[set] Optimize non-const page_for() using last_page_lookup caching

This speeds up SetOrderedInsert tests by 15 to 40 percent, and the
subset_mplus1p benchmarks by 9 to 27 percent.

src/hb-bit-set.hh

@@ -874,7 +874,19 @@ struct hb_bit_set_t
 
   page_t *page_for (hb_codepoint_t g, bool insert = false)
   {
-    page_map_t map = {get_major (g), pages.length};
+    unsigned major = get_major (g);
+
+    /* The extra page_map length is necessary; can't just rely on vector here,
+     * since the next check would be tricked because a null page also has
+     * major==0, which we can't distinguish from an actualy major==0 page... */
+    if (likely (last_page_lookup < page_map.length))
+    {
+      auto &cached_page = page_map.arrayZ[last_page_lookup];
+      if (cached_page.major == major)
+	return &pages[cached_page.index];
+    }
+
+    page_map_t map = {major, pages.length};
     unsigned int i;
     if (!page_map.bfind (map, &i, HB_NOT_FOUND_STORE_CLOSEST))
     {
@@ -890,6 +902,8 @@ struct hb_bit_set_t
 	       (page_map.length - 1 - i) * page_map.item_size);
       page_map[i] = map;
     }
+
+    last_page_lookup = i;
     return &pages[page_map[i].index];
   }
   const page_t *page_for (hb_codepoint_t g) const