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

[Renato Golin via llvm-commits]

rengolin added a comment.

In http://reviews.llvm.org/D17433#358270, @flyingforyou wrote:

> `CortexA9Model `: cortex-a9, a12, a15, a17, krait
>  `SwiftModel `: swift, cyclone
>  `isLikeA9 `: cortex-a9,a15,krait
>
> above targets will be affected by "PredictableSelectIsExpensive".


Right, if the behaviour changes to Swift or Cyclone, than we should think better about this change. Someone form Apple should give an opinion on this before we take any action.

> I don't have access to a9/a15/a17. (What I have is only for cortex-a53/57, x86's i series)


I do. I'll run some benchmarks on A15 and will let you know. You run the benchmarks on A53/57. Feel free to run EEMBC or SPEC, if you have access, too.

@jmolloy, can you run this on an A9, just to be sure?

> And I also don't have access to swift/cyclone.


Me neither, that's why someone from Apple should try it out, too.

> But I don't know how to measure the performance with test-suite.


If you know how to run the test-suite (http://llvm.org/docs/lnt/quickstart.html), you can just add the following options to lnt:

  --threads=1 --build-threads=$CPUS --use-perf --benchmarking-only --multisample=8

Threads = 1 allows the benchmarking to be less noisy, built-threads should be on as much as the number of CPUs, use-perf if you have perf installed (linux-tools), the results are much more accurate, "benchmarking-only" is the key option here, which only runs the tests that are marked as benchmarks and multisample will run the same benchmarks N times (for ARM, you should run at least 3, but 5 is a good number, and 8 a better one).

cheers,
--renato


http://reviews.llvm.org/D17433