After looking at our system call path, Mary Brown suggested that we should put all mfspr SRR* instructions before any mtspr SRR*.
To test this I used a very simple null syscall (actually getppid) testcase at http://ozlabs.org/~anton/junkcode/null_syscall.c I tested with the following changes against the pseries_defconfig: CONFIG_VIRT_CPU_ACCOUNTING=n CONFIG_AUDIT=n to remove the overhead of virtual CPU accounting and syscall auditing. POWER6: baseline: mean = 757.2 cycles sd = 2.108 modified: mean = 759.1 cycles sd = 2.020 POWER7: baseline: mean = 411.4 cycles sd = 0.138 modified: mean = 404.1 cycles sd = 0.109 So we have 1.77% improvement on POWER7 which looks significant. The POWER6 suggest a 0.25% slowdown, but the results are within 1 standard deviation and may be in the noise. Signed-off-by: Anton Blanchard <an...@samba.org> --- Index: linux-powerpc/arch/powerpc/kernel/exceptions-64s.S =================================================================== --- linux-powerpc.orig/arch/powerpc/kernel/exceptions-64s.S 2011-05-07 10:56:04.000000000 +1000 +++ linux-powerpc/arch/powerpc/kernel/exceptions-64s.S 2011-05-09 11:20:23.872220052 +1000 @@ -211,11 +211,11 @@ END_FTR_SECTION_IFSET(CPU_FTR_REAL_LE) mr r9,r13 GET_PACA(r13) mfspr r11,SPRN_SRR0 - ld r12,PACAKBASE(r13) - ld r10,PACAKMSR(r13) - LOAD_HANDLER(r12, system_call_entry) - mtspr SPRN_SRR0,r12 mfspr r12,SPRN_SRR1 + ld r10,PACAKBASE(r13) + LOAD_HANDLER(r10, system_call_entry) + mtspr SPRN_SRR0,r10 + ld r10,PACAKMSR(r13) mtspr SPRN_SRR1,r10 rfid b . /* prevent speculative execution */ _______________________________________________ Linuxppc-dev mailing list Linuxppc-dev@lists.ozlabs.org https://lists.ozlabs.org/listinfo/linuxppc-dev
