diff options
Diffstat (limited to 'src/3rdparty/harfbuzz-ng/src/hb-repacker.hh')
-rw-r--r-- | src/3rdparty/harfbuzz-ng/src/hb-repacker.hh | 918 |
1 files changed, 712 insertions, 206 deletions
diff --git a/src/3rdparty/harfbuzz-ng/src/hb-repacker.hh b/src/3rdparty/harfbuzz-ng/src/hb-repacker.hh index b02128b5c4..2a9e75c45b 100644 --- a/src/3rdparty/harfbuzz-ng/src/hb-repacker.hh +++ b/src/3rdparty/harfbuzz-ng/src/hb-repacker.hh @@ -33,40 +33,71 @@ #include "hb-serialize.hh" #include "hb-vector.hh" - +/* + * For a detailed writeup on the overflow resolution algorithm see: + * docs/repacker.md + */ 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; + 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 incoming_edges > 1; + 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.links.length; + return !obj.real_links.length && !obj.virtual_links.length; } - void raise_priority () + 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 @@ -76,49 +107,29 @@ struct graph_t // it's parent where possible. int64_t modified_distance = - hb_min (hb_max(distance + distance_modifier (), 0), 0x7FFFFFFFFF); - return (modified_distance << 24) | (0x00FFFFFF & order); + 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; - return -(table_size - table_size / (1 << hb_min(priority, 16u))); + + if (priority == 1) + return -table_size / 2; + + return -table_size; } }; struct overflow_record_t { unsigned parent; - const hb_serialize_context_t::object_t::link_t* link; - }; - - struct clone_buffer_t - { - 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; - - void fini () - { - if (!head) return; - hb_free (head); - head = nullptr; - } + unsigned child; }; /* @@ -128,12 +139,14 @@ struct graph_t * 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), + 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++) { @@ -151,21 +164,23 @@ struct graph_t 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--; + // Fix indices to account for removed nil object. + for (auto& l : v->obj.all_links_writer ()) { + l.objidx--; + } } } ~graph_t () { - vertices_.fini_deep (); - clone_buffers_.fini_deep (); + vertices_.fini (); } bool in_error () const { - return !successful || vertices_.in_error () || clone_buffers_.in_error (); + return !successful || + vertices_.in_error () || + num_roots_for_space_.in_error (); } const vertex_t& root () const @@ -189,26 +204,46 @@ struct graph_t /* * serialize graph into the provided serialization buffer. */ - void serialize (hb_serialize_context_t* c) const + hb_blob_t* serialize () const { - c->start_serialize<void> (); + 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 (); + c.push (); size_t size = vertices_[i].obj.tail - vertices_[i].obj.head; - char* start = c->allocate_size <char> (size); - if (!start) return; + 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); - for (const auto& link : vertices_[i].obj.links) - serialize_link (link, start, c); + // 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.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; } - c->end_serialize (); + + return c.copy_blob (); } /* @@ -226,12 +261,13 @@ struct graph_t 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_incoming_edge_count (); + update_parents (); queue.push (root_idx ()); int new_id = vertices_.length - 1; @@ -242,12 +278,12 @@ struct graph_t queue.remove (0); vertex_t& next = vertices_[next_id]; - sorted_graph.push (next); + sorted_graph[new_id] = next; id_map[next_id] = new_id--; - for (const auto& link : next.obj.links) { + for (const auto& link : next.obj.all_links ()) { removed_edges[link.objidx]++; - if (!(vertices_[link.objidx].incoming_edges - removed_edges[link.objidx])) + if (!(vertices_[link.objidx].incoming_edges () - removed_edges[link.objidx])) queue.push (link.objidx); } } @@ -255,15 +291,12 @@ struct graph_t 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); + print_orphaned_nodes (); - sorted_graph.as_array ().reverse (); + remap_all_obj_indices (id_map, &sorted_graph); - vertices_.fini_deep (); - vertices_ = sorted_graph; - sorted_graph.fini_deep (); + hb_swap (vertices_, sorted_graph); + sorted_graph.fini (); } /* @@ -283,12 +316,13 @@ struct graph_t 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_incoming_edge_count (); + update_parents (); queue.insert (root ().modified_distance (0), root_idx ()); int new_id = root_idx (); @@ -298,12 +332,12 @@ struct graph_t unsigned next_id = queue.pop_minimum().second; vertex_t& next = vertices_[next_id]; - sorted_graph.push (next); + sorted_graph[new_id] = next; id_map[next_id] = new_id--; - for (const auto& link : next.obj.links) { + for (const auto& link : next.obj.all_links ()) { removed_edges[link.objidx]++; - if (!(vertices_[link.objidx].incoming_edges - 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 @@ -317,72 +351,282 @@ struct graph_t 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."); + 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. + } - remap_obj_indices (id_map, &sorted_graph); - sorted_graph.as_array ().reverse (); - vertices_.fini_deep (); - vertices_ = sorted_graph; - sorted_graph.fini_deep (); + return true; } /* - * 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. + * 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. */ - void duplicate (unsigned parent_idx, unsigned child_idx) + bool isolate_subgraph (hb_set_t& roots) { - DEBUG_MSG (SUBSET_REPACK, nullptr, " Duplicating %d => %d", - parent_idx, child_idx); + update_parents (); + hb_hashmap_t<unsigned, unsigned> subgraph; - positions_invalid = true; + // 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); + } - 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; + 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); + } } - clone->obj.head = buffer->head; - clone->obj.tail = buffer->tail; - clone->distance = child.distance; + if (!made_changes) + return false; - for (const auto& l : child.obj.links) - clone->obj.links.push (l); + 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); + } - check_success (!clone->obj.links.in_error ()); + auto new_subgraph = + + subgraph.keys () + | hb_map([&] (unsigned node_idx) { + if (index_map.has (node_idx)) return index_map[node_idx]; + return node_idx; + }) + ; - auto& parent = vertices_[parent_idx]; - unsigned clone_idx = vertices_.length - 2; - for (unsigned i = 0; i < parent.obj.links.length; i++) + 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 ()) { - auto& l = parent.obj.links[i]; - if (l.objidx == child_idx) + if (subgraph.has (link.objidx)) { - l.objidx = clone_idx; - clone->incoming_edges++; - child.incoming_edges--; + subgraph.set (link.objidx, subgraph[link.objidx] + 1); + continue; } + subgraph.set (link.objidx, 1); + find_subgraph (link.objidx, subgraph); } + } + + void find_subgraph (unsigned node_idx, hb_set_t& subgraph) + { + 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); + } + + /* + * 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) + { + 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); + } + 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_[vertices_.length - 2] = *clone; + 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; + } + + /* + * 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); + 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. */ - void raise_childrens_priority (unsigned parent_idx) + bool raise_childrens_priority (unsigned parent_idx) { DEBUG_MSG (SUBSET_REPACK, nullptr, " Raising priority of all children of %d", parent_idx); @@ -390,8 +634,10 @@ struct graph_t // 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 (); + bool made_change = false; + for (auto& l : parent.all_links_writer ()) + made_change |= vertices_[l.objidx].raise_priority (); + return made_change; } /* @@ -404,7 +650,8 @@ struct graph_t for (int parent_idx = vertices_.length - 1; parent_idx >= 0; parent_idx--) { - for (const auto& link : vertices_[parent_idx].obj.links) + // 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)) @@ -414,7 +661,7 @@ struct graph_t overflow_record_t r; r.parent = parent_idx; - r.link = &link; + r.child = link.objidx; overflows->push (r); } } @@ -423,48 +670,156 @@ struct graph_t 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_incoming_edge_count (); + update_parents (); + int limit = 10; 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)", + 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, - o.link->objidx, - child.incoming_edges, - child.obj.links.length); + 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; + } + + /* + * Returns the numbers of incoming edges that are 32bits wide. + */ + unsigned wide_parents (unsigned node_idx, hb_set_t& parents) const + { + unsigned count = 0; + 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_incoming_edge_count () + void update_parents () { - if (!edge_count_invalid) return; + if (!parents_invalid) return; for (unsigned i = 0; i < vertices_.length; i++) - vertices_[i].incoming_edges = 0; + vertices_[i].parents.reset (); - for (const vertex_t& v : vertices_) + for (unsigned p = 0; p < vertices_.length; p++) { - for (auto& l : v.obj.links) + for (auto& l : vertices_[p].obj.all_links ()) { - vertices_[l.objidx].incoming_edges++; + vertices_[l.objidx].parents.push (p); } } - edge_count_invalid = false; + parents_invalid = false; } /* @@ -503,7 +858,7 @@ struct graph_t // 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. + // (such as a fibonacci queue) with a fast decrease priority. for (unsigned i = 0; i < vertices_.length; i++) { if (i == vertices_.length - 1) @@ -515,23 +870,25 @@ struct graph_t hb_priority_queue_t queue; queue.insert (0, vertices_.length - 1); - hb_set_t visited; + 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.has (next_idx)) continue; + if (visited[next_idx]) continue; const auto& next = vertices_[next_idx]; int64_t next_distance = vertices_[next_idx].distance; - visited.add (next_idx); + visited[next_idx] = true; - for (const auto& link : next.obj.links) + for (const auto& link : next.obj.all_links ()) { - if (visited.has (link.objidx)) continue; + if (visited[link.objidx]) continue; const auto& child = vertices_[link.objidx].obj; - int64_t child_weight = child.tail - child.head + - ((int64_t) 1 << (link.width * 8)); + 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) @@ -545,7 +902,7 @@ struct graph_t check_success (!queue.in_error ()); if (!check_success (queue.is_empty ())) { - DEBUG_MSG (SUBSET_REPACK, nullptr, "Graph is not fully connected."); + print_orphaned_nodes (); return; } @@ -576,6 +933,10 @@ struct graph_t bool is_valid_offset (int64_t offset, const hb_serialize_context_t::object_t::link_t& link) const { + if (unlikely (!link.width)) + // Virtual links can't overflow. + return link.is_signed || offset >= 0; + if (link.is_signed) { if (link.width == 4) @@ -595,16 +956,51 @@ struct graph_t } /* + * 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]); + } + } + } + + /* * 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 + 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++) { - for (unsigned j = 0; j < (*sorted_graph)[i].obj.links.length; j++) + (*sorted_graph)[i].remap_parents (id_map); + for (auto& link : (*sorted_graph)[i].obj.all_links_writer ()) { - auto& link = (*sorted_graph)[i].obj.links[j]; link.objidx = id_map[link.objidx]; } } @@ -631,6 +1027,9 @@ struct graph_t { switch (link.width) { + case 0: + // Virtual links aren't serialized. + return; case 4: if (link.is_signed) { @@ -656,17 +1055,146 @@ struct graph_t } } + /* + * 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 (targets.has (start_idx)) + { + targets.del (start_idx); + connected.add (start_idx); + } + + 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); + } + 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 parents_invalid; bool distance_invalid; bool positions_invalid; bool successful; + hb_vector_t<unsigned> num_roots_for_space_; }; +static inline +bool _try_isolating_subgraphs (const hb_vector_t<graph_t::overflow_record_t>& overflows, + graph_t& sorted_graph) +{ + unsigned space = 0; + hb_set_t roots_to_isolate; + + for (int i = overflows.length - 1; i >= 0; i--) + { + const graph_t::overflow_record_t& r = overflows[i]; + + 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; + + if (!space) { + space = overflow_space; + } + + if (space == overflow_space) + roots_to_isolate.add(root); + } + + if (!roots_to_isolate) return false; + + 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--) { + unsigned 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.", + space, + sorted_graph.num_roots_for_space (space), + roots_to_isolate.get_population (), + sorted_graph.next_space ()); + + sorted_graph.isolate_subgraph (roots_to_isolate); + sorted_graph.move_to_new_space (roots_to_isolate); + + return true; +} + +static inline +bool _process_overflows (const hb_vector_t<graph_t::overflow_record_t>& overflows, + hb_set_t& priority_bumped_parents, + graph_t& sorted_graph) +{ + bool resolution_attempted = false; + + // 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.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; + return true; + } + + 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): 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; + } + + // TODO(garretrieger): add additional offset resolution strategies + // - Promotion to extension lookups. + // - Table splitting. + } + + return resolution_attempted; +} /* * Attempts to modify the topological sorting of the provided object graph to @@ -677,93 +1205,71 @@ 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) { +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 ()) { - sorted_graph.serialize (c); - return; + return sorted_graph.serialize (); } sorted_graph.sort_shortest_distance (); + if ((table_tag == HB_OT_TAG_GPOS + || table_tag == HB_OT_TAG_GSUB) + && sorted_graph.will_overflow ()) + { + DEBUG_MSG (SUBSET_REPACK, nullptr, "Assigning spaces to 32 bit subgraphs."); + if (sorted_graph.assign_32bit_spaces ()) + 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); + && round++ < max_rounds) { + DEBUG_MSG (SUBSET_REPACK, nullptr, "=== Overflow 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--) + + if (!_try_isolating_subgraphs (overflows, sorted_graph)) { - const graph_t::overflow_record_t& r = overflows[i]; - const auto& child = sorted_graph.vertices_[r.link->objidx]; - if (child.is_shared ()) + if (!_process_overflows (overflows, priority_bumped_parents, sorted_graph)) { - // 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. + DEBUG_MSG (SUBSET_REPACK, nullptr, "No resolution available :("); 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; + sorted_graph.sort_shortest_distance (); } if (sorted_graph.in_error ()) { - c->err (HB_SERIALIZE_ERROR_OTHER); - return; + DEBUG_MSG (SUBSET_REPACK, nullptr, "Sorted graph in error state."); + return nullptr; } - sorted_graph.serialize (c); -} + if (sorted_graph.will_overflow ()) + { + DEBUG_MSG (SUBSET_REPACK, nullptr, "Offset overflow resolution failed."); + return nullptr; + } + + return sorted_graph.serialize (); +} #endif /* HB_REPACKER_HH */ |