diff --git a/src/sna/fb/fb.h b/src/sna/fb/fb.h
index 215aec99..adefaa19 100644
--- a/src/sna/fb/fb.h
+++ b/src/sna/fb/fb.h
@@ -515,7 +515,9 @@ extern RegionPtr
 fbBitmapToRegion(PixmapPtr pixmap);
 
 extern void
-fbPolyPoint(DrawablePtr drawable, GCPtr gc, int mode, int n, xPoint *pt);
+fbPolyPoint(DrawablePtr drawable, GCPtr gc,
+	    int mode, int n, xPoint *pt,
+	    unsigned flags);
 
 extern void
 fbPushImage(DrawablePtr drawable, GCPtr gc,
diff --git a/src/sna/fb/fbpoint.c b/src/sna/fb/fbpoint.c
index e3974fca..3df79a26 100644
--- a/src/sna/fb/fbpoint.c
+++ b/src/sna/fb/fbpoint.c
@@ -25,21 +25,27 @@
 #include <micoord.h>
 
 #define DOTS	    fbDots8
+#define DOTS__SIMPLE fbDots8__simple
 #define BITS	    BYTE
 #include "fbpointbits.h"
 #undef BITS
+#undef DOTS__SIMPLE
 #undef DOTS
 
 #define DOTS	    fbDots16
+#define DOTS__SIMPLE fbDots16__simple
 #define BITS	    CARD16
 #include "fbpointbits.h"
 #undef BITS
+#undef DOTS__SIMPLE
 #undef DOTS
 
 #define DOTS	    fbDots32
+#define DOTS__SIMPLE fbDots32__simple
 #define BITS	    CARD32
 #include "fbpointbits.h"
 #undef BITS
+#undef DOTS__SIMPLE
 #undef DOTS
 
 static void
@@ -74,7 +80,7 @@ fbDots(FbBits *dstOrig, FbStride dstStride, int dstBpp,
 
 void
 fbPolyPoint(DrawablePtr drawable, GCPtr gc,
-	    int mode, int n, xPoint *pt)
+	    int mode, int n, xPoint *pt, unsigned flags)
 {
 	FbBits *dst;
 	FbStride dstStride;
@@ -97,16 +103,30 @@ fbPolyPoint(DrawablePtr drawable, GCPtr gc,
 
 	fbGetDrawable(drawable, dst, dstStride, dstBpp, dstXoff, dstYoff);
 	dots = fbDots;
-	switch (dstBpp) {
-	case 8:
-		dots = fbDots8;
-		break;
-	case 16:
-		dots = fbDots16;
-		break;
-	case 32:
-		dots = fbDots32;
-		break;
+	if ((flags & 2) == 0 && fb_gc(gc)->and == 0) {
+		switch (dstBpp) {
+		case 8:
+			dots = fbDots8__simple;
+			break;
+		case 16:
+			dots = fbDots16__simple;
+			break;
+		case 32:
+			dots = fbDots32__simple;
+			break;
+		}
+	} else {
+		switch (dstBpp) {
+		case 8:
+			dots = fbDots8;
+			break;
+		case 16:
+			dots = fbDots16;
+			break;
+		case 32:
+			dots = fbDots32;
+			break;
+		}
 	}
 	dots(dst, dstStride, dstBpp, gc->pCompositeClip, pt, n,
 	     drawable->x, drawable->y, dstXoff, dstYoff,
diff --git a/src/sna/fb/fbpointbits.h b/src/sna/fb/fbpointbits.h
index 2d71c1d2..60bf4881 100644
--- a/src/sna/fb/fbpointbits.h
+++ b/src/sna/fb/fbpointbits.h
@@ -106,5 +106,43 @@ DOTS(FbBits * dst,
 	}
 }
 
+static void
+DOTS__SIMPLE(FbBits * dst,
+	     FbStride dstStride,
+	     int dstBpp,
+	     RegionPtr region,
+	     xPoint * ptsOrig,
+	     int npt, int xorg, int yorg, int xoff, int yoff,
+	     FbBits and, FbBits xor)
+{
+	uint32_t *pts = (uint32_t *) ptsOrig;
+	BITS *bits = (BITS *) dst, *p;
+	BITS bxor = (BITS) xor;
+	FbStride bitsStride = dstStride * (sizeof(FbBits) / sizeof(BITS));
+
+	bits += bitsStride * (yorg + yoff) + (xorg + xoff);
+	while (npt >= 2) {
+		union {
+			uint32_t pt32[2];
+			uint64_t pt64;
+		} pt;
+		pt.pt64 = *(uint64_t *)pts;
+
+		p = bits + intToY(pt.pt32[0]) * bitsStride + intToX(pt.pt32[0]);
+		WRITE(p, bxor);
+
+		p = bits + intToY(pt.pt32[1]) * bitsStride + intToX(pt.pt32[1]);
+		WRITE(p, bxor);
+
+		pts += 2;
+		npt -= 2;
+	}
+	if (npt) {
+		uint32_t pt = *pts;
+		p = bits + intToY(pt) * bitsStride + intToX(pt);
+		WRITE(p, bxor);
+	}
+}
+
 #undef RROP
 #undef isClipped
diff --git a/src/sna/sna_accel.c b/src/sna/sna_accel.c
index 792d0070..98810502 100644
--- a/src/sna/sna_accel.c
+++ b/src/sna/sna_accel.c
@@ -6696,7 +6696,7 @@ fallback:
 		goto out_gc;
 
 	DBG(("%s: fbPolyPoint\n", __FUNCTION__));
-	fbPolyPoint(drawable, gc, mode, n, pt);
+	fbPolyPoint(drawable, gc, mode, n, pt, flags);
 	FALLBACK_FLUSH(drawable);
 out_gc:
 	sna_gc_move_to_gpu(gc);