s390/preempt: Optimize preempt_count()

s390/preempt: Optimize preempt_count()

Provide an inline assembly using alternatives to avoid the need of
a base register when reading preempt_count() from lowcore. Use the
LLGT instruction, which reads only the least significant 31 bits of
preempt_count. This masks out the encoded PREEMPT_NEED_RESCHED bit.

Generated code is changed from

000000000046e5d0 <vfree>:
  46e5d0:       c0 04 00 00 00 00       jgnop   46e5d0 <vfree>
  46e5d6:       a7 39 00 00             lghi    %r3,0
  46e5da:       58 10 33 a8             l       %r1,936(%r3)
  46e5de:       c0 1b 00 ff ff 00       nilf    %r1,16776960
  46e5e4:       a7 74 00 11             jne     46e606 <vfree+0x36>

to something like this:

000000000046e5d0 <vfree>:
  46e5d0:       c0 04 00 00 00 00       jgnop   46e5d0 <vfree>
  46e5d6:       e3 10 03 a8 00 17       llgt    %r1,936
  46e5dc:       ec 41 28 b7 00 55       risbgz  %r4,%r1,40,55
  46e5e2:       a7 74 00 0f             jne     46e600 <vfree+0x30>

Overall savings are only 82 bytes according to bloat-o-meter. This
is because of different inlining decisions, and there aren't many
preempt_count() users in the kernel.

Reviewed-by: Sven Schnelle <svens@linux.ibm.com>
Signed-off-by: Heiko Carstens <hca@linux.ibm.com>