svn commit: r233684 - head/sys/x86/include

[Andrey Chernov]

On Fri, Mar 30, 2012 at 02:11:21PM +0200, Dimitry Andric wrote:
> In case of the __bswapXX() macros, you can see that the argument to
> __bswapXX_gen() is first explicitly cast to an unsigned type, for
> example with __bswap32():
> 
> #define __bswap32(x)                    \
>          (__builtin_constant_p(x) ?      \
>              __bswap32_gen((__uint32_t)(x)) : __bswap32_var(x))
> 
> Therefore, right shifting will not give any problems.  For the call to
> __bswap32_var(), such casting is not needed, since the argument will be
> implicitly converted to __uint32_t.
> 
> As Bruce has mentioned, you could add more explicit casts and additional
> parentheses, but those would be superfluous.

I mention right shift just as potential problem. I really want to target 
casting vs. masking problem (masking provide more instructions).

bde@ already suggest the same thing I generally mean in more details in 
his recent answer with the patch:

% +         __bswap16_gen((__uint16_t)(x)) : __bswap16_var((__uint16_t)(x))))

-- 
http://ache.vniz.net/