /*
   *  scissor with scbox_xmax, min(scbox_xmax, xsc_scmin)
   *  know that xsc_scmin is positive, 11.2,sticky.
   */
  if(xsc_scmin & 0x2000)
    ge_xmax = 1;
  else
    ge_xmax = (xsc_scmin & 0x1fff) >= (p1->scbox_xmax << 1);

  xsc_scmax = ge_xmax ? (p1->scbox_xmax << 1) & 0x1ffe : xsc_scmin;
  allxgemax = (p1->allxgemax || clear_allxgemax) && ge_xmax && !p1->load_cmd_scissor;

  xsc = p1->load_cmd_scissor ? ((x_unsc & 0xfffc0) >> 5) : xsc_scmax;
  /* xsc_m = (p1->cnt_span_x & 8) ? p1->xval & 0xf : xsc; */
  
  p0->allxlmin = p1->ew_stall_x ? p1->allxlmin : allxlmin;
  p0->allxgemax = p1->ew_stall_x ? p1->allxgemax : allxgemax;

  p0->x_sc = p1->ew_stall_x ? p1->x_sc : xsc;
  p0->ew_cv_d = p1->sel_xval ? p1->xval & 0xf : p1->x_sc & 0x1fff;

  if (ew_dump&0x1) {
    static int dump_num=0,dump_last=0,dump_stall=0,dump_yval;
    static int dump_x[8] = 
	{ 0xDEAD, 0xDEAD, 0xDEAD, 0xDEAD, 0xDEAD, 0xDEAD, 0xDEAD, 0xDEAD };
    dump_num++;
    if (p1->cnt_span_x >0) {
      dump_x[p1->cnt_span_x-1]=p0->ew_cv_d;
      dump_last=0;
      if (p1->cnt_span_x==1) {
        dump_stall++;
        dump_yval=p1->y_cur;
	if (dump_yval & 0x2000) dump_yval |= ~0x3FFF;
      }
    } else {
      if (!dump_last) {
        if (!p1->sel_xval) 
          printf("ew: {FUNNY:sel_xval!=0 when cnt_span==0}");
        printf("ew[%d]: ",dump_num);
	if (dump_stall>2) printf("S ");
	else printf("n ");
	dump_stall=0;
        printf("ycur=%04i.%X   ",dump_yval>>2, (dump_yval&0x3)<<2);
        printf("x0h-x1m:  ");
        printf("%04i(%i)_%04i(%i)  ",(int) (dump_x[0]>>3),(int) (dump_x[0]&7),(int) (dump_x[1]>>3),(int) (dump_x[1]&7));
        printf("%04i(%i)_%04i(%i)  ",(int) (dump_x[2]>>3),(int) (dump_x[2]&7),(int) (dump_x[3]>>3),(int) (dump_x[3]&7));
        printf("%04i(%i)_%04i(%i)  ",(int) (dump_x[4]>>3),(int) (dump_x[4]&7),(int) (dump_x[5]>>3),(int) (dump_x[5]&7));
        printf("%04i(%i)_%04i(%i)  ",(int) (dump_x[6]>>3),(int) (dump_x[6]&7),(int) (dump_x[7]>>3),(int) (dump_x[7]&7));
        printf("val=%X  ",p0->ew_cv_d);
        printf("\n");
        dump_last=1;
      }
    }
  }

  p0->x_major_1d = ld_xmajor ? x_unsc : p1->x_major_1d;
  p0->ew_cv_start_x = (p0->x_major_1d >> 8) & 0xfff; /* no frac bits needed */
  p0->x_frac_m = p1->x_major_1d & 0xff;
  p0->x_frac_mm = p1->x_frac_m;
  p0->x_frac = p1->x_frac_mm;

  /*
   *  Find the scissored x_max(x_right) x_min(x_left)
   *  These values are used in the memory span unit to calculate span address
   */
  if(p1->reset_l == 0)
  {
    p0->x_sc_max = 0;
    p0->x_sc_min = 0;
  }
  else
  {
    /* Compute x_max with no fraction bits */
    x_sc_maxnew = !p1->clear_xminor ? p1->x_sc_max : 0;
    x_sc_max = (((p1->x_sc >> 3) & 0xfff) >= x_sc_maxnew) && !p1->y_invalid ? 
        (p1->x_sc >> 3) & 0xfff : x_sc_maxnew;
    p0->x_sc_max = p1->ew_stall_x ? p1->x_sc_max : x_sc_max;

    /* compute x_min with no fraction bits */
    x_sc_minnew = p1->clear_xminor ? 0xfff : p1->x_sc_min;
    x_sc_min = (((p1->x_sc >> 3) & 0xfff) < x_sc_minnew) && !p1->y_invalid ? 
       (p1->x_sc >> 3) & 0xfff : x_sc_minnew;
    p0->x_sc_min = p1->ew_stall_x ? p1->x_sc_min : x_sc_min;
  } 

  /* mux for x_sc_major, x_minor not implemented here, were used? */ 
}


/* ======================================================================
 *  ew_ms_adrs():
 *
 *  Calculate addresses for span, to mem span unit
 * 
 *  
 * ======================================================================
 */
static void
  ew_ms_adrs( ew_t *p0, ew_t *p1)
{
  int mult;
  int add_zero;
  int sel_mult;
  int ld_addr;
  int sel_addr;
  int yh_cur;

  int a_adder;
  int b_adder_m;
  int b_adder;
  int sum;
  int a_adder_m;
  int b_adder_a;

  int posy_sel;
  int ycur;

  ycur = p1->y_cur >> 2; /* s,11.0 */

  if(p1->reset_l == 0)
  {
    p0->cnt_mult = 0;
    p0->cnt_addr = 0;
    p0->posy_state = 0;
  }
  else
  {
    p0->cnt_mult = (p1->cnt_prim == 2) || !((p1->cnt_mult == 0xa) || (p1->cnt_mult == 0x0)) ?
       p1->cnt_mult + 1 : 0;
    p0->cnt_addr = (((p1->cnt_span_x == 8)) || ((p1->cnt_addr & 0x3) != 0)) ? p1->cnt_addr + 1 : 0;

    posy_sel = ((p1->cnt_prim == 2) << 3) | ((p1->cnt_span_x == 8) << 2) | ((!((ycur & 0x800) ||
	(ycur & 0x400))) << 1) | p1->posy_state;

    switch(posy_sel)
    {
      case 0x0:
      case 0x2:
      case 0x4:
      case 0x8:
      case 0x9:
      case 0xc:
      case 0xd: p0->posy_state = 0; break;
      default:  p0->posy_state = 1; break;
    }

  }

  switch(p1->cnt_mult)
  {
     case 0x1: yh_cur = ycur & 0x200; break;
     case 0x2: yh_cur = ycur & 0x100; break;
     case 0x3: yh_cur = ycur & 0x080; break;
     case 0x4: yh_cur = ycur & 0x040; break;
     case 0x5: yh_cur = ycur & 0x020; break;
     case 0x6: yh_cur = ycur & 0x010; break;
     case 0x7: yh_cur = ycur & 0x008; break;
     case 0x8: yh_cur = ycur & 0x004; break;
     case 0x9: yh_cur = ycur & 0x002; break;
     case 0xa: yh_cur = ycur & 0x001; break;
     default:  yh_cur = ycur & 0x001; break;
  }

  mult = (p1->cnt_mult != 0);
  add_zero = (ycur & 0x800) || (ycur & 0x400) || !yh_cur;
  sel_mult = (p1->cnt_mult != 0x0) && (p1->cnt_mult != 0x1);
  ld_addr = (p1->cnt_mult != 0x0) || (p1->cnt_addr == 0x3);
  sel_addr = p1->posy_state ? p1->cnt_addr : 0;


  if(p1->reset_l == 0)
  {
    p0->addr_scanline = 0;
  }
  else
  {
    a_adder_m = add_zero ? 0 : (p1->load_cmd_image ? p1->width_timage : p1->width_cimage);
    a_adder   = mult ? a_adder_m : p1->addr_scanline;
    b_adder_m = sel_mult ? (p1->addr_scanline & 0x7ffff) << 1 : 0;
    switch (sel_addr)
    {
         case 0: b_adder_a = 0; break;
         case 1: b_adder_a = p1->left_xminor ? p1->x_sc_min : p1->x_sc_max; break;
         case 2: b_adder_a = p1->left_xminor ? p1->x_sc_max : p1->x_sc_min; break;
         case 3: b_adder_a = (p1->load_cmd_image ? p1->width_timage : p1->width_cimage); break;
    }

    b_adder = mult ? b_adder_m : b_adder_a;
    sum = a_adder + b_adder + (1 & ((mult && !add_zero) || (p1->posy_state && ((p1->cnt_addr == 3) && !mult))));

    p0->addr_scanline = ld_addr  ? sum : p1->addr_scanline;
    p0->ew_ms_addr    = sum;
  }

#ifdef DEBUG
  {
    static ew_debug_count=0;
    fprintf(stderr,"ew[%d]: addr_scanline=0x%X ew_ms_addr=0x%X sum=%X a_adder=%X b_adder=%X mult=%X add_zero=%X sel_addr=%X cnt_mult=%X cnt_addr=%X ycur=%X\n",
        ew_debug_count++,p0->addr_scanline,p0->ew_ms_addr,sum,a_adder,b_adder,mult,add_zero,sel_addr,p1->cnt_mult,p1->cnt_addr,ycur);
    fprintf(stderr,"   cnt_prim=%X  cnt_span_x=%X  posy_state%X\n",p1->cnt_prim,p1->cnt_span_x,p1->posy_state);
  }
#endif /* DEBUG */

}




/* ======================================================================
 *  ew():  main interface for the edge walker unit.
 *
 *  WARNING:  This code is order dependent.  Be careful when rearranging
 *  code sections.
 * ======================================================================
 */
void
  ew(ew_t **pp0, ew_t **pp1)
{
  ew_t *p0, *p1;
  int save_clk, save_clk_old;

  /* intermediate variables */
  int ld_y;
  int count_y;
  int end_prim_y;

  int span_valid;
 
  int ld_xmh;
  int ld_dxmdy;
  int ld_dxhdy;
  int ld_dxldy;
  int switch_xl;
  int end_x;
 
  int ld_xmajor;
  int clear_allxgemax;

  int ld_a;
  int attr_adrs; 
  int attr_we;
  int de_adrs;
  int de_we;

  int dx_att;
  int dy_att;

  int get_ew_stall;
  int larger_x;
  int smaller_x;
  int num_shift;
  int num_pixel;
  int num_cycles_new;
  int num_cycles;
  int comp_value;
  int stall_decr;

  /* 
   *  These inputs come from pins which have logic between the output register
   *  and the pin.  So we re-grab these inputs on the non-used (negative) edge
   *  of the clock, and recalculate the module.
   */
  int64 cs_ew_d;
  int   cs_ew_newprim;

  /* get initial pointers */
  p0 = *pp0;
  p1 = *pp1;
  save_clk = p0->gclk;
  save_clk_old = p1->gclk_old;

  
  if(NEGEDGE)
  {
    cs_ew_d = p0->cs_ew_d;
    cs_ew_newprim = p0->cs_ew_newprim;
  }

  if(POSEDGE)
  {
    /* transfer all next-clock register values to register outputs. */
    *pp0 = p1; /* swap */
    *pp1 = p0;
    p0 = *pp0; /* fix pointers */
    p1 = *pp1;
  }

  if(POSEDGE || NEGEDGE)
  {
    /* Update all next-clock register values */


    /* controls for Y's */
    ld_y = p1->cnt_prim == 1;
    count_y = ((p1->cnt_span_x == 1) || (p1->cnt_span_x == 3) ||
                (p1->cnt_span_x == 5) || (p1->cnt_span_x == 8)) &&
	      !( (p1->cnt_prim == 1) || (p1->cnt_prim == 2) ||
		 (p1->cnt_prim == 3) || (p1->cnt_prim == 4)) &&
              !p1->ew_stall_x && p1->span_valid_x;

    /*
     *   Scissor Y coordinates
     *
     */

    ew_scissor_y(p0, p1, ld_y, count_y, &end_prim_y, cs_ew_d);

    /*
     *   Edge Walker Control, derive control signals before
     *   calling other units.
     */
    if(p1->reset_l == 0)
    {
      p0->cnt_prim = 0;
      p0->cnt_span_x = 0;
      p0->cnt_span_attr = 0;
      p0->end_ew = 0;
      p0->span_valid_s = 0;
      p0->span_valid_x = 0;
    }
    else
    {
      if(cs_ew_newprim && (p1->cnt_prim == 0x16))
        p0->cnt_prim = 1;
      else
        p0->cnt_prim = (!cs_ew_newprim) ? 0 : (p1->cnt_prim + cs_ew_newprim);

      p0->cnt_span_x = ((p1->cnt_span_x == 0x8) || (p1->cnt_prim == 0x4)) ? 0 :
	  (p1->cnt_span_x + (!p1->ew_stall_x));
      p0->cnt_span_attr = ((p1->cnt_span_attr == 0x8) || (p1->cnt_prim == 0xd)) ? 0 :
	  (p1->cnt_span_attr + (!p1->ew_stall_attr));

      p0->end_ew = end_prim_y && (p1->cnt_span_x == 8);

      p0->span_valid_x = (p1->cnt_prim == 0x4) || (p1->span_valid_x && 
			   !(end_prim_y && (p1->cnt_span_x == 0x8)));
      p0->span_valid_x_s = p1->span_valid_x;
      p0->sc_field_s = (p1->cnt_prim == 0xc) ? p1->sc_field : p1->sc_field_s;
      p0->odd_line_s = (p1->cnt_prim == 0xc) ? p1->odd_line : p1->odd_line_s;
    }

    /* general control */
    span_valid = p1->span_valid_x && !p1->flush && (p1->xval != 0) && 
	  !p1->allxlmin && !p1->allxgemax &&
	  (!p1->sc_field_s || (p1->load_cmd_scissor & 1) || (p1->sc_field_s && 
	    !(((p1->y_cur & 0x4)>>2) ^ p1->odd_line_s)));


    if(p1->reset_l)
      p0->span_valid_s = span_valid;
    p0->ew_cv_newspan = (p1->cnt_span_x == 8) && span_valid;
    p0->ew_ep_startspan_m = p1->ew_cv_newspan;
    p0->ew_ep_startspan = p1->ew_ep_startspan_m;

    /* controls for X's */
    ld_xmh = (p1->cnt_prim == 3) || (p1->cnt_prim == 4);
    ld_dxmdy = (!(p1->cnt_span_x == 0x7) && !p1->ew_stall_x) || ld_xmh;
    ld_dxhdy = ld_dxmdy;

    /**** DEBUG ****/
    p0->ld_dxhdy = ld_dxhdy;

    ld_dxldy = (p1->cnt_prim == 2);
    switch_xl = (p1->y_cur == p1->ym) && !(p1->cnt_span_x & 1) && !(p1->cnt_span_x & 8) && !ld_xmh;
    end_x = (p1->cnt_span_x == 8);

    /*
     *   EW X coordinate Calculation.  Computes four X values, one for each
     *   subpixel Y increment, for the left and right sides of a span.
     */

    ew_x_calc(p0, p1, ld_xmh, switch_xl, ld_dxmdy, ld_dxldy, ld_dxhdy, cs_ew_d);

    /* controls for X scissor */
    ld_xmajor = (p1->left_xmajor ^ p1->sign_dxhdy_xmajor) ? (p1->cnt_span_x == 0) : (p1->cnt_span_x == 6);
    clear_allxgemax = (p1->cnt_span_x == 0);
    p0->clear_xminor = p1->ew_stall_x ? p1->clear_xminor : (p1->cnt_span_x == 0);

    /*
     *   Scissor X coordinates
     *
     */

    ew_scissor_x(p0, p1, ld_xmajor, clear_allxgemax);

    /*
     *  Register control signals 
     */
    if(p1->reset_l == 0)
    {
      /* generate ew_stall_attr */
      p0->ew_stall_attr0d = 0;
      p0->ew_stall_attr1d = 0;
      p0->ew_stall_attr2d = 0;
      p0->ew_stall_attr3d = 0;
      p0->ew_stall_attr4d = 0;
      p0->ew_stall_attr5d = 0;
      p0->ew_stall_attr6d = 0;
      p0->ew_stall_attr7d = 0;
      p0->ew_stall_attr8d = 0;
      p0->ew_stall_attr   = 0;
      p0->sel_xval = 0;
    }
    else /* no reset_l */
    {
      p0->sel_xval = (p1->cnt_span_x & 0x8) == 0x8;
      p0->wa_addr_m = p1->ew_stall_attr ? p1->wa_addr_m : p1->cnt_span_attr;
      p0->wa_addr_s = p1->ew_stall_attr ? p1->wa_addr_s : p1->wa_addr_m;
      p0->add_clear = p1->ew_stall_attr ? p1->add_clear : 
	!(p1->cnt_span_attr & 1);
      p0->add32b = p1->ew_stall_attr ? p1->add32b : (p1->cnt_span_attr & 8) == 8;
      p0->shuffle_m = p1->ew_stall_attr ? p1->shuffle_m : p1->cnt_span_attr & 1;
      p0->shuffle = p1->shuffle_m;
      p0->noshuffle_m = p1->ew_stall_attr ? p1->noshuffle_m : (p1->cnt_span_attr & 8) == 8;
      p0->noshuffle = p1->noshuffle_m;
      p0->sel_dydx_m = p1->ew_stall_attr ? p1->sel_dydx_m : p1->cnt_span_attr;
      p0->sel_dydx_mm = p1->sel_dydx_m;
      p0->sel_dydx_mmm = p1->sel_dydx_mm;
      p0->sel_dydx      = p1->sel_dydx_mmm;
      p0->ld_x_frac_m = p1->cnt_span_x == 8;
      p0->ld_x_frac_mm = p1->ld_x_frac_m;
      p0->ld_x_frac_mmm = p1->ld_x_frac_mm;
      p0->ld_x_frac      = p1->ld_x_frac_mmm;
  
      /* generate ew_stall_attr */
      p0->ew_stall_attr0d = p1->ew_stall_x;
      p0->ew_stall_attr1d = p1->ew_stall_attr0d;
      p0->ew_stall_attr2d = p1->ew_stall_attr1d;
      p0->ew_stall_attr3d = p1->ew_stall_attr2d;
      p0->ew_stall_attr4d = p1->ew_stall_attr3d;
      p0->ew_stall_attr5d = p1->ew_stall_attr4d;
      p0->ew_stall_attr6d = p1->ew_stall_attr5d;
      p0->ew_stall_attr7d = p1->ew_stall_attr6d;
      p0->ew_stall_attr8d = p1->ew_stall_attr7d;
      p0->ew_stall_attr = p1->ew_stall_attr7d;
    }

    /* ew offset control */
    ld_a = (p1->cnt_prim == 0x5) || (p1->cnt_prim == 0x6) ||
	   (p1->cnt_prim == 0x7) || (p1->cnt_prim == 0x8) ||
	   (p1->cnt_prim == 0x9) || (p1->cnt_prim == 0xa) || 
           (p1->cnt_prim == 0xb) || (p1->cnt_prim == 0xc) ||
	   (p1->cnt_prim == 0xd);


    attr_adrs = ld_a ? (p1->cnt_prim - 0x5) & 0xf : p1->wa_addr_s;
    de_adrs   = (p1->cnt_prim - 0xe) & 0xf;
    de_we     = (p1->cnt_prim == 0xe)  || (p1->cnt_prim == 0xf)  ||
		(p1->cnt_prim == 0x10) || (p1->cnt_prim == 0x11) ||
		(p1->cnt_prim == 0x12) || (p1->cnt_prim == 0x13) ||
		(p1->cnt_prim == 0x14) || (p1->cnt_prim == 0x15) ||
		(p1->cnt_prim == 0x16);
    attr_we   = !((p1->cnt_prim == 0xe) || (p1->cnt_prim == 0xf)) &&
		 (!p1->ew_stall_attr);
    
    /*
     *  EW Attribute Offset Data Path
     */
    ew_reg_file( p0, p1, ld_a, attr_adrs, attr_we, de_adrs, de_we, cs_ew_d );
    
    ew_attr_adder( p0, p1 );

    ew_shuffle( p0, p1 );


    /* select dx/dy attribute */
    switch(p1->sel_dydx)
    {
      case 0: dx_att = p1->dxw; dy_att = p1->dyw; break;
      case 1: dx_att = p1->dxl; dy_att = p1->dyl; break;
      case 2: dx_att = p1->dxs; dy_att = p1->dys; break;
      case 3: dx_att = p1->dxt; dy_att = p1->dyt; break;
      case 4: dx_att = p1->dxr; dy_att = p1->dyr; break;
      case 5: dx_att = p1->dxg; dy_att = p1->dyg; break;
      case 6: dx_att = p1->dxb; dy_att = p1->dyb; break;
      case 7: dx_att = p1->dxa; dy_att = p1->dya; break;
      case 8:
      default: dx_att = p1->dxz; dy_att = p1->dyz; break;
    }

    ew_offset(p0, p1, dx_att, dy_att);

    /*
     *  Edge Walker Stall Logic
     */
    p0->get_new_stall = ((p1->cnt_span_x & 8) == 8) && span_valid;
    get_ew_stall = p1->get_new_stall || p1->ew_stall_x;

    if(p1->reset_l == 0)
    {
      p0->ew_stall_x = 0;
      p0->cnt_stall = 0;
      p0->end_prim_state = 0;
    }
    else
    {
      larger_x = p1->left_xminor ? p1->x_sc_max : (p1->x_major_1d >> 8);
      smaller_x = p1->left_xminor ? (p1->x_major_1d >> 8) : p1->x_sc_min;
      num_pixel = SIGN_EXTEND_12(larger_x) - SIGN_EXTEND_12(smaller_x);
      p0->ew_ms_length = num_pixel;

      if(!p1->load_cmd_ewstall)
      {
	switch(p1->cycle_type)
	{
	  case 0: num_shift = 0; comp_value = 8; break;
	  case 1: num_shift = 1; comp_value = 6; break;
	  case 2: num_shift = 2; comp_value = 7; break;
	  case 3: 
            comp_value = 7;
	    switch(p1->pixel_size)
	    {
	      case 0: num_shift = 4; break;
	      case 1: num_shift = 3; break;
	      case 2: num_shift = 2; break;
	      case 3: num_shift = 1; break;
	    }
	    break;
        }
      }
      else
      {
        if(p1->load_cmd_tlut)
	{
	  comp_value = 8;
          num_shift = 0;
	}
        else
        {
          comp_value = 7;
	  switch(p1->texel_size)
	  {
	    case 0: num_shift = 4; break;
	    case 1: num_shift = 3; break;
	    case 2: num_shift = 2; break;
	    case 3: num_shift = 1; break;
          }
        }
      }

      if(!p1->load_cmd_ewstall && (p1->cycle_type == 1))
	num_cycles_new = num_pixel << num_shift;
      else
	num_cycles_new = num_pixel >> num_shift;

      num_cycles = p1->get_new_stall ?  num_cycles_new : p1->cnt_stall;
      p0->ew_stall_x = get_ew_stall ? (num_cycles > comp_value) : p1->ew_stall_x;

      stall_decr = p1->get_new_stall ? (num_cycles_new - 1) : (p1->cnt_stall - 1);


      if(p1->ew_stall_x || (p1->get_new_stall && (num_cycles != 0)))
        p0->cnt_stall = stall_decr;
      else if(num_cycles == 0)
        p1->cnt_stall = 0;
      else
        p0->cnt_stall = stall_decr;

      switch(p1->end_prim_state)
      {
        case 0:
          if(p0->cs_ew_newprim) /* negedge signal */
            p0->end_prim_state = 3;
          else
            p0->end_prim_state = 0;
          break;
        case 1:
          if(p0->cs_ew_newprim) /* negedge signal */
            p0->end_prim_state = 3;
          else
            p0->end_prim_state = 1;
          break;
        case 2:
          if(p1->cnt_stall == 0)
            p0->end_prim_state = 1;
          else
            p0->end_prim_state = 2;
          break;
        case 3:
	   /* newprim negedge signal */
          if(p0->cs_ew_newprim && p1->end_ew && p1->span_valid_s && (num_cycles > 8))
            p0->end_prim_state = 2;
          else if(p0->cs_ew_newprim)
            p0->end_prim_state = 3;
          else if(!p1->span_valid_x_s)
            p0->end_prim_state = 1;
          else if(p1->end_ew && p1->span_valid_s && (num_cycles > 0))
            p0->end_prim_state = 2;
          else if(p1->end_ew && !p1->span_valid_s)
            p0->end_prim_state = 1;
          else if(p1->end_ew && (num_cycles == 0))
            p0->end_prim_state = 1;
          else
            p0->end_prim_state = 3;
          break;
      }

      p0->ew_cs_busy = (p0->end_prim_state >> 1) & 1;
    }
    
    /* 
     *  Mem Span Address Calculation
     */
    ew_ms_adrs(p0, p1);

  }
  /*
   *  Save clock
   */
  p0->gclk_old = p1->gclk_old = save_clk;
}

/* ======================================================================
 *  ew_init():  main initialization routine for the edge walker unit.
 *
 * ======================================================================
 */
void
  ew_init(ew_t *p0, ew_t *p1)
{
  p0->gclk = p1->gclk = 0;
  p0->gclk_old = p1->gclk_old = 0;  

  {
    char *ew_dump_str;
    if (ew_dump_str=getenv("EDGEWALK_DUMP")) {
      if (!(sscanf(ew_dump_str,"%i",&ew_dump)))
        ew_dump=0;
    } else 
      ew_dump=0;
  }
}