Simple perf tweaks for a rectilinear Hilbert curve.

Some tweaks to avoid stack copies and branches that save ~25% in
_cairo_traps_tessellate_convex_quad().

1 files changed, 65 insertions, 110 deletions

diff --git a/src/cairo-traps.c b/src/cairo-traps.c
index cdfdadd3..fabfecb0 100644
--- a/src/cairo-traps.c
+++ b/src/cairo-traps.c
@@ -41,13 +41,6 @@
 
 /* private functions */
 
-static cairo_status_t
-_cairo_traps_grow (cairo_traps_t *traps);
-
-static void
-_cairo_traps_add_trap (cairo_traps_t *traps, cairo_fixed_t top, cairo_fixed_t bottom,
-		       cairo_line_t *left, cairo_line_t *right);
-
 static int
 _compare_point_fixed_by_y (const void *av, const void *bv);
 
@@ -94,12 +87,8 @@ _cairo_traps_clear (cairo_traps_t *traps)
 void
 _cairo_traps_fini (cairo_traps_t *traps)
 {
-    if (traps->traps && traps->traps != traps->traps_embedded)
+    if (traps->traps != traps->traps_embedded)
 	free (traps->traps);
-
-    traps->traps = NULL;
-    traps->traps_size = 0;
-    traps->num_traps = 0;
 }
 
 /**
@@ -111,9 +100,9 @@ _cairo_traps_fini (cairo_traps_t *traps)
  * Initializes a #cairo_traps_t to contain a single rectangular
  * trapezoid.
  **/
-cairo_status_t
+void
 _cairo_traps_init_box (cairo_traps_t *traps,
-		       cairo_box_t   *box)
+		       const cairo_box_t   *box)
 {
     _cairo_traps_init (traps);
 
@@ -131,25 +120,41 @@ _cairo_traps_init_box (cairo_traps_t *traps,
     traps->traps[0].right.p2 = box->p2;
 
     traps->extents = *box;
-
-    return traps->status;
 }
 
-cairo_status_t
-_cairo_traps_status (cairo_traps_t *traps)
+/* make room for at least one more trap */
+static cairo_bool_t
+_cairo_traps_grow (cairo_traps_t *traps)
 {
-    return traps->status;
+    cairo_trapezoid_t *new_traps;
+    int new_size = 2 * MAX (traps->traps_size, 16);
+
+    if (traps->traps == traps->traps_embedded) {
+	new_traps = _cairo_malloc_ab (new_size, sizeof (cairo_trapezoid_t));
+	if (new_traps != NULL)
+	    memcpy (new_traps, traps->traps, sizeof (traps->traps_embedded));
+    } else {
+	new_traps = _cairo_realloc_ab (traps->traps,
+	                               new_size, sizeof (cairo_trapezoid_t));
+    }
+
+    if (new_traps == NULL) {
+	traps->status = _cairo_error (CAIRO_STATUS_NO_MEMORY);
+	return FALSE;
+    }
+
+    traps->traps = new_traps;
+    traps->traps_size = new_size;
+    return TRUE;
 }
 
-static void
-_cairo_traps_add_trap (cairo_traps_t *traps, cairo_fixed_t top, cairo_fixed_t bottom,
+void
+_cairo_traps_add_trap (cairo_traps_t *traps,
+		       cairo_fixed_t top, cairo_fixed_t bottom,
 		       cairo_line_t *left, cairo_line_t *right)
 {
     cairo_trapezoid_t *trap;
 
-    if (traps->status)
-	return;
-
     /* Note: With the goofy trapezoid specification, (where an
      * arbitrary two points on the lines can specified for the left
      * and right edges), these limit checks would not work in
@@ -209,9 +214,8 @@ _cairo_traps_add_trap (cairo_traps_t *traps, cairo_fixed_t top, cairo_fixed_t bo
 	return;
     }
 
-    if (traps->num_traps >= traps->traps_size) {
-	traps->status = _cairo_traps_grow (traps);
-	if (traps->status)
+    if (traps->num_traps == traps->traps_size) {
+	if (! _cairo_traps_grow (traps))
 	    return;
     }
 
@@ -246,51 +250,6 @@ _cairo_traps_add_trap (cairo_traps_t *traps, cairo_fixed_t top, cairo_fixed_t bo
     traps->num_traps++;
 }
 
-void
-_cairo_traps_add_trap_from_points (cairo_traps_t *traps, cairo_fixed_t top, cairo_fixed_t bottom,
-				   cairo_point_t left_p1, cairo_point_t left_p2,
-				   cairo_point_t right_p1, cairo_point_t right_p2)
-{
-    cairo_line_t left;
-    cairo_line_t right;
-
-    if (traps->status)
-	return;
-
-    left.p1 = left_p1;
-    left.p2 = left_p2;
-
-    right.p1 = right_p1;
-    right.p2 = right_p2;
-
-    _cairo_traps_add_trap (traps, top, bottom, &left, &right);
-}
-
-/* make room for at least one more trap */
-static cairo_status_t
-_cairo_traps_grow (cairo_traps_t *traps)
-{
-    cairo_trapezoid_t *new_traps;
-    int new_size = 2 * MAX (traps->traps_size, 16);
-
-    if (traps->traps == traps->traps_embedded) {
-	new_traps = _cairo_malloc_ab (new_size, sizeof (cairo_trapezoid_t));
-	if (new_traps)
-	    memcpy (new_traps, traps->traps, sizeof (traps->traps_embedded));
-    } else {
-	new_traps = _cairo_realloc_ab (traps->traps,
-	                               new_size, sizeof (cairo_trapezoid_t));
-    }
-
-    if (new_traps == NULL)
-	return _cairo_error (CAIRO_STATUS_NO_MEMORY);
-
-    traps->traps = new_traps;
-    traps->traps_size = new_size;
-
-    return CAIRO_STATUS_SUCCESS;
-}
-
 static int
 _compare_point_fixed_by_y (const void *av, const void *bv)
 {
@@ -379,7 +338,8 @@ _cairo_trapezoid_array_translate_and_scale (cairo_trapezoid_t *offset_traps,
  * quadrilateral. We would not benefit from having any distinct
  * implementation of triangle vs. quadrilateral tessellation here. */
 cairo_status_t
-_cairo_traps_tessellate_triangle (cairo_traps_t *traps, cairo_point_t t[3])
+_cairo_traps_tessellate_triangle (cairo_traps_t *traps,
+				  const cairo_point_t t[3])
 {
     cairo_point_t quad[4];
 
@@ -392,12 +352,15 @@ _cairo_traps_tessellate_triangle (cairo_traps_t *traps, cairo_point_t t[3])
 }
 
 cairo_status_t
-_cairo_traps_tessellate_convex_quad (cairo_traps_t *traps, cairo_point_t q[4])
+_cairo_traps_tessellate_convex_quad (cairo_traps_t *traps,
+				     const cairo_point_t q[4])
 {
     int a, b, c, d;
     int i;
     cairo_slope_t ab, ad;
     cairo_bool_t b_left_of_d;
+    cairo_line_t left;
+    cairo_line_t right;
 
     /* Choose a as a point with minimal y */
     a = 0;
@@ -462,15 +425,13 @@ _cairo_traps_tessellate_convex_quad (cairo_traps_t *traps, cairo_point_t q[4])
 	     *  | /      \|     \ \  c.y d.y  cd   ad
 	     *  d         d       d
 	     */
-	    _cairo_traps_add_trap_from_points (traps,
-					       q[a].y, q[b].y,
-					       q[a], q[b], q[a], q[d]);
-	    _cairo_traps_add_trap_from_points (traps,
-					       q[b].y, q[c].y,
-					       q[b], q[c], q[a], q[d]);
-	    _cairo_traps_add_trap_from_points (traps,
-					       q[c].y, q[d].y,
-					       q[c], q[d], q[a], q[d]);
+	    left.p1  = q[a]; left.p2  = q[b];
+	    right.p1 = q[a]; right.p2 = q[d];
+	    _cairo_traps_add_trap (traps, q[a].y, q[b].y, &left, &right);
+	    left.p1  = q[b]; left.p2  = q[c];
+	    _cairo_traps_add_trap (traps, q[b].y, q[c].y, &left, &right);
+	    left.p1  = q[c]; left.p2  = q[d];
+	    _cairo_traps_add_trap (traps, q[c].y, q[d].y, &left, &right);
 	} else {
 	    /* Y-sort is abcd and b is right of d, (slope(ab) <= slope (ad))
 	     *
@@ -482,15 +443,13 @@ _cairo_traps_tessellate_convex_quad (cairo_traps_t *traps, cairo_point_t q[4])
 	     *  / /     |/      \ | c.y d.y  ad  cd
 	     *  d       d         d
 	     */
-	    _cairo_traps_add_trap_from_points (traps,
-					       q[a].y, q[b].y,
-					       q[a], q[d], q[a], q[b]);
-	    _cairo_traps_add_trap_from_points (traps,
-					       q[b].y, q[c].y,
-					       q[a], q[d], q[b], q[c]);
-	    _cairo_traps_add_trap_from_points (traps,
-					       q[c].y, q[d].y,
-					       q[a], q[d], q[c], q[d]);
+	    left.p1  = q[a]; left.p2  = q[d];
+	    right.p1 = q[a]; right.p2 = q[b];
+	    _cairo_traps_add_trap (traps, q[a].y, q[b].y, &left, &right);
+	    right.p1 = q[b]; right.p2 = q[c];
+	    _cairo_traps_add_trap (traps, q[b].y, q[c].y, &left, &right);
+	    right.p1 = q[c]; right.p2 = q[d];
+	    _cairo_traps_add_trap (traps, q[c].y, q[d].y, &left, &right);
 	}
     } else {
 	if (b_left_of_d) {
@@ -504,15 +463,13 @@ _cairo_traps_tessellate_convex_quad (cairo_traps_t *traps, cairo_point_t q[4])
 	     *  //         \ /     \|  d.y c.y  bc  dc
 	     *  c           c       c
 	     */
-	    _cairo_traps_add_trap_from_points (traps,
-					       q[a].y, q[b].y,
-					       q[a], q[b], q[a], q[d]);
-	    _cairo_traps_add_trap_from_points (traps,
-					       q[b].y, q[d].y,
-					       q[b], q[c], q[a], q[d]);
-	    _cairo_traps_add_trap_from_points (traps,
-					       q[d].y, q[c].y,
-					       q[b], q[c], q[d], q[c]);
+	    left.p1  = q[a]; left.p2  = q[b];
+	    right.p1 = q[a]; right.p2 = q[d];
+	    _cairo_traps_add_trap (traps, q[a].y, q[b].y, &left, &right);
+	    left.p1  = q[b]; left.p2  = q[c];
+	    _cairo_traps_add_trap (traps, q[b].y, q[d].y, &left, &right);
+	    right.p1 = q[d]; right.p2 = q[c];
+	    _cairo_traps_add_trap (traps, q[d].y, q[c].y, &left, &right);
 	} else {
 	    /* Y-sort is abdc and b is right of d, (slope (ab) <= slope (ad))
 	     *
@@ -524,15 +481,13 @@ _cairo_traps_tessellate_convex_quad (cairo_traps_t *traps, cairo_point_t q[4])
 	     *  |/     \ /         \\  d.y c.y  dc  bc
 	     *  c       c	   c
 	     */
-	    _cairo_traps_add_trap_from_points (traps,
-					       q[a].y, q[b].y,
-					       q[a], q[d], q[a], q[b]);
-	    _cairo_traps_add_trap_from_points (traps,
-					       q[b].y, q[d].y,
-					       q[a], q[d], q[b], q[c]);
-	    _cairo_traps_add_trap_from_points (traps,
-					       q[d].y, q[c].y,
-					       q[d], q[c], q[b], q[c]);
+	    left.p1  = q[a]; left.p2  = q[d];
+	    right.p1 = q[a]; right.p2 = q[b];
+	    _cairo_traps_add_trap (traps, q[a].y, q[b].y, &left, &right);
+	    right.p1 = q[b]; right.p2 = q[c];
+	    _cairo_traps_add_trap (traps, q[b].y, q[d].y, &left, &right);
+	    left.p1  = q[d]; left.p2  = q[c];
+	    _cairo_traps_add_trap (traps, q[d].y, q[c].y, &left, &right);
 	}
     }