[PATCH] [AArch64] Improve and enable the SeparateConstOffsetFromGEP for AArch64 backend.

[Hao Liu]

>>! In D5864#45, @Gerolf wrote:
> Hi Hao,
> 
> could you share your current performance data + options you use? With your patch I see small gains on astar + xalancbmk, but a ~4% loss on libquantum.  The data is based on the 11/07 trunk on top of r221549, O3 + LTO, cyclone, ref input.
> In addition to cortex, could you run on x86 also?
> 
> Thanks
> Gerolf

Hi Gerolf,

I use "-O3 -ffast-math -mcpu=cortex-a57" and ref input. My last run was on the 10/24 trunk on top of r220484 (To ignore the impact of trunk changes, I always run and compare to the specific revision). I've ran the performance test for about 4 times on this specific revision. To ignore noise and variances, I remove those benchmark who doesn't exist in every performance runs. Then I find there is no obvious regressions and have stable improvements on following 3 benchmarks (This is just results from one run, other runs have similar results):
    SPEC2006.473.astar          -4.84%
    SPEC2006.444.namd           -4.39%
    SPEC2006.445.gobmk          -2.80%

The percentage number is calculated by: 100% - current_time / reference_time.

It's quite strange that your results have many differences from mine on cyclone. libquantum is a quite stable benchmark on our test infrastructure, and I never see performance changes on it with my patch. So I always think my patch doesn't affect its performance.

I can not test cyclone, so I run on x86 with today's trunk with and without my patch (I enabled the SeparateConstOffsetFromGEP pass on x86 backend similarly to AArch64 backend). I've ran for about 8 times on libquantum. This benchmark is not stable on my x86 machine. The results have +/-1% variances comparing to the average results with or without my patch. But the average results are similar with or without my patch. So I think my patch won't affect the performance of libquantum on x86. I've ran a performance test for all the int/fp benchmarks in SPEC 2006. Just haven't got result yet. As it needs quite a long time for the result. If it is fortunate, I'll get the result tomorrow and show to you.
So I wonder whether libquantum is stable on cyclone. If it is not stable, maybe you can run it several times to ignore the variance. Or if it is real regression, we can then investigate it.

Do you have more details about the results. I'm interested in how many improvements on astar. Also I'm interested in why there is no improvement on namd and gobmk. They are supposed to get some benefits from this patch.

Thanks,
-Hao

http://reviews.llvm.org/D5864