On 16.11.2022 20:39, Eric Biggers wrote: > On Wed, Nov 16, 2022 at 04:52:51PM +0300, Alexey Khoroshilov wrote: >> be32_to_cpu() macro in some cases may be expanded to an expression >> that evaluates its arguments multiple times. > > When is that, exactly? I considered the path #define be32_to_cpu __be32_to_cpu #define __be32_to_cpu(x) __swab32((__force __u32)(__be32)(x)) #define __swab32(x) \ (__builtin_constant_p((__u32)(x)) ? \ ___constant_swab32(x) : \ __fswab32(x)) #define ___constant_swab32(x) ((__u32)( \ (((__u32)(x) & (__u32)0x000000ffUL) << 24) | \ (((__u32)(x) & (__u32)0x0000ff00UL) << 8) | \ (((__u32)(x) & (__u32)0x00ff0000UL) >> 8) | \ (((__u32)(x) & (__u32)0xff000000UL) >> 24))) But you are right, it is protected by __builtin_constant_p() that prevents that for memory access. > If that's true, then lots of other places in the kernel would need to be fixed > too. Try running: > > git grep -E '[bl]e(16|32|64)_to_cpu\([^)]+\+\+\)' > > If true, then it would be much better to fix the macros. And yes, anyway it makes sense to keep for be32_to_cpu()-like function macro a contract to handle its arguments as a plain function. Thanks, Alexey