Index: subsystems/win32/win32k/dib/dib1bpp.c
===================================================================
--- subsystems/win32/win32k/dib/dib1bpp.c	(revision 40280)
+++ subsystems/win32/win32k/dib/dib1bpp.c	(working copy)
@@ -59,177 +59,6 @@
   }
 }
 
-static
-void
-DIB_1BPP_BitBltSrcCopy_From1BPP (
-	SURFOBJ* DestSurf, SURFOBJ* SourceSurf,
-	PRECTL DestRect, POINTL *SourcePoint )
-{
-	// the 'window' in this sense is the x-position that corresponds
-	// to the left-edge of the 8-pixel byte we are currently working with.
-	// dwx is current x-window, dwx2 is the 'last' window we need to process
-	int dwx, dwx2; // destination window x-position
-	int swx; // source window y-position
-
-	// left and right edges of source and dest rectangles
-	int dl = DestRect->left; // dest left
-	int dr = DestRect->right-1; // dest right (inclusive)
-	int sl = SourcePoint->x; // source left
-	int sr = sl + dr - dl; // source right (inclusive)
-
-	// which direction are we going?
-	int xinc;
-	int yinc;
-	int ySrcDelta, yDstDelta;
-
-	// following 4 variables are used for the y-sweep
-	int dy; // dest y
-	int dy1; // dest y start
-	int dy2; // dest y end
-	int sy1; // src y start
-
-    int dx;
-	int shift;
-	BYTE srcmask, dstmask;
-
-	// 'd' and 's' are the dest & src buffer pointers that I use on my x-sweep
-	// 'pd' and 'ps' are the dest & src buffer pointers used on the inner y-sweep
-	PBYTE d, pd; // dest ptrs
-	PBYTE s, ps; // src ptrs
-
-	shift = (dl-sl)&7;
-
-	if ( DestRect->top <= SourcePoint->y )
-	{
-		// moving up ( scan top -> bottom )
-		dy1 = DestRect->top;
-		dy2 = DestRect->bottom - 1;
-		sy1 = SourcePoint->y;
-		yinc = 1;
-		ySrcDelta = SourceSurf->lDelta;
-		yDstDelta = DestSurf->lDelta;
-	}
-	else
-	{
-		// moving down ( scan bottom -> top )
-		dy1 = DestRect->bottom - 1;
-		dy2 = DestRect->top;
-		sy1 = SourcePoint->y + dy1 - dy2;
-		yinc = -1;
-		ySrcDelta = -SourceSurf->lDelta;
-		yDstDelta = -DestSurf->lDelta;
-	}
-	if ( DestRect->left <= SourcePoint->x )
-	{
-		// moving left ( scan left->right )
-		dwx = dl&~7;
-		swx = (sl-(dl&7))&~7;
-		dwx2 = dr&~7;
-		xinc = 1;
-	}
-	else
-	{
-		// moving right ( scan right->left )
-		dwx = dr&~7;
-		swx = (sr-(dr&7))&~7; //(sr-7)&~7; // we need the left edge of this block... thus the -7
-		dwx2 = dl&~7;
-		xinc = -1;
-	}
-	d = &(((PBYTE)DestSurf->pvScan0)[dy1*DestSurf->lDelta + (dwx>>3)]);
-	s = &(((PBYTE)SourceSurf->pvScan0)[sy1*SourceSurf->lDelta + (swx>>3)]);
-	for ( ;; )
-	{
-		dy = dy1;
-		pd = d;
-		ps = s;
-		srcmask = 0xff;
-		dx = dwx; /* dest x for this pass */
-		if ( dwx < dl )
-		{
-			int diff = dl-dwx;
-			srcmask &= (1<<(8-diff))-1;
-			dx = dl;
-		}
-		if ( dwx+7 > dr )
-		{
-			int diff = dr-dwx+1;
-			srcmask &= ~((1<<(8-diff))-1);
-		}
-		dstmask = ~srcmask;
-
-		// we unfortunately *must* have 5 different versions of the inner
-		// loop to be certain we don't try to read from memory that is not
-		// needed and may in fact be invalid
-		if ( !shift )
-		{
-			for ( ;; )
-			{
-				*pd = (BYTE)((*pd & dstmask) | (*ps & srcmask));
-
-				// this *must* be here, because we could be going up *or* down...
-				if ( dy == dy2 )
-					break;
-				dy += yinc;
-				pd += yDstDelta;
-				ps += ySrcDelta;
-			}
-		}
-		else if ( !(0xFF00 & (srcmask<<shift) ) ) // check if ps[0] not needed...
-		{
-			for ( ;; )
-			{
-				*pd = (BYTE)((*pd & dstmask)
-					| ( ( ps[1] >> shift ) & srcmask ));
-
-				// this *must* be here, because we could be going up *or* down...
-				if ( dy == dy2 )
-					break;
-				dy += yinc;
-				pd += yDstDelta;
-				ps += ySrcDelta;
-			}
-		}
-		else if ( !(0xFF & (srcmask<<shift) ) ) // check if ps[1] not needed...
-		{
-			for ( ;; )
-			{
-				*pd = (*pd & dstmask)
-					| ( ( ps[0] << ( 8 - shift ) ) & srcmask );
-
-				// this *must* be here, because we could be going up *or* down...
-				if ( dy == dy2 )
-					break;
-				dy += yinc;
-				pd += yDstDelta;
-				ps += ySrcDelta;
-			}
-		}
-		else // both ps[0] and ps[1] are needed
-		{
-			for ( ;; )
-			{
-				*pd = (*pd & dstmask)
-					| ( ( ( (ps[1])|(ps[0]<<8) ) >> shift ) & srcmask );
-
-				// this *must* be here, because we could be going up *or* down...
-				if ( dy == dy2 )
-					break;
-				dy += yinc;
-				pd += yDstDelta;
-				ps += ySrcDelta;
-			}
-		}
-
-		// this *must* be here, because we could be going right *or* left...
-		if ( dwx == dwx2 )
-			break;
-		d += xinc;
-		s += xinc;
-		dwx += xinc<<3;
-		swx += xinc<<3;
-	}
-}
-
 BOOLEAN
 DIB_1BPP_BitBltSrcCopy(PBLTINFO BltInfo)
 {
@@ -238,7 +67,18 @@
 	switch ( BltInfo->SourceSurface->iBitmapFormat )
 	{
 	case BMF_1BPP:
-		DIB_1BPP_BitBltSrcCopy_From1BPP ( BltInfo->DestSurface, BltInfo->SourceSurface, &BltInfo->DestRect, &BltInfo->SourcePoint );
+		for (j=BltInfo->DestRect.top; j<BltInfo->DestRect.bottom; j++)
+		{
+			sx = BltInfo->SourcePoint.x;
+			for (i=BltInfo->DestRect.left; i<BltInfo->DestRect.right; i++)
+			{
+				DIB_1BPP_PutPixel(BltInfo->DestSurface, i, j, 
+					XLATEOBJ_iXlate(BltInfo->XlateSourceToDest, 
+					DIB_1BPP_GetPixel(BltInfo->SourceSurface, sx, sy)));
+				sx++;
+			}
+			sy++;
+		}
 		break;
 
 	case BMF_4BPP:
@@ -247,12 +87,9 @@
 			sx = BltInfo->SourcePoint.x;
 			for (i=BltInfo->DestRect.left; i<BltInfo->DestRect.right; i++)
 			{
-				if(XLATEOBJ_iXlate(BltInfo->XlateSourceToDest, DIB_4BPP_GetPixel(BltInfo->SourceSurface, sx, sy)) == 0)
-				{
-					DIB_1BPP_PutPixel(BltInfo->DestSurface, i, j, 0);
-				} else {
-					DIB_1BPP_PutPixel(BltInfo->DestSurface, i, j, 1);
-				}
+				DIB_1BPP_PutPixel(BltInfo->DestSurface, i, j, 
+					XLATEOBJ_iXlate(BltInfo->XlateSourceToDest, 
+					DIB_4BPP_GetPixel(BltInfo->SourceSurface, sx, sy)));
 				sx++;
 			}
 			sy++;
@@ -265,12 +102,9 @@
 			sx = BltInfo->SourcePoint.x;
 			for (i=BltInfo->DestRect.left; i<BltInfo->DestRect.right; i++)
 			{
-				if(XLATEOBJ_iXlate(BltInfo->XlateSourceToDest, DIB_8BPP_GetPixel(BltInfo->SourceSurface, sx, sy)) == 0)
-				{
-					DIB_1BPP_PutPixel(BltInfo->DestSurface, i, j, 0);
-				} else {
-					DIB_1BPP_PutPixel(BltInfo->DestSurface, i, j, 1);
-				}
+				DIB_1BPP_PutPixel(BltInfo->DestSurface, i, j, 
+					XLATEOBJ_iXlate(BltInfo->XlateSourceToDest, 
+					DIB_8BPP_GetPixel(BltInfo->SourceSurface, sx, sy)));
 				sx++;
 			}
 			sy++;
@@ -283,12 +117,9 @@
 			sx = BltInfo->SourcePoint.x;
 			for (i=BltInfo->DestRect.left; i<BltInfo->DestRect.right; i++)
 			{
-				if(XLATEOBJ_iXlate(BltInfo->XlateSourceToDest, DIB_16BPP_GetPixel(BltInfo->SourceSurface, sx, sy)) == 0)
-				{
-					DIB_1BPP_PutPixel(BltInfo->DestSurface, i, j, 0);
-				} else {
-					DIB_1BPP_PutPixel(BltInfo->DestSurface, i, j, 1);
-				}
+				DIB_1BPP_PutPixel(BltInfo->DestSurface, i, j, 
+					XLATEOBJ_iXlate(BltInfo->XlateSourceToDest, 
+					DIB_16BPP_GetPixel(BltInfo->SourceSurface, sx, sy)));
 				sx++;
 			}
 			sy++;
@@ -301,12 +132,9 @@
 			sx = BltInfo->SourcePoint.x;
 			for (i=BltInfo->DestRect.left; i<BltInfo->DestRect.right; i++)
 			{
-				if(XLATEOBJ_iXlate(BltInfo->XlateSourceToDest, DIB_24BPP_GetPixel(BltInfo->SourceSurface, sx, sy)) == 0)
-				{
-					DIB_1BPP_PutPixel(BltInfo->DestSurface, i, j, 0);
-				} else {
-					DIB_1BPP_PutPixel(BltInfo->DestSurface, i, j, 1);
-				}
+				DIB_1BPP_PutPixel(BltInfo->DestSurface, i, j, 
+					XLATEOBJ_iXlate(BltInfo->XlateSourceToDest, 
+					DIB_24BPP_GetPixel(BltInfo->SourceSurface, sx, sy)));
 				sx++;
 			}
 			sy++;
@@ -319,12 +147,9 @@
 			sx = BltInfo->SourcePoint.x;
 			for (i=BltInfo->DestRect.left; i<BltInfo->DestRect.right; i++)
 			{
-				if(XLATEOBJ_iXlate(BltInfo->XlateSourceToDest, DIB_32BPP_GetPixel(BltInfo->SourceSurface, sx, sy)) == 0)
-				{
-					DIB_1BPP_PutPixel(BltInfo->DestSurface, i, j, 0);
-				} else {
-					DIB_1BPP_PutPixel(BltInfo->DestSurface, i, j, 1);
-				}
+				DIB_1BPP_PutPixel(BltInfo->DestSurface, i, j, 
+					XLATEOBJ_iXlate(BltInfo->XlateSourceToDest, 
+					DIB_32BPP_GetPixel(BltInfo->SourceSurface, sx, sy)));
 				sx++;
 			}
 			sy++;
Index: subsystems/win32/win32k/objects/dibobj.c
===================================================================
--- subsystems/win32/win32k/objects/dibobj.c	(revision 40280)
+++ subsystems/win32/win32k/objects/dibobj.c	(working copy)
@@ -436,7 +436,7 @@
     NTSTATUS Status = STATUS_SUCCESS;
     PDC pDC;
     HBITMAP hSourceBitmap = NULL;
-    SURFOBJ *pDestSurf, *pSourceSurf = NULL;
+    SURFOBJ *pDestSurf = NULL, *pSourceSurf = NULL;
     RECTL rcDest;
     POINTL ptSource;
     INT DIBWidth;
@@ -476,6 +476,19 @@
         return 0;
     }
 
+    SURFACE *psurf = SURFACE_LockSurface(pDC->rosdc.hBitmap);
+    // Use hDIBPalette if it exists
+    if (psurf->hDIBPalette)
+    {
+        DDBPalette = psurf->hDIBPalette;
+    }
+    else
+    {
+        // Destination palette obtained from the DC
+        DDBPalette = pDC->ppdev->DevInfo.hpalDefault;
+    }
+    SURFACE_UnlockSurface(psurf);
+
     pDestSurf = EngLockSurface((HSURF)pDC->rosdc.hBitmap);
     if (!pDestSurf)
     {
@@ -522,7 +535,7 @@
     }
 
     /* Obtain destination palette from the DC */
-    pDCPalette = PALETTE_LockPalette(pDC->ppdev->DevInfo.hpalDefault);
+    pDCPalette = PALETTE_LockPalette(DDBPalette);
     if (!pDCPalette)
     {
        SetLastWin32Error(ERROR_INVALID_HANDLE);
@@ -531,7 +544,6 @@
     }
 
     DDBPaletteType = pDCPalette->Mode;
-    DDBPalette = pDC->ppdev->DevInfo.hpalDefault;
     PALETTE_UnlockPalette(pDCPalette);
 
     DIBPalette = BuildDIBPalette(bmi, (PINT)&DIBPaletteType);