s390/preempt: Optimize __preempt_count_dec_and_test()

Provide an inline assembly using alternatives to avoid the need of a
base register due to relocatable lowcore when adding or subtracting
small constants from preempt_count.

Main user is preempt_enable(), which subtracts one from preempt_count
and tests if the result is zero.

With this the generated code changes from

  1000b8:       a7 19 00 00             lghi    %r1,0
  1000bc:       eb ff 13 a8 00 6e       alsi    936(%r1),-1
  1000c2:       a7 54 00 05             jnhe    1000cc <__rcu_read_unlock+0x14>

to something like this:

  1000b8:       eb ff 03 a8 00 6e       alsi    936,-1
  1000be:       a7 54 00 05             jnhe    1000c8 <__rcu_read_unlock+0x10>

Kernel image size is reduced by 45kb (bloat-o-meter -t, defconfig, gcc15).

Reviewed-by: Sven Schnelle <svens@linux.ibm.com>
Signed-off-by: Heiko Carstens <hca@linux.ibm.com>
This commit is contained in:
Heiko Carstens 2026-01-13 20:44:00 +01:00
parent 05405b8fd2
commit 48b4790f05

View file

@ -113,7 +113,22 @@ static __always_inline void __preempt_count_sub(int val)
*/
static __always_inline bool __preempt_count_dec_and_test(void)
{
#ifdef __HAVE_ASM_FLAG_OUTPUTS__
unsigned long lc_preempt;
int cc;
lc_preempt = offsetof(struct lowcore, preempt_count);
asm_inline(
ALTERNATIVE("alsi %[offzero](%%r0),%[val]\n",
"alsi %[offalt](%%r0),%[val]\n",
ALT_FEATURE(MFEATURE_LOWCORE))
: "=@cc" (cc), "+m" (((struct lowcore *)0)->preempt_count)
: [offzero] "i" (lc_preempt), [val] "i" (-1),
[offalt] "i" (lc_preempt + LOWCORE_ALT_ADDRESS));
return (cc == 0) || (cc == 2);
#else
return __atomic_add_const_and_test(-1, &get_lowcore()->preempt_count);
#endif
}
/*