From: Toni Wilen <twilen@winuae.net>
Date: Wed, 26 Aug 2009 15:03:08 +0000 (+0300)
Subject: imported winuaesrc1620b11.zip
X-Git-Tag: 2100~62
X-Git-Url: https://git.unchartedbackwaters.co.uk/w/?a=commitdiff_plain;h=76e4bdb6b116f77a4ee1ae3f7eeb9ed458a26526;p=francis%2Fwinuae.git

imported winuaesrc1620b11.zip
---

diff --git a/blitter.c b/blitter.c
index b3f2b13f..b2111e08 100644
--- a/blitter.c
+++ b/blitter.c
@@ -245,9 +245,10 @@ STATIC_INLINE const int *get_ch (void)
 
 STATIC_INLINE int channel_state (int cycles)
 {
-    const int *diag = get_ch ();
+    const int *diag;
     if (cycles < 0)
 	return 0;
+    diag = get_ch ();
     if (cycles < diag[0])
 	return diag[1 + cycles];
     cycles -= diag[0];
@@ -256,9 +257,10 @@ STATIC_INLINE int channel_state (int cycles)
 }
 STATIC_INLINE int channel_pos (int cycles)
 {
-    const int *diag = get_ch ();
+    const int *diag;
     if (cycles < 0)
 	return 0;
+    diag =  get_ch ();
     if (cycles < diag[0])
 	return cycles;
     cycles -= diag[0];
@@ -271,8 +273,6 @@ STATIC_INLINE int canblit (int hpos)
 {
     if (is_bitplane_dma (hpos))
 	return 0;
-    if (cycle_line[hpos] == CYCLE_CPU)
-	return -1;
     if (cycle_line[hpos])
 	return 0;
     return 1;
@@ -657,7 +657,6 @@ static void decide_blitter_line (int hsync, int hpos)
 	    // final 2 idle cycles?
 	    if (blit_final) {
 		if (blit_cyclecounter > get_ch ()[0]) {
-		    bltdpt = bltcpt;
 		    blitter_done (last_blitter_hpos);
 		    return;
 		}
@@ -667,6 +666,7 @@ static void decide_blitter_line (int hsync, int hpos)
 	    /* onedot mode and no pixel = bus write access is skipped */
 	    if (c == 4 && blitsing && blitonedot > 1) {
 		if (vblitsize == 0) {
+		    bltdpt = bltcpt;
 		    blit_final = 1;
 		    blit_cyclecounter = 0;
 		}
@@ -695,6 +695,7 @@ static void decide_blitter_line (int hsync, int hpos)
 		alloc_cycle_ext (last_blitter_hpos, CYCLE_BLITTER);
 		record_dma_blit (0x00, blt_info.bltddat, bltdpt, last_blitter_hpos);
 		if (vblitsize == 0) {
+		    bltdpt = bltcpt;
 		    blit_final = 1;
 		    blit_cyclecounter = 0;
 		    break;
@@ -839,7 +840,6 @@ STATIC_INLINE int blitter_doddma (int hpos)
 	if (blit_ch == 1)
 	    blitter_hcounter1 = blitter_hcounter2;
     }
-    blit_final = 0;
     return wd;
 }
 
@@ -989,6 +989,8 @@ void decide_blitter (int hpos)
 	    if (c == 0) {
 	        blt_info.got_cycle = 1;
 	        blit_cyclecounter++;
+		if (blit_cyclecounter == 0)
+		    blit_final = 0;
 	        blit_totalcyclecounter++;
 	        /* check if blit with zero channels has ended  */
 	        if (blit_ch == 0 && blit_cyclecounter >= blit_maxcyclecounter) {
diff --git a/cpummu.c b/cpummu.c
index 69dd7c77..aca34410 100644
--- a/cpummu.c
+++ b/cpummu.c
@@ -23,7 +23,7 @@
  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  */
 
-#define DEBUG 0
+#define DEBUG 1
 #include "sysconfig.h"
 #include "sysdeps.h"
 
@@ -304,13 +304,15 @@ static void mmu_bus_error(uaecptr addr, int fc, int write, int size)
 	regs.wb3_status = write ? 0x80 | ssw : 0;
 	if (!write)
 		ssw |= MMU_SSW_RW;
+
 	if (regs.t0 || regs.t1)
 	    ssw |= MMU_SSW_CT;
 
 	regs.mmu_fault_addr = addr;
 	regs.mmu_ssw = ssw | MMU_SSW_ATC;
 
-	//write_log (L"BUS ERROR: fc=%d w=%d log=%08x ssw=%04x\n", fc, write, addr, ssw);
+	write_log (L"BUS ERROR: fc=%d w=%d log=%08x ssw=%04x\n", fc, write, addr, ssw);
+	activate_debugger ();
 
 	THROW(2);
 }
@@ -460,6 +462,9 @@ static uaecptr REGPARAM2 mmu_lookup_pagetable(uaecptr addr, int super, int write
 	uae_u32 desc, desc_addr, wp;
 	int i, indirect;
 
+    if (addr == 0xb000)
+	write_log (L"x");
+
 	indirect = 0;
 	wp = 0;
 	desc = super ? regs.srp : regs.urp;
diff --git a/custom.c b/custom.c
index e1473eaa..b67150e9 100644
--- a/custom.c
+++ b/custom.c
@@ -237,7 +237,7 @@ static unsigned int bplcon0d, bplcon0_res, bplcon0_planes, bplcon0_planes_limit;
 static unsigned int diwstrt, diwstop, diwhigh;
 static int diwhigh_written;
 static unsigned int ddfstrt, ddfstop, ddfstrt_old_hpos, ddfstrt_old_vpos;
-static int ddf_change, badmode;
+static int ddf_change, badmode, diw_change;
 static int fmode;
 static int bplcon1_hpos;
 
@@ -260,6 +260,7 @@ static int diwfirstword, diwlastword;
 static enum diw_states diwstate, hdiwstate, ddfstate;
 int first_planes_vpos, last_planes_vpos;
 int diwfirstword_total, diwlastword_total;
+int ddffirstword_total, ddflastword_total;
 int firstword_bplcon1;
 
 static int last_copper_hpos;
@@ -2798,6 +2799,7 @@ static void calcdiw (void)
 	    plfstop = 0xff;
 	plfstrt_start = HARD_DDF_START - 2;
     }
+    diw_change = 2;
 }
 
 /* display mode changed (lores, doubling etc..), recalculate everything */
@@ -4216,12 +4218,18 @@ static void update_copper (int until_hpos)
 
 	    /* Now we know that the comparisons were successful.  We might still
 	       have to wait for the blitter though.  */
-	    if ((cop_state.saved_i2 & 0x8000) == 0 && (DMACONR (old_hpos) & 0x4000)) {
-		/* We need to wait for the blitter.  */
-		cop_state.state = COP_bltwait;
-		copper_enabled_thisline = 0;
-		unset_special (SPCFLAG_COPPER);
-		goto out;
+	    if ((cop_state.saved_i2 & 0x8000) == 0) {
+		decide_blitter (old_hpos);
+		if (bltstate != BLT_done) {
+		    /* We need to wait for the blitter.  */
+		    cop_state.state = COP_bltwait;
+		    copper_enabled_thisline = 0;
+		    unset_special (SPCFLAG_COPPER);
+		    goto out;
+		} else {
+		    if (debug_dma)
+			record_dma_event (DMA_EVENT_COPPERWAKE, old_hpos, vp);
+		}
 	    }
 
 #ifdef DEBUGGER
@@ -4297,8 +4305,9 @@ static void compute_spcflag_copper (int hpos)
 
 /*
  Copper writes to BLTSIZE: 3 blitter idle cycles, blitter normal cycle starts
- BFD=0 wait: blitter interrupt, 4 cycles, copper fetches next word
  (CPU write to BLTSIZE only have 2 idle cycles at start)
+
+ BFD=0 wait: 1 cycle (or 2 if hpos is not aligned) delay before wait ends
 */
 void blitter_done_notify (int hpos)
 {
@@ -4307,7 +4316,7 @@ void blitter_done_notify (int hpos)
     if (cop_state.state != COP_bltwait)
 	return;
 
-    hpos += 4;
+    hpos += 3;
     hpos &= ~1;
     if (hpos >= maxhpos) {
 	hpos -= maxhpos;
@@ -4316,6 +4325,8 @@ void blitter_done_notify (int hpos)
     cop_state.hpos = hpos;
     cop_state.vpos = vp;
     cop_state.state = COP_read1;
+    if (debug_dma)
+	record_dma_event (DMA_EVENT_COPPERWAKE, hpos, vp);
 
     if (dmaen (DMA_COPPER) && vp == vpos) {
 	copper_enabled_thisline = 1;
@@ -4651,6 +4662,8 @@ static void init_hardware_frame (void)
     last_planes_vpos = 0;
     diwfirstword_total = max_diwlastword;
     diwlastword_total = 0;
+    ddffirstword_total = max_diwlastword;
+    ddflastword_total = 0;
     plflastline_total = 0;
     plffirstline_total = maxvpos;
 }
@@ -5250,23 +5263,34 @@ static void hsync_handler (void)
 	    last_planes_vpos = vpos - 1;
 	}
     }
-    if (vpos >= first_planes_vpos && vpos <= last_planes_vpos) {
-	if (diwlastword > diwlastword_total)
-	    diwlastword_total = diwlastword;
-	if (diwfirstword < diwfirstword_total) {
-	    diwfirstword_total = diwfirstword;
+    if (diw_change == 0) {
+	if (vpos >= first_planes_vpos && vpos <= last_planes_vpos) {
+	    if (diwlastword > diwlastword_total)
+		diwlastword_total = diwlastword;
+	    if (diwfirstword < diwfirstword_total) {
+		diwfirstword_total = diwfirstword;
+		firstword_bplcon1 = bplcon1;
+	    }
+	}
+	if (diwstate == DIW_waiting_stop) {
+	    int f = 8 << fetchmode;
+	    if (plfstrt + f < ddffirstword_total + f)
+		ddffirstword_total = plfstrt + f;
+	    if (plfstop + 2 * f > ddflastword_total + 2 * f)
+		ddflastword_total = plfstop + 2 * f;
+	}
+	if ((plffirstline < plffirstline_total || (plffirstline_total == minfirstline && vpos > minfirstline)) && plffirstline < vpos / 2) {
 	    firstword_bplcon1 = bplcon1;
+	    if (plffirstline < minfirstline)
+		plffirstline_total = minfirstline;
+	    else
+		plffirstline_total = plffirstline;
 	}
+	if (plflastline > plflastline_total && plflastline > plffirstline_total && plflastline > maxvpos / 2)
+	    plflastline_total = plflastline;
     }
-    if ((plffirstline < plffirstline_total || (plffirstline_total == minfirstline && vpos > minfirstline)) && plffirstline < vpos / 2) {
-        firstword_bplcon1 = bplcon1;
-	if (plffirstline < minfirstline)
-	    plffirstline_total = minfirstline;
-	else
-	    plffirstline_total = plffirstline;
-    }
-    if (plflastline > plflastline_total && plflastline > plffirstline_total && plflastline > vpos / 2)
-	plflastline_total = plflastline;
+    if (diw_change > 0)
+	diw_change--;
 
 
 #if 0
@@ -6614,18 +6638,21 @@ STATIC_INLINE int dma_cycle (void)
 	decide_fetch_ce (hpos);
 	bpldma = is_bitplane_dma (hpos_old);
 	if (bltstate != BLT_done) {
-	    if (!blitpri && blitter_nasty >= BLIT_NASTY && cycle_line[hpos_old] == 0 && !bpldma)
+	    if (!blitpri && blitter_nasty >= BLIT_NASTY && cycle_line[hpos_old] == 0 && !bpldma) {
+		alloc_cycle (hpos_old, CYCLE_CPUNASTY);
 		break;
+	    }
 	    decide_blitter (hpos);
 	    // copper may have been waiting for the blitter
 	    sync_copper (hpos);
 	}
-	if (cycle_line[hpos_old] == 0 && !bpldma)
+	if (cycle_line[hpos_old] == 0 && !bpldma) {
+	    alloc_cycle (hpos_old, CYCLE_CPU);
 	    break;
+	}
 	do_cycles (1 * CYCLE_UNIT);
 	/* bus was allocated to dma channel, wait for next cycle.. */
     }
-    alloc_cycle (hpos_old, CYCLE_CPU);
     return hpos_old;
 }
 
@@ -6685,7 +6712,14 @@ void wait_cpu_cycle_write (uaecptr addr, int mode, uae_u32 v)
 void do_cycles_ce (long cycles)
 {
     int hpos;
-    while (cycles > 0) {
+    int mask = CYCLE_UNIT - 1;
+    static int extra;
+
+    extra += cycles & mask;
+    cycles &= ~mask;
+    cycles += extra & ~mask;
+    extra &= mask;
+    while (cycles >= CYCLE_UNIT) {
 	hpos = current_hpos () + 1;
 	sync_copper (hpos);
 	decide_line (hpos);
diff --git a/debug.c b/debug.c
index 2612a23c..8e77f65d 100644
--- a/debug.c
+++ b/debug.c
@@ -811,6 +811,8 @@ static void decode_dma_record (int hpos, int vpos, int toggle)
 		l2[cl2++] = 'b';
 	    if (dr->evt & DMA_EVENT_BPLFETCHUPDATE)
 		l2[cl2++] = 'p';
+	    if (dr->evt & DMA_EVENT_COPPERWAKE)
+		l2[cl2++] = 'W';
 	    if (i < cols - 1 && h < maxh - 1) {
 		l1[cl + col - 1] = 32;
 		l2[cl + col - 1] = 32;
diff --git a/disk.c b/disk.c
index 949e0d51..945bb01b 100644
--- a/disk.c
+++ b/disk.c
@@ -2502,7 +2502,7 @@ static void disk_dmafinished (void)
 	int dr, mfmpos = -1;
 	write_log (L"disk dma finished %08X MFMpos=", dskpt);
 	for (dr = 0; dr < MAX_FLOPPY_DRIVES; dr++)
-	    write_log (L"%d%s", floppy[dr].mfmpos, dr < MAX_FLOPPY_DRIVES - 1 ? "," : "");
+	    write_log (L"%d%s", floppy[dr].mfmpos, dr < MAX_FLOPPY_DRIVES - 1 ? L"," : L"");
 	write_log (L"\n");
     }
 }
@@ -3079,12 +3079,12 @@ void DSKLEN (uae_u16 v, int hpos)
     }
     if (dr == 4) {
         write_log (L"disk %s DMA started, drvmask=%x motormask=%x\n",
-	    dskdmaen == 3 ? "write" : "read", selected ^ 15, motormask);
+	    dskdmaen == 3 ? L"write" : L"read", selected ^ 15, motormask);
 	noselected = 1;
     } else {
 	if (disk_debug_logging > 0) {
 	    write_log (L"disk %s DMA started, drvmask=%x track %d mfmpos %d dmaen=%d\n",
-		dskdmaen == 3 ? "write" : "read", selected ^ 15,
+		dskdmaen == 3 ? L"write" : L"read", selected ^ 15,
 		floppy[dr].cyl * 2 + side, floppy[dr].mfmpos, dma_enable);
 	    disk_dma_debugmsg ();
 	}
diff --git a/drawing.c b/drawing.c
index f7008c73..c527846a 100644
--- a/drawing.c
+++ b/drawing.c
@@ -319,6 +319,7 @@ void notice_screen_contents_lost (void)
 extern int plffirstline_total, plflastline_total;
 extern int first_planes_vpos, last_planes_vpos;
 extern int diwfirstword_total, diwlastword_total;
+extern int ddffirstword_total, ddflastword_total;
 extern int firstword_bplcon1;
 
 #define MIN_DISPLAY_W 256
@@ -355,6 +356,18 @@ int get_custom_limits (int *pw, int *ph, int *pdx, int *pdy)
 	    diwfirstword_total = 48 << currprefs.gfx_resolution;
 	if (diwlastword_total > (448 << currprefs.gfx_resolution))
 	    diwlastword_total = 448 << currprefs.gfx_resolution;
+	ddffirstword_total = coord_hw_to_window_x (ddffirstword_total * 2 + DIW_DDF_OFFSET);
+	ddflastword_total = coord_hw_to_window_x (ddflastword_total * 2 + DIW_DDF_OFFSET);
+	if (ddffirstword_total < (48 << currprefs.gfx_resolution))
+	    ddffirstword_total = 48 << currprefs.gfx_resolution;
+	if (ddflastword_total > (448 << currprefs.gfx_resolution))
+	    ddflastword_total = 448 << currprefs.gfx_resolution;
+	if (0 && !(currprefs.chipset_mask & CSMASK_AGA)) {
+	    if (ddffirstword_total > diwfirstword_total)
+		diwfirstword_total = ddffirstword_total;
+	    if (ddflastword_total < diwlastword_total)
+		diwlastword_total = ddflastword_total;
+	}
     }
 
     w = diwlastword_total - diwfirstword_total;
diff --git a/gencpu.c b/gencpu.c
index 7f8a58e8..f420af7c 100644
--- a/gencpu.c
+++ b/gencpu.c
@@ -49,6 +49,7 @@ static int cpulengths[65536];
 #define GF_AA 7
 #define GF_NOREFILL 8
 #define GF_PREFETCH 16
+#define GF_FC 32
 
 /* For the current opcode, the next lower level that will have different code.
  * Initialized to -1 for each opcode. If it remains unchanged, indicates we
@@ -670,11 +671,20 @@ static void genamode2 (amodes mode, char *reg, wordsizes size, char *name, int g
 	    default: abort ();
 	    }
 	} else if (using_mmu) {
-	    switch (size) {
-	    case sz_byte: insn_n_cycles += 4; printf ("\tuae_s8 %s = get_byte_mmu (%sa);\n", name, name); break;
-	    case sz_word: insn_n_cycles += 4; printf ("\tuae_s16 %s = get_word_mmu (%sa);\n", name, name); break;
-	    case sz_long: insn_n_cycles += 8; printf ("\tuae_s32 %s = get_long_mmu (%sa);\n", name, name); break;
-	    default: abort ();
+	    if (flags & GF_FC) {
+		switch (size) {
+		case sz_byte: insn_n_cycles += 4; printf ("\tuae_s8 %s = sfc_get_byte (%sa);\n", name, name); break;
+		case sz_word: insn_n_cycles += 4; printf ("\tuae_s16 %s = sfc_get_word (%sa);\n", name, name); break;
+		case sz_long: insn_n_cycles += 8; printf ("\tuae_s32 %s = sfc_get_long (%sa);\n", name, name); break;
+		default: abort ();
+		}
+	    } else {
+		switch (size) {
+		case sz_byte: insn_n_cycles += 4; printf ("\tuae_s8 %s = get_byte_mmu (%sa);\n", name, name); break;
+		case sz_word: insn_n_cycles += 4; printf ("\tuae_s16 %s = get_word_mmu (%sa);\n", name, name); break;
+		case sz_long: insn_n_cycles += 8; printf ("\tuae_s32 %s = get_long_mmu (%sa);\n", name, name); break;
+		default: abort ();
+		}
 	    }
 	} else {
 	    switch (size) {
@@ -722,7 +732,7 @@ static void genamode_e3 (amodes mode, char *reg, wordsizes size, char *name, int
     genamode2 (mode, reg, size, name, getv, movem, flags, e3fudge);
 }
 
-static void genastore_2 (char *from, amodes mode, char *reg, wordsizes size, char *to, int store_dir)
+static void genastore_2 (char *from, amodes mode, char *reg, wordsizes size, char *to, int store_dir, int flags)
 {
     switch (mode) {
      case Dreg:
@@ -810,19 +820,28 @@ static void genastore_2 (char *from, amodes mode, char *reg, wordsizes size, cha
 	    switch (size) {
 	     case sz_byte:
 		insn_n_cycles += 4;
-		printf ("\tput_byte_mmu (%sa,%s);\n", to, from);
+		if (flags & GF_FC)
+		    printf ("\tdfc_put_byte (%sa,%s);\n", to, from);
+		else
+		    printf ("\tput_byte_mmu (%sa,%s);\n", to, from);
 		break;
 	     case sz_word:
 		insn_n_cycles += 4;
 		if (cpu_level < 2 && (mode == PC16 || mode == PC8r))
 		    abort ();
-		printf ("\tput_word_mmu (%sa,%s);\n", to, from);
+		if (flags & GF_FC)
+		    printf ("\tdfc_put_word (%sa,%s);\n", to, from);
+		else
+		    printf ("\tput_word_mmu (%sa,%s);\n", to, from);
 		break;
 	     case sz_long:
 		insn_n_cycles += 8;
 		if (cpu_level < 2 && (mode == PC16 || mode == PC8r))
 		    abort ();
-		printf ("\tput_long_mmu (%sa,%s);\n", to, from);
+		if (flags & GF_FC)
+		    printf ("\tdfc_put_long (%sa,%s);\n", to, from);
+		else
+		    printf ("\tput_long_mmu (%sa,%s);\n", to, from);
 		break;
 	     default:
 		abort ();
@@ -867,11 +886,15 @@ static void genastore_2 (char *from, amodes mode, char *reg, wordsizes size, cha
 
 static void genastore (char *from, amodes mode, char *reg, wordsizes size, char *to)
 {
-    genastore_2 (from, mode, reg, size, to, 0);
+    genastore_2 (from, mode, reg, size, to, 0, 0);
 }
 static void genastore_rev (char *from, amodes mode, char *reg, wordsizes size, char *to)
 {
-    genastore_2 (from, mode, reg, size, to, 1);
+    genastore_2 (from, mode, reg, size, to, 1, 0);
+}
+static void genastore_dfc (char *from, amodes mode, char *reg, wordsizes size, char *to)
+{
+    genastore_2 (from, mode, reg, size, to, 1, GF_FC);
 }
 
 
@@ -942,20 +965,20 @@ static void genmovemel_ce (uae_u16 opcode)
 
 static void genmovemle (uae_u16 opcode)
 {
-    char putcode[100];
+    char *putcode;
     int size = table68k[opcode].size == sz_long ? 4 : 2;
 
     if (using_mmu) {
 	if (table68k[opcode].size == sz_long) {
-	    strcpy (putcode, "put_long_mmu (srca");
+	    putcode = "put_long_mmu (srca";
 	} else {
-	    strcpy (putcode, "put_word_mmu (srca");
+	    putcode = "put_word_mmu (srca";
 	}
     } else {
 	if (table68k[opcode].size == sz_long) {
-	    strcpy (putcode, "put_long (srca");
+	    putcode = "put_long (srca";
 	} else {
-	    strcpy (putcode, "put_word (srca");
+	    putcode = "put_word (srca";
 	}
     }
     count_write += table68k[opcode].size == sz_long ? 2 : 1;
@@ -2855,23 +2878,20 @@ static void gen_opcode (unsigned long int opcode)
 	    printf ("\t}\n");
 	}
 	break;
-    case i_MOVES:		/* ignore DFC and SFC because we have no MMU */
+    case i_MOVES: /* ignore DFC and SFC when using_mmu == false */
     {
 	int old_brace_level;
-	if (using_mmu) {
-	    printf ("write_log(L\"WARNING: MOVES currently ignores MMU!\");\n");
-	}
 	genamode (curi->smode, "srcreg", curi->size, "extra", 1, 0, 0);
 	printf ("\tif (extra & 0x800)\n");
 	old_brace_level = n_braces;
 	start_brace ();
 	printf ("\tuae_u32 src = regs.regs[(extra >> 12) & 15];\n");
 	genamode (curi->dmode, "dstreg", curi->size, "dst", 2, 0, 0);
-	genastore ("src", curi->dmode, "dstreg", curi->size, "dst");
+	genastore_dfc ("src", curi->dmode, "dstreg", curi->size, "dst");
 	pop_braces (old_brace_level);
-	printf ("else");
+	printf ("\telse");
 	start_brace ();
-	genamode (curi->dmode, "dstreg", curi->size, "src", 1, 0, 0);
+	genamode (curi->dmode, "dstreg", curi->size, "src", 1, 0, GF_FC);
 	printf ("\tif (extra & 0x8000) {\n");
 	switch (curi->size) {
 	case sz_byte: printf ("\tm68k_areg (regs, (extra >> 12) & 7) = (uae_s32)(uae_s8)src;\n"); break;
@@ -3117,64 +3137,57 @@ static void gen_opcode (unsigned long int opcode)
 	break;
 
      case i_MOVE16:
-	 if (using_ce020) {
-	    if ((opcode & 0xfff8) == 0xf620) {
-		 /* MOVE16 (Ax)+,(Ay)+ */
-		 printf ("\tuaecptr mems = m68k_areg (regs, srcreg) & ~15, memd;\n");
-		 printf ("\tdstreg = (%s >> 12) & 7;\n", gen_nextiword (0));
-		 printf ("\tmemd = m68k_areg (regs, dstreg) & ~15;\n");
-		 printf ("\tput_long_ce020 (memd, get_long_ce020 (mems));\n");
-		 printf ("\tput_long_ce020 (memd+4, get_long_ce020 (mems+4));\n");
-		 printf ("\tput_long_ce020 (memd+8, get_long_ce020 (mems+8));\n");
-		 printf ("\tput_long_ce020 (memd+12, get_long_ce020 (mems+12));\n");
-		 printf ("\tif (srcreg != dstreg)\n");
-		 printf ("\tm68k_areg (regs, srcreg) += 16;\n");
-		 printf ("\tm68k_areg (regs, dstreg) += 16;\n");
-	    } else {
-		 /* Other variants */
-		 genamode (curi->smode, "srcreg", curi->size, "mems", 0, 2, 0);
-		 genamode (curi->dmode, "dstreg", curi->size, "memd", 0, 2, 0);
-		 printf ("\tmemsa &= ~15;\n");
-		 printf ("\tmemda &= ~15;\n");
-		 printf ("\tput_long_ce020 (memda, get_long_ce020 (memsa));\n");
-		 printf ("\tput_long_ce020 (memda+4, get_long_ce020 (memsa+4));\n");
-		 printf ("\tput_long_ce020 (memda+8, get_long_ce020 (memsa+8));\n");
-		 printf ("\tput_long_ce020 (memda+12, get_long_ce020 (memsa+12));\n");
-		 if ((opcode & 0xfff8) == 0xf600)
-		     printf ("\tm68k_areg (regs, srcreg) += 16;\n");
-		 else if ((opcode & 0xfff8) == 0xf608)
-		     printf ("\tm68k_areg (regs, dstreg) += 16;\n");
-	    }
-	 } else {
-	    if ((opcode & 0xfff8) == 0xf620) {
-		 /* MOVE16 (Ax)+,(Ay)+ */
-		 printf ("\tuaecptr mems = m68k_areg (regs, srcreg) & ~15, memd;\n");
-		 printf ("\tdstreg = (%s >> 12) & 7;\n", gen_nextiword (0));
-		 printf ("\tmemd = m68k_areg (regs, dstreg) & ~15;\n");
-		 printf ("\tput_long (memd, get_long (mems));\n");
-		 printf ("\tput_long (memd+4, get_long (mems+4));\n");
-		 printf ("\tput_long (memd+8, get_long (mems+8));\n");
-		 printf ("\tput_long (memd+12, get_long (mems+12));\n");
-		 printf ("\tif (srcreg != dstreg)\n");
-		 printf ("\tm68k_areg (regs, srcreg) += 16;\n");
-		 printf ("\tm68k_areg (regs, dstreg) += 16;\n");
-	    } else {
-		 /* Other variants */
-		 genamode (curi->smode, "srcreg", curi->size, "mems", 0, 2, 0);
-		 genamode (curi->dmode, "dstreg", curi->size, "memd", 0, 2, 0);
-		 printf ("\tmemsa &= ~15;\n");
-		 printf ("\tmemda &= ~15;\n");
-		 printf ("\tput_long (memda, get_long (memsa));\n");
-		 printf ("\tput_long (memda+4, get_long (memsa+4));\n");
-		 printf ("\tput_long (memda+8, get_long (memsa+8));\n");
-		 printf ("\tput_long (memda+12, get_long (memsa+12));\n");
-		 if ((opcode & 0xfff8) == 0xf600)
-		     printf ("\tm68k_areg (regs, srcreg) += 16;\n");
-		 else if ((opcode & 0xfff8) == 0xf608)
-		     printf ("\tm68k_areg (regs, dstreg) += 16;\n");
-	    }
-	 }
-	break;
+     {
+	char *src, *dst;
+	if (using_ce020) {
+	    src = "get_long_ce020";
+	    dst = "put_long_ce020";
+	} else if (using_mmu) {
+	    src = "get_long_mmu";
+	    dst = "put_long_mmu";
+	} else {
+	    src = "get_long";
+	    dst = "put_long";
+	}
+	if ((opcode & 0xfff8) == 0xf620) {
+	     /* MOVE16 (Ax)+,(Ay)+ */
+	    printf ("\tuae_u32 v1, v2, v3, v4;\n");
+	    printf ("\tuaecptr mems = m68k_areg (regs, srcreg) & ~15, memd;\n");
+	    printf ("\tdstreg = (%s >> 12) & 7;\n", gen_nextiword (0));
+	    printf ("\tmemd = m68k_areg (regs, dstreg) & ~15;\n");
+	    printf ("\tv1 = %s (mems);\n", src);
+	    printf ("\tv2 = %s (mems + 4);\n", src);
+	    printf ("\tv3 = %s (mems + 8);\n", src);
+	    printf ("\tv4 = %s (mems + 12);\n", src);
+	    printf ("\t%s (memd , v1);\n", dst);
+	    printf ("\t%s (memd + 4, v2);\n", dst);
+	    printf ("\t%s (memd + 8, v3);\n", dst);
+	    printf ("\t%s (memd + 12, v4);\n", dst);
+	    printf ("\tif (srcreg != dstreg)\n");
+	    printf ("\t\tm68k_areg (regs, srcreg) += 16;\n");
+	    printf ("\tm68k_areg (regs, dstreg) += 16;\n");
+	} else {
+	    /* Other variants */
+	    printf ("\tuae_u32 v1, v2, v3, v4;\n");
+	    genamode (curi->smode, "srcreg", curi->size, "mems", 0, 2, 0);
+	    genamode (curi->dmode, "dstreg", curi->size, "memd", 0, 2, 0);
+	    printf ("\tmemsa &= ~15;\n");
+	    printf ("\tmemda &= ~15;\n");
+	    printf ("\tv1 = %s (memsa);\n", src);
+	    printf ("\tv2 = %s (memsa + 4);\n", src);
+	    printf ("\tv3 = %s (memsa + 8);\n", src);
+	    printf ("\tv4 = %s (memsa + 12);\n", src);
+	    printf ("\t%s (memda , v1);\n", dst);
+	    printf ("\t%s (memda + 4, v2);\n", dst);
+	    printf ("\t%s (memda + 8, v3);\n", dst);
+	    printf ("\t%s (memda + 12, v4);\n", dst);
+	    if ((opcode & 0xfff8) == 0xf600)
+	         printf ("\tm68k_areg (regs, srcreg) += 16;\n");
+	     else if ((opcode & 0xfff8) == 0xf608)
+	         printf ("\tm68k_areg (regs, dstreg) += 16;\n");
+	}
+    }
+    break;
 
     case i_PFLUSHN:
     case i_PFLUSH:
diff --git a/include/cpummu.h b/include/cpummu.h
index c2b7e2d6..8c349e5c 100644
--- a/include/cpummu.h
+++ b/include/cpummu.h
@@ -164,7 +164,10 @@ extern void mmu_dump_tables(void);
 #define MMU_SSW_LK		0x0200
 #define MMU_SSW_ATC		0x0400
 #define MMU_SSW_MA		0x0800
+#define MMU_SSW_CM	0x1000
 #define MMU_SSW_CT	0x2000
+#define MMU_SSW_CU	0x4000
+#define MMU_SSW_CP	0x8000
 
 #define TTR_I0	4
 #define TTR_I1	5
diff --git a/include/custom.h b/include/custom.h
index 7367581f..d86b2283 100644
--- a/include/custom.h
+++ b/include/custom.h
@@ -122,10 +122,10 @@ extern frame_time_t syncbase;
 #define CYCLE_STROBE	0x02
 #define CYCLE_MISC	0x04
 #define CYCLE_SPRITE	0x08
-#define CYCLE_BITPLANE	0x10
-#define CYCLE_COPPER	0x20
-#define CYCLE_BLITTER	0x40
-#define CYCLE_CPU	0x80
+#define CYCLE_COPPER	0x10
+#define CYCLE_BLITTER	0x20
+#define CYCLE_CPU	0x40
+#define CYCLE_CPUNASTY	0x80
 
 extern unsigned long frametime, timeframes;
 extern int plfstrt, plfstop, plffirstline, plflastline;
diff --git a/include/debug.h b/include/debug.h
index 93d73ae5..2e0c0821 100644
--- a/include/debug.h
+++ b/include/debug.h
@@ -83,6 +83,7 @@ struct dma_rec
 #define DMA_EVENT_BLITNASTY 2
 #define DMA_EVENT_BLITFINISHED 4
 #define DMA_EVENT_BPLFETCHUPDATE 8
+#define DMA_EVENT_COPPERWAKE 16
 
 extern struct dma_rec *record_dma (uae_u16 reg, uae_u16 dat, uae_u32 addr, int hpos, int vpos);
 extern void record_dma_reset (void);
diff --git a/include/newcpu.h b/include/newcpu.h
index 4519ab95..4b65308d 100644
--- a/include/newcpu.h
+++ b/include/newcpu.h
@@ -93,13 +93,25 @@ typedef double fptype;
 #define CPU020_CLOCK_MULT 4
 
 #define CACHELINES020 64
-#define CACHELINES040 1024 // 040 cache really isn't like this..
 struct cache020
 {
     uae_u32 data;
     uae_u32 tag;
     uae_u32 valid:1;
 };
+#define CACHESETS040 64
+struct cache040set
+{
+    uae_u32 data[4];
+    int valid[4];
+    uae_u32 tag;
+};
+#define CACHELINES040 4
+struct cache040
+{
+    struct cache040set cs[4];
+    int count;
+};
 
 extern struct regstruct
 {
diff --git a/newcpu.c b/newcpu.c
index 2bfc4378..1730da62 100644
--- a/newcpu.c
+++ b/newcpu.c
@@ -79,8 +79,8 @@ static uae_u64 srp_030, crp_030;
 static uae_u32 tt0_030, tt1_030, tc_030;
 static uae_u16 mmusr_030;
 
-static struct cache020 cacheline020[CACHELINES020];
-static struct cache020 cacheline040[CACHELINES040];
+static struct cache020 caches020[CACHELINES020];
+static struct cache040 caches040[CACHELINES040];
 
 #if COUNT_INSTRS
 static unsigned long int instrcount[65536];
@@ -134,16 +134,16 @@ void dump_counts (void)
 
 static void set_cpu_caches (void)
 {
-    int i;
+    int i, j;
 
     if (currprefs.cpu_model < 68040) {
 	if (regs.cacr & 0x08) { // Clear Cache
 	    for (i = 0; i < CACHELINES020; i++)
-		cacheline020[i].valid = 0;
+		caches020[i].valid = 0;
 	    regs.prefetch020addr = 0xff000000;
 	}
 	if (regs.cacr & 0x04) { // Clear Entry
-	    cacheline020[(regs.caar >> 2) & 0x3f].valid = 0;
+	    caches020[(regs.caar >> 2) & 0x3f].valid = 0;
 	}
 #ifdef JIT
 	set_cache_state (regs.cacr & 1);
@@ -157,8 +157,14 @@ static void set_cpu_caches (void)
 	set_cache_state ((regs.cacr & 0x8000) ? 1 : 0);
 #endif
 	if (!(regs.cacr & 0x8000)) {
-	    for (i = 0; i < CACHELINES040; i++)
-		cacheline040[i].valid = 0;
+	    for (i = 0; i < CACHESETS040; i++) {
+		for (j = 0; j < CACHELINES040; j++) {
+		    caches040[i].cs[j].valid[0] = 0;
+		    caches040[i].cs[j].valid[1] = 0;
+		    caches040[i].cs[j].valid[2] = 0;
+		    caches040[i].cs[j].valid[3] = 0;
+		}
+	    }
 	    regs.prefetch020addr = 0xff000000;
 	}
     }
@@ -1212,7 +1218,7 @@ static void Exception_normal (int nr, uaecptr oldpc)
 			m68k_areg (regs, 7) -= 2;
 			put_word (m68k_areg (regs, 7), regs.mmu_ssw);
 			m68k_areg (regs, 7) -= 4;
-			put_word (m68k_areg (regs, 7), regs.mmu_fault_addr);
+			put_long (m68k_areg (regs, 7), regs.mmu_fault_addr);
 
 		    } else {
 
@@ -2587,6 +2593,9 @@ static void m68k_run_mmu040 (void)
 	    }
 #ifdef _WIN32
 	} __except (GetExceptionCode () == MMUEX) {
+	    // movem to memory?
+	    if ((opcode & 0xff80) == 0x4880)
+		regs.mmu_ssw |= MMU_SSW_CM;
 	    // write bus errors restart at next instruction
 	    if (regs.wb3_status & 0x80) {
 //		write_log (L"%08x %04x %d\n", regs.fault_pc, opcode, cpufunclen[opcode]);
@@ -3645,23 +3654,30 @@ int getDivs68kCycles (uae_s32 dividend, uae_s16 divisor)
 
     return mcycles * 2;
 }
-
+#if 0
 STATIC_INLINE void fill_cache040 (uae_u32 addr)
 {
-    int index;
+    int index, i, j;
     uae_u32 tag;
     uae_u32 data;
-    struct cache020 *c;
-
-    addr &= ~3;
-    index = (addr >> 2) & (CACHELINES040 - 1);
-    tag = regs.s | (addr & ~((CACHELINES040 << 2) - 1));
-    c = &cacheline040[index];
-    if (c->valid && c->tag == tag) {
-	// cache hit
-	regs.prefetch020addr = addr;
-	regs.prefetch020data = c->data;
-	return;
+    struct cache040 *c;
+
+    addr &= ~15;
+    index = (addr >> 4) & (CACHESETS040 - 1);
+    tag = regs.s | (addr & ~((CACHESETS040 << 4) - 1));
+    c = &caches040[index];
+    for (i = 0; i < CACHELINES040; i++) {
+	struct cache040set *cs = &c->cs;
+	for (j = 0; j < 4; j++) {
+	    if (cs->valid[j] && c->tag == tag[j]) {
+		// cache hit
+		regs.prefetch020addr = addr;
+		regs.prefetch020data = c->data[j];
+		return;
+	    } if (cs->valid[j] == 0) {
+		inv = &cs->valid[j];
+	    }
+	}
     }
     // cache miss
     data = mem_access_delay_longi_read_020 (addr);
@@ -3673,6 +3689,7 @@ STATIC_INLINE void fill_cache040 (uae_u32 addr)
     regs.prefetch020addr = addr;
     regs.prefetch020data = data;
 }
+#endif
 
 STATIC_INLINE void fill_cache020 (uae_u32 addr)
 {
@@ -3684,7 +3701,7 @@ STATIC_INLINE void fill_cache020 (uae_u32 addr)
     addr &= ~3;
     index = (addr >> 2) & (CACHELINES020 - 1);
     tag = regs.s | (addr & ~((CACHELINES020 << 2) - 1));
-    c = &cacheline020[index];
+    c = &caches020[index];
     if (c->valid && c->tag == tag) {
 	// cache hit
 	regs.prefetch020addr = addr;
@@ -3704,9 +3721,11 @@ STATIC_INLINE void fill_cache020 (uae_u32 addr)
 
 void fill_cache0x0 (uae_u32 addr)
 {
+#if 0
     if (currprefs.cpu_model >= 68040)
 	fill_cache040 (addr);
     else
+#endif
 	fill_cache020 (addr);
 }
 
diff --git a/od-win32/sounddep/sound.c b/od-win32/sounddep/sound.c
index 1cd7a936..12530a0e 100644
--- a/od-win32/sounddep/sound.c
+++ b/od-win32/sounddep/sound.c
@@ -694,7 +694,7 @@ static int open_audio_wasapi (struct sound_data *sd, int index, int exclusive)
 {
     HRESULT hr;
     struct sound_dp *s = sd->data;
-    WAVEFORMATEX *pwfx;
+    WAVEFORMATEX *pwfx, *pwfx_saved;
     WAVEFORMATEXTENSIBLE wavfmt;
     int final;
     LPWSTR name = NULL;
@@ -753,6 +753,7 @@ static int open_audio_wasapi (struct sound_data *sd, int index, int exclusive)
 
     final = 0;
     rncnt = 0;
+    pwfx_saved = NULL;
     for (;;) {
 
 	if (sd->channels == 6) {
@@ -788,9 +789,18 @@ static int open_audio_wasapi (struct sound_data *sd, int index, int exclusive)
 	hr = s->pAudioClient->lpVtbl->IsFormatSupported (s->pAudioClient, sharemode, &wavfmt.Format, &pwfx);
 	if (SUCCEEDED (hr) && hr != S_FALSE)
 	    break;
-	write_log (L"WASAPI: IsFormatSupported(%d,%08X,%d) %08X\n", sd->channels, rn[rncnt], sd->freq, hr);
+	write_log (L"WASAPI: IsFormatSupported(%d,%08X,%d) (%d,%d) %08X\n",
+	    sd->channels, rn[rncnt], sd->freq,
+	    pwfx ? pwfx->nChannels : -1, pwfx ? pwfx->nSamplesPerSec : -1,
+	    hr);
+	if (final && SUCCEEDED (hr))
+	    break;
 	if (hr != AUDCLNT_E_UNSUPPORTED_FORMAT && hr != S_FALSE)
 	    goto error;
+	if (hr == S_FALSE && pwfx_saved == NULL) {
+	    pwfx_saved = pwfx;
+	    pwfx = NULL;
+	}
 	rncnt++;
 	if (rn[rncnt])
 	    continue;
@@ -810,10 +820,10 @@ static int open_audio_wasapi (struct sound_data *sd, int index, int exclusive)
 	    continue;
 	}
 	final = 1;
-	if (pwfx == NULL)
+	if (pwfx_saved == NULL)
 	    goto error;
-	sd->channels = pwfx->nChannels;
-	sd->freq = pwfx->nSamplesPerSec;
+	sd->channels = pwfx_saved->nChannels;
+	sd->freq = pwfx_saved->nSamplesPerSec;
     }
 
     size = sd->sndbufsize * sd->channels * 16 / 8;
@@ -915,12 +925,14 @@ static int open_audio_wasapi (struct sound_data *sd, int index, int exclusive)
 	name, s->wasapiexclusive, sd->channels, sd->freq, sd->sndbufsize, s->bufferFrameCount);
 
     CoTaskMemFree (pwfx);
+    CoTaskMemFree (pwfx_saved);
     CoTaskMemFree (name);
 
     return 1;
 
 error:
     CoTaskMemFree (pwfx);
+    CoTaskMemFree (pwfx_saved);
     CoTaskMemFree (name);
     close_audio_wasapi (sd);
     return 0;
diff --git a/od-win32/win32.h b/od-win32/win32.h
index dffc8353..3502854a 100644
--- a/od-win32/win32.h
+++ b/od-win32/win32.h
@@ -17,8 +17,8 @@
 
 #define WINUAEPUBLICBETA 1
 
-#define WINUAEBETA L"10"
-#define WINUAEDATE MAKEBD(2009, 8, 23)
+#define WINUAEBETA L"11"
+#define WINUAEDATE MAKEBD(2009, 8, 26)
 #define WINUAEEXTRA L""
 #define WINUAEREV L""
 
diff --git a/od-win32/win32gfx.c b/od-win32/win32gfx.c
index 7e94ab8e..373e9e81 100644
--- a/od-win32/win32gfx.c
+++ b/od-win32/win32gfx.c
@@ -1709,9 +1709,9 @@ int vsync_switchmode (int hz, int oldhz)
 	return tempvsync;
 
     if (currprefs.ntscmode)
-	newh = h * 50 / 60;
-    else
 	newh = h * 60 / 50;
+    else
+	newh = h * 50 / 60;
     hz = hz * dbl;
 
     found = NULL;
diff --git a/od-win32/winuaechangelog.txt b/od-win32/winuaechangelog.txt
index f3d06cf9..4bd9bf35 100644
--- a/od-win32/winuaechangelog.txt
+++ b/od-win32/winuaechangelog.txt
@@ -1,4 +1,16 @@
 
+Beta 11:
+
+- fixed 68020 CE-mode cycle counting bug (too slow memory accesses)
+- MOVE16 MMU support update (fetch all data before writing), MOVES MMU
+  support update, MOVEM to memory MMU special handling support
+  (still not working with mmulib, reason unknown)
+- b10 blitter cycles given to CPU update was very wrong..
+- copper BFD=0 wait woke up too early in some cases (and real fix here)
+- calculate autoscale settings only when display window registers have
+  been stable at least 1 full scanline (also fixes NTSC "jumping")
+- WASAPI shared mode closest match update
+
 Beta 10:
 
 - and more reading of write-only register fixes, original Cardiaxx