Currently we use the link register to branch up high in the early MMU on syscall entry path. Unfortunately, this trashes the link stack as the address we are going to is not associated with the earlier mflr.
This patch simply converts us to used the count register (volatile over syscalls anyway) instead. This is much better at predicting in this scenario and doesn't trash link stack causing a bunch of additional branch mispredicts later. Benchmarking this on POWER8 saves a bunch of cycles on Anton's null syscall benchmark here: http://ozlabs.org/~anton/junkcode/null_syscall.c Signed-off-by: Michael Neuling <mi...@neuling.org> diff --git a/arch/powerpc/kernel/exceptions-64s.S b/arch/powerpc/kernel/exceptions-64s.S index a8a5361..87ef8f5 100644 --- a/arch/powerpc/kernel/exceptions-64s.S +++ b/arch/powerpc/kernel/exceptions-64s.S @@ -74,13 +74,13 @@ END_FTR_SECTION_IFSET(CPU_FTR_REAL_LE) \ mflr r10 ; \ ld r12,PACAKBASE(r13) ; \ LOAD_HANDLER(r12, system_call_entry_direct) ; \ - mtlr r12 ; \ + mtctr r12 ; \ mfspr r12,SPRN_SRR1 ; \ /* Re-use of r13... No spare regs to do this */ \ li r13,MSR_RI ; \ mtmsrd r13,1 ; \ GET_PACA(r13) ; /* get r13 back */ \ - blr ; + bctr ; #else /* We can branch directly */ #define SYSCALL_PSERIES_2_DIRECT \ _______________________________________________ Linuxppc-dev mailing list Linuxppc-dev@lists.ozlabs.org https://lists.ozlabs.org/listinfo/linuxppc-dev
