On 9/17/20 9:31 AM, Philippe Mathieu-Daudé wrote:
> cpu_to_endian() and endian_to_cpu() can be extended to handle
> constant expressions. That way the programmer doesn't need to
> remember the const_X() API exists.
>
> Suggested-by: Stefan Hajnoczi <stefa...@gmail.com>
> Signed-off-by: Philippe Mathieu-Daudé <phi...@redhat.com>
> ---
> include/qemu/bswap.h | 8 ++++++--
> 1 file changed, 6 insertions(+), 2 deletions(-)
>
> diff --git a/include/qemu/bswap.h b/include/qemu/bswap.h
> index de256cea3ab..8827e4760b9 100644
> --- a/include/qemu/bswap.h
> +++ b/include/qemu/bswap.h
> @@ -191,12 +191,16 @@ static inline void bswap64s(uint64_t *s)
> #define CPU_CONVERT(endian, size, type)\
> static inline type endian ## size ## _to_cpu(type v)\
> {\
> - return glue(endian, _bswap)(v, size);\
> + return __builtin_constant_p(v) ?\
> + const_ ## endian ## size(v) :\
> + glue(endian, _bswap)(v, size);\
> }\
> \
> static inline type cpu_to_ ## endian ## size(type v)\
> {\
> - return glue(endian, _bswap)(v, size);\
> + return __builtin_constant_p(v) ?\
> + const_ ## endian ## size(v) :\
> + glue(endian, _bswap)(v, size);\
> }\
> \
> static inline void endian ## size ## _to_cpus(type *p)\
>
How does this actually affect code generation?
If it does, then that's a mistake in our definition of bswap{16,32,64}(), which
should have been able to fold constants just fine.
Looking at all of that CONFIG_MACHINE_BSWAP_H stuff, I think we should just
ditch it all in favour of __builin_bswap*. The final piece, __builtin_bswap16,
came in at ac868f29d7e8 in gcc-4.8.
r~
