Add SIMD support for Windows when compiled by Visual C++.

ext/json/ext/json.h

@@ -94,6 +94,8 @@ typedef unsigned char _Bool;
 
 #if defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ && INTPTR_MAX == INT64_MAX
 #define JSON_CPU_LITTLE_ENDIAN_64BITS 1
+#elif defined(_MSC_VER) && (defined(_M_X64) || defined(_M_AMD64))
+#define JSON_CPU_LITTLE_ENDIAN_64BITS 1
 #else
 #define JSON_CPU_LITTLE_ENDIAN_64BITS 0
 #endif

ext/json/ext/parser/parser.c

@@ -86,7 +86,16 @@ static void rvalue_cache_insert_at(rvalue_cache *cache, int index, VALUE rstring
 #define rstring_cache_memcmp memcmp
 
 #if JSON_CPU_LITTLE_ENDIAN_64BITS
-#if __has_builtin(__builtin_bswap64)
+
+#if defined(JSON_SIMD_WINDOWS)
+#define HAVE_BYTE_SWAP64 1
+#define byte_swap64(x) _byteswap_uint64(x)
+#elif __has_builtin(__builtin_bswap64)
+#define HAVE_BYTE_SWAP64 1
+#define byte_swap64(x) __builtin_bswap64(x)
+#endif
+
+#if defined(HAVE_BYTE_SWAP64) && (HAVE_BYTE_SWAP64 == 1)
 #undef rstring_cache_memcmp
 ALWAYS_INLINE(static) int rstring_cache_memcmp(const char *str, const char *rptr, const long length)
 {
@@ -100,8 +109,8 @@ ALWAYS_INLINE(static) int rstring_cache_memcmp(const char *str, const char *rptr
         memcpy(&a, str + i, 8);
         memcpy(&b, rptr + i, 8);
         if (a != b) {
-            a = __builtin_bswap64(a);
-            b = __builtin_bswap64(b);
+            a = byte_swap64(a);
+            b = byte_swap64(b);
             return (a < b) ? -1 : 1;
         }
     }
@@ -114,6 +123,7 @@ ALWAYS_INLINE(static) int rstring_cache_memcmp(const char *str, const char *rptr
 
     return 0;
 }
+#undef byte_swap64
 #endif
 #endif
 

ext/json/ext/simd/conf.rb

@@ -5,6 +5,7 @@
 when /^(x86_64|x64)/
   header, type, init, extra = 'x86intrin.h', '__m128i', '_mm_set1_epi8(32)', 'if (__builtin_cpu_supports("sse2")) { printf("OK"); }'
 end
+# intrin.h
 if header
   if have_header(header) && try_compile(<<~SRC, '-Werror=implicit-function-declaration')
       #{cpp_include(header)}
@@ -17,7 +18,16 @@
     SRC
     $defs.push("-DJSON_ENABLE_SIMD")
   else
-    puts "Disable SIMD"
+    if RbConfig::CONFIG['host_os'] =~ /mswin/i
+        if have_header('intrin.h') && try_compile(<<~SRC, '-arch:SSE2')
+          #include <intrin.h>
+          int main() { __m128i x = _mm_setzero_si128(); return 0; }
+        SRC
+        $defs.push("-DJSON_ENABLE_SIMD")
+      end
+    else
+      puts "Disable SIMD"
+    end
   end
 end
 

ext/json/ext/simd/simd.h

@@ -22,12 +22,24 @@ typedef enum {
 # define HAVE_BUILTIN_CTZLL 0
 #endif
 
+#if defined(_MSC_VER) && defined(HAVE_INTRIN_H)
+#define JSON_SIMD_WINDOWS 1
+#include <intrin.h>
+#endif
+
 static inline uint32_t trailing_zeros64(uint64_t input)
 {
     JSON_ASSERT(input > 0); // __builtin_ctz(0) is undefined behavior
 
 #if HAVE_BUILTIN_CTZLL
     return __builtin_ctzll(input);
+#elif JSON_SIMD_WINDOWS
+    uint32_t index;
+    if (_BitScanForward64(&index, input)) {
+        return index;
+    } else {
+        return 64; // _BitScanForward64 returns 0 if input is 0, so there are 64 zeros.
+    }
 #else
     uint32_t trailing_zeros = 0;
     uint64_t temp = input;
@@ -45,6 +57,13 @@ static inline int trailing_zeros(int input)
 
 #if HAVE_BUILTIN_CTZLL
     return __builtin_ctz(input);
+#elif JSON_SIMD_WINDOWS
+    uint32_t index;
+    if (_BitScanForward(&index, input)) {
+        return index;
+    } else {
+        return 32; // _BitScanForward returns 0 if input is 0, so there are 32 zeros.
+    }
 #else
     int trailing_zeros = 0;
     int temp = input;
@@ -147,13 +166,16 @@ static inline uint8x16x4_t load_uint8x16_4(const unsigned char *table)
 
 #if defined(__amd64__) || defined(__amd64) || defined(__x86_64__) || defined(__x86_64) || defined(_M_X64) || defined(_M_AMD64)
 
+#if defined(HAVE_X86INTRIN_H) || defined(JSON_SIMD_WINDOWS)
+
+// We check for HAVE_X86INTRIN_H again to include the heder as if we have 'intrin.h' it's already been incldued.
 #ifdef HAVE_X86INTRIN_H
 #include <x86intrin.h>
+#endif
 
 #define HAVE_SIMD 1
 #define HAVE_SIMD_SSE2 1
 
-#ifdef HAVE_CPUID_H
 #define FIND_SIMD_IMPLEMENTATION_DEFINED 1
 
 #if defined(__clang__) || defined(__GNUC__)
@@ -192,15 +214,27 @@ ALWAYS_INLINE(static) TARGET_SSE2 int string_scan_simd_sse2(const char **ptr, co
     return 0;
 }
 
+#ifdef HAVE_CPUID_H
 #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 __has_builtin(__builtin_cpu_supports)
     if (__builtin_cpu_supports("sse2")) {
         return SIMD_SSE2;
     }
+#endif
+
+#if defined(JSON_SIMD_WINDOWS)
+    // https://learn.microsoft.com/en-us/cpp/intrinsics/cpuid-cpuidex?view=msvc-170
+    int cpuInfo[4] = {0};
+    __cpuid(cpuInfo, 1);
+    if ((cpuInfo[3] & (1 << 26)) != 0) {
+        return SIMD_SSE2;
+    }
+#endif
 
     return SIMD_NONE;
 }