XFree86 4.3.0 mergeX_4_3_0-20030324-merge

1 files changed, 794 insertions, 253 deletions

diff --git a/xc/programs/Xserver/hw/xfree86/drivers/sis/sis300_accel.c b/xc/programs/Xserver/hw/xfree86/drivers/sis/sis300_accel.c
index 2e0c8ae2c..8db9d49c2 100644
--- a/xc/programs/Xserver/hw/xfree86/drivers/sis/sis300_accel.c
+++ b/xc/programs/Xserver/hw/xfree86/drivers/sis/sis300_accel.c
@@ -1,14 +1,33 @@
-/* $XFree86: xc/programs/Xserver/hw/xfree86/drivers/sis/sis300_accel.c,v 1.12 2002/04/04 14:05:47 eich Exp $ */
-
+/* $XFree86: xc/programs/Xserver/hw/xfree86/drivers/sis/sis300_accel.c,v 1.14 2003/01/29 15:42:16 eich Exp $ */
 /*
+ * 2D Acceleration for SiS300, SiS540, SiS630, SiS730, SiS530, SiS620
+ *
+ * Copyright Xavier Ducoin <x.ducoin@lectra.com>
+ * Copyright 2002 by Thomas Winischhofer, Vienna, Austria
+ *
+ * Permission to use, copy, modify, distribute, and sell this software and its
+ * documentation for any purpose is hereby granted without fee, provided that
+ * the above copyright notice appear in all copies and that both that
+ * copyright notice and this permission notice appear in supporting
+ * documentation, and that the name of the copyright holders not be used in
+ * advertising or publicity pertaining to distribution of the software without
+ * specific, written prior permission.  The copyright holders make no representations
+ * about the suitability of this software for any purpose.  It is provided
+ * "as is" without express or implied warranty.
  *
- *      Acceleration for SiS300 SiS630 SiS540.
- *      It is done in a separate file because the register formats are 
- *      very different from the previous chips.
+ * THE COPYRIGHT HOLDERS DISCLAIM ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
+ * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS, IN NO
+ * EVENT SHALL THE COPYRIGHT HOLDERS BE LIABLE FOR ANY SPECIAL, INDIRECT OR
+ * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF USE,
+ * DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER
+ * TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
+ * PERFORMANCE OF THIS SOFTWARE.
  *
+ *      Authors:
  *
- *      
  *      Xavier Ducoin <x.ducoin@lectra.com>
+ *      Thomas Winischhofer <thomas@winischhofer.net>
+ *
  */
 #if 0
 #define DEBUG
@@ -25,11 +44,19 @@
 #include "sis.h"
 #include "sis300_accel.h"
 
-#ifdef  DEBUG
-static void MMIODump(ScrnInfoPtr pScrn);
+#ifdef SISDUALHEAD
+/* TW: This is the offset to the memory for each head */
+#define HEADOFFSET 	(pSiS->dhmOffset)
 #endif
 
-Bool SiS300AccelInit(ScreenPtr pScreen);
+#undef STSCE    	/* TW: Use/Don't use ScreenToScreenColorExpand - does not work */
+
+#undef TRAP     	/* TW: Use/Don't use Trapezoid Fills - does not work - XAA provides
+		         * illegal trapezoid data (left and right edges cross each other
+			 * sometimes) which causes drawing errors. Further, I have not found
+			 * out how to draw polygones with a height greater than 127...
+                         */
+
 static void SiSInitializeAccelerator(ScrnInfoPtr pScrn);
 static void SiSSync(ScrnInfoPtr pScrn);
 static void SiSSetupForScreenToScreenCopy(ScrnInfoPtr pScrn,
@@ -42,6 +69,11 @@ static void SiSSetupForSolidFill(ScrnInfoPtr pScrn, int color,
                                 int rop, unsigned int planemask);
 static void SiSSubsequentSolidFillRect(ScrnInfoPtr pScrn,
                                 int x, int y, int w, int h);
+#ifdef TRAP
+static void SiSSubsequentSolidFillTrap(ScrnInfoPtr pScrn, int y, int h,
+	                        int left, int dxL, int dyL, int eL,
+	                        int right, int dxR, int dyR, int eR);
+#endif
 static void SiSSetupForSolidLine(ScrnInfoPtr pScrn, int color,
                                 int rop, unsigned int planemask);
 static void SiSSubsequentSolidTwoPointLine(ScrnInfoPtr pScrn, int x1,
@@ -60,6 +92,13 @@ static void SiSSetupForMonoPatternFill(ScrnInfoPtr pScrn,
 static void SiSSubsequentMonoPatternFill(ScrnInfoPtr pScrn,
                                 int patx, int paty,
                                 int x, int y, int w, int h);
+#ifdef TRAP
+static void SiSSubsequentMonoPatternFillTrap(ScrnInfoPtr pScrn,
+                                int patx, int paty,
+                                int y, int h,
+                                int left, int dxL, int dyL, int eL,
+	                        int right, int dxR, int dyR, int eR );
+#endif
 #if 0
 static void SiSSetupForColorPatternFill(ScrnInfoPtr pScrn,
                                 int patx, int paty, int rop,
@@ -68,11 +107,15 @@ static void SiSSetupForColorPatternFill(ScrnInfoPtr pScrn,
 static void SiSSubsequentColorPatternFill(ScrnInfoPtr pScrn,
                                 int patx, int paty,
                                 int x, int y, int w, int h);
+#endif
+#if 0
 static void SiSSetupForCPUToScreenColorExpand(ScrnInfoPtr pScrn,
                                 int fg, int bg,
                                 int rop, unsigned int planemask);
 static void SiSSubsequentCPUToScreenColorExpand(ScrnInfoPtr pScrn,
                                 int x, int y, int w, int h, int skipleft);
+#endif
+#ifdef STSCE
 static void SiSSetupForScreenToScreenColorExpand(ScrnInfoPtr pScrn,
                                 int fg, int bg,
                                 int rop, unsigned int planemask);
@@ -80,20 +123,18 @@ static void SiSSubsequentScreenToScreenColorExpand(ScrnInfoPtr pScrn,
                                 int x, int y, int w, int h,
                                 int srcx, int srcy, int skipleft);
 #endif
-static void SiSSetupForScanlineCPUToScreenColorExpandFill(ScrnInfoPtr pScrn, 
-                                int fg, int bg, int rop, 
+static void SiSSetupForScanlineCPUToScreenColorExpandFill(ScrnInfoPtr pScrn,
+                                int fg, int bg, int rop,
                                 unsigned int planemask);
 static void SiSSubsequentScanlineCPUToScreenColorExpandFill(ScrnInfoPtr pScrn,
-                                int x, int y, int w, int h, 
+                                int x, int y, int w, int h,
                                 int skipleft);
 static void SiSSubsequentColorExpandScanline(ScrnInfoPtr pScrn, int bufno);
-#if 0
-static void SiSSetupForImageWrite(ScrnInfoPtr pScrn, int rop, 
-                                unsigned int planemask, int trans_color, 
-                                int bpp, int depth);
-static void SiSSubsequentImageWriteRect(ScrnInfoPtr pScrn, int x, int y, 
-                                int w, int h, int skipleft);
+
+#ifdef SISDUALHEAD
+static void SiSRestoreAccelState(ScrnInfoPtr pScrn);
 #endif
+
 static void
 SiSInitializeAccelerator(ScrnInfoPtr pScrn)
 {
@@ -116,71 +157,74 @@ SiS300AccelInit(ScreenPtr pScreen)
 	int             i;
 
 	pSiS->AccelInfoPtr = infoPtr = XAACreateInfoRec();
-	if (!infoPtr)  
-		return FALSE;
+	if (!infoPtr)  return FALSE;
 
 	SiSInitializeAccelerator(pScrn);
 
 	infoPtr->Flags = LINEAR_FRAMEBUFFER |
-			   		 OFFSCREEN_PIXMAPS |
-					 PIXMAP_CACHE;
+			 OFFSCREEN_PIXMAPS |
+			 PIXMAP_CACHE;
 
 	/* sync */
 	infoPtr->Sync = SiSSync;
 
-	if ((pScrn->bitsPerPixel != 8) && (pScrn->bitsPerPixel != 16) &&
-		(pScrn->bitsPerPixel != 32))         return FALSE;
+	/* Acceleration only supported at 8, 16 and 32 bpp */
+	if((pScrn->bitsPerPixel != 8) && (pScrn->bitsPerPixel != 16) &&
+	         (pScrn->bitsPerPixel != 32))
+	    return FALSE;
 
-	/* BitBlt */
+	/* screen to screen copy - TW: We now support transparent copies */
 	infoPtr->SetupForScreenToScreenCopy = SiSSetupForScreenToScreenCopy;
 	infoPtr->SubsequentScreenToScreenCopy = SiSSubsequentScreenToScreenCopy;
-	infoPtr->ScreenToScreenCopyFlags = NO_PLANEMASK | NO_TRANSPARENCY;
+	infoPtr->ScreenToScreenCopyFlags = NO_PLANEMASK |
+	                                   TRANSPARENCY_GXCOPY_ONLY;
 
 	/* solid fills */
 	infoPtr->SetupForSolidFill = SiSSetupForSolidFill;
 	infoPtr->SubsequentSolidFillRect = SiSSubsequentSolidFillRect;
+#ifdef TRAP
+	infoPtr->SubsequentSolidFillTrap = SiSSubsequentSolidFillTrap;
+#endif
 	infoPtr->SolidFillFlags = NO_PLANEMASK;
 
 	/* solid line */
 	infoPtr->SetupForSolidLine = SiSSetupForSolidLine;
 	infoPtr->SubsequentSolidTwoPointLine = SiSSubsequentSolidTwoPointLine;
 	infoPtr->SubsequentSolidHorVertLine = SiSSubsequentSolidHorzVertLine;
-	infoPtr->SolidFillFlags = NO_PLANEMASK;
+	infoPtr->SolidLineFlags = NO_PLANEMASK;
 
 	/* dashed line */
 	infoPtr->SetupForDashedLine = SiSSetupForDashedLine;
 	infoPtr->SubsequentDashedTwoPointLine = SiSSubsequentDashedTwoPointLine;
 	infoPtr->DashPatternMaxLength = 64;
-	infoPtr->DashedLineFlags = NO_PLANEMASK | 
-					LINE_PATTERN_MSBFIRST_LSBJUSTIFIED;
+	infoPtr->DashedLineFlags = NO_PLANEMASK |
+				   LINE_PATTERN_MSBFIRST_LSBJUSTIFIED;
 
 	/* 8x8 mono pattern fill */
 	infoPtr->SetupForMono8x8PatternFill = SiSSetupForMonoPatternFill;
-	infoPtr->SubsequentMono8x8PatternFillRect =
-							SiSSubsequentMonoPatternFill;
+	infoPtr->SubsequentMono8x8PatternFillRect = SiSSubsequentMonoPatternFill;
+#ifdef TRAP
+	infoPtr->SubsequentMono8x8PatternFillTrap = SiSSubsequentMonoPatternFillTrap;
+#endif
 	infoPtr->Mono8x8PatternFillFlags = NO_PLANEMASK |
-							HARDWARE_PATTERN_SCREEN_ORIGIN |
-							HARDWARE_PATTERN_PROGRAMMED_BITS |
-							NO_TRANSPARENCY |
-							BIT_ORDER_IN_BYTE_MSBFIRST ;
-
-#if 0
-	/* 8x8 color pattern fill ---seems not useful by xaa */
-	infoPtr->SetupForColor8x8PatternFill =
-	    SiSSetupForColorPatternFill;
-	infoPtr->SubsequentColor8x8PatternFillRect =
-	    SiSSubsequentColorPatternFill;
-	infoPtr->Color8x8PatternFillFlags = NO_PLANEMASK |
-	    HARDWARE_PATTERN_SCREEN_ORIGIN |
-	    HARDWARE_PATTERN_PROGRAMMED_BITS ;
+					   HARDWARE_PATTERN_SCREEN_ORIGIN |
+					   HARDWARE_PATTERN_PROGRAMMED_BITS |
+					   NO_TRANSPARENCY |
+					   BIT_ORDER_IN_BYTE_MSBFIRST ;
 
+#ifdef STSCE
 	/* Screen To Screen Color Expand */
+	/* TW: The hardware does support this the way we need it */
 	infoPtr->SetupForScreenToScreenColorExpandFill =
-	    SiSSetupForScreenToScreenColorExpand;
+	    			SiSSetupForScreenToScreenColorExpand;
 	infoPtr->SubsequentScreenToScreenColorExpandFill =
-	    SiSSubsequentScreenToScreenColorExpand;
-	
-	/* CPU To Screen Color Expand ---implement another instead of this one! */
+	    			SiSSubsequentScreenToScreenColorExpand;
+	infoPtr->ScreenToScreenColorExpandFillFlags = NO_PLANEMASK |
+	                                              BIT_ORDER_IN_BYTE_MSBFIRST ;
+#endif
+
+#if 0
+	/* CPU To Screen Color Expand --- implement another instead of this one! */
 	infoPtr->SetupForCPUToScreenColorExpandFill =
 	    SiSSetupForCPUToScreenColorExpand;
 	infoPtr->SubsequentCPUToScreenColorExpandFill =
@@ -195,62 +239,55 @@ SiS300AccelInit(ScreenPtr pScreen)
 	    HARDWARE_PATTERN_PROGRAMMED_BITS ;
 #endif
 
-	/* per-scanline color expansion*/
-	pSiS->ColorExpandBufferNumber = 16;     
-	pSiS->ColorExpandBufferCountMask = 0x0F;
-	pSiS->PerColorExpandBufferSize = ((pScrn->virtualX + 31)/32) * 4;       
+	/* per-scanline color expansion (using indirect method) */
+	if(pSiS->VGAEngine == SIS_530_VGA) {
+	   pSiS->ColorExpandBufferNumber = 4;
+	   pSiS->ColorExpandBufferCountMask = 0x03;
+	} else {
+	   pSiS->ColorExpandBufferNumber = 16;
+	   pSiS->ColorExpandBufferCountMask = 0x0F;
+	}
+	pSiS->PerColorExpandBufferSize = ((pScrn->virtualX + 31)/32) * 4;
 	infoPtr->NumScanlineColorExpandBuffers = pSiS->ColorExpandBufferNumber;
 	infoPtr->ScanlineColorExpandBuffers = (unsigned char **)&pSiS->ColorExpandBufferAddr[0];
-	infoPtr->SetupForScanlineCPUToScreenColorExpandFill = SiSSetupForScanlineCPUToScreenColorExpandFill;
-	infoPtr->SubsequentScanlineCPUToScreenColorExpandFill = SiSSubsequentScanlineCPUToScreenColorExpandFill;
-	infoPtr->SubsequentColorExpandScanline = SiSSubsequentColorExpandScanline;
+
+	infoPtr->SetupForScanlineCPUToScreenColorExpandFill =
+	                            SiSSetupForScanlineCPUToScreenColorExpandFill;
+	infoPtr->SubsequentScanlineCPUToScreenColorExpandFill =
+	                            SiSSubsequentScanlineCPUToScreenColorExpandFill;
+	infoPtr->SubsequentColorExpandScanline =
+	                            SiSSubsequentColorExpandScanline;
 	infoPtr->ScanlineCPUToScreenColorExpandFillFlags =
 	    NO_PLANEMASK |
 	    CPU_TRANSFER_PAD_DWORD |
 	    SCANLINE_PAD_DWORD |
 	    BIT_ORDER_IN_BYTE_MSBFIRST |
 	    LEFT_EDGE_CLIPPING;
-	
-#if 0
-	divider = ((pScrn->virtualX*pScrn->bitsPerPixel)/8)+8;
-	pSiS->ImageWriteBufferSize = (((12*1024)+divider-1)/divider)*divider;
-	infoPtr->SetupForImageWrite = SiSSetupForImageWrite;
-	infoPtr->SubsequentImageWriteRect = SiSSubsequentImageWriteRect;
-	infoPtr->ImageWriteFlags = CPU_TRANSFER_PAD_DWORD |
-	    SCANLINE_PAD_DWORD |
-	    LEFT_EDGE_CLIPPING |    
-	    NO_PLANEMASK|
-	    NO_TRANSPARENCY |
-	    NO_GXCOPY |
-	    SYNC_AFTER_IMAGE_WRITE;
+
+#ifdef SISDUALHEAD
+	if (pSiS->DualHeadMode) {
+		infoPtr->RestoreAccelState = SiSRestoreAccelState;
+	}
 #endif
 
 	/* init Frame Buffer Manager */
-
-        /* TW: check if maxxfbmem must contain TurboQueue and HWCursor */
 	topFB = pSiS->maxxfbmem;
-	if ((topFB)
-	    >= (pSiS->FbMapSize)
-	    - ((pSiS->TurboQueue) ? (1024*512) : 0)
-	       - ((pSiS->HWCursor) ? 4096 : 0))  {
-	    topFB = pSiS->FbMapSize;
-	    /* TurboQueue len is always 512k */
-	    if (pSiS->TurboQueue) topFB -= 1024*512;  
-            /* HWCursor len is always 4096 */
-	    if (pSiS->HWCursor)  topFB -= 4096;
-	}
-	reservedFbSize = (pSiS->ColorExpandBufferNumber
-			   * pSiS->PerColorExpandBufferSize);
-        /* TW: New for MaxXFBmem Option */
-	UsableFbSize = topFB - reservedFbSize; 
-	/* Layout:
-	 * |--------------++++++++++++++++++++^******==========~~~~~~~~~~~~|
-	 *   UsableFbSize  ColorExpandBuffers | Heap  HWCursor  TurboQueue
-	 *                                 topFB
+
+	reservedFbSize = pSiS->ColorExpandBufferNumber * pSiS->PerColorExpandBufferSize;
+
+	UsableFbSize = topFB - reservedFbSize;
+
+	/* Layout: (Sizes do not reflect correct proportions)
+	 * |--------------++++++++++++++++++++^************==========~~~~~~~~~~~~|
+	 *   UsableFbSize  ColorExpandBuffers | DRI-Heap  |  HWCursor  TurboQueue   300/310/325 series
+	 * |--------------++++++++++++++++++++|  ====================~~~~~~~~~~~~|
+	 *   UsableFbSize  ColorExpandBuffers |        TurboQueue     HWCursor      530/620
+	 *                                  topFB
 	 */
+
 	AvailBufBase = pSiS->FbBase + UsableFbSize;
 	for (i = 0; i < pSiS->ColorExpandBufferNumber; i++) {
-		pSiS->ColorExpandBufferAddr[i] = AvailBufBase + 
+		pSiS->ColorExpandBufferAddr[i] = AvailBufBase +
 		    i * pSiS->PerColorExpandBufferSize;
 		pSiS->ColorExpandBufferScreenOffset[i] = UsableFbSize +
 		    i * pSiS->PerColorExpandBufferSize;
@@ -258,17 +295,30 @@ SiS300AccelInit(ScreenPtr pScreen)
 	Avail.x1 = 0;
 	Avail.y1 = 0;
 	Avail.x2 = pScrn->displayWidth;
-	Avail.y2 = UsableFbSize
-	    / (pScrn->displayWidth * pScrn->bitsPerPixel/8) - 1;
-	if (Avail.y2 < 0)
-		Avail.y2 = 32767;
-	
+	Avail.y2 = (UsableFbSize / (pScrn->displayWidth * pScrn->bitsPerPixel/8)) - 1;
+
+	if(Avail.y2 < 0)  Avail.y2 = 32767;
+
+	if(Avail.y2 < pScrn->currentMode->VDisplay) {
+		xf86DrvMsg(pScrn->scrnIndex, X_ERROR,
+			"Not enough video RAM for accelerator. At least "
+			"%dKB needed, %dKB available\n",
+			((((pScrn->displayWidth * pScrn->bitsPerPixel/8)   /* TW: +8 for make it sure */
+			     * pScrn->currentMode->VDisplay) + reservedFbSize) / 1024) + 8,
+			pSiS->maxxfbmem/1024);
+		pSiS->NoAccel = TRUE;
+		pSiS->NoXvideo = TRUE;
+		XAADestroyInfoRec(pSiS->AccelInfoPtr);
+		pSiS->AccelInfoPtr = NULL;
+		return FALSE;
+	}
+
 	xf86DrvMsg(pScrn->scrnIndex, X_INFO,
 		   "Frame Buffer From (%d,%d) To (%d,%d)\n",
 		   Avail.x1, Avail.y1, Avail.x2, Avail.y2);
-	
+
 	xf86InitFBManager(pScreen, &Avail);
-	
+
 	return(XAAInit(pScreen, infoPtr));
 }
 
@@ -284,7 +334,26 @@ SiSSync(ScrnInfoPtr pScrn)
 	SiSIdle
 }
 
-static int sisALUConv[] =
+#ifdef SISDUALHEAD
+static void
+SiSRestoreAccelState(ScrnInfoPtr pScrn)
+{
+	SISPtr pSiS = SISPTR(pScrn);
+
+	/* TW: We don't need to do anything special here; forcing the
+	 *     other head to re-read the CmdQueLen is not necessary:
+	 *     After the Sync in RestoreAccelState(), the real queue
+	 *     length is always larger than (or at least equal to)
+	 *     the amount stored in CmdQueueLen of the other head,
+	 *     so the only thing that might happen is one unnecessary
+	 *     Sync on the other head. I think we can live with that.
+	 */
+	pSiS->DoColorExpand = FALSE;
+	SiSIdle
+}
+#endif
+
+static const int sisALUConv[] =
 {
     0x00,       /* dest = 0;            0,      GXclear,        0 */
     0x88,       /* dest &= src;         DSa,    GXand,          0x1 */
@@ -304,7 +373,7 @@ static int sisALUConv[] =
     0xFF,       /* dest = 0xFF;         1,      GXset,          0xF */
 };
 /* same ROP but with Pattern as Source */
-static int sisPatALUConv[] =
+static const int sisPatALUConv[] =
 {
     0x00,       /* dest = 0;            0,      GXclear,        0 */
     0xA0,       /* dest &= src;         DPa,    GXand,          0x1 */
@@ -329,58 +398,25 @@ static void SiSSetupForScreenToScreenCopy(ScrnInfoPtr pScrn,
                                 unsigned int planemask, int trans_color)
 {
 	SISPtr  pSiS = SISPTR(pScrn);
-/*      XAAInfoRecPtr   pXAA = XAAPTR(pScrn);*/
 
 	PDEBUG(ErrorF("Setup ScreenCopy(%d, %d, 0x%x, 0x%x, 0x%x)\n",
 					xdir, ydir, rop, planemask, trans_color));
-/*
-	ErrorF("XAAInfoPtr->UsingPixmapCache = %s\n"
-			"XAAInfoPtr->CanDoMono8x8 = %s\n"
-			"XAAInfoPtr->CanDoColor8x8 = %s\n"
-			"XAAInfoPtr->CachePixelGranularity = %d\n"
-			"XAAInfoPtr->MaxCacheableTileWidth = %d\n"
-			"XAAInfoPtr->MaxCacheableTileHeight = %d\n"
-			"XAAInfoPtr->MaxCacheableStippleWidth = %d\n"
-			"XAAInfoPtr->MaxCacheableStippleHeight = %d\n"
-			"XAAInfoPtr->MonoPatternPitch = %d\n"
-			"XAAInfoPtr->CacheWidthMono8x8Pattern = %d\n"
-			"XAAInfoPtr->CacheHeightMono8x8Pattern = %d\n"
-			"XAAInfoPtr->ColorPatternPitch = %d\n"
-			"XAAInfoPtr->CacheWidthColor8x8Pattern = %d\n"
-			"XAAInfoPtr->CacheHeightColor8x8Pattern = %d\n"
-			"XAAInfoPtr->CacheColorExpandDensity = %d\n"
-			"XAAInfoPtr->maxOffPixWidth = %d\n"
-			"XAAInfoPtr->maxOffPixHeight= %d\n"
-			"XAAInfoPtr->NeedToSync = %s\n"
-			"\n",
-					pXAA->UsingPixmapCache ? "True" : "False",
-					pXAA->CanDoMono8x8 ? "True" : "False",
-					pXAA->CanDoColor8x8 ? "True" : "False",
-					pXAA->CachePixelGranularity,
-					pXAA->MaxCacheableTileWidth,
-					pXAA->MaxCacheableTileHeight,
-					pXAA->MaxCacheableStippleWidth,
-					pXAA->MaxCacheableStippleHeight,
-					pXAA->MonoPatternPitch,
-					pXAA->CacheWidthMono8x8Pattern,
-					pXAA->CacheHeightMono8x8Pattern,
-					pXAA->ColorPatternPitch,
-					pXAA->CacheWidthColor8x8Pattern,
-					pXAA->CacheHeightColor8x8Pattern,
-					pXAA->CacheColorExpandDensity,
-					pXAA->maxOffPixWidth,
-					pXAA->maxOffPixHeight,
-					pXAA->NeedToSync ? "True" : "False");
-*/
-					
-	SiSSetupDSTColorDepth(SISPTR(pScrn)->DstColor);
+
+	SiSSetupDSTColorDepth(pSiS->DstColor);
 	SiSSetupSRCPitch(pSiS->scrnOffset)
 	SiSSetupDSTRect(pSiS->scrnOffset, -1)
-	SiSSetupROP(sisALUConv[rop])
-	if (xdir > 0) {
+
+	if(trans_color != -1) {
+		SiSSetupROP(0x0A)
+		SiSSetupSRCTrans(trans_color)
+		SiSSetupCMDFlag(TRANSPARENT_BITBLT)
+	} else {
+	        SiSSetupROP(sisALUConv[rop])
+	}
+	if(xdir > 0) {
 		SiSSetupCMDFlag(X_INC)
 	}
-	if (ydir > 0) {
+	if(ydir > 0) {
 		SiSSetupCMDFlag(Y_INC)
 	}
 }
@@ -396,27 +432,35 @@ static void SiSSubsequentScreenToScreenCopy(ScrnInfoPtr pScrn,
 					src_x, src_y, dst_x, dst_y, width, height));
 
 	srcbase = dstbase = 0;
-	if (src_y >= 2048) { 
+	if(src_y >= 2048) {
 		srcbase = pSiS->scrnOffset * src_y;
 		src_y = 0;
-	}       
-	if (dst_y >= pScrn->virtualY) {
-		dstbase = pSiS->scrnOffset*dst_y;
+	}
+	if( (dst_y >= pScrn->virtualY) || (dst_y >= 2048) ) {
+		dstbase = pSiS->scrnOffset * dst_y;
 		dst_y = 0;
-	}       
+	}
+#ifdef SISDUALHEAD
+	if(pSiS->VGAEngine != SIS_530_VGA) {
+	   	srcbase += HEADOFFSET;
+	   	dstbase += HEADOFFSET;
+	}
+#endif
 	SiSSetupSRCBase(srcbase);
 	SiSSetupDSTBase(dstbase);
-	if (!(pSiS->CommandReg & X_INC))  {
+
+	if(!(pSiS->CommandReg & X_INC))  {
 		src_x += width-1;
 		dst_x += width-1;
 	}
-	if (!(pSiS->CommandReg & Y_INC))  {
+	if(!(pSiS->CommandReg & Y_INC))  {
 		src_y += height-1;
 		dst_y += height-1;
 	}
 	SiSSetupRect(width, height)
 	SiSSetupSRCXY(src_x, src_y)
 	SiSSetupDSTXY(dst_x, dst_y)
+
 	SiSDoCMD
 }
 
@@ -430,11 +474,10 @@ SiSSetupForSolidFill(ScrnInfoPtr pScrn,
 					color, rop, planemask));
 
 	SiSSetupPATFG(color)
-/*      SiSSetupDSTRect(pSiS->scrnOffset, pScrn->virtualY)*/
 	SiSSetupDSTRect(pSiS->scrnOffset, -1)
-	SiSSetupDSTColorDepth(SISPTR(pScrn)->DstColor);
+	SiSSetupDSTColorDepth(pSiS->DstColor);
 	SiSSetupROP(sisPatALUConv[rop])
-	SiSSetupCMDFlag(X_INC | Y_INC | PATFG | BITBLT)
+	/* SiSSetupCMDFlag(PATFG) - is zero */
 }
 
 static void
@@ -447,15 +490,136 @@ SiSSubsequentSolidFillRect(ScrnInfoPtr pScrn,
 	PDEBUG(ErrorF("Subsequent SolidFillRect(%d, %d, %d, %d)\n",
 					x, y, w, h));
 	dstbase = 0;
-	if (y >= 2048) { 
-		dstbase=pSiS->scrnOffset*y;
+
+	if (y >= 2048) {
+		dstbase = pSiS->scrnOffset * y;
 		y = 0;
 	}
+#ifdef SISDUALHEAD
+        if(pSiS->VGAEngine != SIS_530_VGA) {
+	   	dstbase += HEADOFFSET;
+        }
+#endif
 	SiSSetupDSTBase(dstbase)
 	SiSSetupDSTXY(x,y)
 	SiSSetupRect(w,h)
+	/* Clear commandReg because Setup can be used for Rect and Trap */
+	pSiS->CommandReg &= ~(T_XISMAJORL | T_XISMAJORR |
+	                      T_L_X_INC | T_L_Y_INC |
+	                      T_R_X_INC | T_R_Y_INC |
+			      TRAPAZOID_FILL);
+	SiSSetupCMDFlag(X_INC | Y_INC | BITBLT)
+
+	SiSDoCMD
+}
+
+/* TW: Trapezoid */
+/* This would work better if XAA would provide us with valid trapezoids.
+ * In fact, with small trapezoids the left and the right edge often cross
+ * each other or result in a line length of 0 which causes drawing errors
+ * (filling over whole scanline).
+ * Furthermore, I have not found out how to draw trapezoids with a height
+ * greater than 127.
+ */
+#ifdef TRAP
+static void
+SiSSubsequentSolidFillTrap(ScrnInfoPtr pScrn, int y, int h,
+	       int left,  int dxL, int dyL, int eL,
+	       int right, int dxR, int dyR, int eR )
+{
+	SISPtr  pSiS = SISPTR(pScrn);
+	long dstbase;
+#if 0
+	float kL, kR;
+#endif
+
+	dstbase = 0;
+	if (y >= 2048) {
+		dstbase=pSiS->scrnOffset*y;
+		y = 0;
+	}
+#ifdef SISDUALHEAD
+	if(pSiS->VGAEngine != SIS_530_VGA) {
+	   	dstbase += HEADOFFSET;
+	}
+#endif
+	SiSSetupDSTBase(dstbase)
+	/* SiSSetupRect(w,h) */
+
+#if 1
+	SiSSetupPATFG(0xff0000) /* FOR TESTING */
+#endif
+
+	/* Clear CommandReg because SetUp can be used for Rect and Trap */
+	pSiS->CommandReg &= ~(T_L_X_INC | T_L_Y_INC |
+	                      T_R_X_INC | T_R_Y_INC |
+	                      T_XISMAJORL | T_XISMAJORR |
+			      BITBLT);
+
+        xf86DrvMsg(0, X_INFO, "Trap (%d %d %d %d) dxL %d dyL %d eL %d   dxR %d dyR %d eR %d\n",
+		left, right, y, h, dxL, dyL, eL, dxR, dyR, eR);
+
+	/* Unfortunately, we must check if the right and the left edge
+	 * cross each other...  INCOMPLETE (line equation wrong)
+	 */
+#if 0
+	if (dxL == 0) kL = 0;
+	else kL = (float)dyL / (float)dxL;
+	if (dxR == 0) kR = 0;
+	else kR = (float)dyR / (float)dxR;
+	xf86DrvMsg(0, X_INFO, "kL %f kR %f!\n", kL, kR);
+	if ( (kR != kL) &&
+	     (!(kR == 0 && kL == 0)) &&
+	     (!(kR <  0 && kL >  0)) ) {
+	   xf86DrvMsg(0, X_INFO, "Inside if (%f - %d)\n", ( kL * ( ( ((float)right - (float)left) / (kL - kR) ) - left) + y), h+y);
+           if ( ( ( kL * ( ( ((float)right - (float)left) / (kL - kR) ) - (float)left) + (float)y) < (h + y) ) ) {
+	     xf86DrvMsg(0, X_INFO, "Cross detected!\n");
+	   }
+	}
+#endif
+
+	/* Determine egde angles */
+	if (dxL < 0) { dxL = -dxL; }
+	else { SiSSetupCMDFlag(T_L_X_INC) }
+	if (dxR < 0) { dxR = -dxR; }
+	else { SiSSetupCMDFlag(T_R_X_INC) }
+
+	/* (Y direction always positive - do this anyway) */
+	if (dyL < 0) { dyL = -dyL; }
+	else { SiSSetupCMDFlag(T_L_Y_INC) }
+	if (dyR < 0) { dyR = -dyR; }
+	else { SiSSetupCMDFlag(T_R_Y_INC) }
+
+	/* Determine major axis */
+	if (dxL >= dyL) {      /* X is major axis */
+		SiSSetupCMDFlag(T_XISMAJORL)
+	}
+	if (dxR >= dyR) {      /* X is major axis */
+		SiSSetupCMDFlag(T_XISMAJORR)
+	}
+
+	/* Set up deltas */
+	SiSSetupdL(dxL, dyL)
+	SiSSetupdR(dxR, dyR)
+
+#if 0   /* Could it be that this crappy engine can only draw trapezoids up to 127 pixels high? */
+	h &= 0x7F;
+	if (h == 0) h = 10;
+#endif
+
+	/* Set up y, h, left, right */
+	SiSSetupYH(y,h)
+	SiSSetupLR(left,right)
+
+	/* Set up initial error term */
+	SiSSetupEL(eL)
+	SiSSetupER(eR)
+
+	SiSSetupCMDFlag(TRAPAZOID_FILL);
+
 	SiSDoCMD
 }
+#endif
 
 static void
 SiSSetupForSolidLine(ScrnInfoPtr pScrn,
@@ -468,9 +632,8 @@ SiSSetupForSolidLine(ScrnInfoPtr pScrn,
 
 	SiSSetupLineCount(1)
 	SiSSetupPATFG(color)
-/*      SiSSetupDSTRect(pSiS->scrnOffset, pScrn->virtualY)*/
 	SiSSetupDSTRect(pSiS->scrnOffset, -1)
-	SiSSetupDSTColorDepth(SISPTR(pScrn)->DstColor);
+	SiSSetupDSTColorDepth(pSiS->DstColor);
 	SiSSetupROP(sisPatALUConv[rop])
 	SiSSetupCMDFlag(PATFG | LINE)
 }
@@ -488,17 +651,26 @@ SiSSubsequentSolidTwoPointLine(ScrnInfoPtr pScrn,
 	dstbase = 0;
 	miny = (y1 > y2) ? y2 : y1;
 	maxy = (y1 > y2) ? y1 : y2;
-	if (maxy >= 2048) {
-		dstbase = pSiS->scrnOffset*miny;
+	if(maxy >= 2048) {
+		dstbase = pSiS->scrnOffset * miny;
 		y1 -= miny;
 		y2 -= miny;
 	}
+#ifdef SISDUALHEAD
+	if(pSiS->VGAEngine != SIS_530_VGA) {
+	   dstbase += HEADOFFSET;
+	}
+#endif
 	SiSSetupDSTBase(dstbase)
 
 	SiSSetupX0Y0(x1,y1)
 	SiSSetupX1Y1(x2,y2)
-	if (flags & OMIT_LAST)
+	if (flags & OMIT_LAST) {
 		SiSSetupCMDFlag(NO_LAST_PIXEL)
+	} else {
+		pSiS->CommandReg &= ~(NO_LAST_PIXEL);
+	}
+
 	SiSDoCMD
 }
 
@@ -512,20 +684,26 @@ SiSSubsequentSolidHorzVertLine(ScrnInfoPtr pScrn,
 	PDEBUG(ErrorF("Subsequent SolidHorzVertLine(%d, %d, %d, %d)\n",
 					x, y, len, dir));
 	len--; /* starting point is included! */
+
 	dstbase = 0;
-	if ((y >= 2048) || ((y + len) >= 2048)) { 
+	if((y >= 2048) || ((y + len) >= 2048)) {
 		dstbase = pSiS->scrnOffset * y;
 		y = 0;
 	}
+#ifdef SISDUALHEAD
+	if(pSiS->VGAEngine != SIS_530_VGA) {
+	   	dstbase += HEADOFFSET;
+	}
+#endif
 	SiSSetupDSTBase(dstbase)
 
 	SiSSetupX0Y0(x,y)
 	if (dir == DEGREES_0) {
 		SiSSetupX1Y1(x + len, y);
-	}
-	else {
+	} else {
 		SiSSetupX1Y1(x, y + len);
 	}
+
 	SiSDoCMD
 }
 
@@ -540,15 +718,19 @@ SiSSetupForDashedLine(ScrnInfoPtr pScrn,
 			fg, bg, rop, planemask, length, *(pattern+4), *pattern));
 
 	SiSSetupLineCount(1)
-/*      SiSSetupDSTRect(pSiS->scrnOffset, pScrn->virtualY)*/
 	SiSSetupDSTRect(pSiS->scrnOffset, -1)
-	SiSSetupDSTColorDepth(SISPTR(pScrn)->DstColor);
+	SiSSetupDSTColorDepth(pSiS->DstColor);
 	SiSSetupStyleLow(*pattern)
 	SiSSetupStyleHigh(*(pattern+4))
+	SiSSetupStylePeriod(length-1);			/* TW: This was missing!!! */
 	SiSSetupROP(sisPatALUConv[rop])
 	SiSSetupPATFG(fg)
-	if (bg != -1) 
+	SiSSetupCMDFlag(LINE | LINE_STYLE)  		/* TW: This was missing!!! */
+	if(bg != -1) {
 		SiSSetupPATBG(bg)
+	} else {
+	        SiSSetupCMDFlag(TRANSPARENT); 		/* TW: This was missing!!! */
+	}
 }
 
 static void
@@ -563,20 +745,28 @@ SiSSubsequentDashedTwoPointLine(ScrnInfoPtr pScrn,
 			x1, y1, x2, y2, flags, phase));
 
 	dstbase = 0;
-	miny=(y1 > y2) ? y2 : y1;
-	maxy=(y1 > y2) ? y1 : y2;
-	if (maxy >= 2048) {
+	miny = (y1 > y2) ? y2 : y1;
+	maxy = (y1 > y2) ? y1 : y2;
+	if(maxy >= 2048) {
 		dstbase = pSiS->scrnOffset * miny;
 		y1 -= miny;
 		y2 -= miny;
 	}
+#ifdef SISDUALHEAD
+	if(pSiS->VGAEngine != SIS_530_VGA) {
+	   dstbase += HEADOFFSET;
+	}
+#endif
 	SiSSetupDSTBase(dstbase)
 
 	SiSSetupX0Y0(x1,y1)
 	SiSSetupX1Y1(x2,y2)
-	if (flags & OMIT_LAST) {
+	if(flags & OMIT_LAST) {
 		SiSSetupCMDFlag(NO_LAST_PIXEL)
+	} else {
+		pSiS->CommandReg &= ~(NO_LAST_PIXEL);
 	}
+
 	SiSDoCMD
 }
 
@@ -589,13 +779,12 @@ SiSSetupForMonoPatternFill(ScrnInfoPtr pScrn,
 
 	PDEBUG(ErrorF("Setup MonoPatFill(0x%x,0x%x, 0x%x,0x%x, 0x%x, 0x%x)\n",
 					patx, paty, fg, bg, rop, planemask));
-
 	SiSSetupDSTRect(pSiS->scrnOffset, -1)
-	SiSSetupDSTColorDepth(SISPTR(pScrn)->DstColor);
+	SiSSetupDSTColorDepth(pSiS->DstColor);
 	SiSSetupMONOPAT(patx,paty)
 	SiSSetupPATFG(fg)
 	SiSSetupROP(sisPatALUConv[rop])
-	SiSSetupCMDFlag(PATMONO | X_INC | Y_INC)
+	SiSSetupCMDFlag(PATMONO)
 	SiSSetupPATBG(bg)
 }
 
@@ -610,18 +799,103 @@ SiSSubsequentMonoPatternFill(ScrnInfoPtr pScrn,
 	PDEBUG(ErrorF("Subsequent MonoPatFill(0x%x,0x%x, %d,%d, %d,%d)\n",
 							patx, paty, x, y, w, h));
 	dstbase = 0;
-	if (y >= 2048) { 
+
+	if (y >= 2048) {
 		dstbase = pSiS->scrnOffset * y;
 		y = 0;
 	}
+#ifdef SISDUALHEAD
+	if(pSiS->VGAEngine != SIS_530_VGA) {
+	   dstbase += HEADOFFSET;
+	}
+#endif
+	SiSSetupDSTBase(dstbase)
+	SiSSetupDSTXY(x, y)
+	SiSSetupRect(w, h)
+	/* Clear commandReg because Setup can be used for Rect and Trap */
+	pSiS->CommandReg &= ~(T_XISMAJORL | T_XISMAJORR |
+	                      T_L_X_INC | T_L_Y_INC |
+	                      T_R_X_INC | T_R_Y_INC |
+	                      TRAPAZOID_FILL);
+	SiSSetupCMDFlag(X_INC | Y_INC)
+
+	SiSDoCMD
+}
 
+/* Trapezoid */
+#ifdef TRAP
+static void
+SiSSubsequentMonoPatternFillTrap(ScrnInfoPtr pScrn,
+               int patx, int paty,
+               int y, int h,
+	       int left, int dxL, int dyL, int eL,
+	       int right, int dxR, int dyR, int eR )
+{
+	SISPtr  pSiS = SISPTR(pScrn);
+	long dstbase;
+
+	PDEBUG(ErrorF("Subsequent Mono8x8PatternFillTrap(%d, %d, %d - %d %d/%d %d/%d)\n",
+					y, h, left, right, dxL, dxR, eL, eR));
+
+	dstbase = 0;
+	if (y >= 2048) {
+		dstbase=pSiS->scrnOffset*y;
+		y = 0;
+	}
+#ifdef SISDUALHEAD
+	if(pSiS->VGAEngine != SIS_530_VGA) {
+	   dstbase += HEADOFFSET;
+	}
+#endif
 	SiSSetupDSTBase(dstbase)
-	SiSSetupDSTXY(x,y)
-	SiSSetupRect(w,h)
+
+	/* Clear CommandReg because SetUp can be used for Rect and Trap */
+	pSiS->CommandReg &= ~(T_XISMAJORL | T_XISMAJORR |
+	                      T_L_X_INC | T_L_Y_INC |
+			      T_R_X_INC | T_R_Y_INC |
+			      BITBLT);
+
+	if (dxL < 0) { dxL = -dxL;  }
+	else { SiSSetupCMDFlag(T_L_X_INC) }
+	if (dxR < 0) { dxR = -dxR; }
+	else { SiSSetupCMDFlag(T_R_X_INC) }
+
+	if (dyL < 0) { dyL = -dyL; }
+	else { SiSSetupCMDFlag(T_L_Y_INC) }
+	if (dyR < 0) { dyR = -dyR; }
+	else { SiSSetupCMDFlag(T_R_Y_INC) }
+
+	/* Determine major axis */
+	if (dxL >= dyL) {      /* X is major axis */
+		SiSSetupCMDFlag(T_XISMAJORL)
+	}
+	if (dxR >= dyR) {      /* X is major axis */
+		SiSSetupCMDFlag(T_XISMAJORR)
+	}
+
+	SiSSetupYH(y,h)
+	SiSSetupLR(left,right)
+
+	SiSSetupdL(dxL, dyL)
+	SiSSetupdR(dxR, dyR)
+
+	SiSSetupEL(eL)
+	SiSSetupER(eR)
+
+	SiSSetupCMDFlag(TRAPAZOID_FILL);
+
 	SiSDoCMD
 }
+#endif
+
 
 #if 0
+
+/* TW: The following (already commented) functions have NOT been adapted for dual-head mode */
+
+
+/* ------- Color Pattern Fill --- is not useful for XAA -------------- */
+
 static void
 SiSSetupForColorPatternFill(ScrnInfoPtr pScrn,
                                 int patx, int paty, int rop,
@@ -635,7 +909,7 @@ SiSSetupForColorPatternFill(ScrnInfoPtr pScrn,
 
 /*      SiSSetupDSTRect(pSiS->scrnOffset, pScrn->virtualY)*/
 	SiSSetupDSTRect(pSiS->scrnOffset, -1)
-	SiSSetupDSTColorDepth(SISPTR(pScrn)->DstColor);
+	SiSSetupDSTColorDepth(pSiS->DstColor);
 	SiSSetupROP(sisPatALUConv[rop])
 	SiSSetupCMDFlag(PATPATREG | X_INC | Y_INC)
 }
@@ -662,6 +936,11 @@ SiSSubsequentColorPatternFill(ScrnInfoPtr pScrn,
 	SiSDoCMD
 }
 
+/* ----- CPU To Screen Color Expand (single task) ------------------------- */
+
+/* This does not work. Assumingly for the same
+ * reason why STSColorExpand does not work either.
+ */
 static void
 SiSSetupForCPUToScreenColorExpand(ScrnInfoPtr pScrn,
                                 int fg, int bg,
@@ -674,7 +953,7 @@ SiSSetupForCPUToScreenColorExpand(ScrnInfoPtr pScrn,
 
 /*      SiSSetupDSTRect(pSiS->scrnOffset, pScrn->virtualY)*/
 	SiSSetupDSTRect(pSiS->scrnOffset, -1)
-	SiSSetupDSTColorDepth(SISPTR(pScrn)->DstColor);
+	SiSSetupDSTColorDepth(pSiS->DstColor);
 	SiSSetupSRCXY(0,0)
 	SiSSetupSRCFG(fg)
 	SiSSetupROP(sisPatALUConv[rop])
@@ -711,7 +990,13 @@ SiSSubsequentCPUToScreenColorExpand(ScrnInfoPtr pScrn,
 /*      SiSDoCMD*/
 	pSiS->DoColorExpand = TRUE;
 }
+#endif
 
+/* ------ Screen To Screen Color Expand ------------------------------- */
+
+/* TW: The hareware does not seem to support this the way we need it */
+
+#ifdef STSCE
 static void
 SiSSetupForScreenToScreenColorExpand(ScrnInfoPtr pScrn,
                                 int fg, int bg,
@@ -722,135 +1007,391 @@ SiSSetupForScreenToScreenColorExpand(ScrnInfoPtr pScrn,
 	PDEBUG(ErrorF("Setup ScreenToScreen ColorExp(0x%x,0x%x, 0x%x)\n",
 							fg, bg, rop));
 
-	SiSSetupDSTBase(0)
-/*      SiSSetupDSTRect(pSiS->scrnOffset, pScrn->virtualY)*/
+	SiSSetupDSTColorDepth(pSiS->DstColor)
 	SiSSetupDSTRect(pSiS->scrnOffset, -1)
-	SiSSetupDSTColorDepth(SISPTR(pScrn)->DstColor);
-	SiSSetupSRCFG(fg)
-	SiSSetupSRCBG(bg)
-	SiSSetupSRCXY(0,0)
 	SiSSetupROP(sisALUConv[rop])
-	SiSSetupCMDFlag(X_INC | Y_INC | ENCOLOREXP)
+	SiSSetupSRCFG(fg)
+	/* SiSSetupSRCXY(0,0) */
+
+	if(bg == -1) {
+		SiSSetupCMDFlag(TRANSPARENT | ENCOLOREXP | X_INC |
+				Y_INC | SRCVIDEO);
+	} else {
+		SiSSetupSRCBG(bg);
+		SiSSetupCMDFlag(ENCOLOREXP | X_INC | Y_INC |
+				SRCVIDEO);
+	};
 }
+#endif
+
+/* TW. This method blits in a single task; this does not seem to work
+ * because the hardware does not use the source pitch as scanline
+ * offset but only to calculate pattern address from source X and Y.
+ * XAA provides the pattern bitmap with scrnOffset (displayWidth * bpp/8)
+ * offset, but this does not seem to be supported by the hardware.
+ */
+#ifdef STSCE
+
+/* For testing, these are the methods: (use only one at a time!) */
+
+#undef npitch 		/* Normal: Use srcx/y as srcx/y, use scrnOffset as source pitch
+			 * This would work if the hareware used the source pitch for
+			 * incrementing the source address after each scanline - but
+			 * it doesn't do this! The first line of the area is correctly
+			 * color expanded, but since the source pitch is ignored and
+			 * the source address not incremented correctly, the following
+			 * lines are color expanded with any bit pattern that is left
+			 * in the unused space of the source bitmap (which is organized
+			 * with the depth of the screen framebuffer hence with a pitch
+			 * of scrnOffset).
+			 */
+
+#undef pitchdw    	/* Use source pitch "displayWidth / 8" instead
+		   	 * of scrnOffset (=displayWidth * bpp / 8)
+			 * This can't work, because the pitch of the source
+			 * bitmap is scrnoffset!
+		   	 */
+
+#define  nopitch   	/* Calculate srcbase with srcx and srcy, set the
+		   	 * pitch to scrnOffset (which IS the correct pitch
+		   	 * for the source bitmap) and set srcx and srcy both
+		   	 * to 0.
+			 * This would work if the hareware used the source pitch for
+			 * incrementing the source address after each scanline - but
+			 * it doesn't do this! Again: The first line of the area is
+			 * correctly color expanded, but since the source pitch is
+			 * ignored for scanline address incremention, the following
+			 * lines are not correctly color expanded.
+			 * WHATEVER I write to source pitch is ignored!
+		   	 */
 
 static void
 SiSSubsequentScreenToScreenColorExpand(ScrnInfoPtr pScrn,
                                 int x, int y, int w, int h,
                                 int srcx, int srcy, int skipleft)
 {
-	SISPtr          pSiS = SISPTR(pScrn);
+	SISPtr pSiS = SISPTR(pScrn);
+        long srcbase, dstbase;
+#if 0
+	int _x0, _y0, _x1, _y1;
+#endif
+#ifdef pitchdw
+	int newsrcx, newsrcy;
 
-	PDEBUG(ErrorF("Sub ScreenToScreen ColorExp(%d,%d, %d,%d, %d,%d, %d)\n",
-							x, y, w, h, srcx, srcy, skipleft));
+	/* srcx and srcy are provided based on a scrnOffset pitch ( = displayWidth * bpp / 8 )
+	 * We recalulate srcx and srcy based on pitch = displayWidth / 8
+	 */
+        newsrcy = ((pSiS->scrnOffset * srcy) + (srcx * ((pScrn->bitsPerPixel+7)/8))) /
+					  (pScrn->displayWidth/8);
+        newsrcx = ((pSiS->scrnOffset * srcy) + (srcx * ((pScrn->bitsPerPixel+7)/8))) %
+					  (pScrn->displayWidth/8);
+#endif
+	xf86DrvMsg(0, X_INFO, "Sub ScreenToScreen ColorExp(%d,%d, %d,%d, %d,%d, %d)\n",
+					x, y, w, h, srcx, srcy, skipleft);
+
+	srcbase = dstbase = 0;
+
+#ifdef pitchdw
+	if (newsrcy >= 2048) {
+		srcbase = (pScrn->displayWidth / 8) * newsrcy;
+		newsrcy = 0;
+	}
+#endif
+#ifdef nopitch
+	srcbase = (pSiS->scrnOffset * srcy) + (srcx * ((pScrn->bitsPerPixel+7)/8));
+#endif
+#ifdef npitch
+	if (srcy >= 2048) {
+		srcbase = pSiS->scrnOffset * srcy;
+		srcy = 0;
+	}
+#endif
+	if (y >= 2048) {
+		dstbase = pSiS->scrnOffset * y;
+		y = 0;
+	}
+#ifdef SISDUALHEAD
+	if(pSiS->VGAEngine != SIS_530_VGA) {
+	   srcbase += HEADOFFSET;
+	   dstbase += HEADOFFSET;
+	}
+#endif
+	SiSSetupSRCBase(srcbase)
+	SiSSetupDSTBase(dstbase)
+
+#ifdef pitchdw
+	SiSSetupSRCPitch(pScrn->displayWidth/8)
+#endif
+#ifdef nopitch
+	SiSSetupSRCPitch(pSiS->scrnOffset)
+	/* SiSSetupSRCPitch(100) */ /* For test - has NO effect WHATSOEVER */
+#endif
+#ifdef npitch
+	SiSSetupSRCPitch(pSiS->scrnOffset)
+#endif
 
-	SiSSetupSRCPitch(((w+31)&0xFFE0)/8)
-	SiSSetupDSTXY(x,y)
 	SiSSetupRect(w,h)
+
+#if 0   /* How do I implement the offset? Not this way, that's for sure.. */
+	if (skipleft > 0) {
+		_x0 = x+skipleft;
+		_y0 = y;
+		_x1 = x+w;
+		_y1 = y+h;
+		SiSSetupClipLT(_x0, _y0);
+		SiSSetupClipRB(_x1, _y1);
+		SiSSetupCMDFlag(CLIPENABLE);
+	}
+#endif
+#ifdef pitchdw
+	SiSSetupSRCXY(newsrcx, newsrcy)
+#endif
+#ifdef nopitch
+	SiSSetupSRCXY(0,0)
+#endif
+#ifdef npitch
+	SiSSetupSRCXY(srcx, srcy)
+#endif
+
+	SiSSetupDSTXY(x,y)
+
 	SiSDoCMD
 }
 #endif
 
-static void 
-SiSSetupForScanlineCPUToScreenColorExpandFill(ScrnInfoPtr pScrn, int fg, int bg, int rop, unsigned int planemask)
+/* TW: TEST: Do it scanline-wise because the other way does not seem to
+ *     be supported by the hardware. (The source pitch seems to be
+ *     displayWidth * (bbp/8) as opposed by the XAA HOWTO, where
+ *     it is stated that the pitch would be displayWidth pixels;
+ *     besides, the hardware seems to ignore the source pitch
+ *     for address increments.)
+ *     Apart from this (which can be solved by doing the color
+ *     expand scanline-wise), I don't know how to implement the
+ *     offset argument. The current method (which uses hardware
+ *     clipping) does not work.
+ *
+ *     THIS DOES NOT WORK IN THE CURRENT STATE.
+ */
+#if 0
+static void
+SiSSubsequentScreenToScreenColorExpand(ScrnInfoPtr pScrn,
+                                int x, int y, int w, int h,
+                                int srcx, int srcy, int offset)
+{
+	SISPtr pSiS = SISPTR(pScrn);
+        long srcbase, dstbase;
+	int _x0, _y0, _x1, _y1;
+
+	int newsrcx, newsrcy;
+
+        newsrcy = ((pSiS->scrnOffset * srcy) + (srcx * ((pScrn->bitsPerPixel+7)/8))) /
+					  (((((w+7)/8)+3) >> 2) * 4);
+					  /* (pScrn->displayWidth/8); */
+        newsrcx = ((pSiS->scrnOffset * srcy) + (srcx * ((pScrn->bitsPerPixel+7)/8))) %
+					  (((((w+7)/8)+3) >> 2) * 4);
+					  /* (pScrn->displayWidth/8); */
+
+	xf86DrvMsg(0, X_INFO, "Sub STS CE(%d,%d, %d,%d, %d,%d, %d)\n",
+					x, y, w, h, srcx, srcy, skipleft);
+
+	srcbase = dstbase = 0;
+	if (newsrcy >= 2048) {
+		srcbase = (((((w+7)/8)+3) >> 2) * 4) * newsrcy;
+		          /* (pScrn->displayWidth/8) * newsrcy;      */
+		          /* pSiS->scrnOffset * srcy;                */
+		newsrcy = 0;
+	}
+	if (y >= 2048) {
+		dstbase = pSiS->scrnOffset * y;
+		y = 0;
+	}
+#ifdef SISDUALHEAD
+	srcbase += HEADOFFSET;
+	dstbase += HEADOFFSET;
+#endif
+	SiSSetupDSTBase(dstbase)
+
+	SiSSetupRect(w, 1)
+
+	SiSSetupSRCXY(newsrcx, newsrcy)
+
+	/* SiSSetupSRCPitch(pScrn->displayWidth/8) */ /* old: (((w+31)&0xFFE0)/8) */
+	SiSSetupSRCPitch(((((w+7)/8)+3) >> 2) * 4)
+#if 1
+        if (offset > 0) {
+	    SiSSetupCMDFlag(CLIPENABLE)
+        } else
+            pSiS->CommandReg &= ~CLIPENABLE;
+#endif
+
+	while (h) {
+
+	   SiSSetupSRCBase(srcbase)
+#if 1
+	   if (offset > 0) {
+		_x0 = x+skipleft;
+		_y0 = y;
+		_x1 = x+w;
+		_y1 = y+h;
+		SiSSetupClipLT(_x0, _y0);
+		SiSSetupClipRB(_x1, _y1);
+	   }
+#endif
+	   SiSSetupDSTXY(x,y)
+
+	   SiSDoCMD
+
+	   srcbase += ((((w+7)/8)+3) >> 2) * 4 * 8* ((pScrn->bitsPerPixel+7)/8);
+	   	      /* pSiS->scrnOffset;  */
+	   y++;
+	   h--;
+	}
+}
+#endif
+
+
+/* ----- CPU To Screen Color Expand (scanline-wise) ----------------- */
+
+/* We do it using the indirect method */
+
+static void
+SiSSetupForScanlineCPUToScreenColorExpandFill(ScrnInfoPtr pScrn,
+		int fg, int bg, int rop, unsigned int planemask)
 {
 	SISPtr pSiS=SISPTR(pScrn);
-	
-	while ((MMIO_IN16(pSiS->IOBase, 0x8242) & 0x1F00)!=0) {}
+
+        /* TW: Make sure that current CPU-driven BitBlt buffer stage is 0
+	 *     This is required!!! (Otherwise -> drawing errors)
+	 */
+	while((MMIO_IN16(pSiS->IOBase, 0x8242) & 0x1F00) != 0) {} /* WDR: == 0x10 */
+
+#if 0   /* TW: This is obviously not needed */
 	pSiS->ColorExpandRingHead = 0;
 	pSiS->ColorExpandRingTail = pSiS->ColorExpandBufferNumber - 1;
+#endif
+
 	SiSSetupSRCXY(0,0);
 	SiSSetupROP(sisALUConv[rop]);
 	SiSSetupSRCFG(fg);
 	SiSSetupDSTRect(pSiS->scrnOffset, -1);
 	SiSSetupDSTColorDepth(pSiS->DstColor);
-	if (bg == -1) { 
-		SiSSetupCMDFlag(TRANSPARENT | ENCOLOREXP | X_INC |
-						Y_INC|SRCSYSTEM);  
-	}
-	else {
+	if(bg == -1) {
+		SiSSetupCMDFlag(TRANSPARENT |
+		                ENCOLOREXP |
+				X_INC | Y_INC |
+				SRCCPUBLITBUF);
+	} else {
 		SiSSetupSRCBG(bg);
-		SiSSetupCMDFlag(ENCOLOREXP | X_INC | Y_INC |
-						SRCSYSTEM);      
-	};
+		SiSSetupCMDFlag(ENCOLOREXP |
+		                X_INC | Y_INC |
+				SRCCPUBLITBUF);
+	}
 }
 
-static void SiSSubsequentScanlineCPUToScreenColorExpandFill(
-                        ScrnInfoPtr pScrn, int x, int y, int w, 
+
+static void
+SiSSubsequentScanlineCPUToScreenColorExpandFill(
+                        ScrnInfoPtr pScrn, int x, int y, int w,
                         int h, int skipleft)
 {
 	SISPtr pSiS = SISPTR(pScrn);
-static int     srcpitch;
 	int _x0, _y0, _x1, _y1;
 	long dstbase;
 
 	dstbase = 0;
 	if (y >= 2048) {
-		dstbase = pSiS->scrnOffset*y;
+		dstbase = pSiS->scrnOffset * y;
 		y = 0;
 	}
+#ifdef SISDUALHEAD
+	if(pSiS->VGAEngine != SIS_530_VGA) {
+	   	dstbase += HEADOFFSET;
+	}
+#endif
+
+	/* TW: Wait until there is no color expansion command in queue
+	 *     (This solves the OpenOffice.org window-move bug)
+	 *     Added Idle-check - bit 23 is set sometimes, although
+	 *     engine is actually idle!
+	 *     Update: Bit 23 is not reliable. After heavy 3D engine
+	 *     action, this bit never gets cleared again. So do
+	 *     SiSIdle instead.
+	 */
+	if((MMIO_IN16(pSiS->IOBase, 0x8242) & 0xe000) != 0xe000) {
+           /* while ((MMIO_IN16(pSiS->IOBase, 0x8242) & 0x0080) != 0) {} */
+	   SiSIdle
+	}
+
 	SiSSetupDSTBase(dstbase)
 
 	if (skipleft > 0) {
-		_x0 = x+skipleft;
+		_x0 = x + skipleft;
 		_y0 = y;
-		_x1 = x+w;
-		_y1 = y+h;
+		_x1 = x + w;
+		_y1 = y + h;
 		SiSSetupClipLT(_x0, _y0);
 		SiSSetupClipRB(_x1, _y1);
-		SiSSetupCMDFlag(CLIPENABLE);    
-	} 
+		SiSSetupCMDFlag(CLIPENABLE);
+	} else {
+		pSiS->CommandReg &= (~CLIPENABLE);
+	}
+
 	SiSSetupRect(w, 1);
-	srcpitch = ((((w+7)/8)+3) >> 2) * 4;
-	SiSSetupSRCPitch(srcpitch);
-	pSiS->ycurrent = y;
+        SiSSetupSRCPitch(((((w+7)/8)+3) >> 2) * 4);
 	pSiS->xcurrent = x;
+	pSiS->ycurrent = y;
 }
 
-static void SiSSubsequentColorExpandScanline(ScrnInfoPtr pScrn, int bufno)
+static void
+SiSSubsequentColorExpandScanline(ScrnInfoPtr pScrn, int bufno)
 {
-	SISPtr pSiS=SISPTR(pScrn);              
+	SISPtr pSiS=SISPTR(pScrn);
+#if 0
 	int newhead,bltbufstage,newtail;
+#endif
+	long cbo;
 
-	SiSSetupSRCBase(pSiS->ColorExpandBufferScreenOffset[bufno]);
-	SiSSetupDSTXY(pSiS->xcurrent, pSiS->ycurrent);
-	pSiS->ycurrent++;
-	SiSDoCMD
-	newhead = pSiS->ColorExpandRingHead = (bufno + 1) & 
-							pSiS->ColorExpandBufferCountMask;
-	while (newhead == pSiS->ColorExpandRingTail) {
-		bltbufstage = (int)((MMIO_IN16(pSiS->IOBase,0x8242) & 
-							0x1F00)>>8);
-		newtail = newhead - (bltbufstage + 1);
-		pSiS->ColorExpandRingTail = (newtail >= 0) ? 
-							newtail: (pSiS->ColorExpandBufferNumber+newtail);
+	cbo = pSiS->ColorExpandBufferScreenOffset[bufno];
+#ifdef SISDUALHEAD
+	if(pSiS->VGAEngine != SIS_530_VGA) {
+	   cbo += HEADOFFSET;
 	}
-}
+#endif
 
-#if 0
-static void SiSSetupForImageWrite(ScrnInfoPtr pScrn, int rop, 
-            unsigned int planemask, int trans_color, int bpp, int depth)
-{
-	return;
-}
+	/* TW: Wait until there is no color expansion command in queue
+	 *     (This solves the GTK-big-font bug)
+	 *     Added Idle-check - bit 23 is set sometimes, although
+	 *     engine is actually idle!
+	 *     Update: Bit 23 is not reliable. After heavy 3D engine
+	 *     action, this bit never gets cleared again. So do
+	 *     SiSIdle instead.
+	 */
+	if((MMIO_IN16(pSiS->IOBase, 0x8242) & 0xe000) != 0xe000) {
+	  /* while ((MMIO_IN16(pSiS->IOBase, 0x8242) & 0x0080) != 0) {} */
+	    SiSIdle
+	}
 
-static void SiSSubsequentImageWriteRect(ScrnInfoPtr pScrn, 
-            int x, int y, int w, int h, int skipleft)
-{
-	return;
-}
-#endif
+	SiSSetupSRCBase(cbo);
 
-#ifdef DEBUG
-static void
-MMIODump(ScrnInfoPtr pScrn)
-{
-	SISPtr  pSiS = SISPTR(pScrn);
-	int     i;
+	SiSSetupDSTXY(pSiS->xcurrent, pSiS->ycurrent);
 
-	SiSIdle
-	for (i=0x8200; i <= 0x823c; i += 4)  {
-		ErrorF("[%x] %0X \n", i,
-				MMIO_IN32(pSiS->IOBase,i));
+	SiSDoCMD
+
+	pSiS->ycurrent++;
+
+	if(pSiS->VGAEngine == SIS_530_VGA) {
+	   while(MMIO_IN8(pSiS->IOBase, 0x8242) & 0x80) {}
+	}
+
+#if 0	/* TW: What is this good for? The Head/Tail data is never ever used elsewhere! */
+	pSiS->ColorExpandRingHead = newhead =
+			(bufno + 1) & pSiS->ColorExpandBufferCountMask;
+	while (newhead == pSiS->ColorExpandRingTail) {
+		bltbufstage = (int)((MMIO_IN16(pSiS->IOBase,0x8242) & 0x1F00) >> 8);
+		newtail = newhead - (bltbufstage + 1);
+		pSiS->ColorExpandRingTail = (newtail >= 0) ?
+				newtail : (pSiS->ColorExpandBufferNumber + newtail);
 	}
-}
 #endif
+}
+