oj 2.18.5 → 3.16.11

data/ext/oj/dump.c

@@ -1,118 +1,46 @@
-/* dump.c
- * Copyright (c) 2012, Peter Ohler
- * All rights reserved.
- * 
- * 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.
- * 
- *  - Neither the name of Peter Ohler nor the names of its contributors may be
- *    used to endorse or promote products derived from this software without
- *    specific prior written permission.
- * 
- * 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 HOLDER 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.
- */
+// Copyright (c) 2012, 2017 Peter Ohler. All rights reserved.
+// Licensed under the MIT License. See LICENSE file in the project root for license details.
+
+#include "dump.h"
 
-#include <stdlib.h>
 #include <errno.h>
-#if !IS_WINDOWS
-#include <sys/time.h>
-#endif
-#include <time.h>
+#include <math.h>
+#include <stdint.h>
 #include <stdio.h>
+#include <stdlib.h>
 #include <string.h>
-#include <math.h>
 #include <unistd.h>
-#include <errno.h>
+#if !IS_WINDOWS
+#include <poll.h>
+#endif
 
-#include "oj.h"
 #include "cache8.h"
+#include "mem.h"
 #include "odd.h"
-
-#if !HAS_ENCODING_SUPPORT || defined(RUBINIUS_RUBY)
-#define rb_eEncodingError	rb_eException
-#endif
+#include "oj.h"
+#include "trace.h"
+#include "util.h"
 
 // Workaround in case INFINITY is not defined in math.h or if the OS is CentOS
-#define OJ_INFINITY (1.0/0.0)
-
-// Extra padding at end of buffer.
-#define BUFFER_EXTRA 10
+#define OJ_INFINITY (1.0 / 0.0)
 
 #define MAX_DEPTH 1000
 
-typedef unsigned long	ulong;
-
-static void	raise_strict(VALUE obj);
-static void	dump_val(VALUE obj, int depth, Out out, int argc, VALUE *argv, bool as_ok);
-static void	dump_nil(Out out);
-static void	dump_true(Out out);
-static void	dump_false(Out out);
-static void	dump_fixnum(VALUE obj, Out out);
-static void	dump_bignum(VALUE obj, Out out);
-static void	dump_float(VALUE obj, Out out);
-static void	dump_raw(const char *str, size_t cnt, Out out);
-static void	dump_cstr(const char *str, size_t cnt, int is_sym, int escape1, Out out);
-static void	dump_hex(uint8_t c, Out out);
-static void	dump_str_comp(VALUE obj, Out out);
-static void	dump_str_obj(VALUE obj, VALUE clas, int depth, Out out);
-static void	dump_sym_comp(VALUE obj, Out out);
-static void	dump_sym_obj(VALUE obj, Out out);
-static void	dump_class_comp(VALUE obj, Out out);
-static void	dump_class_obj(VALUE obj, Out out);
-static void	dump_array(VALUE obj, VALUE clas, int depth, Out out);
-static int	hash_cb_strict(VALUE key, VALUE value, Out out);
-static int	hash_cb_compat(VALUE key, VALUE value, Out out);
-static int	hash_cb_object(VALUE key, VALUE value, Out out);
-static void	dump_hash(VALUE obj, VALUE clas, int depth, int mode, Out out);
-static void	dump_time(VALUE obj, Out out, int withZone);
-static void	dump_ruby_time(VALUE obj, Out out);
-static void	dump_xml_time(VALUE obj, Out out);
-static void	dump_data_strict(VALUE obj, Out out);
-static void	dump_data_null(VALUE obj, Out out);
-static void	dump_data_comp(VALUE obj, int depth, Out out, int argc, VALUE *argv, bool as_ok);
-static void	dump_data_obj(VALUE obj, int depth, Out out);
-static void	dump_obj_comp(VALUE obj, int depth, Out out, int argc, VALUE *argv, bool as_ok);
-static void	dump_obj_obj(VALUE obj, int depth, Out out);
-static void	dump_struct_comp(VALUE obj, int depth, Out out, int argc, VALUE *argv, bool as_ok);
-static void	dump_struct_obj(VALUE obj, int depth, Out out);
-#if HAS_IVAR_HELPERS
-static int	dump_attr_cb(ID key, VALUE value, Out out);
-#endif
-static void	dump_obj_attrs(VALUE obj, VALUE clas, slot_t id, int depth, Out out);
-static void	dump_odd(VALUE obj, Odd odd, VALUE clas, int depth, Out out);
+static const char inf_val[]  = INF_VAL;
+static const char ninf_val[] = NINF_VAL;
+static const char nan_val[]  = NAN_VAL;
 
-static void	grow(Out out, size_t len);
-static size_t	hibit_friendly_size(const uint8_t *str, size_t len);
-static size_t	xss_friendly_size(const uint8_t *str, size_t len);
-static size_t	ascii_friendly_size(const uint8_t *str, size_t len);
+typedef unsigned long ulong;
 
-static void	dump_leaf(Leaf leaf, int depth, Out out);
-static void	dump_leaf_str(Leaf leaf, Out out);
-static void	dump_leaf_fixnum(Leaf leaf, Out out);
-static void	dump_leaf_float(Leaf leaf, Out out);
-static void	dump_leaf_array(Leaf leaf, int depth, Out out);
-static void	dump_leaf_hash(Leaf leaf, int depth, Out out);
+static size_t hibit_friendly_size(const uint8_t *str, size_t len);
+static size_t slash_friendly_size(const uint8_t *str, size_t len);
+static size_t xss_friendly_size(const uint8_t *str, size_t len);
+static size_t ascii_friendly_size(const uint8_t *str, size_t len);
 
-static const char	hex_chars[17] = "0123456789abcdef";
+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[256] = "\
 66666666221622666666666666666666\
 11211111111111111111111111111111\
 11111111111111111111111111112111\
@@ -123,7 +51,7 @@ static char	newline_friendly_chars[256] = "\
 11111111111111111111111111111111";
 
 // JSON standard
-static char	hibit_friendly_chars[256] = "\
+static char hibit_friendly_chars[256] = "\
 66666666222622666666666666666666\
 11211111111111111111111111111111\
 11111111111111111111111111112111\
@@ -133,9 +61,20 @@ static char	hibit_friendly_chars[256] = "\
 11111111111111111111111111111111\
 11111111111111111111111111111111";
 
+// JSON standard but escape forward slashes `/`
+static char slash_friendly_chars[256] = "\
+66666666222622666666666666666666\
+11211111111111121111111111111111\
+11111111111111111111111111112111\
+11111111111111111111111111111111\
+11111111111111111111111111111111\
+11111111111111111111111111111111\
+11111111111111111111111111111111\
+11111111111111111111111111111111";
+
 // 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[256] = "\
 66666666222622666666666666666666\
 11211111111111111111111111111111\
 11111111111111111111111111112111\
@@ -146,7 +85,7 @@ static char	ascii_friendly_chars[256] = "\
 33333333333333333333333333333333";
 
 // XSS safe mode
-static char	xss_friendly_chars[256] = "\
+static char xss_friendly_chars[256] = "\
 66666666222622666666666666666666\
 11211161111111121111111111116161\
 11111111111111111111111111112111\
@@ -156,2616 +95,1459 @@ static char	xss_friendly_chars[256] = "\
 33333333333333333333333333333333\
 33333333333333333333333333333333";
 
-inline static size_t
-newline_friendly_size(const uint8_t *str, size_t len) {
-    size_t	size = 0;
+// JSON XSS combo
+static char hixss_friendly_chars[256] = "\
+66666666222622666666666666666666\
+11211111111111111111111111111111\
+11111111111111111111111111112111\
+11111111111111111111111111111111\
+11111111111111111111111111111111\
+11111111111111111111111111111111\
+11111111111111111111111111111111\
+11611111111111111111111111111111";
 
-    for (; 0 < len; str++, len--) {
-	size += newline_friendly_chars[*str];
-    }
-    return size - len * (size_t)'0';
-}
+// Rails XSS combo
+static char rails_xss_friendly_chars[256] = "\
+66666666222622666666666666666666\
+11211161111111111111111111116161\
+11111111111111111111111111112111\
+11111111111111111111111111111111\
+11111111111111111111111111111111\
+11111111111111111111111111111111\
+11111111111111111111111111111111\
+11611111111111111111111111111111";
 
-inline static size_t
-hibit_friendly_size(const uint8_t *str, size_t len) {
-    size_t	size = 0;
+// Rails HTML non-escape
+static char rails_friendly_chars[256] = "\
+66666666222622666666666666666666\
+11211111111111111111111111111111\
+11111111111111111111111111112111\
+11111111111111111111111111111111\
+11111111111111111111111111111111\
+11111111111111111111111111111111\
+11111111111111111111111111111111\
+11111111111111111111111111111111";
 
-    for (; 0 < len; str++, len--) {
-	size += hibit_friendly_chars[*str];
-    }
-    return size - len * (size_t)'0';
+static void raise_strict(VALUE obj) {
+    rb_raise(rb_eTypeError, "Failed to dump %s Object to JSON in strict mode.", rb_class2name(rb_obj_class(obj)));
 }
 
-inline static size_t
-ascii_friendly_size(const uint8_t *str, size_t len) {
-    size_t	size = 0;
+inline static size_t calculate_string_size(const uint8_t *str, size_t len, const char *table) {
+    size_t size = 0;
+    size_t i    = len;
 
-    for (; 0 < len; str++, len--) {
-	size += ascii_friendly_chars[*str];
+    for (; 3 < i; i -= 4) {
+        size += table[*str++];
+        size += table[*str++];
+        size += table[*str++];
+        size += table[*str++];
+    }
+    for (; 0 < i; i--) {
+        size += table[*str++];
     }
     return size - len * (size_t)'0';
 }
 
-inline static size_t
-xss_friendly_size(const uint8_t *str, size_t len) {
-    size_t	size = 0;
-
-    for (; 0 < len; str++, len--) {
-	size += xss_friendly_chars[*str];
-    }
-    return size - len * (size_t)'0';
+inline static size_t newline_friendly_size(const uint8_t *str, size_t len) {
+    return calculate_string_size(str, len, newline_friendly_chars);
 }
 
-inline static void
-fill_indent(Out out, int cnt) {
-    if (0 < out->indent) {
-	cnt *= out->indent;
-	*out->cur++ = '\n';
-	for (; 0 < cnt; cnt--) {
-	    *out->cur++ = ' ';
-	}
-    }
+#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;
 }
 
-inline static void
-dump_ulong(unsigned long num, Out out) {
-    char	buf[32];
-    char	*b = buf + sizeof(buf) - 1;
+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];
 
-    *b-- = '\0';
-    if (0 < num) {
-	for (; 0 < num; num /= 10, b--) {
-	    *b = (num % 10) + '0';
-	}
-	b++;
-    } else {
-	*b = '0';
-    }
-    for (; '\0' != *b; b++) {
-	*out->cur++ = *b;
-    }
-    *out->cur = '\0';
-}
+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);
 
-static void
-grow(Out out, size_t len) {
-    size_t  size = out->end - out->buf;
-    long    pos = out->cur - out->buf;
-    char    *buf;
-	
-    size *= 2;
-    if (size <= len * 2 + pos) {
-	size += len;
-    }
-    if (out->allocated) {
-	buf = REALLOC_N(out->buf, char, (size + BUFFER_EXTRA));
-    } else {
-	buf = ALLOC_N(char, (size + BUFFER_EXTRA));
-	out->allocated = 1;
-	memcpy(buf, out->buf, out->end - out->buf + BUFFER_EXTRA);
+    // 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);
+        }
     }
-    if (0 == buf) {
-	rb_raise(rb_eNoMemError, "Failed to create string. [%d:%s]\n", ENOSPC, strerror(ENOSPC));
+
+    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);
+        }
     }
-    out->buf = buf;
-    out->end = buf + size;
-    out->cur = out->buf + pos;
 }
+#endif
 
-inline static void
-dump_hex(uint8_t c, Out out) {
-    uint8_t	d = (c >> 4) & 0x0F;
+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;
 
-    *out->cur++ = hex_chars[d];
-    d = c & 0x0F;
-    *out->cur++ = hex_chars[d];
-}
+    for (; i + sizeof(uint8x16_t) <= len; i += sizeof(uint8x16_t), str += sizeof(uint8x16_t)) {
+        size += sizeof(uint8x16_t);
 
-static void
-dump_raw(const char *str, size_t cnt, Out out) {
-    if (out->end - out->cur <= (long)cnt + 10) {
-	grow(out, cnt + 10);
+        // 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;
     }
-    memcpy(out->cur, str, cnt);
-    out->cur += cnt;
-    *out->cur = '\0';
-}
 
-const char*
-dump_unicode(const char *str, const char *end, Out out) {
-    uint32_t	code = 0;
-    uint8_t	b = *(uint8_t*)str;
-    int		i, cnt;
-    
-    if (0xC0 == (0xE0 & b)) {
-	cnt = 1;
-	code = b & 0x0000001F;
-    } else if (0xE0 == (0xF0 & b)) {
-	cnt = 2;
-	code = b & 0x0000000F;
-    } else if (0xF0 == (0xF8 & b)) {
-	cnt = 3;
-	code = b & 0x00000007;
-    } else if (0xF8 == (0xFC & b)) {
-	cnt = 4;
-	code = b & 0x00000003;
-    } else if (0xFC == (0xFE & b)) {
-	cnt = 5;
-	code = b & 0x00000001;
-    } else {
-	cnt = 0;
-	rb_raise(rb_eEncodingError, "Invalid Unicode\n");
-    }
-    str++;
-    for (; 0 < cnt; cnt--, str++) {
-	b = *(uint8_t*)str;
-	if (end <= str || 0x80 != (0xC0 & b)) {
-	    rb_raise(rb_eEncodingError, "Invalid Unicode\n");
-	}
-	code = (code << 6) | (b & 0x0000003F);
-    }
-    if (0x0000FFFF < code) {
-	uint32_t	c1;
-
-	code -= 0x00010000;
-	c1 = ((code >> 10) & 0x000003FF) + 0x0000D800;
-	code = (code & 0x000003FF) + 0x0000DC00;
-	*out->cur++ = '\\';
-	*out->cur++ = 'u';
-	for (i = 3; 0 <= i; i--) {
-	    *out->cur++ = hex_chars[(uint8_t)(c1 >> (i * 4)) & 0x0F];
-	}
-    }
-    *out->cur++ = '\\';
-    *out->cur++ = 'u';
-    for (i = 3; 0 <= i; i--) {
-	*out->cur++ = hex_chars[(uint8_t)(code >> (i * 4)) & 0x0F];
-    }	
-    return str - 1;
+    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
 }
 
-// returns 0 if not using circular references, -1 if not further writing is
-// needed (duplicate), and a positive value if the object was added to the cache.
-static long
-check_circular(VALUE obj, Out out) {
-    slot_t	id = 0;
-    slot_t	*slot;
-
-    if (ObjectMode == out->opts->mode && Yes == out->opts->circular) {
-	if (0 == (id = oj_cache8_get(out->circ_cache, obj, &slot))) {
-	    out->circ_cnt++;
-	    id = out->circ_cnt;
-	    *slot = id;
-	} else {
-	    if (out->end - out->cur <= 18) {
-		grow(out, 18);
-	    }
-	    *out->cur++ = '"';
-	    *out->cur++ = '^';
-	    *out->cur++ = 'r';
-	    dump_ulong(id, out);
-	    *out->cur++ = '"';
-
-	    return -1;
-	}
-    }
-    return (long)id;
+inline static size_t slash_friendly_size(const uint8_t *str, size_t len) {
+    return calculate_string_size(str, len, slash_friendly_chars);
 }
 
-static void
-dump_nil(Out out) {
-    size_t	size = 4;
+inline static size_t ascii_friendly_size(const uint8_t *str, size_t len) {
+    return calculate_string_size(str, len, ascii_friendly_chars);
+}
 
-    if (out->end - out->cur <= (long)size) {
-	grow(out, size);
-    }
-    *out->cur++ = 'n';
-    *out->cur++ = 'u';
-    *out->cur++ = 'l';
-    *out->cur++ = 'l';
-    *out->cur = '\0';
+inline static size_t xss_friendly_size(const uint8_t *str, size_t len) {
+    return calculate_string_size(str, len, xss_friendly_chars);
 }
 
-static void
-dump_true(Out out) {
-    size_t	size = 4;
+inline static size_t hixss_friendly_size(const uint8_t *str, size_t len) {
+    size_t size  = 0;
+    size_t i     = len;
+    bool   check = false;
 
-    if (out->end - out->cur <= (long)size) {
-	grow(out, size);
+    for (; 0 < i; str++, i--) {
+        size += hixss_friendly_chars[*str];
+        if (0 != (0x80 & *str)) {
+            check = true;
+        }
     }
-    *out->cur++ = 't';
-    *out->cur++ = 'r';
-    *out->cur++ = 'u';
-    *out->cur++ = 'e';
-    *out->cur = '\0';
+    return size - len * (size_t)'0' + check;
 }
 
-static void
-dump_false(Out out) {
-    size_t	size = 5;
+inline static long rails_xss_friendly_size(const uint8_t *str, size_t len) {
+    long    size = 0;
+    uint8_t hi   = 0;
 
-    if (out->end - out->cur <= (long)size) {
-	grow(out, size);
-    }
-    *out->cur++ = 'f';
-    *out->cur++ = 'a';
-    *out->cur++ = 'l';
-    *out->cur++ = 's';
-    *out->cur++ = 'e';
-    *out->cur = '\0';
-}
+#ifdef HAVE_SIMD_NEON
+    size_t i = 0;
 
-static void
-dump_fixnum(VALUE obj, Out out) {
-    char	buf[32];
-    char	*b = buf + sizeof(buf) - 1;
-    long long	num = rb_num2ll(obj);
-    int		neg = 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);
 
-    if (0 > num) {
-	neg = 1;
-	num = -num;
-    }
-    *b-- = '\0';
-    if (0 < num) {
-	for (; 0 < num; num /= 10, b--) {
-	    *b = (num % 10) + '0';
-	}
-	if (neg) {
-	    *b = '-';
-	} else {
-	    b++;
-	}
-    } else {
-	*b = '0';
-    }
-    if (out->end - out->cur <= (long)(sizeof(buf) - (b - buf))) {
-	grow(out, sizeof(buf) - (b - buf));
-    }
-    for (; '\0' != *b; b++) {
-	*out->cur++ = *b;
-    }
-    *out->cur = '\0';
-}
+        uint8x16_t chunk = vld1q_u8(str);
 
-static void
-dump_bignum(VALUE obj, Out out) {
-    volatile VALUE	rs = rb_big2str(obj, 10);
-    int			cnt = (int)RSTRING_LEN(rs);
+        // 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));
 
-    if (out->end - out->cur <= (long)cnt) {
-	grow(out, cnt);
+        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;
     }
-    memcpy(out->cur, rb_string_value_ptr((VALUE*)&rs), cnt);
-    out->cur += cnt;
-    *out->cur = '\0';
-}
 
-static const char	inf_val[] = INF_VAL;
-static const char	ninf_val[] = NINF_VAL;
-static const char	nan_val[] = NAN_VAL;
+    // '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;
 
-// Removed dependencies on math due to problems with CentOS 5.4.
-static void
-dump_float(VALUE obj, Out out) {
-    char	buf[64];
-    char	*b;
-    double	d = rb_num2dbl(obj);
-    int		cnt = 0;
-
-    if (0.0 == d) {
-	b = buf;
-	*b++ = '0';
-	*b++ = '.';
-	*b++ = '0';
-	*b++ = '\0';
-	cnt = 3;
-    } else if (OJ_INFINITY == d) {
-	if (ObjectMode == out->opts->mode) {
-	    strcpy(buf, inf_val);
-	    cnt = sizeof(inf_val) - 1;
-	} else {
-	    NanDump	nd = out->opts->dump_opts.nan_dump;
-	    
-	    if (AutoNan == nd) {
-		switch (out->opts->mode) {
-		case CompatMode:	nd = WordNan;	break;
-		case StrictMode:	nd = RaiseNan;	break;
-		case NullMode:		nd = NullNan;	break;
-		default:				break;
-		}
-	    }
-	    switch (nd) {
-	    case RaiseNan:
-		raise_strict(obj);
-		break;
-	    case WordNan:
-		strcpy(buf, "Infinity");
-		cnt = 8;
-		break;
-	    case NullNan:
-		strcpy(buf, "null");
-		cnt = 4;
-		break;
-	    case HugeNan:
-	    default:
-		strcpy(buf, inf_val);
-		cnt = sizeof(inf_val) - 1;
-		break;
-	    }
-	}
-    } else if (-OJ_INFINITY == d) {
-	if (ObjectMode == out->opts->mode) {
-	    strcpy(buf, ninf_val);
-	    cnt = sizeof(ninf_val) - 1;
-	} else {
-	    NanDump	nd = out->opts->dump_opts.nan_dump;
-	    
-	    if (AutoNan == nd) {
-		switch (out->opts->mode) {
-		case CompatMode:	nd = WordNan;	break;
-		case StrictMode:	nd = RaiseNan;	break;
-		case NullMode:		nd = NullNan;	break;
-		default:				break;
-		}
-	    }
-	    switch (nd) {
-	    case RaiseNan:
-		raise_strict(obj);
-		break;
-	    case WordNan:
-		strcpy(buf, "-Infinity");
-		cnt = 9;
-		break;
-	    case NullNan:
-		strcpy(buf, "null");
-		cnt = 4;
-		break;
-	    case HugeNan:
-	    default:
-		strcpy(buf, ninf_val);
-		cnt = sizeof(ninf_val) - 1;
-		break;
-	    }
-	}
-    } else if (isnan(d)) {
-	if (ObjectMode == out->opts->mode) {
-	    strcpy(buf, nan_val);
-	    cnt = sizeof(nan_val) - 1;
-	} else {
-	    NanDump	nd = out->opts->dump_opts.nan_dump;
-	    
-	    if (AutoNan == nd) {
-		switch (out->opts->mode) {
-		case CompatMode:	nd = WordNan;	break;
-		case StrictMode:	nd = RaiseNan;	break;
-		case NullMode:		nd = NullNan;	break;
-		default:				break;
-		}
-	    }
-	    switch (nd) {
-	    case RaiseNan:
-		raise_strict(obj);
-		break;
-	    case WordNan:
-		strcpy(buf, "NaN");
-		cnt = 3;
-		break;
-	    case NullNan:
-		strcpy(buf, "null");
-		cnt = 4;
-		break;
-	    case HugeNan:
-	    default:
-		strcpy(buf, nan_val);
-		cnt = sizeof(nan_val) - 1;
-		break;
-	    }
-	}
-    } else if (d == (double)(long long int)d) {
-	cnt = snprintf(buf, sizeof(buf), "%.1f", d);
-    } else if (0 == out->opts->float_prec) {
-	volatile VALUE	rstr = rb_funcall(obj, oj_to_s_id, 0);
-
-	cnt = (int)RSTRING_LEN(rstr);
-	if ((int)sizeof(buf) <= cnt) {
-	    cnt = sizeof(buf) - 1;
-	}
-	strncpy(buf, rb_string_value_ptr((VALUE*)&rstr), cnt);
-	buf[cnt] = '\0';
-    } else {
-	cnt = snprintf(buf, sizeof(buf), out->opts->float_fmt, d);
+    for (; i < len; str++, i++) {
+        size += rails_xss_friendly_chars[*str] - '0';
+        hi |= *str & 0x80;
+    }
+    if (0 == hi) {
+        return size;
     }
-    if (out->end - out->cur <= (long)cnt) {
-	grow(out, cnt);
+    return -(size);
+#else
+    size_t i = len;
+    for (; 0 < i; str++, i--) {
+        size += rails_xss_friendly_chars[*str];
+        hi |= *str & 0x80;
     }
-    for (b = buf; '\0' != *b; b++) {
-	*out->cur++ = *b;
+    if (0 == hi) {
+        return size - len * (size_t)'0';
     }
-    *out->cur = '\0';
+    return -(size - len * (size_t)'0');
+#endif /* HAVE_SIMD_NEON */
 }
 
-static void
-dump_cstr(const char *str, size_t cnt, int is_sym, int escape1, Out out) {
-    size_t	size;
-    char	*cmap;
+inline static size_t rails_friendly_size(const uint8_t *str, size_t len) {
+    long    size = 0;
+    uint8_t hi   = 0;
+#ifdef HAVE_SIMD_NEON
+    size_t i = 0;
 
-    switch (out->opts->escape_mode) {
-    case NLEsc:
-	cmap = newline_friendly_chars;
-	size = newline_friendly_size((uint8_t*)str, cnt);
-	break;
-    case ASCIIEsc:
-	cmap = ascii_friendly_chars;
-	size = ascii_friendly_size((uint8_t*)str, cnt);
-	break;
-    case XSSEsc:
-	cmap = xss_friendly_chars;
-	size = xss_friendly_size((uint8_t*)str, cnt);
-	break;
-    case JSONEsc:
-    default:
-	cmap = hibit_friendly_chars;
-	size = hibit_friendly_size((uint8_t*)str, cnt);
-    }
-    if (out->end - out->cur <= (long)size + BUFFER_EXTRA) { // extra 10 for escaped first char, quotes, and sym
-	grow(out, size + BUFFER_EXTRA);
-    }
-    *out->cur++ = '"';
-    if (escape1) {
-	*out->cur++ = '\\';
-	*out->cur++ = 'u';
-	*out->cur++ = '0';
-	*out->cur++ = '0';
-	dump_hex((uint8_t)*str, out);
-	cnt--;
-	size--;
-	str++;
-	is_sym = 0; // just to make sure
-    }
-    if (cnt == size) {
-	if (is_sym) {
-	    *out->cur++ = ':';
-	}
-	for (; '\0' != *str; str++) {
-	    *out->cur++ = *str;
-	}
-	*out->cur++ = '"';
-    } else {
-	const char	*end = str + cnt;
-
-	if (is_sym) {
-	    *out->cur++ = ':';
-	}
-	for (; str < end; str++) {
-	    switch (cmap[(uint8_t)*str]) {
-	    case '1':
-		*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
-		str = dump_unicode(str, end, out);
-		break;
-	    case '6': // control characters
-		*out->cur++ = '\\';
-		*out->cur++ = 'u';
-		*out->cur++ = '0';
-		*out->cur++ = '0';
-		dump_hex((uint8_t)*str, out);
-		break;
-	    default:
-		break; // ignore, should never happen if the table is correct
-	    }
-	}
-	*out->cur++ = '"';
-    }
-    *out->cur = '\0';
-}
+    uint8x16_t has_some_hibit = vdupq_n_u8(0);
+    uint8x16_t hibit          = vdupq_n_u8(0x80);
 
-static void
-dump_str_comp(VALUE obj, Out out) {
-    dump_cstr(rb_string_value_ptr((VALUE*)&obj), (int)RSTRING_LEN(obj), 0, 0, out);
-}
+    for (; i + sizeof(uint8x16_t) <= len; i += sizeof(uint8x16_t), str += sizeof(uint8x16_t)) {
+        size += sizeof(uint8x16_t);
 
-static void
-dump_str_obj(VALUE obj, VALUE clas, int depth, Out out) {
-    if (Qundef != clas && rb_cString != clas) {
-	dump_obj_attrs(obj, clas, 0, depth, out);
-    } else {
-	const char	*s = rb_string_value_ptr((VALUE*)&obj);
-	size_t		len = (size_t)RSTRING_LEN(obj);
-	char		s1 = s[1];
+        // See https://lemire.me/blog/2019/07/23/arbitrary-byte-to-byte-maps-using-arm-neon/
+        uint8x16_t chunk = vld1q_u8(str);
 
-	dump_cstr(s, len, 0, (':' == *s || ('^' == *s && ('r' == s1 || 'i' == s1))), out);
+        // 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;
     }
-}
 
-static void
-dump_sym_comp(VALUE obj, Out out) {
-    const char	*sym = rb_id2name(SYM2ID(obj));
-    
-    dump_cstr(sym, strlen(sym), 0, 0, out);
-}
+    // '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;
 
-static void
-dump_sym_obj(VALUE obj, Out out) {
-    const char	*sym = rb_id2name(SYM2ID(obj));
-    size_t	len = strlen(sym);
-    
-    dump_cstr(sym, len, 1, 0, out);
+    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;
+    }
+    if (0 == hi) {
+        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, size_t *lenp) {
+    const char *str = NULL;
+
+    if (AutoNan == opt) {
+        switch (mode) {
+        case CompatMode: opt = WordNan; break;
+        case StrictMode: opt = RaiseNan; break;
+        default: break;
+        }
+    }
+    switch (opt) {
+    case RaiseNan: raise_strict(obj); break;
+    case WordNan:
+        if (plus) {
+            str   = "Infinity";
+            *lenp = 8;
+        } else {
+            str   = "-Infinity";
+            *lenp = 9;
+        }
+        break;
+    case NullNan:
+        str   = "null";
+        *lenp = 4;
+        break;
+    case HugeNan:
+    default:
+        if (plus) {
+            str   = inf_val;
+            *lenp = sizeof(inf_val) - 1;
+        } else {
+            str   = ninf_val;
+            *lenp = sizeof(ninf_val) - 1;
+        }
+        break;
+    }
+    return str;
 }
 
-static void
-dump_class_comp(VALUE obj, Out out) {
-    const char	*s = rb_class2name(obj);
+inline static void dump_hex(uint8_t c, Out out) {
+    uint8_t d = (c >> 4) & 0x0F;
 
-    dump_cstr(s, strlen(s), 0, 0, out);
+    *out->cur++ = hex_chars[d];
+    d           = c & 0x0F;
+    *out->cur++ = hex_chars[d];
 }
 
-static void
-dump_class_obj(VALUE obj, Out out) {
-    const char	*s = rb_class2name(obj);
-    size_t	len = strlen(s);
+static void raise_invalid_unicode(const char *str, int len, int pos) {
+    char    c;
+    char    code[32];
+    char   *cp = code;
+    int     i;
+    uint8_t d;
 
-    if (out->end - out->cur <= 6) {
-	grow(out, 6);
+    *cp++ = '[';
+    for (i = pos; i < len && i - pos < 5; i++) {
+        c     = str[i];
+        d     = (c >> 4) & 0x0F;
+        *cp++ = hex_chars[d];
+        d     = c & 0x0F;
+        *cp++ = hex_chars[d];
+        *cp++ = ' ';
     }
-    *out->cur++ = '{';
-    *out->cur++ = '"';
-    *out->cur++ = '^';
-    *out->cur++ = 'c';
-    *out->cur++ = '"';
-    *out->cur++ = ':';
-    dump_cstr(s, len, 0, 0, out);
-    *out->cur++ = '}';
-    *out->cur = '\0';
+    cp--;
+    *cp++ = ']';
+    *cp   = '\0';
+    rb_raise(oj_json_generator_error_class, "Invalid Unicode %s at %d", code, pos);
 }
 
-static void
-dump_array(VALUE a, VALUE clas, int depth, Out out) {
-    size_t	size;
-    int		i, cnt;
-    int		d2 = depth + 1;
-    long	id = check_circular(a, out);
-
-    if (id < 0) {
-	return;
-    }
-    if (Qundef != clas && rb_cArray != clas && ObjectMode == out->opts->mode) {
-	dump_obj_attrs(a, clas, 0, depth, out);
-	return;
-    }
-    cnt = (int)RARRAY_LEN(a);
-    *out->cur++ = '[';
-    if (0 < id) {
-	size = d2 * out->indent + 16;
-	if (out->end - out->cur <= (long)size) {
-	    grow(out, size);
-	}
-	fill_indent(out, d2);
-	*out->cur++ = '"';
-	*out->cur++ = '^';
-	*out->cur++ = 'i';
-	dump_ulong(id, out);
-	*out->cur++ = '"';
-    }
-    size = 2;
-    if (out->end - out->cur <= (long)size) {
-	grow(out, size);
-    }
-    if (0 == cnt) {
-	*out->cur++ = ']';
-    } else {
-	if (0 < id) {
-	    *out->cur++ = ',';
-	}
-	if (out->opts->dump_opts.use) {
-	    size = d2 * out->opts->dump_opts.indent_size + out->opts->dump_opts.array_size + 1;
-	} else {
-	    size = d2 * out->indent + 2;
-	}
-	cnt--;
-	for (i = 0; i <= cnt; i++) {
-	    if (out->end - out->cur <= (long)size) {
-		grow(out, size);
-	    }
-	    if (out->opts->dump_opts.use) {
-		if (0 < out->opts->dump_opts.array_size) {
-		    strcpy(out->cur, out->opts->dump_opts.array_nl);
-		    out->cur += out->opts->dump_opts.array_size;
-		}
-		if (0 < out->opts->dump_opts.indent_size) {
-		    int	i;
-		    for (i = d2; 0 < i; i--) {
-			strcpy(out->cur, out->opts->dump_opts.indent_str);
-			out->cur += out->opts->dump_opts.indent_size;
-		    }
-		}
-	    } else {
-		fill_indent(out, d2);
-	    }
-	    dump_val(rb_ary_entry(a, i), d2, out, 0, 0, true);
-	    if (i < cnt) {
-		*out->cur++ = ',';
-	    }
-	}
-	size = depth * out->indent + 1;
-	if (out->end - out->cur <= (long)size) {
-	    grow(out, size);
-	}
-	if (out->opts->dump_opts.use) {
-	    //printf("*** d2: %u  indent: %u '%s'\n", d2, out->opts->dump_opts->indent_size, out->opts->dump_opts->indent);
-	    if (0 < out->opts->dump_opts.array_size) {
-		strcpy(out->cur, out->opts->dump_opts.array_nl);
-		out->cur += out->opts->dump_opts.array_size;
-	    }
-	    if (0 < out->opts->dump_opts.indent_size) {
-		int	i;
-
-		for (i = depth; 0 < i; i--) {
-		    strcpy(out->cur, out->opts->dump_opts.indent_str);
-		    out->cur += out->opts->dump_opts.indent_size;
-		}
-	    }
-	} else {
-	    fill_indent(out, depth);
-	}
-	*out->cur++ = ']';
-    }
-    *out->cur = '\0';
-}
+static const char *dump_unicode(const char *str, const char *end, Out out, const char *orig) {
+    uint32_t code = 0;
+    uint8_t  b    = *(uint8_t *)str;
+    int      i, cnt;
 
-static int
-hash_cb_strict(VALUE key, VALUE value, Out out) {
-    int		depth = out->depth;
-    long	size;
-    int		rtype = rb_type(key);
-    
-    if (rtype != T_STRING && rtype != T_SYMBOL) {
-	rb_raise(rb_eTypeError, "In :strict mode all Hash keys must be Strings or Symbols, not %s.\n", rb_class2name(rb_obj_class(key)));
-    }
-    if (out->omit_nil && Qnil == value) {
-	return ST_CONTINUE;
-    }
-    if (!out->opts->dump_opts.use) {
-	size = depth * out->indent + 1;
-	if (out->end - out->cur <= size) {
-	    grow(out, size);
-	}
-	fill_indent(out, depth);
-	if (rtype == T_STRING) {
-	    dump_str_comp(key, out);
-	} else {
-	    dump_sym_comp(key, out);
-	}
-	*out->cur++ = ':';
+    if (0xC0 == (0xE0 & b)) {
+        cnt  = 1;
+        code = b & 0x0000001F;
+    } else if (0xE0 == (0xF0 & b)) {
+        cnt  = 2;
+        code = b & 0x0000000F;
+    } else if (0xF0 == (0xF8 & b)) {
+        cnt  = 3;
+        code = b & 0x00000007;
+    } else if (0xF8 == (0xFC & b)) {
+        cnt  = 4;
+        code = b & 0x00000003;
+    } else if (0xFC == (0xFE & b)) {
+        cnt  = 5;
+        code = b & 0x00000001;
     } else {
-	size = depth * out->opts->dump_opts.indent_size + out->opts->dump_opts.hash_size + 1;
-	if (out->end - out->cur <= size) {
-	    grow(out, size);
-	}
-	if (0 < out->opts->dump_opts.hash_size) {
-	    strcpy(out->cur, out->opts->dump_opts.hash_nl);
-	    out->cur += out->opts->dump_opts.hash_size;
-	}
-	if (0 < out->opts->dump_opts.indent_size) {
-	    int	i;
-	    for (i = depth; 0 < i; i--) {
-		strcpy(out->cur, out->opts->dump_opts.indent_str);
-		out->cur += out->opts->dump_opts.indent_size;
-	    }
-	}
-	if (rtype == T_STRING) {
-	    dump_str_comp(key, out);
-	} else {
-	    dump_sym_comp(key, out);
-	}
-	size = out->opts->dump_opts.before_size + out->opts->dump_opts.after_size + 2;
-	if (out->end - out->cur <= size) {
-	    grow(out, size);
-	}
-	if (0 < out->opts->dump_opts.before_size) {
-	    strcpy(out->cur, out->opts->dump_opts.before_sep);
-	    out->cur += out->opts->dump_opts.before_size;
-	}
-	*out->cur++ = ':';
-	if (0 < out->opts->dump_opts.after_size) {
-	    strcpy(out->cur, out->opts->dump_opts.after_sep);
-	    out->cur += out->opts->dump_opts.after_size;
-	}
-    }
-    dump_val(value, depth, out, 0, 0, false);
-    out->depth = depth;
-    *out->cur++ = ',';
-
-    return ST_CONTINUE;
-}
-
-static int
-hash_cb_compat(VALUE key, VALUE value, Out out) {
-    int		depth = out->depth;
-    long	size;
-
-    if (out->omit_nil && Qnil == value) {
-	return ST_CONTINUE;
+        cnt = 0;
+        raise_invalid_unicode(orig, (int)(end - orig), (int)(str - orig));
     }
-    if (!out->opts->dump_opts.use) {
-	size = depth * out->indent + 1;
-	if (out->end - out->cur <= size) {
-	    grow(out, size);
-	}
-	fill_indent(out, depth);
-    } else {
-	size = depth * out->opts->dump_opts.indent_size + out->opts->dump_opts.hash_size + 1;
-	if (out->end - out->cur <= size) {
-	    grow(out, size);
-	}
-	if (0 < out->opts->dump_opts.hash_size) {
-	    strcpy(out->cur, out->opts->dump_opts.hash_nl);
-	    out->cur += out->opts->dump_opts.hash_size;
-	}
-	if (0 < out->opts->dump_opts.indent_size) {
-	    int	i;
-	    for (i = depth; 0 < i; i--) {
-		strcpy(out->cur, out->opts->dump_opts.indent_str);
-		out->cur += out->opts->dump_opts.indent_size;
-	    }
-	}
-    }
-    switch (rb_type(key)) {
-    case T_STRING:
-	dump_str_comp(key, out);
-	break;
-    case T_SYMBOL:
-	dump_sym_comp(key, out);
-	break;
-    default:
-	/*rb_raise(rb_eTypeError, "In :compat mode all Hash keys must be Strings or Symbols, not %s.\n", rb_class2name(rb_obj_class(key)));*/
-	dump_str_comp(rb_funcall(key, oj_to_s_id, 0), out);
-	break;
+    str++;
+    for (; 0 < cnt; cnt--, str++) {
+        b = *(uint8_t *)str;
+        if (end <= str || 0x80 != (0xC0 & b)) {
+            raise_invalid_unicode(orig, (int)(end - orig), (int)(str - orig));
+        }
+        code = (code << 6) | (b & 0x0000003F);
     }
-    if (!out->opts->dump_opts.use) {
-	*out->cur++ = ':';
-    } else {
-	size = out->opts->dump_opts.before_size + out->opts->dump_opts.after_size + 2;
-	if (out->end - out->cur <= size) {
-	    grow(out, size);
-	}
-	if (0 < out->opts->dump_opts.before_size) {
-	    strcpy(out->cur, out->opts->dump_opts.before_sep);
-	    out->cur += out->opts->dump_opts.before_size;
-	}
-	*out->cur++ = ':';
-	if (0 < out->opts->dump_opts.after_size) {
-	    strcpy(out->cur, out->opts->dump_opts.after_sep);
-	    out->cur += out->opts->dump_opts.after_size;
-	}
-    }
-    dump_val(value, depth, out, 0, 0, true);
-    out->depth = depth;
-    *out->cur++ = ',';
-
-    return ST_CONTINUE;
-}
+    if (0x0000FFFF < code) {
+        uint32_t c1;
 
-static int
-hash_cb_object(VALUE key, VALUE value, Out out) {
-    int		depth = out->depth;
-    long	size = depth * out->indent + 1;
-
-    if (out->omit_nil && Qnil == value) {
-	return ST_CONTINUE;
-    }
-    if (out->end - out->cur <= size) {
-	grow(out, size);
-    }
-    fill_indent(out, depth);
-    if (rb_type(key) == T_STRING) {
-	dump_str_obj(key, Qundef, depth, out);
-	*out->cur++ = ':';
-	dump_val(value, depth, out, 0, 0, true);
-    } else if (rb_type(key) == T_SYMBOL) {
-	dump_sym_obj(key, out);
-	*out->cur++ = ':';
-	dump_val(value, depth, out, 0, 0, true);
-    } else {
-	int	d2 = depth + 1;
-	long	s2 = size + out->indent + 1;
-	int	i;
-	int	started = 0;
-	uint8_t	b;
-
-	if (out->end - out->cur <= s2 + 15) {
-	    grow(out, s2 + 15);
-	}
-	*out->cur++ = '"';
-	*out->cur++ = '^';
-	*out->cur++ = '#';
-	out->hash_cnt++;
-	for (i = 28; 0 <= i; i -= 4) {
-	    b = (uint8_t)((out->hash_cnt >> i) & 0x0000000F);
-	    if ('\0' != b) {
-		started = 1;
-	    }
-	    if (started) {
-		*out->cur++ = hex_chars[b];
-	    }
-	}
-	*out->cur++ = '"';
-	*out->cur++ = ':';
-	*out->cur++ = '[';
-	fill_indent(out, d2);
-	dump_val(key, d2, out, 0, 0, true);
-	if (out->end - out->cur <= (long)s2) {
-	    grow(out, s2);
-	}
-	*out->cur++ = ',';
-	fill_indent(out, d2);
-	dump_val(value, d2, out, 0, 0, true);
-	if (out->end - out->cur <= (long)size) {
-	    grow(out, size);
-	}
-	fill_indent(out, depth);
-	*out->cur++ = ']';
-    }
-    out->depth = depth;
-    *out->cur++ = ',';
-    
-    return ST_CONTINUE;
+        code -= 0x00010000;
+        c1   = ((code >> 10) & 0x000003FF) + 0x0000D800;
+        code = (code & 0x000003FF) + 0x0000DC00;
+        APPEND_CHARS(out->cur, "\\u", 2);
+        for (i = 3; 0 <= i; i--) {
+            *out->cur++ = hex_chars[(uint8_t)(c1 >> (i * 4)) & 0x0F];
+        }
+    }
+    APPEND_CHARS(out->cur, "\\u", 2);
+    for (i = 3; 0 <= i; i--) {
+        *out->cur++ = hex_chars[(uint8_t)(code >> (i * 4)) & 0x0F];
+    }
+    return str - 1;
 }
 
-static void
-dump_hash(VALUE obj, VALUE clas, int depth, int mode, Out out) {
-    int		cnt;
-    size_t	size;
+static const char *check_unicode(const char *str, const char *end, const char *orig) {
+    uint8_t b   = *(uint8_t *)str;
+    int     cnt = 0;
 
-    if (Qundef != clas && rb_cHash != clas && ObjectMode == mode) {
-	dump_obj_attrs(obj, clas, 0, depth, out);
-	return;
+    if (0xC0 == (0xE0 & b)) {
+        cnt = 1;
+    } else if (0xE0 == (0xF0 & b)) {
+        cnt = 2;
+    } else if (0xF0 == (0xF8 & b)) {
+        cnt = 3;
+    } else if (0xF8 == (0xFC & b)) {
+        cnt = 4;
+    } else if (0xFC == (0xFE & b)) {
+        cnt = 5;
+    } else {
+        raise_invalid_unicode(orig, (int)(end - orig), (int)(str - orig));
     }
-    cnt = (int)RHASH_SIZE(obj);
-    size = depth * out->indent + 2;
-    if (out->end - out->cur <= 2) {
-	grow(out, 2);
+    str++;
+    for (; 0 < cnt; cnt--, str++) {
+        b = *(uint8_t *)str;
+        if (end <= str || 0x80 != (0xC0 & b)) {
+            raise_invalid_unicode(orig, (int)(end - orig), (int)(str - orig));
+        }
+    }
+    return str;
+}
+
+// Returns 0 if not using circular references, -1 if no further writing is
+// needed (duplicate), and a positive value if the object was added to the
+// cache.
+long oj_check_circular(VALUE obj, Out out) {
+    slot_t  id = 0;
+    slot_t *slot;
+
+    if (Yes == out->opts->circular) {
+        if (0 == (id = oj_cache8_get(out->circ_cache, obj, &slot))) {
+            out->circ_cnt++;
+            id    = out->circ_cnt;
+            *slot = id;
+        } else {
+            if (ObjectMode == out->opts->mode) {
+                assure_size(out, 18);
+                APPEND_CHARS(out->cur, "\"^r", 3);
+                dump_ulong(id, out);
+                *out->cur++ = '"';
+            }
+            return -1;
+        }
     }
-    if (0 == cnt) {
-	*out->cur++ = '{';
-	*out->cur++ = '}';
+    return (long)id;
+}
+
+void oj_dump_time(VALUE obj, Out out, int withZone) {
+    char      buf[64];
+    char     *b = buf + sizeof(buf) - 1;
+    long      size;
+    char     *dot;
+    int       neg = 0;
+    long      one = 1000000000;
+    long long sec;
+    long long nsec;
+
+    // rb_time_timespec as well as rb_time_timeeval have a bug that causes an
+    // exception to be raised if a time is before 1970 on 32 bit systems so
+    // check the timespec size and use the ruby calls if a 32 bit system.
+    if (16 <= sizeof(struct timespec)) {
+        struct timespec ts = rb_time_timespec(obj);
+
+        sec  = (long long)ts.tv_sec;
+        nsec = ts.tv_nsec;
     } else {
-	long	id = check_circular(obj, out);
-
-	if (0 > id) {
-	    return;
-	}
-	*out->cur++ = '{';
-	if (0 < id) {
-	    if (out->end - out->cur <= (long)size + 16) {
-		grow(out, size + 16);
-	    }
-	    fill_indent(out, depth + 1);
-	    *out->cur++ = '"';
-	    *out->cur++ = '^';
-	    *out->cur++ = 'i';
-	    *out->cur++ = '"';
-	    *out->cur++ = ':';
-	    dump_ulong(id, out);
-	    *out->cur++ = ',';
-	}
-	out->depth = depth + 1;
-	if (ObjectMode == mode) {
-	    rb_hash_foreach(obj, hash_cb_object, (VALUE)out);
-	} else if (CompatMode == mode) {
-	    rb_hash_foreach(obj, hash_cb_compat, (VALUE)out);
-	} else {
-	    rb_hash_foreach(obj, hash_cb_strict, (VALUE)out);
-	}
-	if (',' == *(out->cur - 1)) {
-	    out->cur--; // backup to overwrite last comma
-	}
-	if (!out->opts->dump_opts.use) {
-	    if (out->end - out->cur <= (long)size) {
-		grow(out, size);
-	    }
-	    fill_indent(out, depth);
-	} else {
-	    size = depth * out->opts->dump_opts.indent_size + out->opts->dump_opts.hash_size + 1;
-	    if (out->end - out->cur <= (long)size) {
-		grow(out, size);
-	    }
-	    if (0 < out->opts->dump_opts.hash_size) {
-		strcpy(out->cur, out->opts->dump_opts.hash_nl);
-		out->cur += out->opts->dump_opts.hash_size;
-	    }
-	    if (0 < out->opts->dump_opts.indent_size) {
-		int	i;
-
-		for (i = depth; 0 < i; i--) {
-		    strcpy(out->cur, out->opts->dump_opts.indent_str);
-		    out->cur += out->opts->dump_opts.indent_size;
-		}
-	    }
-	}
-	*out->cur++ = '}';
+        sec  = NUM2LL(rb_funcall2(obj, oj_tv_sec_id, 0, 0));
+        nsec = NUM2LL(rb_funcall2(obj, oj_tv_nsec_id, 0, 0));
     }
-    *out->cur = '\0';
-}
 
-static void
-dump_time(VALUE obj, Out out, int withZone) {
-    char		buf[64];
-    char		*b = buf + sizeof(buf) - 1;
-    long		size;
-    char		*dot;
-    int			neg = 0;
-    long		one = 1000000000;
-#if HAS_RB_TIME_TIMESPEC
-    struct timespec	ts = rb_time_timespec(obj);
-    time_t		sec = ts.tv_sec;
-    long		nsec = ts.tv_nsec;
-#else
-    time_t		sec = NUM2LONG(rb_funcall2(obj, oj_tv_sec_id, 0, 0));
-#if HAS_NANO_TIME
-    long long		nsec = rb_num2ll(rb_funcall2(obj, oj_tv_nsec_id, 0, 0));
-#else
-    long long		nsec = rb_num2ll(rb_funcall2(obj, oj_tv_usec_id, 0, 0)) * 1000;
-#endif
-#endif
-    
     *b-- = '\0';
     if (withZone) {
-	long	tzsecs = NUM2LONG(rb_funcall2(obj, oj_utc_offset_id, 0, 0));
-	int	zneg = (0 > tzsecs);
-
-	if (0 == tzsecs && rb_funcall2(obj, oj_utcq_id, 0, 0)) {
-	    tzsecs = 86400;
-	}
-	if (zneg) {
-	    tzsecs = -tzsecs;
-	}
-	if (0 == tzsecs) {
-	    *b-- = '0';
-	} else {
-	    for (; 0 < tzsecs; b--, tzsecs /= 10) {
-		*b = '0' + (tzsecs % 10);
-	    }
-	    if (zneg) {
-		*b-- = '-';
-	    }
-	}
-	*b-- = 'e';
+        long tzsecs = NUM2LONG(rb_funcall2(obj, oj_utc_offset_id, 0, 0));
+        int  zneg   = (0 > tzsecs);
+
+        if (0 == tzsecs && rb_funcall2(obj, oj_utcq_id, 0, 0)) {
+            tzsecs = 86400;
+        }
+        if (zneg) {
+            tzsecs = -tzsecs;
+        }
+        if (0 == tzsecs) {
+            *b-- = '0';
+        } else {
+            for (; 0 < tzsecs; b--, tzsecs /= 10) {
+                *b = '0' + (tzsecs % 10);
+            }
+            if (zneg) {
+                *b-- = '-';
+            }
+        }
+        *b-- = 'e';
     }
     if (0 > sec) {
-	neg = 1;
-	sec = -sec;
-	if (0 < nsec) {
-	    nsec = 1000000000 - nsec;
-	    sec--;
-	}
+        neg = 1;
+        sec = -sec;
+        if (0 < nsec) {
+            nsec = 1000000000 - nsec;
+            sec--;
+        }
     }
     dot = b - 9;
     if (0 < out->opts->sec_prec) {
-	if (9 > out->opts->sec_prec) {
-	    int	i;
-
-	    for (i = 9 - out->opts->sec_prec; 0 < i; i--) {
-		dot++;
-		nsec = (nsec + 5) / 10;
-		one /= 10;
-	    }
-	}
-	if (one <= nsec) {
-	    nsec -= one;
-	    sec++;
-	}
-	for (; dot < b; b--, nsec /= 10) {
-	    *b = '0' + (nsec % 10);
-	}
-	*b-- = '.';
+        if (9 > out->opts->sec_prec) {
+            int i;
+
+            for (i = 9 - out->opts->sec_prec; 0 < i; i--) {
+                dot++;
+                nsec = (nsec + 5) / 10;
+                one /= 10;
+            }
+        }
+        if (one <= nsec) {
+            nsec -= one;
+            sec++;
+        }
+        for (; dot < b; b--, nsec /= 10) {
+            *b = '0' + (nsec % 10);
+        }
+        *b-- = '.';
     }
     if (0 == sec) {
-	*b-- = '0';
+        *b-- = '0';
     } else {
-	for (; 0 < sec; b--, sec /= 10) {
-	    *b = '0' + (sec % 10);
-	}
+        for (; 0 < sec; b--, sec /= 10) {
+            *b = '0' + (sec % 10);
+        }
     }
     if (neg) {
-	*b-- = '-';
+        *b-- = '-';
     }
     b++;
     size = sizeof(buf) - (b - buf) - 1;
-    if (out->end - out->cur <= size) {
-	grow(out, size);
-    }
-    memcpy(out->cur, b, size);
-    out->cur += size;
+    assure_size(out, size);
+    APPEND_CHARS(out->cur, b, size);
     *out->cur = '\0';
 }
 
-static void
-dump_ruby_time(VALUE obj, Out out) {
-    volatile VALUE	rstr = rb_funcall(obj, oj_to_s_id, 0);
+void oj_dump_ruby_time(VALUE obj, Out out) {
+    volatile VALUE rstr = oj_safe_string_convert(obj);
 
-    dump_cstr(rb_string_value_ptr((VALUE*)&rstr), (int)RSTRING_LEN(rstr), 0, 0, out);
+    oj_dump_cstr(RSTRING_PTR(rstr), RSTRING_LEN(rstr), 0, 0, out);
 }
 
-static void
-dump_xml_time(VALUE obj, Out out) {
-    char		buf[64];
-    struct tm		*tm;
-    long		one = 1000000000;
-#if HAS_RB_TIME_TIMESPEC
-    struct timespec	ts = rb_time_timespec(obj);
-    time_t		sec = ts.tv_sec;
-    long		nsec = ts.tv_nsec;
-#else
-    time_t		sec = NUM2LONG(rb_funcall2(obj, oj_tv_sec_id, 0, 0));
-#if HAS_NANO_TIME
-    long long		nsec = rb_num2ll(rb_funcall2(obj, oj_tv_nsec_id, 0, 0));
-#else
-    long long		nsec = rb_num2ll(rb_funcall2(obj, oj_tv_usec_id, 0, 0)) * 1000;
-#endif
-#endif
-    long		tzsecs = NUM2LONG(rb_funcall2(obj, oj_utc_offset_id, 0, 0));
-    int			tzhour, tzmin;
-    char		tzsign = '+';
+void oj_dump_xml_time(VALUE obj, Out out) {
+    char             buf[64];
+    struct _timeInfo ti;
+    long             one = 1000000000;
+    int64_t          sec;
+    long long        nsec;
+    long             tzsecs = NUM2LONG(rb_funcall2(obj, oj_utc_offset_id, 0, 0));
+    int              tzhour, tzmin;
+    char             tzsign = '+';
 
-    if (out->end - out->cur <= 36) {
-	grow(out, 36);
+    if (16 <= sizeof(struct timespec)) {
+        struct timespec ts = rb_time_timespec(obj);
+
+        sec  = ts.tv_sec;
+        nsec = ts.tv_nsec;
+    } else {
+        sec  = NUM2LL(rb_funcall2(obj, oj_tv_sec_id, 0, 0));
+        nsec = NUM2LL(rb_funcall2(obj, oj_tv_nsec_id, 0, 0));
     }
-    if (9 > out->opts->sec_prec) {
-	int	i;
 
-	for (i = 9 - out->opts->sec_prec; 0 < i; i--) {
-	    nsec = (nsec + 5) / 10;
-	    one /= 10;
-	}
-	if (one <= nsec) {
-	    nsec -= one;
-	    sec++;
-	}
+    assure_size(out, 36);
+    if (9 > out->opts->sec_prec) {
+        int i;
+
+        // This is pretty lame but to be compatible with rails and active
+        // support rounding is not done but instead a floor is done when
+        // second precision is 3 just to be like rails. sigh.
+        if (3 == out->opts->sec_prec) {
+            nsec /= 1000000;
+            one = 1000;
+        } else {
+            for (i = 9 - out->opts->sec_prec; 0 < i; i--) {
+                nsec = (nsec + 5) / 10;
+                one /= 10;
+            }
+            if (one <= nsec) {
+                nsec -= one;
+                sec++;
+            }
+        }
     }
     // 2012-01-05T23:58:07.123456000+09:00
-    //tm = localtime(&sec);
+    // tm = localtime(&sec);
     sec += tzsecs;
-    tm = gmtime(&sec);
-#if 1
+    sec_as_time((int64_t)sec, &ti);
     if (0 > tzsecs) {
         tzsign = '-';
         tzhour = (int)(tzsecs / -3600);
-        tzmin = (int)(tzsecs / -60) - (tzhour * 60);
+        tzmin  = (int)(tzsecs / -60) - (tzhour * 60);
     } else {
         tzhour = (int)(tzsecs / 3600);
-        tzmin = (int)(tzsecs / 60) - (tzhour * 60);
-    }
-#else
-    if (0 > tm->tm_gmtoff) {
-        tzsign = '-';
-        tzhour = (int)(tm->tm_gmtoff / -3600);
-        tzmin = (int)(tm->tm_gmtoff / -60) - (tzhour * 60);
-    } else {
-        tzhour = (int)(tm->tm_gmtoff / 3600);
-        tzmin = (int)(tm->tm_gmtoff / 60) - (tzhour * 60);
-    }
-#endif
-    if (0 == nsec || 0 == out->opts->sec_prec) {
-	if (0 == tzsecs && rb_funcall2(obj, oj_utcq_id, 0, 0)) {
-	    sprintf(buf, "%04d-%02d-%02dT%02d:%02d:%02dZ",
-		    tm->tm_year + 1900, tm->tm_mon + 1, tm->tm_mday,
-		    tm->tm_hour, tm->tm_min, tm->tm_sec);
-	    dump_cstr(buf, 20, 0, 0, out);
-	} else {
-	    sprintf(buf, "%04d-%02d-%02dT%02d:%02d:%02d%c%02d:%02d",
-		    tm->tm_year + 1900, tm->tm_mon + 1, tm->tm_mday,
-		    tm->tm_hour, tm->tm_min, tm->tm_sec,
-		    tzsign, tzhour, tzmin);
-	    dump_cstr(buf, 25, 0, 0, out);
-	}
+        tzmin  = (int)(tzsecs / 60) - (tzhour * 60);
+    }
+    if ((0 == nsec && !out->opts->sec_prec_set) || 0 == out->opts->sec_prec) {
+        if (0 == tzsecs && rb_funcall2(obj, oj_utcq_id, 0, 0)) {
+            int len = sprintf(buf, "%04d-%02d-%02dT%02d:%02d:%02dZ", ti.year, ti.mon, ti.day, ti.hour, ti.min, ti.sec);
+            oj_dump_cstr(buf, len, 0, 0, out);
+        } else {
+            int len = sprintf(buf,
+                              "%04d-%02d-%02dT%02d:%02d:%02d%c%02d:%02d",
+                              ti.year,
+                              ti.mon,
+                              ti.day,
+                              ti.hour,
+                              ti.min,
+                              ti.sec,
+                              tzsign,
+                              tzhour,
+                              tzmin);
+            oj_dump_cstr(buf, len, 0, 0, out);
+        }
     } else if (0 == tzsecs && rb_funcall2(obj, oj_utcq_id, 0, 0)) {
-	char	format[64] = "%04d-%02d-%02dT%02d:%02d:%02d.%09ldZ";
-	int	len = 30;
-
-	if (9 > out->opts->sec_prec) {
-	    format[32] = '0' + out->opts->sec_prec;
-	    len -= 9 - out->opts->sec_prec;
-	}
-	sprintf(buf, format,
-		tm->tm_year + 1900, tm->tm_mon + 1, tm->tm_mday,
-		tm->tm_hour, tm->tm_min, tm->tm_sec, nsec);
-	dump_cstr(buf, len, 0, 0, out);
-    } else {
-	char	format[64] = "%04d-%02d-%02dT%02d:%02d:%02d.%09ld%c%02d:%02d";
-	int	len = 35;
-
-	if (9 > out->opts->sec_prec) {
-	    format[32] = '0' + out->opts->sec_prec;
-	    len -= 9 - out->opts->sec_prec;
-	}
-	sprintf(buf, format,
-		tm->tm_year + 1900, tm->tm_mon + 1, tm->tm_mday,
-		tm->tm_hour, tm->tm_min, tm->tm_sec, nsec,
-		tzsign, tzhour, tzmin);
-	dump_cstr(buf, len, 0, 0, out);
-    }
-}
-
-static void
-dump_data_strict(VALUE obj, Out out) {
-    VALUE	clas = rb_obj_class(obj);
-
-    if (oj_bigdecimal_class == clas) {
-	volatile VALUE	rstr = rb_funcall(obj, oj_to_s_id, 0);
-
-	dump_raw(rb_string_value_ptr((VALUE*)&rstr), (int)RSTRING_LEN(rstr), out);
-    } else {
-	raise_strict(obj);
-    }
-}
-
-static void
-dump_data_null(VALUE obj, Out out) {
-    VALUE	clas = rb_obj_class(obj);
-
-    if (oj_bigdecimal_class == clas) {
-	volatile VALUE	rstr = rb_funcall(obj, oj_to_s_id, 0);
-
-	dump_raw(rb_string_value_ptr((VALUE*)&rstr), (int)RSTRING_LEN(rstr), out);
-    } else {
-	dump_nil(out);
-    }
-}
-
-static void
-dump_data_comp(VALUE obj, int depth, Out out, int argc, VALUE *argv, bool as_ok) {
-    VALUE	clas = rb_obj_class(obj);
-
-    if (as_ok && Yes == out->opts->to_json && rb_respond_to(obj, oj_to_hash_id)) {
-	volatile VALUE	h = rb_funcall(obj, oj_to_hash_id, 0);
- 
-	if (T_HASH != rb_type(h)) {
-	    // It seems that ActiveRecord implemented to_hash so that it returns
-	    // an Array and not a Hash. To get around that any value returned
-	    // will be dumped.
-
-	    //rb_raise(rb_eTypeError, "%s.to_hash() did not return a Hash.\n", rb_class2name(rb_obj_class(obj)));
-	    dump_val(h, depth, out, 0, 0, false);
-	}
-	dump_hash(h, Qundef, depth, out->opts->mode, out);
-    } else if (Yes == out->opts->bigdec_as_num && oj_bigdecimal_class == clas) {
-	volatile VALUE	rstr = rb_funcall(obj, oj_to_s_id, 0);
-
-	dump_raw(rb_string_value_ptr((VALUE*)&rstr), (int)RSTRING_LEN(rstr), out);
-    } else if (as_ok && Yes == out->opts->as_json && rb_respond_to(obj, oj_as_json_id)) {
-	volatile VALUE	aj;
-
-#if HAS_METHOD_ARITY
-	// Some classes elect to not take an options argument so check the arity
-	// of as_json.
-	switch (rb_obj_method_arity(obj, oj_as_json_id)) {
-	case 0:
-	    aj = rb_funcall2(obj, oj_as_json_id, 0, 0);
-	    break;
-	case 1:
-	    if (1 <= argc) {
-		aj = rb_funcall2(obj, oj_as_json_id, 1, argv);
-	    } else {
-		VALUE	nothing [1];
-
-		nothing[0] = Qnil;
-		aj = rb_funcall2(obj, oj_as_json_id, 1, nothing);
-	    }
-	    break;
-	default:
-	    aj = rb_funcall2(obj, oj_as_json_id, argc, argv);
-	    break;
-	}
-#else
-	aj = rb_funcall2(obj, oj_as_json_id, argc, argv);
-#endif
-	// Catch the obvious brain damaged recursive dumping.
-	if (aj == obj) {
-	    volatile VALUE	rstr = rb_funcall(obj, oj_to_s_id, 0);
-
-	    dump_cstr(rb_string_value_ptr((VALUE*)&rstr), (int)RSTRING_LEN(rstr), 0, 0, out);
-	} else {
-	    dump_val(aj, depth, out, 0, 0, false);
-	}
-    } else if (Yes == out->opts->to_json && rb_respond_to(obj, oj_to_json_id)) {
-	volatile VALUE	rs;
-	const char	*s;
-	int		len;
-
-	rs = rb_funcall(obj, oj_to_json_id, 0);
-	s = rb_string_value_ptr((VALUE*)&rs);
-	len = (int)RSTRING_LEN(rs);
-
-	if (out->end - out->cur <= len + 1) {
-	    grow(out, len);
-	}
-	memcpy(out->cur, s, len);
-	out->cur += len;
-	*out->cur = '\0';
-    } else {
-	if (rb_cTime == clas) {
-	    switch (out->opts->time_format) {
-	    case RubyTime:	dump_ruby_time(obj, out);	break;
-	    case XmlTime:	dump_xml_time(obj, out);	break;
-	    case UnixZTime:	dump_time(obj, out, 1);		break;
-	    case UnixTime:
-	    default:		dump_time(obj, out, 0);		break;
-	    }
-	} else if (oj_bigdecimal_class == clas) {
-	    volatile VALUE	rstr = rb_funcall(obj, oj_to_s_id, 0);
-
-	    dump_cstr(rb_string_value_ptr((VALUE*)&rstr), (int)RSTRING_LEN(rstr), 0, 0, out);
-	} else {
-	    volatile VALUE	rstr = rb_funcall(obj, oj_to_s_id, 0);
-
-	    dump_cstr(rb_string_value_ptr((VALUE*)&rstr), (int)RSTRING_LEN(rstr), 0, 0, out);
-	}
-    }
-}
-
-static void
-dump_data_obj(VALUE obj, int depth, Out out) {
-    VALUE	clas = rb_obj_class(obj);
-
-    if (rb_cTime == clas) {
-	if (out->end - out->cur <= 6) {
-	    grow(out, 6);
-	}
-	*out->cur++ = '{';
-	*out->cur++ = '"';
-	*out->cur++ = '^';
-	*out->cur++ = 't';
-	*out->cur++ = '"';
-	*out->cur++ = ':';
-	switch (out->opts->time_format) {
-	case RubyTime: // Does not output fractional seconds
-	case XmlTime:
-	    dump_xml_time(obj, out);
-	    break;
-	case UnixZTime:
-	    dump_time(obj, out, 1);
-	    break;
-	case UnixTime:
-	default:
-	    dump_time(obj, out, 0);
-	    break;
-	}
-	*out->cur++ = '}';
-	*out->cur = '\0';
-    } else {
-	if (oj_bigdecimal_class == clas) {
-	    volatile VALUE	rstr = rb_funcall(obj, oj_to_s_id, 0);
-		
-	    if (Yes == out->opts->bigdec_as_num) {
-		dump_raw(rb_string_value_ptr((VALUE*)&rstr), (int)RSTRING_LEN(rstr), out);
-	    } else {
-		dump_cstr(rb_string_value_ptr((VALUE*)&rstr), (int)RSTRING_LEN(rstr), 0, 0, out);
-	    }
-	} else {
-	    dump_nil(out);
-	}
-    }
-}
-
-static void
-dump_obj_comp(VALUE obj, int depth, Out out, int argc, VALUE *argv, bool as_ok) {
-    if (as_ok && Yes == out->opts->to_json && rb_respond_to(obj, oj_to_hash_id)) {
-	volatile VALUE	h = rb_funcall(obj, oj_to_hash_id, 0);
-
-	if (T_HASH != rb_type(h)) {
-	    // It seems that ActiveRecord implemented to_hash so that it returns
-	    // an Array and not a Hash. To get around that any value returned
-	    // will be dumped.
-
-	    //rb_raise(rb_eTypeError, "%s.to_hash() did not return a Hash.\n", rb_class2name(rb_obj_class(obj)));
-	    dump_val(h, depth, out, 0, 0, false);
-	} else {
-	    dump_hash(h, Qundef, depth, out->opts->mode, out);
-	}
-    } else if (as_ok && Yes == out->opts->as_json && rb_respond_to(obj, oj_as_json_id)) {
-	volatile VALUE	aj;
-
-#if HAS_METHOD_ARITY
-	// Some classes elect to not take an options argument so check the arity
-	// of as_json.
-	switch (rb_obj_method_arity(obj, oj_as_json_id)) {
-	case 0:
-	    aj = rb_funcall2(obj, oj_as_json_id, 0, 0);
-	    break;
-	case 1:
-	    if (1 <= argc) {
-		aj = rb_funcall2(obj, oj_as_json_id, 1, argv);
-	    } else {
-		VALUE	nothing [1];
-
-		nothing[0] = Qnil;
-		aj = rb_funcall2(obj, oj_as_json_id, 1, nothing);
-	    }
-	    break;
-	default:
-	    aj = rb_funcall2(obj, oj_as_json_id, argc, argv);
-	    break;
-	}
-#else
-	aj = rb_funcall2(obj, oj_as_json_id, argc, argv);
-#endif
-	// Catch the obvious brain damaged recursive dumping.
-	if (aj == obj) {
-	    volatile VALUE	rstr = rb_funcall(obj, oj_to_s_id, 0);
-
-	    dump_cstr(rb_string_value_ptr((VALUE*)&rstr), (int)RSTRING_LEN(rstr), 0, 0, out);
-	} else {
-	    dump_val(aj, depth, out, 0, 0, false);
-	}
-    } else if (Yes == out->opts->to_json && rb_respond_to(obj, oj_to_json_id)) {
-	volatile VALUE	rs;
-	const char	*s;
-	int		len;
-
-	rs = rb_funcall(obj, oj_to_json_id, 0);
-	s = rb_string_value_ptr((VALUE*)&rs);
-	len = (int)RSTRING_LEN(rs);
-
-	if (out->end - out->cur <= len + 1) {
-	    grow(out, len);
-	}
-	memcpy(out->cur, s, len);
-	out->cur += len;
-	*out->cur = '\0';
-    } else {
-	VALUE	clas = rb_obj_class(obj);
-
-	if (oj_bigdecimal_class == clas) {
-	    volatile VALUE	rstr = rb_funcall(obj, oj_to_s_id, 0);
-
-	    if (Yes == out->opts->bigdec_as_num) {
-		dump_raw(rb_string_value_ptr((VALUE*)&rstr), (int)RSTRING_LEN(rstr), out);
-	    } else {
-		dump_cstr(rb_string_value_ptr((VALUE*)&rstr), (int)RSTRING_LEN(rstr), 0, 0, out);
-	    }
-#if (defined T_RATIONAL && defined RRATIONAL)
-	} else if (oj_datetime_class == clas || oj_date_class == clas || rb_cRational == clas) {
-#else
-	} else if (oj_datetime_class == clas || oj_date_class == clas) {
-#endif
-	    volatile VALUE	rstr = rb_funcall(obj, oj_to_s_id, 0);
-
-	    dump_cstr(rb_string_value_ptr((VALUE*)&rstr), (int)RSTRING_LEN(rstr), 0, 0, out);
-	} else {
-	    dump_obj_attrs(obj, Qundef, 0, depth, out);
-	}
-    }
-    *out->cur = '\0';
-}
-
-inline static void
-dump_obj_obj(VALUE obj, int depth, Out out) {
-    long	id = check_circular(obj, out);
-
-    if (0 <= id) {
-	VALUE	clas = rb_obj_class(obj);
-
-	if (oj_bigdecimal_class == clas) {
-	    volatile VALUE	rstr = rb_funcall(obj, oj_to_s_id, 0);
-
-	    dump_raw(rb_string_value_ptr((VALUE*)&rstr), (int)RSTRING_LEN(rstr), out);
-	} else {
-	    dump_obj_attrs(obj, clas, id, depth, out);
-	}
-    }
-}
-
-#ifdef RUBINIUS_RUBY
-static int
-isRbxHashAttr(const char *attr) {
-    const char	*hashAttrs[] = {
-	"@capacity",
-	"@max_entries",
-	"@state",
-	"@mask",
-	"@size",
-	"@entries",
-	"@default_proc",
-	0 };
-    const char	**ap;
-
-    for (ap = hashAttrs; 0 != *ap; ap++) {
-	if (0 == strcmp(attr, *ap)) {
-	    return 1;
-	}
-    }
-    return 0;
-}
-#endif
-
-#if HAS_IVAR_HELPERS
-static int
-dump_attr_cb(ID key, VALUE value, Out out) {
-    int		depth = out->depth;
-    size_t	size = depth * out->indent + 1;
-    const char	*attr = rb_id2name(key);
-
-    if (out->omit_nil && Qnil == value) {
-	return ST_CONTINUE;
-    }
-#if HAS_EXCEPTION_MAGIC
-    if (0 == strcmp("bt", attr) || 0 == strcmp("mesg", attr)) {
-	return ST_CONTINUE;
-    }
-#endif
-    if (out->end - out->cur <= (long)size) {
-	grow(out, size);
-    }
-    fill_indent(out, depth);
-    if ('@' == *attr) {
-	attr++;
-	dump_cstr(attr, strlen(attr), 0, 0, out);
-    } else {
-	char	buf[32];
-
-	*buf = '~';
-	strncpy(buf + 1, attr, sizeof(buf) - 2);
-	buf[sizeof(buf) - 1] = '\0';
-	dump_cstr(buf, strlen(buf), 0, 0, out);
-    }
-    *out->cur++ = ':';
-    dump_val(value, depth, out, 0, 0, true);
-    out->depth = depth;
-    *out->cur++ = ',';
-    
-    return ST_CONTINUE;
-}
-#endif
-
-static void
-dump_obj_attrs(VALUE obj, VALUE clas, slot_t id, int depth, Out out) {
-    size_t	size = 0;
-    int		d2 = depth + 1;
-    int		type = rb_type(obj);
-
-    if (out->end - out->cur <= 2) {
-	grow(out, 2);
-    }
-    *out->cur++ = '{';
-    if (Qundef != clas) {
-	const char	*class_name = rb_class2name(clas);
-	int		clen = (int)strlen(class_name);
-
-	size = d2 * out->indent + clen + 10;
-	if (out->end - out->cur <= (long)size) {
-	    grow(out, size);
-	}
-	fill_indent(out, d2);
-	*out->cur++ = '"';
-	*out->cur++ = '^';
-	*out->cur++ = 'o';
-	*out->cur++ = '"';
-	*out->cur++ = ':';
-	dump_cstr(class_name, clen, 0, 0, out);
-    }
-    if (0 < id) {
-	size = d2 * out->indent + 16;
-	if (out->end - out->cur <= (long)size) {
-	    grow(out, size);
-	}
-	*out->cur++ = ',';
-	fill_indent(out, d2);
-	*out->cur++ = '"';
-	*out->cur++ = '^';
-	*out->cur++ = 'i';
-	*out->cur++ = '"';
-	*out->cur++ = ':';
-	dump_ulong(id, out);
-    }
-    switch (type) {
-    case T_STRING:
-	size = d2 * out->indent + 14;
-	if (out->end - out->cur <= (long)size) {
-	    grow(out, size);
-	}
-	*out->cur++ = ',';
-	fill_indent(out, d2);
-	*out->cur++ = '"';
-	*out->cur++ = 's';
-	*out->cur++ = 'e';
-	*out->cur++ = 'l';
-	*out->cur++ = 'f';
-	*out->cur++ = '"';
-	*out->cur++ = ':';
-	dump_cstr(rb_string_value_ptr((VALUE*)&obj), (int)RSTRING_LEN(obj), 0, 0, out);
-	break;
-    case T_ARRAY:
-	size = d2 * out->indent + 14;
-	if (out->end - out->cur <= (long)size) {
-	    grow(out, size);
-	}
-	*out->cur++ = ',';
-	fill_indent(out, d2);
-	*out->cur++ = '"';
-	*out->cur++ = 's';
-	*out->cur++ = 'e';
-	*out->cur++ = 'l';
-	*out->cur++ = 'f';
-	*out->cur++ = '"';
-	*out->cur++ = ':';
-	dump_array(obj, Qundef, depth + 1, out);
-	break;
-    case T_HASH:
-	size = d2 * out->indent + 14;
-	if (out->end - out->cur <= (long)size) {
-	    grow(out, size);
-	}
-	*out->cur++ = ',';
-	fill_indent(out, d2);
-	*out->cur++ = '"';
-	*out->cur++ = 's';
-	*out->cur++ = 'e';
-	*out->cur++ = 'l';
-	*out->cur++ = 'f';
-	*out->cur++ = '"';
-	*out->cur++ = ':';
-	dump_hash(obj, Qundef, depth + 1, out->opts->mode, out);
-	break;
-    default:
-	break;
-    }
-    {
-	int	cnt;
-#if HAS_IVAR_HELPERS
-	cnt = (int)rb_ivar_count(obj);
-#else
-	volatile VALUE	vars = rb_funcall2(obj, oj_instance_variables_id, 0, 0);
-	VALUE		*np = RARRAY_PTR(vars);
-	ID		vid;
-	const char	*attr;
-	int		i;
-	int		first = 1;
-
-	cnt = (int)RARRAY_LEN(vars);
-#endif
-	if (Qundef != clas && 0 < cnt) {
-	    *out->cur++ = ',';
-	}
-	out->depth = depth + 1;
-#if HAS_IVAR_HELPERS
-	rb_ivar_foreach(obj, dump_attr_cb, (VALUE)out);
-	if (',' == *(out->cur - 1)) {
-	    out->cur--; // backup to overwrite last comma
-	}
-#else
-	size = d2 * out->indent + 1;
-	for (i = cnt; 0 < i; i--, np++) {
-	    VALUE	value;
-	    
-	    vid = rb_to_id(*np);
-	    attr = rb_id2name(vid);
-#ifdef RUBINIUS_RUBY
-	    if (T_HASH == type && isRbxHashAttr(attr)) {
-		continue;
-	    }
-#endif
-	    value = rb_ivar_get(obj, vid);
-	    if (out->omit_nil && Qnil == value) {
-		continue;
-	    }
-	    if (first) {
-		first = 0;
-	    } else {
-		*out->cur++ = ',';
-	    }
-	    if (out->end - out->cur <= (long)size) {
-		grow(out, size);
-	    }
-	    fill_indent(out, d2);
-	    if ('@' == *attr) {
-		attr++;
-		dump_cstr(attr, strlen(attr), 0, 0, out);
-	    } else {
-		char	buf[32];
-
-		*buf = '~';
-		strncpy(buf + 1, attr, sizeof(buf) - 2);
-		buf[sizeof(buf) - 1] = '\0';
-		dump_cstr(buf, strlen(attr) + 1, 0, 0, out);
-	    }
-	    *out->cur++ = ':';
-	    dump_val(value, d2, out, 0, 0, true);
-	    if (out->end - out->cur <= 2) {
-		grow(out, 2);
-	    }
-	}
-#endif
-#if HAS_EXCEPTION_MAGIC
-	if (rb_obj_is_kind_of(obj, rb_eException)) {
-	    volatile VALUE	rv;
-
-	    if (',' != *(out->cur - 1)) {
-		*out->cur++ = ',';
-	    }
-	    // message
-	    if (out->end - out->cur <= (long)size) {
-		grow(out, size);
-	    }
-	    fill_indent(out, d2);
-	    dump_cstr("~mesg", 5, 0, 0, out);
-	    *out->cur++ = ':';
-	    rv = rb_funcall2(obj, rb_intern("message"), 0, 0);
-	    dump_val(rv, d2, out, 0, 0, true);
-	    if (out->end - out->cur <= 2) {
-		grow(out, 2);
-	    }
-	    *out->cur++ = ',';
-	    // backtrace
-	    if (out->end - out->cur <= (long)size) {
-		grow(out, size);
-	    }
-	    fill_indent(out, d2);
-	    dump_cstr("~bt", 3, 0, 0, out);
-	    *out->cur++ = ':';
-	    rv = rb_funcall2(obj, rb_intern("backtrace"), 0, 0);
-	    dump_val(rv, d2, out, 0, 0, true);
-	    if (out->end - out->cur <= 2) {
-		grow(out, 2);
-	    }
-	}
-#endif
-	out->depth = depth;
-    }
-    fill_indent(out, depth);
-    *out->cur++ = '}';
-    *out->cur = '\0';
-}
-
-static void
-dump_struct_comp(VALUE obj, int depth, Out out, int argc, VALUE *argv, bool as_ok) {
-    if (as_ok && Yes == out->opts->to_json && rb_respond_to(obj, oj_to_hash_id)) {
-	volatile VALUE	h = rb_funcall(obj, oj_to_hash_id, 0);
- 
-	if (T_HASH != rb_type(h)) {
-	    // It seems that ActiveRecord implemented to_hash so that it returns
-	    // an Array and not a Hash. To get around that any value returned
-	    // will be dumped.
-
-	    //rb_raise(rb_eTypeError, "%s.to_hash() did not return a Hash.\n", rb_class2name(rb_obj_class(obj)));
-	    dump_val(h, depth, out, 0, 0, false);
-	}
-	dump_hash(h, Qundef, depth, out->opts->mode, out);
-    } else if (as_ok && Yes == out->opts->as_json && rb_respond_to(obj, oj_as_json_id)) {
-	volatile VALUE	aj;
-
-#if HAS_METHOD_ARITY
-	// Some classes elect to not take an options argument so check the arity
-	// of as_json.
-	switch (rb_obj_method_arity(obj, oj_as_json_id)) {
-	case 0:
-	    aj = rb_funcall2(obj, oj_as_json_id, 0, 0);
-	    break;
-	case 1:
-	    if (1 <= argc) {
-		aj = rb_funcall2(obj, oj_as_json_id, 1, argv);
-	    } else {
-		VALUE	nothing [1];
-
-		nothing[0] = Qnil;
-		aj = rb_funcall2(obj, oj_as_json_id, 1, nothing);
-	    }
-	    break;
-	default:
-	    aj = rb_funcall2(obj, oj_as_json_id, argc, argv);
-	    break;
-	}
-#else
-	aj = rb_funcall2(obj, oj_as_json_id, argc, argv);
-#endif
-	// Catch the obvious brain damaged recursive dumping.
-	if (aj == obj) {
-	    volatile VALUE	rstr = rb_funcall(obj, oj_to_s_id, 0);
-
-	    dump_cstr(rb_string_value_ptr((VALUE*)&rstr), (int)RSTRING_LEN(rstr), 0, 0, out);
-	} else {
-	    dump_val(aj, depth, out, 0, 0, false);
-	}
-    } else if (Yes == out->opts->to_json && rb_respond_to(obj, oj_to_json_id)) {
-	volatile VALUE	rs = rb_funcall(obj, oj_to_json_id, 0);
-	const char	*s;
-	int		len;
-
-	s = rb_string_value_ptr((VALUE*)&rs);
-	len = (int)RSTRING_LEN(rs);
-	if (out->end - out->cur <= len) {
-	    grow(out, len);
-	}
-	memcpy(out->cur, s, len);
-	out->cur += len;
-    } else {
-	volatile VALUE	rstr = rb_funcall(obj, oj_to_s_id, 0);
-
-	dump_cstr(rb_string_value_ptr((VALUE*)&rstr), (int)RSTRING_LEN(rstr), 0, 0, out);
-    }
-}
-
-static void
-dump_struct_obj(VALUE obj, int depth, Out out) {
-    VALUE	clas = rb_obj_class(obj);
-    const char	*class_name = rb_class2name(clas);
-    int		i;
-    int		d2 = depth + 1;
-    int		d3 = d2 + 1;
-    size_t	len = strlen(class_name);
-    size_t	size = d2 * out->indent + d3 * out->indent + 10 + len;
-
-    if (out->end - out->cur <= (long)size) {
-	grow(out, size);
-    }
-    *out->cur++ = '{';
-    fill_indent(out, d2);
-    *out->cur++ = '"';
-    *out->cur++ = '^';
-    *out->cur++ = 'u';
-    *out->cur++ = '"';
-    *out->cur++ = ':';
-    *out->cur++ = '[';
-#if HAS_STRUCT_MEMBERS
-    if ('#' == *class_name) {
-	VALUE		ma = rb_struct_s_members(clas);
-	const char	*name;
-	int		cnt = (int)RARRAY_LEN(ma);
-
-	*out->cur++ = '[';
-	for (i = 0; i < cnt; i++) {
-	    name = rb_id2name(SYM2ID(rb_ary_entry(ma, i)));
-	    len = strlen(name);
-	    size = len + 3;
-	    if (out->end - out->cur <= (long)size) {
-		grow(out, size);
-	    }
-	    if (0 < i) {
-		*out->cur++ = ',';
-	    }
-	    *out->cur++ = '"';
-	    memcpy(out->cur, name, len);
-	    out->cur += len;
-	    *out->cur++ = '"';
-	}
-	*out->cur++ = ']';
-    } else {
-#else
-    if (true) {
-#endif
-	fill_indent(out, d3);
-	*out->cur++ = '"';
-	memcpy(out->cur, class_name, len);
-	out->cur += len;
-	*out->cur++ = '"';
-    }
-    *out->cur++ = ',';
-    size = d3 * out->indent + 2;
-#ifdef RSTRUCT_LEN
-    {
-	VALUE	v;
-	int	cnt;
-#if UNIFY_FIXNUM_AND_BIGNUM
-	cnt = (int)NUM2LONG(RSTRUCT_LEN(obj));
-#else // UNIFY_FIXNUM_AND_INTEGER
-	cnt = (int)RSTRUCT_LEN(obj);
-#endif // UNIFY_FIXNUM_AND_INTEGER
-	
-	for (i = 0; i < cnt; i++) {
-	    v = RSTRUCT_GET(obj, i);
-	    if (out->end - out->cur <= (long)size) {
-		grow(out, size);
-	    }
-	    fill_indent(out, d3);
-	    dump_val(v, d3, out, 0, 0, true);
-	    *out->cur++ = ',';
-	}
-    }
-#else
-    {
-	// This is a bit risky as a struct in C ruby is not the same as a Struct
-	// class in interpreted Ruby so length() may not be defined.
-	int	slen = FIX2INT(rb_funcall2(obj, oj_length_id, 0, 0));
-
-	for (i = 0; i < slen; i++) {
-	    if (out->end - out->cur <= (long)size) {
-		grow(out, size);
-	    }
-	    fill_indent(out, d3);
-	    dump_val(rb_struct_aref(obj, INT2FIX(i)), d3, out, 0, 0, true);
-	    *out->cur++ = ',';
-	}
-    }
-#endif
-    out->cur--;
-    *out->cur++ = ']';
-    *out->cur++ = '}';
-    *out->cur = '\0';
-}
-
-static void
-dump_odd(VALUE obj, Odd odd, VALUE clas, int depth, Out out) {
-    ID			*idp;
-    AttrGetFunc		*fp;
-    volatile VALUE	v;
-    const char		*name;
-    size_t		size;
-    int			d2 = depth + 1;
-
-    if (out->end - out->cur <= 2) {
-	grow(out, 2);
-    }
-    *out->cur++ = '{';
-    if (Qundef != clas) {
-	const char	*class_name = rb_class2name(clas);
-	int		clen = (int)strlen(class_name);
-
-	size = d2 * out->indent + clen + 10;
-	if (out->end - out->cur <= (long)size) {
-	    grow(out, size);
-	}
-	fill_indent(out, d2);
-	*out->cur++ = '"';
-	*out->cur++ = '^';
-	*out->cur++ = 'O';
-	*out->cur++ = '"';
-	*out->cur++ = ':';
-	dump_cstr(class_name, clen, 0, 0, out);
-	*out->cur++ = ',';
-    }
-    if (odd->raw) {
-	v = rb_funcall(obj, *odd->attrs, 0);
-	if (Qundef == v || T_STRING != rb_type(v)) {
-	    rb_raise(rb_eEncodingError, "Invalid type for raw JSON.\n");
-	} else {	    
-	    const char	*s = rb_string_value_ptr((VALUE*)&v);
-	    int		len = (int)RSTRING_LEN(v);
-	    const char	*name = rb_id2name(*odd->attrs);
-	    size_t	nlen = strlen(name);
-
-	    size = len + d2 * out->indent + nlen + 10;
-	    if (out->end - out->cur <= (long)size) {
-		grow(out, size);
-	    }
-	    fill_indent(out, d2);
-	    *out->cur++ = '"';
-	    memcpy(out->cur, name, nlen);
-	    out->cur += nlen;
-	    *out->cur++ = '"';
-	    *out->cur++ = ':';
-	    memcpy(out->cur, s, len);
-	    out->cur += len;
-	    *out->cur = '\0';
-	}
+        char format[64] = "%04d-%02d-%02dT%02d:%02d:%02d.%09ldZ";
+        int  len;
+
+        if (9 > out->opts->sec_prec) {
+            format[32] = '0' + out->opts->sec_prec;
+        }
+        len = sprintf(buf, format, ti.year, ti.mon, ti.day, ti.hour, ti.min, ti.sec, (long)nsec);
+        oj_dump_cstr(buf, len, 0, 0, out);
     } else {
-	size = d2 * out->indent + 1;
-	for (idp = odd->attrs, fp = odd->attrFuncs; 0 != *idp; idp++, fp++) {
-	    size_t	nlen;
-
-	    if (out->end - out->cur <= (long)size) {
-		grow(out, size);
-	    }
-	    name = rb_id2name(*idp);
-	    nlen = strlen(name);
-	    if (0 != *fp) {
-		v = (*fp)(obj);
-	    } else if (0 == strchr(name, '.')) {
-		v = rb_funcall(obj, *idp, 0);
-	    } else {
-		char	nbuf[256];
-		char	*n2 = nbuf;
-		char	*n;
-		char	*end;
-		ID	i;
-	    
-		if (sizeof(nbuf) <= nlen) {
-		    n2 = strdup(name);
-		} else {
-		    strcpy(n2, name);
-		}
-		n = n2;
-		v = obj;
-		while (0 != (end = strchr(n, '.'))) {
-		    *end = '\0';
-		    i = rb_intern(n);
-		    v = rb_funcall(v, i, 0);
-		    n = end + 1;
-		}
-		i = rb_intern(n);
-		v = rb_funcall(v, i, 0);
-		if (nbuf != n2) {
-		    free(n2);
-		}
-	    }
-	    fill_indent(out, d2);
-	    dump_cstr(name, nlen, 0, 0, out);
-	    *out->cur++ = ':';
-	    dump_val(v, d2, out, 0, 0, true);
-	    if (out->end - out->cur <= 2) {
-		grow(out, 2);
-	    }
-	    *out->cur++ = ',';
-	}
-	out->cur--;
-    }
-    *out->cur++ = '}';
-    *out->cur = '\0';
-}
+        char format[64] = "%04d-%02d-%02dT%02d:%02d:%02d.%09ld%c%02d:%02d";
+        int  len;
 
-static void
-raise_strict(VALUE obj) {
-    rb_raise(rb_eTypeError, "Failed to dump %s Object to JSON in strict mode.\n", rb_class2name(rb_obj_class(obj)));
-}
-
-static void
-dump_val(VALUE obj, int depth, Out out, int argc, VALUE *argv, bool as_ok) {
-    int	type = rb_type(obj);
-
-    if (MAX_DEPTH < depth) {
-	rb_raise(rb_eNoMemError, "Too deeply nested.\n");
-    }
-#ifdef OJ_DEBUG
-    printf("Oj-debug: dump_val %s\n", rb_class2name(rb_obj_class(obj)));
-#endif
-    switch (type) {
-    case T_NIL:		dump_nil(out);				break;
-    case T_TRUE:	dump_true(out);				break;
-    case T_FALSE:	dump_false(out);			break;
-    case T_FIXNUM:	dump_fixnum(obj, out);			break;
-    case T_FLOAT:	dump_float(obj, out);			break;
-    case T_MODULE:
-    case T_CLASS:
-	switch (out->opts->mode) {
-	case StrictMode:	raise_strict(obj);		break;
-	case NullMode:		dump_nil(out);			break;
-	case CompatMode:	dump_class_comp(obj, out);	break;
-	case ObjectMode:
-	default:		dump_class_obj(obj, out);	break;
-	}
-	break;
-    case T_SYMBOL:
-	switch (out->opts->mode) {
-	case StrictMode:	raise_strict(obj);		break;
-	case NullMode:		dump_nil(out);			break;
-	case CompatMode:	dump_sym_comp(obj, out);	break;
-	case ObjectMode:
-	default:		dump_sym_obj(obj, out);		break;
-	}
-	break;
-    case T_STRUCT: // for Range
-	switch (out->opts->mode) {
-	case StrictMode:	raise_strict(obj);		break;
-	case NullMode:		dump_nil(out);			break;
-	case CompatMode:	dump_struct_comp(obj, depth, out, argc, argv, as_ok);	break;
-	case ObjectMode:
-	default:		dump_struct_obj(obj, depth, out);	break;
-	}
-	break;
-    default:
-	// Most developers have enough sense not to subclass primitive types but
-	// since these classes could potentially be subclassed a check for odd
-	// classes is performed.
-	{
-	    VALUE	clas = rb_obj_class(obj);
-	    Odd		odd;
-
-	    if (ObjectMode == out->opts->mode && 0 != (odd = oj_get_odd(clas))) {
-		dump_odd(obj, odd, clas, depth + 1, out);
-		return;
-	    }
-	    switch (type) {
-	    case T_BIGNUM:		dump_bignum(obj, out);		break;
-	    case T_STRING:
-		switch (out->opts->mode) {
-		case StrictMode:
-		case NullMode:
-		case CompatMode:	dump_str_comp(obj, out);	break;
-		case ObjectMode:
-		default:		dump_str_obj(obj, clas, depth, out);	break;
-		}
-		break;
-	    case T_ARRAY:		dump_array(obj, clas, depth, out);	break;
-	    case T_HASH:		dump_hash(obj, clas, depth, out->opts->mode, out);	break;
-#if (defined T_RATIONAL && defined RRATIONAL)
-	    case T_RATIONAL:
-#endif
-	    case T_OBJECT:
-		switch (out->opts->mode) {
-		case StrictMode:	dump_data_strict(obj, out);	break;
-		case NullMode:		dump_data_null(obj, out);	break;
-		case CompatMode:	dump_obj_comp(obj, depth, out, argc, argv, as_ok);	break;
-		case ObjectMode:
-		default:		dump_obj_obj(obj, depth, out);	break;
-		}
-		break;
-	    case T_DATA:
-		switch (out->opts->mode) {
-		case StrictMode:	dump_data_strict(obj, out);	break;
-		case NullMode:		dump_data_null(obj, out);	break;
-		case CompatMode:	dump_data_comp(obj, depth, out, argc, argv, as_ok);break;
-		case ObjectMode:
-		default:		dump_data_obj(obj, depth, out);	break;
-		}
-		break;
-#if (defined T_COMPLEX && defined RCOMPLEX)
-	    case T_COMPLEX:
-#endif
-	    case T_REGEXP:
-		switch (out->opts->mode) {
-		case StrictMode:	raise_strict(obj);		break;
-		case NullMode:		dump_nil(out);			break;
-		case CompatMode:
-		case ObjectMode:
-		default:		dump_obj_comp(obj, depth, out, argc, argv, as_ok);	break;
-		}
-		break;
-	    default:
-		switch (out->opts->mode) {
-		case StrictMode:	raise_strict(obj);		break;
-		case NullMode:		dump_nil(out);			break;
-		case CompatMode: {
-		    volatile VALUE	rstr = rb_funcall(obj, oj_to_s_id, 0);
-		    dump_cstr(rb_string_value_ptr((VALUE*)&rstr), (int)RSTRING_LEN(rstr), 0, 0, out);
-		    break;
-		}
-		case ObjectMode:
-		default:
-		    rb_raise(rb_eNotImpError, "Failed to dump '%s' Object (%02x)\n",
-			     rb_class2name(rb_obj_class(obj)), rb_type(obj));
-		    break;
-		}
-		break;
-	    }
-	}
+        if (9 > out->opts->sec_prec) {
+            format[32] = '0' + out->opts->sec_prec;
+        }
+        len = sprintf(buf, format, ti.year, ti.mon, ti.day, ti.hour, ti.min, ti.sec, (long)nsec, tzsign, tzhour, tzmin);
+        oj_dump_cstr(buf, len, 0, 0, out);
     }
 }
 
-void
-oj_dump_obj_to_json(VALUE obj, Options copts, Out out) {
+void oj_dump_obj_to_json(VALUE obj, Options copts, Out out) {
     oj_dump_obj_to_json_using_params(obj, copts, out, 0, 0);
 }
 
-void
-oj_dump_obj_to_json_using_params(VALUE obj, Options copts, Out out, int argc, VALUE *argv) {
+void oj_dump_obj_to_json_using_params(VALUE obj, Options copts, Out out, int argc, VALUE *argv) {
     if (0 == out->buf) {
-	out->buf = ALLOC_N(char, 4096);
-	out->end = out->buf + 4095 - BUFFER_EXTRA; // 1 less than end plus extra for possible errors
-	out->allocated = 1;
+        oj_out_init(out);
     }
-    out->cur = out->buf;
     out->circ_cnt = 0;
-    out->opts = copts;
+    out->opts     = copts;
     out->hash_cnt = 0;
+    out->indent   = copts->indent;
+    out->argc     = argc;
+    out->argv     = argv;
+    out->ropts    = NULL;
     if (Yes == copts->circular) {
-	oj_cache8_new(&out->circ_cache);
+        oj_cache8_new(&out->circ_cache);
+    }
+    switch (copts->mode) {
+    case StrictMode: oj_dump_strict_val(obj, 0, out); break;
+    case NullMode: oj_dump_null_val(obj, 0, out); break;
+    case ObjectMode: oj_dump_obj_val(obj, 0, out); break;
+    case CompatMode: oj_dump_compat_val(obj, 0, out, Yes == copts->to_json); break;
+    case RailsMode: oj_dump_rails_val(obj, 0, out); break;
+    case CustomMode: oj_dump_custom_val(obj, 0, out, true); break;
+    case WabMode: oj_dump_wab_val(obj, 0, out); break;
+    default: oj_dump_custom_val(obj, 0, out, true); break;
     }
-    out->indent = copts->indent;
-    dump_val(obj, 0, out, argc, argv, true);
     if (0 < out->indent) {
-	switch (*(out->cur - 1)) {
-	case ']':
-	case '}':
-	    grow(out, 1);
-	    *out->cur++ = '\n';
-	default:
-	    break;
-	}
+        switch (*(out->cur - 1)) {
+        case ']':
+        case '}': assure_size(out, 1); *out->cur++ = '\n';
+        default: break;
+        }
     }
     *out->cur = '\0';
     if (Yes == copts->circular) {
-	oj_cache8_delete(out->circ_cache);
+        oj_cache8_delete(out->circ_cache);
     }
 }
 
-void
-oj_write_obj_to_file(VALUE obj, const char *path, Options copts) {
-    char	buf[4096];
-    struct _Out out;
-    size_t	size;
-    FILE	*f;
-    int		ok;
-
-    out.buf = buf;
-    out.end = buf + sizeof(buf) - BUFFER_EXTRA;
-    out.allocated = 0;
+void oj_write_obj_to_file(VALUE obj, const char *path, Options copts) {
+    struct _out out;
+    size_t      size;
+    FILE       *f;
+    int         ok;
+
+    oj_out_init(&out);
+
     out.omit_nil = copts->dump_opts.omit_nil;
     oj_dump_obj_to_json(obj, copts, &out);
     size = out.cur - out.buf;
     if (0 == (f = fopen(path, "w"))) {
-	if (out.allocated) {
-	    xfree(out.buf);
-	}
-	rb_raise(rb_eIOError, "%s\n", strerror(errno));
+        oj_out_free(&out);
+        rb_raise(rb_eIOError, "%s", strerror(errno));
     }
     ok = (size == fwrite(out.buf, 1, size, f));
-    if (out.allocated) {
-	xfree(out.buf);
+
+    oj_out_free(&out);
+
+    if (!ok) {
+        int err = ferror(f);
+        fclose(f);
+
+        rb_raise(rb_eIOError, "Write failed. [%d:%s]", err, strerror(err));
     }
     fclose(f);
-    if (!ok) {
-	int	err = ferror(f);
+}
 
-	rb_raise(rb_eIOError, "Write failed. [%d:%s]\n", err, strerror(err));
+#if !IS_WINDOWS
+static void write_ready(int fd) {
+    struct pollfd pp;
+    int           i;
+
+    pp.fd      = fd;
+    pp.events  = POLLERR | POLLOUT;
+    pp.revents = 0;
+    if (0 >= (i = poll(&pp, 1, 5000))) {
+        if (0 == i || EAGAIN == errno) {
+            rb_raise(rb_eIOError, "write timed out");
+        }
+        rb_raise(rb_eIOError, "write failed. %d %s.", errno, strerror(errno));
     }
 }
+#endif
 
-void
-oj_write_obj_to_stream(VALUE obj, VALUE stream, Options copts) {
-    char	buf[4096];
-    struct _Out out;
-    ssize_t	size;
-    VALUE	clas = rb_obj_class(stream);
+void oj_write_obj_to_stream(VALUE obj, VALUE stream, Options copts) {
+    struct _out out;
+    ssize_t     size;
+    VALUE       clas = rb_obj_class(stream);
 #if !IS_WINDOWS
-    int		fd;
-    VALUE	s;
+    int   fd;
+    VALUE s;
 #endif
 
-    out.buf = buf;
-    out.end = buf + sizeof(buf) - BUFFER_EXTRA;
-    out.allocated = 0;
+    oj_out_init(&out);
+
     out.omit_nil = copts->dump_opts.omit_nil;
     oj_dump_obj_to_json(obj, copts, &out);
     size = out.cur - out.buf;
     if (oj_stringio_class == clas) {
-	rb_funcall(stream, oj_write_id, 1, rb_str_new(out.buf, size));
+        rb_funcall(stream, oj_write_id, 1, rb_str_new(out.buf, size));
 #if !IS_WINDOWS
-    } else if (rb_respond_to(stream, oj_fileno_id) &&
-	       Qnil != (s = rb_funcall(stream, oj_fileno_id, 0)) &&
-	       0 != (fd = FIX2INT(s))) {
-	if (size != write(fd, out.buf, size)) {
-	    if (out.allocated) {
-		xfree(out.buf);
-	    }
-	    rb_raise(rb_eIOError, "Write failed. [%d:%s]\n", errno, strerror(errno));
-	}
+    } else if (rb_respond_to(stream, oj_fileno_id) && Qnil != (s = rb_funcall(stream, oj_fileno_id, 0)) &&
+               0 != (fd = FIX2INT(s))) {
+        ssize_t cnt;
+        ssize_t total = 0;
+
+        while (true) {
+            if (0 > (cnt = write(fd, out.buf + total, size - total))) {
+                if (EAGAIN != errno) {
+                    rb_raise(rb_eIOError, "write failed. %d %s.", errno, strerror(errno));
+                    break;
+                }
+            }
+            total += cnt;
+            if (size <= total) {
+                // Completed
+                break;
+            }
+            write_ready(fd);
+        }
 #endif
     } else if (rb_respond_to(stream, oj_write_id)) {
-	rb_funcall(stream, oj_write_id, 1, rb_str_new(out.buf, size));
+        rb_funcall(stream, oj_write_id, 1, rb_str_new(out.buf, size));
     } else {
-	if (out.allocated) {
-	    xfree(out.buf);
-	}
-	rb_raise(rb_eArgError, "to_stream() expected an IO Object.");
-    }
-    if (out.allocated) {
-	xfree(out.buf);
+        oj_out_free(&out);
+        rb_raise(rb_eArgError, "to_stream() expected an IO Object.");
     }
+    oj_out_free(&out);
 }
 
-// dump leaf functions
+void oj_dump_str(VALUE obj, int depth, Out out, bool as_ok) {
+    int idx = RB_ENCODING_GET(obj);
 
-inline static void
-dump_chars(const char *s, size_t size, Out out) {
-    if (out->end - out->cur <= (long)size) {
-	grow(out, size);
+    if (oj_utf8_encoding_index != idx) {
+        rb_encoding *enc = rb_enc_from_index(idx);
+        obj              = rb_str_conv_enc(obj, enc, oj_utf8_encoding);
     }
-    memcpy(out->cur, s, size);
-    out->cur += size;
-    *out->cur = '\0';
+    oj_dump_cstr(RSTRING_PTR(obj), RSTRING_LEN(obj), 0, 0, out);
 }
 
-static void
-dump_leaf_str(Leaf leaf, Out out) {
-    switch (leaf->value_type) {
-    case STR_VAL:
-	dump_cstr(leaf->str, strlen(leaf->str), 0, 0, out);
-	break;
-    case RUBY_VAL:
-	dump_cstr(rb_string_value_cstr(&leaf->value), (int)RSTRING_LEN(leaf->value), 0, 0, out);
-	break;
-    case COL_VAL:
-    default:
-	rb_raise(rb_eTypeError, "Unexpected value type %02x.\n", leaf->value_type);
-	break;
-    }
+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), RSTRING_LEN(s), 0, 0, out);
 }
 
-static void
-dump_leaf_fixnum(Leaf leaf, Out out) {
-    switch (leaf->value_type) {
-    case STR_VAL:
-	dump_chars(leaf->str, strlen(leaf->str), out);
-	break;
-    case RUBY_VAL:
-	if (T_BIGNUM == rb_type(leaf->value)) {
-	    dump_bignum(leaf->value, out);
-	} else {
-	    dump_fixnum(leaf->value, out);
-	}
-	break;
-    case COL_VAL:
-    default:
-	rb_raise(rb_eTypeError, "Unexpected value type %02x.\n", leaf->value_type);
-	break;
+static void debug_raise(const char *orig, size_t cnt, int line) {
+    char        buf[1024];
+    char       *b     = buf;
+    const char *s     = orig;
+    const char *s_end = s + cnt;
+
+    if (32 < s_end - s) {
+        s_end = s + 32;
     }
+    for (; s < s_end; s++) {
+        b += sprintf(b, " %02x", *s);
+    }
+    *b = '\0';
+    rb_raise(oj_json_generator_error_class, "Partial character in string. %s @ %d", buf, line);
 }
 
-static void
-dump_leaf_float(Leaf leaf, Out out) {
-    switch (leaf->value_type) {
-    case STR_VAL:
-	dump_chars(leaf->str, strlen(leaf->str), out);
-	break;
-    case RUBY_VAL:
-	dump_float(leaf->value, out);
-	break;
-    case COL_VAL:
-    default:
-	rb_raise(rb_eTypeError, "Unexpected value type %02x.\n", leaf->value_type);
-	break;
+void oj_dump_raw_json(VALUE obj, int depth, Out out) {
+    if (oj_string_writer_class == rb_obj_class(obj)) {
+        StrWriter sw;
+        size_t    len;
+
+        sw  = oj_str_writer_unwrap(obj);
+        len = sw->out.cur - sw->out.buf;
+
+        if (0 < len) {
+            len--;
+        }
+        oj_dump_raw(sw->out.buf, len, out);
+    } else {
+        volatile VALUE jv;
+
+        TRACE(out->opts->trace, "raw_json", obj, depth + 1, TraceRubyIn);
+        jv = rb_funcall(obj, oj_raw_json_id, 2, RB_INT2NUM(depth), RB_INT2NUM(out->indent));
+        TRACE(out->opts->trace, "raw_json", obj, depth + 1, TraceRubyOut);
+        oj_dump_raw(RSTRING_PTR(jv), (size_t)RSTRING_LEN(jv), out);
     }
 }
 
-static void
-dump_leaf_array(Leaf leaf, int depth, Out out) {
-    size_t	size;
-    int		d2 = depth + 1;
+#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};
 
-    size = 2;
-    if (out->end - out->cur <= (long)size) {
-	grow(out, size);
+    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));
     }
-    *out->cur++ = '[';
-    if (0 == leaf->elements) {
-	*out->cur++ = ']';
-    } else {
-	Leaf	first = leaf->elements->next;
-	Leaf	e = first;
-
-	size = d2 * out->indent + 2;
-	do {
-	    if (out->end - out->cur <= (long)size) {
-		grow(out, size);
-	    }
-	    fill_indent(out, d2);
-	    dump_leaf(e, d2, out);
-	    if (e->next != first) {
-		*out->cur++ = ',';
-	    }
-	    e = e->next;
-	} while (e != first);
-	size = depth * out->indent + 1;
-	if (out->end - out->cur <= (long)size) {
-	    grow(out, size);
-	}
-	fill_indent(out, depth);
-	*out->cur++ = ']';
+    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;
     }
-    *out->cur = '\0';
+    return result;
 }
 
-static void
-dump_leaf_hash(Leaf leaf, int depth, Out out) {
-    size_t	size;
-    int		d2 = depth + 1;
+#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;
+#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;
 
-    size = 2;
-    if (out->end - out->cur <= (long)size) {
-	grow(out, size);
+    switch (out->opts->escape_mode) {
+    case NLEsc:
+        cmap = newline_friendly_chars;
+        size = newline_friendly_size((uint8_t *)str, cnt);
+        break;
+    case ASCIIEsc:
+        cmap = ascii_friendly_chars;
+        size = ascii_friendly_size((uint8_t *)str, cnt);
+        break;
+    case SlashEsc:
+        has_hi = true;
+        cmap   = slash_friendly_chars;
+        size   = slash_friendly_size((uint8_t *)str, cnt);
+        break;
+    case XSSEsc:
+        cmap = xss_friendly_chars;
+        size = xss_friendly_size((uint8_t *)str, cnt);
+        break;
+    case JXEsc:
+        cmap                  = hixss_friendly_chars;
+        size                  = hixss_friendly_size((uint8_t *)str, cnt);
+        do_unicode_validation = true;
+        break;
+    case RailsXEsc: {
+        long sz;
+
+        cmap = rails_xss_friendly_chars;
+#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;
+        } else {
+            size = (size_t)sz;
+        }
+        do_unicode_validation = true;
+        break;
+    }
+    case RailsEsc: {
+        long sz;
+        cmap = rails_friendly_chars;
+#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;
+        } else {
+            size = (size_t)sz;
+        }
+        do_unicode_validation = true;
+        break;
     }
-    *out->cur++ = '{';
-    if (0 == leaf->elements) {
-	*out->cur++ = '}';
+    case JSONEsc:
+    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++ = '"';
+
+    if (escape1) {
+        APPEND_CHARS(out->cur, "\\u00", 4);
+        dump_hex((uint8_t)*str, out);
+        cnt--;
+        size--;
+        str++;
+        is_sym = 0;  // just to make sure
+    }
+    if (cnt == size && !has_hi) {
+        if (is_sym) {
+            *out->cur++ = ':';
+        }
+        APPEND_CHARS(out->cur, str, cnt);
+        *out->cur++ = '"';
     } else {
-	Leaf	first = leaf->elements->next;
-	Leaf	e = first;
-
-	size = d2 * out->indent + 2;
-	do {
-	    if (out->end - out->cur <= (long)size) {
-		grow(out, size);
-	    }
-	    fill_indent(out, d2);
-	    dump_cstr(e->key, strlen(e->key), 0, 0, out);
-	    *out->cur++ = ':';
-	    dump_leaf(e, d2, out);
-	    if (e->next != first) {
-		*out->cur++ = ',';
-	    }
-	    e = e->next;
-	} while (e != first);
-	size = depth * out->indent + 1;
-	if (out->end - out->cur <= (long)size) {
-	    grow(out, size);
-	}
-	fill_indent(out, depth);
-	*out->cur++ = '}';
+        const char *end         = str + cnt;
+        const char *check_start = str;
+
+        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++) {
+            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)) {
+                        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
+            }
+        }
+        *out->cur++ = '"';
+    }
+    if (do_unicode_validation && 0 < str - orig && 0 != (0x80 & *(str - 1))) {
+        uint8_t c = (uint8_t)*(str - 1);
+        int     i;
+        int     scnt = (int)(str - orig);
+
+        // Last utf-8 characters must be 0x10xxxxxx. The start must be
+        // 0x110xxxxx for 2 characters, 0x1110xxxx for 3, and 0x11110xxx for
+        // 4.
+        if (0 != (0x40 & c)) {
+            debug_raise(orig, cnt, __LINE__);
+        }
+        for (i = 1; i < (int)scnt && i < 4; i++) {
+            c = str[-1 - i];
+            if (0x80 != (0xC0 & c)) {
+                switch (i) {
+                case 1:
+                    if (0xC0 != (0xE0 & c)) {
+                        debug_raise(orig, cnt, __LINE__);
+                    }
+                    break;
+                case 2:
+                    if (0xE0 != (0xF0 & c)) {
+                        debug_raise(orig, cnt, __LINE__);
+                    }
+                    break;
+                case 3:
+                    if (0xF0 != (0xF8 & c)) {
+                        debug_raise(orig, cnt, __LINE__);
+                    }
+                    break;
+                default:  // can't get here
+                    break;
+                }
+                break;
+            }
+        }
+        if (i == (int)scnt || 4 <= i) {
+            debug_raise(orig, cnt, __LINE__);
+        }
     }
     *out->cur = '\0';
 }
 
-static void
-dump_leaf(Leaf leaf, int depth, Out out) {
-    switch (leaf->rtype) {
-    case T_NIL:
-	dump_nil(out);
-	break;
-    case T_TRUE:
-	dump_true(out);
-	break;
-    case T_FALSE:
-	dump_false(out);
-	break;
-    case T_STRING:
-	dump_leaf_str(leaf, out);
-	break;
-    case T_FIXNUM:
-	dump_leaf_fixnum(leaf, out);
-	break;
-    case T_FLOAT:
-	dump_leaf_float(leaf, out);
-	break;
-    case T_ARRAY:
-	dump_leaf_array(leaf, depth, out);
-	break;
-    case T_HASH:
-	dump_leaf_hash(leaf, depth, out);
-	break;
-    default:
-	rb_raise(rb_eTypeError, "Unexpected type %02x.\n", leaf->rtype);
-	break;
-    }
+void oj_dump_class(VALUE obj, int depth, Out out, bool as_ok) {
+    const char *s = rb_class2name(obj);
+
+    oj_dump_cstr(s, strlen(s), 0, 0, out);
 }
 
-void
-oj_dump_leaf_to_json(Leaf leaf, Options copts, Out out) {
-    if (0 == out->buf) {
-	out->buf = ALLOC_N(char, 4096);
-	out->end = out->buf + 4095 - BUFFER_EXTRA; // 1 less than end plus extra for possible errors
-	out->allocated = 1;
-    }
-    out->cur = out->buf;
-    out->circ_cnt = 0;
-    out->opts = copts;
-    out->hash_cnt = 0;
-    out->indent = copts->indent;
-    dump_leaf(leaf, 0, out);
+void oj_dump_obj_to_s(VALUE obj, Out out) {
+    volatile VALUE rstr = oj_safe_string_convert(obj);
+
+    oj_dump_cstr(RSTRING_PTR(rstr), RSTRING_LEN(rstr), 0, 0, out);
 }
 
-void
-oj_write_leaf_to_file(Leaf leaf, const char *path, Options copts) {
-    char	buf[4096];
-    struct _Out out;
-    size_t	size;
-    FILE	*f;
+void oj_dump_raw(const char *str, size_t cnt, Out out) {
+    assure_size(out, cnt + 10);
+    APPEND_CHARS(out->cur, str, cnt);
+    *out->cur = '\0';
+}
 
-    out.buf = buf;
-    out.end = buf + sizeof(buf) - BUFFER_EXTRA;
-    out.allocated = 0;
-    out.omit_nil = copts->dump_opts.omit_nil;
-    oj_dump_leaf_to_json(leaf, copts, &out);
-    size = out.cur - out.buf;
-    if (0 == (f = fopen(path, "w"))) {
-	rb_raise(rb_eIOError, "%s\n", strerror(errno));
+void oj_out_init(Out out) {
+    out->buf       = out->stack_buffer;
+    out->cur       = out->buf;
+    out->end       = out->buf + sizeof(out->stack_buffer) - BUFFER_EXTRA;
+    out->allocated = false;
+}
+
+void oj_out_free(Out out) {
+    if (out->allocated) {
+        OJ_R_FREE(out->buf);  // TBD
     }
-    if (size != fwrite(out.buf, 1, size, f)) {
-	int	err = ferror(f);
+}
 
-	rb_raise(rb_eIOError, "Write failed. [%d:%s]\n", err, strerror(err));
+void oj_grow_out(Out out, size_t len) {
+    size_t size = out->end - out->buf;
+    long   pos  = out->cur - out->buf;
+    char  *buf  = out->buf;
+
+    size *= 2;
+    if (size <= len * 2 + pos) {
+        size += len;
     }
-    if (out.allocated) {
-	xfree(out.buf);
+    if (out->allocated) {
+        OJ_R_REALLOC_N(buf, char, (size + BUFFER_EXTRA));
+    } else {
+        buf            = OJ_R_ALLOC_N(char, (size + BUFFER_EXTRA));
+        out->allocated = true;
+        memcpy(buf, out->buf, out->end - out->buf + BUFFER_EXTRA);
     }
-    fclose(f);
+    if (0 == buf) {
+        rb_raise(rb_eNoMemError, "Failed to create string. [%d:%s]", ENOSPC, strerror(ENOSPC));
+    }
+    out->buf = buf;
+    out->end = buf + size;
+    out->cur = out->buf + pos;
 }
 
-// string writer functions
-
-static void
-key_check(StrWriter sw, const char *key) {
-    DumpType	type = sw->types[sw->depth];
+void oj_dump_nil(VALUE obj, int depth, Out out, bool as_ok) {
+    assure_size(out, 4);
+    APPEND_CHARS(out->cur, "null", 4);
+    *out->cur = '\0';
+}
 
-    if (0 == key && (ObjectNew == type || ObjectType == type)) {
-	rb_raise(rb_eStandardError, "Can not push onto an Object without a key.");
-    }
+void oj_dump_true(VALUE obj, int depth, Out out, bool as_ok) {
+    assure_size(out, 4);
+    APPEND_CHARS(out->cur, "true", 4);
+    *out->cur = '\0';
 }
 
-static void
-push_type(StrWriter sw, DumpType type) {
-    if (sw->types_end <= sw->types + sw->depth + 1) {
-	size_t	size = (sw->types_end - sw->types) * 2;
+void oj_dump_false(VALUE obj, int depth, Out out, bool as_ok) {
+    assure_size(out, 5);
+    APPEND_CHARS(out->cur, "false", 5);
+    *out->cur = '\0';
+}
 
-	REALLOC_N(sw->types, char, size);
-	sw->types_end = sw->types + size;
+static const char digits_table[] = "\
+00010203040506070809\
+10111213141516171819\
+20212223242526272829\
+30313233343536373839\
+40414243444546474849\
+50515253545556575859\
+60616263646566676869\
+70717273747576777879\
+80818283848586878889\
+90919293949596979899";
+
+char *oj_longlong_to_string(long long num, bool negative, char *buf) {
+    while (100 <= num) {
+        unsigned idx = num % 100 * 2;
+        *buf--       = digits_table[idx + 1];
+        *buf--       = digits_table[idx];
+        num /= 100;
+    }
+    if (num < 10) {
+        *buf-- = num + '0';
+    } else {
+        *buf-- = digits_table[num * 2 + 1];
+        *buf-- = digits_table[num * 2];
     }
-    sw->depth++;
-    sw->types[sw->depth] = type;
-}
 
-static void
-maybe_comma(StrWriter sw) {
-    switch (sw->types[sw->depth]) {
-    case ObjectNew:
-	sw->types[sw->depth] = ObjectType;
-	break;
-    case ArrayNew:
-	sw->types[sw->depth] = ArrayType;
-	break;
-    case ObjectType:
-    case ArrayType:
-	// Always have a few characters available in the out.buf.
-	*sw->out.cur++ = ',';
-	break;
+    if (negative) {
+        *buf = '-';
+    } else {
+        buf++;
     }
+    return buf;
 }
 
-void
-oj_str_writer_push_key(StrWriter sw, const char *key) {
-    DumpType	type = sw->types[sw->depth];
-    long	size;
+void oj_dump_fixnum(VALUE obj, int depth, Out out, bool as_ok) {
+    char      buf[32];
+    char     *b              = buf + sizeof(buf) - 1;
+    long long num            = NUM2LL(obj);
+    bool      neg            = false;
+    size_t    cnt            = 0;
+    bool      dump_as_string = false;
 
-    if (sw->keyWritten) {
-	rb_raise(rb_eStandardError, "Can not push more than one key before pushing a non-key.");
+    if (out->opts->int_range_max != 0 && out->opts->int_range_min != 0 &&
+        (out->opts->int_range_max < num || out->opts->int_range_min > num)) {
+        dump_as_string = true;
     }
-    if (ObjectNew != type && ObjectType != type) {
-	rb_raise(rb_eStandardError, "Can only push a key onto an Object.");
+    if (0 > num) {
+        neg = true;
+        num = -num;
     }
-    size = sw->depth * sw->out.indent + 3;
-    if (sw->out.end - sw->out.cur <= (long)size) {
-	grow(&sw->out, size);
+    *b-- = '\0';
+
+    if (dump_as_string) {
+        *b-- = '"';
     }
-    maybe_comma(sw);
-    if (0 < sw->depth) {
-	fill_indent(&sw->out, sw->depth);
+    if (0 < num) {
+        b = oj_longlong_to_string(num, neg, b);
+    } else {
+        *b = '0';
     }
-    dump_cstr(key, strlen(key), 0, 0, &sw->out);
-    *sw->out.cur++ = ':';
-    sw->keyWritten = 1;
+    if (dump_as_string) {
+        *--b = '"';
+    }
+    cnt = sizeof(buf) - (b - buf) - 1;
+    assure_size(out, cnt);
+    APPEND_CHARS(out->cur, b, cnt);
+    *out->cur = '\0';
 }
 
-void
-oj_str_writer_push_object(StrWriter sw, const char *key) {
-    if (sw->keyWritten) {
-	sw->keyWritten = 0;
-	if (sw->out.end - sw->out.cur <= 1) {
-	    grow(&sw->out, 1);
-	}
-    } else {
-	long	size;
-
-	key_check(sw, key);
-	size = sw->depth * sw->out.indent + 3;
-	if (sw->out.end - sw->out.cur <= (long)size) {
-	    grow(&sw->out, size);
-	}
-	maybe_comma(sw);
-	if (0 < sw->depth) {
-	    fill_indent(&sw->out, sw->depth);
-	}
-	if (0 != key) {
-	    dump_cstr(key, strlen(key), 0, 0, &sw->out);
-	    *sw->out.cur++ = ':';
-	}
-    }
-    *sw->out.cur++ = '{';
-    push_type(sw, ObjectNew);
-}
+void oj_dump_bignum(VALUE obj, int depth, Out out, bool as_ok) {
+    volatile VALUE rs             = rb_big2str(obj, 10);
+    size_t         cnt            = RSTRING_LEN(rs);
+    bool           dump_as_string = false;
 
-void
-oj_str_writer_push_array(StrWriter sw, const char *key) {
-    if (sw->keyWritten) {
-	sw->keyWritten = 0;
-	if (sw->out.end - sw->out.cur <= 1) {
-	    grow(&sw->out, 1);
-	}
+    if (out->opts->int_range_max != 0 || out->opts->int_range_min != 0) {  // Bignum cannot be inside of Fixnum range
+        dump_as_string = true;
+        assure_size(out, cnt + 2);
+        *out->cur++ = '"';
     } else {
-	long	size;
-
-	key_check(sw, key);
-	size = sw->depth * sw->out.indent + 3;
-	if (sw->out.end - sw->out.cur <= size) {
-	    grow(&sw->out, size);
-	}
-	maybe_comma(sw);
-	if (0 < sw->depth) {
-	    fill_indent(&sw->out, sw->depth);
-	}
-	if (0 != key) {
-	    dump_cstr(key, strlen(key), 0, 0, &sw->out);
-	    *sw->out.cur++ = ':';
-	}
-    }
-    *sw->out.cur++ = '[';
-    push_type(sw, ArrayNew);
+        assure_size(out, cnt);
+    }
+    APPEND_CHARS(out->cur, RSTRING_PTR(rs), cnt);
+    if (dump_as_string) {
+        *out->cur++ = '"';
+    }
+    *out->cur = '\0';
 }
 
-void
-oj_str_writer_push_value(StrWriter sw, VALUE val, const char *key) {
-    if (sw->keyWritten) {
-	sw->keyWritten = 0;
-    } else {
-	long	size;
-
-	key_check(sw, key);
-	size = sw->depth * sw->out.indent + 3;
-	if (sw->out.end - sw->out.cur <= size) {
-	    grow(&sw->out, size);
-	}
-	maybe_comma(sw);
-	if (0 < sw->depth) {
-	    fill_indent(&sw->out, sw->depth);
-	}
-	if (0 != key) {
-	    dump_cstr(key, strlen(key), 0, 0, &sw->out);
-	    *sw->out.cur++ = ':';
-	}
-    }
-    dump_val(val, sw->depth, &sw->out, 0, 0, true);
-}
+// Removed dependencies on math due to problems with CentOS 5.4.
+void oj_dump_float(VALUE obj, int depth, Out out, bool as_ok) {
+    char   buf[64];
+    char  *b;
+    double d   = rb_num2dbl(obj);
+    size_t cnt = 0;
 
-void
-oj_str_writer_push_json(StrWriter sw, const char *json, const char *key) {
-    if (sw->keyWritten) {
-	sw->keyWritten = 0;
+    if (0.0 == d) {
+        b    = buf;
+        *b++ = '0';
+        *b++ = '.';
+        *b++ = '0';
+        *b++ = '\0';
+        cnt  = 3;
+    } else if (OJ_INFINITY == d) {
+        if (ObjectMode == out->opts->mode) {
+            strcpy(buf, inf_val);
+            cnt = sizeof(inf_val) - 1;
+        } else {
+            NanDump nd = out->opts->dump_opts.nan_dump;
+
+            if (AutoNan == nd) {
+                switch (out->opts->mode) {
+                case CompatMode: nd = WordNan; break;
+                case StrictMode: nd = RaiseNan; break;
+                case NullMode: nd = NullNan; break;
+                case CustomMode: nd = NullNan; break;
+                default: break;
+                }
+            }
+            switch (nd) {
+            case RaiseNan: raise_strict(obj); break;
+            case WordNan:
+                strcpy(buf, "Infinity");
+                cnt = 8;
+                break;
+            case NullNan:
+                strcpy(buf, "null");
+                cnt = 4;
+                break;
+            case HugeNan:
+            default:
+                strcpy(buf, inf_val);
+                cnt = sizeof(inf_val) - 1;
+                break;
+            }
+        }
+    } else if (-OJ_INFINITY == d) {
+        if (ObjectMode == out->opts->mode) {
+            strcpy(buf, ninf_val);
+            cnt = sizeof(ninf_val) - 1;
+        } else {
+            NanDump nd = out->opts->dump_opts.nan_dump;
+
+            if (AutoNan == nd) {
+                switch (out->opts->mode) {
+                case CompatMode: nd = WordNan; break;
+                case StrictMode: nd = RaiseNan; break;
+                case NullMode: nd = NullNan; break;
+                default: break;
+                }
+            }
+            switch (nd) {
+            case RaiseNan: raise_strict(obj); break;
+            case WordNan:
+                strcpy(buf, "-Infinity");
+                cnt = 9;
+                break;
+            case NullNan:
+                strcpy(buf, "null");
+                cnt = 4;
+                break;
+            case HugeNan:
+            default:
+                strcpy(buf, ninf_val);
+                cnt = sizeof(ninf_val) - 1;
+                break;
+            }
+        }
+    } else if (isnan(d)) {
+        if (ObjectMode == out->opts->mode) {
+            strcpy(buf, nan_val);
+            cnt = sizeof(nan_val) - 1;
+        } else {
+            NanDump nd = out->opts->dump_opts.nan_dump;
+
+            if (AutoNan == nd) {
+                switch (out->opts->mode) {
+                case ObjectMode: nd = HugeNan; break;
+                case StrictMode: nd = RaiseNan; break;
+                case NullMode: nd = NullNan; break;
+                default: break;
+                }
+            }
+            switch (nd) {
+            case RaiseNan: raise_strict(obj); break;
+            case WordNan:
+                strcpy(buf, "NaN");
+                cnt = 3;
+                break;
+            case NullNan:
+                strcpy(buf, "null");
+                cnt = 4;
+                break;
+            case HugeNan:
+            default:
+                strcpy(buf, nan_val);
+                cnt = sizeof(nan_val) - 1;
+                break;
+            }
+        }
+    } else if (d == (double)(long long int)d) {
+        cnt = snprintf(buf, sizeof(buf), "%.1f", d);
+    } else if (0 == out->opts->float_prec) {
+        volatile VALUE rstr = oj_safe_string_convert(obj);
+
+        cnt = RSTRING_LEN(rstr);
+        if ((int)sizeof(buf) <= cnt) {
+            cnt = sizeof(buf) - 1;
+        }
+        memcpy(buf, RSTRING_PTR(rstr), cnt);
+        buf[cnt] = '\0';
     } else {
-	long	size;
-
-	key_check(sw, key);
-	size = sw->depth * sw->out.indent + 3;
-	if (sw->out.end - sw->out.cur <= size) {
-	    grow(&sw->out, size);
-	}
-	maybe_comma(sw);
-	if (0 < sw->depth) {
-	    fill_indent(&sw->out, sw->depth);
-	}
-	if (0 != key) {
-	    dump_cstr(key, strlen(key), 0, 0, &sw->out);
-	    *sw->out.cur++ = ':';
-	}
-    }
-    dump_raw(json, strlen(json), &sw->out);
-}
-
-void
-oj_str_writer_pop(StrWriter sw) {
-    long	size;
-    DumpType	type = sw->types[sw->depth];
-
-    if (sw->keyWritten) {
-	sw->keyWritten = 0;
-	rb_raise(rb_eStandardError, "Can not pop after writing a key but no value.");
-    }
-    sw->depth--;
-    if (0 > sw->depth) {
-	rb_raise(rb_eStandardError, "Can not pop with no open array or object.");
-    }
-    size = sw->depth * sw->out.indent + 2;
-    if (sw->out.end - sw->out.cur <= size) {
-	grow(&sw->out, size);
-    }
-    fill_indent(&sw->out, sw->depth);
-    switch (type) {
-    case ObjectNew:
-    case ObjectType:
-	*sw->out.cur++ = '}';
-	break;
-    case ArrayNew:
-    case ArrayType:
-	*sw->out.cur++ = ']';
-	break;
-    }
-    if (0 == sw->depth && 0 <= sw->out.indent) {
-	*sw->out.cur++ = '\n';
+        cnt = oj_dump_float_printf(buf, sizeof(buf), obj, d, out->opts->float_fmt);
     }
+    assure_size(out, cnt);
+    APPEND_CHARS(out->cur, buf, cnt);
+    *out->cur = '\0';
 }
 
-void
-oj_str_writer_pop_all(StrWriter sw) {
-    while (0 < sw->depth) {
-	oj_str_writer_pop(sw);
+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.
+    if (17 <= cnt && (0 == strcmp("0001", buf + cnt - 4) || 0 == strcmp("9999", buf + cnt - 4))) {
+        volatile VALUE rstr = oj_safe_string_convert(obj);
+
+        strcpy(buf, RSTRING_PTR(rstr));
+        cnt = RSTRING_LEN(rstr);
     }
+    return cnt;
 }