glamor: Scissor Render composite operations to the bounds of the drawing.

Unlike the previous two fixes, this one introduces new GL calls and
statechanges of the scissor.  However, given that our Render drawing
already does CPU side transformation and inefficient box upload, this
shouldn't be a limiting factor for Render acceleration.

Surprisingly, it improves x11perf -comppixwin10 -repeat 1 -reps 10000
on i965 by 3.21191% +/- 1.79977% (n=50).

v2: Make the jump to the exit land after scissor disable.
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Keith Packard <keithp@keithp.com>

glamor/glamor_render.c

@@ -1198,6 +1198,29 @@ glamor_composite_with_shader(CARD8 op,
 
     nrect_max = MIN(nrect, GLAMOR_COMPOSITE_VBO_VERT_CNT / 4);
 
+    if (nrect < 100) {
+        BoxRec bounds = glamor_start_rendering_bounds();
+
+        for (int i = 0; i < nrect; i++) {
+            BoxRec box = {
+                .x1 = rects[i].x_dst,
+                .y1 = rects[i].y_dst,
+                .x2 = rects[i].x_dst + rects[i].width,
+                .y2 = rects[i].y_dst + rects[i].height,
+            };
+            glamor_bounds_union_box(&bounds, &box);
+        }
+
+        if (bounds.x1 >= bounds.x2 || bounds.y1 >= bounds.y2)
+            goto disable_va;
+
+        glEnable(GL_SCISSOR_TEST);
+        glScissor(bounds.x1 + dest_x_off,
+                  bounds.y1 + dest_y_off,
+                  bounds.x2 - bounds.x1,
+                  bounds.y2 - bounds.y1);
+    }
+
     while (nrect) {
         int mrect, rect_processed;
         int vb_stride;
@@ -1279,6 +1302,8 @@ glamor_composite_with_shader(CARD8 op,
         }
     }
 
+    glDisable(GL_SCISSOR_TEST);
+disable_va:
     glDisableVertexAttribArray(GLAMOR_VERTEX_POS);
     glDisableVertexAttribArray(GLAMOR_VERTEX_SOURCE);
     glDisableVertexAttribArray(GLAMOR_VERTEX_MASK);