[ARM] 3252/1: help gcc do the best with ___arch__swab32

Patch from Nicolas Pitre

Depending on your gcc version, the current C-only implementation would
produce suboptimal code, ranging from a bad register selection forcing
an additional mov instruction to a failure to merge the eor and the ror
in a single instruction.  With a little help gcc always produces the
best code.

Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
This commit is contained in:
Nicolas Pitre 2006-01-10 19:48:02 +00:00 committed by Russell King
parent b016450f9f
commit a3e4943686

View file

@ -22,7 +22,16 @@ static inline __attribute_const__ __u32 ___arch__swab32(__u32 x)
{
__u32 t;
t = x ^ ((x << 16) | (x >> 16)); /* eor r1,r0,r0,ror #16 */
if (__builtin_constant_p(x)) {
t = x ^ ((x << 16) | (x >> 16)); /* eor r1,r0,r0,ror #16 */
} else {
/*
* The compiler needs a bit of a hint here to always do the
* right thing and not screw it up to different degrees
* depending on the gcc version.
*/
asm ("eor\t%0, %1, %1, ror #16" : "=r" (t) : "r" (x));
}
x = (x << 24) | (x >> 8); /* mov r0,r0,ror #8 */
t &= ~0x00FF0000; /* bic r1,r1,#0x00FF0000 */
x ^= (t >> 8); /* eor r0,r0,r1,lsr #8 */