Merge pull request #3787 from harfbuzz/split_mark_base_bos

[repacker] Add support for splitting MarkBasePosFormat1 in the packer.

src/Makefile.sources

@@ -351,10 +351,10 @@ HB_SUBSET_sources = \
 	graph/gsubgpos-graph.hh \
 	graph/gsubgpos-context.hh \
 	graph/gsubgpos-context.cc \
-	graph/pairpos-graph.hh \
 	graph/coverage-graph.hh \
 	graph/classdef-graph.hh \
 	graph/pairpos-graph.hh \
+	graph/markbasepos-graph.hh \
 	graph/split-helpers.hh \
 	graph/serialize.hh \
 	$(NULL)

src/OT/Layout/GPOS/MarkRecord.hh

@@ -9,7 +9,7 @@ struct MarkRecord
 {
   friend struct MarkArray;
 
-  protected:
+  public:
   HBUINT16      klass;                  /* Class defined for this mark */
   Offset16To<Anchor>
                 markAnchor;             /* Offset to Anchor table--from

src/graph/coverage-graph.hh

@@ -109,7 +109,7 @@ struct Coverage : public OT::Layout::Common::Coverage
   {
     char* buffer = (char*) hb_calloc (1, max_size);
     hb_serialize_context_t serializer (buffer, max_size);
-    Coverage_serialize (&serializer, glyphs);
+    OT::Layout::Common::Coverage_serialize (&serializer, glyphs);
     serializer.end_serialize ();
     if (serializer.in_error ())
     {

src/graph/graph.hh

@@ -49,6 +49,51 @@ struct graph_t
     unsigned end = 0;
     unsigned priority = 0;
 
+    void normalize ()
+    {
+      obj.real_links.qsort ();
+      for (auto& l : obj.real_links)
+      {
+        for (unsigned i = 0; i < l.width; i++)
+        {
+          obj.head[l.position + i] = 0;
+        }
+      }
+    }
+
+    bool equals (const vertex_t& other,
+                 const graph_t& graph,
+                 const graph_t& other_graph,
+                 unsigned depth) const
+    {
+      if (!(as_bytes () == other.as_bytes ()))
+      {
+        DEBUG_MSG (SUBSET_REPACK, nullptr,
+                   "vertex [%lu] bytes != [%lu] bytes, depth = %u",
+                   table_size (),
+                   other.table_size (),
+                   depth);
+
+        auto a = as_bytes ();
+        auto b = other.as_bytes ();
+        while (a || b)
+        {
+          DEBUG_MSG (SUBSET_REPACK, nullptr,
+                     "  0x%x %s 0x%x", *a, (*a == *b) ? "==" : "!=", *b);
+          a++;
+          b++;
+        }
+        return false;
+      }
+
+      return links_equal (obj.real_links, other.obj.real_links, graph, other_graph, depth);
+    }
+
+    hb_bytes_t as_bytes () const
+    {
+      return hb_bytes_t (obj.head, table_size ());
+    }
+
     friend void swap (vertex_t& a, vertex_t& b)
     {
       hb_swap (a.obj, b.obj);
@@ -60,6 +105,18 @@ struct graph_t
       hb_swap (a.priority, b.priority);
     }
 
+    hb_hashmap_t<unsigned, unsigned>
+    position_to_index_map () const
+    {
+      hb_hashmap_t<unsigned, unsigned> result;
+
+      for (const auto& l : obj.real_links) {
+        result.set (l.position, l.objidx);
+      }
+
+      return result;
+    }
+
     bool is_shared () const
     {
       return parents.length > 1;
@@ -84,7 +141,7 @@ struct graph_t
     {
       for (unsigned i = 0; i < obj.real_links.length; i++)
       {
-        auto& link = obj.real_links[i];
+        auto& link = obj.real_links.arrayZ[i];
         if (link.objidx != child_index)
           continue;
 
@@ -155,6 +212,57 @@ struct graph_t
 
       return -table_size;
     }
+
+   private:
+    bool links_equal (const hb_vector_t<hb_serialize_context_t::object_t::link_t>& this_links,
+                      const hb_vector_t<hb_serialize_context_t::object_t::link_t>& other_links,
+                      const graph_t& graph,
+                      const graph_t& other_graph,
+                      unsigned depth) const
+    {
+      auto a = this_links.iter ();
+      auto b = other_links.iter ();
+
+      while (a && b)
+      {
+        const auto& link_a = *a;
+        const auto& link_b = *b;
+
+        if (link_a.width != link_b.width ||
+            link_a.is_signed != link_b.is_signed ||
+            link_a.whence != link_b.whence ||
+            link_a.position != link_b.position ||
+            link_a.bias != link_b.bias)
+          return false;
+
+        if (!graph.vertices_[link_a.objidx].equals (
+                other_graph.vertices_[link_b.objidx], graph, other_graph, depth + 1))
+          return false;
+
+        a++;
+        b++;
+      }
+
+      if (bool (a) != bool (b))
+        return false;
+
+      return true;
+    }
+  };
+
+  template <typename T>
+  struct vertex_and_table_t
+  {
+    vertex_and_table_t () : index (0), vertex (nullptr), table (nullptr)
+    {}
+
+    unsigned index;
+    vertex_t* vertex;
+    T* table;
+
+    operator bool () {
+       return table && vertex;
+    }
   };
 
   /*
@@ -169,7 +277,8 @@ struct graph_t
       : parents_invalid (true),
         distance_invalid (true),
         positions_invalid (true),
-        successful (true)
+        successful (true),
+        buffers ()
   {
     num_roots_for_space_.push (1);
     bool removed_nil = false;
@@ -201,6 +310,20 @@ struct graph_t
   ~graph_t ()
   {
     vertices_.fini ();
+    for (char* b : buffers)
+      hb_free (b);
+  }
+
+  bool operator== (const graph_t& other) const
+  {
+    return root ().equals (other.root (), *this, other, 0);
+  }
+
+  // Sorts links of all objects in a consistent manner and zeroes all offsets.
+  void normalize ()
+  {
+    for (auto& v : vertices_.writer ())
+      v.normalize ();
   }
 
   bool in_error () const
@@ -228,6 +351,27 @@ struct graph_t
     return vertices_[i].obj;
   }
 
+  void add_buffer (char* buffer)
+  {
+    buffers.push (buffer);
+  }
+
+  /*
+   * Adds a 16 bit link from parent_id to child_id
+   */
+  template<typename T>
+  void add_link (T* offset,
+                 unsigned parent_id,
+                 unsigned child_id)
+  {
+    auto& v = vertices_[parent_id];
+    auto* link = v.obj.real_links.push ();
+    link->width = 2;
+    link->objidx = child_id;
+    link->position = (char*) offset - (char*) v.obj.head;
+    vertices_[child_id].parents.push (parent_id);
+  }
+
   /*
    * Generates a new topological sorting of graph ordered by the shortest
    * distance to each node if positions are marked as invalid.
@@ -345,6 +489,31 @@ struct graph_t
     }
   }
 
+  template <typename T, typename ...Ts>
+  vertex_and_table_t<T> as_table (unsigned parent, const void* offset, Ts... ds)
+  {
+    return as_table_from_index<T> (index_for_offset (parent, offset), std::forward<Ts>(ds)...);
+  }
+
+  template <typename T, typename ...Ts>
+  vertex_and_table_t<T> as_table_from_index (unsigned index, Ts... ds)
+  {
+    if (index >= vertices_.length)
+      return vertex_and_table_t<T> ();
+
+    vertex_and_table_t<T> r;
+    r.vertex = &vertices_[index];
+    r.table = (T*) r.vertex->obj.head;
+    r.index = index;
+    if (!r.table)
+      return vertex_and_table_t<T> ();
+
+    if (!r.table->sanitize (*(r.vertex), std::forward<Ts>(ds)...))
+      return vertex_and_table_t<T> ();
+
+    return r;
+  }
+
   // Finds the object id of the object pointed to by the offset at 'offset'
   // within object[node_idx].
   unsigned index_for_offset (unsigned node_idx, const void* offset) const
@@ -352,8 +521,11 @@ struct graph_t
     const auto& node = object (node_idx);
     if (offset < node.head || offset >= node.tail) return -1;
 
-    for (const auto& link : node.real_links)
+    unsigned length = node.real_links.length;
+    for (unsigned i = 0; i < length; i++)
     {
+      // Use direct access for increased performance, this is a hot method.
+      const auto& link = node.real_links.arrayZ[i];
       if (offset != node.head + link.position)
         continue;
       return link.objidx;
@@ -1059,6 +1231,7 @@ struct graph_t
   bool positions_invalid;
   bool successful;
   hb_vector_t<unsigned> num_roots_for_space_;
+  hb_vector_t<char*> buffers;
 };
 
 }

src/graph/gsubgpos-context.cc

@@ -33,8 +33,7 @@ gsubgpos_graph_context_t::gsubgpos_graph_context_t (hb_tag_t table_tag_,
     : table_tag (table_tag_),
       graph (graph_),
       lookup_list_index (0),
-      lookups (),
-      buffers ()
+      lookups ()
 {
   if (table_tag_ != HB_OT_TAG_GPOS
       &&  table_tag_ != HB_OT_TAG_GSUB)
@@ -53,7 +52,7 @@ unsigned gsubgpos_graph_context_t::create_node (unsigned size)
   if (!buffer)
     return -1;
 
-  buffers.push (buffer);
+  add_buffer (buffer);
 
   return graph.new_node (buffer, buffer + size);
 }

src/graph/gsubgpos-context.hh

@@ -40,22 +40,16 @@ struct gsubgpos_graph_context_t
   graph_t& graph;
   unsigned lookup_list_index;
   hb_hashmap_t<unsigned, graph::Lookup*> lookups;
-  hb_vector_t<char*> buffers;
+
 
   HB_INTERNAL gsubgpos_graph_context_t (hb_tag_t table_tag_,
                                         graph_t& graph_);
 
-  ~gsubgpos_graph_context_t ()
-  {
-    for (char* b : buffers)
-      hb_free (b);
-  }
-
   HB_INTERNAL unsigned create_node (unsigned size);
 
   void add_buffer (char* buffer)
   {
-    buffers.push (buffer);
+    graph.add_buffer (buffer);
   }
 
  private:

src/graph/gsubgpos-graph.hh

@@ -29,6 +29,7 @@
 #include "../OT/Layout/GSUB/ExtensionSubst.hh"
 #include "gsubgpos-context.hh"
 #include "pairpos-graph.hh"
+#include "markbasepos-graph.hh"
 
 #ifndef GRAPH_GSUBGPOS_GRAPH_HH
 #define GRAPH_GSUBGPOS_GRAPH_HH
@@ -121,7 +122,9 @@ struct Lookup : public OT::Lookup
     if (c.table_tag != HB_OT_TAG_GPOS)
       return true;
 
-    if (!is_ext && type != OT::Layout::GPOS_impl::PosLookupSubTable::Type::Pair)
+    if (!is_ext &&
+        type != OT::Layout::GPOS_impl::PosLookupSubTable::Type::Pair &&
+        type != OT::Layout::GPOS_impl::PosLookupSubTable::Type::MarkBase)
       return true;
 
     hb_vector_t<hb_pair_t<unsigned, hb_vector_t<unsigned>>> all_new_subtables;
@@ -138,15 +141,23 @@ struct Lookup : public OT::Lookup
 
         subtable_index = extension->get_subtable_index (c.graph, ext_subtable_index);
         type = extension->get_lookup_type ();
-        if (type != OT::Layout::GPOS_impl::PosLookupSubTable::Type::Pair)
+        if (type != OT::Layout::GPOS_impl::PosLookupSubTable::Type::Pair
+            && type != OT::Layout::GPOS_impl::PosLookupSubTable::Type::MarkBase)
           continue;
       }
 
-      PairPos* pairPos = (PairPos*) c.graph.object (subtable_index).head;
-      if (!pairPos || !pairPos->sanitize (c.graph.vertices_[subtable_index])) continue;
-
-      hb_vector_t<unsigned> new_sub_tables = pairPos->split_subtables (c, subtable_index);
+      hb_vector_t<unsigned> new_sub_tables;
+      switch (type)
+      {
+      case 2:
+        new_sub_tables = split_subtable<PairPos> (c, subtable_index); break;
+      case 4:
+        new_sub_tables = split_subtable<MarkBasePos> (c, subtable_index); break;
+      default:
+        break;
+      }
       if (new_sub_tables.in_error ()) return false;
+      if (!new_sub_tables) continue;
       hb_pair_t<unsigned, hb_vector_t<unsigned>>* entry = all_new_subtables.push ();
       entry->first = i;
       entry->second = std::move (new_sub_tables);
@@ -159,6 +170,17 @@ struct Lookup : public OT::Lookup
     return true;
   }
 
+  template<typename T>
+  hb_vector_t<unsigned> split_subtable (gsubgpos_graph_context_t& c,
+                                        unsigned objidx)
+  {
+    T* sub_table = (T*) c.graph.object (objidx).head;
+    if (!sub_table || !sub_table->sanitize (c.graph.vertices_[objidx]))
+      return hb_vector_t<unsigned> ();
+
+    return sub_table->split_subtables (c, objidx);
+  }
+
   void add_sub_tables (gsubgpos_graph_context_t& c,
                        unsigned this_index,
                        unsigned type,

src/graph/markbasepos-graph.hh

@@ -0,0 +1,507 @@
+/*
+ * Copyright © 2022  Google, Inc.
+ *
+ *  This is part of HarfBuzz, a text shaping library.
+ *
+ * Permission is hereby granted, without written agreement and without
+ * license or royalty fees, to use, copy, modify, and distribute this
+ * software and its documentation for any purpose, provided that the
+ * above copyright notice and the following two paragraphs appear in
+ * all copies of this software.
+ *
+ * IN NO EVENT SHALL THE COPYRIGHT HOLDER BE LIABLE TO ANY PARTY FOR
+ * DIRECT, INDIRECT, SPECIAL, INCIDENTAL, OR CONSEQUENTIAL DAMAGES
+ * ARISING OUT OF THE USE OF THIS SOFTWARE AND ITS DOCUMENTATION, EVEN
+ * IF THE COPYRIGHT HOLDER HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH
+ * DAMAGE.
+ *
+ * THE COPYRIGHT HOLDER SPECIFICALLY DISCLAIMS ANY WARRANTIES, INCLUDING,
+ * BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
+ * FITNESS FOR A PARTICULAR PURPOSE.  THE SOFTWARE PROVIDED HEREUNDER IS
+ * ON AN "AS IS" BASIS, AND THE COPYRIGHT HOLDER HAS NO OBLIGATION TO
+ * PROVIDE MAINTENANCE, SUPPORT, UPDATES, ENHANCEMENTS, OR MODIFICATIONS.
+ *
+ * Google Author(s): Garret Rieger
+ */
+
+#ifndef GRAPH_MARKBASEPOS_GRAPH_HH
+#define GRAPH_MARKBASEPOS_GRAPH_HH
+
+#include "split-helpers.hh"
+#include "coverage-graph.hh"
+#include "../OT/Layout/GPOS/MarkBasePos.hh"
+#include "../OT/Layout/GPOS/PosLookupSubTable.hh"
+
+namespace graph {
+
+struct AnchorMatrix : public OT::Layout::GPOS_impl::AnchorMatrix
+{
+  bool sanitize (graph_t::vertex_t& vertex, unsigned class_count) const
+  {
+    int64_t vertex_len = vertex.obj.tail - vertex.obj.head;
+    if (vertex_len < AnchorMatrix::min_size) return false;
+
+    return vertex_len >= AnchorMatrix::min_size +
+        OT::Offset16::static_size * class_count * this->rows;
+  }
+
+  bool shrink (gsubgpos_graph_context_t& c,
+               unsigned this_index,
+               unsigned old_class_count,
+               unsigned new_class_count)
+  {
+    if (new_class_count >= old_class_count) return false;
+    auto& o = c.graph.vertices_[this_index].obj;
+    unsigned base_count = rows;
+    o.tail = o.head +
+             AnchorMatrix::min_size +
+             OT::Offset16::static_size * base_count * new_class_count;
+
+    // Reposition links into the new indexing scheme.
+    for (auto& link : o.real_links.writer ())
+    {
+      unsigned index = (link.position - 2) / 2;
+      unsigned base = index / old_class_count;
+      unsigned klass = index % old_class_count;
+      if (klass >= new_class_count)
+        // should have already been removed
+        return false;
+
+      unsigned new_index = base * new_class_count + klass;
+
+      link.position = (char*) &(this->matrixZ[new_index]) - (char*) this;
+    }
+
+    return true;
+  }
+
+  unsigned clone (gsubgpos_graph_context_t& c,
+                  unsigned this_index,
+                  unsigned start,
+                  unsigned end,
+                  unsigned class_count)
+  {
+    unsigned base_count = rows;
+    unsigned new_class_count = end - start;
+    unsigned size = AnchorMatrix::min_size +
+                    OT::Offset16::static_size * new_class_count * rows;
+    unsigned prime_id = c.create_node (size);
+    if (prime_id == (unsigned) -1) return -1;
+    AnchorMatrix* prime = (AnchorMatrix*) c.graph.object (prime_id).head;
+    prime->rows = base_count;
+
+    auto& o = c.graph.vertices_[this_index].obj;
+    int num_links = o.real_links.length;
+    for (int i = 0; i < num_links; i++)
+    {
+      const auto& link = o.real_links[i];
+      unsigned old_index = (link.position - 2) / OT::Offset16::static_size;
+      unsigned klass = old_index % class_count;
+      if (klass < start || klass >= end) continue;
+
+      unsigned base = old_index / class_count;
+      unsigned new_klass = klass - start;
+      unsigned new_index = base * new_class_count + new_klass;
+
+
+      unsigned child_idx = link.objidx;
+      c.graph.add_link (&(prime->matrixZ[new_index]),
+                        prime_id,
+                        child_idx);
+
+      auto& child = c.graph.vertices_[child_idx];
+      child.remove_parent (this_index);
+
+      o.real_links.remove (i);
+      num_links--;
+      i--;
+    }
+
+    return prime_id;
+  }
+};
+
+struct MarkArray : public OT::Layout::GPOS_impl::MarkArray
+{
+  bool sanitize (graph_t::vertex_t& vertex) const
+  {
+    int64_t vertex_len = vertex.obj.tail - vertex.obj.head;
+    unsigned min_size = MarkArray::min_size;
+    if (vertex_len < min_size) return false;
+
+    return vertex_len >= get_size ();
+  }
+
+  bool shrink (gsubgpos_graph_context_t& c,
+               const hb_hashmap_t<unsigned, unsigned>& mark_array_links,
+               unsigned this_index,
+               unsigned new_class_count)
+  {
+    auto& o = c.graph.vertices_[this_index].obj;
+    for (const auto& link : o.real_links)
+      c.graph.vertices_[link.objidx].remove_parent (this_index);
+    o.real_links.reset ();
+
+    unsigned new_index = 0;
+    for (const auto& record : this->iter ())
+    {
+      unsigned klass = record.klass;
+      if (klass >= new_class_count) continue;
+
+      (*this)[new_index].klass = klass;
+      unsigned position = (char*) &record.markAnchor - (char*) this;
+      unsigned* objidx;
+      if (!mark_array_links.has (position, &objidx))
+      {
+        new_index++;
+        continue;
+      }
+
+      c.graph.add_link (&(*this)[new_index].markAnchor, this_index, *objidx);
+      new_index++;
+    }
+
+    this->len = new_index;
+    o.tail = o.head + MarkArray::min_size +
+             OT::Layout::GPOS_impl::MarkRecord::static_size * new_index;
+    return true;
+  }
+
+  unsigned clone (gsubgpos_graph_context_t& c,
+                  unsigned this_index,
+                  const hb_hashmap_t<unsigned, unsigned>& pos_to_index,
+                  hb_set_t& marks,
+                  unsigned start_class)
+  {
+    unsigned size = MarkArray::min_size +
+                    OT::Layout::GPOS_impl::MarkRecord::static_size *
+                    marks.get_population ();
+    unsigned prime_id = c.create_node (size);
+    if (prime_id == (unsigned) -1) return -1;
+    MarkArray* prime = (MarkArray*) c.graph.object (prime_id).head;
+    prime->len = marks.get_population ();
+
+
+    unsigned i = 0;
+    for (hb_codepoint_t mark : marks)
+    {
+      (*prime)[i].klass = (*this)[mark].klass - start_class;
+      unsigned offset_pos = (char*) &((*this)[mark].markAnchor) - (char*) this;
+      unsigned* anchor_index;
+      if (pos_to_index.has (offset_pos, &anchor_index))
+        c.graph.move_child (this_index,
+                            &((*this)[mark].markAnchor),
+                            prime_id,
+                            &((*prime)[i].markAnchor));
+
+      i++;
+    }
+
+    return prime_id;
+  }
+};
+
+struct MarkBasePosFormat1 : public OT::Layout::GPOS_impl::MarkBasePosFormat1_2<SmallTypes>
+{
+  bool sanitize (graph_t::vertex_t& vertex) const
+  {
+    int64_t vertex_len = vertex.obj.tail - vertex.obj.head;
+    return vertex_len >= MarkBasePosFormat1::static_size;
+  }
+
+  hb_vector_t<unsigned> split_subtables (gsubgpos_graph_context_t& c, unsigned this_index)
+  {
+    hb_set_t visited;
+
+    const unsigned base_coverage_id = c.graph.index_for_offset (this_index, &baseCoverage);
+    const unsigned base_size =
+        OT::Layout::GPOS_impl::PairPosFormat1_3<SmallTypes>::min_size +
+        MarkArray::min_size +
+        AnchorMatrix::min_size +
+        c.graph.vertices_[base_coverage_id].table_size ();
+
+    hb_vector_t<class_info_t> class_to_info = get_class_info (c, this_index);
+
+    unsigned class_count = classCount;
+    auto base_array = c.graph.as_table<AnchorMatrix> (this_index,
+                                                      &baseArray,
+                                                      class_count);
+    if (!base_array) return hb_vector_t<unsigned> ();
+    unsigned base_count = base_array.table->rows;
+
+    unsigned partial_coverage_size = 4;
+    unsigned accumulated = base_size;
+    hb_vector_t<unsigned> split_points;
+
+    for (unsigned klass = 0; klass < class_count; klass++)
+    {
+      class_info_t& info = class_to_info[klass];
+      partial_coverage_size += OT::HBUINT16::static_size * info.marks.get_population ();
+      unsigned accumulated_delta =
+          OT::Layout::GPOS_impl::MarkRecord::static_size * info.marks.get_population () +
+          OT::Offset16::static_size * base_count;
+
+      for (unsigned objidx : info.child_indices)
+        accumulated_delta += c.graph.find_subgraph_size (objidx, visited);
+
+      accumulated += accumulated_delta;
+      unsigned total = accumulated + partial_coverage_size;
+
+      if (total >= (1 << 16))
+      {
+        split_points.push (klass);
+        accumulated = base_size + accumulated_delta;
+        partial_coverage_size = 4 + OT::HBUINT16::static_size * info.marks.get_population ();
+        visited.clear (); // node sharing isn't allowed between splits.
+      }
+    }
+
+
+    const unsigned mark_array_id = c.graph.index_for_offset (this_index, &markArray);
+    split_context_t split_context {
+      c,
+      this,
+      this_index,
+      std::move (class_to_info),
+      c.graph.vertices_[mark_array_id].position_to_index_map (),
+    };
+
+    return actuate_subtable_split<split_context_t> (split_context, split_points);
+  }
+
+ private:
+
+  struct class_info_t {
+    hb_set_t marks;
+    hb_vector_t<unsigned> child_indices;
+  };
+
+  struct split_context_t {
+    gsubgpos_graph_context_t& c;
+    MarkBasePosFormat1* thiz;
+    unsigned this_index;
+    hb_vector_t<class_info_t> class_to_info;
+    hb_hashmap_t<unsigned, unsigned> mark_array_links;
+
+    hb_set_t marks_for (unsigned start, unsigned end)
+    {
+      hb_set_t marks;
+      for (unsigned klass = start; klass < end; klass++)
+      {
+        + class_to_info[klass].marks.iter ()
+        | hb_sink (marks)
+        ;
+      }
+      return marks;
+    }
+
+    unsigned original_count ()
+    {
+      return thiz->classCount;
+    }
+
+    unsigned clone_range (unsigned start, unsigned end)
+    {
+      return thiz->clone_range (*this, this->this_index, start, end);
+    }
+
+    bool shrink (unsigned count)
+    {
+      return thiz->shrink (*this, this->this_index, count);
+    }
+  };
+
+  hb_vector_t<class_info_t> get_class_info (gsubgpos_graph_context_t& c,
+                                            unsigned this_index)
+  {
+    hb_vector_t<class_info_t> class_to_info;
+
+    unsigned class_count= classCount;
+    class_to_info.resize (class_count);
+
+    auto mark_array = c.graph.as_table<MarkArray> (this_index, &markArray);
+    if (!mark_array) return hb_vector_t<class_info_t> ();
+    unsigned mark_count = mark_array.table->len;
+    for (unsigned mark = 0; mark < mark_count; mark++)
+    {
+      unsigned klass = (*mark_array.table)[mark].get_class ();
+      class_to_info[klass].marks.add (mark);
+    }
+
+    for (const auto& link : mark_array.vertex->obj.real_links)
+    {
+      unsigned mark = (link.position - 2) /
+                     OT::Layout::GPOS_impl::MarkRecord::static_size;
+      unsigned klass = (*mark_array.table)[mark].get_class ();
+      class_to_info[klass].child_indices.push (link.objidx);
+    }
+
+    unsigned base_array_id =
+        c.graph.index_for_offset (this_index, &baseArray);
+    auto& base_array_v = c.graph.vertices_[base_array_id];
+
+    for (const auto& link : base_array_v.obj.real_links)
+    {
+      unsigned index = (link.position - 2) / OT::Offset16::static_size;
+      unsigned klass = index % class_count;
+      class_to_info[klass].child_indices.push (link.objidx);
+    }
+
+    return class_to_info;
+  }
+
+  bool shrink (split_context_t& sc,
+               unsigned this_index,
+               unsigned count)
+  {
+    DEBUG_MSG (SUBSET_REPACK, nullptr,
+               "  Shrinking MarkBasePosFormat1 (%u) to [0, %u).",
+               this_index,
+               count);
+
+    unsigned old_count = classCount;
+    if (count >= old_count)
+      return true;
+
+    classCount = count;
+
+    auto mark_coverage = sc.c.graph.as_table<Coverage> (this_index,
+                                                        &markCoverage);
+    if (!mark_coverage) return false;
+    hb_set_t marks = sc.marks_for (0, count);
+    auto new_coverage =
+        + hb_zip (hb_range (), mark_coverage.table->iter ())
+        | hb_filter (marks, hb_first)
+        | hb_map_retains_sorting (hb_second)
+        ;
+    if (!Coverage::make_coverage (sc.c, + new_coverage,
+                                  mark_coverage.index,
+                                  4 + 2 * marks.get_population ()))
+      return false;
+
+
+    auto base_array = sc.c.graph.as_table<AnchorMatrix> (this_index,
+                                                         &baseArray,
+                                                         old_count);
+    if (!base_array || !base_array.table->shrink (sc.c,
+                                                  base_array.index,
+                                                  old_count,
+                                                  count))
+      return false;
+
+    auto mark_array = sc.c.graph.as_table<MarkArray> (this_index,
+                                                      &markArray);
+    if (!mark_array || !mark_array.table->shrink (sc.c,
+                                                  sc.mark_array_links,
+                                                  mark_array.index,
+                                                  count))
+      return false;
+
+    return true;
+  }
+
+  // Create a new MarkBasePos that has all of the data for classes from [start, end).
+  unsigned clone_range (split_context_t& sc,
+                        unsigned this_index,
+                        unsigned start, unsigned end) const
+  {
+    DEBUG_MSG (SUBSET_REPACK, nullptr,
+               "  Cloning MarkBasePosFormat1 (%u) range [%u, %u).", this_index, start, end);
+
+    graph_t& graph = sc.c.graph;
+    unsigned prime_size = OT::Layout::GPOS_impl::MarkBasePosFormat1_2<SmallTypes>::static_size;
+
+    unsigned prime_id = sc.c.create_node (prime_size);
+    if (prime_id == (unsigned) -1) return -1;
+
+    MarkBasePosFormat1* prime = (MarkBasePosFormat1*) graph.object (prime_id).head;
+    prime->format = this->format;
+    unsigned new_class_count = end - start;
+    prime->classCount = new_class_count;
+
+    unsigned base_coverage_id =
+        graph.index_for_offset (sc.this_index, &baseCoverage);
+    graph.add_link (&(prime->baseCoverage), prime_id, base_coverage_id);
+    graph.duplicate (prime_id, base_coverage_id);
+
+    auto mark_coverage = sc.c.graph.as_table<Coverage> (this_index,
+                                                        &markCoverage);
+    if (!mark_coverage) return false;
+    hb_set_t marks = sc.marks_for (start, end);
+    auto new_coverage =
+        + hb_zip (hb_range (), mark_coverage.table->iter ())
+        | hb_filter (marks, hb_first)
+        | hb_map_retains_sorting (hb_second)
+        ;
+    if (!Coverage::add_coverage (sc.c,
+                                 prime_id,
+                                 2,
+                                 + new_coverage,
+                                 marks.get_population () * 2 + 4))
+      return -1;
+
+    auto mark_array =
+        graph.as_table <MarkArray> (sc.this_index, &markArray);
+    if (!mark_array) return -1;
+    unsigned new_mark_array =
+        mark_array.table->clone (sc.c,
+                                 mark_array.index,
+                                 sc.mark_array_links,
+                                 marks,
+                                 start);
+    graph.add_link (&(prime->markArray), prime_id, new_mark_array);
+
+    unsigned class_count = classCount;
+    auto base_array =
+        graph.as_table<AnchorMatrix> (sc.this_index, &baseArray, class_count);
+    if (!base_array) return -1;
+    unsigned new_base_array =
+        base_array.table->clone (sc.c,
+                                 base_array.index,
+                                 start, end, this->classCount);
+    graph.add_link (&(prime->baseArray), prime_id, new_base_array);
+
+    return prime_id;
+  }
+};
+
+
+struct MarkBasePos : public OT::Layout::GPOS_impl::MarkBasePos
+{
+  hb_vector_t<unsigned> split_subtables (gsubgpos_graph_context_t& c,
+                                         unsigned this_index)
+  {
+    switch (u.format) {
+    case 1:
+      return ((MarkBasePosFormat1*)(&u.format1))->split_subtables (c, this_index);
+#ifndef HB_NO_BORING_EXPANSION
+    case 2: HB_FALLTHROUGH;
+      // Don't split 24bit PairPos's.
+#endif
+    default:
+      return hb_vector_t<unsigned> ();
+    }
+  }
+
+  bool sanitize (graph_t::vertex_t& vertex) const
+  {
+    int64_t vertex_len = vertex.obj.tail - vertex.obj.head;
+    if (vertex_len < u.format.get_size ()) return false;
+
+    switch (u.format) {
+    case 1:
+      return ((MarkBasePosFormat1*)(&u.format1))->sanitize (vertex);
+#ifndef HB_NO_BORING_EXPANSION
+    case 2: HB_FALLTHROUGH;
+#endif
+    default:
+      // We don't handle format 3 and 4 here.
+      return false;
+    }
+  }
+};
+
+
+}
+
+#endif  // GRAPH_MARKBASEPOS_GRAPH_HH

src/graph/pairpos-graph.hh

@@ -548,14 +548,8 @@ struct PairPosFormat2 : public OT::Layout::GPOS_impl::PairPosFormat2_4<SmallType
   get_all_device_tables (gsubgpos_graph_context_t& c,
                          unsigned this_index) const
   {
-    hb_hashmap_t<unsigned, unsigned> result;
-
-    const auto& o = c.graph.object (this_index);
-    for (const auto& l : o.real_links) {
-      result.set (l.position, l.objidx);
-    }
-
-    return result;
+    const auto& v = c.graph.vertices_[this_index];
+    return v.position_to_index_map ();
   }
 
   const Coverage* get_coverage (gsubgpos_graph_context_t& c,

src/hb-repacker.hh

@@ -276,33 +276,17 @@ bool _process_overflows (const hb_vector_t<graph::overflow_record_t>& overflows,
   return resolution_attempted;
 }
 
-/*
- * Attempts to modify the topological sorting of the provided object graph to
- * eliminate offset overflows in the links between objects of the graph. If a
- * non-overflowing ordering is found the updated graph is serialized it into the
- * provided serialization context.
- *
- * If necessary the structure of the graph may be modified in ways that do not
- * affect the functionality of the graph. For example shared objects may be
- * duplicated.
- *
- * For a detailed writeup describing how the algorithm operates see:
- * docs/repacker.md
- */
-template<typename T>
-inline hb_blob_t*
-hb_resolve_overflows (const T& packed,
-                      hb_tag_t table_tag,
-                      unsigned max_rounds = 20,
-                      bool recalculate_extensions = false) {
-  graph_t sorted_graph (packed);
+inline bool
+hb_resolve_graph_overflows (hb_tag_t table_tag,
+                            unsigned max_rounds ,
+                            bool recalculate_extensions,
+                            graph_t& sorted_graph /* IN/OUT */)
+{
   sorted_graph.sort_shortest_distance ();
 
   bool will_overflow = graph::will_overflow (sorted_graph);
   if (!will_overflow)
-  {
-    return graph::serialize (sorted_graph);
-  }
+    return true;
 
   graph::gsubgpos_graph_context_t ext_context (table_tag, sorted_graph);
   if ((table_tag == HB_OT_TAG_GPOS
@@ -314,13 +298,13 @@ hb_resolve_overflows (const T& packed,
       DEBUG_MSG (SUBSET_REPACK, nullptr, "Splitting subtables if needed.");
       if (!_presplit_subtables_if_needed (ext_context)) {
         DEBUG_MSG (SUBSET_REPACK, nullptr, "Subtable splitting failed.");
-        return nullptr;
+        return false;
       }
 
       DEBUG_MSG (SUBSET_REPACK, nullptr, "Promoting lookups to extensions if needed.");
       if (!_promote_extensions_if_needed (ext_context)) {
         DEBUG_MSG (SUBSET_REPACK, nullptr, "Extensions promotion failed.");
-        return nullptr;
+        return false;
       }
     }
 
@@ -360,15 +344,41 @@ hb_resolve_overflows (const T& packed,
   if (sorted_graph.in_error ())
   {
     DEBUG_MSG (SUBSET_REPACK, nullptr, "Sorted graph in error state.");
-    return nullptr;
+    return false;
   }
 
   if (graph::will_overflow (sorted_graph))
   {
     DEBUG_MSG (SUBSET_REPACK, nullptr, "Offset overflow resolution failed.");
-    return nullptr;
+    return false;
   }
 
+  return true;
+}
+
+/*
+ * Attempts to modify the topological sorting of the provided object graph to
+ * eliminate offset overflows in the links between objects of the graph. If a
+ * non-overflowing ordering is found the updated graph is serialized it into the
+ * provided serialization context.
+ *
+ * If necessary the structure of the graph may be modified in ways that do not
+ * affect the functionality of the graph. For example shared objects may be
+ * duplicated.
+ *
+ * For a detailed writeup describing how the algorithm operates see:
+ * docs/repacker.md
+ */
+template<typename T>
+inline hb_blob_t*
+hb_resolve_overflows (const T& packed,
+                      hb_tag_t table_tag,
+                      unsigned max_rounds = 20,
+                      bool recalculate_extensions = false) {
+  graph_t sorted_graph (packed);
+  if (!hb_resolve_graph_overflows (table_tag, max_rounds, recalculate_extensions, sorted_graph))
+    return nullptr;
+
   return graph::serialize (sorted_graph);
 }
 

src/hb-serialize.hh

@@ -142,7 +142,10 @@ struct hb_serialize_context_t
 
       HB_INTERNAL static int cmp (const void* a, const void* b)
       {
-        return ((const link_t*)a)->position - ((const link_t*)b)->position;
+        int cmp = ((const link_t*)a)->position - ((const link_t*)b)->position;
+        if (cmp) return cmp;
+
+        return ((const link_t*)a)->objidx - ((const link_t*)b)->objidx;
       }
     };
 

src/meson.build

@@ -349,6 +349,7 @@ hb_subset_sources = files(
   'graph/gsubgpos-context.hh',
   'graph/gsubgpos-graph.hh',
   'graph/pairpos-graph.hh',
+  'graph/markbasepos-graph.hh',
   'graph/coverage-graph.hh',
   'graph/classdef-graph.hh',
   'graph/split-helpers.hh',

src/test-repacker.cc

@@ -177,6 +177,16 @@ static unsigned add_coverage (unsigned start, unsigned end,
   return add_object ((char*) coverage, 10, c);
 }
 
+
+template<typename It>
+static unsigned add_coverage (It it,
+                              hb_serialize_context_t* c)
+{
+  c->push ();
+  OT::Layout::Common::Coverage_serialize (c, it);
+  return c->pop_pack (false);
+}
+
 // Adds a class that maps glyphs from [start_glyph, end_glyph)
 // to classes 1...n
 static unsigned add_class_def (uint16_t start_glyph,
@@ -297,56 +307,173 @@ static unsigned add_pair_pos_2 (unsigned starting_class,
   return c->pop_pack (false);
 }
 
-
-static void run_resolve_overflow_test (const char* name,
-                                       hb_serialize_context_t& overflowing,
-                                       hb_serialize_context_t& expected,
-                                       unsigned num_iterations = 0,
-                                       bool recalculate_extensions = false,
-                                       hb_tag_t tag = HB_TAG ('G', 'S', 'U', 'B'))
+static unsigned add_mark_base_pos_1 (unsigned mark_coverage,
+                                     unsigned base_coverage,
+                                     unsigned mark_array,
+                                     unsigned base_array,
+                                     unsigned class_count,
+                                     hb_serialize_context_t* c)
 {
-  printf (">>> Testing overflowing resolution for %s\n",
-          name);
+  uint8_t format[] = {
+    0, 1
+  };
 
-  graph_t graph (overflowing.object_graph ());
+  start_object ((char*) format, 2, c);
+  add_offset (mark_coverage, c);
+  add_offset (base_coverage, c);
 
+  uint8_t count[] = {
+    (uint8_t) ((class_count >> 8) & 0xFF),
+    (uint8_t) (class_count & 0xFF),
+  };
+  extend ((char*) count, 2, c);
 
-  assert (overflowing.offset_overflow ());
-  hb_blob_t* out = hb_resolve_overflows (overflowing.object_graph (),
-                                         tag,
-                                         num_iterations,
-                                         recalculate_extensions);
-  assert (out);
+  add_offset (mark_array, c);
+  add_offset (base_array, c);
 
-  hb_bytes_t result = out->as_bytes ();
+  return c->pop_pack (false);
+}
 
-  assert (!expected.offset_overflow ());
-  hb_bytes_t expected_result = expected.copy_bytes ();
+template<int mark_count,
+    int class_count,
+    int base_count,
+    int table_count>
+struct MarkBasePosBuffers
+{
+  unsigned base_anchors[class_count * base_count];
+  unsigned mark_anchors[mark_count];
+  uint8_t anchor_buffers[class_count * base_count + 100];
+  uint8_t class_buffer[class_count * 2];
 
-  if (result.length != expected_result.length)
+  MarkBasePosBuffers(hb_serialize_context_t* c)
   {
-    printf("result.length (%u) != expected.length (%u).\n",
-           result.length,
-           expected_result.length);
+    for (unsigned i = 0; i < sizeof(anchor_buffers) / 2; i++)
+    {
+      OT::HBUINT16* value = (OT::HBUINT16*) (&anchor_buffers[2*i]);
+      *value = i;
+    }
+
+    for (unsigned i = 0; i < class_count * base_count; i++)
+    {
+      base_anchors[i] = add_object ((char*) &anchor_buffers[i], 100, c);
+      if (i < class_count) {
+        class_buffer[i*2] = (uint8_t) ((i >> 8) & 0xFF);
+        class_buffer[i*2 + 1] = (uint8_t) (i & 0xFF);
+      }
+    }
+
+    for (unsigned i = 0; i < mark_count; i++)
+    {
+      mark_anchors[i] = add_object ((char*) &anchor_buffers[i], 4, c);
+    }
   }
-  assert (result.length == expected_result.length);
 
-  bool equal = true;
-  for (unsigned i = 0; i < expected_result.length; i++)
+  unsigned create_mark_base_pos_1 (unsigned table_index, hb_serialize_context_t* c)
   {
-    if (result[i] != expected_result[i])
+    unsigned class_per_table = class_count / table_count;
+    unsigned mark_per_class = mark_count / class_count;
+    unsigned start_class = class_per_table * table_index;
+    unsigned end_class = class_per_table * (table_index + 1) - 1;
+
+    // baseArray
+    uint8_t base_count_buffer[] = {
+      (uint8_t) ((base_count >> 8) & 0xFF),
+      (uint8_t) (base_count & 0xFF),
+
+    };
+    start_object ((char*) base_count_buffer, 2, c);
+    for (unsigned base = 0; base < base_count; base++)
     {
-      equal = false;
-      uint8_t a = result[i];
-      uint8_t b = expected_result[i];
-      printf("%08u: %x != %x\n", i, a, b);
+      for (unsigned klass = start_class; klass <= end_class; klass++)
+      {
+        unsigned i = base * class_count + klass;
+        add_offset (base_anchors[i], c);
+      }
     }
+    unsigned base_array = c->pop_pack (false);
+
+    // markArray
+    unsigned num_marks = class_per_table * mark_per_class;
+    uint8_t mark_count_buffer[] = {
+      (uint8_t) ((num_marks >> 8) & 0xFF),
+      (uint8_t) (num_marks & 0xFF),
+    };
+    start_object ((char*) mark_count_buffer, 2, c);
+    for (unsigned mark = 0; mark < mark_count; mark++)
+    {
+      unsigned klass = mark % class_count;
+      if (klass < start_class || klass > end_class) continue;
+      klass -= start_class;
+
+      extend ((char*) &class_buffer[2 * klass], 2, c);
+      add_offset (mark_anchors[mark], c);
+    }
+    unsigned mark_array = c->pop_pack (false);
+
+    // markCoverage
+    auto it =
+        + hb_range ((hb_codepoint_t) mark_count)
+        | hb_filter ([&] (hb_codepoint_t mark) {
+          unsigned klass = mark % class_count;
+          return klass >= class_per_table * table_index &&
+              klass < class_per_table * (table_index + 1);
+        })
+        ;
+    unsigned mark_coverage = add_coverage (it, c);
+
+    // baseCoverage
+    unsigned base_coverage = add_coverage (10, 10 + base_count - 1, c);
+
+    return add_mark_base_pos_1 (mark_coverage,
+                                base_coverage,
+                                mark_array,
+                                base_array,
+                                class_per_table,
+                                c);
   }
+};
+
 
-  assert (equal);
 
-  expected_result.fini ();
+
+
+static void run_resolve_overflow_test (const char* name,
+                                       hb_serialize_context_t& overflowing,
+                                       hb_serialize_context_t& expected,
+                                       unsigned num_iterations = 0,
+                                       bool recalculate_extensions = false,
+                                       hb_tag_t tag = HB_TAG ('G', 'S', 'U', 'B'))
+{
+  printf (">>> Testing overflowing resolution for %s\n",
+          name);
+
+  graph_t graph (overflowing.object_graph ());
+  graph_t expected_graph (expected.object_graph ());
+  if (graph::will_overflow (expected_graph))
+  {
+    expected_graph.assign_spaces ();
+    expected_graph.sort_shortest_distance ();
+  }
+
+  // Check that overflow resolution succeeds
+  assert (overflowing.offset_overflow ());
+  assert (hb_resolve_graph_overflows (tag,
+                                      num_iterations,
+                                      recalculate_extensions,
+                                      graph));
+
+  // Check the graphs can be serialized.
+  hb_blob_t* out = graph::serialize (graph);
+  assert (out);
   hb_blob_destroy (out);
+  out = graph::serialize (expected_graph);
+  assert (out);
+  hb_blob_destroy (out);
+
+  // Check the graphs are equivalent
+  graph.normalize ();
+  expected_graph.normalize ();
+  assert (graph == expected_graph);
 }
 
 static void add_virtual_offset (unsigned id,
@@ -1308,6 +1435,38 @@ populate_serializer_with_large_pair_pos_2 (hb_serialize_context_t* c,
   free (device_tables);
 }
 
+template<int mark_count,
+    int class_count,
+    int base_count,
+    int table_count>
+static void
+populate_serializer_with_large_mark_base_pos_1 (hb_serialize_context_t* c)
+{
+  c->start_serialize<char> ();
+
+  MarkBasePosBuffers<mark_count, class_count, base_count, table_count> buffers (c);
+
+  unsigned mark_base_pos[table_count];
+  for (unsigned i = 0; i < table_count; i++)
+    mark_base_pos[i] = buffers.create_mark_base_pos_1 (i, c);
+
+  for (int i = 0; i < table_count; i++)
+    mark_base_pos[i] = add_extension (mark_base_pos[i], 4, c);
+
+  start_lookup (9, table_count, c);
+
+  for (int i = 0; i < table_count; i++)
+    add_offset (mark_base_pos[i], c);
+
+  unsigned lookup = finish_lookup (c);
+
+  unsigned lookup_list = add_lookup_list (&lookup, 1, c);
+
+  add_gsubgpos_header (lookup_list, c);
+
+  c->end_serialize();
+}
+
 static void test_sort_shortest ()
 {
   size_t buffer_size = 100;
@@ -1792,6 +1951,29 @@ static void test_resolve_with_pair_pos_2_split_with_device_tables ()
   free (expected_buffer);
 }
 
+static void test_resolve_with_basic_mark_base_pos_1_split ()
+{
+  size_t buffer_size = 200000;
+  void* buffer = malloc (buffer_size);
+  assert (buffer);
+  hb_serialize_context_t c (buffer, buffer_size);
+  populate_serializer_with_large_mark_base_pos_1 <40, 10, 110, 1>(&c);
+
+  void* expected_buffer = malloc (buffer_size);
+  assert (expected_buffer);
+  hb_serialize_context_t e (expected_buffer, buffer_size);
+  populate_serializer_with_large_mark_base_pos_1 <40, 10, 110, 2>(&e);
+
+  run_resolve_overflow_test ("test_resolve_with_basic_mark_base_pos_1_split",
+                             c,
+                             e,
+                             20,
+                             true,
+                             HB_TAG('G', 'P', 'O', 'S'));
+  free (buffer);
+  free (expected_buffer);
+}
+
 static void test_resolve_overflows_via_splitting_spaces ()
 {
   size_t buffer_size = 160000;
@@ -1944,6 +2126,7 @@ main (int argc, char **argv)
   test_resolve_with_basic_pair_pos_2_split ();
   test_resolve_with_pair_pos_2_split_with_device_tables ();
   test_resolve_with_close_to_limit_pair_pos_2_split ();
+  test_resolve_with_basic_mark_base_pos_1_split ();
 
   // TODO(grieger): have run overflow tests compare graph equality not final packed binary.
   // TODO(grieger): split test where multiple subtables in one lookup are split to test link ordering.