diff --git a/CMakeLists.txt b/CMakeLists.txt
index 8d269969..61a35cde 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -19,6 +19,7 @@ option (OPENDHT_SHARED "Build shared library" ON)
 option (OPENDHT_LOG "Build with logs" ON)
 option (OPENDHT_PYTHON "Build Python bindings" OFF)
 option (OPENDHT_TOOLS "Build DHT tools" ON)
+option (OPENDHT_ARGON2 "Use included argon2 sources" OFF)
 option (OPENDHT_LTO "Build with LTO" OFF)
 
 find_package (GnuTLS 3.3 REQUIRED)
@@ -26,6 +27,14 @@ find_package (Msgpack 1.2 REQUIRED)
 if (OPENDHT_TOOLS)
     find_package (Readline 6 REQUIRED)
 endif ()
+if (NOT OPENDHT_ARGON2)
+    find_package(PkgConfig)
+    pkg_search_module(argon2 libargon2)
+    if (NOT argon2_FOUND)
+        message("Argon2 not found, using included version.")
+        set(OPENDHT_ARGON2 ON)
+    endif()
+endif ()
 
 set (CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -pthread -std=c++11 -Wno-return-type -Wall -Wextra -Wnon-virtual-dtor -pedantic-errors -fvisibility=hidden")
 set (CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DMSGPACK_DISABLE_LEGACY_NIL -DMSGPACK_DISABLE_LEGACY_CONVERT")
@@ -61,16 +70,21 @@ list (APPEND opendht_SOURCES
     src/network_engine.cpp
     src/securedht.cpp
     src/dhtrunner.cpp
-    src/argon2/argon2.c
-    src/argon2/core.c
-    src/argon2/blake2/blake2b.c
-    src/argon2/thread.c
-    src/argon2/encoding.c
-    src/argon2/ref.c
     src/indexation/pht.cpp
     src/log.cpp
 )
 
+if (OPENDHT_ARGON2)
+    list (APPEND opendht_SOURCES
+        argon2/argon2.c
+        argon2/core.c
+        argon2/blake2/blake2b.c
+        argon2/thread.c
+        argon2/encoding.c
+        argon2/ref.c
+    )
+endif ()
+
 list (APPEND opendht_HEADERS
     include/opendht/def.h
     include/opendht/utils.h
@@ -122,6 +136,12 @@ if (OPENDHT_STATIC)
     if (OPENDHT_LTO)
         target_link_libraries(opendht-static -flto -fuse-linker-plugin)
     endif ()
+    if (OPENDHT_ARGON2)
+        target_include_directories(opendht-static SYSTEM PRIVATE argon2)
+    else ()
+        target_link_libraries(opendht-static ${argon2_LIBRARIES})
+        target_include_directories(opendht-static SYSTEM PRIVATE ${argon2_INCLUDE_DIRS})
+    endif ()
     target_link_libraries(opendht-static gnutls nettle)
     install (TARGETS opendht-static DESTINATION ${CMAKE_INSTALL_LIBDIR} EXPORT opendht)
 endif ()
@@ -137,6 +157,12 @@ if (OPENDHT_SHARED)
     if (OPENDHT_LTO)
         target_link_libraries(opendht -flto -fuse-linker-plugin)
     endif ()
+    if (OPENDHT_ARGON2)
+        target_include_directories(opendht SYSTEM PRIVATE argon2)
+    else ()
+        target_link_libraries(opendht ${argon2_LIBRARIES})
+        target_include_directories(opendht SYSTEM PRIVATE ${argon2_INCLUDE_DIRS})
+    endif ()
     target_link_libraries(opendht gnutls nettle)
     install (TARGETS opendht DESTINATION ${CMAKE_INSTALL_LIBDIR} EXPORT opendht)
 endif ()
diff --git a/Makefile.am b/Makefile.am
index a8835d8a..1bf729e0 100644
--- a/Makefile.am
+++ b/Makefile.am
@@ -1,6 +1,12 @@
 AM_CXXFLAGS = -pthread
 
-SUBDIRS = src
+SUBDIRS =
+
+if WITH_INCLUDED_ARGON2
+SUBDIRS += argon2
+endif
+
+SUBDIRS += src
 
 if ENABLE_TOOLS
 SUBDIRS += tools
diff --git a/src/argon2/LICENSE b/argon2/LICENSE
similarity index 100%
rename from src/argon2/LICENSE
rename to argon2/LICENSE
diff --git a/src/argon2/Makefile.am b/argon2/Makefile.am
similarity index 100%
rename from src/argon2/Makefile.am
rename to argon2/Makefile.am
diff --git a/src/argon2/argon2.c b/argon2/argon2.c
similarity index 86%
rename from src/argon2/argon2.c
rename to argon2/argon2.c
index 52fbae9b..7e2ec8ff 100644
--- a/src/argon2/argon2.c
+++ b/argon2/argon2.c
@@ -104,6 +104,14 @@ int argon2_hash(const uint32_t t_cost, const uint32_t m_cost,
     int result;
     uint8_t *out;
 
+    if (pwdlen > ARGON2_MAX_PWD_LENGTH) {
+        return ARGON2_PWD_TOO_LONG;
+    }
+
+    if (saltlen > ARGON2_MAX_SALT_LENGTH) {
+        return ARGON2_SALT_TOO_LONG;
+    }
+
     if (hashlen > ARGON2_MAX_OUTLEN) {
         return ARGON2_OUTPUT_TOO_LONG;
     }
@@ -139,7 +147,7 @@ int argon2_hash(const uint32_t t_cost, const uint32_t m_cost,
     result = argon2_ctx(&context, type);
 
     if (result != ARGON2_OK) {
-        secure_wipe_memory(out, hashlen);
+        clear_internal_memory(out, hashlen);
         free(out);
         return result;
     }
@@ -152,13 +160,13 @@ int argon2_hash(const uint32_t t_cost, const uint32_t m_cost,
     /* if encoding requested, write it */
     if (encoded && encodedlen) {
         if (encode_string(encoded, encodedlen, &context, type) != ARGON2_OK) {
-            secure_wipe_memory(out, hashlen); /* wipe buffers if error */
-            secure_wipe_memory(encoded, encodedlen);
+            clear_internal_memory(out, hashlen); /* wipe buffers if error */
+            clear_internal_memory(encoded, encodedlen);
             free(out);
             return ARGON2_ENCODING_FAIL;
         }
     }
-    secure_wipe_memory(out, hashlen);
+    clear_internal_memory(out, hashlen);
     free(out);
 
     return ARGON2_OK;
@@ -238,69 +246,64 @@ int argon2_verify(const char *encoded, const void *pwd, const size_t pwdlen,
                   argon2_type type) {
 
     argon2_context ctx;
-    uint8_t *out;
-    int ret;
-    int decode_result;
-    uint32_t encoded_len;
-    size_t encoded_len_tmp;
+    uint8_t *desired_result = NULL;
 
-    if(encoded == NULL) {
+    int ret = ARGON2_OK;
+
+    size_t encoded_len;
+    uint32_t max_field_len;
+
+    if (pwdlen > ARGON2_MAX_PWD_LENGTH) {
+        return ARGON2_PWD_TOO_LONG;
+    }
+
+    if (encoded == NULL) {
         return ARGON2_DECODING_FAIL;
     }
 
-    encoded_len_tmp = strlen(encoded);
-    /* max values, to be updated in decode_string */
-    if (UINT32_MAX < encoded_len_tmp) {
+    encoded_len = strlen(encoded);
+    if (encoded_len > UINT32_MAX) {
         return ARGON2_DECODING_FAIL;
     }
 
-    encoded_len = (uint32_t)encoded_len_tmp;
-    ctx.adlen = encoded_len;
-    ctx.saltlen = encoded_len;
-    ctx.outlen = encoded_len;
-    ctx.allocate_cbk = NULL;
-    ctx.free_cbk = NULL;
-    ctx.secret = NULL;
-    ctx.secretlen = 0;
-    ctx.pwdlen = 0;
-    ctx.pwd = NULL;
-    ctx.ad = malloc(ctx.adlen);
+    /* No field can be longer than the encoded length */
+    max_field_len = (uint32_t)encoded_len;
+
+    ctx.saltlen = max_field_len;
+    ctx.outlen = max_field_len;
+
     ctx.salt = malloc(ctx.saltlen);
     ctx.out = malloc(ctx.outlen);
-    if (!ctx.out || !ctx.salt || !ctx.ad) {
-        free(ctx.ad);
-        free(ctx.salt);
-        free(ctx.out);
-        return ARGON2_MEMORY_ALLOCATION_ERROR;
-    }
-    out = malloc(ctx.outlen);
-    if (!out) {
-        free(ctx.ad);
-        free(ctx.salt);
-        free(ctx.out);
-        return ARGON2_MEMORY_ALLOCATION_ERROR;
-    }
-    decode_result = decode_string(&ctx, encoded, type);
-    if (decode_result != ARGON2_OK) {
-        free(ctx.ad);
-        free(ctx.salt);
-        free(ctx.out);
-        free(out);
-        return decode_result;
+    if (!ctx.salt || !ctx.out) {
+        ret = ARGON2_MEMORY_ALLOCATION_ERROR;
+        goto fail;
     }
 
-    ret = argon2_hash(ctx.t_cost, ctx.m_cost, ctx.threads, pwd, pwdlen,
-                      ctx.salt, ctx.saltlen, out, ctx.outlen, NULL, 0, type,
-                      ctx.version);
+    ctx.pwd = (uint8_t *)pwd;
+    ctx.pwdlen = (uint32_t)pwdlen;
 
-    free(ctx.ad);
+    ret = decode_string(&ctx, encoded, type);
+    if (ret != ARGON2_OK) {
+        goto fail;
+    }
+
+    /* Set aside the desired result, and get a new buffer. */
+    desired_result = ctx.out;
+    ctx.out = malloc(ctx.outlen);
+    if (!ctx.out) {
+        ret = ARGON2_MEMORY_ALLOCATION_ERROR;
+        goto fail;
+    }
+
+    ret = argon2_verify_ctx(&ctx, (char *)desired_result, type);
+    if (ret != ARGON2_OK) {
+        goto fail;
+    }
+
+fail:
     free(ctx.salt);
-
-    if (ret == ARGON2_OK && argon2_compare(out, ctx.out, ctx.outlen)) {
-        ret = ARGON2_VERIFY_MISMATCH;
-    }
-    free(out);
     free(ctx.out);
+    free(desired_result);
 
     return ret;
 }
@@ -334,18 +337,16 @@ int argon2id_ctx(argon2_context *context) {
 
 int argon2_verify_ctx(argon2_context *context, const char *hash,
                       argon2_type type) {
-    int result;
-    if (0 == context->outlen || NULL == hash) {
-        return ARGON2_OUT_PTR_MISMATCH;
+    int ret = argon2_ctx(context, type);
+    if (ret != ARGON2_OK) {
+        return ret;
     }
 
-    result = argon2_ctx(context, type);
-
-    if (ARGON2_OK != result) {
-        return result;
+    if (argon2_compare((uint8_t *)hash, context->out, context->outlen)) {
+        return ARGON2_VERIFY_MISMATCH;
     }
 
-    return 0 == memcmp(hash, context->out, context->outlen);
+    return ARGON2_OK;
 }
 
 int argon2d_verify_ctx(argon2_context *context, const char *hash) {
diff --git a/src/argon2/argon2.h b/argon2/argon2.h
similarity index 100%
rename from src/argon2/argon2.h
rename to argon2/argon2.h
diff --git a/src/argon2/blake2/blake2-impl.h b/argon2/blake2/blake2-impl.h
similarity index 98%
rename from src/argon2/blake2/blake2-impl.h
rename to argon2/blake2/blake2-impl.h
index 083bef48..80f97037 100644
--- a/src/argon2/blake2/blake2-impl.h
+++ b/argon2/blake2/blake2-impl.h
@@ -151,6 +151,6 @@ static BLAKE2_INLINE uint64_t rotr64(const uint64_t w, const unsigned c) {
     return (w >> c) | (w << (64 - c));
 }
 
-void secure_wipe_memory(void *v, size_t n);
+void clear_internal_memory(void *v, size_t n);
 
 #endif
diff --git a/src/argon2/blake2/blake2.h b/argon2/blake2/blake2.h
similarity index 100%
rename from src/argon2/blake2/blake2.h
rename to argon2/blake2/blake2.h
diff --git a/src/argon2/blake2/blake2b.c b/argon2/blake2/blake2b.c
similarity index 96%
rename from src/argon2/blake2/blake2b.c
rename to argon2/blake2/blake2b.c
index 603bc819..156476e7 100644
--- a/src/argon2/blake2/blake2b.c
+++ b/argon2/blake2/blake2b.c
@@ -61,7 +61,7 @@ static BLAKE2_INLINE void blake2b_increment_counter(blake2b_state *S,
 }
 
 static BLAKE2_INLINE void blake2b_invalidate_state(blake2b_state *S) {
-    secure_wipe_memory(S, sizeof(*S));      /* wipe */
+    clear_internal_memory(S, sizeof(*S));      /* wipe */
     blake2b_set_lastblock(S); /* invalidate for further use */
 }
 
@@ -157,7 +157,8 @@ int blake2b_init_key(blake2b_state *S, size_t outlen, const void *key,
         memset(block, 0, BLAKE2B_BLOCKBYTES);
         memcpy(block, key, keylen);
         blake2b_update(S, block, BLAKE2B_BLOCKBYTES);
-        secure_wipe_memory(block, BLAKE2B_BLOCKBYTES); /* Burn the key from stack */
+        /* Burn the key from stack */
+        clear_internal_memory(block, BLAKE2B_BLOCKBYTES);
     }
     return 0;
 }
@@ -284,9 +285,9 @@ int blake2b_final(blake2b_state *S, void *out, size_t outlen) {
     }
 
     memcpy(out, buffer, S->outlen);
-    secure_wipe_memory(buffer, sizeof(buffer));
-    secure_wipe_memory(S->buf, sizeof(S->buf));
-    secure_wipe_memory(S->h, sizeof(S->h));
+    clear_internal_memory(buffer, sizeof(buffer));
+    clear_internal_memory(S->buf, sizeof(S->buf));
+    clear_internal_memory(S->h, sizeof(S->h));
     return 0;
 }
 
@@ -324,7 +325,7 @@ int blake2b(void *out, size_t outlen, const void *in, size_t inlen,
     ret = blake2b_final(&S, out, outlen);
 
 fail:
-    secure_wipe_memory(&S, sizeof(S));
+    clear_internal_memory(&S, sizeof(S));
     return ret;
 }
 
@@ -382,7 +383,7 @@ int blake2b_long(void *pout, size_t outlen, const void *in, size_t inlen) {
         memcpy(out, out_buffer, toproduce);
     }
 fail:
-    secure_wipe_memory(&blake_state, sizeof(blake_state));
+    clear_internal_memory(&blake_state, sizeof(blake_state));
     return ret;
 #undef TRY
 }
diff --git a/src/argon2/blake2/blamka-round-opt.h b/argon2/blake2/blamka-round-opt.h
similarity index 100%
rename from src/argon2/blake2/blamka-round-opt.h
rename to argon2/blake2/blamka-round-opt.h
diff --git a/src/argon2/blake2/blamka-round-ref.h b/argon2/blake2/blamka-round-ref.h
similarity index 100%
rename from src/argon2/blake2/blamka-round-ref.h
rename to argon2/blake2/blamka-round-ref.h
diff --git a/src/argon2/core.c b/argon2/core.c
similarity index 81%
rename from src/argon2/core.c
rename to argon2/core.c
index ed02494a..69b2320c 100644
--- a/src/argon2/core.c
+++ b/argon2/core.c
@@ -82,25 +82,43 @@ static void store_block(void *output, const block *src) {
     }
 }
 
-/***************Memory allocators*****************/
-int allocate_memory(block **memory, uint32_t m_cost) {
-    if (memory != NULL) {
-        size_t memory_size = sizeof(block) * m_cost;
-        if (m_cost != 0 &&
-            memory_size / m_cost !=
-                sizeof(block)) { /*1. Check for multiplication overflow*/
-            return ARGON2_MEMORY_ALLOCATION_ERROR;
-        }
+/***************Memory functions*****************/
 
-        *memory = (block *)malloc(memory_size); /*2. Try to allocate*/
-
-        if (!*memory) {
-            return ARGON2_MEMORY_ALLOCATION_ERROR;
-        }
-        return ARGON2_OK;
-    } else {
+int allocate_memory(const argon2_context *context, uint8_t **memory,
+                    size_t num, size_t size) {
+    size_t memory_size = num*size;
+    if (memory == NULL) {
         return ARGON2_MEMORY_ALLOCATION_ERROR;
     }
+
+    /* 1. Check for multiplication overflow */
+    if (size != 0 && memory_size / size != num) {
+        return ARGON2_MEMORY_ALLOCATION_ERROR;
+    }
+
+    /* 2. Try to allocate with appropriate allocator */
+    if (context->allocate_cbk) {
+        (context->allocate_cbk)(memory, memory_size);
+    } else {
+        *memory = malloc(memory_size);
+    }
+
+    if (*memory == NULL) {
+        return ARGON2_MEMORY_ALLOCATION_ERROR;
+    }
+
+    return ARGON2_OK;
+}
+
+void free_memory(const argon2_context *context, uint8_t *memory,
+                 size_t num, size_t size) {
+    size_t memory_size = num*size;
+    clear_internal_memory(memory, memory_size);
+    if (context->free_cbk) {
+        (context->free_cbk)(memory, memory_size);
+    } else {
+        free(memory);
+    }
 }
 
 void NOT_OPTIMIZED secure_wipe_memory(void *v, size_t n) {
@@ -116,17 +134,14 @@ void NOT_OPTIMIZED secure_wipe_memory(void *v, size_t n) {
 #endif
 }
 
-/*********Memory functions*/
-
-void clear_memory(argon2_instance_t *instance, int clear) {
-    if (instance->memory != NULL && clear) {
-        secure_wipe_memory(instance->memory,
-                           sizeof(block) * instance->memory_blocks);
-    }
+/* Memory clear flag defaults to true. */
+int FLAG_clear_internal_memory = 1;
+void clear_internal_memory(void *v, size_t n) {
+  if (FLAG_clear_internal_memory && v) {
+    secure_wipe_memory(v, n);
+  }
 }
 
-void free_memory(block *memory) { free(memory); }
-
 void finalize(const argon2_context *context, argon2_instance_t *instance) {
     if (context != NULL && instance != NULL) {
         block blockhash;
@@ -147,26 +162,17 @@ void finalize(const argon2_context *context, argon2_instance_t *instance) {
             store_block(blockhash_bytes, &blockhash);
             blake2b_long(context->out, context->outlen, blockhash_bytes,
                          ARGON2_BLOCK_SIZE);
-            secure_wipe_memory(blockhash.v,
-                               ARGON2_BLOCK_SIZE); /* clear blockhash */
-            secure_wipe_memory(blockhash_bytes,
-                               ARGON2_BLOCK_SIZE); /* clear blockhash_bytes */
+            /* clear blockhash and blockhash_bytes */
+            clear_internal_memory(blockhash.v, ARGON2_BLOCK_SIZE);
+            clear_internal_memory(blockhash_bytes, ARGON2_BLOCK_SIZE);
         }
 
 #ifdef GENKAT
         print_tag(context->out, context->outlen);
 #endif
 
-        /* Clear memory */
-        clear_memory(instance, context->flags & ARGON2_FLAG_CLEAR_PASSWORD);
-
-        /* Deallocate the memory */
-        if (NULL != context->free_cbk) {
-            context->free_cbk((uint8_t *)instance->memory,
-                              instance->memory_blocks * sizeof(block));
-        } else {
-            free_memory(instance->memory);
-        }
+        free_memory(context, (uint8_t *)instance->memory,
+                    instance->memory_blocks, sizeof(block));
     }
 }
 
@@ -240,29 +246,45 @@ uint32_t index_alpha(const argon2_instance_t *instance,
     return absolute_position;
 }
 
+/* Single-threaded version for p=1 case */
+static int fill_memory_blocks_st(argon2_instance_t *instance) {
+    uint32_t r, s, l;
+
+    for (r = 0; r < instance->passes; ++r) {
+        for (s = 0; s < ARGON2_SYNC_POINTS; ++s) {
+            for (l = 0; l < instance->lanes; ++l) {
+                argon2_position_t position = {r, l, (uint8_t)s, 0};
+                fill_segment(instance, position);
+            }
+        }
+#ifdef GENKAT
+        internal_kat(instance, r); /* Print all memory blocks */
+#endif
+    }
+    return ARGON2_OK;
+}
+
+#if !defined(ARGON2_NO_THREADS)
+
 #ifdef _WIN32
 static unsigned __stdcall fill_segment_thr(void *thread_data)
 #else
 static void *fill_segment_thr(void *thread_data)
 #endif
 {
-    argon2_thread_data *my_data = (argon2_thread_data *)thread_data;
+    argon2_thread_data *my_data = thread_data;
     fill_segment(my_data->instance_ptr, my_data->pos);
     argon2_thread_exit();
     return 0;
 }
 
-int fill_memory_blocks(argon2_instance_t *instance) {
+/* Multi-threaded version for p > 1 case */
+static int fill_memory_blocks_mt(argon2_instance_t *instance) {
     uint32_t r, s;
     argon2_thread_handle_t *thread = NULL;
     argon2_thread_data *thr_data = NULL;
     int rc = ARGON2_OK;
 
-    if (instance == NULL || instance->lanes == 0) {
-        rc = ARGON2_THREAD_FAIL;
-        goto fail;
-    }
-
     /* 1. Allocating space for threads */
     thread = calloc(instance->lanes, sizeof(argon2_thread_handle_t));
     if (thread == NULL) {
@@ -336,6 +358,20 @@ fail:
     return rc;
 }
 
+#endif /* ARGON2_NO_THREADS */
+
+int fill_memory_blocks(argon2_instance_t *instance) {
+	if (instance == NULL || instance->lanes == 0) {
+	    return ARGON2_INCORRECT_PARAMETER;
+    }
+#if defined(ARGON2_NO_THREADS)
+    return fill_memory_blocks_st(instance);
+#else
+    return instance->threads == 1 ?
+			fill_memory_blocks_st(instance) : fill_memory_blocks_mt(instance);
+#endif
+}
+
 int validate_inputs(const argon2_context *context) {
     if (NULL == context) {
         return ARGON2_INCORRECT_PARAMETER;
@@ -354,37 +390,37 @@ int validate_inputs(const argon2_context *context) {
         return ARGON2_OUTPUT_TOO_LONG;
     }
 
-    /* Validate password length */
+    /* Validate password (required param) */
     if (NULL == context->pwd) {
         if (0 != context->pwdlen) {
             return ARGON2_PWD_PTR_MISMATCH;
         }
-    } else {
-        if (ARGON2_MIN_PWD_LENGTH > context->pwdlen) {
-            return ARGON2_PWD_TOO_SHORT;
-        }
-
-        if (ARGON2_MAX_PWD_LENGTH < context->pwdlen) {
-            return ARGON2_PWD_TOO_LONG;
-        }
     }
 
-    /* Validate salt length */
+    if (ARGON2_MIN_PWD_LENGTH > context->pwdlen) {
+      return ARGON2_PWD_TOO_SHORT;
+    }
+
+    if (ARGON2_MAX_PWD_LENGTH < context->pwdlen) {
+        return ARGON2_PWD_TOO_LONG;
+    }
+
+    /* Validate salt (required param) */
     if (NULL == context->salt) {
         if (0 != context->saltlen) {
             return ARGON2_SALT_PTR_MISMATCH;
         }
-    } else {
-        if (ARGON2_MIN_SALT_LENGTH > context->saltlen) {
-            return ARGON2_SALT_TOO_SHORT;
-        }
-
-        if (ARGON2_MAX_SALT_LENGTH < context->saltlen) {
-            return ARGON2_SALT_TOO_LONG;
-        }
     }
 
-    /* Validate secret length */
+    if (ARGON2_MIN_SALT_LENGTH > context->saltlen) {
+        return ARGON2_SALT_TOO_SHORT;
+    }
+
+    if (ARGON2_MAX_SALT_LENGTH < context->saltlen) {
+        return ARGON2_SALT_TOO_LONG;
+    }
+
+    /* Validate secret (optional param) */
     if (NULL == context->secret) {
         if (0 != context->secretlen) {
             return ARGON2_SECRET_PTR_MISMATCH;
@@ -393,13 +429,12 @@ int validate_inputs(const argon2_context *context) {
         if (ARGON2_MIN_SECRET > context->secretlen) {
             return ARGON2_SECRET_TOO_SHORT;
         }
-
         if (ARGON2_MAX_SECRET < context->secretlen) {
             return ARGON2_SECRET_TOO_LONG;
         }
     }
 
-    /* Validate associated data */
+    /* Validate associated data (optional param) */
     if (NULL == context->ad) {
         if (0 != context->adlen) {
             return ARGON2_AD_PTR_MISMATCH;
@@ -408,7 +443,6 @@ int validate_inputs(const argon2_context *context) {
         if (ARGON2_MIN_AD_LENGTH > context->adlen) {
             return ARGON2_AD_TOO_SHORT;
         }
-
         if (ARGON2_MAX_AD_LENGTH < context->adlen) {
             return ARGON2_AD_TOO_LONG;
         }
@@ -467,8 +501,8 @@ int validate_inputs(const argon2_context *context) {
 
 void fill_first_blocks(uint8_t *blockhash, const argon2_instance_t *instance) {
     uint32_t l;
-    /* Make the first and second block in each lane as G(H0||i||0) or
-       G(H0||i||1) */
+    /* Make the first and second block in each lane as G(H0||0||i) or
+       G(H0||1||i) */
     uint8_t blockhash_bytes[ARGON2_BLOCK_SIZE];
     for (l = 0; l < instance->lanes; ++l) {
 
@@ -485,7 +519,7 @@ void fill_first_blocks(uint8_t *blockhash, const argon2_instance_t *instance) {
         load_block(&instance->memory[l * instance->lane_length + 1],
                    blockhash_bytes);
     }
-    secure_wipe_memory(blockhash_bytes, ARGON2_BLOCK_SIZE);
+    clear_internal_memory(blockhash_bytes, ARGON2_BLOCK_SIZE);
 }
 
 void initial_hash(uint8_t *blockhash, argon2_context *context,
@@ -568,22 +602,13 @@ int initialize(argon2_instance_t *instance, argon2_context *context) {
 
     if (instance == NULL || context == NULL)
         return ARGON2_INCORRECT_PARAMETER;
+    instance->context_ptr = context;
 
     /* 1. Memory allocation */
-
-    if (NULL != context->allocate_cbk) {
-        uint8_t *p;
-        result = context->allocate_cbk(&p, instance->memory_blocks *
-                                               ARGON2_BLOCK_SIZE);
-        if (ARGON2_OK != result) {
-            return result;
-        }
-        instance->memory = (block *)p;
-    } else {
-        result = allocate_memory(&(instance->memory), instance->memory_blocks);
-        if (ARGON2_OK != result) {
-            return result;
-        }
+    result = allocate_memory(context, (uint8_t **)&(instance->memory),
+                             instance->memory_blocks, sizeof(block));
+    if (result != ARGON2_OK) {
+        return result;
     }
 
     /* 2. Initial hashing */
@@ -592,9 +617,9 @@ int initialize(argon2_instance_t *instance, argon2_context *context) {
     /* Hashing all inputs */
     initial_hash(blockhash, context, instance->type);
     /* Zeroing 8 extra bytes */
-    secure_wipe_memory(blockhash + ARGON2_PREHASH_DIGEST_LENGTH,
-                       ARGON2_PREHASH_SEED_LENGTH -
-                           ARGON2_PREHASH_DIGEST_LENGTH);
+    clear_internal_memory(blockhash + ARGON2_PREHASH_DIGEST_LENGTH,
+                          ARGON2_PREHASH_SEED_LENGTH -
+                              ARGON2_PREHASH_DIGEST_LENGTH);
 
 #ifdef GENKAT
     initial_kat(blockhash, context, instance->type);
@@ -604,7 +629,7 @@ int initialize(argon2_instance_t *instance, argon2_context *context) {
      */
     fill_first_blocks(blockhash, instance);
     /* Clearing the hash */
-    secure_wipe_memory(blockhash, ARGON2_PREHASH_SEED_LENGTH);
+    clear_internal_memory(blockhash, ARGON2_PREHASH_SEED_LENGTH);
 
     return ARGON2_OK;
 }
diff --git a/src/argon2/core.h b/argon2/core.h
similarity index 80%
rename from src/argon2/core.h
rename to argon2/core.h
index f5c2d437..a73a01f3 100644
--- a/src/argon2/core.h
+++ b/argon2/core.h
@@ -30,8 +30,7 @@
 
 #define CONST_CAST(x) (x)(uintptr_t)
 
-/*************************Argon2 internal
- * constants**************************************************/
+/**********************Argon2 internal constants*******************************/
 
 enum argon2_core_constants {
     /* Memory block size in bytes */
@@ -49,8 +48,7 @@ enum argon2_core_constants {
     ARGON2_PREHASH_SEED_LENGTH = 72
 };
 
-/*************************Argon2 internal data
- * types**************************************************/
+/*************************Argon2 internal data types***********************/
 
 /*
  * Structure for the (1KB) memory block implemented as 128 64-bit words.
@@ -87,6 +85,7 @@ typedef struct Argon2_instance_t {
     uint32_t threads;
     argon2_type type;
     int print_internals; /* whether to print the memory blocks */
+    argon2_context *context_ptr; /* points back to original context */
 } argon2_instance_t;
 
 /*
@@ -106,32 +105,43 @@ typedef struct Argon2_thread_data {
     argon2_position_t pos;
 } argon2_thread_data;
 
-/*************************Argon2 core
- * functions**************************************************/
+/*************************Argon2 core functions********************************/
 
-/* Allocates memory to the given pointer
+/* Allocates memory to the given pointer, uses the appropriate allocator as
+ * specified in the context. Total allocated memory is num*size.
+ * @param context argon2_context which specifies the allocator
  * @param memory pointer to the pointer to the memory
- * @param m_cost number of blocks to allocate in the memory
+ * @param size the size in bytes for each element to be allocated
+ * @param num the number of elements to be allocated
  * @return ARGON2_OK if @memory is a valid pointer and memory is allocated
  */
-int allocate_memory(block **memory, uint32_t m_cost);
+int allocate_memory(const argon2_context *context, uint8_t **memory,
+                    size_t num, size_t size);
 
-/* Function that securely cleans the memory
+/*
+ * Frees memory at the given pointer, uses the appropriate deallocator as
+ * specified in the context. Also cleans the memory using clear_internal_memory.
+ * @param context argon2_context which specifies the deallocator
+ * @param memory pointer to buffer to be freed
+ * @param size the size in bytes for each element to be deallocated
+ * @param num the number of elements to be deallocated
+ */
+void free_memory(const argon2_context *context, uint8_t *memory,
+                 size_t num, size_t size);
+
+/* Function that securely cleans the memory. This ignores any flags set
+ * regarding clearing memory. Usually one just calls clear_internal_memory.
  * @param mem Pointer to the memory
  * @param s Memory size in bytes
  */
 void secure_wipe_memory(void *v, size_t n);
 
-/* Clears memory
- * @param instance pointer to the current instance
- * @param clear_memory indicates if we clear the memory with zeros.
+/* Function that securely clears the memory if FLAG_clear_internal_memory is
+ * set. If the flag isn't set, this function does nothing.
+ * @param mem Pointer to the memory
+ * @param s Memory size in bytes
  */
-void clear_memory(argon2_instance_t *instance, int clear);
-
-/* Deallocates memory
- * @param memory pointer to the blocks
- */
-void free_memory(block *memory);
+void clear_internal_memory(void *v, size_t n);
 
 /*
  * Computes absolute position of reference block in the lane following a skewed
@@ -205,6 +215,7 @@ void finalize(const argon2_context *context, argon2_instance_t *instance);
 /*
  * Function that fills the segment using previous segments also from other
  * threads
+ * @param context current context
  * @param instance Pointer to the current instance
  * @param position Current position
  * @pre all block pointers must be valid
diff --git a/src/argon2/encoding.c b/argon2/encoding.c
similarity index 89%
rename from src/argon2/encoding.c
rename to argon2/encoding.c
index 2fc8d547..4fd965b2 100644
--- a/src/argon2/encoding.c
+++ b/argon2/encoding.c
@@ -39,11 +39,6 @@
  *   the parameters, salts and outputs. It does not compute the hash
  *   itself.
  *
- *   -- The third section is test code, with a main() function. With
- *   this section, the whole file compiles as a stand-alone program
- *   that exercises the encoding and decoding functions with some
- *   test vectors.
- *
  * The code was originally written by Thomas Pornin <pornin@bolet.org>,
  * to whom comments and remarks may be sent. It is released under what
  * should amount to Public Domain or its closest equivalent; the
@@ -246,19 +241,18 @@ static const char *decode_decimal(const char *str, unsigned long *v) {
  *
  * The code below applies the following format:
  *
- *  $argon2<T>[$v=<num>]$m=<num>,t=<num>,p=<num>[,keyid=<bin>][,data=<bin>][$<bin>[$<bin>]]
+ *  $argon2<T>[$v=<num>]$m=<num>,t=<num>,p=<num>$<bin>$<bin>
  *
- * where <T> is either 'd' or 'i', <num> is a decimal integer (positive, fits in
- * an 'unsigned long'), and <bin> is Base64-encoded data (no '=' padding
+ * where <T> is either 'd', 'id', or 'i', <num> is a decimal integer (positive,
+ * fits in an 'unsigned long'), and <bin> is Base64-encoded data (no '=' padding
  * characters, no newline or whitespace).
- * The "keyid" is a binary identifier for a key (up to 8 bytes);
- * "data" is associated data (up to 32 bytes). When the 'keyid'
- * (resp. the 'data') is empty, then it is ommitted from the output.
  *
  * The last two binary chunks (encoded in Base64) are, in that order,
- * the salt and the output. Both are optional, but you cannot have an
- * output without a salt. The binary salt length is between 8 and 48 bytes.
- * The output length is always exactly 32 bytes.
+ * the salt and the output. Both are required. The binary salt length and the
+ * output length must be in the allowed ranges defined in argon2.h.
+ *
+ * The ctx struct must contain buffers large enough to hold the salt and pwd
+ * when it is fed into decode_string.
  */
 
 int decode_string(argon2_context *ctx, const char *str, argon2_type type) {
@@ -273,7 +267,7 @@ int decode_string(argon2_context *ctx, const char *str, argon2_type type) {
         str += cc_len;                                                         \
     } while ((void)0, 0)
 
-/* prefix checking with supplied code */
+/* optional prefix checking with supplied code */
 #define CC_opt(prefix, code)                                                   \
     do {                                                                       \
         size_t cc_len = strlen(prefix);                                        \
@@ -283,7 +277,7 @@ int decode_string(argon2_context *ctx, const char *str, argon2_type type) {
         }                                                                      \
     } while ((void)0, 0)
 
-/* Decoding  prefix into decimal */
+/* Decoding prefix into decimal */
 #define DECIMAL(x)                                                             \
     do {                                                                       \
         unsigned long dec_x;                                                   \
@@ -294,6 +288,7 @@ int decode_string(argon2_context *ctx, const char *str, argon2_type type) {
         (x) = dec_x;                                                           \
     } while ((void)0, 0)
 
+/* Decoding base64 into a binary buffer */
 #define BIN(buf, max_len, len)                                                 \
     do {                                                                       \
         size_t bin_len = (max_len);                                            \
@@ -304,26 +299,20 @@ int decode_string(argon2_context *ctx, const char *str, argon2_type type) {
         (len) = (uint32_t)bin_len;                                             \
     } while ((void)0, 0)
 
-    size_t maxadlen = ctx->adlen;
     size_t maxsaltlen = ctx->saltlen;
     size_t maxoutlen = ctx->outlen;
     int validation_result;
     const char* type_string;
 
-    ctx->adlen = 0;
-    ctx->saltlen = 0;
-    ctx->outlen = 0;
-    ctx->pwdlen = 0;
-
     /* We should start with the argon2_type we are using */
-    CC("$");
     type_string = argon2_type2string(type, 0);
-    if (type_string) {
-        CC(type_string);
-    } else {
+    if (!type_string) {
         return ARGON2_INCORRECT_TYPE;
     }
 
+    CC("$");
+    CC(type_string);
+
     /* Reading the version number if the default is suppressed */
     ctx->version = ARGON2_VERSION_10;
     CC_opt("$v=", DECIMAL(ctx->version));
@@ -336,21 +325,27 @@ int decode_string(argon2_context *ctx, const char *str, argon2_type type) {
     DECIMAL(ctx->lanes);
     ctx->threads = ctx->lanes;
 
-    CC_opt(",data=", BIN(ctx->ad, maxadlen, ctx->adlen));
-    if (*str == 0) {
-        return ARGON2_OK;
-    }
     CC("$");
     BIN(ctx->salt, maxsaltlen, ctx->saltlen);
-    if (*str == 0) {
-        return ARGON2_OK;
-    }
     CC("$");
     BIN(ctx->out, maxoutlen, ctx->outlen);
+
+    /* The rest of the fields get the default values */
+    ctx->secret = NULL;
+    ctx->secretlen = 0;
+    ctx->ad = NULL;
+    ctx->adlen = 0;
+    ctx->allocate_cbk = NULL;
+    ctx->free_cbk = NULL;
+    ctx->flags = ARGON2_DEFAULT_FLAGS;
+
+    /* On return, must have valid context */
     validation_result = validate_inputs(ctx);
     if (validation_result != ARGON2_OK) {
         return validation_result;
     }
+
+    /* Can't have any additional characters */
     if (*str == 0) {
         return ARGON2_OK;
     } else {
@@ -393,19 +388,23 @@ int encode_string(char *dst, size_t dst_len, argon2_context *ctx,
     } while ((void)0, 0)
 
     const char* type_string = argon2_type2string(type, 0);
-    SS("$");
-    if (type_string) {
-        SS(type_string);
-    } else {
-        return ARGON2_ENCODING_FAIL;
+    int validation_result = validate_inputs(ctx);
+
+    if (!type_string) {
+      return ARGON2_ENCODING_FAIL;
     }
 
-    if (validate_inputs(ctx) != ARGON2_OK) {
-        return validate_inputs(ctx);
+    if (validation_result != ARGON2_OK) {
+      return validation_result;
     }
 
+
+    SS("$");
+    SS(type_string);
+
     SS("$v=");
     SX(ctx->version);
+
     SS("$m=");
     SX(ctx->m_cost);
     SS(",t=");
@@ -413,20 +412,9 @@ int encode_string(char *dst, size_t dst_len, argon2_context *ctx,
     SS(",p=");
     SX(ctx->lanes);
 
-    if (ctx->adlen > 0) {
-        SS(",data=");
-        SB(ctx->ad, ctx->adlen);
-    }
-
-    if (ctx->saltlen == 0)
-        return ARGON2_OK;
-
     SS("$");
     SB(ctx->salt, ctx->saltlen);
 
-    if (ctx->outlen == 0)
-        return ARGON2_OK;
-
     SS("$");
     SB(ctx->out, ctx->outlen);
     return ARGON2_OK;
diff --git a/src/argon2/encoding.h b/argon2/encoding.h
similarity index 80%
rename from src/argon2/encoding.h
rename to argon2/encoding.h
index f652dfa0..1c1ec8b9 100644
--- a/src/argon2/encoding.h
+++ b/argon2/encoding.h
@@ -29,21 +29,21 @@
 * is less than the number of required characters (including the
 * terminating 0), then this function returns ARGON2_ENCODING_ERROR.
 *
-* if ctx->outlen is 0, then the hash string will be a salt string
-* (no output). if ctx->saltlen is also 0, then the string will be a
-* parameter-only string (no salt and no output).
-*
 * on success, ARGON2_OK is returned.
-*
-* No other parameters are checked
 */
 int encode_string(char *dst, size_t dst_len, argon2_context *ctx,
                   argon2_type type);
 
 /*
 * Decodes an Argon2 hash string into the provided structure 'ctx'.
-* The fields ctx.saltlen, ctx.adlen, ctx.outlen set the maximal salt, ad, out
-* length values that are allowed; invalid input string causes an error.
+* The only fields that must be set prior to this call are ctx.saltlen and
+* ctx.outlen (which must be the maximal salt and out length values that are
+* allowed), ctx.salt and ctx.out (which must be buffers of the specified
+* length), and ctx.pwd and ctx.pwdlen which must hold a valid password.
+*
+* Invalid input string causes an error. On success, the ctx is valid and all
+* fields have been initialized.
+*
 * Returned value is ARGON2_OK on success, other ARGON2_ codes on error.
 */
 int decode_string(argon2_context *ctx, const char *str, argon2_type type);
diff --git a/src/argon2/opt.c b/argon2/opt.c
similarity index 54%
rename from src/argon2/opt.c
rename to argon2/opt.c
index ab2cb6a2..2315a087 100644
--- a/src/argon2/opt.c
+++ b/argon2/opt.c
@@ -20,18 +20,37 @@
 #include <stdlib.h>
 
 #include "argon2.h"
-#include "opt.h"
+#include "core.h"
 
 #include "blake2/blake2.h"
 #include "blake2/blamka-round-opt.h"
 
-void fill_block(__m128i *state, const uint8_t *ref_block, uint8_t *next_block) {
+/*
+ * Function fills a new memory block and optionally XORs the old block over the new one.
+ * Memory must be initialized.
+ * @param state Pointer to the just produced block. Content will be updated(!)
+ * @param ref_block Pointer to the reference block
+ * @param next_block Pointer to the block to be XORed over. May coincide with @ref_block
+ * @param with_xor Whether to XOR into the new block (1) or just overwrite (0)
+ * @pre all block pointers must be valid
+ */
+static void fill_block(__m128i *state, const block *ref_block,
+                       block *next_block, int with_xor) {
     __m128i block_XY[ARGON2_OWORDS_IN_BLOCK];
-    uint32_t i;
+    unsigned int i;
 
-    for (i = 0; i < ARGON2_OWORDS_IN_BLOCK; i++) {
-        block_XY[i] = state[i] = _mm_xor_si128(
-            state[i], _mm_loadu_si128((__m128i const *)(&ref_block[16 * i])));
+    if (with_xor) {
+        for (i = 0; i < ARGON2_OWORDS_IN_BLOCK; i++) {
+            state[i] = _mm_xor_si128(
+                state[i], _mm_loadu_si128((const __m128i *)ref_block->v + i));
+            block_XY[i] = _mm_xor_si128(
+                state[i], _mm_loadu_si128((const __m128i *)next_block->v + i));
+        }
+    } else {
+        for (i = 0; i < ARGON2_OWORDS_IN_BLOCK; i++) {
+            block_XY[i] = state[i] = _mm_xor_si128(
+                state[i], _mm_loadu_si128((const __m128i *)ref_block->v + i));
+        }
     }
 
     for (i = 0; i < 8; ++i) {
@@ -48,93 +67,38 @@ void fill_block(__m128i *state, const uint8_t *ref_block, uint8_t *next_block) {
 
     for (i = 0; i < ARGON2_OWORDS_IN_BLOCK; i++) {
         state[i] = _mm_xor_si128(state[i], block_XY[i]);
-        _mm_storeu_si128((__m128i *)(&next_block[16 * i]), state[i]);
+        _mm_storeu_si128((__m128i *)next_block->v + i, state[i]);
     }
 }
 
-void fill_block_with_xor(__m128i *state, const uint8_t *ref_block,
-                         uint8_t *next_block) {
-    __m128i block_XY[ARGON2_OWORDS_IN_BLOCK];
-    uint32_t i;
+static void next_addresses(block *address_block, block *input_block) {
+    /*Temporary zero-initialized blocks*/
+    __m128i zero_block[ARGON2_OWORDS_IN_BLOCK];
+    __m128i zero2_block[ARGON2_OWORDS_IN_BLOCK];
 
-    for (i = 0; i < ARGON2_OWORDS_IN_BLOCK; i++) { 
-       state[i] = _mm_xor_si128(
-            state[i], _mm_loadu_si128((__m128i const *)(&ref_block[16 * i])));
-        block_XY[i] =  _mm_xor_si128(
-            state[i], _mm_loadu_si128((__m128i const *)(&next_block[16 * i])));
-    }
+    memset(zero_block, 0, sizeof(zero_block));
+    memset(zero2_block, 0, sizeof(zero2_block));
 
-    for (i = 0; i < 8; ++i) {
-        BLAKE2_ROUND(state[8 * i + 0], state[8 * i + 1], state[8 * i + 2],
-                     state[8 * i + 3], state[8 * i + 4], state[8 * i + 5],
-                     state[8 * i + 6], state[8 * i + 7]);
-    }
+    /*Increasing index counter*/
+    input_block->v[6]++;
 
-    for (i = 0; i < 8; ++i) {
-        BLAKE2_ROUND(state[8 * 0 + i], state[8 * 1 + i], state[8 * 2 + i],
-                     state[8 * 3 + i], state[8 * 4 + i], state[8 * 5 + i],
-                     state[8 * 6 + i], state[8 * 7 + i]);
-    }
+    /*First iteration of G*/
+    fill_block(zero_block, input_block, address_block, 0);
 
-    for (i = 0; i < ARGON2_OWORDS_IN_BLOCK; i++) {
-        state[i] = _mm_xor_si128(state[i], block_XY[i]);
-        _mm_storeu_si128((__m128i *)(&next_block[16 * i]), state[i]);
-    }
-}
-
-void generate_addresses(const argon2_instance_t *instance,
-                        const argon2_position_t *position,
-                        uint64_t *pseudo_rands) {
-    block address_block, input_block, tmp_block;
-    uint32_t i;
-
-    init_block_value(&address_block, 0);
-    init_block_value(&input_block, 0);
-
-    if (instance != NULL && position != NULL) {
-        input_block.v[0] = position->pass;
-        input_block.v[1] = position->lane;
-        input_block.v[2] = position->slice;
-        input_block.v[3] = instance->memory_blocks;
-        input_block.v[4] = instance->passes;
-        input_block.v[5] = instance->type;
-
-        for (i = 0; i < instance->segment_length; ++i) {
-            if (i % ARGON2_ADDRESSES_IN_BLOCK == 0) {
-                /*Temporary zero-initialized blocks*/
-                __m128i zero_block[ARGON2_OWORDS_IN_BLOCK];
-                __m128i zero2_block[ARGON2_OWORDS_IN_BLOCK];
-                memset(zero_block, 0, sizeof(zero_block));
-                memset(zero2_block, 0, sizeof(zero2_block));
-                init_block_value(&address_block, 0);
-                init_block_value(&tmp_block, 0);
-                /*Increasing index counter*/
-                input_block.v[6]++;
-                /*First iteration of G*/
-                fill_block_with_xor(zero_block, (uint8_t *)&input_block.v,
-                           (uint8_t *)&tmp_block.v);
-                /*Second iteration of G*/
-                fill_block_with_xor(zero2_block, (uint8_t *)&tmp_block.v,
-                           (uint8_t *)&address_block.v);
-            }
-
-            pseudo_rands[i] = address_block.v[i % ARGON2_ADDRESSES_IN_BLOCK];
-        }
-    }
+    /*Second iteration of G*/
+    fill_block(zero2_block, address_block, address_block, 0);
 }
 
 void fill_segment(const argon2_instance_t *instance,
                   argon2_position_t position) {
     block *ref_block = NULL, *curr_block = NULL;
+    block address_block, input_block;
     uint64_t pseudo_rand, ref_index, ref_lane;
     uint32_t prev_offset, curr_offset;
     uint32_t starting_index, i;
     __m128i state[64];
     int data_independent_addressing;
 
-    /* Pseudo-random values that determine the reference block position */
-    uint64_t *pseudo_rands = NULL;
-
     if (instance == NULL) {
         return;
     }
@@ -144,20 +108,26 @@ void fill_segment(const argon2_instance_t *instance,
         (instance->type == Argon2_id && (position.pass == 0) &&
          (position.slice < ARGON2_SYNC_POINTS / 2));
 
-    pseudo_rands =
-        (uint64_t *)malloc(sizeof(uint64_t) * instance->segment_length);
-    if (pseudo_rands == NULL) {
-        return;
-    }
-
     if (data_independent_addressing) {
-        generate_addresses(instance, &position, pseudo_rands);
+        init_block_value(&input_block, 0);
+
+        input_block.v[0] = position.pass;
+        input_block.v[1] = position.lane;
+        input_block.v[2] = position.slice;
+        input_block.v[3] = instance->memory_blocks;
+        input_block.v[4] = instance->passes;
+        input_block.v[5] = instance->type;
     }
 
     starting_index = 0;
 
     if ((0 == position.pass) && (0 == position.slice)) {
         starting_index = 2; /* we have already generated the first two blocks */
+
+        /* Don't forget to generate the first block of addresses: */
+        if (data_independent_addressing) {
+            next_addresses(&address_block, &input_block);
+        }
     }
 
     /* Offset of the current block */
@@ -184,7 +154,10 @@ void fill_segment(const argon2_instance_t *instance,
         /* 1.2 Computing the index of the reference block */
         /* 1.2.1 Taking pseudo-random value from the previous block */
         if (data_independent_addressing) {
-            pseudo_rand = pseudo_rands[i];
+            if (i % ARGON2_ADDRESSES_IN_BLOCK == 0) {
+                next_addresses(&address_block, &input_block);
+            }
+            pseudo_rand = address_block.v[i % ARGON2_ADDRESSES_IN_BLOCK];
         } else {
             pseudo_rand = instance->memory[prev_offset].v[0];
         }
@@ -210,18 +183,13 @@ void fill_segment(const argon2_instance_t *instance,
         curr_block = instance->memory + curr_offset;
         if (ARGON2_VERSION_10 == instance->version) {
             /* version 1.2.1 and earlier: overwrite, not XOR */
-            fill_block(state, (uint8_t *)ref_block->v,
-                       (uint8_t *)curr_block->v);
+            fill_block(state, ref_block, curr_block, 0);
         } else {
             if(0 == position.pass) {
-                fill_block(state, (uint8_t *)ref_block->v,
-                           (uint8_t *)curr_block->v);
+                fill_block(state, ref_block, curr_block, 0);
             } else {
-                fill_block_with_xor(state, (uint8_t *)ref_block->v,
-                                    (uint8_t *)curr_block->v);
+                fill_block(state, ref_block, curr_block, 1);
             }
         }
     }
-
-    free(pseudo_rands);
 }
diff --git a/src/argon2/opt.h b/argon2/opt.h
similarity index 100%
rename from src/argon2/opt.h
rename to argon2/opt.h
diff --git a/src/argon2/ref.c b/argon2/ref.c
similarity index 60%
rename from src/argon2/ref.c
rename to argon2/ref.c
index 0b086c38..4dbd1e5a 100644
--- a/src/argon2/ref.c
+++ b/argon2/ref.c
@@ -20,61 +20,38 @@
 #include <stdlib.h>
 
 #include "argon2.h"
-#include "ref.h"
+#include "core.h"
 
 #include "blake2/blamka-round-ref.h"
 #include "blake2/blake2-impl.h"
 #include "blake2/blake2.h"
 
 
-void fill_block(const block *prev_block, const block *ref_block,
-    block *next_block) {
+/*
+ * Function fills a new memory block and optionally XORs the old block over the new one.
+ * @next_block must be initialized.
+ * @param prev_block Pointer to the previous block
+ * @param ref_block Pointer to the reference block
+ * @param next_block Pointer to the block to be constructed
+ * @param with_xor Whether to XOR into the new block (1) or just overwrite (0)
+ * @pre all block pointers must be valid
+ */
+static void fill_block(const block *prev_block, const block *ref_block,
+                       block *next_block, int with_xor) {
     block blockR, block_tmp;
     unsigned i;
 
     copy_block(&blockR, ref_block);
     xor_block(&blockR, prev_block);
     copy_block(&block_tmp, &blockR);
-            /*Now blockR = ref_block + prev_block and bloc_tmp = ref_block + prev_block */
-                /* Apply Blake2 on columns of 64-bit words: (0,1,...,15) , then
-                (16,17,..31)... finally (112,113,...127) */
-    for (i = 0; i < 8; ++i) {
-        BLAKE2_ROUND_NOMSG(
-            blockR.v[16 * i], blockR.v[16 * i + 1], blockR.v[16 * i + 2],
-            blockR.v[16 * i + 3], blockR.v[16 * i + 4], blockR.v[16 * i + 5],
-            blockR.v[16 * i + 6], blockR.v[16 * i + 7], blockR.v[16 * i + 8],
-            blockR.v[16 * i + 9], blockR.v[16 * i + 10], blockR.v[16 * i + 11],
-            blockR.v[16 * i + 12], blockR.v[16 * i + 13], blockR.v[16 * i + 14],
-            blockR.v[16 * i + 15]);
+    /* Now blockR = ref_block + prev_block and block_tmp = ref_block + prev_block */
+    if (with_xor) {
+        /* Saving the next block contents for XOR over: */
+        xor_block(&block_tmp, next_block);
+        /* Now blockR = ref_block + prev_block and
+           block_tmp = ref_block + prev_block + next_block */
     }
 
-    /* Apply Blake2 on rows of 64-bit words: (0,1,16,17,...112,113), then
-    (2,3,18,19,...,114,115).. finally (14,15,30,31,...,126,127) */
-    for (i = 0; i < 8; i++) {
-        BLAKE2_ROUND_NOMSG(
-            blockR.v[2 * i], blockR.v[2 * i + 1], blockR.v[2 * i + 16],
-            blockR.v[2 * i + 17], blockR.v[2 * i + 32], blockR.v[2 * i + 33],
-            blockR.v[2 * i + 48], blockR.v[2 * i + 49], blockR.v[2 * i + 64],
-            blockR.v[2 * i + 65], blockR.v[2 * i + 80], blockR.v[2 * i + 81],
-            blockR.v[2 * i + 96], blockR.v[2 * i + 97], blockR.v[2 * i + 112],
-            blockR.v[2 * i + 113]);
-    }
-
-    copy_block(next_block, &block_tmp);
-    xor_block(next_block, &blockR);
-}
-
-
-void fill_block_with_xor(const block *prev_block, const block *ref_block,
-                block *next_block) {
-    block blockR, block_tmp;
-    unsigned i;
-
-    copy_block(&blockR, ref_block);
-    xor_block(&blockR, prev_block);
-    copy_block(&block_tmp, &blockR);
-    xor_block(&block_tmp, next_block); /*Saving the next block contents for XOR over*/
-    /*Now blockR = ref_block + prev_block and bloc_tmp = ref_block + prev_block + next_block*/
     /* Apply Blake2 on columns of 64-bit words: (0,1,...,15) , then
        (16,17,..31)... finally (112,113,...127) */
     for (i = 0; i < 8; ++i) {
@@ -103,47 +80,22 @@ void fill_block_with_xor(const block *prev_block, const block *ref_block,
     xor_block(next_block, &blockR);
 }
 
-void generate_addresses(const argon2_instance_t *instance,
-                        const argon2_position_t *position,
-                        uint64_t *pseudo_rands) {
-    block zero_block, input_block, address_block,tmp_block;
-    uint32_t i;
-
-    init_block_value(&zero_block, 0);
-    init_block_value(&input_block, 0);
-
-    if (instance != NULL && position != NULL) {
-        input_block.v[0] = position->pass;
-        input_block.v[1] = position->lane;
-        input_block.v[2] = position->slice;
-        input_block.v[3] = instance->memory_blocks;
-        input_block.v[4] = instance->passes;
-        input_block.v[5] = instance->type;
-
-        for (i = 0; i < instance->segment_length; ++i) {
-            if (i % ARGON2_ADDRESSES_IN_BLOCK == 0) {
-                input_block.v[6]++;
-                init_block_value(&tmp_block, 0);
-                init_block_value(&address_block, 0);
-                fill_block_with_xor(&zero_block, &input_block, &tmp_block);
-                fill_block_with_xor(&zero_block, &tmp_block, &address_block);
-            }
-
-            pseudo_rands[i] = address_block.v[i % ARGON2_ADDRESSES_IN_BLOCK];
-        }
-    }
+static void next_addresses(block *address_block, block *input_block,
+                           const block *zero_block) {
+    input_block->v[6]++;
+    fill_block(zero_block, input_block, address_block, 0);
+    fill_block(zero_block, address_block, address_block, 0);
 }
 
 void fill_segment(const argon2_instance_t *instance,
                   argon2_position_t position) {
     block *ref_block = NULL, *curr_block = NULL;
+    block address_block, input_block, zero_block;
     uint64_t pseudo_rand, ref_index, ref_lane;
     uint32_t prev_offset, curr_offset;
     uint32_t starting_index;
     uint32_t i;
     int data_independent_addressing;
-    /* Pseudo-random values that determine the reference block position */
-    uint64_t *pseudo_rands = NULL;
 
     if (instance == NULL) {
         return;
@@ -154,21 +106,27 @@ void fill_segment(const argon2_instance_t *instance,
         (instance->type == Argon2_id && (position.pass == 0) &&
          (position.slice < ARGON2_SYNC_POINTS / 2));
 
-    pseudo_rands =
-        (uint64_t *)malloc(sizeof(uint64_t) * (instance->segment_length));
-
-    if (pseudo_rands == NULL) {
-        return;
-    }
-
     if (data_independent_addressing) {
-        generate_addresses(instance, &position, pseudo_rands);
+        init_block_value(&zero_block, 0);
+        init_block_value(&input_block, 0);
+
+        input_block.v[0] = position.pass;
+        input_block.v[1] = position.lane;
+        input_block.v[2] = position.slice;
+        input_block.v[3] = instance->memory_blocks;
+        input_block.v[4] = instance->passes;
+        input_block.v[5] = instance->type;
     }
 
     starting_index = 0;
 
     if ((0 == position.pass) && (0 == position.slice)) {
         starting_index = 2; /* we have already generated the first two blocks */
+
+        /* Don't forget to generate the first block of addresses: */
+        if (data_independent_addressing) {
+            next_addresses(&address_block, &input_block, &zero_block);
+        }
     }
 
     /* Offset of the current block */
@@ -193,7 +151,10 @@ void fill_segment(const argon2_instance_t *instance,
         /* 1.2 Computing the index of the reference block */
         /* 1.2.1 Taking pseudo-random value from the previous block */
         if (data_independent_addressing) {
-            pseudo_rand = pseudo_rands[i];
+            if (i % ARGON2_ADDRESSES_IN_BLOCK == 0) {
+                next_addresses(&address_block, &input_block, &zero_block);
+            }
+            pseudo_rand = address_block.v[i % ARGON2_ADDRESSES_IN_BLOCK];
         } else {
             pseudo_rand = instance->memory[prev_offset].v[0];
         }
@@ -219,17 +180,15 @@ void fill_segment(const argon2_instance_t *instance,
         curr_block = instance->memory + curr_offset;
         if (ARGON2_VERSION_10 == instance->version) {
             /* version 1.2.1 and earlier: overwrite, not XOR */
-            fill_block(instance->memory + prev_offset, ref_block, curr_block);
+            fill_block(instance->memory + prev_offset, ref_block, curr_block, 0);
         } else {
             if(0 == position.pass) {
                 fill_block(instance->memory + prev_offset, ref_block,
-                           curr_block);
+                           curr_block, 0);
             } else {
-                fill_block_with_xor(instance->memory + prev_offset, ref_block,
-                                    curr_block);
+                fill_block(instance->memory + prev_offset, ref_block,
+                           curr_block, 1);
             }
         }
     }
-
-    free(pseudo_rands);
 }
diff --git a/src/argon2/ref.h b/argon2/ref.h
similarity index 100%
rename from src/argon2/ref.h
rename to argon2/ref.h
diff --git a/src/argon2/thread.c b/argon2/thread.c
similarity index 95%
rename from src/argon2/thread.c
rename to argon2/thread.c
index 5de2c133..13a9b426 100644
--- a/src/argon2/thread.c
+++ b/argon2/thread.c
@@ -15,6 +15,8 @@
  * software. If not, they may be obtained at the above URLs.
  */
 
+#if !defined(ARGON2_NO_THREADS)
+
 #include "thread.h"
 #if defined(_WIN32)
 #include <windows.h>
@@ -51,3 +53,5 @@ void argon2_thread_exit(void) {
     pthread_exit(NULL);
 #endif
 }
+
+#endif /* ARGON2_NO_THREADS */
diff --git a/src/argon2/thread.h b/argon2/thread.h
similarity index 97%
rename from src/argon2/thread.h
rename to argon2/thread.h
index f04bb407..ecbd41dc 100644
--- a/src/argon2/thread.h
+++ b/argon2/thread.h
@@ -17,6 +17,9 @@
 
 #ifndef ARGON2_THREAD_H
 #define ARGON2_THREAD_H
+
+#if !defined(ARGON2_NO_THREADS)
+
 /*
         Here we implement an abstraction layer for the simpĺe requirements
         of the Argon2 code. We only require 3 primitives---thread creation,
@@ -60,4 +63,5 @@ int argon2_thread_join(argon2_thread_handle_t handle);
 */
 void argon2_thread_exit(void);
 
+#endif /* ARGON2_NO_THREADS */
 #endif
diff --git a/configure.ac b/configure.ac
index ddef5aa6..6f4fb127 100644
--- a/configure.ac
+++ b/configure.ac
@@ -99,6 +99,24 @@ PKG_CHECK_MODULES([GnuTLS], [gnutls >= 3.3])
 PKG_CHECK_MODULES([MsgPack], [msgpack >= 1.2])
 CXXFLAGS="${CXXFLAGS} -DMSGPACK_DISABLE_LEGACY_NIL -DMSGPACK_DISABLE_LEGACY_CONVERT"
 
+dnl Check for Argon2 				
+AC_ARG_WITH([argon2], AS_HELP_STRING([--without-argon2], [Use included Argon2]))
+AS_IF([test "x$with_argon2" != "xno"],
+			[PKG_CHECK_MODULES([Argon2], [libargon2], [have_argon2=yes], [have_argon2=no])],
+			[have_argon2=no])
+AS_IF([test "x$have_argon2" = "xyes"], [], [
+		AS_IF([test "x$with_argon2" = "xyes"], [
+				AC_MSG_ERROR([Argon2 requested but not found])
+		],[
+				AC_MSG_NOTICE([Using included Argon2])
+				AC_CONFIG_FILES(argon2/Makefile)
+				AC_SUBST(Argon2_CFLAGS, "-I\${top_srcdir}/argon2")
+				AC_SUBST(Argon2_LIBS, "\${top_srcdir}/argon2/libargon2.la")
+		])
+])
+
+AM_CONDITIONAL([WITH_INCLUDED_ARGON2], [test "x$have_argon2" = "xno"])
+
 AC_ARG_ENABLE([tools], AS_HELP_STRING([--disable-tools],[Disable tools (CLI DHT node)]),,build_tools=yes)
 AM_CONDITIONAL(ENABLE_TOOLS, test x$build_tools == xyes)
 AM_COND_IF([ENABLE_TOOLS], [
@@ -120,7 +138,6 @@ AM_COND_IF([USE_CYTHON], [
 
 AC_CONFIG_FILES([Makefile
                  src/Makefile
-                 src/argon2/Makefile
                  tools/Makefile
                  opendht.pc])
 AC_OUTPUT
diff --git a/src/Makefile.am b/src/Makefile.am
index d0d77ed1..29545131 100644
--- a/src/Makefile.am
+++ b/src/Makefile.am
@@ -1,11 +1,7 @@
-SUBDIRS = argon2
-
 lib_LTLIBRARIES = libopendht.la
 
-AM_CPPFLAGS = -I../include/opendht
-libopendht_la_CXXFLAGS = @CXXFLAGS@
-libopendht_la_LDFLAGS = @LDFLAGS@ @GnuTLS_LIBS@ @Nettle_LIBS@
-libopendht_la_LIBADD = ./argon2/libargon2.la
+libopendht_la_CPPFLAGS = @CPPFLAGS@ -I$(top_srcdir)/include/opendht @Argon2_CFLAGS@
+libopendht_la_LIBADD = @Argon2_LIBS@ @GnuTLS_LIBS@ @Nettle_LIBS@
 
 libopendht_la_SOURCES = \
         dht.cpp \
diff --git a/src/crypto.cpp b/src/crypto.cpp
index 14337ad3..fee9a394 100644
--- a/src/crypto.cpp
+++ b/src/crypto.cpp
@@ -27,7 +27,7 @@ extern "C" {
 #include <nettle/gcm.h>
 #include <nettle/aes.h>
 
-#include "argon2/argon2.h"
+#include <argon2.h>
 }
 
 #include <random>
diff --git a/tools/CMakeLists.txt b/tools/CMakeLists.txt
index b4f21563..0539e3b0 100644
--- a/tools/CMakeLists.txt
+++ b/tools/CMakeLists.txt
@@ -3,9 +3,9 @@ add_executable (dhtnode dhtnode.cpp tools_common.h)
 add_executable (dhtscanner dhtscanner.cpp tools_common.h)
 add_executable (dhtchat dhtchat.cpp tools_common.h)
 
-target_link_libraries (dhtnode LINK_PUBLIC gnutls readline)
-target_link_libraries (dhtscanner LINK_PUBLIC gnutls readline)
-target_link_libraries (dhtchat LINK_PUBLIC gnutls readline)
+target_link_libraries (dhtnode LINK_PUBLIC readline)
+target_link_libraries (dhtscanner LINK_PUBLIC readline)
+target_link_libraries (dhtchat LINK_PUBLIC readline)
 
 if (OPENDHT_SHARED)
 	target_link_libraries (dhtnode LINK_PUBLIC opendht)