nilanjana87 wrote: > We evaluated profi internally (at Google) last year. Our configuration uses > AutoFDO with [flow-sensitive discriminators > ](https://lists.llvm.org/pipermail/llvm-dev/2020-November/146694.html) > (FS-AFDO). We found slight regressions with this configuration and didn't > enable it for our workloads. With FS-AFDO we see performance improvements of > ~1%. Can you share your configuration and performance numbers?
Our configuration uses SamplePGO with non-FS discriminators. We experimented with some benchmarks & 2 real-world workloads. Our experiments show better perf gains for 9/13 cases with this flag enabled, with 3 of them showing > 1% extra gain. On the downside of using this flag, we have seen >1% net loss in perf. gain for 3 benchmarks. For the real-world use-cases, we typically see 7-11% perf gain with SamplePGO & ~1% improvement on top of that with this flag enabled. My argument in favor of enabling this flag by default is that increasing coverage of samples is ideally a good thing. If increased coverage shows reduced performance, then that can be addressed separately by disabling this flag for those particular cases. Empirically too, we see more cases with wins than losses on enabling this by default. https://github.com/llvm/llvm-project/pull/145957 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
