[PATCH] D17433: [ARM] fix initialization of PredictableSelectIsExpensive

Renato Golin via llvm-commits llvm-commits at lists.llvm.org
Tue Feb 23 01:19:54 PST 2016

rengolin accepted this revision.
rengolin added a comment.
This revision is now accepted and ready to land.

In http://reviews.llvm.org/D17433#359481, @jmolloy wrote:

> My twopenniworth on this: cortex-a15, a12, a17 and krait have been lazy and inherited all the behaviours of cortex-a9. Therefore it is fine to choose a behaviour for a9 that doesn't take into account those other cores. They can change if it's an issue for them.

Good point.

> I feel that this flag will cause improvements and regressions whatever way it is set. The only real way to do this is using PGO. Therefore I feel that it should be set to whatever the most appropriate setting is for the core without real regard to benchmark results. Someone's going to be unhappy either way.

I've run a quick benchmark on A15 by turning on and off that flag, and strangely, with the flag on, the overall performance is only slightly worse (0.5%), but consistently so, across all benchmarks.

Since my methodology was less than kosher, the impact is less than meaningful and the code looks better as it is than before, I think this should go in.



More information about the llvm-commits mailing list