On Fri, 12 Nov 2010, Joerg Sonnenberger wrote:
> On Fri, Nov 12, 2010 at 06:35:32PM +0000, Eduardo Horvath wrote:
> > Take, for example, the simple act of byte swapping network data. x86 has
> > a bswap instruction. SPARC and PowerPC architectures have a multiplexer
> > on the load/store. (Don't know about MIPS, don't remember about ARM.) In
> > order to swap a 64-bit value in registers you need to issue a series of
> > shift, mask, and or instrutions, usually about 8 of them, most of which
> > are depenent (cannot be issued in parallel on a superscalar machine).
> > The macro NetBSD uses here is bswap64(). This maps very well to the x86
> > instruction, but is really not usable for any machine that does the
> > swapping in the load/store path. Instead we could use something like
> > load_be64()/load_le64()/store_be64()/store_le64() which could be mapped to
> > a reverse-endian load/store instruction on architetures that support it
> > and bswap on x86.
>
> If you look at sys/endian.h, that is exactly the kind of abstraction the
> API makes easy to use. Just provide the proper inline functions for your
> platform.
No it doesen't because all those macros assume the value is being
transferred from one register to another rather than regiser to memory.
The assignment:
foo.size = htole64(size);
Cannot be replaced with:
__inline __asm("stxa %1, [%0] ASI_LITLE" : &foo.size : size);
Eduardo
