[S390] convert/optimize csum_fold() to C

In the meantime gcc generates better code than the old inline
assemblies do. Original inline assembly results in:

lr	%r1,%r2
sr	%r3,%r3
lr	%r2,%r1
srdl	%r2,16
alr	%r2,%r3
alr	%r1,%r2
srl	%r1,16
xilf	%r1,65535
llghr	%r2,%r1
br	%r14

Out of the C code gcc generates this:

rll	%r1,%r2,16
ar	%r1,%r2
srl	%r1,16
xilf	%r1,65535
llghr	%r2,%r1
br	%r14

In addition we don't have any static register allocations anymore and
gcc is free to shuffle instructions around for better pipeline usage.

Signed-off-by: Heiko Carstens <heiko.carstens@de.ibm.com>
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>
This commit is contained in:
Heiko Carstens 2009-09-11 10:28:32 +02:00 committed by Martin Schwidefsky
parent 05e7ff7da7
commit 04efc3be76
1 changed files with 4 additions and 21 deletions

View File

@ -78,28 +78,11 @@ csum_partial_copy_nocheck (const void *src, void *dst, int len, __wsum sum)
*/
static inline __sum16 csum_fold(__wsum sum)
{
#ifndef __s390x__
register_pair rp;
u32 csum = (__force u32) sum;
asm volatile(
" slr %N1,%N1\n" /* %0 = H L */
" lr %1,%0\n" /* %0 = H L, %1 = H L 0 0 */
" srdl %1,16\n" /* %0 = H L, %1 = 0 H L 0 */
" alr %1,%N1\n" /* %0 = H L, %1 = L H L 0 */
" alr %0,%1\n" /* %0 = H+L+C L+H */
" srl %0,16\n" /* %0 = H+L+C */
: "+&d" (sum), "=d" (rp) : : "cc");
#else /* __s390x__ */
asm volatile(
" sr 3,3\n" /* %0 = H*65536 + L */
" lr 2,%0\n" /* %0 = H L, 2/3 = H L / 0 0 */
" srdl 2,16\n" /* %0 = H L, 2/3 = 0 H / L 0 */
" alr 2,3\n" /* %0 = H L, 2/3 = L H / L 0 */
" alr %0,2\n" /* %0 = H+L+C L+H */
" srl %0,16\n" /* %0 = H+L+C */
: "+&d" (sum) : : "cc", "2", "3");
#endif /* __s390x__ */
return (__force __sum16) ~sum;
csum += (csum >> 16) + (csum << 16);
csum >>= 16;
return (__force __sum16) ~csum;
}
/*