oj 3.16.13 → 3.16.14

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: fc0290fa1cfe6af1094de1d7188836e0c09cb04f2f08401de118253026604650
-  data.tar.gz: de5258e96984a21afb2fac946fe28ad255926893a6157f2874445edf10aa8bbe
+  metadata.gz: c97d24d950284c4f108b8ff933e4358250d53b69f9300b3faf2b8b7f29fdb5b7
+  data.tar.gz: 3112c763244b2f558e4f2a8f0ae79c01dcea9ec4281a915a2e39a404a34bf8f1
 SHA512:
-  metadata.gz: d7870818fd86043a17b834756b67a4009a6f7ef60baf53b02a0b0d4431ccba723d9e533553bea04ae46ae9f233e447b79c4106e6827782bd0c2ffb9c332081a3
-  data.tar.gz: fd3966ac7fb5da9f1a5ebb68f4a8f5b9a5f9fa1a1255e93dfef078f66f00a6af5bb7e37676441f7d6229b29222741a2bc7b75164fd445a39b906ef904946d41b
+  metadata.gz: 3f32303a7e78478137b76fbadf4bcab93d164d60918fd5b8c20bccdd7cf0476e290e1290e4bc114f3fd6a3afaac0fb3c9df61c91ede88bfb9a9762d098731688
+  data.tar.gz: bcc7411ae7cde8ff0457563f67d99093b09c3d70a36a13c5db1144dbf983d5f864f204f421a966be30d3a2c3610f82033ed08c6c0c86d7ff2e8b391b4cd5b02f

data/CHANGELOG.md

@@ -1,5 +1,13 @@
 # CHANGELOG
 
+## 3.16.14 - 2026-02-04
+
+-  Fixed SSE issue #989.
+
+- Removed ostruct dependency.
+
+- Removed generic object JSON gem tests.
+
 ## 3.16.13 - 2025-12-05
 
 - Fixed rails encoding for Hash and Array subclasses.

data/ext/oj/dump.c

@@ -201,6 +201,45 @@ void initialize_neon(void) {
 }
 #endif
 
+#ifdef HAVE_SIMD_SSE4_2
+
+static __m128i hibit_friendly_chars_sse42[8];
+
+// From: https://stackoverflow.com/questions/36998538/fastest-way-to-horizontally-sum-sse-unsigned-byte-vector
+inline uint32_t _mm_sum_epu8(const __m128i v) {
+    __m128i vsum = _mm_sad_epu8(v, _mm_setzero_si128());
+    return _mm_cvtsi128_si32(vsum) + _mm_extract_epi16(vsum, 4);
+}
+
+inline static OJ_TARGET_SSE42 size_t hibit_friendly_size_sse42(const uint8_t *str, size_t len) {
+    size_t size = 0;
+    size_t i    = 0;
+
+    for (; i + sizeof(__m128i) <= len; i += sizeof(__m128i), str += sizeof(__m128i)) {
+        size += sizeof(__m128i);
+
+        __m128i chunk = _mm_loadu_si128((__m128i *)str);
+        __m128i tmp   = vector_lookup_sse42(chunk, hibit_friendly_chars_sse42, 8);
+        size += _mm_sum_epu8(tmp);
+    }
+    size_t total = size + calculate_string_size(str, len - i, hibit_friendly_chars);
+    return total;
+}
+
+void OJ_TARGET_SSE42 initialize_sse42(void) {
+    for (int i = 0; i < 8; i++) {
+        hibit_friendly_chars_sse42[i] = _mm_sub_epi8(
+            _mm_loadu_si128((__m128i *)(hibit_friendly_chars + i * sizeof(__m128i))),
+            _mm_set1_epi8('1'));
+    }
+}
+
+#else
+
+#define SIMD_TARGET
+
+#endif /* HAVE_SIMD_SSE4_2 */
+
 inline static size_t hibit_friendly_size(const uint8_t *str, size_t len) {
 #ifdef HAVE_SIMD_NEON
     size_t size = 0;
@@ -220,6 +259,13 @@ inline static size_t hibit_friendly_size(const uint8_t *str, size_t len) {
 
     size_t total = size + calculate_string_size(str, len - i, hibit_friendly_chars);
     return total;
+#elif defined(HAVE_SIMD_SSE4_2)
+    if (SIMD_Impl == SIMD_SSE42) {
+        if (len >= sizeof(__m128i)) {
+            return hibit_friendly_size_sse42(str, len);
+        }
+    }
+    return calculate_string_size(str, len, hibit_friendly_chars);
 #else
     return calculate_string_size(str, len, hibit_friendly_chars);
 #endif
@@ -944,6 +990,34 @@ neon_update(const char *str, uint8x16x4_t *cmap_neon, int neon_table_size, bool
     return result;
 }
 
+#elif defined(HAVE_SIMD_SSE4_2)
+typedef struct _sse42_match_result {
+    __m128i actions;
+    bool    needs_escape;
+    int     escape_mask;
+    bool    has_some_hibit;
+    bool    do_unicode_validation;
+} sse42_match_result;
+
+static inline OJ_TARGET_SSE42 sse42_match_result
+sse42_update(const char *str, __m128i *cmap_sse42, int sse42_tab_size, bool do_unicode_validation, bool has_hi) {
+    sse42_match_result result = {.has_some_hibit = false, .do_unicode_validation = false};
+
+    __m128i chunk        = _mm_loadu_si128((__m128i *)str);
+    __m128i actions      = vector_lookup_sse42(chunk, cmap_sse42, sse42_tab_size);
+    __m128i needs_escape = _mm_xor_si128(_mm_cmpeq_epi8(actions, _mm_setzero_si128()), _mm_set1_epi8(0xFF));
+    result.actions       = _mm_add_epi8(actions, _mm_set1_epi8('1'));
+
+    result.escape_mask  = _mm_movemask_epi8(needs_escape);
+    result.needs_escape = result.escape_mask != 0;
+    if (has_hi && do_unicode_validation) {
+        __m128i has_some_hibit       = _mm_and_si128(chunk, _mm_set1_epi8(0x80));
+        result.has_some_hibit        = _mm_movemask_epi8(has_some_hibit) != 0;
+        result.do_unicode_validation = has_hi && do_unicode_validation && result.has_some_hibit;
+    }
+    return result;
+}
+
 #endif /* HAVE_SIMD_NEON */
 
 static inline FORCE_INLINE const char *process_character(char         action,
@@ -1023,6 +1097,9 @@ void oj_dump_cstr(const char *str, size_t cnt, bool is_sym, bool escape1, Out ou
 #ifdef HAVE_SIMD_NEON
     uint8x16x4_t *cmap_neon       = NULL;
     int           neon_table_size = 0;
+#elif defined(HAVE_SIMD_SSE4_2)
+    __m128i *cmap_sse42 = NULL;
+    int      sse42_tab_size;
 #endif /* HAVE_SIMD_NEON */
     const char *orig                  = str;
     bool        has_hi                = false;
@@ -1091,6 +1168,9 @@ void oj_dump_cstr(const char *str, size_t cnt, bool is_sym, bool escape1, Out ou
 #ifdef HAVE_SIMD_NEON
         cmap_neon       = hibit_friendly_chars_neon;
         neon_table_size = 2;
+#elif defined(HAVE_SIMD_SSE4_2)
+        cmap_sse42     = hibit_friendly_chars_sse42;
+        sse42_tab_size = 8;
 #endif /* HAVE_NEON_SIMD */
         size = hibit_friendly_size((uint8_t *)str, cnt);
     }
@@ -1118,21 +1198,32 @@ void oj_dump_cstr(const char *str, size_t cnt, bool is_sym, bool escape1, Out ou
         if (is_sym) {
             *out->cur++ = ':';
         }
-#ifdef HAVE_SIMD_NEON
-        const char *chunk_start;
-        const char *chunk_end;
-        const char *cursor   = str;
-        bool        use_neon = (cmap_neon != NULL && cnt >= (sizeof(uint8x16_t))) ? true : false;
-        char        matches[16];
+
+#if defined(HAVE_SIMD_NEON) || defined(HAVE_SIMD_SSE4_2)
+
 #define SEARCH_FLUSH                                  \
     if (str > cursor) {                               \
         APPEND_CHARS(out->cur, cursor, str - cursor); \
         cursor = str;                                 \
     }
 
-#endif /* HAVE_SIMD_NEON */
+        const char *chunk_start;
+        const char *chunk_end;
+        const char *cursor = str;
+        char        matches[16];
+#endif /* HAVE_SIMD_NEON || HAVE_SIMD_SSE4_2 */
+
+#if defined(HAVE_SIMD_NEON)
+        bool use_simd = (cmap_neon != NULL && cnt >= (sizeof(uint8x16_t))) ? true : false;
+#elif defined(HAVE_SIMD_SSE4_2)
+        bool use_simd = false;
+        if (SIMD_Impl == SIMD_SSE42) {
+            use_simd = (cmap_sse42 != NULL && cnt >= (sizeof(__m128i))) ? true : false;
+        }
+#endif
+
 #ifdef HAVE_SIMD_NEON
-        if (use_neon) {
+        if (use_simd) {
             while (str < end) {
                 const char *chunk_ptr = NULL;
                 if (str + sizeof(uint8x16_t) <= end) {
@@ -1195,7 +1286,55 @@ void oj_dump_cstr(const char *str, size_t cnt, bool is_sym, bool escape1, Out ou
             }
             SEARCH_FLUSH;
         }
-#endif /* HAVE_SIMD_NEON */
+#endif
+
+#ifdef HAVE_SIMD_SSE4_2
+        if (SIMD_Impl == SIMD_SSE42) {
+            if (use_simd) {
+                while (str < end) {
+                    const char *chunk_ptr = NULL;
+                    if (str + sizeof(__m128i) <= end) {
+                        chunk_ptr   = str;
+                        chunk_start = str;
+                        chunk_end   = str + sizeof(__m128i);
+                    } else if ((end - str) >= SIMD_MINIMUM_THRESHOLD) {
+                        memset(out->cur, 'A', sizeof(__m128i));
+                        memcpy(out->cur, str, (end - str));
+                        chunk_ptr   = out->cur;
+                        chunk_start = str;
+                        chunk_end   = end;
+                    } else {
+                        break;
+                    }
+                    sse42_match_result result = sse42_update(chunk_ptr,
+                                                             cmap_sse42,
+                                                             sse42_tab_size,
+                                                             do_unicode_validation,
+                                                             has_hi);
+                    if ((result.do_unicode_validation) || result.needs_escape) {
+                        SEARCH_FLUSH;
+                        _mm_storeu_si128((__m128i *)matches, result.actions);
+                        while (str < chunk_end) {
+                            long match_index = str - chunk_start;
+                            str              = process_character(matches[match_index],
+                                                    str,
+                                                    end,
+                                                    out,
+                                                    orig,
+                                                    do_unicode_validation,
+                                                    &check_start);
+                            str++;
+                        }
+                        cursor = str;
+                        continue;
+                    }
+                    str = chunk_end;
+                }
+                SEARCH_FLUSH;
+            }
+        }
+#endif /* HAVE_SIMD_SSE4_2 */
+
         for (; str < end; str++) {
             str = process_character(cmap[(uint8_t)*str], str, end, out, orig, do_unicode_validation, &check_start);
         }

data/ext/oj/extconf.rb

@@ -35,11 +35,15 @@ have_func('rb_ext_ractor_safe', 'ruby.h')
 
 dflags['OJ_DEBUG'] = true unless ENV['OJ_DEBUG'].nil?
 
-# Enable SIMD optimizations - try SSE4.2 on x86_64 for best performance
-# Falls back to SSE2 or compiler defaults if not available
-if try_cflags('-msse4.2')
-  $CPPFLAGS += ' -msse4.2'
-elsif try_cflags('-msse2')
+# SIMD optimizations use runtime CPU detection and function-level target attributes
+# We do NOT add global -msse4.2/-msse2 flags here because:
+# 1. It would cause illegal instruction errors on CPUs without SSE4.2
+# 2. The code uses __attribute__((target("sse4.2"))) for SSE4.2 functions
+# 3. Runtime detection in oj_get_simd_implementation() selects the right path
+#
+# We only add -msse2 if available, since SSE2 is baseline for all x86_64 CPUs
+# and needed for compiling the SSE2 fallback code on 32-bit x86
+if try_cflags('-msse2')
   $CPPFLAGS += ' -msse2'
 end
 

data/ext/oj/oj.c

@@ -167,6 +167,8 @@ pthread_mutex_t oj_cache_mutex;
 VALUE oj_cache_mutex = Qnil;
 #endif
 
+SIMD_Implementation SIMD_Impl = SIMD_NONE;
+
 extern void oj_parser_init();
 
 const char oj_json_class[] = "json_class";
@@ -1780,6 +1782,78 @@ static VALUE mem_report(VALUE self) {
  *
  * - *:wab* specifically for WAB data exchange.
  */
+
+// =============================================================================
+// Runtime SIMD CPU detection
+// Cross-platform support for Windows (MSVC), Linux, and macOS (GCC/Clang)
+// =============================================================================
+SIMD_Implementation oj_get_simd_implementation(void) {
+#ifdef HAVE_SIMD_X86
+    // x86/x86_64 runtime detection
+
+#if defined(_MSC_VER)
+    // MSVC: Use __cpuid intrinsic
+    int cpu_info[4];
+    __cpuid(cpu_info, 1);
+
+    // Check for SSE4.2 (bit 20 of ECX)
+    if (cpu_info[2] & (1 << 20)) {
+        return SIMD_SSE42;
+    }
+    // Check for SSE2 (bit 26 of EDX)
+    if (cpu_info[3] & (1 << 26)) {
+        return SIMD_SSE2;
+    }
+
+#elif defined(__GNUC__) || defined(__clang__)
+    // GCC/Clang: Use __builtin_cpu_supports if available
+#if defined(__has_builtin)
+#if __has_builtin(__builtin_cpu_supports)
+#define OJ_HAS_BUILTIN_CPU_SUPPORTS 1
+#endif
+#elif defined(__GNUC__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 8))
+    // GCC 4.8+ has __builtin_cpu_supports
+#define OJ_HAS_BUILTIN_CPU_SUPPORTS 1
+#endif
+
+#ifdef OJ_HAS_BUILTIN_CPU_SUPPORTS
+#ifdef HAVE_SIMD_SSE4_2
+    if (__builtin_cpu_supports("sse4.2")) {
+        return SIMD_SSE42;
+    }
+#endif
+#ifdef HAVE_SIMD_SSE2
+    if (__builtin_cpu_supports("sse2")) {
+        return SIMD_SSE2;
+    }
+#endif
+#else
+    // Fallback: Use CPUID instruction directly
+    unsigned int eax, ebx, ecx, edx;
+    if (__get_cpuid(1, &eax, &ebx, &ecx, &edx)) {
+        // Check for SSE4.2 (bit 20 of ECX)
+        if (ecx & (1 << 20)) {
+            return SIMD_SSE42;
+        }
+        // Check for SSE2 (bit 26 of EDX)
+        if (edx & (1 << 26)) {
+            return SIMD_SSE2;
+        }
+    }
+#endif  // OJ_HAS_BUILTIN_CPU_SUPPORTS
+
+#endif  // _MSC_VER vs GCC/Clang
+
+#endif  // HAVE_SIMD_X86
+
+#ifdef HAVE_SIMD_NEON
+    // ARM NEON is always available on ARM64 and detected at compile time
+    return SIMD_NEON;
+#endif
+
+    return SIMD_NONE;
+}
+
 void Init_oj(void) {
     int err = 0;
 
@@ -2080,10 +2154,18 @@ void Init_oj(void) {
 #endif
     oj_init_doc();
 
+    SIMD_Impl = oj_get_simd_implementation();
+
     oj_parser_init();
     oj_scanner_init();
 
 #ifdef HAVE_SIMD_NEON
     initialize_neon();
 #endif /* HAVE_SIMD_NEON */
+
+#ifdef HAVE_SIMD_SSE4_2
+    if (SIMD_Impl == SIMD_SSE42) {
+        initialize_sse42();
+    }
+#endif /* HAVE_SIMD_SSE4_2 */
 }

data/ext/oj/parse.c

@@ -202,7 +202,8 @@ static inline const char *scan_string_noSIMD(const char *str, const char *end) {
 #ifdef HAVE_SIMD_SSE4_2
 // Optimized SIMD string scanner using SSE4.2 instructions
 // Uses prefetching and processes multiple chunks in parallel to reduce latency
-static inline const char *scan_string_SSE42(const char *str, const char *end) {
+// Note: OJ_TARGET_SSE42 attribute allows this to compile even without global -msse4.2
+static OJ_TARGET_SSE42 const char *scan_string_SSE42(const char *str, const char *end) {
     static const char chars[16]   = "\x00\\\"";
     const __m128i     terminate   = _mm_loadu_si128((const __m128i *)&chars[0]);
     const char       *safe_end_64 = end - 64;
@@ -212,7 +213,7 @@ static inline const char *scan_string_SSE42(const char *str, const char *end) {
     // This reduces pipeline stalls and improves instruction-level parallelism
     while (str <= safe_end_64) {
         // Prefetch next cache line for better memory throughput
-        __builtin_prefetch(str + 64, 0, 0);
+        OJ_PREFETCH(str + 64);
 
         // Load and compare 4 chunks in parallel
         const __m128i chunk0 = _mm_loadu_si128((const __m128i *)(str));
@@ -225,7 +226,7 @@ static inline const char *scan_string_SSE42(const char *str, const char *end) {
                                     chunk0,
                                     16,
                                     _SIDD_UBYTE_OPS | _SIDD_CMP_EQUAL_ANY | _SIDD_LEAST_SIGNIFICANT);
-        if (__builtin_expect(r0 != 16, 0))
+        if (OJ_UNLIKELY(r0 != 16))
             return str + r0;
 
         const int r1 = _mm_cmpestri(terminate,
@@ -233,7 +234,7 @@ static inline const char *scan_string_SSE42(const char *str, const char *end) {
                                     chunk1,
                                     16,
                                     _SIDD_UBYTE_OPS | _SIDD_CMP_EQUAL_ANY | _SIDD_LEAST_SIGNIFICANT);
-        if (__builtin_expect(r1 != 16, 0))
+        if (OJ_UNLIKELY(r1 != 16))
             return str + 16 + r1;
 
         const int r2 = _mm_cmpestri(terminate,
@@ -241,7 +242,7 @@ static inline const char *scan_string_SSE42(const char *str, const char *end) {
                                     chunk2,
                                     16,
                                     _SIDD_UBYTE_OPS | _SIDD_CMP_EQUAL_ANY | _SIDD_LEAST_SIGNIFICANT);
-        if (__builtin_expect(r2 != 16, 0))
+        if (OJ_UNLIKELY(r2 != 16))
             return str + 32 + r2;
 
         const int r3 = _mm_cmpestri(terminate,
@@ -249,7 +250,7 @@ static inline const char *scan_string_SSE42(const char *str, const char *end) {
                                     chunk3,
                                     16,
                                     _SIDD_UBYTE_OPS | _SIDD_CMP_EQUAL_ANY | _SIDD_LEAST_SIGNIFICANT);
-        if (__builtin_expect(r3 != 16, 0))
+        if (OJ_UNLIKELY(r3 != 16))
             return str + 48 + r3;
 
         str += 64;
@@ -274,7 +275,8 @@ static inline const char *scan_string_SSE42(const char *str, const char *end) {
 #ifdef HAVE_SIMD_SSE2
 // Optimized SSE2 string scanner (fallback for older x86_64 CPUs)
 // Uses SSE2 instructions with prefetching and parallel processing
-static inline const char *scan_string_SSE2(const char *str, const char *end) {
+// Note: OJ_TARGET_SSE2 attribute allows this to compile even without global -msse2
+static OJ_TARGET_SSE2 const char *scan_string_SSE2(const char *str, const char *end) {
     const char *safe_end_64 = end - 64;
     const char *safe_end_16 = end - 16;
 
@@ -285,7 +287,7 @@ static inline const char *scan_string_SSE2(const char *str, const char *end) {
 
     // Process 64 bytes at a time for better throughput
     while (str <= safe_end_64) {
-        __builtin_prefetch(str + 64, 0, 0);
+        OJ_PREFETCH(str + 64);
 
         // Load 4 chunks
         const __m128i chunk0 = _mm_loadu_si128((const __m128i *)(str));
@@ -309,20 +311,20 @@ static inline const char *scan_string_SSE2(const char *str, const char *end) {
 
         // Convert to masks
         int mask0 = _mm_movemask_epi8(cmp0);
-        if (__builtin_expect(mask0 != 0, 0))
-            return str + __builtin_ctz(mask0);
+        if (OJ_UNLIKELY(mask0 != 0))
+            return str + OJ_CTZ(mask0);
 
         int mask1 = _mm_movemask_epi8(cmp1);
-        if (__builtin_expect(mask1 != 0, 0))
-            return str + 16 + __builtin_ctz(mask1);
+        if (OJ_UNLIKELY(mask1 != 0))
+            return str + 16 + OJ_CTZ(mask1);
 
         int mask2 = _mm_movemask_epi8(cmp2);
-        if (__builtin_expect(mask2 != 0, 0))
-            return str + 32 + __builtin_ctz(mask2);
+        if (OJ_UNLIKELY(mask2 != 0))
+            return str + 32 + OJ_CTZ(mask2);
 
         int mask3 = _mm_movemask_epi8(cmp3);
-        if (__builtin_expect(mask3 != 0, 0))
-            return str + 48 + __builtin_ctz(mask3);
+        if (OJ_UNLIKELY(mask3 != 0))
+            return str + 48 + OJ_CTZ(mask3);
 
         str += 64;
     }
@@ -335,7 +337,7 @@ static inline const char *scan_string_SSE2(const char *str, const char *end) {
             _mm_cmpeq_epi8(chunk, quote));
         int mask = _mm_movemask_epi8(matches);
         if (mask != 0)
-            return str + __builtin_ctz(mask);
+            return str + OJ_CTZ(mask);
     }
 
     return scan_string_noSIMD(str, end);
@@ -345,11 +347,19 @@ static inline const char *scan_string_SSE2(const char *str, const char *end) {
 static const char *(*scan_func)(const char *str, const char *end) = scan_string_noSIMD;
 
 void oj_scanner_init(void) {
+    // Use runtime CPU detection to select the best SIMD implementation
+    // This ensures we don't crash on CPUs that don't support SSE4.2
+    SIMD_Implementation impl = oj_get_simd_implementation();
+
+    switch (impl) {
 #ifdef HAVE_SIMD_SSE4_2
-    scan_func = scan_string_SSE42;
-#elif defined(HAVE_SIMD_SSE2)
-    scan_func = scan_string_SSE2;
+    case SIMD_SSE42: scan_func = scan_string_SSE42; break;
+#endif
+#ifdef HAVE_SIMD_SSE2
+    case SIMD_SSE2: scan_func = scan_string_SSE2; break;
 #endif
+    default: scan_func = scan_string_noSIMD; break;
+    }
     // Note: ARM NEON string scanning would be added here if needed
 }
 

data/ext/oj/rails.c

@@ -668,7 +668,7 @@ static VALUE encoder_new(int argc, VALUE *argv, VALUE self) {
     } else {
         e->arg = rb_hash_new();
     }
-    oj_parse_options(*argv, &e->opts);
+    oj_parse_options(e->arg, &e->opts);
 
     return TypedData_Wrap_Struct(encoder_class, &oj_encoder_type, e);
 }

data/ext/oj/simd.h

@@ -3,45 +3,170 @@
 
 // SIMD architecture detection and configuration
 // This header provides unified SIMD support across different CPU architectures
+// with cross-platform runtime detection (Windows/Linux/Mac)
 
+// SIMD implementation enum - used for runtime selection
+typedef enum _simd_implementation { SIMD_NONE, SIMD_NEON, SIMD_SSE2, SIMD_SSE42 } SIMD_Implementation;
+
+// Define in oj.c.
+extern SIMD_Implementation SIMD_Impl;
+
+// Runtime CPU detection function (implemented in oj.c)
+SIMD_Implementation oj_get_simd_implementation(void);
+
+// =============================================================================
+// Compiler compatibility macros
+// =============================================================================
+
+// Branch prediction hints
+#if defined(__GNUC__) || defined(__clang__)
+#define OJ_LIKELY(x) __builtin_expect(!!(x), 1)
+#define OJ_UNLIKELY(x) __builtin_expect(!!(x), 0)
+#else
+#define OJ_LIKELY(x) (x)
+#define OJ_UNLIKELY(x) (x)
+#endif
+
+// Prefetch hints
+#if defined(__GNUC__) || defined(__clang__)
+#define OJ_PREFETCH(addr) __builtin_prefetch(addr, 0, 0)
+#elif defined(_MSC_VER)
+#include <intrin.h>
+#define OJ_PREFETCH(addr) _mm_prefetch((const char *)(addr), _MM_HINT_T0)
+#else
+#define OJ_PREFETCH(addr) ((void)0)
+#endif
+
+// Count trailing zeros (for SSE2 mask scanning)
+#if defined(__GNUC__) || defined(__clang__)
+#define OJ_CTZ(x) __builtin_ctz(x)
+#elif defined(_MSC_VER)
+#include <intrin.h>
+static __inline int oj_ctz_msvc(unsigned int x) {
+    unsigned long index;
+    _BitScanForward(&index, x);
+    return (int)index;
+}
+#define OJ_CTZ(x) oj_ctz_msvc(x)
+#else
+// Fallback: naive implementation
+static inline int oj_ctz_fallback(unsigned int x) {
+    int count = 0;
+    while ((x & 1) == 0 && count < 32) {
+        x >>= 1;
+        count++;
+    }
+    return count;
+}
+#define OJ_CTZ(x) oj_ctz_fallback(x)
+#endif
+
+// =============================================================================
 // x86/x86_64 SIMD detection
+// =============================================================================
 #if defined(__x86_64__) || defined(__i386__) || defined(_M_IX86) || defined(_M_X64)
 #define HAVE_SIMD_X86 1
 
-// SSE4.2 support (Intel Core i7+, AMD Bulldozer+)
-// Enabled automatically when compiler has -msse4.2 flag
-#if defined(__SSE4_2__)
+// Include appropriate SIMD headers
+#if defined(_MSC_VER)
+// MSVC: use intrin.h for all intrinsics
+#include <intrin.h>
 #define HAVE_SIMD_SSE4_2 1
-#include <nmmintrin.h>
+#define HAVE_SIMD_SSE2 1
+#elif defined(__GNUC__) || defined(__clang__)
+// GCC/Clang: check for header availability and include them
+// We include headers but use target attributes to enable instructions per-function
+// Include cpuid.h for __get_cpuid fallback when __builtin_cpu_supports is unavailable
+#if __has_include(<cpuid.h>)
+#include <cpuid.h>
 #endif
-
-// SSE2 support (fallback for older x86_64 CPUs - all x86_64 CPUs support SSE2)
-#if defined(__SSE2__) && !defined(HAVE_SIMD_SSE4_2)
+#if defined(__SSE4_2__) || defined(__SSE2__)
+// If any SSE is enabled globally, x86intrin.h should be available
+#include <x86intrin.h>
+#define HAVE_SIMD_SSE4_2 1
+#define HAVE_SIMD_SSE2 1
+#else
+// Try to include headers anyway for target attribute functions
+#if __has_include(<x86intrin.h>)
+#include <x86intrin.h>
+#define HAVE_SIMD_SSE4_2 1
+#define HAVE_SIMD_SSE2 1
+#elif __has_include(<nmmintrin.h>)
+#include <nmmintrin.h>
+#define HAVE_SIMD_SSE4_2 1
 #define HAVE_SIMD_SSE2 1
+#elif __has_include(<emmintrin.h>)
 #include <emmintrin.h>
+#define HAVE_SIMD_SSE2 1
+#endif
+#endif
+#endif
+
+// Target attribute macros for function-level SIMD enabling
+#if defined(__clang__) || defined(__GNUC__)
+#define OJ_TARGET_SSE42 __attribute__((target("sse4.2")))
+#define OJ_TARGET_SSE2 __attribute__((target("sse2")))
+#else
+// MSVC doesn't need target attributes - intrinsics are always available
+#define OJ_TARGET_SSE42
+#define OJ_TARGET_SSE2
 #endif
 
 #endif  // x86/x86_64
 
+// =============================================================================
 // ARM NEON detection
+// =============================================================================
 #if defined(__ARM_NEON) || defined(__ARM_NEON__) || defined(__aarch64__) || defined(_M_ARM64)
 #define HAVE_SIMD_NEON 1
 #define SIMD_MINIMUM_THRESHOLD 6
 #include <arm_neon.h>
 #endif
 
-// Define which SIMD implementation to use (priority order: SSE4.2 > NEON > SSE2)
-#if defined(HAVE_SIMD_SSE4_2)
+// =============================================================================
+// SIMD type string for debugging/logging
+// =============================================================================
+#if defined(HAVE_SIMD_SSE4_2) || defined(HAVE_SIMD_SSE2)
 #define HAVE_SIMD_STRING_SCAN 1
-#define SIMD_TYPE "SSE4.2"
+#define SIMD_TYPE "x86 (runtime detected)"
 #elif defined(HAVE_SIMD_NEON)
 #define HAVE_SIMD_STRING_SCAN 1
 #define SIMD_TYPE "NEON"
-#elif defined(HAVE_SIMD_SSE2)
-#define HAVE_SIMD_STRING_SCAN 1
-#define SIMD_TYPE "SSE2"
 #else
 #define SIMD_TYPE "none"
 #endif
 
+#if defined(HAVE_SIMD_SSE4_2)
+
+#define SIMD_MINIMUM_THRESHOLD 6
+
+extern void initialize_sse42(void);
+
+static inline OJ_TARGET_SSE42 __m128i vector_lookup_sse42(__m128i input, __m128i *lookup_table, int tab_size) {
+    // Extract high 4 bits to determine which 16-byte chunk (0-15)
+    __m128i hi_index = _mm_and_si128(_mm_srli_epi32(input, 4), _mm_set1_epi8(0x0F));
+
+    // Extract low 4 bits for index within the chunk (0-15)
+    __m128i low_index = _mm_and_si128(input, _mm_set1_epi8(0x0F));
+
+    // Perform lookups in all 16 tables
+    __m128i results[16];
+    for (int i = 0; i < tab_size; i++) {
+        results[i] = _mm_shuffle_epi8(lookup_table[i], low_index);
+    }
+
+    // Create masks for each chunk and blend results
+    __m128i final_result = _mm_setzero_si128();
+
+    for (int i = 0; i < tab_size; i++) {
+        __m128i mask          = _mm_cmpeq_epi8(hi_index, _mm_set1_epi8(i));
+        __m128i masked_result = _mm_and_si128(mask, results[i]);
+        final_result          = _mm_or_si128(final_result, masked_result);
+    }
+
+    return final_result;
+}
+
+#endif
+
 #endif /* OJ_SIMD_H */

data/lib/oj/version.rb

@@ -1,4 +1,4 @@
 module Oj
   # Current version of the module.
-  VERSION = '3.16.13'
+  VERSION = '3.16.14'
 end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: oj
 version: !ruby/object:Gem::Version
-  version: 3.16.13
+  version: 3.16.14
 platform: ruby
 authors:
 - Peter Ohler
@@ -23,20 +23,6 @@ dependencies:
     - - ">="
       - !ruby/object:Gem::Version
         version: '3.0'
-- !ruby/object:Gem::Dependency
-  name: ostruct
-  requirement: !ruby/object:Gem::Requirement
-    requirements:
-    - - ">="
-      - !ruby/object:Gem::Version
-        version: '0.2'
-  type: :runtime
-  prerelease: false
-  version_requirements: !ruby/object:Gem::Requirement
-    requirements:
-    - - ">="
-      - !ruby/object:Gem::Version
-        version: '0.2'
 - !ruby/object:Gem::Dependency
   name: minitest
   requirement: !ruby/object:Gem::Requirement
@@ -229,7 +215,7 @@ required_rubygems_version: !ruby/object:Gem::Requirement
     - !ruby/object:Gem::Version
       version: '0'
 requirements: []
-rubygems_version: 3.6.9
+rubygems_version: 4.0.3
 specification_version: 4
 summary: A fast JSON parser and serializer.
 test_files: []