18 January 2017

Polly gained a new ARM64 Buildbot sponsored by Qualcomm Innovation Center, which adds ARM64 platform coverage to Polly’s continious integration testing:

ARM64 Buildbot

With this buildbot Polly increases its buildbot farm to nine builders. Polly runs the following builders:

Besides the standard -O3 -mllvm -polly configuration, Polly regularly tests its ability to generate parallel code, and also runs in an “extended coverage mode” called -mllvm -polly-unprofitable which forces Polly to transform each trivial loop it encounteres to provide higher test coverage on Polly’s transformation engine by transforming even code where we do not expect any performance gains. Finally, with our Polly Performance bot, we regularly report performance results to the LLVM performance tracking infrastructure hosted at http://llvm.org/perf/.

Besides the above set of buildbots, we also host a set of “Before-Vectorizer” build bots, which keep track of our efforts to move Polly into the LLVM pass pipeline with the goal to remove the need for additional canonicalization passes when running Polly. Such passes always pre-transform the IR that is fed to the LLVM standard pass pipeline and can consequently result in random difficult to avoid performance changes. After moving Polly into the Pass pipeline, we expect Polly to only affect performance in cases where our performance model predicts a clear benefit. You can find more information about this topic in the Polly Documentation.

Polly is also regularly tested externally. Publicly accessible are the FFmpeg FATE tests. More tests are run in-house by different users.