Merge pull request #5042 from harfbuzz/aat-regression

[aat] Fix performance regression

src/hb-aat-layout-common.hh

@@ -66,6 +66,7 @@ struct hb_aat_apply_context_t :
   const ankr *ankr_table;
   const OT::GDEF *gdef_table;
   const hb_sorted_vector_t<hb_aat_map_t::range_flags_t> *range_flags = nullptr;
+  bool using_buffer_glyph_set = false;
   hb_bit_set_t buffer_glyph_set;
   const hb_bit_set_t *left_set = nullptr;
   const hb_bit_set_t *right_set = nullptr;
@@ -86,6 +87,30 @@ struct hb_aat_apply_context_t :
   HB_INTERNAL void set_ankr_table (const AAT::ankr *ankr_table_);
 
   void set_lookup_index (unsigned int i) { lookup_index = i; }
+
+#define HB_MALLOC_COST 48
+#define HB_BIT_SET_HAS_COST 8
+  void setup_buffer_glyph_set (unsigned subchain_count)
+  {
+    // Using buffer_glyph_set has at least two mallocs. Avoid it for small workloads.
+    unsigned malloced_cost = HB_MALLOC_COST * 2 + subchain_count * HB_BIT_SET_HAS_COST;
+    unsigned unmalloced_cost = subchain_count * buffer->len;
+    using_buffer_glyph_set = malloced_cost < unmalloced_cost;
+
+    if (using_buffer_glyph_set)
+      buffer_glyph_set = buffer->bit_set ();
+  }
+  bool buffer_intersects_machine () const
+  {
+    if (using_buffer_glyph_set)
+      return buffer_glyph_set.intersects (*machine_glyph_set);
+
+    // Faster for shorter buffers.
+    for (unsigned i = 0; i < buffer->len; i++)
+      if (machine_glyph_set->has (buffer->info[i].codepoint))
+	return true;
+    return false;
+  }
 };
 
 

src/hb-aat-layout-morx-table.hh

@@ -177,7 +177,7 @@ struct RearrangementSubtable
 
     StateTableDriver<Types, EntryData, Flags> driver (machine, c->face);
 
-    if (!c->buffer_glyph_set.intersects (*c->machine_glyph_set))
+    if (!c->buffer_intersects_machine ())
     {
       (void) c->buffer->message (c->font, "skipped chainsubtable because no glyph matches");
       return_trace (false);
@@ -347,7 +347,7 @@ struct ContextualSubtable
 
     StateTableDriver<Types, EntryData, Flags> driver (machine, c->face);
 
-    if (!c->buffer_glyph_set.intersects (*c->machine_glyph_set))
+    if (!c->buffer_intersects_machine ())
     {
       (void) c->buffer->message (c->font, "skipped chainsubtable because no glyph matches");
       return_trace (false);
@@ -623,7 +623,7 @@ struct LigatureSubtable
 
     StateTableDriver<Types, EntryData, Flags> driver (machine, c->face);
 
-    if (!c->buffer_glyph_set.intersects (*c->machine_glyph_set))
+    if (!c->buffer_intersects_machine ())
     {
       (void) c->buffer->message (c->font, "skipped chainsubtable because no glyph matches");
       return_trace (false);
@@ -664,8 +664,11 @@ struct NoncontextualSubtable
   {
     TRACE_APPLY (this);
 
-    if (!c->buffer_glyph_set.intersects (*c->machine_glyph_set))
+    if (!c->buffer_intersects_machine ())
+    {
+      (void) c->buffer->message (c->font, "skipped chainsubtable because no glyph matches");
       return_trace (false);
+    }
 
     const OT::GDEF &gdef (*c->gdef_table);
     bool has_glyph_classes = gdef.has_glyph_classes ();
@@ -917,7 +920,7 @@ struct InsertionSubtable
 
     StateTableDriver<Types, EntryData, Flags> driver (machine, c->face);
 
-    if (!c->buffer_glyph_set.intersects (*c->machine_glyph_set))
+    if (!c->buffer_intersects_machine ())
     {
       (void) c->buffer->message (c->font, "skipped chainsubtable because no glyph matches");
       return_trace (false);
@@ -1371,11 +1374,13 @@ struct mortmorx
       }
 
       this->chain_count = table->get_chain_count ();
+      this->subchain_count = table->get_subchain_count ();
 
       this->accels = (hb_atomic_ptr_t<hb_aat_layout_chain_accelerator_t> *) hb_calloc (this->chain_count, sizeof (*accels));
       if (unlikely (!this->accels))
       {
 	this->chain_count = 0;
+	this->subchain_count = 0;
 	this->table.destroy ();
 	this->table = hb_blob_get_empty ();
       }
@@ -1419,6 +1424,7 @@ struct mortmorx
 
     hb_blob_ptr_t<T> table;
     unsigned int chain_count;
+    unsigned int subchain_count;
     hb_atomic_ptr_t<hb_aat_layout_chain_accelerator_t> *accels;
   };
 
@@ -1443,6 +1449,18 @@ struct mortmorx
   {
     return chainCount;
   }
+  unsigned get_subchain_count () const
+  {
+    const Chain<Types> *chain = &firstChain;
+    unsigned int count = chainCount;
+    unsigned int subchain_count = 0;
+    for (unsigned int i = 0; i < count; i++)
+    {
+      subchain_count += chain->get_subtable_count ();
+      chain = &StructAfter<Chain<Types>> (*chain);
+    }
+    return subchain_count;
+  }
 
   void apply (hb_aat_apply_context_t *c,
 	      const hb_aat_map_t &map,
@@ -1452,7 +1470,7 @@ struct mortmorx
 
     c->buffer->unsafe_to_concat ();
 
-    c->buffer_glyph_set = c->buffer->bit_set ();
+    c->setup_buffer_glyph_set (accel.subchain_count);
 
     c->set_lookup_index (0);
     const Chain<Types> *chain = &firstChain;