Adding upstream version 1.43.0.upstream/1.43.0

Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>

38 files changed, 11048 insertions, 760 deletions

diff --git a/libnetdata/Makefile.am b/libnetdata/Makefile.am
index e85f4abe..f01a1d34 100644
--- a/libnetdata/Makefile.am
+++ b/libnetdata/Makefile.am
@@ -15,6 +15,7 @@ SUBDIRS = \
     ebpf \
     eval \
     facets \
+    functions_evloop \
     json \
     july \
     health \
diff --git a/libnetdata/aral/aral.c b/libnetdata/aral/aral.c
index 16328db6..7223ee35 100644
--- a/libnetdata/aral/aral.c
+++ b/libnetdata/aral/aral.c
@@ -1057,7 +1057,7 @@ int aral_stress_test(size_t threads, size_t elements, size_t seconds) {
     }
 
     netdata_log_info("ARAL: did %zu malloc, %zu free, "
-         "using %zu threads, in %llu usecs",
+         "using %zu threads, in %"PRIu64" usecs",
          auc.ar->aral_lock.user_malloc_operations,
          auc.ar->aral_lock.user_free_operations,
          threads,
diff --git a/libnetdata/buffer/buffer.c b/libnetdata/buffer/buffer.c
index 2d09bb1f..feee72fc 100644
--- a/libnetdata/buffer/buffer.c
+++ b/libnetdata/buffer/buffer.c
@@ -20,16 +20,6 @@ void buffer_reset(BUFFER *wb) {
     buffer_overflow_check(wb);
 }
 
-const char *buffer_tostring(BUFFER *wb)
-{
-    buffer_need_bytes(wb, 1);
-    wb->buffer[wb->len] = '\0';
-
-    buffer_overflow_check(wb);
-
-    return(wb->buffer);
-}
-
 void buffer_char_replace(BUFFER *wb, char from, char to) {
     char *s = wb->buffer, *end = &wb->buffer[wb->len];
 
diff --git a/libnetdata/buffer/buffer.h b/libnetdata/buffer/buffer.h
index 0f154028..26efe007 100644
--- a/libnetdata/buffer/buffer.h
+++ b/libnetdata/buffer/buffer.h
@@ -97,7 +97,6 @@ typedef struct web_buffer {
 #define buffer_no_cacheable(wb) do { (wb)->options |= WB_CONTENT_NO_CACHEABLE; if((wb)->options & WB_CONTENT_CACHEABLE)    (wb)->options &= ~WB_CONTENT_CACHEABLE;  (wb)->expires = 0; } while(0)
 
 #define buffer_strlen(wb) ((wb)->len)
-const char *buffer_tostring(BUFFER *wb);
 
 #define BUFFER_OVERFLOW_EOF "EOF"
 
@@ -158,6 +157,16 @@ void buffer_json_initialize(BUFFER *wb, const char *key_quote, const char *value
 
 void buffer_json_finalize(BUFFER *wb);
 
+static const char *buffer_tostring(BUFFER *wb)
+{
+    buffer_need_bytes(wb, 1);
+    wb->buffer[wb->len] = '\0';
+
+    buffer_overflow_check(wb);
+
+    return(wb->buffer);
+}
+
 static inline void _buffer_json_depth_push(BUFFER *wb, BUFFER_JSON_NODE_TYPE type) {
 #ifdef NETDATA_INTERNAL_CHECKS
     assert(wb->json.depth <= BUFFER_JSON_MAX_DEPTH && "BUFFER JSON: max nesting reached");
@@ -249,20 +258,25 @@ static inline void buffer_strcat(BUFFER *wb, const char *txt) {
     buffer_overflow_check(wb);
 }
 
+static inline void buffer_contents_replace(BUFFER *wb, const char *txt, size_t len) {
+    wb->len = 0;
+    buffer_need_bytes(wb, len + 1);
+
+    memcpy(wb->buffer, txt, len);
+    wb->len = len;
+    wb->buffer[wb->len] = '\0';
+
+    buffer_overflow_check(wb);
+}
+
 static inline void buffer_strncat(BUFFER *wb, const char *txt, size_t len) {
     if(unlikely(!txt || !*txt)) return;
 
-    const char *t = txt;
     buffer_need_bytes(wb, len + 1);
-    char *s = &wb->buffer[wb->len];
-    char *d = s;
-    const char *e = &wb->buffer[wb->len + len];
 
-    while(*t && d < e)
-        *d++ = *t++;
-
-    wb->len += d - s;
+    memcpy(&wb->buffer[wb->len], txt, len);
 
+    wb->len += len;
     wb->buffer[wb->len] = '\0';
 
     buffer_overflow_check(wb);
@@ -944,10 +958,12 @@ typedef enum __attribute__((packed)) {
     RRDF_FIELD_OPTS_VISIBLE      = (1 << 1), // the field should be visible by default
     RRDF_FIELD_OPTS_STICKY       = (1 << 2), // the field should be sticky
     RRDF_FIELD_OPTS_FULL_WIDTH   = (1 << 3), // the field should get full width
-    RRDF_FIELD_OPTS_WRAP         = (1 << 4), // the field should get full width
+    RRDF_FIELD_OPTS_WRAP         = (1 << 4), // the field should wrap
+    RRDR_FIELD_OPTS_DUMMY        = (1 << 5), // not a presentable field
 } RRDF_FIELD_OPTIONS;
 
 typedef enum __attribute__((packed)) {
+    RRDF_FIELD_TYPE_NONE,
     RRDF_FIELD_TYPE_INTEGER,
     RRDF_FIELD_TYPE_STRING,
     RRDF_FIELD_TYPE_DETAIL_STRING,
@@ -960,6 +976,9 @@ typedef enum __attribute__((packed)) {
 static inline const char *rrdf_field_type_to_string(RRDF_FIELD_TYPE type) {
     switch(type) {
         default:
+        case RRDF_FIELD_TYPE_NONE:
+            return "none";
+
         case RRDF_FIELD_TYPE_INTEGER:
             return "integer";
 
@@ -984,10 +1003,11 @@ static inline const char *rrdf_field_type_to_string(RRDF_FIELD_TYPE type) {
 }
 
 typedef enum __attribute__((packed)) {
-    RRDF_FIELD_VISUAL_VALUE,    // show the value, possibly applying a transformation
-    RRDF_FIELD_VISUAL_BAR,      // show the value and a bar, respecting the max field to fill the bar at 100%
-    RRDF_FIELD_VISUAL_PILL,     //
-    RRDF_FIELD_VISUAL_MARKDOC,  //
+    RRDF_FIELD_VISUAL_VALUE,        // show the value, possibly applying a transformation
+    RRDF_FIELD_VISUAL_BAR,          // show the value and a bar, respecting the max field to fill the bar at 100%
+    RRDF_FIELD_VISUAL_PILL,         //
+    RRDF_FIELD_VISUAL_RICH,         //
+    RRDR_FIELD_VISUAL_ROW_OPTIONS,  // this is a dummy column that is used for row options
 } RRDF_FIELD_VISUAL;
 
 static inline const char *rrdf_field_visual_to_string(RRDF_FIELD_VISUAL visual) {
@@ -1002,8 +1022,11 @@ static inline const char *rrdf_field_visual_to_string(RRDF_FIELD_VISUAL visual)
         case RRDF_FIELD_VISUAL_PILL:
             return "pill";
 
-        case RRDF_FIELD_VISUAL_MARKDOC:
-            return "markdoc";
+        case RRDF_FIELD_VISUAL_RICH:
+            return "richValue";
+
+        case RRDR_FIELD_VISUAL_ROW_OPTIONS:
+            return "rowOptions";
     }
 }
 
@@ -1150,6 +1173,9 @@ buffer_rrdf_table_add_field(BUFFER *wb, size_t field_id, const char *key, const
 
         buffer_json_member_add_boolean(wb, "full_width", options & RRDF_FIELD_OPTS_FULL_WIDTH);
         buffer_json_member_add_boolean(wb, "wrap", options & RRDF_FIELD_OPTS_WRAP);
+
+        if(options & RRDR_FIELD_OPTS_DUMMY)
+            buffer_json_member_add_boolean(wb, "dummy", true);
     }
     buffer_json_object_close(wb);
 }
diff --git a/libnetdata/clocks/clocks.c b/libnetdata/clocks/clocks.c
index 806dc06a..489e9685 100644
--- a/libnetdata/clocks/clocks.c
+++ b/libnetdata/clocks/clocks.c
@@ -299,7 +299,7 @@ usec_t heartbeat_next(heartbeat_t *hb, usec_t tick) {
         usec_t tmp = (now_realtime_usec() * clock_realtime_resolution) % (tick / 2);
 
         error_limit_static_global_var(erl, 10, 0);
-        error_limit(&erl, "heartbeat randomness of %llu is too big for a tick of %llu - setting it to %llu", hb->randomness, tick, tmp);
+        error_limit(&erl, "heartbeat randomness of %"PRIu64" is too big for a tick of %"PRIu64" - setting it to %"PRIu64"", hb->randomness, tick, tmp);
         hb->randomness = tmp;
     }
 
@@ -326,12 +326,12 @@ usec_t heartbeat_next(heartbeat_t *hb, usec_t tick) {
     if(unlikely(now < next)) {
         errno = 0;
         error_limit_static_global_var(erl, 10, 0);
-        error_limit(&erl, "heartbeat clock: woke up %llu microseconds earlier than expected (can be due to the CLOCK_REALTIME set to the past).", next - now);
+        error_limit(&erl, "heartbeat clock: woke up %"PRIu64" microseconds earlier than expected (can be due to the CLOCK_REALTIME set to the past).", next - now);
     }
     else if(unlikely(now - next >  tick / 2)) {
         errno = 0;
         error_limit_static_global_var(erl, 10, 0);
-        error_limit(&erl, "heartbeat clock: woke up %llu microseconds later than expected (can be due to system load or the CLOCK_REALTIME set to the future).", now - next);
+        error_limit(&erl, "heartbeat clock: woke up %"PRIu64" microseconds later than expected (can be due to system load or the CLOCK_REALTIME set to the future).", now - next);
     }
 
     if(unlikely(!hb->realtime)) {
@@ -381,7 +381,7 @@ void sleep_usec_with_now(usec_t usec, usec_t started_ut) {
             }
         }
         else {
-            netdata_log_error("Cannot nanosleep() for %llu microseconds.", usec);
+            netdata_log_error("Cannot nanosleep() for %"PRIu64" microseconds.", usec);
             break;
         }
     }
diff --git a/libnetdata/clocks/clocks.h b/libnetdata/clocks/clocks.h
index b050b625..5b88a457 100644
--- a/libnetdata/clocks/clocks.h
+++ b/libnetdata/clocks/clocks.h
@@ -16,10 +16,10 @@ struct timespec {
 typedef int clockid_t;
 #endif
 
-typedef unsigned long long nsec_t;
-typedef unsigned long long msec_t;
-typedef unsigned long long usec_t;
-typedef long long susec_t;
+typedef uint64_t nsec_t;
+typedef uint64_t msec_t;
+typedef uint64_t usec_t;
+typedef int64_t  susec_t;
 
 typedef struct heartbeat {
     usec_t realtime;
diff --git a/libnetdata/dictionary/dictionary.c b/libnetdata/dictionary/dictionary.c
index 05da5534..a74a5958 100644
--- a/libnetdata/dictionary/dictionary.c
+++ b/libnetdata/dictionary/dictionary.c
@@ -250,6 +250,7 @@ static inline void pointer_del(DICTIONARY *dict __maybe_unused, DICTIONARY_ITEM
 // ----------------------------------------------------------------------------
 // memory statistics
 
+#ifdef DICT_WITH_STATS
 static inline void DICTIONARY_STATS_PLUS_MEMORY(DICTIONARY *dict, size_t key_size, size_t item_size, size_t value_size) {
     if(key_size)
         __atomic_fetch_add(&dict->stats->memory.index, (long)JUDYHS_INDEX_SIZE_ESTIMATE(key_size), __ATOMIC_RELAXED);
@@ -260,6 +261,7 @@ static inline void DICTIONARY_STATS_PLUS_MEMORY(DICTIONARY *dict, size_t key_siz
     if(value_size)
         __atomic_fetch_add(&dict->stats->memory.values, (long)value_size, __ATOMIC_RELAXED);
 }
+
 static inline void DICTIONARY_STATS_MINUS_MEMORY(DICTIONARY *dict, size_t key_size, size_t item_size, size_t value_size) {
     if(key_size)
         __atomic_fetch_sub(&dict->stats->memory.index, (long)JUDYHS_INDEX_SIZE_ESTIMATE(key_size), __ATOMIC_RELAXED);
@@ -270,6 +272,10 @@ static inline void DICTIONARY_STATS_MINUS_MEMORY(DICTIONARY *dict, size_t key_si
     if(value_size)
         __atomic_fetch_sub(&dict->stats->memory.values, (long)value_size, __ATOMIC_RELAXED);
 }
+#else
+#define DICTIONARY_STATS_PLUS_MEMORY(dict, key_size, item_size, value_size) do {;} while(0)
+#define DICTIONARY_STATS_MINUS_MEMORY(dict, key_size, item_size, value_size) do {;} while(0)
+#endif
 
 // ----------------------------------------------------------------------------
 // callbacks registration
@@ -376,14 +382,21 @@ void dictionary_version_increment(DICTIONARY *dict) {
 // ----------------------------------------------------------------------------
 // internal statistics API
 
+#ifdef DICT_WITH_STATS
 static inline void DICTIONARY_STATS_SEARCHES_PLUS1(DICTIONARY *dict) {
     __atomic_fetch_add(&dict->stats->ops.searches, 1, __ATOMIC_RELAXED);
 }
+#else
+#define DICTIONARY_STATS_SEARCHES_PLUS1(dict) do {;} while(0)
+#endif
+
 static inline void DICTIONARY_ENTRIES_PLUS1(DICTIONARY *dict) {
+#ifdef DICT_WITH_STATS
     // statistics
     __atomic_fetch_add(&dict->stats->items.entries, 1, __ATOMIC_RELAXED);
     __atomic_fetch_add(&dict->stats->items.referenced, 1, __ATOMIC_RELAXED);
     __atomic_fetch_add(&dict->stats->ops.inserts, 1, __ATOMIC_RELAXED);
+#endif
 
     if(unlikely(is_dictionary_single_threaded(dict))) {
         dict->version++;
@@ -397,10 +410,13 @@ static inline void DICTIONARY_ENTRIES_PLUS1(DICTIONARY *dict) {
         __atomic_fetch_add(&dict->referenced_items, 1, __ATOMIC_RELAXED);
     }
 }
+
 static inline void DICTIONARY_ENTRIES_MINUS1(DICTIONARY *dict) {
+#ifdef DICT_WITH_STATS
     // statistics
     __atomic_fetch_add(&dict->stats->ops.deletes, 1, __ATOMIC_RELAXED);
     __atomic_fetch_sub(&dict->stats->items.entries, 1, __ATOMIC_RELAXED);
+#endif
 
     size_t entries; (void)entries;
     if(unlikely(is_dictionary_single_threaded(dict))) {
@@ -418,14 +434,19 @@ static inline void DICTIONARY_ENTRIES_MINUS1(DICTIONARY *dict) {
                    dict->creation_line,
                    dict->creation_file);
 }
+
 static inline void DICTIONARY_VALUE_RESETS_PLUS1(DICTIONARY *dict) {
+#ifdef DICT_WITH_STATS
     __atomic_fetch_add(&dict->stats->ops.resets, 1, __ATOMIC_RELAXED);
+#endif
 
     if(unlikely(is_dictionary_single_threaded(dict)))
         dict->version++;
     else
         __atomic_fetch_add(&dict->version, 1, __ATOMIC_RELAXED);
 }
+
+#ifdef DICT_WITH_STATS
 static inline void DICTIONARY_STATS_TRAVERSALS_PLUS1(DICTIONARY *dict) {
     __atomic_fetch_add(&dict->stats->ops.traversals, 1, __ATOMIC_RELAXED);
 }
@@ -476,9 +497,29 @@ static inline void DICTIONARY_STATS_DICT_DESTROY_QUEUED_MINUS1(DICTIONARY *dict)
 static inline void DICTIONARY_STATS_DICT_FLUSHES_PLUS1(DICTIONARY *dict) {
     __atomic_fetch_add(&dict->stats->ops.flushes, 1, __ATOMIC_RELAXED);
 }
+#else
+#define DICTIONARY_STATS_TRAVERSALS_PLUS1(dict) do {;} while(0)
+#define DICTIONARY_STATS_WALKTHROUGHS_PLUS1(dict) do {;} while(0)
+#define DICTIONARY_STATS_CHECK_SPINS_PLUS(dict, count) do {;} while(0)
+#define DICTIONARY_STATS_INSERT_SPINS_PLUS(dict, count) do {;} while(0)
+#define DICTIONARY_STATS_DELETE_SPINS_PLUS(dict, count) do {;} while(0)
+#define DICTIONARY_STATS_SEARCH_IGNORES_PLUS1(dict) do {;} while(0)
+#define DICTIONARY_STATS_CALLBACK_INSERTS_PLUS1(dict) do {;} while(0)
+#define DICTIONARY_STATS_CALLBACK_CONFLICTS_PLUS1(dict) do {;} while(0)
+#define DICTIONARY_STATS_CALLBACK_REACTS_PLUS1(dict) do {;} while(0)
+#define DICTIONARY_STATS_CALLBACK_DELETES_PLUS1(dict) do {;} while(0)
+#define DICTIONARY_STATS_GARBAGE_COLLECTIONS_PLUS1(dict) do {;} while(0)
+#define DICTIONARY_STATS_DICT_CREATIONS_PLUS1(dict) do {;} while(0)
+#define DICTIONARY_STATS_DICT_DESTRUCTIONS_PLUS1(dict) do {;} while(0)
+#define DICTIONARY_STATS_DICT_DESTROY_QUEUED_PLUS1(dict) do {;} while(0)
+#define DICTIONARY_STATS_DICT_DESTROY_QUEUED_MINUS1(dict) do {;} while(0)
+#define DICTIONARY_STATS_DICT_FLUSHES_PLUS1(dict) do {;} while(0)
+#endif
 
 static inline void DICTIONARY_REFERENCED_ITEMS_PLUS1(DICTIONARY *dict) {
+#ifdef DICT_WITH_STATS
     __atomic_fetch_add(&dict->stats->items.referenced, 1, __ATOMIC_RELAXED);
+#endif
 
     if(unlikely(is_dictionary_single_threaded(dict)))
         ++dict->referenced_items;
@@ -487,7 +528,9 @@ static inline void DICTIONARY_REFERENCED_ITEMS_PLUS1(DICTIONARY *dict) {
 }
 
 static inline void DICTIONARY_REFERENCED_ITEMS_MINUS1(DICTIONARY *dict) {
+#ifdef DICT_WITH_STATS
     __atomic_fetch_sub(&dict->stats->items.referenced, 1, __ATOMIC_RELAXED);
+#endif
 
     long int referenced_items; (void)referenced_items;
     if(unlikely(is_dictionary_single_threaded(dict)))
@@ -504,7 +547,9 @@ static inline void DICTIONARY_REFERENCED_ITEMS_MINUS1(DICTIONARY *dict) {
 }
 
 static inline void DICTIONARY_PENDING_DELETES_PLUS1(DICTIONARY *dict) {
+#ifdef DICT_WITH_STATS
     __atomic_fetch_add(&dict->stats->items.pending_deletion, 1, __ATOMIC_RELAXED);
+#endif
 
     if(unlikely(is_dictionary_single_threaded(dict)))
         ++dict->pending_deletion_items;
@@ -513,7 +558,9 @@ static inline void DICTIONARY_PENDING_DELETES_PLUS1(DICTIONARY *dict) {
 }
 
 static inline long int DICTIONARY_PENDING_DELETES_MINUS1(DICTIONARY *dict) {
+#ifdef DICT_WITH_STATS
     __atomic_fetch_sub(&dict->stats->items.pending_deletion, 1, __ATOMIC_RELEASE);
+#endif
 
     if(unlikely(is_dictionary_single_threaded(dict)))
         return --dict->pending_deletion_items;
@@ -977,7 +1024,7 @@ static int item_check_and_acquire_advanced(DICTIONARY *dict, DICTIONARY_ITEM *it
             DICTIONARY_REFERENCED_ITEMS_PLUS1(dict);
     }
 
-    if(unlikely(spins > 1 && dict->stats))
+    if(unlikely(spins > 1))
         DICTIONARY_STATS_CHECK_SPINS_PLUS(dict, spins - 1);
 
     return ret;
@@ -1022,7 +1069,7 @@ static inline int item_is_not_referenced_and_can_be_removed_advanced(DICTIONARY
         item->deleter_pid = gettid();
 #endif
 
-    if(unlikely(spins > 1 && dict->stats))
+    if(unlikely(spins > 1))
         DICTIONARY_STATS_DELETE_SPINS_PLUS(dict, spins - 1);
 
     return ret;
@@ -1535,22 +1582,6 @@ static inline void dict_item_release_and_check_if_it_is_deleted_and_can_be_remov
 }
 
 static bool dict_item_del(DICTIONARY *dict, const char *name, ssize_t name_len) {
-    if(unlikely(!name || !*name)) {
-        internal_error(
-            true,
-            "DICTIONARY: attempted to %s() without a name on a dictionary created from %s() %zu@%s.",
-            __FUNCTION__,
-            dict->creation_function,
-            dict->creation_line,
-            dict->creation_file);
-        return false;
-    }
-
-    if(unlikely(is_dictionary_destroyed(dict))) {
-        internal_error(true, "DICTIONARY: attempted to dictionary_del() on a destroyed dictionary");
-        return false;
-    }
-
     if(name_len == -1)
         name_len = (ssize_t)strlen(name) + 1; // we need the terminating null too
 
@@ -1695,7 +1726,7 @@ static DICTIONARY_ITEM *dict_item_add_or_reset_value_and_acquire(DICTIONARY *dic
     } while(!item);
 
 
-    if(unlikely(spins > 0 && dict->stats))
+    if(unlikely(spins > 0))
         DICTIONARY_STATS_INSERT_SPINS_PLUS(dict, spins);
 
     if(is_master_dictionary(dict) && added_or_updated)
@@ -2064,11 +2095,15 @@ void dictionary_flush(DICTIONARY *dict) {
     if(unlikely(!dict))
         return;
 
-    void *value;
-    dfe_start_write(dict, value) {
-        dictionary_del_advanced(dict, item_get_name(value_dfe.item), (ssize_t)item_get_name_len(value_dfe.item) + 1);
+    ll_recursive_lock(dict, DICTIONARY_LOCK_WRITE);
+
+    DICTIONARY_ITEM *item, *next = NULL;
+    for(item = dict->items.list; item ;item = next) {
+        next = item->next;
+        dict_item_del(dict, item_get_name(item), (ssize_t) item_get_name_len(item) + 1);
     }
-    dfe_done(value);
+
+    ll_recursive_unlock(dict, DICTIONARY_LOCK_WRITE);
 
     DICTIONARY_STATS_DICT_FLUSHES_PLUS1(dict);
 }
@@ -2251,6 +2286,12 @@ bool dictionary_del_advanced(DICTIONARY *dict, const char *name, ssize_t name_le
         return false;
 
     api_internal_check(dict, NULL, false, true);
+
+    if(unlikely(is_dictionary_destroyed(dict))) {
+        internal_error(true, "DICTIONARY: attempted to delete item on a destroyed dictionary");
+        return false;
+    }
+
     return dict_item_del(dict, name, name_len);
 }
 
@@ -2260,6 +2301,8 @@ bool dictionary_del_advanced(DICTIONARY *dict, const char *name, ssize_t name_le
 void *dictionary_foreach_start_rw(DICTFE *dfe, DICTIONARY *dict, char rw) {
     if(unlikely(!dfe || !dict)) return NULL;
 
+    DICTIONARY_STATS_TRAVERSALS_PLUS1(dict);
+
     if(unlikely(is_dictionary_destroyed(dict))) {
         internal_error(true, "DICTIONARY: attempted to dictionary_foreach_start_rw() on a destroyed dictionary");
         dfe->counter = 0;
@@ -2275,8 +2318,6 @@ void *dictionary_foreach_start_rw(DICTFE *dfe, DICTIONARY *dict, char rw) {
     dfe->locked = true;
     ll_recursive_lock(dict, dfe->rw);
 
-    DICTIONARY_STATS_TRAVERSALS_PLUS1(dict);
-
     // get the first item from the list
     DICTIONARY_ITEM *item = dict->items.list;
 
@@ -2836,7 +2877,7 @@ static usec_t dictionary_unittest_run_and_measure_time(DICTIONARY *dict, char *m
         }
     }
 
-    fprintf(stderr, " %zu errors, %d (found %ld) items in dictionary, %d (found %ld) referenced, %d (found %ld) deleted, %llu usec \n",
+    fprintf(stderr, " %zu errors, %d (found %ld) items in dictionary, %d (found %ld) referenced, %d (found %ld) deleted, %"PRIu64" usec \n",
             errs, dict?dict->entries:0, found_ok, dict?dict->referenced_items:0, found_referenced, dict?dict->pending_deletion_items:0, found_deleted, dt);
     *errors += errs;
     return dt;
@@ -3129,6 +3170,9 @@ struct thread_unittest {
     int join;
     DICTIONARY *dict;
     int dups;
+
+    netdata_thread_t thread;
+    struct dictionary_stats stats;
 };
 
 static void *unittest_dict_thread(void *arg) {
@@ -3140,46 +3184,59 @@ static void *unittest_dict_thread(void *arg) {
         DICT_ITEM_CONST DICTIONARY_ITEM *item =
             dictionary_set_and_acquire_item_advanced(tu->dict, "dict thread checking 1234567890",
                                                      -1, NULL, 0, NULL);
-
+        tu->stats.ops.inserts++;
 
         dictionary_get(tu->dict, dictionary_acquired_item_name(item));
+        tu->stats.ops.searches++;
 
         void *t1;
         dfe_start_write(tu->dict, t1) {
 
             // this should delete the referenced item
             dictionary_del(tu->dict, t1_dfe.name);
+            tu->stats.ops.deletes++;
 
             void *t2;
             dfe_start_write(tu->dict, t2) {
                 // this should add another
                 dictionary_set(tu->dict, t2_dfe.name, NULL, 0);
+                tu->stats.ops.inserts++;
 
                 dictionary_get(tu->dict, dictionary_acquired_item_name(item));
+                tu->stats.ops.searches++;
 
                 // and this should delete it again
                 dictionary_del(tu->dict, t2_dfe.name);
+                tu->stats.ops.deletes++;
             }
             dfe_done(t2);
+            tu->stats.ops.traversals++;
 
             // this should fail to add it
             dictionary_set(tu->dict, t1_dfe.name, NULL, 0);
+            tu->stats.ops.inserts++;
+
             dictionary_del(tu->dict, t1_dfe.name);
+            tu->stats.ops.deletes++;
         }
         dfe_done(t1);
+        tu->stats.ops.traversals++;
 
         for(int i = 0; i < tu->dups ; i++) {
             dictionary_acquired_item_dup(tu->dict, item);
             dictionary_get(tu->dict, dictionary_acquired_item_name(item));
+            tu->stats.ops.searches++;
         }
 
         for(int i = 0; i < tu->dups ; i++) {
             dictionary_acquired_item_release(tu->dict, item);
             dictionary_del(tu->dict, dictionary_acquired_item_name(item));
+            tu->stats.ops.deletes++;
         }
 
         dictionary_acquired_item_release(tu->dict, item);
         dictionary_del(tu->dict, "dict thread checking 1234567890");
+        tu->stats.ops.deletes++;
 
         // test concurrent deletions and flushes
         {
@@ -3189,16 +3246,19 @@ static void *unittest_dict_thread(void *arg) {
                 for (int i = 0; i < 1000; i++) {
                     snprintfz(buf, 256, "del/flush test %d", i);
                     dictionary_set(tu->dict, buf, NULL, 0);
+                    tu->stats.ops.inserts++;
                 }
 
                 for (int i = 0; i < 1000; i++) {
                     snprintfz(buf, 256, "del/flush test %d", i);
                     dictionary_del(tu->dict, buf);
+                    tu->stats.ops.deletes++;
                 }
             }
             else {
                 for (int i = 0; i < 10; i++) {
                     dictionary_flush(tu->dict);
+                    tu->stats.ops.flushes++;
                 }
             }
         }
@@ -3208,47 +3268,75 @@ static void *unittest_dict_thread(void *arg) {
 }
 
 static int dictionary_unittest_threads() {
-
-    struct thread_unittest tu = {
-        .join = 0,
-        .dict = NULL,
-        .dups = 1,
-    };
-
-    // threads testing of dictionary
-    tu.dict = dictionary_create(DICT_OPTION_DONT_OVERWRITE_VALUE);
     time_t seconds_to_run = 5;
     int threads_to_create = 2;
+
+    struct thread_unittest tu[threads_to_create];
+    memset(tu, 0, sizeof(struct thread_unittest) * threads_to_create);
+
     fprintf(
-        stderr,
-        "\nChecking dictionary concurrency with %d threads for %lld seconds...\n",
-        threads_to_create,
-        (long long)seconds_to_run);
+            stderr,
+            "\nChecking dictionary concurrency with %d threads for %lld seconds...\n",
+            threads_to_create,
+            (long long)seconds_to_run);
+
+    // threads testing of dictionary
+    struct dictionary_stats stats = {};
+    tu[0].join = 0;
+    tu[0].dups = 1;
+    tu[0].dict = dictionary_create_advanced(DICT_OPTION_DONT_OVERWRITE_VALUE, &stats, 0);
 
-    netdata_thread_t threads[threads_to_create];
-    tu.join = 0;
     for (int i = 0; i < threads_to_create; i++) {
+        if(i)
+            tu[i] = tu[0];
+
         char buf[100 + 1];
         snprintf(buf, 100, "dict%d", i);
         netdata_thread_create(
-            &threads[i],
+            &tu[i].thread,
             buf,
             NETDATA_THREAD_OPTION_DONT_LOG | NETDATA_THREAD_OPTION_JOINABLE,
             unittest_dict_thread,
-            &tu);
+            &tu[i]);
     }
+
     sleep_usec(seconds_to_run * USEC_PER_SEC);
 
-    __atomic_store_n(&tu.join, 1, __ATOMIC_RELAXED);
     for (int i = 0; i < threads_to_create; i++) {
+        __atomic_store_n(&tu[i].join, 1, __ATOMIC_RELAXED);
+
         void *retval;
-        netdata_thread_join(threads[i], &retval);
+        netdata_thread_join(tu[i].thread, &retval);
+
+        if(i) {
+            tu[0].stats.ops.inserts += tu[i].stats.ops.inserts;
+            tu[0].stats.ops.deletes += tu[i].stats.ops.deletes;
+            tu[0].stats.ops.searches += tu[i].stats.ops.searches;
+            tu[0].stats.ops.flushes += tu[i].stats.ops.flushes;
+            tu[0].stats.ops.traversals += tu[i].stats.ops.traversals;
+        }
     }
 
     fprintf(stderr,
-            "inserts %zu"
+            "CALLS : inserts %zu"
             ", deletes %zu"
             ", searches %zu"
+            ", traversals %zu"
+            ", flushes %zu"
+            "\n",
+            tu[0].stats.ops.inserts,
+            tu[0].stats.ops.deletes,
+            tu[0].stats.ops.searches,
+            tu[0].stats.ops.traversals,
+            tu[0].stats.ops.flushes
+    );
+
+#ifdef DICT_WITH_STATS
+    fprintf(stderr,
+            "ACTUAL: inserts %zu"
+            ", deletes %zu"
+            ", searches %zu"
+            ", traversals %zu"
             ", resets %zu"
             ", flushes %zu"
             ", entries %d"
@@ -3259,22 +3347,23 @@ static int dictionary_unittest_threads() {
             ", delete spins %zu"
             ", search ignores %zu"
             "\n",
-            tu.dict->stats->ops.inserts,
-            tu.dict->stats->ops.deletes,
-            tu.dict->stats->ops.searches,
-            tu.dict->stats->ops.resets,
-            tu.dict->stats->ops.flushes,
-            tu.dict->entries,
-            tu.dict->referenced_items,
-            tu.dict->pending_deletion_items,
-            tu.dict->stats->spin_locks.use_spins,
-            tu.dict->stats->spin_locks.insert_spins,
-            tu.dict->stats->spin_locks.delete_spins,
-            tu.dict->stats->spin_locks.search_spins
+            stats.ops.inserts,
+            stats.ops.deletes,
+            stats.ops.searches,
+            stats.ops.traversals,
+            stats.ops.resets,
+            stats.ops.flushes,
+            tu[0].dict->entries,
+            tu[0].dict->referenced_items,
+            tu[0].dict->pending_deletion_items,
+            stats.spin_locks.use_spins,
+            stats.spin_locks.insert_spins,
+            stats.spin_locks.delete_spins,
+            stats.spin_locks.search_spins
     );
-    dictionary_destroy(tu.dict);
-    tu.dict = NULL;
+#endif
 
+    dictionary_destroy(tu[0].dict);
     return 0;
 }
 
@@ -3407,6 +3496,7 @@ static int dictionary_unittest_view_threads() {
     netdata_thread_join(view_thread, &retval);
     netdata_thread_join(master_thread, &retval);
 
+#ifdef DICT_WITH_STATS
     fprintf(stderr,
             "MASTER: inserts %zu"
             ", deletes %zu"
@@ -3457,6 +3547,8 @@ static int dictionary_unittest_view_threads() {
             stats_view.spin_locks.delete_spins,
             stats_view.spin_locks.search_spins
     );
+#endif
+
     dictionary_destroy(tv.master);
     dictionary_destroy(tv.view);
 
diff --git a/libnetdata/dictionary/dictionary.h b/libnetdata/dictionary/dictionary.h
index eea14d3f..72efe1d0 100644
--- a/libnetdata/dictionary/dictionary.h
+++ b/libnetdata/dictionary/dictionary.h
@@ -35,6 +35,10 @@
  *
  */
 
+#ifdef NETDATA_INTERNAL_CHECKS
+#define DICT_WITH_STATS 1
+#endif
+
 #ifdef DICTIONARY_INTERNALS
 #define DICTFE_CONST
 #define DICT_ITEM_CONST
@@ -92,7 +96,7 @@ struct dictionary_stats {
 
     // memory
     struct {
-        long index;               // bytes of keys indexed (indication of the index size)
+        long index;                 // bytes of keys indexed (indication of the index size)
         long values;                // bytes of caller structures
         long dict;                  // bytes of the structures dictionary needs
     } memory;
diff --git a/libnetdata/dyn_conf/README.md b/libnetdata/dyn_conf/README.md
new file mode 100644
index 00000000..6c812740
--- /dev/null
+++ b/libnetdata/dyn_conf/README.md
@@ -0,0 +1,167 @@
+# Netdata Dynamic Configuration
+
+Purpose of Netdata Dynamic Configuration is to allow configuration of select Netdata plugins and options through the
+Netdata API and by extension by UI.
+
+## HTTP API documentation
+
+### Summary API
+
+For summary of all jobs and their statuses (for all children that stream to parent) use the following URL:
+
+| Method  | Endpoint                      | Description                                                |
+|:-------:|-------------------------------|------------------------------------------------------------|
+| **GET** | `api/v2/job_statuses`         | list of Jobs                                               |
+| **GET** | `api/v2/job_statuses?grouped` | list of Jobs (hierarchical, grouped by host/plugin/module) |
+
+### Dyncfg API
+
+### Top level
+
+| Method  | Endpoint         | Description                             |
+|:-------:|------------------|-----------------------------------------|
+| **GET** | `/api/v2/config` | registered Plugins (sent DYNCFG_ENABLE) |
+
+### Plugin level
+
+| Method  | Endpoint                          | Description                  |
+|:-------:|-----------------------------------|------------------------------|
+| **GET** | `/api/v2/config/[plugin]`         | Plugin config                |
+| **PUT** | `/api/v2/config/[plugin]`         | update Plugin config         |
+| **GET** | `/api/v2/config/[plugin]/modules` | Modules registered by Plugin |
+| **GET** | `/api/v2/config/[plugin]/schema`  | Plugin config schema         |
+
+### Module level
+
+| Method  | Endpoint                                      | Description               |
+|:-------:|-----------------------------------------------|---------------------------|
+| **GET** | `/api/v2/config/<plugin>/[module]`            | Module config             |
+| **PUT** | `/api/v2/config/[plugin]/[module]`            | update Module config      |
+| **GET** | `/api/v2/config/[plugin]/[module]/jobs`       | Jobs registered by Module |
+| **GET** | `/api/v2/config/[plugin]/[module]/job_schema` | Job config schema         |
+| **GET** | `/api/v2/config/[plugin]/[module]/schema`     | Module config schema      |
+
+### Job level - only for modules where `module_type == job_array`
+
+|   Method   | Endpoint                                 | Description                    |
+|:----------:|------------------------------------------|--------------------------------|
+|  **GET**   | `/api/v2/config/[plugin]/[module]/[job]` | Job config                     |
+|  **PUT**   | `/api/v2/config/[plugin]/[module]/[job]` | update Job config              |
+|  **POST**  | `/api/v2/config/[plugin]/[module]/[job]` | create Job                     |
+| **DELETE** | `/api/v2/config/[plugin]/[module]/[job]` | delete Job (created by Dyncfg) |
+
+## AGENT<->PLUGIN interface documentation
+
+### 1. DYNCFG_ENABLE
+
+Plugin signifies to agent its ability to use new dynamic config and the name it wishes to use by sending
+
+```
+plugin->agent:
+=============
+DYNCFG_ENABLE [plugin_url_name]
+```
+
+This can be sent only once per lifetime of the plugin (at startup or later) sending it multiple times is considered a
+protocol violation and plugin might get terminated.
+After this command is sent the plugin has to be ready to accept all the new commands/keywords related to dynamic
+configuration (this command lets agent know this plugin is dyncfg capable and wishes to use dyncfg functionality).
+
+After this command agent can call
+
+```
+agent->plugin:
+=============
+FUNCTION_PAYLOAD [UUID] 1 "set_plugin_config"
+the new configuration
+blah blah blah
+FUNCTION_PAYLOAD_END
+
+plugin->agent:
+=============
+FUNCTION_RESULT_BEGIN [UUID] [(1/0)(accept/reject)] [text/plain] 5
+FUNCTION_RESULT_END
+```
+
+to set the new config which can be accepted/rejected by plugin by sending answer for this FUNCTION as it would with any
+other regular function.
+
+The new `FUNCTION_PAYLOAD` command differs from regular `FUNCTION` command exclusively in its ability to send bigger
+payloads (configuration file contents) to the plugin (not just parameters list).
+
+Agent can also call (after `DYNCFG_ENABLE`)
+
+```
+Agent->plugin:
+=============
+FUNCTION [UID] 1 "get_plugin_config"
+
+Plugin->agent:
+=============
+FUNCTION_RESULT_BEGIN [UID] 1 text/plain 5
+{
+	"the currently used config from plugin" : "nice"
+}
+FUNCTION_RESULT_END
+```
+
+and
+
+```
+Agent->plugin:
+=============
+FUNCTION [UID] 1 "get_plugin_config_schema"
+
+Plugin->agent:
+=============
+FUNCTION_RESULT_BEGIN [UID] 1 text/plain 5
+{
+	"the schema of plugin configuration" : "splendid"
+}
+FUNCTION_RESULT_END
+```
+
+Plugin can also register zero, one or more configurable modules using:
+
+```
+plugin->agent:
+=============
+DYNCFG_REGISTER_MODULE [module_url_name] (job_array|single)
+```
+
+modules can be added any time during plugins lifetime (you are not required to add them all at startup).
+
+### 2. DYNCFG_REGISTER_MODULE
+
+Module has to choose one of following types at registration:
+
+- `single` - module itself has configuration but does not accept any jobs *(this is useful mainly for internal netdata
+  configurable things like webserver etc.)*
+- `job_array`  - module itself **can** *(not must)* have configuration and it has an array of jobs which can be added,
+  modified and deleted. **this is what plugin developer needs in most cases**
+
+After module has been registered agent can call
+
+- `set_module_config [module]` FUNCTION_PAYLOAD
+- `get_module_config [module]` FUNCTION
+- `get_module_config_schema [module]` FUNCTION
+
+with same syntax as `set_plugin_config` and `get_plugin_config`. In case of `set` command the plugin has ability to
+reject the new configuration pushed to it.
+
+In a case the module was registered as `job_array` type following commands can be used to manage jobs:
+
+### 3. Job interface for job_array modules
+
+- `get_job_config_schema [module]` - FUNCTION
+- `get_job_config [module] [job]` - FUNCTION
+- `set_job_config [module] [job]` - FUNCTION_PAYLOAD
+- `delete_job_name [module] [job]` - FUNCTION
+
+### 4. Streaming
+
+When above commands are transferred trough streaming additionally `plugin_name` is prefixed as first parameter. This is
+done to allow routing to appropriate plugin @child.
+
+As a plugin developer you don't need to concern yourself with this detail as that parameter is stripped when sent to the
+plugin *(and added when sent trough streaming)* automagically.
diff --git a/libnetdata/dyn_conf/dyn_conf.c b/libnetdata/dyn_conf/dyn_conf.c
index 00289fdf..ee4b4733 100644
--- a/libnetdata/dyn_conf/dyn_conf.c
+++ b/libnetdata/dyn_conf/dyn_conf.c
@@ -3,7 +3,7 @@
 #include "dyn_conf.h"
 
 #define DYN_CONF_PATH_MAX (4096)
-#define DYN_CONF_DIR VARLIB_DIR "/etc"
+#define DYN_CONF_DIR VARLIB_DIR "/dynconf"
 
 #define DYN_CONF_JOB_SCHEMA "job_schema"
 #define DYN_CONF_SCHEMA "schema"
@@ -11,9 +11,20 @@
 #define DYN_CONF_JOB_LIST "jobs"
 #define DYN_CONF_CFG_EXT ".cfg"
 
-DICTIONARY *plugins_dict = NULL;
+void job_flags_wallkthrough(dyncfg_job_flg_t flags, void (*cb)(const char *str, void *data), void *data)
+{
+    if (flags & JOB_FLG_PS_LOADED)
+        cb("JOB_FLG_PS_LOADED", data);
+    if (flags & JOB_FLG_PLUGIN_PUSHED)
+        cb("JOB_FLG_PLUGIN_PUSHED", data);
+    if (flags & JOB_FLG_STREAMING_PUSHED)
+        cb("JOB_FLG_STREAMING_PUSHED", data);
+    if (flags & JOB_FLG_USER_CREATED)
+        cb("JOB_FLG_USER_CREATED", data);
+}
 
 struct deferred_cfg_send {
+    DICTIONARY *plugins_dict;
     char *plugin_name;
     char *module_name;
     char *job_name;
@@ -33,7 +44,7 @@ static void deferred_config_free(struct deferred_cfg_send *dcs)
     freez(dcs);
 }
 
-static void deferred_config_push_back(const char *plugin_name, const char *module_name, const char *job_name)
+static void deferred_config_push_back(DICTIONARY *plugins_dict, const char *plugin_name, const char *module_name, const char *job_name)
 {
     struct deferred_cfg_send *deferred = callocz(1, sizeof(struct deferred_cfg_send));
     deferred->plugin_name = strdupz(plugin_name);
@@ -42,6 +53,7 @@ static void deferred_config_push_back(const char *plugin_name, const char *modul
         if (job_name != NULL)
             deferred->job_name = strdupz(job_name);
     }
+    deferred->plugins_dict = plugins_dict;
     pthread_mutex_lock(&deferred_configs_lock);
     if (dyncfg_shutdown) {
         pthread_mutex_unlock(&deferred_configs_lock);
@@ -95,7 +107,7 @@ static int _get_list_of_plugins_json_cb(const DICTIONARY_ITEM *item, void *entry
     return 0;
 }
 
-json_object *get_list_of_plugins_json()
+json_object *get_list_of_plugins_json(DICTIONARY *plugins_dict)
 {
     json_object *obj = json_object_new_array();
 
@@ -114,18 +126,7 @@ static int _get_list_of_modules_json_cb(const DICTIONARY_ITEM *item, void *entry
 
     json_object *json_item = json_object_new_string(module->name);
     json_object_object_add(json_module, "name", json_item);
-    const char *module_type;
-    switch (module->type) {
-        case MOD_TYPE_SINGLE:
-            module_type = "single";
-            break;
-        case MOD_TYPE_ARRAY:
-            module_type = "job_array";
-            break;
-        default:
-            module_type = "unknown";
-            break;
-    }
+    const char *module_type = module_type2str(module->type);
     json_item = json_object_new_string(module_type);
     json_object_object_add(json_module, "type", json_item);
 
@@ -163,17 +164,31 @@ const char *job_status2str(enum job_status status)
     }
 }
 
-static int _get_list_of_jobs_json_cb(const DICTIONARY_ITEM *item, void *entry, void *data)
+static void _job_flags2str_cb(const char *str, void *data)
 {
-    UNUSED(item);
-    json_object *obj = (json_object *)data;
-    struct job *job = (struct job *)entry;
+    json_object *json_item = json_object_new_string(str);
+    json_object_array_add((json_object *)data, json_item);
+}
 
+json_object *job2json(struct job *job) {
     json_object *json_job = json_object_new_object();
+
     json_object *json_item = json_object_new_string(job->name);
     json_object_object_add(json_job, "name", json_item);
+
+    json_item = json_object_new_string(job_type2str(job->type));
+    json_object_object_add(json_job, "type", json_item);
+
+    netdata_mutex_lock(&job->lock);
     json_item = json_object_new_string(job_status2str(job->status));
+    json_object_object_add(json_job, "status", json_item);
+
+    json_item = json_object_new_int(job->state);
     json_object_object_add(json_job, "state", json_item);
+
+    json_item = job->reason == NULL ? NULL : json_object_new_string(job->reason);
+    json_object_object_add(json_job, "reason", json_item);
+
     int64_t last_state_update_s  = job->last_state_update / USEC_PER_SEC;
     int64_t last_state_update_us = job->last_state_update % USEC_PER_SEC;
 
@@ -183,6 +198,22 @@ static int _get_list_of_jobs_json_cb(const DICTIONARY_ITEM *item, void *entry, v
     json_item = json_object_new_int64(last_state_update_us);
     json_object_object_add(json_job, "last_state_update_us", json_item);
 
+    json_item = json_object_new_array();
+    job_flags_wallkthrough(job->flags, _job_flags2str_cb, json_item);
+    json_object_object_add(json_job, "flags", json_item);
+
+    netdata_mutex_unlock(&job->lock);
+
+    return json_job;
+}
+
+static int _get_list_of_jobs_json_cb(const DICTIONARY_ITEM *item, void *entry, void *data)
+{
+    UNUSED(item);
+    json_object *obj = (json_object *)data;
+
+    json_object *json_job = job2json((struct job *)entry);
+
     json_object_array_add(obj, json_job);
 
     return 0;
@@ -206,24 +237,59 @@ struct job *get_job_by_name(struct module *module, const char *job_name)
     return dictionary_get(module->jobs, job_name);
 }
 
-int remove_job(struct module *module, struct job *job)
+void unlink_job(const char *plugin_name, const char *module_name, const char *job_name)
 {
     // as we are going to do unlink here we better make sure we have all to build proper path
-    if (unlikely(job->name == NULL || module == NULL || module->name == NULL || module->plugin == NULL || module->plugin->name == NULL))
-        return 0;
+    if (unlikely(job_name == NULL || module_name == NULL || plugin_name == NULL))
+        return;
+    BUFFER *buffer = buffer_create(DYN_CONF_PATH_MAX, NULL);
+    buffer_sprintf(buffer, DYN_CONF_DIR "/%s/%s/%s" DYN_CONF_CFG_EXT, plugin_name, module_name, job_name);
+    if (unlink(buffer_tostring(buffer)))
+        netdata_log_error("Cannot remove file %s", buffer_tostring(buffer));
 
-    enum set_config_result rc = module->delete_job_cb(module->job_config_cb_usr_ctx, module->name, job->name);
+    buffer_free(buffer);
+}
+
+void delete_job(struct configurable_plugin *plugin, const char *module_name, const char *job_name)
+{
+    struct module *module = get_module_by_name(plugin, module_name);
+    if (module == NULL) {
+        error_report("DYNCFG module \"%s\" not found", module_name);
+        return;
+    }
+
+    struct job  *job_item = get_job_by_name(module, job_name);
+    if (job_item == NULL) {
+        error_report("DYNCFG job \"%s\" not found", job_name);
+        return;
+    }
+
+    dictionary_del(module->jobs, job_name);
+}
+
+void delete_job_pname(DICTIONARY *plugins_dict, const char *plugin_name, const char *module_name, const char *job_name)
+{
+    const DICTIONARY_ITEM *plugin_item = dictionary_get_and_acquire_item(plugins_dict, plugin_name);
+    if (plugin_item == NULL) {
+        error_report("DYNCFG plugin \"%s\" not found", plugin_name);
+        return;
+    }
+    struct configurable_plugin *plugin  = dictionary_acquired_item_value(plugin_item);
+
+    delete_job(plugin, module_name, job_name);
+
+    dictionary_acquired_item_release(plugins_dict, plugin_item);
+}
+
+int remove_job(struct module *module, struct job *job)
+{
+    enum set_config_result rc = module->delete_job_cb(module->job_config_cb_usr_ctx, module->plugin->name, module->name, job->name);
 
     if (rc != SET_CONFIG_ACCEPTED) {
         error_report("DYNCFG module \"%s\" rejected delete job for \"%s\"", module->name, job->name);
         return 0;
     }
-
-    BUFFER *buffer = buffer_create(DYN_CONF_PATH_MAX, NULL);
-    buffer_sprintf(buffer, DYN_CONF_DIR "/%s/%s/%s" DYN_CONF_CFG_EXT, module->plugin->name, module->name, job->name);
-    unlink(buffer_tostring(buffer));
-    buffer_free(buffer);
-    return dictionary_del(module->jobs, job->name);
+    return 1;
 }
 
 struct module *get_module_by_name(struct configurable_plugin *plugin, const char *module_name)
@@ -231,7 +297,7 @@ struct module *get_module_by_name(struct configurable_plugin *plugin, const char
     return dictionary_get(plugin->modules, module_name);
 }
 
-inline struct configurable_plugin *get_plugin_by_name(const char *name)
+inline struct configurable_plugin *get_plugin_by_name(DICTIONARY *plugins_dict, const char *name)
 {
     return dictionary_get(plugins_dict, name);
 }
@@ -282,6 +348,59 @@ static int store_config(const char *module_name, const char *submodule_name, con
     return 0;
 }
 
+#ifdef NETDATA_DEV_MODE
+#define netdata_dev_fatal(...) fatal(__VA_ARGS__)
+#else
+#define netdata_dev_fatal(...) error_report(__VA_ARGS__)
+#endif
+
+void dyn_conf_store_config(const char *function, const char *payload, struct configurable_plugin *plugin) {
+    dyncfg_config_t config = {
+        .data = (char*)payload,
+        .data_size = strlen(payload)
+    };
+
+    char *fnc = strdupz(function);
+    // split fnc to words
+    char *words[DYNCFG_MAX_WORDS];
+    size_t words_c = quoted_strings_splitter(fnc, words, DYNCFG_MAX_WORDS, isspace_map_pluginsd);
+
+    const char *fnc_name = get_word(words, words_c, 0);
+    if (fnc_name == NULL) {
+        error_report("Function name expected \"%s\"", function);
+        goto CLEANUP;
+    }
+    if (strncmp(fnc_name, FUNCTION_NAME_SET_PLUGIN_CONFIG, strlen(FUNCTION_NAME_SET_PLUGIN_CONFIG)) == 0) {
+        store_config(plugin->name, NULL, NULL, config);
+        goto CLEANUP;
+    }
+
+    if (words_c < 2) {
+        error_report("Module name expected \"%s\"", function);
+        goto CLEANUP;
+    }
+    const char *module_name = get_word(words, words_c, 1);
+    if (strncmp(fnc_name, FUNCTION_NAME_SET_MODULE_CONFIG, strlen(FUNCTION_NAME_SET_MODULE_CONFIG)) == 0) {
+        store_config(plugin->name, module_name, NULL, config);
+        goto CLEANUP;
+    }
+
+    if (words_c < 3) {
+        error_report("Job name expected \"%s\"", function);
+        goto CLEANUP;
+    }
+    const char *job_name = get_word(words, words_c, 2);
+    if (strncmp(fnc_name, FUNCTION_NAME_SET_JOB_CONFIG, strlen(FUNCTION_NAME_SET_JOB_CONFIG)) == 0) {
+        store_config(plugin->name, module_name, job_name, config);
+        goto CLEANUP;
+    }
+
+    netdata_dev_fatal("Unknown function \"%s\"", function);
+
+CLEANUP:
+    freez(fnc);
+}
+
 dyncfg_config_t load_config(const char *plugin_name, const char *module_name, const char *job_id)
 {
     BUFFER *filename = buffer_create(DYN_CONF_PATH_MAX, NULL);
@@ -310,16 +429,12 @@ dyncfg_config_t load_config(const char *plugin_name, const char *module_name, co
 
 char *set_plugin_config(struct configurable_plugin *plugin, dyncfg_config_t cfg)
 {
-    enum set_config_result rc = plugin->set_config_cb(plugin->cb_usr_ctx, &cfg);
+    enum set_config_result rc = plugin->set_config_cb(plugin->cb_usr_ctx, plugin->name, &cfg);
     if (rc != SET_CONFIG_ACCEPTED) {
         error_report("DYNCFG plugin \"%s\" rejected config", plugin->name);
         return "plugin rejected config";
     }
 
-    if (store_config(plugin->name, NULL, NULL, cfg)) {
-        error_report("DYNCFG could not store config for module \"%s\"", plugin->name);
-        return "could not store config on disk";
-    }
     return NULL;
 }
 
@@ -327,62 +442,38 @@ static char *set_module_config(struct module *mod, dyncfg_config_t cfg)
 {
     struct configurable_plugin *plugin = mod->plugin;
 
-    enum set_config_result rc = mod->set_config_cb(mod->config_cb_usr_ctx, mod->name, &cfg);
+    enum set_config_result rc = mod->set_config_cb(mod->config_cb_usr_ctx, plugin->name, mod->name, &cfg);
     if (rc != SET_CONFIG_ACCEPTED) {
         error_report("DYNCFG module \"%s\" rejected config", plugin->name);
         return "module rejected config";
     }
 
-    if (store_config(plugin->name, mod->name, NULL, cfg)) {
-        error_report("DYNCFG could not store config for module \"%s\"", mod->name);
-        return "could not store config on disk";
-    }
-
     return NULL;
 }
 
-struct job *job_new()
+struct job *job_new(const char *job_id)
 {
     struct job *job = callocz(1, sizeof(struct job));
     job->state = JOB_STATUS_UNKNOWN;
     job->last_state_update = now_realtime_usec();
+    job->name = strdupz(job_id);
+    netdata_mutex_init(&job->lock);
     return job;
 }
 
-static int set_job_config(struct job *job, dyncfg_config_t cfg)
+static inline void job_del(struct job *job)
 {
-    struct module *mod = job->module;
-    enum set_config_result rt = mod->set_job_config_cb(mod->job_config_cb_usr_ctx, mod->name, job->name, &cfg);
-
-    if (rt != SET_CONFIG_ACCEPTED) {
-        error_report("DYNCFG module \"%s\" rejected config for job \"%s\"", mod->name, job->name);
-        return 1;
-    }
-
-    if (store_config(mod->plugin->name, mod->name, job->name, cfg)) {
-        error_report("DYNCFG could not store config for module \"%s\"", mod->name);
-        return 1;
-    }
-
-    return 0;
+    netdata_mutex_destroy(&job->lock);
+    freez(job->reason);
+    freez((void*)job->name);
+    freez(job);
 }
 
-struct job *add_job(struct module *mod, const char *job_id, dyncfg_config_t cfg)
+void job_del_cb(const DICTIONARY_ITEM *item, void *value, void *data)
 {
-    struct job *job = job_new();
-    job->name = strdupz(job_id);
-    job->module = mod;
-
-    if (set_job_config(job, cfg)) {
-        freez(job->name);
-        freez(job);
-        return NULL;
-    }
-
-    dictionary_set(mod->jobs, job->name, job, sizeof(job));
-
-    return job;
-
+    UNUSED(item);
+    UNUSED(data);
+    job_del((struct job *)value);
 }
 
 void module_del_cb(const DICTIONARY_ITEM *item, void *value, void *data)
@@ -395,10 +486,9 @@ void module_del_cb(const DICTIONARY_ITEM *item, void *value, void *data)
     freez(mod);
 }
 
-
-const DICTIONARY_ITEM *register_plugin(struct configurable_plugin *plugin)
+const DICTIONARY_ITEM *register_plugin(DICTIONARY *plugins_dict, struct configurable_plugin *plugin, bool localhost)
 {
-    if (get_plugin_by_name(plugin->name) != NULL) {
+    if (get_plugin_by_name(plugins_dict, plugin->name) != NULL) {
         error_report("DYNCFG plugin \"%s\" already registered", plugin->name);
         return NULL;
     }
@@ -413,7 +503,8 @@ const DICTIONARY_ITEM *register_plugin(struct configurable_plugin *plugin)
     plugin->modules = dictionary_create(DICT_OPTION_VALUE_LINK_DONT_CLONE);
     dictionary_register_delete_callback(plugin->modules, module_del_cb, NULL);
 
-    deferred_config_push_back(plugin->name, NULL, NULL);
+    if (localhost)
+        deferred_config_push_back(plugins_dict, plugin->name, NULL, NULL);
 
     dictionary_set(plugins_dict, plugin->name, plugin, sizeof(plugin));
 
@@ -421,24 +512,14 @@ const DICTIONARY_ITEM *register_plugin(struct configurable_plugin *plugin)
     return dictionary_get_and_acquire_item(plugins_dict, plugin->name);
 }
 
-void unregister_plugin(const DICTIONARY_ITEM *plugin)
+void unregister_plugin(DICTIONARY *plugins_dict, const DICTIONARY_ITEM *plugin)
 {
     struct configurable_plugin *plug = dictionary_acquired_item_value(plugin);
     dictionary_acquired_item_release(plugins_dict, plugin);
     dictionary_del(plugins_dict, plug->name);
 }
 
-void job_del_cb(const DICTIONARY_ITEM *item, void *value, void *data)
-{
-    UNUSED(item);
-    UNUSED(data);
-    struct job *job = (struct job *)value;
-    freez(job->reason);
-    freez(job->name);
-    freez(job);
-}
-
-int register_module(struct configurable_plugin *plugin, struct module *module)
+int register_module(DICTIONARY *plugins_dict, struct configurable_plugin *plugin, struct module *module, bool localhost)
 {
     if (get_module_by_name(plugin, module->name) != NULL) {
         error_report("DYNCFG module \"%s\" already registered", module->name);
@@ -447,7 +528,8 @@ int register_module(struct configurable_plugin *plugin, struct module *module)
 
     pthread_mutex_init(&module->lock, NULL);
 
-    deferred_config_push_back(plugin->name, module->name, NULL);
+    if (localhost)
+        deferred_config_push_back(plugins_dict, plugin->name, module->name, NULL);
 
     module->plugin = plugin;
 
@@ -455,34 +537,38 @@ int register_module(struct configurable_plugin *plugin, struct module *module)
         module->jobs = dictionary_create(DICT_OPTION_VALUE_LINK_DONT_CLONE);
         dictionary_register_delete_callback(module->jobs, job_del_cb, NULL);
 
-        // load all jobs from disk
-        BUFFER *path = buffer_create(DYN_CONF_PATH_MAX, NULL);
-        buffer_sprintf(path, "%s/%s/%s", DYN_CONF_DIR, plugin->name, module->name);
-        DIR *dir = opendir(buffer_tostring(path));
-        if (dir != NULL) {
-            struct dirent *ent;
-            while ((ent = readdir(dir)) != NULL) {
-                if (ent->d_name[0] == '.')
-                    continue;
-                if (ent->d_type != DT_REG)
-                    continue;
-                size_t len = strnlen(ent->d_name, NAME_MAX);
-                if (len <= strlen(DYN_CONF_CFG_EXT))
-                    continue;
-                if (strcmp(ent->d_name + len - strlen(DYN_CONF_CFG_EXT), DYN_CONF_CFG_EXT) != 0)
-                    continue;
-                ent->d_name[len - strlen(DYN_CONF_CFG_EXT)] = '\0';
-
-                struct job *job = job_new();
-                job->name = strdupz(ent->d_name);
-                job->module = module;
-                dictionary_set(module->jobs, job->name, job, sizeof(job));
-
-                deferred_config_push_back(plugin->name, module->name, job->name);
+        if (localhost) {
+            // load all jobs from disk
+            BUFFER *path = buffer_create(DYN_CONF_PATH_MAX, NULL);
+            buffer_sprintf(path, "%s/%s/%s", DYN_CONF_DIR, plugin->name, module->name);
+            DIR *dir = opendir(buffer_tostring(path));
+            if (dir != NULL) {
+                struct dirent *ent;
+                while ((ent = readdir(dir)) != NULL) {
+                    if (ent->d_name[0] == '.')
+                        continue;
+                    if (ent->d_type != DT_REG)
+                        continue;
+                    size_t len = strnlen(ent->d_name, NAME_MAX);
+                    if (len <= strlen(DYN_CONF_CFG_EXT))
+                        continue;
+                    if (strcmp(ent->d_name + len - strlen(DYN_CONF_CFG_EXT), DYN_CONF_CFG_EXT) != 0)
+                        continue;
+                    ent->d_name[len - strlen(DYN_CONF_CFG_EXT)] = '\0';
+
+                    struct job *job = job_new(ent->d_name);
+                    job->module = module;
+                    job->flags = JOB_FLG_PS_LOADED;
+                    job->type = JOB_TYPE_USER;
+
+                    dictionary_set(module->jobs, job->name, job, sizeof(job));
+
+                    deferred_config_push_back(plugins_dict, plugin->name, module->name, ent->d_name);
+                }
+                closedir(dir);
             }
-            closedir(dir);
+            buffer_free(path);
         }
-        buffer_free(path);
     }
 
     dictionary_set(plugin->modules, module->name, module, sizeof(module));
@@ -490,11 +576,49 @@ int register_module(struct configurable_plugin *plugin, struct module *module)
     return 0;
 }
 
+int register_job(DICTIONARY *plugins_dict, const char *plugin_name, const char *module_name, const char *job_name, enum job_type job_type, dyncfg_job_flg_t flags, int ignore_existing)
+{
+    int rc = 1;
+    const DICTIONARY_ITEM *plugin_item = dictionary_get_and_acquire_item(plugins_dict, plugin_name);
+    if (plugin_item == NULL) {
+        error_report("plugin \"%s\" not registered", plugin_name);
+        return rc;
+    }
+    struct configurable_plugin *plugin = dictionary_acquired_item_value(plugin_item);
+    struct module *mod = get_module_by_name(plugin, module_name);
+    if (mod == NULL) {
+        error_report("module \"%s\" not registered", module_name);
+        goto ERR_EXIT;
+    }
+    if (mod->type != MOD_TYPE_ARRAY) {
+        error_report("module \"%s\" is not an array", module_name);
+        goto ERR_EXIT;
+    }
+    if (get_job_by_name(mod, job_name) != NULL) {
+        if (!ignore_existing)
+            error_report("job \"%s\" already registered", job_name);
+        goto ERR_EXIT;
+    }
+
+    struct job *job = job_new(job_name);
+    job->module = mod;
+    job->flags = flags;
+    job->type = job_type;
+
+    dictionary_set(mod->jobs, job->name, job, sizeof(job));
+
+    rc = 0;
+ERR_EXIT:
+    dictionary_acquired_item_release(plugins_dict, plugin_item);
+    return rc;
+}
+
 void freez_dyncfg(void *ptr) {
     freez(ptr);
 }
 
-static void handle_dyncfg_root(struct uni_http_response *resp, int method)
+#ifdef NETDATA_TEST_DYNCFG
+static void handle_dyncfg_root(DICTIONARY *plugins_dict, struct uni_http_response *resp, int method)
 {
     if (method != HTTP_METHOD_GET) {
         resp->content = "method not allowed";
@@ -502,7 +626,7 @@ static void handle_dyncfg_root(struct uni_http_response *resp, int method)
         resp->status = HTTP_RESP_METHOD_NOT_ALLOWED;
         return;
     }
-    json_object *obj = get_list_of_plugins_json();
+    json_object *obj = get_list_of_plugins_json(plugins_dict);
     json_object *wrapper = json_object_new_object();
     json_object_object_add(wrapper, "configurable_plugins", obj);
     resp->content = strdupz(json_object_to_json_string_ext(wrapper, JSON_C_TO_STRING_PRETTY));
@@ -518,7 +642,7 @@ static void handle_plugin_root(struct uni_http_response *resp, int method, struc
     switch(method) {
         case HTTP_METHOD_GET:
         {
-            dyncfg_config_t cfg = plugin->get_config_cb(plugin->cb_usr_ctx);
+            dyncfg_config_t cfg = plugin->get_config_cb(plugin->cb_usr_ctx, plugin->name);
             resp->content = mallocz(cfg.data_size);
             memcpy(resp->content, cfg.data, cfg.data_size);
             resp->status = HTTP_RESP_OK;
@@ -558,11 +682,12 @@ static void handle_plugin_root(struct uni_http_response *resp, int method, struc
             return;
     }
 }
+#endif
 
 void handle_module_root(struct uni_http_response *resp, int method, struct configurable_plugin *plugin, const char *module, void *post_payload, size_t post_payload_size)
 {
-    if (strncmp(module, DYN_CONF_SCHEMA, strlen(DYN_CONF_SCHEMA)) == 0) {
-        dyncfg_config_t cfg = plugin->get_config_schema_cb(plugin->cb_usr_ctx);
+    if (strncmp(module, DYN_CONF_SCHEMA, sizeof(DYN_CONF_SCHEMA)) == 0) {
+        dyncfg_config_t cfg = plugin->get_config_schema_cb(plugin->cb_usr_ctx, plugin->name);
         resp->content = mallocz(cfg.data_size);
         memcpy(resp->content, cfg.data, cfg.data_size);
         resp->status = HTTP_RESP_OK;
@@ -570,7 +695,7 @@ void handle_module_root(struct uni_http_response *resp, int method, struct confi
         resp->content_length = cfg.data_size;
         return;
     }
-    if (strncmp(module, DYN_CONF_MODULE_LIST, strlen(DYN_CONF_MODULE_LIST)) == 0) {
+    if (strncmp(module, DYN_CONF_MODULE_LIST, sizeof(DYN_CONF_MODULE_LIST)) == 0) {
         if (method != HTTP_METHOD_GET) {
             resp->content = "method not allowed (only GET)";
             resp->content_length = strlen(resp->content);
@@ -596,7 +721,7 @@ void handle_module_root(struct uni_http_response *resp, int method, struct confi
         return;
     }
     if (method == HTTP_METHOD_GET) {
-        dyncfg_config_t cfg = mod->get_config_cb(mod->config_cb_usr_ctx, mod->name);
+        dyncfg_config_t cfg = mod->get_config_cb(mod->config_cb_usr_ctx, plugin->name, mod->name);
         resp->content = mallocz(cfg.data_size);
         memcpy(resp->content, cfg.data, cfg.data_size);
         resp->status = HTTP_RESP_OK;
@@ -651,8 +776,7 @@ static inline void _handle_job_root(struct uni_http_response *resp, int method,
             .data = post_payload,
             .data_size = post_payload_size
         };
-        job = add_job(mod, job_id, cont);
-        if (job == NULL) {
+        if (mod->set_job_config_cb(mod->job_config_cb_usr_ctx, mod->plugin->name, mod->name, job_id, &cont)) {
             resp->content = "failed to add job";
             resp->content_length = strlen(resp->content);
             resp->status = HTTP_RESP_INTERNAL_SERVER_ERROR;
@@ -672,7 +796,7 @@ static inline void _handle_job_root(struct uni_http_response *resp, int method,
     switch (method) {
         case HTTP_METHOD_GET:
         {
-            dyncfg_config_t cfg = mod->get_job_config_cb(mod->job_config_cb_usr_ctx, mod->name, job->name);
+            dyncfg_config_t cfg = mod->get_job_config_cb(mod->job_config_cb_usr_ctx, mod->plugin->name, mod->name, job->name);
             resp->content = mallocz(cfg.data_size);
             memcpy(resp->content, cfg.data, cfg.data_size);
             resp->status = HTTP_RESP_OK;
@@ -692,10 +816,11 @@ static inline void _handle_job_root(struct uni_http_response *resp, int method,
                 .data = post_payload,
                 .data_size = post_payload_size
             };
-            if(set_job_config(job, cont)) {
-                resp->status = HTTP_RESP_BAD_REQUEST;
+            if (mod->set_job_config_cb(mod->job_config_cb_usr_ctx, mod->plugin->name, mod->name, job->name, &cont) != SET_CONFIG_ACCEPTED) {
+                error_report("DYNCFG module \"%s\" rejected config for job \"%s\"", mod->name, job->name);
                 resp->content = "failed to set job config";
                 resp->content_length = strlen(resp->content);
+                resp->status = HTTP_RESP_INTERNAL_SERVER_ERROR;
                 return;
             }
             resp->status = HTTP_RESP_OK;
@@ -726,8 +851,8 @@ static inline void _handle_job_root(struct uni_http_response *resp, int method,
 
 void handle_job_root(struct uni_http_response *resp, int method, struct module *mod, const char *job_id, void *post_payload, size_t post_payload_size)
 {
-    if (strncmp(job_id, DYN_CONF_SCHEMA, strlen(DYN_CONF_SCHEMA)) == 0) {
-        dyncfg_config_t cfg = mod->get_config_schema_cb(mod->config_cb_usr_ctx, mod->name);
+    if (strncmp(job_id, DYN_CONF_SCHEMA, sizeof(DYN_CONF_SCHEMA)) == 0) {
+        dyncfg_config_t cfg = mod->get_config_schema_cb(mod->config_cb_usr_ctx, mod->plugin->name, mod->name);
         resp->content = mallocz(cfg.data_size);
         memcpy(resp->content, cfg.data, cfg.data_size);
         resp->status = HTTP_RESP_OK;
@@ -735,8 +860,8 @@ void handle_job_root(struct uni_http_response *resp, int method, struct module *
         resp->content_length = cfg.data_size;
         return;
     }
-    if (strncmp(job_id, DYN_CONF_JOB_SCHEMA, strlen(DYN_CONF_JOB_SCHEMA)) == 0) {
-        dyncfg_config_t cfg = mod->get_job_config_schema_cb(mod->job_config_cb_usr_ctx, mod->name);
+    if (strncmp(job_id, DYN_CONF_JOB_SCHEMA, sizeof(DYN_CONF_JOB_SCHEMA)) == 0) {
+        dyncfg_config_t cfg = mod->get_job_config_schema_cb(mod->job_config_cb_usr_ctx, mod->plugin->name, mod->name);
         resp->content = mallocz(cfg.data_size);
         memcpy(resp->content, cfg.data, cfg.data_size);
         resp->status = HTTP_RESP_OK;
@@ -744,7 +869,7 @@ void handle_job_root(struct uni_http_response *resp, int method, struct module *
         resp->content_length = cfg.data_size;
         return;
     }
-    if (strncmp(job_id, DYN_CONF_JOB_LIST, strlen(DYN_CONF_JOB_LIST)) == 0) {
+    if (strncmp(job_id, DYN_CONF_JOB_LIST, sizeof(DYN_CONF_JOB_LIST)) == 0) {
         if (mod->type != MOD_TYPE_ARRAY) {
             resp->content = "module type is not job_array (can't get the list of jobs)";
             resp->content_length = strlen(resp->content);
@@ -776,7 +901,14 @@ void handle_job_root(struct uni_http_response *resp, int method, struct module *
     dictionary_acquired_item_release(mod->jobs, job_item);
 }
 
-struct uni_http_response dyn_conf_process_http_request(int method, const char *plugin, const char *module, const char *job_id, void *post_payload, size_t post_payload_size)
+struct uni_http_response dyn_conf_process_http_request(
+    DICTIONARY *plugins_dict __maybe_unused,
+    int method __maybe_unused,
+    const char *plugin __maybe_unused,
+    const char *module __maybe_unused,
+    const char *job_id __maybe_unused,
+    void *post_payload __maybe_unused,
+    size_t post_payload_size __maybe_unused)
 {
     struct uni_http_response resp = {
         .status = HTTP_RESP_INTERNAL_SERVER_ERROR,
@@ -785,8 +917,14 @@ struct uni_http_response dyn_conf_process_http_request(int method, const char *p
         .content_free = NULL,
         .content_length = 0
     };
+#ifndef NETDATA_TEST_DYNCFG
+    resp.content = "DYNCFG is disabled (as it is for now developer only feature). This will be enabled by default when ready for technical preview.";
+    resp.content_length = strlen(resp.content);
+    resp.status = HTTP_RESP_PRECOND_FAIL;
+    return resp;
+#else
     if (plugin == NULL) {
-        handle_dyncfg_root(&resp, method);
+        handle_dyncfg_root(plugins_dict, &resp, method);
         return resp;
     }
     const DICTIONARY_ITEM *plugin_item = dictionary_get_and_acquire_item(plugins_dict, plugin);
@@ -814,9 +952,9 @@ struct uni_http_response dyn_conf_process_http_request(int method, const char *p
         goto EXIT_PLUGIN;
     }
     if (mod->type != MOD_TYPE_ARRAY) {
-        resp.content = "module is not array";
+        resp.content = "400 - this module is not array type";
         resp.content_length = strlen(resp.content);
-        resp.status = HTTP_RESP_NOT_FOUND;
+        resp.status = HTTP_RESP_BAD_REQUEST;
         goto EXIT_PLUGIN;
     }
     handle_job_root(&resp, method, mod, job_id, post_payload, post_payload_size);
@@ -824,6 +962,7 @@ struct uni_http_response dyn_conf_process_http_request(int method, const char *p
 EXIT_PLUGIN:
     dictionary_acquired_item_release(plugins_dict, plugin_item);
     return resp;
+#endif
 }
 
 void plugin_del_cb(const DICTIONARY_ITEM *item, void *value, void *data)
@@ -836,41 +975,53 @@ void plugin_del_cb(const DICTIONARY_ITEM *item, void *value, void *data)
     freez(plugin);
 }
 
-void report_job_status(struct configurable_plugin *plugin, const char *module_name, const char *job_name, enum job_status status, int status_code, char *reason)
+// on failure - return NULL - all unlocked, nothing acquired
+// on success - return pointer to job item - keep job and plugin acquired and locked!!!
+//              for caller convenience (to prevent another lock and races)
+//              caller is responsible to unlock the job and release it when not needed anymore
+//              this also avoids dependency creep
+const DICTIONARY_ITEM *report_job_status_acq_lock(DICTIONARY *plugins_dict, const DICTIONARY_ITEM **plugin_acq_item, DICTIONARY **job_dict, const char *plugin_name, const char *module_name, const char *job_name, enum job_status status, int status_code, char *reason)
 {
-    const DICTIONARY_ITEM *item = dictionary_get_and_acquire_item(plugins_dict, plugin->name);
-    if (item == NULL) {
-        netdata_log_error("plugin %s not found", plugin->name);
-        return;
+    *plugin_acq_item = dictionary_get_and_acquire_item(plugins_dict, plugin_name);
+    if (*plugin_acq_item == NULL) {
+        netdata_log_error("plugin %s not found", plugin_name);
+        return NULL;
     }
-    struct configurable_plugin *plug = dictionary_acquired_item_value(item);
+
+    struct configurable_plugin *plug = dictionary_acquired_item_value(*plugin_acq_item);
     struct module *mod = get_module_by_name(plug, module_name);
     if (mod == NULL) {
         netdata_log_error("module %s not found", module_name);
-        goto EXIT_PLUGIN;
+        dictionary_acquired_item_release(plugins_dict, *plugin_acq_item);
+        return NULL;
     }
     if (mod->type != MOD_TYPE_ARRAY) {
         netdata_log_error("module %s is not array", module_name);
-        goto EXIT_PLUGIN;
+        dictionary_acquired_item_release(plugins_dict, *plugin_acq_item);
+        return NULL;
     }
+    *job_dict = mod->jobs;
     const DICTIONARY_ITEM *job_item = dictionary_get_and_acquire_item(mod->jobs, job_name);
     if (job_item == NULL) {
         netdata_log_error("job %s not found", job_name);
-        goto EXIT_PLUGIN;
+        dictionary_acquired_item_release(plugins_dict, *plugin_acq_item);
+        return NULL;
     }
     struct job *job = dictionary_acquired_item_value(job_item);
+
+    pthread_mutex_lock(&job->lock);
     job->status = status;
     job->state = status_code;
     if (job->reason != NULL) {
         freez(job->reason);
     }
-    job->reason = reason;
+    job->reason = reason != NULL ? strdupz(reason) : NULL; // reason is optional
     job->last_state_update = now_realtime_usec();
 
-    dictionary_acquired_item_release(mod->jobs, job_item);
+    job->dirty = true;
 
-EXIT_PLUGIN:
-    dictionary_acquired_item_release(plugins_dict, item);
+    // no unlock and acquired_item_release on success on purpose
+    return job_item;
 }
 
 int dyn_conf_init(void)
@@ -882,9 +1033,6 @@ int dyn_conf_init(void)
         }
     }
 
-    plugins_dict = dictionary_create(DICT_OPTION_VALUE_LINK_DONT_CLONE);
-    dictionary_register_delete_callback(plugins_dict, plugin_del_cb, NULL);
-
     return 0;
 }
 
@@ -912,6 +1060,15 @@ void *dyncfg_main(void *ptr)
 
     while (!netdata_exit) {
         struct deferred_cfg_send *dcs = deferred_config_pop(ptr);
+        DICTIONARY *plugins_dict = dcs->plugins_dict;
+#ifdef NETDATA_INTERNAL_CHECKS
+        if (plugins_dict == NULL) {
+            fatal("DYNCFG, plugins_dict is NULL");
+            deferred_config_free(dcs);
+            continue;
+        }
+#endif
+
         const DICTIONARY_ITEM *plugin_item = dictionary_get_and_acquire_item(plugins_dict, dcs->plugin_name);
         if (plugin_item == NULL) {
             error_report("DYNCFG, plugin %s not found", dcs->plugin_name);
@@ -922,21 +1079,21 @@ void *dyncfg_main(void *ptr)
         if (dcs->module_name == NULL) {
             dyncfg_config_t cfg = load_config(dcs->plugin_name, NULL, NULL);
             if (cfg.data != NULL) {
-                plugin->set_config_cb(plugin->cb_usr_ctx, &cfg);
+                plugin->set_config_cb(plugin->cb_usr_ctx, plugin->name, &cfg);
                 freez(cfg.data);
             }
         } else if (dcs->job_name == NULL) {
             dyncfg_config_t cfg = load_config(dcs->plugin_name, dcs->module_name, NULL);
             if (cfg.data != NULL) {
                 struct module *mod = get_module_by_name(plugin, dcs->module_name);
-                mod->set_config_cb(mod->config_cb_usr_ctx, mod->name, &cfg);
+                mod->set_config_cb(mod->config_cb_usr_ctx, plugin->name, mod->name, &cfg);
                 freez(cfg.data);
             }
         } else {
             dyncfg_config_t cfg = load_config(dcs->plugin_name, dcs->module_name, dcs->job_name);
             if (cfg.data != NULL) {
                 struct module *mod = get_module_by_name(plugin, dcs->module_name);
-                mod->set_job_config_cb(mod->job_config_cb_usr_ctx, mod->name, dcs->job_name, &cfg);
+                mod->set_job_config_cb(mod->job_config_cb_usr_ctx, plugin->name, mod->name, dcs->job_name, &cfg);
                 freez(cfg.data);
             }
         }
@@ -947,3 +1104,37 @@ void *dyncfg_main(void *ptr)
     netdata_thread_cleanup_pop(1);
     return NULL;
 }
+
+bool is_dyncfg_function(const char *function_name, uint8_t type) {
+    // TODO add hash to speed things up
+    if (type & (DYNCFG_FUNCTION_TYPE_GET | DYNCFG_FUNCTION_TYPE_REGULAR)) {
+        if (strncmp(function_name, FUNCTION_NAME_GET_PLUGIN_CONFIG, strlen(FUNCTION_NAME_GET_PLUGIN_CONFIG)) == 0)
+            return true;
+        if (strncmp(function_name, FUNCTION_NAME_GET_PLUGIN_CONFIG_SCHEMA, strlen(FUNCTION_NAME_GET_PLUGIN_CONFIG_SCHEMA)) == 0)
+            return true;
+        if (strncmp(function_name, FUNCTION_NAME_GET_MODULE_CONFIG, strlen(FUNCTION_NAME_GET_MODULE_CONFIG)) == 0)
+            return true;
+        if (strncmp(function_name, FUNCTION_NAME_GET_MODULE_CONFIG_SCHEMA, strlen(FUNCTION_NAME_GET_MODULE_CONFIG_SCHEMA)) == 0)
+            return true;
+        if (strncmp(function_name, FUNCTION_NAME_GET_JOB_CONFIG, strlen(FUNCTION_NAME_GET_JOB_CONFIG)) == 0)
+            return true;
+        if (strncmp(function_name, FUNCTION_NAME_GET_JOB_CONFIG_SCHEMA, strlen(FUNCTION_NAME_GET_JOB_CONFIG_SCHEMA)) == 0)
+            return true;
+    }
+
+    if (type & (DYNCFG_FUNCTION_TYPE_SET | DYNCFG_FUNCTION_TYPE_PAYLOAD)) {
+        if (strncmp(function_name, FUNCTION_NAME_SET_PLUGIN_CONFIG, strlen(FUNCTION_NAME_SET_PLUGIN_CONFIG)) == 0)
+            return true;
+        if (strncmp(function_name, FUNCTION_NAME_SET_MODULE_CONFIG, strlen(FUNCTION_NAME_SET_MODULE_CONFIG)) == 0)
+            return true;
+        if (strncmp(function_name, FUNCTION_NAME_SET_JOB_CONFIG, strlen(FUNCTION_NAME_SET_JOB_CONFIG)) == 0)
+            return true;
+    }
+
+    if (type & (DYNCFG_FUNCTION_TYPE_DELETE | DYNCFG_FUNCTION_TYPE_REGULAR)) {
+        if (strncmp(function_name, FUNCTION_NAME_DELETE_JOB, strlen(FUNCTION_NAME_DELETE_JOB)) == 0)
+            return true;
+    }
+
+    return false;
+}
diff --git a/libnetdata/dyn_conf/dyn_conf.h b/libnetdata/dyn_conf/dyn_conf.h
index f10ae6a1..f6a5fe49 100644
--- a/libnetdata/dyn_conf/dyn_conf.h
+++ b/libnetdata/dyn_conf/dyn_conf.h
@@ -5,6 +5,21 @@
 
 #include "../libnetdata.h"
 
+#define FUNCTION_NAME_GET_PLUGIN_CONFIG "get_plugin_config"
+#define FUNCTION_NAME_GET_PLUGIN_CONFIG_SCHEMA "get_plugin_config_schema"
+#define FUNCTION_NAME_GET_MODULE_CONFIG "get_module_config"
+#define FUNCTION_NAME_GET_MODULE_CONFIG_SCHEMA "get_module_config_schema"
+#define FUNCTION_NAME_GET_JOB_CONFIG "get_job_config"
+#define FUNCTION_NAME_GET_JOB_CONFIG_SCHEMA "get_job_config_schema"
+#define FUNCTION_NAME_SET_PLUGIN_CONFIG "set_plugin_config"
+#define FUNCTION_NAME_SET_MODULE_CONFIG "set_module_config"
+#define FUNCTION_NAME_SET_JOB_CONFIG "set_job_config"
+#define FUNCTION_NAME_DELETE_JOB "delete_job"
+
+#define DYNCFG_MAX_WORDS 5
+
+#define DYNCFG_VFNC_RET_CFG_ACCEPTED (1)
+
 enum module_type {
     MOD_TYPE_UNKNOWN = 0,
     MOD_TYPE_ARRAY,
@@ -20,6 +35,18 @@ static inline enum module_type str2_module_type(const char *type_name)
     return MOD_TYPE_UNKNOWN;
 }
 
+static inline const char *module_type2str(enum module_type type)
+{
+    switch (type) {
+        case MOD_TYPE_ARRAY:
+            return "job_array";
+        case MOD_TYPE_SINGLE:
+            return "single";
+        default:
+            return "unknown";
+    }
+}
+
 struct dyncfg_config {
     void *data;
     size_t data_size;
@@ -37,7 +64,7 @@ enum job_status {
     JOB_STATUS_ERROR
 };
 
-inline enum job_status str2job_state(const char *state_name) {
+static inline enum job_status str2job_state(const char *state_name) {
     if (strcmp(state_name, "stopped") == 0)
         return JOB_STATUS_STOPPED;
     else if (strcmp(state_name, "running") == 0)
@@ -47,24 +74,75 @@ inline enum job_status str2job_state(const char *state_name) {
     return JOB_STATUS_UNKNOWN;
 }
 
+const char *job_status2str(enum job_status status);
+
 enum set_config_result {
     SET_CONFIG_ACCEPTED = 0,
     SET_CONFIG_REJECTED,
     SET_CONFIG_DEFFER
 };
 
+typedef uint32_t dyncfg_job_flg_t;
+enum job_flags {
+    JOB_FLG_PS_LOADED        = 1 << 0, // PS abbr. Persistent Storage
+    JOB_FLG_PLUGIN_PUSHED    = 1 << 1, // got it from plugin (e.g. autodiscovered job)
+    JOB_FLG_STREAMING_PUSHED = 1 << 2, // got it through streaming
+    JOB_FLG_USER_CREATED     = 1 << 3, // user created this job during agent runtime
+};
+
+enum job_type {
+    JOB_TYPE_UNKNOWN = 0,
+    JOB_TYPE_STOCK = 1,
+    JOB_TYPE_USER = 2,
+    JOB_TYPE_AUTODISCOVERED = 3,
+};
+
+static inline const char* job_type2str(enum job_type type)
+{
+    switch (type) {
+        case JOB_TYPE_STOCK:
+            return "stock";
+        case JOB_TYPE_USER:
+            return "user";
+        case JOB_TYPE_AUTODISCOVERED:
+            return "autodiscovered";
+        case JOB_TYPE_UNKNOWN:
+        default:
+            return "unknown";
+    }
+}
+
+static inline enum job_type str2job_type(const char *type_name)
+{
+    if (strcmp(type_name, "stock") == 0)
+        return JOB_TYPE_STOCK;
+    else if (strcmp(type_name, "user") == 0)
+        return JOB_TYPE_USER;
+    else if (strcmp(type_name, "autodiscovered") == 0)
+        return JOB_TYPE_AUTODISCOVERED;
+    error_report("Unknown job type: %s", type_name);
+    return JOB_TYPE_UNKNOWN;
+}
+
 struct job
 {
-    char *name;
+    const char *name;
+    enum job_type type;
+    struct module *module;
+
+    pthread_mutex_t lock;
+    // lock protexts only fields below (which are modified during job existence)
+    // others are static during lifetime of job
 
-    //state reported by config
+    int dirty; // this relates to rrdpush, true if parent has different data than us
+
+    // state reported by plugin
+    usec_t last_state_update;
     enum job_status status; // reported by plugin, enum as this has to be interpreted by UI
     int state; // code reported by plugin which can mean anything plugin wants
     char *reason; // reported by plugin, can be NULL (optional)
 
-    usec_t last_state_update;
-
-    struct module *module;
+    dyncfg_job_flg_t flags;
 };
 
 struct module
@@ -76,18 +154,18 @@ struct module
     struct configurable_plugin *plugin;
 
     // module config
-    enum set_config_result (*set_config_cb)(void *usr_ctx, const char *module_name, dyncfg_config_t *cfg);
-    dyncfg_config_t (*get_config_cb)(void *usr_ctx, const char *name);
-    dyncfg_config_t (*get_config_schema_cb)(void *usr_ctx, const char *name);
+    enum set_config_result (*set_config_cb)(void *usr_ctx, const char *plugin_name, const char *module_name, dyncfg_config_t *cfg);
+    dyncfg_config_t (*get_config_cb)(void *usr_ctx, const char *plugin_name, const char *module_name);
+    dyncfg_config_t (*get_config_schema_cb)(void *usr_ctx, const char *plugin_name, const char *module_name);
     void *config_cb_usr_ctx;
 
     DICTIONARY *jobs;
 
     // jobs config
-    dyncfg_config_t (*get_job_config_cb)(void *usr_ctx, const char *module_name, const char *job_name);
-    dyncfg_config_t (*get_job_config_schema_cb)(void *usr_ctx, const char *module_name);
-    enum set_config_result (*set_job_config_cb)(void *usr_ctx, const char *module_name, const char *job_name, dyncfg_config_t *cfg);
-    enum set_config_result (*delete_job_cb)(void *usr_ctx, const char *module_name, const char *job_name);
+    dyncfg_config_t (*get_job_config_cb)(void *usr_ctx, const char *plugin_name, const char *module_name, const char *job_name);
+    dyncfg_config_t (*get_job_config_schema_cb)(void *usr_ctx, const char *plugin_name, const char *module_name);
+    enum set_config_result (*set_job_config_cb)(void *usr_ctx, const char *plugin_name, const char *module_name, const char *job_name, dyncfg_config_t *cfg);
+    enum set_config_result (*delete_job_cb)(void *usr_ctx, const char *plugin_name, const char *module_name, const char *job_name);
     void *job_config_cb_usr_ctx;
 };
 
@@ -97,26 +175,34 @@ struct configurable_plugin {
     DICTIONARY *modules;
     const char *schema;
 
-    dyncfg_config_t (*get_config_cb)(void *usr_ctx);
-    dyncfg_config_t (*get_config_schema_cb)(void *usr_ctx);
-    enum set_config_result (*set_config_cb)(void *usr_ctx, dyncfg_config_t *cfg);
+    dyncfg_config_t (*get_config_cb)(void *usr_ctx, const char *plugin_name);
+    dyncfg_config_t (*get_config_schema_cb)(void *usr_ctx, const char *plugin_name);
+    enum set_config_result (*set_config_cb)(void *usr_ctx, const char *plugin_name, dyncfg_config_t *cfg);
     void *cb_usr_ctx; // context for all callbacks (split if needed in future)
 };
 
 // API to be used by plugins
-const DICTIONARY_ITEM *register_plugin(struct configurable_plugin *plugin);
-void unregister_plugin(const DICTIONARY_ITEM *plugin);
-int register_module(struct configurable_plugin *plugin, struct module *module);
+const DICTIONARY_ITEM *register_plugin(DICTIONARY *plugins_dict, struct configurable_plugin *plugin, bool localhost);
+void unregister_plugin(DICTIONARY *plugins_dict, const DICTIONARY_ITEM *plugin);
+int register_module(DICTIONARY *plugins_dict, struct configurable_plugin *plugin, struct module *module, bool localhost);
+int register_job(DICTIONARY *plugins_dict, const char *plugin_name, const char *module_name, const char *job_name, enum job_type job_type, dyncfg_job_flg_t flags, int ignore_existing);
+
+const DICTIONARY_ITEM *report_job_status_acq_lock(DICTIONARY *plugins_dict, const DICTIONARY_ITEM **plugin_acq_item, DICTIONARY **job_dict, const char *plugin_name, const char *module_name, const char *job_name, enum job_status status, int status_code, char *reason);
 
-void report_job_status(struct configurable_plugin *plugin, const char *module_name, const char *job_name, enum job_status status, int status_code, char *reason);
+void dyn_conf_store_config(const char *function, const char *payload, struct configurable_plugin *plugin);
+void unlink_job(const char *plugin_name, const char *module_name, const char *job_name);
+void delete_job(struct configurable_plugin *plugin, const char *module_name, const char *job_name);
+void delete_job_pname(DICTIONARY *plugins_dict, const char *plugin_name, const char *module_name, const char *job_name);
 
 // API to be used by the web server(s)
-json_object *get_list_of_plugins_json();
-struct configurable_plugin *get_plugin_by_name(const char *name);
+json_object *get_list_of_plugins_json(DICTIONARY *plugins_dict);
+struct configurable_plugin *get_plugin_by_name(DICTIONARY *plugins_dict, const char *name);
 
 json_object *get_list_of_modules_json(struct configurable_plugin *plugin);
 struct module *get_module_by_name(struct configurable_plugin *plugin, const char *module_name);
 
+json_object *job2json(struct job *job);
+
 // helper struct to make interface between internal webserver and h2o same
 struct uni_http_response {
     int status;
@@ -126,11 +212,26 @@ struct uni_http_response {
     void (*content_free)(void *);
 };
 
-struct uni_http_response dyn_conf_process_http_request(int method, const char *plugin, const char *module, const char *job_id, void *payload, size_t payload_size);
+struct uni_http_response dyn_conf_process_http_request(DICTIONARY *plugins_dict, int method, const char *plugin, const char *module, const char *job_id, void *payload, size_t payload_size);
 
 // API to be used by main netdata process, initialization and destruction etc.
 int dyn_conf_init(void);
 void freez_dyncfg(void *ptr);
+
+#define dyncfg_dictionary_create() dictionary_create(DICT_OPTION_VALUE_LINK_DONT_CLONE)
+
+void plugin_del_cb(const DICTIONARY_ITEM *item, void *value, void *data);
+
 void *dyncfg_main(void *in);
 
+#define DYNCFG_FUNCTION_TYPE_REGULAR (1 << 0)
+#define DYNCFG_FUNCTION_TYPE_PAYLOAD (1 << 1)
+#define DYNCFG_FUNCTION_TYPE_GET     (1 << 2)
+#define DYNCFG_FUNCTION_TYPE_SET     (1 << 3)
+#define DYNCFG_FUNCTION_TYPE_DELETE  (1 << 4)
+#define DYNCFG_FUNCTION_TYPE_ALL \
+    (DYNCFG_FUNCTION_TYPE_REGULAR | DYNCFG_FUNCTION_TYPE_PAYLOAD | DYNCFG_FUNCTION_TYPE_GET | DYNCFG_FUNCTION_TYPE_SET | DYNCFG_FUNCTION_TYPE_DELETE)
+
+bool is_dyncfg_function(const char *function_name, uint8_t type);
+
 #endif //DYN_CONF_H
diff --git a/libnetdata/dyn_conf/tests/sample_test_config.json b/libnetdata/dyn_conf/tests/sample_test_config.json
new file mode 100644
index 00000000..a6595f12
--- /dev/null
+++ b/libnetdata/dyn_conf/tests/sample_test_config.json
@@ -0,0 +1,22 @@
+{
+    "http_endpoints": {
+        "parent": {
+            "host": "127.0.0.1",
+            "mguid": null,
+            "port": 20001,
+            "ssl": false
+        },
+        "child": {
+            "host": "127.0.0.1",
+            "mguid": "3bc2f7de-1445-11ee-9ed7-3c7c3f21784c",
+            "port": 19999,
+            "ssl": false
+        }
+    },
+    "global": {
+        "test_plugin_name": "external_plugin",
+        "test_array_module_name": "module_of_the_future",
+        "test_single_module_name": "module_of_the_future_single_type",
+        "test_job_name": "fixed_job"
+    }
+}
diff --git a/libnetdata/dyn_conf/tests/sub_tests/test_parent_child.rb b/libnetdata/dyn_conf/tests/sub_tests/test_parent_child.rb
new file mode 100644
index 00000000..820db77f
--- /dev/null
+++ b/libnetdata/dyn_conf/tests/sub_tests/test_parent_child.rb
@@ -0,0 +1,192 @@
+class ParentChildTest
+    @@plugin_cfg = <<~HEREDOC
+{ "test" : "true" }
+HEREDOC
+    @@plugin_cfg2 = <<~HEREDOC
+{ "asdfgh" : "asdfgh" }
+HEREDOC
+
+    @@job_cfg = <<~HEREDOC
+{ "i am newly created job" : "true" }
+HEREDOC
+
+    def initialize
+        @parent = $config[:http_endpoints][:parent]
+        @child  = $config[:http_endpoints][:child]
+        @plugin = $config[:global][:test_plugin_name]
+        @arry_mod = $config[:global][:test_array_module_name]
+        @single_mod = $config[:global][:test_single_module_name]
+        @test_job = $config[:global][:test_job_name]
+    end
+    def check_test_plugin_modules_list(host, child = nil)
+        rc = DynCfgHttpClient.get_plugin_module_list(host, @plugin, child)
+        assert_eq(rc.code, 200, "as HTTP code for get_module_list request on plugin \"#{@plugin}\"")
+        modules = nil
+        assert_nothing_raised do
+            modules = JSON.parse(rc.parsed_response, symbolize_names: true)
+        end
+        assert_has_key?(modules, :modules)
+        assert_eq(modules[:modules].count, 2, "as number of modules in plugin \"#{@plugin}\"")
+        modules[:modules].each do |m|
+            assert_has_key?(m, :name)
+            assert_has_key?(m, :type)
+            assert_is_one_of(m[:type], "job_array", "single")
+        end
+        assert_eq_str(modules[:modules][0][:name], @arry_mod, "name of first module in plugin \"#{@plugin}\"")
+        assert_eq_str(modules[:modules][1][:name], @single_mod, "name of second module in plugin \"#{@plugin}\"")
+    end
+    def run
+        TEST_SUITE("Parent/Child plugin config")
+
+        TEST("parent/child/get_plugin_list", "Get child (hops:1) plugin list trough parent")
+        plugins = DynCfgHttpClient.get_plugin_list(@parent, @child)
+        assert_eq(plugins.code, 200, "as HTTP code for get_plugin_list request")
+        assert_nothing_raised do
+            plugins = JSON.parse(plugins.parsed_response, symbolize_names: true)
+        end
+        assert_has_key?(plugins, :configurable_plugins)
+        assert_array_include?(plugins[:configurable_plugins], @plugin)
+        PASS()
+
+        TEST("parent/child/(set/get)plugin_config", "Set then get and compare child (hops:1) plugin config trough parent")
+        rc = DynCfgHttpClient.set_plugin_config(@parent, @plugin, @@plugin_cfg, @child)
+        assert_eq(rc.code, 200, "as HTTP code for set_plugin_config request")
+
+        rc = DynCfgHttpClient.get_plugin_config(@parent, @plugin, @child)
+        assert_eq(rc.code, 200, "as HTTP code for get_plugin_config request")
+        assert_eq_str(rc.parsed_response.chomp!, @@plugin_cfg, "as plugin config")
+
+        # We do this twice with different configs to ensure first config was not loaded from persistent storage (from previous tests)
+        rc = DynCfgHttpClient.set_plugin_config(@parent, @plugin, @@plugin_cfg2, @child)
+        assert_eq(rc.code, 200, "as HTTP code for set_plugin_config request 2")
+
+        rc = DynCfgHttpClient.get_plugin_config(@parent, @plugin, @child)
+        assert_eq(rc.code, 200, "as HTTP code for get_plugin_config request 2")
+        assert_eq_str(rc.parsed_response.chomp!, @@plugin_cfg2, "set/get plugin config 2")
+        PASS()
+
+        TEST("child/get_plugin_config", "Get child (hops:0) plugin config and compare with what we got trough parent (set_plugin_config from previous test)")
+        rc = DynCfgHttpClient.get_plugin_config(@child, @plugin, nil)
+        assert_eq(rc.code, 200, "as HTTP code for get_plugin_config request")
+        assert_eq_str(rc.parsed_response.chomp!, @@plugin_cfg2.chomp, "as plugin config")
+        PASS()
+
+        TEST("parent/child/plugin_module_list", "Get child (hops:1) plugin module list trough parent and check its contents")
+        check_test_plugin_modules_list(@parent, @child)
+        PASS()
+
+        TEST("child/plugin_module_list", "Get child (hops:0) plugin module list directly and check its contents")
+        check_test_plugin_modules_list(@child, nil)
+        PASS()
+
+        TEST("parent/child/module/jobs", "Get list of jobs from child (hops:1) trough parent and check its contents, check job updates")
+        rc = DynCfgHttpClient.get_job_list(@parent, @plugin, @arry_mod, @child)
+        assert_eq(rc.code, 200, "as HTTP code for get_jobs request")
+        jobs = nil
+        assert_nothing_raised do
+            jobs = JSON.parse(rc.parsed_response, symbolize_names: true)
+        end
+        assert_has_key?(jobs, :jobs)
+        new_job = jobs[:jobs].find {|i| i[:name] == @test_job}
+        assert_not_nil(new_job)
+        assert_has_key?(new_job, :status)
+        assert_not_eq_str(new_job[:status], "unknown", "job status is other than unknown")
+        assert_has_key?(new_job, :flags)
+        assert_array_include?(new_job[:flags], "JOB_FLG_STREAMING_PUSHED")
+        PASS()
+
+        TEST("child/module/jobs", "Get list of jobs direct from child (hops:0) and check its contents, check job updates")
+        rc = DynCfgHttpClient.get_job_list(@child, @plugin, @arry_mod, nil)
+        assert_eq(rc.code, 200, "as HTTP code for get_jobs request")
+        jobs = nil
+        assert_nothing_raised do
+            jobs = JSON.parse(rc.parsed_response, symbolize_names: true)
+        end
+        assert_has_key?(jobs, :jobs)
+        new_job = jobs[:jobs].find {|i| i[:name] == @test_job}
+        assert_not_nil(new_job)
+        assert_has_key?(new_job, :status)
+        assert_not_eq_str(new_job[:status], "unknown", "job status is other than unknown")
+        assert_has_key?(new_job, :flags)
+
+        assert_array_not_include?(new_job[:flags], "JOB_FLG_STREAMING_PUSHED") # this is plugin directly at child so it should not show this flag
+        PASS()
+
+        TEST("parent/child/single_module/jobs", "Attempt getting list of jobs from child (hops:1) trough parent on single module. Check it fails properly")
+        rc = DynCfgHttpClient.get_job_list(@parent, @plugin, @single_mod, @child)
+        assert_eq(rc.code, 400, "as HTTP code for get_jobs request")
+        assert_eq_str(rc.parsed_response, '400 - this module is not array type', "as HTTP code for get_jobs request on single module")
+        PASS()
+
+        created_job = SecureRandom.uuid
+        TEST("parent/child/module/cr_del_job", "Create and delete job on child (hops:1) trough parent")
+        # create new job
+        rc = DynCfgHttpClient.create_job(@parent, @plugin, @arry_mod, created_job, @@job_cfg, @child)
+        assert_eq_http_code(rc, 200, "as HTTP code for create_job request")
+        # check this job is in job list @parent
+        rc = DynCfgHttpClient.get_job_list(@parent, @plugin, @arry_mod, @child)
+        assert_eq(rc.code, 200, "as HTTP code for get_jobs request")
+        jobs = nil
+        assert_nothing_raised do
+            jobs = JSON.parse(rc.parsed_response, symbolize_names: true)
+        end
+        assert_has_key?(jobs, :jobs)
+        new_job = jobs[:jobs].find {|i| i[:name] == created_job}
+        assert_not_nil(new_job)
+        # check this job is in job list @child
+        rc = DynCfgHttpClient.get_job_list(@child, @plugin, @arry_mod, nil)
+        assert_eq(rc.code, 200, "as HTTP code for get_jobs request")
+        jobs = nil
+        assert_nothing_raised do
+            jobs = JSON.parse(rc.parsed_response, symbolize_names: true)
+        end
+        assert_has_key?(jobs, :jobs)
+        new_job = jobs[:jobs].find {|i| i[:name] == created_job}
+        assert_not_nil(new_job)
+        # check we can get job config back
+        rc = DynCfgHttpClient.get_job_config(@parent, @plugin, @arry_mod, created_job, @child)
+        assert_eq(rc.code, 200, "as HTTP code for get_job_config request")
+        assert_eq_str(rc.parsed_response.chomp!, @@job_cfg, "as job config")
+        # delete job
+        rc = DynCfgHttpClient.delete_job(@parent, @plugin, @arry_mod, created_job, @child)
+        assert_eq(rc.code, 200, "as HTTP code for delete_job request")
+        # Check it is not in parents job list anymore
+        rc = DynCfgHttpClient.get_job_list(@parent, @plugin, @arry_mod, @child)
+        assert_eq(rc.code, 200, "as HTTP code for get_jobs request")
+        jobs = nil
+        assert_nothing_raised do
+            jobs = JSON.parse(rc.parsed_response, symbolize_names: true)
+        end
+        assert_has_key?(jobs, :jobs)
+        new_job = jobs[:jobs].find {|i| i[:name] == created_job}
+        assert_nil(new_job)
+        # Check it is not in childs job list anymore
+        rc = DynCfgHttpClient.get_job_list(@child, @plugin, @arry_mod, nil)
+        assert_eq(rc.code, 200, "as HTTP code for get_jobs request")
+        jobs = nil
+        assert_nothing_raised do
+            jobs = JSON.parse(rc.parsed_response, symbolize_names: true)
+        end
+        assert_has_key?(jobs, :jobs)
+        new_job = jobs[:jobs].find {|i| i[:name] == created_job}
+        assert_nil(new_job)
+        PASS()
+
+        TEST("parent/child/module/del_undeletable_job", "Try delete job on child (child rejects), check failure case works (hops:1)")
+        # test if plugin rejects job deletion the job still remains in list as it should
+        rc = DynCfgHttpClient.delete_job(@parent, @plugin, @arry_mod, @test_job, @child)
+        assert_eq(rc.code, 500, "as HTTP code for delete_job request")
+        rc = DynCfgHttpClient.get_job_list(@parent, @plugin, @arry_mod, @child)
+        assert_eq(rc.code, 200, "as HTTP code for get_jobs request")
+        jobs = nil
+        assert_nothing_raised do
+            jobs = JSON.parse(rc.parsed_response, symbolize_names: true)
+        end
+        assert_has_key?(jobs, :jobs)
+        job = jobs[:jobs].find {|i| i[:name] == @test_job}
+        assert_not_nil(job)
+        PASS()
+    end
+end
+
+ParentChildTest.new.run()
diff --git a/libnetdata/dyn_conf/tests/test_dyncfg.rb b/libnetdata/dyn_conf/tests/test_dyncfg.rb
new file mode 100755
index 00000000..1b4b3a06
--- /dev/null
+++ b/libnetdata/dyn_conf/tests/test_dyncfg.rb
@@ -0,0 +1,266 @@
+#!/usr/bin/env ruby
+
+require 'json'
+require 'httparty'
+require 'pastel'
+require 'securerandom'
+
+ARGV.length == 1 or raise "Usage: #{$0} <config file>"
+config_file = ARGV[0]
+
+File.exist?(config_file) or raise "File not found: #{config_file}"
+
+$config = JSON.parse(File.read(config_file), symbolize_names: true)
+
+$plugin_name = $config[:global][:test_plugin_name]
+$pastel = Pastel.new
+
+class TestRunner
+    attr_reader :stats
+    def initialize
+        @stats = {
+            :suites => 0,
+            :tests  => 0,
+            :assertions => 0
+        }
+        @test = nil
+    end
+    def add_assertion()
+        @stats[:assertions] += 1
+    end
+    def FAIL(msg, exception = nil, loc = nil)
+        puts $pastel.red.bold(" ✕ FAIL")
+        STDERR.print "    "
+        if loc
+            STDERR.print $pastel.yellow("@#{loc.path}:#{loc.lineno}: ")
+        else
+            STDERR.print $pastel.yellow("@#{caller_locations(1, 1).first.path}:#{caller_locations(1, 1).first.lineno}: ")
+        end
+        STDERR.puts msg
+        STDERR.puts exception.full_message(:highlight => true) if exception
+        STDERR.puts $pastel.yellow("    Backtrace:")
+        caller.each do |line|
+            STDERR.puts "        #{line}"
+        end
+        exit 1
+    end
+    def PASS()
+        STDERR.puts $pastel.green.bold(" ✓ PASS")
+        @stats[:tests] += 1
+        @test = nil
+    end
+    def TEST_SUITE(name)
+        puts $pastel.bold("• TEST SUITE: \"#{name}\"")
+        @stats[:suites] += 1
+    end
+    def assert_no_test_running()
+        unless @test.nil?
+            STDERR.puts $pastel.red("\nFATAL: Test \"#{@test}\" did not call PASS() or FAIL()!")
+            exit 1
+        end
+    end
+    def TEST(name, description = nil)
+        assert_no_test_running()
+        @test = name
+        col = 0
+        txt = "  ├─ T: #{name} "
+        col += txt.length
+        print $pastel.bold(txt)
+
+        tab = 50
+        rem = tab - (col % tab)
+        rem.times do putc ' ' end
+        col += rem
+
+        if (description)
+            txt = " - #{description} "
+            col += txt.length
+            print txt
+
+            tab = 180
+            rem = tab - (col % tab)
+            rem.times do putc '.' end
+        end
+    end
+    def FINALIZE()
+        assert_no_test_running()
+    end
+end
+
+$test_runner = TestRunner.new
+def FAIL(msg, exception = nil, loc = nil)
+    $test_runner.FAIL(msg, exception, loc)
+end
+def PASS()
+    $test_runner.PASS()
+end
+def TEST_SUITE(name)
+    $test_runner.TEST_SUITE(name)
+end
+def TEST(name, description = nil)
+    $test_runner.TEST(name, description)
+end
+
+def assert_eq(got, expected, msg = nil)
+    unless got == expected
+        FAIL("Expected #{expected}, got #{got} #{msg ? "(#{msg})" : ""}", nil, caller_locations(1, 1).first)
+    end
+    $test_runner.add_assertion()
+end
+def assert_eq_http_code(got, expected, msg = nil)
+    unless got.code == expected
+        FAIL("Expected #{expected}, got #{got}. Server \"#{got.parsed_response}\" #{msg ? "(#{msg})" : ""}", nil, caller_locations(1, 1).first)
+    end
+    $test_runner.add_assertion()
+end
+def assert_eq_str(got, expected, msg = nil)
+    unless got == expected
+        FAIL("Strings do not match #{msg ? "(#{msg})" : ""}", nil, caller_locations(1, 1).first)
+    end
+    $test_runner.add_assertion()
+end
+def assert_not_eq_str(got, expected, msg = nil)
+    unless got != expected
+        FAIL("Strings shoud not match #{msg ? "(#{msg})" : ""}", nil, caller_locations(1, 1).first)
+    end
+    $test_runner.add_assertion()
+end
+def assert_nothing_raised()
+    begin
+        yield
+    rescue Exception => e
+        FAIL("Unexpected exception of type #{e.class} raised. Msg: \"#{e.message}\"", e, caller_locations(1, 1).first)
+    end
+    $test_runner.add_assertion()
+end
+def assert_has_key?(hash, key)
+    unless hash.has_key?(key)
+        FAIL("Expected key \"#{key}\" in hash", nil, caller_locations(1, 1).first)
+    end
+    $test_runner.add_assertion()
+end
+def assert_array_include?(array, value)
+    unless array.include?(value)
+        FAIL("Expected array to include \"#{value}\"", nil, caller_locations(1, 1).first)
+    end
+    $test_runner.add_assertion()
+end
+def assert_array_not_include?(array, value)
+    if array.include?(value)
+        FAIL("Expected array to not include \"#{value}\"", nil, caller_locations(1, 1).first)
+    end
+    $test_runner.add_assertion()
+end
+def assert_is_one_of(value, *values)
+    unless values.include?(value)
+        FAIL("Expected value to be one of #{values.join(", ")}", nil, caller_locations(1, 1).first)
+    end
+    $test_runner.add_assertion()
+end
+def assert_not_nil(value)
+    if value.nil?
+        FAIL("Expected value to not be nil", nil, caller_locations(1, 1).first)
+    end
+    $test_runner.add_assertion()
+end
+def assert_nil(value)
+    unless value.nil?
+        FAIL("Expected value to be nil", nil, caller_locations(1, 1).first)
+    end
+    $test_runner.add_assertion()
+end
+
+
+class DynCfgHttpClient
+    def self.protocol(cfg)
+        return cfg[:ssl] ? 'https://' : 'http://'
+    end
+    def self.url_base(host)
+        return "#{protocol(host)}#{host[:host]}:#{host[:port]}"
+    end
+    def self.get_url_cfg_base(host, child = nil)
+        url = url_base(host)
+        url += "/host/#{child[:mguid]}" if child
+        url += "/api/v2/config"
+        return url
+    end
+    def self.get_url_cfg_plugin(host, plugin, child = nil)
+        return get_url_cfg_base(host, child) + '/' + plugin
+    end
+    def self.get_url_cfg_module(host, plugin, mod, child = nil)
+        return get_url_cfg_plugin(host, plugin, child) + '/' + mod
+    end
+    def self.get_url_cfg_job(host, plugin, mod, job_id, child = nil)
+        return get_url_cfg_module(host, plugin, mod, child) + "/#{job_id}"
+    end
+    def self.get_plugin_list(host, child = nil)
+        begin
+            return HTTParty.get(get_url_cfg_base(host, child), verify: false, format: :plain)
+        rescue => e
+            FAIL(e.message, e)
+        end
+    end
+    def self.get_plugin_config(host, plugin, child = nil)
+        begin
+            return HTTParty.get(get_url_cfg_plugin(host, plugin, child), verify: false)
+        rescue => e
+            FAIL(e.message, e)
+        end
+    end
+    def self.set_plugin_config(host, plugin, cfg, child = nil)
+        begin
+            return HTTParty.put(get_url_cfg_plugin(host, plugin, child), verify: false, body: cfg)
+        rescue => e
+            FAIL(e.message, e)
+        end
+    end
+    def self.get_plugin_module_list(host, plugin, child = nil)
+        begin
+            return HTTParty.get(get_url_cfg_plugin(host, plugin, child) + "/modules", verify: false, format: :plain)
+        rescue => e
+            FAIL(e.message, e)
+        end
+    end
+    def self.get_job_list(host, plugin, mod, child = nil)
+        begin
+            return HTTParty.get(get_url_cfg_module(host, plugin, mod, child) + "/jobs", verify: false, format: :plain)
+        rescue => e
+            FAIL(e.message, e)
+        end
+    end
+    def self.create_job(host, plugin, mod, job_id, job_cfg, child = nil)
+        begin
+            return HTTParty.post(get_url_cfg_job(host, plugin, mod, job_id, child), verify: false, body: job_cfg)
+        rescue => e
+            FAIL(e.message, e)
+        end
+    end
+    def self.delete_job(host, plugin, mod, job_id, child = nil)
+        begin
+            return HTTParty.delete(get_url_cfg_job(host, plugin, mod, job_id, child), verify: false)
+        rescue => e
+            FAIL(e.message, e)
+        end
+    end
+    def self.get_job_config(host, plugin, mod, job_id, child = nil)
+        begin
+            return HTTParty.get(get_url_cfg_job(host, plugin, mod, job_id, child), verify: false, format: :plain)
+        rescue => e
+            FAIL(e.message, e)
+        end
+    end
+    def self.set_job_config(host, plugin, mod, job_id, job_cfg, child = nil)
+        begin
+            return HTTParty.put(get_url_cfg_job(host, plugin, mod, job_id, child), verify: false, body: job_cfg)
+        rescue => e
+            FAIL(e.message, e)
+        end
+    end
+end
+
+require_relative 'sub_tests/test_parent_child.rb'
+
+$test_runner.FINALIZE()
+puts $pastel.green.bold("All tests passed!")
+puts ("Total #{$test_runner.stats[:assertions]} assertions, #{$test_runner.stats[:tests]} tests in #{$test_runner.stats[:suites]} suites")
+exit 0
diff --git a/libnetdata/dyn_conf/tests/test_plugin/test.plugin b/libnetdata/dyn_conf/tests/test_plugin/test.plugin
new file mode 100755
index 00000000..b890ab31
--- /dev/null
+++ b/libnetdata/dyn_conf/tests/test_plugin/test.plugin
@@ -0,0 +1,250 @@
+#!/usr/bin/env ruby
+
+# bogus chart that we create just so there is at least one chart
+CHART_TYPE = 'lines'
+UPDATE_EVERY = 1
+PRIORITY = 100000
+CHART_NAME = 'number_of_processes'
+DIMENSION_NAME = 'running'
+
+$plugin_name = "external_plugin"
+$plugin_version = "0.0.1"
+$plugin_config = <<-HEREDOC
+test_plugin_config
+hableba hableba hableba
+HEREDOC
+
+$array_module_name = 'module_of_the_future'
+$fixed_job_name = 'fixed_job'
+
+$modules = {
+  $array_module_name => {
+    :type => :job_array,
+    :jobs => {
+      $fixed_job_name => {
+        :type => :fixed,
+        :config => <<-HEREDOC
+fixed_job_config
+HEREDOC
+      },
+    },
+    :config => <<-HEREDOC
+module_of_the_future_config
+HEREDOC
+  },
+  "module_of_the_future_single_type" => {
+    :type => :single,
+    :jobs => {},
+    :config => <<-HEREDOC
+module_of_the_future_single_type_config
+HEREDOC
+  }
+}
+
+def out(str)
+  $log.puts "2 NETDATA> #{str}"
+  $stdout.puts str
+  $stdout.flush
+  $log.flush
+end
+
+def log(str)
+  $log.puts "LOG      > #{str}"
+  $log.flush
+end
+
+#TODO this is AI code, verify
+def split_with_quotes(str)
+  result = []
+  current_word = ""
+  in_quotes = false
+  escaped = false
+
+  str.each_char do |char|
+    if char == '\\' && !escaped
+      escaped = true
+      next
+    end
+
+    if char == '"' && !escaped
+      in_quotes = !in_quotes
+      current_word << char
+    elsif char == ' ' && !in_quotes
+      result << current_word unless current_word.empty?
+      current_word = ""
+    else
+      current_word << char
+    end
+
+    escaped = false
+  end
+
+  result << current_word unless current_word.empty?
+
+  result
+end
+
+
+def print_startup_messages
+  out "DYNCFG_ENABLE #{$plugin_name}"
+  $modules.each do |name, module_config|
+    out "DYNCFG_REGISTER_MODULE #{name} #{module_config[:type]}"
+  end
+  out "CHART system.#{CHART_NAME} '' 'Number of running processes' 'processes' processes processes.#{CHART_NAME} #{CHART_TYPE} #{PRIORITY} #{UPDATE_EVERY}"
+  out "DIMENSION #{DIMENSION_NAME} '' absolute 1 1"
+
+  $modules.each do |mod_name, mod|
+    next unless mod[:type] == :job_array
+    mod[:jobs].each do |job_name, job|
+      next unless job[:type] == :fixed
+      out "DYNCFG_REGISTER_JOB #{mod_name} #{job_name} stock 0"
+      out "REPORT_JOB_STATUS #{$array_module_name} #{$fixed_job_name} running 0"
+    end
+  end
+end
+
+def function_result(txid, msg, result)
+  out "FUNCTION_RESULT_BEGIN #{txid} #{result} text/plain 5"
+  out msg
+  out "FUNCTION_RESULT_END"
+end
+
+def process_payload_function(params)
+  log "payload function #{params[:fncname]}, #{params[:fncparams]}"
+  fnc_name, mod_name, job_name = params[:fncparams]
+  case fnc_name
+  when 'set_plugin_config'
+    $plugin_config = params[:payload]
+    function_result(params[:txid], "plugin config set", 1)
+  when 'set_module_config'
+    mod = $modules[mod_name]
+    return function_result(params[:txid], "no such module", 0) if mod.nil?
+    mod[:config] = params[:payload]
+    function_result(params[:txid], "module config set", 1)
+  when 'set_job_config'
+    mod = $modules[mod_name]
+    return function_result(params[:txid], "no such module", 0) if mod.nil?
+    job = mod[:jobs][job_name]
+    if job.nil?
+      job = Hash.new if job.nil?
+      job[:type] = :dynamic
+      mod[:jobs][job_name] = job
+    end
+    job[:config] = params[:payload]
+    function_result(params[:txid], "job config set", 1)
+  end
+end
+
+def process_function(params)
+  log "normal function #{params[:fncname]}, #{params[:fncparams]}"
+  fnc_name, mod_name, job_name = params[:fncparams]
+  case fnc_name
+  when 'get_plugin_config'
+    function_result(params[:txid], $plugin_config, 1)
+  when 'get_module_config'
+    return function_result(params[:txid], "no such module", 0) unless $modules.has_key?(mod_name)
+    function_result(params[:txid], $modules[mod_name][:config], 1)
+  when 'get_job_config'
+    mod = $modules[mod_name]
+    return function_result(params[:txid], "no such module", 0) if mod.nil?
+    job = mod[:jobs][job_name]
+    return function_result(params[:txid], "no such job", 0) if job.nil?
+    function_result(params[:txid], job[:config], 1)
+  when 'delete_job'
+    mod = $modules[mod_name]
+    return function_result(params[:txid], "no such module", 0) if mod.nil?
+    job = mod[:jobs][job_name]
+    return function_result(params[:txid], "no such job", 0) if job.nil?
+    if job[:type] == :fixed
+      return function_result(params[:txid], "this job can't be deleted", 0) 
+    else
+      mod[:jobs].delete(job_name)
+      function_result(params[:txid], "job deleted", 1)
+    end
+  end
+end
+
+$inflight_incoming = nil
+def process_input(input)
+  words = split_with_quotes(input)
+
+  unless $inflight_incoming.nil?
+    if input == "FUNCTION_PAYLOAD_END"
+      log $inflight_incoming[:payload]
+      process_payload_function($inflight_incoming)
+      $inflight_incoming = nil
+    else
+      $inflight_incoming[:payload] << input
+      $inflight_incoming[:payload] << "\n"
+    end
+    return
+  end
+
+  case words[0]
+  when "FUNCTION", "FUNCTION_PAYLOAD"
+    params = {}
+    params[:command] = words[0]
+    params[:txid] = words[1]
+    params[:timeout] = words[2].to_i
+    params[:fncname] = words[3]
+    params[:fncname] = params[:fncname][1..-2] if params[:fncname].start_with?('"') && params[:fncname].end_with?('"')
+    if params[:command] == "FUNCTION_PAYLOAD"
+      $inflight_incoming = Hash.new
+      params[:fncparams] = split_with_quotes(params[:fncname])
+      params[:fncname] = params[:fncparams][0]
+      $inflight_incoming[:txid] = params[:txid]
+      $inflight_incoming[:fncname] = params[:fncname]
+      $inflight_incoming[:params] = params
+      $inflight_incoming[:fncparams] = params[:fncparams]
+      $inflight_incoming[:payload] = ""
+    else
+      params[:fncparams] = split_with_quotes(params[:fncname])
+      params[:fncname] = params[:fncparams][0]
+      process_function(params)
+    end
+  end
+end
+
+def read_and_output_metric
+  processes = `ps -e | wc -l`.to_i - 1 # -1 to exclude the header line
+  timestamp = Time.now.to_i
+
+  puts "BEGIN system.#{CHART_NAME}"
+  puts "SET #{DIMENSION_NAME} = #{processes}"
+  puts "END"
+end
+
+def the_main
+  $stderr.reopen("/tmp/test_plugin_err.log", "w")
+  $log = File.open("/tmp/test_plugin.log", "w")
+  $log.puts "Starting plugin"
+  print_startup_messages
+  $log.puts "init done"
+  $log.flush
+
+  last_metric_time = Time.now
+
+  loop do
+    time_since_last_metric = Time.now - last_metric_time
+
+    # If it's been more than 1 second since we collected metrics, collect them now
+    if time_since_last_metric >= 1
+      read_and_output_metric
+      last_metric_time = Time.now
+    end
+
+    # Use select to wait for input, but only wait up to the time remaining until we need to collect metrics again
+    remaining_time = [1 - time_since_last_metric, 0].max
+    if select([$stdin], nil, nil, remaining_time)
+      input = $stdin.gets
+      next if input.class != String
+      input.chomp!
+      $log.puts "RAW INPUT< #{input}"
+      $log.flush
+      process_input(input)
+    end
+  end
+end
+
+
+the_main if __FILE__ == $PROGRAM_NAME
diff --git a/libnetdata/ebpf/ebpf.c b/libnetdata/ebpf/ebpf.c
index 6793f403..1bd45ef2 100644
--- a/libnetdata/ebpf/ebpf.c
+++ b/libnetdata/ebpf/ebpf.c
@@ -792,13 +792,13 @@ void ebpf_update_controller(int fd, ebpf_module_t *em)
 {
     uint32_t values[NETDATA_CONTROLLER_END] = {
         (em->apps_charts & NETDATA_EBPF_APPS_FLAG_YES) | em->cgroup_charts,
-        em->apps_level
+        em->apps_level, 0, 0, 0, 0
     };
     uint32_t key;
-    uint32_t end = (em->apps_level != NETDATA_APPS_NOT_SET) ? NETDATA_CONTROLLER_END : NETDATA_CONTROLLER_APPS_LEVEL;
+    uint32_t end = NETDATA_CONTROLLER_PID_TABLE_ADD;
 
     for (key = NETDATA_CONTROLLER_APPS_ENABLED; key < end; key++) {
-        int ret = bpf_map_update_elem(fd, &key, &values[key], 0);
+        int ret = bpf_map_update_elem(fd, &key, &values[key], BPF_ANY);
         if (ret)
             netdata_log_error("Add key(%u) for controller table failed.", key);
     }
@@ -855,7 +855,7 @@ struct bpf_link **ebpf_load_program(char *plugins_dir, ebpf_module_t *em, int kv
 
     uint32_t idx = ebpf_select_index(em->kernels, is_rhf, kver);
 
-    ebpf_mount_name(lpath, 4095, plugins_dir, idx, em->thread_name, em->mode, is_rhf);
+    ebpf_mount_name(lpath, 4095, plugins_dir, idx, em->info.thread_name, em->mode, is_rhf);
 
     // When this function is called ebpf.plugin is using legacy code, so we should reset the variable
     em->load &= ~ NETDATA_EBPF_LOAD_METHODS;
@@ -1269,7 +1269,7 @@ void ebpf_update_module_using_config(ebpf_module_t *modules, netdata_ebpf_load_m
 
 #ifdef NETDATA_DEV_MODE
     netdata_log_info("The thread %s was configured with: mode = %s; update every = %d; apps = %s; cgroup = %s; ebpf type format = %s; ebpf co-re tracing = %s; collect pid = %s; maps per core = %s, lifetime=%u",
-         modules->thread_name,
+         modules->info.thread_name,
          load_mode,
          modules->update_every,
          (modules->apps_charts)?"enabled":"disabled",
diff --git a/libnetdata/ebpf/ebpf.h b/libnetdata/ebpf/ebpf.h
index 691a4c26..6708f669 100644
--- a/libnetdata/ebpf/ebpf.h
+++ b/libnetdata/ebpf/ebpf.h
@@ -301,11 +301,27 @@ enum ebpf_global_table_values {
 typedef uint64_t netdata_idx_t;
 
 typedef struct ebpf_module {
-    const char *thread_name;
-    const char *config_name;
-    const char *thread_description;
+    // Constants used with module
+    struct {
+        const char *thread_name;
+        const char *config_name;
+        const char *thread_description;
+    } info;
+
+    // Helpers used with plugin
+    struct {
+        void *(*start_routine)(void *);                             // the thread function
+        void (*apps_routine)(struct ebpf_module *em, void *ptr);    // the apps charts
+        void (*fnct_routine)(BUFFER *bf, struct ebpf_module *em);   // the function used for exteernal requests
+        const char *fcnt_name;                                      // name given to cloud
+        const char *fcnt_desc;                                      // description given about function
+        const char *fcnt_thread_chart_name;
+        int order_thread_chart;
+        const char *fcnt_thread_lifetime_name;
+        int order_thread_lifetime;
+    } functions;
+
     enum ebpf_threads_status enabled;
-    void *(*start_routine)(void *);
     int update_every;
     int global_charts;
     netdata_apps_integration_flags_t apps_charts;
@@ -314,7 +330,6 @@ typedef struct ebpf_module {
     netdata_run_mode_t mode;
     uint32_t thread_id;
     int optional;
-    void (*apps_routine)(struct ebpf_module *em, void *ptr);
     ebpf_local_maps_t *maps;
     ebpf_specify_name_t *names;
     uint32_t pid_map_size;
diff --git a/libnetdata/facets/facets.c b/libnetdata/facets/facets.c
index b285baf0..52898feb 100644
--- a/libnetdata/facets/facets.c
+++ b/libnetdata/facets/facets.c
@@ -1,77 +1,201 @@
+// SPDX-License-Identifier: GPL-3.0-or-later
 #include "facets.h"
 
-#define HISTOGRAM_COLUMNS 60
+#define HISTOGRAM_COLUMNS 150               // the target number of points in a histogram
+#define FACETS_KEYS_WITH_VALUES_MAX 200     // the max number of keys that can be facets
+#define FACETS_KEYS_IN_ROW_MAX 500          // the max number of keys in a row
 
-static void facets_row_free(FACETS *facets __maybe_unused, FACET_ROW *row);
+#define FACETS_KEYS_HASHTABLE_ENTRIES 127
+#define FACETS_VALUES_HASHTABLE_ENTRIES 31
 
 // ----------------------------------------------------------------------------
 
-time_t calculate_bar_width(time_t before, time_t after) {
-    // Array of valid durations in seconds
-    static time_t valid_durations[] = {
-            1,
-            15,
-            30,
-            1 * 60, 2 * 60, 3 * 60, 5 * 60, 10 * 60, 15 * 60, 30 * 60,          // minutes
-            1 * 3600, 2 * 3600, 6 * 3600, 8 * 3600, 12 * 3600,                  // hours
-            1 * 86400, 2 * 86400, 3 * 86400, 5 * 86400, 7 * 86400, 14 * 86400,  // days
-            1 * (30*86400)                                                      // months
-    };
-    static int array_size = sizeof(valid_durations) / sizeof(valid_durations[0]);
+static const char id_encoding_characters[64 + 1] = "ABCDEFGHIJKLMNOPQRSTUVWXYZ.abcdefghijklmnopqrstuvwxyz_0123456789";
+static const uint8_t id_encoding_characters_reverse[256] = {
+        ['A'] = 0,  ['B'] = 1,  ['C'] = 2,  ['D'] = 3,
+        ['E'] = 4,  ['F'] = 5,  ['G'] = 6,  ['H'] = 7,
+        ['I'] = 8,  ['J'] = 9,  ['K'] = 10, ['L'] = 11,
+        ['M'] = 12, ['N'] = 13, ['O'] = 14, ['P'] = 15,
+        ['Q'] = 16, ['R'] = 17, ['S'] = 18, ['T'] = 19,
+        ['U'] = 20, ['V'] = 21, ['W'] = 22, ['X'] = 23,
+        ['Y'] = 24, ['Z'] = 25, ['.'] = 26, ['a'] = 27,
+        ['b'] = 28, ['c'] = 29, ['d'] = 30, ['e'] = 31,
+        ['f'] = 32, ['g'] = 33, ['h'] = 34, ['i'] = 35,
+        ['j'] = 36, ['k'] = 37, ['l'] = 38, ['m'] = 39,
+        ['n'] = 40, ['o'] = 41, ['p'] = 42, ['q'] = 43,
+        ['r'] = 44, ['s'] = 45, ['t'] = 46, ['u'] = 47,
+        ['v'] = 48, ['w'] = 49, ['x'] = 50, ['y'] = 51,
+        ['z'] = 52, ['_'] = 53, ['0'] = 54, ['1'] = 55,
+        ['2'] = 56, ['3'] = 57, ['4'] = 58, ['5'] = 59,
+        ['6'] = 60, ['7'] = 61, ['8'] = 62, ['9'] = 63
+};
 
-    time_t duration = before - after;
-    time_t bar_width = 1;
+__attribute__((constructor)) void initialize_facets_id_encoding_characters_reverse(void) {
 
-    for (int i = array_size - 1; i >= 0; --i) {
-        if (duration / valid_durations[i] >= HISTOGRAM_COLUMNS) {
-            bar_width = valid_durations[i];
-            break;
+}
+
+#define FACET_STRING_HASH_SIZE 12
+#define FACETS_HASH XXH64_hash_t
+#define FACETS_HASH_FUNCTION(src, len) XXH3_64bits(src, len)
+#define FACETS_HASH_ZERO (FACETS_HASH)0
+
+static inline void facets_hash_to_str(FACETS_HASH num, char *out) {
+    out[11] = '\0';
+    out[10] = id_encoding_characters[num & 63]; num >>= 6;
+    out[9]  = id_encoding_characters[num & 63]; num >>= 6;
+    out[8]  = id_encoding_characters[num & 63]; num >>= 6;
+    out[7]  = id_encoding_characters[num & 63]; num >>= 6;
+    out[6]  = id_encoding_characters[num & 63]; num >>= 6;
+    out[5]  = id_encoding_characters[num & 63]; num >>= 6;
+    out[4]  = id_encoding_characters[num & 63]; num >>= 6;
+    out[3]  = id_encoding_characters[num & 63]; num >>= 6;
+    out[2]  = id_encoding_characters[num & 63]; num >>= 6;
+    out[1]  = id_encoding_characters[num & 63]; num >>= 6;
+    out[0]  = id_encoding_characters[num & 63];
+}
+
+static inline FACETS_HASH str_to_facets_hash(const char *str) {
+    FACETS_HASH num = 0;
+    int shifts = 6 * (FACET_STRING_HASH_SIZE - 2);
+
+    num |= ((FACETS_HASH)(id_encoding_characters_reverse[(uint8_t)(str[0])])) << shifts; shifts -= 6;
+    num |= ((FACETS_HASH)(id_encoding_characters_reverse[(uint8_t)(str[1])])) << shifts; shifts -= 6;
+    num |= ((FACETS_HASH)(id_encoding_characters_reverse[(uint8_t)(str[2])])) << shifts; shifts -= 6;
+    num |= ((FACETS_HASH)(id_encoding_characters_reverse[(uint8_t)(str[3])])) << shifts; shifts -= 6;
+    num |= ((FACETS_HASH)(id_encoding_characters_reverse[(uint8_t)(str[4])])) << shifts; shifts -= 6;
+    num |= ((FACETS_HASH)(id_encoding_characters_reverse[(uint8_t)(str[5])])) << shifts; shifts -= 6;
+    num |= ((FACETS_HASH)(id_encoding_characters_reverse[(uint8_t)(str[6])])) << shifts; shifts -= 6;
+    num |= ((FACETS_HASH)(id_encoding_characters_reverse[(uint8_t)(str[7])])) << shifts; shifts -= 6;
+    num |= ((FACETS_HASH)(id_encoding_characters_reverse[(uint8_t)(str[8])])) << shifts; shifts -= 6;
+    num |= ((FACETS_HASH)(id_encoding_characters_reverse[(uint8_t)(str[9])])) << shifts; shifts -= 6;
+    num |= ((FACETS_HASH)(id_encoding_characters_reverse[(uint8_t)(str[10])])) << shifts;
+
+    return num;
+}
+
+static const char *hash_to_static_string(FACETS_HASH hash) {
+    static __thread char hash_str[FACET_STRING_HASH_SIZE];
+    facets_hash_to_str(hash, hash_str);
+    return hash_str;
+}
+
+static inline bool is_valid_string_hash(const char *s) {
+    if(strlen(s) != FACET_STRING_HASH_SIZE - 1) {
+        netdata_log_error("The user supplied key '%s' does not have the right length for a facets hash.", s);
+        return false;
+    }
+
+    uint8_t *t = (uint8_t *)s;
+    while(*t) {
+        if(id_encoding_characters_reverse[*t] == 0 && *t != id_encoding_characters[0]) {
+            netdata_log_error("The user supplied key '%s' contains invalid characters for a facets hash.", s);
+            return false;
         }
+
+        t++;
     }
 
-    return bar_width;
+    return true;
 }
 
 // ----------------------------------------------------------------------------
 
-static inline void uint32_to_char(uint32_t num, char *out) {
-    static char id_encoding_characters[64 + 1] = "ABCDEFGHIJKLMNOPQRSTUVWXYZ.abcdefghijklmnopqrstuvwxyz_0123456789";
+typedef uint64_t SIMPLE_HASHTABLE_HASH;
+#define SIMPLE_HASHTABLE_HASH_SECOND_HASH_SHIFTS 32
 
-    int i;
-    for(i = 5; i >= 0; --i) {
-        out[i] = id_encoding_characters[num & 63];
-        num >>= 6;
-    }
-    out[6] = '\0';
+typedef struct simple_hashtable_slot {
+    SIMPLE_HASHTABLE_HASH hash;
+    void *data;
+} SIMPLE_HASHTABLE_SLOT;
+
+typedef struct simple_hashtable {
+    size_t size;
+    SIMPLE_HASHTABLE_SLOT *hashtable;
+} SIMPLE_HASHTABLE;
+
+static void simple_hashtable_init(SIMPLE_HASHTABLE *ht, size_t size) {
+    ht->size = size;
+    ht->hashtable = callocz(ht->size, sizeof(*ht->hashtable));
 }
 
-inline void facets_string_hash(const char *src, char *out) {
-    uint32_t hash1 = fnv1a_hash32(src);
-    uint32_t hash2 = djb2_hash32(src);
-    uint32_t hash3 = larson_hash32(src);
+static void simple_hashtable_free(SIMPLE_HASHTABLE *ht) {
+    freez(ht->hashtable);
+    ht->hashtable = NULL;
+    ht->size = 0;
+}
+
+static inline SIMPLE_HASHTABLE_SLOT *simple_hashtable_get_slot(SIMPLE_HASHTABLE *ht, SIMPLE_HASHTABLE_HASH hash) {
+    // IMPORTANT:
+    // If the hashtable supported deletions, we would need to have a special slot.data value
+    // to mark deleted values and assume they are occupied during lookup, but empty during insert.
+    // But for our case, we don't need it, since we never delete items from the hashtable.
+
+    size_t slot = hash % ht->size;
+    if(!ht->hashtable[slot].data || ht->hashtable[slot].hash == hash)
+        return &ht->hashtable[slot];
+
+    slot = ((hash >> SIMPLE_HASHTABLE_HASH_SECOND_HASH_SHIFTS) + 1) % ht->size;
+    // Linear probing until we find it
+    while (ht->hashtable[slot].data && ht->hashtable[slot].hash != hash)
+        slot = (slot + 1) % ht->size;  // Wrap around if necessary
+
+    return &ht->hashtable[slot];
+}
 
-    uint32_to_char(hash1, out);
-    uint32_to_char(hash2, &out[6]);
-    uint32_to_char(hash3, &out[12]);
+static void simple_hashtable_resize_double(SIMPLE_HASHTABLE *ht) {
+    SIMPLE_HASHTABLE_SLOT *old = ht->hashtable;
+    size_t old_size = ht->size;
 
-    out[18] = '\0';
+    ht->size = (ht->size * 2) + 1;
+    ht->hashtable = callocz(ht->size, sizeof(*ht->hashtable));
+    for(size_t i = 0 ; i < old_size ; i++) {
+        if(!old[i].data)
+            continue;
+
+        SIMPLE_HASHTABLE_SLOT *slot = simple_hashtable_get_slot(ht, old[i].hash);
+        *slot = old[i];
+    }
+
+    freez(old);
 }
 
+
 // ----------------------------------------------------------------------------
 
 typedef struct facet_value {
+    FACETS_HASH hash;
     const char *name;
+    uint32_t name_len;
 
     bool selected;
+    bool empty;
 
     uint32_t rows_matching_facet_value;
     uint32_t final_facet_value_counter;
+    uint32_t order;
+
+    uint32_t *histogram;
+    uint32_t min, max, sum;
+
+    struct facet_value *prev, *next;
 } FACET_VALUE;
 
+typedef enum {
+    FACET_KEY_VALUE_NONE    = 0,
+    FACET_KEY_VALUE_UPDATED = (1 << 0),
+    FACET_KEY_VALUE_EMPTY   = (1 << 1),
+    FACET_KEY_VALUE_COPIED  = (1 << 2),
+} FACET_KEY_VALUE_FLAGS;
+
+#define facet_key_value_updated(k) ((k)->current_value.flags & FACET_KEY_VALUE_UPDATED)
+#define facet_key_value_empty(k) ((k)->current_value.flags & FACET_KEY_VALUE_EMPTY)
+#define facet_key_value_copied(k) ((k)->current_value.flags & FACET_KEY_VALUE_COPIED)
+
 struct facet_key {
-    const char *name;
+    FACETS *facets;
 
-    DICTIONARY *values;
+    FACETS_HASH hash;
+    const char *name;
 
     FACET_KEY_OPTIONS options;
 
@@ -80,14 +204,27 @@ struct facet_key {
     // members about the current row
     uint32_t key_found_in_row;
     uint32_t key_values_selected_in_row;
+    uint32_t order;
 
     struct {
-        char hash[FACET_STRING_HASH_SIZE];
-        bool updated;
+        bool enabled;
+        uint32_t used;
+        FACET_VALUE *ll;
+        SIMPLE_HASHTABLE ht;
+    } values;
+
+    struct {
+        FACETS_HASH hash;
+        FACET_KEY_VALUE_FLAGS flags;
+        const char *raw;
+        uint32_t raw_len;
         BUFFER *b;
+        FACET_VALUE *v;
     } current_value;
 
-    uint32_t order;
+    struct {
+        FACET_VALUE *v;
+    } empty_value;
 
     struct {
         facet_dynamic_row_t cb;
@@ -95,6 +232,7 @@ struct facet_key {
     } dynamic;
 
     struct {
+        bool view_only;
         facets_key_transformer_t cb;
         void *data;
     } transform;
@@ -108,16 +246,36 @@ struct facets {
     SIMPLE_PATTERN *included_keys;
 
     FACETS_OPTIONS options;
-    usec_t anchor;
+
+    struct {
+        usec_t start_ut;
+        usec_t stop_ut;
+        FACETS_ANCHOR_DIRECTION direction;
+    } anchor;
 
     SIMPLE_PATTERN *query;          // the full text search pattern
     size_t keys_filtered_by_query;  // the number of fields we do full text search (constant)
-    size_t keys_matched_by_query;   // the number of fields matched the full text search (per row)
 
     DICTIONARY *accepted_params;
 
-    FACET_KEY *keys_ll;
-    DICTIONARY *keys;
+    struct {
+        size_t count;
+        FACET_KEY *ll;
+        SIMPLE_HASHTABLE ht;
+    } keys;
+
+    struct {
+        // this is like a stack, of the keys that are used as facets
+        size_t used;
+        FACET_KEY *array[FACETS_KEYS_WITH_VALUES_MAX];
+    } keys_with_values;
+
+    struct {
+        // this is like a stack, of the keys that need to clean up between each row
+        size_t used;
+        FACET_KEY *array[FACETS_KEYS_IN_ROW_MAX];
+    } keys_in_row;
+
     FACET_ROW *base;    // double linked list of the selected facets rows
 
     uint32_t items_to_return;
@@ -125,10 +283,33 @@ struct facets {
     uint32_t order;
 
     struct {
-        FACET_ROW *last_added;
+        FACET_ROW_SEVERITY severity;
+        size_t keys_matched_by_query;   // the number of fields matched the full text search (per row)
+    } current_row;
+
+    struct {
+        usec_t after_ut;
+        usec_t before_ut;
+    } timeframe;
 
-        size_t evaluated;
-        size_t matched;
+    struct {
+        FACET_KEY *key;
+        FACETS_HASH hash;
+        char *chart;
+        bool enabled;
+        uint32_t slots;
+        usec_t slot_width_ut;
+        usec_t after_ut;
+        usec_t before_ut;
+    } histogram;
+
+    struct {
+        facet_row_severity_t cb;
+        void *data;
+    } severity;
+
+    struct {
+        FACET_ROW *last_added;
 
         size_t first;
         size_t forwards;
@@ -138,117 +319,338 @@ struct facets {
         size_t prepends;
         size_t appends;
         size_t shifts;
+
+        struct {
+            size_t evaluated;
+            size_t matched;
+            size_t created;
+            size_t reused;
+        } rows;
+
+        struct {
+            size_t registered;
+            size_t unique;
+            size_t hashtable_increases;
+        } keys;
+
+        struct {
+            size_t registered;
+            size_t transformed;
+            size_t dynamic;
+            size_t empty;
+            size_t indexed;
+            size_t inserts;
+            size_t conflicts;
+            size_t hashtable_increases;
+        } values;
+
+        struct {
+            size_t searches;
+        } fts;
     } operations;
 };
 
-// ----------------------------------------------------------------------------
+usec_t facets_row_oldest_ut(FACETS *facets) {
+    if(facets->base)
+        return facets->base->prev->usec;
 
-static inline void facet_value_is_used(FACET_KEY *k, FACET_VALUE *v) {
-    if(!k->key_found_in_row)
-        v->rows_matching_facet_value++;
+    return 0;
+}
 
-    k->key_found_in_row++;
+usec_t facets_row_newest_ut(FACETS *facets) {
+    if(facets->base)
+        return facets->base->usec;
 
-    if(v->selected)
-        k->key_values_selected_in_row++;
+    return 0;
 }
 
-static inline bool facets_key_is_facet(FACETS *facets, FACET_KEY *k) {
-    bool included = true, excluded = false;
+uint32_t facets_rows(FACETS *facets) {
+    return facets->items_to_return;
+}
 
-    if(k->options & (FACET_KEY_OPTION_FACET | FACET_KEY_OPTION_NO_FACET)) {
-        if(k->options & FACET_KEY_OPTION_FACET) {
-            included = true;
-            excluded = false;
-        }
-        else if(k->options & FACET_KEY_OPTION_NO_FACET) {
-            included = false;
-            excluded = true;
-        }
+// ----------------------------------------------------------------------------
+
+static void facets_row_free(FACETS *facets __maybe_unused, FACET_ROW *row);
+static inline void facet_value_is_used(FACET_KEY *k, FACET_VALUE *v);
+static inline bool facets_key_is_facet(FACETS *facets, FACET_KEY *k);
+
+// ----------------------------------------------------------------------------
+// The FACET_VALUE index within each FACET_KEY
+
+#define foreach_value_in_key(k, v) \
+    for((v) = (k)->values.ll; (v) ;(v) = (v)->next)
+
+#define foreach_value_in_key_done(v) do { ; } while(0)
+
+static inline void FACETS_VALUES_INDEX_CREATE(FACET_KEY *k) {
+    k->values.ll = NULL;
+    k->values.used = 0;
+    simple_hashtable_init(&k->values.ht, FACETS_VALUES_HASHTABLE_ENTRIES);
+}
+
+static inline void FACETS_VALUES_INDEX_DESTROY(FACET_KEY *k) {
+    FACET_VALUE *v = k->values.ll;
+    while(v) {
+        FACET_VALUE *next = v->next;
+        freez(v->histogram);
+        freez((void *)v->name);
+        freez(v);
+        v = next;
     }
-    else {
-        if (facets->included_keys) {
-            if (!simple_pattern_matches(facets->included_keys, k->name))
-                included = false;
-        }
+    k->values.ll = NULL;
+    k->values.used = 0;
+    k->values.enabled = false;
 
-        if (facets->excluded_keys) {
-            if (simple_pattern_matches(facets->excluded_keys, k->name))
-                excluded = true;
-        }
+    simple_hashtable_free(&k->values.ht);
+}
+
+static inline const char *facets_key_get_value(FACET_KEY *k) {
+    return facet_key_value_copied(k) ? buffer_tostring(k->current_value.b) : k->current_value.raw;
+}
+
+static inline uint32_t facets_key_get_value_length(FACET_KEY *k) {
+    return facet_key_value_copied(k) ? buffer_strlen(k->current_value.b) : k->current_value.raw_len;
+}
+
+static inline void facets_key_value_copy_to_buffer(FACET_KEY *k) {
+    if(!facet_key_value_copied(k)) {
+        buffer_contents_replace(k->current_value.b, k->current_value.raw, k->current_value.raw_len);
+        k->current_value.flags |= FACET_KEY_VALUE_COPIED;
     }
+}
 
-    if(included && !excluded) {
-        k->options |= FACET_KEY_OPTION_FACET;
-        k->options &= ~FACET_KEY_OPTION_NO_FACET;
-        return true;
+static const char *facets_value_dup(const char *s, uint32_t len) {
+    char *d = mallocz(len + 1);
+
+    if(len)
+        memcpy(d, s, len);
+
+    d[len] = '\0';
+
+    return d;
+}
+
+static inline void FACET_VALUE_ADD_CONFLICT(FACET_KEY *k, FACET_VALUE *v, const FACET_VALUE * const nv) {
+    if(!v->name && !v->name_len && nv->name && nv->name_len) {
+        // an actual value, not a filter
+        v->name = facets_value_dup(nv->name, nv->name_len);
+        v->name_len = nv->name_len;
     }
 
-    k->options |= FACET_KEY_OPTION_NO_FACET;
-    k->options &= ~FACET_KEY_OPTION_FACET;
-    return false;
+    if(v->name && v->name_len)
+        facet_value_is_used(k, v);
+
+    internal_fatal(v->name && nv->name && v->name_len == nv->name_len && memcmp(v->name, nv->name, v->name_len) != 0,
+                   "value hash conflict: '%s' and '%s' have the same hash '%s'",
+                   v->name, nv->name, hash_to_static_string(v->hash));
+
+    k->facets->operations.values.conflicts++;
 }
 
-// ----------------------------------------------------------------------------
-// FACET_VALUE dictionary hooks
+static inline FACET_VALUE *FACET_VALUE_GET_FROM_INDEX(FACET_KEY *k, FACETS_HASH hash) {
+    SIMPLE_HASHTABLE_SLOT *slot = simple_hashtable_get_slot(&k->values.ht, hash);
+    return slot->data;
+}
 
-static void facet_value_insert_callback(const DICTIONARY_ITEM *item __maybe_unused, void *value, void *data) {
-    FACET_VALUE *v = value;
-    FACET_KEY *k = data;
+static inline FACET_VALUE *FACET_VALUE_ADD_TO_INDEX(FACET_KEY *k, const FACET_VALUE * const tv) {
+    SIMPLE_HASHTABLE_SLOT *slot = simple_hashtable_get_slot(&k->values.ht, tv->hash);
+
+    if(slot->data) {
+        // already exists
+
+        FACET_VALUE *v = slot->data;
+        FACET_VALUE_ADD_CONFLICT(k, v, tv);
+        return v;
+    }
+
+    // we have to add it
+
+    FACET_VALUE *v = mallocz(sizeof(*v));
+    slot->hash = tv->hash;
+    slot->data = v;
+
+    memcpy(v, tv, sizeof(*v));
+
+    DOUBLE_LINKED_LIST_APPEND_ITEM_UNSAFE(k->values.ll, v, prev, next);
+    k->values.used++;
 
     if(!v->selected)
         v->selected = k->default_selected_for_values;
 
-    if(v->name) {
+    if(v->name && v->name_len) {
         // an actual value, not a filter
-        v->name = strdupz(v->name);
+        v->name = facets_value_dup(v->name, v->name_len);
         facet_value_is_used(k, v);
     }
+    else {
+        v->name = NULL;
+        v->name_len = 0;
+    }
+
+    k->facets->operations.values.inserts++;
+
+    if(unlikely(k->values.used > k->values.ht.size / 2)) {
+        simple_hashtable_resize_double(&k->values.ht);
+        k->facets->operations.values.hashtable_increases++;
+    }
+
+    return v;
 }
 
-static bool facet_value_conflict_callback(const DICTIONARY_ITEM *item __maybe_unused, void *old_value, void *new_value, void *data) {
-    FACET_VALUE *v = old_value;
-    FACET_VALUE *nv = new_value;
-    FACET_KEY *k = data;
+static inline void FACET_VALUE_ADD_EMPTY_VALUE_TO_INDEX(FACET_KEY *k) {
+    static const FACET_VALUE tv = {
+            .hash = FACETS_HASH_ZERO,
+            .name = FACET_VALUE_UNSET,
+            .name_len = sizeof(FACET_VALUE_UNSET) - 1,
+    };
+
+    k->current_value.hash = FACETS_HASH_ZERO;
 
-    if(!v->name && nv->name)
-        // an actual value, not a filter
-        v->name = strdupz(nv->name);
+    if(k->empty_value.v) {
+        FACET_VALUE_ADD_CONFLICT(k, k->empty_value.v, &tv);
+        k->current_value.v = k->empty_value.v;
+    }
+    else {
+        FACET_VALUE *v = FACET_VALUE_ADD_TO_INDEX(k, &tv);
+        v->empty = true;
+        k->empty_value.v = v;
+        k->current_value.v = v;
+    }
+}
 
-    if(v->name)
-        facet_value_is_used(k, v);
+static inline void FACET_VALUE_ADD_CURRENT_VALUE_TO_INDEX(FACET_KEY *k) {
+    static __thread FACET_VALUE tv = { 0 };
 
-    internal_fatal(v->name && strcmp(v->name, nv->name) != 0, "hash conflict: '%s' and '%s' have the same hash '%s'", v->name, nv->name,
-                   dictionary_acquired_item_name(item));
+    internal_fatal(!facet_key_value_updated(k), "trying to add a non-updated value to the index");
 
-    return false;
+    tv.name = facets_key_get_value(k);
+    tv.name_len = facets_key_get_value_length(k);
+    tv.hash = FACETS_HASH_FUNCTION(tv.name, tv.name_len);
+
+    k->current_value.v = FACET_VALUE_ADD_TO_INDEX(k, &tv);
+    k->facets->operations.values.indexed++;
 }
 
-static void facet_value_delete_callback(const DICTIONARY_ITEM *item __maybe_unused, void *value, void *data __maybe_unused) {
-    FACET_VALUE *v = value;
-    freez((char *)v->name);
+static inline void FACET_VALUE_ADD_OR_UPDATE_SELECTED(FACET_KEY *k, FACETS_HASH hash) {
+    FACET_VALUE tv = {
+            .hash = hash,
+            .selected = true,
+            .name = NULL,
+            .name_len = 0,
+    };
+    FACET_VALUE_ADD_TO_INDEX(k, &tv);
 }
 
 // ----------------------------------------------------------------------------
-// FACET_KEY dictionary hooks
+// The FACET_KEY index within each FACET
+
+#define foreach_key_in_facets(facets, k) \
+    for((k) = (facets)->keys.ll; (k) ;(k) = (k)->next)
+
+#define foreach_key_in_facets_done(k) do { ; } while(0)
 
 static inline void facet_key_late_init(FACETS *facets, FACET_KEY *k) {
-    if(k->values)
+    if(k->values.enabled)
         return;
 
     if(facets_key_is_facet(facets, k)) {
-        k->values = dictionary_create_advanced(
-                DICT_OPTION_SINGLE_THREADED | DICT_OPTION_DONT_OVERWRITE_VALUE | DICT_OPTION_FIXED_SIZE,
-                NULL, sizeof(FACET_VALUE));
-        dictionary_register_insert_callback(k->values, facet_value_insert_callback, k);
-        dictionary_register_conflict_callback(k->values, facet_value_conflict_callback, k);
-        dictionary_register_delete_callback(k->values, facet_value_delete_callback, k);
+        FACETS_VALUES_INDEX_CREATE(k);
+        k->values.enabled = true;
+        if(facets->keys_with_values.used < FACETS_KEYS_WITH_VALUES_MAX)
+            facets->keys_with_values.array[facets->keys_with_values.used++] = k;
     }
 }
 
-static void facet_key_insert_callback(const DICTIONARY_ITEM *item __maybe_unused, void *value, void *data) {
-    FACET_KEY *k = value;
-    FACETS *facets = data;
+static inline void FACETS_KEYS_INDEX_CREATE(FACETS *facets) {
+    facets->keys.ll = NULL;
+    facets->keys.count = 0;
+    facets->keys_with_values.used = 0;
+
+    simple_hashtable_init(&facets->keys.ht, FACETS_KEYS_HASHTABLE_ENTRIES);
+}
+
+static inline void FACETS_KEYS_INDEX_DESTROY(FACETS *facets) {
+    FACET_KEY *k = facets->keys.ll;
+    while(k) {
+        FACET_KEY *next = k->next;
+
+        FACETS_VALUES_INDEX_DESTROY(k);
+        buffer_free(k->current_value.b);
+        freez((void *)k->name);
+        freez(k);
+
+        k = next;
+    }
+    facets->keys.ll = NULL;
+    facets->keys.count = 0;
+    facets->keys_with_values.used = 0;
+
+    simple_hashtable_free(&facets->keys.ht);
+}
+
+static inline FACET_KEY *FACETS_KEY_GET_FROM_INDEX(FACETS *facets, FACETS_HASH hash) {
+    SIMPLE_HASHTABLE_SLOT *slot = simple_hashtable_get_slot(&facets->keys.ht, hash);
+    return slot->data;
+}
+
+bool facets_key_name_value_length_is_selected(FACETS *facets, const char *key, size_t key_length, const char *value, size_t value_length) {
+    FACETS_HASH hash = FACETS_HASH_FUNCTION(key, key_length);
+    FACET_KEY *k = FACETS_KEY_GET_FROM_INDEX(facets, hash);
+    if(!k || k->default_selected_for_values)
+        return false;
+
+    hash = FACETS_HASH_FUNCTION(value, value_length);
+    FACET_VALUE *v = FACET_VALUE_GET_FROM_INDEX(k, hash);
+    return (v && v->selected) ? true : false;
+}
+
+void facets_add_possible_value_name_to_key(FACETS *facets, const char *key, size_t key_length, const char *value, size_t value_length) {
+    FACETS_HASH hash = FACETS_HASH_FUNCTION(key, key_length);
+    FACET_KEY *k = FACETS_KEY_GET_FROM_INDEX(facets, hash);
+    if(!k) return;
+
+    hash = FACETS_HASH_FUNCTION(value, value_length);
+    FACET_VALUE *v = FACET_VALUE_GET_FROM_INDEX(k, hash);
+    if(v && v->name && v->name_len) return;
+
+    FACET_VALUE tv = {
+            .hash = hash,
+            .name = value,
+            .name_len = value_length,
+    };
+    FACET_VALUE_ADD_TO_INDEX(k, &tv);
+}
+
+static void facet_key_set_name(FACET_KEY *k, const char *name, size_t name_length) {
+    internal_fatal(k->name && name && (strncmp(k->name, name, name_length) != 0 || k->name[name_length] != '\0'),
+            "key hash conflict: '%s' and '%s' have the same hash",
+            k->name, name);
+
+    if(likely(k->name || !name || !name_length))
+        return;
+
+    // an actual value, not a filter
+
+    char buf[name_length + 1];
+    memcpy(buf, name, name_length);
+    buf[name_length] = '\0';
+
+    internal_fatal(strchr(buf, '='), "found = in key");
+
+    k->name = strdupz(buf);
+    facet_key_late_init(k->facets, k);
+}
+
+static inline FACET_KEY *FACETS_KEY_CREATE(FACETS *facets, FACETS_HASH hash, const char *name, size_t name_length, FACET_KEY_OPTIONS options) {
+    facets->operations.keys.unique++;
+
+    FACET_KEY *k = callocz(1, sizeof(*k));
+
+    k->hash = hash;
+    k->facets = facets;
+    k->options = options;
+    k->current_value.b = buffer_create(sizeof(FACET_VALUE_UNSET), NULL);
+    k->default_selected_for_values = true;
 
     if(!(k->options & FACET_KEY_OPTION_REORDER))
         k->order = facets->order++;
@@ -256,56 +658,735 @@ static void facet_key_insert_callback(const DICTIONARY_ITEM *item __maybe_unused
     if((k->options & FACET_KEY_OPTION_FTS) || (facets->options & FACETS_OPTION_ALL_KEYS_FTS))
         facets->keys_filtered_by_query++;
 
-    if(k->name) {
-        // an actual value, not a filter
-        k->name = strdupz(k->name);
-        facet_key_late_init(facets, k);
-    }
+    facet_key_set_name(k, name, name_length);
 
-    k->current_value.b = buffer_create(0, NULL);
+    DOUBLE_LINKED_LIST_APPEND_ITEM_UNSAFE(facets->keys.ll, k, prev, next);
+    facets->keys.count++;
 
-    DOUBLE_LINKED_LIST_APPEND_ITEM_UNSAFE(facets->keys_ll, k, prev, next);
+    return k;
 }
 
-static bool facet_key_conflict_callback(const DICTIONARY_ITEM *item __maybe_unused, void *old_value, void *new_value, void *data) {
-    FACET_KEY *k = old_value;
-    FACET_KEY *nk = new_value;
-    FACETS *facets = data;
+static inline FACET_KEY *FACETS_KEY_ADD_TO_INDEX(FACETS *facets, FACETS_HASH hash, const char *name, size_t name_length, FACET_KEY_OPTIONS options) {
+    facets->operations.keys.registered++;
 
-    if(!k->name && nk->name) {
-        // an actual value, not a filter
-        k->name = strdupz(nk->name);
-        facet_key_late_init(facets, k);
+    SIMPLE_HASHTABLE_SLOT *slot = simple_hashtable_get_slot(&facets->keys.ht, hash);
+
+    if(unlikely(!slot->data)) {
+        // we have to add it
+        FACET_KEY *k = FACETS_KEY_CREATE(facets, hash, name, name_length, options);
+
+        slot->hash = hash;
+        slot->data = k;
+
+        if(facets->keys.count > facets->keys.ht.size / 2) {
+            simple_hashtable_resize_double(&facets->keys.ht);
+            facets->operations.keys.hashtable_increases++;
+        }
+
+        return k;
     }
 
-    if(k->options & FACET_KEY_OPTION_REORDER) {
+    // already in the index
+
+    FACET_KEY *k = slot->data;
+
+    facet_key_set_name(k, name, name_length);
+
+    if(unlikely(k->options & FACET_KEY_OPTION_REORDER)) {
         k->order = facets->order++;
         k->options &= ~FACET_KEY_OPTION_REORDER;
     }
 
-    return false;
+    return k;
+}
+
+bool facets_key_name_is_filter(FACETS *facets, const char *key) {
+    FACETS_HASH hash = FACETS_HASH_FUNCTION(key, strlen(key));
+    FACET_KEY *k = FACETS_KEY_GET_FROM_INDEX(facets, hash);
+    return (!k || k->default_selected_for_values) ? false : true;
+}
+
+bool facets_key_name_is_facet(FACETS *facets, const char *key) {
+    size_t key_len = strlen(key);
+    FACETS_HASH hash = FACETS_HASH_FUNCTION(key, key_len);
+    FACET_KEY *k = FACETS_KEY_ADD_TO_INDEX(facets, hash, key, key_len, 0);
+    return (k && (k->options & FACET_KEY_OPTION_FACET));
+}
+
+// ----------------------------------------------------------------------------
+
+static usec_t calculate_histogram_bar_width(usec_t after_ut, usec_t before_ut) {
+    // Array of valid durations in seconds
+    static time_t valid_durations_s[] = {
+            1, 2, 5, 10, 15, 30,                                                // seconds
+            1 * 60, 2 * 60, 3 * 60, 5 * 60, 10 * 60, 15 * 60, 30 * 60,          // minutes
+            1 * 3600, 2 * 3600, 6 * 3600, 8 * 3600, 12 * 3600,                  // hours
+            1 * 86400, 2 * 86400, 3 * 86400, 5 * 86400, 7 * 86400, 14 * 86400,  // days
+            1 * (30*86400)                                                      // months
+    };
+    static int array_size = sizeof(valid_durations_s) / sizeof(valid_durations_s[0]);
+
+    usec_t duration_ut = before_ut - after_ut;
+    usec_t bar_width_ut = 1 * USEC_PER_SEC;
+
+    for (int i = array_size - 1; i >= 0; --i) {
+        if (duration_ut / (valid_durations_s[i] * USEC_PER_SEC) >= HISTOGRAM_COLUMNS) {
+            bar_width_ut = valid_durations_s[i] * USEC_PER_SEC;
+            break;
+        }
+    }
+
+    return bar_width_ut;
+}
+
+static inline usec_t facets_histogram_slot_baseline_ut(FACETS *facets, usec_t ut) {
+    usec_t delta_ut = ut % facets->histogram.slot_width_ut;
+    return ut - delta_ut;
+}
+
+void facets_set_timeframe_and_histogram_by_id(FACETS *facets, const char *key_id, usec_t after_ut, usec_t before_ut) {
+    if(after_ut > before_ut) {
+        usec_t t = after_ut;
+        after_ut = before_ut;
+        before_ut = t;
+    }
+
+    facets->histogram.enabled = true;
+
+    if(key_id && *key_id && strlen(key_id) == FACET_STRING_HASH_SIZE - 1) {
+        facets->histogram.chart = strdupz(key_id);
+        facets->histogram.hash = str_to_facets_hash(facets->histogram.chart);
+    }
+    else {
+        freez(facets->histogram.chart);
+        facets->histogram.chart = NULL;
+        facets->histogram.hash = FACETS_HASH_ZERO;
+    }
+
+    facets->timeframe.after_ut = after_ut;
+    facets->timeframe.before_ut = before_ut;
+
+    facets->histogram.slot_width_ut = calculate_histogram_bar_width(after_ut, before_ut);
+    facets->histogram.after_ut = facets_histogram_slot_baseline_ut(facets, after_ut);
+    facets->histogram.before_ut = facets_histogram_slot_baseline_ut(facets, before_ut) + facets->histogram.slot_width_ut;
+    facets->histogram.slots = (facets->histogram.before_ut - facets->histogram.after_ut) / facets->histogram.slot_width_ut + 1;
+
+    internal_fatal(after_ut < facets->histogram.after_ut, "histogram after_ut is not less or equal to wanted after_ut");
+    internal_fatal(before_ut > facets->histogram.before_ut, "histogram before_ut is not more or equal to wanted before_ut");
+
+    if(facets->histogram.slots > 1000) {
+        facets->histogram.slots = 1000 + 1;
+        facets->histogram.slot_width_ut = (facets->histogram.before_ut - facets->histogram.after_ut) / 1000;
+    }
+}
+
+void facets_set_timeframe_and_histogram_by_name(FACETS *facets, const char *key_name, usec_t after_ut, usec_t before_ut) {
+    char hash_str[FACET_STRING_HASH_SIZE];
+    FACETS_HASH hash = FACETS_HASH_FUNCTION(key_name, strlen(key_name));
+    facets_hash_to_str(hash, hash_str);
+    facets_set_timeframe_and_histogram_by_id(facets, hash_str, after_ut, before_ut);
+}
+
+static inline void facets_histogram_update_value(FACETS *facets, usec_t usec) {
+    if(!facets->histogram.enabled ||
+            !facets->histogram.key ||
+            !facets->histogram.key->values.enabled ||
+            !facet_key_value_updated(facets->histogram.key) ||
+            usec < facets->histogram.after_ut ||
+            usec > facets->histogram.before_ut)
+        return;
+
+    FACET_VALUE *v = facets->histogram.key->current_value.v;
+
+    if(unlikely(!v->histogram))
+        v->histogram = callocz(facets->histogram.slots, sizeof(*v->histogram));
+
+    usec_t base_ut = facets_histogram_slot_baseline_ut(facets, usec);
+
+    if(unlikely(base_ut < facets->histogram.after_ut))
+        base_ut = facets->histogram.after_ut;
+
+    if(unlikely(base_ut > facets->histogram.before_ut))
+        base_ut = facets->histogram.before_ut;
+
+    uint32_t slot = (base_ut - facets->histogram.after_ut) / facets->histogram.slot_width_ut;
+
+    if(unlikely(slot >= facets->histogram.slots))
+        slot = facets->histogram.slots - 1;
+
+    v->histogram[slot]++;
+}
+
+static inline void facets_histogram_value_names(BUFFER *wb, FACETS *facets __maybe_unused, FACET_KEY *k, const char *key, const char *first_key) {
+    BUFFER *tb = NULL;
+
+    buffer_json_member_add_array(wb, key);
+    {
+        if(first_key)
+            buffer_json_add_array_item_string(wb, first_key);
+
+        if(k && k->values.enabled) {
+            FACET_VALUE *v;
+            foreach_value_in_key(k, v) {
+                if (unlikely(!v->histogram))
+                    continue;
+
+                if(!v->empty && k->transform.cb && k->transform.view_only) {
+                    if(!tb)
+                        tb = buffer_create(0, NULL);
+
+                    buffer_contents_replace(tb, v->name, v->name_len);
+                    k->transform.cb(facets, tb, FACETS_TRANSFORM_HISTOGRAM, k->transform.data);
+                    buffer_json_add_array_item_string(wb, buffer_tostring(tb));
+                }
+                else
+                    buffer_json_add_array_item_string(wb, v->name);
+            }
+            foreach_value_in_key_done(v);
+        }
+    }
+    buffer_json_array_close(wb); // key
+
+    buffer_free(tb);
+}
+
+static inline void facets_histogram_value_units(BUFFER *wb, FACETS *facets __maybe_unused, FACET_KEY *k, const char *key) {
+    buffer_json_member_add_array(wb, key);
+    {
+        if(k && k->values.enabled) {
+            FACET_VALUE *v;
+            foreach_value_in_key(k, v) {
+                if (unlikely(!v->histogram))
+                    continue;
+
+                buffer_json_add_array_item_string(wb, "events");
+            }
+            foreach_value_in_key_done(v);
+        }
+    }
+    buffer_json_array_close(wb); // key
+}
+
+static inline void facets_histogram_value_min(BUFFER *wb, FACETS *facets __maybe_unused, FACET_KEY *k, const char *key) {
+    buffer_json_member_add_array(wb, key);
+    {
+        if(k && k->values.enabled) {
+            FACET_VALUE *v;
+            foreach_value_in_key(k, v) {
+                if (unlikely(!v->histogram))
+                    continue;
+
+                buffer_json_add_array_item_uint64(wb, v->min);
+            }
+            foreach_value_in_key_done(v);
+        }
+    }
+    buffer_json_array_close(wb); // key
+}
+
+static inline void facets_histogram_value_max(BUFFER *wb, FACETS *facets __maybe_unused, FACET_KEY *k, const char *key) {
+    buffer_json_member_add_array(wb, key);
+    {
+        if(k && k->values.enabled) {
+            FACET_VALUE *v;
+            foreach_value_in_key(k, v) {
+                if (unlikely(!v->histogram))
+                    continue;
+
+                buffer_json_add_array_item_uint64(wb, v->max);
+            }
+            foreach_value_in_key_done(v);
+        }
+    }
+    buffer_json_array_close(wb); // key
+}
+
+static inline void facets_histogram_value_avg(BUFFER *wb, FACETS *facets __maybe_unused, FACET_KEY *k, const char *key) {
+    buffer_json_member_add_array(wb, key);
+    {
+        if(k && k->values.enabled) {
+            FACET_VALUE *v;
+            foreach_value_in_key(k, v) {
+                if (unlikely(!v->histogram))
+                    continue;
+
+                buffer_json_add_array_item_double(wb, (double) v->sum / (double) facets->histogram.slots);
+            }
+            foreach_value_in_key_done(v);
+        }
+    }
+    buffer_json_array_close(wb); // key
+}
+
+static inline void facets_histogram_value_arp(BUFFER *wb, FACETS *facets __maybe_unused, FACET_KEY *k, const char *key) {
+    buffer_json_member_add_array(wb, key);
+    {
+        if(k && k->values.enabled) {
+            FACET_VALUE *v;
+            foreach_value_in_key(k, v) {
+                if (unlikely(!v->histogram))
+                    continue;
+
+                buffer_json_add_array_item_uint64(wb, 0);
+            }
+            foreach_value_in_key_done(v);
+        }
+    }
+    buffer_json_array_close(wb); // key
+}
+
+static inline void facets_histogram_value_con(BUFFER *wb, FACETS *facets __maybe_unused, FACET_KEY *k, const char *key, uint32_t sum) {
+    buffer_json_member_add_array(wb, key);
+    {
+        if(k && k->values.enabled) {
+            FACET_VALUE *v;
+            foreach_value_in_key(k, v) {
+                if (unlikely(!v->histogram))
+                    continue;
+
+                buffer_json_add_array_item_double(wb, (double) v->sum * 100.0 / (double) sum);
+            }
+            foreach_value_in_key_done(v);
+        }
+    }
+    buffer_json_array_close(wb); // key
+}
+
+static void facets_histogram_generate(FACETS *facets, FACET_KEY *k, BUFFER *wb) {
+    size_t dimensions = 0;
+    uint32_t min = UINT32_MAX, max = 0, sum = 0, count = 0;
+
+    if(k && k->values.enabled) {
+        FACET_VALUE *v;
+        foreach_value_in_key(k, v) {
+            if (unlikely(!v->histogram))
+                continue;
+
+            dimensions++;
+
+            v->min = UINT32_MAX;
+            v->max = 0;
+            v->sum = 0;
+
+            for(uint32_t i = 0; i < facets->histogram.slots ;i++) {
+                uint32_t n = v->histogram[i];
+
+                if(n < min)
+                    min = n;
+
+                if(n > max)
+                    max = n;
+
+                sum += n;
+                count++;
+
+                if(n < v->min)
+                    v->min = n;
+
+                if(n > v->max)
+                    v->max = n;
+
+                v->sum += n;
+            }
+        }
+        foreach_value_in_key_done(v);
+    }
+
+    buffer_json_member_add_object(wb, "summary");
+    {
+        buffer_json_member_add_array(wb, "nodes");
+        {
+            buffer_json_add_array_item_object(wb); // node
+            {
+                buffer_json_member_add_string(wb, "mg", "default");
+                buffer_json_member_add_string(wb, "nm", "facets.histogram");
+                buffer_json_member_add_uint64(wb, "ni", 0);
+                buffer_json_member_add_object(wb, "st");
+                {
+                    buffer_json_member_add_uint64(wb, "ai", 0);
+                    buffer_json_member_add_uint64(wb, "code", 200);
+                    buffer_json_member_add_string(wb, "msg", "");
+                }
+                buffer_json_object_close(wb); // st
+
+                if(dimensions) {
+                    buffer_json_member_add_object(wb, "is");
+                    {
+                        buffer_json_member_add_uint64(wb, "sl", 1);
+                        buffer_json_member_add_uint64(wb, "qr", 1);
+                    }
+                    buffer_json_object_close(wb); // is
+
+                    buffer_json_member_add_object(wb, "ds");
+                    {
+                        buffer_json_member_add_uint64(wb, "sl", dimensions);
+                        buffer_json_member_add_uint64(wb, "qr", dimensions);
+                    }
+                    buffer_json_object_close(wb); // ds
+                }
+
+                if(count) {
+                    buffer_json_member_add_object(wb, "sts");
+                    {
+                        buffer_json_member_add_uint64(wb, "min", min);
+                        buffer_json_member_add_uint64(wb, "max", max);
+                        buffer_json_member_add_double(wb, "avg", (double) sum / (double) count);
+                        buffer_json_member_add_double(wb, "con", 100.0);
+                    }
+                    buffer_json_object_close(wb); // sts
+                }
+            }
+            buffer_json_object_close(wb); // node
+        }
+        buffer_json_array_close(wb); // nodes
+
+        buffer_json_member_add_array(wb, "contexts");
+        {
+            buffer_json_add_array_item_object(wb); // context
+            {
+                buffer_json_member_add_string(wb, "id", "facets.histogram");
+
+                if(dimensions) {
+                    buffer_json_member_add_object(wb, "is");
+                    {
+                        buffer_json_member_add_uint64(wb, "sl", 1);
+                        buffer_json_member_add_uint64(wb, "qr", 1);
+                    }
+                    buffer_json_object_close(wb); // is
+
+                    buffer_json_member_add_object(wb, "ds");
+                    {
+                        buffer_json_member_add_uint64(wb, "sl", dimensions);
+                        buffer_json_member_add_uint64(wb, "qr", dimensions);
+                    }
+                    buffer_json_object_close(wb); // ds
+                }
+
+                if(count) {
+                    buffer_json_member_add_object(wb, "sts");
+                    {
+                        buffer_json_member_add_uint64(wb, "min", min);
+                        buffer_json_member_add_uint64(wb, "max", max);
+                        buffer_json_member_add_double(wb, "avg", (double) sum / (double) count);
+                        buffer_json_member_add_double(wb, "con", 100.0);
+                    }
+                    buffer_json_object_close(wb); // sts
+                }
+            }
+            buffer_json_object_close(wb); // context
+        }
+        buffer_json_array_close(wb); // contexts
+
+        buffer_json_member_add_array(wb, "instances");
+        {
+            buffer_json_add_array_item_object(wb); // instance
+            {
+                buffer_json_member_add_string(wb, "id", "facets.histogram");
+                buffer_json_member_add_uint64(wb, "ni", 0);
+
+                if(dimensions) {
+                    buffer_json_member_add_object(wb, "ds");
+                    {
+                        buffer_json_member_add_uint64(wb, "sl", dimensions);
+                        buffer_json_member_add_uint64(wb, "qr", dimensions);
+                    }
+                    buffer_json_object_close(wb); // ds
+                }
+
+                if(count) {
+                    buffer_json_member_add_object(wb, "sts");
+                    {
+                        buffer_json_member_add_uint64(wb, "min", min);
+                        buffer_json_member_add_uint64(wb, "max", max);
+                        buffer_json_member_add_double(wb, "avg", (double) sum / (double) count);
+                        buffer_json_member_add_double(wb, "con", 100.0);
+                    }
+                    buffer_json_object_close(wb); // sts
+                }
+            }
+            buffer_json_object_close(wb); // instance
+        }
+        buffer_json_array_close(wb); // instances
+
+        buffer_json_member_add_array(wb, "dimensions");
+        if(dimensions && k && k->values.enabled) {
+            size_t pri = 0;
+            FACET_VALUE *v;
+            foreach_value_in_key(k, v) {
+                if(unlikely(!v->histogram))
+                    continue;
+
+                buffer_json_add_array_item_object(wb); // dimension
+                {
+                    buffer_json_member_add_string(wb, "id", v->name);
+                    buffer_json_member_add_object(wb, "ds");
+                    {
+                        buffer_json_member_add_uint64(wb, "sl", 1);
+                        buffer_json_member_add_uint64(wb, "qr", 1);
+                    }
+                    buffer_json_object_close(wb); // ds
+                    buffer_json_member_add_object(wb, "sts");
+                    {
+                        buffer_json_member_add_uint64(wb, "min", v->min);
+                        buffer_json_member_add_uint64(wb, "max", v->max);
+                        buffer_json_member_add_double(wb, "avg", (double)v->sum / (double)facets->histogram.slots);
+                        buffer_json_member_add_double(wb, "con", (double)v->sum * 100.0 / (double)sum);
+                    }
+                    buffer_json_object_close(wb); // sts
+                    buffer_json_member_add_uint64(wb, "pri", pri++);
+                }
+                buffer_json_object_close(wb); // dimension
+            }
+            foreach_value_in_key_done(v);
+        }
+        buffer_json_array_close(wb); // dimensions
+
+        buffer_json_member_add_array(wb, "labels");
+        buffer_json_array_close(wb); // labels
+
+        buffer_json_member_add_array(wb, "alerts");
+        buffer_json_array_close(wb); // alerts
+    }
+    buffer_json_object_close(wb); // summary
+
+    buffer_json_member_add_object(wb, "totals");
+    {
+        buffer_json_member_add_object(wb, "nodes");
+        {
+            buffer_json_member_add_uint64(wb, "sl", 1);
+            buffer_json_member_add_uint64(wb, "qr", 1);
+        }
+        buffer_json_object_close(wb); // nodes
+
+        if(dimensions) {
+            buffer_json_member_add_object(wb, "contexts");
+            {
+                buffer_json_member_add_uint64(wb, "sl", 1);
+                buffer_json_member_add_uint64(wb, "qr", 1);
+            }
+            buffer_json_object_close(wb); // contexts
+            buffer_json_member_add_object(wb, "instances");
+            {
+                buffer_json_member_add_uint64(wb, "sl", 1);
+                buffer_json_member_add_uint64(wb, "qr", 1);
+            }
+            buffer_json_object_close(wb); // instances
+
+            buffer_json_member_add_object(wb, "dimensions");
+            {
+                buffer_json_member_add_uint64(wb, "sl", dimensions);
+                buffer_json_member_add_uint64(wb, "qr", dimensions);
+            }
+            buffer_json_object_close(wb); // dimension
+        }
+    }
+    buffer_json_object_close(wb); // totals
+
+    buffer_json_member_add_object(wb, "result");
+    {
+        facets_histogram_value_names(wb, facets, k, "labels", "time");
+
+        buffer_json_member_add_object(wb, "point");
+        {
+            buffer_json_member_add_uint64(wb, "value", 0);
+            buffer_json_member_add_uint64(wb, "arp", 1);
+            buffer_json_member_add_uint64(wb, "pa", 2);
+        }
+        buffer_json_object_close(wb); // point
+
+        buffer_json_member_add_array(wb, "data");
+        if(k && k->values.enabled) {
+            usec_t t = facets->histogram.after_ut;
+            for(uint32_t i = 0; i < facets->histogram.slots ;i++) {
+                buffer_json_add_array_item_array(wb); // row
+                {
+                    buffer_json_add_array_item_time_ms(wb, t / USEC_PER_SEC);
+
+                    FACET_VALUE *v;
+                    foreach_value_in_key(k, v) {
+                        if (unlikely(!v->histogram))
+                            continue;
+
+                        buffer_json_add_array_item_array(wb); // point
+
+                        buffer_json_add_array_item_uint64(wb, v->histogram[i]);
+                        buffer_json_add_array_item_uint64(wb, 0); // arp - anomaly rate
+                        buffer_json_add_array_item_uint64(wb, 0); // pa - point annotation
+
+                        buffer_json_array_close(wb); // point
+                    }
+                    foreach_value_in_key_done(v);
+                }
+                buffer_json_array_close(wb); // row
+
+                t += facets->histogram.slot_width_ut;
+            }
+        }
+        buffer_json_array_close(wb); //data
+    }
+    buffer_json_object_close(wb); // result
+
+    buffer_json_member_add_object(wb, "db");
+    {
+        buffer_json_member_add_uint64(wb, "tiers", 1);
+        buffer_json_member_add_uint64(wb, "update_every", facets->histogram.slot_width_ut / USEC_PER_SEC);
+//        we should add these only when we know the retention of the db
+//        buffer_json_member_add_time_t(wb, "first_entry", facets->histogram.after_ut / USEC_PER_SEC);
+//        buffer_json_member_add_time_t(wb, "last_entry", facets->histogram.before_ut / USEC_PER_SEC);
+        buffer_json_member_add_string(wb, "units", "events");
+        buffer_json_member_add_object(wb, "dimensions");
+        {
+            facets_histogram_value_names(wb, facets, k, "ids", NULL);
+            facets_histogram_value_units(wb, facets, k, "units");
+
+            buffer_json_member_add_object(wb, "sts");
+            {
+                facets_histogram_value_min(wb, facets, k, "min");
+                facets_histogram_value_max(wb, facets, k, "max");
+                facets_histogram_value_avg(wb, facets, k, "avg");
+                facets_histogram_value_arp(wb, facets, k, "arp");
+                facets_histogram_value_con(wb, facets, k, "con", sum);
+            }
+            buffer_json_object_close(wb); // sts
+        }
+        buffer_json_object_close(wb); // dimensions
+
+        buffer_json_member_add_array(wb, "per_tier");
+        {
+            buffer_json_add_array_item_object(wb); // tier0
+            {
+                buffer_json_member_add_uint64(wb, "tier", 0);
+                buffer_json_member_add_uint64(wb, "queries", 1);
+                buffer_json_member_add_uint64(wb, "points", count);
+                buffer_json_member_add_time_t(wb, "update_every", facets->histogram.slot_width_ut / USEC_PER_SEC);
+//                we should add these only when we know the retention of the db
+//                buffer_json_member_add_time_t(wb, "first_entry", facets->histogram.after_ut / USEC_PER_SEC);
+//                buffer_json_member_add_time_t(wb, "last_entry", facets->histogram.before_ut / USEC_PER_SEC);
+            }
+            buffer_json_object_close(wb); // tier0
+        }
+        buffer_json_array_close(wb); // per_tier
+    }
+    buffer_json_object_close(wb); // db
+
+    buffer_json_member_add_object(wb, "view");
+    {
+        char title[1024 + 1] = "Events Distribution";
+        FACET_KEY *kt = FACETS_KEY_GET_FROM_INDEX(facets, facets->histogram.hash);
+        if(kt && kt->name)
+            snprintfz(title, 1024, "Events Distribution by %s", kt->name);
+
+        buffer_json_member_add_string(wb, "title", title);
+        buffer_json_member_add_time_t(wb, "update_every", facets->histogram.slot_width_ut / USEC_PER_SEC);
+        buffer_json_member_add_time_t(wb, "after", facets->histogram.after_ut / USEC_PER_SEC);
+        buffer_json_member_add_time_t(wb, "before", facets->histogram.before_ut / USEC_PER_SEC);
+        buffer_json_member_add_string(wb, "units", "events");
+        buffer_json_member_add_string(wb, "chart_type", "stackedBar");
+        buffer_json_member_add_object(wb, "dimensions");
+        {
+            buffer_json_member_add_array(wb, "grouped_by");
+            {
+                buffer_json_add_array_item_string(wb, "dimension");
+            }
+            buffer_json_array_close(wb); // grouped_by
+
+            facets_histogram_value_names(wb, facets, k, "ids", NULL);
+            facets_histogram_value_names(wb, facets, k, "names", NULL);
+            facets_histogram_value_units(wb, facets, k, "units");
+
+            buffer_json_member_add_object(wb, "sts");
+            {
+                facets_histogram_value_min(wb, facets, k, "min");
+                facets_histogram_value_max(wb, facets, k, "max");
+                facets_histogram_value_avg(wb, facets, k, "avg");
+                facets_histogram_value_arp(wb, facets, k, "arp");
+                facets_histogram_value_con(wb, facets, k, "con", sum);
+            }
+            buffer_json_object_close(wb); // sts
+        }
+        buffer_json_object_close(wb); // dimensions
+
+        buffer_json_member_add_uint64(wb, "min", min);
+        buffer_json_member_add_uint64(wb, "max", max);
+    }
+    buffer_json_object_close(wb); // view
+
+    buffer_json_member_add_array(wb, "agents");
+    {
+        buffer_json_add_array_item_object(wb); // agent
+        {
+            buffer_json_member_add_string(wb, "mg", "default");
+            buffer_json_member_add_string(wb, "nm", "facets.histogram");
+            buffer_json_member_add_time_t(wb, "now", now_realtime_sec());
+            buffer_json_member_add_uint64(wb, "ai", 0);
+        }
+        buffer_json_object_close(wb); // agent
+    }
+    buffer_json_array_close(wb); // agents
 }
 
-static void facet_key_delete_callback(const DICTIONARY_ITEM *item __maybe_unused, void *value, void *data __maybe_unused) {
-    FACET_KEY *k = value;
-    FACETS *facets = data;
+// ----------------------------------------------------------------------------
+
+static inline void facet_value_is_used(FACET_KEY *k, FACET_VALUE *v) {
+    if(!k->key_found_in_row)
+        v->rows_matching_facet_value++;
+
+    k->key_found_in_row++;
+
+    if(v->selected)
+        k->key_values_selected_in_row++;
+}
+
+static inline bool facets_key_is_facet(FACETS *facets, FACET_KEY *k) {
+    bool included = true, excluded = false, never = false;
+
+    if(k->options & (FACET_KEY_OPTION_FACET | FACET_KEY_OPTION_NO_FACET | FACET_KEY_OPTION_NEVER_FACET)) {
+        if(k->options & FACET_KEY_OPTION_FACET) {
+            included = true;
+            excluded = false;
+            never = false;
+        }
+        else if(k->options & (FACET_KEY_OPTION_NO_FACET | FACET_KEY_OPTION_NEVER_FACET)) {
+            included = false;
+            excluded = true;
+            never = true;
+        }
+    }
+    else {
+        if (facets->included_keys) {
+            if (!simple_pattern_matches(facets->included_keys, k->name))
+                included = false;
+        }
+
+        if (facets->excluded_keys) {
+            if (simple_pattern_matches(facets->excluded_keys, k->name)) {
+                excluded = true;
+                never = true;
+            }
+        }
+    }
 
-    DOUBLE_LINKED_LIST_REMOVE_ITEM_UNSAFE(facets->keys_ll, k, prev, next);
+    if(included && !excluded) {
+        k->options |= FACET_KEY_OPTION_FACET;
+        k->options &= ~FACET_KEY_OPTION_NO_FACET;
+        return true;
+    }
 
-    dictionary_destroy(k->values);
-    buffer_free(k->current_value.b);
-    freez((char *)k->name);
+    k->options |= FACET_KEY_OPTION_NO_FACET;
+    k->options &= ~FACET_KEY_OPTION_FACET;
+
+    if(never)
+        k->options |= FACET_KEY_OPTION_NEVER_FACET;
+
+    return false;
 }
 
 // ----------------------------------------------------------------------------
 
-FACETS *facets_create(uint32_t items_to_return, usec_t anchor, FACETS_OPTIONS options, const char *visible_keys, const char *facet_keys, const char *non_facet_keys) {
+FACETS *facets_create(uint32_t items_to_return, FACETS_OPTIONS options, const char *visible_keys, const char *facet_keys, const char *non_facet_keys) {
     FACETS *facets = callocz(1, sizeof(FACETS));
     facets->options = options;
-    facets->keys = dictionary_create_advanced(DICT_OPTION_SINGLE_THREADED|DICT_OPTION_DONT_OVERWRITE_VALUE|DICT_OPTION_FIXED_SIZE, NULL, sizeof(FACET_KEY));
-    dictionary_register_insert_callback(facets->keys, facet_key_insert_callback, facets);
-    dictionary_register_conflict_callback(facets->keys, facet_key_conflict_callback, facets);
-    dictionary_register_delete_callback(facets->keys, facet_key_delete_callback, facets);
+    FACETS_KEYS_INDEX_CREATE(facets);
 
     if(facet_keys && *facet_keys)
         facets->included_keys = simple_pattern_create(facet_keys, "|", SIMPLE_PATTERN_EXACT, true);
@@ -316,8 +1397,10 @@ FACETS *facets_create(uint32_t items_to_return, usec_t anchor, FACETS_OPTIONS op
     if(visible_keys && *visible_keys)
         facets->visible_keys = simple_pattern_create(visible_keys, "|", SIMPLE_PATTERN_EXACT, true);
 
-    facets->max_items_to_return = items_to_return;
-    facets->anchor = anchor;
+    facets->max_items_to_return = items_to_return > 1 ? items_to_return : 2;
+    facets->anchor.start_ut = 0;
+    facets->anchor.stop_ut = 0;
+    facets->anchor.direction = FACETS_ANCHOR_DIRECTION_BACKWARD;
     facets->order = 1;
 
     return facets;
@@ -325,7 +1408,7 @@ FACETS *facets_create(uint32_t items_to_return, usec_t anchor, FACETS_OPTIONS op
 
 void facets_destroy(FACETS *facets) {
     dictionary_destroy(facets->accepted_params);
-    dictionary_destroy(facets->keys);
+    FACETS_KEYS_INDEX_DESTROY(facets);
     simple_pattern_free(facets->visible_keys);
     simple_pattern_free(facets->included_keys);
     simple_pattern_free(facets->excluded_keys);
@@ -337,6 +1420,7 @@ void facets_destroy(FACETS *facets) {
         facets_row_free(facets, r);
     }
 
+    freez(facets->histogram.chart);
     freez(facets);
 }
 
@@ -347,26 +1431,24 @@ void facets_accepted_param(FACETS *facets, const char *param) {
     dictionary_set(facets->accepted_params, param, NULL, 0);
 }
 
-inline FACET_KEY *facets_register_key(FACETS *facets, const char *key, FACET_KEY_OPTIONS options) {
-    FACET_KEY tk = {
-            .name = key,
-            .options = options,
-            .default_selected_for_values = true,
-    };
-    char hash[FACET_STRING_HASH_SIZE];
-    facets_string_hash(tk.name, hash);
-    return dictionary_set(facets->keys, hash, &tk, sizeof(tk));
+static inline FACET_KEY *facets_register_key_name_length(FACETS *facets, const char *key, size_t key_length, FACET_KEY_OPTIONS options) {
+    return FACETS_KEY_ADD_TO_INDEX(facets, FACETS_HASH_FUNCTION(key, key_length), key, key_length, options);
 }
 
-inline FACET_KEY *facets_register_key_transformation(FACETS *facets, const char *key, FACET_KEY_OPTIONS options, facets_key_transformer_t cb, void *data) {
-    FACET_KEY *k = facets_register_key(facets, key, options);
+inline FACET_KEY *facets_register_key_name(FACETS *facets, const char *key, FACET_KEY_OPTIONS options) {
+    return facets_register_key_name_length(facets, key, strlen(key), options);
+}
+
+inline FACET_KEY *facets_register_key_name_transformation(FACETS *facets, const char *key, FACET_KEY_OPTIONS options, facets_key_transformer_t cb, void *data) {
+    FACET_KEY *k = facets_register_key_name(facets, key, options);
     k->transform.cb = cb;
     k->transform.data = data;
+    k->transform.view_only = (options & FACET_KEY_OPTION_TRANSFORM_VIEW) ? true : false;
     return k;
 }
 
-inline FACET_KEY *facets_register_dynamic_key(FACETS *facets, const char *key, FACET_KEY_OPTIONS options, facet_dynamic_row_t cb, void *data) {
-    FACET_KEY *k = facets_register_key(facets, key, options);
+inline FACET_KEY *facets_register_dynamic_key_name(FACETS *facets, const char *key, FACET_KEY_OPTIONS options, facet_dynamic_row_t cb, void *data) {
+    FACET_KEY *k = facets_register_key_name(facets, key, options);
     k->dynamic.cb = cb;
     k->dynamic.data = data;
     return k;
@@ -376,64 +1458,124 @@ void facets_set_query(FACETS *facets, const char *query) {
     if(!query)
         return;
 
-    facets->query = simple_pattern_create(query, " \t", SIMPLE_PATTERN_SUBSTRING, false);
+    facets->query = simple_pattern_create(query, "|", SIMPLE_PATTERN_SUBSTRING, false);
 }
 
 void facets_set_items(FACETS *facets, uint32_t items) {
-    facets->max_items_to_return = items;
+    facets->max_items_to_return = items > 1 ? items : 2;
+}
+
+void facets_set_anchor(FACETS *facets, usec_t start_ut, usec_t stop_ut, FACETS_ANCHOR_DIRECTION direction) {
+    facets->anchor.start_ut = start_ut;
+    facets->anchor.stop_ut = stop_ut;
+    facets->anchor.direction = direction;
+
+    if((facets->anchor.direction == FACETS_ANCHOR_DIRECTION_BACKWARD && facets->anchor.start_ut && facets->anchor.start_ut < facets->anchor.stop_ut) ||
+            (facets->anchor.direction == FACETS_ANCHOR_DIRECTION_FORWARD && facets->anchor.stop_ut && facets->anchor.stop_ut < facets->anchor.start_ut)) {
+        internal_error(true, "start and stop anchors are flipped");
+        facets->anchor.start_ut = stop_ut;
+        facets->anchor.stop_ut = start_ut;
+    }
 }
 
-void facets_set_anchor(FACETS *facets, usec_t anchor) {
-    facets->anchor = anchor;
+void facets_enable_slice_mode(FACETS *facets) {
+    facets->options |= FACETS_OPTION_DONT_SEND_EMPTY_VALUE_FACETS | FACETS_OPTION_SORT_FACETS_ALPHABETICALLY;
 }
 
-void facets_register_facet_filter(FACETS *facets, const char *key_id, char *value_ids, FACET_KEY_OPTIONS options) {
-    FACET_KEY tk = {
-            .options = options,
-    };
-    FACET_KEY *k = dictionary_set(facets->keys, key_id, &tk, sizeof(tk));
+inline FACET_KEY *facets_register_facet_id(FACETS *facets, const char *key_id, FACET_KEY_OPTIONS options) {
+    if(!is_valid_string_hash(key_id))
+        return NULL;
 
-    k->default_selected_for_values = false;
+    FACETS_HASH hash = str_to_facets_hash(key_id);
+
+    internal_error(strcmp(hash_to_static_string(hash), key_id) != 0,
+                   "Regenerating the user supplied key, does not produce the same hash string");
+
+    FACET_KEY *k = FACETS_KEY_ADD_TO_INDEX(facets, hash, NULL, 0, options);
     k->options |= FACET_KEY_OPTION_FACET;
     k->options &= ~FACET_KEY_OPTION_NO_FACET;
     facet_key_late_init(facets, k);
 
-    FACET_VALUE tv = {
-            .selected = true,
-    };
-    dictionary_set(k->values, value_ids, &tv, sizeof(tv));
+    return k;
+}
+
+void facets_register_facet_id_filter(FACETS *facets, const char *key_id, char *value_id, FACET_KEY_OPTIONS options) {
+    FACET_KEY *k = facets_register_facet_id(facets, key_id, options);
+    if(k) {
+        if(is_valid_string_hash(value_id)) {
+            k->default_selected_for_values = false;
+            FACET_VALUE_ADD_OR_UPDATE_SELECTED(k, str_to_facets_hash(value_id));
+        }
+    }
+}
+
+void facets_set_current_row_severity(FACETS *facets, FACET_ROW_SEVERITY severity) {
+    facets->current_row.severity = severity;
+}
+
+void facets_register_row_severity(FACETS *facets, facet_row_severity_t cb, void *data) {
+    facets->severity.cb = cb;
+    facets->severity.data = data;
+}
+
+void facets_set_additional_options(FACETS *facets, FACETS_OPTIONS options) {
+    facets->options |= options;
 }
 
 // ----------------------------------------------------------------------------
 
-static inline void facets_check_value(FACETS *facets __maybe_unused, FACET_KEY *k) {
-    if(!k->current_value.updated)
-        buffer_flush(k->current_value.b);
+static inline void facets_key_set_empty_value(FACETS *facets, FACET_KEY *k) {
+    if(likely(!facet_key_value_updated(k) && facets->keys_in_row.used < FACETS_KEYS_IN_ROW_MAX))
+        facets->keys_in_row.array[facets->keys_in_row.used++] = k;
+
+    k->current_value.flags |= FACET_KEY_VALUE_UPDATED | FACET_KEY_VALUE_EMPTY;
 
-    if(k->transform.cb)
-        k->transform.cb(facets, k->current_value.b, k->transform.data);
+    facets->operations.values.registered++;
+    facets->operations.values.empty++;
 
-    if(!k->current_value.updated) {
-        buffer_strcat(k->current_value.b, FACET_VALUE_UNSET);
-        k->current_value.updated = true;
+    // no need to copy the UNSET value
+    // empty values are exported as empty
+    k->current_value.raw = NULL;
+    k->current_value.raw_len = 0;
+    k->current_value.b->len = 0;
+    k->current_value.flags &= ~FACET_KEY_VALUE_COPIED;
+
+    if(unlikely(k->values.enabled))
+        FACET_VALUE_ADD_EMPTY_VALUE_TO_INDEX(k);
+    else {
+        k->key_found_in_row++;
+        k->key_values_selected_in_row++;
+    }
+}
+
+static inline void facets_key_check_value(FACETS *facets, FACET_KEY *k) {
+    if(likely(!facet_key_value_updated(k) && facets->keys_in_row.used < FACETS_KEYS_IN_ROW_MAX))
+        facets->keys_in_row.array[facets->keys_in_row.used++] = k;
+
+    k->current_value.flags |= FACET_KEY_VALUE_UPDATED;
+    k->current_value.flags &= ~FACET_KEY_VALUE_EMPTY;
+
+    facets->operations.values.registered++;
+
+    if(k->transform.cb && !k->transform.view_only) {
+        facets->operations.values.transformed++;
+        facets_key_value_copy_to_buffer(k);
+        k->transform.cb(facets, k->current_value.b, FACETS_TRANSFORM_VALUE, k->transform.data);
     }
 
 //    bool found = false;
 //    if(strstr(buffer_tostring(k->current_value), "fprintd") != NULL)
 //        found = true;
 
-    if(facets->query && ((k->options & FACET_KEY_OPTION_FTS) || facets->options & FACETS_OPTION_ALL_KEYS_FTS)) {
+    if(facets->query && !facet_key_value_empty(k) && ((k->options & FACET_KEY_OPTION_FTS) || facets->options & FACETS_OPTION_ALL_KEYS_FTS)) {
+        facets->operations.fts.searches++;
+        facets_key_value_copy_to_buffer(k);
         if(simple_pattern_matches(facets->query, buffer_tostring(k->current_value.b)))
-            facets->keys_matched_by_query++;
+            facets->current_row.keys_matched_by_query++;
     }
 
-    if(k->values) {
-        FACET_VALUE tk = {
-            .name = buffer_tostring(k->current_value.b),
-        };
-        facets_string_hash(tk.name, k->current_value.hash);
-        dictionary_set(k->values, k->current_value.hash, &tk, sizeof(tk));
-    }
+    if(k->values.enabled)
+        FACET_VALUE_ADD_CURRENT_VALUE_TO_INDEX(k);
     else {
         k->key_found_in_row++;
         k->key_values_selected_in_row++;
@@ -441,21 +1583,19 @@ static inline void facets_check_value(FACETS *facets __maybe_unused, FACET_KEY *
 }
 
 void facets_add_key_value(FACETS *facets, const char *key, const char *value) {
-    FACET_KEY *k = facets_register_key(facets, key, 0);
-    buffer_flush(k->current_value.b);
-    buffer_strcat(k->current_value.b, value);
-    k->current_value.updated = true;
+    FACET_KEY *k = facets_register_key_name(facets, key, 0);
+    k->current_value.raw = value;
+    k->current_value.raw_len = strlen(value);
 
-    facets_check_value(facets, k);
+    facets_key_check_value(facets, k);
 }
 
-void facets_add_key_value_length(FACETS *facets, const char *key, const char *value, size_t value_len) {
-    FACET_KEY *k = facets_register_key(facets, key, 0);
-    buffer_flush(k->current_value.b);
-    buffer_strncat(k->current_value.b, value, value_len);
-    k->current_value.updated = true;
+void facets_add_key_value_length(FACETS *facets, const char *key, size_t key_len, const char *value, size_t value_len) {
+    FACET_KEY *k = facets_register_key_name_length(facets, key, key_len, 0);
+    k->current_value.raw = value;
+    k->current_value.raw_len = value_len;
 
-    facets_check_value(facets, k);
+    facets_key_check_value(facets, k);
 }
 
 // ----------------------------------------------------------------------------
@@ -466,7 +1606,8 @@ static void facet_row_key_value_insert_callback(const DICTIONARY_ITEM *item __ma
     FACET_ROW *row = data; (void)row;
 
     rkv->wb = buffer_create(0, NULL);
-    buffer_strcat(rkv->wb, rkv->tmp && *rkv->tmp ? rkv->tmp : FACET_VALUE_UNSET);
+    if(!rkv->empty)
+        buffer_contents_replace(rkv->wb, rkv->tmp, rkv->tmp_len);
 }
 
 static bool facet_row_key_value_conflict_callback(const DICTIONARY_ITEM *item __maybe_unused, void *old_value, void *new_value, void *data) {
@@ -474,8 +1615,12 @@ static bool facet_row_key_value_conflict_callback(const DICTIONARY_ITEM *item __
     FACET_ROW_KEY_VALUE *n_rkv = new_value;
     FACET_ROW *row = data; (void)row;
 
-    buffer_flush(rkv->wb);
-    buffer_strcat(rkv->wb, n_rkv->tmp && *n_rkv->tmp ? n_rkv->tmp : FACET_VALUE_UNSET);
+    rkv->empty = n_rkv->empty;
+
+    if(!rkv->empty)
+        buffer_contents_replace(rkv->wb, n_rkv->tmp, n_rkv->tmp_len);
+    else
+        buffer_flush(rkv->wb);
 
     return false;
 }
@@ -498,52 +1643,52 @@ static void facets_row_free(FACETS *facets __maybe_unused, FACET_ROW *row) {
 static FACET_ROW *facets_row_create(FACETS *facets, usec_t usec, FACET_ROW *into) {
     FACET_ROW *row;
 
-    if(into)
+    if(into) {
         row = into;
+        facets->operations.rows.reused++;
+    }
     else {
         row = callocz(1, sizeof(FACET_ROW));
         row->dict = dictionary_create_advanced(DICT_OPTION_SINGLE_THREADED|DICT_OPTION_DONT_OVERWRITE_VALUE|DICT_OPTION_FIXED_SIZE, NULL, sizeof(FACET_ROW_KEY_VALUE));
         dictionary_register_insert_callback(row->dict, facet_row_key_value_insert_callback, row);
         dictionary_register_conflict_callback(row->dict, facet_row_key_value_conflict_callback, row);
         dictionary_register_delete_callback(row->dict, facet_row_key_value_delete_callback, row);
+        facets->operations.rows.created++;
     }
 
+    row->severity = facets->current_row.severity;
     row->usec = usec;
 
     FACET_KEY *k;
-    dfe_start_read(facets->keys, k) {
+    foreach_key_in_facets(facets, k) {
         FACET_ROW_KEY_VALUE t = {
-                .tmp = (k->current_value.updated && buffer_strlen(k->current_value.b)) ?
-                        buffer_tostring(k->current_value.b) : FACET_VALUE_UNSET,
+                .tmp = NULL,
+                .tmp_len = 0,
                 .wb = NULL,
+                .empty = true,
         };
+
+        if(facet_key_value_updated(k) && !facet_key_value_empty(k)) {
+            t.tmp = facets_key_get_value(k);
+            t.tmp_len = facets_key_get_value_length(k);
+            t.empty = false;
+        }
+
         dictionary_set(row->dict, k->name, &t, sizeof(t));
     }
-    dfe_done(k);
+    foreach_key_in_facets_done(k);
 
     return row;
 }
 
 // ----------------------------------------------------------------------------
 
-static void facets_row_keep(FACETS *facets, usec_t usec) {
-    facets->operations.matched++;
-
-    if(usec < facets->anchor) {
-        facets->operations.skips_before++;
-        return;
-    }
-
-    if(unlikely(!facets->base)) {
-        facets->operations.last_added = facets_row_create(facets, usec, NULL);
-        DOUBLE_LINKED_LIST_APPEND_ITEM_UNSAFE(facets->base, facets->operations.last_added, prev, next);
-        facets->items_to_return++;
-        facets->operations.first++;
-        return;
-    }
+static inline FACET_ROW *facets_row_keep_seek_to_position(FACETS *facets, usec_t usec) {
+    if(usec < facets->base->prev->usec)
+        return facets->base->prev;
 
-    if(likely(usec > facets->base->prev->usec))
-        facets->operations.last_added = facets->base->prev;
+    if(usec > facets->base->usec)
+        return facets->base;
 
     FACET_ROW *last = facets->operations.last_added;
     while(last->prev != facets->base->prev && usec > last->prev->usec) {
@@ -556,137 +1701,248 @@ static void facets_row_keep(FACETS *facets, usec_t usec) {
         facets->operations.forwards++;
     }
 
-    if(facets->items_to_return >= facets->max_items_to_return) {
-        if(last == facets->base->prev && usec < last->usec) {
-            facets->operations.skips_after++;
-            return;
+    return last;
+}
+
+static void facets_row_keep_first_entry(FACETS *facets, usec_t usec) {
+    facets->operations.last_added = facets_row_create(facets, usec, NULL);
+    DOUBLE_LINKED_LIST_APPEND_ITEM_UNSAFE(facets->base, facets->operations.last_added, prev, next);
+    facets->items_to_return++;
+    facets->operations.first++;
+}
+
+static inline bool facets_is_entry_within_anchor(FACETS *facets, usec_t usec) {
+    if(facets->anchor.start_ut || facets->anchor.stop_ut) {
+        // we have an anchor key
+        // we don't want to keep rows on the other side of the direction
+
+        switch (facets->anchor.direction) {
+            default:
+            case FACETS_ANCHOR_DIRECTION_BACKWARD:
+                // we need to keep only the smaller timestamps
+                if (facets->anchor.start_ut && usec >= facets->anchor.start_ut) {
+                    facets->operations.skips_before++;
+                    return false;
+                }
+                if (facets->anchor.stop_ut && usec <= facets->anchor.stop_ut) {
+                    facets->operations.skips_after++;
+                    return false;
+                }
+                break;
+
+            case FACETS_ANCHOR_DIRECTION_FORWARD:
+                // we need to keep only the bigger timestamps
+                if (facets->anchor.start_ut && usec <= facets->anchor.start_ut) {
+                    facets->operations.skips_after++;
+                    return false;
+                }
+                if (facets->anchor.stop_ut && usec >= facets->anchor.stop_ut) {
+                    facets->operations.skips_before++;
+                    return false;
+                }
+                break;
         }
     }
 
-    facets->items_to_return++;
+    return true;
+}
+
+static void facets_row_keep(FACETS *facets, usec_t usec) {
+    facets->operations.rows.matched++;
+
+    if(unlikely(!facets->base)) {
+        // the first row to keep
+        facets_row_keep_first_entry(facets, usec);
+        return;
+    }
+
+    FACET_ROW *closest = facets_row_keep_seek_to_position(facets, usec);
+    FACET_ROW *to_replace = NULL;
+
+    if(likely(facets->items_to_return >= facets->max_items_to_return)) {
+        // we have enough items to return already
+
+        switch(facets->anchor.direction) {
+            default:
+            case FACETS_ANCHOR_DIRECTION_BACKWARD:
+                if(closest == facets->base->prev && usec < closest->usec) {
+                    // this is to the end of the list, belonging to the next page
+                    facets->operations.skips_after++;
+                    return;
+                }
+
+                // it seems we need to remove an item - the last one
+                to_replace = facets->base->prev;
+                if(closest == to_replace)
+                    closest = to_replace->prev;
+
+                break;
+
+            case FACETS_ANCHOR_DIRECTION_FORWARD:
+                if(closest == facets->base && usec > closest->usec) {
+                    // this is to the beginning of the list, belonging to the next page
+                    facets->operations.skips_before++;
+                    return;
+                }
+
+                // it seems we need to remove an item - the first one
+                to_replace = facets->base;
+                if(closest == to_replace)
+                    closest = to_replace->next;
 
-    if(usec > last->usec) {
-        if(facets->items_to_return > facets->max_items_to_return) {
-            facets->items_to_return--;
-            facets->operations.shifts++;
-            facets->operations.last_added = facets->base->prev;
-            DOUBLE_LINKED_LIST_REMOVE_ITEM_UNSAFE(facets->base, facets->operations.last_added, prev, next);
-            facets->operations.last_added = facets_row_create(facets, usec, facets->operations.last_added);
+                break;
         }
-        DOUBLE_LINKED_LIST_PREPEND_ITEM_UNSAFE(facets->base, facets->operations.last_added, prev, next);
-        facets->operations.prepends++;
+
+        facets->operations.shifts++;
+        facets->items_to_return--;
+        DOUBLE_LINKED_LIST_REMOVE_ITEM_UNSAFE(facets->base, to_replace, prev, next);
     }
-    else {
-        facets->operations.last_added = facets_row_create(facets, usec, NULL);
-        DOUBLE_LINKED_LIST_APPEND_ITEM_UNSAFE(facets->base, facets->operations.last_added, prev, next);
+
+    internal_fatal(!closest, "FACETS: closest cannot be NULL");
+    internal_fatal(closest == to_replace, "FACETS: closest cannot be the same as to_replace");
+
+    facets->operations.last_added = facets_row_create(facets, usec, to_replace);
+
+    if(usec < closest->usec) {
+        DOUBLE_LINKED_LIST_INSERT_ITEM_AFTER_UNSAFE(facets->base, closest, facets->operations.last_added, prev, next);
         facets->operations.appends++;
     }
+    else {
+        DOUBLE_LINKED_LIST_INSERT_ITEM_BEFORE_UNSAFE(facets->base, closest, facets->operations.last_added, prev, next);
+        facets->operations.prepends++;
+    }
 
-    while(facets->items_to_return > facets->max_items_to_return) {
-        // we have to remove something
+    facets->items_to_return++;
+}
 
-        FACET_ROW *tmp = facets->base->prev;
-        DOUBLE_LINKED_LIST_REMOVE_ITEM_UNSAFE(facets->base, tmp, prev, next);
-        facets->items_to_return--;
+static inline void facets_reset_key(FACET_KEY *k) {
+    k->key_found_in_row = 0;
+    k->key_values_selected_in_row = 0;
+    k->current_value.flags = FACET_KEY_VALUE_NONE;
+    k->current_value.hash = FACETS_HASH_ZERO;
+}
 
-        if(unlikely(facets->operations.last_added == tmp))
-            facets->operations.last_added = facets->base->prev;
+static void facets_reset_keys_with_value_and_row(FACETS *facets) {
+    size_t entries = facets->keys_in_row.used;
 
-        facets_row_free(facets, tmp);
-        facets->operations.shifts++;
+    for(size_t p = 0; p < entries ;p++) {
+        FACET_KEY *k = facets->keys_in_row.array[p];
+        facets_reset_key(k);
     }
+
+    facets->current_row.severity = FACET_ROW_SEVERITY_NORMAL;
+    facets->current_row.keys_matched_by_query = 0;
+    facets->keys_in_row.used = 0;
 }
 
 void facets_rows_begin(FACETS *facets) {
     FACET_KEY *k;
-    // dfe_start_read(facets->keys, k) {
-    for(k = facets->keys_ll ; k ; k = k->next) {
-        k->key_found_in_row = 0;
-        k->key_values_selected_in_row = 0;
-        k->current_value.updated = false;
-        k->current_value.hash[0] = '\0';
+    foreach_key_in_facets(facets, k) {
+        facets_reset_key(k);
     }
-    // dfe_done(k);
+    foreach_key_in_facets_done(k);
 
-    facets->keys_matched_by_query = 0;
+    facets->keys_in_row.used = 0;
+    facets_reset_keys_with_value_and_row(facets);
 }
 
-void facets_row_finished(FACETS *facets, usec_t usec) {
-    if(facets->query && facets->keys_filtered_by_query && !facets->keys_matched_by_query)
-        goto cleanup;
+bool facets_row_finished(FACETS *facets, usec_t usec) {
+    facets->operations.rows.evaluated++;
+
+    if(unlikely((facets->query && facets->keys_filtered_by_query && !facets->current_row.keys_matched_by_query) ||
+            (facets->timeframe.before_ut && usec > facets->timeframe.before_ut) ||
+            (facets->timeframe.after_ut && usec < facets->timeframe.after_ut))) {
+        // this row is not useful
+        // 1. not matched by full text search, or
+        // 2. not in our timeframe
+        facets_reset_keys_with_value_and_row(facets);
+        return false;
+    }
 
-    facets->operations.evaluated++;
+    bool within_anchor = facets_is_entry_within_anchor(facets, usec);
+    if(unlikely(!within_anchor && (facets->options & FACETS_OPTION_DATA_ONLY))) {
+        facets_reset_keys_with_value_and_row(facets);
+        return false;
+    }
 
-    uint32_t total_keys = 0;
-    uint32_t selected_by = 0;
+    size_t entries = facets->keys_with_values.used;
+    size_t total_keys = 0;
+    size_t selected_keys = 0;
 
-    FACET_KEY *k;
-    // dfe_start_read(facets->keys, k) {
-    for(k = facets->keys_ll ; k ; k = k->next) {
-        if(!k->key_found_in_row) {
+    for(size_t p = 0; p < entries ;p++) {
+        FACET_KEY *k = facets->keys_with_values.array[p];
+
+        if(!facet_key_value_updated(k))
             // put the FACET_VALUE_UNSET value into it
-            facets_check_value(facets, k);
-        }
+            facets_key_set_empty_value(facets, k);
 
-        internal_fatal(!k->key_found_in_row, "all keys should be found in the row at this point");
-        internal_fatal(k->key_found_in_row != 1, "all keys should be matched exactly once at this point");
-        internal_fatal(k->key_values_selected_in_row > 1, "key values are selected in row more than once");
+        total_keys++;
 
-        k->key_found_in_row = 1;
+        if(k->key_values_selected_in_row)
+            selected_keys++;
 
-        total_keys += k->key_found_in_row;
-        selected_by += (k->key_values_selected_in_row) ? 1 : 0;
+        if(unlikely(!facets->histogram.key && facets->histogram.hash == k->hash))
+            facets->histogram.key = k;
     }
-    // dfe_done(k);
 
-    if(selected_by >= total_keys - 1) {
-        uint32_t found = 0;
+    if(selected_keys >= total_keys - 1) {
+        size_t found = 0;
+        (void) found;
+
+        for(size_t p = 0; p < entries; p++) {
+            FACET_KEY *k = facets->keys_with_values.array[p];
 
-        // dfe_start_read(facets->keys, k){
-        for(k = facets->keys_ll ; k ; k = k->next) {
-            uint32_t counted_by = selected_by;
+            size_t counted_by = selected_keys;
 
-            if (counted_by != total_keys && !k->key_values_selected_in_row)
+            if(counted_by != total_keys && !k->key_values_selected_in_row)
                 counted_by++;
 
             if(counted_by == total_keys) {
-                if(k->values) {
-                    if(!k->current_value.hash[0])
-                        facets_string_hash(buffer_tostring(k->current_value.b), k->current_value.hash);
-
-                    FACET_VALUE *v = dictionary_get(k->values, k->current_value.hash);
-                    v->final_facet_value_counter++;
-                }
-
+                k->current_value.v->final_facet_value_counter++;
                 found++;
             }
         }
-        // dfe_done(k);
 
         internal_fatal(!found, "We should find at least one facet to count this row");
-        (void)found;
     }
 
-    if(selected_by == total_keys)
-        facets_row_keep(facets, usec);
+    if(selected_keys == total_keys) {
+        // we need to keep this row
+        facets_histogram_update_value(facets, usec);
+
+        if(within_anchor)
+            facets_row_keep(facets, usec);
+    }
+
+    facets_reset_keys_with_value_and_row(facets);
 
-cleanup:
-    facets_rows_begin(facets);
+    return selected_keys == total_keys;
 }
 
 // ----------------------------------------------------------------------------
 // output
 
-void facets_report(FACETS *facets, BUFFER *wb) {
-    buffer_json_member_add_boolean(wb, "show_ids", false);
-    buffer_json_member_add_boolean(wb, "has_history", true);
+static const char *facets_severity_to_string(FACET_ROW_SEVERITY severity) {
+    switch(severity) {
+        default:
+        case FACET_ROW_SEVERITY_NORMAL:
+            return "normal";
 
-    buffer_json_member_add_object(wb, "pagination");
-    buffer_json_member_add_boolean(wb, "enabled", true);
-    buffer_json_member_add_string(wb, "key", "anchor");
-    buffer_json_member_add_string(wb, "column", "timestamp");
-    buffer_json_object_close(wb);
+        case FACET_ROW_SEVERITY_DEBUG:
+            return "debug";
+
+        case FACET_ROW_SEVERITY_NOTICE:
+            return "notice";
+
+        case FACET_ROW_SEVERITY_WARNING:
+            return "warning";
+
+        case FACET_ROW_SEVERITY_CRITICAL:
+            return "critical";
+    }
+}
 
+void facets_accepted_parameters_to_json_array(FACETS *facets, BUFFER *wb, bool with_keys) {
     buffer_json_member_add_array(wb, "accepted_params");
     {
         if(facets->accepted_params) {
@@ -697,54 +1953,338 @@ void facets_report(FACETS *facets, BUFFER *wb) {
             dfe_done(t);
         }
 
-        FACET_KEY *k;
-        dfe_start_read(facets->keys, k) {
-            if(!k->values)
-                continue;
+        if(with_keys) {
+            FACET_KEY *k;
+            foreach_key_in_facets(facets, k){
+                if (!k->values.enabled)
+                    continue;
 
-            buffer_json_add_array_item_string(wb, k_dfe.name);
+                buffer_json_add_array_item_string(wb, hash_to_static_string(k->hash));
+            }
+            foreach_key_in_facets_done(k);
         }
-        dfe_done(k);
     }
     buffer_json_array_close(wb); // accepted_params
+}
+
+static int facets_keys_reorder_compar(const void *a, const void *b) {
+    const FACET_KEY *ak = *((const FACET_KEY **)a);
+    const FACET_KEY *bk = *((const FACET_KEY **)b);
+
+    const char *an = ak->name;
+    const char *bn = bk->name;
+
+    if(!an) an = "0";
+    if(!bn) bn = "0";
+
+    while(*an && ispunct(*an)) an++;
+    while(*bn && ispunct(*bn)) bn++;
+
+    return strcasecmp(an, bn);
+}
+
+void facets_sort_and_reorder_keys(FACETS *facets) {
+    size_t entries = facets->keys_with_values.used;
+    if(!entries)
+        return;
+
+    FACET_KEY *keys[entries];
+    memcpy(keys, facets->keys_with_values.array, sizeof(FACET_KEY *) * entries);
+
+    qsort(keys, entries, sizeof(FACET_KEY *), facets_keys_reorder_compar);
+
+    for(size_t i = 0; i < entries ;i++)
+        keys[i]->order = i + 1;
+}
+
+static int facets_key_values_reorder_by_name_compar(const void *a, const void *b) {
+    const FACET_VALUE *av = *((const FACET_VALUE **)a);
+    const FACET_VALUE *bv = *((const FACET_VALUE **)b);
+
+    const char *an = (av->name && av->name_len) ? av->name : "0";
+    const char *bn = (bv->name && bv->name_len) ? bv->name : "0";
+
+    while(*an && ispunct(*an)) an++;
+    while(*bn && ispunct(*bn)) bn++;
+
+    int ret = strcasecmp(an, bn);
+    return ret;
+}
+
+static int facets_key_values_reorder_by_count_compar(const void *a, const void *b) {
+    const FACET_VALUE *av = *((const FACET_VALUE **)a);
+    const FACET_VALUE *bv = *((const FACET_VALUE **)b);
+
+    if(av->final_facet_value_counter < bv->final_facet_value_counter)
+        return 1;
+
+    if(av->final_facet_value_counter > bv->final_facet_value_counter)
+        return -1;
+
+    return facets_key_values_reorder_by_name_compar(a, b);
+}
+
+static int facets_key_values_reorder_by_name_numeric_compar(const void *a, const void *b) {
+    const FACET_VALUE *av = *((const FACET_VALUE **)a);
+    const FACET_VALUE *bv = *((const FACET_VALUE **)b);
 
-    buffer_json_member_add_array(wb, "facets");
+    const char *an = (av->name && av->name_len) ? av->name : "0";
+    const char *bn = (bv->name && bv->name_len) ? bv->name : "0";
+
+    if(strcmp(an, FACET_VALUE_UNSET) == 0) an = "0";
+    if(strcmp(bn, FACET_VALUE_UNSET) == 0) bn = "0";
+
+    int64_t ad = str2ll(an, NULL);
+    int64_t bd = str2ll(bn, NULL);
+
+    if(ad < bd)
+        return -1;
+
+    if(ad > bd)
+        return 1;
+
+    return facets_key_values_reorder_by_name_compar(a, b);
+}
+
+static uint32_t facets_sort_and_reorder_values_internal(FACET_KEY *k) {
+    bool all_values_numeric = true;
+    size_t entries = k->values.used;
+    FACET_VALUE *values[entries], *v;
+    uint32_t used = 0;
+    foreach_value_in_key(k, v) {
+        if((k->facets->options & FACETS_OPTION_DONT_SEND_EMPTY_VALUE_FACETS) && v->empty)
+            continue;
+
+        if(all_values_numeric && !v->empty && v->name && v->name_len) {
+            const char *s = v->name;
+            while(isdigit(*s)) s++;
+            if(*s != '\0')
+                all_values_numeric = false;
+        }
+
+        values[used++] = v;
+
+        if(used >= entries)
+            break;
+    }
+    foreach_value_in_key_done(v);
+
+    if(!used)
+        return 0;
+
+    if(k->facets->options & FACETS_OPTION_SORT_FACETS_ALPHABETICALLY) {
+        if(all_values_numeric)
+            qsort(values, used, sizeof(FACET_VALUE *), facets_key_values_reorder_by_name_numeric_compar);
+        else
+            qsort(values, used, sizeof(FACET_VALUE *), facets_key_values_reorder_by_name_compar);
+    }
+    else
+        qsort(values, used, sizeof(FACET_VALUE *), facets_key_values_reorder_by_count_compar);
+
+    for(size_t i = 0; i < used; i++)
+        values[i]->order = i + 1;
+
+    return used;
+}
+
+static uint32_t facets_sort_and_reorder_values(FACET_KEY *k) {
+    if(!k->values.enabled || !k->values.ll || !k->values.used)
+        return 0;
+
+    if(!k->transform.cb || !(k->facets->options & FACETS_OPTION_SORT_FACETS_ALPHABETICALLY))
+        return facets_sort_and_reorder_values_internal(k);
+
+    // we have a transformation and has to be sorted alphabetically
+
+    BUFFER *tb = buffer_create(0, NULL);
+    uint32_t ret = 0;
+
+    size_t entries = k->values.used;
+    struct {
+        const char *name;
+        uint32_t name_len;
+    } values[entries];
+    FACET_VALUE *v;
+    uint32_t used = 0;
+
+    foreach_value_in_key(k, v) {
+        if(used >= entries)
+            break;
+
+        values[used].name = v->name;
+        values[used].name_len = v->name_len;
+        used++;
+
+        buffer_contents_replace(tb, v->name, v->name_len);
+        k->transform.cb(k->facets, tb, FACETS_TRANSFORM_FACET_SORT, k->transform.data);
+        v->name = strdupz(buffer_tostring(tb));
+        v->name_len = buffer_strlen(tb);
+    }
+    foreach_value_in_key_done(v);
+
+    ret = facets_sort_and_reorder_values_internal(k);
+
+    used = 0;
+    foreach_value_in_key(k, v) {
+        if(used >= entries)
+            break;
+
+        freez((void *)v->name);
+        v->name = values[used].name;
+        v->name_len = values[used].name_len;
+        used++;
+    }
+    foreach_value_in_key_done(v);
+
+    buffer_free(tb);
+    return ret;
+}
+
+void facets_table_config(BUFFER *wb) {
+    buffer_json_member_add_boolean(wb, "show_ids", false);   // do not show the column ids to the user
+    buffer_json_member_add_boolean(wb, "has_history", true); // enable date-time picker with after-before
+
+    buffer_json_member_add_object(wb, "pagination");
     {
-        FACET_KEY *k;
-        dfe_start_read(facets->keys, k) {
-            if(!k->values)
-                continue;
+        buffer_json_member_add_boolean(wb, "enabled", true);
+        buffer_json_member_add_string(wb, "key", "anchor");
+        buffer_json_member_add_string(wb, "column", "timestamp");
+        buffer_json_member_add_string(wb, "units", "timestamp_usec");
+    }
+    buffer_json_object_close(wb); // pagination
+}
 
-            buffer_json_add_array_item_object(wb); // key
-            {
-                buffer_json_member_add_string(wb, "id", k_dfe.name);
-                buffer_json_member_add_string(wb, "name", k->name);
+static const char *facets_json_key_name_string(FACET_KEY *k, DICTIONARY *used_hashes_registry) {
+    if(k->name) {
+        if(used_hashes_registry && !k->default_selected_for_values) {
+            char hash_str[FACET_STRING_HASH_SIZE];
+            facets_hash_to_str(k->hash, hash_str);
+            dictionary_set(used_hashes_registry, hash_str, (void *)k->name, strlen(k->name) + 1);
+        }
 
-                if(!k->order)
-                    k->order = facets->order++;
+        return k->name;
+    }
 
-                buffer_json_member_add_uint64(wb, "order", k->order);
-                buffer_json_member_add_array(wb, "options");
+    // key has no name
+    const char *name = "[UNAVAILABLE_FIELD]";
+
+    if(used_hashes_registry) {
+        char hash_str[FACET_STRING_HASH_SIZE];
+        facets_hash_to_str(k->hash, hash_str);
+        const char *s = dictionary_get(used_hashes_registry, hash_str);
+        if(s) name = s;
+    }
+
+    return name;
+}
+
+static const char *facets_json_key_value_string(FACET_KEY *k, FACET_VALUE *v, DICTIONARY *used_hashes_registry) {
+    if(v->name && v->name_len) {
+        if(used_hashes_registry && !k->default_selected_for_values && v->selected) {
+            char hash_str[FACET_STRING_HASH_SIZE];
+            facets_hash_to_str(v->hash, hash_str);
+            dictionary_set(used_hashes_registry, hash_str, (void *)v->name, v->name_len + 1);
+        }
+
+        return v->name;
+    }
+
+    // key has no name
+    const char *name = "[unavailable field]";
+
+    if(used_hashes_registry) {
+        char hash_str[FACET_STRING_HASH_SIZE];
+        facets_hash_to_str(v->hash, hash_str);
+        const char *s = dictionary_get(used_hashes_registry, hash_str);
+        if(s) name = s;
+    }
+
+    return name;
+}
+
+void facets_report(FACETS *facets, BUFFER *wb, DICTIONARY *used_hashes_registry) {
+    if(!(facets->options & FACETS_OPTION_DATA_ONLY)) {
+        facets_table_config(wb);
+        facets_accepted_parameters_to_json_array(facets, wb, true);
+    }
+
+    // ------------------------------------------------------------------------
+    // facets
+
+    if(!(facets->options & FACETS_OPTION_DONT_SEND_FACETS)) {
+        bool show_facets = false;
+
+        if(facets->options & FACETS_OPTION_DATA_ONLY) {
+            if(facets->options & FACETS_OPTION_SHOW_DELTAS) {
+                buffer_json_member_add_array(wb, "facets_delta");
+                show_facets = true;
+            }
+        }
+        else {
+            buffer_json_member_add_array(wb, "facets");
+            show_facets = true;
+        }
+
+        if(show_facets) {
+            BUFFER *tb = NULL;
+            FACET_KEY *k;
+            foreach_key_in_facets(facets, k) {
+                if(!k->values.enabled)
+                    continue;
+
+                if(!facets_sort_and_reorder_values(k))
+                    // no values for this key
+                    continue;
+
+                buffer_json_add_array_item_object(wb); // key
                 {
-                    FACET_VALUE *v;
-                    dfe_start_read(k->values, v) {
-                        buffer_json_add_array_item_object(wb);
-                        {
-                            buffer_json_member_add_string(wb, "id", v_dfe.name);
-                            buffer_json_member_add_string(wb, "name", v->name);
-                            buffer_json_member_add_uint64(wb, "count", v->final_facet_value_counter);
+                    buffer_json_member_add_string(wb, "id", hash_to_static_string(k->hash));
+                    buffer_json_member_add_string(wb, "name", facets_json_key_name_string(k, used_hashes_registry));
+
+                    if(!k->order) k->order = facets->order++;
+                    buffer_json_member_add_uint64(wb, "order", k->order);
+
+                    buffer_json_member_add_array(wb, "options");
+                    {
+                        FACET_VALUE *v;
+                        foreach_value_in_key(k, v) {
+                            if((facets->options & FACETS_OPTION_DONT_SEND_EMPTY_VALUE_FACETS) && v->empty)
+                                continue;
+
+                            buffer_json_add_array_item_object(wb);
+                            {
+                                buffer_json_member_add_string(wb, "id", hash_to_static_string(v->hash));
+
+                                if(!v->empty && k->transform.cb && k->transform.view_only) {
+                                    if(!tb)
+                                        tb = buffer_create(0, NULL);
+
+                                    buffer_contents_replace(tb, v->name, v->name_len);
+                                    k->transform.cb(facets, tb, FACETS_TRANSFORM_FACET, k->transform.data);
+                                    buffer_json_member_add_string(wb, "name", buffer_tostring(tb));
+                                }
+                                else
+                                    buffer_json_member_add_string(wb, "name", facets_json_key_value_string(k, v, used_hashes_registry));
+
+                                buffer_json_member_add_uint64(wb, "count", v->final_facet_value_counter);
+                                buffer_json_member_add_uint64(wb, "order", v->order);
+                            }
+                            buffer_json_object_close(wb);
                         }
-                        buffer_json_object_close(wb);
+                        foreach_value_in_key_done(v);
                     }
-                    dfe_done(v);
+                    buffer_json_array_close(wb); // options
                 }
-                buffer_json_array_close(wb); // options
+                buffer_json_object_close(wb); // key
             }
-            buffer_json_object_close(wb); // key
+            foreach_key_in_facets_done(k);
+            buffer_free(tb);
+            buffer_json_array_close(wb); // facets
         }
-        dfe_done(k);
     }
-    buffer_json_array_close(wb); // facets
+
+    // ------------------------------------------------------------------------
+    // columns
 
     buffer_json_member_add_object(wb, "columns");
     {
@@ -755,56 +2295,94 @@ void facets_report(FACETS *facets, BUFFER *wb) {
                 RRDF_FIELD_TYPE_TIMESTAMP,
                 RRDF_FIELD_VISUAL_VALUE,
                 RRDF_FIELD_TRANSFORM_DATETIME_USEC, 0, NULL, NAN,
-                RRDF_FIELD_SORT_DESCENDING,
+                RRDF_FIELD_SORT_DESCENDING|RRDF_FIELD_SORT_FIXED,
                 NULL,
                 RRDF_FIELD_SUMMARY_COUNT,
                 RRDF_FIELD_FILTER_RANGE,
-                RRDF_FIELD_OPTS_VISIBLE | RRDF_FIELD_OPTS_UNIQUE_KEY,
+                RRDF_FIELD_OPTS_WRAP | RRDF_FIELD_OPTS_VISIBLE | RRDF_FIELD_OPTS_UNIQUE_KEY,
                 NULL);
 
-        FACET_KEY *k;
-        dfe_start_read(facets->keys, k) {
-            RRDF_FIELD_OPTIONS options = RRDF_FIELD_OPTS_NONE;
-            bool visible = k->options & (FACET_KEY_OPTION_VISIBLE|FACET_KEY_OPTION_STICKY);
-
-            if((facets->options & FACETS_OPTION_ALL_FACETS_VISIBLE && k->values))
-                visible = true;
-
-            if(!visible)
-                visible = simple_pattern_matches(facets->visible_keys, k->name);
-
-            if(visible)
-                options |= RRDF_FIELD_OPTS_VISIBLE;
-
-            if(k->options & FACET_KEY_OPTION_MAIN_TEXT)
-                options |= RRDF_FIELD_OPTS_FULL_WIDTH | RRDF_FIELD_OPTS_WRAP;
+        buffer_rrdf_table_add_field(
+                wb, field_id++,
+                "rowOptions", "rowOptions",
+                RRDF_FIELD_TYPE_NONE,
+                RRDR_FIELD_VISUAL_ROW_OPTIONS,
+                RRDF_FIELD_TRANSFORM_NONE, 0, NULL, NAN,
+                RRDF_FIELD_SORT_FIXED,
+                NULL,
+                RRDF_FIELD_SUMMARY_COUNT,
+                RRDF_FIELD_FILTER_NONE,
+                RRDR_FIELD_OPTS_DUMMY,
+                NULL);
 
-            buffer_rrdf_table_add_field(
-                    wb, field_id++,
-                    k_dfe.name, k->name ? k->name : k_dfe.name,
-                    RRDF_FIELD_TYPE_STRING,
-                    RRDF_FIELD_VISUAL_VALUE,
-                    RRDF_FIELD_TRANSFORM_NONE, 0, NULL, NAN,
-                    RRDF_FIELD_SORT_ASCENDING,
-                    NULL,
-                    RRDF_FIELD_SUMMARY_COUNT,
-                    k->values ? RRDF_FIELD_FILTER_FACET : RRDF_FIELD_FILTER_NONE,
-                    options,
-                    FACET_VALUE_UNSET);
-        }
-        dfe_done(k);
+        FACET_KEY *k;
+        foreach_key_in_facets(facets, k) {
+                    RRDF_FIELD_OPTIONS options = RRDF_FIELD_OPTS_WRAP;
+                    bool visible = k->options & (FACET_KEY_OPTION_VISIBLE | FACET_KEY_OPTION_STICKY);
+
+                    if ((facets->options & FACETS_OPTION_ALL_FACETS_VISIBLE && k->values.enabled))
+                        visible = true;
+
+                    if (!visible)
+                        visible = simple_pattern_matches(facets->visible_keys, k->name);
+
+                    if (visible)
+                        options |= RRDF_FIELD_OPTS_VISIBLE;
+
+                    if (k->options & FACET_KEY_OPTION_MAIN_TEXT)
+                        options |= RRDF_FIELD_OPTS_FULL_WIDTH | RRDF_FIELD_OPTS_WRAP;
+
+                    const char *hash_str = hash_to_static_string(k->hash);
+
+                    buffer_rrdf_table_add_field(
+                            wb, field_id++,
+                            hash_str, k->name ? k->name : hash_str,
+                            RRDF_FIELD_TYPE_STRING,
+                            (k->options & FACET_KEY_OPTION_RICH_TEXT) ? RRDF_FIELD_VISUAL_RICH : RRDF_FIELD_VISUAL_VALUE,
+                            RRDF_FIELD_TRANSFORM_NONE, 0, NULL, NAN,
+                            RRDF_FIELD_SORT_FIXED,
+                            NULL,
+                            RRDF_FIELD_SUMMARY_COUNT,
+                            (k->options & FACET_KEY_OPTION_NEVER_FACET) ? RRDF_FIELD_FILTER_NONE
+                                                                        : RRDF_FIELD_FILTER_FACET,
+                            options, FACET_VALUE_UNSET);
+                }
+        foreach_key_in_facets_done(k);
     }
     buffer_json_object_close(wb); // columns
 
+    // ------------------------------------------------------------------------
+    // rows data
+
     buffer_json_member_add_array(wb, "data");
     {
+        usec_t last_usec = 0; (void)last_usec;
+
         for(FACET_ROW *row = facets->base ; row ;row = row->next) {
+
+            internal_fatal(
+                    facets->anchor.start_ut && (
+                        (facets->anchor.direction == FACETS_ANCHOR_DIRECTION_BACKWARD && row->usec >= facets->anchor.start_ut) ||
+                        (facets->anchor.direction == FACETS_ANCHOR_DIRECTION_FORWARD && row->usec <= facets->anchor.start_ut)
+                    ), "Wrong data returned related to %s start anchor!", facets->anchor.direction == FACETS_ANCHOR_DIRECTION_FORWARD ? "forward" : "backward");
+
+            internal_fatal(last_usec && row->usec > last_usec, "Wrong order of data returned!");
+
+            last_usec = row->usec;
+
             buffer_json_add_array_item_array(wb); // each row
             buffer_json_add_array_item_uint64(wb, row->usec);
+            buffer_json_add_array_item_object(wb);
+            {
+                if(facets->severity.cb)
+                    row->severity = facets->severity.cb(facets, row, facets->severity.data);
+
+                buffer_json_member_add_string(wb, "severity", facets_severity_to_string(row->severity));
+            }
+            buffer_json_object_close(wb);
 
             FACET_KEY *k;
-            dfe_start_read(facets->keys, k)
-            {
+            foreach_key_in_facets(facets, k) {
                 FACET_ROW_KEY_VALUE *rkv = dictionary_get(row->dict, k->name);
 
                 if(unlikely(k->dynamic.cb)) {
@@ -812,32 +2390,117 @@ void facets_report(FACETS *facets, BUFFER *wb) {
                         rkv = dictionary_set(row->dict, k->name, NULL, sizeof(*rkv));
 
                     k->dynamic.cb(facets, wb, rkv, row, k->dynamic.data);
+                    facets->operations.values.dynamic++;
+                }
+                else {
+                    if(!rkv || rkv->empty) {
+                        buffer_json_add_array_item_string(wb, NULL);
+                    }
+                    else if(unlikely(k->transform.cb && k->transform.view_only)) {
+                        k->transform.cb(facets, rkv->wb, FACETS_TRANSFORM_DATA, k->transform.data);
+                        buffer_json_add_array_item_string(wb, buffer_tostring(rkv->wb));
+                    }
+                    else
+                        buffer_json_add_array_item_string(wb, buffer_tostring(rkv->wb));
                 }
-                else
-                	buffer_json_add_array_item_string(wb, rkv ? buffer_tostring(rkv->wb) : FACET_VALUE_UNSET);
             }
-            dfe_done(k);
+            foreach_key_in_facets_done(k);
             buffer_json_array_close(wb); // each row
         }
     }
     buffer_json_array_close(wb); // data
 
-    buffer_json_member_add_string(wb, "default_sort_column", "timestamp");
-    buffer_json_member_add_array(wb, "default_charts");
-    buffer_json_array_close(wb);
+    if(!(facets->options & FACETS_OPTION_DATA_ONLY)) {
+        buffer_json_member_add_string(wb, "default_sort_column", "timestamp");
+        buffer_json_member_add_array(wb, "default_charts");
+        buffer_json_array_close(wb);
+    }
 
-    buffer_json_member_add_object(wb, "items");
-    {
-        buffer_json_member_add_uint64(wb, "evaluated", facets->operations.evaluated);
-        buffer_json_member_add_uint64(wb, "matched", facets->operations.matched);
+    // ------------------------------------------------------------------------
+    // histogram
+
+    if(facets->histogram.enabled && !(facets->options & FACETS_OPTION_DONT_SEND_HISTOGRAM)) {
+        FACETS_HASH first_histogram_hash = 0;
+        buffer_json_member_add_array(wb, "available_histograms");
+        {
+            FACET_KEY *k;
+            foreach_key_in_facets(facets, k) {
+                if (!k->values.enabled)
+                    continue;
+
+                if(unlikely(!first_histogram_hash))
+                    first_histogram_hash = k->hash;
+
+                buffer_json_add_array_item_object(wb);
+                buffer_json_member_add_string(wb, "id", hash_to_static_string(k->hash));
+                buffer_json_member_add_string(wb, "name", k->name);
+                buffer_json_member_add_uint64(wb, "order", k->order);
+                buffer_json_object_close(wb);
+            }
+            foreach_key_in_facets_done(k);
+        }
+        buffer_json_array_close(wb);
+
+        {
+            FACET_KEY *k = FACETS_KEY_GET_FROM_INDEX(facets, facets->histogram.hash);
+            if(!k || !k->values.enabled)
+                k = FACETS_KEY_GET_FROM_INDEX(facets, first_histogram_hash);
+
+            bool show_histogram = false;
+
+            if(facets->options & FACETS_OPTION_DATA_ONLY) {
+                if(facets->options & FACETS_OPTION_SHOW_DELTAS) {
+                    buffer_json_member_add_object(wb, "histogram_delta");
+                    show_histogram = true;
+                }
+            }
+            else {
+                buffer_json_member_add_object(wb, "histogram");
+                show_histogram = true;
+            }
+
+            if(show_histogram) {
+                buffer_json_member_add_string(wb, "id", k ? hash_to_static_string(k->hash) : "");
+                buffer_json_member_add_string(wb, "name", k ? k->name : "");
+                buffer_json_member_add_object(wb, "chart");
+                {
+                    facets_histogram_generate(facets, k, wb);
+                }
+                buffer_json_object_close(wb); // chart
+                buffer_json_object_close(wb); // histogram
+            }
+        }
+    }
+
+    // ------------------------------------------------------------------------
+    // items
+
+    bool show_items = false;
+    if(facets->options & FACETS_OPTION_DATA_ONLY) {
+        if(facets->options & FACETS_OPTION_SHOW_DELTAS) {
+            buffer_json_member_add_object(wb, "items_delta");
+            show_items = true;
+        }
+    }
+    else {
+        buffer_json_member_add_object(wb, "items");
+        show_items = true;
+    }
+
+    if(show_items) {
+        buffer_json_member_add_uint64(wb, "evaluated", facets->operations.rows.evaluated);
+        buffer_json_member_add_uint64(wb, "matched", facets->operations.rows.matched);
         buffer_json_member_add_uint64(wb, "returned", facets->items_to_return);
         buffer_json_member_add_uint64(wb, "max_to_return", facets->max_items_to_return);
         buffer_json_member_add_uint64(wb, "before", facets->operations.skips_before);
         buffer_json_member_add_uint64(wb, "after", facets->operations.skips_after + facets->operations.shifts);
+        buffer_json_object_close(wb); // items
     }
-    buffer_json_object_close(wb); // items
 
-    buffer_json_member_add_object(wb, "stats");
+    // ------------------------------------------------------------------------
+    // stats
+
+    buffer_json_member_add_object(wb, "_stats");
     {
         buffer_json_member_add_uint64(wb, "first", facets->operations.first);
         buffer_json_member_add_uint64(wb, "forwards", facets->operations.forwards);
@@ -847,7 +2510,38 @@ void facets_report(FACETS *facets, BUFFER *wb) {
         buffer_json_member_add_uint64(wb, "prepends", facets->operations.prepends);
         buffer_json_member_add_uint64(wb, "appends", facets->operations.appends);
         buffer_json_member_add_uint64(wb, "shifts", facets->operations.shifts);
+        buffer_json_member_add_object(wb, "rows");
+        {
+            buffer_json_member_add_uint64(wb, "created", facets->operations.rows.created);
+            buffer_json_member_add_uint64(wb, "reused", facets->operations.rows.reused);
+            buffer_json_member_add_uint64(wb, "evaluated", facets->operations.rows.evaluated);
+            buffer_json_member_add_uint64(wb, "matched", facets->operations.rows.matched);
+        }
+        buffer_json_object_close(wb); // rows
+        buffer_json_member_add_object(wb, "keys");
+        {
+            buffer_json_member_add_uint64(wb, "registered", facets->operations.keys.registered);
+            buffer_json_member_add_uint64(wb, "unique", facets->operations.keys.unique);
+            buffer_json_member_add_uint64(wb, "hashtable_increases", facets->operations.keys.hashtable_increases);
+        }
+        buffer_json_object_close(wb); // keys
+        buffer_json_member_add_object(wb, "values");
+        {
+            buffer_json_member_add_uint64(wb, "registered", facets->operations.values.registered);
+            buffer_json_member_add_uint64(wb, "transformed", facets->operations.values.transformed);
+            buffer_json_member_add_uint64(wb, "dynamic", facets->operations.values.dynamic);
+            buffer_json_member_add_uint64(wb, "empty", facets->operations.values.empty);
+            buffer_json_member_add_uint64(wb, "indexed", facets->operations.values.indexed);
+            buffer_json_member_add_uint64(wb, "inserts", facets->operations.values.inserts);
+            buffer_json_member_add_uint64(wb, "conflicts", facets->operations.values.conflicts);
+            buffer_json_member_add_uint64(wb, "hashtable_increases", facets->operations.values.hashtable_increases);
+        }
+        buffer_json_object_close(wb); // values
+        buffer_json_member_add_object(wb, "fts");
+        {
+            buffer_json_member_add_uint64(wb, "searches", facets->operations.fts.searches);
+        }
+        buffer_json_object_close(wb); // fts
     }
     buffer_json_object_close(wb); // items
-
 }
diff --git a/libnetdata/facets/facets.h b/libnetdata/facets/facets.h
index 796d15a0..75972561 100644
--- a/libnetdata/facets/facets.h
+++ b/libnetdata/facets/facets.h
@@ -1,3 +1,5 @@
+// SPDX-License-Identifier: GPL-3.0-or-later
+
 #ifndef FACETS_H
 #define FACETS_H 1
 
@@ -6,59 +8,111 @@
 #define FACET_VALUE_UNSET "-"
 
 typedef enum __attribute__((packed)) {
-    FACET_KEY_OPTION_FACET      = (1 << 0), // filterable values
-    FACET_KEY_OPTION_NO_FACET   = (1 << 1), // non-filterable value
-    FACET_KEY_OPTION_STICKY     = (1 << 2), // should be sticky in the table
-    FACET_KEY_OPTION_VISIBLE    = (1 << 3), // should be in the default table
-    FACET_KEY_OPTION_FTS        = (1 << 4), // the key is filterable by full text search (FTS)
-    FACET_KEY_OPTION_MAIN_TEXT  = (1 << 5), // full width and wrap
-    FACET_KEY_OPTION_REORDER    = (1 << 6), // give the key a new order id on first encounter
+    FACETS_ANCHOR_DIRECTION_FORWARD,
+    FACETS_ANCHOR_DIRECTION_BACKWARD,
+} FACETS_ANCHOR_DIRECTION;
+
+typedef enum __attribute__((packed)) {
+    FACETS_TRANSFORM_VALUE,
+    FACETS_TRANSFORM_HISTOGRAM,
+    FACETS_TRANSFORM_FACET,
+    FACETS_TRANSFORM_DATA,
+    FACETS_TRANSFORM_FACET_SORT,
+} FACETS_TRANSFORMATION_SCOPE;
+
+typedef enum __attribute__((packed)) {
+    FACET_KEY_OPTION_FACET          = (1 << 0), // filterable values
+    FACET_KEY_OPTION_NO_FACET       = (1 << 1), // non-filterable value
+    FACET_KEY_OPTION_NEVER_FACET    = (1 << 2), // never enable this field as facet
+    FACET_KEY_OPTION_STICKY         = (1 << 3), // should be sticky in the table
+    FACET_KEY_OPTION_VISIBLE        = (1 << 4), // should be in the default table
+    FACET_KEY_OPTION_FTS            = (1 << 5), // the key is filterable by full text search (FTS)
+    FACET_KEY_OPTION_MAIN_TEXT      = (1 << 6), // full width and wrap
+    FACET_KEY_OPTION_RICH_TEXT      = (1 << 7),
+    FACET_KEY_OPTION_REORDER        = (1 << 8), // give the key a new order id on first encounter
+    FACET_KEY_OPTION_TRANSFORM_VIEW = (1 << 9), // when registering the transformation, do it only at the view, not on all data
 } FACET_KEY_OPTIONS;
 
+typedef enum __attribute__((packed)) {
+    FACET_ROW_SEVERITY_DEBUG,       // lowest - not important
+    FACET_ROW_SEVERITY_NORMAL,      // the default
+    FACET_ROW_SEVERITY_NOTICE,      // bold
+    FACET_ROW_SEVERITY_WARNING,     // yellow + bold
+    FACET_ROW_SEVERITY_CRITICAL,    // red + bold
+} FACET_ROW_SEVERITY;
+
 typedef struct facet_row_key_value {
     const char *tmp;
+    uint32_t tmp_len;
+    bool empty;
     BUFFER *wb;
 } FACET_ROW_KEY_VALUE;
 
 typedef struct facet_row {
     usec_t usec;
     DICTIONARY *dict;
+    FACET_ROW_SEVERITY severity;
     struct facet_row *prev, *next;
 } FACET_ROW;
 
 typedef struct facets FACETS;
 typedef struct facet_key FACET_KEY;
 
-#define FACET_STRING_HASH_SIZE 19
-void facets_string_hash(const char *src, char *out);
-
-typedef void (*facets_key_transformer_t)(FACETS *facets __maybe_unused, BUFFER *wb, void *data);
+typedef void (*facets_key_transformer_t)(FACETS *facets __maybe_unused, BUFFER *wb, FACETS_TRANSFORMATION_SCOPE scope, void *data);
 typedef void (*facet_dynamic_row_t)(FACETS *facets, BUFFER *json_array, FACET_ROW_KEY_VALUE *rkv, FACET_ROW *row, void *data);
-FACET_KEY *facets_register_dynamic_key(FACETS *facets, const char *key, FACET_KEY_OPTIONS options, facet_dynamic_row_t cb, void *data);
-FACET_KEY *facets_register_key_transformation(FACETS *facets, const char *key, FACET_KEY_OPTIONS options, facets_key_transformer_t cb, void *data);
+typedef FACET_ROW_SEVERITY (*facet_row_severity_t)(FACETS *facets, FACET_ROW *row, void *data);
+FACET_KEY *facets_register_dynamic_key_name(FACETS *facets, const char *key, FACET_KEY_OPTIONS options, facet_dynamic_row_t cb, void *data);
+FACET_KEY *facets_register_key_name_transformation(FACETS *facets, const char *key, FACET_KEY_OPTIONS options, facets_key_transformer_t cb, void *data);
+void facets_register_row_severity(FACETS *facets, facet_row_severity_t cb, void *data);
 
 typedef enum __attribute__((packed)) {
-    FACETS_OPTION_ALL_FACETS_VISIBLE    = (1 << 0), // all facets, should be visible by default in the table
-    FACETS_OPTION_ALL_KEYS_FTS          = (1 << 1), // all keys are searchable by full text search
+    FACETS_OPTION_ALL_FACETS_VISIBLE            = (1 << 0), // all facets should be visible by default in the table
+    FACETS_OPTION_ALL_KEYS_FTS                  = (1 << 1), // all keys are searchable by full text search
+    FACETS_OPTION_DONT_SEND_FACETS              = (1 << 2), // "facets" object will not be included in the report
+    FACETS_OPTION_DONT_SEND_HISTOGRAM           = (1 << 3), // "histogram" object will not be included in the report
+    FACETS_OPTION_DATA_ONLY                     = (1 << 4),
+    FACETS_OPTION_DONT_SEND_EMPTY_VALUE_FACETS  = (1 << 5), // empty facet values will not be included in the report
+    FACETS_OPTION_SORT_FACETS_ALPHABETICALLY    = (1 << 6),
+    FACETS_OPTION_SHOW_DELTAS                   = (1 << 7),
 } FACETS_OPTIONS;
 
-FACETS *facets_create(uint32_t items_to_return, usec_t anchor, FACETS_OPTIONS options, const char *visible_keys, const char *facet_keys, const char *non_facet_keys);
+FACETS *facets_create(uint32_t items_to_return, FACETS_OPTIONS options, const char *visible_keys, const char *facet_keys, const char *non_facet_keys);
 void facets_destroy(FACETS *facets);
 
 void facets_accepted_param(FACETS *facets, const char *param);
 
 void facets_rows_begin(FACETS *facets);
-void facets_row_finished(FACETS *facets, usec_t usec);
+bool facets_row_finished(FACETS *facets, usec_t usec);
 
-FACET_KEY *facets_register_key(FACETS *facets, const char *param, FACET_KEY_OPTIONS options);
+FACET_KEY *facets_register_key_name(FACETS *facets, const char *key, FACET_KEY_OPTIONS options);
 void facets_set_query(FACETS *facets, const char *query);
 void facets_set_items(FACETS *facets, uint32_t items);
-void facets_set_anchor(FACETS *facets, usec_t anchor);
-void facets_register_facet_filter(FACETS *facets, const char *key_id, char *value_ids, FACET_KEY_OPTIONS options);
+void facets_set_anchor(FACETS *facets, usec_t start_ut, usec_t stop_ut, FACETS_ANCHOR_DIRECTION direction);
+void facets_enable_slice_mode(FACETS *facets);
+
+FACET_KEY *facets_register_facet_id(FACETS *facets, const char *key_id, FACET_KEY_OPTIONS options);
+void facets_register_facet_id_filter(FACETS *facets, const char *key_id, char *value_id, FACET_KEY_OPTIONS options);
+void facets_set_timeframe_and_histogram_by_id(FACETS *facets, const char *key_id, usec_t after_ut, usec_t before_ut);
+void facets_set_timeframe_and_histogram_by_name(FACETS *facets, const char *key_name, usec_t after_ut, usec_t before_ut);
 
 void facets_add_key_value(FACETS *facets, const char *key, const char *value);
-void facets_add_key_value_length(FACETS *facets, const char *key, const char *value, size_t value_len);
+void facets_add_key_value_length(FACETS *facets, const char *key, size_t key_len, const char *value, size_t value_len);
+
+void facets_report(FACETS *facets, BUFFER *wb, DICTIONARY *used_hashes_registry);
+void facets_accepted_parameters_to_json_array(FACETS *facets, BUFFER *wb, bool with_keys);
+void facets_set_current_row_severity(FACETS *facets, FACET_ROW_SEVERITY severity);
+void facets_set_additional_options(FACETS *facets, FACETS_OPTIONS options);
+
+bool facets_key_name_is_filter(FACETS *facets, const char *key);
+bool facets_key_name_is_facet(FACETS *facets, const char *key);
+bool facets_key_name_value_length_is_selected(FACETS *facets, const char *key, size_t key_length, const char *value, size_t value_length);
+void facets_add_possible_value_name_to_key(FACETS *facets, const char *key, size_t key_length, const char *value, size_t value_length);
+
+void facets_sort_and_reorder_keys(FACETS *facets);
+usec_t facets_row_oldest_ut(FACETS *facets);
+usec_t facets_row_newest_ut(FACETS *facets);
+uint32_t facets_rows(FACETS *facets);
 
-void facets_report(FACETS *facets, BUFFER *wb);
+void facets_table_config(BUFFER *wb);
 
 #endif
diff --git a/libnetdata/functions_evloop/Makefile.am b/libnetdata/functions_evloop/Makefile.am
new file mode 100644
index 00000000..161784b8
--- /dev/null
+++ b/libnetdata/functions_evloop/Makefile.am
@@ -0,0 +1,8 @@
+# SPDX-License-Identifier: GPL-3.0-or-later
+
+AUTOMAKE_OPTIONS = subdir-objects
+MAINTAINERCLEANFILES = $(srcdir)/Makefile.in
+
+dist_noinst_DATA = \
+    README.md \
+    $(NULL)
diff --git a/libnetdata/functions_evloop/README.md b/libnetdata/functions_evloop/README.md
new file mode 100644
index 00000000..e69de29b
--- /dev/null
+++ b/libnetdata/functions_evloop/README.md
diff --git a/libnetdata/functions_evloop/functions_evloop.c b/libnetdata/functions_evloop/functions_evloop.c
new file mode 100644
index 00000000..3fcd70aa
--- /dev/null
+++ b/libnetdata/functions_evloop/functions_evloop.c
@@ -0,0 +1,210 @@
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#include "functions_evloop.h"
+
+#define MAX_FUNCTION_PARAMETERS 1024
+
+struct functions_evloop_worker_job {
+    bool used;
+    bool running;
+    bool cancelled;
+    char *cmd;
+    const char *transaction;
+    time_t timeout;
+    functions_evloop_worker_execute_t cb;
+};
+
+struct rrd_functions_expectation {
+    const char *function;
+    size_t function_length;
+    functions_evloop_worker_execute_t cb;
+    time_t default_timeout;
+    struct rrd_functions_expectation *prev, *next;
+};
+
+struct functions_evloop_globals {
+    const char *tag;
+
+    DICTIONARY *worker_queue;
+    pthread_mutex_t worker_mutex;
+    pthread_cond_t worker_cond_var;
+    size_t workers;
+
+    netdata_mutex_t *stdout_mutex;
+    bool *plugin_should_exit;
+
+    netdata_thread_t reader_thread;
+    netdata_thread_t *worker_threads;
+
+    struct rrd_functions_expectation *expectations;
+};
+
+static void *rrd_functions_worker_globals_worker_main(void *arg) {
+    struct functions_evloop_globals *wg = arg;
+
+    bool last_acquired = true;
+    while (true) {
+        pthread_mutex_lock(&wg->worker_mutex);
+
+        if(dictionary_entries(wg->worker_queue) == 0 || !last_acquired)
+            pthread_cond_wait(&wg->worker_cond_var, &wg->worker_mutex);
+
+        const DICTIONARY_ITEM *acquired = NULL;
+        struct functions_evloop_worker_job *j;
+        dfe_start_write(wg->worker_queue, j) {
+            if(j->running || j->cancelled)
+                continue;
+
+            acquired = dictionary_acquired_item_dup(wg->worker_queue, j_dfe.item);
+            j->running = true;
+            break;
+        }
+        dfe_done(j);
+
+        pthread_mutex_unlock(&wg->worker_mutex);
+
+        if(acquired) {
+            last_acquired = true;
+            j = dictionary_acquired_item_value(acquired);
+            j->cb(j->transaction, j->cmd, j->timeout, &j->cancelled);
+            dictionary_del(wg->worker_queue, j->transaction);
+            dictionary_acquired_item_release(wg->worker_queue, acquired);
+            dictionary_garbage_collect(wg->worker_queue);
+        }
+        else
+            last_acquired = false;
+    }
+    return NULL;
+}
+
+static void *rrd_functions_worker_globals_reader_main(void *arg) {
+    struct functions_evloop_globals *wg = arg;
+
+    char buffer[PLUGINSD_LINE_MAX + 1];
+
+    char *s = NULL;
+    while(!(*wg->plugin_should_exit) && (s = fgets(buffer, PLUGINSD_LINE_MAX, stdin))) {
+
+        char *words[MAX_FUNCTION_PARAMETERS] = { NULL };
+        size_t num_words = quoted_strings_splitter_pluginsd(buffer, words, MAX_FUNCTION_PARAMETERS);
+
+        const char *keyword = get_word(words, num_words, 0);
+
+        if(keyword && strcmp(keyword, PLUGINSD_KEYWORD_FUNCTION) == 0) {
+            char *transaction = get_word(words, num_words, 1);
+            char *timeout_s = get_word(words, num_words, 2);
+            char *function = get_word(words, num_words, 3);
+
+            if(!transaction || !*transaction || !timeout_s || !*timeout_s || !function || !*function) {
+                netdata_log_error("Received incomplete %s (transaction = '%s', timeout = '%s', function = '%s'). Ignoring it.",
+                                  keyword,
+                                  transaction?transaction:"(unset)",
+                                  timeout_s?timeout_s:"(unset)",
+                                  function?function:"(unset)");
+            }
+            else {
+                int timeout = str2i(timeout_s);
+
+                bool found = false;
+                struct rrd_functions_expectation *we;
+                for(we = wg->expectations; we ;we = we->next) {
+                    if(strncmp(function, we->function, we->function_length) == 0) {
+                        struct functions_evloop_worker_job t = {
+                                .cmd = strdupz(function),
+                                .transaction = strdupz(transaction),
+                                .running = false,
+                                .cancelled = false,
+                                .timeout = timeout > 0 ? timeout : we->default_timeout,
+                                .used = false,
+                                .cb = we->cb,
+                        };
+                        struct functions_evloop_worker_job *j = dictionary_set(wg->worker_queue, transaction, &t, sizeof(t));
+                        if(j->used) {
+                            netdata_log_error("Received duplicate function transaction '%s'", transaction);
+                            freez((void *)t.cmd);
+                            freez((void *)t.transaction);
+                        }
+                        else {
+                            found = true;
+                            j->used = true;
+                            pthread_cond_signal(&wg->worker_cond_var);
+                        }
+                    }
+                }
+
+                if(!found) {
+                    netdata_mutex_lock(wg->stdout_mutex);
+                    pluginsd_function_json_error_to_stdout(transaction, HTTP_RESP_NOT_FOUND,
+                                                           "No function with this name found.");
+                    netdata_mutex_unlock(wg->stdout_mutex);
+                }
+            }
+        }
+        else if(keyword && strcmp(keyword, PLUGINSD_KEYWORD_FUNCTION_CANCEL) == 0) {
+            char *transaction = get_word(words, num_words, 1);
+            const DICTIONARY_ITEM *acquired = dictionary_get_and_acquire_item(wg->worker_queue, transaction);
+            if(acquired) {
+                struct functions_evloop_worker_job *j = dictionary_acquired_item_value(acquired);
+                __atomic_store_n(&j->cancelled, true, __ATOMIC_RELAXED);
+                dictionary_acquired_item_release(wg->worker_queue, acquired);
+                dictionary_del(wg->worker_queue, transaction);
+                dictionary_garbage_collect(wg->worker_queue);
+            }
+            else
+                netdata_log_error("Received CANCEL for transaction '%s', but it not available here", transaction);
+        }
+        else
+            netdata_log_error("Received unknown command: %s", keyword?keyword:"(unset)");
+    }
+
+    if(!s || feof(stdin) || ferror(stdin)) {
+        *wg->plugin_should_exit = true;
+        netdata_log_error("Received error on stdin.");
+    }
+
+    exit(1);
+}
+
+void worker_queue_delete_cb(const DICTIONARY_ITEM *item __maybe_unused, void *value, void *data __maybe_unused) {
+    struct functions_evloop_worker_job *j = value;
+    freez((void *)j->cmd);
+    freez((void *)j->transaction);
+}
+
+struct functions_evloop_globals *functions_evloop_init(size_t worker_threads, const char *tag, netdata_mutex_t *stdout_mutex, bool *plugin_should_exit) {
+    struct functions_evloop_globals *wg = callocz(1, sizeof(struct functions_evloop_globals));
+
+    wg->worker_queue = dictionary_create(DICT_OPTION_DONT_OVERWRITE_VALUE);
+    dictionary_register_delete_callback(wg->worker_queue, worker_queue_delete_cb, NULL);
+
+    pthread_mutex_init(&wg->worker_mutex, NULL);
+    pthread_cond_init(&wg->worker_cond_var, NULL);
+
+    wg->plugin_should_exit = plugin_should_exit;
+    wg->stdout_mutex = stdout_mutex;
+    wg->workers = worker_threads;
+    wg->worker_threads = callocz(wg->workers, sizeof(netdata_thread_t ));
+    wg->tag = tag;
+
+    char tag_buffer[NETDATA_THREAD_TAG_MAX + 1];
+    snprintfz(tag_buffer, NETDATA_THREAD_TAG_MAX, "%s_READER", wg->tag);
+    netdata_thread_create(&wg->reader_thread, tag_buffer, NETDATA_THREAD_OPTION_DONT_LOG,
+                          rrd_functions_worker_globals_reader_main, wg);
+
+    for(size_t i = 0; i < wg->workers ; i++) {
+        snprintfz(tag_buffer, NETDATA_THREAD_TAG_MAX, "%s_WORK[%zu]", wg->tag, i+1);
+        netdata_thread_create(&wg->worker_threads[i], tag_buffer, NETDATA_THREAD_OPTION_DONT_LOG,
+                              rrd_functions_worker_globals_worker_main, wg);
+    }
+
+    return wg;
+}
+
+void functions_evloop_add_function(struct functions_evloop_globals *wg, const char *function, functions_evloop_worker_execute_t cb, time_t default_timeout) {
+    struct rrd_functions_expectation *we = callocz(1, sizeof(*we));
+    we->function = function;
+    we->function_length = strlen(we->function);
+    we->cb = cb;
+    we->default_timeout = default_timeout;
+    DOUBLE_LINKED_LIST_APPEND_ITEM_UNSAFE(wg->expectations, we, prev, next);
+}
diff --git a/libnetdata/functions_evloop/functions_evloop.h b/libnetdata/functions_evloop/functions_evloop.h
new file mode 100644
index 00000000..ee0f72cb
--- /dev/null
+++ b/libnetdata/functions_evloop/functions_evloop.h
@@ -0,0 +1,98 @@
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#ifndef NETDATA_FUNCTIONS_EVLOOP_H
+#define NETDATA_FUNCTIONS_EVLOOP_H
+
+#include "../libnetdata.h"
+
+#define PLUGINSD_KEYWORD_CHART                  "CHART"
+#define PLUGINSD_KEYWORD_CHART_DEFINITION_END   "CHART_DEFINITION_END"
+#define PLUGINSD_KEYWORD_DIMENSION              "DIMENSION"
+#define PLUGINSD_KEYWORD_BEGIN                  "BEGIN"
+#define PLUGINSD_KEYWORD_SET                    "SET"
+#define PLUGINSD_KEYWORD_END                    "END"
+#define PLUGINSD_KEYWORD_FLUSH                  "FLUSH"
+#define PLUGINSD_KEYWORD_DISABLE                "DISABLE"
+#define PLUGINSD_KEYWORD_VARIABLE               "VARIABLE"
+#define PLUGINSD_KEYWORD_LABEL                  "LABEL"
+#define PLUGINSD_KEYWORD_OVERWRITE              "OVERWRITE"
+#define PLUGINSD_KEYWORD_CLABEL                 "CLABEL"
+#define PLUGINSD_KEYWORD_CLABEL_COMMIT          "CLABEL_COMMIT"
+#define PLUGINSD_KEYWORD_FUNCTION               "FUNCTION"
+#define PLUGINSD_KEYWORD_FUNCTION_CANCEL        "FUNCTION_CANCEL"
+#define PLUGINSD_KEYWORD_FUNCTION_RESULT_BEGIN  "FUNCTION_RESULT_BEGIN"
+#define PLUGINSD_KEYWORD_FUNCTION_RESULT_END    "FUNCTION_RESULT_END"
+
+#define PLUGINSD_KEYWORD_REPLAY_CHART           "REPLAY_CHART"
+#define PLUGINSD_KEYWORD_REPLAY_BEGIN           "RBEGIN"
+#define PLUGINSD_KEYWORD_REPLAY_SET             "RSET"
+#define PLUGINSD_KEYWORD_REPLAY_RRDDIM_STATE    "RDSTATE"
+#define PLUGINSD_KEYWORD_REPLAY_RRDSET_STATE    "RSSTATE"
+#define PLUGINSD_KEYWORD_REPLAY_END             "REND"
+
+#define PLUGINSD_KEYWORD_BEGIN_V2               "BEGIN2"
+#define PLUGINSD_KEYWORD_SET_V2                 "SET2"
+#define PLUGINSD_KEYWORD_END_V2                 "END2"
+
+#define PLUGINSD_KEYWORD_HOST_DEFINE            "HOST_DEFINE"
+#define PLUGINSD_KEYWORD_HOST_DEFINE_END        "HOST_DEFINE_END"
+#define PLUGINSD_KEYWORD_HOST_LABEL             "HOST_LABEL"
+#define PLUGINSD_KEYWORD_HOST                   "HOST"
+
+#define PLUGINSD_KEYWORD_DYNCFG_ENABLE          "DYNCFG_ENABLE"
+#define PLUGINSD_KEYWORD_DYNCFG_REGISTER_MODULE "DYNCFG_REGISTER_MODULE"
+
+#define PLUGINSD_KEYWORD_REPORT_JOB_STATUS      "REPORT_JOB_STATUS"
+
+#define PLUGINSD_KEYWORD_EXIT                   "EXIT"
+
+#define PLUGINS_FUNCTIONS_TIMEOUT_DEFAULT 10 // seconds
+
+typedef void (*functions_evloop_worker_execute_t)(const char *transaction, char *function, int timeout, bool *cancelled);
+struct functions_evloop_worker_job;
+struct functions_evloop_globals *functions_evloop_init(size_t worker_threads, const char *tag, netdata_mutex_t *stdout_mutex, bool *plugin_should_exit);
+void functions_evloop_add_function(struct functions_evloop_globals *wg, const char *function, functions_evloop_worker_execute_t cb, time_t default_timeout);
+
+
+#define pluginsd_function_result_begin_to_buffer(wb, transaction, code, content_type, expires)      \
+    buffer_sprintf(wb                                                                               \
+                    , PLUGINSD_KEYWORD_FUNCTION_RESULT_BEGIN " \"%s\" %d \"%s\" %ld\n"              \
+                    , (transaction) ? (transaction) : ""                                            \
+                    , (int)(code)                                                                   \
+                    , (content_type) ? (content_type) : ""                                          \
+                    , (long int)(expires)                                                           \
+    )
+
+#define pluginsd_function_result_end_to_buffer(wb) \
+    buffer_strcat(wb, "\n" PLUGINSD_KEYWORD_FUNCTION_RESULT_END "\n")
+
+#define pluginsd_function_result_begin_to_stdout(transaction, code, content_type, expires)          \
+    fprintf(stdout                                                                                  \
+                    , PLUGINSD_KEYWORD_FUNCTION_RESULT_BEGIN " \"%s\" %d \"%s\" %ld\n"              \
+                    , (transaction) ? (transaction) : ""                                            \
+                    , (int)(code)                                                                   \
+                    , (content_type) ? (content_type) : ""                                          \
+                    , (long int)(expires)                                                           \
+    )
+
+#define pluginsd_function_result_end_to_stdout() \
+    fprintf(stdout, "\n" PLUGINSD_KEYWORD_FUNCTION_RESULT_END "\n")
+
+static inline void pluginsd_function_json_error_to_stdout(const char *transaction, int code, const char *msg) {
+    char buffer[PLUGINSD_LINE_MAX + 1];
+    json_escape_string(buffer, msg, PLUGINSD_LINE_MAX);
+
+    pluginsd_function_result_begin_to_stdout(transaction, code, "application/json", now_realtime_sec());
+    fprintf(stdout, "{\"status\":%d,\"error_message\":\"%s\"}", code, buffer);
+    pluginsd_function_result_end_to_stdout();
+    fflush(stdout);
+}
+
+static inline void pluginsd_function_result_to_stdout(const char *transaction, int code, const char *content_type, time_t expires, BUFFER *result) {
+    pluginsd_function_result_begin_to_stdout(transaction, code, content_type, expires);
+    fwrite(buffer_tostring(result), buffer_strlen(result), 1, stdout);
+    pluginsd_function_result_end_to_stdout();
+    fflush(stdout);
+}
+
+#endif //NETDATA_FUNCTIONS_EVLOOP_H
diff --git a/libnetdata/health/health.c b/libnetdata/health/health.c
index 53ebecb4..f2dc46e3 100644
--- a/libnetdata/health/health.c
+++ b/libnetdata/health/health.c
@@ -1,3 +1,5 @@
+// SPDX-License-Identifier: GPL-3.0-or-later
+
 #include "health.h"
 
 SILENCERS *silencers;
@@ -27,8 +29,8 @@ void health_silencers_add(SILENCER *silencer) {
     // Add the created instance to the linked list in silencers
     silencer->next = silencers->silencers;
     silencers->silencers = silencer;
-    netdata_log_debug(D_HEALTH, "HEALTH command API: Added silencer %s:%s:%s:%s:%s", silencer->alarms,
-          silencer->charts, silencer->contexts, silencer->hosts, silencer->families
+    netdata_log_debug(D_HEALTH, "HEALTH command API: Added silencer %s:%s:%s:%s", silencer->alarms,
+          silencer->charts, silencer->contexts, silencer->hosts
     );
 }
 
@@ -49,8 +51,7 @@ SILENCER *health_silencers_addparam(SILENCER *silencer, char *key, char *value)
             hash_template = 0,
             hash_chart = 0,
             hash_context = 0,
-            hash_host = 0,
-            hash_families = 0;
+            hash_host = 0;
 
     if (unlikely(!hash_alarm)) {
         hash_alarm = simple_uhash(HEALTH_ALARM_KEY);
@@ -58,7 +59,6 @@ SILENCER *health_silencers_addparam(SILENCER *silencer, char *key, char *value)
         hash_chart = simple_uhash(HEALTH_CHART_KEY);
         hash_context = simple_uhash(HEALTH_CONTEXT_KEY);
         hash_host = simple_uhash(HEALTH_HOST_KEY);
-        hash_families = simple_uhash(HEALTH_FAMILIES_KEY);
     }
 
     uint32_t hash = simple_uhash(key);
@@ -68,8 +68,7 @@ SILENCER *health_silencers_addparam(SILENCER *silencer, char *key, char *value)
                 (hash == hash_template && !strcasecmp(key, HEALTH_TEMPLATE_KEY)) ||
                 (hash == hash_chart && !strcasecmp(key, HEALTH_CHART_KEY)) ||
                 (hash == hash_context && !strcasecmp(key, HEALTH_CONTEXT_KEY)) ||
-                (hash == hash_host && !strcasecmp(key, HEALTH_HOST_KEY)) ||
-                (hash == hash_families && !strcasecmp(key, HEALTH_FAMILIES_KEY))
+                (hash == hash_host && !strcasecmp(key, HEALTH_HOST_KEY))
                 ) {
             silencer = create_silencer();
             if(!silencer) {
@@ -91,9 +90,6 @@ SILENCER *health_silencers_addparam(SILENCER *silencer, char *key, char *value)
     } else if (hash == hash_host && !strcasecmp(key, HEALTH_HOST_KEY)) {
         silencer->hosts = strdupz(value);
         silencer->hosts_pattern = simple_pattern_create(silencer->hosts, NULL, SIMPLE_PATTERN_EXACT, true);
-    } else if (hash == hash_families && !strcasecmp(key, HEALTH_FAMILIES_KEY)) {
-        silencer->families = strdupz(value);
-        silencer->families_pattern = simple_pattern_create(silencer->families, NULL, SIMPLE_PATTERN_EXACT, true);
     }
 
     return silencer;
@@ -170,4 +166,4 @@ int health_initialize_global_silencers() {
     silencers->silencers=NULL;
 
     return 0;
-}
\ No newline at end of file
+}
diff --git a/libnetdata/health/health.h b/libnetdata/health/health.h
index 6b8f9b38..31173fe6 100644
--- a/libnetdata/health/health.h
+++ b/libnetdata/health/health.h
@@ -1,3 +1,5 @@
+// SPDX-License-Identifier: GPL-3.0-or-later
+
 #ifndef NETDATA_HEALTH_LIB
 # define NETDATA_HEALTH_LIB 1
 
@@ -9,7 +11,6 @@
 #define HEALTH_CHART_KEY "chart"
 #define HEALTH_HOST_KEY "hosts"
 #define HEALTH_OS_KEY "os"
-#define HEALTH_FAMILIES_KEY "families"
 #define HEALTH_LOOKUP_KEY "lookup"
 #define HEALTH_CALC_KEY "calc"
 
@@ -26,9 +27,6 @@ typedef struct silencer {
     char *charts;
     SIMPLE_PATTERN *charts_pattern;
 
-    char *families;
-    SIMPLE_PATTERN *families_pattern;
-
     struct silencer *next;
 } SILENCER;
 
diff --git a/libnetdata/http/http_defs.h b/libnetdata/http/http_defs.h
index ffd1c344..71bd442c 100644
--- a/libnetdata/http/http_defs.h
+++ b/libnetdata/http/http_defs.h
@@ -8,6 +8,7 @@
 
 // HTTP_CODES 3XX Redirections
 #define HTTP_RESP_MOVED_PERM 301
+#define HTTP_RESP_NOT_MODIFIED 304
 #define HTTP_RESP_REDIR_TEMP 307
 #define HTTP_RESP_REDIR_PERM 308
 #define HTTP_RESP_HTTPS_UPGRADE 399
@@ -22,12 +23,12 @@
 #define HTTP_RESP_CONFLICT 409
 #define HTTP_RESP_PRECOND_FAIL 412
 #define HTTP_RESP_CONTENT_TOO_LONG 413
+#define HTTP_RESP_CLIENT_CLOSED_REQUEST 499 // nginx's enxtension to the standard
 
 // HTTP_CODES 5XX Server Errors
 #define HTTP_RESP_INTERNAL_SERVER_ERROR 500
 #define HTTP_RESP_INTERNAL_SERVER_ERROR_STR "Internal Server Error"
-#define HTTP_RESP_BACKEND_FETCH_FAILED 503  // 503 is right
-#define HTTP_RESP_SERVICE_UNAVAILABLE 503   // 503 is right
+#define HTTP_RESP_SERVICE_UNAVAILABLE 503
 #define HTTP_RESP_GATEWAY_TIMEOUT 504
 #define HTTP_RESP_BACKEND_RESPONSE_INVALID 591
 
diff --git a/libnetdata/inlined.h b/libnetdata/inlined.h
index eb55f0fe..9c07d97b 100644
--- a/libnetdata/inlined.h
+++ b/libnetdata/inlined.h
@@ -394,10 +394,10 @@ static inline NETDATA_DOUBLE str2ndd_encoded(const char *src, char **endptr) {
     return str2ndd(src, endptr) * sign;
 }
 
-static inline char *strncpyz(char *dst, const char *src, size_t n) {
+static inline char *strncpyz(char *dst, const char *src, size_t dst_size_minus_1) {
     char *p = dst;
 
-    while (*src && n--)
+    while (*src && dst_size_minus_1--)
         *dst++ = *src++;
 
     *dst = '\0';
diff --git a/libnetdata/libnetdata.c b/libnetdata/libnetdata.c
index 26582ffe..159e8995 100644
--- a/libnetdata/libnetdata.c
+++ b/libnetdata/libnetdata.c
@@ -1942,6 +1942,7 @@ void timing_action(TIMING_ACTION action, TIMING_STEP step) {
     }
 }
 
+#ifdef ENABLE_HTTPS
 int hash256_string(const unsigned char *string, size_t size, char *hash) {
     EVP_MD_CTX *ctx;
     ctx = EVP_MD_CTX_create();
@@ -1966,19 +1967,15 @@ int hash256_string(const unsigned char *string, size_t size, char *hash) {
     EVP_MD_CTX_destroy(ctx);
     return 1;
 }
+#endif
 
-// Returns 1 if an absolute period was requested or 0 if it was a relative period
-bool rrdr_relative_window_to_absolute(time_t *after, time_t *before, time_t *now_ptr, bool unittest_running) {
-    time_t now = now_realtime_sec() - 1;
 
-    if(now_ptr)
-        *now_ptr = now;
+bool rrdr_relative_window_to_absolute(time_t *after, time_t *before, time_t now) {
+    if(!now) now = now_realtime_sec();
 
     int absolute_period_requested = -1;
-    long long after_requested, before_requested;
-
-    before_requested = *before;
-    after_requested = *after;
+    time_t before_requested = *before;
+    time_t after_requested = *after;
 
     // allow relative for before (smaller than API_RELATIVE_TIME_MAX)
     if(ABS(before_requested) <= API_RELATIVE_TIME_MAX) {
@@ -2023,11 +2020,29 @@ bool rrdr_relative_window_to_absolute(time_t *after, time_t *before, time_t *now
     // shift the query back to be in the present time
     // (this may also happen because of the rules above)
     if(before_requested > now) {
-        long long delta = before_requested - now;
+        time_t delta = before_requested - now;
         before_requested -= delta;
         after_requested  -= delta;
     }
 
+    *before = before_requested;
+    *after = after_requested;
+
+    return (absolute_period_requested != 1);
+}
+
+// Returns 1 if an absolute period was requested or 0 if it was a relative period
+bool rrdr_relative_window_to_absolute_query(time_t *after, time_t *before, time_t *now_ptr, bool unittest_running) {
+    time_t now = now_realtime_sec() - 1;
+
+    if(now_ptr)
+        *now_ptr = now;
+
+    time_t before_requested = *before;
+    time_t after_requested = *after;
+
+    int absolute_period_requested = rrdr_relative_window_to_absolute(&after_requested, &before_requested, now);
+
     time_t absolute_minimum_time = now - (10 * 365 * 86400);
     time_t absolute_maximum_time = now + (1 * 365 * 86400);
 
diff --git a/libnetdata/libnetdata.h b/libnetdata/libnetdata.h
index b8dedf51..c337b3b5 100644
--- a/libnetdata/libnetdata.h
+++ b/libnetdata/libnetdata.h
@@ -798,6 +798,10 @@ void for_each_open_fd(OPEN_FD_ACTION action, OPEN_FD_EXCLUDE excluded_fds);
 void netdata_cleanup_and_exit(int ret) NORETURN;
 void send_statistics(const char *action, const char *action_result, const char *action_data);
 extern char *netdata_configured_host_prefix;
+
+#define XXH_INLINE_ALL
+#include "xxhash.h"
+
 #include "libjudy/src/Judy.h"
 #include "july/july.h"
 #include "os.h"
@@ -838,6 +842,7 @@ extern char *netdata_configured_host_prefix;
 #include "gorilla/gorilla.h"
 #include "facets/facets.h"
 #include "dyn_conf/dyn_conf.h"
+#include "functions_evloop/functions_evloop.h"
 
 // BEWARE: this exists in alarm-notify.sh
 #define DEFAULT_CLOUD_BASE_URL "https://app.netdata.cloud"
@@ -984,7 +989,8 @@ int hash256_string(const unsigned char *string, size_t size, char *hash);
 extern bool unittest_running;
 #define API_RELATIVE_TIME_MAX (3 * 365 * 86400)
 
-bool rrdr_relative_window_to_absolute(time_t *after, time_t *before, time_t *now_ptr, bool unittest_running);
+bool rrdr_relative_window_to_absolute(time_t *after, time_t *before, time_t now);
+bool rrdr_relative_window_to_absolute_query(time_t *after, time_t *before, time_t *now_ptr, bool unittest_running);
 
 int netdata_base64_decode(const char *encoded, char *decoded, size_t decoded_size);
 
diff --git a/libnetdata/log/README.md b/libnetdata/log/README.md
index 3684abd6..f811bb4b 100644
--- a/libnetdata/log/README.md
+++ b/libnetdata/log/README.md
@@ -12,4 +12,3 @@ learn_rel_path: "Developers/libnetdata"
 The netdata log library supports debug, info, error and fatal error logging. 
 By default we have an access log, an error log and a collectors log. 
 
-
diff --git a/libnetdata/log/log.c b/libnetdata/log/log.c
index e43a4f46..02bb776c 100644
--- a/libnetdata/log/log.c
+++ b/libnetdata/log/log.c
@@ -12,11 +12,11 @@ int web_server_is_multithreaded = 1;
 const char *program_name = "";
 uint64_t debug_flags = 0;
 
-int access_log_syslog = 1;
-int error_log_syslog = 1;
-int collector_log_syslog = 1;
-int output_log_syslog = 1;  // debug log
-int health_log_syslog = 1;
+int access_log_syslog = 0;
+int error_log_syslog = 0;
+int collector_log_syslog = 0;
+int output_log_syslog = 0;  // debug log
+int health_log_syslog = 0;
 
 int stdaccess_fd = -1;
 FILE *stdaccess = NULL;
@@ -34,6 +34,8 @@ const char *facility_log = NULL;
 const char *stdhealth_filename = NULL;
 const char *stdcollector_filename = NULL;
 
+netdata_log_level_t global_log_severity_level = NETDATA_LOG_LEVEL_INFO;
+
 #ifdef ENABLE_ACLK
 const char *aclklog_filename = NULL;
 int aclklog_fd = -1;
@@ -780,6 +782,11 @@ void debug_int( const char *file, const char *function, const unsigned long line
 
 void info_int( int is_collector, const char *file __maybe_unused, const char *function __maybe_unused, const unsigned long line __maybe_unused, const char *fmt, ... )
 {
+#if !defined(NETDATA_INTERNAL_CHECKS) && !defined(NETDATA_DEV_MODE)
+    if (NETDATA_LOG_LEVEL_INFO > global_log_severity_level)
+        return;
+#endif
+
     va_list args;
     FILE *fp = (is_collector || !stderror) ? stderr : stderror;
 
@@ -890,7 +897,7 @@ void error_limit_int(ERROR_LIMIT *erl, const char *prefix, const char *file __ma
                 erl->count, (unsigned long long)(erl->last_logged ? now - erl->last_logged : 0));
 
     if(erl->sleep_ut)
-        fprintf(fp, " (sleeping for %llu microseconds every time this happens)", erl->sleep_ut);
+        fprintf(fp, " (sleeping for %"PRIu64" microseconds every time this happens)", erl->sleep_ut);
 
     if(__errno) {
         char buf[1024];
@@ -908,6 +915,11 @@ void error_limit_int(ERROR_LIMIT *erl, const char *prefix, const char *file __ma
 }
 
 void error_int(int is_collector, const char *prefix, const char *file __maybe_unused, const char *function __maybe_unused, const unsigned long line __maybe_unused, const char *fmt, ... ) {
+#if !defined(NETDATA_INTERNAL_CHECKS) && !defined(NETDATA_DEV_MODE)
+    if (NETDATA_LOG_LEVEL_ERROR > global_log_severity_level)
+        return;
+#endif
+
     // save a copy of errno - just in case this function generates a new error
     int __errno = errno;
     FILE *fp = (is_collector || !stderror) ? stderr : stderror;
@@ -1125,3 +1137,37 @@ void log_aclk_message_bin( const char *data, const size_t data_len, int tx, cons
     }
 }
 #endif
+
+void log_set_global_severity_level(netdata_log_level_t value)
+{
+    global_log_severity_level = value;
+}
+
+netdata_log_level_t log_severity_string_to_severity_level(char *level)
+{
+    if (!strcmp(level, NETDATA_LOG_LEVEL_INFO_STR))
+        return NETDATA_LOG_LEVEL_INFO;
+    if (!strcmp(level, NETDATA_LOG_LEVEL_ERROR_STR) || !strcmp(level, NETDATA_LOG_LEVEL_ERROR_SHORT_STR))
+        return NETDATA_LOG_LEVEL_ERROR;
+
+    return NETDATA_LOG_LEVEL_INFO;
+}
+
+char *log_severity_level_to_severity_string(netdata_log_level_t level)
+{
+    switch (level) {
+        case NETDATA_LOG_LEVEL_ERROR:
+            return NETDATA_LOG_LEVEL_ERROR_STR;
+        case NETDATA_LOG_LEVEL_INFO:
+        default:
+            return NETDATA_LOG_LEVEL_INFO_STR;
+    }
+}
+
+void log_set_global_severity_for_external_plugins() {
+    char *s = getenv("NETDATA_LOG_SEVERITY_LEVEL");
+    if (!s)
+        return;
+    netdata_log_level_t level = log_severity_string_to_severity_level(s);
+    log_set_global_severity_level(level);
+}
diff --git a/libnetdata/log/log.h b/libnetdata/log/log.h
index 9ced07a9..cf0865cf 100644
--- a/libnetdata/log/log.h
+++ b/libnetdata/log/log.h
@@ -43,9 +43,6 @@ extern "C" {
 #define D_ANALYTICS         0x0000000080000000
 #define D_RRDENGINE         0x0000000100000000
 #define D_ACLK              0x0000000200000000
-#define D_METADATALOG       0x0000000400000000
-#define D_ACLK_SYNC         0x0000000800000000
-#define D_META_SYNC         0x0000001000000000
 #define D_REPLICATION       0x0000002000000000
 #define D_SYSTEM            0x8000000000000000
 
@@ -105,6 +102,23 @@ typedef struct error_with_limit {
     usec_t sleep_ut;
 } ERROR_LIMIT;
 
+typedef enum netdata_log_level {
+    NETDATA_LOG_LEVEL_ERROR,
+    NETDATA_LOG_LEVEL_INFO,
+
+    NETDATA_LOG_LEVEL_END
+} netdata_log_level_t;
+
+#define NETDATA_LOG_LEVEL_INFO_STR "info"
+#define NETDATA_LOG_LEVEL_ERROR_STR "error"
+#define NETDATA_LOG_LEVEL_ERROR_SHORT_STR "err"
+
+extern netdata_log_level_t global_log_severity_level;
+netdata_log_level_t log_severity_string_to_severity_level(char *level);
+char *log_severity_level_to_severity_string(netdata_log_level_t level);
+void log_set_global_severity_level(netdata_log_level_t value);
+void log_set_global_severity_for_external_plugins();
+
 #define error_limit_static_global_var(var, log_every_secs, sleep_usecs) static ERROR_LIMIT var = { .last_logged = 0, .count = 0, .log_every = (log_every_secs), .sleep_ut = (sleep_usecs) }
 #define error_limit_static_thread_var(var, log_every_secs, sleep_usecs) static __thread ERROR_LIMIT var = { .last_logged = 0, .count = 0, .log_every = (log_every_secs), .sleep_ut = (sleep_usecs) }
 
diff --git a/libnetdata/required_dummies.h b/libnetdata/required_dummies.h
index 5a0d4e05..1ffe1e9e 100644
--- a/libnetdata/required_dummies.h
+++ b/libnetdata/required_dummies.h
@@ -37,6 +37,7 @@ void rrdset_thread_rda_free(void){};
 void sender_thread_buffer_free(void){};
 void query_target_free(void){};
 void service_exits(void){};
+void rrd_collector_finished(void){};
 
 // required by get_system_cpus()
 char *netdata_configured_host_prefix = "";
diff --git a/libnetdata/socket/socket.c b/libnetdata/socket/socket.c
index e7d0b480..67dc4c71 100644
--- a/libnetdata/socket/socket.c
+++ b/libnetdata/socket/socket.c
@@ -10,6 +10,40 @@
 
 #include "../libnetdata.h"
 
+bool ip_to_hostname(const char *ip, char *dst, size_t dst_len) {
+    if(!dst || !dst_len)
+        return false;
+
+    struct sockaddr_in sa;
+    struct sockaddr_in6 sa6;
+    struct sockaddr *sa_ptr;
+    int sa_len;
+
+    // Try to convert the IP address to sockaddr_in (IPv4)
+    if (inet_pton(AF_INET, ip, &(sa.sin_addr)) == 1) {
+        sa.sin_family = AF_INET;
+        sa_ptr = (struct sockaddr *)&sa;
+        sa_len = sizeof(sa);
+    }
+        // Try to convert the IP address to sockaddr_in6 (IPv6)
+    else if (inet_pton(AF_INET6, ip, &(sa6.sin6_addr)) == 1) {
+        sa6.sin6_family = AF_INET6;
+        sa_ptr = (struct sockaddr *)&sa6;
+        sa_len = sizeof(sa6);
+    }
+
+    else {
+        dst[0] = '\0';
+        return false;
+    }
+
+    // Perform the reverse lookup
+    int res = getnameinfo(sa_ptr, sa_len, dst, dst_len, NULL, 0, NI_NAMEREQD);
+    if(res != 0)
+        return false;
+
+    return true;
+}
 
 SOCKET_PEERS socket_peers(int sock_fd) {
     SOCKET_PEERS peers;
@@ -810,7 +844,7 @@ int connect_to_this_ip46(int protocol, int socktype, const char *host, uint32_t
             errno = 0;
             if(connect(fd, ai->ai_addr, ai->ai_addrlen) < 0) {
                 if(errno == EALREADY || errno == EINPROGRESS) {
-                    netdata_log_info("Waiting for connection to ip %s port %s to be established", hostBfr, servBfr);
+                    internal_error(true, "Waiting for connection to ip %s port %s to be established", hostBfr, servBfr);
 
                     // Convert 'struct timeval' to milliseconds for poll():
                     int timeout_milliseconds = timeout->tv_sec * 1000 + timeout->tv_usec / 1000;
@@ -835,6 +869,7 @@ int connect_to_this_ip46(int protocol, int socktype, const char *host, uint32_t
                     }
                     else if (ret == 0) {
                         // poll() timed out, the connection is not established within the specified timeout.
+                        errno = 0;
                         netdata_log_error("Timed out while connecting to '%s', port '%s'.", hostBfr, servBfr);
                         close(fd);
                         fd = -1;
diff --git a/libnetdata/socket/socket.h b/libnetdata/socket/socket.h
index c4bd4736..e4ca08d4 100644
--- a/libnetdata/socket/socket.h
+++ b/libnetdata/socket/socket.h
@@ -243,5 +243,6 @@ typedef struct socket_peers {
 } SOCKET_PEERS;
 
 SOCKET_PEERS socket_peers(int sock_fd);
+bool ip_to_hostname(const char *ip, char *dst, size_t dst_len);
 
 #endif //NETDATA_SOCKET_H
diff --git a/libnetdata/string/string.c b/libnetdata/string/string.c
index 373d0c24..54b8f171 100644
--- a/libnetdata/string/string.c
+++ b/libnetdata/string/string.c
@@ -28,19 +28,22 @@ static struct string_partition {
 
     Pvoid_t JudyHSArray;        // the Judy array - hashtable
 
-    size_t searches;            // the number of successful searches in the index
-    size_t duplications;        // when a string is referenced
-    size_t releases;            // when a string is unreferenced
-
     size_t inserts;             // the number of successful inserts to the index
     size_t deletes;             // the number of successful deleted from the index
 
     long int entries;           // the number of entries in the index
-    long int active_references; // the number of active references alive
     long int memory;            // the memory used, without the JudyHS index
 
 #ifdef NETDATA_INTERNAL_CHECKS
     // internal statistics
+
+    struct {
+        size_t searches;            // the number of successful searches in the index
+        size_t releases;            // when a string is unreferenced
+        size_t duplications;        // when a string is referenced
+        long int active_references; // the number of active references alive
+    } atomic;
+
     size_t found_deleted_on_search;
     size_t found_available_on_search;
     size_t found_deleted_on_insert;
@@ -51,14 +54,15 @@ static struct string_partition {
 } string_base[STRING_PARTITIONS] = { 0 };
 
 #ifdef NETDATA_INTERNAL_CHECKS
+#define string_stats_atomic_increment(partition, var) __atomic_add_fetch(&string_base[partition].atomic.var, 1, __ATOMIC_RELAXED)
+#define string_stats_atomic_decrement(partition, var) __atomic_sub_fetch(&string_base[partition].atomic.var, 1, __ATOMIC_RELAXED)
 #define string_internal_stats_add(partition, var, val) __atomic_add_fetch(&string_base[partition].var, val, __ATOMIC_RELAXED)
 #else
+#define string_stats_atomic_increment(partition, var) do {;} while(0)
+#define string_stats_atomic_decrement(partition, var) do {;} while(0)
 #define string_internal_stats_add(partition, var, val) do {;} while(0)
 #endif
 
-#define string_stats_atomic_increment(partition, var) __atomic_add_fetch(&string_base[partition].var, 1, __ATOMIC_RELAXED)
-#define string_stats_atomic_decrement(partition, var) __atomic_sub_fetch(&string_base[partition].var, 1, __ATOMIC_RELAXED)
-
 void string_statistics(size_t *inserts, size_t *deletes, size_t *searches, size_t *entries, size_t *references, size_t *memory, size_t *duplications, size_t *releases) {
     if (inserts) *inserts = 0;
     if (deletes) *deletes = 0;
@@ -72,12 +76,15 @@ void string_statistics(size_t *inserts, size_t *deletes, size_t *searches, size_
     for(size_t i = 0; i < STRING_PARTITIONS ;i++) {
         if (inserts)        *inserts        += string_base[i].inserts;
         if (deletes)        *deletes        += string_base[i].deletes;
-        if (searches)       *searches       += string_base[i].searches;
         if (entries)        *entries        += (size_t) string_base[i].entries;
-        if (references)     *references     += (size_t) string_base[i].active_references;
         if (memory)         *memory         += (size_t) string_base[i].memory;
-        if (duplications)   *duplications   += string_base[i].duplications;
-        if (releases)       *releases       += string_base[i].releases;
+
+#ifdef NETDATA_INTERNAL_CHECKS
+        if (searches)       *searches       += string_base[i].atomic.searches;
+        if (references)     *references     += (size_t) string_base[i].atomic.active_references;
+        if (duplications)   *duplications   += string_base[i].atomic.duplications;
+        if (releases)       *releases       += string_base[i].atomic.releases;
+#endif
     }
 }
 
@@ -85,7 +92,9 @@ void string_statistics(size_t *inserts, size_t *deletes, size_t *searches, size_
 #define string_entry_release(se) __atomic_sub_fetch(&((se)->refcount), 1, __ATOMIC_SEQ_CST);
 
 static inline bool string_entry_check_and_acquire(STRING *se) {
+#ifdef NETDATA_INTERNAL_CHECKS
     uint8_t partition = string_partition(se);
+#endif
 
     REFCOUNT expected, desired, count = 0;
 
@@ -125,7 +134,9 @@ STRING *string_dup(STRING *string) {
 
     string_entry_acquire(string);
 
+#ifdef NETDATA_INTERNAL_CHECKS
     uint8_t partition = string_partition(string);
+#endif
 
     // statistics
     string_stats_atomic_increment(partition, active_references);
@@ -275,7 +286,9 @@ static inline void string_index_delete(STRING *string) {
 STRING *string_strdupz(const char *str) {
     if(unlikely(!str || !*str)) return NULL;
 
+#ifdef NETDATA_INTERNAL_CHECKS
     uint8_t partition = string_partition_str(str);
+#endif
 
     size_t length = strlen(str) + 1;
     STRING *string = string_index_search(str, length);
@@ -297,7 +310,9 @@ STRING *string_strdupz(const char *str) {
 void string_freez(STRING *string) {
     if(unlikely(!string)) return;
 
+#ifdef NETDATA_INTERNAL_CHECKS
     uint8_t partition = string_partition(string);
+#endif
     REFCOUNT refcount = string_entry_release(string);
 
 #ifdef NETDATA_INTERNAL_CHECKS
@@ -518,42 +533,42 @@ int string_unittest(size_t entries) {
             strings[i] = string_strdupz(names[i]);
         }
         end_ut = now_realtime_usec();
-        fprintf(stderr, "Created %zu strings in %llu usecs\n", entries, end_ut - start_ut);
+        fprintf(stderr, "Created %zu strings in %"PRIu64" usecs\n", entries, end_ut - start_ut);
 
         start_ut = now_realtime_usec();
         for(size_t i = 0; i < entries ;i++) {
             strings[i] = string_dup(strings[i]);
         }
         end_ut = now_realtime_usec();
-        fprintf(stderr, "Cloned %zu strings in %llu usecs\n", entries, end_ut - start_ut);
+        fprintf(stderr, "Cloned %zu strings in %"PRIu64" usecs\n", entries, end_ut - start_ut);
 
         start_ut = now_realtime_usec();
         for(size_t i = 0; i < entries ;i++) {
             strings[i] = string_strdupz(string2str(strings[i]));
         }
         end_ut = now_realtime_usec();
-        fprintf(stderr, "Found %zu existing strings in %llu usecs\n", entries, end_ut - start_ut);
+        fprintf(stderr, "Found %zu existing strings in %"PRIu64" usecs\n", entries, end_ut - start_ut);
 
         start_ut = now_realtime_usec();
         for(size_t i = 0; i < entries ;i++) {
             string_freez(strings[i]);
         }
         end_ut = now_realtime_usec();
-        fprintf(stderr, "Released %zu referenced strings in %llu usecs\n", entries, end_ut - start_ut);
+        fprintf(stderr, "Released %zu referenced strings in %"PRIu64" usecs\n", entries, end_ut - start_ut);
 
         start_ut = now_realtime_usec();
         for(size_t i = 0; i < entries ;i++) {
             string_freez(strings[i]);
         }
         end_ut = now_realtime_usec();
-        fprintf(stderr, "Released (again) %zu referenced strings in %llu usecs\n", entries, end_ut - start_ut);
+        fprintf(stderr, "Released (again) %zu referenced strings in %"PRIu64" usecs\n", entries, end_ut - start_ut);
 
         start_ut = now_realtime_usec();
         for(size_t i = 0; i < entries ;i++) {
             string_freez(strings[i]);
         }
         end_ut = now_realtime_usec();
-        fprintf(stderr, "Freed %zu strings in %llu usecs\n", entries, end_ut - start_ut);
+        fprintf(stderr, "Freed %zu strings in %"PRIu64" usecs\n", entries, end_ut - start_ut);
 
         freez(strings);
 
diff --git a/libnetdata/threads/threads.c b/libnetdata/threads/threads.c
index ae3c7106..cc569060 100644
--- a/libnetdata/threads/threads.c
+++ b/libnetdata/threads/threads.c
@@ -9,8 +9,8 @@ static pthread_attr_t *netdata_threads_attr = NULL;
 
 typedef struct {
     void *arg;
-    pthread_t *thread;
     char tag[NETDATA_THREAD_NAME_MAX + 1];
+    SPINLOCK detach_lock;
     void *(*start_routine) (void *);
     NETDATA_THREAD_OPTIONS options;
 } NETDATA_THREAD;
@@ -178,16 +178,19 @@ void rrdset_thread_rda_free(void);
 void sender_thread_buffer_free(void);
 void query_target_free(void);
 void service_exits(void);
+void rrd_collector_finished(void);
 
 static void thread_cleanup(void *ptr) {
     if(netdata_thread != ptr) {
         NETDATA_THREAD *info = (NETDATA_THREAD *)ptr;
         netdata_log_error("THREADS: internal error - thread local variable does not match the one passed to this function. Expected thread '%s', passed thread '%s'", netdata_thread->tag, info->tag);
     }
+    spinlock_lock(&netdata_thread->detach_lock);
 
     if(!(netdata_thread->options & NETDATA_THREAD_OPTION_DONT_LOG_CLEANUP))
         netdata_log_info("thread with task id %d finished", gettid());
 
+    rrd_collector_finished();
     sender_thread_buffer_free();
     rrdset_thread_rda_free();
     query_target_free();
@@ -197,6 +200,7 @@ static void thread_cleanup(void *ptr) {
 
     netdata_thread->tag[0] = '\0';
 
+    spinlock_unlock(&netdata_thread->detach_lock);
     freez(netdata_thread);
     netdata_thread = NULL;
 }
@@ -232,12 +236,12 @@ void uv_thread_set_name_np(uv_thread_t ut, const char* name) {
     strncpyz(threadname, name, NETDATA_THREAD_NAME_MAX);
 
 #if defined(__FreeBSD__)
-    pthread_set_name_np(ut, threadname);
+    pthread_set_name_np(ut ? ut : pthread_self(), threadname);
 #elif defined(__APPLE__)
     // Apple can only set its own name
     UNUSED(ut);
 #else
-    ret = pthread_setname_np(ut, threadname);
+    ret = pthread_setname_np(ut ? ut : pthread_self(), threadname);
 #endif
 
     thread_name_get(true);
@@ -279,13 +283,15 @@ static void *netdata_thread_init(void *ptr) {
 }
 
 int netdata_thread_create(netdata_thread_t *thread, const char *tag, NETDATA_THREAD_OPTIONS options, void *(*start_routine) (void *), void *arg) {
-    NETDATA_THREAD *info = mallocz(sizeof(NETDATA_THREAD));
+    NETDATA_THREAD *info = callocz(1, sizeof(NETDATA_THREAD));
     info->arg = arg;
-    info->thread = thread;
     info->start_routine = start_routine;
     info->options = options;
     strncpyz(info->tag, tag, NETDATA_THREAD_NAME_MAX);
 
+    spinlock_init(&info->detach_lock);
+    spinlock_lock(&info->detach_lock);
+
     int ret = pthread_create(thread, netdata_threads_attr, netdata_thread_init, info);
     if(ret != 0)
         netdata_log_error("failed to create new thread for %s. pthread_create() failed with code %d", tag, ret);
@@ -298,6 +304,7 @@ int netdata_thread_create(netdata_thread_t *thread, const char *tag, NETDATA_THR
         }
     }
 
+    spinlock_unlock(&info->detach_lock);
     return ret;
 }
 
diff --git a/libnetdata/xxhash.h b/libnetdata/xxhash.h
new file mode 100644
index 00000000..5e2c0ed2
--- /dev/null
+++ b/libnetdata/xxhash.h
@@ -0,0 +1,6773 @@
+/*
+ * xxHash - Extremely Fast Hash algorithm
+ * Header File
+ * Copyright (C) 2012-2023 Yann Collet
+ *
+ * BSD 2-Clause License (https://www.opensource.org/licenses/bsd-license.php)
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are
+ * met:
+ *
+ *    * Redistributions of source code must retain the above copyright
+ *      notice, this list of conditions and the following disclaimer.
+ *    * Redistributions in binary form must reproduce the above
+ *      copyright notice, this list of conditions and the following disclaimer
+ *      in the documentation and/or other materials provided with the
+ *      distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ * You can contact the author at:
+ *   - xxHash homepage: https://www.xxhash.com
+ *   - xxHash source repository: https://github.com/Cyan4973/xxHash
+ */
+
+/*!
+ * @mainpage xxHash
+ *
+ * xxHash is an extremely fast non-cryptographic hash algorithm, working at RAM speed
+ * limits.
+ *
+ * It is proposed in four flavors, in three families:
+ * 1. @ref XXH32_family
+ *   - Classic 32-bit hash function. Simple, compact, and runs on almost all
+ *     32-bit and 64-bit systems.
+ * 2. @ref XXH64_family
+ *   - Classic 64-bit adaptation of XXH32. Just as simple, and runs well on most
+ *     64-bit systems (but _not_ 32-bit systems).
+ * 3. @ref XXH3_family
+ *   - Modern 64-bit and 128-bit hash function family which features improved
+ *     strength and performance across the board, especially on smaller data.
+ *     It benefits greatly from SIMD and 64-bit without requiring it.
+ *
+ * Benchmarks
+ * ---
+ * The reference system uses an Intel i7-9700K CPU, and runs Ubuntu x64 20.04.
+ * The open source benchmark program is compiled with clang v10.0 using -O3 flag.
+ *
+ * | Hash Name            | ISA ext | Width | Large Data Speed | Small Data Velocity |
+ * | -------------------- | ------- | ----: | ---------------: | ------------------: |
+ * | XXH3_64bits()        | @b AVX2 |    64 |        59.4 GB/s |               133.1 |
+ * | MeowHash             | AES-NI  |   128 |        58.2 GB/s |                52.5 |
+ * | XXH3_128bits()       | @b AVX2 |   128 |        57.9 GB/s |               118.1 |
+ * | CLHash               | PCLMUL  |    64 |        37.1 GB/s |                58.1 |
+ * | XXH3_64bits()        | @b SSE2 |    64 |        31.5 GB/s |               133.1 |
+ * | XXH3_128bits()       | @b SSE2 |   128 |        29.6 GB/s |               118.1 |
+ * | RAM sequential read  |         |   N/A |        28.0 GB/s |                 N/A |
+ * | ahash                | AES-NI  |    64 |        22.5 GB/s |               107.2 |
+ * | City64               |         |    64 |        22.0 GB/s |                76.6 |
+ * | T1ha2                |         |    64 |        22.0 GB/s |                99.0 |
+ * | City128              |         |   128 |        21.7 GB/s |                57.7 |
+ * | FarmHash             | AES-NI  |    64 |        21.3 GB/s |                71.9 |
+ * | XXH64()              |         |    64 |        19.4 GB/s |                71.0 |
+ * | SpookyHash           |         |    64 |        19.3 GB/s |                53.2 |
+ * | Mum                  |         |    64 |        18.0 GB/s |                67.0 |
+ * | CRC32C               | SSE4.2  |    32 |        13.0 GB/s |                57.9 |
+ * | XXH32()              |         |    32 |         9.7 GB/s |                71.9 |
+ * | City32               |         |    32 |         9.1 GB/s |                66.0 |
+ * | Blake3*              | @b AVX2 |   256 |         4.4 GB/s |                 8.1 |
+ * | Murmur3              |         |    32 |         3.9 GB/s |                56.1 |
+ * | SipHash*             |         |    64 |         3.0 GB/s |                43.2 |
+ * | Blake3*              | @b SSE2 |   256 |         2.4 GB/s |                 8.1 |
+ * | HighwayHash          |         |    64 |         1.4 GB/s |                 6.0 |
+ * | FNV64                |         |    64 |         1.2 GB/s |                62.7 |
+ * | Blake2*              |         |   256 |         1.1 GB/s |                 5.1 |
+ * | SHA1*                |         |   160 |         0.8 GB/s |                 5.6 |
+ * | MD5*                 |         |   128 |         0.6 GB/s |                 7.8 |
+ * @note
+ *   - Hashes which require a specific ISA extension are noted. SSE2 is also noted,
+ *     even though it is mandatory on x64.
+ *   - Hashes with an asterisk are cryptographic. Note that MD5 is non-cryptographic
+ *     by modern standards.
+ *   - Small data velocity is a rough average of algorithm's efficiency for small
+ *     data. For more accurate information, see the wiki.
+ *   - More benchmarks and strength tests are found on the wiki:
+ *         https://github.com/Cyan4973/xxHash/wiki
+ *
+ * Usage
+ * ------
+ * All xxHash variants use a similar API. Changing the algorithm is a trivial
+ * substitution.
+ *
+ * @pre
+ *    For functions which take an input and length parameter, the following
+ *    requirements are assumed:
+ *    - The range from [`input`, `input + length`) is valid, readable memory.
+ *      - The only exception is if the `length` is `0`, `input` may be `NULL`.
+ *    - For C++, the objects must have the *TriviallyCopyable* property, as the
+ *      functions access bytes directly as if it was an array of `unsigned char`.
+ *
+ * @anchor single_shot_example
+ * **Single Shot**
+ *
+ * These functions are stateless functions which hash a contiguous block of memory,
+ * immediately returning the result. They are the easiest and usually the fastest
+ * option.
+ *
+ * XXH32(), XXH64(), XXH3_64bits(), XXH3_128bits()
+ *
+ * @code{.c}
+ *   #include <string.h>
+ *   #include "xxhash.h"
+ *
+ *   // Example for a function which hashes a null terminated string with XXH32().
+ *   XXH32_hash_t hash_string(const char* string, XXH32_hash_t seed)
+ *   {
+ *       // NULL pointers are only valid if the length is zero
+ *       size_t length = (string == NULL) ? 0 : strlen(string);
+ *       return XXH32(string, length, seed);
+ *   }
+ * @endcode
+ *
+ * @anchor streaming_example
+ * **Streaming**
+ *
+ * These groups of functions allow incremental hashing of unknown size, even
+ * more than what would fit in a size_t.
+ *
+ * XXH32_reset(), XXH64_reset(), XXH3_64bits_reset(), XXH3_128bits_reset()
+ *
+ * @code{.c}
+ *   #include <stdio.h>
+ *   #include <assert.h>
+ *   #include "xxhash.h"
+ *   // Example for a function which hashes a FILE incrementally with XXH3_64bits().
+ *   XXH64_hash_t hashFile(FILE* f)
+ *   {
+ *       // Allocate a state struct. Do not just use malloc() or new.
+ *       XXH3_state_t* state = XXH3_createState();
+ *       assert(state != NULL && "Out of memory!");
+ *       // Reset the state to start a new hashing session.
+ *       XXH3_64bits_reset(state);
+ *       char buffer[4096];
+ *       size_t count;
+ *       // Read the file in chunks
+ *       while ((count = fread(buffer, 1, sizeof(buffer), f)) != 0) {
+ *           // Run update() as many times as necessary to process the data
+ *           XXH3_64bits_update(state, buffer, count);
+ *       }
+ *       // Retrieve the finalized hash. This will not change the state.
+ *       XXH64_hash_t result = XXH3_64bits_digest(state);
+ *       // Free the state. Do not use free().
+ *       XXH3_freeState(state);
+ *       return result;
+ *   }
+ * @endcode
+ *
+ * @file xxhash.h
+ * xxHash prototypes and implementation
+ */
+
+#if defined (__cplusplus)
+extern "C" {
+#endif
+
+/* ****************************
+ *  INLINE mode
+ ******************************/
+/*!
+ * @defgroup public Public API
+ * Contains details on the public xxHash functions.
+ * @{
+ */
+#ifdef XXH_DOXYGEN
+/*!
+ * @brief Gives access to internal state declaration, required for static allocation.
+ *
+ * Incompatible with dynamic linking, due to risks of ABI changes.
+ *
+ * Usage:
+ * @code{.c}
+ *     #define XXH_STATIC_LINKING_ONLY
+ *     #include "xxhash.h"
+ * @endcode
+ */
+#  define XXH_STATIC_LINKING_ONLY
+/* Do not undef XXH_STATIC_LINKING_ONLY for Doxygen */
+
+/*!
+ * @brief Gives access to internal definitions.
+ *
+ * Usage:
+ * @code{.c}
+ *     #define XXH_STATIC_LINKING_ONLY
+ *     #define XXH_IMPLEMENTATION
+ *     #include "xxhash.h"
+ * @endcode
+ */
+#  define XXH_IMPLEMENTATION
+/* Do not undef XXH_IMPLEMENTATION for Doxygen */
+
+/*!
+ * @brief Exposes the implementation and marks all functions as `inline`.
+ *
+ * Use these build macros to inline xxhash into the target unit.
+ * Inlining improves performance on small inputs, especially when the length is
+ * expressed as a compile-time constant:
+ *
+ *  https://fastcompression.blogspot.com/2018/03/xxhash-for-small-keys-impressive-power.html
+ *
+ * It also keeps xxHash symbols private to the unit, so they are not exported.
+ *
+ * Usage:
+ * @code{.c}
+ *     #define XXH_INLINE_ALL
+ *     #include "xxhash.h"
+ * @endcode
+ * Do not compile and link xxhash.o as a separate object, as it is not useful.
+ */
+#  define XXH_INLINE_ALL
+#  undef XXH_INLINE_ALL
+/*!
+ * @brief Exposes the implementation without marking functions as inline.
+ */
+#  define XXH_PRIVATE_API
+#  undef XXH_PRIVATE_API
+/*!
+ * @brief Emulate a namespace by transparently prefixing all symbols.
+ *
+ * If you want to include _and expose_ xxHash functions from within your own
+ * library, but also want to avoid symbol collisions with other libraries which
+ * may also include xxHash, you can use @ref XXH_NAMESPACE to automatically prefix
+ * any public symbol from xxhash library with the value of @ref XXH_NAMESPACE
+ * (therefore, avoid empty or numeric values).
+ *
+ * Note that no change is required within the calling program as long as it
+ * includes `xxhash.h`: Regular symbol names will be automatically translated
+ * by this header.
+ */
+#  define XXH_NAMESPACE /* YOUR NAME HERE */
+#  undef XXH_NAMESPACE
+#endif
+
+#if (defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API)) \
+    && !defined(XXH_INLINE_ALL_31684351384)
+   /* this section should be traversed only once */
+#  define XXH_INLINE_ALL_31684351384
+   /* give access to the advanced API, required to compile implementations */
+#  undef XXH_STATIC_LINKING_ONLY   /* avoid macro redef */
+#  define XXH_STATIC_LINKING_ONLY
+   /* make all functions private */
+#  undef XXH_PUBLIC_API
+#  if defined(__GNUC__)
+#    define XXH_PUBLIC_API static __inline __attribute__((unused))
+#  elif defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
+#    define XXH_PUBLIC_API static inline
+#  elif defined(_MSC_VER)
+#    define XXH_PUBLIC_API static __inline
+#  else
+     /* note: this version may generate warnings for unused static functions */
+#    define XXH_PUBLIC_API static
+#  endif
+
+   /*
+    * This part deals with the special case where a unit wants to inline xxHash,
+    * but "xxhash.h" has previously been included without XXH_INLINE_ALL,
+    * such as part of some previously included *.h header file.
+    * Without further action, the new include would just be ignored,
+    * and functions would effectively _not_ be inlined (silent failure).
+    * The following macros solve this situation by prefixing all inlined names,
+    * avoiding naming collision with previous inclusions.
+    */
+   /* Before that, we unconditionally #undef all symbols,
+    * in case they were already defined with XXH_NAMESPACE.
+    * They will then be redefined for XXH_INLINE_ALL
+    */
+#  undef XXH_versionNumber
+    /* XXH32 */
+#  undef XXH32
+#  undef XXH32_createState
+#  undef XXH32_freeState
+#  undef XXH32_reset
+#  undef XXH32_update
+#  undef XXH32_digest
+#  undef XXH32_copyState
+#  undef XXH32_canonicalFromHash
+#  undef XXH32_hashFromCanonical
+    /* XXH64 */
+#  undef XXH64
+#  undef XXH64_createState
+#  undef XXH64_freeState
+#  undef XXH64_reset
+#  undef XXH64_update
+#  undef XXH64_digest
+#  undef XXH64_copyState
+#  undef XXH64_canonicalFromHash
+#  undef XXH64_hashFromCanonical
+    /* XXH3_64bits */
+#  undef XXH3_64bits
+#  undef XXH3_64bits_withSecret
+#  undef XXH3_64bits_withSeed
+#  undef XXH3_64bits_withSecretandSeed
+#  undef XXH3_createState
+#  undef XXH3_freeState
+#  undef XXH3_copyState
+#  undef XXH3_64bits_reset
+#  undef XXH3_64bits_reset_withSeed
+#  undef XXH3_64bits_reset_withSecret
+#  undef XXH3_64bits_update
+#  undef XXH3_64bits_digest
+#  undef XXH3_generateSecret
+    /* XXH3_128bits */
+#  undef XXH128
+#  undef XXH3_128bits
+#  undef XXH3_128bits_withSeed
+#  undef XXH3_128bits_withSecret
+#  undef XXH3_128bits_reset
+#  undef XXH3_128bits_reset_withSeed
+#  undef XXH3_128bits_reset_withSecret
+#  undef XXH3_128bits_reset_withSecretandSeed
+#  undef XXH3_128bits_update
+#  undef XXH3_128bits_digest
+#  undef XXH128_isEqual
+#  undef XXH128_cmp
+#  undef XXH128_canonicalFromHash
+#  undef XXH128_hashFromCanonical
+    /* Finally, free the namespace itself */
+#  undef XXH_NAMESPACE
+
+    /* employ the namespace for XXH_INLINE_ALL */
+#  define XXH_NAMESPACE XXH_INLINE_
+   /*
+    * Some identifiers (enums, type names) are not symbols,
+    * but they must nonetheless be renamed to avoid redeclaration.
+    * Alternative solution: do not redeclare them.
+    * However, this requires some #ifdefs, and has a more dispersed impact.
+    * Meanwhile, renaming can be achieved in a single place.
+    */
+#  define XXH_IPREF(Id)   XXH_NAMESPACE ## Id
+#  define XXH_OK XXH_IPREF(XXH_OK)
+#  define XXH_ERROR XXH_IPREF(XXH_ERROR)
+#  define XXH_errorcode XXH_IPREF(XXH_errorcode)
+#  define XXH32_canonical_t  XXH_IPREF(XXH32_canonical_t)
+#  define XXH64_canonical_t  XXH_IPREF(XXH64_canonical_t)
+#  define XXH128_canonical_t XXH_IPREF(XXH128_canonical_t)
+#  define XXH32_state_s XXH_IPREF(XXH32_state_s)
+#  define XXH32_state_t XXH_IPREF(XXH32_state_t)
+#  define XXH64_state_s XXH_IPREF(XXH64_state_s)
+#  define XXH64_state_t XXH_IPREF(XXH64_state_t)
+#  define XXH3_state_s  XXH_IPREF(XXH3_state_s)
+#  define XXH3_state_t  XXH_IPREF(XXH3_state_t)
+#  define XXH128_hash_t XXH_IPREF(XXH128_hash_t)
+   /* Ensure the header is parsed again, even if it was previously included */
+#  undef XXHASH_H_5627135585666179
+#  undef XXHASH_H_STATIC_13879238742
+#endif /* XXH_INLINE_ALL || XXH_PRIVATE_API */
+
+/* ****************************************************************
+ *  Stable API
+ *****************************************************************/
+#ifndef XXHASH_H_5627135585666179
+#define XXHASH_H_5627135585666179 1
+
+/*! @brief Marks a global symbol. */
+#if !defined(XXH_INLINE_ALL) && !defined(XXH_PRIVATE_API)
+#  if defined(WIN32) && defined(_MSC_VER) && (defined(XXH_IMPORT) || defined(XXH_EXPORT))
+#    ifdef XXH_EXPORT
+#      define XXH_PUBLIC_API __declspec(dllexport)
+#    elif XXH_IMPORT
+#      define XXH_PUBLIC_API __declspec(dllimport)
+#    endif
+#  else
+#    define XXH_PUBLIC_API   /* do nothing */
+#  endif
+#endif
+
+#ifdef XXH_NAMESPACE
+#  define XXH_CAT(A,B) A##B
+#  define XXH_NAME2(A,B) XXH_CAT(A,B)
+#  define XXH_versionNumber XXH_NAME2(XXH_NAMESPACE, XXH_versionNumber)
+/* XXH32 */
+#  define XXH32 XXH_NAME2(XXH_NAMESPACE, XXH32)
+#  define XXH32_createState XXH_NAME2(XXH_NAMESPACE, XXH32_createState)
+#  define XXH32_freeState XXH_NAME2(XXH_NAMESPACE, XXH32_freeState)
+#  define XXH32_reset XXH_NAME2(XXH_NAMESPACE, XXH32_reset)
+#  define XXH32_update XXH_NAME2(XXH_NAMESPACE, XXH32_update)
+#  define XXH32_digest XXH_NAME2(XXH_NAMESPACE, XXH32_digest)
+#  define XXH32_copyState XXH_NAME2(XXH_NAMESPACE, XXH32_copyState)
+#  define XXH32_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH32_canonicalFromHash)
+#  define XXH32_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH32_hashFromCanonical)
+/* XXH64 */
+#  define XXH64 XXH_NAME2(XXH_NAMESPACE, XXH64)
+#  define XXH64_createState XXH_NAME2(XXH_NAMESPACE, XXH64_createState)
+#  define XXH64_freeState XXH_NAME2(XXH_NAMESPACE, XXH64_freeState)
+#  define XXH64_reset XXH_NAME2(XXH_NAMESPACE, XXH64_reset)
+#  define XXH64_update XXH_NAME2(XXH_NAMESPACE, XXH64_update)
+#  define XXH64_digest XXH_NAME2(XXH_NAMESPACE, XXH64_digest)
+#  define XXH64_copyState XXH_NAME2(XXH_NAMESPACE, XXH64_copyState)
+#  define XXH64_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH64_canonicalFromHash)
+#  define XXH64_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH64_hashFromCanonical)
+/* XXH3_64bits */
+#  define XXH3_64bits XXH_NAME2(XXH_NAMESPACE, XXH3_64bits)
+#  define XXH3_64bits_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSecret)
+#  define XXH3_64bits_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSeed)
+#  define XXH3_64bits_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSecretandSeed)
+#  define XXH3_createState XXH_NAME2(XXH_NAMESPACE, XXH3_createState)
+#  define XXH3_freeState XXH_NAME2(XXH_NAMESPACE, XXH3_freeState)
+#  define XXH3_copyState XXH_NAME2(XXH_NAMESPACE, XXH3_copyState)
+#  define XXH3_64bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset)
+#  define XXH3_64bits_reset_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSeed)
+#  define XXH3_64bits_reset_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSecret)
+#  define XXH3_64bits_reset_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSecretandSeed)
+#  define XXH3_64bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_update)
+#  define XXH3_64bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_digest)
+#  define XXH3_generateSecret XXH_NAME2(XXH_NAMESPACE, XXH3_generateSecret)
+#  define XXH3_generateSecret_fromSeed XXH_NAME2(XXH_NAMESPACE, XXH3_generateSecret_fromSeed)
+/* XXH3_128bits */
+#  define XXH128 XXH_NAME2(XXH_NAMESPACE, XXH128)
+#  define XXH3_128bits XXH_NAME2(XXH_NAMESPACE, XXH3_128bits)
+#  define XXH3_128bits_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSeed)
+#  define XXH3_128bits_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSecret)
+#  define XXH3_128bits_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSecretandSeed)
+#  define XXH3_128bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset)
+#  define XXH3_128bits_reset_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSeed)
+#  define XXH3_128bits_reset_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSecret)
+#  define XXH3_128bits_reset_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSecretandSeed)
+#  define XXH3_128bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_update)
+#  define XXH3_128bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_digest)
+#  define XXH128_isEqual XXH_NAME2(XXH_NAMESPACE, XXH128_isEqual)
+#  define XXH128_cmp     XXH_NAME2(XXH_NAMESPACE, XXH128_cmp)
+#  define XXH128_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH128_canonicalFromHash)
+#  define XXH128_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH128_hashFromCanonical)
+#endif
+
+
+/* *************************************
+*  Compiler specifics
+***************************************/
+
+/* specific declaration modes for Windows */
+#if !defined(XXH_INLINE_ALL) && !defined(XXH_PRIVATE_API)
+#  if defined(WIN32) && defined(_MSC_VER) && (defined(XXH_IMPORT) || defined(XXH_EXPORT))
+#    ifdef XXH_EXPORT
+#      define XXH_PUBLIC_API __declspec(dllexport)
+#    elif XXH_IMPORT
+#      define XXH_PUBLIC_API __declspec(dllimport)
+#    endif
+#  else
+#    define XXH_PUBLIC_API   /* do nothing */
+#  endif
+#endif
+
+#if defined (__GNUC__)
+# define XXH_CONSTF  __attribute__((const))
+# define XXH_PUREF   __attribute__((pure))
+# define XXH_MALLOCF __attribute__((malloc))
+#else
+# define XXH_CONSTF  /* disable */
+# define XXH_PUREF
+# define XXH_MALLOCF
+#endif
+
+/* *************************************
+*  Version
+***************************************/
+#define XXH_VERSION_MAJOR    0
+#define XXH_VERSION_MINOR    8
+#define XXH_VERSION_RELEASE  2
+/*! @brief Version number, encoded as two digits each */
+#define XXH_VERSION_NUMBER  (XXH_VERSION_MAJOR *100*100 + XXH_VERSION_MINOR *100 + XXH_VERSION_RELEASE)
+
+/*!
+ * @brief Obtains the xxHash version.
+ *
+ * This is mostly useful when xxHash is compiled as a shared library,
+ * since the returned value comes from the library, as opposed to header file.
+ *
+ * @return @ref XXH_VERSION_NUMBER of the invoked library.
+ */
+XXH_PUBLIC_API XXH_CONSTF unsigned XXH_versionNumber (void);
+
+
+/* ****************************
+*  Common basic types
+******************************/
+#include <stddef.h>   /* size_t */
+/*!
+ * @brief Exit code for the streaming API.
+ */
+typedef enum {
+    XXH_OK = 0, /*!< OK */
+    XXH_ERROR   /*!< Error */
+} XXH_errorcode;
+
+
+/*-**********************************************************************
+*  32-bit hash
+************************************************************************/
+#if defined(XXH_DOXYGEN) /* Don't show <stdint.h> include */
+/*!
+ * @brief An unsigned 32-bit integer.
+ *
+ * Not necessarily defined to `uint32_t` but functionally equivalent.
+ */
+typedef uint32_t XXH32_hash_t;
+
+#elif !defined (__VMS) \
+  && (defined (__cplusplus) \
+  || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+#   include <stdint.h>
+    typedef uint32_t XXH32_hash_t;
+
+#else
+#   include <limits.h>
+#   if UINT_MAX == 0xFFFFFFFFUL
+      typedef unsigned int XXH32_hash_t;
+#   elif ULONG_MAX == 0xFFFFFFFFUL
+      typedef unsigned long XXH32_hash_t;
+#   else
+#     error "unsupported platform: need a 32-bit type"
+#   endif
+#endif
+
+/*!
+ * @}
+ *
+ * @defgroup XXH32_family XXH32 family
+ * @ingroup public
+ * Contains functions used in the classic 32-bit xxHash algorithm.
+ *
+ * @note
+ *   XXH32 is useful for older platforms, with no or poor 64-bit performance.
+ *   Note that the @ref XXH3_family provides competitive speed for both 32-bit
+ *   and 64-bit systems, and offers true 64/128 bit hash results.
+ *
+ * @see @ref XXH64_family, @ref XXH3_family : Other xxHash families
+ * @see @ref XXH32_impl for implementation details
+ * @{
+ */
+
+/*!
+ * @brief Calculates the 32-bit hash of @p input using xxHash32.
+ *
+ * Speed on Core 2 Duo @ 3 GHz (single thread, SMHasher benchmark): 5.4 GB/s
+ *
+ * See @ref single_shot_example "Single Shot Example" for an example.
+ *
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ * @param seed The 32-bit seed to alter the hash's output predictably.
+ *
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return The calculated 32-bit hash value.
+ *
+ * @see
+ *    XXH64(), XXH3_64bits_withSeed(), XXH3_128bits_withSeed(), XXH128():
+ *    Direct equivalents for the other variants of xxHash.
+ * @see
+ *    XXH32_createState(), XXH32_update(), XXH32_digest(): Streaming version.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32 (const void* input, size_t length, XXH32_hash_t seed);
+
+#ifndef XXH_NO_STREAM
+/*!
+ * Streaming functions generate the xxHash value from an incremental input.
+ * This method is slower than single-call functions, due to state management.
+ * For small inputs, prefer `XXH32()` and `XXH64()`, which are better optimized.
+ *
+ * An XXH state must first be allocated using `XXH*_createState()`.
+ *
+ * Start a new hash by initializing the state with a seed using `XXH*_reset()`.
+ *
+ * Then, feed the hash state by calling `XXH*_update()` as many times as necessary.
+ *
+ * The function returns an error code, with 0 meaning OK, and any other value
+ * meaning there is an error.
+ *
+ * Finally, a hash value can be produced anytime, by using `XXH*_digest()`.
+ * This function returns the nn-bits hash as an int or long long.
+ *
+ * It's still possible to continue inserting input into the hash state after a
+ * digest, and generate new hash values later on by invoking `XXH*_digest()`.
+ *
+ * When done, release the state using `XXH*_freeState()`.
+ *
+ * @see streaming_example at the top of @ref xxhash.h for an example.
+ */
+
+/*!
+ * @typedef struct XXH32_state_s XXH32_state_t
+ * @brief The opaque state struct for the XXH32 streaming API.
+ *
+ * @see XXH32_state_s for details.
+ */
+typedef struct XXH32_state_s XXH32_state_t;
+
+/*!
+ * @brief Allocates an @ref XXH32_state_t.
+ *
+ * Must be freed with XXH32_freeState().
+ * @return An allocated XXH32_state_t on success, `NULL` on failure.
+ */
+XXH_PUBLIC_API XXH_MALLOCF XXH32_state_t* XXH32_createState(void);
+/*!
+ * @brief Frees an @ref XXH32_state_t.
+ *
+ * Must be allocated with XXH32_createState().
+ * @param statePtr A pointer to an @ref XXH32_state_t allocated with @ref XXH32_createState().
+ * @return XXH_OK.
+ */
+XXH_PUBLIC_API XXH_errorcode  XXH32_freeState(XXH32_state_t* statePtr);
+/*!
+ * @brief Copies one @ref XXH32_state_t to another.
+ *
+ * @param dst_state The state to copy to.
+ * @param src_state The state to copy from.
+ * @pre
+ *   @p dst_state and @p src_state must not be `NULL` and must not overlap.
+ */
+XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dst_state, const XXH32_state_t* src_state);
+
+/*!
+ * @brief Resets an @ref XXH32_state_t to begin a new hash.
+ *
+ * This function resets and seeds a state. Call it before @ref XXH32_update().
+ *
+ * @param statePtr The state struct to reset.
+ * @param seed The 32-bit seed to alter the hash result predictably.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success, @ref XXH_ERROR on failure.
+ */
+XXH_PUBLIC_API XXH_errorcode XXH32_reset  (XXH32_state_t* statePtr, XXH32_hash_t seed);
+
+/*!
+ * @brief Consumes a block of @p input to an @ref XXH32_state_t.
+ *
+ * Call this to incrementally consume blocks of data.
+ *
+ * @param statePtr The state struct to update.
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return @ref XXH_OK on success, @ref XXH_ERROR on failure.
+ */
+XXH_PUBLIC_API XXH_errorcode XXH32_update (XXH32_state_t* statePtr, const void* input, size_t length);
+
+/*!
+ * @brief Returns the calculated hash value from an @ref XXH32_state_t.
+ *
+ * @note
+ *   Calling XXH32_digest() will not affect @p statePtr, so you can update,
+ *   digest, and update again.
+ *
+ * @param statePtr The state struct to calculate the hash from.
+ *
+ * @pre
+ *  @p statePtr must not be `NULL`.
+ *
+ * @return The calculated xxHash32 value from that state.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32_digest (const XXH32_state_t* statePtr);
+#endif /* !XXH_NO_STREAM */
+
+/*******   Canonical representation   *******/
+
+/*
+ * The default return values from XXH functions are unsigned 32 and 64 bit
+ * integers.
+ * This the simplest and fastest format for further post-processing.
+ *
+ * However, this leaves open the question of what is the order on the byte level,
+ * since little and big endian conventions will store the same number differently.
+ *
+ * The canonical representation settles this issue by mandating big-endian
+ * convention, the same convention as human-readable numbers (large digits first).
+ *
+ * When writing hash values to storage, sending them over a network, or printing
+ * them, it's highly recommended to use the canonical representation to ensure
+ * portability across a wider range of systems, present and future.
+ *
+ * The following functions allow transformation of hash values to and from
+ * canonical format.
+ */
+
+/*!
+ * @brief Canonical (big endian) representation of @ref XXH32_hash_t.
+ */
+typedef struct {
+    unsigned char digest[4]; /*!< Hash bytes, big endian */
+} XXH32_canonical_t;
+
+/*!
+ * @brief Converts an @ref XXH32_hash_t to a big endian @ref XXH32_canonical_t.
+ *
+ * @param dst The @ref XXH32_canonical_t pointer to be stored to.
+ * @param hash The @ref XXH32_hash_t to be converted.
+ *
+ * @pre
+ *   @p dst must not be `NULL`.
+ */
+XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash);
+
+/*!
+ * @brief Converts an @ref XXH32_canonical_t to a native @ref XXH32_hash_t.
+ *
+ * @param src The @ref XXH32_canonical_t to convert.
+ *
+ * @pre
+ *   @p src must not be `NULL`.
+ *
+ * @return The converted hash.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src);
+
+
+/*! @cond Doxygen ignores this part */
+#ifdef __has_attribute
+# define XXH_HAS_ATTRIBUTE(x) __has_attribute(x)
+#else
+# define XXH_HAS_ATTRIBUTE(x) 0
+#endif
+/*! @endcond */
+
+/*! @cond Doxygen ignores this part */
+/*
+ * C23 __STDC_VERSION__ number hasn't been specified yet. For now
+ * leave as `201711L` (C17 + 1).
+ * TODO: Update to correct value when its been specified.
+ */
+#define XXH_C23_VN 201711L
+/*! @endcond */
+
+/*! @cond Doxygen ignores this part */
+/* C-language Attributes are added in C23. */
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= XXH_C23_VN) && defined(__has_c_attribute)
+# define XXH_HAS_C_ATTRIBUTE(x) __has_c_attribute(x)
+#else
+# define XXH_HAS_C_ATTRIBUTE(x) 0
+#endif
+/*! @endcond */
+
+/*! @cond Doxygen ignores this part */
+#if defined(__cplusplus) && defined(__has_cpp_attribute)
+# define XXH_HAS_CPP_ATTRIBUTE(x) __has_cpp_attribute(x)
+#else
+# define XXH_HAS_CPP_ATTRIBUTE(x) 0
+#endif
+/*! @endcond */
+
+/*! @cond Doxygen ignores this part */
+/*
+ * Define XXH_FALLTHROUGH macro for annotating switch case with the 'fallthrough' attribute
+ * introduced in CPP17 and C23.
+ * CPP17 : https://en.cppreference.com/w/cpp/language/attributes/fallthrough
+ * C23   : https://en.cppreference.com/w/c/language/attributes/fallthrough
+ */
+#if XXH_HAS_C_ATTRIBUTE(fallthrough) || XXH_HAS_CPP_ATTRIBUTE(fallthrough)
+# define XXH_FALLTHROUGH [[fallthrough]]
+#elif XXH_HAS_ATTRIBUTE(__fallthrough__)
+# define XXH_FALLTHROUGH __attribute__ ((__fallthrough__))
+#else
+# define XXH_FALLTHROUGH /* fallthrough */
+#endif
+/*! @endcond */
+
+/*! @cond Doxygen ignores this part */
+/*
+ * Define XXH_NOESCAPE for annotated pointers in public API.
+ * https://clang.llvm.org/docs/AttributeReference.html#noescape
+ * As of writing this, only supported by clang.
+ */
+#if XXH_HAS_ATTRIBUTE(noescape)
+# define XXH_NOESCAPE __attribute__((noescape))
+#else
+# define XXH_NOESCAPE
+#endif
+/*! @endcond */
+
+
+/*!
+ * @}
+ * @ingroup public
+ * @{
+ */
+
+#ifndef XXH_NO_LONG_LONG
+/*-**********************************************************************
+*  64-bit hash
+************************************************************************/
+#if defined(XXH_DOXYGEN) /* don't include <stdint.h> */
+/*!
+ * @brief An unsigned 64-bit integer.
+ *
+ * Not necessarily defined to `uint64_t` but functionally equivalent.
+ */
+typedef uint64_t XXH64_hash_t;
+#elif !defined (__VMS) \
+  && (defined (__cplusplus) \
+  || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+#  include <stdint.h>
+   typedef uint64_t XXH64_hash_t;
+#else
+#  include <limits.h>
+#  if defined(__LP64__) && ULONG_MAX == 0xFFFFFFFFFFFFFFFFULL
+     /* LP64 ABI says uint64_t is unsigned long */
+     typedef unsigned long XXH64_hash_t;
+#  else
+     /* the following type must have a width of 64-bit */
+     typedef unsigned long long XXH64_hash_t;
+#  endif
+#endif
+
+/*!
+ * @}
+ *
+ * @defgroup XXH64_family XXH64 family
+ * @ingroup public
+ * @{
+ * Contains functions used in the classic 64-bit xxHash algorithm.
+ *
+ * @note
+ *   XXH3 provides competitive speed for both 32-bit and 64-bit systems,
+ *   and offers true 64/128 bit hash results.
+ *   It provides better speed for systems with vector processing capabilities.
+ */
+
+/*!
+ * @brief Calculates the 64-bit hash of @p input using xxHash64.
+ *
+ * This function usually runs faster on 64-bit systems, but slower on 32-bit
+ * systems (see benchmark).
+ *
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ * @param seed The 64-bit seed to alter the hash's output predictably.
+ *
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return The calculated 64-bit hash.
+ *
+ * @see
+ *    XXH32(), XXH3_64bits_withSeed(), XXH3_128bits_withSeed(), XXH128():
+ *    Direct equivalents for the other variants of xxHash.
+ * @see
+ *    XXH64_createState(), XXH64_update(), XXH64_digest(): Streaming version.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed);
+
+/*******   Streaming   *******/
+#ifndef XXH_NO_STREAM
+/*!
+ * @brief The opaque state struct for the XXH64 streaming API.
+ *
+ * @see XXH64_state_s for details.
+ */
+typedef struct XXH64_state_s XXH64_state_t;   /* incomplete type */
+
+/*!
+ * @brief Allocates an @ref XXH64_state_t.
+ *
+ * Must be freed with XXH64_freeState().
+ * @return An allocated XXH64_state_t on success, `NULL` on failure.
+ */
+XXH_PUBLIC_API XXH_MALLOCF XXH64_state_t* XXH64_createState(void);
+
+/*!
+ * @brief Frees an @ref XXH64_state_t.
+ *
+ * Must be allocated with XXH64_createState().
+ * @param statePtr A pointer to an @ref XXH64_state_t allocated with @ref XXH64_createState().
+ * @return XXH_OK.
+ */
+XXH_PUBLIC_API XXH_errorcode  XXH64_freeState(XXH64_state_t* statePtr);
+
+/*!
+ * @brief Copies one @ref XXH64_state_t to another.
+ *
+ * @param dst_state The state to copy to.
+ * @param src_state The state to copy from.
+ * @pre
+ *   @p dst_state and @p src_state must not be `NULL` and must not overlap.
+ */
+XXH_PUBLIC_API void XXH64_copyState(XXH_NOESCAPE XXH64_state_t* dst_state, const XXH64_state_t* src_state);
+
+/*!
+ * @brief Resets an @ref XXH64_state_t to begin a new hash.
+ *
+ * This function resets and seeds a state. Call it before @ref XXH64_update().
+ *
+ * @param statePtr The state struct to reset.
+ * @param seed The 64-bit seed to alter the hash result predictably.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success, @ref XXH_ERROR on failure.
+ */
+XXH_PUBLIC_API XXH_errorcode XXH64_reset  (XXH_NOESCAPE XXH64_state_t* statePtr, XXH64_hash_t seed);
+
+/*!
+ * @brief Consumes a block of @p input to an @ref XXH64_state_t.
+ *
+ * Call this to incrementally consume blocks of data.
+ *
+ * @param statePtr The state struct to update.
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return @ref XXH_OK on success, @ref XXH_ERROR on failure.
+ */
+XXH_PUBLIC_API XXH_errorcode XXH64_update (XXH_NOESCAPE XXH64_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length);
+
+/*!
+ * @brief Returns the calculated hash value from an @ref XXH64_state_t.
+ *
+ * @note
+ *   Calling XXH64_digest() will not affect @p statePtr, so you can update,
+ *   digest, and update again.
+ *
+ * @param statePtr The state struct to calculate the hash from.
+ *
+ * @pre
+ *  @p statePtr must not be `NULL`.
+ *
+ * @return The calculated xxHash64 value from that state.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64_digest (XXH_NOESCAPE const XXH64_state_t* statePtr);
+#endif /* !XXH_NO_STREAM */
+/*******   Canonical representation   *******/
+
+/*!
+ * @brief Canonical (big endian) representation of @ref XXH64_hash_t.
+ */
+typedef struct { unsigned char digest[sizeof(XXH64_hash_t)]; } XXH64_canonical_t;
+
+/*!
+ * @brief Converts an @ref XXH64_hash_t to a big endian @ref XXH64_canonical_t.
+ *
+ * @param dst The @ref XXH64_canonical_t pointer to be stored to.
+ * @param hash The @ref XXH64_hash_t to be converted.
+ *
+ * @pre
+ *   @p dst must not be `NULL`.
+ */
+XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH_NOESCAPE XXH64_canonical_t* dst, XXH64_hash_t hash);
+
+/*!
+ * @brief Converts an @ref XXH64_canonical_t to a native @ref XXH64_hash_t.
+ *
+ * @param src The @ref XXH64_canonical_t to convert.
+ *
+ * @pre
+ *   @p src must not be `NULL`.
+ *
+ * @return The converted hash.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64_hashFromCanonical(XXH_NOESCAPE const XXH64_canonical_t* src);
+
+#ifndef XXH_NO_XXH3
+
+/*!
+ * @}
+ * ************************************************************************
+ * @defgroup XXH3_family XXH3 family
+ * @ingroup public
+ * @{
+ *
+ * XXH3 is a more recent hash algorithm featuring:
+ *  - Improved speed for both small and large inputs
+ *  - True 64-bit and 128-bit outputs
+ *  - SIMD acceleration
+ *  - Improved 32-bit viability
+ *
+ * Speed analysis methodology is explained here:
+ *
+ *    https://fastcompression.blogspot.com/2019/03/presenting-xxh3.html
+ *
+ * Compared to XXH64, expect XXH3 to run approximately
+ * ~2x faster on large inputs and >3x faster on small ones,
+ * exact differences vary depending on platform.
+ *
+ * XXH3's speed benefits greatly from SIMD and 64-bit arithmetic,
+ * but does not require it.
+ * Most 32-bit and 64-bit targets that can run XXH32 smoothly can run XXH3
+ * at competitive speeds, even without vector support. Further details are
+ * explained in the implementation.
+ *
+ * XXH3 has a fast scalar implementation, but it also includes accelerated SIMD
+ * implementations for many common platforms:
+ *   - AVX512
+ *   - AVX2
+ *   - SSE2
+ *   - ARM NEON
+ *   - WebAssembly SIMD128
+ *   - POWER8 VSX
+ *   - s390x ZVector
+ * This can be controlled via the @ref XXH_VECTOR macro, but it automatically
+ * selects the best version according to predefined macros. For the x86 family, an
+ * automatic runtime dispatcher is included separately in @ref xxh_x86dispatch.c.
+ *
+ * XXH3 implementation is portable:
+ * it has a generic C90 formulation that can be compiled on any platform,
+ * all implementations generate exactly the same hash value on all platforms.
+ * Starting from v0.8.0, it's also labelled "stable", meaning that
+ * any future version will also generate the same hash value.
+ *
+ * XXH3 offers 2 variants, _64bits and _128bits.
+ *
+ * When only 64 bits are needed, prefer invoking the _64bits variant, as it
+ * reduces the amount of mixing, resulting in faster speed on small inputs.
+ * It's also generally simpler to manipulate a scalar return type than a struct.
+ *
+ * The API supports one-shot hashing, streaming mode, and custom secrets.
+ */
+/*-**********************************************************************
+*  XXH3 64-bit variant
+************************************************************************/
+
+/*!
+ * @brief 64-bit unseeded variant of XXH3.
+ *
+ * This is equivalent to @ref XXH3_64bits_withSeed() with a seed of 0, however
+ * it may have slightly better performance due to constant propagation of the
+ * defaults.
+ *
+ * @see
+ *    XXH32(), XXH64(), XXH3_128bits(): equivalent for the other xxHash algorithms
+ * @see
+ *    XXH3_64bits_withSeed(), XXH3_64bits_withSecret(): other seeding variants
+ * @see
+ *    XXH3_64bits_reset(), XXH3_64bits_update(), XXH3_64bits_digest(): Streaming version.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits(XXH_NOESCAPE const void* input, size_t length);
+
+/*!
+ * @brief 64-bit seeded variant of XXH3
+ *
+ * This variant generates a custom secret on the fly based on default secret
+ * altered using the `seed` value.
+ *
+ * While this operation is decently fast, note that it's not completely free.
+ *
+ * @note
+ *    seed == 0 produces the same results as @ref XXH3_64bits().
+ *
+ * @param input The data to hash
+ * @param length The length
+ * @param seed The 64-bit seed to alter the state.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_withSeed(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed);
+
+/*!
+ * The bare minimum size for a custom secret.
+ *
+ * @see
+ *  XXH3_64bits_withSecret(), XXH3_64bits_reset_withSecret(),
+ *  XXH3_128bits_withSecret(), XXH3_128bits_reset_withSecret().
+ */
+#define XXH3_SECRET_SIZE_MIN 136
+
+/*!
+ * @brief 64-bit variant of XXH3 with a custom "secret".
+ *
+ * It's possible to provide any blob of bytes as a "secret" to generate the hash.
+ * This makes it more difficult for an external actor to prepare an intentional collision.
+ * The main condition is that secretSize *must* be large enough (>= XXH3_SECRET_SIZE_MIN).
+ * However, the quality of the secret impacts the dispersion of the hash algorithm.
+ * Therefore, the secret _must_ look like a bunch of random bytes.
+ * Avoid "trivial" or structured data such as repeated sequences or a text document.
+ * Whenever in doubt about the "randomness" of the blob of bytes,
+ * consider employing "XXH3_generateSecret()" instead (see below).
+ * It will generate a proper high entropy secret derived from the blob of bytes.
+ * Another advantage of using XXH3_generateSecret() is that
+ * it guarantees that all bits within the initial blob of bytes
+ * will impact every bit of the output.
+ * This is not necessarily the case when using the blob of bytes directly
+ * because, when hashing _small_ inputs, only a portion of the secret is employed.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_withSecret(XXH_NOESCAPE const void* data, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize);
+
+
+/*******   Streaming   *******/
+#ifndef XXH_NO_STREAM
+/*
+ * Streaming requires state maintenance.
+ * This operation costs memory and CPU.
+ * As a consequence, streaming is slower than one-shot hashing.
+ * For better performance, prefer one-shot functions whenever applicable.
+ */
+
+/*!
+ * @brief The state struct for the XXH3 streaming API.
+ *
+ * @see XXH3_state_s for details.
+ */
+typedef struct XXH3_state_s XXH3_state_t;
+XXH_PUBLIC_API XXH_MALLOCF XXH3_state_t* XXH3_createState(void);
+XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr);
+
+/*!
+ * @brief Copies one @ref XXH3_state_t to another.
+ *
+ * @param dst_state The state to copy to.
+ * @param src_state The state to copy from.
+ * @pre
+ *   @p dst_state and @p src_state must not be `NULL` and must not overlap.
+ */
+XXH_PUBLIC_API void XXH3_copyState(XXH_NOESCAPE XXH3_state_t* dst_state, XXH_NOESCAPE const XXH3_state_t* src_state);
+
+/*!
+ * @brief Resets an @ref XXH3_state_t to begin a new hash.
+ *
+ * This function resets `statePtr` and generate a secret with default parameters. Call it before @ref XXH3_64bits_update().
+ * Digest will be equivalent to `XXH3_64bits()`.
+ *
+ * @param statePtr The state struct to reset.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success, @ref XXH_ERROR on failure.
+ *
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr);
+
+/*!
+ * @brief Resets an @ref XXH3_state_t with 64-bit seed to begin a new hash.
+ *
+ * This function resets `statePtr` and generate a secret from `seed`. Call it before @ref XXH3_64bits_update().
+ * Digest will be equivalent to `XXH3_64bits_withSeed()`.
+ *
+ * @param statePtr The state struct to reset.
+ * @param seed     The 64-bit seed to alter the state.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success, @ref XXH_ERROR on failure.
+ *
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed);
+
+/*!
+ * XXH3_64bits_reset_withSecret():
+ * `secret` is referenced, it _must outlive_ the hash streaming session.
+ * Similar to one-shot API, `secretSize` must be >= `XXH3_SECRET_SIZE_MIN`,
+ * and the quality of produced hash values depends on secret's entropy
+ * (secret's content should look like a bunch of random bytes).
+ * When in doubt about the randomness of a candidate `secret`,
+ * consider employing `XXH3_generateSecret()` instead (see below).
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize);
+
+/*!
+ * @brief Consumes a block of @p input to an @ref XXH3_state_t.
+ *
+ * Call this to incrementally consume blocks of data.
+ *
+ * @param statePtr The state struct to update.
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return @ref XXH_OK on success, @ref XXH_ERROR on failure.
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_64bits_update (XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length);
+
+/*!
+ * @brief Returns the calculated XXH3 64-bit hash value from an @ref XXH3_state_t.
+ *
+ * @note
+ *   Calling XXH3_64bits_digest() will not affect @p statePtr, so you can update,
+ *   digest, and update again.
+ *
+ * @param statePtr The state struct to calculate the hash from.
+ *
+ * @pre
+ *  @p statePtr must not be `NULL`.
+ *
+ * @return The calculated XXH3 64-bit hash value from that state.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t  XXH3_64bits_digest (XXH_NOESCAPE const XXH3_state_t* statePtr);
+#endif /* !XXH_NO_STREAM */
+
+/* note : canonical representation of XXH3 is the same as XXH64
+ * since they both produce XXH64_hash_t values */
+
+
+/*-**********************************************************************
+*  XXH3 128-bit variant
+************************************************************************/
+
+/*!
+ * @brief The return value from 128-bit hashes.
+ *
+ * Stored in little endian order, although the fields themselves are in native
+ * endianness.
+ */
+typedef struct {
+    XXH64_hash_t low64;   /*!< `value & 0xFFFFFFFFFFFFFFFF` */
+    XXH64_hash_t high64;  /*!< `value >> 64` */
+} XXH128_hash_t;
+
+/*!
+ * @brief Unseeded 128-bit variant of XXH3
+ *
+ * The 128-bit variant of XXH3 has more strength, but it has a bit of overhead
+ * for shorter inputs.
+ *
+ * This is equivalent to @ref XXH3_128bits_withSeed() with a seed of 0, however
+ * it may have slightly better performance due to constant propagation of the
+ * defaults.
+ *
+ * @see
+ *    XXH32(), XXH64(), XXH3_64bits(): equivalent for the other xxHash algorithms
+ * @see
+ *    XXH3_128bits_withSeed(), XXH3_128bits_withSecret(): other seeding variants
+ * @see
+ *    XXH3_128bits_reset(), XXH3_128bits_update(), XXH3_128bits_digest(): Streaming version.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits(XXH_NOESCAPE const void* data, size_t len);
+/*! @brief Seeded 128-bit variant of XXH3. @see XXH3_64bits_withSeed(). */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_withSeed(XXH_NOESCAPE const void* data, size_t len, XXH64_hash_t seed);
+/*! @brief Custom secret 128-bit variant of XXH3. @see XXH3_64bits_withSecret(). */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_withSecret(XXH_NOESCAPE const void* data, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize);
+
+/*******   Streaming   *******/
+#ifndef XXH_NO_STREAM
+/*
+ * Streaming requires state maintenance.
+ * This operation costs memory and CPU.
+ * As a consequence, streaming is slower than one-shot hashing.
+ * For better performance, prefer one-shot functions whenever applicable.
+ *
+ * XXH3_128bits uses the same XXH3_state_t as XXH3_64bits().
+ * Use already declared XXH3_createState() and XXH3_freeState().
+ *
+ * All reset and streaming functions have same meaning as their 64-bit counterpart.
+ */
+
+/*!
+ * @brief Resets an @ref XXH3_state_t to begin a new hash.
+ *
+ * This function resets `statePtr` and generate a secret with default parameters. Call it before @ref XXH3_128bits_update().
+ * Digest will be equivalent to `XXH3_128bits()`.
+ *
+ * @param statePtr The state struct to reset.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success, @ref XXH_ERROR on failure.
+ *
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr);
+
+/*!
+ * @brief Resets an @ref XXH3_state_t with 64-bit seed to begin a new hash.
+ *
+ * This function resets `statePtr` and generate a secret from `seed`. Call it before @ref XXH3_128bits_update().
+ * Digest will be equivalent to `XXH3_128bits_withSeed()`.
+ *
+ * @param statePtr The state struct to reset.
+ * @param seed     The 64-bit seed to alter the state.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ *
+ * @return @ref XXH_OK on success, @ref XXH_ERROR on failure.
+ *
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed);
+/*! @brief Custom secret 128-bit variant of XXH3. @see XXH_64bits_reset_withSecret(). */
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize);
+
+/*!
+ * @brief Consumes a block of @p input to an @ref XXH3_state_t.
+ *
+ * Call this to incrementally consume blocks of data.
+ *
+ * @param statePtr The state struct to update.
+ * @param input The block of data to be hashed, at least @p length bytes in size.
+ * @param length The length of @p input, in bytes.
+ *
+ * @pre
+ *   @p statePtr must not be `NULL`.
+ * @pre
+ *   The memory between @p input and @p input + @p length must be valid,
+ *   readable, contiguous memory. However, if @p length is `0`, @p input may be
+ *   `NULL`. In C++, this also must be *TriviallyCopyable*.
+ *
+ * @return @ref XXH_OK on success, @ref XXH_ERROR on failure.
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_128bits_update (XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length);
+
+/*!
+ * @brief Returns the calculated XXH3 128-bit hash value from an @ref XXH3_state_t.
+ *
+ * @note
+ *   Calling XXH3_128bits_digest() will not affect @p statePtr, so you can update,
+ *   digest, and update again.
+ *
+ * @param statePtr The state struct to calculate the hash from.
+ *
+ * @pre
+ *  @p statePtr must not be `NULL`.
+ *
+ * @return The calculated XXH3 128-bit hash value from that state.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_digest (XXH_NOESCAPE const XXH3_state_t* statePtr);
+#endif /* !XXH_NO_STREAM */
+
+/* Following helper functions make it possible to compare XXH128_hast_t values.
+ * Since XXH128_hash_t is a structure, this capability is not offered by the language.
+ * Note: For better performance, these functions can be inlined using XXH_INLINE_ALL */
+
+/*!
+ * XXH128_isEqual():
+ * Return: 1 if `h1` and `h2` are equal, 0 if they are not.
+ */
+XXH_PUBLIC_API XXH_PUREF int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2);
+
+/*!
+ * @brief Compares two @ref XXH128_hash_t
+ * This comparator is compatible with stdlib's `qsort()`/`bsearch()`.
+ *
+ * @return: >0 if *h128_1  > *h128_2
+ *          =0 if *h128_1 == *h128_2
+ *          <0 if *h128_1  < *h128_2
+ */
+XXH_PUBLIC_API XXH_PUREF int XXH128_cmp(XXH_NOESCAPE const void* h128_1, XXH_NOESCAPE const void* h128_2);
+
+
+/*******   Canonical representation   *******/
+typedef struct { unsigned char digest[sizeof(XXH128_hash_t)]; } XXH128_canonical_t;
+
+
+/*!
+ * @brief Converts an @ref XXH128_hash_t to a big endian @ref XXH128_canonical_t.
+ *
+ * @param dst The @ref XXH128_canonical_t pointer to be stored to.
+ * @param hash The @ref XXH128_hash_t to be converted.
+ *
+ * @pre
+ *   @p dst must not be `NULL`.
+ */
+XXH_PUBLIC_API void XXH128_canonicalFromHash(XXH_NOESCAPE XXH128_canonical_t* dst, XXH128_hash_t hash);
+
+/*!
+ * @brief Converts an @ref XXH128_canonical_t to a native @ref XXH128_hash_t.
+ *
+ * @param src The @ref XXH128_canonical_t to convert.
+ *
+ * @pre
+ *   @p src must not be `NULL`.
+ *
+ * @return The converted hash.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH128_hashFromCanonical(XXH_NOESCAPE const XXH128_canonical_t* src);
+
+
+#endif  /* !XXH_NO_XXH3 */
+#endif  /* XXH_NO_LONG_LONG */
+
+/*!
+ * @}
+ */
+#endif /* XXHASH_H_5627135585666179 */
+
+
+
+#if defined(XXH_STATIC_LINKING_ONLY) && !defined(XXHASH_H_STATIC_13879238742)
+#define XXHASH_H_STATIC_13879238742
+/* ****************************************************************************
+ * This section contains declarations which are not guaranteed to remain stable.
+ * They may change in future versions, becoming incompatible with a different
+ * version of the library.
+ * These declarations should only be used with static linking.
+ * Never use them in association with dynamic linking!
+ ***************************************************************************** */
+
+/*
+ * These definitions are only present to allow static allocation
+ * of XXH states, on stack or in a struct, for example.
+ * Never **ever** access their members directly.
+ */
+
+/*!
+ * @internal
+ * @brief Structure for XXH32 streaming API.
+ *
+ * @note This is only defined when @ref XXH_STATIC_LINKING_ONLY,
+ * @ref XXH_INLINE_ALL, or @ref XXH_IMPLEMENTATION is defined. Otherwise it is
+ * an opaque type. This allows fields to safely be changed.
+ *
+ * Typedef'd to @ref XXH32_state_t.
+ * Do not access the members of this struct directly.
+ * @see XXH64_state_s, XXH3_state_s
+ */
+struct XXH32_state_s {
+   XXH32_hash_t total_len_32; /*!< Total length hashed, modulo 2^32 */
+   XXH32_hash_t large_len;    /*!< Whether the hash is >= 16 (handles @ref total_len_32 overflow) */
+   XXH32_hash_t v[4];         /*!< Accumulator lanes */
+   XXH32_hash_t mem32[4];     /*!< Internal buffer for partial reads. Treated as unsigned char[16]. */
+   XXH32_hash_t memsize;      /*!< Amount of data in @ref mem32 */
+   XXH32_hash_t reserved;     /*!< Reserved field. Do not read nor write to it. */
+};   /* typedef'd to XXH32_state_t */
+
+
+#ifndef XXH_NO_LONG_LONG  /* defined when there is no 64-bit support */
+
+/*!
+ * @internal
+ * @brief Structure for XXH64 streaming API.
+ *
+ * @note This is only defined when @ref XXH_STATIC_LINKING_ONLY,
+ * @ref XXH_INLINE_ALL, or @ref XXH_IMPLEMENTATION is defined. Otherwise it is
+ * an opaque type. This allows fields to safely be changed.
+ *
+ * Typedef'd to @ref XXH64_state_t.
+ * Do not access the members of this struct directly.
+ * @see XXH32_state_s, XXH3_state_s
+ */
+struct XXH64_state_s {
+   XXH64_hash_t total_len;    /*!< Total length hashed. This is always 64-bit. */
+   XXH64_hash_t v[4];         /*!< Accumulator lanes */
+   XXH64_hash_t mem64[4];     /*!< Internal buffer for partial reads. Treated as unsigned char[32]. */
+   XXH32_hash_t memsize;      /*!< Amount of data in @ref mem64 */
+   XXH32_hash_t reserved32;   /*!< Reserved field, needed for padding anyways*/
+   XXH64_hash_t reserved64;   /*!< Reserved field. Do not read or write to it. */
+};   /* typedef'd to XXH64_state_t */
+
+#ifndef XXH_NO_XXH3
+
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) /* >= C11 */
+#  include <stdalign.h>
+#  define XXH_ALIGN(n)      alignas(n)
+#elif defined(__cplusplus) && (__cplusplus >= 201103L) /* >= C++11 */
+/* In C++ alignas() is a keyword */
+#  define XXH_ALIGN(n)      alignas(n)
+#elif defined(__GNUC__)
+#  define XXH_ALIGN(n)      __attribute__ ((aligned(n)))
+#elif defined(_MSC_VER)
+#  define XXH_ALIGN(n)      __declspec(align(n))
+#else
+#  define XXH_ALIGN(n)   /* disabled */
+#endif
+
+/* Old GCC versions only accept the attribute after the type in structures. */
+#if !(defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L))   /* C11+ */ \
+    && ! (defined(__cplusplus) && (__cplusplus >= 201103L)) /* >= C++11 */ \
+    && defined(__GNUC__)
+#   define XXH_ALIGN_MEMBER(align, type) type XXH_ALIGN(align)
+#else
+#   define XXH_ALIGN_MEMBER(align, type) XXH_ALIGN(align) type
+#endif
+
+/*!
+ * @brief The size of the internal XXH3 buffer.
+ *
+ * This is the optimal update size for incremental hashing.
+ *
+ * @see XXH3_64b_update(), XXH3_128b_update().
+ */
+#define XXH3_INTERNALBUFFER_SIZE 256
+
+/*!
+ * @internal
+ * @brief Default size of the secret buffer (and @ref XXH3_kSecret).
+ *
+ * This is the size used in @ref XXH3_kSecret and the seeded functions.
+ *
+ * Not to be confused with @ref XXH3_SECRET_SIZE_MIN.
+ */
+#define XXH3_SECRET_DEFAULT_SIZE 192
+
+/*!
+ * @internal
+ * @brief Structure for XXH3 streaming API.
+ *
+ * @note This is only defined when @ref XXH_STATIC_LINKING_ONLY,
+ * @ref XXH_INLINE_ALL, or @ref XXH_IMPLEMENTATION is defined.
+ * Otherwise it is an opaque type.
+ * Never use this definition in combination with dynamic library.
+ * This allows fields to safely be changed in the future.
+ *
+ * @note ** This structure has a strict alignment requirement of 64 bytes!! **
+ * Do not allocate this with `malloc()` or `new`,
+ * it will not be sufficiently aligned.
+ * Use @ref XXH3_createState() and @ref XXH3_freeState(), or stack allocation.
+ *
+ * Typedef'd to @ref XXH3_state_t.
+ * Do never access the members of this struct directly.
+ *
+ * @see XXH3_INITSTATE() for stack initialization.
+ * @see XXH3_createState(), XXH3_freeState().
+ * @see XXH32_state_s, XXH64_state_s
+ */
+struct XXH3_state_s {
+   XXH_ALIGN_MEMBER(64, XXH64_hash_t acc[8]);
+       /*!< The 8 accumulators. See @ref XXH32_state_s::v and @ref XXH64_state_s::v */
+   XXH_ALIGN_MEMBER(64, unsigned char customSecret[XXH3_SECRET_DEFAULT_SIZE]);
+       /*!< Used to store a custom secret generated from a seed. */
+   XXH_ALIGN_MEMBER(64, unsigned char buffer[XXH3_INTERNALBUFFER_SIZE]);
+       /*!< The internal buffer. @see XXH32_state_s::mem32 */
+   XXH32_hash_t bufferedSize;
+       /*!< The amount of memory in @ref buffer, @see XXH32_state_s::memsize */
+   XXH32_hash_t useSeed;
+       /*!< Reserved field. Needed for padding on 64-bit. */
+   size_t nbStripesSoFar;
+       /*!< Number or stripes processed. */
+   XXH64_hash_t totalLen;
+       /*!< Total length hashed. 64-bit even on 32-bit targets. */
+   size_t nbStripesPerBlock;
+       /*!< Number of stripes per block. */
+   size_t secretLimit;
+       /*!< Size of @ref customSecret or @ref extSecret */
+   XXH64_hash_t seed;
+       /*!< Seed for _withSeed variants. Must be zero otherwise, @see XXH3_INITSTATE() */
+   XXH64_hash_t reserved64;
+       /*!< Reserved field. */
+   const unsigned char* extSecret;
+       /*!< Reference to an external secret for the _withSecret variants, NULL
+        *   for other variants. */
+   /* note: there may be some padding at the end due to alignment on 64 bytes */
+}; /* typedef'd to XXH3_state_t */
+
+#undef XXH_ALIGN_MEMBER
+
+/*!
+ * @brief Initializes a stack-allocated `XXH3_state_s`.
+ *
+ * When the @ref XXH3_state_t structure is merely emplaced on stack,
+ * it should be initialized with XXH3_INITSTATE() or a memset()
+ * in case its first reset uses XXH3_NNbits_reset_withSeed().
+ * This init can be omitted if the first reset uses default or _withSecret mode.
+ * This operation isn't necessary when the state is created with XXH3_createState().
+ * Note that this doesn't prepare the state for a streaming operation,
+ * it's still necessary to use XXH3_NNbits_reset*() afterwards.
+ */
+#define XXH3_INITSTATE(XXH3_state_ptr)                       \
+    do {                                                     \
+        XXH3_state_t* tmp_xxh3_state_ptr = (XXH3_state_ptr); \
+        tmp_xxh3_state_ptr->seed = 0;                        \
+        tmp_xxh3_state_ptr->extSecret = NULL;                \
+    } while(0)
+
+
+/*!
+ * simple alias to pre-selected XXH3_128bits variant
+ */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH128(XXH_NOESCAPE const void* data, size_t len, XXH64_hash_t seed);
+
+
+/* ===   Experimental API   === */
+/* Symbols defined below must be considered tied to a specific library version. */
+
+/*!
+ * XXH3_generateSecret():
+ *
+ * Derive a high-entropy secret from any user-defined content, named customSeed.
+ * The generated secret can be used in combination with `*_withSecret()` functions.
+ * The `_withSecret()` variants are useful to provide a higher level of protection
+ * than 64-bit seed, as it becomes much more difficult for an external actor to
+ * guess how to impact the calculation logic.
+ *
+ * The function accepts as input a custom seed of any length and any content,
+ * and derives from it a high-entropy secret of length @p secretSize into an
+ * already allocated buffer @p secretBuffer.
+ *
+ * The generated secret can then be used with any `*_withSecret()` variant.
+ * The functions @ref XXH3_128bits_withSecret(), @ref XXH3_64bits_withSecret(),
+ * @ref XXH3_128bits_reset_withSecret() and @ref XXH3_64bits_reset_withSecret()
+ * are part of this list. They all accept a `secret` parameter
+ * which must be large enough for implementation reasons (>= @ref XXH3_SECRET_SIZE_MIN)
+ * _and_ feature very high entropy (consist of random-looking bytes).
+ * These conditions can be a high bar to meet, so @ref XXH3_generateSecret() can
+ * be employed to ensure proper quality.
+ *
+ * @p customSeed can be anything. It can have any size, even small ones,
+ * and its content can be anything, even "poor entropy" sources such as a bunch
+ * of zeroes. The resulting `secret` will nonetheless provide all required qualities.
+ *
+ * @pre
+ *   - @p secretSize must be >= @ref XXH3_SECRET_SIZE_MIN
+ *   - When @p customSeedSize > 0, supplying NULL as customSeed is undefined behavior.
+ *
+ * Example code:
+ * @code{.c}
+ *    #include <stdio.h>
+ *    #include <stdlib.h>
+ *    #include <string.h>
+ *    #define XXH_STATIC_LINKING_ONLY // expose unstable API
+ *    #include "xxhash.h"
+ *    // Hashes argv[2] using the entropy from argv[1].
+ *    int main(int argc, char* argv[])
+ *    {
+ *        char secret[XXH3_SECRET_SIZE_MIN];
+ *        if (argv != 3) { return 1; }
+ *        XXH3_generateSecret(secret, sizeof(secret), argv[1], strlen(argv[1]));
+ *        XXH64_hash_t h = XXH3_64bits_withSecret(
+ *             argv[2], strlen(argv[2]),
+ *             secret, sizeof(secret)
+ *        );
+ *        printf("%016llx\n", (unsigned long long) h);
+ *    }
+ * @endcode
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_generateSecret(XXH_NOESCAPE void* secretBuffer, size_t secretSize, XXH_NOESCAPE const void* customSeed, size_t customSeedSize);
+
+/*!
+ * @brief Generate the same secret as the _withSeed() variants.
+ *
+ * The generated secret can be used in combination with
+ *`*_withSecret()` and `_withSecretandSeed()` variants.
+ *
+ * Example C++ `std::string` hash class:
+ * @code{.cpp}
+ *    #include <string>
+ *    #define XXH_STATIC_LINKING_ONLY // expose unstable API
+ *    #include "xxhash.h"
+ *    // Slow, seeds each time
+ *    class HashSlow {
+ *        XXH64_hash_t seed;
+ *    public:
+ *        HashSlow(XXH64_hash_t s) : seed{s} {}
+ *        size_t operator()(const std::string& x) const {
+ *            return size_t{XXH3_64bits_withSeed(x.c_str(), x.length(), seed)};
+ *        }
+ *    };
+ *    // Fast, caches the seeded secret for future uses.
+ *    class HashFast {
+ *        unsigned char secret[XXH3_SECRET_SIZE_MIN];
+ *    public:
+ *        HashFast(XXH64_hash_t s) {
+ *            XXH3_generateSecret_fromSeed(secret, seed);
+ *        }
+ *        size_t operator()(const std::string& x) const {
+ *            return size_t{
+ *                XXH3_64bits_withSecret(x.c_str(), x.length(), secret, sizeof(secret))
+ *            };
+ *        }
+ *    };
+ * @endcode
+ * @param secretBuffer A writable buffer of @ref XXH3_SECRET_SIZE_MIN bytes
+ * @param seed The seed to seed the state.
+ */
+XXH_PUBLIC_API void XXH3_generateSecret_fromSeed(XXH_NOESCAPE void* secretBuffer, XXH64_hash_t seed);
+
+/*!
+ * These variants generate hash values using either
+ * @p seed for "short" keys (< XXH3_MIDSIZE_MAX = 240 bytes)
+ * or @p secret for "large" keys (>= XXH3_MIDSIZE_MAX).
+ *
+ * This generally benefits speed, compared to `_withSeed()` or `_withSecret()`.
+ * `_withSeed()` has to generate the secret on the fly for "large" keys.
+ * It's fast, but can be perceptible for "not so large" keys (< 1 KB).
+ * `_withSecret()` has to generate the masks on the fly for "small" keys,
+ * which requires more instructions than _withSeed() variants.
+ * Therefore, _withSecretandSeed variant combines the best of both worlds.
+ *
+ * When @p secret has been generated by XXH3_generateSecret_fromSeed(),
+ * this variant produces *exactly* the same results as `_withSeed()` variant,
+ * hence offering only a pure speed benefit on "large" input,
+ * by skipping the need to regenerate the secret for every large input.
+ *
+ * Another usage scenario is to hash the secret to a 64-bit hash value,
+ * for example with XXH3_64bits(), which then becomes the seed,
+ * and then employ both the seed and the secret in _withSecretandSeed().
+ * On top of speed, an added benefit is that each bit in the secret
+ * has a 50% chance to swap each bit in the output, via its impact to the seed.
+ *
+ * This is not guaranteed when using the secret directly in "small data" scenarios,
+ * because only portions of the secret are employed for small data.
+ */
+XXH_PUBLIC_API XXH_PUREF XXH64_hash_t
+XXH3_64bits_withSecretandSeed(XXH_NOESCAPE const void* data, size_t len,
+                              XXH_NOESCAPE const void* secret, size_t secretSize,
+                              XXH64_hash_t seed);
+/*! @copydoc XXH3_64bits_withSecretandSeed() */
+XXH_PUBLIC_API XXH_PUREF XXH128_hash_t
+XXH3_128bits_withSecretandSeed(XXH_NOESCAPE const void* input, size_t length,
+                               XXH_NOESCAPE const void* secret, size_t secretSize,
+                               XXH64_hash_t seed64);
+#ifndef XXH_NO_STREAM
+/*! @copydoc XXH3_64bits_withSecretandSeed() */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr,
+                                    XXH_NOESCAPE const void* secret, size_t secretSize,
+                                    XXH64_hash_t seed64);
+/*! @copydoc XXH3_64bits_withSecretandSeed() */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr,
+                                     XXH_NOESCAPE const void* secret, size_t secretSize,
+                                     XXH64_hash_t seed64);
+#endif /* !XXH_NO_STREAM */
+
+#endif  /* !XXH_NO_XXH3 */
+#endif  /* XXH_NO_LONG_LONG */
+#if defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API)
+#  define XXH_IMPLEMENTATION
+#endif
+
+#endif  /* defined(XXH_STATIC_LINKING_ONLY) && !defined(XXHASH_H_STATIC_13879238742) */
+
+
+/* ======================================================================== */
+/* ======================================================================== */
+/* ======================================================================== */
+
+
+/*-**********************************************************************
+ * xxHash implementation
+ *-**********************************************************************
+ * xxHash's implementation used to be hosted inside xxhash.c.
+ *
+ * However, inlining requires implementation to be visible to the compiler,
+ * hence be included alongside the header.
+ * Previously, implementation was hosted inside xxhash.c,
+ * which was then #included when inlining was activated.
+ * This construction created issues with a few build and install systems,
+ * as it required xxhash.c to be stored in /include directory.
+ *
+ * xxHash implementation is now directly integrated within xxhash.h.
+ * As a consequence, xxhash.c is no longer needed in /include.
+ *
+ * xxhash.c is still available and is still useful.
+ * In a "normal" setup, when xxhash is not inlined,
+ * xxhash.h only exposes the prototypes and public symbols,
+ * while xxhash.c can be built into an object file xxhash.o
+ * which can then be linked into the final binary.
+ ************************************************************************/
+
+#if ( defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API) \
+   || defined(XXH_IMPLEMENTATION) ) && !defined(XXH_IMPLEM_13a8737387)
+#  define XXH_IMPLEM_13a8737387
+
+/* *************************************
+*  Tuning parameters
+***************************************/
+
+/*!
+ * @defgroup tuning Tuning parameters
+ * @{
+ *
+ * Various macros to control xxHash's behavior.
+ */
+#ifdef XXH_DOXYGEN
+/*!
+ * @brief Define this to disable 64-bit code.
+ *
+ * Useful if only using the @ref XXH32_family and you have a strict C90 compiler.
+ */
+#  define XXH_NO_LONG_LONG
+#  undef XXH_NO_LONG_LONG /* don't actually */
+/*!
+ * @brief Controls how unaligned memory is accessed.
+ *
+ * By default, access to unaligned memory is controlled by `memcpy()`, which is
+ * safe and portable.
+ *
+ * Unfortunately, on some target/compiler combinations, the generated assembly
+ * is sub-optimal.
+ *
+ * The below switch allow selection of a different access method
+ * in the search for improved performance.
+ *
+ * @par Possible options:
+ *
+ *  - `XXH_FORCE_MEMORY_ACCESS=0` (default): `memcpy`
+ *   @par
+ *     Use `memcpy()`. Safe and portable. Note that most modern compilers will
+ *     eliminate the function call and treat it as an unaligned access.
+ *
+ *  - `XXH_FORCE_MEMORY_ACCESS=1`: `__attribute__((aligned(1)))`
+ *   @par
+ *     Depends on compiler extensions and is therefore not portable.
+ *     This method is safe _if_ your compiler supports it,
+ *     and *generally* as fast or faster than `memcpy`.
+ *
+ *  - `XXH_FORCE_MEMORY_ACCESS=2`: Direct cast
+ *  @par
+ *     Casts directly and dereferences. This method doesn't depend on the
+ *     compiler, but it violates the C standard as it directly dereferences an
+ *     unaligned pointer. It can generate buggy code on targets which do not
+ *     support unaligned memory accesses, but in some circumstances, it's the
+ *     only known way to get the most performance.
+ *
+ *  - `XXH_FORCE_MEMORY_ACCESS=3`: Byteshift
+ *  @par
+ *     Also portable. This can generate the best code on old compilers which don't
+ *     inline small `memcpy()` calls, and it might also be faster on big-endian
+ *     systems which lack a native byteswap instruction. However, some compilers
+ *     will emit literal byteshifts even if the target supports unaligned access.
+ *
+ *
+ * @warning
+ *   Methods 1 and 2 rely on implementation-defined behavior. Use these with
+ *   care, as what works on one compiler/platform/optimization level may cause
+ *   another to read garbage data or even crash.
+ *
+ * See https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html for details.
+ *
+ * Prefer these methods in priority order (0 > 3 > 1 > 2)
+ */
+#  define XXH_FORCE_MEMORY_ACCESS 0
+
+/*!
+ * @def XXH_SIZE_OPT
+ * @brief Controls how much xxHash optimizes for size.
+ *
+ * xxHash, when compiled, tends to result in a rather large binary size. This
+ * is mostly due to heavy usage to forced inlining and constant folding of the
+ * @ref XXH3_family to increase performance.
+ *
+ * However, some developers prefer size over speed. This option can
+ * significantly reduce the size of the generated code. When using the `-Os`
+ * or `-Oz` options on GCC or Clang, this is defined to 1 by default,
+ * otherwise it is defined to 0.
+ *
+ * Most of these size optimizations can be controlled manually.
+ *
+ * This is a number from 0-2.
+ *  - `XXH_SIZE_OPT` == 0: Default. xxHash makes no size optimizations. Speed
+ *    comes first.
+ *  - `XXH_SIZE_OPT` == 1: Default for `-Os` and `-Oz`. xxHash is more
+ *    conservative and disables hacks that increase code size. It implies the
+ *    options @ref XXH_NO_INLINE_HINTS == 1, @ref XXH_FORCE_ALIGN_CHECK == 0,
+ *    and @ref XXH3_NEON_LANES == 8 if they are not already defined.
+ *  - `XXH_SIZE_OPT` == 2: xxHash tries to make itself as small as possible.
+ *    Performance may cry. For example, the single shot functions just use the
+ *    streaming API.
+ */
+#  define XXH_SIZE_OPT 0
+
+/*!
+ * @def XXH_FORCE_ALIGN_CHECK
+ * @brief If defined to non-zero, adds a special path for aligned inputs (XXH32()
+ * and XXH64() only).
+ *
+ * This is an important performance trick for architectures without decent
+ * unaligned memory access performance.
+ *
+ * It checks for input alignment, and when conditions are met, uses a "fast
+ * path" employing direct 32-bit/64-bit reads, resulting in _dramatically
+ * faster_ read speed.
+ *
+ * The check costs one initial branch per hash, which is generally negligible,
+ * but not zero.
+ *
+ * Moreover, it's not useful to generate an additional code path if memory
+ * access uses the same instruction for both aligned and unaligned
+ * addresses (e.g. x86 and aarch64).
+ *
+ * In these cases, the alignment check can be removed by setting this macro to 0.
+ * Then the code will always use unaligned memory access.
+ * Align check is automatically disabled on x86, x64, ARM64, and some ARM chips
+ * which are platforms known to offer good unaligned memory accesses performance.
+ *
+ * It is also disabled by default when @ref XXH_SIZE_OPT >= 1.
+ *
+ * This option does not affect XXH3 (only XXH32 and XXH64).
+ */
+#  define XXH_FORCE_ALIGN_CHECK 0
+
+/*!
+ * @def XXH_NO_INLINE_HINTS
+ * @brief When non-zero, sets all functions to `static`.
+ *
+ * By default, xxHash tries to force the compiler to inline almost all internal
+ * functions.
+ *
+ * This can usually improve performance due to reduced jumping and improved
+ * constant folding, but significantly increases the size of the binary which
+ * might not be favorable.
+ *
+ * Additionally, sometimes the forced inlining can be detrimental to performance,
+ * depending on the architecture.
+ *
+ * XXH_NO_INLINE_HINTS marks all internal functions as static, giving the
+ * compiler full control on whether to inline or not.
+ *
+ * When not optimizing (-O0), using `-fno-inline` with GCC or Clang, or if
+ * @ref XXH_SIZE_OPT >= 1, this will automatically be defined.
+ */
+#  define XXH_NO_INLINE_HINTS 0
+
+/*!
+ * @def XXH3_INLINE_SECRET
+ * @brief Determines whether to inline the XXH3 withSecret code.
+ *
+ * When the secret size is known, the compiler can improve the performance
+ * of XXH3_64bits_withSecret() and XXH3_128bits_withSecret().
+ *
+ * However, if the secret size is not known, it doesn't have any benefit. This
+ * happens when xxHash is compiled into a global symbol. Therefore, if
+ * @ref XXH_INLINE_ALL is *not* defined, this will be defined to 0.
+ *
+ * Additionally, this defaults to 0 on GCC 12+, which has an issue with function pointers
+ * that are *sometimes* force inline on -Og, and it is impossible to automatically
+ * detect this optimization level.
+ */
+#  define XXH3_INLINE_SECRET 0
+
+/*!
+ * @def XXH32_ENDJMP
+ * @brief Whether to use a jump for `XXH32_finalize`.
+ *
+ * For performance, `XXH32_finalize` uses multiple branches in the finalizer.
+ * This is generally preferable for performance,
+ * but depending on exact architecture, a jmp may be preferable.
+ *
+ * This setting is only possibly making a difference for very small inputs.
+ */
+#  define XXH32_ENDJMP 0
+
+/*!
+ * @internal
+ * @brief Redefines old internal names.
+ *
+ * For compatibility with code that uses xxHash's internals before the names
+ * were changed to improve namespacing. There is no other reason to use this.
+ */
+#  define XXH_OLD_NAMES
+#  undef XXH_OLD_NAMES /* don't actually use, it is ugly. */
+
+/*!
+ * @def XXH_NO_STREAM
+ * @brief Disables the streaming API.
+ *
+ * When xxHash is not inlined and the streaming functions are not used, disabling
+ * the streaming functions can improve code size significantly, especially with
+ * the @ref XXH3_family which tends to make constant folded copies of itself.
+ */
+#  define XXH_NO_STREAM
+#  undef XXH_NO_STREAM /* don't actually */
+#endif /* XXH_DOXYGEN */
+/*!
+ * @}
+ */
+
+#ifndef XXH_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
+   /* prefer __packed__ structures (method 1) for GCC
+    * < ARMv7 with unaligned access (e.g. Raspbian armhf) still uses byte shifting, so we use memcpy
+    * which for some reason does unaligned loads. */
+#  if defined(__GNUC__) && !(defined(__ARM_ARCH) && __ARM_ARCH < 7 && defined(__ARM_FEATURE_UNALIGNED))
+#    define XXH_FORCE_MEMORY_ACCESS 1
+#  endif
+#endif
+
+#ifndef XXH_SIZE_OPT
+   /* default to 1 for -Os or -Oz */
+#  if (defined(__GNUC__) || defined(__clang__)) && defined(__OPTIMIZE_SIZE__)
+#    define XXH_SIZE_OPT 1
+#  else
+#    define XXH_SIZE_OPT 0
+#  endif
+#endif
+
+#ifndef XXH_FORCE_ALIGN_CHECK  /* can be defined externally */
+   /* don't check on sizeopt, x86, aarch64, or arm when unaligned access is available */
+#  if XXH_SIZE_OPT >= 1 || \
+      defined(__i386)  || defined(__x86_64__) || defined(__aarch64__) || defined(__ARM_FEATURE_UNALIGNED) \
+   || defined(_M_IX86) || defined(_M_X64)     || defined(_M_ARM64)    || defined(_M_ARM) /* visual */
+#    define XXH_FORCE_ALIGN_CHECK 0
+#  else
+#    define XXH_FORCE_ALIGN_CHECK 1
+#  endif
+#endif
+
+#ifndef XXH_NO_INLINE_HINTS
+#  if XXH_SIZE_OPT >= 1 || defined(__NO_INLINE__)  /* -O0, -fno-inline */
+#    define XXH_NO_INLINE_HINTS 1
+#  else
+#    define XXH_NO_INLINE_HINTS 0
+#  endif
+#endif
+
+#ifndef XXH3_INLINE_SECRET
+#  if (defined(__GNUC__) && !defined(__clang__) && __GNUC__ >= 12) \
+     || !defined(XXH_INLINE_ALL)
+#    define XXH3_INLINE_SECRET 0
+#  else
+#    define XXH3_INLINE_SECRET 1
+#  endif
+#endif
+
+#ifndef XXH32_ENDJMP
+/* generally preferable for performance */
+#  define XXH32_ENDJMP 0
+#endif
+
+/*!
+ * @defgroup impl Implementation
+ * @{
+ */
+
+
+/* *************************************
+*  Includes & Memory related functions
+***************************************/
+#if defined(XXH_NO_STREAM)
+/* nothing */
+#elif defined(XXH_NO_STDLIB)
+
+/* When requesting to disable any mention of stdlib,
+ * the library loses the ability to invoked malloc / free.
+ * In practice, it means that functions like `XXH*_createState()`
+ * will always fail, and return NULL.
+ * This flag is useful in situations where
+ * xxhash.h is integrated into some kernel, embedded or limited environment
+ * without access to dynamic allocation.
+ */
+
+static XXH_CONSTF void* XXH_malloc(size_t s) { (void)s; return NULL; }
+static void XXH_free(void* p) { (void)p; }
+
+#else
+
+/*
+ * Modify the local functions below should you wish to use
+ * different memory routines for malloc() and free()
+ */
+#include <stdlib.h>
+
+/*!
+ * @internal
+ * @brief Modify this function to use a different routine than malloc().
+ */
+static XXH_MALLOCF void* XXH_malloc(size_t s) { return malloc(s); }
+
+/*!
+ * @internal
+ * @brief Modify this function to use a different routine than free().
+ */
+static void XXH_free(void* p) { free(p); }
+
+#endif  /* XXH_NO_STDLIB */
+
+#include <string.h>
+
+/*!
+ * @internal
+ * @brief Modify this function to use a different routine than memcpy().
+ */
+static void* XXH_memcpy(void* dest, const void* src, size_t size)
+{
+    return memcpy(dest,src,size);
+}
+
+#include <limits.h>   /* ULLONG_MAX */
+
+
+/* *************************************
+*  Compiler Specific Options
+***************************************/
+#ifdef _MSC_VER /* Visual Studio warning fix */
+#  pragma warning(disable : 4127) /* disable: C4127: conditional expression is constant */
+#endif
+
+#if XXH_NO_INLINE_HINTS  /* disable inlining hints */
+#  if defined(__GNUC__) || defined(__clang__)
+#    define XXH_FORCE_INLINE static __attribute__((unused))
+#  else
+#    define XXH_FORCE_INLINE static
+#  endif
+#  define XXH_NO_INLINE static
+/* enable inlining hints */
+#elif defined(__GNUC__) || defined(__clang__)
+#  define XXH_FORCE_INLINE static __inline__ __attribute__((always_inline, unused))
+#  define XXH_NO_INLINE static __attribute__((noinline))
+#elif defined(_MSC_VER)  /* Visual Studio */
+#  define XXH_FORCE_INLINE static __forceinline
+#  define XXH_NO_INLINE static __declspec(noinline)
+#elif defined (__cplusplus) \
+  || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L))   /* C99 */
+#  define XXH_FORCE_INLINE static inline
+#  define XXH_NO_INLINE static
+#else
+#  define XXH_FORCE_INLINE static
+#  define XXH_NO_INLINE static
+#endif
+
+#if XXH3_INLINE_SECRET
+#  define XXH3_WITH_SECRET_INLINE XXH_FORCE_INLINE
+#else
+#  define XXH3_WITH_SECRET_INLINE XXH_NO_INLINE
+#endif
+
+
+/* *************************************
+*  Debug
+***************************************/
+/*!
+ * @ingroup tuning
+ * @def XXH_DEBUGLEVEL
+ * @brief Sets the debugging level.
+ *
+ * XXH_DEBUGLEVEL is expected to be defined externally, typically via the
+ * compiler's command line options. The value must be a number.
+ */
+#ifndef XXH_DEBUGLEVEL
+#  ifdef DEBUGLEVEL /* backwards compat */
+#    define XXH_DEBUGLEVEL DEBUGLEVEL
+#  else
+#    define XXH_DEBUGLEVEL 0
+#  endif
+#endif
+
+#if (XXH_DEBUGLEVEL>=1)
+#  include <assert.h>   /* note: can still be disabled with NDEBUG */
+#  define XXH_ASSERT(c)   assert(c)
+#else
+#  if defined(__INTEL_COMPILER)
+#    define XXH_ASSERT(c)   XXH_ASSUME((unsigned char) (c))
+#  else
+#    define XXH_ASSERT(c)   XXH_ASSUME(c)
+#  endif
+#endif
+
+/* note: use after variable declarations */
+#ifndef XXH_STATIC_ASSERT
+#  if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)    /* C11 */
+#    define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { _Static_assert((c),m); } while(0)
+#  elif defined(__cplusplus) && (__cplusplus >= 201103L)            /* C++11 */
+#    define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { static_assert((c),m); } while(0)
+#  else
+#    define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { struct xxh_sa { char x[(c) ? 1 : -1]; }; } while(0)
+#  endif
+#  define XXH_STATIC_ASSERT(c) XXH_STATIC_ASSERT_WITH_MESSAGE((c),#c)
+#endif
+
+/*!
+ * @internal
+ * @def XXH_COMPILER_GUARD(var)
+ * @brief Used to prevent unwanted optimizations for @p var.
+ *
+ * It uses an empty GCC inline assembly statement with a register constraint
+ * which forces @p var into a general purpose register (eg eax, ebx, ecx
+ * on x86) and marks it as modified.
+ *
+ * This is used in a few places to avoid unwanted autovectorization (e.g.
+ * XXH32_round()). All vectorization we want is explicit via intrinsics,
+ * and _usually_ isn't wanted elsewhere.
+ *
+ * We also use it to prevent unwanted constant folding for AArch64 in
+ * XXH3_initCustomSecret_scalar().
+ */
+#if defined(__GNUC__) || defined(__clang__)
+#  define XXH_COMPILER_GUARD(var) __asm__("" : "+r" (var))
+#else
+#  define XXH_COMPILER_GUARD(var) ((void)0)
+#endif
+
+/* Specifically for NEON vectors which use the "w" constraint, on
+ * Clang. */
+#if defined(__clang__) && defined(__ARM_ARCH) && !defined(__wasm__)
+#  define XXH_COMPILER_GUARD_CLANG_NEON(var) __asm__("" : "+w" (var))
+#else
+#  define XXH_COMPILER_GUARD_CLANG_NEON(var) ((void)0)
+#endif
+
+/* *************************************
+*  Basic Types
+***************************************/
+#if !defined (__VMS) \
+ && (defined (__cplusplus) \
+ || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+# include <stdint.h>
+  typedef uint8_t xxh_u8;
+#else
+  typedef unsigned char xxh_u8;
+#endif
+typedef XXH32_hash_t xxh_u32;
+
+#ifdef XXH_OLD_NAMES
+#  warning "XXH_OLD_NAMES is planned to be removed starting v0.9. If the program depends on it, consider moving away from it by employing newer type names directly"
+#  define BYTE xxh_u8
+#  define U8   xxh_u8
+#  define U32  xxh_u32
+#endif
+
+/* ***   Memory access   *** */
+
+/*!
+ * @internal
+ * @fn xxh_u32 XXH_read32(const void* ptr)
+ * @brief Reads an unaligned 32-bit integer from @p ptr in native endianness.
+ *
+ * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
+ *
+ * @param ptr The pointer to read from.
+ * @return The 32-bit native endian integer from the bytes at @p ptr.
+ */
+
+/*!
+ * @internal
+ * @fn xxh_u32 XXH_readLE32(const void* ptr)
+ * @brief Reads an unaligned 32-bit little endian integer from @p ptr.
+ *
+ * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
+ *
+ * @param ptr The pointer to read from.
+ * @return The 32-bit little endian integer from the bytes at @p ptr.
+ */
+
+/*!
+ * @internal
+ * @fn xxh_u32 XXH_readBE32(const void* ptr)
+ * @brief Reads an unaligned 32-bit big endian integer from @p ptr.
+ *
+ * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
+ *
+ * @param ptr The pointer to read from.
+ * @return The 32-bit big endian integer from the bytes at @p ptr.
+ */
+
+/*!
+ * @internal
+ * @fn xxh_u32 XXH_readLE32_align(const void* ptr, XXH_alignment align)
+ * @brief Like @ref XXH_readLE32(), but has an option for aligned reads.
+ *
+ * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
+ * Note that when @ref XXH_FORCE_ALIGN_CHECK == 0, the @p align parameter is
+ * always @ref XXH_alignment::XXH_unaligned.
+ *
+ * @param ptr The pointer to read from.
+ * @param align Whether @p ptr is aligned.
+ * @pre
+ *   If @p align == @ref XXH_alignment::XXH_aligned, @p ptr must be 4 byte
+ *   aligned.
+ * @return The 32-bit little endian integer from the bytes at @p ptr.
+ */
+
+#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
+/*
+ * Manual byteshift. Best for old compilers which don't inline memcpy.
+ * We actually directly use XXH_readLE32 and XXH_readBE32.
+ */
+#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==2))
+
+/*
+ * Force direct memory access. Only works on CPU which support unaligned memory
+ * access in hardware.
+ */
+static xxh_u32 XXH_read32(const void* memPtr) { return *(const xxh_u32*) memPtr; }
+
+#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1))
+
+/*
+ * __attribute__((aligned(1))) is supported by gcc and clang. Originally the
+ * documentation claimed that it only increased the alignment, but actually it
+ * can decrease it on gcc, clang, and icc:
+ * https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69502,
+ * https://gcc.godbolt.org/z/xYez1j67Y.
+ */
+#ifdef XXH_OLD_NAMES
+typedef union { xxh_u32 u32; } __attribute__((packed)) unalign;
+#endif
+static xxh_u32 XXH_read32(const void* ptr)
+{
+    typedef __attribute__((aligned(1))) xxh_u32 xxh_unalign32;
+    return *((const xxh_unalign32*)ptr);
+}
+
+#else
+
+/*
+ * Portable and safe solution. Generally efficient.
+ * see: https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html
+ */
+static xxh_u32 XXH_read32(const void* memPtr)
+{
+    xxh_u32 val;
+    XXH_memcpy(&val, memPtr, sizeof(val));
+    return val;
+}
+
+#endif   /* XXH_FORCE_DIRECT_MEMORY_ACCESS */
+
+
+/* ***   Endianness   *** */
+
+/*!
+ * @ingroup tuning
+ * @def XXH_CPU_LITTLE_ENDIAN
+ * @brief Whether the target is little endian.
+ *
+ * Defined to 1 if the target is little endian, or 0 if it is big endian.
+ * It can be defined externally, for example on the compiler command line.
+ *
+ * If it is not defined,
+ * a runtime check (which is usually constant folded) is used instead.
+ *
+ * @note
+ *   This is not necessarily defined to an integer constant.
+ *
+ * @see XXH_isLittleEndian() for the runtime check.
+ */
+#ifndef XXH_CPU_LITTLE_ENDIAN
+/*
+ * Try to detect endianness automatically, to avoid the nonstandard behavior
+ * in `XXH_isLittleEndian()`
+ */
+#  if defined(_WIN32) /* Windows is always little endian */ \
+     || defined(__LITTLE_ENDIAN__) \
+     || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
+#    define XXH_CPU_LITTLE_ENDIAN 1
+#  elif defined(__BIG_ENDIAN__) \
+     || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
+#    define XXH_CPU_LITTLE_ENDIAN 0
+#  else
+/*!
+ * @internal
+ * @brief Runtime check for @ref XXH_CPU_LITTLE_ENDIAN.
+ *
+ * Most compilers will constant fold this.
+ */
+static int XXH_isLittleEndian(void)
+{
+    /*
+     * Portable and well-defined behavior.
+     * Don't use static: it is detrimental to performance.
+     */
+    const union { xxh_u32 u; xxh_u8 c[4]; } one = { 1 };
+    return one.c[0];
+}
+#   define XXH_CPU_LITTLE_ENDIAN   XXH_isLittleEndian()
+#  endif
+#endif
+
+
+
+
+/* ****************************************
+*  Compiler-specific Functions and Macros
+******************************************/
+#define XXH_GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__)
+
+#ifdef __has_builtin
+#  define XXH_HAS_BUILTIN(x) __has_builtin(x)
+#else
+#  define XXH_HAS_BUILTIN(x) 0
+#endif
+
+
+
+/*
+ * C23 and future versions have standard "unreachable()".
+ * Once it has been implemented reliably we can add it as an
+ * additional case:
+ *
+ * ```
+ * #if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= XXH_C23_VN)
+ * #  include <stddef.h>
+ * #  ifdef unreachable
+ * #    define XXH_UNREACHABLE() unreachable()
+ * #  endif
+ * #endif
+ * ```
+ *
+ * Note C++23 also has std::unreachable() which can be detected
+ * as follows:
+ * ```
+ * #if defined(__cpp_lib_unreachable) && (__cpp_lib_unreachable >= 202202L)
+ * #  include <utility>
+ * #  define XXH_UNREACHABLE() std::unreachable()
+ * #endif
+ * ```
+ * NB: `__cpp_lib_unreachable` is defined in the `<version>` header.
+ * We don't use that as including `<utility>` in `extern "C"` blocks
+ * doesn't work on GCC12
+ */
+
+#if XXH_HAS_BUILTIN(__builtin_unreachable)
+#  define XXH_UNREACHABLE() __builtin_unreachable()
+
+#elif defined(_MSC_VER)
+#  define XXH_UNREACHABLE() __assume(0)
+
+#else
+#  define XXH_UNREACHABLE()
+#endif
+
+#if XXH_HAS_BUILTIN(__builtin_assume)
+#  define XXH_ASSUME(c) __builtin_assume(c)
+#else
+#  define XXH_ASSUME(c) if (!(c)) { XXH_UNREACHABLE(); }
+#endif
+
+/*!
+ * @internal
+ * @def XXH_rotl32(x,r)
+ * @brief 32-bit rotate left.
+ *
+ * @param x The 32-bit integer to be rotated.
+ * @param r The number of bits to rotate.
+ * @pre
+ *   @p r > 0 && @p r < 32
+ * @note
+ *   @p x and @p r may be evaluated multiple times.
+ * @return The rotated result.
+ */
+#if !defined(NO_CLANG_BUILTIN) && XXH_HAS_BUILTIN(__builtin_rotateleft32) \
+                               && XXH_HAS_BUILTIN(__builtin_rotateleft64)
+#  define XXH_rotl32 __builtin_rotateleft32
+#  define XXH_rotl64 __builtin_rotateleft64
+/* Note: although _rotl exists for minGW (GCC under windows), performance seems poor */
+#elif defined(_MSC_VER)
+#  define XXH_rotl32(x,r) _rotl(x,r)
+#  define XXH_rotl64(x,r) _rotl64(x,r)
+#else
+#  define XXH_rotl32(x,r) (((x) << (r)) | ((x) >> (32 - (r))))
+#  define XXH_rotl64(x,r) (((x) << (r)) | ((x) >> (64 - (r))))
+#endif
+
+/*!
+ * @internal
+ * @fn xxh_u32 XXH_swap32(xxh_u32 x)
+ * @brief A 32-bit byteswap.
+ *
+ * @param x The 32-bit integer to byteswap.
+ * @return @p x, byteswapped.
+ */
+#if defined(_MSC_VER)     /* Visual Studio */
+#  define XXH_swap32 _byteswap_ulong
+#elif XXH_GCC_VERSION >= 403
+#  define XXH_swap32 __builtin_bswap32
+#else
+static xxh_u32 XXH_swap32 (xxh_u32 x)
+{
+    return  ((x << 24) & 0xff000000 ) |
+            ((x <<  8) & 0x00ff0000 ) |
+            ((x >>  8) & 0x0000ff00 ) |
+            ((x >> 24) & 0x000000ff );
+}
+#endif
+
+
+/* ***************************
+*  Memory reads
+*****************************/
+
+/*!
+ * @internal
+ * @brief Enum to indicate whether a pointer is aligned.
+ */
+typedef enum {
+    XXH_aligned,  /*!< Aligned */
+    XXH_unaligned /*!< Possibly unaligned */
+} XXH_alignment;
+
+/*
+ * XXH_FORCE_MEMORY_ACCESS==3 is an endian-independent byteshift load.
+ *
+ * This is ideal for older compilers which don't inline memcpy.
+ */
+#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
+
+XXH_FORCE_INLINE xxh_u32 XXH_readLE32(const void* memPtr)
+{
+    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
+    return bytePtr[0]
+         | ((xxh_u32)bytePtr[1] << 8)
+         | ((xxh_u32)bytePtr[2] << 16)
+         | ((xxh_u32)bytePtr[3] << 24);
+}
+
+XXH_FORCE_INLINE xxh_u32 XXH_readBE32(const void* memPtr)
+{
+    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
+    return bytePtr[3]
+         | ((xxh_u32)bytePtr[2] << 8)
+         | ((xxh_u32)bytePtr[1] << 16)
+         | ((xxh_u32)bytePtr[0] << 24);
+}
+
+#else
+XXH_FORCE_INLINE xxh_u32 XXH_readLE32(const void* ptr)
+{
+    return XXH_CPU_LITTLE_ENDIAN ? XXH_read32(ptr) : XXH_swap32(XXH_read32(ptr));
+}
+
+static xxh_u32 XXH_readBE32(const void* ptr)
+{
+    return XXH_CPU_LITTLE_ENDIAN ? XXH_swap32(XXH_read32(ptr)) : XXH_read32(ptr);
+}
+#endif
+
+XXH_FORCE_INLINE xxh_u32
+XXH_readLE32_align(const void* ptr, XXH_alignment align)
+{
+    if (align==XXH_unaligned) {
+        return XXH_readLE32(ptr);
+    } else {
+        return XXH_CPU_LITTLE_ENDIAN ? *(const xxh_u32*)ptr : XXH_swap32(*(const xxh_u32*)ptr);
+    }
+}
+
+
+/* *************************************
+*  Misc
+***************************************/
+/*! @ingroup public */
+XXH_PUBLIC_API unsigned XXH_versionNumber (void) { return XXH_VERSION_NUMBER; }
+
+
+/* *******************************************************************
+*  32-bit hash functions
+*********************************************************************/
+/*!
+ * @}
+ * @defgroup XXH32_impl XXH32 implementation
+ * @ingroup impl
+ *
+ * Details on the XXH32 implementation.
+ * @{
+ */
+ /* #define instead of static const, to be used as initializers */
+#define XXH_PRIME32_1  0x9E3779B1U  /*!< 0b10011110001101110111100110110001 */
+#define XXH_PRIME32_2  0x85EBCA77U  /*!< 0b10000101111010111100101001110111 */
+#define XXH_PRIME32_3  0xC2B2AE3DU  /*!< 0b11000010101100101010111000111101 */
+#define XXH_PRIME32_4  0x27D4EB2FU  /*!< 0b00100111110101001110101100101111 */
+#define XXH_PRIME32_5  0x165667B1U  /*!< 0b00010110010101100110011110110001 */
+
+#ifdef XXH_OLD_NAMES
+#  define PRIME32_1 XXH_PRIME32_1
+#  define PRIME32_2 XXH_PRIME32_2
+#  define PRIME32_3 XXH_PRIME32_3
+#  define PRIME32_4 XXH_PRIME32_4
+#  define PRIME32_5 XXH_PRIME32_5
+#endif
+
+/*!
+ * @internal
+ * @brief Normal stripe processing routine.
+ *
+ * This shuffles the bits so that any bit from @p input impacts several bits in
+ * @p acc.
+ *
+ * @param acc The accumulator lane.
+ * @param input The stripe of input to mix.
+ * @return The mixed accumulator lane.
+ */
+static xxh_u32 XXH32_round(xxh_u32 acc, xxh_u32 input)
+{
+    acc += input * XXH_PRIME32_2;
+    acc  = XXH_rotl32(acc, 13);
+    acc *= XXH_PRIME32_1;
+#if (defined(__SSE4_1__) || defined(__aarch64__) || defined(__wasm_simd128__)) && !defined(XXH_ENABLE_AUTOVECTORIZE)
+    /*
+     * UGLY HACK:
+     * A compiler fence is the only thing that prevents GCC and Clang from
+     * autovectorizing the XXH32 loop (pragmas and attributes don't work for some
+     * reason) without globally disabling SSE4.1.
+     *
+     * The reason we want to avoid vectorization is because despite working on
+     * 4 integers at a time, there are multiple factors slowing XXH32 down on
+     * SSE4:
+     * - There's a ridiculous amount of lag from pmulld (10 cycles of latency on
+     *   newer chips!) making it slightly slower to multiply four integers at
+     *   once compared to four integers independently. Even when pmulld was
+     *   fastest, Sandy/Ivy Bridge, it is still not worth it to go into SSE
+     *   just to multiply unless doing a long operation.
+     *
+     * - Four instructions are required to rotate,
+     *      movqda tmp,  v // not required with VEX encoding
+     *      pslld  tmp, 13 // tmp <<= 13
+     *      psrld  v,   19 // x >>= 19
+     *      por    v,  tmp // x |= tmp
+     *   compared to one for scalar:
+     *      roll   v, 13    // reliably fast across the board
+     *      shldl  v, v, 13 // Sandy Bridge and later prefer this for some reason
+     *
+     * - Instruction level parallelism is actually more beneficial here because
+     *   the SIMD actually serializes this operation: While v1 is rotating, v2
+     *   can load data, while v3 can multiply. SSE forces them to operate
+     *   together.
+     *
+     * This is also enabled on AArch64, as Clang is *very aggressive* in vectorizing
+     * the loop. NEON is only faster on the A53, and with the newer cores, it is less
+     * than half the speed.
+     *
+     * Additionally, this is used on WASM SIMD128 because it JITs to the same
+     * SIMD instructions and has the same issue.
+     */
+    XXH_COMPILER_GUARD(acc);
+#endif
+    return acc;
+}
+
+/*!
+ * @internal
+ * @brief Mixes all bits to finalize the hash.
+ *
+ * The final mix ensures that all input bits have a chance to impact any bit in
+ * the output digest, resulting in an unbiased distribution.
+ *
+ * @param hash The hash to avalanche.
+ * @return The avalanched hash.
+ */
+static xxh_u32 XXH32_avalanche(xxh_u32 hash)
+{
+    hash ^= hash >> 15;
+    hash *= XXH_PRIME32_2;
+    hash ^= hash >> 13;
+    hash *= XXH_PRIME32_3;
+    hash ^= hash >> 16;
+    return hash;
+}
+
+#define XXH_get32bits(p) XXH_readLE32_align(p, align)
+
+/*!
+ * @internal
+ * @brief Processes the last 0-15 bytes of @p ptr.
+ *
+ * There may be up to 15 bytes remaining to consume from the input.
+ * This final stage will digest them to ensure that all input bytes are present
+ * in the final mix.
+ *
+ * @param hash The hash to finalize.
+ * @param ptr The pointer to the remaining input.
+ * @param len The remaining length, modulo 16.
+ * @param align Whether @p ptr is aligned.
+ * @return The finalized hash.
+ * @see XXH64_finalize().
+ */
+static XXH_PUREF xxh_u32
+XXH32_finalize(xxh_u32 hash, const xxh_u8* ptr, size_t len, XXH_alignment align)
+{
+#define XXH_PROCESS1 do {                             \
+    hash += (*ptr++) * XXH_PRIME32_5;                 \
+    hash = XXH_rotl32(hash, 11) * XXH_PRIME32_1;      \
+} while (0)
+
+#define XXH_PROCESS4 do {                             \
+    hash += XXH_get32bits(ptr) * XXH_PRIME32_3;       \
+    ptr += 4;                                         \
+    hash  = XXH_rotl32(hash, 17) * XXH_PRIME32_4;     \
+} while (0)
+
+    if (ptr==NULL) XXH_ASSERT(len == 0);
+
+    /* Compact rerolled version; generally faster */
+    if (!XXH32_ENDJMP) {
+        len &= 15;
+        while (len >= 4) {
+            XXH_PROCESS4;
+            len -= 4;
+        }
+        while (len > 0) {
+            XXH_PROCESS1;
+            --len;
+        }
+        return XXH32_avalanche(hash);
+    } else {
+         switch(len&15) /* or switch(bEnd - p) */ {
+           case 12:      XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 8:       XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 4:       XXH_PROCESS4;
+                         return XXH32_avalanche(hash);
+
+           case 13:      XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 9:       XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 5:       XXH_PROCESS4;
+                         XXH_PROCESS1;
+                         return XXH32_avalanche(hash);
+
+           case 14:      XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 10:      XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 6:       XXH_PROCESS4;
+                         XXH_PROCESS1;
+                         XXH_PROCESS1;
+                         return XXH32_avalanche(hash);
+
+           case 15:      XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 11:      XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 7:       XXH_PROCESS4;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 3:       XXH_PROCESS1;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 2:       XXH_PROCESS1;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 1:       XXH_PROCESS1;
+                         XXH_FALLTHROUGH;  /* fallthrough */
+           case 0:       return XXH32_avalanche(hash);
+        }
+        XXH_ASSERT(0);
+        return hash;   /* reaching this point is deemed impossible */
+    }
+}
+
+#ifdef XXH_OLD_NAMES
+#  define PROCESS1 XXH_PROCESS1
+#  define PROCESS4 XXH_PROCESS4
+#else
+#  undef XXH_PROCESS1
+#  undef XXH_PROCESS4
+#endif
+
+/*!
+ * @internal
+ * @brief The implementation for @ref XXH32().
+ *
+ * @param input , len , seed Directly passed from @ref XXH32().
+ * @param align Whether @p input is aligned.
+ * @return The calculated hash.
+ */
+XXH_FORCE_INLINE XXH_PUREF xxh_u32
+XXH32_endian_align(const xxh_u8* input, size_t len, xxh_u32 seed, XXH_alignment align)
+{
+    xxh_u32 h32;
+
+    if (input==NULL) XXH_ASSERT(len == 0);
+
+    if (len>=16) {
+        const xxh_u8* const bEnd = input + len;
+        const xxh_u8* const limit = bEnd - 15;
+        xxh_u32 v1 = seed + XXH_PRIME32_1 + XXH_PRIME32_2;
+        xxh_u32 v2 = seed + XXH_PRIME32_2;
+        xxh_u32 v3 = seed + 0;
+        xxh_u32 v4 = seed - XXH_PRIME32_1;
+
+        do {
+            v1 = XXH32_round(v1, XXH_get32bits(input)); input += 4;
+            v2 = XXH32_round(v2, XXH_get32bits(input)); input += 4;
+            v3 = XXH32_round(v3, XXH_get32bits(input)); input += 4;
+            v4 = XXH32_round(v4, XXH_get32bits(input)); input += 4;
+        } while (input < limit);
+
+        h32 = XXH_rotl32(v1, 1)  + XXH_rotl32(v2, 7)
+            + XXH_rotl32(v3, 12) + XXH_rotl32(v4, 18);
+    } else {
+        h32  = seed + XXH_PRIME32_5;
+    }
+
+    h32 += (xxh_u32)len;
+
+    return XXH32_finalize(h32, input, len&15, align);
+}
+
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t len, XXH32_hash_t seed)
+{
+#if !defined(XXH_NO_STREAM) && XXH_SIZE_OPT >= 2
+    /* Simple version, good for code maintenance, but unfortunately slow for small inputs */
+    XXH32_state_t state;
+    XXH32_reset(&state, seed);
+    XXH32_update(&state, (const xxh_u8*)input, len);
+    return XXH32_digest(&state);
+#else
+    if (XXH_FORCE_ALIGN_CHECK) {
+        if ((((size_t)input) & 3) == 0) {   /* Input is 4-bytes aligned, leverage the speed benefit */
+            return XXH32_endian_align((const xxh_u8*)input, len, seed, XXH_aligned);
+    }   }
+
+    return XXH32_endian_align((const xxh_u8*)input, len, seed, XXH_unaligned);
+#endif
+}
+
+
+
+/*******   Hash streaming   *******/
+#ifndef XXH_NO_STREAM
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH32_state_t* XXH32_createState(void)
+{
+    return (XXH32_state_t*)XXH_malloc(sizeof(XXH32_state_t));
+}
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH_errorcode XXH32_freeState(XXH32_state_t* statePtr)
+{
+    XXH_free(statePtr);
+    return XXH_OK;
+}
+
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dstState, const XXH32_state_t* srcState)
+{
+    XXH_memcpy(dstState, srcState, sizeof(*dstState));
+}
+
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH_errorcode XXH32_reset(XXH32_state_t* statePtr, XXH32_hash_t seed)
+{
+    XXH_ASSERT(statePtr != NULL);
+    memset(statePtr, 0, sizeof(*statePtr));
+    statePtr->v[0] = seed + XXH_PRIME32_1 + XXH_PRIME32_2;
+    statePtr->v[1] = seed + XXH_PRIME32_2;
+    statePtr->v[2] = seed + 0;
+    statePtr->v[3] = seed - XXH_PRIME32_1;
+    return XXH_OK;
+}
+
+
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH32_update(XXH32_state_t* state, const void* input, size_t len)
+{
+    if (input==NULL) {
+        XXH_ASSERT(len == 0);
+        return XXH_OK;
+    }
+
+    {   const xxh_u8* p = (const xxh_u8*)input;
+        const xxh_u8* const bEnd = p + len;
+
+        state->total_len_32 += (XXH32_hash_t)len;
+        state->large_len |= (XXH32_hash_t)((len>=16) | (state->total_len_32>=16));
+
+        if (state->memsize + len < 16)  {   /* fill in tmp buffer */
+            XXH_memcpy((xxh_u8*)(state->mem32) + state->memsize, input, len);
+            state->memsize += (XXH32_hash_t)len;
+            return XXH_OK;
+        }
+
+        if (state->memsize) {   /* some data left from previous update */
+            XXH_memcpy((xxh_u8*)(state->mem32) + state->memsize, input, 16-state->memsize);
+            {   const xxh_u32* p32 = state->mem32;
+                state->v[0] = XXH32_round(state->v[0], XXH_readLE32(p32)); p32++;
+                state->v[1] = XXH32_round(state->v[1], XXH_readLE32(p32)); p32++;
+                state->v[2] = XXH32_round(state->v[2], XXH_readLE32(p32)); p32++;
+                state->v[3] = XXH32_round(state->v[3], XXH_readLE32(p32));
+            }
+            p += 16-state->memsize;
+            state->memsize = 0;
+        }
+
+        if (p <= bEnd-16) {
+            const xxh_u8* const limit = bEnd - 16;
+
+            do {
+                state->v[0] = XXH32_round(state->v[0], XXH_readLE32(p)); p+=4;
+                state->v[1] = XXH32_round(state->v[1], XXH_readLE32(p)); p+=4;
+                state->v[2] = XXH32_round(state->v[2], XXH_readLE32(p)); p+=4;
+                state->v[3] = XXH32_round(state->v[3], XXH_readLE32(p)); p+=4;
+            } while (p<=limit);
+
+        }
+
+        if (p < bEnd) {
+            XXH_memcpy(state->mem32, p, (size_t)(bEnd-p));
+            state->memsize = (unsigned)(bEnd-p);
+        }
+    }
+
+    return XXH_OK;
+}
+
+
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH32_hash_t XXH32_digest(const XXH32_state_t* state)
+{
+    xxh_u32 h32;
+
+    if (state->large_len) {
+        h32 = XXH_rotl32(state->v[0], 1)
+            + XXH_rotl32(state->v[1], 7)
+            + XXH_rotl32(state->v[2], 12)
+            + XXH_rotl32(state->v[3], 18);
+    } else {
+        h32 = state->v[2] /* == seed */ + XXH_PRIME32_5;
+    }
+
+    h32 += state->total_len_32;
+
+    return XXH32_finalize(h32, (const xxh_u8*)state->mem32, state->memsize, XXH_aligned);
+}
+#endif /* !XXH_NO_STREAM */
+
+/*******   Canonical representation   *******/
+
+/*!
+ * @ingroup XXH32_family
+ * The default return values from XXH functions are unsigned 32 and 64 bit
+ * integers.
+ *
+ * The canonical representation uses big endian convention, the same convention
+ * as human-readable numbers (large digits first).
+ *
+ * This way, hash values can be written into a file or buffer, remaining
+ * comparable across different systems.
+ *
+ * The following functions allow transformation of hash values to and from their
+ * canonical format.
+ */
+XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash)
+{
+    XXH_STATIC_ASSERT(sizeof(XXH32_canonical_t) == sizeof(XXH32_hash_t));
+    if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap32(hash);
+    XXH_memcpy(dst, &hash, sizeof(*dst));
+}
+/*! @ingroup XXH32_family */
+XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src)
+{
+    return XXH_readBE32(src);
+}
+
+
+#ifndef XXH_NO_LONG_LONG
+
+/* *******************************************************************
+*  64-bit hash functions
+*********************************************************************/
+/*!
+ * @}
+ * @ingroup impl
+ * @{
+ */
+/*******   Memory access   *******/
+
+typedef XXH64_hash_t xxh_u64;
+
+#ifdef XXH_OLD_NAMES
+#  define U64 xxh_u64
+#endif
+
+#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
+/*
+ * Manual byteshift. Best for old compilers which don't inline memcpy.
+ * We actually directly use XXH_readLE64 and XXH_readBE64.
+ */
+#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==2))
+
+/* Force direct memory access. Only works on CPU which support unaligned memory access in hardware */
+static xxh_u64 XXH_read64(const void* memPtr)
+{
+    return *(const xxh_u64*) memPtr;
+}
+
+#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1))
+
+/*
+ * __attribute__((aligned(1))) is supported by gcc and clang. Originally the
+ * documentation claimed that it only increased the alignment, but actually it
+ * can decrease it on gcc, clang, and icc:
+ * https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69502,
+ * https://gcc.godbolt.org/z/xYez1j67Y.
+ */
+#ifdef XXH_OLD_NAMES
+typedef union { xxh_u32 u32; xxh_u64 u64; } __attribute__((packed)) unalign64;
+#endif
+static xxh_u64 XXH_read64(const void* ptr)
+{
+    typedef __attribute__((aligned(1))) xxh_u64 xxh_unalign64;
+    return *((const xxh_unalign64*)ptr);
+}
+
+#else
+
+/*
+ * Portable and safe solution. Generally efficient.
+ * see: https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html
+ */
+static xxh_u64 XXH_read64(const void* memPtr)
+{
+    xxh_u64 val;
+    XXH_memcpy(&val, memPtr, sizeof(val));
+    return val;
+}
+
+#endif   /* XXH_FORCE_DIRECT_MEMORY_ACCESS */
+
+#if defined(_MSC_VER)     /* Visual Studio */
+#  define XXH_swap64 _byteswap_uint64
+#elif XXH_GCC_VERSION >= 403
+#  define XXH_swap64 __builtin_bswap64
+#else
+static xxh_u64 XXH_swap64(xxh_u64 x)
+{
+    return  ((x << 56) & 0xff00000000000000ULL) |
+            ((x << 40) & 0x00ff000000000000ULL) |
+            ((x << 24) & 0x0000ff0000000000ULL) |
+            ((x << 8)  & 0x000000ff00000000ULL) |
+            ((x >> 8)  & 0x00000000ff000000ULL) |
+            ((x >> 24) & 0x0000000000ff0000ULL) |
+            ((x >> 40) & 0x000000000000ff00ULL) |
+            ((x >> 56) & 0x00000000000000ffULL);
+}
+#endif
+
+
+/* XXH_FORCE_MEMORY_ACCESS==3 is an endian-independent byteshift load. */
+#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
+
+XXH_FORCE_INLINE xxh_u64 XXH_readLE64(const void* memPtr)
+{
+    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
+    return bytePtr[0]
+         | ((xxh_u64)bytePtr[1] << 8)
+         | ((xxh_u64)bytePtr[2] << 16)
+         | ((xxh_u64)bytePtr[3] << 24)
+         | ((xxh_u64)bytePtr[4] << 32)
+         | ((xxh_u64)bytePtr[5] << 40)
+         | ((xxh_u64)bytePtr[6] << 48)
+         | ((xxh_u64)bytePtr[7] << 56);
+}
+
+XXH_FORCE_INLINE xxh_u64 XXH_readBE64(const void* memPtr)
+{
+    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
+    return bytePtr[7]
+         | ((xxh_u64)bytePtr[6] << 8)
+         | ((xxh_u64)bytePtr[5] << 16)
+         | ((xxh_u64)bytePtr[4] << 24)
+         | ((xxh_u64)bytePtr[3] << 32)
+         | ((xxh_u64)bytePtr[2] << 40)
+         | ((xxh_u64)bytePtr[1] << 48)
+         | ((xxh_u64)bytePtr[0] << 56);
+}
+
+#else
+XXH_FORCE_INLINE xxh_u64 XXH_readLE64(const void* ptr)
+{
+    return XXH_CPU_LITTLE_ENDIAN ? XXH_read64(ptr) : XXH_swap64(XXH_read64(ptr));
+}
+
+static xxh_u64 XXH_readBE64(const void* ptr)
+{
+    return XXH_CPU_LITTLE_ENDIAN ? XXH_swap64(XXH_read64(ptr)) : XXH_read64(ptr);
+}
+#endif
+
+XXH_FORCE_INLINE xxh_u64
+XXH_readLE64_align(const void* ptr, XXH_alignment align)
+{
+    if (align==XXH_unaligned)
+        return XXH_readLE64(ptr);
+    else
+        return XXH_CPU_LITTLE_ENDIAN ? *(const xxh_u64*)ptr : XXH_swap64(*(const xxh_u64*)ptr);
+}
+
+
+/*******   xxh64   *******/
+/*!
+ * @}
+ * @defgroup XXH64_impl XXH64 implementation
+ * @ingroup impl
+ *
+ * Details on the XXH64 implementation.
+ * @{
+ */
+/* #define rather that static const, to be used as initializers */
+#define XXH_PRIME64_1  0x9E3779B185EBCA87ULL  /*!< 0b1001111000110111011110011011000110000101111010111100101010000111 */
+#define XXH_PRIME64_2  0xC2B2AE3D27D4EB4FULL  /*!< 0b1100001010110010101011100011110100100111110101001110101101001111 */
+#define XXH_PRIME64_3  0x165667B19E3779F9ULL  /*!< 0b0001011001010110011001111011000110011110001101110111100111111001 */
+#define XXH_PRIME64_4  0x85EBCA77C2B2AE63ULL  /*!< 0b1000010111101011110010100111011111000010101100101010111001100011 */
+#define XXH_PRIME64_5  0x27D4EB2F165667C5ULL  /*!< 0b0010011111010100111010110010111100010110010101100110011111000101 */
+
+#ifdef XXH_OLD_NAMES
+#  define PRIME64_1 XXH_PRIME64_1
+#  define PRIME64_2 XXH_PRIME64_2
+#  define PRIME64_3 XXH_PRIME64_3
+#  define PRIME64_4 XXH_PRIME64_4
+#  define PRIME64_5 XXH_PRIME64_5
+#endif
+
+/*! @copydoc XXH32_round */
+static xxh_u64 XXH64_round(xxh_u64 acc, xxh_u64 input)
+{
+    acc += input * XXH_PRIME64_2;
+    acc  = XXH_rotl64(acc, 31);
+    acc *= XXH_PRIME64_1;
+    return acc;
+}
+
+static xxh_u64 XXH64_mergeRound(xxh_u64 acc, xxh_u64 val)
+{
+    val  = XXH64_round(0, val);
+    acc ^= val;
+    acc  = acc * XXH_PRIME64_1 + XXH_PRIME64_4;
+    return acc;
+}
+
+/*! @copydoc XXH32_avalanche */
+static xxh_u64 XXH64_avalanche(xxh_u64 hash)
+{
+    hash ^= hash >> 33;
+    hash *= XXH_PRIME64_2;
+    hash ^= hash >> 29;
+    hash *= XXH_PRIME64_3;
+    hash ^= hash >> 32;
+    return hash;
+}
+
+
+#define XXH_get64bits(p) XXH_readLE64_align(p, align)
+
+/*!
+ * @internal
+ * @brief Processes the last 0-31 bytes of @p ptr.
+ *
+ * There may be up to 31 bytes remaining to consume from the input.
+ * This final stage will digest them to ensure that all input bytes are present
+ * in the final mix.
+ *
+ * @param hash The hash to finalize.
+ * @param ptr The pointer to the remaining input.
+ * @param len The remaining length, modulo 32.
+ * @param align Whether @p ptr is aligned.
+ * @return The finalized hash
+ * @see XXH32_finalize().
+ */
+static XXH_PUREF xxh_u64
+XXH64_finalize(xxh_u64 hash, const xxh_u8* ptr, size_t len, XXH_alignment align)
+{
+    if (ptr==NULL) XXH_ASSERT(len == 0);
+    len &= 31;
+    while (len >= 8) {
+        xxh_u64 const k1 = XXH64_round(0, XXH_get64bits(ptr));
+        ptr += 8;
+        hash ^= k1;
+        hash  = XXH_rotl64(hash,27) * XXH_PRIME64_1 + XXH_PRIME64_4;
+        len -= 8;
+    }
+    if (len >= 4) {
+        hash ^= (xxh_u64)(XXH_get32bits(ptr)) * XXH_PRIME64_1;
+        ptr += 4;
+        hash = XXH_rotl64(hash, 23) * XXH_PRIME64_2 + XXH_PRIME64_3;
+        len -= 4;
+    }
+    while (len > 0) {
+        hash ^= (*ptr++) * XXH_PRIME64_5;
+        hash = XXH_rotl64(hash, 11) * XXH_PRIME64_1;
+        --len;
+    }
+    return  XXH64_avalanche(hash);
+}
+
+#ifdef XXH_OLD_NAMES
+#  define PROCESS1_64 XXH_PROCESS1_64
+#  define PROCESS4_64 XXH_PROCESS4_64
+#  define PROCESS8_64 XXH_PROCESS8_64
+#else
+#  undef XXH_PROCESS1_64
+#  undef XXH_PROCESS4_64
+#  undef XXH_PROCESS8_64
+#endif
+
+/*!
+ * @internal
+ * @brief The implementation for @ref XXH64().
+ *
+ * @param input , len , seed Directly passed from @ref XXH64().
+ * @param align Whether @p input is aligned.
+ * @return The calculated hash.
+ */
+XXH_FORCE_INLINE XXH_PUREF xxh_u64
+XXH64_endian_align(const xxh_u8* input, size_t len, xxh_u64 seed, XXH_alignment align)
+{
+    xxh_u64 h64;
+    if (input==NULL) XXH_ASSERT(len == 0);
+
+    if (len>=32) {
+        const xxh_u8* const bEnd = input + len;
+        const xxh_u8* const limit = bEnd - 31;
+        xxh_u64 v1 = seed + XXH_PRIME64_1 + XXH_PRIME64_2;
+        xxh_u64 v2 = seed + XXH_PRIME64_2;
+        xxh_u64 v3 = seed + 0;
+        xxh_u64 v4 = seed - XXH_PRIME64_1;
+
+        do {
+            v1 = XXH64_round(v1, XXH_get64bits(input)); input+=8;
+            v2 = XXH64_round(v2, XXH_get64bits(input)); input+=8;
+            v3 = XXH64_round(v3, XXH_get64bits(input)); input+=8;
+            v4 = XXH64_round(v4, XXH_get64bits(input)); input+=8;
+        } while (input<limit);
+
+        h64 = XXH_rotl64(v1, 1) + XXH_rotl64(v2, 7) + XXH_rotl64(v3, 12) + XXH_rotl64(v4, 18);
+        h64 = XXH64_mergeRound(h64, v1);
+        h64 = XXH64_mergeRound(h64, v2);
+        h64 = XXH64_mergeRound(h64, v3);
+        h64 = XXH64_mergeRound(h64, v4);
+
+    } else {
+        h64  = seed + XXH_PRIME64_5;
+    }
+
+    h64 += (xxh_u64) len;
+
+    return XXH64_finalize(h64, input, len, align);
+}
+
+
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH64_hash_t XXH64 (XXH_NOESCAPE const void* input, size_t len, XXH64_hash_t seed)
+{
+#if !defined(XXH_NO_STREAM) && XXH_SIZE_OPT >= 2
+    /* Simple version, good for code maintenance, but unfortunately slow for small inputs */
+    XXH64_state_t state;
+    XXH64_reset(&state, seed);
+    XXH64_update(&state, (const xxh_u8*)input, len);
+    return XXH64_digest(&state);
+#else
+    if (XXH_FORCE_ALIGN_CHECK) {
+        if ((((size_t)input) & 7)==0) {  /* Input is aligned, let's leverage the speed advantage */
+            return XXH64_endian_align((const xxh_u8*)input, len, seed, XXH_aligned);
+    }   }
+
+    return XXH64_endian_align((const xxh_u8*)input, len, seed, XXH_unaligned);
+
+#endif
+}
+
+/*******   Hash Streaming   *******/
+#ifndef XXH_NO_STREAM
+/*! @ingroup XXH64_family*/
+XXH_PUBLIC_API XXH64_state_t* XXH64_createState(void)
+{
+    return (XXH64_state_t*)XXH_malloc(sizeof(XXH64_state_t));
+}
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH_errorcode XXH64_freeState(XXH64_state_t* statePtr)
+{
+    XXH_free(statePtr);
+    return XXH_OK;
+}
+
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API void XXH64_copyState(XXH_NOESCAPE XXH64_state_t* dstState, const XXH64_state_t* srcState)
+{
+    XXH_memcpy(dstState, srcState, sizeof(*dstState));
+}
+
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH_errorcode XXH64_reset(XXH_NOESCAPE XXH64_state_t* statePtr, XXH64_hash_t seed)
+{
+    XXH_ASSERT(statePtr != NULL);
+    memset(statePtr, 0, sizeof(*statePtr));
+    statePtr->v[0] = seed + XXH_PRIME64_1 + XXH_PRIME64_2;
+    statePtr->v[1] = seed + XXH_PRIME64_2;
+    statePtr->v[2] = seed + 0;
+    statePtr->v[3] = seed - XXH_PRIME64_1;
+    return XXH_OK;
+}
+
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH64_update (XXH_NOESCAPE XXH64_state_t* state, XXH_NOESCAPE const void* input, size_t len)
+{
+    if (input==NULL) {
+        XXH_ASSERT(len == 0);
+        return XXH_OK;
+    }
+
+    {   const xxh_u8* p = (const xxh_u8*)input;
+        const xxh_u8* const bEnd = p + len;
+
+        state->total_len += len;
+
+        if (state->memsize + len < 32) {  /* fill in tmp buffer */
+            XXH_memcpy(((xxh_u8*)state->mem64) + state->memsize, input, len);
+            state->memsize += (xxh_u32)len;
+            return XXH_OK;
+        }
+
+        if (state->memsize) {   /* tmp buffer is full */
+            XXH_memcpy(((xxh_u8*)state->mem64) + state->memsize, input, 32-state->memsize);
+            state->v[0] = XXH64_round(state->v[0], XXH_readLE64(state->mem64+0));
+            state->v[1] = XXH64_round(state->v[1], XXH_readLE64(state->mem64+1));
+            state->v[2] = XXH64_round(state->v[2], XXH_readLE64(state->mem64+2));
+            state->v[3] = XXH64_round(state->v[3], XXH_readLE64(state->mem64+3));
+            p += 32 - state->memsize;
+            state->memsize = 0;
+        }
+
+        if (p+32 <= bEnd) {
+            const xxh_u8* const limit = bEnd - 32;
+
+            do {
+                state->v[0] = XXH64_round(state->v[0], XXH_readLE64(p)); p+=8;
+                state->v[1] = XXH64_round(state->v[1], XXH_readLE64(p)); p+=8;
+                state->v[2] = XXH64_round(state->v[2], XXH_readLE64(p)); p+=8;
+                state->v[3] = XXH64_round(state->v[3], XXH_readLE64(p)); p+=8;
+            } while (p<=limit);
+
+        }
+
+        if (p < bEnd) {
+            XXH_memcpy(state->mem64, p, (size_t)(bEnd-p));
+            state->memsize = (unsigned)(bEnd-p);
+        }
+    }
+
+    return XXH_OK;
+}
+
+
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH64_hash_t XXH64_digest(XXH_NOESCAPE const XXH64_state_t* state)
+{
+    xxh_u64 h64;
+
+    if (state->total_len >= 32) {
+        h64 = XXH_rotl64(state->v[0], 1) + XXH_rotl64(state->v[1], 7) + XXH_rotl64(state->v[2], 12) + XXH_rotl64(state->v[3], 18);
+        h64 = XXH64_mergeRound(h64, state->v[0]);
+        h64 = XXH64_mergeRound(h64, state->v[1]);
+        h64 = XXH64_mergeRound(h64, state->v[2]);
+        h64 = XXH64_mergeRound(h64, state->v[3]);
+    } else {
+        h64  = state->v[2] /*seed*/ + XXH_PRIME64_5;
+    }
+
+    h64 += (xxh_u64) state->total_len;
+
+    return XXH64_finalize(h64, (const xxh_u8*)state->mem64, (size_t)state->total_len, XXH_aligned);
+}
+#endif /* !XXH_NO_STREAM */
+
+/******* Canonical representation   *******/
+
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH_NOESCAPE XXH64_canonical_t* dst, XXH64_hash_t hash)
+{
+    XXH_STATIC_ASSERT(sizeof(XXH64_canonical_t) == sizeof(XXH64_hash_t));
+    if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap64(hash);
+    XXH_memcpy(dst, &hash, sizeof(*dst));
+}
+
+/*! @ingroup XXH64_family */
+XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(XXH_NOESCAPE const XXH64_canonical_t* src)
+{
+    return XXH_readBE64(src);
+}
+
+#ifndef XXH_NO_XXH3
+
+/* *********************************************************************
+*  XXH3
+*  New generation hash designed for speed on small keys and vectorization
+************************************************************************ */
+/*!
+ * @}
+ * @defgroup XXH3_impl XXH3 implementation
+ * @ingroup impl
+ * @{
+ */
+
+/* ===   Compiler specifics   === */
+
+#if ((defined(sun) || defined(__sun)) && __cplusplus) /* Solaris includes __STDC_VERSION__ with C++. Tested with GCC 5.5 */
+#  define XXH_RESTRICT   /* disable */
+#elif defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L   /* >= C99 */
+#  define XXH_RESTRICT   restrict
+#elif (defined (__GNUC__) && ((__GNUC__ > 3) || (__GNUC__ == 3 && __GNUC_MINOR__ >= 1))) \
+   || (defined (__clang__)) \
+   || (defined (_MSC_VER) && (_MSC_VER >= 1400)) \
+   || (defined (__INTEL_COMPILER) && (__INTEL_COMPILER >= 1300))
+/*
+ * There are a LOT more compilers that recognize __restrict but this
+ * covers the major ones.
+ */
+#  define XXH_RESTRICT   __restrict
+#else
+#  define XXH_RESTRICT   /* disable */
+#endif
+
+#if (defined(__GNUC__) && (__GNUC__ >= 3))  \
+  || (defined(__INTEL_COMPILER) && (__INTEL_COMPILER >= 800)) \
+  || defined(__clang__)
+#    define XXH_likely(x) __builtin_expect(x, 1)
+#    define XXH_unlikely(x) __builtin_expect(x, 0)
+#else
+#    define XXH_likely(x) (x)
+#    define XXH_unlikely(x) (x)
+#endif
+
+#ifndef XXH_HAS_INCLUDE
+#  ifdef __has_include
+#    define XXH_HAS_INCLUDE(x) __has_include(x)
+#  else
+#    define XXH_HAS_INCLUDE(x) 0
+#  endif
+#endif
+
+#if defined(__GNUC__) || defined(__clang__)
+#  if defined(__ARM_FEATURE_SVE)
+#    include <arm_sve.h>
+#  endif
+#  if defined(__ARM_NEON__) || defined(__ARM_NEON) \
+   || (defined(_M_ARM) && _M_ARM >= 7) \
+   || defined(_M_ARM64) || defined(_M_ARM64EC) \
+   || (defined(__wasm_simd128__) && XXH_HAS_INCLUDE(<arm_neon.h>)) /* WASM SIMD128 via SIMDe */
+#    define inline __inline__  /* circumvent a clang bug */
+#    include <arm_neon.h>
+#    undef inline
+#  elif defined(__AVX2__)
+#    include <immintrin.h>
+#  elif defined(__SSE2__)
+#    include <emmintrin.h>
+#  endif
+#endif
+
+#if defined(_MSC_VER)
+#  include <intrin.h>
+#endif
+
+/*
+ * One goal of XXH3 is to make it fast on both 32-bit and 64-bit, while
+ * remaining a true 64-bit/128-bit hash function.
+ *
+ * This is done by prioritizing a subset of 64-bit operations that can be
+ * emulated without too many steps on the average 32-bit machine.
+ *
+ * For example, these two lines seem similar, and run equally fast on 64-bit:
+ *
+ *   xxh_u64 x;
+ *   x ^= (x >> 47); // good
+ *   x ^= (x >> 13); // bad
+ *
+ * However, to a 32-bit machine, there is a major difference.
+ *
+ * x ^= (x >> 47) looks like this:
+ *
+ *   x.lo ^= (x.hi >> (47 - 32));
+ *
+ * while x ^= (x >> 13) looks like this:
+ *
+ *   // note: funnel shifts are not usually cheap.
+ *   x.lo ^= (x.lo >> 13) | (x.hi << (32 - 13));
+ *   x.hi ^= (x.hi >> 13);
+ *
+ * The first one is significantly faster than the second, simply because the
+ * shift is larger than 32. This means:
+ *  - All the bits we need are in the upper 32 bits, so we can ignore the lower
+ *    32 bits in the shift.
+ *  - The shift result will always fit in the lower 32 bits, and therefore,
+ *    we can ignore the upper 32 bits in the xor.
+ *
+ * Thanks to this optimization, XXH3 only requires these features to be efficient:
+ *
+ *  - Usable unaligned access
+ *  - A 32-bit or 64-bit ALU
+ *      - If 32-bit, a decent ADC instruction
+ *  - A 32 or 64-bit multiply with a 64-bit result
+ *  - For the 128-bit variant, a decent byteswap helps short inputs.
+ *
+ * The first two are already required by XXH32, and almost all 32-bit and 64-bit
+ * platforms which can run XXH32 can run XXH3 efficiently.
+ *
+ * Thumb-1, the classic 16-bit only subset of ARM's instruction set, is one
+ * notable exception.
+ *
+ * First of all, Thumb-1 lacks support for the UMULL instruction which
+ * performs the important long multiply. This means numerous __aeabi_lmul
+ * calls.
+ *
+ * Second of all, the 8 functional registers are just not enough.
+ * Setup for __aeabi_lmul, byteshift loads, pointers, and all arithmetic need
+ * Lo registers, and this shuffling results in thousands more MOVs than A32.
+ *
+ * A32 and T32 don't have this limitation. They can access all 14 registers,
+ * do a 32->64 multiply with UMULL, and the flexible operand allowing free
+ * shifts is helpful, too.
+ *
+ * Therefore, we do a quick sanity check.
+ *
+ * If compiling Thumb-1 for a target which supports ARM instructions, we will
+ * emit a warning, as it is not a "sane" platform to compile for.
+ *
+ * Usually, if this happens, it is because of an accident and you probably need
+ * to specify -march, as you likely meant to compile for a newer architecture.
+ *
+ * Credit: large sections of the vectorial and asm source code paths
+ *         have been contributed by @easyaspi314
+ */
+#if defined(__thumb__) && !defined(__thumb2__) && defined(__ARM_ARCH_ISA_ARM)
+#   warning "XXH3 is highly inefficient without ARM or Thumb-2."
+#endif
+
+/* ==========================================
+ * Vectorization detection
+ * ========================================== */
+
+#ifdef XXH_DOXYGEN
+/*!
+ * @ingroup tuning
+ * @brief Overrides the vectorization implementation chosen for XXH3.
+ *
+ * Can be defined to 0 to disable SIMD or any of the values mentioned in
+ * @ref XXH_VECTOR_TYPE.
+ *
+ * If this is not defined, it uses predefined macros to determine the best
+ * implementation.
+ */
+#  define XXH_VECTOR XXH_SCALAR
+/*!
+ * @ingroup tuning
+ * @brief Possible values for @ref XXH_VECTOR.
+ *
+ * Note that these are actually implemented as macros.
+ *
+ * If this is not defined, it is detected automatically.
+ * internal macro XXH_X86DISPATCH overrides this.
+ */
+enum XXH_VECTOR_TYPE /* fake enum */ {
+    XXH_SCALAR = 0,  /*!< Portable scalar version */
+    XXH_SSE2   = 1,  /*!<
+                      * SSE2 for Pentium 4, Opteron, all x86_64.
+                      *
+                      * @note SSE2 is also guaranteed on Windows 10, macOS, and
+                      * Android x86.
+                      */
+    XXH_AVX2   = 2,  /*!< AVX2 for Haswell and Bulldozer */
+    XXH_AVX512 = 3,  /*!< AVX512 for Skylake and Icelake */
+    XXH_NEON   = 4,  /*!<
+                       * NEON for most ARMv7-A, all AArch64, and WASM SIMD128
+                       * via the SIMDeverywhere polyfill provided with the
+                       * Emscripten SDK.
+                       */
+    XXH_VSX    = 5,  /*!< VSX and ZVector for POWER8/z13 (64-bit) */
+    XXH_SVE    = 6,  /*!< SVE for some ARMv8-A and ARMv9-A */
+};
+/*!
+ * @ingroup tuning
+ * @brief Selects the minimum alignment for XXH3's accumulators.
+ *
+ * When using SIMD, this should match the alignment required for said vector
+ * type, so, for example, 32 for AVX2.
+ *
+ * Default: Auto detected.
+ */
+#  define XXH_ACC_ALIGN 8
+#endif
+
+/* Actual definition */
+#ifndef XXH_DOXYGEN
+#  define XXH_SCALAR 0
+#  define XXH_SSE2   1
+#  define XXH_AVX2   2
+#  define XXH_AVX512 3
+#  define XXH_NEON   4
+#  define XXH_VSX    5
+#  define XXH_SVE    6
+#endif
+
+#ifndef XXH_VECTOR    /* can be defined on command line */
+#  if defined(__ARM_FEATURE_SVE)
+#    define XXH_VECTOR XXH_SVE
+#  elif ( \
+        defined(__ARM_NEON__) || defined(__ARM_NEON) /* gcc */ \
+     || defined(_M_ARM) || defined(_M_ARM64) || defined(_M_ARM64EC) /* msvc */ \
+     || (defined(__wasm_simd128__) && XXH_HAS_INCLUDE(<arm_neon.h>)) /* wasm simd128 via SIMDe */ \
+   ) && ( \
+        defined(_WIN32) || defined(__LITTLE_ENDIAN__) /* little endian only */ \
+    || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) \
+   )
+#    define XXH_VECTOR XXH_NEON
+#  elif defined(__AVX512F__)
+#    define XXH_VECTOR XXH_AVX512
+#  elif defined(__AVX2__)
+#    define XXH_VECTOR XXH_AVX2
+#  elif defined(__SSE2__) || defined(_M_AMD64) || defined(_M_X64) || (defined(_M_IX86_FP) && (_M_IX86_FP == 2))
+#    define XXH_VECTOR XXH_SSE2
+#  elif (defined(__PPC64__) && defined(__POWER8_VECTOR__)) \
+     || (defined(__s390x__) && defined(__VEC__)) \
+     && defined(__GNUC__) /* TODO: IBM XL */
+#    define XXH_VECTOR XXH_VSX
+#  else
+#    define XXH_VECTOR XXH_SCALAR
+#  endif
+#endif
+
+/* __ARM_FEATURE_SVE is only supported by GCC & Clang. */
+#if (XXH_VECTOR == XXH_SVE) && !defined(__ARM_FEATURE_SVE)
+#  ifdef _MSC_VER
+#    pragma warning(once : 4606)
+#  else
+#    warning "__ARM_FEATURE_SVE isn't supported. Use SCALAR instead."
+#  endif
+#  undef XXH_VECTOR
+#  define XXH_VECTOR XXH_SCALAR
+#endif
+
+/*
+ * Controls the alignment of the accumulator,
+ * for compatibility with aligned vector loads, which are usually faster.
+ */
+#ifndef XXH_ACC_ALIGN
+#  if defined(XXH_X86DISPATCH)
+#     define XXH_ACC_ALIGN 64  /* for compatibility with avx512 */
+#  elif XXH_VECTOR == XXH_SCALAR  /* scalar */
+#     define XXH_ACC_ALIGN 8
+#  elif XXH_VECTOR == XXH_SSE2  /* sse2 */
+#     define XXH_ACC_ALIGN 16
+#  elif XXH_VECTOR == XXH_AVX2  /* avx2 */
+#     define XXH_ACC_ALIGN 32
+#  elif XXH_VECTOR == XXH_NEON  /* neon */
+#     define XXH_ACC_ALIGN 16
+#  elif XXH_VECTOR == XXH_VSX   /* vsx */
+#     define XXH_ACC_ALIGN 16
+#  elif XXH_VECTOR == XXH_AVX512  /* avx512 */
+#     define XXH_ACC_ALIGN 64
+#  elif XXH_VECTOR == XXH_SVE   /* sve */
+#     define XXH_ACC_ALIGN 64
+#  endif
+#endif
+
+#if defined(XXH_X86DISPATCH) || XXH_VECTOR == XXH_SSE2 \
+    || XXH_VECTOR == XXH_AVX2 || XXH_VECTOR == XXH_AVX512
+#  define XXH_SEC_ALIGN XXH_ACC_ALIGN
+#elif XXH_VECTOR == XXH_SVE
+#  define XXH_SEC_ALIGN XXH_ACC_ALIGN
+#else
+#  define XXH_SEC_ALIGN 8
+#endif
+
+#if defined(__GNUC__) || defined(__clang__)
+#  define XXH_ALIASING __attribute__((may_alias))
+#else
+#  define XXH_ALIASING /* nothing */
+#endif
+
+/*
+ * UGLY HACK:
+ * GCC usually generates the best code with -O3 for xxHash.
+ *
+ * However, when targeting AVX2, it is overzealous in its unrolling resulting
+ * in code roughly 3/4 the speed of Clang.
+ *
+ * There are other issues, such as GCC splitting _mm256_loadu_si256 into
+ * _mm_loadu_si128 + _mm256_inserti128_si256. This is an optimization which
+ * only applies to Sandy and Ivy Bridge... which don't even support AVX2.
+ *
+ * That is why when compiling the AVX2 version, it is recommended to use either
+ *   -O2 -mavx2 -march=haswell
+ * or
+ *   -O2 -mavx2 -mno-avx256-split-unaligned-load
+ * for decent performance, or to use Clang instead.
+ *
+ * Fortunately, we can control the first one with a pragma that forces GCC into
+ * -O2, but the other one we can't control without "failed to inline always
+ * inline function due to target mismatch" warnings.
+ */
+#if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \
+  && defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
+  && defined(__OPTIMIZE__) && XXH_SIZE_OPT <= 0 /* respect -O0 and -Os */
+#  pragma GCC push_options
+#  pragma GCC optimize("-O2")
+#endif
+
+#if XXH_VECTOR == XXH_NEON
+
+/*
+ * UGLY HACK: While AArch64 GCC on Linux does not seem to care, on macOS, GCC -O3
+ * optimizes out the entire hashLong loop because of the aliasing violation.
+ *
+ * However, GCC is also inefficient at load-store optimization with vld1q/vst1q,
+ * so the only option is to mark it as aliasing.
+ */
+typedef uint64x2_t xxh_aliasing_uint64x2_t XXH_ALIASING;
+
+/*!
+ * @internal
+ * @brief `vld1q_u64` but faster and alignment-safe.
+ *
+ * On AArch64, unaligned access is always safe, but on ARMv7-a, it is only
+ * *conditionally* safe (`vld1` has an alignment bit like `movdq[ua]` in x86).
+ *
+ * GCC for AArch64 sees `vld1q_u8` as an intrinsic instead of a load, so it
+ * prohibits load-store optimizations. Therefore, a direct dereference is used.
+ *
+ * Otherwise, `vld1q_u8` is used with `vreinterpretq_u8_u64` to do a safe
+ * unaligned load.
+ */
+#if defined(__aarch64__) && defined(__GNUC__) && !defined(__clang__)
+XXH_FORCE_INLINE uint64x2_t XXH_vld1q_u64(void const* ptr) /* silence -Wcast-align */
+{
+    return *(xxh_aliasing_uint64x2_t const *)ptr;
+}
+#else
+XXH_FORCE_INLINE uint64x2_t XXH_vld1q_u64(void const* ptr)
+{
+    return vreinterpretq_u64_u8(vld1q_u8((uint8_t const*)ptr));
+}
+#endif
+
+/*!
+ * @internal
+ * @brief `vmlal_u32` on low and high halves of a vector.
+ *
+ * This is a workaround for AArch64 GCC < 11 which implemented arm_neon.h with
+ * inline assembly and were therefore incapable of merging the `vget_{low, high}_u32`
+ * with `vmlal_u32`.
+ */
+#if defined(__aarch64__) && defined(__GNUC__) && !defined(__clang__) && __GNUC__ < 11
+XXH_FORCE_INLINE uint64x2_t
+XXH_vmlal_low_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
+{
+    /* Inline assembly is the only way */
+    __asm__("umlal   %0.2d, %1.2s, %2.2s" : "+w" (acc) : "w" (lhs), "w" (rhs));
+    return acc;
+}
+XXH_FORCE_INLINE uint64x2_t
+XXH_vmlal_high_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
+{
+    /* This intrinsic works as expected */
+    return vmlal_high_u32(acc, lhs, rhs);
+}
+#else
+/* Portable intrinsic versions */
+XXH_FORCE_INLINE uint64x2_t
+XXH_vmlal_low_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
+{
+    return vmlal_u32(acc, vget_low_u32(lhs), vget_low_u32(rhs));
+}
+/*! @copydoc XXH_vmlal_low_u32
+ * Assume the compiler converts this to vmlal_high_u32 on aarch64 */
+XXH_FORCE_INLINE uint64x2_t
+XXH_vmlal_high_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
+{
+    return vmlal_u32(acc, vget_high_u32(lhs), vget_high_u32(rhs));
+}
+#endif
+
+/*!
+ * @ingroup tuning
+ * @brief Controls the NEON to scalar ratio for XXH3
+ *
+ * This can be set to 2, 4, 6, or 8.
+ *
+ * ARM Cortex CPUs are _very_ sensitive to how their pipelines are used.
+ *
+ * For example, the Cortex-A73 can dispatch 3 micro-ops per cycle, but only 2 of those
+ * can be NEON. If you are only using NEON instructions, you are only using 2/3 of the CPU
+ * bandwidth.
+ *
+ * This is even more noticeable on the more advanced cores like the Cortex-A76 which
+ * can dispatch 8 micro-ops per cycle, but still only 2 NEON micro-ops at once.
+ *
+ * Therefore, to make the most out of the pipeline, it is beneficial to run 6 NEON lanes
+ * and 2 scalar lanes, which is chosen by default.
+ *
+ * This does not apply to Apple processors or 32-bit processors, which run better with
+ * full NEON. These will default to 8. Additionally, size-optimized builds run 8 lanes.
+ *
+ * This change benefits CPUs with large micro-op buffers without negatively affecting
+ * most other CPUs:
+ *
+ *  | Chipset               | Dispatch type       | NEON only | 6:2 hybrid | Diff. |
+ *  |:----------------------|:--------------------|----------:|-----------:|------:|
+ *  | Snapdragon 730 (A76)  | 2 NEON/8 micro-ops  |  8.8 GB/s |  10.1 GB/s |  ~16% |
+ *  | Snapdragon 835 (A73)  | 2 NEON/3 micro-ops  |  5.1 GB/s |   5.3 GB/s |   ~5% |
+ *  | Marvell PXA1928 (A53) | In-order dual-issue |  1.9 GB/s |   1.9 GB/s |    0% |
+ *  | Apple M1              | 4 NEON/8 micro-ops  | 37.3 GB/s |  36.1 GB/s |  ~-3% |
+ *
+ * It also seems to fix some bad codegen on GCC, making it almost as fast as clang.
+ *
+ * When using WASM SIMD128, if this is 2 or 6, SIMDe will scalarize 2 of the lanes meaning
+ * it effectively becomes worse 4.
+ *
+ * @see XXH3_accumulate_512_neon()
+ */
+# ifndef XXH3_NEON_LANES
+#  if (defined(__aarch64__) || defined(__arm64__) || defined(_M_ARM64) || defined(_M_ARM64EC)) \
+   && !defined(__APPLE__) && XXH_SIZE_OPT <= 0
+#   define XXH3_NEON_LANES 6
+#  else
+#   define XXH3_NEON_LANES XXH_ACC_NB
+#  endif
+# endif
+#endif  /* XXH_VECTOR == XXH_NEON */
+
+/*
+ * VSX and Z Vector helpers.
+ *
+ * This is very messy, and any pull requests to clean this up are welcome.
+ *
+ * There are a lot of problems with supporting VSX and s390x, due to
+ * inconsistent intrinsics, spotty coverage, and multiple endiannesses.
+ */
+#if XXH_VECTOR == XXH_VSX
+/* Annoyingly, these headers _may_ define three macros: `bool`, `vector`,
+ * and `pixel`. This is a problem for obvious reasons.
+ *
+ * These keywords are unnecessary; the spec literally says they are
+ * equivalent to `__bool`, `__vector`, and `__pixel` and may be undef'd
+ * after including the header.
+ *
+ * We use pragma push_macro/pop_macro to keep the namespace clean. */
+#  pragma push_macro("bool")
+#  pragma push_macro("vector")
+#  pragma push_macro("pixel")
+/* silence potential macro redefined warnings */
+#  undef bool
+#  undef vector
+#  undef pixel
+
+#  if defined(__s390x__)
+#    include <s390intrin.h>
+#  else
+#    include <altivec.h>
+#  endif
+
+/* Restore the original macro values, if applicable. */
+#  pragma pop_macro("pixel")
+#  pragma pop_macro("vector")
+#  pragma pop_macro("bool")
+
+typedef __vector unsigned long long xxh_u64x2;
+typedef __vector unsigned char xxh_u8x16;
+typedef __vector unsigned xxh_u32x4;
+
+/*
+ * UGLY HACK: Similar to aarch64 macOS GCC, s390x GCC has the same aliasing issue.
+ */
+typedef xxh_u64x2 xxh_aliasing_u64x2 XXH_ALIASING;
+
+# ifndef XXH_VSX_BE
+#  if defined(__BIG_ENDIAN__) \
+  || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
+#    define XXH_VSX_BE 1
+#  elif defined(__VEC_ELEMENT_REG_ORDER__) && __VEC_ELEMENT_REG_ORDER__ == __ORDER_BIG_ENDIAN__
+#    warning "-maltivec=be is not recommended. Please use native endianness."
+#    define XXH_VSX_BE 1
+#  else
+#    define XXH_VSX_BE 0
+#  endif
+# endif /* !defined(XXH_VSX_BE) */
+
+# if XXH_VSX_BE
+#  if defined(__POWER9_VECTOR__) || (defined(__clang__) && defined(__s390x__))
+#    define XXH_vec_revb vec_revb
+#  else
+/*!
+ * A polyfill for POWER9's vec_revb().
+ */
+XXH_FORCE_INLINE xxh_u64x2 XXH_vec_revb(xxh_u64x2 val)
+{
+    xxh_u8x16 const vByteSwap = { 0x07, 0x06, 0x05, 0x04, 0x03, 0x02, 0x01, 0x00,
+                                  0x0F, 0x0E, 0x0D, 0x0C, 0x0B, 0x0A, 0x09, 0x08 };
+    return vec_perm(val, val, vByteSwap);
+}
+#  endif
+# endif /* XXH_VSX_BE */
+
+/*!
+ * Performs an unaligned vector load and byte swaps it on big endian.
+ */
+XXH_FORCE_INLINE xxh_u64x2 XXH_vec_loadu(const void *ptr)
+{
+    xxh_u64x2 ret;
+    XXH_memcpy(&ret, ptr, sizeof(xxh_u64x2));
+# if XXH_VSX_BE
+    ret = XXH_vec_revb(ret);
+# endif
+    return ret;
+}
+
+/*
+ * vec_mulo and vec_mule are very problematic intrinsics on PowerPC
+ *
+ * These intrinsics weren't added until GCC 8, despite existing for a while,
+ * and they are endian dependent. Also, their meaning swap depending on version.
+ * */
+# if defined(__s390x__)
+ /* s390x is always big endian, no issue on this platform */
+#  define XXH_vec_mulo vec_mulo
+#  define XXH_vec_mule vec_mule
+# elif defined(__clang__) && XXH_HAS_BUILTIN(__builtin_altivec_vmuleuw) && !defined(__ibmxl__)
+/* Clang has a better way to control this, we can just use the builtin which doesn't swap. */
+ /* The IBM XL Compiler (which defined __clang__) only implements the vec_* operations */
+#  define XXH_vec_mulo __builtin_altivec_vmulouw
+#  define XXH_vec_mule __builtin_altivec_vmuleuw
+# else
+/* gcc needs inline assembly */
+/* Adapted from https://github.com/google/highwayhash/blob/master/highwayhash/hh_vsx.h. */
+XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mulo(xxh_u32x4 a, xxh_u32x4 b)
+{
+    xxh_u64x2 result;
+    __asm__("vmulouw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b));
+    return result;
+}
+XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mule(xxh_u32x4 a, xxh_u32x4 b)
+{
+    xxh_u64x2 result;
+    __asm__("vmuleuw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b));
+    return result;
+}
+# endif /* XXH_vec_mulo, XXH_vec_mule */
+#endif /* XXH_VECTOR == XXH_VSX */
+
+#if XXH_VECTOR == XXH_SVE
+#define ACCRND(acc, offset) \
+do { \
+    svuint64_t input_vec = svld1_u64(mask, xinput + offset);         \
+    svuint64_t secret_vec = svld1_u64(mask, xsecret + offset);       \
+    svuint64_t mixed = sveor_u64_x(mask, secret_vec, input_vec);     \
+    svuint64_t swapped = svtbl_u64(input_vec, kSwap);                \
+    svuint64_t mixed_lo = svextw_u64_x(mask, mixed);                 \
+    svuint64_t mixed_hi = svlsr_n_u64_x(mask, mixed, 32);            \
+    svuint64_t mul = svmad_u64_x(mask, mixed_lo, mixed_hi, swapped); \
+    acc = svadd_u64_x(mask, acc, mul);                               \
+} while (0)
+#endif /* XXH_VECTOR == XXH_SVE */
+
+/* prefetch
+ * can be disabled, by declaring XXH_NO_PREFETCH build macro */
+#if defined(XXH_NO_PREFETCH)
+#  define XXH_PREFETCH(ptr)  (void)(ptr)  /* disabled */
+#else
+#  if XXH_SIZE_OPT >= 1
+#    define XXH_PREFETCH(ptr) (void)(ptr)
+#  elif defined(_MSC_VER) && (defined(_M_X64) || defined(_M_IX86))  /* _mm_prefetch() not defined outside of x86/x64 */
+#    include <mmintrin.h>   /* https://msdn.microsoft.com/fr-fr/library/84szxsww(v=vs.90).aspx */
+#    define XXH_PREFETCH(ptr)  _mm_prefetch((const char*)(ptr), _MM_HINT_T0)
+#  elif defined(__GNUC__) && ( (__GNUC__ >= 4) || ( (__GNUC__ == 3) && (__GNUC_MINOR__ >= 1) ) )
+#    define XXH_PREFETCH(ptr)  __builtin_prefetch((ptr), 0 /* rw==read */, 3 /* locality */)
+#  else
+#    define XXH_PREFETCH(ptr) (void)(ptr)  /* disabled */
+#  endif
+#endif  /* XXH_NO_PREFETCH */
+
+
+/* ==========================================
+ * XXH3 default settings
+ * ========================================== */
+
+#define XXH_SECRET_DEFAULT_SIZE 192   /* minimum XXH3_SECRET_SIZE_MIN */
+
+#if (XXH_SECRET_DEFAULT_SIZE < XXH3_SECRET_SIZE_MIN)
+#  error "default keyset is not large enough"
+#endif
+
+/*! Pseudorandom secret taken directly from FARSH. */
+XXH_ALIGN(64) static const xxh_u8 XXH3_kSecret[XXH_SECRET_DEFAULT_SIZE] = {
+    0xb8, 0xfe, 0x6c, 0x39, 0x23, 0xa4, 0x4b, 0xbe, 0x7c, 0x01, 0x81, 0x2c, 0xf7, 0x21, 0xad, 0x1c,
+    0xde, 0xd4, 0x6d, 0xe9, 0x83, 0x90, 0x97, 0xdb, 0x72, 0x40, 0xa4, 0xa4, 0xb7, 0xb3, 0x67, 0x1f,
+    0xcb, 0x79, 0xe6, 0x4e, 0xcc, 0xc0, 0xe5, 0x78, 0x82, 0x5a, 0xd0, 0x7d, 0xcc, 0xff, 0x72, 0x21,
+    0xb8, 0x08, 0x46, 0x74, 0xf7, 0x43, 0x24, 0x8e, 0xe0, 0x35, 0x90, 0xe6, 0x81, 0x3a, 0x26, 0x4c,
+    0x3c, 0x28, 0x52, 0xbb, 0x91, 0xc3, 0x00, 0xcb, 0x88, 0xd0, 0x65, 0x8b, 0x1b, 0x53, 0x2e, 0xa3,
+    0x71, 0x64, 0x48, 0x97, 0xa2, 0x0d, 0xf9, 0x4e, 0x38, 0x19, 0xef, 0x46, 0xa9, 0xde, 0xac, 0xd8,
+    0xa8, 0xfa, 0x76, 0x3f, 0xe3, 0x9c, 0x34, 0x3f, 0xf9, 0xdc, 0xbb, 0xc7, 0xc7, 0x0b, 0x4f, 0x1d,
+    0x8a, 0x51, 0xe0, 0x4b, 0xcd, 0xb4, 0x59, 0x31, 0xc8, 0x9f, 0x7e, 0xc9, 0xd9, 0x78, 0x73, 0x64,
+    0xea, 0xc5, 0xac, 0x83, 0x34, 0xd3, 0xeb, 0xc3, 0xc5, 0x81, 0xa0, 0xff, 0xfa, 0x13, 0x63, 0xeb,
+    0x17, 0x0d, 0xdd, 0x51, 0xb7, 0xf0, 0xda, 0x49, 0xd3, 0x16, 0x55, 0x26, 0x29, 0xd4, 0x68, 0x9e,
+    0x2b, 0x16, 0xbe, 0x58, 0x7d, 0x47, 0xa1, 0xfc, 0x8f, 0xf8, 0xb8, 0xd1, 0x7a, 0xd0, 0x31, 0xce,
+    0x45, 0xcb, 0x3a, 0x8f, 0x95, 0x16, 0x04, 0x28, 0xaf, 0xd7, 0xfb, 0xca, 0xbb, 0x4b, 0x40, 0x7e,
+};
+
+static const xxh_u64 PRIME_MX1 = 0x165667919E3779F9ULL;  /*!< 0b0001011001010110011001111001000110011110001101110111100111111001 */
+static const xxh_u64 PRIME_MX2 = 0x9FB21C651E98DF25ULL;  /*!< 0b1001111110110010000111000110010100011110100110001101111100100101 */
+
+#ifdef XXH_OLD_NAMES
+#  define kSecret XXH3_kSecret
+#endif
+
+#ifdef XXH_DOXYGEN
+/*!
+ * @brief Calculates a 32-bit to 64-bit long multiply.
+ *
+ * Implemented as a macro.
+ *
+ * Wraps `__emulu` on MSVC x86 because it tends to call `__allmul` when it doesn't
+ * need to (but it shouldn't need to anyways, it is about 7 instructions to do
+ * a 64x64 multiply...). Since we know that this will _always_ emit `MULL`, we
+ * use that instead of the normal method.
+ *
+ * If you are compiling for platforms like Thumb-1 and don't have a better option,
+ * you may also want to write your own long multiply routine here.
+ *
+ * @param x, y Numbers to be multiplied
+ * @return 64-bit product of the low 32 bits of @p x and @p y.
+ */
+XXH_FORCE_INLINE xxh_u64
+XXH_mult32to64(xxh_u64 x, xxh_u64 y)
+{
+   return (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF);
+}
+#elif defined(_MSC_VER) && defined(_M_IX86)
+#    define XXH_mult32to64(x, y) __emulu((unsigned)(x), (unsigned)(y))
+#else
+/*
+ * Downcast + upcast is usually better than masking on older compilers like
+ * GCC 4.2 (especially 32-bit ones), all without affecting newer compilers.
+ *
+ * The other method, (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF), will AND both operands
+ * and perform a full 64x64 multiply -- entirely redundant on 32-bit.
+ */
+#    define XXH_mult32to64(x, y) ((xxh_u64)(xxh_u32)(x) * (xxh_u64)(xxh_u32)(y))
+#endif
+
+/*!
+ * @brief Calculates a 64->128-bit long multiply.
+ *
+ * Uses `__uint128_t` and `_umul128` if available, otherwise uses a scalar
+ * version.
+ *
+ * @param lhs , rhs The 64-bit integers to be multiplied
+ * @return The 128-bit result represented in an @ref XXH128_hash_t.
+ */
+static XXH128_hash_t
+XXH_mult64to128(xxh_u64 lhs, xxh_u64 rhs)
+{
+    /*
+     * GCC/Clang __uint128_t method.
+     *
+     * On most 64-bit targets, GCC and Clang define a __uint128_t type.
+     * This is usually the best way as it usually uses a native long 64-bit
+     * multiply, such as MULQ on x86_64 or MUL + UMULH on aarch64.
+     *
+     * Usually.
+     *
+     * Despite being a 32-bit platform, Clang (and emscripten) define this type
+     * despite not having the arithmetic for it. This results in a laggy
+     * compiler builtin call which calculates a full 128-bit multiply.
+     * In that case it is best to use the portable one.
+     * https://github.com/Cyan4973/xxHash/issues/211#issuecomment-515575677
+     */
+#if (defined(__GNUC__) || defined(__clang__)) && !defined(__wasm__) \
+    && defined(__SIZEOF_INT128__) \
+    || (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128)
+
+    __uint128_t const product = (__uint128_t)lhs * (__uint128_t)rhs;
+    XXH128_hash_t r128;
+    r128.low64  = (xxh_u64)(product);
+    r128.high64 = (xxh_u64)(product >> 64);
+    return r128;
+
+    /*
+     * MSVC for x64's _umul128 method.
+     *
+     * xxh_u64 _umul128(xxh_u64 Multiplier, xxh_u64 Multiplicand, xxh_u64 *HighProduct);
+     *
+     * This compiles to single operand MUL on x64.
+     */
+#elif (defined(_M_X64) || defined(_M_IA64)) && !defined(_M_ARM64EC)
+
+#ifndef _MSC_VER
+#   pragma intrinsic(_umul128)
+#endif
+    xxh_u64 product_high;
+    xxh_u64 const product_low = _umul128(lhs, rhs, &product_high);
+    XXH128_hash_t r128;
+    r128.low64  = product_low;
+    r128.high64 = product_high;
+    return r128;
+
+    /*
+     * MSVC for ARM64's __umulh method.
+     *
+     * This compiles to the same MUL + UMULH as GCC/Clang's __uint128_t method.
+     */
+#elif defined(_M_ARM64) || defined(_M_ARM64EC)
+
+#ifndef _MSC_VER
+#   pragma intrinsic(__umulh)
+#endif
+    XXH128_hash_t r128;
+    r128.low64  = lhs * rhs;
+    r128.high64 = __umulh(lhs, rhs);
+    return r128;
+
+#else
+    /*
+     * Portable scalar method. Optimized for 32-bit and 64-bit ALUs.
+     *
+     * This is a fast and simple grade school multiply, which is shown below
+     * with base 10 arithmetic instead of base 0x100000000.
+     *
+     *           9 3 // D2 lhs = 93
+     *         x 7 5 // D2 rhs = 75
+     *     ----------
+     *           1 5 // D2 lo_lo = (93 % 10) * (75 % 10) = 15
+     *         4 5 | // D2 hi_lo = (93 / 10) * (75 % 10) = 45
+     *         2 1 | // D2 lo_hi = (93 % 10) * (75 / 10) = 21
+     *     + 6 3 | | // D2 hi_hi = (93 / 10) * (75 / 10) = 63
+     *     ---------
+     *         2 7 | // D2 cross = (15 / 10) + (45 % 10) + 21 = 27
+     *     + 6 7 | | // D2 upper = (27 / 10) + (45 / 10) + 63 = 67
+     *     ---------
+     *       6 9 7 5 // D4 res = (27 * 10) + (15 % 10) + (67 * 100) = 6975
+     *
+     * The reasons for adding the products like this are:
+     *  1. It avoids manual carry tracking. Just like how
+     *     (9 * 9) + 9 + 9 = 99, the same applies with this for UINT64_MAX.
+     *     This avoids a lot of complexity.
+     *
+     *  2. It hints for, and on Clang, compiles to, the powerful UMAAL
+     *     instruction available in ARM's Digital Signal Processing extension
+     *     in 32-bit ARMv6 and later, which is shown below:
+     *
+     *         void UMAAL(xxh_u32 *RdLo, xxh_u32 *RdHi, xxh_u32 Rn, xxh_u32 Rm)
+     *         {
+     *             xxh_u64 product = (xxh_u64)*RdLo * (xxh_u64)*RdHi + Rn + Rm;
+     *             *RdLo = (xxh_u32)(product & 0xFFFFFFFF);
+     *             *RdHi = (xxh_u32)(product >> 32);
+     *         }
+     *
+     *     This instruction was designed for efficient long multiplication, and
+     *     allows this to be calculated in only 4 instructions at speeds
+     *     comparable to some 64-bit ALUs.
+     *
+     *  3. It isn't terrible on other platforms. Usually this will be a couple
+     *     of 32-bit ADD/ADCs.
+     */
+
+    /* First calculate all of the cross products. */
+    xxh_u64 const lo_lo = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs & 0xFFFFFFFF);
+    xxh_u64 const hi_lo = XXH_mult32to64(lhs >> 32,        rhs & 0xFFFFFFFF);
+    xxh_u64 const lo_hi = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs >> 32);
+    xxh_u64 const hi_hi = XXH_mult32to64(lhs >> 32,        rhs >> 32);
+
+    /* Now add the products together. These will never overflow. */
+    xxh_u64 const cross = (lo_lo >> 32) + (hi_lo & 0xFFFFFFFF) + lo_hi;
+    xxh_u64 const upper = (hi_lo >> 32) + (cross >> 32)        + hi_hi;
+    xxh_u64 const lower = (cross << 32) | (lo_lo & 0xFFFFFFFF);
+
+    XXH128_hash_t r128;
+    r128.low64  = lower;
+    r128.high64 = upper;
+    return r128;
+#endif
+}
+
+/*!
+ * @brief Calculates a 64-bit to 128-bit multiply, then XOR folds it.
+ *
+ * The reason for the separate function is to prevent passing too many structs
+ * around by value. This will hopefully inline the multiply, but we don't force it.
+ *
+ * @param lhs , rhs The 64-bit integers to multiply
+ * @return The low 64 bits of the product XOR'd by the high 64 bits.
+ * @see XXH_mult64to128()
+ */
+static xxh_u64
+XXH3_mul128_fold64(xxh_u64 lhs, xxh_u64 rhs)
+{
+    XXH128_hash_t product = XXH_mult64to128(lhs, rhs);
+    return product.low64 ^ product.high64;
+}
+
+/*! Seems to produce slightly better code on GCC for some reason. */
+XXH_FORCE_INLINE XXH_CONSTF xxh_u64 XXH_xorshift64(xxh_u64 v64, int shift)
+{
+    XXH_ASSERT(0 <= shift && shift < 64);
+    return v64 ^ (v64 >> shift);
+}
+
+/*
+ * This is a fast avalanche stage,
+ * suitable when input bits are already partially mixed
+ */
+static XXH64_hash_t XXH3_avalanche(xxh_u64 h64)
+{
+    h64 = XXH_xorshift64(h64, 37);
+    h64 *= PRIME_MX1;
+    h64 = XXH_xorshift64(h64, 32);
+    return h64;
+}
+
+/*
+ * This is a stronger avalanche,
+ * inspired by Pelle Evensen's rrmxmx
+ * preferable when input has not been previously mixed
+ */
+static XXH64_hash_t XXH3_rrmxmx(xxh_u64 h64, xxh_u64 len)
+{
+    /* this mix is inspired by Pelle Evensen's rrmxmx */
+    h64 ^= XXH_rotl64(h64, 49) ^ XXH_rotl64(h64, 24);
+    h64 *= PRIME_MX2;
+    h64 ^= (h64 >> 35) + len ;
+    h64 *= PRIME_MX2;
+    return XXH_xorshift64(h64, 28);
+}
+
+
+/* ==========================================
+ * Short keys
+ * ==========================================
+ * One of the shortcomings of XXH32 and XXH64 was that their performance was
+ * sub-optimal on short lengths. It used an iterative algorithm which strongly
+ * favored lengths that were a multiple of 4 or 8.
+ *
+ * Instead of iterating over individual inputs, we use a set of single shot
+ * functions which piece together a range of lengths and operate in constant time.
+ *
+ * Additionally, the number of multiplies has been significantly reduced. This
+ * reduces latency, especially when emulating 64-bit multiplies on 32-bit.
+ *
+ * Depending on the platform, this may or may not be faster than XXH32, but it
+ * is almost guaranteed to be faster than XXH64.
+ */
+
+/*
+ * At very short lengths, there isn't enough input to fully hide secrets, or use
+ * the entire secret.
+ *
+ * There is also only a limited amount of mixing we can do before significantly
+ * impacting performance.
+ *
+ * Therefore, we use different sections of the secret and always mix two secret
+ * samples with an XOR. This should have no effect on performance on the
+ * seedless or withSeed variants because everything _should_ be constant folded
+ * by modern compilers.
+ *
+ * The XOR mixing hides individual parts of the secret and increases entropy.
+ *
+ * This adds an extra layer of strength for custom secrets.
+ */
+XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
+XXH3_len_1to3_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+    XXH_ASSERT(input != NULL);
+    XXH_ASSERT(1 <= len && len <= 3);
+    XXH_ASSERT(secret != NULL);
+    /*
+     * len = 1: combined = { input[0], 0x01, input[0], input[0] }
+     * len = 2: combined = { input[1], 0x02, input[0], input[1] }
+     * len = 3: combined = { input[2], 0x03, input[0], input[1] }
+     */
+    {   xxh_u8  const c1 = input[0];
+        xxh_u8  const c2 = input[len >> 1];
+        xxh_u8  const c3 = input[len - 1];
+        xxh_u32 const combined = ((xxh_u32)c1 << 16) | ((xxh_u32)c2  << 24)
+                               | ((xxh_u32)c3 <<  0) | ((xxh_u32)len << 8);
+        xxh_u64 const bitflip = (XXH_readLE32(secret) ^ XXH_readLE32(secret+4)) + seed;
+        xxh_u64 const keyed = (xxh_u64)combined ^ bitflip;
+        return XXH64_avalanche(keyed);
+    }
+}
+
+XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
+XXH3_len_4to8_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+    XXH_ASSERT(input != NULL);
+    XXH_ASSERT(secret != NULL);
+    XXH_ASSERT(4 <= len && len <= 8);
+    seed ^= (xxh_u64)XXH_swap32((xxh_u32)seed) << 32;
+    {   xxh_u32 const input1 = XXH_readLE32(input);
+        xxh_u32 const input2 = XXH_readLE32(input + len - 4);
+        xxh_u64 const bitflip = (XXH_readLE64(secret+8) ^ XXH_readLE64(secret+16)) - seed;
+        xxh_u64 const input64 = input2 + (((xxh_u64)input1) << 32);
+        xxh_u64 const keyed = input64 ^ bitflip;
+        return XXH3_rrmxmx(keyed, len);
+    }
+}
+
+XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
+XXH3_len_9to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+    XXH_ASSERT(input != NULL);
+    XXH_ASSERT(secret != NULL);
+    XXH_ASSERT(9 <= len && len <= 16);
+    {   xxh_u64 const bitflip1 = (XXH_readLE64(secret+24) ^ XXH_readLE64(secret+32)) + seed;
+        xxh_u64 const bitflip2 = (XXH_readLE64(secret+40) ^ XXH_readLE64(secret+48)) - seed;
+        xxh_u64 const input_lo = XXH_readLE64(input)           ^ bitflip1;
+        xxh_u64 const input_hi = XXH_readLE64(input + len - 8) ^ bitflip2;
+        xxh_u64 const acc = len
+                          + XXH_swap64(input_lo) + input_hi
+                          + XXH3_mul128_fold64(input_lo, input_hi);
+        return XXH3_avalanche(acc);
+    }
+}
+
+XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
+XXH3_len_0to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+    XXH_ASSERT(len <= 16);
+    {   if (XXH_likely(len >  8)) return XXH3_len_9to16_64b(input, len, secret, seed);
+        if (XXH_likely(len >= 4)) return XXH3_len_4to8_64b(input, len, secret, seed);
+        if (len) return XXH3_len_1to3_64b(input, len, secret, seed);
+        return XXH64_avalanche(seed ^ (XXH_readLE64(secret+56) ^ XXH_readLE64(secret+64)));
+    }
+}
+
+/*
+ * DISCLAIMER: There are known *seed-dependent* multicollisions here due to
+ * multiplication by zero, affecting hashes of lengths 17 to 240.
+ *
+ * However, they are very unlikely.
+ *
+ * Keep this in mind when using the unseeded XXH3_64bits() variant: As with all
+ * unseeded non-cryptographic hashes, it does not attempt to defend itself
+ * against specially crafted inputs, only random inputs.
+ *
+ * Compared to classic UMAC where a 1 in 2^31 chance of 4 consecutive bytes
+ * cancelling out the secret is taken an arbitrary number of times (addressed
+ * in XXH3_accumulate_512), this collision is very unlikely with random inputs
+ * and/or proper seeding:
+ *
+ * This only has a 1 in 2^63 chance of 8 consecutive bytes cancelling out, in a
+ * function that is only called up to 16 times per hash with up to 240 bytes of
+ * input.
+ *
+ * This is not too bad for a non-cryptographic hash function, especially with
+ * only 64 bit outputs.
+ *
+ * The 128-bit variant (which trades some speed for strength) is NOT affected
+ * by this, although it is always a good idea to use a proper seed if you care
+ * about strength.
+ */
+XXH_FORCE_INLINE xxh_u64 XXH3_mix16B(const xxh_u8* XXH_RESTRICT input,
+                                     const xxh_u8* XXH_RESTRICT secret, xxh_u64 seed64)
+{
+#if defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
+  && defined(__i386__) && defined(__SSE2__)  /* x86 + SSE2 */ \
+  && !defined(XXH_ENABLE_AUTOVECTORIZE)      /* Define to disable like XXH32 hack */
+    /*
+     * UGLY HACK:
+     * GCC for x86 tends to autovectorize the 128-bit multiply, resulting in
+     * slower code.
+     *
+     * By forcing seed64 into a register, we disrupt the cost model and
+     * cause it to scalarize. See `XXH32_round()`
+     *
+     * FIXME: Clang's output is still _much_ faster -- On an AMD Ryzen 3600,
+     * XXH3_64bits @ len=240 runs at 4.6 GB/s with Clang 9, but 3.3 GB/s on
+     * GCC 9.2, despite both emitting scalar code.
+     *
+     * GCC generates much better scalar code than Clang for the rest of XXH3,
+     * which is why finding a more optimal codepath is an interest.
+     */
+    XXH_COMPILER_GUARD(seed64);
+#endif
+    {   xxh_u64 const input_lo = XXH_readLE64(input);
+        xxh_u64 const input_hi = XXH_readLE64(input+8);
+        return XXH3_mul128_fold64(
+            input_lo ^ (XXH_readLE64(secret)   + seed64),
+            input_hi ^ (XXH_readLE64(secret+8) - seed64)
+        );
+    }
+}
+
+/* For mid range keys, XXH3 uses a Mum-hash variant. */
+XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
+XXH3_len_17to128_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
+                     const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+                     XXH64_hash_t seed)
+{
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
+    XXH_ASSERT(16 < len && len <= 128);
+
+    {   xxh_u64 acc = len * XXH_PRIME64_1;
+#if XXH_SIZE_OPT >= 1
+        /* Smaller and cleaner, but slightly slower. */
+        unsigned int i = (unsigned int)(len - 1) / 32;
+        do {
+            acc += XXH3_mix16B(input+16 * i, secret+32*i, seed);
+            acc += XXH3_mix16B(input+len-16*(i+1), secret+32*i+16, seed);
+        } while (i-- != 0);
+#else
+        if (len > 32) {
+            if (len > 64) {
+                if (len > 96) {
+                    acc += XXH3_mix16B(input+48, secret+96, seed);
+                    acc += XXH3_mix16B(input+len-64, secret+112, seed);
+                }
+                acc += XXH3_mix16B(input+32, secret+64, seed);
+                acc += XXH3_mix16B(input+len-48, secret+80, seed);
+            }
+            acc += XXH3_mix16B(input+16, secret+32, seed);
+            acc += XXH3_mix16B(input+len-32, secret+48, seed);
+        }
+        acc += XXH3_mix16B(input+0, secret+0, seed);
+        acc += XXH3_mix16B(input+len-16, secret+16, seed);
+#endif
+        return XXH3_avalanche(acc);
+    }
+}
+
+#define XXH3_MIDSIZE_MAX 240
+
+XXH_NO_INLINE XXH_PUREF XXH64_hash_t
+XXH3_len_129to240_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
+                      const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+                      XXH64_hash_t seed)
+{
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
+    XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX);
+
+    #define XXH3_MIDSIZE_STARTOFFSET 3
+    #define XXH3_MIDSIZE_LASTOFFSET  17
+
+    {   xxh_u64 acc = len * XXH_PRIME64_1;
+        xxh_u64 acc_end;
+        unsigned int const nbRounds = (unsigned int)len / 16;
+        unsigned int i;
+        XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX);
+        for (i=0; i<8; i++) {
+            acc += XXH3_mix16B(input+(16*i), secret+(16*i), seed);
+        }
+        /* last bytes */
+        acc_end = XXH3_mix16B(input + len - 16, secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET, seed);
+        XXH_ASSERT(nbRounds >= 8);
+        acc = XXH3_avalanche(acc);
+#if defined(__clang__)                                /* Clang */ \
+    && (defined(__ARM_NEON) || defined(__ARM_NEON__)) /* NEON */ \
+    && !defined(XXH_ENABLE_AUTOVECTORIZE)             /* Define to disable */
+        /*
+         * UGLY HACK:
+         * Clang for ARMv7-A tries to vectorize this loop, similar to GCC x86.
+         * In everywhere else, it uses scalar code.
+         *
+         * For 64->128-bit multiplies, even if the NEON was 100% optimal, it
+         * would still be slower than UMAAL (see XXH_mult64to128).
+         *
+         * Unfortunately, Clang doesn't handle the long multiplies properly and
+         * converts them to the nonexistent "vmulq_u64" intrinsic, which is then
+         * scalarized into an ugly mess of VMOV.32 instructions.
+         *
+         * This mess is difficult to avoid without turning autovectorization
+         * off completely, but they are usually relatively minor and/or not
+         * worth it to fix.
+         *
+         * This loop is the easiest to fix, as unlike XXH32, this pragma
+         * _actually works_ because it is a loop vectorization instead of an
+         * SLP vectorization.
+         */
+        #pragma clang loop vectorize(disable)
+#endif
+        for (i=8 ; i < nbRounds; i++) {
+            /*
+             * Prevents clang for unrolling the acc loop and interleaving with this one.
+             */
+            XXH_COMPILER_GUARD(acc);
+            acc_end += XXH3_mix16B(input+(16*i), secret+(16*(i-8)) + XXH3_MIDSIZE_STARTOFFSET, seed);
+        }
+        return XXH3_avalanche(acc + acc_end);
+    }
+}
+
+
+/* =======     Long Keys     ======= */
+
+#define XXH_STRIPE_LEN 64
+#define XXH_SECRET_CONSUME_RATE 8   /* nb of secret bytes consumed at each accumulation */
+#define XXH_ACC_NB (XXH_STRIPE_LEN / sizeof(xxh_u64))
+
+#ifdef XXH_OLD_NAMES
+#  define STRIPE_LEN XXH_STRIPE_LEN
+#  define ACC_NB XXH_ACC_NB
+#endif
+
+#ifndef XXH_PREFETCH_DIST
+#  ifdef __clang__
+#    define XXH_PREFETCH_DIST 320
+#  else
+#    if (XXH_VECTOR == XXH_AVX512)
+#      define XXH_PREFETCH_DIST 512
+#    else
+#      define XXH_PREFETCH_DIST 384
+#    endif
+#  endif  /* __clang__ */
+#endif  /* XXH_PREFETCH_DIST */
+
+/*
+ * These macros are to generate an XXH3_accumulate() function.
+ * The two arguments select the name suffix and target attribute.
+ *
+ * The name of this symbol is XXH3_accumulate_<name>() and it calls
+ * XXH3_accumulate_512_<name>().
+ *
+ * It may be useful to hand implement this function if the compiler fails to
+ * optimize the inline function.
+ */
+#define XXH3_ACCUMULATE_TEMPLATE(name)                      \
+void                                                        \
+XXH3_accumulate_##name(xxh_u64* XXH_RESTRICT acc,           \
+                       const xxh_u8* XXH_RESTRICT input,    \
+                       const xxh_u8* XXH_RESTRICT secret,   \
+                       size_t nbStripes)                    \
+{                                                           \
+    size_t n;                                               \
+    for (n = 0; n < nbStripes; n++ ) {                      \
+        const xxh_u8* const in = input + n*XXH_STRIPE_LEN;  \
+        XXH_PREFETCH(in + XXH_PREFETCH_DIST);               \
+        XXH3_accumulate_512_##name(                         \
+                 acc,                                       \
+                 in,                                        \
+                 secret + n*XXH_SECRET_CONSUME_RATE);       \
+    }                                                       \
+}
+
+
+XXH_FORCE_INLINE void XXH_writeLE64(void* dst, xxh_u64 v64)
+{
+    if (!XXH_CPU_LITTLE_ENDIAN) v64 = XXH_swap64(v64);
+    XXH_memcpy(dst, &v64, sizeof(v64));
+}
+
+/* Several intrinsic functions below are supposed to accept __int64 as argument,
+ * as documented in https://software.intel.com/sites/landingpage/IntrinsicsGuide/ .
+ * However, several environments do not define __int64 type,
+ * requiring a workaround.
+ */
+#if !defined (__VMS) \
+  && (defined (__cplusplus) \
+  || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
+    typedef int64_t xxh_i64;
+#else
+    /* the following type must have a width of 64-bit */
+    typedef long long xxh_i64;
+#endif
+
+
+/*
+ * XXH3_accumulate_512 is the tightest loop for long inputs, and it is the most optimized.
+ *
+ * It is a hardened version of UMAC, based off of FARSH's implementation.
+ *
+ * This was chosen because it adapts quite well to 32-bit, 64-bit, and SIMD
+ * implementations, and it is ridiculously fast.
+ *
+ * We harden it by mixing the original input to the accumulators as well as the product.
+ *
+ * This means that in the (relatively likely) case of a multiply by zero, the
+ * original input is preserved.
+ *
+ * On 128-bit inputs, we swap 64-bit pairs when we add the input to improve
+ * cross-pollination, as otherwise the upper and lower halves would be
+ * essentially independent.
+ *
+ * This doesn't matter on 64-bit hashes since they all get merged together in
+ * the end, so we skip the extra step.
+ *
+ * Both XXH3_64bits and XXH3_128bits use this subroutine.
+ */
+
+#if (XXH_VECTOR == XXH_AVX512) \
+     || (defined(XXH_DISPATCH_AVX512) && XXH_DISPATCH_AVX512 != 0)
+
+#ifndef XXH_TARGET_AVX512
+# define XXH_TARGET_AVX512  /* disable attribute target */
+#endif
+
+XXH_FORCE_INLINE XXH_TARGET_AVX512 void
+XXH3_accumulate_512_avx512(void* XXH_RESTRICT acc,
+                     const void* XXH_RESTRICT input,
+                     const void* XXH_RESTRICT secret)
+{
+    __m512i* const xacc = (__m512i *) acc;
+    XXH_ASSERT((((size_t)acc) & 63) == 0);
+    XXH_STATIC_ASSERT(XXH_STRIPE_LEN == sizeof(__m512i));
+
+    {
+        /* data_vec    = input[0]; */
+        __m512i const data_vec    = _mm512_loadu_si512   (input);
+        /* key_vec     = secret[0]; */
+        __m512i const key_vec     = _mm512_loadu_si512   (secret);
+        /* data_key    = data_vec ^ key_vec; */
+        __m512i const data_key    = _mm512_xor_si512     (data_vec, key_vec);
+        /* data_key_lo = data_key >> 32; */
+        __m512i const data_key_lo = _mm512_srli_epi64 (data_key, 32);
+        /* product     = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
+        __m512i const product     = _mm512_mul_epu32     (data_key, data_key_lo);
+        /* xacc[0] += swap(data_vec); */
+        __m512i const data_swap = _mm512_shuffle_epi32(data_vec, (_MM_PERM_ENUM)_MM_SHUFFLE(1, 0, 3, 2));
+        __m512i const sum       = _mm512_add_epi64(*xacc, data_swap);
+        /* xacc[0] += product; */
+        *xacc = _mm512_add_epi64(product, sum);
+    }
+}
+XXH_FORCE_INLINE XXH_TARGET_AVX512 XXH3_ACCUMULATE_TEMPLATE(avx512)
+
+/*
+ * XXH3_scrambleAcc: Scrambles the accumulators to improve mixing.
+ *
+ * Multiplication isn't perfect, as explained by Google in HighwayHash:
+ *
+ *  // Multiplication mixes/scrambles bytes 0-7 of the 64-bit result to
+ *  // varying degrees. In descending order of goodness, bytes
+ *  // 3 4 2 5 1 6 0 7 have quality 228 224 164 160 100 96 36 32.
+ *  // As expected, the upper and lower bytes are much worse.
+ *
+ * Source: https://github.com/google/highwayhash/blob/0aaf66b/highwayhash/hh_avx2.h#L291
+ *
+ * Since our algorithm uses a pseudorandom secret to add some variance into the
+ * mix, we don't need to (or want to) mix as often or as much as HighwayHash does.
+ *
+ * This isn't as tight as XXH3_accumulate, but still written in SIMD to avoid
+ * extraction.
+ *
+ * Both XXH3_64bits and XXH3_128bits use this subroutine.
+ */
+
+XXH_FORCE_INLINE XXH_TARGET_AVX512 void
+XXH3_scrambleAcc_avx512(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 63) == 0);
+    XXH_STATIC_ASSERT(XXH_STRIPE_LEN == sizeof(__m512i));
+    {   __m512i* const xacc = (__m512i*) acc;
+        const __m512i prime32 = _mm512_set1_epi32((int)XXH_PRIME32_1);
+
+        /* xacc[0] ^= (xacc[0] >> 47) */
+        __m512i const acc_vec     = *xacc;
+        __m512i const shifted     = _mm512_srli_epi64    (acc_vec, 47);
+        /* xacc[0] ^= secret; */
+        __m512i const key_vec     = _mm512_loadu_si512   (secret);
+        __m512i const data_key    = _mm512_ternarylogic_epi32(key_vec, acc_vec, shifted, 0x96 /* key_vec ^ acc_vec ^ shifted */);
+
+        /* xacc[0] *= XXH_PRIME32_1; */
+        __m512i const data_key_hi = _mm512_srli_epi64 (data_key, 32);
+        __m512i const prod_lo     = _mm512_mul_epu32     (data_key, prime32);
+        __m512i const prod_hi     = _mm512_mul_epu32     (data_key_hi, prime32);
+        *xacc = _mm512_add_epi64(prod_lo, _mm512_slli_epi64(prod_hi, 32));
+    }
+}
+
+XXH_FORCE_INLINE XXH_TARGET_AVX512 void
+XXH3_initCustomSecret_avx512(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
+{
+    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 63) == 0);
+    XXH_STATIC_ASSERT(XXH_SEC_ALIGN == 64);
+    XXH_ASSERT(((size_t)customSecret & 63) == 0);
+    (void)(&XXH_writeLE64);
+    {   int const nbRounds = XXH_SECRET_DEFAULT_SIZE / sizeof(__m512i);
+        __m512i const seed_pos = _mm512_set1_epi64((xxh_i64)seed64);
+        __m512i const seed     = _mm512_mask_sub_epi64(seed_pos, 0xAA, _mm512_set1_epi8(0), seed_pos);
+
+        const __m512i* const src  = (const __m512i*) ((const void*) XXH3_kSecret);
+              __m512i* const dest = (      __m512i*) customSecret;
+        int i;
+        XXH_ASSERT(((size_t)src & 63) == 0); /* control alignment */
+        XXH_ASSERT(((size_t)dest & 63) == 0);
+        for (i=0; i < nbRounds; ++i) {
+            dest[i] = _mm512_add_epi64(_mm512_load_si512(src + i), seed);
+    }   }
+}
+
+#endif
+
+#if (XXH_VECTOR == XXH_AVX2) \
+    || (defined(XXH_DISPATCH_AVX2) && XXH_DISPATCH_AVX2 != 0)
+
+#ifndef XXH_TARGET_AVX2
+# define XXH_TARGET_AVX2  /* disable attribute target */
+#endif
+
+XXH_FORCE_INLINE XXH_TARGET_AVX2 void
+XXH3_accumulate_512_avx2( void* XXH_RESTRICT acc,
+                    const void* XXH_RESTRICT input,
+                    const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 31) == 0);
+    {   __m256i* const xacc    =       (__m256i *) acc;
+        /* Unaligned. This is mainly for pointer arithmetic, and because
+         * _mm256_loadu_si256 requires  a const __m256i * pointer for some reason. */
+        const         __m256i* const xinput  = (const __m256i *) input;
+        /* Unaligned. This is mainly for pointer arithmetic, and because
+         * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */
+        const         __m256i* const xsecret = (const __m256i *) secret;
+
+        size_t i;
+        for (i=0; i < XXH_STRIPE_LEN/sizeof(__m256i); i++) {
+            /* data_vec    = xinput[i]; */
+            __m256i const data_vec    = _mm256_loadu_si256    (xinput+i);
+            /* key_vec     = xsecret[i]; */
+            __m256i const key_vec     = _mm256_loadu_si256   (xsecret+i);
+            /* data_key    = data_vec ^ key_vec; */
+            __m256i const data_key    = _mm256_xor_si256     (data_vec, key_vec);
+            /* data_key_lo = data_key >> 32; */
+            __m256i const data_key_lo = _mm256_srli_epi64 (data_key, 32);
+            /* product     = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
+            __m256i const product     = _mm256_mul_epu32     (data_key, data_key_lo);
+            /* xacc[i] += swap(data_vec); */
+            __m256i const data_swap = _mm256_shuffle_epi32(data_vec, _MM_SHUFFLE(1, 0, 3, 2));
+            __m256i const sum       = _mm256_add_epi64(xacc[i], data_swap);
+            /* xacc[i] += product; */
+            xacc[i] = _mm256_add_epi64(product, sum);
+    }   }
+}
+XXH_FORCE_INLINE XXH_TARGET_AVX2 XXH3_ACCUMULATE_TEMPLATE(avx2)
+
+XXH_FORCE_INLINE XXH_TARGET_AVX2 void
+XXH3_scrambleAcc_avx2(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 31) == 0);
+    {   __m256i* const xacc = (__m256i*) acc;
+        /* Unaligned. This is mainly for pointer arithmetic, and because
+         * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */
+        const         __m256i* const xsecret = (const __m256i *) secret;
+        const __m256i prime32 = _mm256_set1_epi32((int)XXH_PRIME32_1);
+
+        size_t i;
+        for (i=0; i < XXH_STRIPE_LEN/sizeof(__m256i); i++) {
+            /* xacc[i] ^= (xacc[i] >> 47) */
+            __m256i const acc_vec     = xacc[i];
+            __m256i const shifted     = _mm256_srli_epi64    (acc_vec, 47);
+            __m256i const data_vec    = _mm256_xor_si256     (acc_vec, shifted);
+            /* xacc[i] ^= xsecret; */
+            __m256i const key_vec     = _mm256_loadu_si256   (xsecret+i);
+            __m256i const data_key    = _mm256_xor_si256     (data_vec, key_vec);
+
+            /* xacc[i] *= XXH_PRIME32_1; */
+            __m256i const data_key_hi = _mm256_srli_epi64 (data_key, 32);
+            __m256i const prod_lo     = _mm256_mul_epu32     (data_key, prime32);
+            __m256i const prod_hi     = _mm256_mul_epu32     (data_key_hi, prime32);
+            xacc[i] = _mm256_add_epi64(prod_lo, _mm256_slli_epi64(prod_hi, 32));
+        }
+    }
+}
+
+XXH_FORCE_INLINE XXH_TARGET_AVX2 void XXH3_initCustomSecret_avx2(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
+{
+    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 31) == 0);
+    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE / sizeof(__m256i)) == 6);
+    XXH_STATIC_ASSERT(XXH_SEC_ALIGN <= 64);
+    (void)(&XXH_writeLE64);
+    XXH_PREFETCH(customSecret);
+    {   __m256i const seed = _mm256_set_epi64x((xxh_i64)(0U - seed64), (xxh_i64)seed64, (xxh_i64)(0U - seed64), (xxh_i64)seed64);
+
+        const __m256i* const src  = (const __m256i*) ((const void*) XXH3_kSecret);
+              __m256i*       dest = (      __m256i*) customSecret;
+
+#       if defined(__GNUC__) || defined(__clang__)
+        /*
+         * On GCC & Clang, marking 'dest' as modified will cause the compiler:
+         *   - do not extract the secret from sse registers in the internal loop
+         *   - use less common registers, and avoid pushing these reg into stack
+         */
+        XXH_COMPILER_GUARD(dest);
+#       endif
+        XXH_ASSERT(((size_t)src & 31) == 0); /* control alignment */
+        XXH_ASSERT(((size_t)dest & 31) == 0);
+
+        /* GCC -O2 need unroll loop manually */
+        dest[0] = _mm256_add_epi64(_mm256_load_si256(src+0), seed);
+        dest[1] = _mm256_add_epi64(_mm256_load_si256(src+1), seed);
+        dest[2] = _mm256_add_epi64(_mm256_load_si256(src+2), seed);
+        dest[3] = _mm256_add_epi64(_mm256_load_si256(src+3), seed);
+        dest[4] = _mm256_add_epi64(_mm256_load_si256(src+4), seed);
+        dest[5] = _mm256_add_epi64(_mm256_load_si256(src+5), seed);
+    }
+}
+
+#endif
+
+/* x86dispatch always generates SSE2 */
+#if (XXH_VECTOR == XXH_SSE2) || defined(XXH_X86DISPATCH)
+
+#ifndef XXH_TARGET_SSE2
+# define XXH_TARGET_SSE2  /* disable attribute target */
+#endif
+
+XXH_FORCE_INLINE XXH_TARGET_SSE2 void
+XXH3_accumulate_512_sse2( void* XXH_RESTRICT acc,
+                    const void* XXH_RESTRICT input,
+                    const void* XXH_RESTRICT secret)
+{
+    /* SSE2 is just a half-scale version of the AVX2 version. */
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+    {   __m128i* const xacc    =       (__m128i *) acc;
+        /* Unaligned. This is mainly for pointer arithmetic, and because
+         * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
+        const         __m128i* const xinput  = (const __m128i *) input;
+        /* Unaligned. This is mainly for pointer arithmetic, and because
+         * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
+        const         __m128i* const xsecret = (const __m128i *) secret;
+
+        size_t i;
+        for (i=0; i < XXH_STRIPE_LEN/sizeof(__m128i); i++) {
+            /* data_vec    = xinput[i]; */
+            __m128i const data_vec    = _mm_loadu_si128   (xinput+i);
+            /* key_vec     = xsecret[i]; */
+            __m128i const key_vec     = _mm_loadu_si128   (xsecret+i);
+            /* data_key    = data_vec ^ key_vec; */
+            __m128i const data_key    = _mm_xor_si128     (data_vec, key_vec);
+            /* data_key_lo = data_key >> 32; */
+            __m128i const data_key_lo = _mm_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
+            /* product     = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
+            __m128i const product     = _mm_mul_epu32     (data_key, data_key_lo);
+            /* xacc[i] += swap(data_vec); */
+            __m128i const data_swap = _mm_shuffle_epi32(data_vec, _MM_SHUFFLE(1,0,3,2));
+            __m128i const sum       = _mm_add_epi64(xacc[i], data_swap);
+            /* xacc[i] += product; */
+            xacc[i] = _mm_add_epi64(product, sum);
+    }   }
+}
+XXH_FORCE_INLINE XXH_TARGET_SSE2 XXH3_ACCUMULATE_TEMPLATE(sse2)
+
+XXH_FORCE_INLINE XXH_TARGET_SSE2 void
+XXH3_scrambleAcc_sse2(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+    {   __m128i* const xacc = (__m128i*) acc;
+        /* Unaligned. This is mainly for pointer arithmetic, and because
+         * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
+        const         __m128i* const xsecret = (const __m128i *) secret;
+        const __m128i prime32 = _mm_set1_epi32((int)XXH_PRIME32_1);
+
+        size_t i;
+        for (i=0; i < XXH_STRIPE_LEN/sizeof(__m128i); i++) {
+            /* xacc[i] ^= (xacc[i] >> 47) */
+            __m128i const acc_vec     = xacc[i];
+            __m128i const shifted     = _mm_srli_epi64    (acc_vec, 47);
+            __m128i const data_vec    = _mm_xor_si128     (acc_vec, shifted);
+            /* xacc[i] ^= xsecret[i]; */
+            __m128i const key_vec     = _mm_loadu_si128   (xsecret+i);
+            __m128i const data_key    = _mm_xor_si128     (data_vec, key_vec);
+
+            /* xacc[i] *= XXH_PRIME32_1; */
+            __m128i const data_key_hi = _mm_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
+            __m128i const prod_lo     = _mm_mul_epu32     (data_key, prime32);
+            __m128i const prod_hi     = _mm_mul_epu32     (data_key_hi, prime32);
+            xacc[i] = _mm_add_epi64(prod_lo, _mm_slli_epi64(prod_hi, 32));
+        }
+    }
+}
+
+XXH_FORCE_INLINE XXH_TARGET_SSE2 void XXH3_initCustomSecret_sse2(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
+{
+    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 15) == 0);
+    (void)(&XXH_writeLE64);
+    {   int const nbRounds = XXH_SECRET_DEFAULT_SIZE / sizeof(__m128i);
+
+#       if defined(_MSC_VER) && defined(_M_IX86) && _MSC_VER < 1900
+        /* MSVC 32bit mode does not support _mm_set_epi64x before 2015 */
+        XXH_ALIGN(16) const xxh_i64 seed64x2[2] = { (xxh_i64)seed64, (xxh_i64)(0U - seed64) };
+        __m128i const seed = _mm_load_si128((__m128i const*)seed64x2);
+#       else
+        __m128i const seed = _mm_set_epi64x((xxh_i64)(0U - seed64), (xxh_i64)seed64);
+#       endif
+        int i;
+
+        const void* const src16 = XXH3_kSecret;
+        __m128i* dst16 = (__m128i*) customSecret;
+#       if defined(__GNUC__) || defined(__clang__)
+        /*
+         * On GCC & Clang, marking 'dest' as modified will cause the compiler:
+         *   - do not extract the secret from sse registers in the internal loop
+         *   - use less common registers, and avoid pushing these reg into stack
+         */
+        XXH_COMPILER_GUARD(dst16);
+#       endif
+        XXH_ASSERT(((size_t)src16 & 15) == 0); /* control alignment */
+        XXH_ASSERT(((size_t)dst16 & 15) == 0);
+
+        for (i=0; i < nbRounds; ++i) {
+            dst16[i] = _mm_add_epi64(_mm_load_si128((const __m128i *)src16+i), seed);
+    }   }
+}
+
+#endif
+
+#if (XXH_VECTOR == XXH_NEON)
+
+/* forward declarations for the scalar routines */
+XXH_FORCE_INLINE void
+XXH3_scalarRound(void* XXH_RESTRICT acc, void const* XXH_RESTRICT input,
+                 void const* XXH_RESTRICT secret, size_t lane);
+
+XXH_FORCE_INLINE void
+XXH3_scalarScrambleRound(void* XXH_RESTRICT acc,
+                         void const* XXH_RESTRICT secret, size_t lane);
+
+/*!
+ * @internal
+ * @brief The bulk processing loop for NEON and WASM SIMD128.
+ *
+ * The NEON code path is actually partially scalar when running on AArch64. This
+ * is to optimize the pipelining and can have up to 15% speedup depending on the
+ * CPU, and it also mitigates some GCC codegen issues.
+ *
+ * @see XXH3_NEON_LANES for configuring this and details about this optimization.
+ *
+ * NEON's 32-bit to 64-bit long multiply takes a half vector of 32-bit
+ * integers instead of the other platforms which mask full 64-bit vectors,
+ * so the setup is more complicated than just shifting right.
+ *
+ * Additionally, there is an optimization for 4 lanes at once noted below.
+ *
+ * Since, as stated, the most optimal amount of lanes for Cortexes is 6,
+ * there needs to be *three* versions of the accumulate operation used
+ * for the remaining 2 lanes.
+ *
+ * WASM's SIMD128 uses SIMDe's arm_neon.h polyfill because the intrinsics overlap
+ * nearly perfectly.
+ */
+
+XXH_FORCE_INLINE void
+XXH3_accumulate_512_neon( void* XXH_RESTRICT acc,
+                    const void* XXH_RESTRICT input,
+                    const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+    XXH_STATIC_ASSERT(XXH3_NEON_LANES > 0 && XXH3_NEON_LANES <= XXH_ACC_NB && XXH3_NEON_LANES % 2 == 0);
+    {   /* GCC for darwin arm64 does not like aliasing here */
+        xxh_aliasing_uint64x2_t* const xacc = (xxh_aliasing_uint64x2_t*) acc;
+        /* We don't use a uint32x4_t pointer because it causes bus errors on ARMv7. */
+        uint8_t const* xinput = (const uint8_t *) input;
+        uint8_t const* xsecret  = (const uint8_t *) secret;
+
+        size_t i;
+#ifdef __wasm_simd128__
+        /*
+         * On WASM SIMD128, Clang emits direct address loads when XXH3_kSecret
+         * is constant propagated, which results in it converting it to this
+         * inside the loop:
+         *
+         *    a = v128.load(XXH3_kSecret +  0 + $secret_offset, offset = 0)
+         *    b = v128.load(XXH3_kSecret + 16 + $secret_offset, offset = 0)
+         *    ...
+         *
+         * This requires a full 32-bit address immediate (and therefore a 6 byte
+         * instruction) as well as an add for each offset.
+         *
+         * Putting an asm guard prevents it from folding (at the cost of losing
+         * the alignment hint), and uses the free offset in `v128.load` instead
+         * of adding secret_offset each time which overall reduces code size by
+         * about a kilobyte and improves performance.
+         */
+        XXH_COMPILER_GUARD(xsecret);
+#endif
+        /* Scalar lanes use the normal scalarRound routine */
+        for (i = XXH3_NEON_LANES; i < XXH_ACC_NB; i++) {
+            XXH3_scalarRound(acc, input, secret, i);
+        }
+        i = 0;
+        /* 4 NEON lanes at a time. */
+        for (; i+1 < XXH3_NEON_LANES / 2; i+=2) {
+            /* data_vec = xinput[i]; */
+            uint64x2_t data_vec_1 = XXH_vld1q_u64(xinput  + (i * 16));
+            uint64x2_t data_vec_2 = XXH_vld1q_u64(xinput  + ((i+1) * 16));
+            /* key_vec  = xsecret[i];  */
+            uint64x2_t key_vec_1  = XXH_vld1q_u64(xsecret + (i * 16));
+            uint64x2_t key_vec_2  = XXH_vld1q_u64(xsecret + ((i+1) * 16));
+            /* data_swap = swap(data_vec) */
+            uint64x2_t data_swap_1 = vextq_u64(data_vec_1, data_vec_1, 1);
+            uint64x2_t data_swap_2 = vextq_u64(data_vec_2, data_vec_2, 1);
+            /* data_key = data_vec ^ key_vec; */
+            uint64x2_t data_key_1 = veorq_u64(data_vec_1, key_vec_1);
+            uint64x2_t data_key_2 = veorq_u64(data_vec_2, key_vec_2);
+
+            /*
+             * If we reinterpret the 64x2 vectors as 32x4 vectors, we can use a
+             * de-interleave operation for 4 lanes in 1 step with `vuzpq_u32` to
+             * get one vector with the low 32 bits of each lane, and one vector
+             * with the high 32 bits of each lane.
+             *
+             * The intrinsic returns a double vector because the original ARMv7-a
+             * instruction modified both arguments in place. AArch64 and SIMD128 emit
+             * two instructions from this intrinsic.
+             *
+             *  [ dk11L | dk11H | dk12L | dk12H ] -> [ dk11L | dk12L | dk21L | dk22L ]
+             *  [ dk21L | dk21H | dk22L | dk22H ] -> [ dk11H | dk12H | dk21H | dk22H ]
+             */
+            uint32x4x2_t unzipped = vuzpq_u32(
+                vreinterpretq_u32_u64(data_key_1),
+                vreinterpretq_u32_u64(data_key_2)
+            );
+            /* data_key_lo = data_key & 0xFFFFFFFF */
+            uint32x4_t data_key_lo = unzipped.val[0];
+            /* data_key_hi = data_key >> 32 */
+            uint32x4_t data_key_hi = unzipped.val[1];
+            /*
+             * Then, we can split the vectors horizontally and multiply which, as for most
+             * widening intrinsics, have a variant that works on both high half vectors
+             * for free on AArch64. A similar instruction is available on SIMD128.
+             *
+             * sum = data_swap + (u64x2) data_key_lo * (u64x2) data_key_hi
+             */
+            uint64x2_t sum_1 = XXH_vmlal_low_u32(data_swap_1, data_key_lo, data_key_hi);
+            uint64x2_t sum_2 = XXH_vmlal_high_u32(data_swap_2, data_key_lo, data_key_hi);
+            /*
+             * Clang reorders
+             *    a += b * c;     // umlal   swap.2d, dkl.2s, dkh.2s
+             *    c += a;         // add     acc.2d, acc.2d, swap.2d
+             * to
+             *    c += a;         // add     acc.2d, acc.2d, swap.2d
+             *    c += b * c;     // umlal   acc.2d, dkl.2s, dkh.2s
+             *
+             * While it would make sense in theory since the addition is faster,
+             * for reasons likely related to umlal being limited to certain NEON
+             * pipelines, this is worse. A compiler guard fixes this.
+             */
+            XXH_COMPILER_GUARD_CLANG_NEON(sum_1);
+            XXH_COMPILER_GUARD_CLANG_NEON(sum_2);
+            /* xacc[i] = acc_vec + sum; */
+            xacc[i]   = vaddq_u64(xacc[i], sum_1);
+            xacc[i+1] = vaddq_u64(xacc[i+1], sum_2);
+        }
+        /* Operate on the remaining NEON lanes 2 at a time. */
+        for (; i < XXH3_NEON_LANES / 2; i++) {
+            /* data_vec = xinput[i]; */
+            uint64x2_t data_vec = XXH_vld1q_u64(xinput  + (i * 16));
+            /* key_vec  = xsecret[i];  */
+            uint64x2_t key_vec  = XXH_vld1q_u64(xsecret + (i * 16));
+            /* acc_vec_2 = swap(data_vec) */
+            uint64x2_t data_swap = vextq_u64(data_vec, data_vec, 1);
+            /* data_key = data_vec ^ key_vec; */
+            uint64x2_t data_key = veorq_u64(data_vec, key_vec);
+            /* For two lanes, just use VMOVN and VSHRN. */
+            /* data_key_lo = data_key & 0xFFFFFFFF; */
+            uint32x2_t data_key_lo = vmovn_u64(data_key);
+            /* data_key_hi = data_key >> 32; */
+            uint32x2_t data_key_hi = vshrn_n_u64(data_key, 32);
+            /* sum = data_swap + (u64x2) data_key_lo * (u64x2) data_key_hi; */
+            uint64x2_t sum = vmlal_u32(data_swap, data_key_lo, data_key_hi);
+            /* Same Clang workaround as before */
+            XXH_COMPILER_GUARD_CLANG_NEON(sum);
+            /* xacc[i] = acc_vec + sum; */
+            xacc[i] = vaddq_u64 (xacc[i], sum);
+        }
+    }
+}
+XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(neon)
+
+XXH_FORCE_INLINE void
+XXH3_scrambleAcc_neon(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+
+    {   xxh_aliasing_uint64x2_t* xacc       = (xxh_aliasing_uint64x2_t*) acc;
+        uint8_t const* xsecret = (uint8_t const*) secret;
+
+        size_t i;
+        /* WASM uses operator overloads and doesn't need these. */
+#ifndef __wasm_simd128__
+        /* { prime32_1, prime32_1 } */
+        uint32x2_t const kPrimeLo = vdup_n_u32(XXH_PRIME32_1);
+        /* { 0, prime32_1, 0, prime32_1 } */
+        uint32x4_t const kPrimeHi = vreinterpretq_u32_u64(vdupq_n_u64((xxh_u64)XXH_PRIME32_1 << 32));
+#endif
+
+        /* AArch64 uses both scalar and neon at the same time */
+        for (i = XXH3_NEON_LANES; i < XXH_ACC_NB; i++) {
+            XXH3_scalarScrambleRound(acc, secret, i);
+        }
+        for (i=0; i < XXH3_NEON_LANES / 2; i++) {
+            /* xacc[i] ^= (xacc[i] >> 47); */
+            uint64x2_t acc_vec  = xacc[i];
+            uint64x2_t shifted  = vshrq_n_u64(acc_vec, 47);
+            uint64x2_t data_vec = veorq_u64(acc_vec, shifted);
+
+            /* xacc[i] ^= xsecret[i]; */
+            uint64x2_t key_vec  = XXH_vld1q_u64(xsecret + (i * 16));
+            uint64x2_t data_key = veorq_u64(data_vec, key_vec);
+            /* xacc[i] *= XXH_PRIME32_1 */
+#ifdef __wasm_simd128__
+            /* SIMD128 has multiply by u64x2, use it instead of expanding and scalarizing */
+            xacc[i] = data_key * XXH_PRIME32_1;
+#else
+            /*
+             * Expanded version with portable NEON intrinsics
+             *
+             *    lo(x) * lo(y) + (hi(x) * lo(y) << 32)
+             *
+             * prod_hi = hi(data_key) * lo(prime) << 32
+             *
+             * Since we only need 32 bits of this multiply a trick can be used, reinterpreting the vector
+             * as a uint32x4_t and multiplying by { 0, prime, 0, prime } to cancel out the unwanted bits
+             * and avoid the shift.
+             */
+            uint32x4_t prod_hi = vmulq_u32 (vreinterpretq_u32_u64(data_key), kPrimeHi);
+            /* Extract low bits for vmlal_u32  */
+            uint32x2_t data_key_lo = vmovn_u64(data_key);
+            /* xacc[i] = prod_hi + lo(data_key) * XXH_PRIME32_1; */
+            xacc[i] = vmlal_u32(vreinterpretq_u64_u32(prod_hi), data_key_lo, kPrimeLo);
+#endif
+        }
+    }
+}
+#endif
+
+#if (XXH_VECTOR == XXH_VSX)
+
+XXH_FORCE_INLINE void
+XXH3_accumulate_512_vsx(  void* XXH_RESTRICT acc,
+                    const void* XXH_RESTRICT input,
+                    const void* XXH_RESTRICT secret)
+{
+    /* presumed aligned */
+    xxh_aliasing_u64x2* const xacc = (xxh_aliasing_u64x2*) acc;
+    xxh_u8 const* const xinput   = (xxh_u8 const*) input;   /* no alignment restriction */
+    xxh_u8 const* const xsecret  = (xxh_u8 const*) secret;    /* no alignment restriction */
+    xxh_u64x2 const v32 = { 32, 32 };
+    size_t i;
+    for (i = 0; i < XXH_STRIPE_LEN / sizeof(xxh_u64x2); i++) {
+        /* data_vec = xinput[i]; */
+        xxh_u64x2 const data_vec = XXH_vec_loadu(xinput + 16*i);
+        /* key_vec = xsecret[i]; */
+        xxh_u64x2 const key_vec  = XXH_vec_loadu(xsecret + 16*i);
+        xxh_u64x2 const data_key = data_vec ^ key_vec;
+        /* shuffled = (data_key << 32) | (data_key >> 32); */
+        xxh_u32x4 const shuffled = (xxh_u32x4)vec_rl(data_key, v32);
+        /* product = ((xxh_u64x2)data_key & 0xFFFFFFFF) * ((xxh_u64x2)shuffled & 0xFFFFFFFF); */
+        xxh_u64x2 const product  = XXH_vec_mulo((xxh_u32x4)data_key, shuffled);
+        /* acc_vec = xacc[i]; */
+        xxh_u64x2 acc_vec        = xacc[i];
+        acc_vec += product;
+
+        /* swap high and low halves */
+#ifdef __s390x__
+        acc_vec += vec_permi(data_vec, data_vec, 2);
+#else
+        acc_vec += vec_xxpermdi(data_vec, data_vec, 2);
+#endif
+        xacc[i] = acc_vec;
+    }
+}
+XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(vsx)
+
+XXH_FORCE_INLINE void
+XXH3_scrambleAcc_vsx(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    XXH_ASSERT((((size_t)acc) & 15) == 0);
+
+    {   xxh_aliasing_u64x2* const xacc = (xxh_aliasing_u64x2*) acc;
+        const xxh_u8* const xsecret = (const xxh_u8*) secret;
+        /* constants */
+        xxh_u64x2 const v32  = { 32, 32 };
+        xxh_u64x2 const v47 = { 47, 47 };
+        xxh_u32x4 const prime = { XXH_PRIME32_1, XXH_PRIME32_1, XXH_PRIME32_1, XXH_PRIME32_1 };
+        size_t i;
+        for (i = 0; i < XXH_STRIPE_LEN / sizeof(xxh_u64x2); i++) {
+            /* xacc[i] ^= (xacc[i] >> 47); */
+            xxh_u64x2 const acc_vec  = xacc[i];
+            xxh_u64x2 const data_vec = acc_vec ^ (acc_vec >> v47);
+
+            /* xacc[i] ^= xsecret[i]; */
+            xxh_u64x2 const key_vec  = XXH_vec_loadu(xsecret + 16*i);
+            xxh_u64x2 const data_key = data_vec ^ key_vec;
+
+            /* xacc[i] *= XXH_PRIME32_1 */
+            /* prod_lo = ((xxh_u64x2)data_key & 0xFFFFFFFF) * ((xxh_u64x2)prime & 0xFFFFFFFF);  */
+            xxh_u64x2 const prod_even  = XXH_vec_mule((xxh_u32x4)data_key, prime);
+            /* prod_hi = ((xxh_u64x2)data_key >> 32) * ((xxh_u64x2)prime >> 32);  */
+            xxh_u64x2 const prod_odd  = XXH_vec_mulo((xxh_u32x4)data_key, prime);
+            xacc[i] = prod_odd + (prod_even << v32);
+    }   }
+}
+
+#endif
+
+#if (XXH_VECTOR == XXH_SVE)
+
+XXH_FORCE_INLINE void
+XXH3_accumulate_512_sve( void* XXH_RESTRICT acc,
+                   const void* XXH_RESTRICT input,
+                   const void* XXH_RESTRICT secret)
+{
+    uint64_t *xacc = (uint64_t *)acc;
+    const uint64_t *xinput = (const uint64_t *)(const void *)input;
+    const uint64_t *xsecret = (const uint64_t *)(const void *)secret;
+    svuint64_t kSwap = sveor_n_u64_z(svptrue_b64(), svindex_u64(0, 1), 1);
+    uint64_t element_count = svcntd();
+    if (element_count >= 8) {
+        svbool_t mask = svptrue_pat_b64(SV_VL8);
+        svuint64_t vacc = svld1_u64(mask, xacc);
+        ACCRND(vacc, 0);
+        svst1_u64(mask, xacc, vacc);
+    } else if (element_count == 2) {   /* sve128 */
+        svbool_t mask = svptrue_pat_b64(SV_VL2);
+        svuint64_t acc0 = svld1_u64(mask, xacc + 0);
+        svuint64_t acc1 = svld1_u64(mask, xacc + 2);
+        svuint64_t acc2 = svld1_u64(mask, xacc + 4);
+        svuint64_t acc3 = svld1_u64(mask, xacc + 6);
+        ACCRND(acc0, 0);
+        ACCRND(acc1, 2);
+        ACCRND(acc2, 4);
+        ACCRND(acc3, 6);
+        svst1_u64(mask, xacc + 0, acc0);
+        svst1_u64(mask, xacc + 2, acc1);
+        svst1_u64(mask, xacc + 4, acc2);
+        svst1_u64(mask, xacc + 6, acc3);
+    } else {
+        svbool_t mask = svptrue_pat_b64(SV_VL4);
+        svuint64_t acc0 = svld1_u64(mask, xacc + 0);
+        svuint64_t acc1 = svld1_u64(mask, xacc + 4);
+        ACCRND(acc0, 0);
+        ACCRND(acc1, 4);
+        svst1_u64(mask, xacc + 0, acc0);
+        svst1_u64(mask, xacc + 4, acc1);
+    }
+}
+
+XXH_FORCE_INLINE void
+XXH3_accumulate_sve(xxh_u64* XXH_RESTRICT acc,
+               const xxh_u8* XXH_RESTRICT input,
+               const xxh_u8* XXH_RESTRICT secret,
+               size_t nbStripes)
+{
+    if (nbStripes != 0) {
+        uint64_t *xacc = (uint64_t *)acc;
+        const uint64_t *xinput = (const uint64_t *)(const void *)input;
+        const uint64_t *xsecret = (const uint64_t *)(const void *)secret;
+        svuint64_t kSwap = sveor_n_u64_z(svptrue_b64(), svindex_u64(0, 1), 1);
+        uint64_t element_count = svcntd();
+        if (element_count >= 8) {
+            svbool_t mask = svptrue_pat_b64(SV_VL8);
+            svuint64_t vacc = svld1_u64(mask, xacc + 0);
+            do {
+                /* svprfd(svbool_t, void *, enum svfprop); */
+                svprfd(mask, xinput + 128, SV_PLDL1STRM);
+                ACCRND(vacc, 0);
+                xinput += 8;
+                xsecret += 1;
+                nbStripes--;
+           } while (nbStripes != 0);
+
+           svst1_u64(mask, xacc + 0, vacc);
+        } else if (element_count == 2) { /* sve128 */
+            svbool_t mask = svptrue_pat_b64(SV_VL2);
+            svuint64_t acc0 = svld1_u64(mask, xacc + 0);
+            svuint64_t acc1 = svld1_u64(mask, xacc + 2);
+            svuint64_t acc2 = svld1_u64(mask, xacc + 4);
+            svuint64_t acc3 = svld1_u64(mask, xacc + 6);
+            do {
+                svprfd(mask, xinput + 128, SV_PLDL1STRM);
+                ACCRND(acc0, 0);
+                ACCRND(acc1, 2);
+                ACCRND(acc2, 4);
+                ACCRND(acc3, 6);
+                xinput += 8;
+                xsecret += 1;
+                nbStripes--;
+           } while (nbStripes != 0);
+
+           svst1_u64(mask, xacc + 0, acc0);
+           svst1_u64(mask, xacc + 2, acc1);
+           svst1_u64(mask, xacc + 4, acc2);
+           svst1_u64(mask, xacc + 6, acc3);
+        } else {
+            svbool_t mask = svptrue_pat_b64(SV_VL4);
+            svuint64_t acc0 = svld1_u64(mask, xacc + 0);
+            svuint64_t acc1 = svld1_u64(mask, xacc + 4);
+            do {
+                svprfd(mask, xinput + 128, SV_PLDL1STRM);
+                ACCRND(acc0, 0);
+                ACCRND(acc1, 4);
+                xinput += 8;
+                xsecret += 1;
+                nbStripes--;
+           } while (nbStripes != 0);
+
+           svst1_u64(mask, xacc + 0, acc0);
+           svst1_u64(mask, xacc + 4, acc1);
+       }
+    }
+}
+
+#endif
+
+/* scalar variants - universal */
+
+#if defined(__aarch64__) && (defined(__GNUC__) || defined(__clang__))
+/*
+ * In XXH3_scalarRound(), GCC and Clang have a similar codegen issue, where they
+ * emit an excess mask and a full 64-bit multiply-add (MADD X-form).
+ *
+ * While this might not seem like much, as AArch64 is a 64-bit architecture, only
+ * big Cortex designs have a full 64-bit multiplier.
+ *
+ * On the little cores, the smaller 32-bit multiplier is used, and full 64-bit
+ * multiplies expand to 2-3 multiplies in microcode. This has a major penalty
+ * of up to 4 latency cycles and 2 stall cycles in the multiply pipeline.
+ *
+ * Thankfully, AArch64 still provides the 32-bit long multiply-add (UMADDL) which does
+ * not have this penalty and does the mask automatically.
+ */
+XXH_FORCE_INLINE xxh_u64
+XXH_mult32to64_add64(xxh_u64 lhs, xxh_u64 rhs, xxh_u64 acc)
+{
+    xxh_u64 ret;
+    /* note: %x = 64-bit register, %w = 32-bit register */
+    __asm__("umaddl %x0, %w1, %w2, %x3" : "=r" (ret) : "r" (lhs), "r" (rhs), "r" (acc));
+    return ret;
+}
+#else
+XXH_FORCE_INLINE xxh_u64
+XXH_mult32to64_add64(xxh_u64 lhs, xxh_u64 rhs, xxh_u64 acc)
+{
+    return XXH_mult32to64((xxh_u32)lhs, (xxh_u32)rhs) + acc;
+}
+#endif
+
+/*!
+ * @internal
+ * @brief Scalar round for @ref XXH3_accumulate_512_scalar().
+ *
+ * This is extracted to its own function because the NEON path uses a combination
+ * of NEON and scalar.
+ */
+XXH_FORCE_INLINE void
+XXH3_scalarRound(void* XXH_RESTRICT acc,
+                 void const* XXH_RESTRICT input,
+                 void const* XXH_RESTRICT secret,
+                 size_t lane)
+{
+    xxh_u64* xacc = (xxh_u64*) acc;
+    xxh_u8 const* xinput  = (xxh_u8 const*) input;
+    xxh_u8 const* xsecret = (xxh_u8 const*) secret;
+    XXH_ASSERT(lane < XXH_ACC_NB);
+    XXH_ASSERT(((size_t)acc & (XXH_ACC_ALIGN-1)) == 0);
+    {
+        xxh_u64 const data_val = XXH_readLE64(xinput + lane * 8);
+        xxh_u64 const data_key = data_val ^ XXH_readLE64(xsecret + lane * 8);
+        xacc[lane ^ 1] += data_val; /* swap adjacent lanes */
+        xacc[lane] = XXH_mult32to64_add64(data_key /* & 0xFFFFFFFF */, data_key >> 32, xacc[lane]);
+    }
+}
+
+/*!
+ * @internal
+ * @brief Processes a 64 byte block of data using the scalar path.
+ */
+XXH_FORCE_INLINE void
+XXH3_accumulate_512_scalar(void* XXH_RESTRICT acc,
+                     const void* XXH_RESTRICT input,
+                     const void* XXH_RESTRICT secret)
+{
+    size_t i;
+    /* ARM GCC refuses to unroll this loop, resulting in a 24% slowdown on ARMv6. */
+#if defined(__GNUC__) && !defined(__clang__) \
+  && (defined(__arm__) || defined(__thumb2__)) \
+  && defined(__ARM_FEATURE_UNALIGNED) /* no unaligned access just wastes bytes */ \
+  && XXH_SIZE_OPT <= 0
+#  pragma GCC unroll 8
+#endif
+    for (i=0; i < XXH_ACC_NB; i++) {
+        XXH3_scalarRound(acc, input, secret, i);
+    }
+}
+XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(scalar)
+
+/*!
+ * @internal
+ * @brief Scalar scramble step for @ref XXH3_scrambleAcc_scalar().
+ *
+ * This is extracted to its own function because the NEON path uses a combination
+ * of NEON and scalar.
+ */
+XXH_FORCE_INLINE void
+XXH3_scalarScrambleRound(void* XXH_RESTRICT acc,
+                         void const* XXH_RESTRICT secret,
+                         size_t lane)
+{
+    xxh_u64* const xacc = (xxh_u64*) acc;   /* presumed aligned */
+    const xxh_u8* const xsecret = (const xxh_u8*) secret;   /* no alignment restriction */
+    XXH_ASSERT((((size_t)acc) & (XXH_ACC_ALIGN-1)) == 0);
+    XXH_ASSERT(lane < XXH_ACC_NB);
+    {
+        xxh_u64 const key64 = XXH_readLE64(xsecret + lane * 8);
+        xxh_u64 acc64 = xacc[lane];
+        acc64 = XXH_xorshift64(acc64, 47);
+        acc64 ^= key64;
+        acc64 *= XXH_PRIME32_1;
+        xacc[lane] = acc64;
+    }
+}
+
+/*!
+ * @internal
+ * @brief Scrambles the accumulators after a large chunk has been read
+ */
+XXH_FORCE_INLINE void
+XXH3_scrambleAcc_scalar(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
+{
+    size_t i;
+    for (i=0; i < XXH_ACC_NB; i++) {
+        XXH3_scalarScrambleRound(acc, secret, i);
+    }
+}
+
+XXH_FORCE_INLINE void
+XXH3_initCustomSecret_scalar(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
+{
+    /*
+     * We need a separate pointer for the hack below,
+     * which requires a non-const pointer.
+     * Any decent compiler will optimize this out otherwise.
+     */
+    const xxh_u8* kSecretPtr = XXH3_kSecret;
+    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 15) == 0);
+
+#if defined(__GNUC__) && defined(__aarch64__)
+    /*
+     * UGLY HACK:
+     * GCC and Clang generate a bunch of MOV/MOVK pairs for aarch64, and they are
+     * placed sequentially, in order, at the top of the unrolled loop.
+     *
+     * While MOVK is great for generating constants (2 cycles for a 64-bit
+     * constant compared to 4 cycles for LDR), it fights for bandwidth with
+     * the arithmetic instructions.
+     *
+     *   I   L   S
+     * MOVK
+     * MOVK
+     * MOVK
+     * MOVK
+     * ADD
+     * SUB      STR
+     *          STR
+     * By forcing loads from memory (as the asm line causes the compiler to assume
+     * that XXH3_kSecretPtr has been changed), the pipelines are used more
+     * efficiently:
+     *   I   L   S
+     *      LDR
+     *  ADD LDR
+     *  SUB     STR
+     *          STR
+     *
+     * See XXH3_NEON_LANES for details on the pipsline.
+     *
+     * XXH3_64bits_withSeed, len == 256, Snapdragon 835
+     *   without hack: 2654.4 MB/s
+     *   with hack:    3202.9 MB/s
+     */
+    XXH_COMPILER_GUARD(kSecretPtr);
+#endif
+    {   int const nbRounds = XXH_SECRET_DEFAULT_SIZE / 16;
+        int i;
+        for (i=0; i < nbRounds; i++) {
+            /*
+             * The asm hack causes the compiler to assume that kSecretPtr aliases with
+             * customSecret, and on aarch64, this prevented LDP from merging two
+             * loads together for free. Putting the loads together before the stores
+             * properly generates LDP.
+             */
+            xxh_u64 lo = XXH_readLE64(kSecretPtr + 16*i)     + seed64;
+            xxh_u64 hi = XXH_readLE64(kSecretPtr + 16*i + 8) - seed64;
+            XXH_writeLE64((xxh_u8*)customSecret + 16*i,     lo);
+            XXH_writeLE64((xxh_u8*)customSecret + 16*i + 8, hi);
+    }   }
+}
+
+
+typedef void (*XXH3_f_accumulate)(xxh_u64* XXH_RESTRICT, const xxh_u8* XXH_RESTRICT, const xxh_u8* XXH_RESTRICT, size_t);
+typedef void (*XXH3_f_scrambleAcc)(void* XXH_RESTRICT, const void*);
+typedef void (*XXH3_f_initCustomSecret)(void* XXH_RESTRICT, xxh_u64);
+
+
+#if (XXH_VECTOR == XXH_AVX512)
+
+#define XXH3_accumulate_512 XXH3_accumulate_512_avx512
+#define XXH3_accumulate     XXH3_accumulate_avx512
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_avx512
+#define XXH3_initCustomSecret XXH3_initCustomSecret_avx512
+
+#elif (XXH_VECTOR == XXH_AVX2)
+
+#define XXH3_accumulate_512 XXH3_accumulate_512_avx2
+#define XXH3_accumulate     XXH3_accumulate_avx2
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_avx2
+#define XXH3_initCustomSecret XXH3_initCustomSecret_avx2
+
+#elif (XXH_VECTOR == XXH_SSE2)
+
+#define XXH3_accumulate_512 XXH3_accumulate_512_sse2
+#define XXH3_accumulate     XXH3_accumulate_sse2
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_sse2
+#define XXH3_initCustomSecret XXH3_initCustomSecret_sse2
+
+#elif (XXH_VECTOR == XXH_NEON)
+
+#define XXH3_accumulate_512 XXH3_accumulate_512_neon
+#define XXH3_accumulate     XXH3_accumulate_neon
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_neon
+#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
+
+#elif (XXH_VECTOR == XXH_VSX)
+
+#define XXH3_accumulate_512 XXH3_accumulate_512_vsx
+#define XXH3_accumulate     XXH3_accumulate_vsx
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_vsx
+#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
+
+#elif (XXH_VECTOR == XXH_SVE)
+#define XXH3_accumulate_512 XXH3_accumulate_512_sve
+#define XXH3_accumulate     XXH3_accumulate_sve
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_scalar
+#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
+
+#else /* scalar */
+
+#define XXH3_accumulate_512 XXH3_accumulate_512_scalar
+#define XXH3_accumulate     XXH3_accumulate_scalar
+#define XXH3_scrambleAcc    XXH3_scrambleAcc_scalar
+#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
+
+#endif
+
+#if XXH_SIZE_OPT >= 1 /* don't do SIMD for initialization */
+#  undef XXH3_initCustomSecret
+#  define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
+#endif
+
+XXH_FORCE_INLINE void
+XXH3_hashLong_internal_loop(xxh_u64* XXH_RESTRICT acc,
+                      const xxh_u8* XXH_RESTRICT input, size_t len,
+                      const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+                            XXH3_f_accumulate f_acc,
+                            XXH3_f_scrambleAcc f_scramble)
+{
+    size_t const nbStripesPerBlock = (secretSize - XXH_STRIPE_LEN) / XXH_SECRET_CONSUME_RATE;
+    size_t const block_len = XXH_STRIPE_LEN * nbStripesPerBlock;
+    size_t const nb_blocks = (len - 1) / block_len;
+
+    size_t n;
+
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
+
+    for (n = 0; n < nb_blocks; n++) {
+        f_acc(acc, input + n*block_len, secret, nbStripesPerBlock);
+        f_scramble(acc, secret + secretSize - XXH_STRIPE_LEN);
+    }
+
+    /* last partial block */
+    XXH_ASSERT(len > XXH_STRIPE_LEN);
+    {   size_t const nbStripes = ((len - 1) - (block_len * nb_blocks)) / XXH_STRIPE_LEN;
+        XXH_ASSERT(nbStripes <= (secretSize / XXH_SECRET_CONSUME_RATE));
+        f_acc(acc, input + nb_blocks*block_len, secret, nbStripes);
+
+        /* last stripe */
+        {   const xxh_u8* const p = input + len - XXH_STRIPE_LEN;
+#define XXH_SECRET_LASTACC_START 7  /* not aligned on 8, last secret is different from acc & scrambler */
+            XXH3_accumulate_512(acc, p, secret + secretSize - XXH_STRIPE_LEN - XXH_SECRET_LASTACC_START);
+    }   }
+}
+
+XXH_FORCE_INLINE xxh_u64
+XXH3_mix2Accs(const xxh_u64* XXH_RESTRICT acc, const xxh_u8* XXH_RESTRICT secret)
+{
+    return XXH3_mul128_fold64(
+               acc[0] ^ XXH_readLE64(secret),
+               acc[1] ^ XXH_readLE64(secret+8) );
+}
+
+static XXH64_hash_t
+XXH3_mergeAccs(const xxh_u64* XXH_RESTRICT acc, const xxh_u8* XXH_RESTRICT secret, xxh_u64 start)
+{
+    xxh_u64 result64 = start;
+    size_t i = 0;
+
+    for (i = 0; i < 4; i++) {
+        result64 += XXH3_mix2Accs(acc+2*i, secret + 16*i);
+#if defined(__clang__)                                /* Clang */ \
+    && (defined(__arm__) || defined(__thumb__))       /* ARMv7 */ \
+    && (defined(__ARM_NEON) || defined(__ARM_NEON__)) /* NEON */  \
+    && !defined(XXH_ENABLE_AUTOVECTORIZE)             /* Define to disable */
+        /*
+         * UGLY HACK:
+         * Prevent autovectorization on Clang ARMv7-a. Exact same problem as
+         * the one in XXH3_len_129to240_64b. Speeds up shorter keys > 240b.
+         * XXH3_64bits, len == 256, Snapdragon 835:
+         *   without hack: 2063.7 MB/s
+         *   with hack:    2560.7 MB/s
+         */
+        XXH_COMPILER_GUARD(result64);
+#endif
+    }
+
+    return XXH3_avalanche(result64);
+}
+
+#define XXH3_INIT_ACC { XXH_PRIME32_3, XXH_PRIME64_1, XXH_PRIME64_2, XXH_PRIME64_3, \
+                        XXH_PRIME64_4, XXH_PRIME32_2, XXH_PRIME64_5, XXH_PRIME32_1 }
+
+XXH_FORCE_INLINE XXH64_hash_t
+XXH3_hashLong_64b_internal(const void* XXH_RESTRICT input, size_t len,
+                           const void* XXH_RESTRICT secret, size_t secretSize,
+                           XXH3_f_accumulate f_acc,
+                           XXH3_f_scrambleAcc f_scramble)
+{
+    XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[XXH_ACC_NB] = XXH3_INIT_ACC;
+
+    XXH3_hashLong_internal_loop(acc, (const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize, f_acc, f_scramble);
+
+    /* converge into final hash */
+    XXH_STATIC_ASSERT(sizeof(acc) == 64);
+    /* do not align on 8, so that the secret is different from the accumulator */
+#define XXH_SECRET_MERGEACCS_START 11
+    XXH_ASSERT(secretSize >= sizeof(acc) + XXH_SECRET_MERGEACCS_START);
+    return XXH3_mergeAccs(acc, (const xxh_u8*)secret + XXH_SECRET_MERGEACCS_START, (xxh_u64)len * XXH_PRIME64_1);
+}
+
+/*
+ * It's important for performance to transmit secret's size (when it's static)
+ * so that the compiler can properly optimize the vectorized loop.
+ * This makes a big performance difference for "medium" keys (<1 KB) when using AVX instruction set.
+ * When the secret size is unknown, or on GCC 12 where the mix of NO_INLINE and FORCE_INLINE
+ * breaks -Og, this is XXH_NO_INLINE.
+ */
+XXH3_WITH_SECRET_INLINE XXH64_hash_t
+XXH3_hashLong_64b_withSecret(const void* XXH_RESTRICT input, size_t len,
+                             XXH64_hash_t seed64, const xxh_u8* XXH_RESTRICT secret, size_t secretLen)
+{
+    (void)seed64;
+    return XXH3_hashLong_64b_internal(input, len, secret, secretLen, XXH3_accumulate, XXH3_scrambleAcc);
+}
+
+/*
+ * It's preferable for performance that XXH3_hashLong is not inlined,
+ * as it results in a smaller function for small data, easier to the instruction cache.
+ * Note that inside this no_inline function, we do inline the internal loop,
+ * and provide a statically defined secret size to allow optimization of vector loop.
+ */
+XXH_NO_INLINE XXH_PUREF XXH64_hash_t
+XXH3_hashLong_64b_default(const void* XXH_RESTRICT input, size_t len,
+                          XXH64_hash_t seed64, const xxh_u8* XXH_RESTRICT secret, size_t secretLen)
+{
+    (void)seed64; (void)secret; (void)secretLen;
+    return XXH3_hashLong_64b_internal(input, len, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_accumulate, XXH3_scrambleAcc);
+}
+
+/*
+ * XXH3_hashLong_64b_withSeed():
+ * Generate a custom key based on alteration of default XXH3_kSecret with the seed,
+ * and then use this key for long mode hashing.
+ *
+ * This operation is decently fast but nonetheless costs a little bit of time.
+ * Try to avoid it whenever possible (typically when seed==0).
+ *
+ * It's important for performance that XXH3_hashLong is not inlined. Not sure
+ * why (uop cache maybe?), but the difference is large and easily measurable.
+ */
+XXH_FORCE_INLINE XXH64_hash_t
+XXH3_hashLong_64b_withSeed_internal(const void* input, size_t len,
+                                    XXH64_hash_t seed,
+                                    XXH3_f_accumulate f_acc,
+                                    XXH3_f_scrambleAcc f_scramble,
+                                    XXH3_f_initCustomSecret f_initSec)
+{
+#if XXH_SIZE_OPT <= 0
+    if (seed == 0)
+        return XXH3_hashLong_64b_internal(input, len,
+                                          XXH3_kSecret, sizeof(XXH3_kSecret),
+                                          f_acc, f_scramble);
+#endif
+    {   XXH_ALIGN(XXH_SEC_ALIGN) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE];
+        f_initSec(secret, seed);
+        return XXH3_hashLong_64b_internal(input, len, secret, sizeof(secret),
+                                          f_acc, f_scramble);
+    }
+}
+
+/*
+ * It's important for performance that XXH3_hashLong is not inlined.
+ */
+XXH_NO_INLINE XXH64_hash_t
+XXH3_hashLong_64b_withSeed(const void* XXH_RESTRICT input, size_t len,
+                           XXH64_hash_t seed, const xxh_u8* XXH_RESTRICT secret, size_t secretLen)
+{
+    (void)secret; (void)secretLen;
+    return XXH3_hashLong_64b_withSeed_internal(input, len, seed,
+                XXH3_accumulate, XXH3_scrambleAcc, XXH3_initCustomSecret);
+}
+
+
+typedef XXH64_hash_t (*XXH3_hashLong64_f)(const void* XXH_RESTRICT, size_t,
+                                          XXH64_hash_t, const xxh_u8* XXH_RESTRICT, size_t);
+
+XXH_FORCE_INLINE XXH64_hash_t
+XXH3_64bits_internal(const void* XXH_RESTRICT input, size_t len,
+                     XXH64_hash_t seed64, const void* XXH_RESTRICT secret, size_t secretLen,
+                     XXH3_hashLong64_f f_hashLong)
+{
+    XXH_ASSERT(secretLen >= XXH3_SECRET_SIZE_MIN);
+    /*
+     * If an action is to be taken if `secretLen` condition is not respected,
+     * it should be done here.
+     * For now, it's a contract pre-condition.
+     * Adding a check and a branch here would cost performance at every hash.
+     * Also, note that function signature doesn't offer room to return an error.
+     */
+    if (len <= 16)
+        return XXH3_len_0to16_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, seed64);
+    if (len <= 128)
+        return XXH3_len_17to128_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64);
+    if (len <= XXH3_MIDSIZE_MAX)
+        return XXH3_len_129to240_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64);
+    return f_hashLong(input, len, seed64, (const xxh_u8*)secret, secretLen);
+}
+
+
+/* ===   Public entry point   === */
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH64_hash_t XXH3_64bits(XXH_NOESCAPE const void* input, size_t length)
+{
+    return XXH3_64bits_internal(input, length, 0, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_default);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH64_hash_t
+XXH3_64bits_withSecret(XXH_NOESCAPE const void* input, size_t length, XXH_NOESCAPE const void* secret, size_t secretSize)
+{
+    return XXH3_64bits_internal(input, length, 0, secret, secretSize, XXH3_hashLong_64b_withSecret);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH64_hash_t
+XXH3_64bits_withSeed(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed)
+{
+    return XXH3_64bits_internal(input, length, seed, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_withSeed);
+}
+
+XXH_PUBLIC_API XXH64_hash_t
+XXH3_64bits_withSecretandSeed(XXH_NOESCAPE const void* input, size_t length, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed)
+{
+    if (length <= XXH3_MIDSIZE_MAX)
+        return XXH3_64bits_internal(input, length, seed, XXH3_kSecret, sizeof(XXH3_kSecret), NULL);
+    return XXH3_hashLong_64b_withSecret(input, length, seed, (const xxh_u8*)secret, secretSize);
+}
+
+
+/* ===   XXH3 streaming   === */
+#ifndef XXH_NO_STREAM
+/*
+ * Malloc's a pointer that is always aligned to align.
+ *
+ * This must be freed with `XXH_alignedFree()`.
+ *
+ * malloc typically guarantees 16 byte alignment on 64-bit systems and 8 byte
+ * alignment on 32-bit. This isn't enough for the 32 byte aligned loads in AVX2
+ * or on 32-bit, the 16 byte aligned loads in SSE2 and NEON.
+ *
+ * This underalignment previously caused a rather obvious crash which went
+ * completely unnoticed due to XXH3_createState() not actually being tested.
+ * Credit to RedSpah for noticing this bug.
+ *
+ * The alignment is done manually: Functions like posix_memalign or _mm_malloc
+ * are avoided: To maintain portability, we would have to write a fallback
+ * like this anyways, and besides, testing for the existence of library
+ * functions without relying on external build tools is impossible.
+ *
+ * The method is simple: Overallocate, manually align, and store the offset
+ * to the original behind the returned pointer.
+ *
+ * Align must be a power of 2 and 8 <= align <= 128.
+ */
+static XXH_MALLOCF void* XXH_alignedMalloc(size_t s, size_t align)
+{
+    XXH_ASSERT(align <= 128 && align >= 8); /* range check */
+    XXH_ASSERT((align & (align-1)) == 0);   /* power of 2 */
+    XXH_ASSERT(s != 0 && s < (s + align));  /* empty/overflow */
+    {   /* Overallocate to make room for manual realignment and an offset byte */
+        xxh_u8* base = (xxh_u8*)XXH_malloc(s + align);
+        if (base != NULL) {
+            /*
+             * Get the offset needed to align this pointer.
+             *
+             * Even if the returned pointer is aligned, there will always be
+             * at least one byte to store the offset to the original pointer.
+             */
+            size_t offset = align - ((size_t)base & (align - 1)); /* base % align */
+            /* Add the offset for the now-aligned pointer */
+            xxh_u8* ptr = base + offset;
+
+            XXH_ASSERT((size_t)ptr % align == 0);
+
+            /* Store the offset immediately before the returned pointer. */
+            ptr[-1] = (xxh_u8)offset;
+            return ptr;
+        }
+        return NULL;
+    }
+}
+/*
+ * Frees an aligned pointer allocated by XXH_alignedMalloc(). Don't pass
+ * normal malloc'd pointers, XXH_alignedMalloc has a specific data layout.
+ */
+static void XXH_alignedFree(void* p)
+{
+    if (p != NULL) {
+        xxh_u8* ptr = (xxh_u8*)p;
+        /* Get the offset byte we added in XXH_malloc. */
+        xxh_u8 offset = ptr[-1];
+        /* Free the original malloc'd pointer */
+        xxh_u8* base = ptr - offset;
+        XXH_free(base);
+    }
+}
+/*! @ingroup XXH3_family */
+/*!
+ * @brief Allocate an @ref XXH3_state_t.
+ *
+ * Must be freed with XXH3_freeState().
+ * @return An allocated XXH3_state_t on success, `NULL` on failure.
+ */
+XXH_PUBLIC_API XXH3_state_t* XXH3_createState(void)
+{
+    XXH3_state_t* const state = (XXH3_state_t*)XXH_alignedMalloc(sizeof(XXH3_state_t), 64);
+    if (state==NULL) return NULL;
+    XXH3_INITSTATE(state);
+    return state;
+}
+
+/*! @ingroup XXH3_family */
+/*!
+ * @brief Frees an @ref XXH3_state_t.
+ *
+ * Must be allocated with XXH3_createState().
+ * @param statePtr A pointer to an @ref XXH3_state_t allocated with @ref XXH3_createState().
+ * @return XXH_OK.
+ */
+XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr)
+{
+    XXH_alignedFree(statePtr);
+    return XXH_OK;
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API void
+XXH3_copyState(XXH_NOESCAPE XXH3_state_t* dst_state, XXH_NOESCAPE const XXH3_state_t* src_state)
+{
+    XXH_memcpy(dst_state, src_state, sizeof(*dst_state));
+}
+
+static void
+XXH3_reset_internal(XXH3_state_t* statePtr,
+                    XXH64_hash_t seed,
+                    const void* secret, size_t secretSize)
+{
+    size_t const initStart = offsetof(XXH3_state_t, bufferedSize);
+    size_t const initLength = offsetof(XXH3_state_t, nbStripesPerBlock) - initStart;
+    XXH_ASSERT(offsetof(XXH3_state_t, nbStripesPerBlock) > initStart);
+    XXH_ASSERT(statePtr != NULL);
+    /* set members from bufferedSize to nbStripesPerBlock (excluded) to 0 */
+    memset((char*)statePtr + initStart, 0, initLength);
+    statePtr->acc[0] = XXH_PRIME32_3;
+    statePtr->acc[1] = XXH_PRIME64_1;
+    statePtr->acc[2] = XXH_PRIME64_2;
+    statePtr->acc[3] = XXH_PRIME64_3;
+    statePtr->acc[4] = XXH_PRIME64_4;
+    statePtr->acc[5] = XXH_PRIME32_2;
+    statePtr->acc[6] = XXH_PRIME64_5;
+    statePtr->acc[7] = XXH_PRIME32_1;
+    statePtr->seed = seed;
+    statePtr->useSeed = (seed != 0);
+    statePtr->extSecret = (const unsigned char*)secret;
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
+    statePtr->secretLimit = secretSize - XXH_STRIPE_LEN;
+    statePtr->nbStripesPerBlock = statePtr->secretLimit / XXH_SECRET_CONSUME_RATE;
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr)
+{
+    if (statePtr == NULL) return XXH_ERROR;
+    XXH3_reset_internal(statePtr, 0, XXH3_kSecret, XXH_SECRET_DEFAULT_SIZE);
+    return XXH_OK;
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize)
+{
+    if (statePtr == NULL) return XXH_ERROR;
+    XXH3_reset_internal(statePtr, 0, secret, secretSize);
+    if (secret == NULL) return XXH_ERROR;
+    if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR;
+    return XXH_OK;
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed)
+{
+    if (statePtr == NULL) return XXH_ERROR;
+    if (seed==0) return XXH3_64bits_reset(statePtr);
+    if ((seed != statePtr->seed) || (statePtr->extSecret != NULL))
+        XXH3_initCustomSecret(statePtr->customSecret, seed);
+    XXH3_reset_internal(statePtr, seed, NULL, XXH_SECRET_DEFAULT_SIZE);
+    return XXH_OK;
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed64)
+{
+    if (statePtr == NULL) return XXH_ERROR;
+    if (secret == NULL) return XXH_ERROR;
+    if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR;
+    XXH3_reset_internal(statePtr, seed64, secret, secretSize);
+    statePtr->useSeed = 1; /* always, even if seed64==0 */
+    return XXH_OK;
+}
+
+/*!
+ * @internal
+ * @brief Processes a large input for XXH3_update() and XXH3_digest_long().
+ *
+ * Unlike XXH3_hashLong_internal_loop(), this can process data that overlaps a block.
+ *
+ * @param acc                Pointer to the 8 accumulator lanes
+ * @param nbStripesSoFarPtr  In/out pointer to the number of leftover stripes in the block*
+ * @param nbStripesPerBlock  Number of stripes in a block
+ * @param input              Input pointer
+ * @param nbStripes          Number of stripes to process
+ * @param secret             Secret pointer
+ * @param secretLimit        Offset of the last block in @p secret
+ * @param f_acc              Pointer to an XXH3_accumulate implementation
+ * @param f_scramble         Pointer to an XXH3_scrambleAcc implementation
+ * @return                   Pointer past the end of @p input after processing
+ */
+XXH_FORCE_INLINE const xxh_u8 *
+XXH3_consumeStripes(xxh_u64* XXH_RESTRICT acc,
+                    size_t* XXH_RESTRICT nbStripesSoFarPtr, size_t nbStripesPerBlock,
+                    const xxh_u8* XXH_RESTRICT input, size_t nbStripes,
+                    const xxh_u8* XXH_RESTRICT secret, size_t secretLimit,
+                    XXH3_f_accumulate f_acc,
+                    XXH3_f_scrambleAcc f_scramble)
+{
+    const xxh_u8* initialSecret = secret + *nbStripesSoFarPtr * XXH_SECRET_CONSUME_RATE;
+    /* Process full blocks */
+    if (nbStripes >= (nbStripesPerBlock - *nbStripesSoFarPtr)) {
+        /* Process the initial partial block... */
+        size_t nbStripesThisIter = nbStripesPerBlock - *nbStripesSoFarPtr;
+
+        do {
+            /* Accumulate and scramble */
+            f_acc(acc, input, initialSecret, nbStripesThisIter);
+            f_scramble(acc, secret + secretLimit);
+            input += nbStripesThisIter * XXH_STRIPE_LEN;
+            nbStripes -= nbStripesThisIter;
+            /* Then continue the loop with the full block size */
+            nbStripesThisIter = nbStripesPerBlock;
+            initialSecret = secret;
+        } while (nbStripes >= nbStripesPerBlock);
+        *nbStripesSoFarPtr = 0;
+    }
+    /* Process a partial block */
+    if (nbStripes > 0) {
+        f_acc(acc, input, initialSecret, nbStripes);
+        input += nbStripes * XXH_STRIPE_LEN;
+        *nbStripesSoFarPtr += nbStripes;
+    }
+    /* Return end pointer */
+    return input;
+}
+
+#ifndef XXH3_STREAM_USE_STACK
+# if XXH_SIZE_OPT <= 0 && !defined(__clang__) /* clang doesn't need additional stack space */
+#   define XXH3_STREAM_USE_STACK 1
+# endif
+#endif
+/*
+ * Both XXH3_64bits_update and XXH3_128bits_update use this routine.
+ */
+XXH_FORCE_INLINE XXH_errorcode
+XXH3_update(XXH3_state_t* XXH_RESTRICT const state,
+            const xxh_u8* XXH_RESTRICT input, size_t len,
+            XXH3_f_accumulate f_acc,
+            XXH3_f_scrambleAcc f_scramble)
+{
+    if (input==NULL) {
+        XXH_ASSERT(len == 0);
+        return XXH_OK;
+    }
+
+    XXH_ASSERT(state != NULL);
+    {   const xxh_u8* const bEnd = input + len;
+        const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret;
+#if defined(XXH3_STREAM_USE_STACK) && XXH3_STREAM_USE_STACK >= 1
+        /* For some reason, gcc and MSVC seem to suffer greatly
+         * when operating accumulators directly into state.
+         * Operating into stack space seems to enable proper optimization.
+         * clang, on the other hand, doesn't seem to need this trick */
+        XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[8];
+        XXH_memcpy(acc, state->acc, sizeof(acc));
+#else
+        xxh_u64* XXH_RESTRICT const acc = state->acc;
+#endif
+        state->totalLen += len;
+        XXH_ASSERT(state->bufferedSize <= XXH3_INTERNALBUFFER_SIZE);
+
+        /* small input : just fill in tmp buffer */
+        if (len <= XXH3_INTERNALBUFFER_SIZE - state->bufferedSize) {
+            XXH_memcpy(state->buffer + state->bufferedSize, input, len);
+            state->bufferedSize += (XXH32_hash_t)len;
+            return XXH_OK;
+        }
+
+        /* total input is now > XXH3_INTERNALBUFFER_SIZE */
+        #define XXH3_INTERNALBUFFER_STRIPES (XXH3_INTERNALBUFFER_SIZE / XXH_STRIPE_LEN)
+        XXH_STATIC_ASSERT(XXH3_INTERNALBUFFER_SIZE % XXH_STRIPE_LEN == 0);   /* clean multiple */
+
+        /*
+         * Internal buffer is partially filled (always, except at beginning)
+         * Complete it, then consume it.
+         */
+        if (state->bufferedSize) {
+            size_t const loadSize = XXH3_INTERNALBUFFER_SIZE - state->bufferedSize;
+            XXH_memcpy(state->buffer + state->bufferedSize, input, loadSize);
+            input += loadSize;
+            XXH3_consumeStripes(acc,
+                               &state->nbStripesSoFar, state->nbStripesPerBlock,
+                                state->buffer, XXH3_INTERNALBUFFER_STRIPES,
+                                secret, state->secretLimit,
+                                f_acc, f_scramble);
+            state->bufferedSize = 0;
+        }
+        XXH_ASSERT(input < bEnd);
+        if (bEnd - input > XXH3_INTERNALBUFFER_SIZE) {
+            size_t nbStripes = (size_t)(bEnd - 1 - input) / XXH_STRIPE_LEN;
+            input = XXH3_consumeStripes(acc,
+                                       &state->nbStripesSoFar, state->nbStripesPerBlock,
+                                       input, nbStripes,
+                                       secret, state->secretLimit,
+                                       f_acc, f_scramble);
+            XXH_memcpy(state->buffer + sizeof(state->buffer) - XXH_STRIPE_LEN, input - XXH_STRIPE_LEN, XXH_STRIPE_LEN);
+
+        }
+        /* Some remaining input (always) : buffer it */
+        XXH_ASSERT(input < bEnd);
+        XXH_ASSERT(bEnd - input <= XXH3_INTERNALBUFFER_SIZE);
+        XXH_ASSERT(state->bufferedSize == 0);
+        XXH_memcpy(state->buffer, input, (size_t)(bEnd-input));
+        state->bufferedSize = (XXH32_hash_t)(bEnd-input);
+#if defined(XXH3_STREAM_USE_STACK) && XXH3_STREAM_USE_STACK >= 1
+        /* save stack accumulators into state */
+        XXH_memcpy(state->acc, acc, sizeof(acc));
+#endif
+    }
+
+    return XXH_OK;
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_64bits_update(XXH_NOESCAPE XXH3_state_t* state, XXH_NOESCAPE const void* input, size_t len)
+{
+    return XXH3_update(state, (const xxh_u8*)input, len,
+                       XXH3_accumulate, XXH3_scrambleAcc);
+}
+
+
+XXH_FORCE_INLINE void
+XXH3_digest_long (XXH64_hash_t* acc,
+                  const XXH3_state_t* state,
+                  const unsigned char* secret)
+{
+    xxh_u8 lastStripe[XXH_STRIPE_LEN];
+    const xxh_u8* lastStripePtr;
+
+    /*
+     * Digest on a local copy. This way, the state remains unaltered, and it can
+     * continue ingesting more input afterwards.
+     */
+    XXH_memcpy(acc, state->acc, sizeof(state->acc));
+    if (state->bufferedSize >= XXH_STRIPE_LEN) {
+        /* Consume remaining stripes then point to remaining data in buffer */
+        size_t const nbStripes = (state->bufferedSize - 1) / XXH_STRIPE_LEN;
+        size_t nbStripesSoFar = state->nbStripesSoFar;
+        XXH3_consumeStripes(acc,
+                           &nbStripesSoFar, state->nbStripesPerBlock,
+                            state->buffer, nbStripes,
+                            secret, state->secretLimit,
+                            XXH3_accumulate, XXH3_scrambleAcc);
+        lastStripePtr = state->buffer + state->bufferedSize - XXH_STRIPE_LEN;
+    } else {  /* bufferedSize < XXH_STRIPE_LEN */
+        /* Copy to temp buffer */
+        size_t const catchupSize = XXH_STRIPE_LEN - state->bufferedSize;
+        XXH_ASSERT(state->bufferedSize > 0);  /* there is always some input buffered */
+        XXH_memcpy(lastStripe, state->buffer + sizeof(state->buffer) - catchupSize, catchupSize);
+        XXH_memcpy(lastStripe + catchupSize, state->buffer, state->bufferedSize);
+        lastStripePtr = lastStripe;
+    }
+    /* Last stripe */
+    XXH3_accumulate_512(acc,
+                        lastStripePtr,
+                        secret + state->secretLimit - XXH_SECRET_LASTACC_START);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_digest (XXH_NOESCAPE const XXH3_state_t* state)
+{
+    const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret;
+    if (state->totalLen > XXH3_MIDSIZE_MAX) {
+        XXH_ALIGN(XXH_ACC_ALIGN) XXH64_hash_t acc[XXH_ACC_NB];
+        XXH3_digest_long(acc, state, secret);
+        return XXH3_mergeAccs(acc,
+                              secret + XXH_SECRET_MERGEACCS_START,
+                              (xxh_u64)state->totalLen * XXH_PRIME64_1);
+    }
+    /* totalLen <= XXH3_MIDSIZE_MAX: digesting a short input */
+    if (state->useSeed)
+        return XXH3_64bits_withSeed(state->buffer, (size_t)state->totalLen, state->seed);
+    return XXH3_64bits_withSecret(state->buffer, (size_t)(state->totalLen),
+                                  secret, state->secretLimit + XXH_STRIPE_LEN);
+}
+#endif /* !XXH_NO_STREAM */
+
+
+/* ==========================================
+ * XXH3 128 bits (a.k.a XXH128)
+ * ==========================================
+ * XXH3's 128-bit variant has better mixing and strength than the 64-bit variant,
+ * even without counting the significantly larger output size.
+ *
+ * For example, extra steps are taken to avoid the seed-dependent collisions
+ * in 17-240 byte inputs (See XXH3_mix16B and XXH128_mix32B).
+ *
+ * This strength naturally comes at the cost of some speed, especially on short
+ * lengths. Note that longer hashes are about as fast as the 64-bit version
+ * due to it using only a slight modification of the 64-bit loop.
+ *
+ * XXH128 is also more oriented towards 64-bit machines. It is still extremely
+ * fast for a _128-bit_ hash on 32-bit (it usually clears XXH64).
+ */
+
+XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
+XXH3_len_1to3_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+    /* A doubled version of 1to3_64b with different constants. */
+    XXH_ASSERT(input != NULL);
+    XXH_ASSERT(1 <= len && len <= 3);
+    XXH_ASSERT(secret != NULL);
+    /*
+     * len = 1: combinedl = { input[0], 0x01, input[0], input[0] }
+     * len = 2: combinedl = { input[1], 0x02, input[0], input[1] }
+     * len = 3: combinedl = { input[2], 0x03, input[0], input[1] }
+     */
+    {   xxh_u8 const c1 = input[0];
+        xxh_u8 const c2 = input[len >> 1];
+        xxh_u8 const c3 = input[len - 1];
+        xxh_u32 const combinedl = ((xxh_u32)c1 <<16) | ((xxh_u32)c2 << 24)
+                                | ((xxh_u32)c3 << 0) | ((xxh_u32)len << 8);
+        xxh_u32 const combinedh = XXH_rotl32(XXH_swap32(combinedl), 13);
+        xxh_u64 const bitflipl = (XXH_readLE32(secret) ^ XXH_readLE32(secret+4)) + seed;
+        xxh_u64 const bitfliph = (XXH_readLE32(secret+8) ^ XXH_readLE32(secret+12)) - seed;
+        xxh_u64 const keyed_lo = (xxh_u64)combinedl ^ bitflipl;
+        xxh_u64 const keyed_hi = (xxh_u64)combinedh ^ bitfliph;
+        XXH128_hash_t h128;
+        h128.low64  = XXH64_avalanche(keyed_lo);
+        h128.high64 = XXH64_avalanche(keyed_hi);
+        return h128;
+    }
+}
+
+XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
+XXH3_len_4to8_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+    XXH_ASSERT(input != NULL);
+    XXH_ASSERT(secret != NULL);
+    XXH_ASSERT(4 <= len && len <= 8);
+    seed ^= (xxh_u64)XXH_swap32((xxh_u32)seed) << 32;
+    {   xxh_u32 const input_lo = XXH_readLE32(input);
+        xxh_u32 const input_hi = XXH_readLE32(input + len - 4);
+        xxh_u64 const input_64 = input_lo + ((xxh_u64)input_hi << 32);
+        xxh_u64 const bitflip = (XXH_readLE64(secret+16) ^ XXH_readLE64(secret+24)) + seed;
+        xxh_u64 const keyed = input_64 ^ bitflip;
+
+        /* Shift len to the left to ensure it is even, this avoids even multiplies. */
+        XXH128_hash_t m128 = XXH_mult64to128(keyed, XXH_PRIME64_1 + (len << 2));
+
+        m128.high64 += (m128.low64 << 1);
+        m128.low64  ^= (m128.high64 >> 3);
+
+        m128.low64   = XXH_xorshift64(m128.low64, 35);
+        m128.low64  *= PRIME_MX2;
+        m128.low64   = XXH_xorshift64(m128.low64, 28);
+        m128.high64  = XXH3_avalanche(m128.high64);
+        return m128;
+    }
+}
+
+XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
+XXH3_len_9to16_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+    XXH_ASSERT(input != NULL);
+    XXH_ASSERT(secret != NULL);
+    XXH_ASSERT(9 <= len && len <= 16);
+    {   xxh_u64 const bitflipl = (XXH_readLE64(secret+32) ^ XXH_readLE64(secret+40)) - seed;
+        xxh_u64 const bitfliph = (XXH_readLE64(secret+48) ^ XXH_readLE64(secret+56)) + seed;
+        xxh_u64 const input_lo = XXH_readLE64(input);
+        xxh_u64       input_hi = XXH_readLE64(input + len - 8);
+        XXH128_hash_t m128 = XXH_mult64to128(input_lo ^ input_hi ^ bitflipl, XXH_PRIME64_1);
+        /*
+         * Put len in the middle of m128 to ensure that the length gets mixed to
+         * both the low and high bits in the 128x64 multiply below.
+         */
+        m128.low64 += (xxh_u64)(len - 1) << 54;
+        input_hi   ^= bitfliph;
+        /*
+         * Add the high 32 bits of input_hi to the high 32 bits of m128, then
+         * add the long product of the low 32 bits of input_hi and XXH_PRIME32_2 to
+         * the high 64 bits of m128.
+         *
+         * The best approach to this operation is different on 32-bit and 64-bit.
+         */
+        if (sizeof(void *) < sizeof(xxh_u64)) { /* 32-bit */
+            /*
+             * 32-bit optimized version, which is more readable.
+             *
+             * On 32-bit, it removes an ADC and delays a dependency between the two
+             * halves of m128.high64, but it generates an extra mask on 64-bit.
+             */
+            m128.high64 += (input_hi & 0xFFFFFFFF00000000ULL) + XXH_mult32to64((xxh_u32)input_hi, XXH_PRIME32_2);
+        } else {
+            /*
+             * 64-bit optimized (albeit more confusing) version.
+             *
+             * Uses some properties of addition and multiplication to remove the mask:
+             *
+             * Let:
+             *    a = input_hi.lo = (input_hi & 0x00000000FFFFFFFF)
+             *    b = input_hi.hi = (input_hi & 0xFFFFFFFF00000000)
+             *    c = XXH_PRIME32_2
+             *
+             *    a + (b * c)
+             * Inverse Property: x + y - x == y
+             *    a + (b * (1 + c - 1))
+             * Distributive Property: x * (y + z) == (x * y) + (x * z)
+             *    a + (b * 1) + (b * (c - 1))
+             * Identity Property: x * 1 == x
+             *    a + b + (b * (c - 1))
+             *
+             * Substitute a, b, and c:
+             *    input_hi.hi + input_hi.lo + ((xxh_u64)input_hi.lo * (XXH_PRIME32_2 - 1))
+             *
+             * Since input_hi.hi + input_hi.lo == input_hi, we get this:
+             *    input_hi + ((xxh_u64)input_hi.lo * (XXH_PRIME32_2 - 1))
+             */
+            m128.high64 += input_hi + XXH_mult32to64((xxh_u32)input_hi, XXH_PRIME32_2 - 1);
+        }
+        /* m128 ^= XXH_swap64(m128 >> 64); */
+        m128.low64  ^= XXH_swap64(m128.high64);
+
+        {   /* 128x64 multiply: h128 = m128 * XXH_PRIME64_2; */
+            XXH128_hash_t h128 = XXH_mult64to128(m128.low64, XXH_PRIME64_2);
+            h128.high64 += m128.high64 * XXH_PRIME64_2;
+
+            h128.low64   = XXH3_avalanche(h128.low64);
+            h128.high64  = XXH3_avalanche(h128.high64);
+            return h128;
+    }   }
+}
+
+/*
+ * Assumption: `secret` size is >= XXH3_SECRET_SIZE_MIN
+ */
+XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
+XXH3_len_0to16_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
+{
+    XXH_ASSERT(len <= 16);
+    {   if (len > 8) return XXH3_len_9to16_128b(input, len, secret, seed);
+        if (len >= 4) return XXH3_len_4to8_128b(input, len, secret, seed);
+        if (len) return XXH3_len_1to3_128b(input, len, secret, seed);
+        {   XXH128_hash_t h128;
+            xxh_u64 const bitflipl = XXH_readLE64(secret+64) ^ XXH_readLE64(secret+72);
+            xxh_u64 const bitfliph = XXH_readLE64(secret+80) ^ XXH_readLE64(secret+88);
+            h128.low64 = XXH64_avalanche(seed ^ bitflipl);
+            h128.high64 = XXH64_avalanche( seed ^ bitfliph);
+            return h128;
+    }   }
+}
+
+/*
+ * A bit slower than XXH3_mix16B, but handles multiply by zero better.
+ */
+XXH_FORCE_INLINE XXH128_hash_t
+XXH128_mix32B(XXH128_hash_t acc, const xxh_u8* input_1, const xxh_u8* input_2,
+              const xxh_u8* secret, XXH64_hash_t seed)
+{
+    acc.low64  += XXH3_mix16B (input_1, secret+0, seed);
+    acc.low64  ^= XXH_readLE64(input_2) + XXH_readLE64(input_2 + 8);
+    acc.high64 += XXH3_mix16B (input_2, secret+16, seed);
+    acc.high64 ^= XXH_readLE64(input_1) + XXH_readLE64(input_1 + 8);
+    return acc;
+}
+
+
+XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
+XXH3_len_17to128_128b(const xxh_u8* XXH_RESTRICT input, size_t len,
+                      const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+                      XXH64_hash_t seed)
+{
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
+    XXH_ASSERT(16 < len && len <= 128);
+
+    {   XXH128_hash_t acc;
+        acc.low64 = len * XXH_PRIME64_1;
+        acc.high64 = 0;
+
+#if XXH_SIZE_OPT >= 1
+        {
+            /* Smaller, but slightly slower. */
+            unsigned int i = (unsigned int)(len - 1) / 32;
+            do {
+                acc = XXH128_mix32B(acc, input+16*i, input+len-16*(i+1), secret+32*i, seed);
+            } while (i-- != 0);
+        }
+#else
+        if (len > 32) {
+            if (len > 64) {
+                if (len > 96) {
+                    acc = XXH128_mix32B(acc, input+48, input+len-64, secret+96, seed);
+                }
+                acc = XXH128_mix32B(acc, input+32, input+len-48, secret+64, seed);
+            }
+            acc = XXH128_mix32B(acc, input+16, input+len-32, secret+32, seed);
+        }
+        acc = XXH128_mix32B(acc, input, input+len-16, secret, seed);
+#endif
+        {   XXH128_hash_t h128;
+            h128.low64  = acc.low64 + acc.high64;
+            h128.high64 = (acc.low64    * XXH_PRIME64_1)
+                        + (acc.high64   * XXH_PRIME64_4)
+                        + ((len - seed) * XXH_PRIME64_2);
+            h128.low64  = XXH3_avalanche(h128.low64);
+            h128.high64 = (XXH64_hash_t)0 - XXH3_avalanche(h128.high64);
+            return h128;
+        }
+    }
+}
+
+XXH_NO_INLINE XXH_PUREF XXH128_hash_t
+XXH3_len_129to240_128b(const xxh_u8* XXH_RESTRICT input, size_t len,
+                       const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+                       XXH64_hash_t seed)
+{
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
+    XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX);
+
+    {   XXH128_hash_t acc;
+        unsigned i;
+        acc.low64 = len * XXH_PRIME64_1;
+        acc.high64 = 0;
+        /*
+         *  We set as `i` as offset + 32. We do this so that unchanged
+         * `len` can be used as upper bound. This reaches a sweet spot
+         * where both x86 and aarch64 get simple agen and good codegen
+         * for the loop.
+         */
+        for (i = 32; i < 160; i += 32) {
+            acc = XXH128_mix32B(acc,
+                                input  + i - 32,
+                                input  + i - 16,
+                                secret + i - 32,
+                                seed);
+        }
+        acc.low64 = XXH3_avalanche(acc.low64);
+        acc.high64 = XXH3_avalanche(acc.high64);
+        /*
+         * NB: `i <= len` will duplicate the last 32-bytes if
+         * len % 32 was zero. This is an unfortunate necessity to keep
+         * the hash result stable.
+         */
+        for (i=160; i <= len; i += 32) {
+            acc = XXH128_mix32B(acc,
+                                input + i - 32,
+                                input + i - 16,
+                                secret + XXH3_MIDSIZE_STARTOFFSET + i - 160,
+                                seed);
+        }
+        /* last bytes */
+        acc = XXH128_mix32B(acc,
+                            input + len - 16,
+                            input + len - 32,
+                            secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET - 16,
+                            (XXH64_hash_t)0 - seed);
+
+        {   XXH128_hash_t h128;
+            h128.low64  = acc.low64 + acc.high64;
+            h128.high64 = (acc.low64    * XXH_PRIME64_1)
+                        + (acc.high64   * XXH_PRIME64_4)
+                        + ((len - seed) * XXH_PRIME64_2);
+            h128.low64  = XXH3_avalanche(h128.low64);
+            h128.high64 = (XXH64_hash_t)0 - XXH3_avalanche(h128.high64);
+            return h128;
+        }
+    }
+}
+
+XXH_FORCE_INLINE XXH128_hash_t
+XXH3_hashLong_128b_internal(const void* XXH_RESTRICT input, size_t len,
+                            const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
+                            XXH3_f_accumulate f_acc,
+                            XXH3_f_scrambleAcc f_scramble)
+{
+    XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[XXH_ACC_NB] = XXH3_INIT_ACC;
+
+    XXH3_hashLong_internal_loop(acc, (const xxh_u8*)input, len, secret, secretSize, f_acc, f_scramble);
+
+    /* converge into final hash */
+    XXH_STATIC_ASSERT(sizeof(acc) == 64);
+    XXH_ASSERT(secretSize >= sizeof(acc) + XXH_SECRET_MERGEACCS_START);
+    {   XXH128_hash_t h128;
+        h128.low64  = XXH3_mergeAccs(acc,
+                                     secret + XXH_SECRET_MERGEACCS_START,
+                                     (xxh_u64)len * XXH_PRIME64_1);
+        h128.high64 = XXH3_mergeAccs(acc,
+                                     secret + secretSize
+                                            - sizeof(acc) - XXH_SECRET_MERGEACCS_START,
+                                     ~((xxh_u64)len * XXH_PRIME64_2));
+        return h128;
+    }
+}
+
+/*
+ * It's important for performance that XXH3_hashLong() is not inlined.
+ */
+XXH_NO_INLINE XXH_PUREF XXH128_hash_t
+XXH3_hashLong_128b_default(const void* XXH_RESTRICT input, size_t len,
+                           XXH64_hash_t seed64,
+                           const void* XXH_RESTRICT secret, size_t secretLen)
+{
+    (void)seed64; (void)secret; (void)secretLen;
+    return XXH3_hashLong_128b_internal(input, len, XXH3_kSecret, sizeof(XXH3_kSecret),
+                                       XXH3_accumulate, XXH3_scrambleAcc);
+}
+
+/*
+ * It's important for performance to pass @p secretLen (when it's static)
+ * to the compiler, so that it can properly optimize the vectorized loop.
+ *
+ * When the secret size is unknown, or on GCC 12 where the mix of NO_INLINE and FORCE_INLINE
+ * breaks -Og, this is XXH_NO_INLINE.
+ */
+XXH3_WITH_SECRET_INLINE XXH128_hash_t
+XXH3_hashLong_128b_withSecret(const void* XXH_RESTRICT input, size_t len,
+                              XXH64_hash_t seed64,
+                              const void* XXH_RESTRICT secret, size_t secretLen)
+{
+    (void)seed64;
+    return XXH3_hashLong_128b_internal(input, len, (const xxh_u8*)secret, secretLen,
+                                       XXH3_accumulate, XXH3_scrambleAcc);
+}
+
+XXH_FORCE_INLINE XXH128_hash_t
+XXH3_hashLong_128b_withSeed_internal(const void* XXH_RESTRICT input, size_t len,
+                                XXH64_hash_t seed64,
+                                XXH3_f_accumulate f_acc,
+                                XXH3_f_scrambleAcc f_scramble,
+                                XXH3_f_initCustomSecret f_initSec)
+{
+    if (seed64 == 0)
+        return XXH3_hashLong_128b_internal(input, len,
+                                           XXH3_kSecret, sizeof(XXH3_kSecret),
+                                           f_acc, f_scramble);
+    {   XXH_ALIGN(XXH_SEC_ALIGN) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE];
+        f_initSec(secret, seed64);
+        return XXH3_hashLong_128b_internal(input, len, (const xxh_u8*)secret, sizeof(secret),
+                                           f_acc, f_scramble);
+    }
+}
+
+/*
+ * It's important for performance that XXH3_hashLong is not inlined.
+ */
+XXH_NO_INLINE XXH128_hash_t
+XXH3_hashLong_128b_withSeed(const void* input, size_t len,
+                            XXH64_hash_t seed64, const void* XXH_RESTRICT secret, size_t secretLen)
+{
+    (void)secret; (void)secretLen;
+    return XXH3_hashLong_128b_withSeed_internal(input, len, seed64,
+                XXH3_accumulate, XXH3_scrambleAcc, XXH3_initCustomSecret);
+}
+
+typedef XXH128_hash_t (*XXH3_hashLong128_f)(const void* XXH_RESTRICT, size_t,
+                                            XXH64_hash_t, const void* XXH_RESTRICT, size_t);
+
+XXH_FORCE_INLINE XXH128_hash_t
+XXH3_128bits_internal(const void* input, size_t len,
+                      XXH64_hash_t seed64, const void* XXH_RESTRICT secret, size_t secretLen,
+                      XXH3_hashLong128_f f_hl128)
+{
+    XXH_ASSERT(secretLen >= XXH3_SECRET_SIZE_MIN);
+    /*
+     * If an action is to be taken if `secret` conditions are not respected,
+     * it should be done here.
+     * For now, it's a contract pre-condition.
+     * Adding a check and a branch here would cost performance at every hash.
+     */
+    if (len <= 16)
+        return XXH3_len_0to16_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, seed64);
+    if (len <= 128)
+        return XXH3_len_17to128_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64);
+    if (len <= XXH3_MIDSIZE_MAX)
+        return XXH3_len_129to240_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64);
+    return f_hl128(input, len, seed64, secret, secretLen);
+}
+
+
+/* ===   Public XXH128 API   === */
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(XXH_NOESCAPE const void* input, size_t len)
+{
+    return XXH3_128bits_internal(input, len, 0,
+                                 XXH3_kSecret, sizeof(XXH3_kSecret),
+                                 XXH3_hashLong_128b_default);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t
+XXH3_128bits_withSecret(XXH_NOESCAPE const void* input, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize)
+{
+    return XXH3_128bits_internal(input, len, 0,
+                                 (const xxh_u8*)secret, secretSize,
+                                 XXH3_hashLong_128b_withSecret);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t
+XXH3_128bits_withSeed(XXH_NOESCAPE const void* input, size_t len, XXH64_hash_t seed)
+{
+    return XXH3_128bits_internal(input, len, seed,
+                                 XXH3_kSecret, sizeof(XXH3_kSecret),
+                                 XXH3_hashLong_128b_withSeed);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t
+XXH3_128bits_withSecretandSeed(XXH_NOESCAPE const void* input, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed)
+{
+    if (len <= XXH3_MIDSIZE_MAX)
+        return XXH3_128bits_internal(input, len, seed, XXH3_kSecret, sizeof(XXH3_kSecret), NULL);
+    return XXH3_hashLong_128b_withSecret(input, len, seed, secret, secretSize);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t
+XXH128(XXH_NOESCAPE const void* input, size_t len, XXH64_hash_t seed)
+{
+    return XXH3_128bits_withSeed(input, len, seed);
+}
+
+
+/* ===   XXH3 128-bit streaming   === */
+#ifndef XXH_NO_STREAM
+/*
+ * All initialization and update functions are identical to 64-bit streaming variant.
+ * The only difference is the finalization routine.
+ */
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr)
+{
+    return XXH3_64bits_reset(statePtr);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize)
+{
+    return XXH3_64bits_reset_withSecret(statePtr, secret, secretSize);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed)
+{
+    return XXH3_64bits_reset_withSeed(statePtr, seed);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed)
+{
+    return XXH3_64bits_reset_withSecretandSeed(statePtr, secret, secretSize, seed);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_128bits_update(XXH_NOESCAPE XXH3_state_t* state, XXH_NOESCAPE const void* input, size_t len)
+{
+    return XXH3_64bits_update(state, input, len);
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest (XXH_NOESCAPE const XXH3_state_t* state)
+{
+    const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret;
+    if (state->totalLen > XXH3_MIDSIZE_MAX) {
+        XXH_ALIGN(XXH_ACC_ALIGN) XXH64_hash_t acc[XXH_ACC_NB];
+        XXH3_digest_long(acc, state, secret);
+        XXH_ASSERT(state->secretLimit + XXH_STRIPE_LEN >= sizeof(acc) + XXH_SECRET_MERGEACCS_START);
+        {   XXH128_hash_t h128;
+            h128.low64  = XXH3_mergeAccs(acc,
+                                         secret + XXH_SECRET_MERGEACCS_START,
+                                         (xxh_u64)state->totalLen * XXH_PRIME64_1);
+            h128.high64 = XXH3_mergeAccs(acc,
+                                         secret + state->secretLimit + XXH_STRIPE_LEN
+                                                - sizeof(acc) - XXH_SECRET_MERGEACCS_START,
+                                         ~((xxh_u64)state->totalLen * XXH_PRIME64_2));
+            return h128;
+        }
+    }
+    /* len <= XXH3_MIDSIZE_MAX : short code */
+    if (state->seed)
+        return XXH3_128bits_withSeed(state->buffer, (size_t)state->totalLen, state->seed);
+    return XXH3_128bits_withSecret(state->buffer, (size_t)(state->totalLen),
+                                   secret, state->secretLimit + XXH_STRIPE_LEN);
+}
+#endif /* !XXH_NO_STREAM */
+/* 128-bit utility functions */
+
+#include <string.h>   /* memcmp, memcpy */
+
+/* return : 1 is equal, 0 if different */
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2)
+{
+    /* note : XXH128_hash_t is compact, it has no padding byte */
+    return !(memcmp(&h1, &h2, sizeof(h1)));
+}
+
+/* This prototype is compatible with stdlib's qsort().
+ * @return : >0 if *h128_1  > *h128_2
+ *           <0 if *h128_1  < *h128_2
+ *           =0 if *h128_1 == *h128_2  */
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API int XXH128_cmp(XXH_NOESCAPE const void* h128_1, XXH_NOESCAPE const void* h128_2)
+{
+    XXH128_hash_t const h1 = *(const XXH128_hash_t*)h128_1;
+    XXH128_hash_t const h2 = *(const XXH128_hash_t*)h128_2;
+    int const hcmp = (h1.high64 > h2.high64) - (h2.high64 > h1.high64);
+    /* note : bets that, in most cases, hash values are different */
+    if (hcmp) return hcmp;
+    return (h1.low64 > h2.low64) - (h2.low64 > h1.low64);
+}
+
+
+/*======   Canonical representation   ======*/
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API void
+XXH128_canonicalFromHash(XXH_NOESCAPE XXH128_canonical_t* dst, XXH128_hash_t hash)
+{
+    XXH_STATIC_ASSERT(sizeof(XXH128_canonical_t) == sizeof(XXH128_hash_t));
+    if (XXH_CPU_LITTLE_ENDIAN) {
+        hash.high64 = XXH_swap64(hash.high64);
+        hash.low64  = XXH_swap64(hash.low64);
+    }
+    XXH_memcpy(dst, &hash.high64, sizeof(hash.high64));
+    XXH_memcpy((char*)dst + sizeof(hash.high64), &hash.low64, sizeof(hash.low64));
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH128_hash_t
+XXH128_hashFromCanonical(XXH_NOESCAPE const XXH128_canonical_t* src)
+{
+    XXH128_hash_t h;
+    h.high64 = XXH_readBE64(src);
+    h.low64  = XXH_readBE64(src->digest + 8);
+    return h;
+}
+
+
+
+/* ==========================================
+ * Secret generators
+ * ==========================================
+ */
+#define XXH_MIN(x, y) (((x) > (y)) ? (y) : (x))
+
+XXH_FORCE_INLINE void XXH3_combine16(void* dst, XXH128_hash_t h128)
+{
+    XXH_writeLE64( dst, XXH_readLE64(dst) ^ h128.low64 );
+    XXH_writeLE64( (char*)dst+8, XXH_readLE64((char*)dst+8) ^ h128.high64 );
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API XXH_errorcode
+XXH3_generateSecret(XXH_NOESCAPE void* secretBuffer, size_t secretSize, XXH_NOESCAPE const void* customSeed, size_t customSeedSize)
+{
+#if (XXH_DEBUGLEVEL >= 1)
+    XXH_ASSERT(secretBuffer != NULL);
+    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
+#else
+    /* production mode, assert() are disabled */
+    if (secretBuffer == NULL) return XXH_ERROR;
+    if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR;
+#endif
+
+    if (customSeedSize == 0) {
+        customSeed = XXH3_kSecret;
+        customSeedSize = XXH_SECRET_DEFAULT_SIZE;
+    }
+#if (XXH_DEBUGLEVEL >= 1)
+    XXH_ASSERT(customSeed != NULL);
+#else
+    if (customSeed == NULL) return XXH_ERROR;
+#endif
+
+    /* Fill secretBuffer with a copy of customSeed - repeat as needed */
+    {   size_t pos = 0;
+        while (pos < secretSize) {
+            size_t const toCopy = XXH_MIN((secretSize - pos), customSeedSize);
+            memcpy((char*)secretBuffer + pos, customSeed, toCopy);
+            pos += toCopy;
+    }   }
+
+    {   size_t const nbSeg16 = secretSize / 16;
+        size_t n;
+        XXH128_canonical_t scrambler;
+        XXH128_canonicalFromHash(&scrambler, XXH128(customSeed, customSeedSize, 0));
+        for (n=0; n<nbSeg16; n++) {
+            XXH128_hash_t const h128 = XXH128(&scrambler, sizeof(scrambler), n);
+            XXH3_combine16((char*)secretBuffer + n*16, h128);
+        }
+        /* last segment */
+        XXH3_combine16((char*)secretBuffer + secretSize - 16, XXH128_hashFromCanonical(&scrambler));
+    }
+    return XXH_OK;
+}
+
+/*! @ingroup XXH3_family */
+XXH_PUBLIC_API void
+XXH3_generateSecret_fromSeed(XXH_NOESCAPE void* secretBuffer, XXH64_hash_t seed)
+{
+    XXH_ALIGN(XXH_SEC_ALIGN) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE];
+    XXH3_initCustomSecret(secret, seed);
+    XXH_ASSERT(secretBuffer != NULL);
+    memcpy(secretBuffer, secret, XXH_SECRET_DEFAULT_SIZE);
+}
+
+
+
+/* Pop our optimization override from above */
+#if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \
+  && defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
+  && defined(__OPTIMIZE__) && XXH_SIZE_OPT <= 0 /* respect -O0 and -Os */
+#  pragma GCC pop_options
+#endif
+
+#endif  /* XXH_NO_LONG_LONG */
+
+#endif  /* XXH_NO_XXH3 */
+
+/*!
+ * @}
+ */
+#endif  /* XXH_IMPLEMENTATION */
+
+
+#if defined (__cplusplus)
+} /* extern "C" */
+#endif