fix: optimize for bare-metal HashFunSegFind, DCBlotColor8, GrCalcScreenUpdates, GrUpdateScreen32; inc patch num

src/Kernel/FunSeg.ZC

@@ -12,19 +12,24 @@ U0 HashFunSegFind(CHashTable *h, U8 *addr, Bool *_has_lower, U64 *_best, CHash *
 {
 	Bool		*has_lower = *_has_lower;
 	CHashExport	*tmpex;
-	U64			 i, j, best = *_best;
+	U64			 i, j, best = *_best, reg mask = h->mask;
-	CHash		*res = *_res;
+	CHash		*res = *_res, **body = h->body;;
+	CHashFun	*temp_hashfun;
+	U16			 hf_flags;
 
-	for (i = 0; i <= h->mask; i++)
+	for (i = 0; i <= mask; i++)
 	{
-		tmpex = h->body[i];
+		tmpex = body[i];
+		
 		while (tmpex)
 		{
 			j = 0;
 			if (tmpex->type & HTT_FUN)
 			{
-				if (!Bt(&tmpex(CHashFun *)->flags, Cf_EXTERN) && !Bt(&tmpex(CHashFun *)->flags, Ff_INTERNAL))
+				temp_hashfun = tmpex; // cast
-					j = tmpex(CHashFun *)->exe_addr;
+				hf_flags = temp_hashfun->flags;
+				if (!Bt(&hf_flags, Cf_EXTERN) && !Bt(&hf_flags, Ff_INTERNAL))
+					j = temp_hashfun->exe_addr;
 			}
 			else if (tmpex->type & HTT_EXPORT_SYS_SYM)
 				j = tmpex->val;

src/Kernel/KGlobals.ZC

@@ -14,7 +14,7 @@ U8		*rev_bits_table; //Table with U8 bits reversed
 CDate	 local_time_offset;
 F64		*pow10_I64,
 		 sys_os_version = 2.03;
-U64		 sys_os_version_sub = 117;
+U64		 sys_os_version_sub = 118;
 U8		*sys_os_version_str;
 U8		*sys_os_version_full;
 U8		*sys_os_version_nice;

src/System/Gr/GrScreen.ZC

@@ -345,22 +345,29 @@ U0 GrUpdateTextFG()
 U0 DCBlotColor8(CDC *dc, CDC *img)
 {
 	U8	*src = img->body, *b0 = dc->body;
-	I64  j, k, d0 = img->width_internal * img->height;
+	I64  reg j, reg jj, v = 0, k, d0 = img->width_internal * img->height;
 
-	for (k = 0; k < d0; k++)
+	for (k = 0; k < d0; k += 8)
 	{
-		j = *src++;
+		j = *(src(U64 *));
-		if (j != TRANSPARENT)
+		src += 8;
-			*b0++ = j;
+		do
-		else
+		{
+			jj = j & 0xFF;
+
+			if (jj != TRANSPARENT)
+				*b0 = jj;
+
 			b0++;
+			j >>= 8;
+		} while (v++ < 8);
 	}
 }
 
 U0 GrCalcScreenUpdates()
 {
-	U16 *screen, *last_screen = gr.screen_cache;
+	U64 *screen, *last_screen = gr.screen_cache;
-	U64 i, *src = text.raw_screen, *dst = text.fb_alias, diffs_size = GR_WIDTH * GR_HEIGHT / 2;
+	U64 i, *src = text.raw_screen, *dst = text.fb_alias, diffs_size = GR_WIDTH * GR_HEIGHT / 8;
 
 	if (gr.screen_zoom == 1)
 		screen = gr.dc2->body;
@@ -368,17 +375,16 @@ U0 GrCalcScreenUpdates()
 		screen = gr.zoomed_dc->body;
 
 	for (i = 0; i < diffs_size; i++)
-	{
 		if (screen[i] != last_screen[i])
-			dst[i] = src[i];
+			MemCopy(dst + i * 4, src + i * 4, 4 * 64);
-	}
+
-	MemCopy(gr.screen_cache, screen, diffs_size * 2);
+	MemCopy(gr.screen_cache, screen, diffs_size * 8);
 }
 
 U0 GrUpdateScreen32()
 {
-	U64 size, *dst;
+	U64 size, *dst, reg src64val;
-	U8 *src;
+	U8 *src, reg a, reg b, c, reg d;
 
 	if (gr.screen_zoom == 1)
 	{
@@ -392,9 +398,17 @@ U0 GrUpdateScreen32()
 		size = src + gr.zoomed_dc->height * gr.zoomed_dc->width_internal;
 	}
 	dst = text.raw_screen;
-	while (src < size) //draw 2 pixels at a time
+	while (src < size) // draw 4 pixels at a time
-		*dst++ = gr_palette[*src++ & 0xFF] | gr_palette[*src++ & 0xFF] << 32;
+	{
-
+		src64val = *(src(U64 *));
+		a = src64val & 0xFF;
+		b = (src64val >>= 8) & 0xFF;
+		c = (src64val >>= 8) & 0xFF;
+		d = (src64val >>= 8) & 0xFF;
+		src += 4;
+		*dst++ = gr_palette[a] | gr_palette[b] << 32;
+		*dst++ = gr_palette[c] | gr_palette[d] << 32;
+	}
 
 	GrCalcScreenUpdates;