Reduce bloat in defines.hpp

Author: ZeroMemes

include/libhat/defines.hpp

@@ -2,32 +2,10 @@
 
 #include "platform.h"
 
-// Macros wrapping intrinsics
-#if defined(LIBHAT_X86_64)
-    #define LIBHAT_TZCNT64(num) _tzcnt_u64(num)
-    #define LIBHAT_BLSR64(num) _blsr_u64(num)
-#elif defined(LIBHAT_X86)
-    #include <bit>
-    #define LIBHAT_TZCNT64(num) std::countl_zero(num)
-    #define LIBHAT_BLSR64(num) num & (num - 1)
-#endif
-
 #ifdef _MSC_VER
-    #include <intrin.h>
-
-    namespace hat::detail {
-        inline unsigned long bsf(unsigned long num) noexcept {
-            unsigned long offset;
-            _BitScanForward(&offset, num);
-            return offset;
-        }
-    }
-
     #define LIBHAT_RETURN_ADDRESS() _ReturnAddress()
-    #define LIBHAT_BSF32(num) hat::detail::bsf(num)
 #else
     #define LIBHAT_RETURN_ADDRESS() __builtin_extract_return_addr(__builtin_return_address(0))
-    #define LIBHAT_BSF32(num) __builtin_ctz(num)
 #endif
 
 #if __cpp_if_consteval >= 202106L

src/arch/x86/AVX512.cpp

@@ -63,7 +63,7 @@ namespace hat::detail {
             }
 
             while (mask) {
-                const auto offset = LIBHAT_TZCNT64(mask);
+                const auto offset = _tzcnt_u64(mask);
                 const auto i = reinterpret_cast<const std::byte*>(&it) + offset - cmpIndex;
                 if constexpr (veccmp) {
                     const auto data = _mm512_loadu_si512(i);
@@ -79,7 +79,7 @@ namespace hat::detail {
                         return i;
                     }
                 }
-                mask = LIBHAT_BLSR64(mask);
+                mask = _blsr_u64(mask);
             }
         }
 

src/arch/x86/SSE.cpp

@@ -6,6 +6,22 @@
 
 #include <immintrin.h>
 
+#ifdef _MSC_VER
+    #include <intrin.h>
+
+    namespace hat::detail {
+        inline unsigned long bsf(unsigned long num) noexcept {
+            unsigned long offset;
+            _BitScanForward(&offset, num);
+            return offset;
+        }
+    }
+
+    #define LIBHAT_BSF32(num) hat::detail::bsf(num)
+#else
+    #define LIBHAT_BSF32(num) __builtin_ctz(num)
+#endif
+
 namespace hat::detail {
 
     inline void load_signature_128(const signature_view signature, __m128i& bytes, __m128i& mask) {