diff options
Diffstat (limited to 'src/3rdparty/harfbuzz-ng/src/hb-repacker.hh')
| -rw-r--r-- | src/3rdparty/harfbuzz-ng/src/hb-repacker.hh | 1289 |
1 files changed, 217 insertions, 1072 deletions
diff --git a/src/3rdparty/harfbuzz-ng/src/hb-repacker.hh b/src/3rdparty/harfbuzz-ng/src/hb-repacker.hh index 2a9e75c45b..e9cd376ad3 100644 --- a/src/3rdparty/harfbuzz-ng/src/hb-repacker.hh +++ b/src/3rdparty/harfbuzz-ng/src/hb-repacker.hh @@ -29,1077 +29,167 @@ #include "hb-open-type.hh" #include "hb-map.hh" -#include "hb-priority-queue.hh" -#include "hb-serialize.hh" #include "hb-vector.hh" +#include "graph/graph.hh" +#include "graph/gsubgpos-graph.hh" +#include "graph/serialize.hh" + +using graph::graph_t; /* * For a detailed writeup on the overflow resolution algorithm see: * docs/repacker.md */ -struct graph_t -{ - struct vertex_t - { - hb_serialize_context_t::object_t obj; - int64_t distance = 0 ; - int64_t space = 0 ; - hb_vector_t<unsigned> parents; - unsigned start = 0; - unsigned end = 0; - unsigned priority = 0; - - bool is_shared () const - { - return parents.length > 1; - } - - unsigned incoming_edges () const - { - return parents.length; - } - - void remove_parent (unsigned parent_index) - { - for (unsigned i = 0; i < parents.length; i++) - { - if (parents[i] != parent_index) continue; - parents.remove (i); - break; - } - } - - void remap_parents (const hb_vector_t<unsigned>& id_map) - { - for (unsigned i = 0; i < parents.length; i++) - parents[i] = id_map[parents[i]]; - } - - void remap_parent (unsigned old_index, unsigned new_index) - { - for (unsigned i = 0; i < parents.length; i++) - { - if (parents[i] == old_index) - parents[i] = new_index; - } - } - - bool is_leaf () const - { - return !obj.real_links.length && !obj.virtual_links.length; - } - - bool raise_priority () - { - if (has_max_priority ()) return false; - priority++; - return true; - } - - bool has_max_priority () const { - return priority >= 3; - } - - int64_t modified_distance (unsigned order) const - { - // TODO(garretrieger): once priority is high enough, should try - // setting distance = 0 which will force to sort immediately after - // it's parent where possible. - - int64_t modified_distance = - hb_min (hb_max(distance + distance_modifier (), 0), 0x7FFFFFFFFFF); - if (has_max_priority ()) { - modified_distance = 0; - } - return (modified_distance << 18) | (0x003FFFF & order); - } - - int64_t distance_modifier () const - { - if (!priority) return 0; - int64_t table_size = obj.tail - obj.head; - - if (priority == 1) - return -table_size / 2; - - return -table_size; - } - }; - - struct overflow_record_t - { - unsigned parent; - unsigned child; - }; - - /* - * A topological sorting of an object graph. Ordered - * in reverse serialization order (first object in the - * serialization is at the end of the list). This matches - * the 'packed' object stack used internally in the - * serializer - */ - template<typename T> - graph_t (const T& objects) - : parents_invalid (true), - distance_invalid (true), - positions_invalid (true), - successful (true) - { - num_roots_for_space_.push (1); - bool removed_nil = false; - for (unsigned i = 0; i < objects.length; i++) - { - // TODO(grieger): check all links point to valid objects. - - // If this graph came from a serialization buffer object 0 is the - // nil object. We don't need it for our purposes here so drop it. - if (i == 0 && !objects[i]) - { - removed_nil = true; - continue; - } - - vertex_t* v = vertices_.push (); - if (check_success (!vertices_.in_error ())) - v->obj = *objects[i]; - if (!removed_nil) continue; - // Fix indices to account for removed nil object. - for (auto& l : v->obj.all_links_writer ()) { - l.objidx--; - } - } - } - - ~graph_t () - { - vertices_.fini (); - } - - bool in_error () const - { - return !successful || - vertices_.in_error () || - num_roots_for_space_.in_error (); - } - - const vertex_t& root () const - { - return vertices_[root_idx ()]; - } - - unsigned root_idx () const - { - // Object graphs are in reverse order, the first object is at the end - // of the vector. Since the graph is topologically sorted it's safe to - // assume the first object has no incoming edges. - return vertices_.length - 1; - } - - const hb_serialize_context_t::object_t& object(unsigned i) const - { - return vertices_[i].obj; - } - - /* - * serialize graph into the provided serialization buffer. - */ - hb_blob_t* serialize () const - { - hb_vector_t<char> buffer; - size_t size = serialized_length (); - if (!buffer.alloc (size)) { - DEBUG_MSG (SUBSET_REPACK, nullptr, "Unable to allocate output buffer."); - return nullptr; - } - hb_serialize_context_t c((void *) buffer, size); - - c.start_serialize<void> (); - for (unsigned i = 0; i < vertices_.length; i++) { - c.push (); - - size_t size = vertices_[i].obj.tail - vertices_[i].obj.head; - char* start = c.allocate_size <char> (size); - if (!start) { - DEBUG_MSG (SUBSET_REPACK, nullptr, "Buffer out of space."); - return nullptr; - } - - memcpy (start, vertices_[i].obj.head, size); - - // Only real links needs to be serialized. - for (const auto& link : vertices_[i].obj.real_links) - serialize_link (link, start, &c); - - // All duplications are already encoded in the graph, so don't - // enable sharing during packing. - c.pop_pack (false); - } - c.end_serialize (); - - if (c.in_error ()) { - DEBUG_MSG (SUBSET_REPACK, nullptr, "Error during serialization. Err flag: %d", - c.errors); - return nullptr; - } - - return c.copy_blob (); - } - - /* - * Generates a new topological sorting of graph using Kahn's - * algorithm: https://en.wikipedia.org/wiki/Topological_sorting#Algorithms - */ - void sort_kahn () - { - positions_invalid = true; - - if (vertices_.length <= 1) { - // Graph of 1 or less doesn't need sorting. - return; - } - - hb_vector_t<unsigned> queue; - hb_vector_t<vertex_t> sorted_graph; - if (unlikely (!check_success (sorted_graph.resize (vertices_.length)))) return; - hb_vector_t<unsigned> id_map; - if (unlikely (!check_success (id_map.resize (vertices_.length)))) return; - - hb_vector_t<unsigned> removed_edges; - if (unlikely (!check_success (removed_edges.resize (vertices_.length)))) return; - update_parents (); - - queue.push (root_idx ()); - int new_id = vertices_.length - 1; - - while (!queue.in_error () && queue.length) - { - unsigned next_id = queue[0]; - queue.remove (0); - - vertex_t& next = vertices_[next_id]; - sorted_graph[new_id] = next; - id_map[next_id] = new_id--; - - for (const auto& link : next.obj.all_links ()) { - removed_edges[link.objidx]++; - if (!(vertices_[link.objidx].incoming_edges () - removed_edges[link.objidx])) - queue.push (link.objidx); - } - } - - check_success (!queue.in_error ()); - check_success (!sorted_graph.in_error ()); - if (!check_success (new_id == -1)) - print_orphaned_nodes (); - - remap_all_obj_indices (id_map, &sorted_graph); - - hb_swap (vertices_, sorted_graph); - sorted_graph.fini (); - } - - /* - * Generates a new topological sorting of graph ordered by the shortest - * distance to each node. - */ - void sort_shortest_distance () - { - positions_invalid = true; - - if (vertices_.length <= 1) { - // Graph of 1 or less doesn't need sorting. - return; - } - - update_distances (); - hb_priority_queue_t queue; - hb_vector_t<vertex_t> sorted_graph; - if (unlikely (!check_success (sorted_graph.resize (vertices_.length)))) return; - hb_vector_t<unsigned> id_map; - if (unlikely (!check_success (id_map.resize (vertices_.length)))) return; - - hb_vector_t<unsigned> removed_edges; - if (unlikely (!check_success (removed_edges.resize (vertices_.length)))) return; - update_parents (); - - queue.insert (root ().modified_distance (0), root_idx ()); - int new_id = root_idx (); - unsigned order = 1; - while (!queue.in_error () && !queue.is_empty ()) - { - unsigned next_id = queue.pop_minimum().second; - - vertex_t& next = vertices_[next_id]; - sorted_graph[new_id] = next; - id_map[next_id] = new_id--; - - for (const auto& link : next.obj.all_links ()) { - removed_edges[link.objidx]++; - if (!(vertices_[link.objidx].incoming_edges () - removed_edges[link.objidx])) - // Add the order that the links were encountered to the priority. - // This ensures that ties between priorities objects are broken in a consistent - // way. More specifically this is set up so that if a set of objects have the same - // distance they'll be added to the topological order in the order that they are - // referenced from the parent object. - queue.insert (vertices_[link.objidx].modified_distance (order++), - link.objidx); - } - } - - check_success (!queue.in_error ()); - check_success (!sorted_graph.in_error ()); - if (!check_success (new_id == -1)) - print_orphaned_nodes (); - - remap_all_obj_indices (id_map, &sorted_graph); - - hb_swap (vertices_, sorted_graph); - sorted_graph.fini (); - } - - /* - * Assign unique space numbers to each connected subgraph of 32 bit offset(s). - */ - bool assign_32bit_spaces () - { - unsigned root_index = root_idx (); - hb_set_t visited; - hb_set_t roots; - for (unsigned i = 0; i <= root_index; i++) - { - // Only real links can form 32 bit spaces - for (auto& l : vertices_[i].obj.real_links) - { - if (l.width == 4 && !l.is_signed) - { - roots.add (l.objidx); - find_subgraph (l.objidx, visited); - } - } - } - - // Mark everything not in the subgraphs of 32 bit roots as visited. - // This prevents 32 bit subgraphs from being connected via nodes not in the 32 bit subgraphs. - visited.invert (); - - if (!roots) return false; - - while (roots) - { - unsigned next = HB_SET_VALUE_INVALID; - if (unlikely (!check_success (!roots.in_error ()))) break; - if (!roots.next (&next)) break; - - hb_set_t connected_roots; - find_connected_nodes (next, roots, visited, connected_roots); - if (unlikely (!check_success (!connected_roots.in_error ()))) break; - - isolate_subgraph (connected_roots); - if (unlikely (!check_success (!connected_roots.in_error ()))) break; - - unsigned next_space = this->next_space (); - num_roots_for_space_.push (0); - for (unsigned root : connected_roots) - { - DEBUG_MSG (SUBSET_REPACK, nullptr, "Subgraph %u gets space %u", root, next_space); - vertices_[root].space = next_space; - num_roots_for_space_[next_space] = num_roots_for_space_[next_space] + 1; - distance_invalid = true; - positions_invalid = true; - } - - // TODO(grieger): special case for GSUB/GPOS use extension promotions to move 16 bit space - // into the 32 bit space as needed, instead of using isolation. - } - - - - return true; - } - - /* - * Isolates the subgraph of nodes reachable from root. Any links to nodes in the subgraph - * that originate from outside of the subgraph will be removed by duplicating the linked to - * object. - * - * Indices stored in roots will be updated if any of the roots are duplicated to new indices. - */ - bool isolate_subgraph (hb_set_t& roots) - { - update_parents (); - hb_hashmap_t<unsigned, unsigned> subgraph; - - // incoming edges to root_idx should be all 32 bit in length so we don't need to de-dup these - // set the subgraph incoming edge count to match all of root_idx's incoming edges - hb_set_t parents; - for (unsigned root_idx : roots) - { - subgraph.set (root_idx, wide_parents (root_idx, parents)); - find_subgraph (root_idx, subgraph); - } - - unsigned original_root_idx = root_idx (); - hb_hashmap_t<unsigned, unsigned> index_map; - bool made_changes = false; - for (auto entry : subgraph.iter ()) - { - const auto& node = vertices_[entry.first]; - unsigned subgraph_incoming_edges = entry.second; - - if (subgraph_incoming_edges < node.incoming_edges ()) - { - // Only de-dup objects with incoming links from outside the subgraph. - made_changes = true; - duplicate_subgraph (entry.first, index_map); - } - } - - if (!made_changes) - return false; - - if (original_root_idx != root_idx () - && parents.has (original_root_idx)) - { - // If the root idx has changed since parents was determined, update root idx in parents - parents.add (root_idx ()); - parents.del (original_root_idx); - } - - auto new_subgraph = - + subgraph.keys () - | hb_map([&] (unsigned node_idx) { - if (index_map.has (node_idx)) return index_map[node_idx]; - return node_idx; - }) - ; - - remap_obj_indices (index_map, new_subgraph); - remap_obj_indices (index_map, parents.iter (), true); - - // Update roots set with new indices as needed. - unsigned next = HB_SET_VALUE_INVALID; - while (roots.next (&next)) - { - if (index_map.has (next)) - { - roots.del (next); - roots.add (index_map[next]); - } - } - - return true; - } - - void find_subgraph (unsigned node_idx, hb_hashmap_t<unsigned, unsigned>& subgraph) - { - for (const auto& link : vertices_[node_idx].obj.all_links ()) - { - if (subgraph.has (link.objidx)) - { - subgraph.set (link.objidx, subgraph[link.objidx] + 1); - continue; - } - subgraph.set (link.objidx, 1); - find_subgraph (link.objidx, subgraph); - } - } +struct lookup_size_t +{ + unsigned lookup_index; + size_t size; + unsigned num_subtables; - void find_subgraph (unsigned node_idx, hb_set_t& subgraph) + static int cmp (const void* a, const void* b) { - if (subgraph.has (node_idx)) return; - subgraph.add (node_idx); - for (const auto& link : vertices_[node_idx].obj.all_links ()) - find_subgraph (link.objidx, subgraph); + return cmp ((const lookup_size_t*) a, + (const lookup_size_t*) b); } - /* - * duplicates all nodes in the subgraph reachable from node_idx. Does not re-assign - * links. index_map is updated with mappings from old id to new id. If a duplication has already - * been performed for a given index, then it will be skipped. - */ - void duplicate_subgraph (unsigned node_idx, hb_hashmap_t<unsigned, unsigned>& index_map) + static int cmp (const lookup_size_t* a, const lookup_size_t* b) { - if (index_map.has (node_idx)) - return; - - index_map.set (node_idx, duplicate (node_idx)); - for (const auto& l : object (node_idx).all_links ()) { - duplicate_subgraph (l.objidx, index_map); - } - } - - /* - * Creates a copy of node_idx and returns it's new index. - */ - unsigned duplicate (unsigned node_idx) - { - positions_invalid = true; - distance_invalid = true; - - auto* clone = vertices_.push (); - auto& child = vertices_[node_idx]; - if (vertices_.in_error ()) { - return -1; - } - - clone->obj.head = child.obj.head; - clone->obj.tail = child.obj.tail; - clone->distance = child.distance; - clone->space = child.space; - clone->parents.reset (); - - unsigned clone_idx = vertices_.length - 2; - for (const auto& l : child.obj.real_links) - { - clone->obj.real_links.push (l); - vertices_[l.objidx].parents.push (clone_idx); + double subtables_per_byte_a = (double) a->num_subtables / (double) a->size; + double subtables_per_byte_b = (double) b->num_subtables / (double) b->size; + if (subtables_per_byte_a == subtables_per_byte_b) { + return b->lookup_index - a->lookup_index; } - for (const auto& l : child.obj.virtual_links) - { - clone->obj.virtual_links.push (l); - vertices_[l.objidx].parents.push (clone_idx); - } - - check_success (!clone->obj.real_links.in_error ()); - check_success (!clone->obj.virtual_links.in_error ()); - - // The last object is the root of the graph, so swap back the root to the end. - // The root's obj idx does change, however since it's root nothing else refers to it. - // all other obj idx's will be unaffected. - vertex_t root = vertices_[vertices_.length - 2]; - vertices_[clone_idx] = *clone; - vertices_[vertices_.length - 1] = root; - // Since the root moved, update the parents arrays of all children on the root. - for (const auto& l : root.obj.all_links ()) - vertices_[l.objidx].remap_parent (root_idx () - 1, root_idx ()); - - return clone_idx; + double cmp = subtables_per_byte_b - subtables_per_byte_a; + if (cmp < 0) return -1; + if (cmp > 0) return 1; + return 0; } +}; - /* - * Creates a copy of child and re-assigns the link from - * parent to the clone. The copy is a shallow copy, objects - * linked from child are not duplicated. - */ - bool duplicate (unsigned parent_idx, unsigned child_idx) - { - update_parents (); - - unsigned links_to_child = 0; - for (const auto& l : vertices_[parent_idx].obj.all_links ()) - { - if (l.objidx == child_idx) links_to_child++; - } - - if (vertices_[child_idx].incoming_edges () <= links_to_child) - { - // Can't duplicate this node, doing so would orphan the original one as all remaining links - // to child are from parent. - DEBUG_MSG (SUBSET_REPACK, nullptr, " Not duplicating %d => %d", - parent_idx, child_idx); +static inline +bool _presplit_subtables_if_needed (graph::gsubgpos_graph_context_t& ext_context) +{ + // For each lookup this will check the size of subtables and split them as needed + // so that no subtable is at risk of overflowing. (where we support splitting for + // that subtable type). + // + // TODO(grieger): de-dup newly added nodes as necessary. Probably just want a full de-dup + // pass after this processing is done. Not super necessary as splits are + // only done where overflow is likely, so de-dup probably will get undone + // later anyways. + + // The loop below can modify the contents of ext_context.lookups if new subtables are added + // to a lookup during a split. So save the initial set of lookup indices so the iteration doesn't + // risk access free'd memory if ext_context.lookups gets resized. + hb_set_t lookup_indices(ext_context.lookups.keys ()); + for (unsigned lookup_index : lookup_indices) + { + graph::Lookup* lookup = ext_context.lookups.get(lookup_index); + if (!lookup->split_subtables_if_needed (ext_context, lookup_index)) return false; - } - - DEBUG_MSG (SUBSET_REPACK, nullptr, " Duplicating %d => %d", - parent_idx, child_idx); - - unsigned clone_idx = duplicate (child_idx); - if (clone_idx == (unsigned) -1) return false; - // duplicate shifts the root node idx, so if parent_idx was root update it. - if (parent_idx == clone_idx) parent_idx++; - - auto& parent = vertices_[parent_idx]; - for (auto& l : parent.obj.all_links_writer ()) - { - if (l.objidx != child_idx) - continue; - - reassign_link (l, parent_idx, clone_idx); - } - - return true; - } - - /* - * Raises the sorting priority of all children. - */ - bool raise_childrens_priority (unsigned parent_idx) - { - DEBUG_MSG (SUBSET_REPACK, nullptr, " Raising priority of all children of %d", - parent_idx); - // This operation doesn't change ordering until a sort is run, so no need - // to invalidate positions. It does not change graph structure so no need - // to update distances or edge counts. - auto& parent = vertices_[parent_idx].obj; - bool made_change = false; - for (auto& l : parent.all_links_writer ()) - made_change |= vertices_[l.objidx].raise_priority (); - return made_change; } - /* - * Will any offsets overflow on graph when it's serialized? - */ - bool will_overflow (hb_vector_t<overflow_record_t>* overflows = nullptr) - { - if (overflows) overflows->resize (0); - update_positions (); - - for (int parent_idx = vertices_.length - 1; parent_idx >= 0; parent_idx--) - { - // Don't need to check virtual links for overflow - for (const auto& link : vertices_[parent_idx].obj.real_links) - { - int64_t offset = compute_offset (parent_idx, link); - if (is_valid_offset (offset, link)) - continue; - - if (!overflows) return true; - - overflow_record_t r; - r.parent = parent_idx; - r.child = link.objidx; - overflows->push (r); - } - } - - if (!overflows) return false; - return overflows->length; - } - - void print_orphaned_nodes () - { - if (!DEBUG_ENABLED(SUBSET_REPACK)) return; - - DEBUG_MSG (SUBSET_REPACK, nullptr, "Graph is not fully connected."); - parents_invalid = true; - update_parents(); - - for (unsigned i = 0; i < root_idx (); i++) - { - const auto& v = vertices_[i]; - if (!v.parents) - DEBUG_MSG (SUBSET_REPACK, nullptr, "Node %u is orphaned.", i); - } - } - - void print_overflows (const hb_vector_t<overflow_record_t>& overflows) - { - if (!DEBUG_ENABLED(SUBSET_REPACK)) return; - - update_parents (); - int limit = 10; - for (const auto& o : overflows) - { - if (!limit--) break; - const auto& parent = vertices_[o.parent]; - const auto& child = vertices_[o.child]; - DEBUG_MSG (SUBSET_REPACK, nullptr, - " overflow from " - "%4d (%4d in, %4d out, space %2d) => " - "%4d (%4d in, %4d out, space %2d)", - o.parent, - parent.incoming_edges (), - parent.obj.real_links.length + parent.obj.virtual_links.length, - space_for (o.parent), - o.child, - child.incoming_edges (), - child.obj.real_links.length + child.obj.virtual_links.length, - space_for (o.child)); - } - if (overflows.length > 10) { - DEBUG_MSG (SUBSET_REPACK, nullptr, " ... plus %d more overflows.", overflows.length - 10); - } - } - - unsigned num_roots_for_space (unsigned space) const - { - return num_roots_for_space_[space]; - } - - unsigned next_space () const - { - return num_roots_for_space_.length; - } - - void move_to_new_space (const hb_set_t& indices) - { - num_roots_for_space_.push (0); - unsigned new_space = num_roots_for_space_.length - 1; - - for (unsigned index : indices) { - auto& node = vertices_[index]; - num_roots_for_space_[node.space] = num_roots_for_space_[node.space] - 1; - num_roots_for_space_[new_space] = num_roots_for_space_[new_space] + 1; - node.space = new_space; - distance_invalid = true; - positions_invalid = true; - } - } - - unsigned space_for (unsigned index, unsigned* root = nullptr) const - { - const auto& node = vertices_[index]; - if (node.space) - { - if (root != nullptr) - *root = index; - return node.space; - } - - if (!node.parents) - { - if (root) - *root = index; - return 0; - } - - return space_for (node.parents[0], root); - } - - void err_other_error () { this->successful = false; } - - private: - - size_t serialized_length () const { - size_t total_size = 0; - for (unsigned i = 0; i < vertices_.length; i++) { - size_t size = vertices_[i].obj.tail - vertices_[i].obj.head; - total_size += size; - } - return total_size; - } + return true; +} - /* - * Returns the numbers of incoming edges that are 32bits wide. - */ - unsigned wide_parents (unsigned node_idx, hb_set_t& parents) const - { - unsigned count = 0; +/* + * Analyze the lookups in a GSUB/GPOS table and decide if any should be promoted + * to extension lookups. + */ +static inline +bool _promote_extensions_if_needed (graph::gsubgpos_graph_context_t& ext_context) +{ + // Simple Algorithm (v1, current): + // 1. Calculate how many bytes each non-extension lookup consumes. + // 2. Select up to 64k of those to remain as non-extension (greedy, highest subtables per byte first) + // 3. Promote the rest. + // + // Advanced Algorithm (v2, not implemented): + // 1. Perform connected component analysis using lookups as roots. + // 2. Compute size of each connected component. + // 3. Select up to 64k worth of connected components to remain as non-extensions. + // (greedy, highest subtables per byte first) + // 4. Promote the rest. + + // TODO(garretrieger): support extension demotion, then consider all lookups. Requires advanced algo. + // TODO(garretrieger): also support extension promotion during iterative resolution phase, then + // we can use a less conservative threshold here. + // TODO(grieger): skip this for the 24 bit case. + if (!ext_context.lookups) return true; + + unsigned total_lookup_table_sizes = 0; + hb_vector_t<lookup_size_t> lookup_sizes; + lookup_sizes.alloc (ext_context.lookups.get_population (), true); + + for (unsigned lookup_index : ext_context.lookups.keys ()) + { + const auto& lookup_v = ext_context.graph.vertices_[lookup_index]; + total_lookup_table_sizes += lookup_v.table_size (); + + const graph::Lookup* lookup = ext_context.lookups.get(lookup_index); hb_set_t visited; - for (unsigned p : vertices_[node_idx].parents) - { - if (visited.has (p)) continue; - visited.add (p); - - // Only real links can be wide - for (const auto& l : vertices_[p].obj.real_links) - { - if (l.objidx == node_idx && l.width == 4 && !l.is_signed) - { - count++; - parents.add (p); - } - } - } - return count; - } - - bool check_success (bool success) - { return this->successful && (success || (err_other_error (), false)); } - - /* - * Creates a map from objid to # of incoming edges. - */ - void update_parents () - { - if (!parents_invalid) return; - - for (unsigned i = 0; i < vertices_.length; i++) - vertices_[i].parents.reset (); - - for (unsigned p = 0; p < vertices_.length; p++) - { - for (auto& l : vertices_[p].obj.all_links ()) - { - vertices_[l.objidx].parents.push (p); - } - } - - parents_invalid = false; - } - - /* - * compute the serialized start and end positions for each vertex. - */ - void update_positions () - { - if (!positions_invalid) return; - - unsigned current_pos = 0; - for (int i = root_idx (); i >= 0; i--) - { - auto& v = vertices_[i]; - v.start = current_pos; - current_pos += v.obj.tail - v.obj.head; - v.end = current_pos; - } - - positions_invalid = false; + lookup_sizes.push (lookup_size_t { + lookup_index, + ext_context.graph.find_subgraph_size (lookup_index, visited), + lookup->number_of_subtables (), + }); } - /* - * Finds the distance to each object in the graph - * from the initial node. - */ - void update_distances () - { - if (!distance_invalid) return; - - // Uses Dijkstra's algorithm to find all of the shortest distances. - // https://en.wikipedia.org/wiki/Dijkstra%27s_algorithm - // - // Implementation Note: - // Since our priority queue doesn't support fast priority decreases - // we instead just add new entries into the queue when a priority changes. - // Redundant ones are filtered out later on by the visited set. - // According to https://www3.cs.stonybrook.edu/~rezaul/papers/TR-07-54.pdf - // for practical performance this is faster then using a more advanced queue - // (such as a fibonacci queue) with a fast decrease priority. - for (unsigned i = 0; i < vertices_.length; i++) - { - if (i == vertices_.length - 1) - vertices_[i].distance = 0; - else - vertices_[i].distance = hb_int_max (int64_t); - } - - hb_priority_queue_t queue; - queue.insert (0, vertices_.length - 1); - - hb_vector_t<bool> visited; - visited.resize (vertices_.length); - - while (!queue.in_error () && !queue.is_empty ()) - { - unsigned next_idx = queue.pop_minimum ().second; - if (visited[next_idx]) continue; - const auto& next = vertices_[next_idx]; - int64_t next_distance = vertices_[next_idx].distance; - visited[next_idx] = true; - - for (const auto& link : next.obj.all_links ()) - { - if (visited[link.objidx]) continue; - - const auto& child = vertices_[link.objidx].obj; - unsigned link_width = link.width ? link.width : 4; // treat virtual offsets as 32 bits wide - int64_t child_weight = (child.tail - child.head) + - ((int64_t) 1 << (link_width * 8)) * (vertices_[link.objidx].space + 1); - int64_t child_distance = next_distance + child_weight; - - if (child_distance < vertices_[link.objidx].distance) - { - vertices_[link.objidx].distance = child_distance; - queue.insert (child_distance, link.objidx); - } - } - } + lookup_sizes.qsort (); - check_success (!queue.in_error ()); - if (!check_success (queue.is_empty ())) - { - print_orphaned_nodes (); - return; - } + size_t lookup_list_size = ext_context.graph.vertices_[ext_context.lookup_list_index].table_size (); + size_t l2_l3_size = lookup_list_size + total_lookup_table_sizes; // Lookup List + Lookups + size_t l3_l4_size = total_lookup_table_sizes; // Lookups + SubTables + size_t l4_plus_size = 0; // SubTables + their descendants - distance_invalid = false; - } - - int64_t compute_offset ( - unsigned parent_idx, - const hb_serialize_context_t::object_t::link_t& link) const + // Start by assuming all lookups are using extension subtables, this size will be removed later + // if it's decided to not make a lookup extension. + for (auto p : lookup_sizes) { - const auto& parent = vertices_[parent_idx]; - const auto& child = vertices_[link.objidx]; - int64_t offset = 0; - switch ((hb_serialize_context_t::whence_t) link.whence) { - case hb_serialize_context_t::whence_t::Head: - offset = child.start - parent.start; break; - case hb_serialize_context_t::whence_t::Tail: - offset = child.start - parent.end; break; - case hb_serialize_context_t::whence_t::Absolute: - offset = child.start; break; - } - - assert (offset >= link.bias); - offset -= link.bias; - return offset; + // TODO(garretrieger): this overestimates the extension subtables size because some extension subtables may be + // reused. However, we can't correct this until we have connected component analysis in place. + unsigned subtables_size = p.num_subtables * 8; + l3_l4_size += subtables_size; + l4_plus_size += subtables_size; } - bool is_valid_offset (int64_t offset, - const hb_serialize_context_t::object_t::link_t& link) const + bool layers_full = false; + for (auto p : lookup_sizes) { - if (unlikely (!link.width)) - // Virtual links can't overflow. - return link.is_signed || offset >= 0; - - if (link.is_signed) - { - if (link.width == 4) - return offset >= -((int64_t) 1 << 31) && offset < ((int64_t) 1 << 31); - else - return offset >= -(1 << 15) && offset < (1 << 15); - } - else - { - if (link.width == 4) - return offset >= 0 && offset < ((int64_t) 1 << 32); - else if (link.width == 3) - return offset >= 0 && offset < ((int32_t) 1 << 24); - else - return offset >= 0 && offset < (1 << 16); - } - } - - /* - * Updates a link in the graph to point to a different object. Corrects the - * parents vector on the previous and new child nodes. - */ - void reassign_link (hb_serialize_context_t::object_t::link_t& link, - unsigned parent_idx, - unsigned new_idx) - { - unsigned old_idx = link.objidx; - link.objidx = new_idx; - vertices_[old_idx].remove_parent (parent_idx); - vertices_[new_idx].parents.push (parent_idx); - } - - /* - * Updates all objidx's in all links using the provided mapping. Corrects incoming edge counts. - */ - template<typename Iterator, hb_requires (hb_is_iterator (Iterator))> - void remap_obj_indices (const hb_hashmap_t<unsigned, unsigned>& id_map, - Iterator subgraph, - bool only_wide = false) - { - if (!id_map) return; - for (unsigned i : subgraph) - { - for (auto& link : vertices_[i].obj.all_links_writer ()) - { - if (!id_map.has (link.objidx)) continue; - if (only_wide && !(link.width == 4 && !link.is_signed)) continue; - - reassign_link (link, i, id_map[link.objidx]); - } - } - } + const graph::Lookup* lookup = ext_context.lookups.get(p.lookup_index); + if (lookup->is_extension (ext_context.table_tag)) + // already an extension so size is counted by the loop above. + continue; - /* - * Updates all objidx's in all links using the provided mapping. - */ - void remap_all_obj_indices (const hb_vector_t<unsigned>& id_map, - hb_vector_t<vertex_t>* sorted_graph) const - { - for (unsigned i = 0; i < sorted_graph->length; i++) + if (!layers_full) { - (*sorted_graph)[i].remap_parents (id_map); - for (auto& link : (*sorted_graph)[i].obj.all_links_writer ()) - { - link.objidx = id_map[link.objidx]; - } - } - } + size_t lookup_size = ext_context.graph.vertices_[p.lookup_index].table_size (); + hb_set_t visited; + size_t subtables_size = ext_context.graph.find_subgraph_size (p.lookup_index, visited, 1) - lookup_size; + size_t remaining_size = p.size - subtables_size - lookup_size; - template <typename O> void - serialize_link_of_type (const hb_serialize_context_t::object_t::link_t& link, - char* head, - hb_serialize_context_t* c) const - { - OT::Offset<O>* offset = reinterpret_cast<OT::Offset<O>*> (head + link.position); - *offset = 0; - c->add_link (*offset, - // serializer has an extra nil object at the start of the - // object array. So all id's are +1 of what our id's are. - link.objidx + 1, - (hb_serialize_context_t::whence_t) link.whence, - link.bias); - } - - void serialize_link (const hb_serialize_context_t::object_t::link_t& link, - char* head, - hb_serialize_context_t* c) const - { - switch (link.width) - { - case 0: - // Virtual links aren't serialized. - return; - case 4: - if (link.is_signed) - { - serialize_link_of_type<OT::HBINT32> (link, head, c); - } else { - serialize_link_of_type<OT::HBUINT32> (link, head, c); - } - return; - case 2: - if (link.is_signed) - { - serialize_link_of_type<OT::HBINT16> (link, head, c); - } else { - serialize_link_of_type<OT::HBUINT16> (link, head, c); - } - return; - case 3: - serialize_link_of_type<OT::HBUINT24> (link, head, c); - return; - default: - // Unexpected link width. - assert (0); - } - } + l3_l4_size += subtables_size; + l3_l4_size -= p.num_subtables * 8; + l4_plus_size += subtables_size + remaining_size; - /* - * Finds all nodes in targets that are reachable from start_idx, nodes in visited will be skipped. - * For this search the graph is treated as being undirected. - * - * Connected targets will be added to connected and removed from targets. All visited nodes - * will be added to visited. - */ - void find_connected_nodes (unsigned start_idx, - hb_set_t& targets, - hb_set_t& visited, - hb_set_t& connected) - { - if (unlikely (!check_success (!visited.in_error ()))) return; - if (visited.has (start_idx)) return; - visited.add (start_idx); + if (l2_l3_size < (1 << 16) + && l3_l4_size < (1 << 16) + && l4_plus_size < (1 << 16)) continue; // this lookup fits within all layers groups - if (targets.has (start_idx)) - { - targets.del (start_idx); - connected.add (start_idx); + layers_full = true; } - const auto& v = vertices_[start_idx]; - - // Graph is treated as undirected so search children and parents of start_idx - for (const auto& l : v.obj.all_links ()) - find_connected_nodes (l.objidx, targets, visited, connected); - - for (unsigned p : v.parents) - find_connected_nodes (p, targets, visited, connected); + if (!ext_context.lookups.get(p.lookup_index)->make_extension (ext_context, p.lookup_index)) + return false; } - public: - // TODO(garretrieger): make private, will need to move most of offset overflow code into graph. - hb_vector_t<vertex_t> vertices_; - private: - bool parents_invalid; - bool distance_invalid; - bool positions_invalid; - bool successful; - hb_vector_t<unsigned> num_roots_for_space_; -}; + return true; +} static inline -bool _try_isolating_subgraphs (const hb_vector_t<graph_t::overflow_record_t>& overflows, +bool _try_isolating_subgraphs (const hb_vector_t<graph::overflow_record_t>& overflows, graph_t& sorted_graph) { unsigned space = 0; @@ -1107,7 +197,7 @@ bool _try_isolating_subgraphs (const hb_vector_t<graph_t::overflow_record_t>& ov for (int i = overflows.length - 1; i >= 0; i--) { - const graph_t::overflow_record_t& r = overflows[i]; + const graph::overflow_record_t& r = overflows[i]; unsigned root; unsigned overflow_space = sorted_graph.space_for (r.parent, &root); @@ -1129,14 +219,14 @@ bool _try_isolating_subgraphs (const hb_vector_t<graph_t::overflow_record_t>& ov // Only move at most half of the roots in a space at a time. unsigned extra = roots_to_isolate.get_population () - maximum_to_move; while (extra--) { - unsigned root = HB_SET_VALUE_INVALID; + uint32_t root = HB_SET_VALUE_INVALID; roots_to_isolate.previous (&root); roots_to_isolate.del (root); } } DEBUG_MSG (SUBSET_REPACK, nullptr, - "Overflow in space %d (%d roots). Moving %d roots to space %d.", + "Overflow in space %u (%u roots). Moving %u roots to space %u.", space, sorted_graph.num_roots_for_space (space), roots_to_isolate.get_population (), @@ -1149,7 +239,7 @@ bool _try_isolating_subgraphs (const hb_vector_t<graph_t::overflow_record_t>& ov } static inline -bool _process_overflows (const hb_vector_t<graph_t::overflow_record_t>& overflows, +bool _process_overflows (const hb_vector_t<graph::overflow_record_t>& overflows, hb_set_t& priority_bumped_parents, graph_t& sorted_graph) { @@ -1158,13 +248,13 @@ bool _process_overflows (const hb_vector_t<graph_t::overflow_record_t>& overflow // Try resolving the furthest overflows first. for (int i = overflows.length - 1; i >= 0; i--) { - const graph_t::overflow_record_t& r = overflows[i]; + const graph::overflow_record_t& r = overflows[i]; const auto& child = sorted_graph.vertices_[r.child]; if (child.is_shared ()) { // The child object is shared, we may be able to eliminate the overflow // by duplicating it. - if (!sorted_graph.duplicate (r.parent, r.child)) continue; + if (sorted_graph.duplicate (r.parent, r.child) == (unsigned) -1) continue; return true; } @@ -1196,57 +286,66 @@ bool _process_overflows (const hb_vector_t<graph_t::overflow_record_t>& overflow 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) { - // Kahn sort is ~twice as fast as shortest distance sort and works for many fonts - // so try it first to save time. - graph_t sorted_graph (packed); - sorted_graph.sort_kahn (); - if (!sorted_graph.will_overflow ()) +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 (); + if (sorted_graph.in_error ()) { - return sorted_graph.serialize (); + DEBUG_MSG (SUBSET_REPACK, nullptr, "Sorted graph in error state after initial sort."); + return false; } - sorted_graph.sort_shortest_distance (); + bool will_overflow = graph::will_overflow (sorted_graph); + if (!will_overflow) + return true; + graph::gsubgpos_graph_context_t ext_context (table_tag, sorted_graph); if ((table_tag == HB_OT_TAG_GPOS || table_tag == HB_OT_TAG_GSUB) - && sorted_graph.will_overflow ()) + && will_overflow) { + if (recalculate_extensions) + { + 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 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 false; + } + } + DEBUG_MSG (SUBSET_REPACK, nullptr, "Assigning spaces to 32 bit subgraphs."); - if (sorted_graph.assign_32bit_spaces ()) + if (sorted_graph.assign_spaces ()) sorted_graph.sort_shortest_distance (); + else + sorted_graph.sort_shortest_distance_if_needed (); } unsigned round = 0; - hb_vector_t<graph_t::overflow_record_t> overflows; + hb_vector_t<graph::overflow_record_t> overflows; // TODO(garretrieger): select a good limit for max rounds. while (!sorted_graph.in_error () - && sorted_graph.will_overflow (&overflows) - && round++ < max_rounds) { - DEBUG_MSG (SUBSET_REPACK, nullptr, "=== Overflow resolution round %d ===", round); - sorted_graph.print_overflows (overflows); + && graph::will_overflow (sorted_graph, &overflows) + && round < max_rounds) { + DEBUG_MSG (SUBSET_REPACK, nullptr, "=== Overflow resolution round %u ===", round); + print_overflows (sorted_graph, overflows); hb_set_t priority_bumped_parents; if (!_try_isolating_subgraphs (overflows, sorted_graph)) { + // Don't count space isolation towards round limit. Only increment + // round counter if space isolation made no changes. + round++; if (!_process_overflows (overflows, priority_bumped_parents, sorted_graph)) { DEBUG_MSG (SUBSET_REPACK, nullptr, "No resolution available :("); @@ -1260,16 +359,62 @@ 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 (sorted_graph.will_overflow ()) + if (graph::will_overflow (sorted_graph)) { DEBUG_MSG (SUBSET_REPACK, nullptr, "Offset overflow resolution failed."); + 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 (sorted_graph.in_error ()) + { + // Invalid graph definition. return nullptr; } - return sorted_graph.serialize (); + if (!sorted_graph.is_fully_connected ()) + { + sorted_graph.print_orphaned_nodes (); + return nullptr; + } + + if (sorted_graph.in_error ()) + { + // Allocations failed somewhere + DEBUG_MSG (SUBSET_REPACK, nullptr, + "Graph is in error, likely due to a memory allocation error."); + return nullptr; + } + + if (!hb_resolve_graph_overflows (table_tag, max_rounds, recalculate_extensions, sorted_graph)) + return nullptr; + + return graph::serialize (sorted_graph); } #endif /* HB_REPACKER_HH */ |