sna: Don't unroll BLT points

The compiler is smarter than I am; unrolling hurts here.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

2 files changed, 19 insertions, 89 deletions

diff --git a/src/sna/sna_accel.c b/src/sna/sna_accel.c
index 6fa15b29..a11a77d5 100644
--- a/src/sna/sna_accel.c
+++ b/src/sna/sna_accel.c
@@ -4593,7 +4593,7 @@ static inline bool box32_trim_and_translate(Box32Rec *box, DrawablePtr d, GCPtr
 	return box32_clip(box, gc);
 }
 
-static inline void box_add_pt(BoxPtr box, int16_t x, int16_t y)
+static inline void box_add_xy(BoxPtr box, int16_t x, int16_t y)
 {
 	if (box->x1 > x)
 		box->x1 = x;
@@ -4606,6 +4606,11 @@ static inline void box_add_pt(BoxPtr box, int16_t x, int16_t y)
 		box->y2 = y;
 }
 
+static inline void box_add_pt(BoxPtr box, const DDXPointRec *pt)
+{
+	box_add_xy(box, pt->x, pt->y);
+}
+
 static inline bool box32_to_box16(const Box32Rec *b32, BoxRec *b16)
 {
 	b16->x1 = b32->x1;
@@ -8968,36 +8973,11 @@ sna_poly_point_extents(DrawablePtr drawable, GCPtr gc,
 			last.x += pt->x;
 			last.y += pt->y;
 			pt++;
-			box_add_pt(&box, last.x, last.y);
+			box_add_xy(&box, last.x, last.y);
 		}
 	} else {
-		--n; ++pt;
-		while (n >= 8) {
-			box_add_pt(&box, pt[0].x, pt[0].y);
-			box_add_pt(&box, pt[1].x, pt[1].y);
-			box_add_pt(&box, pt[2].x, pt[2].y);
-			box_add_pt(&box, pt[3].x, pt[3].y);
-			box_add_pt(&box, pt[4].x, pt[4].y);
-			box_add_pt(&box, pt[5].x, pt[5].y);
-			box_add_pt(&box, pt[6].x, pt[6].y);
-			box_add_pt(&box, pt[7].x, pt[7].y);
-			pt += 8;
-			n -= 8;
-		}
-		if (n & 4) {
-			box_add_pt(&box, pt[0].x, pt[0].y);
-			box_add_pt(&box, pt[1].x, pt[1].y);
-			box_add_pt(&box, pt[2].x, pt[2].y);
-			box_add_pt(&box, pt[3].x, pt[3].y);
-			pt += 4;
-		}
-		if (n & 2) {
-			box_add_pt(&box, pt[0].x, pt[0].y);
-			box_add_pt(&box, pt[1].x, pt[1].y);
-			pt += 2;
-		}
-		if (n & 1)
-			box_add_pt(&box, pt[0].x, pt[0].y);
+		while (--n)
+			box_add_pt(&box, ++pt);
 	}
 	box.x2++;
 	box.y2++;
@@ -9709,7 +9689,7 @@ sna_poly_line_extents(DrawablePtr drawable, GCPtr gc,
 			y += pt->y;
 			if (blt)
 				blt &= pt->x == 0 || pt->y == 0;
-			box_add_pt(&box, x, y);
+			box_add_xy(&box, x, y);
 		}
 	} else {
 		int x = box.x1;
@@ -9721,7 +9701,7 @@ sna_poly_line_extents(DrawablePtr drawable, GCPtr gc,
 				x = pt->x;
 				y = pt->y;
 			}
-			box_add_pt(&box, pt->x, pt->y);
+			box_add_pt(&box, pt);
 		}
 	}
 	box.x2++;
diff --git a/src/sna/sna_blt.c b/src/sna/sna_blt.c
index 59b8141c..b62d7fd7 100644
--- a/src/sna/sna_blt.c
+++ b/src/sna/sna_blt.c
@@ -3185,65 +3185,15 @@ fastcall static void sna_blt_fill_op_points(struct sna *sna,
 		assert(kgem->nbatch < kgem->surface);
 
 		if ((dx|dy) == 0) {
-			while (n_this_time >= 8) {
-				*((uint64_t *)b + 0) = pt_add(cmd, p+0, 0, 0);
-				*((uint64_t *)b + 1) = pt_add(cmd, p+1, 0, 0);
-				*((uint64_t *)b + 2) = pt_add(cmd, p+2, 0, 0);
-				*((uint64_t *)b + 3) = pt_add(cmd, p+3, 0, 0);
-				*((uint64_t *)b + 4) = pt_add(cmd, p+4, 0, 0);
-				*((uint64_t *)b + 5) = pt_add(cmd, p+5, 0, 0);
-				*((uint64_t *)b + 6) = pt_add(cmd, p+6, 0, 0);
-				*((uint64_t *)b + 7) = pt_add(cmd, p+7, 0, 0);
-				b += 16;
-				n_this_time -= 8;
-				p += 8;
-			}
-			if (n_this_time & 4) {
-				*((uint64_t *)b + 0) = pt_add(cmd, p+0, 0, 0);
-				*((uint64_t *)b + 1) = pt_add(cmd, p+1, 0, 0);
-				*((uint64_t *)b + 2) = pt_add(cmd, p+2, 0, 0);
-				*((uint64_t *)b + 3) = pt_add(cmd, p+3, 0, 0);
-				b += 8;
-				p += 4;
-			}
-			if (n_this_time & 2) {
-				*((uint64_t *)b + 0) = pt_add(cmd, p+0, 0, 0);
-				*((uint64_t *)b + 1) = pt_add(cmd, p+1, 0, 0);
-				b += 4;
-				p += 2;
-			}
-			if (n_this_time & 1)
-				*((uint64_t *)b + 0) = pt_add(cmd, p++, 0, 0);
+			do {
+				*(uint64_t *)b = pt_add(cmd, p++, 0, 0);
+				b += 2;
+			} while (--n_this_time);
 		} else {
-			while (n_this_time >= 8) {
-				*((uint64_t *)b + 0) = pt_add(cmd, p+0, dx, dy);
-				*((uint64_t *)b + 1) = pt_add(cmd, p+1, dx, dy);
-				*((uint64_t *)b + 2) = pt_add(cmd, p+2, dx, dy);
-				*((uint64_t *)b + 3) = pt_add(cmd, p+3, dx, dy);
-				*((uint64_t *)b + 4) = pt_add(cmd, p+4, dx, dy);
-				*((uint64_t *)b + 5) = pt_add(cmd, p+5, dx, dy);
-				*((uint64_t *)b + 6) = pt_add(cmd, p+6, dx, dy);
-				*((uint64_t *)b + 7) = pt_add(cmd, p+7, dx, dy);
-				b += 16;
-				n_this_time -= 8;
-				p += 8;
-			}
-			if (n_this_time & 4) {
-				*((uint64_t *)b + 0) = pt_add(cmd, p+0, dx, dy);
-				*((uint64_t *)b + 1) = pt_add(cmd, p+1, dx, dy);
-				*((uint64_t *)b + 2) = pt_add(cmd, p+2, dx, dy);
-				*((uint64_t *)b + 3) = pt_add(cmd, p+3, dx, dy);
-				b += 8;
-				p += 8;
-			}
-			if (n_this_time & 2) {
-				*((uint64_t *)b + 0) = pt_add(cmd, p+0, dx, dy);
-				*((uint64_t *)b + 1) = pt_add(cmd, p+1, dx, dy);
-				b += 4;
-				p += 2;
-			}
-			if (n_this_time & 1)
-				*((uint64_t *)b + 0) = pt_add(cmd, p++, dx, dy);
+			do {
+				*(uint64_t *)b = pt_add(cmd, p++, dx, dy);
+				b += 2;
+			} while (--n_this_time);
 		}
 
 		if (!n)