json 2.12.2 → 2.15.0

data/ext/json/ext/parser/extconf.rb

@@ -1,11 +1,15 @@
 # frozen_string_literal: true
 require 'mkmf'
 
-have_func("rb_enc_interned_str", "ruby.h") # RUBY_VERSION >= 3.0
+have_func("rb_enc_interned_str", "ruby/encoding.h") # RUBY_VERSION >= 3.0
 have_func("rb_hash_new_capa", "ruby.h") # RUBY_VERSION >= 3.2
 have_func("rb_hash_bulk_insert", "ruby.h") # Missing on TruffleRuby
 have_func("strnlen", "string.h") # Missing on Solaris 10
 
 append_cflags("-std=c99")
 
+if enable_config('parser-use-simd', default=!ENV["JSON_DISABLE_SIMD"])
+  load __dir__ + "/../simd/conf.rb"
+end
+
 create_makefile 'json/ext/parser'

data/ext/json/ext/parser/parser.c

@@ -20,6 +20,8 @@ typedef unsigned char _Bool;
 #endif
 #endif
 
+#include "../simd/simd.h"
+
 #ifndef RB_UNLIKELY
 #define RB_UNLIKELY(expr) expr
 #endif
@@ -35,7 +37,7 @@ static ID i_chr, i_aset, i_aref,
           i_leftshift, i_new, i_try_convert, i_uminus, i_encode;
 
 static VALUE sym_max_nesting, sym_allow_nan, sym_allow_trailing_comma, sym_symbolize_names, sym_freeze,
-             sym_decimal_class, sym_on_load;
+             sym_decimal_class, sym_on_load, sym_allow_duplicate_key;
 
 static int binary_encindex;
 static int utf8_encindex;
@@ -363,10 +365,17 @@ static int convert_UTF32_to_UTF8(char *buf, uint32_t ch)
     return len;
 }
 
+enum duplicate_key_action {
+    JSON_DEPRECATED = 0,
+    JSON_IGNORE,
+    JSON_RAISE,
+};
+
 typedef struct JSON_ParserStruct {
     VALUE on_load_proc;
     VALUE decimal_class;
     ID decimal_method_id;
+    enum duplicate_key_action on_duplicate_key;
     int max_nesting;
     bool allow_nan;
     bool allow_trailing_comma;
@@ -386,15 +395,8 @@ typedef struct JSON_ParserStateStruct {
     int current_nesting;
 } JSON_ParserState;
 
-
-#define PARSE_ERROR_FRAGMENT_LEN 32
-#ifdef RBIMPL_ATTR_NORETURN
-RBIMPL_ATTR_NORETURN()
-#endif
-static void raise_parse_error(const char *format, JSON_ParserState *state)
+static void cursor_position(JSON_ParserState *state, long *line_out, long *column_out)
 {
-    unsigned char buffer[PARSE_ERROR_FRAGMENT_LEN + 3];
-
     const char *cursor = state->cursor;
     long column = 0;
     long line = 1;
@@ -411,6 +413,29 @@ static void raise_parse_error(const char *format, JSON_ParserState *state)
             line++;
         }
     }
+    *line_out = line;
+    *column_out = column;
+}
+
+static void emit_parse_warning(const char *message, JSON_ParserState *state)
+{
+    long line, column;
+    cursor_position(state, &line, &column);
+
+    VALUE warning = rb_sprintf("%s at line %ld column %ld", message, line, column);
+    rb_funcall(mJSON, rb_intern("deprecation_warning"), 1, warning);
+}
+
+#define PARSE_ERROR_FRAGMENT_LEN 32
+
+#ifdef RBIMPL_ATTR_NORETURN
+RBIMPL_ATTR_NORETURN()
+#endif
+static void raise_parse_error(const char *format, JSON_ParserState *state)
+{
+    unsigned char buffer[PARSE_ERROR_FRAGMENT_LEN + 3];
+    long line, column;
+    cursor_position(state, &line, &column);
 
     const char *ptr = "EOF";
     if (state->cursor && state->cursor < state->end) {
@@ -517,7 +542,7 @@ static void
 json_eat_comments(JSON_ParserState *state)
 {
     if (state->cursor + 1 < state->end) {
-        switch(state->cursor[1]) {
+        switch (state->cursor[1]) {
             case '/': {
                 state->cursor = memchr(state->cursor, '\n', state->end - state->cursor);
                 if (!state->cursor) {
@@ -688,11 +713,16 @@ static VALUE json_string_unescape(JSON_ParserState *state, const char *string, c
                         }
                         if (pe[0] == '\\' && pe[1] == 'u') {
                             uint32_t sur = unescape_unicode(state, (unsigned char *) pe + 2);
+
+                            if ((sur & 0xFC00) != 0xDC00) {
+                                raise_parse_error_at("invalid surrogate pair at %s", state, p);
+                            }
+
                             ch = (((ch & 0x3F) << 10) | ((((ch >> 6) & 0xF) + 1) << 16)
                                     | (sur & 0x3FF));
                             pe += 5;
                         } else {
-                            unescape = (char *) "?";
+                            raise_parse_error_at("incomplete surrogate pair at %s", state, p);
                             break;
                         }
                     }
@@ -807,10 +837,67 @@ static inline VALUE json_decode_array(JSON_ParserState *state, JSON_ParserConfig
     return array;
 }
 
-static inline VALUE json_decode_object(JSON_ParserState *state, JSON_ParserConfig *config, long count)
+static VALUE json_find_duplicated_key(size_t count, const VALUE *pairs)
+{
+    VALUE set = rb_hash_new_capa(count / 2);
+    for (size_t index = 0; index < count; index += 2) {
+        size_t before = RHASH_SIZE(set);
+        VALUE key = pairs[index];
+        rb_hash_aset(set, key, Qtrue);
+        if (RHASH_SIZE(set) == before) {
+            if (RB_SYMBOL_P(key)) {
+                return rb_sym2str(key);
+            }
+            return key;
+        }
+    }
+    return Qfalse;
+}
+
+static void emit_duplicate_key_warning(JSON_ParserState *state, VALUE duplicate_key)
 {
-    VALUE object = rb_hash_new_capa(count);
-    rb_hash_bulk_insert(count, rvalue_stack_peek(state->stack, count), object);
+    VALUE message = rb_sprintf(
+        "detected duplicate key %"PRIsVALUE" in JSON object. This will raise an error in json 3.0 unless enabled via `allow_duplicate_key: true`",
+        rb_inspect(duplicate_key)
+    );
+
+    emit_parse_warning(RSTRING_PTR(message), state);
+    RB_GC_GUARD(message);
+}
+
+#ifdef RBIMPL_ATTR_NORETURN
+RBIMPL_ATTR_NORETURN()
+#endif
+static void raise_duplicate_key_error(JSON_ParserState *state, VALUE duplicate_key)
+{
+    VALUE message = rb_sprintf(
+        "duplicate key %"PRIsVALUE,
+        rb_inspect(duplicate_key)
+    );
+
+    raise_parse_error(RSTRING_PTR(message), state);
+    RB_GC_GUARD(message);
+}
+
+static inline VALUE json_decode_object(JSON_ParserState *state, JSON_ParserConfig *config, size_t count)
+{
+    size_t entries_count = count / 2;
+    VALUE object = rb_hash_new_capa(entries_count);
+    const VALUE *pairs = rvalue_stack_peek(state->stack, count);
+    rb_hash_bulk_insert(count, pairs, object);
+
+    if (RB_UNLIKELY(RHASH_SIZE(object) < entries_count)) {
+        switch (config->on_duplicate_key) {
+            case JSON_IGNORE:
+                break;
+            case JSON_DEPRECATED:
+                emit_duplicate_key_warning(state, json_find_duplicated_key(count, pairs));
+                break;
+            case JSON_RAISE:
+                raise_duplicate_key_error(state, json_find_duplicated_key(count, pairs));
+                break;
+        }
+    }
 
     rvalue_stack_pop(state->stack, count);
 
@@ -844,7 +931,7 @@ static inline VALUE json_push_value(JSON_ParserState *state, JSON_ParserConfig *
     return value;
 }
 
-static const bool string_scan[256] = {
+static const bool string_scan_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,
@@ -857,32 +944,71 @@ static const bool string_scan[256] = {
      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
 };
 
+#if (defined(__GNUC__ ) || defined(__clang__))
+#define FORCE_INLINE __attribute__((always_inline))
+#else
+#define FORCE_INLINE
+#endif
+
+#ifdef HAVE_SIMD
+static SIMD_Implementation simd_impl = SIMD_NONE;
+#endif /* HAVE_SIMD */
+
+static inline bool FORCE_INLINE string_scan(JSON_ParserState *state)
+{
+#ifdef HAVE_SIMD
+#if defined(HAVE_SIMD_NEON)
+
+    uint64_t mask = 0;
+    if (string_scan_simd_neon(&state->cursor, state->end, &mask)) {
+        state->cursor += trailing_zeros64(mask) >> 2;
+        return 1;
+    }
+
+#elif defined(HAVE_SIMD_SSE2)
+    if (simd_impl == SIMD_SSE2) {
+        int mask = 0;
+        if (string_scan_simd_sse2(&state->cursor, state->end, &mask)) {
+            state->cursor += trailing_zeros(mask);
+            return 1;
+        }
+    }
+#endif /* HAVE_SIMD_NEON or HAVE_SIMD_SSE2 */
+#endif /* HAVE_SIMD */
+
+    while (state->cursor < state->end) {
+        if (RB_UNLIKELY(string_scan_table[(unsigned char)*state->cursor])) {
+            return 1;
+        }
+        state->cursor++;
+    }
+    return 0;
+}
+
 static inline VALUE json_parse_string(JSON_ParserState *state, JSON_ParserConfig *config, bool is_name)
 {
     state->cursor++;
     const char *start = state->cursor;
     bool escaped = false;
 
-    while (state->cursor < state->end) {
-        if (RB_UNLIKELY(string_scan[(unsigned char)*state->cursor])) {
-            switch (*state->cursor) {
-                case '"': {
-                    VALUE string = json_decode_string(state, config, start, state->cursor, escaped, is_name);
-                    state->cursor++;
-                    return json_push_value(state, config, string);
-                }
-                case '\\': {
-                    state->cursor++;
-                    escaped = true;
-                    if ((unsigned char)*state->cursor < 0x20) {
-                        raise_parse_error("invalid ASCII control character in string: %s", state);
-                    }
-                    break;
-                }
-                default:
+    while (RB_UNLIKELY(string_scan(state))) {
+        switch (*state->cursor) {
+            case '"': {
+                VALUE string = json_decode_string(state, config, start, state->cursor, escaped, is_name);
+                state->cursor++;
+                return json_push_value(state, config, string);
+            }
+            case '\\': {
+                state->cursor++;
+                escaped = true;
+                if ((unsigned char)*state->cursor < 0x20) {
                     raise_parse_error("invalid ASCII control character in string: %s", state);
-                    break;
+                }
+                break;
             }
+            default:
+                raise_parse_error("invalid ASCII control character in string: %s", state);
+                break;
         }
 
         state->cursor++;
@@ -1060,6 +1186,8 @@ static VALUE json_parse_any(JSON_ParserState *state, JSON_ParserConfig *config)
             break;
         }
         case '{': {
+            const char *object_start_cursor = state->cursor;
+
             state->cursor++;
             json_eat_whitespace(state);
             long stack_head = state->stack->head;
@@ -1094,8 +1222,15 @@ static VALUE json_parse_any(JSON_ParserState *state, JSON_ParserConfig *config)
                     if (*state->cursor == '}') {
                         state->cursor++;
                         state->current_nesting--;
-                        long count = state->stack->head - stack_head;
-                        return json_push_value(state, config, json_decode_object(state, config, count));
+                        size_t count = state->stack->head - stack_head;
+
+                        // Temporary rewind cursor in case an error is raised
+                        const char *final_cursor = state->cursor;
+                        state->cursor = object_start_cursor;
+                        VALUE object = json_decode_object(state, config, count);
+                        state->cursor = final_cursor;
+
+                        return json_push_value(state, config, object);
                     }
 
                     if (*state->cursor == ',') {
@@ -1135,7 +1270,7 @@ static VALUE json_parse_any(JSON_ParserState *state, JSON_ParserConfig *config)
             break;
     }
 
-    raise_parse_error("unreacheable: %s", state);
+    raise_parse_error("unreachable: %s", state);
 }
 
 static void json_ensure_eof(JSON_ParserState *state)
@@ -1184,6 +1319,7 @@ static int parser_config_init_i(VALUE key, VALUE val, VALUE data)
     else if (key == sym_symbolize_names)      { config->symbolize_names = RTEST(val); }
     else if (key == sym_freeze)               { config->freeze = RTEST(val); }
     else if (key == sym_on_load)              { config->on_load_proc = RTEST(val) ? val : Qfalse; }
+    else if (key == sym_allow_duplicate_key)  { config->on_duplicate_key = RTEST(val) ? JSON_IGNORE : JSON_RAISE; }
     else if (key == sym_decimal_class)        {
         if (RTEST(val)) {
             if (rb_respond_to(val, i_try_convert)) {
@@ -1400,6 +1536,7 @@ void Init_parser(void)
     sym_freeze = ID2SYM(rb_intern("freeze"));
     sym_on_load = ID2SYM(rb_intern("on_load"));
     sym_decimal_class = ID2SYM(rb_intern("decimal_class"));
+    sym_allow_duplicate_key = ID2SYM(rb_intern("allow_duplicate_key"));
 
     i_chr = rb_intern("chr");
     i_aset = rb_intern("[]=");
@@ -1413,4 +1550,8 @@ void Init_parser(void)
     binary_encindex = rb_ascii8bit_encindex();
     utf8_encindex = rb_utf8_encindex();
     enc_utf8 = rb_utf8_encoding();
+
+#ifdef HAVE_SIMD
+    simd_impl = find_simd_implementation();
+#endif
 }

data/ext/json/ext/simd/conf.rb

@@ -0,0 +1,24 @@
+case RbConfig::CONFIG['host_cpu']
+when /^(arm|aarch64)/
+  # Try to compile a small program using NEON instructions
+  header, type, init, extra = 'arm_neon.h', 'uint8x16_t', 'vdupq_n_u8(32)', nil
+when /^(x86_64|x64)/
+  header, type, init, extra = 'x86intrin.h', '__m128i', '_mm_set1_epi8(32)', 'if (__builtin_cpu_supports("sse2")) { printf("OK"); }'
+end
+if header
+  if have_header(header) && try_compile(<<~SRC, '-Werror=implicit-function-declaration')
+      #{cpp_include(header)}
+      int main(int argc, char **argv) {
+        #{type} test = #{init};
+        #{extra}
+        if (argc > 100000) printf("%p", &test);
+        return 0;
+      }
+    SRC
+    $defs.push("-DJSON_ENABLE_SIMD")
+  else
+    puts "Disable SIMD"
+  end
+end
+
+have_header('cpuid.h')

data/ext/json/ext/simd/simd.h

@@ -0,0 +1,188 @@
+typedef enum {
+    SIMD_NONE,
+    SIMD_NEON,
+    SIMD_SSE2
+} SIMD_Implementation;
+
+#ifdef JSON_ENABLE_SIMD
+
+#ifdef __clang__
+# if __has_builtin(__builtin_ctzll)
+#   define HAVE_BUILTIN_CTZLL 1
+# else
+#   define HAVE_BUILTIN_CTZLL 0
+# endif
+#elif defined(__GNUC__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 3))
+# define HAVE_BUILTIN_CTZLL 1
+#else
+# define HAVE_BUILTIN_CTZLL 0
+#endif
+
+static inline uint32_t trailing_zeros64(uint64_t input)
+{
+#if HAVE_BUILTIN_CTZLL
+    return __builtin_ctzll(input);
+#else
+    uint32_t trailing_zeros = 0;
+    uint64_t temp = input;
+    while ((temp & 1) == 0 && temp > 0) {
+        trailing_zeros++;
+        temp >>= 1;
+    }
+    return trailing_zeros;
+#endif
+}
+
+static inline int trailing_zeros(int input)
+{
+#if HAVE_BUILTIN_CTZLL
+    return __builtin_ctz(input);
+#else
+    int trailing_zeros = 0;
+    int temp = input;
+    while ((temp & 1) == 0 && temp > 0) {
+        trailing_zeros++;
+        temp >>= 1;
+    }
+    return trailing_zeros;
+#endif
+}
+
+#if (defined(__GNUC__ ) || defined(__clang__))
+#define FORCE_INLINE __attribute__((always_inline))
+#else
+#define FORCE_INLINE
+#endif
+
+
+#define SIMD_MINIMUM_THRESHOLD 6
+
+#if defined(__ARM_NEON) || defined(__ARM_NEON__) || defined(__aarch64__) || defined(_M_ARM64)
+#include <arm_neon.h>
+
+#define FIND_SIMD_IMPLEMENTATION_DEFINED 1
+static inline SIMD_Implementation find_simd_implementation(void)
+{
+    return SIMD_NEON;
+}
+
+#define HAVE_SIMD 1
+#define HAVE_SIMD_NEON 1
+
+// See: https://community.arm.com/arm-community-blogs/b/servers-and-cloud-computing-blog/posts/porting-x86-vector-bitmask-optimizations-to-arm-neon
+static inline FORCE_INLINE uint64_t neon_match_mask(uint8x16_t matches)
+{
+    const uint8x8_t res = vshrn_n_u16(vreinterpretq_u16_u8(matches), 4);
+    const uint64_t mask = vget_lane_u64(vreinterpret_u64_u8(res), 0);
+    return mask & 0x8888888888888888ull;
+}
+
+static inline FORCE_INLINE uint64_t compute_chunk_mask_neon(const char *ptr)
+{
+    uint8x16_t chunk = vld1q_u8((const unsigned char *)ptr);
+
+    // Trick: c < 32 || c == 34 can be factored as c ^ 2 < 33
+    // https://lemire.me/blog/2025/04/13/detect-control-characters-quotes-and-backslashes-efficiently-using-swar/
+    const uint8x16_t too_low_or_dbl_quote = vcltq_u8(veorq_u8(chunk, vdupq_n_u8(2)), vdupq_n_u8(33));
+
+    uint8x16_t has_backslash = vceqq_u8(chunk, vdupq_n_u8('\\'));
+    uint8x16_t needs_escape  = vorrq_u8(too_low_or_dbl_quote, has_backslash);
+    return neon_match_mask(needs_escape);
+}
+
+static inline FORCE_INLINE int string_scan_simd_neon(const char **ptr, const char *end, uint64_t *mask)
+{
+    while (*ptr + sizeof(uint8x16_t) <= end) {
+        uint64_t chunk_mask = compute_chunk_mask_neon(*ptr);
+        if (chunk_mask) {
+            *mask = chunk_mask;
+            return 1;
+        }
+        *ptr += sizeof(uint8x16_t);
+    }
+    return 0;
+}
+
+static inline 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;
+}
+
+#endif /* ARM Neon Support.*/
+
+#if defined(__amd64__) || defined(__amd64) || defined(__x86_64__) || defined(__x86_64) || defined(_M_X64) || defined(_M_AMD64)
+
+#ifdef HAVE_X86INTRIN_H
+#include <x86intrin.h>
+
+#define HAVE_SIMD 1
+#define HAVE_SIMD_SSE2 1
+
+#ifdef HAVE_CPUID_H
+#define FIND_SIMD_IMPLEMENTATION_DEFINED 1
+
+#if defined(__clang__) || defined(__GNUC__)
+#define TARGET_SSE2 __attribute__((target("sse2")))
+#else
+#define TARGET_SSE2
+#endif
+
+#define _mm_cmpge_epu8(a, b) _mm_cmpeq_epi8(_mm_max_epu8(a, b), a)
+#define _mm_cmple_epu8(a, b) _mm_cmpge_epu8(b, a)
+#define _mm_cmpgt_epu8(a, b) _mm_xor_si128(_mm_cmple_epu8(a, b), _mm_set1_epi8(-1))
+#define _mm_cmplt_epu8(a, b) _mm_cmpgt_epu8(b, a)
+
+static inline TARGET_SSE2 FORCE_INLINE int compute_chunk_mask_sse2(const char *ptr)
+{
+    __m128i chunk         = _mm_loadu_si128((__m128i const*)ptr);
+    // Trick: c < 32 || c == 34 can be factored as c ^ 2 < 33
+    // https://lemire.me/blog/2025/04/13/detect-control-characters-quotes-and-backslashes-efficiently-using-swar/
+    __m128i too_low_or_dbl_quote = _mm_cmplt_epu8(_mm_xor_si128(chunk, _mm_set1_epi8(2)), _mm_set1_epi8(33));
+    __m128i has_backslash = _mm_cmpeq_epi8(chunk, _mm_set1_epi8('\\'));
+    __m128i needs_escape  = _mm_or_si128(too_low_or_dbl_quote, has_backslash);
+    return _mm_movemask_epi8(needs_escape);
+}
+
+static inline TARGET_SSE2 FORCE_INLINE int string_scan_simd_sse2(const char **ptr, const char *end, int *mask)
+{
+    while (*ptr + sizeof(__m128i) <= end) {
+        int chunk_mask = compute_chunk_mask_sse2(*ptr);
+        if (chunk_mask) {
+            *mask = chunk_mask;
+            return 1;
+        }
+        *ptr += sizeof(__m128i);
+    }
+
+    return 0;
+}
+
+#include <cpuid.h>
+#endif /* HAVE_CPUID_H */
+
+static inline SIMD_Implementation find_simd_implementation(void)
+{
+    // TODO Revisit. I think the SSE version now only uses SSE2 instructions.
+    if (__builtin_cpu_supports("sse2")) {
+        return SIMD_SSE2;
+    }
+
+    return SIMD_NONE;
+}
+
+#endif /* HAVE_X86INTRIN_H */
+#endif /* X86_64 Support */
+
+#endif /* JSON_ENABLE_SIMD */
+
+#ifndef FIND_SIMD_IMPLEMENTATION_DEFINED
+static inline SIMD_Implementation find_simd_implementation(void)
+{
+    return SIMD_NONE;
+}
+#endif

data/ext/json/ext/vendor/fpconv.c

@@ -29,6 +29,10 @@
 #include <string.h>
 #include <stdint.h>
 
+#ifdef JSON_DEBUG
+#include <assert.h>
+#endif
+
 #define npowers     87
 #define steppowers  8
 #define firstpower -348 /* 10 ^ -348 */
@@ -320,15 +324,7 @@ static int emit_digits(char* digits, int ndigits, char* dest, int K, bool neg)
 {
     int exp = absv(K + ndigits - 1);
 
-    int max_trailing_zeros = 7;
-
-    if(neg) {
-        max_trailing_zeros -= 1;
-    }
-
-    /* write plain integer */
-    if(K >= 0 && (exp < (ndigits + max_trailing_zeros))) {
-
+    if(K >= 0 && exp < 15) {
         memcpy(dest, digits, ndigits);
         memset(dest + ndigits, '0', K);
 
@@ -432,10 +428,12 @@ static int filter_special(double fp, char* dest)
  *
  * Input:
  * fp -> the double to convert, dest -> destination buffer.
- * The generated string will never be longer than 28 characters.
- * Make sure to pass a pointer to at least 28 bytes of memory.
+ * The generated string will never be longer than 32 characters.
+ * Make sure to pass a pointer to at least 32 bytes of memory.
  * The emitted string will not be null terminated.
  *
+ *
+ *
  * Output:
  * The number of written characters.
  *
@@ -474,6 +472,9 @@ static int fpconv_dtoa(double d, char dest[28])
     int ndigits = grisu2(d, digits, &K);
 
     str_len += emit_digits(digits, ndigits, dest + str_len, K, neg);
+#ifdef JSON_DEBUG
+    assert(str_len <= 32);
+#endif
 
     return str_len;
 }

data/json.gemspec

@@ -44,15 +44,14 @@ spec = Gem::Specification.new do |s|
     "LEGAL",
     "README.md",
     "json.gemspec",
-    *Dir["lib/**/*.rb"],
-  ]
+  ] + Dir.glob("lib/**/*.rb", base: File.expand_path("..", __FILE__))
 
   if java_ext
     s.platform = 'java'
     s.files += Dir["lib/json/ext/**/*.jar"]
   else
     s.extensions = Dir["ext/json/**/extconf.rb"]
-    s.files += Dir["ext/json/**/*.{c,h}"]
+    s.files += Dir["ext/json/**/*.{c,h,rb}"]
   end
 end
 

data/lib/json/add/core.rb

@@ -7,6 +7,7 @@ require 'json/add/date_time'
 require 'json/add/exception'
 require 'json/add/range'
 require 'json/add/regexp'
+require 'json/add/string'
 require 'json/add/struct'
 require 'json/add/symbol'
 require 'json/add/time'

data/lib/json/add/string.rb

@@ -0,0 +1,35 @@
+# frozen_string_literal: true
+unless defined?(::JSON::JSON_LOADED) and ::JSON::JSON_LOADED
+  require 'json'
+end
+
+class String
+  # 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 class method.
+  def self.json_create(object)
+    object["raw"].pack("C*")
+  end
+
+  # 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.
+  def to_json_raw_object
+    {
+      JSON.create_id => self.class.name,
+      "raw" => unpack("C*"),
+    }
+  end
+
+  # 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.
+  def to_json_raw(...)
+    to_json_raw_object.to_json(...)
+  end
+end