Monoblocks: optimize blocks that contain a single type of node (#16293)

Reduces memory usage on the server, especially with many user and/or large viewing distances.
Currently disabled on the client due to known data races on a block's data.

src/client/client.h

@@ -371,6 +371,7 @@ public:
 	scene::ISceneManager *getSceneManager();
 
 	// IGameDef interface
+	bool isClient() override { return true; }
 	IItemDefManager* getItemDefManager() override;
 	const NodeDefManager* getNodeDefManager() override;
 	ICraftDefManager* getCraftDefManager() override;

src/dummygamedef.h

@@ -30,6 +30,7 @@ public:
 		delete m_itemdef;
 	}
 
+	bool isClient() override { return false; }
 	IItemDefManager *getItemDefManager() override { return m_itemdef; }
 	const NodeDefManager *getNodeDefManager() override { return m_nodedef; }
 	NodeDefManager* getWritableNodeDefManager() { return m_nodedef; }

src/gamedef.h

@@ -61,4 +61,5 @@ public:
 	virtual bool sendModChannelMessage(const std::string &channel,
 		const std::string &message) = 0;
 	virtual ModChannel *getModChannel(const std::string &channel) = 0;
+	virtual bool isClient() = 0;
 };

src/mapblock.cpp

@@ -4,6 +4,7 @@
 
 #include "mapblock.h"
 
+#include <memory>
 #include <sstream>
 #include "map.h"
 #include "light.h"
@@ -101,11 +102,13 @@ static const char *modified_reason_strings[] = {
 MapBlock::MapBlock(v3s16 pos, IGameDef *gamedef):
 		m_pos(pos),
 		m_pos_relative(pos * MAP_BLOCKSIZE),
-		data(new MapNode[nodecount]),
+		m_gamedef(gamedef),
-		m_gamedef(gamedef)
+		m_is_mono_block(false)
 {
-	reallocate();
+	// We start with nodecount nodes, because in the vast
-	assert(m_modified > MOD_STATE_CLEAN);
+	// majority of the cases a block is created just before
+	// it is de-serialized or generated.
+	reallocate(nodecount, MapNode(CONTENT_IGNORE));
 }
 
 MapBlock::~MapBlock()
@@ -118,6 +121,7 @@ MapBlock::~MapBlock()
 #endif
 
 	delete[] data;
+	if (!m_is_mono_block)
 		porting::TrackFreedMemory(sizeof(MapNode) * nodecount);
 }
 
@@ -214,7 +218,7 @@ void MapBlock::copyTo(VoxelManipulator &dst)
 	VoxelArea data_area(v3s16(0,0,0), data_size - v3s16(1,1,1));
 
 	// Copy from data to VoxelManipulator
-	dst.copyFrom(data, data_area, v3s16(0,0,0),
+	dst.copyFrom(data, m_is_mono_block, data_area, v3s16(0,0,0),
 			getPosRelative(), data_size);
 }
 
@@ -223,9 +227,61 @@ void MapBlock::copyFrom(const VoxelManipulator &src)
 	v3s16 data_size(MAP_BLOCKSIZE, MAP_BLOCKSIZE, MAP_BLOCKSIZE);
 	VoxelArea data_area(v3s16(0,0,0), data_size - v3s16(1,1,1));
 
+	expandNodesIfNeeded();
 	// Copy from VoxelManipulator to data
 	src.copyTo(data, data_area, v3s16(0,0,0),
 			getPosRelative(), data_size);
+	tryShrinkNodes();
+}
+
+void MapBlock::reallocate(u32 count, MapNode n)
+{
+	assert(count == 1 || count == nodecount);
+	// For now monoblocks are disabled on the client.
+	// The client has known data races on the block's data (FIXME).
+	assert(!m_gamedef->isClient() || count == nodecount);
+
+	delete[] data;
+	if (data && !m_is_mono_block && count == 1)
+		porting::TrackFreedMemory(sizeof(MapNode) * nodecount);
+
+	data = new MapNode[count];
+	std::fill_n(data, count, n);
+
+	m_is_mono_block = (count == 1);
+}
+
+void MapBlock::tryShrinkNodes()
+{
+	// For now monoblocks are disabled on the client.
+	// The client has known data races on the block's data (FIXME).
+	if (m_gamedef->isClient())
+		return;
+
+	if (m_is_mono_block)
+		return;
+
+	MapNode n = data[0];
+	bool is_mono_block = true;
+	for (u32 i=1; i<nodecount; i++) {
+		if (n != data[i]) {
+			is_mono_block = false;
+			break;
+		}
+	}
+
+	if (is_mono_block) {
+		reallocate(1, n);
+		m_is_air = n.getContent() == CONTENT_AIR;
+		m_is_air_expired = false;
+	}
+}
+
+void MapBlock::expandNodesIfNeeded()
+{
+	if (m_is_mono_block) {
+		reallocate(nodecount, data[0]);
+	}
 }
 
 void MapBlock::actuallyUpdateIsAir()
@@ -233,6 +289,10 @@ void MapBlock::actuallyUpdateIsAir()
 	// Running this function un-expires m_is_air
 	m_is_air_expired = false;
 
+	if (m_is_mono_block) {
+		m_is_air = data[0].getContent() == CONTENT_AIR;
+		return;
+	}
 	bool only_air = true;
 	for (u32 i = 0; i < nodecount; i++) {
 		MapNode &n = data[i];
@@ -260,12 +320,12 @@ void MapBlock::expireIsAirCache()
 // Note that there's no technical reason why we *have to* renumber the IDs,
 // but we do it anyway as it also helps compressability.
 void MapBlock::getBlockNodeIdMapping(NameIdMapping *nimap, MapNode *nodes,
-	const NodeDefManager *nodedef)
+	u32 count, const NodeDefManager *nodedef)
 {
 	IdIdMapping &mapping = IdIdMapping::giveClearedThreadLocalInstance();
 
 	content_t id_counter = 0;
-	for (u32 i = 0; i < MapBlock::nodecount; i++) {
+	for (u32 i = 0; i < count; i++) {
 		content_t global_id = nodes[i].getContent();
 		content_t id = CONTENT_IGNORE;
 
@@ -378,13 +438,13 @@ void MapBlock::serialize(std::ostream &os_compressed, u8 version, bool disk, int
 	const u8 params_width = 2;
 	if(disk)
 	{
-		MapNode *tmp_nodes = new MapNode[nodecount];
+		const size_t size = m_is_mono_block ? 1 : nodecount;
-		memcpy(tmp_nodes, data, nodecount * sizeof(MapNode));
+		std::unique_ptr<MapNode[]> tmp_nodes(new MapNode[size]);
-		getBlockNodeIdMapping(&nimap, tmp_nodes, m_gamedef->ndef());
+		std::copy_n(data, size, tmp_nodes.get());
+		getBlockNodeIdMapping(&nimap, tmp_nodes.get(), size, m_gamedef->ndef());
 
-		buf = MapNode::serializeBulk(version, tmp_nodes, nodecount,
+		buf = MapNode::serializeBulk(version, tmp_nodes.get(), nodecount,
-				content_width, params_width);
+				content_width, params_width, m_is_mono_block);
-		delete[] tmp_nodes;
 
 		// write timestamp and node/id mapping first
 		if (version >= 29) {
@@ -396,7 +456,7 @@ void MapBlock::serialize(std::ostream &os_compressed, u8 version, bool disk, int
 	else
 	{
 		buf = MapNode::serializeBulk(version, data, nodecount,
-				content_width, params_width);
+				content_width, params_width, m_is_mono_block);
 	}
 
 	writeU8(os, content_width);
@@ -465,6 +525,7 @@ void MapBlock::deSerialize(std::istream &in_compressed, u8 version, bool disk)
 	TRACESTREAM(<<"MapBlock::deSerialize "<<getPos()<<std::endl);
 
 	m_is_air_expired = true;
+	expandNodesIfNeeded();
 
 	if(version <= 21)
 	{
@@ -593,10 +654,14 @@ void MapBlock::deSerialize(std::istream &in_compressed, u8 version, bool disk)
 			m_node_timers.deSerialize(is, version);
 		}
 
+		if (nimap.size() == 1) {
+			tryShrinkNodes();
+
 			u16 dummy;
-		m_is_air = nimap.size() == 1 && nimap.getId("air", dummy);
+			m_is_air = nimap.getId("air", dummy);
 			m_is_air_expired = false;
 		}
+	}
 
 	TRACESTREAM(<<"MapBlock::deSerialize "<<getPos()
 			<<": Done."<<std::endl);

src/mapblock.h

@@ -22,6 +22,7 @@ class IGameDef;
 class MapBlockMesh;
 class VoxelManipulator;
 class NameIdMapping;
+class TestMapBlock;
 
 #define BLOCK_TIMESTAMP_UNDEFINED 0xffffffff
 
@@ -30,7 +31,7 @@ class NameIdMapping;
 ////
 
 enum ModReason : u32 {
-	MOD_REASON_REALLOCATE                 = 1 << 0,
+//	UNUSED                                = 1 << 0,
 	MOD_REASON_SET_IS_UNDERGROUND         = 1 << 1,
 	MOD_REASON_SET_LIGHTING_COMPLETE      = 1 << 2,
 	MOD_REASON_SET_GENERATED              = 1 << 3,
@@ -74,13 +75,6 @@ public:
 		m_orphan = true;
 	}
 
-	void reallocate()
-	{
-		for (u32 i = 0; i < nodecount; i++)
-			data[i] = MapNode(CONTENT_IGNORE);
-		raiseModified(MOD_STATE_WRITE_NEEDED, MOD_REASON_REALLOCATE);
-	}
-
 	////
 	//// Modification tracking methods
 	////
@@ -234,7 +228,7 @@ public:
 		if (!*valid_position)
 			return {CONTENT_IGNORE};
 
-		return data[z * zstride + y * ystride + x];
+		return data[m_is_mono_block ? 0 : z * zstride + y * ystride + x];
 	}
 
 	inline MapNode getNode(v3s16 p, bool *valid_position)
@@ -253,6 +247,7 @@ public:
 		if (!isValidPosition(x, y, z))
 			throw InvalidPositionException();
 
+		expandNodesIfNeeded();
 		data[z * zstride + y * ystride + x] = n;
 		raiseModified(MOD_STATE_WRITE_NEEDED, MOD_REASON_SET_NODE);
 	}
@@ -268,7 +263,7 @@ public:
 
 	inline MapNode getNodeNoCheck(s16 x, s16 y, s16 z)
 	{
-		return data[z * zstride + y * ystride + x];
+		return data[m_is_mono_block ? 0 : z * zstride + y * ystride + x];
 	}
 
 	inline MapNode getNodeNoCheck(v3s16 p)
@@ -278,6 +273,7 @@ public:
 
 	inline void setNodeNoCheck(s16 x, s16 y, s16 z, MapNode n)
 	{
+		expandNodesIfNeeded();
 		data[z * zstride + y * ystride + x] = n;
 		raiseModified(MOD_STATE_WRITE_NEEDED, MOD_REASON_SET_NODE);
 	}
@@ -431,14 +427,25 @@ private:
 
 	static const u32 nodecount = MAP_BLOCKSIZE * MAP_BLOCKSIZE * MAP_BLOCKSIZE;
 
+private:
+#if BUILD_UNITTESTS
+	// access to data, tryConvertToMonoBlock, deconvertMonoblock
+	friend class TestMapBlock;
+#endif
+
 	/*
 		Private methods
 	*/
 
 	void deSerialize_pre22(std::istream &is, u8 version, bool disk);
+	// check if all nodes are identical, if so convert to monoblock
+	void tryShrinkNodes();
+	// if a monoblock, expand storage back to the full array
+	void expandNodesIfNeeded();
+	void reallocate(u32 count, MapNode n);
 
 	static void getBlockNodeIdMapping(NameIdMapping *nimap, MapNode *nodes,
-		const NodeDefManager *nodedef);
+		u32 count, const NodeDefManager *nodedef);
 	static void correctBlockNodeIds(const NameIdMapping *nimap, MapNode *nodes,
 			IGameDef *gamedef);
 
@@ -475,11 +482,11 @@ private:
 	short m_refcount = 0;
 
 	/*
-	 * Note that this is not an inline array because that has implications for
+	 * Note that this is not an inline array because that has implications for heap
-	 * heap fragmentation (the array is exactly 16K), CPU caches and/or
+	 * fragmentation (the array is exactly 16K, or exactly 4 bytes for a "monoblock"),
-	 * optimizability of algorithms working on this array.
+	 * CPU caches and/or optimizability of algorithms working on this array.
 	 */
-	MapNode *const data; // of `nodecount` elements
+	MapNode *data = nullptr;
 
 	// provides the item and node definitions
 	IGameDef *m_gamedef;
@@ -490,6 +497,11 @@ private:
 	*/
 	float m_usage_timer = 0;
 
+	/*
+	 * For "monoblocks", the whole block is filled with the same node, only this node is stored.
+	 * (For reduced memory usage)
+	 */
+	bool m_is_mono_block;
 public:
 	//// ABM optimizations ////
 	// True if we never want to cache content types for this block

src/mapnode.cpp

@@ -589,7 +589,7 @@ void MapNode::deSerialize(const u8 *source, u8 version)
 
 Buffer<u8> MapNode::serializeBulk(int version,
 		const MapNode *nodes, u32 nodecount,
-		u8 content_width, u8 params_width)
+		u8 content_width, u8 params_width, bool is_mono_block)
 {
 	if (!ser_ver_supported_write(version))
 		throw VersionMismatchException("ERROR: MapNode format not supported");
@@ -601,14 +601,22 @@ Buffer<u8> MapNode::serializeBulk(int version,
 
 	// Writing to the buffer linearly is faster
 	u8 *p = &databuf[0];
+	if (is_mono_block) {
+		MapNode n = nodes[0];
+		for (u32 i = 0; i < nodecount; i++, p += 2)
+			writeU16(p, n.param0);
+		for (u32 i = 0; i < nodecount; i++, p++)
+			writeU8(p, n.param1);
+		for (u32 i = 0; i < nodecount; i++, p++)
+			writeU8(p, n.param2);
+	} else {
 		for (u32 i = 0; i < nodecount; i++, p += 2)
 			writeU16(p, nodes[i].param0);
-
 		for (u32 i = 0; i < nodecount; i++, p++)
 			writeU8(p, nodes[i].param1);
-
 		for (u32 i = 0; i < nodecount; i++, p++)
 			writeU8(p, nodes[i].param2);
+	}
 
 	return databuf;
 }

src/mapnode.h

@@ -305,9 +305,10 @@ struct alignas(u32) MapNode
 	//   content_width = the number of bytes of content per node
 	//   params_width = the number of bytes of params per node
 	//   compressed = true to zlib-compress output
+	//   is_mono_block = if true, nodes is array of size 1
 	static Buffer<u8> serializeBulk(int version,
 			const MapNode *nodes, u32 nodecount,
-			u8 content_width, u8 params_width);
+			u8 content_width, u8 params_width, bool is_mono_block = false);
 	static void deSerializeBulk(std::istream &is, int version,
 			MapNode *nodes, u32 nodecount,
 			u8 content_width, u8 params_width);

src/server.h

@@ -319,6 +319,7 @@ public:
 			std::list<std::string> *log);
 
 	// IGameDef interface
+	bool isClient() override { return false; }
 	// Under envlock
 	virtual IItemDefManager* getItemDefManager();
 	virtual const NodeDefManager* getNodeDefManager();

src/unittest/test_mapblock.cpp

@@ -10,6 +10,7 @@
 #include "serialization.h"
 #include "noise.h"
 #include "inventory.h"
+#include "voxel.h"
 
 class TestMapBlock : public TestBase
 {
@@ -33,6 +34,9 @@ public:
 
 	// Tests loading a non-standard MapBlock
 	void testLoadNonStd(IGameDef *gamedef);
+
+	// Tests blocks with a single recurring node
+	void testMonoblock(IGameDef *gamedef);
 };
 
 static TestMapBlock g_test_instance;
@@ -45,10 +49,110 @@ void TestMapBlock::runTests(IGameDef *gamedef)
 	TEST(testLoad29, gamedef);
 	TEST(testLoad20, gamedef);
 	TEST(testLoadNonStd, gamedef);
+	TEST(testMonoblock, gamedef);
 }
 
 ////////////////////////////////////////////////////////////////////////////////
 
+void TestMapBlock::testMonoblock(IGameDef *gamedef)
+{
+	MapBlock block({}, gamedef);
+	UASSERT(!block.m_is_mono_block);
+
+	// make the array is expanded
+	block.expandNodesIfNeeded();
+	UASSERT(std::all_of(block.data, block.data + MapBlock::nodecount, [](MapNode &n) { return n == MapNode(CONTENT_IGNORE); }));
+
+	// covert to monoblock
+	block.tryShrinkNodes();
+	UASSERT(block.m_is_mono_block);
+	UASSERT(block.data[0].param0 == CONTENT_IGNORE);
+
+	block.data[0] = MapNode(CONTENT_AIR);
+	UASSERT(block.m_is_mono_block);
+	UASSERT(block.data[0].param0 == CONTENT_AIR);
+
+	block.expandNodesIfNeeded();
+	UASSERT(!block.m_is_mono_block);
+	UASSERT(std::all_of(block.data, block.data + MapBlock::nodecount, [](MapNode &n) { return n == MapNode(CONTENT_AIR); }));
+
+	// covert back to mono block
+	block.tryShrinkNodes();
+	UASSERT(block.m_is_mono_block);
+
+	// deconvert explicitly
+	block.expandNodesIfNeeded();
+	UASSERT(!block.m_is_mono_block);
+
+	// covert back to mono block
+	block.tryShrinkNodes();
+	UASSERT(block.m_is_mono_block);
+
+	static_assert(CONTENT_AIR != 42);
+	// set a node, should deconvert the block
+	block.setNode(5,5,5, MapNode(42));
+	UASSERT(!block.m_is_mono_block);
+
+	// cannot covert to mono block
+	block.tryShrinkNodes();
+	UASSERT(!block.m_is_mono_block);
+
+	// set all nodes to 42
+	for (size_t i = 0; i < MapBlock::nodecount; ++i) {
+		block.data[i] = MapNode(42);
+	}
+
+	// can covert to mono block
+	block.tryShrinkNodes();
+	UASSERT(block.m_is_mono_block);
+	UASSERT(block.data[0].param0 == 42);
+
+	VoxelManipulator vmm;
+	v3s16 data_size(MAP_BLOCKSIZE, MAP_BLOCKSIZE, MAP_BLOCKSIZE);
+	vmm.addArea(VoxelArea(block.getPosRelative(), block.getPosRelative() + data_size + v3s16(1,1,1)));
+	block.copyTo(vmm);
+	UASSERT(block.m_is_mono_block);
+	UASSERT(vmm.getNode({5,5,5}).param0 == 42);
+
+	block.setNode(5,5,5,MapNode(23));
+
+	block.copyFrom(vmm);
+	UASSERT(block.m_is_mono_block);
+	UASSERT(block.data[0].param0 == 42);
+
+	vmm.setNode({5,5,5}, MapNode(23));
+	block.copyFrom(vmm);
+	UASSERT(!block.m_is_mono_block);
+
+	vmm.setNode({5,5,5}, MapNode(42,1,0));
+	block.copyFrom(vmm);
+	UASSERT(!block.m_is_mono_block);
+
+	vmm.setNode({5,5,5}, MapNode(42,0,1));
+	block.copyFrom(vmm);
+	UASSERT(!block.m_is_mono_block);
+
+	vmm.setNode({5,5,5}, MapNode(42));
+	block.copyFrom(vmm);
+	UASSERT(block.m_is_mono_block);
+
+	block.setNode(5,5,5,MapNode(23));
+	block.tryShrinkNodes();
+	UASSERT(!block.m_is_mono_block);
+
+	block.setNode(5,5,5,MapNode(42, 1, 0));
+	block.tryShrinkNodes();
+	UASSERT(!block.m_is_mono_block);
+
+	block.setNode(5,5,5,MapNode(42, 0, 1));
+	block.tryShrinkNodes();
+	UASSERT(!block.m_is_mono_block);
+
+	block.setNode(5,5,5,MapNode(42));
+	block.tryShrinkNodes();
+	UASSERT(block.m_is_mono_block);
+}
+
 void TestMapBlock::testSaveLoad(IGameDef *gamedef, const u8 version)
 {
 	// Use the bottom node ids for this test

src/voxel.cpp

@@ -177,7 +177,7 @@ void VoxelManipulator::addArea(const VoxelArea &area)
 	delete[] old_flags;
 }
 
-void VoxelManipulator::copyFrom(MapNode *src, const VoxelArea& src_area,
+void VoxelManipulator::copyFrom(MapNode *src, bool is_mono_block, const VoxelArea& src_area,
 		v3s16 from_pos, v3s16 to_pos, const v3s16 &size)
 {
 	/* The reason for this optimised code is that we're a member function
@@ -216,8 +216,12 @@ void VoxelManipulator::copyFrom(MapNode *src, const VoxelArea& src_area,
 
 	for (s16 z = 0; z < size.Z; z++) {
 		for (s16 y = 0; y < size.Y; y++) {
-			memcpy(&m_data[i_local], &src[i_src], size.X * sizeof(*m_data));
+			if (is_mono_block) {
-			memset(&m_flags[i_local], 0, size.X);
+				std::fill_n(m_data + i_local, size.X, src[0]);
+			} else {
+				std::copy_n(src + i_src, size.X, m_data + i_local);
+			}
+			std::fill_n(m_flags + i_local, size.X, 0);
 			i_src += src_step;
 			i_local += dest_step;
 		}

src/voxel.h

@@ -482,7 +482,7 @@ public:
 		Copy data and set flags to 0
 		dst_area.getExtent() <= src_area.getExtent()
 	*/
-	void copyFrom(MapNode *src, const VoxelArea& src_area,
+	void copyFrom(MapNode *src, bool is_mono_block, const VoxelArea& src_area,
 			v3s16 from_pos, v3s16 to_pos, const v3s16 &size);
 
 	// Copy data