use Windows intrinsics when the CPU supports them

Author: burgerrg

c/alloc.c

@@ -15,8 +15,9 @@
  */
 
 #include "system.h"
-#if defined(_MSC_VER)
+#ifdef _MSC_VER
 #include <intrin.h> /* for Spopcount below */
+static int has_popcnt = 0;
 #endif
 
 /* locally defined functions */
@@ -26,6 +27,11 @@ void S_alloc_init(void) {
     ISPC s; IGEN g; UINT i;
 
     if (S_boot_time) {
+#ifdef _MSC_VER
+      int cpuinfo[4];
+      __cpuid(cpuinfo, 1);
+      has_popcnt = (cpuinfo[2] & (1 << 23)) != 0; /* ECX bit 23 */
+#endif
       ptr tc = TO_PTR(S_G.thread_context);
 
       GCDATA(tc) = TO_PTR(&S_G.main_thread_gc);
@@ -802,18 +808,21 @@ ptr S_null_immutable_string(void) {
 
 int Spopcount(uptr x) {
 #if defined(__clang__) || defined(__GNUC__)
-#if ptr_bits == 32
-  return __builtin_popcount(x);
-#else
-  return __builtin_popcountl(x);
-#endif
-#elif defined(_MSC_VER)
-#if ptr_bits == 32
-  return (int)__popcnt(x);
-#else
-  return (int)__popcnt64(x);
-#endif
+  if (sizeof(x) <= sizeof(unsigned long))
+    return __builtin_popcountl((unsigned long)x);
+  else
+    return __builtin_popcountll((unsigned long long)x);
 #else
+# if defined(_MSC_VER)
+  if (has_popcnt) {
+#   if defined(_WIN64)
+    return (int)__popcnt64((unsigned __int64)x);
+#   else
+    return (int)__popcnt((unsigned int)x);
+#   endif
+# endif
+  }
+
   /* Kernighan's method */
   int count = 0;
   while (x != 0) {