oj 3.16.9 → 3.16.11

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: d3e57c02a1fe6782596953f34b8e2a2b729a09b5d8e7128dd4633d430ca7aa0c
-  data.tar.gz: 7698f8c0203459d62f4421f11a9c3637b04b5b808ecf87ce4ca057e2fd073762
+  metadata.gz: 30aea721380a4e3edc306dd19906d8777f230a639ba3427e9394dd543a3a7e3b
+  data.tar.gz: b024a9d4513c16c1bfe4fc3c4adeacb1afd9a0e670987476d32cfa2fa74e9b1e
 SHA512:
-  metadata.gz: 417ada5b645a6ba48e81b52bb72cec97bb4a64595a61252989346a975c1026b3cbc039cbf7cef166b5dbfbdd79554d5c9e5786da3299ce1fd3f2ec70d3ef479f
-  data.tar.gz: ceffb29c6732b107d42091bc8754e8b4174e26e6d8eb9c4162ed8a12a65d37aa2450f9f5398d39242d92fabd46d63be5fc14e0dc003774378e4e6211b02e3f0d
+  metadata.gz: 527ea1162cb135bbe16eefc10a7cb05444182767aca6fa0b6986622e52d7082bcec020c43e663251406c81602018f7d0842c2c5cee37aeca0269560e502d99dd
+  data.tar.gz: e49e9f63e373cb0ec21f604f97899f87815b86ef5a5eafad30e7bddbd11e71156f92beaa1259c83609c2d45d2a8aac87c8b27e3e266fcc2bd99a1908327c796d

data/CHANGELOG.md

@@ -1,5 +1,15 @@
 # CHANGELOG
 
+## 3.16.11 - 2025-05-29
+
+- Fixed range encoding with the :circular option
+
+## 3.16.10 - 2025-02-24
+
+- Changed oj_parser_type to be non-static.
+
+- Changed ARM versions to used Neon instructions thanks to @samyron.
+
 ## 3.16.9 - 2024-12-28
 
 - Fixed `Oj::Parser` create_id size issue #931.
@@ -64,6 +74,7 @@
 ## 3.14.3 - 2023-04-07
 
 - Fixed compat parse with optimized Hash when parsing a JSON::GenericObject.
+- Deprecated Ruby <= 2.6.10
 
 ## 3.14.2 - 2023-02-10
 

data/ext/oj/custom.c

@@ -40,7 +40,7 @@ static void dump_obj_as_str(VALUE obj, int depth, Out out) {
 static void bigdecimal_dump(VALUE obj, int depth, Out out) {
     volatile VALUE rstr = oj_safe_string_convert(obj);
     const char    *str  = RSTRING_PTR(rstr);
-    int            len  = (int)RSTRING_LEN(rstr);
+    size_t         len  = RSTRING_LEN(rstr);
 
     if (0 == strcasecmp("Infinity", str)) {
         str = oj_nan_str(obj, out->opts->dump_opts.nan_dump, out->opts->mode, true, &len);
@@ -123,7 +123,7 @@ static void date_dump(VALUE obj, int depth, Out out) {
         case RubyTime:
         case XmlTime:
             v = rb_funcall(obj, rb_intern("iso8601"), 0);
-            oj_dump_cstr(RSTRING_PTR(v), (int)RSTRING_LEN(v), 0, 0, out);
+            oj_dump_cstr(RSTRING_PTR(v), RSTRING_LEN(v), 0, 0, out);
             break;
         case UnixZTime:
             v = rb_funcall(obj, rb_intern("to_time"), 0);
@@ -405,7 +405,7 @@ static void dump_odd(VALUE obj, Odd odd, VALUE clas, int depth, Out out) {
             rb_raise(rb_eEncodingError, "Invalid type for raw JSON.\n");
         } else {
             const char *s    = RSTRING_PTR(v);
-            int         len  = (int)RSTRING_LEN(v);
+            size_t      len  = RSTRING_LEN(v);
             const char *name = rb_id2name(*odd->attrs);
             size_t      nlen = strlen(name);
 
@@ -478,7 +478,7 @@ static VALUE dump_common(VALUE obj, int depth, Out out) {
     } else if (Yes == out->opts->to_json && rb_respond_to(obj, oj_to_json_id)) {
         volatile VALUE rs;
         const char    *s;
-        int            len;
+        size_t         len;
 
         TRACE(out->opts->trace, "to_json", obj, depth + 1, TraceRubyIn);
         if (0 == rb_obj_method_arity(obj, oj_to_json_id)) {
@@ -488,7 +488,7 @@ static VALUE dump_common(VALUE obj, int depth, Out out) {
         }
         TRACE(out->opts->trace, "to_json", obj, depth + 1, TraceRubyOut);
         s   = RSTRING_PTR(rs);
-        len = (int)RSTRING_LEN(rs);
+        len = RSTRING_LEN(rs);
 
         assure_size(out, len + 1);
         APPEND_CHARS(out->cur, s, len);
@@ -509,7 +509,7 @@ static VALUE dump_common(VALUE obj, int depth, Out out) {
         if (aj == obj) {
             volatile VALUE rstr = oj_safe_string_convert(obj);
 
-            oj_dump_cstr(RSTRING_PTR(rstr), (int)RSTRING_LEN(rstr), false, false, out);
+            oj_dump_cstr(RSTRING_PTR(rstr), RSTRING_LEN(rstr), false, false, out);
         } else {
             oj_dump_custom_val(aj, depth, out, true);
         }
@@ -676,7 +676,8 @@ static void dump_obj(VALUE obj, int depth, Out out, bool as_ok) {
 
 static void dump_array(VALUE a, int depth, Out out, bool as_ok) {
     size_t size;
-    int    i, cnt;
+    size_t i;
+    size_t cnt;
     int    d2 = depth + 1;
     long   id = oj_check_circular(a, out);
 
@@ -684,7 +685,7 @@ static void dump_array(VALUE a, int depth, Out out, bool as_ok) {
         oj_dump_nil(Qnil, depth, out, false);
         return;
     }
-    cnt         = (int)RARRAY_LEN(a);
+    cnt         = RARRAY_LEN(a);
     *out->cur++ = '[';
     assure_size(out, 2);
     if (0 == cnt) {
@@ -795,7 +796,7 @@ static void dump_struct(VALUE obj, int depth, Out out, bool as_ok) {
                 volatile VALUE s = rb_sym2str(RARRAY_AREF(ma, i));
 
                 name = RSTRING_PTR(s);
-                len  = (int)RSTRING_LEN(s);
+                len  = RSTRING_LEN(s);
             } else {
                 len  = snprintf(num_id, sizeof(num_id), "%d", i);
                 name = num_id;

data/ext/oj/dump.c

@@ -152,8 +152,77 @@ inline static size_t newline_friendly_size(const uint8_t *str, size_t len) {
     return calculate_string_size(str, len, newline_friendly_chars);
 }
 
+#ifdef HAVE_SIMD_NEON
+inline static uint8x16x4_t load_uint8x16_4(const unsigned char *table) {
+    uint8x16x4_t tab;
+    tab.val[0] = vld1q_u8(table);
+    tab.val[1] = vld1q_u8(table + 16);
+    tab.val[2] = vld1q_u8(table + 32);
+    tab.val[3] = vld1q_u8(table + 48);
+    return tab;
+}
+
+static uint8x16x4_t hibit_friendly_chars_neon[2];
+static uint8x16x4_t rails_friendly_chars_neon[2];
+static uint8x16x4_t rails_xss_friendly_chars_neon[4];
+
+void initialize_neon(void) {
+    // We only need the first 128 bytes of the hibit friendly chars table. Everything above 127 is
+    // set to 1. If that ever changes, the code will need to be updated.
+    hibit_friendly_chars_neon[0] = load_uint8x16_4((const unsigned char *)hibit_friendly_chars);
+    hibit_friendly_chars_neon[1] = load_uint8x16_4((const unsigned char *)hibit_friendly_chars + 64);
+
+    // rails_friendly_chars is the same as hibit_friendly_chars. Only the first 128 bytes have values
+    // that are not '1'. If that ever changes, the code will need to be updated.
+    rails_friendly_chars_neon[0] = load_uint8x16_4((const unsigned char *)rails_friendly_chars);
+    rails_friendly_chars_neon[1] = load_uint8x16_4((const unsigned char *)rails_friendly_chars + 64);
+
+    rails_xss_friendly_chars_neon[0] = load_uint8x16_4((const unsigned char *)rails_xss_friendly_chars);
+    rails_xss_friendly_chars_neon[1] = load_uint8x16_4((const unsigned char *)rails_xss_friendly_chars + 64);
+    rails_xss_friendly_chars_neon[2] = load_uint8x16_4((const unsigned char *)rails_xss_friendly_chars + 128);
+    rails_xss_friendly_chars_neon[3] = load_uint8x16_4((const unsigned char *)rails_xss_friendly_chars + 192);
+
+    // All bytes should be 0 except for those that need more than 1 byte of output. This will allow the
+    // code to limit the lookups to the first 128 bytes (values 0 - 127). Bytes above 127 will result
+    // in 0 with the vqtbl4q_u8 instruction.
+    uint8x16_t one = vdupq_n_u8('1');
+    for (int i = 0; i < 2; i++) {
+        for (int j = 0; j < 4; j++) {
+            hibit_friendly_chars_neon[i].val[j] = vsubq_u8(hibit_friendly_chars_neon[i].val[j], one);
+            rails_friendly_chars_neon[i].val[j] = vsubq_u8(rails_friendly_chars_neon[i].val[j], one);
+        }
+    }
+
+    for (int i = 0; i < 4; i++) {
+        for (int j = 0; j < 4; j++) {
+            rails_xss_friendly_chars_neon[i].val[j] = vsubq_u8(rails_xss_friendly_chars_neon[i].val[j], one);
+        }
+    }
+}
+#endif
+
 inline static size_t hibit_friendly_size(const uint8_t *str, size_t len) {
+#ifdef HAVE_SIMD_NEON
+    size_t size = 0;
+    size_t i    = 0;
+
+    for (; i + sizeof(uint8x16_t) <= len; i += sizeof(uint8x16_t), str += sizeof(uint8x16_t)) {
+        size += sizeof(uint8x16_t);
+
+        // See https://lemire.me/blog/2019/07/23/arbitrary-byte-to-byte-maps-using-arm-neon/
+        uint8x16_t chunk  = vld1q_u8(str);
+        uint8x16_t tmp1   = vqtbl4q_u8(hibit_friendly_chars_neon[0], chunk);
+        uint8x16_t tmp2   = vqtbl4q_u8(hibit_friendly_chars_neon[1], veorq_u8(chunk, vdupq_n_u8(0x40)));
+        uint8x16_t result = vorrq_u8(tmp1, tmp2);
+        uint8_t    tmp    = vaddvq_u8(result);
+        size += tmp;
+    }
+
+    size_t total = size + calculate_string_size(str, len - i, hibit_friendly_chars);
+    return total;
+#else
     return calculate_string_size(str, len, hibit_friendly_chars);
+#endif
 }
 
 inline static size_t slash_friendly_size(const uint8_t *str, size_t len) {
@@ -184,9 +253,43 @@ inline static size_t hixss_friendly_size(const uint8_t *str, size_t len) {
 
 inline static long rails_xss_friendly_size(const uint8_t *str, size_t len) {
     long    size = 0;
-    size_t  i    = len;
     uint8_t hi   = 0;
 
+#ifdef HAVE_SIMD_NEON
+    size_t i = 0;
+
+    uint8x16_t has_some_hibit = vdupq_n_u8(0);
+    uint8x16_t hibit          = vdupq_n_u8(0x80);
+    for (; i + sizeof(uint8x16_t) <= len; i += sizeof(uint8x16_t), str += sizeof(uint8x16_t)) {
+        size += sizeof(uint8x16_t);
+
+        uint8x16_t chunk = vld1q_u8(str);
+
+        // Check to see if any of these bytes have the high bit set.
+        has_some_hibit = vorrq_u8(has_some_hibit, vandq_u8(chunk, hibit));
+
+        uint8x16_t tmp1   = vqtbl4q_u8(rails_xss_friendly_chars_neon[0], chunk);
+        uint8x16_t tmp2   = vqtbl4q_u8(rails_xss_friendly_chars_neon[1], veorq_u8(chunk, vdupq_n_u8(0x40)));
+        uint8x16_t tmp3   = vqtbl4q_u8(rails_xss_friendly_chars_neon[2], veorq_u8(chunk, vdupq_n_u8(0x80)));
+        uint8x16_t tmp4   = vqtbl4q_u8(rails_xss_friendly_chars_neon[3], veorq_u8(chunk, vdupq_n_u8(0xc0)));
+        uint8x16_t result = vorrq_u8(tmp4, vorrq_u8(tmp3, vorrq_u8(tmp1, tmp2)));
+        uint8_t    tmp    = vaddvq_u8(result);
+        size += tmp;
+    }
+
+    // 'hi' should be set if any of the bytes we processed have the high bit set. It doesn't matter which ones.
+    hi = vmaxvq_u8(has_some_hibit) != 0;
+
+    for (; i < len; str++, i++) {
+        size += rails_xss_friendly_chars[*str] - '0';
+        hi |= *str & 0x80;
+    }
+    if (0 == hi) {
+        return size;
+    }
+    return -(size);
+#else
+    size_t i = len;
     for (; 0 < i; str++, i--) {
         size += rails_xss_friendly_chars[*str];
         hi |= *str & 0x80;
@@ -195,13 +298,47 @@ inline static long rails_xss_friendly_size(const uint8_t *str, size_t len) {
         return size - len * (size_t)'0';
     }
     return -(size - len * (size_t)'0');
+#endif /* HAVE_SIMD_NEON */
 }
 
 inline static size_t rails_friendly_size(const uint8_t *str, size_t len) {
     long    size = 0;
-    size_t  i    = len;
     uint8_t hi   = 0;
+#ifdef HAVE_SIMD_NEON
+    size_t i = 0;
+
+    uint8x16_t has_some_hibit = vdupq_n_u8(0);
+    uint8x16_t hibit          = vdupq_n_u8(0x80);
+
+    for (; i + sizeof(uint8x16_t) <= len; i += sizeof(uint8x16_t), str += sizeof(uint8x16_t)) {
+        size += sizeof(uint8x16_t);
+
+        // See https://lemire.me/blog/2019/07/23/arbitrary-byte-to-byte-maps-using-arm-neon/
+        uint8x16_t chunk = vld1q_u8(str);
+
+        // Check to see if any of these bytes have the high bit set.
+        has_some_hibit = vorrq_u8(has_some_hibit, vandq_u8(chunk, hibit));
 
+        uint8x16_t tmp1   = vqtbl4q_u8(rails_friendly_chars_neon[0], chunk);
+        uint8x16_t tmp2   = vqtbl4q_u8(rails_friendly_chars_neon[1], veorq_u8(chunk, vdupq_n_u8(0x40)));
+        uint8x16_t result = vorrq_u8(tmp1, tmp2);
+        uint8_t    tmp    = vaddvq_u8(result);
+        size += tmp;
+    }
+
+    // 'hi' should be set if any of the bytes we processed have the high bit set. It doesn't matter which ones.
+    hi = vmaxvq_u8(has_some_hibit) != 0;
+
+    for (; i < len; str++, i++) {
+        size += rails_friendly_chars[*str] - '0';
+        hi |= *str & 0x80;
+    }
+    if (0 == hi) {
+        return size;
+    }
+    return -(size);
+#else
+    size_t i = len;
     for (; 0 < i; str++, i--) {
         size += rails_friendly_chars[*str];
         hi |= *str & 0x80;
@@ -210,9 +347,10 @@ inline static size_t rails_friendly_size(const uint8_t *str, size_t len) {
         return size - len * (size_t)'0';
     }
     return -(size - len * (size_t)'0');
+#endif /* HAVE_SIMD_NEON */
 }
 
-const char *oj_nan_str(VALUE obj, int opt, int mode, bool plus, int *lenp) {
+const char *oj_nan_str(VALUE obj, int opt, int mode, bool plus, size_t *lenp) {
     const char *str = NULL;
 
     if (AutoNan == opt) {
@@ -477,7 +615,7 @@ void oj_dump_time(VALUE obj, Out out, int withZone) {
 void oj_dump_ruby_time(VALUE obj, Out out) {
     volatile VALUE rstr = oj_safe_string_convert(obj);
 
-    oj_dump_cstr(RSTRING_PTR(rstr), (int)RSTRING_LEN(rstr), 0, 0, out);
+    oj_dump_cstr(RSTRING_PTR(rstr), RSTRING_LEN(rstr), 0, 0, out);
 }
 
 void oj_dump_xml_time(VALUE obj, Out out) {
@@ -711,13 +849,13 @@ void oj_dump_str(VALUE obj, int depth, Out out, bool as_ok) {
         rb_encoding *enc = rb_enc_from_index(idx);
         obj              = rb_str_conv_enc(obj, enc, oj_utf8_encoding);
     }
-    oj_dump_cstr(RSTRING_PTR(obj), (int)RSTRING_LEN(obj), 0, 0, out);
+    oj_dump_cstr(RSTRING_PTR(obj), RSTRING_LEN(obj), 0, 0, out);
 }
 
 void oj_dump_sym(VALUE obj, int depth, Out out, bool as_ok) {
     volatile VALUE s = rb_sym2str(obj);
 
-    oj_dump_cstr(RSTRING_PTR(s), (int)RSTRING_LEN(s), 0, 0, out);
+    oj_dump_cstr(RSTRING_PTR(s), RSTRING_LEN(s), 0, 0, out);
 }
 
 static void debug_raise(const char *orig, size_t cnt, int line) {
@@ -758,9 +896,49 @@ void oj_dump_raw_json(VALUE obj, int depth, Out out) {
     }
 }
 
+#ifdef HAVE_SIMD_NEON
+typedef struct _neon_match_result {
+    uint8x16_t needs_escape;
+    bool       has_some_hibit;
+    bool       do_unicode_validation;
+} neon_match_result;
+
+#if defined(__clang__) || defined(__GNUC__)
+#define FORCE_INLINE __attribute__((always_inline))
+#else
+#define FORCE_INLINE
+#endif
+
+static inline FORCE_INLINE neon_match_result
+neon_update(const char *str, uint8x16x4_t *cmap_neon, int neon_table_size, bool do_unicode_validation, bool has_hi) {
+    neon_match_result result = {.has_some_hibit = false, .do_unicode_validation = false};
+
+    uint8x16_t chunk    = vld1q_u8((const unsigned char *)str);
+    uint8x16_t tmp1     = vqtbl4q_u8(cmap_neon[0], chunk);
+    uint8x16_t tmp2     = vqtbl4q_u8(cmap_neon[1], veorq_u8(chunk, vdupq_n_u8(0x40)));
+    result.needs_escape = vorrq_u8(tmp1, tmp2);
+    if (neon_table_size > 2) {
+        uint8x16_t tmp3     = vqtbl4q_u8(cmap_neon[2], veorq_u8(chunk, vdupq_n_u8(0x80)));
+        uint8x16_t tmp4     = vqtbl4q_u8(cmap_neon[3], veorq_u8(chunk, vdupq_n_u8(0xc0)));
+        result.needs_escape = vorrq_u8(result.needs_escape, vorrq_u8(tmp4, tmp3));
+    }
+    if (has_hi && do_unicode_validation) {
+        uint8x16_t has_some_hibit    = vandq_u8(chunk, vdupq_n_u8(0x80));
+        result.has_some_hibit        = vmaxvq_u8(has_some_hibit) != 0;
+        result.do_unicode_validation = has_hi && do_unicode_validation && result.has_some_hibit;
+    }
+    return result;
+}
+
+#endif /* HAVE_SIMD_NEON */
+
 void oj_dump_cstr(const char *str, size_t cnt, bool is_sym, bool escape1, Out out) {
-    size_t      size;
-    char       *cmap;
+    size_t size;
+    char  *cmap;
+#ifdef HAVE_SIMD_NEON
+    uint8x16x4_t *cmap_neon = NULL;
+    int           neon_table_size;
+#endif /* HAVE_SIMD_NEON */
     const char *orig                  = str;
     bool        has_hi                = false;
     bool        do_unicode_validation = false;
@@ -792,7 +970,11 @@ void oj_dump_cstr(const char *str, size_t cnt, bool is_sym, bool escape1, Out ou
         long sz;
 
         cmap = rails_xss_friendly_chars;
-        sz   = rails_xss_friendly_size((uint8_t *)str, cnt);
+#ifdef HAVE_SIMD_NEON
+        cmap_neon       = rails_xss_friendly_chars_neon;
+        neon_table_size = 4;
+#endif /* HAVE_NEON_SIMD */
+        sz = rails_xss_friendly_size((uint8_t *)str, cnt);
         if (sz < 0) {
             has_hi = true;
             size   = (size_t)-sz;
@@ -805,7 +987,11 @@ void oj_dump_cstr(const char *str, size_t cnt, bool is_sym, bool escape1, Out ou
     case RailsEsc: {
         long sz;
         cmap = rails_friendly_chars;
-        sz   = rails_friendly_size((uint8_t *)str, cnt);
+#ifdef HAVE_SIMD_NEON
+        cmap_neon       = rails_friendly_chars_neon;
+        neon_table_size = 2;
+#endif /* HAVE_NEON_SIMD */
+        sz = rails_friendly_size((uint8_t *)str, cnt);
         if (sz < 0) {
             has_hi = true;
             size   = (size_t)-sz;
@@ -816,7 +1002,12 @@ void oj_dump_cstr(const char *str, size_t cnt, bool is_sym, bool escape1, Out ou
         break;
     }
     case JSONEsc:
-    default: cmap = hibit_friendly_chars; size = hibit_friendly_size((uint8_t *)str, cnt);
+    default: cmap = hibit_friendly_chars;
+#ifdef HAVE_SIMD_NEON
+        cmap_neon       = hibit_friendly_chars_neon;
+        neon_table_size = 2;
+#endif /* HAVE_NEON_SIMD */
+        size = hibit_friendly_size((uint8_t *)str, cnt);
     }
     assure_size(out, size + BUFFER_EXTRA);
     *out->cur++ = '"';
@@ -842,8 +1033,116 @@ void oj_dump_cstr(const char *str, size_t cnt, bool is_sym, bool escape1, Out ou
         if (is_sym) {
             *out->cur++ = ':';
         }
+#ifdef HAVE_SIMD_NEON
+        const char *chunk_start;
+        const char *chunk_end;
+        const char *cursor     = str;
+        int         neon_state = (cmap_neon != NULL) ? 1 : 4;
+        char        matches[16];
+        bool        do_hi_validation = false;
+        // uint64_t neon_match_mask = 0;
+#define SEARCH_FLUSH                                  \
+    if (str > cursor) {                               \
+        APPEND_CHARS(out->cur, cursor, str - cursor); \
+        cursor = str;                                 \
+    }
+
+    loop:
+#endif /* HAVE_SIMD_NEON */
         for (; str < end; str++) {
-            switch (cmap[(uint8_t)*str]) {
+            char action = 0;
+#ifdef HAVE_SIMD_NEON
+            /* neon_state:
+             * 1: Scanning for matches. There must be at least
+                  sizeof(uint8x16_t) bytes of input data to use SIMD and
+                  cmap_neon must be non-null.
+             * 2: Matches have been found. Will set str to the position of the
+             *    next match and set the state to 3.
+             *    If there are no more matches it will transition to state 1.
+             * 4: Fallback to the scalar algorithm. Not enough data to use
+             *    SIMD.
+             */
+#define NEON_SET_STATE(state) \
+    neon_state = state;       \
+    goto loop;
+#define NEON_RETURN_TO_STATE(state) neon_state = state;
+            switch (neon_state) {
+            case 1: {
+                while (true) {
+                    const char *chunk_ptr = NULL;
+                    if (str + sizeof(uint8x16_t) <= end) {
+                        chunk_ptr   = str;
+                        chunk_start = str;
+                        chunk_end   = str + sizeof(uint8x16_t);
+                    } else if ((end - str) >= SIMD_MINIMUM_THRESHOLD) {
+                        memset(out->cur, 'A', sizeof(uint8x16_t));
+                        memcpy(out->cur, str, (end - str));
+                        chunk_ptr   = out->cur;
+                        chunk_start = str;
+                        chunk_end   = end;
+                    } else {
+                        SEARCH_FLUSH;
+                        NEON_SET_STATE(4);
+                        break; /* Unreachable */
+                    }
+                    neon_match_result result = neon_update(chunk_ptr,
+                                                           cmap_neon,
+                                                           neon_table_size,
+                                                           do_unicode_validation,
+                                                           has_hi);
+                    if ((result.do_unicode_validation) || vmaxvq_u8(result.needs_escape) != 0) {
+                        SEARCH_FLUSH;
+                        uint8x16_t actions = vaddq_u8(result.needs_escape, vdupq_n_u8('1'));
+                        do_hi_validation   = result.do_unicode_validation;
+                        vst1q_u8((unsigned char *)matches, actions);
+                        NEON_SET_STATE(2);
+                        break; /* Unreachable */
+                    }
+                    str = chunk_end;
+                }
+                // We must have run out of data to use SIMD. Go to state 4.
+                SEARCH_FLUSH;
+                NEON_SET_STATE(4);
+            } break;
+            case 3:
+                cursor = str;
+                // This fall through is intentional. We return to state 3 after we process
+                // a byte (or multiple). We return to this state to ensure the cursor is
+                // pointing to the correct location. We then resume looking for matches
+                // within the previously processed chunk.
+            case 2:
+                if (str >= chunk_end) {
+                    NEON_SET_STATE(1);
+                }
+                if (!do_hi_validation) {
+                    long i = str - chunk_start;
+                    for (; str < chunk_end; i++) {
+                        if ((action = matches[i]) != '1') {
+                            break;
+                        }
+                        *out->cur++ = *str++;
+                    }
+                    // The loop above may have advanced str and directly output them to out->cur.
+                    // Ensure cursor is set appropriately.
+                    cursor = str;
+                    if (str >= chunk_end) {
+                        // We must have advanced past the end... we are done.
+                        NEON_SET_STATE(1);
+                    }
+                } else {
+                    long match_index = str - chunk_start;
+                    action           = matches[match_index];
+                }
+                NEON_RETURN_TO_STATE(3);
+                break;
+            case 4: action = cmap[(uint8_t)*str];
+            }
+#undef NEON_SET_STATE
+#undef NEON_RETURN_TO_STATE
+#else
+            action = cmap[(uint8_t)*str];
+#endif /* HAVE_SIMD_NEON */
+            switch (action) {
             case '1':
                 if (do_unicode_validation && check_start <= str) {
                     if (0 != (0x80 & (uint8_t)*str)) {
@@ -906,7 +1205,7 @@ void oj_dump_cstr(const char *str, size_t cnt, bool is_sym, bool escape1, Out ou
         *out->cur++ = '"';
     }
     if (do_unicode_validation && 0 < str - orig && 0 != (0x80 & *(str - 1))) {
-        uint8_t c = (uint8_t) * (str - 1);
+        uint8_t c = (uint8_t)*(str - 1);
         int     i;
         int     scnt = (int)(str - orig);
 
@@ -957,7 +1256,7 @@ void oj_dump_class(VALUE obj, int depth, Out out, bool as_ok) {
 void oj_dump_obj_to_s(VALUE obj, Out out) {
     volatile VALUE rstr = oj_safe_string_convert(obj);
 
-    oj_dump_cstr(RSTRING_PTR(rstr), (int)RSTRING_LEN(rstr), 0, 0, out);
+    oj_dump_cstr(RSTRING_PTR(rstr), RSTRING_LEN(rstr), 0, 0, out);
 }
 
 void oj_dump_raw(const char *str, size_t cnt, Out out) {
@@ -1092,7 +1391,7 @@ void oj_dump_fixnum(VALUE obj, int depth, Out out, bool as_ok) {
 
 void oj_dump_bignum(VALUE obj, int depth, Out out, bool as_ok) {
     volatile VALUE rs             = rb_big2str(obj, 10);
-    int            cnt            = (int)RSTRING_LEN(rs);
+    size_t         cnt            = RSTRING_LEN(rs);
     bool           dump_as_string = false;
 
     if (out->opts->int_range_max != 0 || out->opts->int_range_min != 0) {  // Bignum cannot be inside of Fixnum range
@@ -1114,7 +1413,7 @@ void oj_dump_float(VALUE obj, int depth, Out out, bool as_ok) {
     char   buf[64];
     char  *b;
     double d   = rb_num2dbl(obj);
-    int    cnt = 0;
+    size_t cnt = 0;
 
     if (0.0 == d) {
         b    = buf;
@@ -1225,7 +1524,7 @@ void oj_dump_float(VALUE obj, int depth, Out out, bool as_ok) {
     } else if (0 == out->opts->float_prec) {
         volatile VALUE rstr = oj_safe_string_convert(obj);
 
-        cnt = (int)RSTRING_LEN(rstr);
+        cnt = RSTRING_LEN(rstr);
         if ((int)sizeof(buf) <= cnt) {
             cnt = sizeof(buf) - 1;
         }
@@ -1239,8 +1538,8 @@ void oj_dump_float(VALUE obj, int depth, Out out, bool as_ok) {
     *out->cur = '\0';
 }
 
-int oj_dump_float_printf(char *buf, size_t blen, VALUE obj, double d, const char *format) {
-    int cnt = snprintf(buf, blen, format, d);
+size_t oj_dump_float_printf(char *buf, size_t blen, VALUE obj, double d, const char *format) {
+    size_t cnt = snprintf(buf, blen, format, d);
 
     // Round off issues at 16 significant digits so check for obvious ones of
     // 0001 and 9999.
@@ -1248,7 +1547,7 @@ int oj_dump_float_printf(char *buf, size_t blen, VALUE obj, double d, const char
         volatile VALUE rstr = oj_safe_string_convert(obj);
 
         strcpy(buf, RSTRING_PTR(rstr));
-        cnt = (int)RSTRING_LEN(rstr);
+        cnt = RSTRING_LEN(rstr);
     }
     return cnt;
 }

data/ext/oj/dump.h

@@ -7,12 +7,17 @@
 #include <ruby.h>
 
 #include "oj.h"
+#include "simd.h"
 
 #define MAX_DEPTH 1000
 
 // Extra padding at end of buffer.
 #define BUFFER_EXTRA 64
 
+#ifdef HAVE_SIMD_NEON
+extern void initialize_neon(void);
+#endif /* HAVE_SIMD_NEON */
+
 extern void oj_dump_nil(VALUE obj, int depth, Out out, bool as_ok);
 extern void oj_dump_true(VALUE obj, int depth, Out out, bool as_ok);
 extern void oj_dump_false(VALUE obj, int depth, Out out, bool as_ok);
@@ -30,7 +35,7 @@ extern void oj_dump_xml_time(VALUE obj, Out out);
 extern void oj_dump_time(VALUE obj, Out out, int withZone);
 extern void oj_dump_obj_to_s(VALUE obj, Out out);
 
-extern const char *oj_nan_str(VALUE obj, int opt, int mode, bool plus, int *lenp);
+extern const char *oj_nan_str(VALUE obj, int opt, int mode, bool plus, size_t *lenp);
 
 // initialize an out buffer with the provided stack allocated memory
 extern void oj_out_init(Out out);
@@ -53,7 +58,7 @@ extern void oj_dump_raw_json(VALUE obj, int depth, Out out);
 extern VALUE oj_add_to_json(int argc, VALUE *argv, VALUE self);
 extern VALUE oj_remove_to_json(int argc, VALUE *argv, VALUE self);
 
-extern int oj_dump_float_printf(char *buf, size_t blen, VALUE obj, double d, const char *format);
+extern size_t oj_dump_float_printf(char *buf, size_t blen, VALUE obj, double d, const char *format);
 
 extern time_t oj_sec_from_time_hard_way(VALUE obj);