Looks like working block setting and removal

worldedit-bukkit/src/main/java/com/boydti/fawe/bukkit/FaweBukkit.java

@@ -58,6 +58,7 @@ public class FaweBukkit implements IFawe, Listener {
     private boolean listeningImages;
     private BukkitImageListener imageListener;
     private CFIPacketListener packetListener;
+    private final boolean chunksStretched;
 
     public VaultUtil getVault() {
         return this.vault;
@@ -99,6 +100,8 @@ public class FaweBukkit implements IFawe, Listener {
             // The tick limiter
             new ChunkListener_9();
         });
+
+        chunksStretched = Integer.parseInt(Bukkit.getMinecraftVersion().split("\\.")[1]) >= 16;
     }
 
     @Override // Please don't delete this again, it's WIP
@@ -302,6 +305,11 @@ public class FaweBukkit implements IFawe, Listener {
         return null;
     }
 
+    @Override
+    public boolean isChunksStretched() {
+        return chunksStretched;
+    }
+
     private void setupPlotSquared() {
         Plugin plotSquared = this.plugin.getServer().getPluginManager().getPlugin("PlotSquared");
         if (plotSquared == null)

worldedit-bukkit/src/main/java/com/boydti/fawe/bukkit/adapter/mc1_16_1/BukkitAdapter_1_16_1.java

@@ -5,7 +5,7 @@ import com.boydti.fawe.FaweCache;
 import com.boydti.fawe.bukkit.adapter.DelegateLock;
 import com.boydti.fawe.bukkit.adapter.NMSAdapter;
 import com.boydti.fawe.config.Settings;
-import com.boydti.fawe.object.collection.BitArray;
+import com.boydti.fawe.object.collection.BitArrayUnstretched;
 import com.boydti.fawe.util.MathMan;
 import com.boydti.fawe.util.ReflectionUtils;
 import com.boydti.fawe.util.TaskManager;
@@ -232,11 +232,13 @@ public final class BukkitAdapter_1_16_1 extends NMSAdapter {
                 bitsPerEntry = Math.max(bitsPerEntry, 1); // For some reason minecraft needs 4096 bits to store 0 entries
             }
 
-            final int blockBitArrayEnd = (bitsPerEntry * 4096) >> 6;
+            final int blocksPerLong = MathMan.floorZero((double) 64 / bitsPerEntry);
+            final int blockBitArrayEnd = MathMan.ceilZero((float) 4096 / blocksPerLong);
+
             if (num_palette == 1) {
                 for (int i = 0; i < blockBitArrayEnd; i++) blockStates[i] = 0;
             } else {
-                final BitArray bitArray = new BitArray(bitsPerEntry, 4096, blockStates);
+                final BitArrayUnstretched bitArray = new BitArrayUnstretched(bitsPerEntry, blockStates);
                 bitArray.fromRaw(blocksCopy);
             }
 

worldedit-bukkit/src/main/java/com/boydti/fawe/bukkit/adapter/mc1_16_1/BukkitGetBlocks_1_16_1.java

@@ -10,7 +10,7 @@ import com.boydti.fawe.bukkit.adapter.DelegateLock;
 import com.boydti.fawe.bukkit.adapter.mc1_16_1.nbt.LazyCompoundTag_1_16_1;
 import com.boydti.fawe.config.Settings;
 import com.boydti.fawe.object.collection.AdaptedMap;
-import com.boydti.fawe.object.collection.BitArray;
+import com.boydti.fawe.object.collection.BitArrayUnstretched;
 import com.google.common.base.Suppliers;
 import com.google.common.collect.Iterables;
 import com.sk89q.jnbt.Tag;
@@ -553,7 +553,7 @@ public class BukkitGetBlocks_1_16_1 extends CharGetBlocks {
                 final int bitsPerEntry = (int) BukkitAdapter_1_16_1.fieldBitsPerEntry.get(bits);
                 final long[] blockStates = bits.a();
 
-                new BitArray(bitsPerEntry, 4096, blockStates).toRaw(data);
+                new BitArrayUnstretched(bitsPerEntry, blockStates).toRaw(data);
 
                 int num_palette;
                 if (palette instanceof DataPaletteLinear) {

worldedit-core/src/main/java/com/boydti/fawe/FaweCache.java

@@ -1,7 +1,6 @@
 package com.boydti.fawe;
 
 import static com.google.common.base.Preconditions.checkNotNull;
-import static org.slf4j.LoggerFactory.getLogger;
 
 import com.boydti.fawe.beta.IChunkSet;
 import com.boydti.fawe.beta.Trimable;
@@ -9,6 +8,7 @@ import com.boydti.fawe.beta.implementation.queue.Pool;
 import com.boydti.fawe.beta.implementation.queue.QueuePool;
 import com.boydti.fawe.config.Settings;
 import com.boydti.fawe.object.collection.BitArray;
+import com.boydti.fawe.object.collection.BitArrayUnstretched;
 import com.boydti.fawe.object.collection.CleanableThreadLocal;
 import com.boydti.fawe.object.collection.VariableThreadLocal;
 import com.boydti.fawe.object.exception.FaweBlockBagException;
@@ -299,6 +299,102 @@ public enum FaweCache implements Trimable {
         }
     }
 
+    /**
+     * Convert raw int array to unstretched palette (1.16)
+     * @param layerOffset
+     * @param blocks
+     * @return palette
+     */
+    public Palette toPaletteUnstretched(int layerOffset, char[] blocks) {
+        return toPaletteUnstretched(layerOffset, null, blocks);
+    }
+
+    /**
+     * Convert raw int array to unstretched palette (1.16)
+     * @param layerOffset
+     * @param blocks
+     * @return palette
+     */
+    public Palette toPaletteUnstretched(int layerOffset, int[] blocks) {
+        return toPaletteUnstretched(layerOffset, blocks, null);
+    }
+
+    private Palette toPaletteUnstretched(int layerOffset, int[] blocksInts, char[] blocksChars) {
+        int[] blockToPalette = BLOCK_TO_PALETTE.get();
+        int[] paletteToBlock = PALETTE_TO_BLOCK.get();
+        long[] blockStates = BLOCK_STATES.get();
+        int[] blocksCopy = SECTION_BLOCKS.get();
+
+        try {
+            int num_palette = 0;
+            int blockIndexStart = layerOffset << 12;
+            int blockIndexEnd = blockIndexStart + 4096;
+            if (blocksChars != null) {
+                for (int i = blockIndexStart, j = 0; i < blockIndexEnd; i++, j++) {
+                    int ordinal = blocksChars[i];
+                    int palette = blockToPalette[ordinal];
+                    if (palette == Integer.MAX_VALUE) {
+                        blockToPalette[ordinal] = palette = num_palette;
+                        paletteToBlock[num_palette] = ordinal;
+                        num_palette++;
+                    }
+                    blocksCopy[j] = palette;
+                }
+            } else if (blocksInts != null) {
+                for (int i = blockIndexStart, j = 0; i < blockIndexEnd; i++, j++) {
+                    int ordinal = blocksInts[i];
+                    int palette = blockToPalette[ordinal];
+                    if (palette == Integer.MAX_VALUE) {
+                        //                        BlockState state = BlockTypesCache.states[ordinal];
+                        blockToPalette[ordinal] = palette = num_palette;
+                        paletteToBlock[num_palette] = ordinal;
+                        num_palette++;
+                    }
+                    blocksCopy[j] = palette;
+                }
+            } else {
+                throw new IllegalArgumentException();
+            }
+
+            for (int i = 0; i < num_palette; i++) {
+                blockToPalette[paletteToBlock[i]] = Integer.MAX_VALUE;
+            }
+
+            // BlockStates
+            int bitsPerEntry = MathMan.log2nlz(num_palette - 1);
+            if (Settings.IMP.PROTOCOL_SUPPORT_FIX || num_palette != 1) {
+                bitsPerEntry = Math.max(bitsPerEntry, 4); // Protocol support breaks <4 bits per entry
+            } else {
+                bitsPerEntry = Math.max(bitsPerEntry, 1); // For some reason minecraft needs 4096 bits to store 0 entries
+            }
+            int blocksPerLong = MathMan.floorZero((double) 64 / bitsPerEntry);
+            int blockBitArrayEnd = MathMan.ceilZero((float) 4096 / blocksPerLong);
+            if (num_palette == 1) {
+                // Set a value, because minecraft needs it for some  reason
+                blockStates[0] = 0;
+                blockBitArrayEnd = 1;
+            } else {
+                BitArrayUnstretched bitArray = new BitArrayUnstretched(bitsPerEntry, blockStates);
+                bitArray.fromRaw(blocksCopy);
+            }
+
+            // Construct palette
+            Palette palette = PALETTE_CACHE.get();
+            palette.bitsPerEntry = bitsPerEntry;
+            palette.paletteToBlockLength = num_palette;
+            palette.paletteToBlock = paletteToBlock;
+
+            palette.blockStatesLength = blockBitArrayEnd;
+            palette.blockStates = blockStates;
+
+            return palette;
+        } catch (Throwable e) {
+            e.printStackTrace();
+            Arrays.fill(blockToPalette, Integer.MAX_VALUE);
+            throw e;
+        }
+    }
+
     /*
      * Vector cache
      */

worldedit-core/src/main/java/com/boydti/fawe/IFawe.java

@@ -41,4 +41,8 @@ public interface IFawe {
 
     Preloader getPreloader();
 
+    default boolean isChunksStretched() {
+        return true;
+    }
+
 }

worldedit-core/src/main/java/com/boydti/fawe/beta/IBlocks.java

@@ -45,11 +45,11 @@ public interface IBlocks extends Trimable {
 
     IBlocks reset();
 
-    default byte[] toByteArray(boolean full) {
+    default byte[] toByteArray(boolean full, boolean stretched) {
-        return toByteArray(null, getBitMask(), full);
+        return toByteArray(null, getBitMask(), full, stretched);
     }
 
-    default byte[] toByteArray(byte[] buffer, int bitMask, boolean full) {
+    default byte[] toByteArray(byte[] buffer, int bitMask, boolean full, boolean stretched) {
         if (buffer == null) {
             buffer = new byte[1024];
         }
@@ -81,7 +81,12 @@ public interface IBlocks extends Trimable {
                 }
 
                 sectionWriter.writeShort(nonEmpty); // non empty
-                FaweCache.Palette palette = FaweCache.IMP.toPalette(0, ids);
+                FaweCache.Palette palette;
+                if (stretched) {
+                     palette = FaweCache.IMP.toPalette(0, ids);
+                } else {
+                    palette = FaweCache.IMP.toPaletteUnstretched(0, ids);
+                }
 
                 sectionWriter.writeByte(palette.bitsPerEntry); // bits per block
                 sectionWriter.writeVarInt(palette.paletteToBlockLength);

worldedit-core/src/main/java/com/boydti/fawe/beta/implementation/packet/ChunkPacket.java

@@ -1,10 +1,12 @@
 package com.boydti.fawe.beta.implementation.packet;
 
+import com.boydti.fawe.Fawe;
 import com.boydti.fawe.FaweCache;
 import com.boydti.fawe.beta.IBlocks;
 import com.boydti.fawe.object.FaweOutputStream;
 import com.boydti.fawe.object.io.FastByteArrayOutputStream;
 import com.sk89q.jnbt.CompoundTag;
+import com.sk89q.worldedit.WorldEdit;
 
 import java.util.HashMap;
 import java.util.function.Function;
@@ -64,7 +66,7 @@ public class ChunkPacket implements Function<byte[], byte[]>, Supplier<byte[]> {
                 if (sectionBytes == null) {
                     IBlocks tmpChunk = getChunk();
                     byte[] buf = FaweCache.IMP.BYTE_BUFFER_8192.get();
-                    sectionBytes = tmpChunk.toByteArray(buf, tmpChunk.getBitMask(), this.full);
+                    sectionBytes = tmpChunk.toByteArray(buf, tmpChunk.getBitMask(), this.full, Fawe.imp().isChunksStretched());
                 }
                 tmp = sectionBytes;
             }
@@ -72,6 +74,7 @@ public class ChunkPacket implements Function<byte[], byte[]>, Supplier<byte[]> {
         return tmp;
     }
 
+
     public Object getNativePacket() {
         return nativePacket;
     }

worldedit-core/src/main/java/com/boydti/fawe/object/collection/BitArrayUnstretched.java

@@ -0,0 +1,115 @@
+package com.boydti.fawe.object.collection;
+
+import com.boydti.fawe.util.MathMan;
+
+public final class BitArrayUnstretched {
+
+    private final long[] data;
+    private final int bitsPerEntry;
+    private final int maxSeqLocIndex;
+    private final int emptyBitCount;
+    private final long mask;
+    private final int longLen;
+
+    public BitArrayUnstretched(int bitsPerEntry, long[] buffer) {
+        this.bitsPerEntry = bitsPerEntry;
+        this.mask = (1L << bitsPerEntry) - 1L;
+        this.emptyBitCount = 64 % bitsPerEntry;
+        this.maxSeqLocIndex = 64 - (bitsPerEntry + emptyBitCount);
+        final int blocksPerLong = MathMan.floorZero((double) 64 / bitsPerEntry);
+        this.longLen = MathMan.ceilZero((float) 4096 / blocksPerLong);
+        if (buffer.length < longLen) {
+            this.data = new long[longLen];
+        } else {
+            this.data = buffer;
+        }
+    }
+
+    public long[] getData() {
+        return data;
+    }
+
+    public final void set(int index, int value) {
+        if (longLen == 0) return;
+        int bitIndexStart = index * bitsPerEntry + MathMan.floorZero((double) index / longLen) * emptyBitCount;
+        int longIndexStart = bitIndexStart >> 6;
+        int localBitIndexStart = bitIndexStart & 63;
+        this.data[longIndexStart] = this.data[longIndexStart] & ~(mask << localBitIndexStart) | (long) value << localBitIndexStart;
+    }
+
+    public final int get(int index) {
+        if (longLen == 0) return 0;
+        int bitIndexStart = index * bitsPerEntry + MathMan.floorZero((double) index / longLen) * emptyBitCount;
+
+        int longIndexStart = bitIndexStart >> 6;
+
+        int localBitIndexStart = bitIndexStart & 63;
+        return (int)(this.data[longIndexStart] >>> localBitIndexStart & mask);
+    }
+
+    public int getLength() {
+        return longLen;
+    }
+    
+    public final void fromRaw(int[] arr) {
+        final long[] data = this.data;
+        final int bitsPerEntry = this.bitsPerEntry;
+        final int maxSeqLocIndex = this.maxSeqLocIndex;
+
+        int localStart = 0;
+        int arrI = 0;
+        long l = 0;
+        for (int i = 0; i < longLen; i++) {
+            int lastVal;
+            for (; localStart <= maxSeqLocIndex && arrI < 4096; localStart += bitsPerEntry) {
+                lastVal = arr[arrI++];
+                l |= ((long) lastVal << localStart);
+            }
+            localStart = 0;
+            data[i] = l;
+            l = 0;
+        }
+    }
+
+    public final int[] toRaw() {
+        return toRaw(new int[4096]);
+    }
+
+    public final int[] toRaw(int[] buffer) {
+        final long[] data = this.data;
+        final int bitsPerEntry = this.bitsPerEntry;
+        final int maxSeqLocIndex = this.maxSeqLocIndex;
+
+        int localStart = 0;
+        int arrI = 0;
+        for (int i = 0; i < longLen; i++) {
+            long l = data[i];
+            char lastVal;
+            for (; localStart <= maxSeqLocIndex && arrI < 4096; localStart += bitsPerEntry) {
+                lastVal = (char) (l >>> localStart & this.mask);
+                buffer[arrI++] = lastVal;
+            }
+            localStart = 0;
+        }
+        return buffer;
+    }
+
+    public final char[] toRaw(char[] buffer) {
+        final long[] data = this.data;
+        final int bitsPerEntry = this.bitsPerEntry;
+        final int maxSeqLocIndex = this.maxSeqLocIndex;
+
+        int localStart = 0;
+        int arrI = 0;
+        for (int i = 0; i < longLen; i++) {
+            long l = data[i];
+            char lastVal;
+            for (; localStart <= maxSeqLocIndex && arrI < 4096; localStart += bitsPerEntry) {
+                lastVal = (char) (l >>> localStart & this.mask);
+                buffer[arrI++] = lastVal;
+            }
+            localStart = 0;
+        }
+        return buffer;
+    }
+}