Merge 4bb8f27723 into c98153bf07

Minecraft.Client/Chunk.cpp

@@ -22,24 +22,21 @@
 int Chunk::updates = 0;
 
 #ifdef _LARGE_WORLDS
-DWORD Chunk::tlsIdx = TlsAlloc();
+static thread_local unsigned char s_tlsTileIds[16 * 16 * Level::maxBuildHeight];
 
 void Chunk::CreateNewThreadStorage()
 {
-	unsigned char *tileIds = new unsigned char[16 * 16 * Level::maxBuildHeight];
+	// No-op: thread_local handles per-thread allocation automatically
-	TlsSetValue(tlsIdx, tileIds);
 }
 
 void Chunk::ReleaseThreadStorage()
 {
-	unsigned char *tileIds = static_cast<unsigned char *>(TlsGetValue(tlsIdx));
+	// No-op: thread_local handles per-thread cleanup automatically
-	delete tileIds;
 }
 
 unsigned char *Chunk::GetTileIdsStorage()
 {
-	unsigned char *tileIds = static_cast<unsigned char *>(TlsGetValue(tlsIdx));
+	return s_tlsTileIds;
-	return tileIds;
 }
 #else
 // 4J Stu - Don't want this when multi-threaded
@@ -234,26 +231,51 @@ void Chunk::rebuild()
 	static unsigned char tileIds[16 * 16 * Level::maxBuildHeight];
 #endif
 	byteArray tileArray = byteArray(tileIds, 16 * 16 * Level::maxBuildHeight);
-	level->getChunkAt(x,z)->getBlockData(tileArray);		// 4J - TODO - now our data has been re-arranged, we could just extra the vertical slice of this chunk rather than the whole thing
+	LevelChunk *sourceChunk = level->getChunkAt(x,z);
-
+	if( sourceChunk == nullptr )
-	LevelSource *region = new Region(level, x0 - r, y0 - r, z0 - r, x1 + r, y1 + r, z1 + r, r);
+	{
-	TileRenderer *tileRenderer = new TileRenderer(region, this->x, this->y, this->z, tileIds);
+		// Level chunk not loaded yet - treat as empty
-
+		for (int currentLayer = 0; currentLayer < 2; currentLayer++)
-	// AP - added a caching system for Chunk::rebuild to take advantage of
+		{
-	// Basically we're storing of copy of the tileIDs array inside the region so that calls to Region::getTile can grab data
+			levelRenderer->setGlobalChunkFlag(this->x, this->y, this->z, level, LevelRenderer::CHUNK_FLAG_EMPTY0, currentLayer);
-	// more quickly from this array rather than calling CompressedTileStorage. On the Vita the total thread time spent in
+			RenderManager.CBuffClear(lists + currentLayer);
-	// Region::getTile went from 20% to 4%.
+		}
-#ifdef __PSVITA__
+		levelRenderer->setGlobalChunkFlag(this->x, this->y, this->z, level, LevelRenderer::CHUNK_FLAG_NOTSKYLIT);
-	int xc = x >> 4;
+		levelRenderer->setGlobalChunkFlag(this->x, this->y, this->z, level, LevelRenderer::CHUNK_FLAG_COMPILED);
-	int zc = z >> 4;
+		PIXEndNamedEvent();
-	((Region*)region)->setCachedTiles(tileIds, xc, zc);
+		PIXEndNamedEvent();
-#endif
+		return;
+	}
+	sourceChunk->getBlockData(tileArray);		// 4J - TODO - now our data has been re-arranged, we could just extra the vertical slice of this chunk rather than the whole thing
 
 	// We now go through the vertical section of this level chunk that we are interested in and try and establish
 	// (1) if it is completely empty
 	// (2) if any of the tiles can be quickly determined to not need rendering because they are in the middle of other tiles and
 	//     so can't be seen. A large amount (> 60% in tests) of tiles that call tesselateInWorld in the unoptimised version
 	//     of this function fall into this category. By far the largest category of these are tiles in solid regions of rock.
+	// Build the occluder lookup from Tile::isSolidRender() so that ALL opaque full-cube blocks act as occluders,
+	// not just stone/dirt/bedrock. This catches sand, gravel, ores, cobblestone, sandstone, netherrack, wool, etc.
+	// which dramatically reduces tesselation work for underground/terrain-heavy chunks.
+	//
+	// thread_local: the table is built once per thread and reused across all subsequent rebuild() calls on that
+	// thread. Without it, a plain local would redo 255 virtual dispatches (isSolidRender) every single call -
+	// rebuild() runs up to MAX_CONCURRENT_CHUNK_REBUILDS times per updateDirtyChunks(), ~10 times per frame,
+	// so hundreds of redundant vtable lookups per second for data that never changes at runtime. A plain static
+	// would avoid that but introduces a data race: multiple rebuild threads run concurrently and the C++ static
+	// init guard (mutex) would serialize them on every entry just to check the "already initialized" flag.
+	// thread_local gives each thread its own copy with zero synchronization after the first call.
+	static thread_local auto isOccluder = [] {
+		std::array<bool, 256> table{};
+		for (int id = 1; id < 256; id++)
+		{
+			Tile *tile = Tile::tiles[id];
+			if (tile != nullptr && tile->isSolidRender())
+				table[id] = true;
+		}
+		table[255] = true;  // already-marked-invisible sentinel
+		return table;
+	}();
+
 	bool empty = true;
 	for( int yy = y0; yy < y1; yy++ )
 	{
@@ -279,17 +301,12 @@ void Chunk::rebuild()
 				if(( xx == 0 ) || ( xx == 15 )) continue;
 				if(( zz == 0 ) || ( zz == 15 )) continue;
 
-				// Establish whether this tile and its neighbours are all made of rock, dirt, unbreakable tiles, or have already
+				// Establish whether this tile and its neighbours are all occluders using lookup table
-				// been determined to meet this criteria themselves and have a tile of 255 set.
+				if( !isOccluder[tileId] ) continue;
-				if( !( ( tileId == Tile::stone_Id ) || ( tileId == Tile::dirt_Id ) || ( tileId == Tile::unbreakable_Id ) || ( tileId == 255) ) ) continue;
+				if( !isOccluder[ tileIds[ offset + ( ( ( xx - 1 ) << 11 ) | ( ( zz + 0 ) << 7 ) | ( indexY + 0 )) ] ] ) continue;
-				tileId = tileIds[ offset + ( ( ( xx - 1 ) << 11 ) | ( ( zz + 0 ) << 7 ) | ( indexY + 0 )) ];
+				if( !isOccluder[ tileIds[ offset + ( ( ( xx + 1 ) << 11 ) | ( ( zz + 0 ) << 7 ) | ( indexY + 0 )) ] ] ) continue;
-				if( !( ( tileId == Tile::stone_Id ) || ( tileId == Tile::dirt_Id ) || ( tileId == Tile::unbreakable_Id ) || ( tileId == 255) ) ) continue;
+				if( !isOccluder[ tileIds[ offset + ( ( ( xx + 0 ) << 11 ) | ( ( zz - 1 ) << 7 ) | ( indexY + 0 )) ] ] ) continue;
-				tileId = tileIds[ offset + ( ( ( xx + 1 ) << 11 ) | ( ( zz + 0 ) << 7 ) | ( indexY + 0 )) ];
+				if( !isOccluder[ tileIds[ offset + ( ( ( xx + 0 ) << 11 ) | ( ( zz + 1 ) << 7 ) | ( indexY + 0 )) ] ] ) continue;
-				if( !( ( tileId == Tile::stone_Id ) || ( tileId == Tile::dirt_Id ) || ( tileId == Tile::unbreakable_Id ) || ( tileId == 255) ) ) continue;
-				tileId = tileIds[ offset + ( ( ( xx + 0 ) << 11 ) | ( ( zz - 1 ) << 7 ) | ( indexY + 0 )) ];
-				if( !( ( tileId == Tile::stone_Id ) || ( tileId == Tile::dirt_Id ) || ( tileId == Tile::unbreakable_Id ) || ( tileId == 255) ) ) continue;
-				tileId = tileIds[ offset + ( ( ( xx + 0 ) << 11 ) | ( ( zz + 1 ) << 7 ) | ( indexY + 0 )) ];
-				if( !( ( tileId == Tile::stone_Id ) || ( tileId == Tile::dirt_Id ) || ( tileId == Tile::unbreakable_Id ) || ( tileId == 255) ) ) continue;
 				// Treat the bottom of the world differently - we shouldn't ever be able to look up at this, so consider tiles as invisible
 				// if they are surrounded on sides other than the bottom
 				if( yy > 0 )
@@ -301,8 +318,7 @@ void Chunk::rebuild()
 						indexYMinusOne -= Level::COMPRESSED_CHUNK_SECTION_HEIGHT;
 						yMinusOneOffset = Level::COMPRESSED_CHUNK_SECTION_TILES;
 					}
-					tileId = tileIds[ yMinusOneOffset + ( ( ( xx + 0 ) << 11 ) | ( ( zz + 0 ) << 7 ) | indexYMinusOne ) ];
+					if( !isOccluder[ tileIds[ yMinusOneOffset + ( ( ( xx + 0 ) << 11 ) | ( ( zz + 0 ) << 7 ) | indexYMinusOne ) ] ] ) continue;
-					if( !( ( tileId == Tile::stone_Id ) || ( tileId == Tile::dirt_Id ) || ( tileId == Tile::unbreakable_Id ) || ( tileId == 255) ) ) continue;
 				}
 				int indexYPlusOne = yy + 1;
 				int yPlusOneOffset = 0;
@@ -311,8 +327,7 @@ void Chunk::rebuild()
 					indexYPlusOne -= Level::COMPRESSED_CHUNK_SECTION_HEIGHT;
 					yPlusOneOffset = Level::COMPRESSED_CHUNK_SECTION_TILES;
 				}
-				tileId = tileIds[ yPlusOneOffset + ( ( ( xx + 0 ) << 11 ) | ( ( zz + 0 ) << 7 ) | indexYPlusOne ) ];
+				if( !isOccluder[ tileIds[ yPlusOneOffset + ( ( ( xx + 0 ) << 11 ) | ( ( zz + 0 ) << 7 ) | indexYPlusOne ) ] ] ) continue;
-				if( !( ( tileId == Tile::stone_Id ) || ( tileId == Tile::dirt_Id ) || ( tileId == Tile::unbreakable_Id ) || ( tileId == 255) ) ) continue;
 
 				// This tile is surrounded. Flag it as not requiring to be rendered by setting its id to 255.
 				tileIds[ offset + ( ( ( xx + 0 ) << 11 ) | ( ( zz + 0 ) << 7 ) | ( indexY + 0 ) ) ] = 0xff;
@@ -329,14 +344,31 @@ void Chunk::rebuild()
 			levelRenderer->setGlobalChunkFlag(this->x, this->y, this->z, level, LevelRenderer::CHUNK_FLAG_EMPTY0, currentLayer);
 			RenderManager.CBuffClear(lists + currentLayer);
 		}
+		levelRenderer->setGlobalChunkFlag(this->x, this->y, this->z, level, LevelRenderer::CHUNK_FLAG_NOTSKYLIT);
+		levelRenderer->setGlobalChunkFlag(this->x, this->y, this->z, level, LevelRenderer::CHUNK_FLAG_COMPILED);
 
-		delete region;
+		PIXEndNamedEvent();  // match "Rebuilding chunk" event
-		delete tileRenderer;
 		return;
 	}
 	// 4J - optimisation ends
 	////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
 
+	// Construct Region and TileRenderer only for non-empty chunks.
+	// This is deferred to here so that empty chunks (common after teleport) skip
+	// the Region construction (9 chunk lookups) and TileRenderer 128KB cache memset.
+	Region region(level, x0 - r, y0 - r, z0 - r, x1 + r, y1 + r, z1 + r, r);
+	TileRenderer tileRenderer(&region, this->x, this->y, this->z, tileIds);
+
+	// AP - added a caching system for Chunk::rebuild to take advantage of
+    // Basically we're storing of copy of the tileIDs array inside the region so that calls to Region::getTile can grab data
+    // more quickly from this array rather than calling CompressedTileStorage. On the Vita the total thread time spent in
+    // Region::getTile went from 20% to 4%.
+#ifdef __PSVITA__
+	int xc = x >> 4;
+	int zc = z >> 4;
+	region.setCachedTiles(tileIds, xc, zc);
+#endif
+
 	PIXBeginNamedEvent(0,"Rebuild section C");
 	Tesselator::Bounds bounds;	// 4J MGH - added
 	{
@@ -373,11 +405,11 @@ void Chunk::rebuild()
 					}
 
 					// 4J - get tile from those copied into our local array in earlier optimisation
-					unsigned char tileId = tileIds[ offset + ( ( ( x - x0 ) << 11 ) | ( ( z - z0 ) << 7 ) | indexY) ];
+					const unsigned char tileId = tileIds[ offset + ( ( ( x - x0 ) << 11 ) | ( ( z - z0 ) << 7 ) | indexY) ];
 					// If flagged as not visible, drop out straight away
-					if( tileId == 0xff ) continue;
+					if( tileId == 0xff ) [[unlikely]] continue;
 //					int tileId = region->getTile(x,y,z);
-					if (tileId > 0)
+					if (tileId > 0) [[unlikely]]
 					{
 						if (!started)
 						{
@@ -405,7 +437,7 @@ void Chunk::rebuild()
 						Tile *tile = Tile::tiles[tileId];
 						if (currentLayer == 0 && tile->isEntityTile())
 						{
-							shared_ptr<TileEntity> et = region->getTileEntity(x, y, z);
+							shared_ptr<TileEntity> et = region.getTileEntity(x, y, z);
 							if (TileEntityRenderDispatcher::instance->hasRenderer(et))
 							{
 								renderableTileEntities.push_back(et);
@@ -419,7 +451,7 @@ void Chunk::rebuild()
 						}
 						else if (renderLayer == currentLayer)
 						{
-							rendered |= tileRenderer->tesselateInWorld(tile, x, y, z);
+							rendered |= tileRenderer.tesselateInWorld(tile, x, y, z);
 						}
 					}
 				}
@@ -479,9 +511,6 @@ void Chunk::rebuild()
 			   bounds.boundingBox[3], bounds.boundingBox[4], bounds.boundingBox[5]);
 	}
 
-	delete tileRenderer;
-	delete region;
-
 	PIXEndNamedEvent();
 	PIXBeginNamedEvent(0,"Rebuild section D");
 	// 4J - have rewritten the way that tile entities are stored globally to make it work more easily with split screen. Chunks are now

Minecraft.Client/Chunk.h

@@ -30,9 +30,8 @@ public:
 	static LevelRenderer *levelRenderer;
 private:
 #ifndef _LARGE_WORLDS
-    static Tesselator *t;
+	static Tesselator *t;
 #else
-	static DWORD tlsIdx;
 public:
 	static void CreateNewThreadStorage();
 	static void ReleaseThreadStorage();

Minecraft.Client/Common/UI/UIScene_CreateWorldMenu.cpp

@@ -18,8 +18,6 @@
 #endif
 
 #ifdef  _WINDOWS64
-
-#include <windows.h>
 #include "Xbox\Resource.h"
 #endif
 

Minecraft.Client/Durango/Iggy/gdraw/gdraw_d3d11.cpp

@@ -26,15 +26,10 @@
 
 #define GDRAW_ASSERTS
 
-#ifndef WIN32_LEAN_AND_MEAN
-#define WIN32_LEAN_AND_MEAN
-#endif
-
 // We temporarily disable this warning for the shared interface portions
 #pragma warning (push)
 #pragma warning (disable: 4201) // nonstandard extension used : nameless struct/union
 
-#include <windows.h>
 #include <d3d11_x.h>		// 4J changed to use monolithic version
 #include "gdraw.h"
 #include "iggy.h"

Minecraft.Client/LevelRenderer.cpp

@@ -367,19 +367,17 @@ void LevelRenderer::setLevel(int playerIndex, MultiPlayerLevel *level)
 	else
 	{
 		//		printf("NULLing player %d, chunks @ 0x%x\n",playerIndex,chunks[playerIndex]);
-		if( chunks[playerIndex].data != nullptr )
+        if (chunks[playerIndex].data != nullptr)
-		{
+        {
-			for (unsigned int i = 0; i < chunks[playerIndex].length; i++)
+            for (unsigned int i = 0; i < chunks[playerIndex].length; i++)
-			{
+            {
-				chunks[playerIndex][i].chunk->_delete();
+                chunks[playerIndex][i].chunk->_delete();
-				delete chunks[playerIndex][i].chunk;
+                delete chunks[playerIndex][i].chunk;
-			}
+            }
-			delete chunks[playerIndex].data;
+            delete[] chunks[playerIndex].data;
-			chunks[playerIndex].data = nullptr;
+            chunks[playerIndex].data = nullptr;
-			chunks[playerIndex].length = 0;
+            chunks[playerIndex].length = 0;
-			//			delete sortedChunks[playerIndex];	// 4J - removed - not sorting our chunks anymore
+        }
-			//			sortedChunks[playerIndex] = nullptr;	// 4J - removed - not sorting our chunks anymore
-		}
 
 		// 4J Stu - If we do this for splitscreen players leaving, then all the tile entities in the world dissappear
 		// We should only do this when actually exiting the game, so only when the primary player sets there level to nullptr
@@ -451,7 +449,7 @@ void LevelRenderer::allChanged(int playerIndex)
 			chunks[playerIndex][i].chunk->_delete();
 			delete chunks[playerIndex][i].chunk;
 		}
-		delete chunks[playerIndex].data;
+        delete[] chunks[playerIndex].data;
 		//		delete sortedChunks[playerIndex];	// 4J - removed - not sorting our chunks anymore
 	}
 
@@ -823,8 +821,8 @@ int LevelRenderer::renderChunks(int from, int to, int layer, double alpha)
 	unsigned char emptyFlag = LevelRenderer::CHUNK_FLAG_EMPTY0 << layer;
 	for( int i = 0; i < chunks[playerIndex].length; i++, pClipChunk++ )
 	{
-		if( !pClipChunk->visible ) continue;													// This will be set if the chunk isn't visible, or isn't compiled, or has both empty flags set
+		if( !pClipChunk->visible ) [[likely]] continue;											// This will be set if the chunk isn't visible, or isn't compiled, or has both empty flags set
-		if( pClipChunk->globalIdx == -1 ) continue;												// Not sure if we should ever encounter this... TODO check
+		if( pClipChunk->globalIdx == -1 ) [[unlikely]] continue;								// Not sure if we should ever encounter this... TODO check
 		if( ( globalChunkFlags[pClipChunk->globalIdx] & emptyFlag ) == emptyFlag ) continue;	// Check that this particular layer isn't empty
 
 		// List can be calculated directly from the chunk's global idex
@@ -1943,7 +1941,11 @@ bool LevelRenderer::updateDirtyChunks()
 		int maxNearestChunks = MAX_CONCURRENT_CHUNK_REBUILDS;
 		// 4J Stu - On XboxOne we should cut this down if in a constrained state so the saving threads get more time
 #endif
-		// Find nearest chunk that is dirty
+		// Find nearest chunk that is dirty.
+		// Rewritten for performance: linear scan of clip chunks (cache-friendly), with dirty flag
+		// checked BEFORE computing distances. The old triple-nested x/z/y loop computed distances
+		// for every chunk (~20K) regardless of dirty state. Now only dirty chunks (typically <1%)
+		// pay the cost of distance computation and nearest-selection.
 		for( int p = 0; p < XUSER_MAX_COUNT; p++ )
 		{
 			// It's possible that the localplayers member can be set to nullptr on the main thread when a player chooses to exit the game
@@ -1957,96 +1959,91 @@ bool LevelRenderer::updateDirtyChunks()
 			int py = static_cast<int>(player->y);
 			int pz = static_cast<int>(player->z);
 
-			//			app.DebugPrintf("!! %d %d %d, %d %d %d {%d,%d} ",px,py,pz,stackChunkDirty,nonStackChunkDirty,onlyRebuild, xChunks, zChunks);
+			int numClipChunks = static_cast<int>(chunks[p].length);
-
+			ClipChunk *pClipChunk = chunks[p].data;
-			int considered = 0;
+			for( int i = 0; i < numClipChunks; i++, pClipChunk++ )
-			int wouldBeNearButEmpty = 0;
-			for( int x = 0; x < xChunks; x++ )
 			{
-				for( int z = 0; z < zChunks; z++ )
+				// Fast reject: skip non-dirty chunks immediately before any distance work.
+				// globalIdx can be -1 for unassigned chunks.
+				const int gIdx = pClipChunk->globalIdx;
+				if (gIdx < 0)
+					continue;
+				const unsigned char flags = globalChunkFlags[gIdx];
+				if (!(flags & CHUNK_FLAG_DIRTY))
+					continue;
+
+				// Batch-clear empty chunks upfront. After teleport, thousands of sky/void chunks
+				// are dirty. Clearing them all in one scan (rather than ~8 per call) dramatically
+				// reduces the dirty set for subsequent iterations.
+				Chunk *chunk = pClipChunk->chunk;
+				if (chunk == nullptr)
+					continue;
+				const int ySlice = (pClipChunk->ym - (CHUNK_SIZE / 2)) / CHUNK_SIZE;
+				LevelChunk *lc = level[p]->getChunkAt(chunk->x, chunk->z);
+				if (lc == nullptr || lc->isRenderChunkEmpty(ySlice * 16))
 				{
-					for( int y = 0; y < CHUNK_Y_COUNT; y++ )
+					chunk->clearDirty();
+					globalChunkFlags[gIdx] |= CHUNK_FLAG_EMPTYBOTH;
+					continue;
+				}
+
+				// Non-empty dirty chunk — now compute distance
+				const int xd = pClipChunk->xm - px;
+				const int yd = pClipChunk->ym - py;
+				const int zd = pClipChunk->zm - pz;
+				const int distSq = xd * xd + yd * yd + zd * zd;
+				const int distSqWeighted = distSq + 3 * yd * yd;  // Extra y weighting to prioritise things in same x/z plane as player first
+
+				if( (!onlyRebuild) ||
+					(flags & CHUNK_FLAG_COMPILED) ||
+					( distSq < 96 * 96 ) )	// Always rebuild really near things or else building (say) at tower up into empty blocks when we are low on memory will not create render data
+				{
+					// Is this chunk nearer than our nearest?
+#ifdef _LARGE_WORLDS
+					bool isNearer = nearestClipChunks.empty();
+					auto itNearest = nearestClipChunks.begin();
+					for(; itNearest != nearestClipChunks.end(); ++itNearest)
 					{
-						ClipChunk *pClipChunk = &chunks[p][(z * yChunks + y) * xChunks + x];
+						isNearer = distSqWeighted < itNearest->second;
-						// Get distance to this chunk - deliberately not calling the chunk's method of doing this to avoid overheads (passing entitie, type conversion etc.) that this involves
+						if(isNearer) break;
-						int xd = pClipChunk->xm - px;
+					}
-						int yd = pClipChunk->ym - py;
+					isNearer = isNearer || (nearestClipChunks.size() < maxNearestChunks);
-						int zd = pClipChunk->zm - pz;
+#else
-						int distSq = xd * xd + yd * yd + zd * zd;
+					bool isNearer = distSqWeighted < minDistSq;
-						int distSqWeighted = xd * xd + yd * yd * 4 + zd * zd;  // Weighting against y to prioritise things in same x/z plane as player first
+#endif
 
-						if( globalChunkFlags[ pClipChunk->globalIdx ] & CHUNK_FLAG_DIRTY )
+#ifdef _CRITICAL_CHUNKS
+					// AP - this will make sure that if a deferred grouping has started, only critical chunks go into that
+					// grouping, even if a non-critical chunk is closer.
+					if( (!veryNearCount && isNearer) ||
+						(distSq < 20 * 20 && (flags & CHUNK_FLAG_CRITICAL)) )
+#else
+					if( isNearer )
+#endif
+					{
+						// Non-empty (already confirmed above), add to nearest set
+						nearChunk = pClipChunk;
+						minDistSq = distSqWeighted;
+#ifdef _LARGE_WORLDS
+						nearestClipChunks.insert(itNearest, std::make_pair(nearChunk, minDistSq) );
+						if(nearestClipChunks.size() > maxNearestChunks)
 						{
-							if( (!onlyRebuild) ||
+							nearestClipChunks.pop_back();
-								globalChunkFlags[ pClipChunk->globalIdx ] & CHUNK_FLAG_COMPILED ||
-								( distSq < 96 * 96 ) )	// Always rebuild really near things or else building (say) at tower up into empty blocks when we are low on memory will not create render data
-							{	// distSq adjusted from 20 * 20 to 96 * 96 - updated by detectiveren
-								considered++;
-								// Is this chunk nearer than our nearest?
-#ifdef _LARGE_WORLDS
-								bool isNearer = nearestClipChunks.empty();
-                                auto itNearest = nearestClipChunks.begin();
-                                for(; itNearest != nearestClipChunks.end(); ++itNearest)
-								{
-									isNearer = distSqWeighted < itNearest->second;
-									if(isNearer) break;
-								}
-								isNearer = isNearer || (nearestClipChunks.size() < maxNearestChunks);
-#else
-								bool isNearer = distSqWeighted < minDistSq;
-#endif
-
-#ifdef _CRITICAL_CHUNKS
-								// AP - this will make sure that if a deferred grouping has started, only critical chunks go into that
-								// grouping, even if a non-critical chunk is closer.
-								if( (!veryNearCount && isNearer) ||
-									(distSq < 20 * 20 && (globalChunkFlags[ pClipChunk->globalIdx ] & CHUNK_FLAG_CRITICAL)) )
-#else
-								if( isNearer )
-#endif
-								{
-									// At this point we've got a chunk that we would like to consider for rendering, at least based on its proximity to the player(s).
-									// Its *quite* quick to generate empty render data for render chunks, but if we let the rebuilding do that then the after rebuilding we will have
-									// to start searching for the next nearest chunk from scratch again. Instead, its better to detect empty chunks at this stage, flag them up as not dirty
-									// (and empty), and carry on. The levelchunk's isRenderChunkEmpty method can be quite optimal as it can make use of the chunk's data compression to detect
-									// emptiness without actually testing as many data items as uncompressed data would.
-									Chunk *chunk = pClipChunk->chunk;
-									LevelChunk *lc = level[p]->getChunkAt(chunk->x,chunk->z);
-									if( !lc->isRenderChunkEmpty(y * 16) )
-									{
-										nearChunk = pClipChunk;
-										minDistSq = distSqWeighted;
-#ifdef _LARGE_WORLDS
-										nearestClipChunks.insert(itNearest, std::make_pair(nearChunk, minDistSq) );
-										if(nearestClipChunks.size() > maxNearestChunks)
-										{
-											nearestClipChunks.pop_back();
-										}
-#endif
-									}
-									else
-									{
-										chunk->clearDirty();
-										globalChunkFlags[ pClipChunk->globalIdx ] |= CHUNK_FLAG_EMPTYBOTH;
-										wouldBeNearButEmpty++;
-									}
-								}
-
-#ifdef _CRITICAL_CHUNKS
-								// AP - is the chunk near and also critical
-								if( distSq < 20 * 20 && ((globalChunkFlags[ pClipChunk->globalIdx ] & CHUNK_FLAG_CRITICAL)) )
-#else
-								if( distSq < 20 * 20 )
-#endif
-								{
-									veryNearCount++;
-								}
-							}
 						}
+#endif
+					}
+
+#ifdef _CRITICAL_CHUNKS
+					// AP - is the chunk near and also critical
+					if( distSq < 20 * 20 && (flags & CHUNK_FLAG_CRITICAL) )
+#else
+					if( distSq < 20 * 20 )
+#endif
+					{
+						veryNearCount++;
 					}
 				}
 			}
-			//			app.DebugPrintf("[%d,%d,%d]\n",nearestClipChunks.empty(),considered,wouldBeNearButEmpty);
 		}
 #endif // __PS3__
 		PIXEndNamedEvent();
@@ -2193,7 +2190,11 @@ bool LevelRenderer::updateDirtyChunks()
 		return true;
 	}
 
-	if( nearChunk )	destroyedTileManager->updatedChunkAt(chunk->level, chunk->x, chunk->y, chunk->z, veryNearCount );
+	if( nearChunk )
+	{
+		destroyedTileManager->updatedChunkAt(chunk->level, chunk->x, chunk->y, chunk->z, veryNearCount );
+		return dirtyChunkPresent;
+	}
 
 	return false;
 }
@@ -2444,18 +2445,23 @@ void LevelRenderer::setTilesDirty(int x0, int y0, int z0, int x1, int y1, int z1
 	setDirty(x0 - 1, y0 - 1, z0 - 1, x1 + 1, y1 + 1, z1 + 1, level);
 }
 
-bool inline clip(float *bb, float *frustum)
+bool inline clip(float * __restrict bb, float * __restrict frustum)
 {
+	// Pre-load AABB corners to avoid repeated memory loads
+	const float x0 = bb[0], y0 = bb[1], z0 = bb[2];
+	const float x1 = bb[3], y1 = bb[4], z1 = bb[5];
+
 	for (int i = 0; i < 6; ++i, frustum += 4)
 	{
-		if (frustum[0] * (bb[0]) + frustum[1] * (bb[1]) + frustum[2] * (bb[2]) + frustum[3] > 0) continue;
+		const float a = frustum[0], b = frustum[1], c = frustum[2], d = frustum[3];
-		if (frustum[0] * (bb[3]) + frustum[1] * (bb[1]) + frustum[2] * (bb[2]) + frustum[3] > 0) continue;
+		if (a * x0 + b * y0 + c * z0 + d > 0) continue;
-		if (frustum[0] * (bb[0]) + frustum[1] * (bb[4]) + frustum[2] * (bb[2]) + frustum[3] > 0) continue;
+		if (a * x1 + b * y0 + c * z0 + d > 0) continue;
-		if (frustum[0] * (bb[3]) + frustum[1] * (bb[4]) + frustum[2] * (bb[2]) + frustum[3] > 0) continue;
+		if (a * x0 + b * y1 + c * z0 + d > 0) continue;
-		if (frustum[0] * (bb[0]) + frustum[1] * (bb[1]) + frustum[2] * (bb[5]) + frustum[3] > 0) continue;
+		if (a * x1 + b * y1 + c * z0 + d > 0) continue;
-		if (frustum[0] * (bb[3]) + frustum[1] * (bb[1]) + frustum[2] * (bb[5]) + frustum[3] > 0) continue;
+		if (a * x0 + b * y0 + c * z1 + d > 0) continue;
-		if (frustum[0] * (bb[0]) + frustum[1] * (bb[4]) + frustum[2] * (bb[5]) + frustum[3] > 0) continue;
+		if (a * x1 + b * y0 + c * z1 + d > 0) continue;
-		if (frustum[0] * (bb[3]) + frustum[1] * (bb[4]) + frustum[2] * (bb[5]) + frustum[3] > 0) continue;
+		if (a * x0 + b * y1 + c * z1 + d > 0) continue;
+		if (a * x1 + b * y1 + c * z1 + d > 0) continue;
 
 		return false;
 	}

Minecraft.Client/Tesselator.cpp

@@ -24,17 +24,16 @@ int normal;
 
 
 */
-DWORD Tesselator::tlsIdx = TlsAlloc();
+static thread_local std::unique_ptr<Tesselator> tlsInstance;
 
 Tesselator *Tesselator::getInstance()
 {
-	return static_cast<Tesselator *>(TlsGetValue(tlsIdx));
+    return tlsInstance.get();
 }
 
 void Tesselator::CreateNewThreadStorage(int bytes)
 {
-	Tesselator *instance = new Tesselator(bytes/4);
+    tlsInstance = std::make_unique<Tesselator>(bytes / 4);
-	TlsSetValue(tlsIdx, instance);
 }
 
 Tesselator::Tesselator(int size)
@@ -310,20 +309,16 @@ void Tesselator::color(int r, int g, int b)
 
 void Tesselator::color(int r, int g, int b, int a)
 {
-    if (_noColor) return;
+	if (_noColor) return;
 
-    if (r > 255) r = 255;
+	r = std::clamp(r, 0, 255);
-    if (g > 255) g = 255;
+	g = std::clamp(g, 0, 255);
-    if (b > 255) b = 255;
+	b = std::clamp(b, 0, 255);
-    if (a > 255) a = 255;
+	a = std::clamp(a, 0, 255);
-    if (r < 0) r = 0;
-    if (g < 0) g = 0;
-    if (b < 0) b = 0;
-    if (a < 0) a = 0;
 
-    hasColor = true;
+	hasColor = true;
 	// 4J - removed little-endian option
-    col = (r << 24) | (g << 16) | (b << 8) | (a);
+	col = (r << 24) | (g << 16) | (b << 8) | (a);
 }
 
 void Tesselator::color(byte r, byte g, byte b)
@@ -381,26 +376,15 @@ void Tesselator::packCompactQuad()
 		m_iz[i] += 16 * 128;
 	}
 	// Find min x/y/z
-	unsigned int minx = m_ix[0];
+	unsigned int minx = std::min<unsigned int>({m_ix[0], m_ix[1], m_ix[2], m_ix[3]});
-	unsigned int miny = m_iy[0];
+    unsigned int miny = std::min<unsigned int>({m_iy[0], m_iy[1], m_iy[2], m_iy[3]});
-	unsigned int minz = m_iz[0];
+    unsigned int minz = std::min<unsigned int>({m_iz[0], m_iz[1], m_iz[2], m_iz[3]});
-	for( int i = 1; i < 4; i++ )
-	{
-		if( m_ix[i] < minx ) minx = m_ix[i];
-		if( m_iy[i] < miny ) miny = m_iy[i];
-		if( m_iz[i] < minz ) minz = m_iz[i];
-	}
 	// Everything has been scaled by a factor of 128 to get it into an int, and so
 	// the minimum now should be in the range of (0->32) * 128. Get the base x/y/z
 	// that our quad will be referenced from now, which can be stored in 5 bits
-	unsigned int basex = ( minx >> 7 );
+    unsigned int basex = std::min<unsigned int>(minx >> 7, 31u);
-	unsigned int basey = ( miny >> 7 );
+    unsigned int basey = std::min<unsigned int>(miny >> 7, 31u);
-	unsigned int basez = ( minz >> 7 );
+    unsigned int basez = std::min<unsigned int>(minz >> 7, 31u);
-	// If the min is 32, then this whole quad must be in that plane - make the min 15 instead so
-	// we can still offset from that with our delta to get to the exact edge
-	if( basex == 32 ) basex = 31;
-	if( basey == 32 ) basey = 31;
-	if( basez == 32 ) basez = 31;
 	// Now get deltas to each vertex - these have an 8-bit range so they can span a
 	// full unit range from the base position
 	for( int i = 0; i < 4; i++ )
@@ -420,28 +404,18 @@ void Tesselator::packCompactQuad()
 	data[0] |= ( basex << 26 ) | ( basey << 21 )| ( basez << 16 );
 
 	// Now process UVs. First find min & max U & V
-	unsigned int minu = m_u[0];
+    unsigned int minu = std::min<unsigned int>({m_u[0], m_u[1], m_u[2], m_u[3]});
-	unsigned int minv = m_v[0];
+    unsigned int minv = std::min<unsigned int>({m_v[0], m_v[1], m_v[2], m_v[3]});
-	unsigned int maxu = m_u[0];
+    unsigned int maxu = std::max<unsigned int>({m_u[0], m_u[1], m_u[2], m_u[3]});
-	unsigned int maxv = m_v[0];
+    unsigned int maxv = std::max<unsigned int>({m_v[0], m_v[1], m_v[2], m_v[3]});
-
-	for( int i = 1; i < 4; i++ )
-	{
-		if( m_u[i] < minu ) minu = m_u[i];
-		if( m_v[i] < minv ) minv = m_v[i];
-		if( m_u[i] > maxu ) maxu = m_u[i];
-		if( m_v[i] > maxv ) maxv = m_v[i];
-	}
 	// In nearly all cases, all our UVs should be axis aligned for this quad. So the only values they should
 	// have in each dimension should be the min/max. We're going to store:
 	// (1) minu/maxu (16 bits each, only actuall needs to store 14 bits to get a 0 to 2 range for each
 	// (2) du/dv ( ie maxu-minu, maxv-minv) - 8 bits each, to store a range of 0 to 15.9375 texels. This
 	// should be enough to map the full UV range of a single 16x16 region of the terrain texture, since
 	// we always pull UVs in by 1/16th of their range at the sides
-	unsigned int du = maxu - minu;
+    unsigned int du = std::min<unsigned int>(maxu - minu, 255u);
-	unsigned int dv = maxv - minv;
+    unsigned int dv = std::min<unsigned int>(maxv - minv, 255u);
-	if( du > 255 )	du = 255;
-	if( dv > 255 )	dv = 255;
 	// Check if this quad has UVs that can be referenced this way. This should only happen for flowing water
 	// and lava, where the texture coordinates are rotated for the top surface of the tile.
 	bool axisAligned = true;

Minecraft.Client/Tesselator.h

@@ -34,13 +34,9 @@ private:
 	float xoo, yoo, zoo;
     int _normal;
 
-	// 4J - added for thread local storage
+ public:
-public:
+    static void CreateNewThreadStorage(int bytes);
-	static void CreateNewThreadStorage(int bytes);
+    static Tesselator *getInstance();
-private:
-	static DWORD tlsIdx;
-public:
-	static Tesselator *getInstance();
 
 private:
 	bool tesselating;
@@ -52,6 +48,7 @@ private:
     int vboCounts;
     int size;
 
+  public:
     Tesselator(int size);
 public:
     Tesselator *getUniqueInstance(int size);
@@ -93,35 +90,21 @@ public:
 		}
 		void addVert(float x, float y, float z)
 		{
-			if(x < boundingBox[0])
+            boundingBox[0] = std::min<float>(boundingBox[0], x);
-				boundingBox[0] = x;
+            boundingBox[1] = std::min<float>(boundingBox[1], y);
-			if(y < boundingBox[1])
+            boundingBox[2] = std::min<float>(boundingBox[2], z);
-				boundingBox[1] = y;
+            boundingBox[3] = std::max<float>(boundingBox[3], x);
-			if(z < boundingBox[2])
+            boundingBox[4] = std::max<float>(boundingBox[4], y);
-				boundingBox[2] = z;
+            boundingBox[5] = std::max<float>(boundingBox[5], z);
-
-			if(x > boundingBox[3])
-				boundingBox[3] = x;
-			if(y > boundingBox[4])
-				boundingBox[4] = y;
-			if(z > boundingBox[5])
-				boundingBox[5] = z;
 		}
-		void addBounds(Bounds& ob)
+		void addBounds(const Bounds& ob)
 		{
-			if(ob.boundingBox[0] < boundingBox[0])
+            boundingBox[0] = std::min<float>(boundingBox[0], ob.boundingBox[0]);
-				boundingBox[0] = ob.boundingBox[0];
+            boundingBox[1] = std::min<float>(boundingBox[1], ob.boundingBox[1]);
-			if(ob.boundingBox[1] < boundingBox[1])
+            boundingBox[2] = std::min<float>(boundingBox[2], ob.boundingBox[2]);
-				boundingBox[1] = ob.boundingBox[1];
+            boundingBox[3] = std::max<float>(boundingBox[3], ob.boundingBox[3]);
-			if(ob.boundingBox[2] < boundingBox[2])
+            boundingBox[4] = std::max<float>(boundingBox[4], ob.boundingBox[4]);
-				boundingBox[2] = ob.boundingBox[2];
+            boundingBox[5] = std::max<float>(boundingBox[5], ob.boundingBox[5]);
-
-			if(ob.boundingBox[3] > boundingBox[3])
-				boundingBox[3] = ob.boundingBox[3];
-			if(ob.boundingBox[4] > boundingBox[4])
-				boundingBox[4] = ob.boundingBox[4];
-			if(ob.boundingBox[5] > boundingBox[5])
-				boundingBox[5] = ob.boundingBox[5];
 		}
 		float boundingBox[6];	// 4J MGH added
 

Minecraft.Client/TileRenderer.cpp

@@ -142,6 +142,9 @@ int TileRenderer::getLightColor( Tile *tt, LevelSource *level, int x, int y, int
 	return tt->getLightColor(level, x, y, z);
 }
 
+static constexpr int TILE_RENDERER_CACHE_SIZE = 32 * 32 * 32;
+static thread_local unsigned int s_tlsCache[TILE_RENDERER_CACHE_SIZE];
+
 TileRenderer::TileRenderer( LevelSource* level, int xMin, int yMin, int zMin, unsigned char *tileIds )
 {
 	this->level = level;
@@ -153,14 +156,11 @@ TileRenderer::TileRenderer( LevelSource* level, int xMin, int yMin, int zMin, un
 	this->yMin2 = yMin-2;
 	this->zMin2 = zMin-2;
 	this->tileIds = tileIds;
-	cache = new unsigned int[32*32*32];
+	cache = s_tlsCache;
-	XMemSet(cache,0,32*32*32*sizeof(unsigned int));
+	std::memset(cache, 0, TILE_RENDERER_CACHE_SIZE * sizeof(unsigned int));
 }
 
-TileRenderer::~TileRenderer()
+TileRenderer::~TileRenderer() = default;
-{
-	delete cache;
-}
 
 TileRenderer::TileRenderer( LevelSource* level )
 {

Minecraft.Client/Windows64/Iggy/gdraw/gdraw_d3d11.cpp

@@ -26,15 +26,10 @@
 
 #define GDRAW_ASSERTS
 
-#ifndef WIN32_LEAN_AND_MEAN
-#define WIN32_LEAN_AND_MEAN
-#endif
-
 // We temporarily disable this warning for the shared interface portions
 #pragma warning (push)
 #pragma warning (disable: 4201) // nonstandard extension used : nameless struct/union
 
-#include <windows.h>
 #include <d3d11.h>
 #include "gdraw.h"
 #include "iggy.h"

Minecraft.Client/Windows64/KeyboardMouseInput.h

@@ -2,8 +2,6 @@
 
 #ifdef _WINDOWS64
 
-#include <windows.h>
-
 class KeyboardMouseInput
 {
 public:

Minecraft.Client/stdafx.h

@@ -23,6 +23,7 @@
 #ifdef _WINDOWS64
 #define _HAS_STD_BYTE 0     // solve (std::)'byte' ambiguity with windows headers
 #define WIN32_LEAN_AND_MEAN // Exclude rarely-used stuff from Windows headers
+#define NOMINMAX            // Exclude min/max macros from Windows headers
 // Windows Header Files:
 #include <malloc.h>
 #include <tchar.h>
@@ -132,12 +133,14 @@ typedef XUID GameSessionUID;
 #include <list>
 #include <map>
 #include <set>
+#include <array>
 #include <deque>
 #include <algorithm>
 #include <string>
 #include <sstream>
 #include <iostream>
 #include <exception>
+#include <mutex>
 
 #include <assert.h>
 

Minecraft.Server/Minecraft.Server.vcxproj

@@ -71,6 +71,7 @@
       <MultiProcessorCompilation>true</MultiProcessorCompilation>
       <PreprocessorDefinitions>_LARGE_WORLDS;_DEBUG_MENUS_ENABLED;_DEBUG;_CRT_NON_CONFORMING_SWPRINTFS;_CRT_SECURE_NO_WARNINGS;_WINDOWS64;MINECRAFT_SERVER_BUILD;%(PreprocessorDefinitions)</PreprocessorDefinitions>
       <AdditionalIncludeDirectories>..\Minecraft.Client;..\Minecraft.Client\Windows64\Iggy\include;..\Minecraft.Client\Xbox\Sentient\Include;..\Minecraft.World\x64headers;..\include;$(ProjectDir)Windows64;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
+      <LanguageStandard>stdcpp17</LanguageStandard>
     </ClCompile>
     <MASM>
       <UseSafeExceptionHandlers>false</UseSafeExceptionHandlers>
@@ -103,6 +104,7 @@
       <MultiProcessorCompilation>true</MultiProcessorCompilation>
       <PreprocessorDefinitions>_LARGE_WORLDS;_DEBUG_MENUS_ENABLED;_CRT_NON_CONFORMING_SWPRINTFS;_CRT_SECURE_NO_WARNINGS;_WINDOWS64;MINECRAFT_SERVER_BUILD;%(PreprocessorDefinitions)</PreprocessorDefinitions>
       <AdditionalIncludeDirectories>..\Minecraft.Client;..\Minecraft.Client\Windows64\Iggy\include;..\Minecraft.Client\Xbox\Sentient\Include;..\Minecraft.World\x64headers;..\include;$(ProjectDir)Windows64;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
+      <LanguageStandard>stdcpp17</LanguageStandard>
     </ClCompile>
     <MASM>
       <UseSafeExceptionHandlers>false</UseSafeExceptionHandlers>
@@ -746,4 +748,3 @@
     <Import Project="$(VCTargetsPath)\BuildCustomizations\masm.targets" />
   </ImportGroup>
 </Project>
-

Minecraft.World/Region.cpp

@@ -12,20 +12,7 @@
 
 Region::~Region()
 {
-	for(unsigned int i = 0; i < chunks->length; ++i)
+	// flatChunksHeap automatically freed by unique_ptr
-	{
-		LevelChunkArray *lca = (*chunks)[i];
-		delete [] lca->data;
-		delete lca;
-	}
-	delete [] chunks->data;
-	delete chunks;
-
-	// AP - added a caching system for Chunk::rebuild to take advantage of
-	if( CachedTiles )
-	{
-		free(CachedTiles);
-	}
 }
 
 Region::Region(Level *level, int x1, int y1, int z1, int x2, int y2, int z2, int r)
@@ -37,7 +24,21 @@ Region::Region(Level *level, int x1, int y1, int z1, int x2, int y2, int z2, int
 	int xc2 = (x2 + r) >> 4;
 	int zc2 = (z2 + r) >> 4;
 
-	chunks = new LevelChunk2DArray(xc2 - xc1 + 1, zc2 - zc1 + 1);
+	chunksDimX = xc2 - xc1 + 1;
+	chunksDimZ = zc2 - zc1 + 1;
+	int totalChunks = chunksDimX * chunksDimZ;
+
+	// Use stack buffer for common small regions (render chunks), heap for rare large ones (pathfinding)
+	if( totalChunks <= MAX_STACK_CHUNKS )
+	{
+		flatChunks = flatChunks_stack;
+	}
+	else
+	{
+		flatChunksHeap = std::make_unique<LevelChunk*[]>(totalChunks);
+		flatChunks = flatChunksHeap.get();
+	}
+	std::fill_n(flatChunks, totalChunks, nullptr);
 
 	allEmpty = true;
 	for (int xc = xc1; xc <= xc2; xc++)
@@ -47,8 +48,7 @@ Region::Region(Level *level, int x1, int y1, int z1, int x2, int y2, int z2, int
 			LevelChunk *chunk = level->getChunk(xc, zc);
 			if(chunk != nullptr)
 			{
-				LevelChunkArray *lca = (*chunks)[xc - xc1];
+				flatChunks[(xc - xc1) * chunksDimZ + (zc - zc1)] = chunk;
-				lca->data[zc - zc1] = chunk;
 			}
 		}
 	}
@@ -56,8 +56,7 @@ Region::Region(Level *level, int x1, int y1, int z1, int x2, int y2, int z2, int
 	{
 		for (int zc = (z1 >> 4); zc <= (z2 >> 4); zc++)
 		{
-			LevelChunkArray *lca = (*chunks)[xc - xc1];
+			LevelChunk *chunk = flatChunks[(xc - xc1) * chunksDimZ + (zc - zc1)];
-			LevelChunk *chunk = lca->data[zc - zc1];
 			if (chunk != nullptr)
 			{
 				if (!chunk->isYSpaceEmpty(y1, y2))
@@ -105,12 +104,12 @@ int Region::getTile(int x, int y, int z)
 	xc -= xc1;
 	zc -= zc1;
 
-	if (xc < 0 || xc >= static_cast<int>(chunks->length) || zc < 0 || zc >= static_cast<int>((*chunks)[xc]->length))
+	if (xc < 0 || xc >= chunksDimX || zc < 0 || zc >= chunksDimZ)
 	{
 		return 0;
 	}
 
-	LevelChunk *lc = (*chunks)[xc]->data[zc];
+	LevelChunk *lc = flatChunks[xc * chunksDimZ + zc];
 	if (lc == nullptr) return 0;
 
 	return lc->getTile(x & 15, y, z & 15);
@@ -122,11 +121,11 @@ void Region::setCachedTiles(unsigned char *tiles, int xc, int zc)
 	xcCached = xc;
 	zcCached = zc;
 	int size = 16 * 16 * Level::maxBuildHeight;
-	if( CachedTiles == nullptr )
+    if (!CachedTiles)
-	{
+    {
-		CachedTiles = static_cast<unsigned char *>(malloc(size));
+        CachedTiles = std::make_unique<unsigned char[]>(size);
-	}
+    }
-	memcpy(CachedTiles, tiles, size);
+    std::copy(tiles, tiles + size, CachedTiles.get());
 }
 
 LevelChunk* Region::getLevelChunk(int x, int y, int z)
@@ -137,12 +136,12 @@ LevelChunk* Region::getLevelChunk(int x, int y, int z)
 	int xc = (x >> 4) - xc1;
 	int zc = (z >> 4) - zc1;
 
-	if (xc < 0 || xc >= static_cast<int>(chunks->length) || zc < 0 || zc >= static_cast<int>((*chunks)[xc]->length))
+	if (xc < 0 || xc >= chunksDimX || zc < 0 || zc >= chunksDimZ)
 	{
 		return nullptr;
 	}
 
-	LevelChunk *lc = (*chunks)[xc]->data[zc];
+	LevelChunk *lc = flatChunks[xc * chunksDimZ + zc];
 	return lc;
 }
 
@@ -153,7 +152,15 @@ shared_ptr<TileEntity> Region::getTileEntity(int x, int y, int z)
 	int xc = (x >> 4) - xc1;
 	int zc = (z >> 4) - zc1;
 
-	return (*chunks)[xc]->data[zc]->getTileEntity(x & 15, y, z & 15);
+	if (xc < 0 || xc >= chunksDimX || zc < 0 || zc >= chunksDimZ)
+	{
+		return nullptr;
+	}
+
+	LevelChunk *lc = flatChunks[xc * chunksDimZ + zc];
+	if (lc == nullptr) return nullptr;
+
+	return lc->getTileEntity(x & 15, y, z & 15);
 }
 
 int Region::getLightColor(int x, int y, int z, int emitt, int tileId/*=-1*/)
@@ -228,7 +235,15 @@ int Region::getRawBrightness(int x, int y, int z, bool propagate)
 	int xc = (x >> 4) - xc1;
 	int zc = (z >> 4) - zc1;
 
-	return (*chunks)[xc]->data[zc]->getRawBrightness(x & 15, y, z & 15, level->skyDarken);
+	if (xc < 0 || xc >= chunksDimX || zc < 0 || zc >= chunksDimZ)
+	{
+		return 0;
+	}
+
+	LevelChunk *lc = flatChunks[xc * chunksDimZ + zc];
+	if (lc == nullptr) return 0;
+
+	return lc->getRawBrightness(x & 15, y, z & 15, level->skyDarken);
 }
 
 
@@ -239,7 +254,15 @@ int Region::getData(int x, int y, int z)
 	int xc = (x >> 4) - xc1;
 	int zc = (z >> 4) - zc1;
 
-	return (*chunks)[xc]->data[zc]->getData(x & 15, y, z & 15);
+	if (xc < 0 || xc >= chunksDimX || zc < 0 || zc >= chunksDimZ)
+	{
+		return 0;
+	}
+
+	LevelChunk *lc = flatChunks[xc * chunksDimZ + zc];
+	if (lc == nullptr) return 0;
+
+	return lc->getData(x & 15, y, z & 15);
 }
 
 Material *Region::getMaterial(int x, int y, int z)
@@ -321,7 +344,7 @@ int Region::getBrightnessPropagate(LightLayer::variety layer, int x, int y, int
 		// 4J Stu - The java LightLayer was an enum class type with a member "surrounding" which is what we
 		// were returning here. Surrounding has the same value as the enum value in our C++ code, so just cast
 		// it to an int
-		return (int)layer;
+		return static_cast<int>(layer);
 	}
 	if (layer == LightLayer::Sky && level->dimension->hasCeiling)
 	{
@@ -346,7 +369,15 @@ int Region::getBrightnessPropagate(LightLayer::variety layer, int x, int y, int
 	int xc = (x >> 4) - xc1;
 	int zc = (z >> 4) - zc1;
 
-	return (*chunks)[xc]->data[zc]->getBrightness(layer, x & 15, y, z & 15);
+	if (xc < 0 || xc >= chunksDimX || zc < 0 || zc >= chunksDimZ)
+	{
+		return static_cast<int>(layer);
+	}
+
+	LevelChunk *lc = flatChunks[xc * chunksDimZ + zc];
+	if (lc == nullptr) return static_cast<int>(layer);
+
+	return lc->getBrightness(layer, x & 15, y, z & 15);
 }
 
 // 4J - brought forward from 1.8.2
@@ -359,12 +390,20 @@ int Region::getBrightness(LightLayer::variety layer, int x, int y, int z)
 		// 4J Stu - The java LightLayer was an enum class type with a member "surrounding" which is what we
 		// were returning here. Surrounding has the same value as the enum value in our C++ code, so just cast
 		// it to an int
-		return (int)layer;
+		return static_cast<int>(layer);
 	}
 	int xc = (x >> 4) - xc1;
 	int zc = (z >> 4) - zc1;
 
-	return (*chunks)[xc]->data[zc]->getBrightness(layer, x & 15, y, z & 15);
+	if (xc < 0 || xc >= chunksDimX || zc < 0 || zc >= chunksDimZ)
+	{
+		return static_cast<int>(layer);
+	}
+
+	LevelChunk *lc = flatChunks[xc * chunksDimZ + zc];
+	if (lc == nullptr) return static_cast<int>(layer);
+
+	return lc->getBrightness(layer, x & 15, y, z & 15);
 }
 
 int Region::getMaxBuildHeight()

Minecraft.World/Region.h

@@ -9,18 +9,29 @@ class BiomeSource;
 class Region : public LevelSource
 {
 private:
+	static constexpr int MAX_STACK_CHUNKS = 16;  // 4x4 covers the common render-chunk case (r=1)
 	int xc1, zc1;
-	LevelChunk2DArray *chunks;
+	LevelChunk *flatChunks_stack[MAX_STACK_CHUNKS];
+	LevelChunk **flatChunks;  // non-owning: points to flatChunks_stack or flatChunksHeap's buffer
+	std::unique_ptr<LevelChunk*[]> flatChunksHeap;  // owns heap allocation for large regions
+	int chunksDimX, chunksDimZ;
 	Level *level;
 	bool allEmpty;
 
 	// AP - added a caching system for Chunk::rebuild to take advantage of
-	int xcCached, zcCached;
+    int xcCached, zcCached;
-	unsigned char *CachedTiles;
+    std::unique_ptr<unsigned char[]> CachedTiles;
 
 public:
 	Region(Level *level, int x1, int y1, int z1, int x2, int y2, int z2, int r);
 	virtual ~Region();
+
+	// Non-copyable/movable: flatChunks may point to internal stack buffer
+	Region(const Region&) = delete;
+	Region& operator=(const Region&) = delete;
+	Region(Region&&) = delete;
+	Region& operator=(Region&&) = delete;
+
 	bool isAllEmpty();
 	int getTile(int x, int y, int z);
 	shared_ptr<TileEntity> getTileEntity(int x, int y, int z);

Minecraft.World/stdafx.h

@@ -7,6 +7,7 @@
 #ifdef _WINDOWS64
 #define _HAS_STD_BYTE 0					// solve (std::)'byte' ambiguity with windows headers
 #define WIN32_LEAN_AND_MEAN             // Exclude rarely-used stuff from Windows headers
+#define NOMINMAX						// Exclude min/max macros from Windows headers
 // Windows Header Files:
 #include <windows.h>
 #include <malloc.h>
@@ -105,6 +106,7 @@ typedef XUID GameSessionUID;
 #include <sstream>
 #include <iostream>
 #include <exception>
+#include <array>
 
 #ifndef __PS3__			// the PS3 lib assert is rubbish, and aborts the code, we define our own in PS3Types.h
 #include <assert.h>

include/lce_filesystem/lce_filesystem.cpp

@@ -1,10 +1,6 @@
 #include "stdafx.h"
 #include "lce_filesystem.h"
 
-#ifdef _WINDOWS64
-#include <windows.h>
-#endif // TODO: More os' filesystem handling for when the project moves away from only Windows
-
 #include <stdio.h>
 
 bool FileOrDirectoryExists(const char* path)