From 93e2493dea91216f13a582bfa1fc64044cd26d30 Mon Sep 17 00:00:00 2001
From: Toni Wilen <twilen@winuae.net>
Date: Tue, 18 Mar 2008 20:16:54 +0200
Subject: [PATCH] imported winuaesrc1500b9.zip

---
 akiko.c                      |  21 ++--
 cia.c                        |   8 +-
 custom.c                     | 106 +++++++++++--------
 debug.c                      | 193 +++++++++++++++++++----------------
 drawing.c                    |  22 ++--
 gencpu.c                     |  30 ++++--
 genlinetoscr.c               |   2 -
 include/drawing.h            |   8 +-
 include/gfxfilter.h          |   3 +
 include/readcpu.h            |   4 +-
 jit/gencomp.c                |  13 ++-
 memory.c                     |  15 +--
 newcpu.c                     |  54 ++++++----
 od-win32/direct3d.c          | 105 ++++++-------------
 od-win32/dxwrap.c            |  37 ++++---
 od-win32/picasso96_win.c     |  52 +++++-----
 od-win32/win32.h             |   4 +-
 od-win32/win32_scale2x.c     | 120 ++++++++++++----------
 od-win32/win32gfx.c          |   7 +-
 od-win32/win32gui.c          | 176 ++++++++++++++++----------------
 od-win32/winuaechangelog.txt |  27 +++++
 p96_blit.c                   |   2 +-
 readcpu.c                    |  19 +++-
 savestate.c                  |   5 +-
 table68k                     |  21 +++-
 25 files changed, 588 insertions(+), 466 deletions(-)

diff --git a/akiko.c b/akiko.c
index b59cb733..d808262b 100755
--- a/akiko.c
+++ b/akiko.c
@@ -404,7 +404,7 @@ static int cd_hunt;
 static void checkint (void)
 {
     if (cdrom_status1 & cdrom_status2)
-	irq();
+	irq ();
 }
 
 static void set_status(uae_u32 status)
@@ -962,7 +962,7 @@ static void cdrom_run_read (void)
     }
     cdrom_sector_counter++;
     if (cdrom_readmask_w == 0)
-	set_status(CDSTATUS_DATASECTOR);
+	set_status (CDSTATUS_DATASECTOR);
 
 }
 
@@ -1061,14 +1061,14 @@ void AKIKO_hsync_handler (void)
 	    gui_cd_led (0, 1);
 	cdrom_run_read ();
 	framecounter = 1000000 / (74 * 75 * cdrom_speed);
-	set_status(CDSTATUS_FRAME);
+	set_status (CDSTATUS_FRAME);
     }
     if (cdrom_playing) {
 	static int frame2counter;
 	if (cdrom_audiotimeout > 0) {
 	    cdrom_audiotimeout--;
 	    if (cdrom_audiotimeout == 0) {
-		set_status(CDSTATUS_DATA_AVAILABLE);
+		set_status (CDSTATUS_DATA_AVAILABLE);
 		cdrom_playing = 0;
 		cdrom_result_buffer[1] = 0;
 		cdrom_return_data (2);
@@ -1496,6 +1496,8 @@ void akiko_reset (void)
 
     cdrom_speed = 1;
     cdrom_current_sector = -1;
+    cdrom_command_offset_complete = 0;
+    cdrom_command_offset_todo = 0;
 
     if (akiko_thread_running > 0) {
 	akiko_thread_running = 0;
@@ -1631,8 +1633,11 @@ uae_u8 *restore_akiko(uae_u8 *src)
     uae_u32 v;
     int i;
 
-    if (!currprefs.cs_cd32cd)
-	return NULL;
+    if (!currprefs.cs_cd32cd) {
+	changed_prefs.cs_cd32c2p = changed_prefs.cs_cd32cd = changed_prefs.cs_cd32nvram = 1;
+	currprefs.cs_cd32c2p = currprefs.cs_cd32cd = currprefs.cs_cd32nvram = 1;
+	akiko_init ();
+    }
 
     restore_u16 ();
     restore_u16 ();
@@ -1693,8 +1698,10 @@ void restore_akiko_finish(void)
     sys_command_cd_pause (DF_IOCTL, unitnum, 0);
     sys_command_cd_stop (DF_IOCTL, unitnum);
     sys_command_cd_pause (DF_IOCTL, unitnum, 1);
-    if (cdrom_playing)
+    if (cdrom_playing) {
+        sys_command_cd_pause (DF_IOCTL, unitnum, 0);
 	sys_command_cd_play (DF_IOCTL, unitnum, last_play_pos, last_play_end, 0);
+    }
 }
 
 #endif
diff --git a/cia.c b/cia.c
index 2f581ec1..e2b05270 100755
--- a/cia.c
+++ b/cia.c
@@ -1082,9 +1082,11 @@ void CIA_reset (void)
 	DISK_select (ciabprb);
     }
 #ifdef CD32
-    akiko_reset ();
-    if (!akiko_init ())
-	currprefs.cs_cd32cd = changed_prefs.cs_cd32cd = 0;
+    if (savestate_state != STATE_RESTORE) {
+	akiko_reset ();
+	if (!akiko_init ())
+	    currprefs.cs_cd32cd = changed_prefs.cs_cd32cd = 0;
+    }
 #endif
 }
 
diff --git a/custom.c b/custom.c
index 40c16083..744bddaf 100755
--- a/custom.c
+++ b/custom.c
@@ -10,13 +10,15 @@
 
 //#define CUSTOM_DEBUG
 #define SPRITE_DEBUG 0
-#define SPRITE_DEBUG_MINY 0x6a
-#define SPRITE_DEBUG_MAXY 0x70
+#define SPRITE_DEBUG_MINY 0xb0
+#define SPRITE_DEBUG_MAXY 0xf8
 #define SPR0_HPOS 0x15
 #define MAX_SPRITES 8
 #define SPRITE_COLLISIONS
 #define SPEEDUP
 
+#define SPRBORDER 0
+
 #include "sysconfig.h"
 #include "sysdeps.h"
 
@@ -78,7 +80,7 @@ void uae_abort (const char *format,...)
     _vsnprintf(buffer, sizeof (buffer) - 1, format, parms );
     va_end (parms);
     if (nomore) {
-	write_log (buffer);
+	write_log ("%s\n", buffer);
 	return;
     }
     gui_message (buffer);
@@ -1933,7 +1935,7 @@ static void record_sprite (int line, int num, int sprxp, uae_u16 *data, uae_u16
     attachment =  (spr[num & ~1].armed && ((sprctl[num | 1] & 0x80) || (!(currprefs.chipset_mask & CSMASK_AGA) && (sprctl[num & ~1] & 0x80))));
 
     /* Try to coalesce entries if they aren't too far apart  */
-    if (! next_sprite_forced && e[-1].max + sprite_width >= sprxp && (attachment || ((bplcon4 >> 4) & 15) == ((bplcon4 >> 0) & 15))) {
+    if (!next_sprite_forced && e[-1].max + sprite_width >= sprxp && (attachment || ((bplcon4 >> 4) & 15) == ((bplcon4 >> 0) & 15))) {
 	e--;
     } else {
 	next_sprite_entry++;
@@ -1987,22 +1989,43 @@ static void record_sprite (int line, int num, int sprxp, uae_u16 *data, uae_u16
     }
 }
 
-STATIC_INLINE int SSCAN2_H10 (int x)
+static void add_sprite (int *countp, int num, int sprxp, int window_xp, int posns[], int nrs[], int window_width)
 {
-    if (fmode & 0x8000)
-	return x & ~(0x100 << sprite_buffer_res);
-    return x;
+    int count = *countp;
+    int j, bestp;
+#ifdef AGA
+    if ( !(bplcon3 & 2) && /* sprites outside playfields enabled? */
+        ((thisline_decision.diwfirstword >= 0 && window_xp + window_width < thisline_decision.diwfirstword)
+        || (thisline_decision.diwlastword >= 0 && window_xp > thisline_decision.diwlastword)))
+	    return;
+#endif
+    /* Sort the sprites in order of ascending X position before recording them.  */
+    for (bestp = 0; bestp < count; bestp++) {
+        if (posns[bestp] > sprxp)
+	    break;
+	if (posns[bestp] == sprxp && nrs[bestp] < num)
+	    break;
+    }
+    for (j = count; j > bestp; j--) {
+        posns[j] = posns[j - 1];
+        nrs[j] = nrs[j - 1];
+    }
+    posns[j] = sprxp;
+    nrs[j] = num;
+    count++;
+    *countp = count;
 }
 
 static void decide_sprites (int hpos)
 {
-    int nrs[MAX_SPRITES], posns[MAX_SPRITES];
+    int nrs[MAX_SPRITES * 2], posns[MAX_SPRITES * 2];
     int count, i;
      /* apparantly writes to custom registers happen in the 3/4th of cycle
       * and sprite xpos comparator sees it immediately */
     int point = hpos * 2 - 3;
     int width = sprite_width;
     int window_width = (width << lores_shift) >> sprres;
+    int sscanmask = 0x100 << sprite_buffer_res;
 
     if (nodraw () || hpos < 0x14 || nr_armed == 0 || point == last_sprite_point)
 	return;
@@ -2012,10 +2035,12 @@ static void decide_sprites (int hpos)
 
     count = 0;
     for (i = 0; i < MAX_SPRITES; i++) {
-	int sprxp = SSCAN2_H10 (spr[i].xpos);
+	int sprxp = (fmode & 0x8000) ? (spr[i].xpos & ~sscanmask) : spr[i].xpos;
 	int hw_xp = sprxp >> sprite_buffer_res;
 	int window_xp = coord_hw_to_window_x (hw_xp) + (DIW_DDF_OFFSET << lores_shift);
-	int j, bestp;
+
+	if (spr[i].xpos < 0)
+	    continue;
 
 	if (!((debug_sprite_mask) & (1 << i)))
 	    continue;
@@ -2023,33 +2048,22 @@ static void decide_sprites (int hpos)
 	if (! spr[i].armed)
 	    continue;
 
-	if (sprxp < 0 || hw_xp <= last_sprite_point || hw_xp > point)
-	    continue;
+	if (hw_xp > last_sprite_point && hw_xp <= point)
+	    add_sprite (&count, i, sprxp, window_xp, posns, nrs, window_width);
 
-#ifdef AGA
-	if ( !(bplcon3 & 2) && /* sprites outside playfields enabled? */
-	    ((thisline_decision.diwfirstword >= 0 && window_xp + window_width < thisline_decision.diwfirstword)
-	    || (thisline_decision.diwlastword >= 0 && window_xp > thisline_decision.diwlastword)))
-	    continue;
-#endif
-	/* Sort the sprites in order of ascending X position before recording them.  */
-	for (bestp = 0; bestp < count; bestp++) {
-	    if (posns[bestp] > sprxp)
-		break;
-	    if (posns[bestp] == sprxp && nrs[bestp] < i)
-		break;
-	}
-	for (j = count; j > bestp; j--) {
-	    posns[j] = posns[j-1];
-	    nrs[j] = nrs[j-1];
+	/* SSCAN2-bit is fun.. */
+	if ((fmode & 0x8000) && !(sprxp & sscanmask)) {
+	    sprxp |= sscanmask;
+	    hw_xp = sprxp >> sprite_buffer_res;
+	    window_xp = coord_hw_to_window_x (hw_xp) + (DIW_DDF_OFFSET << lores_shift);
+	    if (hw_xp > last_sprite_point && hw_xp <= point)
+		add_sprite (&count, MAX_SPRITES + i, sprxp, window_xp, posns, nrs, window_width);
 	}
-	posns[j] = sprxp;
-	nrs[j] = i;
-	count++;
     }
+
     for (i = 0; i < count; i++) {
-	int nr = nrs[i];
-	record_sprite (next_lineno, nr, SSCAN2_H10 (spr[nr].xpos), sprdata[nr], sprdatb[nr], sprctl[nr]);
+	int nr = nrs[i] & (MAX_SPRITES - 1);
+	record_sprite (next_lineno, nr, posns[i], sprdata[nr], sprdatb[nr], sprctl[nr]);
     }
     last_sprite_point = point;
 }
@@ -2887,6 +2901,9 @@ static void BPLCON0 (int hpos, uae_u16 v)
     else if (! (currprefs.chipset_mask & CSMASK_AGA))
 	v &= ~0x00B1;
 
+#if SPRBORDER
+    v |= 1;
+#endif
     if (bplcon0 == v)
 	return;
 
@@ -2950,6 +2967,9 @@ STATIC_INLINE void BPLCON3 (int hpos, uae_u16 v)
 	v &= 0x3f;
 	v |= 0x0c00;
     }
+#if SPRBORDER
+    v |= 2;
+#endif
     if (bplcon3 == v)
 	return;
     decide_line (hpos);
@@ -3248,7 +3268,7 @@ STATIC_INLINE void SPRxCTL_1 (uae_u16 v, int num, int hpos)
 #if SPRITE_DEBUG > 0
     if (vpos >= SPRITE_DEBUG_MINY && vpos <= SPRITE_DEBUG_MAXY) {
 	write_log ("%d:%d:SPR%dCTL %04.4X P=%06.6X VSTRT=%d VSTOP=%d HSTRT=%d D=%d A=%d CP=%x PC=%x\n",
-	    vpos, hpos, num, v, s->pt, s->vstart, s->vstop, s->xpos, spr[num].dmastate, spr[num].armed, cop_state.ip, m68k_getpc ());
+	    vpos, hpos, num, v, s->pt, s->vstart, s->vstop, s->xpos, spr[num].dmastate, spr[num].armed, cop_state.ip, M68K_GETPC);
     }
 #endif
 
@@ -3261,7 +3281,7 @@ STATIC_INLINE void SPRxPOS_1 (uae_u16 v, int num, int hpos)
 #if SPRITE_DEBUG > 0
     if (vpos >= SPRITE_DEBUG_MINY && vpos <= SPRITE_DEBUG_MAXY) {
 	write_log ("%d:%d:SPR%dPOS %04.4X P=%06.6X VSTRT=%d VSTOP=%d HSTRT=%d D=%d A=%d CP=%x PC=%x\n",
-	    vpos, hpos, num, v, s->pt, s->vstart, s->vstop, s->xpos, spr[num].dmastate, spr[num].armed, cop_state.ip, m68k_getpc ());
+	    vpos, hpos, num, v, s->pt, s->vstart, s->vstop, s->xpos, spr[num].dmastate, spr[num].armed, cop_state.ip, M68K_GETPC);
     }
 #endif
 }
@@ -3277,7 +3297,7 @@ STATIC_INLINE void SPRxDATA_1 (uae_u16 v, int num, int hpos)
 #if SPRITE_DEBUG > 1
     if (vpos >= SPRITE_DEBUG_MINY && vpos <= SPRITE_DEBUG_MAXY) {
 	write_log ("%d:%d:SPR%dDATA %04.4X P=%06.6X D=%d A=%d PC=%x\n",
-	    vpos, hpos, num, v, spr[num].pt, spr[num].dmastate, spr[num].armed, m68k_getpc ());
+	    vpos, hpos, num, v, spr[num].pt, spr[num].dmastate, spr[num].armed, M68K_GETPC);
     }
 #endif
 }
@@ -3292,7 +3312,7 @@ STATIC_INLINE void SPRxDATB_1 (uae_u16 v, int num, int hpos)
 #if SPRITE_DEBUG > 1
     if (vpos >= SPRITE_DEBUG_MINY && vpos <= SPRITE_DEBUG_MAXY) {
 	write_log ("%d:%d:SPR%dDATB %04.4X P=%06.6X D=%d A=%d PC=%x\n",
-	    vpos, hpos, num, v, spr[num].pt, spr[num].dmastate, spr[num].armed, m68k_getpc ());
+	    vpos, hpos, num, v, spr[num].pt, spr[num].dmastate, spr[num].armed, M68K_GETPC);
     }
 #endif
 }
@@ -3855,7 +3875,7 @@ STATIC_INLINE uae_u16 sprite_fetch (struct sprite *s, int dma, int hpos, int cyc
     uae_u16 data = last_custom_value;
     if (dma) {
 	data = last_custom_value = chipmem_agnus_wget (s->pt);
-	alloc_cycle(hpos, CYCLE_SPRITE);
+	alloc_cycle (hpos, CYCLE_SPRITE);
     }
     s->pt += 2;
     return data;
@@ -3893,6 +3913,12 @@ STATIC_INLINE void do_sprites_1 (int num, int cycle, int hpos)
 	    write_log ("%d:%d:SPR%d STOP\n", vpos, hpos, num);
 #endif
 	s->dmastate = 0;
+#if 1
+	if (vpos == s->vstop) {
+	    spr_arm (num, 0);
+	    //return;
+	}
+#endif
     }
     if (!dmaen (DMA_SPRITE))
 	return;
@@ -4018,7 +4044,7 @@ static void do_sprites (int hpos)
 	    cycle = 1;
 	    break;
 	}
-	if (cycle >= 0)
+	if (cycle >= 0 && num >= 0 && num < MAX_SPRITES)
 	    do_sprites_1 (num, cycle, i);
     }
     last_sprite_hpos = hpos;
diff --git a/debug.c b/debug.c
index 05a8274e..b38c7f4a 100755
--- a/debug.c
+++ b/debug.c
@@ -2295,17 +2295,17 @@ static void debugtest_set (char **inptr)
 
 static void debug_sprite (char **inptr)
 {
-    uaecptr addr, addr2;
+    uaecptr saddr, addr, addr2;
     int xpos, xpos_ecs;
     int ypos, ypos_ecs;
     int ypose, ypose_ecs;
     int attach;
     uae_u64 w1, w2, ww1, ww2;
     int size = 1, width;
-    int ecs = 0;
-    int sh10 = 0;
-    int y;
+    int ecs, sh10;
+    int y, i;
     char tmp[80];
+    int max = 2;
 
     addr2 = 0;
     ignore_ws (inptr);
@@ -2321,94 +2321,111 @@ static void debug_sprite (char **inptr)
 	if (size != 1 && size != 2 && size != 4)
 	    size = 1;
     }
-    width = size * 16;
-    w1 = get_word (addr);
-    w2 = get_word (addr + size * 2);
-    ypos = w1 >> 8;
-    xpos = w1 & 255;
-    ypose = w2 >> 8;
-    attach = (w2 & 0x80) ? 1 : 0;
-    if (w2 & 4)
-	ypos |= 256;
-    if (w2 & 2)
-	ypose |= 256;
-    ypos_ecs = ypos;
-    ypose_ecs = ypose;
-    if (w2 & 0x40)
-	ypos_ecs |= 512;
-    if (w2 & 0x20)
-	ypose_ecs |= 512;
-    xpos <<= 1;
-    if (w2 & 0x01)
-	xpos |= 1;
-    xpos_ecs = xpos << 2;
-    if (w2 & 0x10)
-	xpos_ecs |= 2;
-    if (w2 & 0x08)
-	xpos_ecs |= 1;
-    if (w2 & (0x40 | 0x20 | 0x10 | 0x08))
-	ecs = 1;
-    if (w1 & 0x80)
-	sh10 = 1;
-    for (y = ypos; y < ypose; y++) {
-	int x;
-	addr += size * 4;
-	if (addr2)
-	    addr2 += size * 4;
-	if (size == 1) {
-	    w1 = get_word (addr);
-	    w2 = get_word (addr + 2);
-	    if (addr2) {
-		ww1 = get_word (addr2);
-		ww2 = get_word (addr2 + 2);
-	    }
-	} else if (size == 2) {
-	    w1 = get_long (addr);
-	    w2 = get_long (addr + 4);
-	    if (addr2) {
-		ww1 = get_long (addr2);
-		ww2 = get_long (addr2 + 4);
-	    }
-	} else if (size == 4) {
-	    w1 = get_long (addr) << 16;
-	    w2 = get_long (addr + 4) << 16;
-	    w1 <<= 16;
-	    w2 <<= 16;
-	    w1 |= get_long (addr);
-	    w2 |= get_long (addr2 + 4);
-	    if (addr2) {
-		ww1 = get_long (addr2) << 16;
-		ww2 = get_long (addr2 + 4) << 16;
-		ww1 <<= 16;
-		ww2 <<= 16;
-		ww1 |= get_long (addr2);
-		ww2 |= get_long (addr2 + 4);
-	    }
-	}
+    for (;;) {
+	ecs = 0;
+	sh10 = 0;
+	saddr = addr;
 	width = size * 16;
-	for (x = 0; x < width; x++) {
-	    int v1 = (w1 >> (width - x)) & 1;
-	    int v2 = (w2 >> (width - x)) & 1;
-	    int v = v1 * 2 + v2;
-	    if (addr2) {
-		int vv1 = (ww1 >> (width - x)) & 1;
-		int vv2 = (ww2 >> (width - x)) & 1;
-		int vv = vv1 * 2 + vv2;
-		v *= 4;
-		v += vv;
-		tmp[x] = v >= 10 ? 'A' + v - 10 : v + '0';
-	    } else {
-		tmp[x] = v + '0';
+	w1 = get_word (addr);
+	w2 = get_word (addr + size * 2);
+	console_out_f ("    %06X ", addr);
+	for (i = 0; i < size * 2; i++)
+	    console_out_f ("%04X ", get_word (addr + i * 2));
+	console_out_f ("\n");
+	ypos = w1 >> 8;
+	xpos = w1 & 255;
+	ypose = w2 >> 8;
+	attach = (w2 & 0x80) ? 1 : 0;
+	if (w2 & 4)
+	    ypos |= 256;
+	if (w2 & 2)
+	    ypose |= 256;
+	ypos_ecs = ypos;
+	ypose_ecs = ypose;
+	if (w2 & 0x40)
+	    ypos_ecs |= 512;
+	if (w2 & 0x20)
+	    ypose_ecs |= 512;
+	xpos <<= 1;
+	if (w2 & 0x01)
+	    xpos |= 1;
+	xpos_ecs = xpos << 2;
+	if (w2 & 0x10)
+	    xpos_ecs |= 2;
+	if (w2 & 0x08)
+	    xpos_ecs |= 1;
+	if (w2 & (0x40 | 0x20 | 0x10 | 0x08))
+	    ecs = 1;
+	if (w1 & 0x80)
+	    sh10 = 1;
+	for (y = ypos; y < ypose; y++) {
+	    int x;
+	    addr += size * 4;
+	    if (addr2)
+		addr2 += size * 4;
+	    if (size == 1) {
+		w1 = get_word (addr);
+		w2 = get_word (addr + 2);
+		if (addr2) {
+		    ww1 = get_word (addr2);
+		    ww2 = get_word (addr2 + 2);
+		}
+	    } else if (size == 2) {
+		w1 = get_long (addr);
+		w2 = get_long (addr + 4);
+		if (addr2) {
+		    ww1 = get_long (addr2);
+		    ww2 = get_long (addr2 + 4);
+		}
+	    } else if (size == 4) {
+		w1 = get_long (addr) << 16;
+		w2 = get_long (addr + 4) << 16;
+		w1 <<= 16;
+		w2 <<= 16;
+		w1 |= get_long (addr);
+		w2 |= get_long (addr2 + 4);
+		if (addr2) {
+		    ww1 = get_long (addr2) << 16;
+		    ww2 = get_long (addr2 + 4) << 16;
+		    ww1 <<= 16;
+		    ww2 <<= 16;
+		    ww1 |= get_long (addr2);
+		    ww2 |= get_long (addr2 + 4);
+		}
 	    }
+	    width = size * 16;
+	    for (x = 0; x < width; x++) {
+		int v1 = (w1 >> (width - x)) & 1;
+		int v2 = (w2 >> (width - x)) & 1;
+		int v = v1 * 2 + v2;
+		if (addr2) {
+		    int vv1 = (ww1 >> (width - x)) & 1;
+		    int vv2 = (ww2 >> (width - x)) & 1;
+		    int vv = vv1 * 2 + vv2;
+		    v *= 4;
+		    v += vv;
+		    tmp[x] = v >= 10 ? 'A' + v - 10 : v + '0';
+		} else {
+		    tmp[x] = v + '0';
+		}
+	    }
+	    tmp[width] = 0;
+	    console_out_f ("%3d %06X %s\n", y, addr, tmp);
 	}
-	tmp[width] = 0;
-	console_out_f ("%3d: %s\n", y, tmp);
+
+	console_out_f ("Sprite address %08X, width = %d\n", saddr, size * 16);
+	console_out_f ("OCS: StartX=%d StartY=%d EndY=%d\n", xpos, ypos, ypose);
+	console_out_f ("ECS: StartX=%d (%d.%d) StartY=%d EndY=%d%s\n", xpos_ecs, xpos_ecs / 4, xpos_ecs & 3, ypos_ecs, ypose_ecs, ecs ? " (*)" : "");
+	console_out_f ("Attach: %d. AGA SSCAN/SH10 bit: %d\n", attach, sh10);
+
+        addr += size * 4;
+	if (get_word (addr) == 0 && get_word (addr + size * 4) == 0)
+	    break;
+	max--;
+	if (max <= 0)
+	    break;
     }
 
-    console_out_f ("Sprite address %08X, width = %d\n", addr, size * 16);
-    console_out_f ("OCS: StartX=%d StartY=%d EndY=%d\n", xpos, ypos, ypose);
-    console_out_f ("ECS: StartX=%d (%d.%d) StartY=%d EndY=%d%s\n", xpos_ecs, xpos_ecs / 4, xpos_ecs & 3, ypos_ecs, ypose_ecs, ecs ? " (*)" : "");
-    console_out_f ("Attach: %d. AGA SSCAN/SH10 bit: %d\n", attach, sh10);
 }
 
 static void disk_debug (char **inptr)
@@ -2597,7 +2614,7 @@ static void debug_1 (void)
 		} else {
 		    next_char (&inptr);
 		    if (more_params (&inptr))
-			debug_sprite_mask = readint (&inptr);
+			debug_sprite_mask = readhex (&inptr);
 		    console_out_f ("sprite mask: %02.2X\n", debug_sprite_mask);
 		}
 	    } else {
diff --git a/drawing.c b/drawing.c
index 6d8a0372..a496e372 100755
--- a/drawing.c
+++ b/drawing.c
@@ -330,6 +330,7 @@ static int pixels_offset;
 static int src_pixel;
 /* How many pixels in window coordinates which are to the left of the left border.  */
 static int unpainted;
+static int seen_sprites;
 
 /* Initialize the variables necessary for drawing a line.
  * This involves setting up start/stop positions and display window
@@ -374,7 +375,6 @@ static void pfield_init_linetoscr (void)
 #ifdef AGA
     if (brdsprt && dip_for_drawing->nr_sprites) {
 	int min = visible_right_border, max = visible_left_border, i;
-	int posdiff = sprite_buffer_res - bplres;
 	for (i = 0; i < dip_for_drawing->nr_sprites; i++) {
 	    int x;
 	    x = curr_sprite_entries[dip_for_drawing->first_sprite_entry + i].pos;
@@ -384,15 +384,8 @@ static void pfield_init_linetoscr (void)
 	    if (x > max)
 		max = x;
 	}
-	min += (DIW_DDF_OFFSET - DISPLAY_LEFT_SHIFT) << sprite_buffer_res;
-	max += (DIW_DDF_OFFSET - DISPLAY_LEFT_SHIFT) << sprite_buffer_res;
-	if (posdiff < 0) {
-	    min <<= -posdiff;
-	    max <<= -posdiff;
-	} else {
-	    min >>= posdiff;
-	    max >>= posdiff;
-	}
+	min = coord_hw_to_window_x (min >> sprite_buffer_res) + (DIW_DDF_OFFSET << lores_shift);
+	max = coord_hw_to_window_x (max >> sprite_buffer_res) + (DIW_DDF_OFFSET << lores_shift);
 	if (min < playfield_start)
 	    playfield_start = min;
 	if (playfield_start < visible_left_border)
@@ -413,8 +406,16 @@ static void pfield_init_linetoscr (void)
     unpainted = visible_left_border < playfield_start ? 0 : visible_left_border - playfield_start;
     src_pixel = MAX_PIXELS_PER_LINE + res_shift_from_window (playfield_start - native_ddf_left + unpainted);
 
+    if (seen_sprites) {
+	/* clear previous sprite data storage line */
+	memset (spritepixels, 0, sizeof (spritepixels));
+	seen_sprites = 0;
+    }
     if (dip_for_drawing->nr_sprites == 0)
 	return;
+    if (seen_sprites < 0)
+	memset (spritepixels, 0, sizeof (spritepixels));
+    seen_sprites = 1;
     /* Must clear parts of apixels.  */
     if (linetoscr_diw_start < native_ddf_left) {
 	int size = res_shift_from_window (native_ddf_left - linetoscr_diw_start);
@@ -2129,6 +2130,7 @@ static void init_drawing_frame (void)
     thisframe_last_drawn_line = -1;
 
     drawing_color_matches = -1;
+    seen_sprites = -1;
 }
 
 /*
diff --git a/gencpu.c b/gencpu.c
index e3af491d..f351d413 100755
--- a/gencpu.c
+++ b/gencpu.c
@@ -1677,6 +1677,19 @@ static void gen_opcode (unsigned long int opcode)
 	sync_m68k_pc ();
 	fill_prefetch_full ();
 	break;
+    case i_LPSTOP: /* 68060 */
+	printf ("\tuae_u16 sw = get_iword (regs, 2);\n");
+	printf ("\tuae_u16 sr;\n");
+	printf ("\tif (sw != (0x100|0x80|0x40)) { Exception (4, regs, 0); goto %s; }\n", endlabelstr);
+	printf ("\tsr = get_iword (regs, 4);\n");
+	printf ("\tif (!(sr & 0x8000)) { Exception (8, regs, 0); goto %s; }\n", endlabelstr);
+	printf ("\tregs->sr = sr;\n");
+	printf ("\tMakeFromSR (regs);\n");
+	printf ("\tm68k_setstopped(regs, 1);\n");
+	m68k_pc_offset += 4;
+	sync_m68k_pc ();
+	fill_prefetch_full ();
+	break;
     case i_RTE:
 	if (cpu_level == 0) {
 	    genamode (Aipi, "7", sz_word, "sr", 1, 0, GF_NOREFILL);
@@ -2900,12 +2913,19 @@ static void gen_opcode (unsigned long int opcode)
 	}
 	break;
 
-    case i_MMUOP:
+    case i_PFLUSHN:
+    case i_PFLUSH:
+    case i_PFLUSHAN:
+    case i_PFLUSHA:
+    case i_PLPAR:
+    case i_PLPAW:
+    case i_PTESTR:
+    case i_PTESTW:
 	genamode (curi->smode, "srcreg", curi->size, "extra", 1, 0, 0);
 	sync_m68k_pc ();
 	printf ("\tmmu_op (opcode, regs, extra);\n");
 	break;
-    case i_MMUOP30A:
+    case i_MMUOP030:
 	printf ("\tuaecptr pc = m68k_getpc (regs);\n");
 	if (curi->smode == Areg || curi->smode == Dreg)
 	    printf("\tuae_u16 extraa = 0;\n");
@@ -2914,12 +2934,6 @@ static void gen_opcode (unsigned long int opcode)
 	sync_m68k_pc ();
 	printf ("\tmmu_op30 (pc, opcode, regs, 1, extraa);\n");
 	break;
-    case i_MMUOP30B:
-	printf ("\tuaecptr pc = m68k_getpc (regs);\n");
-	genamode (curi->smode, "srcreg", curi->size, "extra", 0, 0, 0);
-	sync_m68k_pc ();
-	printf ("\tmmu_op30 (pc, opcode, regs, 0, 0);\n");
-	break;
     default:
 	abort ();
 	break;
diff --git a/genlinetoscr.c b/genlinetoscr.c
index df092518..c89af7ca 100755
--- a/genlinetoscr.c
+++ b/genlinetoscr.c
@@ -135,7 +135,6 @@ static void out_linetoscr_do_dstpix (DEPTH_T bpp, HMODE_T hmode, int aga, CMODE_
     else if (aga && cmode == CMODE_DUALPF) {
 	outln (		"    if (spritepixels[spix]) {");
 	outln (		"        dpix_val = colors_for_drawing.acolors[spritepixels[spix]];");
-	outln (		"        spritepixels[spix] = 0;");
 	outln (		"    } else {");
 	outln (		"        unsigned int val = lookup[spix_val];");
 	outln (		"        if (lookup_no[spix_val] == 2)");
@@ -179,7 +178,6 @@ static void out_sprite (int off)
 {
     outlnf ( "    if (spritepixels[sprx + MAX_PIXELS_PER_LINE * %d]) {", off);
     outlnf ( "        buf[dpix] = colors_for_drawing.acolors[spritepixels[sprx + MAX_PIXELS_PER_LINE * %d]];", off);
-    outlnf ( "        spritepixels[sprx + MAX_PIXELS_PER_LINE * %d] = 0;", off);
     outlnf ( "    } else {");
     outlnf ( "        buf[dpix] = out_val;");
     outlnf ( "    }");
diff --git a/include/drawing.h b/include/drawing.h
index c510c284..b8bbaa6b 100755
--- a/include/drawing.h
+++ b/include/drawing.h
@@ -197,7 +197,7 @@ extern struct color_change *color_changes[2];
 extern struct color_change color_changes[2][MAX_REG_CHANGE];
 #endif
 
-extern struct color_entry color_tables[2][(MAXVPOS+1) * 2];
+extern struct color_entry color_tables[2][(MAXVPOS + 1) * 2];
 extern struct color_entry *curr_color_tables, *prev_color_tables;
 
 extern struct sprite_entry *curr_sprite_entries, *prev_sprite_entries;
@@ -234,10 +234,10 @@ struct draw_info {
 
 extern int next_sprite_entry;
 
-extern struct decision line_decisions[2 * (MAXVPOS+1) + 1];
-extern struct draw_info line_drawinfo[2][2 * (MAXVPOS+1) + 1];
+extern struct decision line_decisions[2 * (MAXVPOS + 1) + 1];
+extern struct draw_info line_drawinfo[2][2 * (MAXVPOS + 1) + 1];
 
-extern uae_u8 line_data[(MAXVPOS+1) * 2][MAX_PLANES * MAX_WORDS_PER_LINE * 2];
+extern uae_u8 line_data[(MAXVPOS + 1) * 2][MAX_PLANES * MAX_WORDS_PER_LINE * 2];
 
 extern uae_u8 *real_bplpt[8];
 
diff --git a/include/gfxfilter.h b/include/gfxfilter.h
index fdd5daf1..d37fc914 100755
--- a/include/gfxfilter.h
+++ b/include/gfxfilter.h
@@ -67,4 +67,7 @@ struct uae_filter
 extern struct uae_filter uaefilters[];
 extern struct uae_filter *usedfilter;
 
+uae_u8 *getfilterrect1 (RECT *sr, RECT *dr, int dst_depth, int aw, int ah, int scale, int temp_width, int temp_height, uae_u8 *dptr, int pitch);
+void getfilterrect2 (RECT *sr, RECT *dr, int dst_width, int dst_height, int aw, int ah, int scale, int temp_width, int temp_height);
+
 #endif
diff --git a/include/readcpu.h b/include/readcpu.h
index bf6a5597..14195514 100755
--- a/include/readcpu.h
+++ b/include/readcpu.h
@@ -28,7 +28,9 @@ ENUMDECL {
     i_PACK, i_UNPK, i_TAS, i_BKPT, i_CALLM, i_RTM, i_TRAPcc, i_MOVES,
     i_FPP, i_FDBcc, i_FScc, i_FTRAPcc, i_FBcc, i_FSAVE, i_FRESTORE,
     i_CINVL, i_CINVP, i_CINVA, i_CPUSHL, i_CPUSHP, i_CPUSHA, i_MOVE16,
-    i_MMUOP30A, i_MMUOP30B, i_MMUOP
+    i_MMUOP030, i_PFLUSHN, i_PFLUSH, i_PFLUSHAN, i_PFLUSHA,
+    i_PLPAR, i_PLPAW, i_PTESTR, i_PTESTW,
+    i_LPSTOP
 } ENUMNAME (instrmnem);
 
 extern struct mnemolookup {
diff --git a/jit/gencomp.c b/jit/gencomp.c
index 2b065f56..fb1c744c 100755
--- a/jit/gencomp.c
+++ b/jit/gencomp.c
@@ -2843,9 +2843,16 @@ gen_opcode (unsigned long int opcode)
 	//}
 	break;
 
-     case i_MMUOP:
-     case i_MMUOP30A:
-     case i_MMUOP30B:
+     case i_MMUOP030:
+     case i_PFLUSHN:
+     case i_PFLUSH:
+     case i_PFLUSHAN:
+     case i_PFLUSHA:
+     case i_PLPAR:
+     case i_PLPAW:
+     case i_PTESTR:
+     case i_PTESTW:
+     case i_LPSTOP:
 	isjump;
 	failure;
 	break;
diff --git a/memory.c b/memory.c
index e602be4d..c60a751d 100755
--- a/memory.c
+++ b/memory.c
@@ -39,9 +39,12 @@ int special_mem;
 
 static int canjit (void)
 {
+    return 1;
+#if 0
     if (canbang || currprefs.cpu_model >= 68020)
 	return 1;
     return 0;
+#endif
 }
 static void nocanbang(void)
 {
@@ -590,7 +593,7 @@ int decode_cloanto_rom_do (uae_u8 *mem, int size, int real_size)
 		t = keysize - 1;
 	}
     }
-    return get_keyring();
+    return 0;
 }
 
 static int decode_rekick_rom_do (uae_u8 *mem, int size, int real_size)
@@ -656,10 +659,10 @@ int load_keyring (struct uae_prefs *p, char *path)
     int keyid;
     int cnt, i;
 
-    free_keyring();
+    free_keyring ();
     keyid = 0;
-    keybuf = target_load_keyfile(p, path, &keysize, tmp);
-    addkey(&keyid, keybuf, keysize, tmp);
+    keybuf = target_load_keyfile (p, path, &keysize, tmp);
+    addkey (&keyid, keybuf, keysize, tmp);
     for (i = 0; keyids[i] >= 0 && keyid < ROM_KEY_NUM; i++) {
 	struct romdata *rd = getromdatabyid (keyids[i]);
 	char *s;
@@ -2826,7 +2829,7 @@ static shmpiece *find_shmpiece (uae_u8 *base)
     while (x && x->native_address != base)
 	x = x->next;
     if (!x) {
-	write_log ("NATMEM: Failure to find mapping at %p\n",base);
+	write_log ("NATMEM: Failure to find mapping at %08X, %p\n", base - NATMEM_OFFSET, base);
 	dumplist ();
 	nocanbang();
 	return 0;
@@ -3180,7 +3183,7 @@ void memory_reset (void)
 	kickmem_mask = 524288 - 1;
 	if (!load_kickstart ()) {
 	    if (strlen (currprefs.romfile) > 0) {
-		write_log ("%s\n", currprefs.romfile);
+		write_log ("Failed to open '%s'\n", currprefs.romfile);
 		notify_user (NUMSG_NOROM);
 	    }
 #ifdef AUTOCONFIG
diff --git a/newcpu.c b/newcpu.c
index 2f288a79..cc11589d 100755
--- a/newcpu.c
+++ b/newcpu.c
@@ -1869,14 +1869,6 @@ void mmu_op (uae_u32 opcode, struct regstruct *regs, uae_u32 extra)
 	if (currprefs.cpu_model == 68060) {
 #if MMUOP_DEBUG > 0
 	    write_log ("PLPA\n");
-#endif
-	    return;
-	}
-    } else if (opcode == 0xff00 && extra == 0x01c0) {
-	/* LPSTOP */
-	if (currprefs.cpu_model == 68060) {
-#if MMUOP_DEBUG > 0
-	    write_log ("LPSTOP\n");
 #endif
 	    return;
 	}
@@ -2579,15 +2571,40 @@ void m68k_disasm_2 (char *buf, int bufsize, uaecptr addr, uaecptr *nextpc, int c
 	 default: strcat (instrname, "   "); break;
 	}
 
-	if (dp->suse) {
-	    newpc = m68k_getpc (&regs) + m68kpc_offset;
-	    newpc += ShowEA (0, opcode, dp->sreg, dp->smode, dp->size, instrname, seaddr, safemode);
-	}
-	if (dp->suse && dp->duse)
-	    strcat (instrname, ",");
-	if (dp->duse) {
-	    newpc = m68k_getpc (&regs) + m68kpc_offset;
-	    newpc += ShowEA (0, opcode, dp->dreg, dp->dmode, dp->size, instrname, deaddr, safemode);
+	if (lookup->mnemo == i_MOVEC2 || lookup->mnemo == i_MOVE2C) {
+	    uae_u16 imm = get_iword_1 (m68kpc_offset) & 0xffff;
+	    uae_u16 creg = imm & 0x0fff;
+	    uae_u16 r = imm >> 12;
+	    char regs[16], *cname = "?";
+	    int i;
+	    for (i = 0; m2cregs[i].regname; i++) {
+		if (m2cregs[i].regno == creg)
+		    break;
+	    }
+	    sprintf (regs, "%c%d", r >= 8 ? 'A' : 'D', r >= 8 ? r - 8 : r);
+	    if (m2cregs[i].regname)
+		cname = m2cregs[i].regname;
+	    if (lookup->mnemo == i_MOVE2C) {
+		strcat (instrname, regs);
+		strcat (instrname, ",");
+		strcat (instrname, cname);
+	    } else {
+		strcat (instrname, cname);
+		strcat (instrname, ",");
+		strcat (instrname, regs);
+	    }
+	    m68kpc_offset += 2;
+	} else {
+	    if (dp->suse) {
+		newpc = m68k_getpc (&regs) + m68kpc_offset;
+		newpc += ShowEA (0, opcode, dp->sreg, dp->smode, dp->size, instrname, seaddr, safemode);
+	    }
+	    if (dp->suse && dp->duse)
+		strcat (instrname, ",");
+	    if (dp->duse) {
+		newpc = m68k_getpc (&regs) + m68kpc_offset;
+		newpc += ShowEA (0, opcode, dp->dreg, dp->dmode, dp->size, instrname, deaddr, safemode);
+	    }
 	}
 
 	for (i = 0; i < (m68kpc_offset - oldpc) / 2; i++) {
@@ -2595,9 +2612,6 @@ void m68k_disasm_2 (char *buf, int bufsize, uaecptr addr, uaecptr *nextpc, int c
 	}
 	while (i++ < 5)
 	    buf = buf_out (buf, &bufsize, "     ");
-    if (strlen (instrname) > 79) {
-	int i = 0;
-    }
 
 	buf = buf_out (buf, &bufsize, instrname);
 
diff --git a/od-win32/direct3d.c b/od-win32/direct3d.c
index 9754a096..afe1d034 100755
--- a/od-win32/direct3d.c
+++ b/od-win32/direct3d.c
@@ -398,7 +398,7 @@ typedef struct _D3DTLVERTEX {
     float tv;
 } D3DTLVERTEX, *LPD3DTLVERTEX;
 
-static void BlitRect(LPDIRECT3DDEVICE9 dev, LPDIRECT3DTEXTURE9 src,
+static void BlitRect (LPDIRECT3DDEVICE9 dev, LPDIRECT3DTEXTURE9 src,
 		     float left, float top, float right, float bottom, D3DCOLOR col,float z)
 {
     int i;
@@ -420,20 +420,20 @@ static void BlitRect(LPDIRECT3DDEVICE9 dev, LPDIRECT3DTEXTURE9 src,
     verts[3].sx = left - 0.5f; verts[3].sy = bottom - 0.5f; verts[3].sz = z;
 
     // set the texture
-    hr = IDirect3DDevice9_SetTexture(dev, 0, (IDirect3DBaseTexture9*)src);
+    hr = IDirect3DDevice9_SetTexture (dev, 0, (IDirect3DBaseTexture9*)src);
     if (FAILED (hr))
 	write_log ("IDirect3DDevice9_SetTexture failed: %s\n", D3D_ErrorString (hr));
 
-    hr = IDirect3DDevice9_SetVertexShader(dev, NULL);
+    hr = IDirect3DDevice9_SetVertexShader (dev, NULL);
     if (FAILED (hr))
 	write_log ("IDirect3DDevice9_SetVertexShader failed: %s\n", D3D_ErrorString (hr));
     // configure shader for vertex type
-    hr = IDirect3DDevice9_SetFVF(dev, D3DFVF_TLVERTEX);
+    hr = IDirect3DDevice9_SetFVF (dev, D3DFVF_TLVERTEX);
     if (FAILED (hr))
 	write_log ("IDirect3DDevice9_SetFVF failed: %s\n", D3D_ErrorString (hr));
 
     // draw the rectangle
-    hr = IDirect3DDevice9_DrawPrimitiveUP(dev, D3DPT_TRIANGLEFAN, 2, verts, sizeof(D3DTLVERTEX));
+    hr = IDirect3DDevice9_DrawPrimitiveUP (dev, D3DPT_TRIANGLEFAN, 2, verts, sizeof(D3DTLVERTEX));
     if (FAILED (hr))
 	write_log ("IDirect3DDevice9_DrawPrimitiveUP failed: %s\n", D3D_ErrorString (hr));
 }
@@ -443,69 +443,26 @@ static void BlitRect(LPDIRECT3DDEVICE9 dev, LPDIRECT3DTEXTURE9 src,
     tin_ = internal window size
     twidth/theight = texture size
 */
-#if 0
+#if 1
 static void calc (float *xp, float *yp, float *sxp, float *syp)
 {
-    float x, y, sx, sy, tx, ty;
-    double mx, my, fx, fy, mmx, mmy;
-
-    mmx = 1;
-    mmy = 1;//2.0 * window_h / tin_h;
-
-    fx = tin_w / 2;
-    fy = tin_h / 2;
-
-    fx = (twidth * window_w / tin_w) / 2;
-    fy = (theight * window_h / tin_h) / 2;
-
-    tx = ((currprefs.gfx_filter_horiz_zoom_mult + currprefs.gfx_filter_horiz_zoom / 4) / 1000.0);
-    ty = ((currprefs.gfx_filter_vert_zoom_mult + currprefs.gfx_filter_vert_zoom / 4) / 1000.0);
-
-    *sxp = (fx + fx / tx) * mmx;
-    *syp = (fy + fy / ty) * mmy;
-    *xp = (fx - fx / tx) * mmx;
-    *yp = (fy - fy / ty) * mmy;
-
-    mx = (currprefs.gfx_filter_horiz_offset / 1000.0) * fx;
-    my = (currprefs.gfx_filter_vert_offset / 1000.0) * fy;
-
-    *xp += mx;
-    *sxp += mx;
-    *yp += my;
-    *syp += my;
-
-    return;
-
-
-    fx = (twidth * window_w / tin_w) / 2;
-    fy = (theight * window_h / tin_h) / 2;
-
-    tx = fx / ((currprefs.gfx_filter_horiz_zoom_mult + currprefs.gfx_filter_horiz_zoom / 4) / 1000.0);
-    ty = fy / ((currprefs.gfx_filter_vert_zoom_mult + currprefs.gfx_filter_vert_zoom / 4) / 1000.0);
-
-    if (currprefs.gfx_lores)
-	tx /= 2;
-    if (!currprefs.gfx_linedbl)
-	ty /= 2;
-
-    mx = (currprefs.gfx_filter_horiz_offset / 1000.0) * fx;
-    my = (currprefs.gfx_filter_vert_offset / 1000.0) * fy;
-
-    x = -tx;
-    y = -ty;
-    sx = tx;
-    sy = ty;
-
-    x += fx + mx;
-    y += fy + my;
-    sx += tx + mx;
-    sy += ty + my;
-
-    *xp = x; *yp = y;
-    *sxp = sx; *syp = sy;
+    RECT sr, dr;
+    float mx = (float)twidth / tin_w;
+    float my = (float)theight / tin_h;
+    int aw = twidth * window_w / tin_w;
+    int ah = theight * window_h / tin_h;
+
+    mx = 1 / mx;
+    my = 1 / my;
+
+    getfilterrect2 (&sr, &dr, window_w, window_h, tin_w, tin_h, 1, tin_w, tin_h);
+    OffsetRect (&dr, aw / 2, ah / 2);
+    *xp = dr.left;
+    *yp = dr.top;
+    *sxp = *xp + (dr.right - dr.left) * mx;
+    *syp = *yp + (dr.bottom - dr.top) * my;
 }
-#endif
-#if 1
+#else
 static void calc (float *xp, float *yp, float *sxp, float *syp)
 {
     int xm, ym;
@@ -566,18 +523,18 @@ int D3D_locktexture (void)
     D3DLOCKED_RECT locked;
     HRESULT hr;
 
-    hr = IDirect3DDevice9_TestCooperativeLevel(d3ddev);
+    hr = IDirect3DDevice9_TestCooperativeLevel (d3ddev);
     if (FAILED (hr))
 	return 0;
 
-    IDirect3DDevice9_Clear(d3ddev, 0L, NULL, D3DCLEAR_TARGET, 0x00000000, 1.0f, 0L );
+    IDirect3DDevice9_Clear (d3ddev, 0L, NULL, D3DCLEAR_TARGET, 0x00000000, 1.0f, 0L );
 
-    hr = IDirect3DDevice9_BeginScene(d3ddev);
+    hr = IDirect3DDevice9_BeginScene (d3ddev);
     if (FAILED (hr)) {
 	write_log ("IDirect3DDevice9_BeginScene failed: %s\n", D3D_ErrorString (hr));
 	return 0;
     }
-    hr = IDirect3DTexture9_LockRect(texture, 0, &locked, NULL, D3DLOCK_DISCARD | D3DLOCK_NOSYSLOCK);
+    hr = IDirect3DTexture9_LockRect (texture, 0, &locked, NULL, D3DLOCK_DISCARD | D3DLOCK_NOSYSLOCK);
     if (FAILED (hr)) {
 	write_log ("IDirect3DTexture9_LockRect failed: %s\n", D3D_ErrorString (hr));
 	D3D_unlocktexture ();
@@ -596,17 +553,17 @@ void D3D_render (void)
 
     if (!d3d_enabled)
 	return;
-    if (FAILED(IDirect3DDevice9_TestCooperativeLevel(d3ddev)))
+    if (FAILED (IDirect3DDevice9_TestCooperativeLevel (d3ddev)))
 	return;
-    IDirect3DDevice9_Clear(d3ddev, 0L, NULL, D3DCLEAR_TARGET, 0x00000000, 1.0f, 0L );
-    hr = IDirect3DDevice9_BeginScene(d3ddev);
+    IDirect3DDevice9_Clear (d3ddev, 0L, NULL, D3DCLEAR_TARGET, 0x00000000, 1.0f, 0L );
+    hr = IDirect3DDevice9_BeginScene (d3ddev);
     if (FAILED (hr))
 	return;
     calc (&x, &y, &sx, &sy);
-    BlitRect(d3ddev, texture, x, y, sx, sy, 0xffffff, 0.1f);
+    BlitRect (d3ddev, texture, x, y, sx, sy, 0xffffff, 0.1f);
     if (scanlines_ok)
 	BlitRect (d3ddev, sltexture, 0, 0, required_sl_texture_w, required_sl_texture_h, 0xffffff, 0.2f);
-    IDirect3DDevice9_EndScene(d3ddev);
+    IDirect3DDevice9_EndScene (d3ddev);
     IDirect3DDevice9_Present (d3ddev, 0, 0, 0 ,0);
 }
 
diff --git a/od-win32/dxwrap.c b/od-win32/dxwrap.c
index ffb60757..dd04ed65 100755
--- a/od-win32/dxwrap.c
+++ b/od-win32/dxwrap.c
@@ -55,9 +55,9 @@ void DirectDraw_Release (void)
     dxdata.ddinit = 0;
     freemainsurface ();
     if (dxdata.fsmodeset)
-	IDirectDraw7_RestoreDisplayMode(dxdata.maindd);
+	IDirectDraw7_RestoreDisplayMode (dxdata.maindd);
     dxdata.fsmodeset = 0;
-    IDirectDraw7_SetCooperativeLevel(dxdata.maindd, dxdata.hwnd, DDSCL_NORMAL);
+    IDirectDraw7_SetCooperativeLevel (dxdata.maindd, dxdata.hwnd, DDSCL_NORMAL);
     releaser (dxdata.dclip, IDirectDrawClipper_Release);
     releaser (dxdata.maindd, IDirectDraw_Release);
     memset (&dxdata, 0, sizeof (dxdata));
@@ -97,9 +97,10 @@ int DirectDraw_Start (GUID *guid)
 	IDirect3D9_GetDeviceCaps (d3d, 0, D3DDEVTYPE_HAL, &d3dCaps);
 	dxdata.maxwidth = d3dCaps.MaxTextureWidth;
 	dxdata.maxheight = d3dCaps.MaxTextureHeight;
+	write_log ("Max hardware surface size: %dx%d\n", dxdata.maxwidth, dxdata.maxheight);
     }
 
-    if (SUCCEEDED(DirectDraw_GetDisplayMode ())) {
+    if (SUCCEEDED (DirectDraw_GetDisplayMode ())) {
 	dxdata.ddinit = 1;
 	dxdata.ddzeroguid = 1;
 	if (guid) {
@@ -138,7 +139,7 @@ static void clearsurf (LPDIRECTDRAWSURFACE7 surf)
     memset(&ddbltfx, 0, sizeof (ddbltfx));
     ddbltfx.dwFillColor = 0;
     ddbltfx.dwSize = sizeof (ddbltfx);
-    while (FAILED(ddrval = IDirectDrawSurface7_Blt (surf, NULL, NULL, NULL, DDBLT_WAIT | DDBLT_COLORFILL, &ddbltfx))) {
+    while (FAILED (ddrval = IDirectDrawSurface7_Blt (surf, NULL, NULL, NULL, DDBLT_WAIT | DDBLT_COLORFILL, &ddbltfx))) {
 	if (ddrval == DDERR_SURFACELOST) {
 	    ddrval = restoresurface (dxdata.primary);
 	    if (FAILED (ddrval))
@@ -161,10 +162,10 @@ int locksurface (LPDIRECTDRAWSURFACE7 surf, LPDDSURFACEDESC2 desc)
 {
     HRESULT ddrval;
     desc->dwSize = sizeof (*desc);
-    while (FAILED(ddrval = IDirectDrawSurface7_Lock (surf, NULL, desc, DDLOCK_SURFACEMEMORYPTR | DDLOCK_WAIT, NULL))) {
+    while (FAILED (ddrval = IDirectDrawSurface7_Lock (surf, NULL, desc, DDLOCK_SURFACEMEMORYPTR | DDLOCK_WAIT, NULL))) {
 	if (ddrval == DDERR_SURFACELOST) {
 	    ddrval = restoresurface (surf);
-	    if (FAILED(ddrval))
+	    if (FAILED (ddrval))
 	        return 0;
 	} else if (ddrval != DDERR_SURFACEBUSY) {
 	    write_log ("locksurface: %s\n", DXError (ddrval));
@@ -286,7 +287,7 @@ HRESULT DirectDraw_SetDisplayMode(int width, int height, int bits, int freq)
     if (dxdata.fsmodeset && dxdata.width == width && dxdata.height == height &&
 	dxdata.depth == bits && dxdata.freq == freq)
 	return DD_OK;
-    ddrval = IDirectDraw7_SetDisplayMode(dxdata.maindd, width, height, bits, freq, 0);
+    ddrval = IDirectDraw7_SetDisplayMode (dxdata.maindd, width, height, bits, freq, 0);
     if (FAILED (ddrval)) {
 	write_log ("IDirectDraw7_SetDisplayMode: %s\n", DXError (ddrval));
     } else {
@@ -328,7 +329,7 @@ HRESULT DirectDraw_SetClipper(HWND hWnd)
     ddrval = IDirectDrawSurface7_SetClipper (dxdata.primary, hWnd ? dxdata.dclip : NULL);
     if (FAILED (ddrval))
 	write_log ("IDirectDrawSurface7_SetClipper: %s\n", DXError (ddrval));
-    if(hWnd && SUCCEEDED(ddrval)) {
+    if(hWnd && SUCCEEDED (ddrval)) {
 	ddrval = IDirectDrawClipper_SetHWnd (dxdata.dclip, 0, hWnd);
 	if (FAILED (ddrval))
 	    write_log ("IDirectDrawClipper_SetHWnd: %s\n", DXError (ddrval));
@@ -342,7 +343,7 @@ char *outGUID (const GUID *guid)
     static char gb[64];
     if (guid == NULL)
 	return "NULL";
-    sprintf(gb, "%08X-%04X-%04X-%02X%02X%02X%02X%02X%02X%02X%02X",
+    sprintf (gb, "%08X-%04X-%04X-%02X%02X%02X%02X%02X%02X%02X%02X",
 	guid->Data1, guid->Data2, guid->Data3,
 	guid->Data4[0], guid->Data4[1], guid->Data4[2], guid->Data4[3],
 	guid->Data4[4], guid->Data4[5], guid->Data4[6], guid->Data4[7]);
@@ -352,7 +353,7 @@ char *outGUID (const GUID *guid)
 const char *DXError (HRESULT ddrval)
 {
     static char dderr[1000];
-    sprintf(dderr, "%08.8X S=%d F=%04.4X C=%04.4X (%d) (%s)",
+    sprintf (dderr, "%08.8X S=%d F=%04.4X C=%04.4X (%d) (%s)",
 	ddrval, (ddrval & 0x80000000) ? 1 : 0,
 	HRESULT_FACILITY(ddrval),
 	HRESULT_CODE(ddrval),
@@ -371,7 +372,7 @@ RGBFTYPE DirectDraw_GetSurfacePixelFormat(LPDDSURFACEDESC2 surface)
     surf_flags = surface->dwFlags;
     pfp = &surface->ddpfPixelFormat;
 
-    if((surf_flags & DDSD_PIXELFORMAT) == 0x0)
+    if ((surf_flags & DDSD_PIXELFORMAT) == 0x0)
 	return RGBFB_NONE;
 
     if ((pfp->dwFlags & DDPF_RGB) == 0)
@@ -422,14 +423,14 @@ RGBFTYPE DirectDraw_GetSurfacePixelFormat(LPDDSURFACEDESC2 surface)
     return RGBFB_NONE;
 }
 
-HRESULT DirectDraw_EnumDisplayModes(DWORD flags, LPDDENUMMODESCALLBACK2 callback)
+HRESULT DirectDraw_EnumDisplayModes (DWORD flags, LPDDENUMMODESCALLBACK2 callback)
 {
     HRESULT result;
-    result = IDirectDraw7_EnumDisplayModes(dxdata.maindd, flags, NULL, NULL, callback);
+    result = IDirectDraw7_EnumDisplayModes (dxdata.maindd, flags, NULL, NULL, callback);
     return result;
 }
 
-HRESULT DirectDraw_EnumDisplays(LPDDENUMCALLBACKEX callback)
+HRESULT DirectDraw_EnumDisplays (LPDDENUMCALLBACKEX callback)
 {
     HRESULT result;
     result = DirectDrawEnumerateEx (callback, 0, DDENUM_DETACHEDSECONDARYDEVICES | DDENUM_ATTACHEDSECONDARYDEVICES);
@@ -522,7 +523,7 @@ HRESULT DirectDraw_ReleaseDC(HDC hdc)
 int DirectDraw_GetVerticalBlankStatus (void)
 {
     BOOL status;
-    if (FAILED(IDirectDraw7_GetVerticalBlankStatus (dxdata.maindd, &status)))
+    if (FAILED (IDirectDraw7_GetVerticalBlankStatus (dxdata.maindd, &status)))
 	return -1;
     return status;
 }
@@ -538,6 +539,8 @@ DWORD DirectDraw_CurrentRefreshRate (void)
 
 HRESULT DirectDraw_FlipToGDISurface (void)
 {
+    if (!dxdata.ddinit)
+	return DD_OK;
     return IDirectDraw7_FlipToGDISurface (dxdata.maindd);
 }
 
@@ -552,7 +555,7 @@ int DirectDraw_BlitToPrimaryScale (RECT *rect)
     dst = dxdata.primary;
     SetRect (&dstrect, x, y, x + w, y + h);
     centerdstrect (&dstrect);
-    while (FAILED(ddrval = IDirectDrawSurface7_Blt (dst, &dstrect, dxdata.secondary, rect, DDBLT_WAIT, NULL))) {
+    while (FAILED (ddrval = IDirectDrawSurface7_Blt (dst, &dstrect, dxdata.secondary, rect, DDBLT_WAIT, NULL))) {
 	if (ddrval == DDERR_SURFACELOST) {
 	    ddrval = restoresurface (dst);
 	    if (FAILED (ddrval))
@@ -650,7 +653,7 @@ void DirectDraw_Fill (RECT *rect, uae_u32 color)
     ddbltfx.dwFillColor = color;
     ddbltfx.dwSize = sizeof (ddbltfx);
     dst = getlocksurface ();
-    while (FAILED(ddrval = IDirectDrawSurface7_Blt (dst, rect, NULL, NULL, DDBLT_WAIT | DDBLT_COLORFILL, &ddbltfx))) {
+    while (FAILED (ddrval = IDirectDrawSurface7_Blt (dst, rect, NULL, NULL, DDBLT_WAIT | DDBLT_COLORFILL, &ddbltfx))) {
 	if (ddrval == DDERR_SURFACELOST) {
 	    ddrval = restoresurface (dst);
 	    if (FAILED (ddrval))
diff --git a/od-win32/picasso96_win.c b/od-win32/picasso96_win.c
index 539e5229..6057853d 100755
--- a/od-win32/picasso96_win.c
+++ b/od-win32/picasso96_win.c
@@ -46,6 +46,8 @@
 
 #if defined(PICASSO96)
 
+#define NOBLITTER 0
+
 static int hwsprite;
 
 #include "registry.h"
@@ -79,6 +81,7 @@ static void flushpixels(void);
 #define P96TRACE(x)
 #define P96TRACE_SPR(x)
 #endif
+#define P96TRACE2(x) do { write_log x; } while(0)
 
 static void REGPARAM2 gfxmem_lputx (uaecptr, uae_u32) REGPARAM;
 static void REGPARAM2 gfxmem_wputx (uaecptr, uae_u32) REGPARAM;
@@ -868,16 +871,15 @@ void picasso_refresh (void)
 /*
 * Functions to perform an action on the frame-buffer
 */
-STATIC_INLINE void do_blitrect_frame_buffer (struct RenderInfo *ri, struct
+static int do_blitrect_frame_buffer (struct RenderInfo *ri, struct
     RenderInfo *dstri, unsigned long srcx, unsigned long srcy,
     unsigned long dstx, unsigned long dsty, unsigned long width, unsigned
     long height, uae_u8 mask, BLIT_OPCODE opcode)
 {
-    uae_u8 *src, *dst, *tmp, *tmp2, *tmp3;
+    uae_u8 *src, *dst;
     uae_u8 Bpp = GetBytesPerPixel (ri->RGBFormat);
     unsigned long total_width = width * Bpp;
     unsigned long linewidth = (total_width + 15) & ~15;
-    unsigned long lines;
 
     src = ri->Memory + srcx * Bpp + srcy * ri->BytesPerRow;
     dst = dstri->Memory + dstx * Bpp + dsty * dstri->BytesPerRow;
@@ -905,7 +907,7 @@ STATIC_INLINE void do_blitrect_frame_buffer (struct RenderInfo *ri, struct
 		for (i = 0; i < height; i++, src -= ri->BytesPerRow, dst -= dstri->BytesPerRow)
 		    memcpy (dst, src, total_width);
 	    }
-	    return;
+	    return 1;
 
 	} else if (Bpp == 4) {
 
@@ -974,28 +976,9 @@ STATIC_INLINE void do_blitrect_frame_buffer (struct RenderInfo *ri, struct
 	    }
 
 	}
-	return;
-    }
-
-    tmp3 = tmp2 = tmp = xmalloc (linewidth * height); /* allocate enough memory for the src-rect */
-    if (!tmp)
-	return;
-
-    /* copy the src-rect into our temporary buffer space */
-    for (lines = 0; lines < height; lines++, src += ri->BytesPerRow, tmp2 += linewidth) {
-	memcpy (tmp2, src, total_width);
-    }
-
-    /* copy the temporary buffer to the destination */
-    for (lines = 0; lines < height; lines++, dst += dstri->BytesPerRow, tmp += linewidth) {
-	unsigned long cols;
-	for (cols = 0; cols < width; cols++) {
-	    dst[cols] &= ~mask;
-	    dst[cols] |= tmp[cols] & mask;
-	}
+	return 1;
     }
-    /* free the temp-buf */
-    free (tmp3);
+    return 0;
 }
 
 /*
@@ -2126,6 +2109,8 @@ uae_u32 REGPARAM2 picasso_InvertRect (struct regstruct *regs)
     unsigned long width_in_bytes;
     uae_u32 result = 0;
 
+    if (NOBLITTER)
+	return 0;
     if (CopyRenderInfoStructureA2U (renderinfo, &ri)) {
 	P96TRACE(("InvertRect %dbpp 0x%lx\n", Bpp, (long)mask));
 
@@ -2174,6 +2159,8 @@ uae_u32 REGPARAM2 picasso_FillRect (struct regstruct *regs)
     struct RenderInfo ri;
     uae_u32 result = 0;
 
+    if (NOBLITTER)
+	return 0;
     if (CopyRenderInfoStructureA2U (renderinfo, &ri) && Y != 0xFFFF) {
 	if (ri.RGBFormat != RGBFormat)
 	    write_log ("Weird Stuff!\n");
@@ -2294,8 +2281,7 @@ STATIC_INLINE int BlitRectHelper (void)
 	dstri = ri;
     }
     /* Do our virtual frame-buffer memory first */
-    do_blitrect_frame_buffer (ri, dstri, srcx, srcy, dstx, dsty, width, height, mask, opcode);
-    return 1;
+    return do_blitrect_frame_buffer (ri, dstri, srcx, srcy, dstx, dsty, width, height, mask, opcode);
 }
 
 STATIC_INLINE int BlitRect (uaecptr ri, uaecptr dstri,
@@ -2349,6 +2335,8 @@ uae_u32 REGPARAM2 picasso_BlitRect (struct regstruct *regs)
     uae_u8  Mask = (uae_u8)m68k_dreg (regs, 6);
     uae_u32 result = 0;
 
+    if (NOBLITTER)
+	return 0;
     P96TRACE(("BlitRect(%d, %d, %d, %d, %d, %d, 0x%x)\n", srcx, srcy, dstx, dsty, width, height, Mask));
     result = BlitRect (renderinfo, (uaecptr)NULL, srcx, srcy, dstx, dsty, width, height, Mask, BLIT_SRC);
     return result;
@@ -2386,6 +2374,8 @@ uae_u32 REGPARAM2 picasso_BlitRectNoMaskComplete (struct regstruct *regs)
     uae_u32 RGBFmt = m68k_dreg (regs, 7);
     uae_u32 result = 0;
 
+    if (NOBLITTER)
+	return 0;
     {
 	uaecptr a5 = m68k_areg (regs, 5) & ~0xffff;
 	if (a5 != RTAREA_DEFAULT && a5 != RTAREA_BACKUP) {
@@ -2469,6 +2459,8 @@ uae_u32 REGPARAM2 picasso_BlitPattern (struct regstruct *regs)
     unsigned long ysize_mask;
     uae_u32 result = 0;
 
+    if (NOBLITTER)
+	return 0;
     if(CopyRenderInfoStructureA2U (rinf, &ri) && CopyPatternStructureA2U (pinf, &pattern)) {
 	Bpp = GetBytesPerPixel(ri.RGBFormat);
 	uae_mem = ri.Memory + Y * ri.BytesPerRow + X * Bpp; /* offset with address */
@@ -2626,6 +2618,8 @@ uae_u32 REGPARAM2 picasso_BlitTemplate (struct regstruct *regs)
     uae_u8 *tmpl_base;
     uae_u32 result = 0;
 
+    if (NOBLITTER)
+	return 0;
     if (CopyRenderInfoStructureA2U (rinf, &ri) && CopyTemplateStructureA2U (tmpl, &tmp)) {
 	Bpp = GetBytesPerPixel(ri.RGBFormat);
 	uae_mem = ri.Memory + Y * ri.BytesPerRow + X * Bpp; /* offset into address */
@@ -2931,6 +2925,8 @@ uae_u32 REGPARAM2 picasso_BlitPlanar2Chunky (struct regstruct *regs)
     struct BitMap local_bm;
     uae_u32 result = 0;
 
+    if (NOBLITTER)
+	return 0;
     if (minterm != 0x0C) {
 	write_log ("ERROR - BlitPlanar2Chunky() has minterm 0x%x, which I don't handle. Using fall-back routine.\n",
 	    minterm);
@@ -3070,6 +3066,8 @@ uae_u32 REGPARAM2 picasso_BlitPlanar2Direct (struct regstruct *regs)
     struct ColorIndexMapping local_cim;
     uae_u32 result = 0;
 
+    if (NOBLITTER)
+	return 0;
     if (minterm != 0x0C) {
 	write_log ("WARNING - BlitPlanar2Direct() has unhandled op-code 0x%x. Using fall-back routine.\n",
 	    minterm);
diff --git a/od-win32/win32.h b/od-win32/win32.h
index c14cffac..2465f496 100755
--- a/od-win32/win32.h
+++ b/od-win32/win32.h
@@ -15,9 +15,9 @@
 #define GETBDM(x) (((x) - ((x / 10000) * 10000)) / 100)
 #define GETBDD(x) ((x) % 100)
 
-#define WINUAEBETA 8
+#define WINUAEBETA 9
 #define WINUAEPUBLICBETA 1
-#define WINUAEDATE MAKEBD(2008, 3, 14)
+#define WINUAEDATE MAKEBD(2008, 3, 18)
 #define WINUAEEXTRA ""
 #define WINUAEREV ""
 
diff --git a/od-win32/win32_scale2x.c b/od-win32/win32_scale2x.c
index 9cbc5c31..df9f1823 100755
--- a/od-win32/win32_scale2x.c
+++ b/od-win32/win32_scale2x.c
@@ -38,6 +38,61 @@ struct uae_filter uaefilters[] =
 };
 
 
+uae_u8 *getfilterrect1 (RECT *sr, RECT *dr, int dst_depth, int aw, int ah, int scale, int temp_width, int temp_height, uae_u8 *dptr, int pitch)
+{
+    int aws = aw * scale;
+    int ahs = ah * scale;
+    
+    SetRect (sr, 0, 0, 0, 0);
+    dr->left = sr->left + (temp_width - aws) /2;
+    dr->top = sr->top + (temp_height - ahs) / 2;
+    dptr += dr->left * (dst_depth / 8);
+    dptr += dr->top * pitch;
+    return dptr;
+}
+
+void getfilterrect2 (RECT *sr, RECT *dr, int dst_width, int dst_height, int aw, int ah, int scale, int temp_width, int temp_height)
+{
+    int aws = aw * scale;
+    int ahs = ah * scale;
+    int xs, ys;
+    int xmult, ymult;
+    int v;
+
+    SetRect (sr, 0, 0, dst_width, dst_height);
+    dr->left = (temp_width - aws) /2;
+    dr->top =  (temp_height - ahs) / 2;
+    dr->left -= (dst_width - aws) / 2;
+    dr->top -= (dst_height - ahs) / 2;
+    dr->right = dr->left + dst_width;
+    dr->bottom = dr->top + dst_height;
+
+    v = currprefs.gfx_filter ? currprefs.gfx_filter_horiz_offset : 0;
+    OffsetRect (dr, (int)(-v * aws / 1000.0), 0);
+    v = currprefs.gfx_filter ? currprefs.gfx_filter_vert_offset : 0;
+    OffsetRect (dr, 0, (int)(-v * ahs / 1000.0));
+
+    xmult = currprefs.gfx_filter_horiz_zoom_mult;
+    if (xmult <= 0)
+        xmult = aws * 1000 / dst_width;
+    else
+        xmult = xmult + xmult * currprefs.gfx_filter_horiz_zoom / 2000;
+
+    ymult = currprefs.gfx_filter_vert_zoom_mult;
+    if (ymult <= 0)
+        ymult = ahs * 1000 / dst_height;
+    else
+        ymult = ymult + ymult * currprefs.gfx_filter_vert_zoom / 2000;
+    
+    xs = dst_width - dst_width * xmult / 1000;
+    dr->left += xs / 2;
+    dr->right -= xs / 2;
+
+    ys = dst_height - dst_height * ymult / 1000;
+    dr->top += ys / 2;
+    dr->bottom -= ys / 2;
+}
+
 static int dst_width, dst_height, amiga_width, amiga_height, amiga_depth, dst_depth, scale;
 static int temp_width, temp_height;
 uae_u8 *bufmem_ptr;
@@ -130,7 +185,7 @@ void S2X_init (int dw, int dh, int aw, int ah, int mult, int ad, int dd)
 	temp_height = dxdata.maxheight;
     tempsurf = allocsurface (temp_width, temp_height);
     if (!tempsurf)
-	write_log ("DDRAW: failed to create temp surface\n");
+	write_log ("DDRAW: failed to create temp surface (%dx%d)\n", temp_width, temp_height);
 
     if (usedfilter->type == UAE_FILTER_HQ) {
 	int w = amiga_width > dst_width ? amiga_width : dst_width;
@@ -142,11 +197,12 @@ void S2X_init (int dw, int dh, int aw, int ah, int mult, int ad, int dd)
 
 void S2X_render (void)
 {
-    int aw, ah, aws, ahs, pitch;
+    int aw, ah, aws, ahs;
     uae_u8 *dptr, *enddptr, *sptr, *endsptr;
     int ok = 0;
     RECT sr, dr;
     DDSURFACEDESC2 desc;
+    DWORD pitch;
 
     aw = amiga_width;
     ah = amiga_height;
@@ -160,22 +216,16 @@ void S2X_render (void)
     if (tempsurf == NULL)
 	return;
 
-    sr.left = -(aw - amiga_width) / 2;
-    sr.top = -(ah - amiga_height) / 2;
-    sptr = gfxvidinfo.bufmem + sr.left * gfxvidinfo.pixbytes + sr.top * gfxvidinfo.rowbytes;
+    sptr = gfxvidinfo.bufmem;
     endsptr = gfxvidinfo.bufmemend;
     bufmem_ptr = sptr;
 
     if (!locksurface (tempsurf, &desc))
 	return;
-    dptr = (uae_u8*)desc.lpSurface;
     pitch = desc.lPitch;
+    dptr = (uae_u8*)desc.lpSurface;
     enddptr = dptr + pitch * temp_height;
-
-    dr.left = sr.left + (temp_width - aws) /2;
-    dr.top = sr.top + (temp_height - ahs) / 2;
-    dptr += dr.left * (dst_depth / 8);
-    dptr += dr.top * pitch;
+    dptr = getfilterrect1 (&sr, &dr, dst_depth, aw, ah, scale, temp_width, temp_height, dptr, pitch);
 
     if (!dptr) /* weird things can happen */
 	goto end;
@@ -281,50 +331,12 @@ void S2X_render (void)
 end:
     unlocksurface (tempsurf);
 
-    {
-	int xs, ys;
-	int xmult, ymult;
-	int v;
-
-	SetRect (&sr, 0, 0, dst_width, dst_height);
-
-	dr.left -= (dst_width - aws) / 2;
-	dr.top -= (dst_height - ahs) / 2;
-	dr.right = dr.left + dst_width;
-	dr.bottom = dr.top + dst_height;
-
-	v = currprefs.gfx_filter ? currprefs.gfx_filter_horiz_offset : 0;
-	OffsetRect (&dr, (int)(-v * aws / 1000.0), 0);
-	v = currprefs.gfx_filter ? currprefs.gfx_filter_vert_offset : 0;
-	OffsetRect (&dr, 0, (int)(-v * ahs / 1000.0));
-
-	xmult = currprefs.gfx_filter_horiz_zoom_mult;
-	if (xmult <= 0)
-	    xmult = aws * 1000 / dst_width;
-	else
-	    xmult = xmult + xmult * currprefs.gfx_filter_horiz_zoom / 2000;
-
-	ymult = currprefs.gfx_filter_vert_zoom_mult;
-	if (ymult <= 0)
-	    ymult = ahs * 1000 / dst_height;
-	else
-	    ymult = ymult + ymult * currprefs.gfx_filter_vert_zoom / 2000;
-
-	xs = dst_width - dst_width * xmult / 1000;
-
-	dr.left += xs / 2;
-	dr.right -= xs / 2;
-
-	ys = dst_height - dst_height * ymult / 1000;
-	dr.top += ys / 2;
-	dr.bottom -= ys / 2;
-
-	if (dr.left >= 0 && dr.top >= 0 && dr.right < temp_width && dr.bottom < temp_height) {
-	    if (dr.left < dr.right && dr.top < dr.bottom)
-		DirectDraw_BlitRect (NULL, &sr, tempsurf, &dr);
-	}
-	statusline ();
+    getfilterrect2 (&sr, &dr, dst_width, dst_height, amiga_width, amiga_height, scale, temp_width, temp_height);
+    if (dr.left >= 0 && dr.top >= 0 && dr.right < temp_width && dr.bottom < temp_height) {
+        if (dr.left < dr.right && dr.top < dr.bottom)
+    	    DirectDraw_BlitRect (NULL, &sr, tempsurf, &dr);
     }
+    statusline ();
 }
 
 void S2X_refresh (void)
diff --git a/od-win32/win32gfx.c b/od-win32/win32gfx.c
index 58e7839a..4dd4949b 100755
--- a/od-win32/win32gfx.c
+++ b/od-win32/win32gfx.c
@@ -167,7 +167,7 @@ uae_u32 default_freq = 0;
 HWND hStatusWnd = NULL;
 uae_u16 picasso96_pixel_format = RGBFF_CHUNKY;
 
-static char scrlinebuf[4096 * 4]; /* this is too large, but let's rather play on the safe side here */
+static uae_u8 scrlinebuf[4096 * 4]; /* this is too large, but let's rather play on the safe side here */
 
 
 void centerdstrect (RECT *dr)
@@ -1871,6 +1871,11 @@ static BOOL doInit (void)
 #endif
 #endif
     screen_is_initialized = 1;
+    WIN32GFX_SetPalette ();
+#ifdef PICASSO96
+    DX_SetPalette (0, 256);
+#endif
+    picasso_refresh ();
     return 1;
 
 oops:
diff --git a/od-win32/win32gui.c b/od-win32/win32gui.c
index c640ce6f..46df619b 100755
--- a/od-win32/win32gui.c
+++ b/od-win32/win32gui.c
@@ -1146,7 +1146,7 @@ void gui_display (int shortcut)
     scaleresource_setmaxsize (-1, -1);
     if (w > 0 && h > 0)
 	scaleresource_setmaxsize (w, h);
-    WIN32GFX_ClearPalette();
+    WIN32GFX_ClearPalette ();
     manual_painting_needed++; /* So that WM_PAINT will refresh the display */
 
     if (isfullscreen () > 0) {
@@ -1172,7 +1172,7 @@ void gui_display (int shortcut)
     }
     manual_painting_needed--; /* So that WM_PAINT doesn't need to use custom refreshing */
     manual_palette_refresh_needed = 1;
-    resumepaused();
+    resumepaused ();
     inputdevice_copyconfig (&changed_prefs, &currprefs);
     inputdevice_config_change_test ();
     clearallkeys ();
@@ -1185,12 +1185,12 @@ void gui_display (int shortcut)
     AVIOutput_Begin ();
 #endif
     fpscounter_reset ();
-    WIN32GFX_SetPalette();
+    WIN32GFX_SetPalette ();
 #ifdef PICASSO96
     DX_SetPalette (0, 256);
 #endif
-    screenshot_free();
-    write_disk_history();
+    screenshot_free ();
+    write_disk_history ();
     gui_active--;
     here--;
 }
@@ -1325,103 +1325,103 @@ int DiskSelection_2 (HWND hDlg, WPARAM wParam, int flag, struct uae_prefs *prefs
 
     switch (flag) {
     case 0:
-	WIN32GUI_LoadUIString(IDS_SELECTADF, szTitle, MAX_DPATH);
-	WIN32GUI_LoadUIString(IDS_ADF, szFormat, MAX_DPATH);
-	sprintf(szFilter, "%s ", szFormat);
-	memcpy(szFilter + strlen(szFilter), DISK_FORMAT_STRING, sizeof(DISK_FORMAT_STRING) + 1);
+	WIN32GUI_LoadUIString (IDS_SELECTADF, szTitle, MAX_DPATH);
+	WIN32GUI_LoadUIString (IDS_ADF, szFormat, MAX_DPATH);
+	sprintf (szFilter, "%s ", szFormat);
+	memcpy (szFilter + strlen (szFilter), DISK_FORMAT_STRING, sizeof (DISK_FORMAT_STRING) + 1);
 
 	openFileName.lpstrDefExt = "ADF";
 	openFileName.lpstrFilter = szFilter;
 	break;
     case 1:
-	WIN32GUI_LoadUIString(IDS_CHOOSEBLANK, szTitle, MAX_DPATH);
-	WIN32GUI_LoadUIString(IDS_ADF, szFormat, MAX_DPATH);
-	sprintf(szFilter, "%s ", szFormat);
-	memcpy(szFilter + strlen(szFilter), "(*.adf)\0*.adf\0", 15);
+	WIN32GUI_LoadUIString (IDS_CHOOSEBLANK, szTitle, MAX_DPATH);
+	WIN32GUI_LoadUIString (IDS_ADF, szFormat, MAX_DPATH);
+	sprintf (szFilter, "%s ", szFormat);
+	memcpy (szFilter + strlen (szFilter), "(*.adf)\0*.adf\0", 15);
 
 	openFileName.lpstrDefExt = "ADF";
 	openFileName.lpstrFilter = szFilter;
 	break;
     case 2:
     case 3:
-	WIN32GUI_LoadUIString(IDS_SELECTHDF, szTitle, MAX_DPATH);
-	WIN32GUI_LoadUIString(IDS_HDF, szFormat, MAX_DPATH);
-	sprintf(szFilter, "%s ", szFormat);
-	memcpy(szFilter + strlen(szFilter),  HDF_FORMAT_STRING, sizeof (HDF_FORMAT_STRING) + 1);
+	WIN32GUI_LoadUIString (IDS_SELECTHDF, szTitle, MAX_DPATH);
+	WIN32GUI_LoadUIString (IDS_HDF, szFormat, MAX_DPATH);
+	sprintf (szFilter, "%s ", szFormat);
+	memcpy (szFilter + strlen (szFilter),  HDF_FORMAT_STRING, sizeof (HDF_FORMAT_STRING) + 1);
 
 	openFileName.lpstrDefExt = "HDF";
 	openFileName.lpstrFilter = szFilter;
 	break;
     case 4:
     case 5:
-	WIN32GUI_LoadUIString(IDS_SELECTUAE, szTitle, MAX_DPATH);
-	WIN32GUI_LoadUIString(IDS_UAE, szFormat, MAX_DPATH );
-	sprintf(szFilter, "%s ", szFormat);
-	memcpy(szFilter + strlen(szFilter), "(*.uae)\0*.uae\0", 15);
+	WIN32GUI_LoadUIString (IDS_SELECTUAE, szTitle, MAX_DPATH);
+	WIN32GUI_LoadUIString (IDS_UAE, szFormat, MAX_DPATH );
+	sprintf (szFilter, "%s ", szFormat);
+	memcpy (szFilter + strlen (szFilter), "(*.uae)\0*.uae\0", 15);
 
 	openFileName.lpstrDefExt = "UAE";
 	openFileName.lpstrFilter = szFilter;
 	break;
     case 6:
-	WIN32GUI_LoadUIString(IDS_SELECTROM, szTitle, MAX_DPATH);
-	WIN32GUI_LoadUIString(IDS_ROM, szFormat, MAX_DPATH);
-	sprintf(szFilter, "%s ", szFormat);
-	memcpy(szFilter + strlen(szFilter), ROM_FORMAT_STRING, sizeof (ROM_FORMAT_STRING) + 1);
+	WIN32GUI_LoadUIString (IDS_SELECTROM, szTitle, MAX_DPATH);
+	WIN32GUI_LoadUIString (IDS_ROM, szFormat, MAX_DPATH);
+	sprintf (szFilter, "%s ", szFormat);
+	memcpy (szFilter + strlen(szFilter), ROM_FORMAT_STRING, sizeof (ROM_FORMAT_STRING) + 1);
 
 	openFileName.lpstrDefExt = "ROM";
 	openFileName.lpstrFilter = szFilter;
 	break;
     case 7:
-	WIN32GUI_LoadUIString(IDS_SELECTKEY, szTitle, MAX_DPATH);
-	WIN32GUI_LoadUIString(IDS_KEY, szFormat, MAX_DPATH);
-	sprintf(szFilter, "%s ", szFormat);
-	memcpy(szFilter + strlen(szFilter), "(*.key)\0*.key\0", 15);
+	WIN32GUI_LoadUIString (IDS_SELECTKEY, szTitle, MAX_DPATH);
+	WIN32GUI_LoadUIString (IDS_KEY, szFormat, MAX_DPATH);
+	sprintf (szFilter, "%s ", szFormat);
+	memcpy (szFilter + strlen(szFilter), "(*.key)\0*.key\0", 15);
 
 	openFileName.lpstrDefExt = "KEY";
 	openFileName.lpstrFilter = szFilter;
 	break;
     case 15:
     case 16:
-	WIN32GUI_LoadUIString(flag == 15 ? IDS_RESTOREINP : IDS_SAVEINP, szTitle, MAX_DPATH);
-	WIN32GUI_LoadUIString(IDS_INP, szFormat, MAX_DPATH);
-	sprintf(szFilter, "%s ", szFormat);
-	memcpy(szFilter + strlen(szFilter), INP_FORMAT_STRING, sizeof (INP_FORMAT_STRING) + 1);
+	WIN32GUI_LoadUIString (flag == 15 ? IDS_RESTOREINP : IDS_SAVEINP, szTitle, MAX_DPATH);
+	WIN32GUI_LoadUIString (IDS_INP, szFormat, MAX_DPATH);
+	sprintf (szFilter, "%s ", szFormat);
+	memcpy (szFilter + strlen(szFilter), INP_FORMAT_STRING, sizeof (INP_FORMAT_STRING) + 1);
 
 	openFileName.lpstrDefExt = "INP";
 	openFileName.lpstrFilter = szFilter;
 	break;
     case 9:
     case 10:
-	WIN32GUI_LoadUIString(flag == 10 ? IDS_RESTOREUSS : IDS_SAVEUSS, szTitle, MAX_DPATH);
-	WIN32GUI_LoadUIString(IDS_USS, szFormat, MAX_DPATH);
-	sprintf(szFilter, "%s ", szFormat);
+	WIN32GUI_LoadUIString (flag == 10 ? IDS_RESTOREUSS : IDS_SAVEUSS, szTitle, MAX_DPATH);
+	WIN32GUI_LoadUIString (IDS_USS, szFormat, MAX_DPATH);
+	sprintf (szFilter, "%s ", szFormat);
 	if (flag == 10) {
-	    memcpy(szFilter + strlen(szFilter), USS_FORMAT_STRING_RESTORE, sizeof (USS_FORMAT_STRING_RESTORE) + 1);
+	    memcpy (szFilter + strlen(szFilter), USS_FORMAT_STRING_RESTORE, sizeof (USS_FORMAT_STRING_RESTORE) + 1);
 	    all = 1;
 	} else {
 	    char tmp[MAX_DPATH];
-	    memcpy(szFilter + strlen(szFilter), USS_FORMAT_STRING_SAVE, sizeof (USS_FORMAT_STRING_SAVE) + 1);
+	    memcpy (szFilter + strlen(szFilter), USS_FORMAT_STRING_SAVE, sizeof (USS_FORMAT_STRING_SAVE) + 1);
 	    p = szFilter;
 	    while (p[0] != 0 || p[1] !=0 ) p++;
 	    p++;
-	    WIN32GUI_LoadUIString(IDS_STATEFILE_UNCOMPRESSED, tmp, sizeof(tmp));
-	    strcat(p, tmp);
-	    strcat(p, " (*.uss)");
-	    p += strlen(p) + 1;
-	    strcpy(p, "*.uss");
-	    p += strlen(p) + 1;
-	    WIN32GUI_LoadUIString(IDS_STATEFILE_RAMDUMP, tmp, sizeof(tmp));
-	    strcat(p, tmp);
-	    strcat(p, " (*.dat)");
+	    WIN32GUI_LoadUIString (IDS_STATEFILE_UNCOMPRESSED, tmp, sizeof (tmp));
+	    strcat (p, tmp);
+	    strcat (p, " (*.uss)");
+	    p += strlen (p) + 1;
+	    strcpy (p, "*.uss");
+	    p += strlen (p) + 1;
+	    WIN32GUI_LoadUIString (IDS_STATEFILE_RAMDUMP, tmp, sizeof (tmp));
+	    strcat (p, tmp);
+	    strcat (p, " (*.dat)");
 	    p += strlen(p) + 1;
 	    strcpy (p, "*.dat");
-	    p += strlen(p) + 1;
-	    WIN32GUI_LoadUIString(IDS_STATEFILE_WAVE, tmp, sizeof(tmp));
-	    strcat(p, tmp);
-	    strcat(p, " (*.wav)");
-	    p += strlen(p) + 1;
+	    p += strlen (p) + 1;
+	    WIN32GUI_LoadUIString (IDS_STATEFILE_WAVE, tmp, sizeof (tmp));
+	    strcat (p, tmp);
+	    strcat (p, " (*.wav)");
+	    p += strlen (p) + 1;
 	    strcpy (p, "*.wav");
-	    p += strlen(p) + 1;
+	    p += strlen (p) + 1;
 	    *p = 0;
 	    all = 0;
 	    filterindex = statefile_previousfilter;
@@ -1430,30 +1430,30 @@ int DiskSelection_2 (HWND hDlg, WPARAM wParam, int flag, struct uae_prefs *prefs
 	openFileName.lpstrFilter = szFilter;
 	break;
     case 11:
-	WIN32GUI_LoadUIString(IDS_SELECTFLASH, szTitle, MAX_DPATH);
-	WIN32GUI_LoadUIString(IDS_FLASH, szFormat, MAX_DPATH);
-	sprintf(szFilter, "%s ", szFormat);
-	memcpy(szFilter + strlen(szFilter), "(*.nvr)\0*.nvr\0", 15);
+	WIN32GUI_LoadUIString (IDS_SELECTFLASH, szTitle, MAX_DPATH);
+	WIN32GUI_LoadUIString (IDS_FLASH, szFormat, MAX_DPATH);
+	sprintf (szFilter, "%s ", szFormat);
+	memcpy (szFilter + strlen (szFilter), "(*.nvr)\0*.nvr\0", 15);
 
 	openFileName.lpstrDefExt = "NVR";
 	openFileName.lpstrFilter = szFilter;
 	break;
     case 8:
     default:
-	WIN32GUI_LoadUIString(IDS_SELECTINFO, szTitle, MAX_DPATH);
+	WIN32GUI_LoadUIString (IDS_SELECTINFO, szTitle, MAX_DPATH);
 
 	openFileName.lpstrFilter = NULL;
 	openFileName.lpstrDefExt = NULL;
 	break;
     case 12:
-	WIN32GUI_LoadUIString(IDS_SELECTFS, szTitle, MAX_DPATH);
+	WIN32GUI_LoadUIString (IDS_SELECTFS, szTitle, MAX_DPATH);
 
 	openFileName.lpstrFilter = NULL;
 	openFileName.lpstrDefExt = NULL;
 	openFileName.lpstrInitialDir = path_out;
 	break;
     case 13:
-	WIN32GUI_LoadUIString(IDS_SELECTINFO, szTitle, MAX_DPATH);
+	WIN32GUI_LoadUIString (IDS_SELECTINFO, szTitle, MAX_DPATH);
 
 	openFileName.lpstrFilter = NULL;
 	openFileName.lpstrDefExt = NULL;
@@ -1462,7 +1462,7 @@ int DiskSelection_2 (HWND hDlg, WPARAM wParam, int flag, struct uae_prefs *prefs
     case 14:
 	strcpy (szTitle, "Select supported archive file");
 	sprintf (szFilter, "%s (%s)", "Archive", ARCHIVE_STRING);
-	strcpy (szFilter + strlen(szFilter) + 1, ARCHIVE_STRING);
+	strcpy (szFilter + strlen (szFilter) + 1, ARCHIVE_STRING);
 
 	openFileName.lpstrFilter = NULL;
 	openFileName.lpstrFilter = szFilter;
@@ -8929,8 +8929,8 @@ static void enable_for_hw3ddlg (HWND hDlg)
     CheckDlgButton(hDlg, IDC_FILTERENABLE, v);
     ew (hDlg, IDC_FILTERHZ, v);
     ew (hDlg, IDC_FILTERVZ, v);
-    ew (hDlg, IDC_FILTERHZMULT, vv && !vv4);
-    ew (hDlg, IDC_FILTERVZMULT, vv && !vv4);
+    ew (hDlg, IDC_FILTERHZMULT, v);
+    ew (hDlg, IDC_FILTERVZMULT, v);
     ew (hDlg, IDC_FILTERHO, v);
     ew (hDlg, IDC_FILTERVO, v);
     ew (hDlg, IDC_FILTERSLR, vv3);
@@ -9600,11 +9600,11 @@ static INT_PTR CALLBACK AVIOutputDlgProc(HWND hDlg, UINT msg, WPARAM wParam, LPA
 
 		case IDC_AVIOUTPUT_FRAMELIMITER:
 		    avioutput_framelimiter = IsDlgButtonChecked (hDlg, IDC_AVIOUTPUT_FRAMELIMITER) ? 0 : 1;
-		    AVIOutput_SetSettings();
+		    AVIOutput_SetSettings ();
 		break;
 		case IDC_AVIOUTPUT_NOSOUNDOUTPUT:
 		    avioutput_nosoundoutput = IsDlgButtonChecked (hDlg, IDC_AVIOUTPUT_NOSOUNDOUTPUT) ? 1 : 0;
-		    AVIOutput_SetSettings();
+		    AVIOutput_SetSettings ();
 		break;
 
 		case IDC_INPREC_PLAYMODE:
@@ -9613,13 +9613,13 @@ static INT_PTR CALLBACK AVIOutputDlgProc(HWND hDlg, UINT msg, WPARAM wParam, LPA
 		    if (input_recording)
 			inprec_close ();
 		    else
-			DiskSelection(hDlg, wParam, 16, &workprefs, NULL);
+			DiskSelection (hDlg, wParam, 16, &workprefs, NULL);
 		break;
 		case IDC_INPREC_PLAY:
 		    if (input_recording)
 			inprec_close ();
 		    else
-			DiskSelection(hDlg, wParam, 15, &workprefs, NULL);
+			DiskSelection (hDlg, wParam, 15, &workprefs, NULL);
 		break;
 
 #ifdef PROWIZARD
@@ -9644,22 +9644,22 @@ static INT_PTR CALLBACK AVIOutputDlgProc(HWND hDlg, UINT msg, WPARAM wParam, LPA
 		break;
 
 		case IDC_AVIOUTPUT_PAL:
-		    SendDlgItemMessage(hDlg, IDC_AVIOUTPUT_FPS, TBM_SETPOS, TRUE, VBLANK_HZ_PAL);
-		    SendMessage(hDlg, WM_HSCROLL, (WPARAM) NULL, (LPARAM) NULL);
+		    SendDlgItemMessage (hDlg, IDC_AVIOUTPUT_FPS, TBM_SETPOS, TRUE, VBLANK_HZ_PAL);
+		    SendMessage (hDlg, WM_HSCROLL, (WPARAM) NULL, (LPARAM) NULL);
 		break;
 
 		case IDC_AVIOUTPUT_NTSC:
-		    SendDlgItemMessage(hDlg, IDC_AVIOUTPUT_FPS, TBM_SETPOS, TRUE, VBLANK_HZ_NTSC);
-		    SendMessage(hDlg, WM_HSCROLL, (WPARAM) NULL, (LPARAM) NULL);
+		    SendDlgItemMessage (hDlg, IDC_AVIOUTPUT_FPS, TBM_SETPOS, TRUE, VBLANK_HZ_NTSC);
+		    SendMessage (hDlg, WM_HSCROLL, (WPARAM) NULL, (LPARAM) NULL);
 		break;
 
 		case IDC_AVIOUTPUT_AUDIO:
 		{
 		    if (avioutput_enabled)
 			AVIOutput_End ();
-		    if(IsDlgButtonChecked(hDlg, IDC_AVIOUTPUT_AUDIO) == BST_CHECKED) {
-			avioutput_audio = AVIOutput_ChooseAudioCodec(hDlg, tmp, sizeof tmp);
-			enable_for_avioutputdlg(hDlg);
+		    if(IsDlgButtonChecked (hDlg, IDC_AVIOUTPUT_AUDIO) == BST_CHECKED) {
+			avioutput_audio = AVIOutput_ChooseAudioCodec (hDlg, tmp, sizeof tmp);
+			enable_for_avioutputdlg (hDlg);
 		    } else {
 			avioutput_audio = 0;
 		    }
@@ -9670,15 +9670,15 @@ static INT_PTR CALLBACK AVIOutputDlgProc(HWND hDlg, UINT msg, WPARAM wParam, LPA
 		{
 		    if (avioutput_enabled)
 			AVIOutput_End ();
-		    if(IsDlgButtonChecked(hDlg, IDC_AVIOUTPUT_VIDEO) == BST_CHECKED) {
-			avioutput_video = AVIOutput_ChooseVideoCodec(hDlg, tmp, sizeof tmp);
+		    if(IsDlgButtonChecked (hDlg, IDC_AVIOUTPUT_VIDEO) == BST_CHECKED) {
+			avioutput_video = AVIOutput_ChooseVideoCodec (hDlg, tmp, sizeof tmp);
 			if (avioutput_audio = AVIAUDIO_WAV)
 			    avioutput_audio = 0;
-			enable_for_avioutputdlg(hDlg);
+			enable_for_avioutputdlg (hDlg);
 		    } else {
 			avioutput_video = 0;
 		    }
-		    enable_for_avioutputdlg(hDlg);
+		    enable_for_avioutputdlg (hDlg);
 		    break;
 		}
 
@@ -9686,8 +9686,8 @@ static INT_PTR CALLBACK AVIOutputDlgProc(HWND hDlg, UINT msg, WPARAM wParam, LPA
 		{
 		    OPENFILENAME ofn;
 
-		    ZeroMemory(&ofn, sizeof(OPENFILENAME));
-		    ofn.lStructSize = sizeof(OPENFILENAME);
+		    ZeroMemory(&ofn, sizeof (OPENFILENAME));
+		    ofn.lStructSize = sizeof (OPENFILENAME);
 		    ofn.hwndOwner = hDlg;
 		    ofn.hInstance = hInst;
 		    ofn.Flags = OFN_EXTENSIONDIFFERENT | OFN_EXPLORER | OFN_PATHMUSTEXIST | OFN_HIDEREADONLY | OFN_OVERWRITEPROMPT | OFN_NOCHANGEDIR;
@@ -9704,7 +9704,7 @@ static INT_PTR CALLBACK AVIOutputDlgProc(HWND hDlg, UINT msg, WPARAM wParam, LPA
 		    ofn.lCustData = 0;
 		    ofn.lpstrFilter = "Video Clip (*.avi)\0*.avi\0Wave Sound (*.wav)\0";
 
-		    if(!GetSaveFileName(&ofn))
+		    if(!GetSaveFileName (&ofn))
 			break;
 		    if (ofn.nFilterIndex == 2) {
 			avioutput_audio = AVIAUDIO_WAV;
@@ -10712,8 +10712,8 @@ static int GetSettings (int all_options, HWND hwnd)
     if (first)
 	write_log ("Entering GUI idle loop\n");
 
-    scaleresource_setmaxsize(800, 600);
-    tres = scaleresource(panelresource, hwnd);
+    scaleresource_setmaxsize (800, 600);
+    tres = scaleresource (panelresource, hwnd);
     dhwnd = CreateDialogIndirect (tres->inst, tres->resource, hwnd, DialogProc);
     dialog_rect.top = dialog_rect.left = 0;
     dialog_rect.right = tres->width;
@@ -10726,18 +10726,18 @@ static int GetSettings (int all_options, HWND hwnd)
 
 	setguititle (dhwnd);
 	ShowWindow (dhwnd, SW_SHOW);
-	MapDialogRect(dhwnd, &dialog_rect);
+	MapDialogRect (dhwnd, &dialog_rect);
 
 	for (;;) {
 	    HANDLE IPChandle;
-	    IPChandle = geteventhandleIPC();
+	    IPChandle = geteventhandleIPC ();
 	    if (IPChandle != INVALID_HANDLE_VALUE) {
 		MsgWaitForMultipleObjects (1, &IPChandle, FALSE, INFINITE, QS_ALLINPUT);
-		while (checkIPC(&workprefs));
+		while (checkIPC( &workprefs));
 	    } else {
 		WaitMessage();
 	    }
-	    dialogmousemove(dhwnd);
+	    dialogmousemove (dhwnd);
 	    while ((v = PeekMessage (&msg, NULL, 0, 0, PM_REMOVE))) {
 		if (dialogreturn >= 0)
 		    break;
@@ -10957,7 +10957,7 @@ static int fsdialog (HWND *hwnd, DWORD *flags)
     *hwnd = hAmigaWnd;
     if (isfullscreen () <= 0)
 	return 0;
-    hr = DirectDraw_FlipToGDISurface();
+    hr = DirectDraw_FlipToGDISurface ();
     if (FAILED(hr))
 	write_log ("FlipToGDISurface failed, %s\n", DXError (hr));
     *flags |= MB_SETFOREGROUND;
diff --git a/od-win32/winuaechangelog.txt b/od-win32/winuaechangelog.txt
index 9cc8216b..97886474 100755
--- a/od-win32/winuaechangelog.txt
+++ b/od-win32/winuaechangelog.txt
@@ -1,4 +1,31 @@
 
+Beta 9:
+
+- crash when fullscreen enabled and error dialog was shown before
+  display was changed to fullscreen
+- CD32 state restore compatibility improved
+- SSCAN2 'undocumented' feature emulated. "If SSCAN2 bit in FMODE is
+  set, then disable SH10 horizontal coincidence detect". This really
+  means both one AND zero SH10 match = single sprite can appear twice
+  on screen! (Fantastic Dizzy CD32 background)
+- AGA mode only: last line of sprite may have been redrawn if bitplane
+  width was changed and sprite was reused later (introduced in 1.4.5)
+- debugger sp command improvements
+- AGA sprite outside bitplanes position bug fix (introduced in 1.4.5)
+- Picasso96 8-bit mode wrong palette after mode switch
+- Picasso96 8-bit and 16-bit optimized blitter bug fix, too much data
+  was blitted if width was >32 bytes (not pixels) and it was not
+  divisible by 32. 32-bit mode was working properly.
+- emulated LPSTOP instruction (68060 only)
+- MMU instructions didn't check supervisor mode. Note that there is no
+  MMU emulation, I think I have to repeat this every time I talk about
+  MMU :)
+- added 68040/68060 MMU instructions to table68k (no more generic
+  "MMUOP" in debugger's disassembler) 68030 MMU instructions are still
+  shown as "MMUOP030", they don't really fit to table68k format because
+  first instruction word does not specify the instruction type
+- MOVEC weird disassembly updated (MOVEM is still weird)
+
 Beta 8:
 
 NOTE: OGL/D3D filters still using old scaling
diff --git a/p96_blit.c b/p96_blit.c
index 45529b28..62f790d1 100755
--- a/p96_blit.c
+++ b/p96_blit.c
@@ -50,7 +50,7 @@ static void NOINLINE BLT_NAME (unsigned int w, unsigned int h, uae_u8 *src, uae_
     }
 
     ww = w / (8 * BLT_MULT);
-    xxd = w - ww * (8 * BLT_MULT);
+    xxd = (w - ww * (8 * BLT_MULT)) / BLT_MULT;
     for(y = 0; y < h; y++) {
 	uae_u32 *src_32 = (uae_u32*)src2;
 	uae_u32 *dst_32 = (uae_u32*)dst2;
diff --git a/readcpu.c b/readcpu.c
index 2ae68eab..07ec9cb7 100755
--- a/readcpu.c
+++ b/readcpu.c
@@ -137,9 +137,18 @@ struct mnemolookup lookuptab[] = {
     { i_CPUSHA, "CPUSHA" },
     { i_MOVE16, "MOVE16" },
 
-    { i_MMUOP30A, "MMUOP30A" },
-    { i_MMUOP30B, "MMUOP30B" },
-    { i_MMUOP, "MMUOP" },
+    { i_MMUOP030, "MMUOP030" },
+    { i_PFLUSHN, "PFLUSHN" },
+    { i_PFLUSH, "PFLUSH" },
+    { i_PFLUSHAN, "PFLUSHAN" },
+    { i_PFLUSHA, "PFLUSHA" },
+
+    { i_PLPAR, "PLPAR" },
+    { i_PLPAW, "PLPAW" },
+    { i_PTESTR, "PTESTR" },
+    { i_PTESTW, "PTESTW" },
+
+    { i_LPSTOP, "LPSTOP" },
     { i_ILLG, "" },
 };
 
@@ -374,6 +383,7 @@ static void build_insn (int insn)
 	    switch (opcstr[pos]) {
 	    case 'p': srcmode = Apdi; pos++; break;
 	    case 'P': srcmode = Aipi; pos++; break;
+	    case 'a': srcmode = Aind; pos++; break;
 	    }
 	    break;
 	case 'L':
@@ -715,7 +725,8 @@ static void build_insn (int insn)
 		    table68k[opc].mnemo = lookuptab[find].mnemo;
 		    break;
 		}
-		if (strlen(lookuptab[find].name) == 0) abort();
+		if (strlen(lookuptab[find].name) == 0)
+		    abort();
 	    }
 	}
 	else {
diff --git a/savestate.c b/savestate.c
index ab4b282f..8146e4db 100755
--- a/savestate.c
+++ b/savestate.c
@@ -541,7 +541,10 @@ void restore_state (char *filename)
 	    end = chunk + len;
 	    write_log ("unknown chunk '%s' size %d bytes\n", name, len);
 	}
-	if (len != end - chunk)
+	if (end == NULL)
+	    write_log ("Chunk '%s', size %d bytes was not accepted!\n",
+		name, len);
+	else if (len != end - chunk)
 	    write_log ("Chunk '%s' total size %d bytes but read %d bytes!\n",
 		       name, len, end - chunk);
 	xfree (chunk);
diff --git a/table68k b/table68k
index ab4c84ef..eb906ec8 100755
--- a/table68k
+++ b/table68k
@@ -25,6 +25,7 @@
 %
 % Arp: --> -(Ar)
 % ArP: --> (Ar)+
+% Ara: --> (Ar)
 % L: (xxx.L)
 %
 % Fields on a line: 
@@ -248,21 +249,31 @@
 1111 0011 01ss sSSS:22:?????:?????:10: FRESTORE s[!Dreg,Areg,Apdi,Immd]
 
 % 68030 MMU (allowed addressing modes not checked!)
-1111 0000 00ss sSSS:30:?????:?????:11: MMUOP30A s[!Immd],#1
-1111 1000 00ss sSSS:30:?????:?????:11: MMUOP30B s[!Immd]
-% Misc MMU
-1111 0101 iiii iSSS:40:?????:?????:11: MMUOP    #i,s
+1111 0000 00ss sSSS:32:?????:?????:11: MMUOP030 s[!Immd],#1
 
-% 68040 instructions
+% 68040/68060 instructions
 1111 0100 pp00 1rrr:42:-----:-----:02: CINVL    #p,Ar
 1111 0100 pp01 0rrr:42:-----:-----:02: CINVP    #p,Ar
 1111 0100 pp01 1rrr:42:-----:-----:00: CINVA    #p
 1111 0100 pp10 1rrr:42:-----:-----:02: CPUSHL   #p,Ar
 1111 0100 pp11 0rrr:42:-----:-----:02: CPUSHP   #p,Ar
 1111 0100 pp11 1rrr:42:-----:-----:00: CPUSHA   #p
+1111 0101 0000 0rrr:42:-----:-----:00: PFLUSHN  Ara
+1111 0101 0000 1rrr:42:-----:-----:00: PFLUSH   Ara
+1111 0101 0001 0rrr:42:-----:-----:00: PFLUSHAN Ara
+1111 0101 0001 1rrr:42:-----:-----:00: PFLUSHA  Ara
+1111 0101 0100 1rrr:42:-----:-----:00: PTESTR   Ara
+1111 0101 0110 1rrr:42:-----:-----:00: PTESTW   Ara
+
 % destination register number is encoded in the following word
 1111 0110 0010 0rrr:40:-----:-----:12: MOVE16   ArP,AxP
 1111 0110 00ss sSSS:40:-----:-----:12: MOVE16   s[Dreg-Aipi],L
 1111 0110 00dd dDDD:40:-----:-----:12: MOVE16   L,d[Areg-Aipi]
 1111 0110 00ss sSSS:40:-----:-----:12: MOVE16   s[Aind],L
 1111 0110 00dd dDDD:40:-----:-----:12: MOVE16   L,d[Aipi-Aind]
+
+% 68060
+1111 1000 0000 0000:52:?????:?????:10: LPSTOP   #1
+1111 0101 1000 1rrr:52:-----:-----:00: PLPAR    Ara
+1111 0101 1100 1rrr:52:-----:-----:00: PLPAW    Ara
+
-- 
2.47.3