oj 3.16.12 → 3.16.13

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 6e2053531cd4c7c7b49bf16ddafbfda868e67e66c1bfbec1b06daaa0ba3f1c45
-  data.tar.gz: ce029b2f90660922dd8fb335c23463a1fe1d81317d56e48f2cfddc9271de2d15
+  metadata.gz: fc0290fa1cfe6af1094de1d7188836e0c09cb04f2f08401de118253026604650
+  data.tar.gz: de5258e96984a21afb2fac946fe28ad255926893a6157f2874445edf10aa8bbe
 SHA512:
-  metadata.gz: deb7f1447b5022adad6d7387b8a8bfd866d399abc2d9e434f7e6d321fa73cb1738ff9aa7ee22ac064c455d5d3951ba7469a30720c9474af5e96c70eaa5b5303a
-  data.tar.gz: b9d28d76c714947c1e6b133225e348838ad13f1c8b7dc82f0fee261272259cd0e83b911f2f2762c55a719f93e669275ad32d393f371c264a4e587e49b1c3a84b
+  metadata.gz: d7870818fd86043a17b834756b67a4009a6f7ef60baf53b02a0b0d4431ccba723d9e533553bea04ae46ae9f233e447b79c4106e6827782bd0c2ffb9c332081a3
+  data.tar.gz: fd3966ac7fb5da9f1a5ebb68f4a8f5b9a5f9fa1a1255e93dfef078f66f00a6af5bb7e37676441f7d6229b29222741a2bc7b75164fd445a39b906ef904946d41b

data/CHANGELOG.md

@@ -1,5 +1,9 @@
 # CHANGELOG
 
+## 3.16.13 - 2025-12-05
+
+- Fixed rails encoding for Hash and Array subclasses.
+
 ## 3.16.12 - 2025-10-29
 
 - Fixed dump realloc bug that occurred when using the compat mode dump options.

data/ext/oj/extconf.rb

@@ -35,13 +35,12 @@ have_func('rb_ext_ractor_safe', 'ruby.h')
 
 dflags['OJ_DEBUG'] = true unless ENV['OJ_DEBUG'].nil?
 
-if with_config('--with-sse42')
-  if try_cflags('-msse4.2')
-    $CPPFLAGS += ' -msse4.2'
-    dflags['OJ_USE_SSE4_2'] = 1
-  else
-    warn 'SSE 4.2 is not supported on this platform.'
-  end
+# 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')
+  $CPPFLAGS += ' -msse2'
 end
 
 if enable_config('trace-log', false)

data/ext/oj/parse.c

@@ -15,12 +15,9 @@
 #include "mem.h"
 #include "oj.h"
 #include "rxclass.h"
+#include "simd.h"
 #include "val_stack.h"
 
-#ifdef OJ_USE_SSE4_2
-#include <nmmintrin.h>
-#endif
-
 // Workaround in case INFINITY is not defined in math.h or if the OS is CentOS
 #define OJ_INFINITY (1.0 / 0.0)
 
@@ -202,23 +199,143 @@ static inline const char *scan_string_noSIMD(const char *str, const char *end) {
     return str;
 }
 
-#ifdef OJ_USE_SSE4_2
-static inline const char *scan_string_SIMD(const char *str, const char *end) {
-    static const char chars[16] = "\x00\\\"";
-    const __m128i     terminate = _mm_loadu_si128((const __m128i *)&chars[0]);
-    const char       *_end      = (const char *)(end - 16);
+#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) {
+    static const char chars[16]   = "\x00\\\"";
+    const __m128i     terminate   = _mm_loadu_si128((const __m128i *)&chars[0]);
+    const char       *safe_end_64 = end - 64;
+    const char       *safe_end_16 = end - 16;
+
+    // Process 64 bytes at a time with parallel SIMD operations
+    // 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);
+
+        // Load and compare 4 chunks in parallel
+        const __m128i chunk0 = _mm_loadu_si128((const __m128i *)(str));
+        const __m128i chunk1 = _mm_loadu_si128((const __m128i *)(str + 16));
+        const __m128i chunk2 = _mm_loadu_si128((const __m128i *)(str + 32));
+        const __m128i chunk3 = _mm_loadu_si128((const __m128i *)(str + 48));
+
+        const int r0 = _mm_cmpestri(terminate,
+                                    3,
+                                    chunk0,
+                                    16,
+                                    _SIDD_UBYTE_OPS | _SIDD_CMP_EQUAL_ANY | _SIDD_LEAST_SIGNIFICANT);
+        if (__builtin_expect(r0 != 16, 0))
+            return str + r0;
+
+        const int r1 = _mm_cmpestri(terminate,
+                                    3,
+                                    chunk1,
+                                    16,
+                                    _SIDD_UBYTE_OPS | _SIDD_CMP_EQUAL_ANY | _SIDD_LEAST_SIGNIFICANT);
+        if (__builtin_expect(r1 != 16, 0))
+            return str + 16 + r1;
+
+        const int r2 = _mm_cmpestri(terminate,
+                                    3,
+                                    chunk2,
+                                    16,
+                                    _SIDD_UBYTE_OPS | _SIDD_CMP_EQUAL_ANY | _SIDD_LEAST_SIGNIFICANT);
+        if (__builtin_expect(r2 != 16, 0))
+            return str + 32 + r2;
+
+        const int r3 = _mm_cmpestri(terminate,
+                                    3,
+                                    chunk3,
+                                    16,
+                                    _SIDD_UBYTE_OPS | _SIDD_CMP_EQUAL_ANY | _SIDD_LEAST_SIGNIFICANT);
+        if (__builtin_expect(r3 != 16, 0))
+            return str + 48 + r3;
+
+        str += 64;
+    }
 
-    for (; str <= _end; str += 16) {
+    // Handle remaining 16-byte chunks
+    for (; str <= safe_end_16; str += 16) {
         const __m128i string = _mm_loadu_si128((const __m128i *)str);
         const int     r      = _mm_cmpestri(terminate,
                                    3,
                                    string,
                                    16,
                                    _SIDD_UBYTE_OPS | _SIDD_CMP_EQUAL_ANY | _SIDD_LEAST_SIGNIFICANT);
-        if (r != 16) {
-            str = (char *)(str + r);
-            return str;
-        }
+        if (r != 16)
+            return str + r;
+    }
+
+    return scan_string_noSIMD(str, end);
+}
+#endif
+
+#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) {
+    const char *safe_end_64 = end - 64;
+    const char *safe_end_16 = end - 16;
+
+    // Create comparison vectors for our three special characters
+    const __m128i null_char = _mm_setzero_si128();
+    const __m128i backslash = _mm_set1_epi8('\\');
+    const __m128i quote     = _mm_set1_epi8('"');
+
+    // Process 64 bytes at a time for better throughput
+    while (str <= safe_end_64) {
+        __builtin_prefetch(str + 64, 0, 0);
+
+        // Load 4 chunks
+        const __m128i chunk0 = _mm_loadu_si128((const __m128i *)(str));
+        const __m128i chunk1 = _mm_loadu_si128((const __m128i *)(str + 16));
+        const __m128i chunk2 = _mm_loadu_si128((const __m128i *)(str + 32));
+        const __m128i chunk3 = _mm_loadu_si128((const __m128i *)(str + 48));
+
+        // Compare all chunks (allows CPU to parallelize)
+        const __m128i cmp0 = _mm_or_si128(
+            _mm_or_si128(_mm_cmpeq_epi8(chunk0, null_char), _mm_cmpeq_epi8(chunk0, backslash)),
+            _mm_cmpeq_epi8(chunk0, quote));
+        const __m128i cmp1 = _mm_or_si128(
+            _mm_or_si128(_mm_cmpeq_epi8(chunk1, null_char), _mm_cmpeq_epi8(chunk1, backslash)),
+            _mm_cmpeq_epi8(chunk1, quote));
+        const __m128i cmp2 = _mm_or_si128(
+            _mm_or_si128(_mm_cmpeq_epi8(chunk2, null_char), _mm_cmpeq_epi8(chunk2, backslash)),
+            _mm_cmpeq_epi8(chunk2, quote));
+        const __m128i cmp3 = _mm_or_si128(
+            _mm_or_si128(_mm_cmpeq_epi8(chunk3, null_char), _mm_cmpeq_epi8(chunk3, backslash)),
+            _mm_cmpeq_epi8(chunk3, quote));
+
+        // Convert to masks
+        int mask0 = _mm_movemask_epi8(cmp0);
+        if (__builtin_expect(mask0 != 0, 0))
+            return str + __builtin_ctz(mask0);
+
+        int mask1 = _mm_movemask_epi8(cmp1);
+        if (__builtin_expect(mask1 != 0, 0))
+            return str + 16 + __builtin_ctz(mask1);
+
+        int mask2 = _mm_movemask_epi8(cmp2);
+        if (__builtin_expect(mask2 != 0, 0))
+            return str + 32 + __builtin_ctz(mask2);
+
+        int mask3 = _mm_movemask_epi8(cmp3);
+        if (__builtin_expect(mask3 != 0, 0))
+            return str + 48 + __builtin_ctz(mask3);
+
+        str += 64;
+    }
+
+    // Handle remaining 16-byte chunks
+    for (; str <= safe_end_16; str += 16) {
+        const __m128i chunk   = _mm_loadu_si128((const __m128i *)str);
+        const __m128i matches = _mm_or_si128(
+            _mm_or_si128(_mm_cmpeq_epi8(chunk, null_char), _mm_cmpeq_epi8(chunk, backslash)),
+            _mm_cmpeq_epi8(chunk, quote));
+        int mask = _mm_movemask_epi8(matches);
+        if (mask != 0)
+            return str + __builtin_ctz(mask);
     }
 
     return scan_string_noSIMD(str, end);
@@ -228,9 +345,12 @@ static inline const char *scan_string_SIMD(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) {
-#ifdef OJ_USE_SSE4_2
-    scan_func = scan_string_SIMD;
+#ifdef HAVE_SIMD_SSE4_2
+    scan_func = scan_string_SSE42;
+#elif defined(HAVE_SIMD_SSE2)
+    scan_func = scan_string_SSE2;
 #endif
+    // Note: ARM NEON string scanning would be added here if needed
 }
 
 // entered at /

data/ext/oj/rails.c

@@ -661,13 +661,15 @@ static VALUE encoder_new(int argc, VALUE *argv, VALUE self) {
     Encoder e = OJ_R_ALLOC(struct _encoder);
 
     e->opts = oj_default_options;
-    e->arg  = Qnil;
     copy_opts(&ropts, &e->ropts);
 
     if (1 <= argc && Qnil != *argv) {
-        oj_parse_options(*argv, &e->opts);
         e->arg = *argv;
+    } else {
+        e->arg = rb_hash_new();
     }
+    oj_parse_options(*argv, &e->opts);
+
     return TypedData_Wrap_Struct(encoder_class, &oj_encoder_type, e);
 }
 

data/ext/oj/simd.h

@@ -1,10 +1,47 @@
 #ifndef OJ_SIMD_H
 #define OJ_SIMD_H
 
+// SIMD architecture detection and configuration
+// This header provides unified SIMD support across different CPU architectures
+
+// 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__)
+#define HAVE_SIMD_SSE4_2 1
+#include <nmmintrin.h>
+#endif
+
+// SSE2 support (fallback for older x86_64 CPUs - all x86_64 CPUs support SSE2)
+#if defined(__SSE2__) && !defined(HAVE_SIMD_SSE4_2)
+#define HAVE_SIMD_SSE2 1
+#include <emmintrin.h>
+#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
 
-#endif /* OJ_SIMD_H */
+// Define which SIMD implementation to use (priority order: SSE4.2 > NEON > SSE2)
+#if defined(HAVE_SIMD_SSE4_2)
+#define HAVE_SIMD_STRING_SCAN 1
+#define SIMD_TYPE "SSE4.2"
+#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
+
+#endif /* OJ_SIMD_H */

data/lib/oj/version.rb

@@ -1,4 +1,4 @@
 module Oj
   # Current version of the module.
-  VERSION = '3.16.12'
+  VERSION = '3.16.13'
 end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: oj
 version: !ruby/object:Gem::Version
-  version: 3.16.12
+  version: 3.16.13
 platform: ruby
 authors:
 - Peter Ohler