From bbb4db90dd2f24b237c3bbcf6ab24389f970d1b8 Mon Sep 17 00:00:00 2001
From: Behdad Esfahbod <behdad@behdad.org>
Date: Thu, 21 Jul 2022 12:34:46 -0600
Subject: [PATCH] [Coverage/SingleSubst] Move hand-written loop to Coverage

---
 src/OT/Layout/Common/CoverageFormat2.hh  | 35 +++++++++++++++++++----
 src/OT/Layout/GSUB/SingleSubstFormat1.hh | 36 ++++--------------------
 2 files changed, 36 insertions(+), 35 deletions(-)

diff --git a/src/OT/Layout/Common/CoverageFormat2.hh b/src/OT/Layout/Common/CoverageFormat2.hh
index 675a552cd..7548f0b47 100644
--- a/src/OT/Layout/Common/CoverageFormat2.hh
+++ b/src/OT/Layout/Common/CoverageFormat2.hh
@@ -142,12 +142,37 @@ struct CoverageFormat2_4
 	    hb_requires (hb_is_sink_of (IterableOut, hb_codepoint_t))>
   void intersect_set (const hb_set_t &glyphs, IterableOut &intersect_glyphs) const
   {
-    for (const auto& range : rangeRecord)
+    /* Why the second branch is >2x faster I have no idea, but it seems to be. */
+    if (0)
     {
-      hb_codepoint_t last = range.last;
-      for (hb_codepoint_t g = range.first - 1;
-           glyphs.next (&g) && g <= last;)
-        intersect_glyphs << g;
+      for (const auto& range : rangeRecord)
+      {
+	hb_codepoint_t last = range.last;
+	for (hb_codepoint_t g = range.first - 1;
+	     glyphs.next (&g) && g <= last;)
+	  intersect_glyphs << g;
+      }
+    }
+    else
+    {
+      iter_t c_iter;
+      c_iter.init (*this);
+      auto s_iter = hb_iter (glyphs);
+      while (c_iter.__more__ () && s_iter)
+      {
+	hb_codepoint_t cv = c_iter.get_glyph ();
+	hb_codepoint_t sv = *s_iter;
+	if (cv == sv)
+	{
+	  intersect_glyphs << cv;
+	  c_iter.__next__ ();
+	  s_iter++;
+	}
+	else if (cv < sv)
+	  c_iter.__next__ ();
+	else
+	  s_iter++;
+      }
     }
   }
 
diff --git a/src/OT/Layout/GSUB/SingleSubstFormat1.hh b/src/OT/Layout/GSUB/SingleSubstFormat1.hh
index f47e03d3b..dd4459574 100644
--- a/src/OT/Layout/GSUB/SingleSubstFormat1.hh
+++ b/src/OT/Layout/GSUB/SingleSubstFormat1.hh
@@ -42,37 +42,13 @@ struct SingleSubstFormat1_3
     hb_codepoint_t d = deltaGlyphID;
     hb_codepoint_t mask = get_mask ();
 
-    if (1)
-    {
-      /* Somehow this branch is much faster. Can't understand why. */
-      auto c_iter = hb_iter (this+coverage);
-      auto s_iter = hb_iter (c->parent_active_glyphs ());
-      while (c_iter && s_iter)
-      {
-	hb_codepoint_t cv = *c_iter;
-	hb_codepoint_t sv = *s_iter;
-	if (cv == sv)
-	{
-	  c->output->add ((cv + d) & mask);
-	  c_iter++;
-	  s_iter++;
-	}
-	else if (cv < sv)
-	  c_iter++;
-	else
-	  s_iter++;
-      }
-    }
-    else
-    {
-      hb_set_t intersection;
-      (this+coverage).intersect_set (c->parent_active_glyphs (), intersection);
+    hb_set_t intersection;
+    (this+coverage).intersect_set (c->parent_active_glyphs (), intersection);
 
-      + hb_iter (intersection)
-      | hb_map ([d, mask] (hb_codepoint_t g) { return (g + d) & mask; })
-      | hb_sink (c->output)
-      ;
-    }
+    + hb_iter (intersection)
+    | hb_map ([d, mask] (hb_codepoint_t g) { return (g + d) & mask; })
+    | hb_sink (c->output)
+    ;
   }
 
   void closure_lookups (hb_closure_lookups_context_t *c) const {}