Move GCC_ATTRIBUTE out of autoconf

acinclude.m4

@@ -334,9 +334,3 @@ AH_BOTTOM([#if C_ATTRIBUTE_FASTCALL
 #else
 #define DB_FASTCALL
 #endif])
-
-AH_BOTTOM([#if C_HAS_ATTRIBUTE
-#define GCC_ATTRIBUTE(x) __attribute__ ((x))
-#else
-#define GCC_ATTRIBUTE(x) /* attribute not supported */
-#endif])

include/compiler.h

@@ -22,6 +22,25 @@
 // This header wraps compiler-specific features, so they won't need to
 // be hacked into the buildsystem.
 
+// Modern C++ compilers have better support for feature testing using GNU
+// extension __has_attribute but C++20 introduces even better alternative:
+// standard-defined __has_cpp_attribute, which will remove the need for
+// defining C_HAS_* macros on a buildsystem level (at some point).
+
+// The __attribute__ syntax is supported by GCC, Clang, and IBM compilers.
+//
+// TODO: C++11 introduces standard syntax for implementation-defined attributes,
+//       it should allow for removal of C_HAS_ATTRIBUTE from the buildsystem.
+//       However, the vast majority of GCC_ATTRIBUTEs in DOSBox code need
+//       to be reviewed, as many of them seem to be incorrectly/unnecessarily
+//       used.
+
+#if C_HAS_ATTRIBUTE
+#define GCC_ATTRIBUTE(x) __attribute__ ((x))
+#else
+#define GCC_ATTRIBUTE(x) /* attribute not supported */
+#endif
+
 // GCC_LIKELY macro is incorrectly named, because other compilers support
 // this feature as well (e.g. Clang, Intel); leave it be for now, at
 // least until full support for C++20 [[likely]] attribute will start arriving

src/platform/visualc/config.h

@@ -63,8 +63,6 @@
 /* Define to 1 if you want serial passthrough support. */
 #define C_DIRECTSERIAL 1
 
-#define GCC_ATTRIBUTE(x) /* attribute not supported */
-
 #define INLINE __forceinline
 #define DB_FASTCALL __fastcall