Call get_nominal_glyphs() for runs of simple clusters at a time

Even without FT or OT font funcs implementing get_nominal_glyphs(), there's measurable
speedup.

src/hb-buffer.hh

@@ -260,7 +260,8 @@ struct hb_buffer_t
   {
     if (have_output)
     {
-      if (unlikely (out_info != info || out_len != idx)) {
+      if (out_info != info || out_len != idx)
+      {
 	if (unlikely (!make_room_for (1, 1))) return;
 	out_info[out_len] = info[idx];
       }
@@ -269,6 +270,23 @@ struct hb_buffer_t
 
     idx++;
   }
+  /* Copies n glyphs at idx to output and advance idx.
+   * If there's no output, just advance idx. */
+  inline void
+  next_glyphs (unsigned int n)
+  {
+    if (have_output)
+    {
+      if (out_info != info || out_len != idx)
+      {
+	if (unlikely (!make_room_for (n, n))) return;
+	memmove (out_info + out_len, info + idx, n * sizeof (out_info[0]));
+      }
+      out_len += n;
+    }
+
+    idx += n;
+  }
   /* Advance idx without copying to output. */
   inline void skip_glyph (void)
   {

src/hb-ot-shape-normalize.cc

@@ -335,6 +335,15 @@ _hb_ot_shape_normalize (const hb_ot_shape_plan_t *plan,
 	end--; /* Leave one base for the marks to cluster with. */
 
       /* From idx to end are simple clusters. */
+      if (might_short_circuit)
+      {
+        unsigned int done = font->get_nominal_glyphs (end - buffer->idx,
+						      &buffer->cur().codepoint,
+						      sizeof (buffer->info[0]),
+						      &buffer->cur().glyph_index(),
+						      sizeof (buffer->info[0]));
+	buffer->next_glyphs (done);
+      }
       while (buffer->idx < end && buffer->successful)
 	decompose_current_character (&c, might_short_circuit);