cinap's draw optimization for byte-aligned images, and replace
"(x&y)? 1: 0 with "(x&y) != 0". 

Reference: /n/atom/patch/applied/memdrawbytes
Date: Tue Apr 15 16:37:37 CES 2014
Signed-off-by: quanstro@quanstro.net

--- /sys/src/libmemdraw/draw.c	Tue Apr 15 16:35:20 2014
+++ /sys/src/libmemdraw/draw.c	Tue Apr 15 16:35:22 2014
@@ -52,7 +52,6 @@
 void
 memimageinit(void)
 {
-	extern void memdrawallocinit(void);
 	static int didinit = 0;
 
 	if(didinit)
@@ -222,7 +221,7 @@
 	int splitcoords;
 	Rectangle omr;
 
-	if(r->min.x>=r->max.x || r->min.y>=r->max.y)
+	if(badrect(*r))
 		return 0;
 	splitcoords = (p0->x!=p1->x) || (p0->y!=p1->y);
 	/* clip to destination */
@@ -1471,7 +1470,7 @@
 	end = p->bytey0e + y*p->bwidth;
 	cmap = p->img->cmap->cmap2rgb;
 	convgrey = p->convgrey;
-	copyalpha = (p->img->flags&Falpha) ? 1 : 0;
+	copyalpha = (p->img->flags&Falpha) != 0;
 
 	w = buf;
 	dx = p->dx;
@@ -1528,8 +1527,8 @@
 static void
 writecmap(Param *p, uchar *w, Buffer src)
 {
-	uchar *cmap, *red, *grn, *blu;
-	int i, dx, delta;
+	uchar *cmap, *red, *grn, *blu, *alpha;
+	int i, dx, delta, a, m;
 
 	cmap = p->img->cmap->rgb2cmap;
 	
@@ -1539,8 +1538,20 @@
 	blu = src.blu;
 
 	dx = p->dx;
-	for(i=0; i<dx; i++, red+=delta, grn+=delta, blu+=delta)
-		*w++ = cmap[(*red>>4)*256+(*grn>>4)*16+(*blu>>4)];
+	if(p->img->flags&Falpha){
+		alpha = src.alpha;
+		m = p->img->shift[CMap]/8;
+		a = p->img->shift[CAlpha]/8;
+		for(i=0; i<dx; i++, red+=delta, grn+=delta, blu+=delta, w+=2){
+			w[a] = *alpha;
+			if(alpha != &ones)
+				alpha+=delta;
+			w[m] = cmap[(*red>>4)*256+(*grn>>4)*16+(*blu>>4)];
+		}
+	} else {
+		for(i=0; i<dx; i++, red+=delta, grn+=delta, blu+=delta)
+			*w++ = cmap[(*red>>4)*256+(*grn>>4)*16+(*blu>>4)];
+	}
 }
 
 #define DBG if(0)
@@ -1566,7 +1577,7 @@
 	convgrey = p->convgrey;	/* convert rgb to grey */
 	isgrey = img->flags&Fgrey;
 	alphaonly = p->alphaonly;
-	copyalpha = (img->flags&Falpha) ? 1 : 0;
+	copyalpha = (img->flags&Falpha) != 0;
 
 DBG print("copyalpha %d alphaonly %d convgrey %d isgrey %d\n", copyalpha, alphaonly, convgrey, isgrey);
 	/* if we can, avoid processing everything */
@@ -1618,16 +1629,17 @@
 				*w++ = RGB2K(ured, ugrn, ublu);
 DBG print("%x\n", w[-1]);
 			}else{
-				*w++ = brepl[(u >> img->shift[CBlue]) & img->mask[CBlue]];
-				*w++ = grepl[(u >> img->shift[CGreen]) & img->mask[CGreen]];
-				*w++ = rrepl[(u >> img->shift[CRed]) & img->mask[CRed]];
+				w[0] = ublu;
+				w[1] = ugrn;
+				w[2] = ured;
+				w += 3;
 			}
 		}
 		r += nb;
 		if(r == end)
 			r = begin;
 	}
-	
+
 	b.alpha = copyalpha ? buf : &ones;
 	b.rgba = (ulong*)buf;
 	if(alphaonly){
@@ -1671,7 +1683,6 @@
 	dx = p->dx;
 
 	nb = img->depth/8;
-	mask = (nb==4) ? 0 : ~((1<<img->depth)-1);
 
 	isalpha = img->flags&Falpha;
 	isgrey = img->flags&Fgrey;
@@ -1683,6 +1694,37 @@
 		adelta = 0;
 	}
 
+	if((img->flags&Fbytes) != 0){
+		int ogry, ored, ogrn, oblu, oalp;
+
+		ogry = img->shift[CGrey]/8;
+		ored = img->shift[CRed]/8;
+		ogrn = img->shift[CGreen]/8;
+		oblu = img->shift[CBlue]/8;
+		oalp = img->shift[CAlpha]/8;
+
+		for(i=0; i<dx; i++){
+			if(isgrey){
+				w[ogry] = *grey;
+				grey += delta;
+			} else {
+				w[ored] = *red;
+				w[ogrn] = *grn;
+				w[oblu] = *blu;
+				red += delta;
+				grn += delta;
+				blu += delta;
+			}
+			if(isalpha){
+				w[oalp] = *alpha;
+				alpha += adelta;
+			}
+			w += nb;
+		}
+		return;
+	}
+
+	mask = (nb==4) ? 0 : ~((1<<img->depth)-1);
 	for(i=0; i<dx; i++){
 		u = w[0] | (w[1]<<8) | (w[2]<<16) | (w[3]<<24);
 DBG print("u %.8lux...", u);
@@ -1979,17 +2021,6 @@
 /*
  * Optimized draw for filling and scrolling; uses memset and memmove.
  */
-static void
-memsetb(void *vp, uchar val, int n)
-{
-	uchar *p, *ep;
-
-	p = vp;
-	ep = p+n;
-	while(p<ep)
-		*p++ = val;
-}
-
 static void
 memsets(void *vp, ushort val, int n)
 {