[PATCH] D9429: Improve the cost evaluation of LSR

[Wei Mi]

wmi added a comment.

In http://reviews.llvm.org/D9429#213884, @qcolombet wrote:

> Hi Wei,
>
> What is the long-term plan here?
>
> My understanding is that you want to gather benchmark numbers and maybe change the default to the best one.
>  Is what you are aiming for?
>
> Thanks,
> -Quentin


Hi Quentin,

Sorry for not updating this one for a while. For google benchmarks and for x86 platform, aggr-estimate is better than basic mode for several benchmarks. There is no much difference between aggr-unlimited, aggr-few and aggr-estimate. My plan is to compare the performance between various modes and do another round of perf analysis based on llvm testsuite (This part has not been done), refine the logic of aggr-estimate mode if there is regression. Then if the status is satisfying, checkin the patch (if review is ok) and make the aggr-estimate mode available for wider use and further tuning (But still keep the basic mode as default until it becomes good enough on platforms other than x86).

Thanks,
Wei.

      


Repository:
  rL LLVM

http://reviews.llvm.org/D9429