json 2.9.1 → 2.18.1

data/ext/json/ext/generator/generator.c

@@ -1,37 +1,46 @@
-#include "ruby.h"
+#include "../json.h"
 #include "../fbuffer/fbuffer.h"
+#include "../vendor/fpconv.c"
 
 #include <math.h>
 #include <ctype.h>
 
+#include "../simd/simd.h"
+
 /* ruby api and some helpers */
 
+enum duplicate_key_action {
+    JSON_DEPRECATED = 0,
+    JSON_IGNORE,
+    JSON_RAISE,
+};
+
 typedef struct JSON_Generator_StateStruct {
     VALUE indent;
     VALUE space;
     VALUE space_before;
     VALUE object_nl;
     VALUE array_nl;
+    VALUE as_json;
 
     long max_nesting;
     long depth;
     long buffer_initial_length;
 
+    enum duplicate_key_action on_duplicate_key;
+
+    bool as_json_single_arg;
     bool allow_nan;
     bool ascii_only;
     bool script_safe;
     bool strict;
 } JSON_Generator_State;
 
-#ifndef RB_UNLIKELY
-#define RB_UNLIKELY(cond) (cond)
-#endif
-
-static VALUE mJSON, cState, mString_Extend, eGeneratorError, eNestingError, Encoding_UTF_8;
+static VALUE mJSON, cState, cFragment, eGeneratorError, eNestingError, Encoding_UTF_8;
 
-static ID i_to_s, i_to_json, i_new, i_pack, i_unpack, i_create_id, i_extend, i_encode;
-static ID sym_indent, sym_space, sym_space_before, sym_object_nl, sym_array_nl, sym_max_nesting, sym_allow_nan,
-          sym_ascii_only, sym_depth, sym_buffer_initial_length, sym_script_safe, sym_escape_slash, sym_strict;
+static ID i_to_s, i_to_json, i_new, i_encode;
+static VALUE sym_indent, sym_space, sym_space_before, sym_object_nl, sym_array_nl, sym_max_nesting, sym_allow_nan, sym_allow_duplicate_key,
+             sym_ascii_only, sym_depth, sym_buffer_initial_length, sym_script_safe, sym_escape_slash, sym_strict, sym_as_json;
 
 
 #define GET_STATE_TO(self, state) \
@@ -43,7 +52,7 @@ static ID sym_indent, sym_space, sym_space_before, sym_object_nl, sym_array_nl,
 
 struct generate_json_data;
 
-typedef void (*generator_func)(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj);
+typedef void (*generator_func)(FBuffer *buffer, struct generate_json_data *data, VALUE obj);
 
 struct generate_json_data {
     FBuffer *buffer;
@@ -51,43 +60,42 @@ struct generate_json_data {
     JSON_Generator_State *state;
     VALUE obj;
     generator_func func;
+    long depth;
 };
 
+static SIMD_Implementation simd_impl;
+
 static VALUE cState_from_state_s(VALUE self, VALUE opts);
 static VALUE cState_partial_generate(VALUE self, VALUE obj, generator_func, VALUE io);
-static void generate_json(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj);
-static void generate_json_object(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj);
-static void generate_json_array(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj);
-static void generate_json_string(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj);
-static void generate_json_null(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj);
-static void generate_json_false(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj);
-static void generate_json_true(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj);
+static void generate_json(FBuffer *buffer, struct generate_json_data *data, VALUE obj);
+static void generate_json_object(FBuffer *buffer, struct generate_json_data *data, VALUE obj);
+static void generate_json_array(FBuffer *buffer, struct generate_json_data *data, VALUE obj);
+static void generate_json_string(FBuffer *buffer, struct generate_json_data *data, VALUE obj);
+static void generate_json_null(FBuffer *buffer, struct generate_json_data *data, VALUE obj);
+static void generate_json_false(FBuffer *buffer, struct generate_json_data *data, VALUE obj);
+static void generate_json_true(FBuffer *buffer, struct generate_json_data *data, VALUE obj);
 #ifdef RUBY_INTEGER_UNIFICATION
-static void generate_json_integer(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj);
+static void generate_json_integer(FBuffer *buffer, struct generate_json_data *data, VALUE obj);
 #endif
-static void generate_json_fixnum(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj);
-static void generate_json_bignum(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj);
-static void generate_json_float(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj);
+static void generate_json_fixnum(FBuffer *buffer, struct generate_json_data *data, VALUE obj);
+static void generate_json_bignum(FBuffer *buffer, struct generate_json_data *data, VALUE obj);
+static void generate_json_float(FBuffer *buffer, struct generate_json_data *data, VALUE obj);
+static void generate_json_fragment(FBuffer *buffer, struct generate_json_data *data, VALUE obj);
 
 static int usascii_encindex, utf8_encindex, binary_encindex;
 
-#ifdef RBIMPL_ATTR_NORETURN
-RBIMPL_ATTR_NORETURN()
-#endif
-static void raise_generator_error_str(VALUE invalid_object, VALUE str)
+NORETURN(static void) raise_generator_error_str(VALUE invalid_object, VALUE str)
 {
+    rb_enc_associate_index(str, utf8_encindex);
     VALUE exc = rb_exc_new_str(eGeneratorError, str);
     rb_ivar_set(exc, rb_intern("@invalid_object"), invalid_object);
     rb_exc_raise(exc);
 }
 
-#ifdef RBIMPL_ATTR_NORETURN
-RBIMPL_ATTR_NORETURN()
-#endif
 #ifdef RBIMPL_ATTR_FORMAT
 RBIMPL_ATTR_FORMAT(RBIMPL_PRINTF_FORMAT, 2, 3)
 #endif
-static void raise_generator_error(VALUE invalid_object, const char *fmt, ...)
+NORETURN(static void) raise_generator_error(VALUE invalid_object, const char *fmt, ...)
 {
     va_list args;
     va_start(args, fmt);
@@ -96,6 +104,98 @@ static void raise_generator_error(VALUE invalid_object, const char *fmt, ...)
     raise_generator_error_str(invalid_object, str);
 }
 
+// 0 - single byte char that don't need to be escaped.
+// (x | 8) - char that needs to be escaped.
+static const unsigned char CHAR_LENGTH_MASK = 7;
+static const unsigned char ESCAPE_MASK = 8;
+
+typedef struct _search_state {
+    const char *ptr;
+    const char *end;
+    const char *cursor;
+    FBuffer *buffer;
+
+#ifdef HAVE_SIMD
+    const char *chunk_base;
+    const char *chunk_end;
+    bool has_matches;
+
+#if defined(HAVE_SIMD_NEON)
+    uint64_t matches_mask;
+#elif defined(HAVE_SIMD_SSE2)
+    int matches_mask;
+#else
+#error "Unknown SIMD Implementation."
+#endif /* HAVE_SIMD_NEON */
+#endif /* HAVE_SIMD */
+} search_state;
+
+ALWAYS_INLINE(static) void search_flush(search_state *search)
+{
+    // Do not remove this conditional without profiling, specifically escape-heavy text.
+    // escape_UTF8_char_basic will advance search->ptr and search->cursor (effectively a search_flush).
+    // For back-to-back characters that need to be escaped, specifically for the SIMD code paths, this method
+    // will be called just before calling escape_UTF8_char_basic. There will be no characters to append for the
+    // consecutive characters that need to be escaped. While the fbuffer_append is a no-op if
+    // nothing needs to be flushed, we can save a few memory references with this conditional.
+    if (search->ptr > search->cursor) {
+        fbuffer_append(search->buffer, search->cursor, search->ptr - search->cursor);
+        search->cursor = search->ptr;
+    }
+}
+
+static const unsigned char escape_table_basic[256] = {
+    // ASCII Control Characters
+     9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+     9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+    // ASCII Characters
+     0, 0, 9, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, // '"'
+     0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+     0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+     0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 9, 0, 0, 0, // '\\'
+     0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+     0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+};
+
+static inline unsigned char search_escape_basic(search_state *search)
+{
+    while (search->ptr < search->end) {
+        if (RB_UNLIKELY(escape_table_basic[(const unsigned char)*search->ptr])) {
+            search_flush(search);
+            return 1;
+        } else {
+            search->ptr++;
+        }
+    }
+    search_flush(search);
+    return 0;
+}
+
+ALWAYS_INLINE(static) void escape_UTF8_char_basic(search_state *search)
+{
+    const unsigned char ch = (unsigned char)*search->ptr;
+    switch (ch) {
+        case '"':  fbuffer_append(search->buffer, "\\\"", 2); break;
+        case '\\': fbuffer_append(search->buffer, "\\\\", 2); break;
+        case '/':  fbuffer_append(search->buffer, "\\/", 2);  break;
+        case '\b': fbuffer_append(search->buffer, "\\b", 2);  break;
+        case '\f': fbuffer_append(search->buffer, "\\f", 2);  break;
+        case '\n': fbuffer_append(search->buffer, "\\n", 2);  break;
+        case '\r': fbuffer_append(search->buffer, "\\r", 2);  break;
+        case '\t': fbuffer_append(search->buffer, "\\t", 2);  break;
+        default: {
+            const char *hexdig = "0123456789abcdef";
+            char scratch[6] = { '\\', 'u', '0', '0', 0, 0 };
+            scratch[4] = hexdig[(ch >> 4) & 0xf];
+            scratch[5] = hexdig[ch & 0xf];
+            fbuffer_append(search->buffer, scratch, 6);
+            break;
+        }
+    }
+    search->ptr++;
+    search->cursor = search->ptr;
+}
+
 /* Converts in_string to a JSON string (without the wrapping '"'
  * characters) in FBuffer out_buffer.
  *
@@ -106,282 +206,501 @@ static void raise_generator_error(VALUE invalid_object, const char *fmt, ...)
  *
  * - If out_ascii_only: non-ASCII characters (>0x7F)
  *
- * - If out_script_safe: forwardslash, line separator (U+2028), and
+ * - If script_safe: forwardslash (/), line separator (U+2028), and
  *   paragraph separator (U+2029)
  *
  * Everything else (should be UTF-8) is just passed through and
  * appended to the result.
  */
-static void convert_UTF8_to_JSON(FBuffer *out_buffer, VALUE str, const char escape_table[256], bool out_script_safe)
-{
-    const char *hexdig = "0123456789abcdef";
-    char scratch[12] = { '\\', 'u', 0, 0, 0, 0, '\\', 'u' };
 
-    const char *ptr = RSTRING_PTR(str);
-    unsigned long len = RSTRING_LEN(str);
 
-    unsigned long beg = 0, pos = 0;
+#if defined(HAVE_SIMD_NEON)
+static inline unsigned char search_escape_basic_neon(search_state *search);
+#elif defined(HAVE_SIMD_SSE2)
+static inline unsigned char search_escape_basic_sse2(search_state *search);
+#endif
 
-#define FLUSH_POS(bytes) if (pos > beg) { fbuffer_append(out_buffer, &ptr[beg], pos - beg); } pos += bytes; beg = pos;
+static inline unsigned char search_escape_basic(search_state *search);
 
-    while (pos < len) {
-        unsigned char ch = ptr[pos];
-        unsigned char ch_len = escape_table[ch];
-        /* JSON encoding */
+static inline void convert_UTF8_to_JSON(search_state *search)
+{
+#ifdef HAVE_SIMD
+#if defined(HAVE_SIMD_NEON)
+    while (search_escape_basic_neon(search)) {
+        escape_UTF8_char_basic(search);
+    }
+#elif defined(HAVE_SIMD_SSE2)
+    if (simd_impl == SIMD_SSE2) {
+        while (search_escape_basic_sse2(search)) {
+            escape_UTF8_char_basic(search);
+        }
+        return;
+    }
+    while (search_escape_basic(search)) {
+        escape_UTF8_char_basic(search);
+    }
+#endif
+#else
+    while (search_escape_basic(search)) {
+        escape_UTF8_char_basic(search);
+    }
+#endif /* HAVE_SIMD */
+}
 
-        if (RB_UNLIKELY(ch_len)) {
-            switch (ch_len) {
-                case 1: {
-                    FLUSH_POS(1);
-                    switch (ch) {
-                        case '"':  fbuffer_append(out_buffer, "\\\"", 2); break;
-                        case '\\': fbuffer_append(out_buffer, "\\\\", 2); break;
-                        case '/':  fbuffer_append(out_buffer, "\\/", 2); break;
-                        case '\b': fbuffer_append(out_buffer, "\\b", 2); break;
-                        case '\f': fbuffer_append(out_buffer, "\\f", 2); break;
-                        case '\n': fbuffer_append(out_buffer, "\\n", 2); break;
-                        case '\r': fbuffer_append(out_buffer, "\\r", 2); break;
-                        case '\t': fbuffer_append(out_buffer, "\\t", 2); break;
-                        default: {
-                            scratch[2] = '0';
-                            scratch[3] = '0';
-                            scratch[4] = hexdig[(ch >> 4) & 0xf];
-                            scratch[5] = hexdig[ch & 0xf];
-                            fbuffer_append(out_buffer, scratch, 6);
-                            break;
-                        }
-                    }
+static inline void escape_UTF8_char(search_state *search, unsigned char ch_len)
+{
+    const unsigned char ch = (unsigned char)*search->ptr;
+    switch (ch_len) {
+        case 1: {
+            switch (ch) {
+                case '"':  fbuffer_append(search->buffer, "\\\"", 2); break;
+                case '\\': fbuffer_append(search->buffer, "\\\\", 2); break;
+                case '/':  fbuffer_append(search->buffer, "\\/", 2);  break;
+                case '\b': fbuffer_append(search->buffer, "\\b", 2);  break;
+                case '\f': fbuffer_append(search->buffer, "\\f", 2);  break;
+                case '\n': fbuffer_append(search->buffer, "\\n", 2);  break;
+                case '\r': fbuffer_append(search->buffer, "\\r", 2);  break;
+                case '\t': fbuffer_append(search->buffer, "\\t", 2);  break;
+                default: {
+                    const char *hexdig = "0123456789abcdef";
+                    char scratch[6] = { '\\', 'u', '0', '0', 0, 0 };
+                    scratch[4] = hexdig[(ch >> 4) & 0xf];
+                    scratch[5] = hexdig[ch & 0xf];
+                    fbuffer_append(search->buffer, scratch, 6);
                     break;
                 }
-                case 3: {
-                    unsigned char b2 = ptr[pos + 1];
-                    if (RB_UNLIKELY(out_script_safe && ch == 0xE2 && b2 == 0x80)) {
-                        unsigned char b3 = ptr[pos + 2];
-                        if (b3 == 0xA8) {
-                            FLUSH_POS(3);
-                            fbuffer_append(out_buffer, "\\u2028", 6);
-                            break;
-                        } else if (b3 == 0xA9) {
-                            FLUSH_POS(3);
-                            fbuffer_append(out_buffer, "\\u2029", 6);
-                            break;
-                        }
-                    }
-                    // fallthrough
-                }
-                default:
-                    pos += ch_len;
-                    break;
             }
-        } else {
-            pos++;
+            break;
+        }
+        case 3: {
+            if (search->ptr[2] & 1) {
+                fbuffer_append(search->buffer, "\\u2029", 6);
+            } else {
+                fbuffer_append(search->buffer, "\\u2028", 6);
+            }
+            break;
         }
     }
-#undef FLUSH_POS
+    search->cursor = (search->ptr += ch_len);
+}
+
+#ifdef HAVE_SIMD
+
+ALWAYS_INLINE(static) char *copy_remaining_bytes(search_state *search, unsigned long vec_len, unsigned long len)
+{
+    RBIMPL_ASSERT_OR_ASSUME(len < vec_len);
+
+    // Flush the buffer so everything up until the last 'len' characters are unflushed.
+    search_flush(search);
+
+    FBuffer *buf = search->buffer;
+    fbuffer_inc_capa(buf, vec_len);
+
+    char *s = (buf->ptr + buf->len);
 
-    if (beg < len) {
-        fbuffer_append(out_buffer, &ptr[beg], len - beg);
+    // Pad the buffer with dummy characters that won't need escaping.
+    // This seem wasteful at first sight, but memset of vector length is very fast.
+    // This is a space as it can be directly represented as an immediate on AArch64.
+    memset(s, ' ', vec_len);
+
+    // Optimistically copy the remaining 'len' characters to the output FBuffer. If there are no characters
+    // to escape, then everything ends up in the correct spot. Otherwise it was convenient temporary storage.
+    if (vec_len == 16) {
+        RBIMPL_ASSERT_OR_ASSUME(len >= SIMD_MINIMUM_THRESHOLD);
+        json_fast_memcpy16(s, search->ptr, len);
+    } else {
+        MEMCPY(s, search->ptr, char, len);
     }
 
-    RB_GC_GUARD(str);
+    return s;
 }
 
-static const char escape_table[256] = {
-    // ASCII Control Characters
-    1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,
-    1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,
-    // ASCII Characters
-    0,0,1,0,0,0,0,0,0,0,0,0,0,0,0,0, // '"'
-    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
-    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
-    0,0,0,0,0,0,0,0,0,0,0,0,1,0,0,0, // '\\'
-    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
-    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
-    // Continuation byte
-    1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,
-    1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,
-    1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,
-    1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,
-    // First byte of a 2-byte code point
-    2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,
-    2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,
-    // First byte of a 4-byte code point
-    3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,
-    //First byte of a 4+byte code point
-    4,4,4,4,4,4,4,4,5,5,5,5,6,6,1,1,
-};
+#ifdef HAVE_SIMD_NEON
 
-static const char script_safe_escape_table[256] = {
-    // ASCII Control Characters
-    1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,
-    1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,
-    // ASCII Characters
-    0,0,1,0,0,0,0,0,0,0,0,0,0,0,0,1, // '"' and '/'
-    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
-    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
-    0,0,0,0,0,0,0,0,0,0,0,0,1,0,0,0, // '\\'
-    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
-    0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
-    // Continuation byte
-    1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,
-    1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,
-    1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,
-    1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,
-    // First byte of a 2-byte code point
-    2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,
-    2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,
-    // First byte of a 4-byte code point
-    3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,
-    //First byte of a 4+byte code point
-    4,4,4,4,4,4,4,4,5,5,5,5,6,6,1,1,
-};
+ALWAYS_INLINE(static) unsigned char neon_next_match(search_state *search)
+{
+    uint64_t mask = search->matches_mask;
+    uint32_t index = trailing_zeros64(mask) >> 2;
 
-static void convert_ASCII_to_JSON(FBuffer *out_buffer, VALUE str, const char escape_table[256])
+    // It is assumed escape_UTF8_char_basic will only ever increase search->ptr by at most one character.
+    // If we want to use a similar approach for full escaping we'll need to ensure:
+    //     search->chunk_base + index >= search->ptr
+    // However, since we know escape_UTF8_char_basic only increases search->ptr by one, if the next match
+    // is one byte after the previous match then:
+    //     search->chunk_base + index == search->ptr
+    search->ptr = search->chunk_base + index;
+    mask &= mask - 1;
+    search->matches_mask = mask;
+    search_flush(search);
+    return 1;
+}
+
+static inline unsigned char search_escape_basic_neon(search_state *search)
 {
-    const char *hexdig = "0123456789abcdef";
-    char scratch[12] = { '\\', 'u', 0, 0, 0, 0, '\\', 'u' };
+    if (RB_UNLIKELY(search->has_matches)) {
+        // There are more matches if search->matches_mask > 0.
+        if (search->matches_mask > 0) {
+            return neon_next_match(search);
+        } else {
+            // neon_next_match will only advance search->ptr up to the last matching character.
+            // Skip over any characters in the last chunk that occur after the last match.
+            search->has_matches = false;
+            search->ptr = search->chunk_end;
+        }
+    }
 
-    const char *ptr = RSTRING_PTR(str);
-    unsigned long len = RSTRING_LEN(str);
+    /*
+    * The code below implements an SIMD-based algorithm to determine if N bytes at a time
+    * need to be escaped.
+    *
+    * Assume the ptr = "Te\sting!" (the double quotes are included in the string)
+    *
+    * The explanation will be limited to the first 8 bytes of the string for simplicity. However
+    * the vector insructions may work on larger vectors.
+    *
+    * First, we load three constants 'lower_bound', 'backslash' and 'dblquote" in vector registers.
+    *
+    * lower_bound: [20 20 20 20 20 20 20 20]
+    * backslash:   [5C 5C 5C 5C 5C 5C 5C 5C]
+    * dblquote:    [22 22 22 22 22 22 22 22]
+    *
+    * Next we load the first chunk of the ptr:
+    * [22 54 65 5C 73 74 69 6E] ("  T  e  \  s  t  i  n)
+    *
+    * First we check if any byte in chunk is less than 32 (0x20). This returns the following vector
+    * as no bytes are less than 32 (0x20):
+    * [0 0 0 0 0 0 0 0]
+    *
+    * Next, we check if any byte in chunk is equal to a backslash:
+    * [0 0 0 FF 0 0 0 0]
+    *
+    * Finally we check if any byte in chunk is equal to a double quote:
+    * [FF 0 0 0 0 0 0 0]
+    *
+    * Now we have three vectors where each byte indicates if the corresponding byte in chunk
+    * needs to be escaped. We combine these vectors with a series of logical OR instructions.
+    * This is the needs_escape vector and it is equal to:
+    * [FF 0 0 FF 0 0 0 0]
+    *
+    * Next we compute the bitwise AND between each byte and 0x1 and compute the horizontal sum of
+    * the values in the vector. This computes how many bytes need to be escaped within this chunk.
+    *
+    * Finally we compute a mask that indicates which bytes need to be escaped. If the mask is 0 then,
+    * no bytes need to be escaped and we can continue to the next chunk. If the mask is not 0 then we
+    * have at least one byte that needs to be escaped.
+    */
+
+    if (string_scan_simd_neon(&search->ptr, search->end, &search->matches_mask)) {
+        search->has_matches = true;
+        search->chunk_base = search->ptr;
+        search->chunk_end = search->ptr + sizeof(uint8x16_t);
+        return neon_next_match(search);
+    }
 
-    unsigned long beg = 0, pos;
+    // There are fewer than 16 bytes left.
+    unsigned long remaining = (search->end - search->ptr);
+    if (remaining >= SIMD_MINIMUM_THRESHOLD) {
+        char *s = copy_remaining_bytes(search, sizeof(uint8x16_t), remaining);
 
-    for (pos = 0; pos < len;) {
-        unsigned char ch = ptr[pos];
-        /* JSON encoding */
-        if (escape_table[ch]) {
-            if (pos > beg) {
-                fbuffer_append(out_buffer, &ptr[beg], pos - beg);
-            }
+        uint64_t mask = compute_chunk_mask_neon(s);
 
-            beg = pos + 1;
-            switch (ch) {
-                case '"':  fbuffer_append(out_buffer, "\\\"", 2); break;
-                case '\\': fbuffer_append(out_buffer, "\\\\", 2); break;
-                case '/':  fbuffer_append(out_buffer, "\\/", 2); break;
-                case '\b': fbuffer_append(out_buffer, "\\b", 2); break;
-                case '\f': fbuffer_append(out_buffer, "\\f", 2); break;
-                case '\n': fbuffer_append(out_buffer, "\\n", 2); break;
-                case '\r': fbuffer_append(out_buffer, "\\r", 2); break;
-                case '\t': fbuffer_append(out_buffer, "\\t", 2); break;
-                default:
-                    scratch[2] = '0';
-                    scratch[3] = '0';
-                    scratch[4] = hexdig[(ch >> 4) & 0xf];
-                    scratch[5] = hexdig[ch & 0xf];
-                    fbuffer_append(out_buffer, scratch, 6);
-            }
+        if (!mask) {
+            // Nothing to escape, ensure search_flush doesn't do anything by setting
+            // search->cursor to search->ptr.
+            fbuffer_consumed(search->buffer, remaining);
+            search->ptr = search->end;
+            search->cursor = search->end;
+            return 0;
         }
 
-        pos++;
+        search->matches_mask = mask;
+        search->has_matches = true;
+        search->chunk_end = search->end;
+        search->chunk_base = search->ptr;
+        return neon_next_match(search);
     }
 
-    if (beg < len) {
-        fbuffer_append(out_buffer, &ptr[beg], len - beg);
+    if (search->ptr < search->end) {
+        return search_escape_basic(search);
     }
 
-    RB_GC_GUARD(str);
+    search_flush(search);
+    return 0;
+}
+#endif /* HAVE_SIMD_NEON */
+
+#ifdef HAVE_SIMD_SSE2
+
+ALWAYS_INLINE(static) unsigned char sse2_next_match(search_state *search)
+{
+    int mask = search->matches_mask;
+    int index = trailing_zeros(mask);
+
+    // It is assumed escape_UTF8_char_basic will only ever increase search->ptr by at most one character.
+    // If we want to use a similar approach for full escaping we'll need to ensure:
+    //     search->chunk_base + index >= search->ptr
+    // However, since we know escape_UTF8_char_basic only increases search->ptr by one, if the next match
+    // is one byte after the previous match then:
+    //     search->chunk_base + index == search->ptr
+    search->ptr = search->chunk_base + index;
+    mask &= mask - 1;
+    search->matches_mask = mask;
+    search_flush(search);
+    return 1;
 }
 
-static void convert_UTF8_to_ASCII_only_JSON(FBuffer *out_buffer, VALUE str, const char escape_table[256], bool out_script_safe)
+#if defined(__clang__) || defined(__GNUC__)
+#define TARGET_SSE2 __attribute__((target("sse2")))
+#else
+#define TARGET_SSE2
+#endif
+
+ALWAYS_INLINE(static) TARGET_SSE2 unsigned char search_escape_basic_sse2(search_state *search)
 {
-    const char *hexdig = "0123456789abcdef";
-    char scratch[12] = { '\\', 'u', 0, 0, 0, 0, '\\', 'u' };
+    if (RB_UNLIKELY(search->has_matches)) {
+        // There are more matches if search->matches_mask > 0.
+        if (search->matches_mask > 0) {
+            return sse2_next_match(search);
+        } else {
+            // sse2_next_match will only advance search->ptr up to the last matching character.
+            // Skip over any characters in the last chunk that occur after the last match.
+            search->has_matches = false;
+            if (RB_UNLIKELY(search->chunk_base + sizeof(__m128i) >= search->end)) {
+                search->ptr = search->end;
+            } else {
+                search->ptr = search->chunk_base + sizeof(__m128i);
+            }
+        }
+    }
 
-    const char *ptr = RSTRING_PTR(str);
-    unsigned long len = RSTRING_LEN(str);
+    if (string_scan_simd_sse2(&search->ptr, search->end, &search->matches_mask)) {
+        search->has_matches = true;
+        search->chunk_base = search->ptr;
+        search->chunk_end = search->ptr + sizeof(__m128i);
+        return sse2_next_match(search);
+    }
 
-    unsigned long beg = 0, pos = 0;
+    // There are fewer than 16 bytes left.
+    unsigned long remaining = (search->end - search->ptr);
+    if (remaining >= SIMD_MINIMUM_THRESHOLD) {
+        char *s = copy_remaining_bytes(search, sizeof(__m128i), remaining);
 
-#define FLUSH_POS(bytes) if (pos > beg) { fbuffer_append(out_buffer, &ptr[beg], pos - beg); } pos += bytes; beg = pos;
+        int needs_escape_mask = compute_chunk_mask_sse2(s);
 
-    while (pos < len) {
-        unsigned char ch = ptr[pos];
-        unsigned char ch_len = escape_table[ch];
+        if (needs_escape_mask == 0) {
+            // Nothing to escape, ensure search_flush doesn't do anything by setting
+            // search->cursor to search->ptr.
+            fbuffer_consumed(search->buffer, remaining);
+            search->ptr = search->end;
+            search->cursor = search->end;
+            return 0;
+        }
+
+        search->has_matches = true;
+        search->matches_mask = needs_escape_mask;
+        search->chunk_base = search->ptr;
+        return sse2_next_match(search);
+    }
+
+    if (search->ptr < search->end) {
+        return search_escape_basic(search);
+    }
+
+    search_flush(search);
+    return 0;
+}
+
+#endif /* HAVE_SIMD_SSE2 */
+
+#endif /* HAVE_SIMD */
+
+static const unsigned char script_safe_escape_table[256] = {
+    // ASCII Control Characters
+     9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+     9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+    // ASCII Characters
+     0, 0, 9, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 9, // '"' and '/'
+     0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+     0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+     0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 9, 0, 0, 0, // '\\'
+     0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+     0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+    // Continuation byte
+     1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+     1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+     1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+     1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+    // First byte of a 2-byte code point
+     2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
+     2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
+    // First byte of a 3-byte code point
+     3, 3,11, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, // 0xE2 is the start of \u2028 and \u2029
+    //First byte of a 4+ byte code point
+     4, 4, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 6, 6, 9, 9,
+};
+
+static inline unsigned char search_script_safe_escape(search_state *search)
+{
+    while (search->ptr < search->end) {
+        unsigned char ch = (unsigned char)*search->ptr;
+        unsigned char ch_len = script_safe_escape_table[ch];
 
         if (RB_UNLIKELY(ch_len)) {
-            switch (ch_len) {
-                case 1: {
-                    FLUSH_POS(1);
-                    switch (ch) {
-                        case '"':  fbuffer_append(out_buffer, "\\\"", 2); break;
-                        case '\\': fbuffer_append(out_buffer, "\\\\", 2); break;
-                        case '/':  fbuffer_append(out_buffer, "\\/", 2); break;
-                        case '\b': fbuffer_append(out_buffer, "\\b", 2); break;
-                        case '\f': fbuffer_append(out_buffer, "\\f", 2); break;
-                        case '\n': fbuffer_append(out_buffer, "\\n", 2); break;
-                        case '\r': fbuffer_append(out_buffer, "\\r", 2); break;
-                        case '\t': fbuffer_append(out_buffer, "\\t", 2); break;
-                        default: {
-                            scratch[2] = '0';
-                            scratch[3] = '0';
-                            scratch[4] = hexdig[(ch >> 4) & 0xf];
-                            scratch[5] = hexdig[ch & 0xf];
-                            fbuffer_append(out_buffer, scratch, 6);
-                            break;
-                        }
+            if (ch_len & ESCAPE_MASK) {
+                if (RB_UNLIKELY(ch_len == 11)) {
+                    const unsigned char *uptr = (const unsigned char *)search->ptr;
+                    if (!(uptr[1] == 0x80 && (uptr[2] >> 1) == 0x54)) {
+                        search->ptr += 3;
+                        continue;
                     }
-                    break;
                 }
-                default: {
-                    uint32_t wchar = 0;
-                    switch(ch_len) {
-                        case 2:
-                            wchar = ptr[pos] & 0x1F;
-                            break;
-                        case 3:
-                            wchar = ptr[pos] & 0x0F;
-                            break;
-                        case 4:
-                            wchar = ptr[pos] & 0x07;
-                            break;
-                    }
+                search_flush(search);
+                return ch_len & CHAR_LENGTH_MASK;
+            } else {
+                search->ptr += ch_len;
+            }
+        } else {
+            search->ptr++;
+        }
+    }
+    search_flush(search);
+    return 0;
+}
 
-                    for (short i = 1; i < ch_len; i++) {
-                        wchar = (wchar << 6) | (ptr[pos+i] & 0x3F);
-                    }
+static void convert_UTF8_to_script_safe_JSON(search_state *search)
+{
+    unsigned char ch_len;
+    while ((ch_len = search_script_safe_escape(search))) {
+        escape_UTF8_char(search, ch_len);
+    }
+}
+
+static const unsigned char ascii_only_escape_table[256] = {
+    // ASCII Control Characters
+     9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+     9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+    // ASCII Characters
+     0, 0, 9, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, // '"'
+     0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+     0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+     0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 9, 0, 0, 0, // '\\'
+     0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+     0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+    // Continuation byte
+     1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+     1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+     1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+     1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+    // First byte of a  2-byte code point
+     2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
+     2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
+    // First byte of a 3-byte code point
+     3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3,
+    //First byte of a 4+ byte code point
+     4, 4, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 6, 6, 9, 9,
+};
 
-                    FLUSH_POS(ch_len);
+static inline unsigned char search_ascii_only_escape(search_state *search, const unsigned char escape_table[256])
+{
+    while (search->ptr < search->end) {
+        unsigned char ch = (unsigned char)*search->ptr;
+        unsigned char ch_len = escape_table[ch];
 
-                    if (wchar <= 0xFFFF) {
-                        scratch[2] = hexdig[wchar >> 12];
-                        scratch[3] = hexdig[(wchar >> 8) & 0xf];
-                        scratch[4] = hexdig[(wchar >> 4) & 0xf];
-                        scratch[5] = hexdig[wchar & 0xf];
-                        fbuffer_append(out_buffer, scratch, 6);
-                    } else {
-                        uint16_t hi, lo;
-                        wchar -= 0x10000;
-                        hi = 0xD800 + (uint16_t)(wchar >> 10);
-                        lo = 0xDC00 + (uint16_t)(wchar & 0x3FF);
-
-                        scratch[2] = hexdig[hi >> 12];
-                        scratch[3] = hexdig[(hi >> 8) & 0xf];
-                        scratch[4] = hexdig[(hi >> 4) & 0xf];
-                        scratch[5] = hexdig[hi & 0xf];
-
-                        scratch[8] = hexdig[lo >> 12];
-                        scratch[9] = hexdig[(lo >> 8) & 0xf];
-                        scratch[10] = hexdig[(lo >> 4) & 0xf];
-                        scratch[11] = hexdig[lo & 0xf];
-
-                        fbuffer_append(out_buffer, scratch, 12);
-                    }
+        if (RB_UNLIKELY(ch_len)) {
+            search_flush(search);
+            return ch_len & CHAR_LENGTH_MASK;
+        } else {
+            search->ptr++;
+        }
+    }
+    search_flush(search);
+    return 0;
+}
 
+static inline void full_escape_UTF8_char(search_state *search, unsigned char ch_len)
+{
+    const unsigned char ch = (unsigned char)*search->ptr;
+    switch (ch_len) {
+        case 1: {
+            switch (ch) {
+                case '"':  fbuffer_append(search->buffer, "\\\"", 2); break;
+                case '\\': fbuffer_append(search->buffer, "\\\\", 2); break;
+                case '/':  fbuffer_append(search->buffer, "\\/", 2);  break;
+                case '\b': fbuffer_append(search->buffer, "\\b", 2);  break;
+                case '\f': fbuffer_append(search->buffer, "\\f", 2);  break;
+                case '\n': fbuffer_append(search->buffer, "\\n", 2);  break;
+                case '\r': fbuffer_append(search->buffer, "\\r", 2);  break;
+                case '\t': fbuffer_append(search->buffer, "\\t", 2);  break;
+                default: {
+                    const char *hexdig = "0123456789abcdef";
+                    char scratch[6] = { '\\', 'u', '0', '0', 0, 0 };
+                    scratch[4] = hexdig[(ch >> 4) & 0xf];
+                    scratch[5] = hexdig[ch & 0xf];
+                    fbuffer_append(search->buffer, scratch, 6);
                     break;
                 }
             }
-        } else {
-            pos++;
+            break;
         }
-    }
-#undef FLUSH_POS
+        default: {
+            const char *hexdig = "0123456789abcdef";
+            char scratch[12] = { '\\', 'u', 0, 0, 0, 0, '\\', 'u' };
+
+            uint32_t wchar = 0;
+
+            switch (ch_len) {
+                case 2:
+                    wchar = ch & 0x1F;
+                    break;
+                case 3:
+                    wchar = ch & 0x0F;
+                    break;
+                case 4:
+                    wchar = ch & 0x07;
+                    break;
+            }
+
+            for (short i = 1; i < ch_len; i++) {
+                wchar = (wchar << 6) | (search->ptr[i] & 0x3F);
+            }
+
+            if (wchar <= 0xFFFF) {
+                scratch[2] = hexdig[wchar >> 12];
+                scratch[3] = hexdig[(wchar >> 8) & 0xf];
+                scratch[4] = hexdig[(wchar >> 4) & 0xf];
+                scratch[5] = hexdig[wchar & 0xf];
+                fbuffer_append(search->buffer, scratch, 6);
+            } else {
+                uint16_t hi, lo;
+                wchar -= 0x10000;
+                hi = 0xD800 + (uint16_t)(wchar >> 10);
+                lo = 0xDC00 + (uint16_t)(wchar & 0x3FF);
+
+                scratch[2] = hexdig[hi >> 12];
+                scratch[3] = hexdig[(hi >> 8) & 0xf];
+                scratch[4] = hexdig[(hi >> 4) & 0xf];
+                scratch[5] = hexdig[hi & 0xf];
+
+                scratch[8] = hexdig[lo >> 12];
+                scratch[9] = hexdig[(lo >> 8) & 0xf];
+                scratch[10] = hexdig[(lo >> 4) & 0xf];
+                scratch[11] = hexdig[lo & 0xf];
+
+                fbuffer_append(search->buffer, scratch, 12);
+            }
 
-    if (beg < len) {
-        fbuffer_append(out_buffer, &ptr[beg], len - beg);
+            break;
+        }
     }
+    search->cursor = (search->ptr += ch_len);
+}
 
-    RB_GC_GUARD(str);
+static void convert_UTF8_to_ASCII_only_JSON(search_state *search, const unsigned char escape_table[256])
+{
+    unsigned char ch_len;
+    while ((ch_len = search_ascii_only_escape(search, escape_table))) {
+        full_escape_UTF8_char(search, ch_len);
+    }
 }
 
 /*
@@ -489,7 +808,8 @@ static VALUE mHash_to_json(int argc, VALUE *argv, VALUE self)
  * _state_ is a JSON::State object, that can also be used to configure the
  * produced JSON string output further.
  */
-static VALUE mArray_to_json(int argc, VALUE *argv, VALUE self) {
+static VALUE mArray_to_json(int argc, VALUE *argv, VALUE self)
+{
     rb_check_arity(argc, 0, 1);
     VALUE Vstate = cState_from_state_s(cState, argc == 1 ? argv[0] : Qnil);
     return cState_partial_generate(Vstate, self, generate_json_array, Qfalse);
@@ -546,17 +866,6 @@ static VALUE mFloat_to_json(int argc, VALUE *argv, VALUE self)
     return cState_partial_generate(Vstate, self, generate_json_float, Qfalse);
 }
 
-/*
- * call-seq: String.included(modul)
- *
- * Extends _modul_ with the String::Extend module.
- */
-static VALUE mString_included_s(VALUE self, VALUE modul) {
-    VALUE result = rb_funcall(modul, i_extend, 1, mString_Extend);
-    rb_call_super(1, &modul);
-    return result;
-}
-
 /*
  * call-seq: to_json(*)
  *
@@ -571,51 +880,6 @@ static VALUE mString_to_json(int argc, VALUE *argv, VALUE self)
     return cState_partial_generate(Vstate, self, generate_json_string, Qfalse);
 }
 
-/*
- * call-seq: to_json_raw_object()
- *
- * This method creates a raw object hash, that can be nested into
- * other data structures and will be generated as a raw string. This
- * method should be used, if you want to convert raw strings to JSON
- * instead of UTF-8 strings, e. g. binary data.
- */
-static VALUE mString_to_json_raw_object(VALUE self)
-{
-    VALUE ary;
-    VALUE result = rb_hash_new();
-    rb_hash_aset(result, rb_funcall(mJSON, i_create_id, 0), rb_class_name(rb_obj_class(self)));
-    ary = rb_funcall(self, i_unpack, 1, rb_str_new2("C*"));
-    rb_hash_aset(result, rb_utf8_str_new_lit("raw"), ary);
-    return result;
-}
-
-/*
- * call-seq: to_json_raw(*args)
- *
- * This method creates a JSON text from the result of a call to
- * to_json_raw_object of this String.
- */
-static VALUE mString_to_json_raw(int argc, VALUE *argv, VALUE self)
-{
-    VALUE obj = mString_to_json_raw_object(self);
-    Check_Type(obj, T_HASH);
-    return mHash_to_json(argc, argv, obj);
-}
-
-/*
- * call-seq: json_create(o)
- *
- * Raw Strings are JSON Objects (the raw bytes are stored in an array for the
- * key "raw"). The Ruby String can be created by this module method.
- */
-static VALUE mString_Extend_json_create(VALUE self, VALUE o)
-{
-    VALUE ary;
-    Check_Type(o, T_HASH);
-    ary = rb_hash_aref(o, rb_str_new2("raw"));
-    return rb_funcall(ary, i_pack, 1, rb_str_new2("C*"));
-}
-
 /*
  * call-seq: to_json(*)
  *
@@ -674,6 +938,7 @@ static void State_mark(void *ptr)
     rb_gc_mark_movable(state->space_before);
     rb_gc_mark_movable(state->object_nl);
     rb_gc_mark_movable(state->array_nl);
+    rb_gc_mark_movable(state->as_json);
 }
 
 static void State_compact(void *ptr)
@@ -684,6 +949,7 @@ static void State_compact(void *ptr)
     state->space_before = rb_gc_location(state->space_before);
     state->object_nl = rb_gc_location(state->object_nl);
     state->array_nl = rb_gc_location(state->array_nl);
+    state->as_json = rb_gc_location(state->as_json);
 }
 
 static void State_free(void *ptr)
@@ -697,11 +963,6 @@ static size_t State_memsize(const void *ptr)
     return sizeof(JSON_Generator_State);
 }
 
-#ifndef HAVE_RB_EXT_RACTOR_SAFE
-#   undef RUBY_TYPED_FROZEN_SHAREABLE
-#   define RUBY_TYPED_FROZEN_SHAREABLE 0
-#endif
-
 static const rb_data_type_t JSON_Generator_State_type = {
     "JSON/Generator/State",
     {
@@ -740,21 +1001,182 @@ static void vstate_spill(struct generate_json_data *data)
     RB_OBJ_WRITTEN(vstate, Qundef, state->space_before);
     RB_OBJ_WRITTEN(vstate, Qundef, state->object_nl);
     RB_OBJ_WRITTEN(vstate, Qundef, state->array_nl);
+    RB_OBJ_WRITTEN(vstate, Qundef, state->as_json);
 }
 
-static inline VALUE vstate_get(struct generate_json_data *data)
+static inline VALUE json_call_to_json(struct generate_json_data *data, VALUE obj)
 {
     if (RB_UNLIKELY(!data->vstate)) {
         vstate_spill(data);
     }
-    return data->vstate;
+    GET_STATE(data->vstate);
+    state->depth = data->depth;
+    VALUE tmp = rb_funcall(obj, i_to_json, 1, data->vstate);
+    // no need to restore state->depth, vstate is just a temporary State
+    return tmp;
+}
+
+static VALUE
+json_call_as_json(JSON_Generator_State *state, VALUE object, VALUE is_key)
+{
+    VALUE proc_args[2] = {object, is_key};
+    return rb_proc_call_with_block(state->as_json, 2, proc_args, Qnil);
+}
+
+static VALUE
+convert_string_subclass(VALUE key)
+{
+    VALUE key_to_s = rb_funcall(key, i_to_s, 0);
+
+    if (RB_UNLIKELY(!RB_TYPE_P(key_to_s, T_STRING))) {
+        VALUE cname = rb_obj_class(key);
+        rb_raise(rb_eTypeError,
+                 "can't convert %"PRIsVALUE" to %s (%"PRIsVALUE"#%s gives %"PRIsVALUE")",
+                 cname, "String", cname, "to_s", rb_obj_class(key_to_s));
+    }
+
+    return key_to_s;
+}
+
+static bool enc_utf8_compatible_p(int enc_idx)
+{
+    if (enc_idx == usascii_encindex) return true;
+    if (enc_idx == utf8_encindex) return true;
+    return false;
+}
+
+static VALUE encode_json_string_try(VALUE str)
+{
+    return rb_funcall(str, i_encode, 1, Encoding_UTF_8);
+}
+
+static VALUE encode_json_string_rescue(VALUE str, VALUE exception)
+{
+    raise_generator_error_str(str, rb_funcall(exception, rb_intern("message"), 0));
+    return Qundef;
+}
+
+static inline bool valid_json_string_p(VALUE str)
+{
+    int coderange = rb_enc_str_coderange(str);
+
+    if (RB_LIKELY(coderange == ENC_CODERANGE_7BIT)) {
+        return true;
+    }
+
+    if (RB_LIKELY(coderange == ENC_CODERANGE_VALID)) {
+        return enc_utf8_compatible_p(RB_ENCODING_GET_INLINED(str));
+    }
+
+    return false;
+}
+
+static inline VALUE ensure_valid_encoding(struct generate_json_data *data, VALUE str, bool as_json_called, bool is_key)
+{
+    if (RB_LIKELY(valid_json_string_p(str))) {
+        return str;
+    }
+
+    if (!as_json_called && data->state->strict && RTEST(data->state->as_json)) {
+        VALUE coerced_str = json_call_as_json(data->state, str, Qfalse);
+        if (coerced_str != str) {
+            if (RB_TYPE_P(coerced_str, T_STRING)) {
+                if (!valid_json_string_p(coerced_str)) {
+                    raise_generator_error(str, "source sequence is illegal/malformed utf-8");
+                }
+            } else {
+                // as_json could return another type than T_STRING
+                if (is_key) {
+                    raise_generator_error(coerced_str, "%"PRIsVALUE" not allowed as object key in JSON", CLASS_OF(coerced_str));
+                }
+            }
+
+            return coerced_str;
+        }
+    }
+
+    if (RB_ENCODING_GET_INLINED(str) == binary_encindex) {
+        VALUE utf8_string = rb_enc_associate_index(rb_str_dup(str), utf8_encindex);
+        switch (rb_enc_str_coderange(utf8_string)) {
+            case ENC_CODERANGE_7BIT:
+                return utf8_string;
+            case ENC_CODERANGE_VALID:
+                // For historical reason, we silently reinterpret binary strings as UTF-8 if it would work.
+                // TODO: Raise in 3.0.0
+                rb_warn("JSON.generate: UTF-8 string passed as BINARY, this will raise an encoding error in json 3.0");
+                return utf8_string;
+                break;
+        }
+    }
+
+    return rb_rescue(encode_json_string_try, str, encode_json_string_rescue, str);
+}
+
+static void raw_generate_json_string(FBuffer *buffer, struct generate_json_data *data, VALUE obj)
+{
+    fbuffer_append_char(buffer, '"');
+
+    long len;
+    search_state search;
+    search.buffer = buffer;
+    RSTRING_GETMEM(obj, search.ptr, len);
+    search.cursor = search.ptr;
+    search.end = search.ptr + len;
+
+#ifdef HAVE_SIMD
+    search.matches_mask = 0;
+    search.has_matches = false;
+    search.chunk_base = NULL;
+    search.chunk_end = NULL;
+#endif /* HAVE_SIMD */
+
+    switch (rb_enc_str_coderange(obj)) {
+        case ENC_CODERANGE_7BIT:
+        case ENC_CODERANGE_VALID:
+            if (RB_UNLIKELY(data->state->ascii_only)) {
+                convert_UTF8_to_ASCII_only_JSON(&search, data->state->script_safe ? script_safe_escape_table : ascii_only_escape_table);
+            } else if (RB_UNLIKELY(data->state->script_safe)) {
+                convert_UTF8_to_script_safe_JSON(&search);
+            } else {
+                convert_UTF8_to_JSON(&search);
+            }
+            break;
+        default:
+            raise_generator_error(obj, "source sequence is illegal/malformed utf-8");
+            break;
+    }
+    fbuffer_append_char(buffer, '"');
+}
+
+static void generate_json_string(FBuffer *buffer, struct generate_json_data *data, VALUE obj)
+{
+    obj = ensure_valid_encoding(data, obj, false, false);
+    raw_generate_json_string(buffer, data, obj);
 }
 
 struct hash_foreach_arg {
+    VALUE hash;
     struct generate_json_data *data;
-    int iter;
+    int first_key_type;
+    bool first;
+    bool mixed_keys_encountered;
 };
 
+NOINLINE(static) void
+json_inspect_hash_with_mixed_keys(struct hash_foreach_arg *arg)
+{
+    if (arg->mixed_keys_encountered) {
+        return;
+    }
+    arg->mixed_keys_encountered = true;
+
+    JSON_Generator_State *state = arg->data->state;
+    if (state->on_duplicate_key != JSON_IGNORE) {
+        VALUE do_raise = state->on_duplicate_key == JSON_RAISE ? Qtrue : Qfalse;
+        rb_funcall(mJSON, rb_intern("on_mixed_keys_hash"), 2, arg->hash, do_raise);
+    }
+}
+
 static int
 json_object_i(VALUE key, VALUE val, VALUE _arg)
 {
@@ -764,256 +1186,267 @@ json_object_i(VALUE key, VALUE val, VALUE _arg)
     FBuffer *buffer = data->buffer;
     JSON_Generator_State *state = data->state;
 
-    long depth = state->depth;
-    int j;
+    long depth = data->depth;
+    int key_type = rb_type(key);
 
-    if (arg->iter > 0) fbuffer_append_char(buffer, ',');
-    if (RB_UNLIKELY(state->object_nl)) {
-        fbuffer_append_str(buffer, state->object_nl);
+    if (arg->first) {
+        arg->first = false;
+        arg->first_key_type = key_type;
     }
-    if (RB_UNLIKELY(state->indent)) {
-        for (j = 0; j < depth; j++) {
-            fbuffer_append_str(buffer, state->indent);
-        }
+    else {
+        fbuffer_append_char(buffer, ',');
+    }
+
+    if (RB_UNLIKELY(data->state->object_nl)) {
+        fbuffer_append_str(buffer, data->state->object_nl);
+    }
+    if (RB_UNLIKELY(data->state->indent)) {
+        fbuffer_append_str_repeat(buffer, data->state->indent, depth);
     }
 
     VALUE key_to_s;
-    switch(rb_type(key)) {
+    bool as_json_called = false;
+
+  start:
+    switch (key_type) {
         case T_STRING:
+            if (RB_UNLIKELY(arg->first_key_type != T_STRING)) {
+                json_inspect_hash_with_mixed_keys(arg);
+            }
+
             if (RB_LIKELY(RBASIC_CLASS(key) == rb_cString)) {
                 key_to_s = key;
             } else {
-                key_to_s = rb_funcall(key, i_to_s, 0);
+                key_to_s = convert_string_subclass(key);
             }
             break;
         case T_SYMBOL:
+            if (RB_UNLIKELY(arg->first_key_type != T_SYMBOL)) {
+                json_inspect_hash_with_mixed_keys(arg);
+            }
+
             key_to_s = rb_sym2str(key);
             break;
         default:
+            if (data->state->strict) {
+                if (RTEST(data->state->as_json) && !as_json_called) {
+                    key = json_call_as_json(data->state, key, Qtrue);
+                    key_type = rb_type(key);
+                    as_json_called = true;
+                    goto start;
+                } else {
+                    raise_generator_error(key, "%"PRIsVALUE" not allowed as object key in JSON", CLASS_OF(key));
+                }
+            }
             key_to_s = rb_convert_type(key, T_STRING, "String", "to_s");
             break;
     }
 
+    key_to_s = ensure_valid_encoding(data, key_to_s, as_json_called, true);
+
     if (RB_LIKELY(RBASIC_CLASS(key_to_s) == rb_cString)) {
-        generate_json_string(buffer, data, state, key_to_s);
+        raw_generate_json_string(buffer, data, key_to_s);
     } else {
-        generate_json(buffer, data, state, key_to_s);
+        generate_json(buffer, data, key_to_s);
     }
-    if (RB_UNLIKELY(state->space_before)) fbuffer_append_str(buffer, state->space_before);
+    if (RB_UNLIKELY(state->space_before)) fbuffer_append_str(buffer, data->state->space_before);
     fbuffer_append_char(buffer, ':');
-    if (RB_UNLIKELY(state->space)) fbuffer_append_str(buffer, state->space);
-    generate_json(buffer, data, state, val);
+    if (RB_UNLIKELY(state->space)) fbuffer_append_str(buffer, data->state->space);
+    generate_json(buffer, data, val);
 
-    arg->iter++;
     return ST_CONTINUE;
 }
 
-static void generate_json_object(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj)
+static inline long increase_depth(struct generate_json_data *data)
 {
-    long max_nesting = state->max_nesting;
-    long depth = ++state->depth;
-    int j;
-
-    if (max_nesting != 0 && depth > max_nesting) {
-        rb_raise(eNestingError, "nesting of %ld is too deep", --state->depth);
+    JSON_Generator_State *state = data->state;
+    long depth = ++data->depth;
+    if (RB_UNLIKELY(depth > state->max_nesting && state->max_nesting)) {
+        rb_raise(eNestingError, "nesting of %ld is too deep. Did you try to serialize objects with circular references?", --data->depth);
     }
+    return depth;
+}
+
+static void generate_json_object(FBuffer *buffer, struct generate_json_data *data, VALUE obj)
+{
+    long depth = increase_depth(data);
 
     if (RHASH_SIZE(obj) == 0) {
         fbuffer_append(buffer, "{}", 2);
-        --state->depth;
+        --data->depth;
         return;
     }
 
     fbuffer_append_char(buffer, '{');
 
     struct hash_foreach_arg arg = {
+        .hash = obj,
         .data = data,
-        .iter = 0,
+        .first = true,
     };
     rb_hash_foreach(obj, json_object_i, (VALUE)&arg);
 
-    depth = --state->depth;
-    if (RB_UNLIKELY(state->object_nl)) {
-        fbuffer_append_str(buffer, state->object_nl);
-        if (RB_UNLIKELY(state->indent)) {
-            for (j = 0; j < depth; j++) {
-                fbuffer_append_str(buffer, state->indent);
-            }
+    depth = --data->depth;
+    if (RB_UNLIKELY(data->state->object_nl)) {
+        fbuffer_append_str(buffer, data->state->object_nl);
+        if (RB_UNLIKELY(data->state->indent)) {
+            fbuffer_append_str_repeat(buffer, data->state->indent, depth);
         }
     }
     fbuffer_append_char(buffer, '}');
 }
 
-static void generate_json_array(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj)
+static void generate_json_array(FBuffer *buffer, struct generate_json_data *data, VALUE obj)
 {
-    long max_nesting = state->max_nesting;
-    long depth = ++state->depth;
-    int i, j;
-    if (max_nesting != 0 && depth > max_nesting) {
-        rb_raise(eNestingError, "nesting of %ld is too deep", --state->depth);
-    }
+    long depth = increase_depth(data);
 
     if (RARRAY_LEN(obj) == 0) {
         fbuffer_append(buffer, "[]", 2);
-        --state->depth;
+        --data->depth;
         return;
     }
 
     fbuffer_append_char(buffer, '[');
-    if (RB_UNLIKELY(state->array_nl)) fbuffer_append_str(buffer, state->array_nl);
-    for(i = 0; i < RARRAY_LEN(obj); i++) {
+    if (RB_UNLIKELY(data->state->array_nl)) fbuffer_append_str(buffer, data->state->array_nl);
+    for (int i = 0; i < RARRAY_LEN(obj); i++) {
         if (i > 0) {
             fbuffer_append_char(buffer, ',');
-            if (RB_UNLIKELY(state->array_nl)) fbuffer_append_str(buffer, state->array_nl);
+            if (RB_UNLIKELY(data->state->array_nl)) fbuffer_append_str(buffer, data->state->array_nl);
         }
-        if (RB_UNLIKELY(state->indent)) {
-            for (j = 0; j < depth; j++) {
-                fbuffer_append_str(buffer, state->indent);
-            }
+        if (RB_UNLIKELY(data->state->indent)) {
+            fbuffer_append_str_repeat(buffer, data->state->indent, depth);
         }
-        generate_json(buffer, data, state, RARRAY_AREF(obj, i));
+        generate_json(buffer, data, RARRAY_AREF(obj, i));
     }
-    state->depth = --depth;
-    if (RB_UNLIKELY(state->array_nl)) {
-        fbuffer_append_str(buffer, state->array_nl);
-        if (RB_UNLIKELY(state->indent)) {
-            for (j = 0; j < depth; j++) {
-                fbuffer_append_str(buffer, state->indent);
-            }
+    data->depth = --depth;
+    if (RB_UNLIKELY(data->state->array_nl)) {
+        fbuffer_append_str(buffer, data->state->array_nl);
+        if (RB_UNLIKELY(data->state->indent)) {
+            fbuffer_append_str_repeat(buffer, data->state->indent, depth);
         }
     }
     fbuffer_append_char(buffer, ']');
 }
 
-static inline int enc_utf8_compatible_p(int enc_idx)
+static void generate_json_fallback(FBuffer *buffer, struct generate_json_data *data, VALUE obj)
 {
-    if (enc_idx == usascii_encindex) return 1;
-    if (enc_idx == utf8_encindex) return 1;
-    return 0;
-}
-
-static VALUE encode_json_string_try(VALUE str)
-{
-    return rb_funcall(str, i_encode, 1, Encoding_UTF_8);
-}
-
-static VALUE encode_json_string_rescue(VALUE str, VALUE exception)
-{
-    raise_generator_error_str(str, rb_funcall(exception, rb_intern("message"), 0));
-    return Qundef;
-}
-
-static inline VALUE ensure_valid_encoding(VALUE str)
-{
-    int encindex = RB_ENCODING_GET(str);
-    VALUE utf8_string;
-    if (RB_UNLIKELY(!enc_utf8_compatible_p(encindex))) {
-        if (encindex == binary_encindex) {
-            utf8_string = rb_enc_associate_index(rb_str_dup(str), utf8_encindex);
-            switch (rb_enc_str_coderange(utf8_string)) {
-                case ENC_CODERANGE_7BIT:
-                    return utf8_string;
-                case ENC_CODERANGE_VALID:
-                    // For historical reason, we silently reinterpret binary strings as UTF-8 if it would work.
-                    // TODO: Raise in 3.0.0
-                    rb_warn("JSON.generate: UTF-8 string passed as BINARY, this will raise an encoding error in json 3.0");
-                    return utf8_string;
-                    break;
-            }
-        }
-
-        str = rb_rescue(encode_json_string_try, str, encode_json_string_rescue, str);
+    VALUE tmp;
+    if (rb_respond_to(obj, i_to_json)) {
+        tmp = json_call_to_json(data, obj);
+        Check_Type(tmp, T_STRING);
+        fbuffer_append_str(buffer, tmp);
+    } else {
+        tmp = rb_funcall(obj, i_to_s, 0);
+        Check_Type(tmp, T_STRING);
+        generate_json_string(buffer, data, tmp);
     }
-    return str;
 }
 
-static void generate_json_string(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj)
+static inline void generate_json_symbol(FBuffer *buffer, struct generate_json_data *data, VALUE obj)
 {
-    obj = ensure_valid_encoding(obj);
-
-    fbuffer_append_char(buffer, '"');
-
-    switch(rb_enc_str_coderange(obj)) {
-        case ENC_CODERANGE_7BIT:
-            convert_ASCII_to_JSON(buffer, obj, state->script_safe ? script_safe_escape_table : escape_table);
-            break;
-        case ENC_CODERANGE_VALID:
-            if (RB_UNLIKELY(state->ascii_only)) {
-                convert_UTF8_to_ASCII_only_JSON(buffer, obj, state->script_safe ? script_safe_escape_table : escape_table, state->script_safe);
-            } else {
-                convert_UTF8_to_JSON(buffer, obj, state->script_safe ? script_safe_escape_table : escape_table, state->script_safe);
-            }
-            break;
-        default:
-            raise_generator_error(obj, "source sequence is illegal/malformed utf-8");
-            break;
+    if (data->state->strict) {
+        generate_json_string(buffer, data, rb_sym2str(obj));
+    } else {
+        generate_json_fallback(buffer, data, obj);
     }
-    fbuffer_append_char(buffer, '"');
 }
 
-static void generate_json_null(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj)
+static void generate_json_null(FBuffer *buffer, struct generate_json_data *data, VALUE obj)
 {
     fbuffer_append(buffer, "null", 4);
 }
 
-static void generate_json_false(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj)
+static void generate_json_false(FBuffer *buffer, struct generate_json_data *data, VALUE obj)
 {
     fbuffer_append(buffer, "false", 5);
 }
 
-static void generate_json_true(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj)
+static void generate_json_true(FBuffer *buffer, struct generate_json_data *data, VALUE obj)
 {
     fbuffer_append(buffer, "true", 4);
 }
 
-static void generate_json_fixnum(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj)
+static void generate_json_fixnum(FBuffer *buffer, struct generate_json_data *data, VALUE obj)
 {
     fbuffer_append_long(buffer, FIX2LONG(obj));
 }
 
-static void generate_json_bignum(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj)
+static void generate_json_bignum(FBuffer *buffer, struct generate_json_data *data, VALUE obj)
 {
     VALUE tmp = rb_funcall(obj, i_to_s, 0);
-    fbuffer_append_str(buffer, tmp);
+    fbuffer_append_str(buffer, StringValue(tmp));
 }
 
 #ifdef RUBY_INTEGER_UNIFICATION
-static void generate_json_integer(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj)
+static void generate_json_integer(FBuffer *buffer, struct generate_json_data *data, VALUE obj)
 {
     if (FIXNUM_P(obj))
-        generate_json_fixnum(buffer, data, state, obj);
+        generate_json_fixnum(buffer, data, obj);
     else
-        generate_json_bignum(buffer, data, state, obj);
+        generate_json_bignum(buffer, data, obj);
 }
 #endif
 
-static void generate_json_float(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj)
+static void generate_json_float(FBuffer *buffer, struct generate_json_data *data, VALUE obj)
 {
     double value = RFLOAT_VALUE(obj);
-    char allow_nan = state->allow_nan;
-    VALUE tmp = rb_funcall(obj, i_to_s, 0);
-    if (!allow_nan) {
-        if (isinf(value) || isnan(value)) {
-            raise_generator_error(obj, "%"PRIsVALUE" not allowed in JSON", tmp);
+    char allow_nan = data->state->allow_nan;
+    if (isinf(value) || isnan(value)) {
+        /* for NaN and Infinity values we either raise an error or rely on Float#to_s. */
+        if (!allow_nan) {
+            if (data->state->strict && data->state->as_json) {
+                VALUE casted_obj = json_call_as_json(data->state, obj, Qfalse);
+                if (casted_obj != obj) {
+                    increase_depth(data);
+                    generate_json(buffer, data, casted_obj);
+                    data->depth--;
+                    return;
+                }
+            }
+            raise_generator_error(obj, "%"PRIsVALUE" not allowed in JSON", rb_funcall(obj, i_to_s, 0));
         }
+
+        VALUE tmp = rb_funcall(obj, i_to_s, 0);
+        fbuffer_append_str(buffer, tmp);
+        return;
     }
-    fbuffer_append_str(buffer, tmp);
+
+    /* This implementation writes directly into the buffer. We reserve
+     * the 32 characters that fpconv_dtoa states as its maximum.
+     */
+    fbuffer_inc_capa(buffer, 32);
+    char* d = buffer->ptr + buffer->len;
+    int len = fpconv_dtoa(value, d);
+    /* fpconv_dtoa converts a float to its shortest string representation,
+     * but it adds a ".0" if this is a plain integer.
+     */
+    fbuffer_consumed(buffer, len);
 }
 
-static void generate_json(FBuffer *buffer, struct generate_json_data *data, JSON_Generator_State *state, VALUE obj)
+static void generate_json_fragment(FBuffer *buffer, struct generate_json_data *data, VALUE obj)
 {
-    VALUE tmp;
+    VALUE fragment = RSTRUCT_GET(obj, 0);
+    Check_Type(fragment, T_STRING);
+    fbuffer_append_str(buffer, fragment);
+}
+
+static void generate_json(FBuffer *buffer, struct generate_json_data *data, VALUE obj)
+{
+    bool as_json_called = false;
+start:
     if (obj == Qnil) {
-        generate_json_null(buffer, data, state, obj);
+        generate_json_null(buffer, data, obj);
     } else if (obj == Qfalse) {
-        generate_json_false(buffer, data, state, obj);
+        generate_json_false(buffer, data, obj);
     } else if (obj == Qtrue) {
-        generate_json_true(buffer, data, state, obj);
+        generate_json_true(buffer, data, obj);
     } else if (RB_SPECIAL_CONST_P(obj)) {
         if (RB_FIXNUM_P(obj)) {
-            generate_json_fixnum(buffer, data, state, obj);
+            generate_json_fixnum(buffer, data, obj);
         } else if (RB_FLONUM_P(obj)) {
-            generate_json_float(buffer, data, state, obj);
+            generate_json_float(buffer, data, obj);
+        } else if (RB_STATIC_SYM_P(obj)) {
+            generate_json_symbol(buffer, data, obj);
         } else {
             goto general;
         }
@@ -1021,36 +1454,52 @@ static void generate_json(FBuffer *buffer, struct generate_json_data *data, JSON
         VALUE klass = RBASIC_CLASS(obj);
         switch (RB_BUILTIN_TYPE(obj)) {
             case T_BIGNUM:
-                generate_json_bignum(buffer, data, state, obj);
+                generate_json_bignum(buffer, data, obj);
                 break;
             case T_HASH:
                 if (klass != rb_cHash) goto general;
-                generate_json_object(buffer, data, state, obj);
+                generate_json_object(buffer, data, obj);
                 break;
             case T_ARRAY:
                 if (klass != rb_cArray) goto general;
-                generate_json_array(buffer, data, state, obj);
+                generate_json_array(buffer, data, obj);
                 break;
             case T_STRING:
                 if (klass != rb_cString) goto general;
-                generate_json_string(buffer, data, state, obj);
+
+                if (RB_LIKELY(valid_json_string_p(obj))) {
+                    raw_generate_json_string(buffer, data, obj);
+                } else if (as_json_called) {
+                    raise_generator_error(obj, "source sequence is illegal/malformed utf-8");
+                } else {
+                    obj = ensure_valid_encoding(data, obj, false, false);
+                    as_json_called = true;
+                    goto start;
+                }
+                break;
+            case T_SYMBOL:
+                generate_json_symbol(buffer, data, obj);
                 break;
             case T_FLOAT:
                 if (klass != rb_cFloat) goto general;
-                generate_json_float(buffer, data, state, obj);
+                generate_json_float(buffer, data, obj);
+                break;
+            case T_STRUCT:
+                if (klass != cFragment) goto general;
+                generate_json_fragment(buffer, data, obj);
                 break;
             default:
             general:
-                if (state->strict) {
-                    raise_generator_error(obj, "%"PRIsVALUE" not allowed in JSON", CLASS_OF(obj));
-                } else if (rb_respond_to(obj, i_to_json)) {
-                    tmp = rb_funcall(obj, i_to_json, 1, vstate_get(data));
-                    Check_Type(tmp, T_STRING);
-                    fbuffer_append_str(buffer, tmp);
+                if (data->state->strict) {
+                    if (RTEST(data->state->as_json) && !as_json_called) {
+                        obj = json_call_as_json(data->state, obj, Qfalse);
+                        as_json_called = true;
+                        goto start;
+                    } else {
+                        raise_generator_error(obj, "%"PRIsVALUE" not allowed in JSON", CLASS_OF(obj));
+                    }
                 } else {
-                    tmp = rb_funcall(obj, i_to_s, 0);
-                    Check_Type(tmp, T_STRING);
-                    generate_json_string(buffer, data, state, tmp);
+                    generate_json_fallback(buffer, data, obj);
                 }
         }
     }
@@ -1060,18 +1509,16 @@ static VALUE generate_json_try(VALUE d)
 {
     struct generate_json_data *data = (struct generate_json_data *)d;
 
-    data->func(data->buffer, data, data->state, data->obj);
+    data->func(data->buffer, data, data->obj);
 
-    return Qnil;
+    return fbuffer_finalize(data->buffer);
 }
 
-static VALUE generate_json_rescue(VALUE d, VALUE exc)
+static VALUE generate_json_ensure(VALUE d)
 {
     struct generate_json_data *data = (struct generate_json_data *)d;
     fbuffer_free(data->buffer);
 
-    rb_exc_raise(exc);
-
     return Qundef;
 }
 
@@ -1087,22 +1534,31 @@ static VALUE cState_partial_generate(VALUE self, VALUE obj, generator_func func,
 
     struct generate_json_data data = {
         .buffer = &buffer,
-        .vstate = self,
+        .vstate = Qfalse, // don't use self as it may be frozen and its depth is mutated when calling to_json
         .state = state,
+        .depth = state->depth,
         .obj = obj,
         .func = func
     };
-    rb_rescue(generate_json_try, (VALUE)&data, generate_json_rescue, (VALUE)&data);
-
-    return fbuffer_finalize(&buffer);
+    VALUE result = rb_ensure(generate_json_try, (VALUE)&data, generate_json_ensure, (VALUE)&data);
+    RB_GC_GUARD(self);
+    return result;
 }
 
-static VALUE cState_generate(VALUE self, VALUE obj, VALUE io)
+/* call-seq:
+ *   generate(obj) -> String
+ *   generate(obj, anIO) -> anIO
+ *
+ * Generates a valid JSON document from object +obj+ and returns the
+ * result. If no valid JSON document can be created this method raises a
+ * GeneratorError exception.
+ */
+static VALUE cState_generate(int argc, VALUE *argv, VALUE self)
 {
-    VALUE result = cState_partial_generate(self, obj, generate_json, io);
-    GET_STATE(self);
-    (void)state;
-    return result;
+    rb_check_arity(argc, 1, 2);
+    VALUE obj = argv[0];
+    VALUE io = argc > 1 ? argv[1] : Qnil;
+    return cState_partial_generate(self, obj, generate_json, io);
 }
 
 static VALUE cState_initialize(int argc, VALUE *argv, VALUE self)
@@ -1132,6 +1588,7 @@ static VALUE cState_init_copy(VALUE obj, VALUE orig)
     objState->space_before = origState->space_before;
     objState->object_nl = origState->object_nl;
     objState->array_nl = origState->array_nl;
+    objState->as_json = origState->as_json;
     return obj;
 }
 
@@ -1182,6 +1639,7 @@ static VALUE string_config(VALUE config)
  */
 static VALUE cState_indent_set(VALUE self, VALUE indent)
 {
+    rb_check_frozen(self);
     GET_STATE(self);
     RB_OBJ_WRITE(self, &state->indent, string_config(indent));
     return Qnil;
@@ -1207,6 +1665,7 @@ static VALUE cState_space(VALUE self)
  */
 static VALUE cState_space_set(VALUE self, VALUE space)
 {
+    rb_check_frozen(self);
     GET_STATE(self);
     RB_OBJ_WRITE(self, &state->space, string_config(space));
     return Qnil;
@@ -1230,6 +1689,7 @@ static VALUE cState_space_before(VALUE self)
  */
 static VALUE cState_space_before_set(VALUE self, VALUE space_before)
 {
+    rb_check_frozen(self);
     GET_STATE(self);
     RB_OBJ_WRITE(self, &state->space_before, string_config(space_before));
     return Qnil;
@@ -1255,6 +1715,7 @@ static VALUE cState_object_nl(VALUE self)
  */
 static VALUE cState_object_nl_set(VALUE self, VALUE object_nl)
 {
+    rb_check_frozen(self);
     GET_STATE(self);
     RB_OBJ_WRITE(self, &state->object_nl, string_config(object_nl));
     return Qnil;
@@ -1278,11 +1739,35 @@ static VALUE cState_array_nl(VALUE self)
  */
 static VALUE cState_array_nl_set(VALUE self, VALUE array_nl)
 {
+    rb_check_frozen(self);
     GET_STATE(self);
     RB_OBJ_WRITE(self, &state->array_nl, string_config(array_nl));
     return Qnil;
 }
 
+/*
+ * call-seq: as_json()
+ *
+ * This string is put at the end of a line that holds a JSON array.
+ */
+static VALUE cState_as_json(VALUE self)
+{
+    GET_STATE(self);
+    return state->as_json;
+}
+
+/*
+ * call-seq: as_json=(as_json)
+ *
+ * This string is put at the end of a line that holds a JSON array.
+ */
+static VALUE cState_as_json_set(VALUE self, VALUE as_json)
+{
+    rb_check_frozen(self);
+    GET_STATE(self);
+    RB_OBJ_WRITE(self, &state->as_json, rb_convert_type(as_json, T_DATA, "Proc", "to_proc"));
+    return Qnil;
+}
 
 /*
 * call-seq: check_circular?
@@ -1321,6 +1806,7 @@ static long long_config(VALUE num)
  */
 static VALUE cState_max_nesting_set(VALUE self, VALUE depth)
 {
+    rb_check_frozen(self);
     GET_STATE(self);
     state->max_nesting = long_config(depth);
     return Qnil;
@@ -1346,6 +1832,7 @@ static VALUE cState_script_safe(VALUE self)
  */
 static VALUE cState_script_safe_set(VALUE self, VALUE enable)
 {
+    rb_check_frozen(self);
     GET_STATE(self);
     state->script_safe = RTEST(enable);
     return Qnil;
@@ -1377,6 +1864,7 @@ static VALUE cState_strict(VALUE self)
  */
 static VALUE cState_strict_set(VALUE self, VALUE enable)
 {
+    rb_check_frozen(self);
     GET_STATE(self);
     state->strict = RTEST(enable);
     return Qnil;
@@ -1401,6 +1889,7 @@ static VALUE cState_allow_nan_p(VALUE self)
  */
 static VALUE cState_allow_nan_set(VALUE self, VALUE enable)
 {
+    rb_check_frozen(self);
     GET_STATE(self);
     state->allow_nan = RTEST(enable);
     return Qnil;
@@ -1425,11 +1914,25 @@ static VALUE cState_ascii_only_p(VALUE self)
  */
 static VALUE cState_ascii_only_set(VALUE self, VALUE enable)
 {
+    rb_check_frozen(self);
     GET_STATE(self);
     state->ascii_only = RTEST(enable);
     return Qnil;
 }
 
+static VALUE cState_allow_duplicate_key_p(VALUE self)
+{
+    GET_STATE(self);
+    switch (state->on_duplicate_key) {
+        case JSON_IGNORE:
+            return Qtrue;
+        case JSON_DEPRECATED:
+            return Qnil;
+        default:
+            return Qfalse;
+    }
+}
+
 /*
  * call-seq: depth
  *
@@ -1449,6 +1952,7 @@ static VALUE cState_depth(VALUE self)
  */
 static VALUE cState_depth_set(VALUE self, VALUE depth)
 {
+    rb_check_frozen(self);
     GET_STATE(self);
     state->depth = long_config(depth);
     return Qnil;
@@ -1482,20 +1986,36 @@ static void buffer_initial_length_set(JSON_Generator_State *state, VALUE buffer_
  */
 static VALUE cState_buffer_initial_length_set(VALUE self, VALUE buffer_initial_length)
 {
+    rb_check_frozen(self);
     GET_STATE(self);
     buffer_initial_length_set(state, buffer_initial_length);
     return Qnil;
 }
 
+struct configure_state_data {
+    JSON_Generator_State *state;
+    VALUE vstate;  // Ruby object that owns the state, or Qfalse if stack-allocated
+};
+
+static inline void state_write_value(struct configure_state_data *data, VALUE *field, VALUE value)
+{
+    if (RTEST(data->vstate)) {
+        RB_OBJ_WRITE(data->vstate, field, value);
+    } else {
+        *field = value;
+    }
+}
+
 static int configure_state_i(VALUE key, VALUE val, VALUE _arg)
 {
-    JSON_Generator_State *state = (JSON_Generator_State *)_arg;
+    struct configure_state_data *data = (struct configure_state_data *)_arg;
+    JSON_Generator_State *state = data->state;
 
-         if (key == sym_indent)                { state->indent = string_config(val); }
-    else if (key == sym_space)                 { state->space = string_config(val); }
-    else if (key == sym_space_before)          { state->space_before = string_config(val); }
-    else if (key == sym_object_nl)             { state->object_nl = string_config(val); }
-    else if (key == sym_array_nl)              { state->array_nl = string_config(val); }
+         if (key == sym_indent)                { state_write_value(data, &state->indent, string_config(val)); }
+    else if (key == sym_space)                 { state_write_value(data, &state->space, string_config(val)); }
+    else if (key == sym_space_before)          { state_write_value(data, &state->space_before, string_config(val)); }
+    else if (key == sym_object_nl)             { state_write_value(data, &state->object_nl, string_config(val)); }
+    else if (key == sym_array_nl)              { state_write_value(data, &state->array_nl, string_config(val)); }
     else if (key == sym_max_nesting)           { state->max_nesting = long_config(val); }
     else if (key == sym_allow_nan)             { state->allow_nan = RTEST(val); }
     else if (key == sym_ascii_only)            { state->ascii_only = RTEST(val); }
@@ -1504,10 +2024,16 @@ static int configure_state_i(VALUE key, VALUE val, VALUE _arg)
     else if (key == sym_script_safe)           { state->script_safe = RTEST(val); }
     else if (key == sym_escape_slash)          { state->script_safe = RTEST(val); }
     else if (key == sym_strict)                { state->strict = RTEST(val); }
+    else if (key == sym_allow_duplicate_key)   { state->on_duplicate_key = RTEST(val) ? JSON_IGNORE : JSON_RAISE; }
+    else if (key == sym_as_json)               {
+        VALUE proc = RTEST(val) ? rb_convert_type(val, T_DATA, "Proc", "to_proc") : Qfalse;
+        state->as_json_single_arg = proc && rb_proc_arity(proc) == 1;
+        state_write_value(data, &state->as_json, proc);
+    }
     return ST_CONTINUE;
 }
 
-static void configure_state(JSON_Generator_State *state, VALUE config)
+static void configure_state(JSON_Generator_State *state, VALUE vstate, VALUE config)
 {
     if (!RTEST(config)) return;
 
@@ -1515,15 +2041,21 @@ static void configure_state(JSON_Generator_State *state, VALUE config)
 
     if (!RHASH_SIZE(config)) return;
 
+    struct configure_state_data data = {
+        .state = state,
+        .vstate = vstate
+    };
+
     // We assume in most cases few keys are set so it's faster to go over
     // the provided keys than to check all possible keys.
-    rb_hash_foreach(config, configure_state_i, (VALUE)state);
+    rb_hash_foreach(config, configure_state_i, (VALUE)&data);
 }
 
 static VALUE cState_configure(VALUE self, VALUE opts)
 {
+    rb_check_frozen(self);
     GET_STATE(self);
-    configure_state(state, opts);
+    configure_state(state, self, opts);
     return self;
 }
 
@@ -1531,7 +2063,7 @@ static VALUE cState_m_generate(VALUE klass, VALUE obj, VALUE opts, VALUE io)
 {
     JSON_Generator_State state = {0};
     state_init(&state);
-    configure_state(&state, opts);
+    configure_state(&state, Qfalse, opts);
 
     char stack_buffer[FBUFFER_STACK_SIZE];
     FBuffer buffer = {
@@ -1543,12 +2075,11 @@ static VALUE cState_m_generate(VALUE klass, VALUE obj, VALUE opts, VALUE io)
         .buffer = &buffer,
         .vstate = Qfalse,
         .state = &state,
+        .depth = state.depth,
         .obj = obj,
         .func = generate_json,
     };
-    rb_rescue(generate_json_try, (VALUE)&data, generate_json_rescue, (VALUE)&data);
-
-    return fbuffer_finalize(&buffer);
+    return rb_ensure(generate_json_try, (VALUE)&data, generate_json_ensure, (VALUE)&data);
 }
 
 /*
@@ -1564,6 +2095,10 @@ void Init_generator(void)
     rb_require("json/common");
 
     mJSON = rb_define_module("JSON");
+
+    rb_global_variable(&cFragment);
+    cFragment = rb_const_get(mJSON, rb_intern("Fragment"));
+
     VALUE mExt = rb_define_module_under(mJSON, "Ext");
     VALUE mGenerator = rb_define_module_under(mExt, "Generator");
 
@@ -1591,6 +2126,8 @@ void Init_generator(void)
     rb_define_method(cState, "object_nl=", cState_object_nl_set, 1);
     rb_define_method(cState, "array_nl", cState_array_nl, 0);
     rb_define_method(cState, "array_nl=", cState_array_nl_set, 1);
+    rb_define_method(cState, "as_json", cState_as_json, 0);
+    rb_define_method(cState, "as_json=", cState_as_json_set, 1);
     rb_define_method(cState, "max_nesting", cState_max_nesting, 0);
     rb_define_method(cState, "max_nesting=", cState_max_nesting_set, 1);
     rb_define_method(cState, "script_safe", cState_script_safe, 0);
@@ -1611,7 +2148,9 @@ void Init_generator(void)
     rb_define_method(cState, "depth=", cState_depth_set, 1);
     rb_define_method(cState, "buffer_initial_length", cState_buffer_initial_length, 0);
     rb_define_method(cState, "buffer_initial_length=", cState_buffer_initial_length_set, 1);
-    rb_define_private_method(cState, "_generate", cState_generate, 2);
+    rb_define_method(cState, "generate", cState_generate, -1);
+
+    rb_define_private_method(cState, "allow_duplicate_key?", cState_allow_duplicate_key_p, 0);
 
     rb_define_singleton_method(cState, "generate", cState_m_generate, 3);
 
@@ -1640,13 +2179,7 @@ void Init_generator(void)
     rb_define_method(mFloat, "to_json", mFloat_to_json, -1);
 
     VALUE mString = rb_define_module_under(mGeneratorMethods, "String");
-    rb_define_singleton_method(mString, "included", mString_included_s, 1);
     rb_define_method(mString, "to_json", mString_to_json, -1);
-    rb_define_method(mString, "to_json_raw", mString_to_json_raw, -1);
-    rb_define_method(mString, "to_json_raw_object", mString_to_json_raw_object, 0);
-
-    mString_Extend = rb_define_module_under(mString, "Extend");
-    rb_define_method(mString_Extend, "json_create", mString_Extend_json_create, 1);
 
     VALUE mTrueClass = rb_define_module_under(mGeneratorMethods, "TrueClass");
     rb_define_method(mTrueClass, "to_json", mTrueClass_to_json, -1);
@@ -1663,10 +2196,6 @@ void Init_generator(void)
     i_to_s = rb_intern("to_s");
     i_to_json = rb_intern("to_json");
     i_new = rb_intern("new");
-    i_pack = rb_intern("pack");
-    i_unpack = rb_intern("unpack");
-    i_create_id = rb_intern("create_id");
-    i_extend = rb_intern("extend");
     i_encode = rb_intern("encode");
 
     sym_indent = ID2SYM(rb_intern("indent"));
@@ -1682,10 +2211,14 @@ void Init_generator(void)
     sym_script_safe = ID2SYM(rb_intern("script_safe"));
     sym_escape_slash = ID2SYM(rb_intern("escape_slash"));
     sym_strict = ID2SYM(rb_intern("strict"));
+    sym_as_json = ID2SYM(rb_intern("as_json"));
+    sym_allow_duplicate_key = ID2SYM(rb_intern("allow_duplicate_key"));
 
     usascii_encindex = rb_usascii_encindex();
     utf8_encindex = rb_utf8_encindex();
     binary_encindex = rb_ascii8bit_encindex();
 
     rb_require("json/ext/generator/state");
+
+    simd_impl = find_simd_implementation();
 }