From: notaz <notasas@gmail.com>
Date: Sat, 5 Aug 2017 23:03:35 +0000 (+0300)
Subject: eliminate texrels (wip)
X-Git-Tag: v1.92~58
X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=ea38612fad50103e224a3d00492d40b7dcff9e94;p=picodrive.git

eliminate texrels (wip)
---

diff --git a/pico/32x/draw.c b/pico/32x/draw.c
index 66f67a7e..9500e088 100644
--- a/pico/32x/draw.c
+++ b/pico/32x/draw.c
@@ -83,7 +83,7 @@ static void convert_pal555(int invert_prio)
 }
 
 // this is almost never used (Wiz and menu bg gen only)
-void FinalizeLine32xRGB555(int sh, int line)
+void FinalizeLine32xRGB555(int sh, int line, struct PicoEState *est)
 {
   unsigned short *pd = DrawLineDest;
   unsigned short *pal = Pico32xMem->pal_native;
@@ -91,7 +91,7 @@ void FinalizeLine32xRGB555(int sh, int line)
   unsigned short *dram, *p32x;
   unsigned char   mdbg;
 
-  FinalizeLine555(sh, line);
+  FinalizeLine555(sh, line, est);
 
   if ((Pico32x.vdp_regs[0] & P32XV_Mx) == 0 || // 32x blanking
       // XXX: how is 32col mode hadled by real hardware?
diff --git a/pico/cd/gfx_dma.c b/pico/cd/gfx_dma.c
index 44fa7ab2..a24a1c3c 100644
--- a/pico/cd/gfx_dma.c
+++ b/pico/cd/gfx_dma.c
@@ -37,7 +37,7 @@ PICO_INTERNAL void DmaSlowCell(unsigned int source, unsigned int a, int len, uns
         // AutoIncrement
         a=(u16)(a+inc);
       }
-      rendstatus |= PDRAW_SPRITES_MOVED;
+      Pico.est.rendstatus |= PDRAW_SPRITES_MOVED;
       break;
 
     case 3: // cram
diff --git a/pico/debug.c b/pico/debug.c
index 959331f8..91bff566 100644
--- a/pico/debug.c
+++ b/pico/debug.c
@@ -203,7 +203,7 @@ void PDebugShowPalette(unsigned short *screen, int stride)
   if (PicoAHW & PAHW_SMS)
     PicoDoHighPal555M4();
   else
-    PicoDoHighPal555(1);
+    PicoDoHighPal555(1, 0, &Pico.est);
   Pico.m.dirtyPal = 1;
 
   screen += 16*stride+8;
diff --git a/pico/draw.c b/pico/draw.c
index ca9a140b..ff84be57 100644
--- a/pico/draw.c
+++ b/pico/draw.c
@@ -54,9 +54,8 @@ int  HighPreSpr[80*2+1]; // slightly preprocessed sprites
 #define SPRL_LO_ABOVE_HI 0x10 // low priority sprites may be on top of hi
 unsigned char HighLnSpr[240][3 + MAX_LINE_SPRITES]; // sprite_count, ^flags, tile_count, [spritep]...
 
-int rendstatus, rendstatus_old;
+int rendstatus_old;
 int rendlines;
-int DrawScanline;
 int PicoDrawMask = -1;
 
 static int skip_next_line=0;
@@ -75,12 +74,15 @@ struct TileStrip
 
 // stuff available in asm:
 #ifdef _ASM_DRAW_C
-void DrawWindow(int tstart, int tend, int prio, int sh);
-void DrawAllSprites(unsigned char *sprited, int prio, int sh);
-void DrawTilesFromCache(int *hc, int sh, int rlim);
-void DrawSpritesSHi(unsigned char *sprited);
-void DrawLayer(int plane_sh, int *hcache, int cellskip, int maxcells);
-void FinalizeLineBGR444(int sh, int line);
+void DrawWindow(int tstart, int tend, int prio, int sh,
+                struct PicoEState *est);
+void DrawAllSprites(unsigned char *sprited, int prio, int sh,
+                    struct PicoEState *est);
+void DrawTilesFromCache(int *hc, int sh, int rlim,
+                    struct PicoEState *est);
+void DrawSpritesSHi(unsigned char *sprited, struct PicoEState *est);
+void DrawLayer(int plane_sh, int *hcache, int cellskip, int maxcells,
+               struct PicoEState *est);
 void *blockcpy(void *dst, const void *src, size_t n);
 void blockcpy_or(void *dst, void *src, size_t n, int pat);
 #else
@@ -258,7 +260,7 @@ static void DrawStrip(struct TileStrip *ts, int plane_sh, int cellskip)
   // terminate the cache list
   *ts->hc = 0;
   // if oldcode wasn't changed, it means all layer is hi priority
-  if (oldcode == -1) rendstatus |= PDRAW_PLANE_HI_PRIO;
+  if (oldcode == -1) Pico.est.rendstatus |= PDRAW_PLANE_HI_PRIO;
 }
 
 // this is messy
@@ -266,7 +268,7 @@ void DrawStripVSRam(struct TileStrip *ts, int plane_sh, int cellskip)
 {
   int tilex,dx,code=0,addr=0,cell=0;
   int oldcode=-1,blank=-1; // The tile we know is blank
-  int pal=0,scan=DrawScanline;
+  int pal=0,scan=Pico.est.DrawScanline;
 
   // Draw tiles across screen:
   tilex=(-ts->hscroll)>>3;
@@ -317,7 +319,7 @@ void DrawStripVSRam(struct TileStrip *ts, int plane_sh, int cellskip)
 
   // terminate the cache list
   *ts->hc = 0;
-  if (oldcode == -1) rendstatus |= PDRAW_PLANE_HI_PRIO;
+  if (oldcode == -1) Pico.est.rendstatus |= PDRAW_PLANE_HI_PRIO;
 }
 #endif
 
@@ -374,7 +376,8 @@ void DrawStripInterlace(struct TileStrip *ts)
 // --------------------------------------------
 
 #ifndef _ASM_DRAW_C
-static void DrawLayer(int plane_sh, int *hcache, int cellskip, int maxcells)
+static void DrawLayer(int plane_sh, int *hcache, int cellskip, int maxcells,
+  struct PicoEState *est)
 {
   struct PicoVideo *pvid=&Pico.video;
   const char shift[4]={5,6,5,7}; // 32,64 or 128 sized tilemaps (2 is invalid)
@@ -401,7 +404,7 @@ static void DrawLayer(int plane_sh, int *hcache, int cellskip, int maxcells)
   else            ts.nametab=(pvid->reg[2]&0x38)<< 9; // A
 
   htab=pvid->reg[13]<<9; // Horizontal scroll table address
-  if ( pvid->reg[11]&2)     htab+=DrawScanline<<1; // Offset by line
+  if ( pvid->reg[11]&2)     htab+=est->DrawScanline<<1; // Offset by line
   if ((pvid->reg[11]&1)==0) htab&=~0xf; // Offset by tile
   htab+=plane_sh&1; // A or B
 
@@ -413,7 +416,7 @@ static void DrawLayer(int plane_sh, int *hcache, int cellskip, int maxcells)
     vscroll=Pico.vsram[plane_sh&1]; // Get vertical scroll value
 
     // Find the line in the name table
-    ts.line=(vscroll+(DrawScanline<<1))&((ymask<<1)|1);
+    ts.line=(vscroll+(est->DrawScanline<<1))&((ymask<<1)|1);
     ts.nametab+=(ts.line>>4)<<shift[width];
 
     DrawStripInterlace(&ts);
@@ -426,7 +429,7 @@ static void DrawLayer(int plane_sh, int *hcache, int cellskip, int maxcells)
     vscroll=Pico.vsram[plane_sh&1]; // Get vertical scroll value
 
     // Find the line in the name table
-    ts.line=(vscroll+DrawScanline)&ymask;
+    ts.line=(vscroll+est->DrawScanline)&ymask;
     ts.nametab+=(ts.line>>3)<<shift[width];
 
     DrawStrip(&ts, plane_sh, cellskip);
@@ -437,7 +440,8 @@ static void DrawLayer(int plane_sh, int *hcache, int cellskip, int maxcells)
 // --------------------------------------------
 
 // tstart & tend are tile pair numbers
-static void DrawWindow(int tstart, int tend, int prio, int sh) // int *hcache
+static void DrawWindow(int tstart, int tend, int prio, int sh,
+                       struct PicoEState *est)
 {
   struct PicoVideo *pvid=&Pico.video;
   int tilex,ty,nametab,code=0;
@@ -447,17 +451,17 @@ static void DrawWindow(int tstart, int tend, int prio, int sh) // int *hcache
   if (pvid->reg[12]&1)
   {
     nametab=(pvid->reg[3]&0x3c)<<9; // 40-cell mode
-    nametab+=(DrawScanline>>3)<<6;
+    nametab+=(est->DrawScanline>>3)<<6;
   }
   else
   {
     nametab=(pvid->reg[3]&0x3e)<<9; // 32-cell mode
-    nametab+=(DrawScanline>>3)<<5;
+    nametab+=(est->DrawScanline>>3)<<5;
   }
 
   tilex=tstart<<1;
 
-  if (!(rendstatus & PDRAW_WND_DIFF_PRIO)) {
+  if (!(est->rendstatus & PDRAW_WND_DIFF_PRIO)) {
     // check the first tile code
     code=Pico.vram[nametab+tilex];
     // if the whole window uses same priority (what is often the case), we may be able to skip this field
@@ -465,7 +469,7 @@ static void DrawWindow(int tstart, int tend, int prio, int sh) // int *hcache
   }
 
   tend<<=1;
-  ty=(DrawScanline&7)<<1; // Y-Offset into tile
+  ty=(est->DrawScanline&7)<<1; // Y-Offset into tile
 
   // Draw tiles across screen:
   if (!sh)
@@ -478,7 +482,7 @@ static void DrawWindow(int tstart, int tend, int prio, int sh) // int *hcache
       code=Pico.vram[nametab+tilex];
       if (code==blank) continue;
       if ((code>>15) != prio) {
-        rendstatus |= PDRAW_WND_DIFF_PRIO;
+        est->rendstatus |= PDRAW_WND_DIFF_PRIO;
         continue;
       }
 
@@ -504,7 +508,7 @@ static void DrawWindow(int tstart, int tend, int prio, int sh) // int *hcache
       code=Pico.vram[nametab+tilex];
       if(code==blank) continue;
       if((code>>15) != prio) {
-        rendstatus |= PDRAW_WND_DIFF_PRIO;
+        est->rendstatus |= PDRAW_WND_DIFF_PRIO;
         continue;
       }
 
@@ -538,23 +542,23 @@ static void DrawTilesFromCacheShPrep(void)
   // we can process whole line and then act as if sh/hi mode was off,
   // but leave lo pri op sprite markers alone
   int c = 320/4, *zb = (int *)(HighCol+8);
-  rendstatus |= PDRAW_SHHI_DONE;
+  Pico.est.rendstatus |= PDRAW_SHHI_DONE;
   while (c--)
   {
     *zb++ &= 0xbfbfbfbf;
   }
 }
 
-static void DrawTilesFromCache(int *hc, int sh, int rlim)
+static void DrawTilesFromCache(int *hc, int sh, int rlim, struct PicoEState *est)
 {
   int code, addr, dx;
   int pal;
 
   // *ts->hc++ = code | (dx<<16) | (ty<<25); // cache it
 
-  if (sh && (rendstatus & (PDRAW_SHHI_DONE|PDRAW_PLANE_HI_PRIO)))
+  if (sh && (est->rendstatus & (PDRAW_SHHI_DONE|PDRAW_PLANE_HI_PRIO)))
   {
-    if (!(rendstatus & PDRAW_SHHI_DONE))
+    if (!(est->rendstatus & PDRAW_SHHI_DONE))
       DrawTilesFromCacheShPrep();
     sh = 0;
   }
@@ -658,7 +662,7 @@ static void DrawSprite(int *sprite, int sh)
   height=(sy>>24)&7; // Width and height in tiles
   sy=(sy<<16)>>16; // Y
 
-  row=DrawScanline-sy; // Row of the sprite we are on
+  row=Pico.est.DrawScanline-sy; // Row of the sprite we are on
 
   if (code&0x1000) row=(height<<3)-1-row; // Flip Y
 
@@ -706,7 +710,7 @@ static void DrawSpriteInterlace(unsigned int *sprite)
   width=(height>>2)&3; height&=3;
   width++; height++; // Width and height in tiles
 
-  row=(DrawScanline<<1)-sy; // Row of the sprite we are on
+  row=(Pico.est.DrawScanline<<1)-sy; // Row of the sprite we are on
 
   code=sprite[1];
   sx=((code>>16)&0x1ff)-0x78; // X
@@ -738,7 +742,7 @@ static void DrawSpriteInterlace(unsigned int *sprite)
 static void DrawAllSpritesInterlace(int pri, int sh)
 {
   struct PicoVideo *pvid=&Pico.video;
-  int i,u,table,link=0,sline=DrawScanline<<1;
+  int i,u,table,link=0,sline=Pico.est.DrawScanline<<1;
   unsigned int *sprites[80]; // Sprite index
 
   table=pvid->reg[5]&0x7f;
@@ -790,7 +794,7 @@ static void DrawAllSpritesInterlace(int pri, int sh)
  * Index + 0  :    hhhhvvvv ----hhvv yyyyyyyy yyyyyyyy // v, h: vert./horiz. size
  * Index + 4  :    xxxxxxxx xxxxxxxx pccvhnnn nnnnnnnn // x: x coord + 8
  */
-static void DrawSpritesSHi(unsigned char *sprited)
+static void DrawSpritesSHi(unsigned char *sprited, const struct PicoEState *est)
 {
   int (*fTileFunc)(int sx,int addr,int pal);
   unsigned char *p;
@@ -835,7 +839,7 @@ static void DrawSpritesSHi(unsigned char *sprited)
     height=(sy>>24)&7; // Width and height in tiles
     sy=(sy<<16)>>16; // Y
 
-    row=DrawScanline-sy; // Row of the sprite we are on
+    row=est->DrawScanline-sy; // Row of the sprite we are on
 
     if (code&0x1000) row=(height<<3)-1-row; // Flip Y
 
@@ -867,7 +871,7 @@ static void DrawSpritesHiAS(unsigned char *sprited, int sh)
   cnt = sprited[0] & 0x7f;
   if (cnt == 0) return;
 
-  rendstatus |= PDRAW_SPR_LO_ON_HI;
+  Pico.est.rendstatus |= PDRAW_SPR_LO_ON_HI;
 
   p = &sprited[3];
 
@@ -906,7 +910,7 @@ static void DrawSpritesHiAS(unsigned char *sprited, int sh)
     height=(sy>>24)&7; // Width and height in tiles
     sy=(sy<<16)>>16; // Y
 
-    row=DrawScanline-sy; // Row of the sprite we are on
+    row=Pico.est.DrawScanline-sy; // Row of the sprite we are on
 
     if (code&0x1000) row=(height<<3)-1-row; // Flip Y
 
@@ -942,7 +946,7 @@ static void DrawSpritesHiAS(unsigned char *sprited, int sh)
 
   /* nasty 2: sh operator pass */
   sprited[0] = sh_cnt;
-  DrawSpritesSHi(sprited);
+  DrawSpritesSHi(sprited, &Pico.est);
 }
 
 
@@ -954,7 +958,8 @@ static void DrawSpritesHiAS(unsigned char *sprited, int sh)
 
 void PrepareSprites(int full)
 {
-  struct PicoVideo *pvid=&Pico.video;
+  const struct PicoVideo *pvid=&Pico.video;
+  const struct PicoEState *est=&Pico.est;
   int u,link=0,sh;
   int table=0;
   int *pd = HighPreSpr;
@@ -991,10 +996,11 @@ void PrepareSprites(int full)
       sy = (pack << 16) >> 16;
       height = (pack >> 24) & 0xf;
 
-      if (sy < max_lines && sy + (height<<3) > DrawScanline && // sprite onscreen (y)?
+      if (sy < max_lines &&
+	  sy + (height<<3) > est->DrawScanline && // sprite onscreen (y)?
           (sx > -24 || sx < max_width))                   // onscreen x
       {
-        int y = (sy >= DrawScanline) ? sy : DrawScanline;
+        int y = (sy >= est->DrawScanline) ? sy : est->DrawScanline;
         int entry = ((pd - HighPreSpr) / 2) | ((code2>>8)&0x80);
         for (; y < sy + (height<<3) && y < max_lines; y++)
         {
@@ -1047,7 +1053,7 @@ found:;
       sx = (code2>>16)&0x1ff;
       sx -= 0x78; // Get X coordinate + 8
 
-      if (sy < max_lines && sy + (height<<3) > DrawScanline) // sprite onscreen (y)?
+      if (sy < max_lines && sy + (height<<3) > est->DrawScanline) // sprite onscreen (y)?
       {
         int entry, y, sx_min, onscr_x, maybe_op = 0;
 
@@ -1057,7 +1063,7 @@ found:;
           maybe_op = SPRL_MAY_HAVE_OP;
 
         entry = ((pd - HighPreSpr) / 2) | ((code2>>8)&0x80);
-        y = (sy >= DrawScanline) ? sy : DrawScanline;
+        y = (sy >= est->DrawScanline) ? sy : est->DrawScanline;
         for (; y < sy + (height<<3) && y < max_lines; y++)
         {
 	  unsigned char *p = &HighLnSpr[y][0];
@@ -1110,16 +1116,17 @@ found:;
 }
 
 #ifndef _ASM_DRAW_C
-static void DrawAllSprites(unsigned char *sprited, int prio, int sh)
+static void DrawAllSprites(unsigned char *sprited, int prio, int sh,
+                           struct PicoEState *est)
 {
-  int rs = rendstatus;
+  int rs = est->rendstatus;
   unsigned char *p;
   int cnt;
 
   if (rs & (PDRAW_SPRITES_MOVED|PDRAW_DIRTY_SPRITES)) {
     //elprintf(EL_STATUS, "PrepareSprites(%i)", (rs>>4)&1);
     PrepareSprites(rs & PDRAW_DIRTY_SPRITES);
-    rendstatus = rs & ~(PDRAW_SPRITES_MOVED|PDRAW_DIRTY_SPRITES);
+    est->rendstatus = rs & ~(PDRAW_SPRITES_MOVED|PDRAW_DIRTY_SPRITES);
   }
 
   cnt = sprited[0] & 0x7f;
@@ -1159,7 +1166,7 @@ void BackFill(int reg7, int sh)
 unsigned short HighPal[0x100];
 
 #ifndef _ASM_DRAW_C
-void PicoDoHighPal555(int sh)
+void PicoDoHighPal555(int sh, int line, struct PicoEState *est)
 {
   unsigned int *spal, *dpal;
   unsigned int t, i;
@@ -1205,7 +1212,7 @@ void FinalizeLine555(int sh, int line)
   int len;
 
   if (Pico.m.dirtyPal)
-    PicoDoHighPal555(sh);
+    PicoDoHighPal555(sh, line, est);
 
   if (Pico.video.reg[12]&1) {
     len = 320;
@@ -1217,7 +1224,7 @@ void FinalizeLine555(int sh, int line)
   {
 #ifndef PSP
     int i, mask=0xff;
-    if (!sh && (rendstatus & PDRAW_SPR_LO_ON_HI))
+    if (!sh && (est->rendstatus & PDRAW_SPR_LO_ON_HI))
       mask=0x3f; // accurate sprites, upper bits are priority stuff
 
     for (i = 0; i < len; i++)
@@ -1225,7 +1232,7 @@ void FinalizeLine555(int sh, int line)
 #else
     extern void amips_clut(unsigned short *dst, unsigned char *src, unsigned short *pal, int count);
     extern void amips_clut_6bit(unsigned short *dst, unsigned char *src, unsigned short *pal, int count);
-    if (!sh && (rendstatus & PDRAW_SPR_LO_ON_HI))
+    if (!sh && (est->rendstatus & PDRAW_SPR_LO_ON_HI))
          amips_clut_6bit(pd, ps, pal, len);
     else amips_clut(pd, ps, pal, len);
 #endif
@@ -1233,10 +1240,10 @@ void FinalizeLine555(int sh, int line)
 }
 #endif
 
-static void FinalizeLine8bit(int sh, int line)
+static void FinalizeLine8bit(int sh, int line, struct PicoEState *est)
 {
   unsigned char *pd = DrawLineDest;
-  int len, rs = rendstatus;
+  int len, rs = est->rendstatus;
   static int dirty_count;
 
   if (!sh && Pico.m.dirtyPal == 1)
@@ -1246,7 +1253,7 @@ static void FinalizeLine8bit(int sh, int line)
          dirty_count = 1;
     else dirty_count++;
     rs |= PDRAW_SONIC_MODE;
-    rendstatus = rs;
+    est->rendstatus = rs;
     if (dirty_count == 3) {
       blockcpy(HighPal, Pico.cram, 0x40*2);
     } else if (dirty_count == 11) {
@@ -1273,18 +1280,19 @@ static void FinalizeLine8bit(int sh, int line)
   }
 }
 
-static void (*FinalizeLine)(int sh, int line);
+static void (*FinalizeLine)(int sh, int line, struct PicoEState *est);
 
 // --------------------------------------------
 
 static int DrawDisplay(int sh)
 {
-  unsigned char *sprited = &HighLnSpr[DrawScanline][0];
+  struct PicoEState *est=&Pico.est;
+  unsigned char *sprited = &HighLnSpr[est->DrawScanline][0];
   struct PicoVideo *pvid=&Pico.video;
   int win=0,edge=0,hvwind=0;
   int maxw,maxcells;
 
-  rendstatus &= ~(PDRAW_SHHI_DONE|PDRAW_PLANE_HI_PRIO);
+  est->rendstatus &= ~(PDRAW_SHHI_DONE|PDRAW_PLANE_HI_PRIO);
 
   if (pvid->reg[12]&1) {
     maxw = 328; maxcells = 40;
@@ -1296,8 +1304,8 @@ static int DrawDisplay(int sh)
   win=pvid->reg[0x12];
   edge=(win&0x1f)<<3;
 
-  if (win&0x80) { if (DrawScanline>=edge) hvwind=1; }
-  else          { if (DrawScanline< edge) hvwind=1; }
+  if (win&0x80) { if (est->DrawScanline>=edge) hvwind=1; }
+  else          { if (est->DrawScanline< edge) hvwind=1; }
 
   if (!hvwind) // we might have a vertical window here
   {
@@ -1315,53 +1323,56 @@ static int DrawDisplay(int sh)
 
   /* - layer B low - */
   if (PicoDrawMask & PDRAW_LAYERB_ON)
-    DrawLayer(1|(sh<<1), HighCacheB, 0, maxcells);
+    DrawLayer(1|(sh<<1), HighCacheB, 0, maxcells, est);
   /* - layer A low - */
   if (!(PicoDrawMask & PDRAW_LAYERA_ON));
   else if (hvwind == 1)
-    DrawWindow(0, maxcells>>1, 0, sh);
+    DrawWindow(0, maxcells>>1, 0, sh, est);
   else if (hvwind == 2) {
-    DrawLayer(0|(sh<<1), HighCacheA, (win&0x80) ?    0 : edge<<1, (win&0x80) ?     edge<<1 : maxcells);
-    DrawWindow(                      (win&0x80) ? edge :       0, (win&0x80) ? maxcells>>1 : edge, 0, sh);
+    DrawLayer(0|(sh<<1), HighCacheA, (win&0x80) ?    0 : edge<<1, (win&0x80) ?     edge<<1 : maxcells, est);
+    DrawWindow(                      (win&0x80) ? edge :       0, (win&0x80) ? maxcells>>1 : edge, 0, sh, est);
   } else
-    DrawLayer(0|(sh<<1), HighCacheA, 0, maxcells);
+    DrawLayer(0|(sh<<1), HighCacheA, 0, maxcells, est);
   /* - sprites low - */
   if (!(PicoDrawMask & PDRAW_SPRITES_LOW_ON));
-  else if (rendstatus & PDRAW_INTERLACE)
+  else if (Pico.est.rendstatus & PDRAW_INTERLACE)
     DrawAllSpritesInterlace(0, sh);
   else if (sprited[1] & SPRL_HAVE_LO)
-    DrawAllSprites(sprited, 0, sh);
+    DrawAllSprites(sprited, 0, sh, est);
 
   /* - layer B hi - */
   if ((PicoDrawMask & PDRAW_LAYERB_ON) && HighCacheB[0])
-    DrawTilesFromCache(HighCacheB, sh, maxw);
+    DrawTilesFromCache(HighCacheB, sh, maxw, est);
   /* - layer A hi - */
   if (!(PicoDrawMask & PDRAW_LAYERA_ON));
   else if (hvwind == 1)
-    DrawWindow(0, maxcells>>1, 1, sh);
+    DrawWindow(0, maxcells>>1, 1, sh, est);
   else if (hvwind == 2) {
-    if (HighCacheA[0]) DrawTilesFromCache(HighCacheA, sh, (win&0x80) ? edge<<4 : maxw);
-    DrawWindow((win&0x80) ? edge : 0, (win&0x80) ? maxcells>>1 : edge, 1, sh);
+    if (HighCacheA[0])
+      DrawTilesFromCache(HighCacheA, sh, (win&0x80) ? edge<<4 : maxw, est);
+    DrawWindow((win&0x80) ? edge : 0, (win&0x80) ? maxcells>>1 : edge, 1, sh, est);
   } else
-    if (HighCacheA[0]) DrawTilesFromCache(HighCacheA, sh, maxw);
+    if (HighCacheA[0])
+      DrawTilesFromCache(HighCacheA, sh, maxw, est);
   /* - sprites hi - */
   if (!(PicoDrawMask & PDRAW_SPRITES_HI_ON));
-  else if (rendstatus & PDRAW_INTERLACE)
+  else if (Pico.est.rendstatus & PDRAW_INTERLACE)
     DrawAllSpritesInterlace(1, sh);
   // have sprites without layer pri bit ontop of sprites with that bit
   else if ((sprited[1] & 0xd0) == 0xd0 && (PicoOpt & POPT_ACC_SPRITES))
     DrawSpritesHiAS(sprited, sh);
   else if (sh && (sprited[1] & SPRL_MAY_HAVE_OP))
-    DrawSpritesSHi(sprited);
+    DrawSpritesSHi(sprited, est);
   else if (sprited[1] & SPRL_HAVE_HI)
-    DrawAllSprites(sprited, 1, 0);
+    DrawAllSprites(sprited, 1, 0, est);
 
 #if 0
   {
     int *c, a, b;
     for (a = 0, c = HighCacheA; *c; c++, a++);
     for (b = 0, c = HighCacheB; *c; c++, b++);
-    printf("%i:%03i: a=%i, b=%i\n", Pico.m.frame_count, DrawScanline, a, b);
+    printf("%i:%03i: a=%i, b=%i\n", Pico.m.frame_count,
+           Pico.est.DrawScanline, a, b);
   }
 #endif
 
@@ -1374,27 +1385,27 @@ PICO_INTERNAL void PicoFrameStart(void)
   int offs = 8, lines = 224;
 
   // prepare to do this frame
-  rendstatus = 0;
+  Pico.est.rendstatus = 0;
   if ((Pico.video.reg[12] & 6) == 6)
-    rendstatus |= PDRAW_INTERLACE; // interlace mode
+    Pico.est.rendstatus |= PDRAW_INTERLACE; // interlace mode
   if (!(Pico.video.reg[12] & 1))
-    rendstatus |= PDRAW_32_COLS;
+    Pico.est.rendstatus |= PDRAW_32_COLS;
   if (Pico.video.reg[1] & 8) {
     offs = 0;
     lines = 240;
   }
 
-  if (rendstatus != rendstatus_old || lines != rendlines) {
+  if (Pico.est.rendstatus != rendstatus_old || lines != rendlines) {
     rendlines = lines;
     // mode_change() might reset rendstatus_old by calling SetColorFormat
     emu_video_mode_change((lines == 240) ? 0 : 8,
       lines, (Pico.video.reg[12] & 1) ? 0 : 1);
-    rendstatus_old = rendstatus;
+    rendstatus_old = Pico.est.rendstatus;
   }
 
   HighCol = HighColBase + offs * HighColIncrement;
   DrawLineDest = (char *)DrawLineDestBase + offs * DrawLineDestIncrement;
-  DrawScanline = 0;
+  Pico.est.DrawScanline = 0;
   skip_next_line = 0;
 
   if (PicoOpt & POPT_ALT_RENDERER)
@@ -1413,7 +1424,7 @@ static void DrawBlankedLine(int line, int offs, int sh, int bgc)
   BackFill(bgc, sh);
 
   if (FinalizeLine != NULL)
-    FinalizeLine(sh, line);
+    FinalizeLine(sh, line, &Pico.est);
 
   if (PicoScanEnd != NULL)
     PicoScanEnd(line + offs);
@@ -1431,7 +1442,7 @@ static void PicoLine(int line, int offs, int sh, int bgc)
     return;
   }
 
-  DrawScanline = line;
+  Pico.est.DrawScanline = line;
   if (PicoScanBegin != NULL)
     skip = PicoScanBegin(line + offs);
 
@@ -1446,7 +1457,7 @@ static void PicoLine(int line, int offs, int sh, int bgc)
     DrawDisplay(sh);
 
   if (FinalizeLine != NULL)
-    FinalizeLine(sh, line);
+    FinalizeLine(sh, line, &Pico.est);
 
   if (PicoScanEnd != NULL)
     skip_next_line = PicoScanEnd(line + offs);
@@ -1466,7 +1477,7 @@ void PicoDrawSync(int to, int blank_last_line)
   if (rendlines != 240)
     offs = 8;
 
-  for (line = DrawScanline; line < to; line++)
+  for (line = Pico.est.DrawScanline; line < to; line++)
   {
     PicoLine(line, offs, sh, bgc);
   }
@@ -1479,7 +1490,7 @@ void PicoDrawSync(int to, int blank_last_line)
     else PicoLine(line, offs, sh, bgc);
     line++;
   }
-  DrawScanline = line;
+  Pico.est.DrawScanline = line;
 
   pprof_end(draw);
 }
@@ -1491,8 +1502,8 @@ void PicoDrawUpdateHighPal(void)
   if (PicoOpt & POPT_ALT_RENDERER)
     sh = 0; // no s/h support
 
-  PicoDoHighPal555(sh);
-  if (rendstatus & PDRAW_SONIC_MODE) {
+  PicoDoHighPal555(sh, 0, &Pico.est);
+  if (Pico.est.rendstatus & PDRAW_SONIC_MODE) {
     // FIXME?
     memcpy(HighPal + 0x40, HighPal, 0x40*2);
     memcpy(HighPal + 0x80, HighPal, 0x40*2);
@@ -1528,7 +1539,7 @@ void PicoDrawSetOutBuf(void *dest, int increment)
 {
   DrawLineDestBase = dest;
   DrawLineDestIncrement = increment;
-  DrawLineDest = DrawLineDestBase + DrawScanline * increment;
+  DrawLineDest = DrawLineDestBase + Pico.est.DrawScanline * increment;
 }
 
 void PicoDrawSetInternalBuf(void *dest, int increment)
@@ -1536,7 +1547,7 @@ void PicoDrawSetInternalBuf(void *dest, int increment)
   if (dest != NULL) {
     HighColBase = dest;
     HighColIncrement = increment;
-    HighCol = HighColBase + DrawScanline * increment;
+    HighCol = HighColBase + Pico.est.DrawScanline * increment;
   }
   else {
     HighColBase = DefHighCol;
@@ -1560,3 +1571,5 @@ void PicoDrawSetCallbacks(int (*begin)(unsigned int num), int (*end)(unsigned in
     PicoScanEnd = end;
   }
 }
+
+// vim:ts=4:sw=4:expandtab
diff --git a/pico/draw_arm.S b/pico/draw_arm.S
index bbdda5b0..2c1db104 100644
--- a/pico/draw_arm.S
+++ b/pico/draw_arm.S
@@ -1,6 +1,6 @@
 /*
  * assembly optimized versions of most funtions from draw.c
- * (C) notaz, 2006-2010
+ * (C) notaz, 2006-2010,2017
  *
  * This work is licensed under the terms of MAME license.
  * See COPYING file in the top-level directory.
@@ -8,12 +8,11 @@
  * this is highly specialized, be careful if changing related C code!
  */
 
-.extern Pico
+#include "pico_int_o32.h"
+
 .extern PicoOpt
 .extern HighCol
-.extern DrawScanline
 .extern HighSprZ
-.extern rendstatus
 .extern HighPreSpr
 .extern DrawLineDest
 .extern DrawStripInterlace
@@ -287,14 +286,16 @@
 @   int cells;   // 0x14
 @ };
 
-@ void DrawLayer(int plane_sh, int *hcache, int cellskip, int maxcells);
+@ void DrawLayer(int plane_sh, int *hcache, int cellskip, int maxcells,
+@                struct PicoEState *est)
 
 .global DrawLayer
 
 DrawLayer:
+    ldr     r12, [sp]             @ est
     stmfd   sp!, {r4-r11,lr}
 
-    ldr     r11, =(Pico+0x22228)  @ Pico.video
+    ldr     r11, [r12, #OFS_Pico_video]
     mov     r8, #1
 
     ldrb    r7, [r11, #16]        @ ??vv??hh
@@ -316,17 +317,17 @@ DrawLayer:
     cmp     r10, #7
     subge   r10, r10, #1          @ r10=shift[width] (5,6,6,7)
 
-    @ calculate xmask:
-    mov     r5, r8, lsl r10
-    sub     r5, r5, #1            @ r5=xmask
+    ldr     r2, [r12, #OFS_DrawScanline]
+    ldr     lr, [r12, #OFS_Pico_vram]
 
     @ Find name table:
     ands    r0,  r0, #1
     ldreqb  r12, [r11, #2]
     ldrneb  r12, [r11, #4]
 
-    ldr     r2, =DrawScanline     @ trying to make good use of pipeline here
-    ldr     lr, =(Pico+0x10000)   @ lr=Pico.vram
+    @ calculate xmask:
+    mov     r5, r8, lsl r10
+    sub     r5, r5, #1            @ r5=xmask
 
     moveq   r12, r12, lsl #10
     movne   r12, r12, lsl #13
@@ -334,7 +335,6 @@ DrawLayer:
 
     ldrh    r8, [r11, #12]
     ldrb    r7, [r11, #11]
-    ldr     r2, [r2]
 
     mov     r4, r8, lsr #8        @ pvid->reg[13]
     mov     r4, r4, lsl #10       @ htab=pvid->reg[13]<<9; (halfwords)
@@ -504,12 +504,12 @@ DrawLayer:
 
 .dsloop_exit:
     tst     r10, #1<<21 @ seen non hi-prio tile
-    ldreq   r1, =rendstatus
+    ldr     r1, [sp, #9*4]  @ est
     mov     r0, #0
-    ldreq   r2, [r1]
+    ldreq   r2, [r1, #OFS_rendstatus]
     str     r0, [r6]    @ terminate the cache list
     orreq   r2, r2, #PDRAW_PLANE_HI_PRIO @ had a layer with all hi-prio tiles
-    streq   r2, [r1]
+    streq   r2, [r1, #OFS_rendstatus]
 
     ldmfd   sp!, {r4-r11,lr}
     bx      lr
@@ -522,9 +522,9 @@ DrawLayer:
     bic     r8, r8, #0x3fc00000
     orr     r8, r8, r5, lsl #25   @ r8=(xmask[31:25]|had_output[24]|tilex[21:0])
 
-    ldr     r4, =DrawScanline
+    ldr     r11, [sp, #9*4]       @ est
     orr     r5, r1, r10, lsl #24
-    ldr     r4, [r4]
+    ldr     r4, [r11, #OFS_DrawScanline]
     sub     r1, r3, #1
     orr     r5, r5, r4, lsl #16   @ r5=(shift_width[31:24]|scanline[23:16]|ymask[15:0])
     and     r1, r1, #7
@@ -679,12 +679,12 @@ DrawLayer:
 
 .dsloop_vs_exit:
     tst     r8, #(1<<24) @ seen non hi-prio tile
-    ldreq   r1, =rendstatus
+    ldr     r1, [sp, #9*4]  @ est
     mov     r0, #0
-    ldreq   r2, [r1]
+    ldreq   r2, [r1, #OFS_rendstatus]
     str     r0, [r6]    @ terminate the cache list
     orreq   r2, r2, #PDRAW_PLANE_HI_PRIO @ had a layer with all hi-prio tiles
-    streq   r2, [r1]
+    streq   r2, [r1, #OFS_rendstatus]
 
     ldmfd   sp!, {r4-r11,lr}
     bx      lr
@@ -770,11 +770,12 @@ BackFill:
 
 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
 
+@ void DrawTilesFromCache(int *hc, int sh, int rlim, struct PicoEState *est)
 
-.global DrawTilesFromCache @ int *hc, int sh, int rlim
+.global DrawTilesFromCache
 
 DrawTilesFromCache:
-    stmfd   sp!, {r4-r8,r11,lr}
+    stmfd   sp!, {r4-r9,r11,lr}
 
     @ cache some stuff to avoid mem access
 .if OVERRIDE_HIGHCOL
@@ -785,7 +786,8 @@ DrawTilesFromCache:
     ldr     r11,=HighCol
     mov     r12,#0xf
 .endif
-    ldr     lr, =(Pico+0x10000) @ lr=Pico.vram
+    ldr     lr, [r3, #OFS_Pico_vram]
+    mov     r9, r3         @ est
 
     mvn     r5, #0         @ r5=prevcode=-1
     ands    r8, r1, #1
@@ -796,7 +798,7 @@ DrawTilesFromCache:
 .dtfc_loop:
     ldr     r6, [r0], #4    @ read code
     movs    r1, r6, lsr #16 @ r1=dx;
-    ldmeqfd sp!, {r4-r8,r11,pc} @ dx is never zero, this must be a terminator, return
+    ldmeqfd sp!, {r4-r9,r11,pc} @ dx is never zero, this must be a terminator, return
     bic     r4, r1, #0xfe00
     add     r1, r11, r4     @ r1=pdest
 
@@ -915,15 +917,14 @@ DrawTilesFromCache:
 
 @ check if we have detected layer covered with hi-prio tiles:
 .dtfc_check_rendflags:
-    ldr     r1, =rendstatus
-    ldr     r2, [r1]
+    ldr     r2, [r9, #OFS_rendstatus]
     tst     r2, #(PDRAW_PLANE_HI_PRIO|PDRAW_SHHI_DONE)
     beq     .dtfc_loop
     bic     r8, r8, #1      @ sh/hi mode off
     tst     r2, #PDRAW_SHHI_DONE
     bne     .dtfc_loop      @ already processed
     orr     r2, r2, #PDRAW_SHHI_DONE
-    str     r2, [r1]
+    str     r2, [r9, #OFS_rendstatus]
 
     add     r1, r11,#8
     mov     r3, #320/4/4
@@ -948,7 +949,9 @@ DrawTilesFromCache:
 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
 
 
-.global DrawSpritesSHi @ unsigned char *sprited
+@ void DrawSpritesSHi(unsigned char *sprited, struct PicoEState *est)
+
+.global DrawSpritesSHi
 
 DrawSpritesSHi:
     ldr     r3, [r0]
@@ -956,7 +959,7 @@ DrawSpritesSHi:
     ands    r3, r3, #0x7f
     bxeq    lr
 
-    stmfd   sp!, {r4-r11,lr}
+    stmfd   sp!, {r1,r4-r11,lr} @ +est
     strb    r12,[r0,#2]     @ set end marker
     add     r10,r0, #3      @ r10=HighLnSpr end
     add     r10,r10,r3      @ r10=HighLnSpr end
@@ -969,16 +972,16 @@ DrawSpritesSHi:
     ldr     r11,=HighCol
     mov     r12,#0xf
 .endif
-    ldr     lr, =(Pico+0x10000) @ lr=Pico.vram
+    ldr     lr, [r1, #OFS_Pico_vram]
 
 
 DrawSpriteSHi:
     @ draw next sprite
     ldrb    r0, [r10,#-1]!
+    ldr     r7, [sp]        @ est
     ldr     r1, =HighPreSpr
-@    ldr     r8, [sp, #-4]
     cmp     r0, #0xff
-    ldmeqfd sp!, {r4-r11,pc} @ end of list
+    ldmeqfd sp!, {r1,r4-r11,pc} @ end of list
     and     r0, r0, #0x7f
     add     r0, r1, r0, lsl #3
 
@@ -997,15 +1000,14 @@ DrawSpriteSHi:
     bne     DrawSpriteSHi   @ non-operator low sprite, already drawn
 
     ldr     r3, [r0]        @ sprite[0]
-    ldr     r7, =DrawScanline
     mov     r6, r3, lsr #28
     sub     r6, r6, #1      @ r6=width-1 (inc later)
     mov     r5, r3, lsr #24
     and     r5, r5, #7      @ r5=height
 
+    ldr     r7, [r7, #OFS_DrawScanline]
     mov     r0, r3, lsl #16 @ r4=sy<<16 (tmp)
 
-    ldr     r7, [r7]
     sub     r7, r7, r0, asr #16 @ r7=row=DrawScanline-sy
 
     tst     r9, #0x1000
@@ -1128,34 +1130,34 @@ DrawSpriteSHi:
 
 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
 
-.global DrawAllSprites @ unsigned char *sprited, int prio, int sh
+@ void DrawAllSprites(unsigned char *sprited, int prio, int sh,
+@                     struct PicoEState *est)
+
+.global DrawAllSprites
 
 DrawAllSprites:
-    ldr     r3, =rendstatus
     orr     r1, r2, r1, lsl #1
-    ldr     r12,[r3]
+    ldr     r12,[r3, #OFS_rendstatus]
     tst     r12,#(PDRAW_DIRTY_SPRITES|PDRAW_SPRITES_MOVED)
     beq     das_no_prep
-    stmfd   sp!, {r0,r1,lr}
+    stmfd   sp!, {r0,r1,r3,lr}
     and     r0, r12,#PDRAW_DIRTY_SPRITES
     bic     r12,r12,#(PDRAW_DIRTY_SPRITES|PDRAW_SPRITES_MOVED)
-    str     r12,[r3]
+    str     r12,[r3, #OFS_rendstatus]
     bl      PrepareSprites
-    ldmfd   sp!, {r0,r1,lr}
+    ldmfd   sp!, {r0,r1,r3,lr}
 
 das_no_prep:
-    ldr     r3, [r0]
-    ands    r3, r3, #0x7f
+    ldr     r2, [r0]
+    ands    r2, r2, #0x7f
     bxeq    lr
 
     @ time to do some real work
-    stmfd   sp!, {r4-r11,lr}
+    stmfd   sp!, {r1,r3-r11,lr} @ +sh|prio<<1 +est
     mov     r12,#0xff
     strb    r12,[r0,#2]     @ set end marker
     add     r10,r0, #3
-    add     r10,r10,r3      @ r10=HighLnSpr end
-
-    str     r1, [sp, #-4]   @ no calls after this point
+    add     r10,r10,r2      @ r10=HighLnSpr end
 
 .if OVERRIDE_HIGHCOL
     ldr     r11,=HighCol
@@ -1165,29 +1167,27 @@ das_no_prep:
     ldr     r11,=HighCol
     mov     r12,#0xf
 .endif
-    ldr     lr, =(Pico+0x10000) @ lr=Pico.vram
+    ldr     lr, [r3, #OFS_Pico_vram]
 
 @ + 0  :    hhhhvvvv ----hhvv yyyyyyyy yyyyyyyy // v, h: horiz. size
 @ + 4  :    xxxxxxxx xxxxxxxx pccvhnnn nnnnnnnn // x: x coord + 8
 
-DrawSprite: @ was: unsigned int *sprite, int sh, int acc_sprites
+DrawSprite:
     @ draw next sprite
     ldrb    r0, [r10,#-1]!
-    ldr     r1, =HighPreSpr
-    ldr     r8, [sp, #-4]
+    ldr     r8, [sp]        @ sh|prio<<1
+    ldr     r7, [sp, #4]    @ est
     mov     r2, r0, lsr #7
     cmp     r0, #0xff
-    ldmeqfd sp!, {r4-r11,pc} @ end of list
+    ldmeqfd sp!, {r1,r3-r11,pc} @ end of list
     cmp     r2, r8, lsr #1
     bne     DrawSprite      @ wrong priority
+    ldr     r1, =HighPreSpr
     and     r0, r0, #0x7f
     add     r0, r1, r0, lsl #3
 
-@    stmfd   sp!, {r4-r9,r11,lr}
-@    orr     r8, r2, r1, lsl #4
-
     ldr     r3, [r0]        @ sprite[0]
-    ldr     r7, =DrawScanline
+    ldr     r7, [r7, #OFS_DrawScanline]
     mov     r6, r3, lsr #28
     sub     r6, r6, #1      @ r6=width-1 (inc later)
     mov     r5, r3, lsr #24
@@ -1195,7 +1195,6 @@ DrawSprite: @ was: unsigned int *sprite, int sh, int acc_sprites
 
     mov     r4, r3, lsl #16 @ r4=sy<<16 (tmp)
 
-    ldr     r7, [r7]
     ldr     r9, [r0, #4]
     sub     r7, r7, r4, asr #16 @ r7=row=DrawScanline-sy
 
@@ -1307,17 +1306,21 @@ DrawSprite: @ was: unsigned int *sprite, int sh, int acc_sprites
 
 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
 
-.global DrawWindow @ int tstart, int tend, int prio, int sh // int *hcache
+@ void DrawWindow(int tstart, int tend, int prio, int sh
+@                 struct PicoEState *est)
+
+.global DrawWindow
 
 DrawWindow:
+    ldr     r12, [sp]             @ est
     stmfd   sp!, {r4-r11,lr}
 
-    ldr     r11, =(Pico+0x22228)  @ Pico.video
-    ldr     r10, =DrawScanline
-    ldrb    r12, [r11, #3]        @ pvid->reg[3]
+    ldr     r6,  [r12, #OFS_Pico_video]
+    ldr     r10, [r12, #OFS_DrawScanline]
+    mov     r11, r12              @ est
+    ldrb    r12, [r6, #3]         @ pvid->reg[3]
 
-    ldr     r10, [r10]
-    ldr     r4,  [r11, #12]
+    ldr     r4,  [r6, #12]
     mov     r5,  r10, lsr #3
     and     r10, r10, #7
     mov     r10, r10, lsl #1      @ r10=ty
@@ -1331,9 +1334,8 @@ DrawWindow:
     addeq   r12, r12, r5, lsl #6  @ nametab
     add     r12, r12, r0, lsl #2  @ +starttile
 
-    ldr     r6, =rendstatus
-    ldr     lr, =(Pico+0x10000)   @ lr=Pico.vram
-    ldr     r6, [r6]
+    ldr     lr, [r11, #OFS_Pico_vram]
+    ldr     r6, [r11, #OFS_rendstatus]
 
     @ fetch the first code now
     ldrh    r7, [lr, r12]
@@ -1442,14 +1444,14 @@ DrawWindow:
     b       .dw_shadow_done
 
 .dwloop_end:
-    ldr     r0, =rendstatus
-    ldr     r1, [r0]
-    and     r6, r6, #PDRAW_WND_DIFF_PRIO
-    orr     r1, r1, r6
-    str     r1, [r0]
+    and     r2, r6, #PDRAW_WND_DIFF_PRIO
+    ldmfd   sp!, {r4-r11,lr}
+    ldr     r0, [sp]
+    ldr     r1, [r0, #OFS_rendstatus]
+    orr     r1, r1, r2
+    str     r1, [r0, #OFS_rendstatus]
 
-    ldmfd   sp!, {r4-r11,r12}
-    bx      r12
+    bx      lr
 
 
 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@@ -1531,12 +1533,15 @@ vidConvCpyRGB565: @ void *to, void *from, int pixels
     bx      lr
 
 
-.global PicoDoHighPal555 @ int sh
+@ void PicoDoHighPal555(int sh, int line, struct PicoEState *est)
+
+.global PicoDoHighPal555
 
 PicoDoHighPal555:
-    stmfd   sp!, {r4-r9,lr}
+    stmfd   sp!, {r4-r10,lr}
+    mov     r10,r2               @ est
     mov     r1, #0
-    ldr     r8, =(Pico+0x22228)  @ Pico.video
+    ldr     r8, [r10, #OFS_Pico_video]
 
 PicoDoHighPal555_nopush:
     orr     r9, r1, r0, lsl #31  @ 0:called from FinalizeLine555, 31: s/h
@@ -1598,17 +1603,20 @@ PicoDoHighPal555_nopush:
 
 PicoDoHighPal555_end:
     tst     r9, #1
-    ldmeqfd sp!, {r4-r9,pc}
+    ldmeqfd sp!, {r4-r10,pc}
 
-    ldr     r8, =(Pico+0x22228)  @ Pico.video
+    ldr     r8, [r10, #OFS_Pico_video]
     b       FinalizeLineRGB555_pal_done
 
 
-.global FinalizeLine555 @ int sh
+@ void FinalizeLine555(int sh, int line, struct PicoEState *est)
+
+.global FinalizeLine555
 
 FinalizeLine555:
-    stmfd   sp!, {r4-r9,lr}
-    ldr     r8, =(Pico+0x22228)  @ Pico.video
+    stmfd   sp!, {r4-r10,lr}
+    mov     r10,r2               @ est
+    ldr     r8, [r10, #OFS_Pico_video]
 
     ldrb    r2, [r8, #-0x1a]     @ 0x2220e ~ dirtyPal
     mov     r1, #1
@@ -1618,9 +1626,8 @@ FinalizeLine555:
 FinalizeLineRGB555_pal_done:
     ldr     r3, =HighPal
 
-    ldr     r12,=rendstatus
+    ldr     r12, [r10, #OFS_rendstatus]
     eors    r0, r0, #1           @ sh is 0
-    ldr     r12,[r12]
     mov     lr, #0xff
     tstne   r12,#PDRAW_ACC_SPRITES
     movne   lr, #0x3f
@@ -1691,12 +1698,11 @@ FinalizeLineRGB555_pal_done:
     stmia   r0!, {r4,r5,r8,r12}
     bne     .fl_loopRGB555
 
-    ldmfd   sp!, {r4-r9,lr}
+    ldmfd   sp!, {r4-r10,lr}
     bx      lr
 
 
 .fl_32scale_RGB555:
-    stmfd   sp!, {r10}
     mov     r9, #0x3900 @ f800 07e0 001f | e000 0780 001c | 3800 01e0 0007
     orr     r9, r9, #0x00e7
 
@@ -1757,8 +1763,7 @@ FinalizeLineRGB555_pal_done:
     stmia   r0!, {r4,r5,r6,r8,r10}
     bne     .fl_loop32scale_RGB555
 
-    ldmfd   sp!, {r10}
-    ldmfd   sp!, {r4-r9,lr}
+    ldmfd   sp!, {r4-r10,lr}
     bx      lr
 
 #ifdef UNALIGNED_DRAWLINEDEST
@@ -1804,7 +1809,7 @@ FinalizeLineRGB555_pal_done:
 
     strh    r8, [r0], #2
 
-    ldmfd   sp!, {r4-r9,lr}
+    ldmfd   sp!, {r4-r10,lr}
     bx      lr
 
 
@@ -1870,8 +1875,7 @@ FinalizeLineRGB555_pal_done:
 
     strh    r4, [r0], #2
 
-    ldmfd   sp!, {r10}
-    ldmfd   sp!, {r4-r9,lr}
+    ldmfd   sp!, {r4-r10,lr}
     bx      lr
 
 #endif /* UNALIGNED_DRAWLINEDEST */
diff --git a/pico/mode4.c b/pico/mode4.c
index 22e30419..c945f6fe 100644
--- a/pico/mode4.c
+++ b/pico/mode4.c
@@ -200,7 +200,7 @@ void PicoFrameStartMode4(void)
   int lines = 192;
   skip_next_line = 0;
   screen_offset = 24;
-  rendstatus = PDRAW_32_COLS;
+  Pico.est.rendstatus = PDRAW_32_COLS;
 
   if ((Pico.video.reg[0] & 6) == 6 && (Pico.video.reg[1] & 0x18)) {
     if (Pico.video.reg[1] & 0x08) {
@@ -213,9 +213,9 @@ void PicoFrameStartMode4(void)
     }
   }
 
-  if (rendstatus != rendstatus_old || lines != rendlines) {
+  if (Pico.est.rendstatus != rendstatus_old || lines != rendlines) {
     emu_video_mode_change(screen_offset, lines, 1);
-    rendstatus_old = rendstatus;
+    rendstatus_old = Pico.est.rendstatus;
     rendlines = lines;
   }
 
@@ -277,7 +277,7 @@ static void FinalizeLineRGB555M4(int line)
 
   // standard FinalizeLine can finish it for us,
   // with features like scaling and such
-  FinalizeLine555(0, line);
+  FinalizeLine555(0, line, &Pico.est);
 }
 
 static void FinalizeLine8bitM4(int line)
diff --git a/pico/pico.c b/pico/pico.c
index 8535d486..1c58f227 100644
--- a/pico/pico.c
+++ b/pico/pico.c
@@ -36,6 +36,9 @@ void PicoInit(void)
   memset(&PicoPad,0,sizeof(PicoPad));
   memset(&PicoPadInt,0,sizeof(PicoPadInt));
 
+  Pico.est.Pico_video = &Pico.video;
+  Pico.est.Pico_vram = Pico.vram;
+
   // Init CPUs:
   SekInit();
   z80_init(); // init even if we aren't going to use it
diff --git a/pico/pico.h b/pico/pico.h
index d5416727..c033c2da 100644
--- a/pico/pico.h
+++ b/pico/pico.h
@@ -99,6 +99,8 @@ typedef enum { PI_ROM, PI_ISPAL, PI_IS40_CELL, PI_IS240_LINES } pint_t;
 typedef union { int vint; void *vptr; } pint_ret_t;
 void PicoGetInternal(pint_t which, pint_ret_t *ret);
 
+struct PicoEState;
+
 // cd/mcd.c
 extern void (*PicoMCDopenTray)(void);
 extern void (*PicoMCDcloseTray)(void);
@@ -175,7 +177,7 @@ extern unsigned char *HighCol;
 #ifdef _ASM_DRAW_C
 void vidConvCpyRGB565(void *to, void *from, int pixels);
 #endif
-void PicoDoHighPal555(int sh);
+void PicoDoHighPal555(int sh, int line, struct PicoEState *est);
 extern int PicoDrawMask;
 #define PDRAW_LAYERB_ON      (1<<2)
 #define PDRAW_LAYERA_ON      (1<<3)
@@ -192,7 +194,7 @@ extern int PicoDrawMask;
 #define PDRAW_PLANE_HI_PRIO (1<<6) // have layer with all hi prio tiles (mk3)
 #define PDRAW_SHHI_DONE     (1<<7) // layer sh/hi already processed
 #define PDRAW_32_COLS       (1<<8) // 32 column mode
-extern int rendstatus, rendstatus_old;
+extern int rendstatus_old;
 extern int rendlines;
 extern unsigned short HighPal[0x100];
 
diff --git a/pico/pico_cmn.c b/pico/pico_cmn.c
index 341255a8..2de65d02 100644
--- a/pico/pico_cmn.c
+++ b/pico/pico_cmn.c
@@ -168,7 +168,7 @@ static int PicoFrameHints(void)
 
   if (!skip)
   {
-    if (DrawScanline < y)
+    if (Pico.est.DrawScanline < y)
       PicoDrawSync(y - 1, 0);
 #ifdef DRAW_FINISH_FUNC
     DRAW_FINISH_FUNC();
diff --git a/pico/pico_int.h b/pico/pico_int.h
index c85319c0..da49e04b 100644
--- a/pico/pico_int.h
+++ b/pico/pico_int.h
@@ -328,6 +328,15 @@ struct PicoMS
   unsigned char pad[0x4e];
 };
 
+// emu state and data for the asm code
+struct PicoEState
+{
+  int DrawScanline;
+  int rendstatus;
+  void *Pico_video;
+  void *Pico_vram;
+};
+
 // some assembly stuff depend on these, do not touch!
 struct Pico
 {
@@ -348,6 +357,7 @@ struct Pico
   struct PicoMisc m;
   struct PicoVideo video;
   struct PicoMS ms;
+  struct PicoEState est;
 };
 
 // sram
@@ -577,10 +587,9 @@ int CM_compareRun(int cyc, int is_sub);
 PICO_INTERNAL void PicoFrameStart(void);
 void PicoDrawSync(int to, int blank_last_line);
 void BackFill(int reg7, int sh);
-void FinalizeLine555(int sh, int line);
+void FinalizeLine555(int sh, int line, struct PicoEState *est);
 extern int (*PicoScanBegin)(unsigned int num);
 extern int (*PicoScanEnd)(unsigned int num);
-extern int DrawScanline;
 #define MAX_LINE_SPRITES 29
 extern unsigned char HighLnSpr[240][3 + MAX_LINE_SPRITES];
 extern void *DrawLineDestBase;
@@ -856,7 +865,7 @@ void p32x_sh2_poll_event(SH2 *sh2, unsigned int flags, unsigned int m68k_cycles)
 
 // 32x/draw.c
 void PicoDrawSetOutFormat32x(pdso_t which, int use_32x_line_mode);
-void FinalizeLine32xRGB555(int sh, int line);
+void FinalizeLine32xRGB555(int sh, int line, struct PicoEState *est);
 void PicoDraw32xLayer(int offs, int lines, int mdbg);
 void PicoDraw32xLayerMdOnly(int offs, int lines);
 extern int (*PicoScan32xBegin)(unsigned int num);
diff --git a/pico/pico_int_o32.h b/pico/pico_int_o32.h
new file mode 100644
index 00000000..1cdc6b58
--- /dev/null
+++ b/pico/pico_int_o32.h
@@ -0,0 +1,5 @@
+/* autogenerated by ./tools/mkoffsets, do not edit */
+#define OFS_DrawScanline     0x00
+#define OFS_rendstatus       0x04
+#define OFS_Pico_video       0x08
+#define OFS_Pico_vram        0x0c
diff --git a/pico/videoport.c b/pico/videoport.c
index 6c876aea..9ddc1679 100644
--- a/pico/videoport.c
+++ b/pico/videoport.c
@@ -36,7 +36,7 @@ static void VideoWrite(u16 d)
     case 1: if(a&1) d=(u16)((d<<8)|(d>>8)); // If address is odd, bytes are swapped (which game needs this?)
             Pico.vram [(a>>1)&0x7fff]=d;
             if (a - ((unsigned)(Pico.video.reg[5]&0x7f) << 9) < 0x400)
-              rendstatus |= PDRAW_DIRTY_SPRITES;
+              Pico.est.rendstatus |= PDRAW_DIRTY_SPRITES;
             break;
     case 3: Pico.m.dirtyPal = 1;
             Pico.cram [(a>>1)&0x003f]=d; break; // wraps (Desert Strike)
@@ -172,7 +172,7 @@ static void DmaSlow(int len)
           //if(pd >= pdend) pd-=0x8000; // should be good for RAM, bad for ROM
         }
       }
-      rendstatus |= PDRAW_DIRTY_SPRITES;
+      Pico.est.rendstatus |= PDRAW_DIRTY_SPRITES;
       break;
 
     case 3: // cram
@@ -241,7 +241,7 @@ static void DmaCopy(int len)
   }
   // remember addr
   Pico.video.addr=a;
-  rendstatus |= PDRAW_DIRTY_SPRITES;
+  Pico.est.rendstatus |= PDRAW_DIRTY_SPRITES;
 }
 
 // check: Contra, Megaman
@@ -280,7 +280,7 @@ static void DmaFill(int data)
   // update length
   Pico.video.reg[0x13] = Pico.video.reg[0x14] = 0; // Dino Dini's Soccer (E) (by Haze)
 
-  rendstatus |= PDRAW_DIRTY_SPRITES;
+  Pico.est.rendstatus |= PDRAW_DIRTY_SPRITES;
 }
 
 static void CommandDma(void)
@@ -319,7 +319,7 @@ static void CommandChange(void)
 static void DrawSync(int blank_on)
 {
   if (Pico.m.scanline < 224 && !(PicoOpt & POPT_ALT_RENDERER) &&
-      !PicoSkipFrame && DrawScanline <= Pico.m.scanline) {
+      !PicoSkipFrame && Pico.est.DrawScanline <= Pico.m.scanline) {
     //elprintf(EL_ANOMALY, "sync");
     PicoDrawSync(Pico.m.scanline, blank_on);
   }
@@ -412,7 +412,7 @@ PICO_INTERNAL_ASM void PicoVideoWrite(unsigned int a,unsigned short d)
             goto update_irq;
           case 0x05:
             //elprintf(EL_STATUS, "spritep moved to %04x", (unsigned)(Pico.video.reg[5]&0x7f) << 9);
-            if (d^dold) rendstatus |= PDRAW_SPRITES_MOVED;
+            if (d^dold) Pico.est.rendstatus |= PDRAW_SPRITES_MOVED;
             break;
           case 0x0c:
             // renderers should update their palettes if sh/hi mode is changed
diff --git a/platform/gp2x/emu.c b/platform/gp2x/emu.c
index 7741ba60..7e91361f 100644
--- a/platform/gp2x/emu.c
+++ b/platform/gp2x/emu.c
@@ -310,7 +310,7 @@ static int make_local_pal_md(int fast_mode)
 		localPal[0xf0] = 0x00ffffff;
 		pallen = 0x100;
 	}
-	else if (rendstatus & PDRAW_SONIC_MODE) { // mid-frame palette changes
+	else if (Pico.est.rendstatus & PDRAW_SONIC_MODE) { // mid-frame palette changes
 		bgr444_to_rgb32(localPal+0x40, HighPal);
 		bgr444_to_rgb32(localPal+0x80, HighPal+0x40);
 	}
diff --git a/platform/psp/emu.c b/platform/psp/emu.c
index 5a2e8809..8deb7fd3 100644
--- a/platform/psp/emu.c
+++ b/platform/psp/emu.c
@@ -223,7 +223,7 @@ static void do_pal_update(int allow_sh, int allow_as)
 		localPal[0xe0] = 0;
 		localPal[0xf0] = 0x001f;
 	}
-	else if (allow_as && (rendstatus & PDRAW_SPR_LO_ON_HI))
+	else if (allow_as && (Pico.est.rendstatus & PDRAW_SPR_LO_ON_HI))
 	{
 		memcpy32((int *)dpal+0x80/2, (void *)localPal, 0x40*2/4);
 	}
@@ -250,7 +250,7 @@ static void EmuScanPrepare(void)
 
 	if (Pico.m.dirtyPal)
 		do_pal_update(1, 1);
-	if ((rendstatus & PDRAW_SPR_LO_ON_HI) && !(Pico.video.reg[0xC]&8))
+	if ((Pico.est.rendstatus & PDRAW_SPR_LO_ON_HI) && !(Pico.video.reg[0xC]&8))
 	     amips_clut_f = amips_clut_6bit;
 	else amips_clut_f = amips_clut;
 }
diff --git a/tools/Makefile b/tools/Makefile
index 0c126cc1..28b748d4 100644
--- a/tools/Makefile
+++ b/tools/Makefile
@@ -1,6 +1,6 @@
 CFLAGS = -Wall -ggdb
 
-TARGETS = amalgamate textfilter
+TARGETS = amalgamate textfilter mkoffsets
 OBJS = $(addsuffix .o,$(TARGETS))
 
 all: $(TARGETS)
@@ -8,3 +8,6 @@ all: $(TARGETS)
 clean:
 	$(RM) $(TARGETS) $(OBJS)
 
+mkoffsets: CFLAGS += -m32 -I..
+
+.PHONY: clean all
diff --git a/tools/mkoffsets.c b/tools/mkoffsets.c
new file mode 100644
index 00000000..4044ad3d
--- /dev/null
+++ b/tools/mkoffsets.c
@@ -0,0 +1,31 @@
+#include <stdio.h>
+#include <stddef.h>
+
+#include "../pico/pico_int.h"
+
+#define DUMP(f, field) \
+  fprintf(f, "#define %-20s 0x%02x\n", \
+    "OFS_" #field, \
+    (int)offsetof(struct PicoEState, field))
+
+int main(int argc, char *argv[])
+{
+  char buf[128];
+  FILE *f;
+
+  snprintf(buf, sizeof(buf), "pico/pico_int_o%d.h", sizeof(void *) * 8);
+  f = fopen(buf, "w");
+  if (!f) {
+    perror("fopen");
+    return 1;
+  }
+
+  fprintf(f, "/* autogenerated by %s, do not edit */\n", argv[0]);
+  DUMP(f, DrawScanline);
+  DUMP(f, rendstatus);
+  DUMP(f, Pico_video);
+  DUMP(f, Pico_vram);
+  fclose(f);
+
+  return 0;
+}