oj 3.13.17 → 3.13.18

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: b2ba7c6f3821ac59250c2893d01269cd45bb2ec5eb4b06babccf480a07fb8862
-  data.tar.gz: 666b047d4ac8f5637e6b1e8ca995bceb105197eabc4664936cc72986280684c2
+  metadata.gz: e7ca20e7921e7b9a3e2838ecbe3505ecb736bce338113a3db3187f411d1762ca
+  data.tar.gz: 6c09a24ddb7017caa2a2850076a6b0fcbccc6ed27cd6e28e28de7fadacbaa7fe
 SHA512:
-  metadata.gz: 85005fe527745233220a680adf9bf3c24bb711abbe6b6758ccff236398c7f1218c9bc586eea7954388548772b17073e5540595d160948e19fffb94e535363ea4
-  data.tar.gz: 5b8b227c2d7e8156adb1adccff99ca18a8f2735ec3b7af9f8ee3f2b747427459c497ae738fe6c941d3df6e5131342677b1e8c08f208edeb7988697f80499339d
+  metadata.gz: 9e1bb01e19979391421d87b0598ff543ce3035a86fcbb06a68b63d4b56f4194739213d506fdd0492f5cf4a2aabd037872935fcb16f5d083fd6e532c6aa869b8d
+  data.tar.gz: 29db51398d6008cbcb827af666da0b41b750f6aea3e561fe37b2354f48b88d2bf393dcbfed931679072042bd19cf179ec0ae674027527b627391c2c4792f3c63

data/CHANGELOG.md

@@ -1,5 +1,9 @@
 # CHANGELOG
 
+## 3.13.18 - 2022-07-25
+
+- Fixed SSE detection at run time.
+
 ## 3.13.17 - 2022-07-15
 
 - Fixed Oj::Parser to detect unterminated arrays and objects.

data/ext/oj/custom.c

@@ -484,7 +484,7 @@ static VALUE dump_common(VALUE obj, int depth, Out out) {
         const char *   s;
         int            len;
 
-        if (Yes == out->opts->trace) {
+        if (RB_UNLIKELY(Yes == out->opts->trace)) {
             oj_trace("to_json", obj, __FILE__, __LINE__, depth + 1, TraceRubyIn);
         }
         if (0 == rb_obj_method_arity(obj, oj_to_json_id)) {
@@ -492,7 +492,7 @@ static VALUE dump_common(VALUE obj, int depth, Out out) {
         } else {
             rs = rb_funcall2(obj, oj_to_json_id, out->argc, out->argv);
         }
-        if (Yes == out->opts->trace) {
+        if (RB_UNLIKELY(Yes == out->opts->trace)) {
             oj_trace("to_json", obj, __FILE__, __LINE__, depth + 1, TraceRubyOut);
         }
         s   = RSTRING_PTR(rs);
@@ -504,7 +504,7 @@ static VALUE dump_common(VALUE obj, int depth, Out out) {
     } else if (Yes == out->opts->as_json && rb_respond_to(obj, oj_as_json_id)) {
         volatile VALUE aj;
 
-        if (Yes == out->opts->trace) {
+        if (RB_UNLIKELY(Yes == out->opts->trace)) {
             oj_trace("as_json", obj, __FILE__, __LINE__, depth + 1, TraceRubyIn);
         }
         // Some classes elect to not take an options argument so check the arity
@@ -514,7 +514,7 @@ static VALUE dump_common(VALUE obj, int depth, Out out) {
         } else {
             aj = rb_funcall2(obj, oj_as_json_id, out->argc, out->argv);
         }
-        if (Yes == out->opts->trace) {
+        if (RB_UNLIKELY(Yes == out->opts->trace)) {
             oj_trace("as_json", obj, __FILE__, __LINE__, depth + 1, TraceRubyOut);
         }
         // Catch the obvious brain damaged recursive dumping.
@@ -885,7 +885,7 @@ static DumpFunc custom_funcs[] = {
 void oj_dump_custom_val(VALUE obj, int depth, Out out, bool as_ok) {
     int type = rb_type(obj);
 
-    if (Yes == out->opts->trace) {
+    if (RB_UNLIKELY(Yes == out->opts->trace)) {
         oj_trace("dump", obj, __FILE__, __LINE__, depth, TraceIn);
     }
     if (MAX_DEPTH < depth) {
@@ -896,14 +896,14 @@ void oj_dump_custom_val(VALUE obj, int depth, Out out, bool as_ok) {
 
         if (NULL != f) {
             f(obj, depth, out, true);
-            if (Yes == out->opts->trace) {
+            if (RB_UNLIKELY(Yes == out->opts->trace)) {
                 oj_trace("dump", obj, __FILE__, __LINE__, depth, TraceOut);
             }
             return;
         }
     }
     oj_dump_nil(Qnil, depth, out, false);
-    if (Yes == out->opts->trace) {
+    if (RB_UNLIKELY(Yes == out->opts->trace)) {
         oj_trace("dump", Qnil, __FILE__, __LINE__, depth, TraceOut);
     }
 }

data/ext/oj/dump.c

@@ -736,11 +736,11 @@ void oj_dump_raw_json(VALUE obj, int depth, Out out) {
     } else {
         volatile VALUE jv;
 
-        if (Yes == out->opts->trace) {
+        if (RB_UNLIKELY(Yes == out->opts->trace)) {
             oj_trace("raw_json", obj, __FILE__, __LINE__, depth + 1, TraceRubyIn);
         }
         jv = rb_funcall(obj, oj_raw_json_id, 2, RB_INT2NUM(depth), RB_INT2NUM(out->indent));
-        if (Yes == out->opts->trace) {
+        if (RB_UNLIKELY(Yes == out->opts->trace)) {
             oj_trace("raw_json", obj, __FILE__, __LINE__, depth + 1, TraceRubyOut);
         }
         oj_dump_raw(RSTRING_PTR(jv), (size_t)RSTRING_LEN(jv), out);

data/ext/oj/dump_compat.c

@@ -109,7 +109,7 @@ dump_to_json(VALUE obj, Out out) {
     const char		*s;
     int			len;
 
-    if (Yes == out->opts->trace) {
+    if (RB_UNLIKELY(Yes == out->opts->trace)) {
 	oj_trace("to_json", obj, __FILE__, __LINE__, 0, TraceRubyIn);
     }
     if (0 == rb_obj_method_arity(obj, oj_to_json_id)) {
@@ -117,7 +117,7 @@ dump_to_json(VALUE obj, Out out) {
     } else {
 	rs = rb_funcall2(obj, oj_to_json_id, out->argc, out->argv);
     }
-    if (Yes == out->opts->trace) {
+    if (RB_UNLIKELY(Yes == out->opts->trace)) {
 	oj_trace("to_json", obj, __FILE__, __LINE__, 0, TraceRubyOut);
     }
 
@@ -893,7 +893,7 @@ void
 oj_dump_compat_val(VALUE obj, int depth, Out out, bool as_ok) {
     int	type = rb_type(obj);
 
-    if (Yes == out->opts->trace) {
+    if (RB_UNLIKELY(Yes == out->opts->trace)) {
 	oj_trace("dump", obj, __FILE__, __LINE__, depth, TraceIn);
     }
     if (out->opts->dump_opts.max_depth <= depth) {
@@ -918,14 +918,14 @@ oj_dump_compat_val(VALUE obj, int depth, Out out, bool as_ok) {
 
 	if (NULL != f) {
 	    f(obj, depth, out, as_ok);
-	    if (Yes == out->opts->trace) {
+	    if (RB_UNLIKELY(Yes == out->opts->trace)) {
 		oj_trace("dump", obj, __FILE__, __LINE__, depth, TraceOut);
 	    }
 	    return;
 	}
     }
     oj_dump_nil(Qnil, depth, out, false);
-    if (Yes == out->opts->trace) {
+    if (RB_UNLIKELY(Yes == out->opts->trace)) {
 	oj_trace("dump", Qnil, __FILE__, __LINE__, depth, TraceOut);
     }
 }

data/ext/oj/dump_object.c

@@ -682,7 +682,7 @@ static DumpFunc obj_funcs[] = {
 void oj_dump_obj_val(VALUE obj, int depth, Out out) {
     int type = rb_type(obj);
 
-    if (Yes == out->opts->trace) {
+    if (RB_UNLIKELY(Yes == out->opts->trace)) {
         oj_trace("dump", obj, __FILE__, __LINE__, depth, TraceIn);
     }
     if (MAX_DEPTH < depth) {
@@ -693,14 +693,14 @@ void oj_dump_obj_val(VALUE obj, int depth, Out out) {
 
         if (NULL != f) {
             f(obj, depth, out, false);
-            if (Yes == out->opts->trace) {
+            if (RB_UNLIKELY(Yes == out->opts->trace)) {
                 oj_trace("dump", obj, __FILE__, __LINE__, depth, TraceOut);
             }
             return;
         }
     }
     oj_dump_nil(Qnil, depth, out, false);
-    if (Yes == out->opts->trace) {
+    if (RB_UNLIKELY(Yes == out->opts->trace)) {
         oj_trace("dump", Qnil, __FILE__, __LINE__, depth, TraceOut);
     }
 }

data/ext/oj/dump_strict.c

@@ -338,7 +338,7 @@ static DumpFunc strict_funcs[] = {
 void oj_dump_strict_val(VALUE obj, int depth, Out out) {
     int type = rb_type(obj);
 
-    if (Yes == out->opts->trace) {
+    if (RB_UNLIKELY(Yes == out->opts->trace)) {
         oj_trace("dump", obj, __FILE__, __LINE__, depth, TraceIn);
     }
     if (MAX_DEPTH < depth) {
@@ -349,7 +349,7 @@ void oj_dump_strict_val(VALUE obj, int depth, Out out) {
 
         if (NULL != f) {
             f(obj, depth, out, false);
-            if (Yes == out->opts->trace) {
+            if (RB_UNLIKELY(Yes == out->opts->trace)) {
                 oj_trace("dump", obj, __FILE__, __LINE__, depth, TraceOut);
             }
             return;
@@ -386,7 +386,7 @@ static DumpFunc null_funcs[] = {
 void oj_dump_null_val(VALUE obj, int depth, Out out) {
     int type = rb_type(obj);
 
-    if (Yes == out->opts->trace) {
+    if (RB_UNLIKELY(Yes == out->opts->trace)) {
         oj_trace("dump", obj, __FILE__, __LINE__, depth, TraceOut);
     }
     if (MAX_DEPTH < depth) {
@@ -397,14 +397,14 @@ void oj_dump_null_val(VALUE obj, int depth, Out out) {
 
         if (NULL != f) {
             f(obj, depth, out, false);
-            if (Yes == out->opts->trace) {
+            if (RB_UNLIKELY(Yes == out->opts->trace)) {
                 oj_trace("dump", obj, __FILE__, __LINE__, depth, TraceOut);
             }
             return;
         }
     }
     oj_dump_nil(Qnil, depth, out, false);
-    if (Yes == out->opts->trace) {
+    if (RB_UNLIKELY(Yes == out->opts->trace)) {
         oj_trace("dump", Qnil, __FILE__, __LINE__, depth, TraceOut);
     }
 }

data/ext/oj/extconf.rb

@@ -34,9 +34,21 @@ have_func('rb_hash_bulk_insert', 'ruby.h') unless '2' == version[0] && '6' == ve
 
 dflags['OJ_DEBUG'] = true unless ENV['OJ_DEBUG'].nil?
 
-if try_cflags('-msse4.2')
+src =<<~SRC
+#include <string.h>
+#include <nmmintrin.h>
+int main(int argc, char **argv) {
+    const char *str = "hello               ";
+    const char chars[16] = "\x00\\\"";
+    const __m128i terminate = _mm_loadu_si128((const __m128i *)&chars[0]);
+    const __m128i string = _mm_loadu_si128((const __m128i *)str);
+    int r = _mm_cmpestri(terminate, 3, string, 16, _SIDD_UBYTE_OPS | _SIDD_CMP_EQUAL_ANY | _SIDD_LEAST_SIGNIFICANT);
+    return r == 16 ? 0 : 1;
+}
+SRC
+
+if try_run(src, '-msse4.2')
   $CPPFLAGS += ' -msse4.2'
-  dflags['OJ_USE_SSE4_2'] = 1
 end
 
 dflags.each do |k,v|

data/ext/oj/oj.c

@@ -2043,4 +2043,5 @@ void Init_oj(void) {
     oj_init_doc();
 
     oj_parser_init();
+    oj_scanner_init();
 }

data/ext/oj/parse.c

@@ -192,7 +192,26 @@ static inline const char *scan_string_noSIMD(const char *str, const char *end) {
     return str;
 }
 
-#if defined(OJ_USE_SSE4_2)
+// Taken from Tensorflow:
+// https://github.com/tensorflow/tensorflow/blob/5dcfc51118817f27fad5246812d83e5dccdc5f72/tensorflow/core/lib/hash/crc32c_accelerate.cc#L21-L38
+#ifdef __SSE4_2__
+#if defined(__x86_64__) && defined(__GNUC__) && \
+    (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 8))
+#define USE_SSE_DETECT 1
+#elif defined(__x86_64__) && defined(__clang__)
+#if __has_builtin(__builtin_cpu_supports)
+#define USE_SSE_DETECT 1
+#endif
+#endif
+#endif /* __SSE4_2__ */
+
+// This version of Apple clang has a bug:
+// https://llvm.org/bugs/show_bug.cgi?id=25510
+#if defined(__APPLE__) && (__clang_major__ <= 8)
+#undef USE_SSE_DETECT
+#endif
+
+#ifdef USE_SSE_DETECT
 #include <nmmintrin.h>
 
 static inline const char *scan_string_SIMD(const char *str, const char *end) {
@@ -213,6 +232,25 @@ static inline const char *scan_string_SIMD(const char *str, const char *end) {
 }
 #endif
 
+static bool cpu_supports_sse42(void) {
+#if USE_SSE_DETECT
+    __builtin_cpu_init();
+    return (__builtin_cpu_supports("sse4.2"));
+#else
+    return false;
+#endif
+}
+
+static const char *(*scan_func) (const char *str, const char *end) = scan_string_noSIMD;
+
+void oj_scanner_init(void) {
+    if (cpu_supports_sse42()) {
+#if USE_SSE_DETECT
+        scan_func = scan_string_SIMD;
+#endif
+    }
+}
+
 // entered at /
 static void read_escaped_str(ParseInfo pi, const char *start) {
     struct _buf buf;
@@ -225,11 +263,7 @@ static void read_escaped_str(ParseInfo pi, const char *start) {
     buf_append_string(&buf, start, cnt);
 
     for (s = pi->cur; '"' != *s;) {
-#if defined(OJ_USE_SSE4_2)
-        const char *scanned = scan_string_SIMD(s, pi->end);
-#else
-        const char *scanned = scan_string_noSIMD(s, pi->end);
-#endif
+        const char *scanned = scan_func(s, pi->end);
         if (scanned >= pi->end) {
             oj_set_error_at(pi,
                             oj_parse_error_class,
@@ -369,11 +403,7 @@ static void read_str(ParseInfo pi) {
     const char *str    = pi->cur;
     Val         parent = stack_peek(&pi->stack);
 
-#if defined(OJ_USE_SSE4_2)
-    pi->cur = scan_string_SIMD(pi->cur, pi->end);
-#else
-    pi->cur = scan_string_noSIMD(pi->cur, pi->end);
-#endif
+    pi->cur = scan_func(pi->cur, pi->end);
     if (RB_UNLIKELY(pi->end <= pi->cur)) {
         oj_set_error_at(pi,
                         oj_parse_error_class,
@@ -503,33 +533,31 @@ static void read_num(ParseInfo pi) {
         int  dec_cnt = 0;
         bool zero1   = false;
 
+        // Skip leading zeros.
+        for (; '0' == *pi->cur; pi->cur++) {
+            zero1 = true;
+        }
+
         for (; '0' <= *pi->cur && *pi->cur <= '9'; pi->cur++) {
-            if (0 == ni.i && '0' == *pi->cur) {
-                zero1 = true;
-            }
-            if (0 < ni.i) {
-                dec_cnt++;
-            }
-            if (!ni.big) {
-                int d = (*pi->cur - '0');
+            int d = (*pi->cur - '0');
 
-                if (0 < d) {
-                    if (zero1 && CompatMode == pi->options.mode) {
-                        oj_set_error_at(pi,
-                                        oj_parse_error_class,
-                                        __FILE__,
-                                        __LINE__,
-                                        "not a number");
-                        return;
-                    }
-                    zero1 = false;
-                }
-                ni.i = ni.i * 10 + d;
-                if (INT64_MAX <= ni.i || DEC_MAX < dec_cnt) {
-                    ni.big = 1;
-                }
+            if (RB_LIKELY(0 != ni.i)) {
+                dec_cnt++;
             }
+            ni.i = ni.i * 10 + d;
+        }
+        if (RB_UNLIKELY(0 != ni.i && zero1 && CompatMode == pi->options.mode)) {
+            oj_set_error_at(pi,
+                            oj_parse_error_class,
+                            __FILE__,
+                            __LINE__,
+                            "not a number");
+            return;
+        }
+        if (INT64_MAX <= ni.i || DEC_MAX < dec_cnt) {
+            ni.big = true;
         }
+
         if ('.' == *pi->cur) {
             pi->cur++;
             // A trailing . is not a valid decimal but if encountered allow it
@@ -549,25 +577,20 @@ static void read_num(ParseInfo pi) {
             for (; '0' <= *pi->cur && *pi->cur <= '9'; pi->cur++) {
                 int d = (*pi->cur - '0');
 
-                if (0 < ni.num || 0 < ni.i) {
+                if (RB_LIKELY(0 != ni.num || 0 != ni.i)) {
                     dec_cnt++;
                 }
-                if (INT64_MAX <= ni.div) {
-                    if (!ni.no_big) {
-                        ni.big = true;
-                    }
-                } else {
-                    ni.num = ni.num * 10 + d;
-                    ni.div *= 10;
-                    ni.di++;
-                    if (INT64_MAX <= ni.div || DEC_MAX < dec_cnt) {
-                        if (!ni.no_big) {
-                            ni.big = true;
-                        }
-                    }
-                }
+                ni.num = ni.num * 10 + d;
+                ni.div *= 10;
+                ni.di++;
             }
         }
+        if (INT64_MAX <= ni.div || DEC_MAX < dec_cnt) {
+            if (!ni.no_big) {
+                ni.big = true;
+            }
+        }
+
         if ('e' == *pi->cur || 'E' == *pi->cur) {
             int eneg = 0;
 

data/ext/oj/parse.h

@@ -97,6 +97,8 @@ static inline void parse_info_init(ParseInfo pi) {
     memset(pi, 0, sizeof(struct _parseInfo));
 }
 
+extern void oj_scanner_init();
+
 static inline bool empty_ok(Options options) {
     switch (options->mode) {
     case ObjectMode:

data/ext/oj/rails.c

@@ -517,7 +517,7 @@ static void dump_as_string(VALUE obj, int depth, Out out, bool as_ok) {
 static void dump_as_json(VALUE obj, int depth, Out out, bool as_ok) {
     volatile VALUE ja;
 
-    if (Yes == out->opts->trace) {
+    if (RB_UNLIKELY(Yes == out->opts->trace)) {
         oj_trace("as_json", obj, __FILE__, __LINE__, depth + 1, TraceRubyIn);
     }
     // Some classes elect to not take an options argument so check the arity
@@ -527,7 +527,7 @@ static void dump_as_json(VALUE obj, int depth, Out out, bool as_ok) {
     } else {
         ja = rb_funcall2(obj, oj_as_json_id, out->argc, out->argv);
     }
-    if (Yes == out->opts->trace) {
+    if (RB_UNLIKELY(Yes == out->opts->trace)) {
         oj_trace("as_json", obj, __FILE__, __LINE__, depth + 1, TraceRubyOut);
     }
 
@@ -1464,7 +1464,7 @@ static DumpFunc rails_funcs[] = {
 static void dump_rails_val(VALUE obj, int depth, Out out, bool as_ok) {
     int type = rb_type(obj);
 
-    if (Yes == out->opts->trace) {
+    if (RB_UNLIKELY(Yes == out->opts->trace)) {
         oj_trace("dump", obj, __FILE__, __LINE__, depth, TraceIn);
     }
     if (MAX_DEPTH < depth) {
@@ -1475,14 +1475,14 @@ static void dump_rails_val(VALUE obj, int depth, Out out, bool as_ok) {
 
         if (NULL != f) {
             f(obj, depth, out, as_ok);
-            if (Yes == out->opts->trace) {
+            if (RB_UNLIKELY(Yes == out->opts->trace)) {
                 oj_trace("dump", obj, __FILE__, __LINE__, depth, TraceOut);
             }
             return;
         }
     }
     oj_dump_nil(Qnil, depth, out, false);
-    if (Yes == out->opts->trace) {
+    if (RB_UNLIKELY(Yes == out->opts->trace)) {
         oj_trace("dump", Qnil, __FILE__, __LINE__, depth, TraceOut);
     }
 }

data/ext/oj/wab.c

@@ -266,7 +266,7 @@ static DumpFunc wab_funcs[] = {
 void oj_dump_wab_val(VALUE obj, int depth, Out out) {
     int type = rb_type(obj);
 
-    if (Yes == out->opts->trace) {
+    if (RB_UNLIKELY(Yes == out->opts->trace)) {
         oj_trace("dump", obj, __FILE__, __LINE__, depth, TraceIn);
     }
     if (MAX_DEPTH < depth) {
@@ -277,7 +277,7 @@ void oj_dump_wab_val(VALUE obj, int depth, Out out) {
 
         if (NULL != f) {
             f(obj, depth, out, false);
-            if (Yes == out->opts->trace) {
+            if (RB_UNLIKELY(Yes == out->opts->trace)) {
                 oj_trace("dump", obj, __FILE__, __LINE__, depth, TraceOut);
             }
             return;

data/lib/oj/version.rb

@@ -1,5 +1,5 @@
 
 module Oj
   # Current version of the module.
-  VERSION = '3.13.17'
+  VERSION = '3.13.18'
 end

metadata

@@ -1,14 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: oj
 version: !ruby/object:Gem::Version
-  version: 3.13.17
+  version: 3.13.18
 platform: ruby
 authors:
 - Peter Ohler
 autorequire:
 bindir: bin
 cert_chain: []
-date: 2022-07-15 00:00:00.000000000 Z
+date: 2022-07-25 00:00:00.000000000 Z
 dependencies:
 - !ruby/object:Gem::Dependency
   name: rake-compiler