From: Pavel Tatashin <pasha.tatashin@xxxxxxxxxx> Date: Fri, 2 Jun 2017 14:40:50 -0400 > In timer_64.c tick functions are access via pointer (tick_ops), every time > clock is read, there is one extra load to get to the function. > > This patch optimizes it, by accessing functions pointer from value. > > Current ched_clock(): > sethi %hi(0xb9b400), %g1 > ldx [ %g1 + 0x250 ], %g1 ! <tick_ops> > ldx [ %g1 ], %g1 > call %g1 > nop > sethi %hi(0xb9b400), %g1 > ldx [ %g1 + 0x300 ], %g1 ! <timer_ticks_per_nsec_quotient> > mulx %o0, %g1, %g1 > rett %i7 + 8 > srlx %g1, 0xa, %o0 > > New sched_clock(): > sethi %hi(0xb9b400), %g1 > ldx [ %g1 + 0x340 ], %g1 > call %g1 > nop > sethi %hi(0xb9b400), %g1 > ldx [ %g1 + 0x378 ], %g1 > mulx %o0, %g1, %g1 > rett %i7 + 8 > srlx %g1, 0xa, %o0 > > Before three loads, now two loads. > > Signed-off-by: Pavel Tatashin <pasha.tatashin@xxxxxxxxxx> > Reviewed-by: Shannon Nelson <shannon.nelson@xxxxxxxxxx> > Reviewed-by: Steven Sistare <steven.sistare@xxxxxxxxxx> For the tick read itself itt's probably time for code patching, and thus taking the number of loads down to one. It's not that hard, the largest code sequence is for hummingbird which is about 13 or 14 instructions. So just make a tick_get_tick() assembler function that's a software trap instruction and then 14 or 15 nops. Patch in the correct assembler into this function as early as possible, and then just call it directly instead of using the ops. We do this for so many things (TLB flushes etc.) so there are many examples to cull the implementation from. In the end we'll get your early boot timestamps for free, load count wise. Thanks. -- To unsubscribe from this list: send the line "unsubscribe sparclinux" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
