Merge branch 'bench_endoders'

benches/encoders.cm

@@ -0,0 +1,405 @@
+// encoders.cm — nota/wota/json encode+decode benchmark
+// Isolates per-type bottlenecks across all three serializers.
+
+var nota = use('nota')
+var wota = use('wota')
+var json = use('json')
+
+// --- Test data shapes ---
+
+// Small integers: fast path for all encoders
+var integers_small = array(100, function(i) { return i + 1 })
+
+// Floats: stresses nota's snprintf path
+var floats_array = array(100, function(i) {
+  return 3.14159 * (i + 1) + 0.00001 * i
+})
+
+// Short strings in records: per-string overhead, property enumeration
+var strings_short = array(50, function(i) {
+  var r = {}
+  r[`k${i}`] = `value_${i}`
+  return r
+})
+
+// Single long string: throughput test (wota byte loop, nota kim)
+var long_str = ""
+var li = 0
+for (li = 0; li < 100; li++) {
+  long_str = `${long_str}abcdefghijklmnopqrstuvwxyz0123456789ABCDEFGHIJKLMN`
+}
+var strings_long = long_str
+
+// Unicode text: nota's kim encoding, wota's byte packing
+var strings_unicode = "こんにちは世界 🌍🌎🌏 Ñoño café résumé naïve Ω∑∏ 你好世界"
+
+// Nested records: cycle detection, property enumeration
+function make_nested(depth, breadth) {
+  var obj = {}
+  var i = 0
+  var k = null
+  if (depth <= 0) {
+    for (i = 0; i < breadth; i++) {
+      k = `v${i}`
+      obj[k] = i * 2.5
+    }
+    return obj
+  }
+  for (i = 0; i < breadth; i++) {
+    k = `n${i}`
+    obj[k] = make_nested(depth - 1, breadth)
+  }
+  return obj
+}
+var nested_records = make_nested(3, 4)
+
+// Flat record: property enumeration cost
+var flat_record = {}
+var fi = 0
+for (fi = 0; fi < 50; fi++) {
+  flat_record[`prop_${fi}`] = fi * 1.1
+}
+
+// Mixed payload: realistic workload
+var mixed_payload = array(50, function(i) {
+  var r = {}
+  r.id = i
+  r.name = `item_${i}`
+  r.active = i % 2 == 0
+  r.score = i * 3.14
+  r.tags = [`t${i % 5}`, `t${(i + 1) % 5}`]
+  return r
+})
+
+// --- Pre-encode for decode benchmarks ---
+
+var nota_enc_integers = nota.encode(integers_small)
+var nota_enc_floats = nota.encode(floats_array)
+var nota_enc_strings_short = nota.encode(strings_short)
+var nota_enc_strings_long = nota.encode(strings_long)
+var nota_enc_strings_unicode = nota.encode(strings_unicode)
+var nota_enc_nested = nota.encode(nested_records)
+var nota_enc_flat = nota.encode(flat_record)
+var nota_enc_mixed = nota.encode(mixed_payload)
+
+var wota_enc_integers = wota.encode(integers_small)
+var wota_enc_floats = wota.encode(floats_array)
+var wota_enc_strings_short = wota.encode(strings_short)
+var wota_enc_strings_long = wota.encode(strings_long)
+var wota_enc_strings_unicode = wota.encode(strings_unicode)
+var wota_enc_nested = wota.encode(nested_records)
+var wota_enc_flat = wota.encode(flat_record)
+var wota_enc_mixed = wota.encode(mixed_payload)
+
+var json_enc_integers = json.encode(integers_small)
+var json_enc_floats = json.encode(floats_array)
+var json_enc_strings_short = json.encode(strings_short)
+var json_enc_strings_long = json.encode(strings_long)
+var json_enc_strings_unicode = json.encode(strings_unicode)
+var json_enc_nested = json.encode(nested_records)
+var json_enc_flat = json.encode(flat_record)
+var json_enc_mixed = json.encode(mixed_payload)
+
+// --- Benchmark functions ---
+
+return {
+  // NOTA encode
+  nota_encode_integers: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = nota.encode(integers_small) }
+    return r
+  },
+  nota_encode_floats: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = nota.encode(floats_array) }
+    return r
+  },
+  nota_encode_strings_short: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = nota.encode(strings_short) }
+    return r
+  },
+  nota_encode_strings_long: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = nota.encode(strings_long) }
+    return r
+  },
+  nota_encode_strings_unicode: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = nota.encode(strings_unicode) }
+    return r
+  },
+  nota_encode_nested: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = nota.encode(nested_records) }
+    return r
+  },
+  nota_encode_flat: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = nota.encode(flat_record) }
+    return r
+  },
+  nota_encode_mixed: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = nota.encode(mixed_payload) }
+    return r
+  },
+
+  // NOTA decode
+  nota_decode_integers: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = nota.decode(nota_enc_integers) }
+    return r
+  },
+  nota_decode_floats: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = nota.decode(nota_enc_floats) }
+    return r
+  },
+  nota_decode_strings_short: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = nota.decode(nota_enc_strings_short) }
+    return r
+  },
+  nota_decode_strings_long: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = nota.decode(nota_enc_strings_long) }
+    return r
+  },
+  nota_decode_strings_unicode: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = nota.decode(nota_enc_strings_unicode) }
+    return r
+  },
+  nota_decode_nested: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = nota.decode(nota_enc_nested) }
+    return r
+  },
+  nota_decode_flat: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = nota.decode(nota_enc_flat) }
+    return r
+  },
+  nota_decode_mixed: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = nota.decode(nota_enc_mixed) }
+    return r
+  },
+
+  // WOTA encode
+  wota_encode_integers: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = wota.encode(integers_small) }
+    return r
+  },
+  wota_encode_floats: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = wota.encode(floats_array) }
+    return r
+  },
+  wota_encode_strings_short: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = wota.encode(strings_short) }
+    return r
+  },
+  wota_encode_strings_long: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = wota.encode(strings_long) }
+    return r
+  },
+  wota_encode_strings_unicode: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = wota.encode(strings_unicode) }
+    return r
+  },
+  wota_encode_nested: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = wota.encode(nested_records) }
+    return r
+  },
+  wota_encode_flat: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = wota.encode(flat_record) }
+    return r
+  },
+  wota_encode_mixed: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = wota.encode(mixed_payload) }
+    return r
+  },
+
+  // WOTA decode
+  wota_decode_integers: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = wota.decode(wota_enc_integers) }
+    return r
+  },
+  wota_decode_floats: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = wota.decode(wota_enc_floats) }
+    return r
+  },
+  wota_decode_strings_short: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = wota.decode(wota_enc_strings_short) }
+    return r
+  },
+  wota_decode_strings_long: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = wota.decode(wota_enc_strings_long) }
+    return r
+  },
+  wota_decode_strings_unicode: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = wota.decode(wota_enc_strings_unicode) }
+    return r
+  },
+  wota_decode_nested: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = wota.decode(wota_enc_nested) }
+    return r
+  },
+  wota_decode_flat: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = wota.decode(wota_enc_flat) }
+    return r
+  },
+  wota_decode_mixed: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = wota.decode(wota_enc_mixed) }
+    return r
+  },
+
+  // JSON encode
+  json_encode_integers: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = json.encode(integers_small) }
+    return r
+  },
+  json_encode_floats: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = json.encode(floats_array) }
+    return r
+  },
+  json_encode_strings_short: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = json.encode(strings_short) }
+    return r
+  },
+  json_encode_strings_long: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = json.encode(strings_long) }
+    return r
+  },
+  json_encode_strings_unicode: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = json.encode(strings_unicode) }
+    return r
+  },
+  json_encode_nested: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = json.encode(nested_records) }
+    return r
+  },
+  json_encode_flat: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = json.encode(flat_record) }
+    return r
+  },
+  json_encode_mixed: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = json.encode(mixed_payload) }
+    return r
+  },
+
+  // JSON decode
+  json_decode_integers: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = json.decode(json_enc_integers) }
+    return r
+  },
+  json_decode_floats: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = json.decode(json_enc_floats) }
+    return r
+  },
+  json_decode_strings_short: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = json.decode(json_enc_strings_short) }
+    return r
+  },
+  json_decode_strings_long: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = json.decode(json_enc_strings_long) }
+    return r
+  },
+  json_decode_strings_unicode: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = json.decode(json_enc_strings_unicode) }
+    return r
+  },
+  json_decode_nested: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = json.decode(json_enc_nested) }
+    return r
+  },
+  json_decode_flat: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = json.decode(json_enc_flat) }
+    return r
+  },
+  json_decode_mixed: function(n) {
+    var i = 0
+    var r = null
+    for (i = 0; i < n; i++) { r = json.decode(json_enc_mixed) }
+    return r
+  }
+}

source/nota.h

@@ -308,6 +308,30 @@ void nota_write_blob(NotaBuffer *nb, unsigned long long nbits, const char *data)
 
 void nota_write_text(NotaBuffer *nb, const char *s)
 {
+    /* ASCII fast path: if all bytes < 0x80, KIM == UTF-8 and rune count == byte count */
+    size_t slen = strlen(s);
+    const unsigned char *scan = (const unsigned char *)s;
+    int is_ascii = 1;
+    for (size_t k = 0; k < slen; k++) {
+        if (scan[k] >= 0x80) {
+            is_ascii = 0;
+            break;
+        }
+    }
+
+    if (is_ascii) {
+        long long runes = (long long)slen;
+        char *p = nota_buffer_alloc(nb, 1 + 10 + slen);
+        p[0] = NOTA_TEXT;
+        char *end = nota_continue_num(runes, p, 4);
+        memcpy(end, s, slen);
+        size_t used = (size_t)(end - p) + slen;
+        size_t allocated = 1 + 10 + slen;
+        nb->size -= (allocated - used);
+        return;
+    }
+
+    /* Non-ASCII path: full UTF-8 decode + KIM encode */
     long long runes = utf8_count(s);
 
     size_t max_kim = (size_t)(runes * 5);
@@ -367,6 +391,13 @@ static void nota_write_int_buf(NotaBuffer *nb, long long n)
     nb->size -= (10 - used);
 }
 
+static const double nota_pow10_table[29] = {
+    1e0,  1e1,  1e2,  1e3,  1e4,  1e5,  1e6,  1e7,
+    1e8,  1e9,  1e10, 1e11, 1e12, 1e13, 1e14, 1e15,
+    1e16, 1e17, 1e18, 1e19, 1e20, 1e21, 1e22, 1e23,
+    1e24, 1e25, 1e26, 1e27, 1e28
+};
+
 static void extract_mantissa_coefficient(double num, long *coefficient, long *exponent)
 {
     if (num == 0.0) {
@@ -375,32 +406,46 @@ static void extract_mantissa_coefficient(double num, long *coefficient, long *ex
         return;
     }
 
-    /* Round to 12 decimal places to avoid floating artifacts. */
+    double absval = fabs(num);
-    double rounded = floor(fabs(num) * 1e12 + 0.5) / 1e12;
+    int sign = (num < 0) ? -1 : 1;
-    if (num < 0) {
+
-        rounded = -rounded;
+    /* Get decimal exponent via log10 */
+    int dec_exp = (int)floor(log10(absval));
+
+    /* Scale to extract 14-digit coefficient.
+       We want coeff * 10^exp = absval, with coeff having up to 14 digits.
+       So coeff = absval * 10^(13 - dec_exp), exp = dec_exp - 13 */
+    int shift = 13 - dec_exp;
+    double scaled;
+    if (shift >= 0 && shift <= 28) {
+        scaled = absval * nota_pow10_table[shift];
+    } else if (shift < 0 && -shift <= 28) {
+        scaled = absval / nota_pow10_table[-shift];
+    } else {
+        scaled = absval * pow(10.0, (double)shift);
     }
 
-    char buf[64];
+    long long coeff = (long long)(scaled + 0.5);
-    snprintf(buf, sizeof(buf), "%.14g", rounded);
 
-    char *exp_pos = strpbrk(buf, "eE");
+    /* Correct off-by-one from log10 rounding */
-    long exp_from_sci = 0;
+    if (coeff >= 100000000000000LL) {
-    if (exp_pos) {
+        coeff = (coeff + 5) / 10;
-        exp_from_sci = atol(exp_pos + 1);
+        shift--;
-        *exp_pos = '\0';
+    } else if (coeff < 10000000000000LL && coeff > 0) {
+        coeff = (long long)(absval * pow(10.0, (double)(shift + 1)) + 0.5);
+        shift++;
     }
 
-    char *dec_point = strchr(buf, '.');
+    int exp_out = -shift;
-    int digits_after_decimal = 0;
+
-    if (dec_point) {
+    /* Strip trailing zeros */
-        digits_after_decimal = (int)strlen(dec_point + 1);
+    while (coeff != 0 && coeff % 10 == 0) {
-        memmove(dec_point, dec_point + 1, strlen(dec_point));
+        coeff /= 10;
+        exp_out++;
     }
 
-    long long coeff_ll = atoll(buf);
+    *coefficient = (long)(coeff * sign);
-    *coefficient = (long)coeff_ll;
+    *exponent = (long)exp_out;
-    *exponent = exp_from_sci - digits_after_decimal;
 }
 
 static void nota_write_float_buf(NotaBuffer *nb, double d)

source/runtime.c

@@ -11308,38 +11308,41 @@ static int nota_get_arr_len (JSContext *ctx, JSValue arr) {
   return (int)len;
 }
 
+typedef struct NotaVisitedNode {
+  JSGCRef ref;
+  struct NotaVisitedNode *next;
+} NotaVisitedNode;
+
 typedef struct NotaEncodeContext {
   JSContext *ctx;
-  JSGCRef *visitedStack_ref; /* pointer to GC-rooted ref */
+  NotaVisitedNode *visited_list;
   NotaBuffer nb;
   int cycle;
   JSGCRef *replacer_ref; /* pointer to GC-rooted ref */
 } NotaEncodeContext;
 
 static void nota_stack_push (NotaEncodeContext *enc, JSValueConst val) {
-  JSContext *ctx = enc->ctx;
+  NotaVisitedNode *node = (NotaVisitedNode *)sys_malloc (sizeof (NotaVisitedNode));
-  int len = nota_get_arr_len (ctx, enc->visitedStack_ref->val);
+  JS_PushGCRef (enc->ctx, &node->ref);
-  JS_SetPropertyNumber (ctx, enc->visitedStack_ref->val, len, JS_DupValue (ctx, val));
+  node->ref.val = JS_DupValue (enc->ctx, val);
+  node->next = enc->visited_list;
+  enc->visited_list = node;
 }
 
 static void nota_stack_pop (NotaEncodeContext *enc) {
-  JSContext *ctx = enc->ctx;
+  NotaVisitedNode *node = enc->visited_list;
-  int len = nota_get_arr_len (ctx, enc->visitedStack_ref->val);
+  enc->visited_list = node->next;
-  JS_SetPropertyStr (ctx, enc->visitedStack_ref->val, "length", JS_NewUint32 (ctx, len - 1));
+  JS_FreeValue (enc->ctx, node->ref.val);
+  JS_PopGCRef (enc->ctx, &node->ref);
+  sys_free (node);
 }
 
 static int nota_stack_has (NotaEncodeContext *enc, JSValueConst val) {
-  JSContext *ctx = enc->ctx;
+  NotaVisitedNode *node = enc->visited_list;
-  int len = nota_get_arr_len (ctx, enc->visitedStack_ref->val);
+  while (node) {
-  for (int i = 0; i < len; i++) {
+    if (JS_StrictEq (enc->ctx, node->ref.val, val))
-    JSValue elem = JS_GetPropertyNumber (ctx, enc->visitedStack_ref->val, i);
+      return 1;
-    if (mist_is_gc_object (elem) && mist_is_gc_object (val)) {
+    node = node->next;
-      if (JS_StrictEq (ctx, elem, val)) {
-        JS_FreeValue (ctx, elem);
-        return 1;
-      }
-    }
-    JS_FreeValue (ctx, elem);
   }
   return 0;
 }
@@ -11481,6 +11484,13 @@ static void nota_encode_value (NotaEncodeContext *enc, JSValueConst val, JSValue
       nota_write_sym (&enc->nb, NOTA_NULL);
       break;
     case JS_TAG_PTR: {
+      if (JS_IsText (replaced_ref.val)) {
+        const char *str = JS_ToCString (ctx, replaced_ref.val);
+        nota_write_text (&enc->nb, str);
+        JS_FreeCString (ctx, str);
+        break;
+      }
+
       if (js_is_blob (ctx, replaced_ref.val)) {
         size_t buf_len;
         void *buf_data = js_get_blob_data (ctx, &buf_len, replaced_ref.val);
@@ -11592,16 +11602,14 @@ static void nota_encode_value (NotaEncodeContext *enc, JSValueConst val, JSValue
 }
 
 void *value2nota (JSContext *ctx, JSValue v) {
-  JSGCRef val_ref, stack_ref, key_ref;
+  JSGCRef val_ref, key_ref;
   JS_PushGCRef (ctx, &val_ref);
-  JS_PushGCRef (ctx, &stack_ref);
   JS_PushGCRef (ctx, &key_ref);
   val_ref.val = v;
 
   NotaEncodeContext enc_s, *enc = &enc_s;
   enc->ctx = ctx;
-  stack_ref.val = JS_NewArray (ctx);
+  enc->visited_list = NULL;
-  enc->visitedStack_ref = &stack_ref;
   enc->cycle = 0;
   enc->replacer_ref = NULL;
 
@@ -11611,14 +11619,12 @@ void *value2nota (JSContext *ctx, JSValue v) {
 
   if (enc->cycle) {
     JS_PopGCRef (ctx, &key_ref);
-    JS_PopGCRef (ctx, &stack_ref);
     JS_PopGCRef (ctx, &val_ref);
     nota_buffer_free (&enc->nb);
     return NULL;
   }
 
   JS_PopGCRef (ctx, &key_ref);
-  JS_PopGCRef (ctx, &stack_ref);
   JS_PopGCRef (ctx, &val_ref);
   void *data_ptr = enc->nb.data;
   enc->nb.data = NULL;
@@ -11646,18 +11652,16 @@ JSValue nota2value (JSContext *js, void *nota) {
 static JSValue js_nota_encode (JSContext *ctx, JSValueConst this_val, int argc, JSValueConst *argv) {
   if (argc < 1) return JS_ThrowTypeError (ctx, "nota.encode requires at least 1 argument");
 
-  JSGCRef val_ref, stack_ref, replacer_ref, key_ref;
+  JSGCRef val_ref, replacer_ref, key_ref;
   JS_PushGCRef (ctx, &val_ref);
-  JS_PushGCRef (ctx, &stack_ref);
   JS_PushGCRef (ctx, &replacer_ref);
   JS_PushGCRef (ctx, &key_ref);
   val_ref.val = argv[0];
-  stack_ref.val = JS_NewArray (ctx);
   replacer_ref.val = (argc > 1 && JS_IsFunction (argv[1])) ? argv[1] : JS_NULL;
 
   NotaEncodeContext enc_s, *enc = &enc_s;
   enc->ctx = ctx;
-  enc->visitedStack_ref = &stack_ref;
+  enc->visited_list = NULL;
   enc->cycle = 0;
   enc->replacer_ref = &replacer_ref;
 
@@ -11678,7 +11682,6 @@ static JSValue js_nota_encode (JSContext *ctx, JSValueConst this_val, int argc,
 
   JS_PopGCRef (ctx, &key_ref);
   JS_PopGCRef (ctx, &replacer_ref);
-  JS_PopGCRef (ctx, &stack_ref);
   JS_PopGCRef (ctx, &val_ref);
   return ret;
 }

source/wota.h

@@ -429,11 +429,15 @@ char *wota_read_text_len(size_t *byte_len, char **text_utf8, char *wota)
     }
 
     /* Copy bytes from the packed 64-bit words */
-    for (long long i = 0; i < nwords; i++) {
+    size_t full_words = nbytes / 8;
-        uint64_t wval = data_words[i];
+    size_t remainder = nbytes % 8;
-        for (int j = 0; j < 8 && (i * 8 + j) < (long long)nbytes; j++) {
+    for (size_t i = 0; i < full_words; i++) {
-            out[i * 8 + j] = (char)((wval >> (56 - j * 8)) & 0xff);
+        uint64_t wval = wota_bswap64(data_words[i]);
-        }
+        memcpy(out + i * 8, &wval, 8);
+    }
+    if (remainder > 0) {
+        uint64_t wval = wota_bswap64(data_words[full_words]);
+        memcpy(out + full_words * 8, &wval, remainder);
     }
 
     out[nbytes] = '\0';
@@ -609,14 +613,18 @@ void wota_write_text_len(WotaBuffer *wb, const char *utf8, size_t nbytes)
     }
 
     uint64_t *blocks = wota_buffer_alloc(wb, nwords);
-    memset(blocks, 0, nwords * sizeof(uint64_t));
 
-    for (size_t i = 0; i < nwords; i++) {
+    size_t full_words = nbytes / 8;
+    size_t remainder = nbytes % 8;
+    for (size_t i = 0; i < full_words; i++) {
+        uint64_t wval;
+        memcpy(&wval, utf8 + i * 8, 8);
+        blocks[i] = wota_bswap64(wval);
+    }
+    if (remainder > 0) {
         uint64_t wval = 0;
-        for (int j = 0; j < 8 && (i * 8 + j) < nbytes; j++) {
+        memcpy(&wval, utf8 + full_words * 8, remainder);
-            wval |= ((uint64_t)(unsigned char)utf8[i * 8 + j]) << (56 - j * 8);
+        blocks[full_words] = wota_bswap64(wval);
-        }
-        blocks[i] = wval;
     }
 }