oj 3.16.6 → 3.17.3

data/ext/oj/dump.c

@@ -40,7 +40,7 @@ static size_t ascii_friendly_size(const uint8_t *str, size_t len);
 static const char hex_chars[17] = "0123456789abcdef";
 
 // JSON standard except newlines are no escaped
-static char newline_friendly_chars[256] = "\
+static char newline_friendly_chars[257] = "\
 66666666221622666666666666666666\
 11211111111111111111111111111111\
 11111111111111111111111111112111\
@@ -51,7 +51,7 @@ static char newline_friendly_chars[256] = "\
 11111111111111111111111111111111";
 
 // JSON standard
-static char hibit_friendly_chars[256] = "\
+static char hibit_friendly_chars[257] = "\
 66666666222622666666666666666666\
 11211111111111111111111111111111\
 11111111111111111111111111112111\
@@ -62,7 +62,7 @@ static char hibit_friendly_chars[256] = "\
 11111111111111111111111111111111";
 
 // JSON standard but escape forward slashes `/`
-static char slash_friendly_chars[256] = "\
+static char slash_friendly_chars[257] = "\
 66666666222622666666666666666666\
 11211111111111121111111111111111\
 11111111111111111111111111112111\
@@ -74,7 +74,7 @@ static char slash_friendly_chars[256] = "\
 
 // High bit set characters are always encoded as unicode. Worse case is 3
 // bytes per character in the output. That makes this conservative.
-static char ascii_friendly_chars[256] = "\
+static char ascii_friendly_chars[257] = "\
 66666666222622666666666666666666\
 11211111111111111111111111111111\
 11111111111111111111111111112111\
@@ -85,7 +85,7 @@ static char ascii_friendly_chars[256] = "\
 33333333333333333333333333333333";
 
 // XSS safe mode
-static char xss_friendly_chars[256] = "\
+static char xss_friendly_chars[257] = "\
 66666666222622666666666666666666\
 11211161111111121111111111116161\
 11111111111111111111111111112111\
@@ -96,7 +96,7 @@ static char xss_friendly_chars[256] = "\
 33333333333333333333333333333333";
 
 // JSON XSS combo
-static char hixss_friendly_chars[256] = "\
+static char hixss_friendly_chars[257] = "\
 66666666222622666666666666666666\
 11211111111111111111111111111111\
 11111111111111111111111111112111\
@@ -107,7 +107,7 @@ static char hixss_friendly_chars[256] = "\
 11611111111111111111111111111111";
 
 // Rails XSS combo
-static char rails_xss_friendly_chars[256] = "\
+static char rails_xss_friendly_chars[257] = "\
 66666666222622666666666666666666\
 11211161111111111111111111116161\
 11111111111111111111111111112111\
@@ -118,7 +118,7 @@ static char rails_xss_friendly_chars[256] = "\
 11611111111111111111111111111111";
 
 // Rails HTML non-escape
-static char rails_friendly_chars[256] = "\
+static char rails_friendly_chars[257] = "\
 66666666222622666666666666666666\
 11211111111111111111111111111111\
 11111111111111111111111111112111\
@@ -152,8 +152,146 @@ 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
+
+#ifdef HAVE_SIMD_SSE4_2
+
+static __m128i hibit_friendly_chars_sse42[8];
+
+// From: https://stackoverflow.com/questions/36998538/fastest-way-to-horizontally-sum-sse-unsigned-byte-vector
+inline static OJ_TARGET_SSE42 uint32_t _mm_sum_epu8(const __m128i v) {
+    __m128i vsum = _mm_sad_epu8(v, _mm_setzero_si128());
+    return _mm_cvtsi128_si32(vsum) + _mm_extract_epi16(vsum, 4);
+}
+
+inline static OJ_TARGET_SSE42 size_t hibit_friendly_size_sse42(const uint8_t *str, size_t len) {
+    size_t size = 0;
+    size_t i    = 0;
+
+    for (; i + sizeof(__m128i) <= len; i += sizeof(__m128i), str += sizeof(__m128i)) {
+        size += sizeof(__m128i);
+
+        __m128i chunk = _mm_loadu_si128((__m128i *)str);
+        __m128i tmp   = vector_lookup_sse42(chunk, hibit_friendly_chars_sse42, 8);
+        size += _mm_sum_epu8(tmp);
+    }
+    size_t total = size + calculate_string_size(str, len - i, hibit_friendly_chars);
+    return total;
+}
+
+void OJ_TARGET_SSE42 initialize_sse42(void) {
+    for (int i = 0; i < 8; i++) {
+        hibit_friendly_chars_sse42[i] = _mm_sub_epi8(
+            _mm_loadu_si128((__m128i *)(hibit_friendly_chars + i * sizeof(__m128i))),
+            _mm_set1_epi8('1'));
+    }
+}
+
+#else
+
+#define SIMD_TARGET
+
+#endif /* HAVE_SIMD_SSE4_2 */
+
 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;
+    }
+
+#ifdef HAVE_FAST_MEMCPY
+    size_t total = 0;
+    if (len - i > 4) {
+        size += (len - i);
+        unsigned char buf[sizeof(uint8x16_t)];
+        memset(buf, ' ', sizeof(buf));
+        fast_memcpy16(buf, str, len - i);
+
+        uint8x16_t chunk  = vld1q_u8((const unsigned char *)buf);
+        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;
+
+        total = size;
+    } else {
+        total = size + calculate_string_size(str, len - i, hibit_friendly_chars);
+    }
+#else
+    size_t total = size + calculate_string_size(str, len - i, hibit_friendly_chars);
+#endif
+
+    return total;
+#elif defined(HAVE_SIMD_SSE4_2)
+    if (SIMD_Impl == SIMD_SSE42) {
+        if (len >= sizeof(__m128i)) {
+            return hibit_friendly_size_sse42(str, len);
+        }
+    }
     return calculate_string_size(str, len, hibit_friendly_chars);
+#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) {
@@ -183,10 +321,52 @@ 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;
+    long     size = 0;
+    uint32_t hi   = 0;
+
+#ifdef HAVE_SIMD_NEON
+    size_t i = 0;
+
+    if (len >= sizeof(uint8x16_t)) {
+        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;
+    }
+
+    size_t len_remaining = len - i;
+
+    for (; i < len; str++, i++) {
+        size += rails_xss_friendly_chars[*str];
+        hi |= *str & 0x80;
+    }
+
+    size -= (len_remaining * ((size_t)'0'));
+
+    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 +375,53 @@ 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;
+    long     size = 0;
+    uint32_t hi   = 0;
+#ifdef HAVE_SIMD_NEON
+    size_t i     = 0;
+    long   extra = 0;
+
+    if (len >= sizeof(uint8x16_t)) {
+        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++, extra++) {
+        size += rails_friendly_chars[*str];
+        hi |= *str & 0x80;
+    }
+
+    size -= (extra * ((size_t)'0'));
+
+    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 +430,10 @@ 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 +698,7 @@ 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 +932,13 @@ 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 +979,155 @@ void oj_dump_raw_json(VALUE obj, int depth, Out out) {
     }
 }
 
+#if defined(__clang__) || defined(__GNUC__)
+#define FORCE_INLINE __attribute__((always_inline))
+#else
+#define FORCE_INLINE
+#endif
+
+#ifdef HAVE_SIMD_NEON
+typedef struct _neon_match_result {
+    uint8x16_t needs_escape;
+    bool       has_some_hibit;
+    bool       do_unicode_validation;
+    uint64_t   escape_mask;
+} neon_match_result;
+
+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, .escape_mask = 0};
+
+    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;
+    }
+    const uint8x8_t res  = vshrn_n_u16(vreinterpretq_u16_u8(vmvnq_u8(vceqq_u8(result.needs_escape, vdupq_n_u8(0)))), 4);
+    const uint64_t  mask = vget_lane_u64(vreinterpret_u64_u8(res), 0);
+    result.escape_mask   = mask & 0x8888888888888888ull;
+    return result;
+}
+
+#elif defined(HAVE_SIMD_SSE4_2)
+typedef struct _sse42_match_result {
+    __m128i actions;
+    bool    needs_escape;
+    int     escape_mask;
+    bool    has_some_hibit;
+    bool    do_unicode_validation;
+} sse42_match_result;
+
+static inline OJ_TARGET_SSE42 sse42_match_result
+sse42_update(const char *str, __m128i *cmap_sse42, int sse42_tab_size, bool do_unicode_validation, bool has_hi) {
+    sse42_match_result result = {.has_some_hibit = false, .do_unicode_validation = false};
+
+    __m128i chunk        = _mm_loadu_si128((__m128i *)str);
+    __m128i actions      = vector_lookup_sse42(chunk, cmap_sse42, sse42_tab_size);
+    __m128i needs_escape = _mm_xor_si128(_mm_cmpeq_epi8(actions, _mm_setzero_si128()), _mm_set1_epi8(0xFF));
+    result.actions       = _mm_add_epi8(actions, _mm_set1_epi8('1'));
+
+    result.escape_mask  = _mm_movemask_epi8(needs_escape);
+    result.needs_escape = result.escape_mask != 0;
+    if (has_hi && do_unicode_validation) {
+        __m128i has_some_hibit       = _mm_and_si128(chunk, _mm_set1_epi8(0x80));
+        result.has_some_hibit        = _mm_movemask_epi8(has_some_hibit) != 0;
+        result.do_unicode_validation = has_hi && do_unicode_validation && result.has_some_hibit;
+    }
+    return result;
+}
+
+#endif /* HAVE_SIMD_NEON */
+
+static inline FORCE_INLINE const char *process_character(char         action,
+                                                         const char  *str,
+                                                         const char  *end,
+                                                         Out          out,
+                                                         const char  *orig,
+                                                         bool         do_unicode_validation,
+                                                         const char **check_start_) {
+    const char *check_start = *check_start_;
+    switch (action) {
+    case '1':
+        if (do_unicode_validation && check_start <= str) {
+            if (0 != (0x80 & (uint8_t)*str)) {
+                if (0xC0 == (0xC0 & (uint8_t)*str)) {
+                    *check_start_ = check_unicode(str, end, orig);
+                } else {
+                    raise_invalid_unicode(orig, (int)(end - orig), (int)(str - orig));
+                }
+            }
+        }
+        *out->cur++ = *str;
+        break;
+    case '2':
+        *out->cur++ = '\\';
+        switch (*str) {
+        case '\\': *out->cur++ = '\\'; break;
+        case '\b': *out->cur++ = 'b'; break;
+        case '\t': *out->cur++ = 't'; break;
+        case '\n': *out->cur++ = 'n'; break;
+        case '\f': *out->cur++ = 'f'; break;
+        case '\r': *out->cur++ = 'r'; break;
+        default: *out->cur++ = *str; break;
+        }
+        break;
+    case '3':  // Unicode
+        if (0xe2 == (uint8_t)*str && do_unicode_validation && 2 <= end - str) {
+            if (0x80 == (uint8_t)str[1] && (0xa8 == (uint8_t)str[2] || 0xa9 == (uint8_t)str[2])) {
+                str = dump_unicode(str, end, out, orig);
+            } else {
+                *check_start_ = check_unicode(str, end, orig);
+                *out->cur++   = *str;
+            }
+            break;
+        }
+        str = dump_unicode(str, end, out, orig);
+        break;
+    case '6':  // control characters
+        if (*(uint8_t *)str < 0x80) {
+            if (0 == (uint8_t)*str && out->opts->dump_opts.omit_null_byte) {
+                break;
+            }
+            APPEND_CHARS(out->cur, "\\u00", 4);
+            dump_hex((uint8_t)*str, out);
+        } else {
+            if (0xe2 == (uint8_t)*str && do_unicode_validation && 2 <= end - str) {
+                if (0x80 == (uint8_t)str[1] && (0xa8 == (uint8_t)str[2] || 0xa9 == (uint8_t)str[2])) {
+                    str = dump_unicode(str, end, out, orig);
+                } else {
+                    *check_start_ = check_unicode(str, end, orig);
+                    *out->cur++   = *str;
+                }
+                break;
+            }
+            str = dump_unicode(str, end, out, orig);
+        }
+        break;
+    default: break;  // ignore, should never happen if the table is correct
+    }
+
+    return str;
+}
+
 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 = 0;
+#elif defined(HAVE_SIMD_SSE4_2)
+    __m128i *cmap_sse42 = NULL;
+    int      sse42_tab_size;
+#endif /* HAVE_SIMD_NEON */
     const char *orig                  = str;
     bool        has_hi                = false;
     bool        do_unicode_validation = false;
@@ -792,7 +1159,11 @@ 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 +1176,11 @@ 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 +1191,15 @@ 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;
+#elif defined(HAVE_SIMD_SSE4_2)
+        cmap_sse42     = hibit_friendly_chars_sse42;
+        sse42_tab_size = 8;
+#endif /* HAVE_NEON_SIMD */
+        size = hibit_friendly_size((uint8_t *)str, cnt);
     }
     assure_size(out, size + BUFFER_EXTRA);
     *out->cur++ = '"';
@@ -833,7 +1216,16 @@ void oj_dump_cstr(const char *str, size_t cnt, bool is_sym, bool escape1, Out ou
         if (is_sym) {
             *out->cur++ = ':';
         }
+#ifdef HAVE_FAST_MEMCPY
+        if (cnt <= 16) {
+            fast_memcpy16(out->cur, str, cnt);
+            out->cur += size;
+        } else {
+            APPEND_CHARS(out->cur, str, cnt);
+        }
+#else
         APPEND_CHARS(out->cur, str, cnt);
+#endif
         *out->cur++ = '"';
     } else {
         const char *end         = str + cnt;
@@ -842,71 +1234,168 @@ void oj_dump_cstr(const char *str, size_t cnt, bool is_sym, bool escape1, Out ou
         if (is_sym) {
             *out->cur++ = ':';
         }
-        for (; str < end; str++) {
-            switch (cmap[(uint8_t)*str]) {
-            case '1':
-                if (do_unicode_validation && check_start <= str) {
-                    if (0 != (0x80 & (uint8_t)*str)) {
-                        if (0xC0 == (0xC0 & (uint8_t)*str)) {
-                            check_start = check_unicode(str, end, orig);
-                        } else {
-                            raise_invalid_unicode(orig, (int)(end - orig), (int)(str - orig));
-                        }
-                    }
-                }
-                *out->cur++ = *str;
-                break;
-            case '2':
-                *out->cur++ = '\\';
-                switch (*str) {
-                case '\\': *out->cur++ = '\\'; break;
-                case '\b': *out->cur++ = 'b'; break;
-                case '\t': *out->cur++ = 't'; break;
-                case '\n': *out->cur++ = 'n'; break;
-                case '\f': *out->cur++ = 'f'; break;
-                case '\r': *out->cur++ = 'r'; break;
-                default: *out->cur++ = *str; break;
+
+#if defined(HAVE_SIMD_NEON) || defined(HAVE_SIMD_SSE4_2)
+
+#define SEARCH_FLUSH                                  \
+    if (str > cursor) {                               \
+        APPEND_CHARS(out->cur, cursor, str - cursor); \
+        cursor = str;                                 \
+    }
+
+        const char *chunk_start;
+        const char *chunk_end;
+        const char *cursor = str;
+        char        matches[16];
+#endif /* HAVE_SIMD_NEON || HAVE_SIMD_SSE4_2 */
+
+#if defined(HAVE_SIMD_NEON)
+        bool use_simd = (cmap_neon != NULL && cnt >= (sizeof(uint8x16_t))) ? true : false;
+#elif defined(HAVE_SIMD_SSE4_2)
+        bool use_simd = false;
+        if (SIMD_Impl == SIMD_SSE42) {
+            use_simd = (cmap_sse42 != NULL && cnt >= (sizeof(__m128i))) ? true : false;
+        }
+#endif
+
+#ifdef HAVE_SIMD_NEON
+
+#ifdef HAVE_FAST_MEMCPY
+#define APPEND_CHARS_SMALL(dst, src, length) \
+    fast_memcpy16((dst), (src), (length));   \
+    (dst) += (length);
+#define MEMCPY16 fast_memcpy16
+#else
+#define APPEND_CHARS_SMALL(dst, src, length) APPEND_CHARS(dst, str, length);
+#define MEMCPY16 memcpy
+#endif
+
+        if (use_simd) {
+            while (str < end) {
+                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, ' ', sizeof(uint8x16_t));
+                    MEMCPY16(out->cur, str, (end - str));
+                    chunk_ptr   = out->cur;
+                    chunk_start = str;
+                    chunk_end   = end;
+                } else {
+                    break;
                 }
-                break;
-            case '3':  // Unicode
-                if (0xe2 == (uint8_t)*str && do_unicode_validation && 2 <= end - str) {
-                    if (0x80 == (uint8_t)str[1] && (0xa8 == (uint8_t)str[2] || 0xa9 == (uint8_t)str[2])) {
-                        str = dump_unicode(str, end, out, orig);
+                neon_match_result result = neon_update(chunk_ptr,
+                                                       cmap_neon,
+                                                       neon_table_size,
+                                                       do_unicode_validation,
+                                                       has_hi);
+                if ((result.do_unicode_validation) || result.escape_mask != 0) {
+                    SEARCH_FLUSH;
+                    bool       process_each = result.do_unicode_validation;
+                    uint8x16_t actions      = vaddq_u8(result.needs_escape, vdupq_n_u8('1'));
+                    vst1q_u8((unsigned char *)matches, actions);
+                    // If no byte in this chunk had the high bit set then we can skip
+                    // all of the '1' bytes by directly copying them to the output.
+                    if (!process_each) {
+                        while (result.escape_mask) {
+                            int  esc_pos = OJ_CTZ64(result.escape_mask) >> 2;
+                            long run_len = esc_pos - (str - chunk_start);
+                            if (run_len > 0) {
+                                APPEND_CHARS_SMALL(out->cur, str, run_len);
+                                str += run_len;
+                            }
+                            str = process_character(matches[esc_pos],
+                                                    str,
+                                                    end,
+                                                    out,
+                                                    orig,
+                                                    do_unicode_validation,
+                                                    &check_start);
+                            str++;
+                            result.escape_mask &= result.escape_mask - 1;
+                        }
+                        if (str < chunk_end) {
+                            APPEND_CHARS_SMALL(out->cur, str, chunk_end - str);
+                            str = chunk_end;
+                        }
                     } else {
-                        check_start = check_unicode(str, end, orig);
-                        *out->cur++ = *str;
+                        while (str < chunk_end) {
+                            long match_index = str - chunk_start;
+                            str              = process_character(matches[match_index],
+                                                                 str,
+                                                                 end,
+                                                                 out,
+                                                                 orig,
+                                                                 do_unicode_validation,
+                                                                 &check_start);
+                            str++;
+                        }
                     }
-                    break;
+                    cursor = str;
+                    continue;
                 }
-                str = dump_unicode(str, end, out, orig);
-                break;
-            case '6':  // control characters
-                if (*(uint8_t *)str < 0x80) {
-                    if (0 == (uint8_t)*str && out->opts->dump_opts.omit_null_byte) {
+                str = chunk_end;
+            }
+            SEARCH_FLUSH;
+        }
+#endif
+
+#ifdef HAVE_SIMD_SSE4_2
+        if (SIMD_Impl == SIMD_SSE42) {
+            if (use_simd) {
+                while (str < end) {
+                    const char *chunk_ptr = NULL;
+                    if (str + sizeof(__m128i) <= end) {
+                        chunk_ptr   = str;
+                        chunk_start = str;
+                        chunk_end   = str + sizeof(__m128i);
+                    } else if ((end - str) >= SIMD_MINIMUM_THRESHOLD) {
+                        memset(out->cur, 'A', sizeof(__m128i));
+                        memcpy(out->cur, str, (end - str));
+                        chunk_ptr   = out->cur;
+                        chunk_start = str;
+                        chunk_end   = end;
+                    } else {
                         break;
                     }
-                    APPEND_CHARS(out->cur, "\\u00", 4);
-                    dump_hex((uint8_t)*str, out);
-                } else {
-                    if (0xe2 == (uint8_t)*str && do_unicode_validation && 2 <= end - str) {
-                        if (0x80 == (uint8_t)str[1] && (0xa8 == (uint8_t)str[2] || 0xa9 == (uint8_t)str[2])) {
-                            str = dump_unicode(str, end, out, orig);
-                        } else {
-                            check_start = check_unicode(str, end, orig);
-                            *out->cur++ = *str;
+                    sse42_match_result result = sse42_update(chunk_ptr,
+                                                             cmap_sse42,
+                                                             sse42_tab_size,
+                                                             do_unicode_validation,
+                                                             has_hi);
+                    if ((result.do_unicode_validation) || result.needs_escape) {
+                        SEARCH_FLUSH;
+                        _mm_storeu_si128((__m128i *)matches, result.actions);
+                        while (str < chunk_end) {
+                            long match_index = str - chunk_start;
+                            str              = process_character(matches[match_index],
+                                                                 str,
+                                                                 end,
+                                                                 out,
+                                                                 orig,
+                                                                 do_unicode_validation,
+                                                                 &check_start);
+                            str++;
                         }
-                        break;
+                        cursor = str;
+                        continue;
                     }
-                    str = dump_unicode(str, end, out, orig);
+                    str = chunk_end;
                 }
-                break;
-            default: break;  // ignore, should never happen if the table is correct
+                SEARCH_FLUSH;
             }
         }
+#endif /* HAVE_SIMD_SSE4_2 */
+
+        for (; str < end; str++) {
+            str = process_character(cmap[(uint8_t)*str], str, end, out, orig, do_unicode_validation, &check_start);
+        }
         *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 +1446,7 @@ 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 +1581,7 @@ 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 +1603,7 @@ 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 +1714,7 @@ 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 +1728,8 @@ 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 +1737,29 @@ 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;
 }
+
+bool oj_key_skip(VALUE key, const char *only, const char *except) {
+    const char *skey;
+
+    switch (rb_type(key)) {
+    case RUBY_T_STRING: skey = StringValueCStr(key); break;
+    case RUBY_T_SYMBOL: skey = rb_id2name(rb_sym2id(key)); break;
+    default: skey = NULL; break;
+    }
+    if (NULL != skey && '\0' != *skey) {
+        size_t size = strlen(skey);
+        char  *buf  = alloca(size + 3);
+
+        buf[0] = ':';
+        strcpy(buf + 1, skey);
+        buf[size + 1] = ':';
+        buf[size + 2] = '\0';
+
+        return ((NULL != only && NULL == strstr(only, buf)) || (NULL != except && NULL != strstr(except, buf)));
+    }
+    return NULL != only;
+}