--- BasiliskII/src/Unix/video_blit.h	2000/09/23 06:51:46	1.2
+++ BasiliskII/src/Unix/video_blit.h	2002/01/15 14:58:37	1.6
@@ -1,7 +1,7 @@
 /*
  *  video_blit.h - Video/graphics emulation, blitters
  *
- *  Basilisk II (C) 1997-2000 Christian Bauer
+ *  Basilisk II (C) 1997-2002 Christian Bauer
  *
  *  This program is free software; you can redistribute it and/or modify
  *  it under the terms of the GNU General Public License as published by
@@ -18,79 +18,97 @@
  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  */
 
-#ifndef FB_BLIT_1
+#ifndef FB_DEPTH
+# error "Undefined screen depth"
+#endif
+
+#if !defined(FB_BLIT_1) && (FB_DEPTH <= 16)
 # error "Undefined 16-bit word blit function"
 #endif
 
-#ifndef FB_BLIT_2
+#if !defined(FB_BLIT_2)
 # error "Undefined 32-bit word blit function"
 #endif
 
-#ifndef FB_DEPTH
-# error "Undefined screen depth"
+#if !defined(FB_BLIT_4)
+# error "Undefined 64-bit word blit function"
 #endif
 
 static void FB_FUNC_NAME(uint8 * dest, const uint8 * source, uint32 length)
 {
+#define DEREF_WORD_PTR(ptr, ofs) (((uint16 *)(ptr))[(ofs)])
+#define DEREF_LONG_PTR(ptr, ofs) (((uint32 *)(ptr))[(ofs)])
+#define DEREF_QUAD_PTR(ptr, ofs) (((uint64 *)(ptr))[(ofs)])
+	
+#ifndef UNALIGNED_PROFITABLE
 #if FB_DEPTH <= 8
 	// Align source and dest to 16-bit word boundaries
-	if (FB_DEPTH <= 8 && ((unsigned long) source) & 1) {
+	if (((unsigned long) source) & 1) {
 		*dest++ = *source++;
 		length -= 1;
 	}
 #endif
 	
-	// source and dest are mutually aligned
-	uint16 * swp = ((uint16 *)source);
-	uint16 * dwp = ((uint16 *) dest );
-
-#if FB_DEPTH <= 8
-	if (length >= 2) {
-#endif
-	
 #if FB_DEPTH <= 16
 	// Align source and dest to 32-bit word boundaries
 	if (((unsigned long) source) & 2) {
-		const uint16 val = *swp++;
-		FB_BLIT_1(*dwp++, val);
+		FB_BLIT_1(DEREF_WORD_PTR(dest, 0), DEREF_WORD_PTR(source, 0));
+		dest += 2; source += 2;
 		length -= 2;
 	}
 #endif
+#endif
 	
-	// Blit 4-byte words
-	if (length >= 4) {
-		const int remainder = (length / 4) % 8;
-		uint32 * slp = (uint32 *)swp + remainder;
-		uint32 * dlp = (uint32 *)dwp + remainder;
+	// Blit 8-byte words
+	if (length >= 8) {
+		const int remainder = (length / 8) % 8;
+		source += remainder * 8;
+		dest += remainder * 8;
 		
-		int n = ((length / 4) + 7) / 8;
+		int n = ((length / 8) + 7) / 8;
 		switch (remainder) {
 		case 0:	do {
-				slp += 8; dlp += 8;
-				FB_BLIT_2(dlp[-8], slp[-8]);
-		case 7: FB_BLIT_2(dlp[-7], slp[-7]);
-		case 6: FB_BLIT_2(dlp[-6], slp[-6]);
-		case 5: FB_BLIT_2(dlp[-5], slp[-5]);
-		case 4: FB_BLIT_2(dlp[-4], slp[-4]);
-		case 3: FB_BLIT_2(dlp[-3], slp[-3]);
-		case 2: FB_BLIT_2(dlp[-2], slp[-2]);
-		case 1: FB_BLIT_2(dlp[-1], slp[-1]);
+				dest += 64; source += 64;
+				FB_BLIT_4(DEREF_QUAD_PTR(dest, -8), DEREF_QUAD_PTR(source, -8));
+		case 7: FB_BLIT_4(DEREF_QUAD_PTR(dest, -7), DEREF_QUAD_PTR(source, -7));
+		case 6: FB_BLIT_4(DEREF_QUAD_PTR(dest, -6), DEREF_QUAD_PTR(source, -6));
+		case 5: FB_BLIT_4(DEREF_QUAD_PTR(dest, -5), DEREF_QUAD_PTR(source, -5));
+		case 4: FB_BLIT_4(DEREF_QUAD_PTR(dest, -4), DEREF_QUAD_PTR(source, -4));
+		case 3: FB_BLIT_4(DEREF_QUAD_PTR(dest, -3), DEREF_QUAD_PTR(source, -3));
+		case 2: FB_BLIT_4(DEREF_QUAD_PTR(dest, -2), DEREF_QUAD_PTR(source, -2));
+		case 1: FB_BLIT_4(DEREF_QUAD_PTR(dest, -1), DEREF_QUAD_PTR(source, -1));
 				} while (--n > 0);
 		}
 	}
 	
+	// There could be one long left to blit
+	if (length & 4) {
+		FB_BLIT_2(DEREF_LONG_PTR(dest, 0), DEREF_LONG_PTR(source, 0));
+#if FB_DEPTH <= 16
+		dest += 4;
+		source += 4;
+#endif
+	}
+	
 #if FB_DEPTH <= 16
-	// There might remain one word to blit
+	// There could be one word left to blit
 	if (length & 2) {
-		uint16 * const s = (uint16 *)(((uint8 *)swp) + length - 2);
-		uint16 * const d = (uint16 *)(((uint8 *)dwp) + length - 2);
-		FB_BLIT_1(*d, *s);
+		FB_BLIT_1(DEREF_WORD_PTR(dest, 0), DEREF_WORD_PTR(source, 0));
+#if FB_DEPTH <= 8
+		dest += 2;
+		source += 2;
+#endif
 	}
 #endif
 	
 #if FB_DEPTH <= 8
-	}
+	// There could be one byte left to blit
+	if (length & 1)
+		*dest = *source;
 #endif
+	
+#undef DEREF_LONG_PTR
+#undef DEREF_WORD_PTR
 }
 
 #undef FB_FUNC_NAME
@@ -103,6 +121,10 @@ static void FB_FUNC_NAME(uint8 * dest, c
 #undef FB_BLIT_2
 #endif
 
+#ifdef FB_BLIT_4
+#undef FB_BLIT_4
+#endif
+
 #ifdef FB_DEPTH
 #undef FB_DEPTH
 #endif