more optimization

CMakeLists.txt

@@ -31,10 +31,10 @@ elseif(CMAKE_BUILD_TYPE STREQUAL Debug)
 	add_compile_options(-g -Og)
 elseif(CMAKE_BUILD_TYPE STREQUAL DebugRelease)
 	message(STATUS "Build mode: DebugRelease")
-	add_compile_options(-g -O3)
+	add_compile_options(-O3 -flto -DNDEBUG -g)
 elseif(CMAKE_BUILD_TYPE STREQUAL Release)
 	message(STATUS "Build mode: Release")
-	add_compile_options(-O3 -flto)
+	add_compile_options(-O3 -flto -DNDEBUG)
 	set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -flto")
 else()
 	message(FATAL_ERROR "CMAKE_BUILD_TYPE must be Debug or Release.")

libswan/include/swan/Chunk.h

@@ -28,10 +28,25 @@ public:
 		data_.reset(new uint8_t[CHUNK_WIDTH * CHUNK_HEIGHT * sizeof(Tile::ID)]);
 	}
 
-	Tile::ID *getTileData();
-	Tile::ID getTileID(RelPos pos);
-	void setTileID(RelPos pos, Tile::ID id, SDL_Texture *tex);
-	void setTileData(RelPos pos, Tile::ID id);
+	Tile::ID *getTileData() {
+		assert(isActive());
+		return (Tile::ID *)data_.get();
+	}
+
+	Tile::ID getTileID(RelPos pos) {
+		return getTileData()[pos.y * CHUNK_WIDTH + pos.x];
+	}
+
+	void setTileID(RelPos pos, Tile::ID id, SDL_Texture *tex) {
+		getTileData()[pos.y * CHUNK_WIDTH + pos.x] = id;
+		draw_list_.push_back({ pos, tex });
+	}
+
+	void setTileData(RelPos pos, Tile::ID id) {
+		getTileData()[pos.y * CHUNK_WIDTH + pos.x] = id;
+		need_render_ = true;
+	}
+
 	void render(const Context &ctx, SDL_Renderer *rnd);
 
 	void compress();

libswan/include/swan/World.h

@@ -31,7 +31,7 @@ public:
 	WorldPlane &addPlane(const std::string &gen);
 	WorldPlane &addPlane() { return addPlane(default_world_gen_); }
 
-	Tile &getTileByID(Tile::ID id);
+	Tile &getTileByID(Tile::ID id) { return *tiles_[id]; }
 	Tile::ID getTileID(const std::string &name);
 	Tile &getTile(const std::string &name);
 	Item &getItem(const std::string &name);

libswan/include/swan/WorldPlane.h

@@ -35,8 +35,10 @@ public:
 
 	bool hasChunk(ChunkPos pos);
 	Chunk &getChunk(ChunkPos pos);
+	Chunk &slowGetChunk(ChunkPos pos);
 	void setTileID(TilePos pos, Tile::ID id);
 	void setTile(TilePos pos, const std::string &name);
+
 	Tile::ID getTileID(TilePos pos);
 	Tile &getTile(TilePos pos);
 
@@ -68,6 +70,7 @@ public:
 private:
 	std::map<std::pair<int, int>, Chunk> chunks_;
 	std::vector<Chunk *> active_chunks_;
+	std::vector<std::pair<ChunkPos, Chunk *>> tick_chunks_;
 	std::vector<std::unique_ptr<Entity>> entities_;
 
 	std::deque<Chunk *> chunk_init_list_;

libswan/src/Chunk.cc

@@ -15,25 +15,6 @@ namespace Swan {
 
 uint8_t *Chunk::renderbuf = new uint8_t[CHUNK_WIDTH * TILE_SIZE * CHUNK_HEIGHT * TILE_SIZE * 4];
 
-Tile::ID *Chunk::getTileData() {
-	assert(isActive());
-	return (Tile::ID *)data_.get();
-}
-
-Tile::ID Chunk::getTileID(RelPos pos) {
-	return getTileData()[pos.y * CHUNK_WIDTH + pos.x];
-}
-
-void Chunk::setTileID(RelPos pos, Tile::ID id, SDL_Texture *tex) {
-	getTileData()[pos.y * CHUNK_WIDTH + pos.x] = id;
-	draw_list_.push_back({ pos, tex });
-}
-
-void Chunk::setTileData(RelPos pos, Tile::ID id) {
-	getTileData()[pos.y * CHUNK_WIDTH + pos.x] = id;
-	need_render_ = true;
-}
-
 void Chunk::compress() {
 	if (isCompressed())
 		return;

libswan/src/World.cc

@@ -11,7 +11,7 @@ namespace Swan {
 
 static void chunkLine(int l, WorldPlane &plane, ChunkPos &abspos, const Vec2i &dir) {
 	for (int i = 0; i < l; ++i) {
-		plane.getChunk(abspos);
+		plane.slowGetChunk(abspos).keepActive();
 		abspos += dir;
 	}
 }
@@ -122,10 +122,6 @@ Tile::ID World::getTileID(const std::string &name) {
 	return iter->second;
 }
 
-Tile &World::getTileByID(Tile::ID id) {
-	return *tiles_[id];
-}
-
 Tile &World::getTile(const std::string &name) {
 	Tile::ID id = getTileID(name);
 	return getTileByID(id);

libswan/src/WorldPlane.cc

@@ -71,35 +71,18 @@ bool WorldPlane::hasChunk(ChunkPos pos) {
 
 // This function will be a bit weird because it's a really fucking hot function.
 Chunk &WorldPlane::getChunk(ChunkPos pos) {
-
-	// This branch should be really predictable, there should basically never
-	// be no active chunks
-	if (active_chunks_.size() != 0) {
-
-		// The last chunk should be the most actively used chunk
-		Chunk *chunk = active_chunks_.back();
-		if (chunk->pos_ == pos) {
-			chunk->keepActive();
+	// First, look through all chunks which have been in use this tick
+	for (auto [chpos, chunk]: tick_chunks_) {
+		if (chpos == pos)
 			return *chunk;
-		}
-
-		// Linear search through a small array is probably faster than tree lookup.
-		// Loop backwards because the hottest chunks should be at the end.
-		for (ssize_t i = active_chunks_.size() - 2; i >= 0; --i) {
-			chunk = active_chunks_[i];
-			if (chunk->pos_ == pos) {
-				chunk->keepActive();
-
-				// Ensure that the hot chunk is at the end by bubbling this one up
-				active_chunks_[i] = active_chunks_[i + 1];
-				active_chunks_[i + 1] = chunk;
-				return *chunk;
-			}
-		}
 	}
 
-	// Slow path: Find a chunk in our global chunk map,
-	// creating a new one if necessary
+	Chunk &chunk = slowGetChunk(pos);
+	tick_chunks_.push_back({ pos, &chunk });
+	return chunk;
+}
+
+Chunk &WorldPlane::slowGetChunk(ChunkPos pos) {
 	auto iter = chunks_.find(pos);
 
 	// Create chunk if that turns out to be necessary
@@ -111,8 +94,8 @@ Chunk &WorldPlane::getChunk(ChunkPos pos) {
 		active_chunks_.push_back(&chunk);
 		chunk_init_list_.push_back(&chunk);
 
-	// Otherwise, tell it that it should keep itself alive a bit longer
-	} else {
+	// Otherwise, it might not be active, so let's activate it
+	} else if (!iter->second.isActive()) {
 		iter->second.keepActive();
 		active_chunks_.push_back(&iter->second);
 		chunk_init_list_.push_back(&iter->second);
@@ -137,9 +120,7 @@ void WorldPlane::setTile(TilePos pos, const std::string &name) {
 }
 
 Tile::ID WorldPlane::getTileID(TilePos pos) {
-	Chunk &chunk = getChunk(chunkPos(pos));
-
-	return chunk.getTileID(relPos(pos));
+	return getChunk(chunkPos(pos)).getTileID(relPos(pos));
 }
 
 Tile &WorldPlane::getTile(TilePos pos) {
@@ -255,6 +236,11 @@ void WorldPlane::update(float dt) {
 }
 
 void WorldPlane::tick(float dt) {
+	// Any chunk which has been in use since last tick should be kept alive
+	for (std::pair<ChunkPos, Chunk *> &ch: tick_chunks_)
+		ch.second->keepActive();
+	tick_chunks_.clear();
+
 	for (auto &ent: entities_)
 		ent->tick(getContext(), dt);