diff options
Diffstat (limited to 'src/3rdparty/harfbuzz-ng/src/hb-repacker.hh')
-rw-r--r-- | src/3rdparty/harfbuzz-ng/src/hb-repacker.hh | 949 |
1 files changed, 324 insertions, 625 deletions
diff --git a/src/3rdparty/harfbuzz-ng/src/hb-repacker.hh b/src/3rdparty/harfbuzz-ng/src/hb-repacker.hh index b02128b5c4..ed40f271cc 100644 --- a/src/3rdparty/harfbuzz-ng/src/hb-repacker.hh +++ b/src/3rdparty/harfbuzz-ng/src/hb-repacker.hh @@ -29,644 +29,395 @@ #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; -struct graph_t -{ - struct vertex_t - { - vertex_t () : - distance (0), - incoming_edges (0), - start (0), - end (0), - priority(0) {} - - void fini () { obj.fini (); } - - hb_serialize_context_t::object_t obj; - int64_t distance; - unsigned incoming_edges; - unsigned start; - unsigned end; - unsigned priority; - - bool is_shared () const - { - return incoming_edges > 1; - } - - bool is_leaf () const - { - return !obj.links.length; - } - - void raise_priority () - { - priority++; - } - - 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), 0x7FFFFFFFFF); - return (modified_distance << 24) | (0x00FFFFFF & order); - } - - int64_t distance_modifier () const - { - if (!priority) return 0; - int64_t table_size = obj.tail - obj.head; - return -(table_size - table_size / (1 << hb_min(priority, 16u))); - } - }; +/* + * For a detailed writeup on the overflow resolution algorithm see: + * docs/repacker.md + */ - struct overflow_record_t - { - unsigned parent; - const hb_serialize_context_t::object_t::link_t* link; - }; +struct lookup_size_t +{ + unsigned lookup_index; + size_t size; + unsigned num_subtables; - struct clone_buffer_t + static int cmp (const void* a, const void* b) { - clone_buffer_t () : head (nullptr), tail (nullptr) {} - - bool copy (const hb_serialize_context_t::object_t& object) - { - fini (); - unsigned size = object.tail - object.head; - head = (char*) hb_malloc (size); - if (!head) return false; - - memcpy (head, object.head, size); - tail = head + size; - return true; - } - - char* head; - char* tail; + return cmp ((const lookup_size_t*) a, + (const lookup_size_t*) b); + } - void fini () - { - if (!head) return; - hb_free (head); - head = nullptr; - } - }; - - /* - * 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 - */ - graph_t (const hb_vector_t<hb_serialize_context_t::object_t *>& objects) - : edge_count_invalid (true), - distance_invalid (true), - positions_invalid (true), - successful (true) + static int cmp (const lookup_size_t* a, const lookup_size_t* b) { - 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; - for (unsigned i = 0; i < v->obj.links.length; i++) - // Fix indices to account for removed nil object. - v->obj.links[i].objidx--; + 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; } - } - ~graph_t () - { - vertices_.fini_deep (); - clone_buffers_.fini_deep (); + double cmp = subtables_per_byte_b - subtables_per_byte_a; + if (cmp < 0) return -1; + if (cmp > 0) return 1; + return 0; } +}; - bool in_error () const +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) { - return !successful || vertices_.in_error () || clone_buffers_.in_error (); + graph::Lookup* lookup = ext_context.lookups.get(lookup_index); + if (!lookup->split_subtables_if_needed (ext_context, lookup_index)) + return false; } - const vertex_t& root () const - { - return vertices_[root_idx ()]; - } + return true; +} - unsigned root_idx () const +/* + * 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 ()) { - // 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 auto& lookup_v = ext_context.graph.vertices_[lookup_index]; + total_lookup_table_sizes += lookup_v.table_size (); - const hb_serialize_context_t::object_t& object(unsigned i) const - { - return vertices_[i].obj; + const graph::Lookup* lookup = ext_context.lookups.get(lookup_index); + hb_set_t visited; + lookup_sizes.push (lookup_size_t { + lookup_index, + ext_context.graph.find_subgraph_size (lookup_index, visited), + lookup->number_of_subtables (), + }); } - /* - * serialize graph into the provided serialization buffer. - */ - void serialize (hb_serialize_context_t* c) const - { - c->start_serialize<void> (); - for (unsigned i = 0; i < vertices_.length; i++) { - c->push (); + lookup_sizes.qsort (); - size_t size = vertices_[i].obj.tail - vertices_[i].obj.head; - char* start = c->allocate_size <char> (size); - if (!start) 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 - memcpy (start, vertices_[i].obj.head, size); - - for (const auto& link : vertices_[i].obj.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 (); + // 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) + { + // 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; } - /* - * Generates a new topological sorting of graph using Kahn's - * algorithm: https://en.wikipedia.org/wiki/Topological_sorting#Algorithms - */ - void sort_kahn () + bool layers_full = false; + for (auto p : lookup_sizes) { - 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; - 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_incoming_edge_count (); - - queue.push (root_idx ()); - int new_id = vertices_.length - 1; + 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; - while (!queue.in_error () && queue.length) + if (!layers_full) { - unsigned next_id = queue[0]; - queue.remove (0); - - vertex_t& next = vertices_[next_id]; - sorted_graph.push (next); - id_map[next_id] = new_id--; - - for (const auto& link : next.obj.links) { - removed_edges[link.objidx]++; - if (!(vertices_[link.objidx].incoming_edges - removed_edges[link.objidx])) - queue.push (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; - check_success (!queue.in_error ()); - check_success (!sorted_graph.in_error ()); - if (!check_success (new_id == -1)) - DEBUG_MSG (SUBSET_REPACK, nullptr, "Graph is not fully connected."); + l3_l4_size += subtables_size; + l3_l4_size -= p.num_subtables * 8; + l4_plus_size += subtables_size + remaining_size; - remap_obj_indices (id_map, &sorted_graph); + if (l2_l3_size < (1 << 16) + && l3_l4_size < (1 << 16) + && l4_plus_size < (1 << 16)) continue; // this lookup fits within all layers groups - sorted_graph.as_array ().reverse (); + layers_full = true; + } - vertices_.fini_deep (); - vertices_ = sorted_graph; - sorted_graph.fini_deep (); + if (!ext_context.lookups.get(p.lookup_index)->make_extension (ext_context, p.lookup_index)) + return false; } - /* - * 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; - } + return true; +} - update_distances (); +static inline +bool _try_isolating_subgraphs (const hb_vector_t<graph::overflow_record_t>& overflows, + graph_t& sorted_graph) +{ + unsigned space = 0; + hb_set_t roots_to_isolate; - hb_priority_queue_t queue; - hb_vector_t<vertex_t> sorted_graph; - hb_vector_t<unsigned> id_map; - if (unlikely (!check_success (id_map.resize (vertices_.length)))) return; + for (int i = overflows.length - 1; i >= 0; i--) + { + const graph::overflow_record_t& r = overflows[i]; - hb_vector_t<unsigned> removed_edges; - if (unlikely (!check_success (removed_edges.resize (vertices_.length)))) return; - update_incoming_edge_count (); + unsigned root; + unsigned overflow_space = sorted_graph.space_for (r.parent, &root); + if (!overflow_space) continue; + if (sorted_graph.num_roots_for_space (overflow_space) <= 1) continue; - 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.push (next); - id_map[next_id] = new_id--; - - for (const auto& link : next.obj.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); - } + if (!space) { + space = overflow_space; } - check_success (!queue.in_error ()); - check_success (!sorted_graph.in_error ()); - if (!check_success (new_id == -1)) - DEBUG_MSG (SUBSET_REPACK, nullptr, "Graph is not fully connected."); - - remap_obj_indices (id_map, &sorted_graph); - - sorted_graph.as_array ().reverse (); - - vertices_.fini_deep (); - vertices_ = sorted_graph; - sorted_graph.fini_deep (); + if (space == overflow_space) + roots_to_isolate.add(root); } - /* - * 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. - */ - void duplicate (unsigned parent_idx, unsigned child_idx) - { - DEBUG_MSG (SUBSET_REPACK, nullptr, " Duplicating %d => %d", - parent_idx, child_idx); - - positions_invalid = true; - - auto* clone = vertices_.push (); - auto& child = vertices_[child_idx]; - clone_buffer_t* buffer = clone_buffers_.push (); - if (vertices_.in_error () - || clone_buffers_.in_error () - || !check_success (buffer->copy (child.obj))) { - return; - } - - clone->obj.head = buffer->head; - clone->obj.tail = buffer->tail; - clone->distance = child.distance; - - for (const auto& l : child.obj.links) - clone->obj.links.push (l); + if (!roots_to_isolate) return false; - check_success (!clone->obj.links.in_error ()); - - auto& parent = vertices_[parent_idx]; - unsigned clone_idx = vertices_.length - 2; - for (unsigned i = 0; i < parent.obj.links.length; i++) - { - auto& l = parent.obj.links[i]; - if (l.objidx == child_idx) - { - l.objidx = clone_idx; - clone->incoming_edges++; - child.incoming_edges--; - } + unsigned maximum_to_move = hb_max ((sorted_graph.num_roots_for_space (space) / 2u), 1u); + if (roots_to_isolate.get_population () > maximum_to_move) { + // 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--) { + uint32_t root = HB_SET_VALUE_INVALID; + roots_to_isolate.previous (&root); + roots_to_isolate.del (root); } - - // 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_[vertices_.length - 2] = *clone; - vertices_[vertices_.length - 1] = root; - } - - /* - * Raises the sorting priority of all children. - */ - void 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; - for (unsigned i = 0; i < parent.links.length; i++) - vertices_[parent.links[i].objidx].raise_priority (); } - /* - * 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 (); + DEBUG_MSG (SUBSET_REPACK, nullptr, + "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 (), + sorted_graph.next_space ()); - for (int parent_idx = vertices_.length - 1; parent_idx >= 0; parent_idx--) - { - for (const auto& link : vertices_[parent_idx].obj.links) - { - int64_t offset = compute_offset (parent_idx, link); - if (is_valid_offset (offset, link)) - continue; + sorted_graph.isolate_subgraph (roots_to_isolate); + sorted_graph.move_to_new_space (roots_to_isolate); - if (!overflows) return true; + return true; +} - overflow_record_t r; - r.parent = parent_idx; - r.link = &link; - overflows->push (r); - } +static inline +bool _resolve_shared_overflow(const hb_vector_t<graph::overflow_record_t>& overflows, + int overflow_index, + graph_t& sorted_graph) +{ + const graph::overflow_record_t& r = overflows[overflow_index]; + + // Find all of the parents in overflowing links that link to this + // same child node. We will then try duplicating the child node and + // re-assigning all of these parents to the duplicate. + hb_set_t parents; + parents.add(r.parent); + for (int i = overflow_index - 1; i >= 0; i--) { + const graph::overflow_record_t& r2 = overflows[i]; + if (r2.child == r.child) { + parents.add(r2.parent); } + } - if (!overflows) return false; - return overflows->length; + unsigned result = sorted_graph.duplicate(&parents, r.child); + if (result == (unsigned) -1 && parents.get_population() > 2) { + // All links to the child are overflowing, so we can't include all + // in the duplication. Remove one parent from the duplication. + // Remove the lowest index parent, which will be the closest to the child. + parents.del(parents.get_min()); + result = sorted_graph.duplicate(&parents, r.child); } - void print_overflows (const hb_vector_t<overflow_record_t>& overflows) - { - if (!DEBUG_ENABLED(SUBSET_REPACK)) return; + if (result == (unsigned) -1) return result; - update_incoming_edge_count (); - for (const auto& o : overflows) - { - const auto& child = vertices_[o.link->objidx]; - DEBUG_MSG (SUBSET_REPACK, nullptr, " overflow from %d => %d (%d incoming , %d outgoing)", - o.parent, - o.link->objidx, - child.incoming_edges, - child.obj.links.length); - } + if (parents.get_population() > 1) { + // If the duplicated node has more than one parent pre-emptively raise it's priority to the maximum. + // This will place it close to the parents. Node's with only one parent, don't need this as normal overflow + // resolution will raise priority if needed. + // + // Reasoning: most of the parents to this child are likely at the same layer in the graph. Duplicating + // the child will theoretically allow it to be placed closer to it's parents. However, due to the shortest + // distance sort by default it's placement will remain in the same layer, thus it will remain in roughly the + // same position (and distance from parents) as the original child node. The overflow resolution will attempt + // to move nodes closer, but only for non-shared nodes. Since this node is shared, it will simply be given + // further duplication which defeats the attempt to duplicate with multiple parents. To fix this we + // pre-emptively raise priority now which allows the duplicated node to pack into the same layer as it's parents. + sorted_graph.vertices_[result].give_max_priority(); } - void err_other_error () { this->successful = false; } - - private: + return result; +} - bool check_success (bool success) - { return this->successful && (success || (err_other_error (), false)); } +static inline +bool _process_overflows (const hb_vector_t<graph::overflow_record_t>& overflows, + hb_set_t& priority_bumped_parents, + graph_t& sorted_graph) +{ + bool resolution_attempted = false; - /* - * Creates a map from objid to # of incoming edges. - */ - void update_incoming_edge_count () + // Try resolving the furthest overflows first. + for (int i = overflows.length - 1; i >= 0; i--) { - if (!edge_count_invalid) return; - - for (unsigned i = 0; i < vertices_.length; i++) - vertices_[i].incoming_edges = 0; - - for (const vertex_t& v : vertices_) + const graph::overflow_record_t& r = overflows[i]; + const auto& child = sorted_graph.vertices_[r.child]; + if (child.is_shared ()) { - for (auto& l : v.obj.links) - { - vertices_[l.objidx].incoming_edges++; - } + // The child object is shared, we may be able to eliminate the overflow + // by duplicating it. + if (!_resolve_shared_overflow(overflows, i, sorted_graph)) continue; + return true; } - edge_count_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--) + if (child.is_leaf () && !priority_bumped_parents.has (r.parent)) { - auto& v = vertices_[i]; - v.start = current_pos; - current_pos += v.obj.tail - v.obj.head; - v.end = current_pos; + // This object is too far from it's parent, attempt to move it closer. + // + // TODO(garretrieger): initially limiting this to leaf's since they can be + // moved closer with fewer consequences. However, this can + // likely can be used for non-leafs as well. + // TODO(garretrieger): also try lowering priority of the parent. Make it + // get placed further up in the ordering, closer to it's children. + // this is probably preferable if the total size of the parent object + // is < then the total size of the children (and the parent can be moved). + // Since in that case moving the parent will cause a smaller increase in + // the length of other offsets. + if (sorted_graph.raise_childrens_priority (r.parent)) { + priority_bumped_parents.add (r.parent); + resolution_attempted = true; + } + continue; } - positions_invalid = false; + // TODO(garretrieger): add additional offset resolution strategies + // - Promotion to extension lookups. + // - Table splitting. } - /* - * 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 fibonaacci 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); - } + return resolution_attempted; +} - hb_priority_queue_t queue; - queue.insert (0, vertices_.length - 1); +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 ()) + { + DEBUG_MSG (SUBSET_REPACK, nullptr, "Sorted graph in error state after initial sort."); + return false; + } - hb_set_t visited; + bool will_overflow = graph::will_overflow (sorted_graph); + if (!will_overflow) + return true; - while (!queue.in_error () && !queue.is_empty ()) + graph::gsubgpos_graph_context_t ext_context (table_tag, sorted_graph); + if ((table_tag == HB_OT_TAG_GPOS + || table_tag == HB_OT_TAG_GSUB) + && will_overflow) + { + if (recalculate_extensions) { - unsigned next_idx = queue.pop_minimum ().second; - if (visited.has (next_idx)) continue; - const auto& next = vertices_[next_idx]; - int64_t next_distance = vertices_[next_idx].distance; - visited.add (next_idx); - - for (const auto& link : next.obj.links) - { - if (visited.has (link.objidx)) continue; - - const auto& child = vertices_[link.objidx].obj; - int64_t child_weight = child.tail - child.head + - ((int64_t) 1 << (link.width * 8)); - 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); - } + 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; } - } - check_success (!queue.in_error ()); - if (!check_success (queue.is_empty ())) - { - DEBUG_MSG (SUBSET_REPACK, nullptr, "Graph is not fully connected."); - return; + 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; + } } - distance_invalid = false; + DEBUG_MSG (SUBSET_REPACK, nullptr, "Assigning spaces to 32 bit subgraphs."); + if (sorted_graph.assign_spaces ()) + sorted_graph.sort_shortest_distance (); + else + sorted_graph.sort_shortest_distance_if_needed (); } - int64_t compute_offset ( - unsigned parent_idx, - const hb_serialize_context_t::object_t::link_t& link) const - { - 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; - } + unsigned round = 0; + hb_vector_t<graph::overflow_record_t> overflows; + // TODO(garretrieger): select a good limit for max rounds. + while (!sorted_graph.in_error () + && graph::will_overflow (sorted_graph, &overflows) + && round < max_rounds) { + DEBUG_MSG (SUBSET_REPACK, nullptr, "=== Overflow resolution round %u ===", round); + print_overflows (sorted_graph, overflows); - bool is_valid_offset (int64_t offset, - const hb_serialize_context_t::object_t::link_t& link) const - { - 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); - } - } + hb_set_t priority_bumped_parents; - /* - * Updates all objidx's in all links using the provided mapping. - */ - void remap_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 (!_try_isolating_subgraphs (overflows, sorted_graph)) { - for (unsigned j = 0; j < (*sorted_graph)[i].obj.links.length; j++) + // 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)) { - auto& link = (*sorted_graph)[i].obj.links[j]; - link.objidx = id_map[link.objidx]; + DEBUG_MSG (SUBSET_REPACK, nullptr, "No resolution available :("); + break; } } + + sorted_graph.sort_shortest_distance (); } - 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 + if (sorted_graph.in_error ()) { - 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); + DEBUG_MSG (SUBSET_REPACK, nullptr, "Sorted graph in error state."); + return false; } - void serialize_link (const hb_serialize_context_t::object_t::link_t& link, - char* head, - hb_serialize_context_t* c) const + if (graph::will_overflow (sorted_graph)) { - switch (link.width) - { - 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); - } + DEBUG_MSG (SUBSET_REPACK, nullptr, "Offset overflow resolution failed."); + 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: - hb_vector_t<clone_buffer_t> clone_buffers_; - bool edge_count_invalid; - bool distance_invalid; - bool positions_invalid; - bool successful; -}; - + return true; +} /* * Attempts to modify the topological sorting of the provided object graph to @@ -677,93 +428,41 @@ struct graph_t * 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 */ -inline void -hb_resolve_overflows (const hb_vector_t<hb_serialize_context_t::object_t *>& packed, - hb_serialize_context_t* c) { - // Kahn sort is ~twice as fast as shortest distance sort and works for many fonts - // so try it first to save time. +template<typename T> +inline hb_blob_t* +hb_resolve_overflows (const T& packed, + hb_tag_t table_tag, + unsigned max_rounds = 32, + bool recalculate_extensions = false) { graph_t sorted_graph (packed); - sorted_graph.sort_kahn (); - if (!sorted_graph.will_overflow ()) + if (sorted_graph.in_error ()) { - sorted_graph.serialize (c); - return; + // Invalid graph definition. + return nullptr; } - sorted_graph.sort_shortest_distance (); - - unsigned round = 0; - hb_vector_t<graph_t::overflow_record_t> overflows; - // TODO(garretrieger): select a good limit for max rounds. - while (!sorted_graph.in_error () - && sorted_graph.will_overflow (&overflows) - && round++ < 10) { - DEBUG_MSG (SUBSET_REPACK, nullptr, "=== Over flow resolution round %d ===", round); - sorted_graph.print_overflows (overflows); - - bool resolution_attempted = false; - hb_set_t priority_bumped_parents; - // 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 auto& child = sorted_graph.vertices_[r.link->objidx]; - if (child.is_shared ()) - { - // The child object is shared, we may be able to eliminate the overflow - // by duplicating it. - sorted_graph.duplicate (r.parent, r.link->objidx); - resolution_attempted = true; - - // Stop processing overflows for this round so that object order can be - // updated to account for the newly added object. - break; - } - - if (child.is_leaf () && !priority_bumped_parents.has (r.parent)) - { - // This object is too far from it's parent, attempt to move it closer. - // - // TODO(garretrieger): initially limiting this to leaf's since they can be - // moved closer with fewer consequences. However, this can - // likely can be used for non-leafs as well. - // TODO(garretrieger): add a maximum priority, don't try to raise past this. - // TODO(garretrieger): also try lowering priority of the parent. Make it - // get placed further up in the ordering, closer to it's children. - // this is probably preferable if the total size of the parent object - // is < then the total size of the children (and the parent can be moved). - // Since in that case moving the parent will cause a smaller increase in - // the length of other offsets. - sorted_graph.raise_childrens_priority (r.parent); - priority_bumped_parents.add (r.parent); - resolution_attempted = true; - continue; - } - - // TODO(garretrieger): add additional offset resolution strategies - // - Promotion to extension lookups. - // - Table splitting. - } - - if (resolution_attempted) - { - sorted_graph.sort_shortest_distance (); - continue; - } - - DEBUG_MSG (SUBSET_REPACK, nullptr, "No resolution available :("); - c->err (HB_SERIALIZE_ERROR_OFFSET_OVERFLOW); - return; + if (!sorted_graph.is_fully_connected ()) + { + sorted_graph.print_orphaned_nodes (); + return nullptr; } if (sorted_graph.in_error ()) { - c->err (HB_SERIALIZE_ERROR_OTHER); - return; + // Allocations failed somewhere + DEBUG_MSG (SUBSET_REPACK, nullptr, + "Graph is in error, likely due to a memory allocation error."); + return nullptr; } - sorted_graph.serialize (c); -} + if (!hb_resolve_graph_overflows (table_tag, max_rounds, recalculate_extensions, sorted_graph)) + return nullptr; + + return graph::serialize (sorted_graph); +} #endif /* HB_REPACKER_HH */ |