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

[Gerolf Hoflehner]

Hi Hao,

thanks for all your efforts landing this. We’ll check the performance impact on cyclone. Did lto have any impact on your results?

Cheers
Gerolf

> On Nov 18, 2014, at 11:11 PM, Hao Liu <Hao.Liu at arm.com> wrote:
> 
>>> ! In D5864#54, @jingyue wrote:
>> LGTM with some minor fixes. You may want to double check with ARM folks on performance. I remember Gerolf said he couldn't get the same performance numbers as you did. 
>> 
>> Thanks again for all the hard work! 
> 
> Hi Jingyue,
> 
> Thanks very much for your hard work on review, which really help me a lot.
> I've committed the code in http://llvm.org/viewvc/llvm-project?view=revision&revision=222328 and http://llvm.org/viewvc/llvm-project?view=revision&revision=222331.
> 
> BTW, for the performance test, I also tested other revision by some slightly modifications such as:
> (1) Always lower GEPs if there is more than one index despite wheter we found a constant.
> (2) Call SeparateConstOffsetFromGEP pass before calling TargetPassConfig::addIRPasses() in AArch64TargetMachine.cpp.
> 
> The performance results are different. Some benchmarks are better, some benchmarks are worse. So I just think maybe we need further tuning in the future for better performance. Anyway, that's the future work.
> 
> Thanks,
> -Hao
> 
> http://reviews.llvm.org/D5864
> 
>