[PATCH] D69437: [RAGreedy] Enable -consider-local-interval-cost by default

Wed Oct 30 19:17:02 PDT 2019

ZhangKang added a comment.

In D69437#1723646 <https://reviews.llvm.org/D69437#1723646>, @sanwou01 wrote:

> @jsji I have now committed the test case with -consider-local-interval-cost disabled, so the impact should be a bit more obvious now.
>
> As Dave points out, it is practically impossible to predict where this behaviour would come up in other architectures, so finding test cases for it is not straightforward. As much as I'd like to add tests for the other platforms, I don't have any, nor a good way to find them!
>
> @ZhangKang I'm looking forward to seeing your findings on PowerPC, thanks for running those!

@sanwou01 @jsji  , I have done the SPEC2017 test on PPC. 
I have listed the case which upgrade or downgrade more than 1%
The geomean for int base is -0.44%
523.xalancbmk_r      -4.55%
531.deepsjeng_r       -3.22%
548.exchange2_r      +2.71%

The geomean for fp base is -0.34%
507.cactuBSSN_r      -1.21%
511.povray_r              -1.7%

The geomean for int peak is 0.43%
500.perlbench_r       -1.37%
520.omnetpp_r         -3.48%
548.exchange2_r      +8.75%

The geomean for fp peak is -0.09%
507.cactuBSSN_r      +1%
508.namd_r               -1.08%
For base test, this patch has not a good performance on PPC.
For int peak, it can let the geomean improve 0.43%, for the fp peak, the result is similar.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D69437/new/

https://reviews.llvm.org/D69437