Add new constructor to ZstdDictCompress and ZstdDictDecompress that

allows the byReference semantics for the provided byte buffer: If you
set this to true, you avoid the copying of the dict data into a
natively malloc'ed buffer, but then also have to promise that the byte
buffer will not be modified before the CTX has been closed.

src/main/java/com/github/luben/zstd/ZstdDictCompress.java

@@ -10,11 +10,14 @@ public class ZstdDictCompress extends SharedDictBase {
     }
 
     private long nativePtr = 0;
+
+    private ByteBuffer sharedDict = null;
+
     private int level = Zstd.defaultCompressionLevel();
 
     private native void init(byte[] dict, int dict_offset, int dict_size, int level);
 
-    private native void initDirect(ByteBuffer dict, int dict_offset, int dict_size, int level);
+    private native void initDirect(ByteBuffer dict, int dict_offset, int dict_size, int level, int byReference);
 
     private native void free();
 
@@ -59,6 +62,18 @@ public ZstdDictCompress(byte[] dict, int offset, int length, int level) {
      * @param level  compression level
      */
     public ZstdDictCompress(ByteBuffer dict, int level) {
+	this(dict, level, false);
+    }
+
+    /**
+     * Create a new dictionary for use with fast compress.
+     * If byReference is true, then the native code does not copy the data but keeps a reference to the byte buffer, which must then not be modified before this context has been closed.
+     *
+     * @param dict   Direct ByteBuffer containing dictionary using position and limit to define range in buffer.
+     * @param level  compression level
+     * @param byReference tell the native part to use the byte buffer directly and not copy the data when true.
+     */
+    public ZstdDictCompress(ByteBuffer dict, int level, boolean byReference) {
 	this.level = level;
 	int length = dict.limit() - dict.position();
         if (!dict.isDirect()) {
@@ -67,11 +82,14 @@ public ZstdDictCompress(ByteBuffer dict, int level) {
         if (length < 0) {
             throw new IllegalArgumentException("dict cannot be empty.");
         }
-	initDirect(dict, dict.position(), length, level);
+	initDirect(dict, dict.position(), length, level, byReference ? 1 : 0);
 
         if (nativePtr == 0L) {
            throw new IllegalStateException("ZSTD_createCDict failed");
         }
+	if (byReference) {
+	    sharedDict = dict; // ensures the dict is not garbage collected while this object remains, and flags that we should not use native free.
+	}
         // Ensures that even if ZstdDictCompress is created and published through a race, no thread could observe
         // nativePtr == 0.
         storeFence();
@@ -85,7 +103,11 @@ int level() {
     @Override
     void  doClose() {
         if (nativePtr != 0) {
-            free();
+	    if (sharedDict == null) {
+                free();
+	    } else {
+		sharedDict = null;
+	    }
             nativePtr = 0;
         }
     }

src/main/java/com/github/luben/zstd/ZstdDictDecompress.java

@@ -11,9 +11,11 @@ public class ZstdDictDecompress extends SharedDictBase {
 
     private long nativePtr = 0L;
 
+    private ByteBuffer sharedDict = null;
+
     private native void init(byte[] dict, int dict_offset, int dict_size);
 
-    private native void initDirect(ByteBuffer dict, int dict_offset, int dict_size);
+    private native void initDirect(ByteBuffer dict, int dict_offset, int dict_size, int byReference);
 
     private native void free();
 
@@ -52,6 +54,17 @@ public ZstdDictDecompress(byte[] dict, int offset, int length) {
      * @param dict   Direct ByteBuffer containing dictionary using position and limit to define range in buffer.
      */
     public ZstdDictDecompress(ByteBuffer dict) {
+	this(dict, false);
+    }
+
+    /**
+     * Create a new dictionary for use with fast decompress.
+     * If byReference is true, then the native code does not copy the data but keeps a reference to the byte buffer, which must then not be modified before this context has been closed.
+     *
+     * @param dict   Direct ByteBuffer containing dictionary using position and limit to define range in buffer.
+     * @param byReference tell the native part to use the byte buffer directly and not copy the data when true.
+     */
+    public ZstdDictDecompress(ByteBuffer dict, boolean byReference) {
 
 	int length = dict.limit() - dict.position();
         if (!dict.isDirect()) {
@@ -60,11 +73,14 @@ public ZstdDictDecompress(ByteBuffer dict) {
         if (length < 0) {
             throw new IllegalArgumentException("dict cannot be empty.");
         }
-	initDirect(dict, dict.position(), length);
+	initDirect(dict, dict.position(), length, byReference ? 1 : 0);
 
         if (nativePtr == 0L) {
            throw new IllegalStateException("ZSTD_createDDict failed");
         }
+	if (byReference) {
+	    sharedDict = dict; // ensures the dict is not garbage collected while this object remains, and flags that we should not use native free.
+	}
         // Ensures that even if ZstdDictDecompress is created and published through a race, no thread could observe
         // nativePtr == 0.
         storeFence();
@@ -74,7 +90,11 @@ public ZstdDictDecompress(ByteBuffer dict) {
     @Override
      void doClose() {
         if (nativePtr != 0) {
-            free();
+	    if (sharedDict == null) {
+                free();
+	    } else {
+		sharedDict = null;
+	    }
             nativePtr = 0;
         }
     }

src/main/native/jni_fast_zstd.c

@@ -32,17 +32,22 @@ JNIEXPORT void JNICALL Java_com_github_luben_zstd_ZstdDictCompress_init
 /*
  * Class:     com_github_luben_zstd_ZstdDictCompress
  * Method:    init
- * Signature: (Ljava/nio/ByteBuffer;III)V
+ * Signature: (Ljava/nio/ByteBuffer;IIII)V
  */
 JNIEXPORT void JNICALL Java_com_github_luben_zstd_ZstdDictCompress_initDirect
-  (JNIEnv *env, jobject obj, jobject dict, jint dict_offset, jint dict_size, jint level)
+  (JNIEnv *env, jobject obj, jobject dict, jint dict_offset, jint dict_size, jint level, jint byReference)
 {
     jclass clazz = (*env)->GetObjectClass(env, obj);
     compress_dict = (*env)->GetFieldID(env, clazz, "nativePtr", "J");
     if (NULL == dict) return;
     void *dict_buff = (*env)->GetDirectBufferAddress(env, dict);
     if (NULL == dict_buff) return;
-    ZSTD_CDict* cdict = ZSTD_createCDict(((char *)dict_buff) + dict_offset, dict_size, level);
+    ZSTD_CDict* cdict = NULL;
+    if (byReference == 0) {
+      cdict = ZSTD_createCDict(((char *)dict_buff) + dict_offset, dict_size, level);
+    } else {
+      cdict = ZSTD_createCDict_byReference(((char *)dict_buff) + dict_offset, dict_size, level);
+    }
     if (NULL == cdict) return;
     (*env)->SetLongField(env, obj, compress_dict, (jlong)(intptr_t) cdict);
 }
@@ -85,17 +90,22 @@ JNIEXPORT void JNICALL Java_com_github_luben_zstd_ZstdDictDecompress_init
 /*
  * Class:     com_github_luben_zstd_ZstdDictDecompress
  * Method:    initDirect
- * Signature: (Ljava/nio/ByteBuffer;II)V
+ * Signature: (Ljava/nio/ByteBuffer;III)V
  */
 JNIEXPORT void JNICALL Java_com_github_luben_zstd_ZstdDictDecompress_initDirect
-  (JNIEnv *env, jobject obj, jobject dict, jint dict_offset, jint dict_size)
+  (JNIEnv *env, jobject obj, jobject dict, jint dict_offset, jint dict_size, jint byReference)
 {
     jclass clazz = (*env)->GetObjectClass(env, obj);
     decompress_dict = (*env)->GetFieldID(env, clazz, "nativePtr", "J");
     if (NULL == dict) return;
     void *dict_buff = (*env)->GetDirectBufferAddress(env, dict);
 
-    ZSTD_DDict* ddict = ZSTD_createDDict(((char *)dict_buff) + dict_offset, dict_size);
+    ZSTD_DDict* ddict = NULL;
+    if (byReference == 0) {
+      ddict = ZSTD_createDDict(((char *)dict_buff) + dict_offset, dict_size);
+    } else {
+      ddict = ZSTD_createDDict_byReference(((char *)dict_buff) + dict_offset, dict_size);
+    }
 
     if (NULL == ddict) return;
     (*env)->SetLongField(env, obj, decompress_dict, (jlong)(intptr_t) ddict);

src/test/scala/ZstdDict.scala

@@ -104,17 +104,18 @@ class ZstdDictSpec extends AnyFlatSpec {
       assert(input.toSeq == decompressed.toSeq)
     }
 
-    it should s"round-trip compression/decompression ByteBuffers with fast dict at level $level with legacy $legacy" in {
+    it should s"round-trip compression/decompression ByteBuffers with fast dict at level $level with byReference $legacy" in {
+      val byReference = legacy // Reuse the variance flag here.
       val size = input.length
       val inBuf = ByteBuffer.allocateDirect(size)
       inBuf.put(input)
       inBuf.flip()
-      val cdict = new ZstdDictCompress(dictInDirectByteBuffer, level)
+      val cdict = new ZstdDictCompress(dictInDirectByteBuffer, level, byReference)
       val compressed = ByteBuffer.allocateDirect(Zstd.compressBound(size).toInt);
       Zstd.compress(compressed, inBuf, cdict)
       compressed.flip()
       cdict.close
-      val ddict = new ZstdDictDecompress(dictInDirectByteBuffer)
+      val ddict = new ZstdDictDecompress(dictInDirectByteBuffer, byReference)
       val decompressed = ByteBuffer.allocateDirect(size)
       Zstd.decompress(decompressed, compressed, ddict)
       decompressed.flip()