Introduce support for target_clones attribute (#3280)

This requires GLIBC 2.23+, plus either gcc 6+ or clang 14+.

- Provides build-time feature detection
- Use with (un)premultiply for ~10% perf gain on AVX CPUs
- Slightly increases binary size, so best to use sparingly

libvips/conversion/premultiply.c

@@ -121,6 +121,7 @@ G_DEFINE_TYPE( VipsPremultiply, vips_premultiply, VIPS_TYPE_CONVERSION );
 	} \
 }
 
+VIPS_TARGET_CLONES("default,avx")
 static int
 vips_premultiply_gen( VipsRegion *or, void *vseq, void *a, void *b,
 	gboolean *stop )

libvips/conversion/unpremultiply.c

@@ -174,6 +174,7 @@ G_DEFINE_TYPE( VipsUnpremultiply, vips_unpremultiply, VIPS_TYPE_CONVERSION );
 	} \
 }
 
+VIPS_TARGET_CLONES("default,avx")
 static int
 vips_unpremultiply_gen( VipsRegion *or, void *vseq, void *a, void *b,
 	gboolean *stop )

libvips/include/vips/util.h

@@ -215,6 +215,16 @@ G_STMT_START { \
  */
 #define VIPS_PATH_MAX (4096)
 
+/* Create multiple copies of a function targeted at groups of SIMD intrinsics,
+ * with the most suitable selected at runtime via dynamic dispatch.
+ */
+#ifdef HAVE_TARGET_CLONES
+	#define VIPS_TARGET_CLONES( TARGETS ) \
+		__attribute__(( target_clones( TARGETS ) ))
+#else
+	#define VIPS_TARGET_CLONES( TARGETS )
+#endif
+
 VIPS_API
 const char *vips_enum_string( GType enm, int value );
 VIPS_API

meson.build

@@ -131,6 +131,20 @@ if cpp.compiles(vector_arithmetic_check, name: 'Has vector arithmetic', dependen
     endif
 endif
 
+# HAVE_TARGET_CLONES
+target_clones_check = '''
+static int __attribute__((target_clones("default,avx")))
+has_target_clones(void) {
+    return 0;
+}
+int main(void) {
+    return has_target_clones();
+}
+'''
+if cc.compiles(target_clones_check, args: '-Werror', name: 'Has target_clones attribute')
+    cfg_var.set('HAVE_TARGET_CLONES', '1')
+endif
+
 func_names = [ 'vsnprintf', '_aligned_malloc', 'posix_memalign', 'memalign', 'cbrt', 'hypot', 'atan2', 'asinh' ]
 foreach func_name : func_names
     if cc.has_function(func_name, dependencies: m_dep)