include/arch/x86/swab.h badness?

["Paweł Sikora" <pluto@agmk.net>]

Hi,

during compiling some byte-order related code for 32-bit x86 which uses glibc's
#include <byteorder.h> i've noticed that __be*_to_cpu uses a generic i386 (ror+xchg) impl.
quick investigation shows that #include <linux/asm/swab.h> sucks a little ;)

please look at the current linux implementation:

static __inline__ __u32 __arch_swab32(__u32 val)
{
#ifdef __i386__
# ifdef CONFIG_X86_BSWAP
        __asm__("bswap %0" : "=r" (val) : "0" (val));
# else
        __asm__("xchgb %b0,%h0\n\t"     /* swap lower bytes     */
            "rorl $16,%0\n\t"   /* swap words           */
            "xchgb %b0,%h0"     /* swap higher bytes    */
            : "=q" (val)
            : "0" (val));
# endif

#else /* __i386__ */
        __asm__("bswapl %0"
            : "=r" (val)
            : "0" (val));
#endif
        return val;
}

in the userspace, the __i386__ is always defined by gcc for any -march=i486/i586/...
and the CONFIG_X86_BSWAP is never defined, so the userspace always ends with inlined
assembly with generic i386 implementation :(

could you please improve this #ifdefs to get bswap opcodes for i486+ ?
maybe the __builtin_bswap* should be used in recent kernel tree?

BR,
Paweł.