Messages in this thread Patch in this message | | | From | Sasha Levin <> | Subject | [PATCH v2] bitops.h: correctly handle rol32 with 0 byte shift | Date | Thu, 3 Dec 2015 22:04:01 -0500 |
| |
ROL on a 32 bit integer with a shift of 32 or more is undefined and the result is arch-dependent. Avoid this by handling the trivial case of roling by 0 correctly.
The trivial solution of checking if shift is 0 breaks gcc's detection of this code as a ROL instruction, which is unacceptable.
This bug was reported and fixed in GCC (https://gcc.gnu.org/bugzilla/show_bug.cgi?id=57157):
The standard rotate idiom,
(x << n) | (x >> (32 - n))
is recognized by gcc (for concreteness, I discuss only the case that x is an uint32_t here).
However, this is portable C only for n in the range 0 < n < 32. For n == 0, we get x >> 32 which gives undefined behaviour according to the C standard (6.5.7, Bitwise shift operators). To portably support n == 0, one has to write the rotate as something like
(x << n) | (x >> ((-n) & 31))
And this is apparently not recognized by gcc.
Note that this is broken on older GCCs and will result in slower ROL.
Acked-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Sasha Levin <sasha.levin@oracle.com> --- include/linux/bitops.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/include/linux/bitops.h b/include/linux/bitops.h index 2b8ed12..defeaac 100644 --- a/include/linux/bitops.h +++ b/include/linux/bitops.h @@ -107,7 +107,7 @@ static inline __u64 ror64(__u64 word, unsigned int shift) */ static inline __u32 rol32(__u32 word, unsigned int shift) { - return (word << shift) | (word >> (32 - shift)); + return (word << shift) | (word >> ((-shift) & 31)); } /** -- 2.5.0
| |