Microbenchmarking for clrinterpreter for iOS performance

[BrzVlad]

On desktop, the closest configuration to ios would be composite r2r with interpreter fallback. The objective is to benchmark this configuration and fix any significant regressions compared to default JIT configuration.

Automated running of benchmarks

• Add BenchmarkDotNet support for composite r2r configuration Add support for composite r2r runs BenchmarkDotNet#2967
  • fix sdk issue hit by composite r2r compilation of BDN autogenerated project Prevent duplicate entries in _SourceItemsToCopyToPublishDirectory* sdk#52296
  • some benchmarks intermittently return 0 as WorkloadResult when running composite + interpreter fallback. This looks like a bug in the composite/interpreter configuration which needs further investigation.
• fix the packs build when clrinterpreter is enabled. The crossgen2 pack is necessary for BDN to consume
• add support for iOS like runtime configuration on desktop (jit disabled, cached interface dispatch, etc) Make CoreCLR + crossgen2 ioslike via --dynamiccodecompiled false #124168
• dotnet/performance integration for running the microbenchmarks in composite-r2r mode
• script support for building runtime+crossgen2 packs, send to helix and addition of pipeline in azdo. This should allow for reliable comparison with other runtime configurations

Results for running the microbenchmark suite can be obtained locally already. Above automation is nice to have but not strictly necessary for future improvements.

• ioslike composite r2r with interpreter fallback vs JIT. This configuration is meant to be representative of iOS Release scenario. 4500 regressions, mean is 5x regression. Investigate primary sources of improvement
  • most regressions are Intrinsics related, figure out mitigation
  • other low hanging fruits
• InterpMode=1 (SPC jitted, rest interpreted) vs mono interp without optimization. This configuration is meant to be representative of iOS Debug scenario. Given SPC is jitted and mono interpreter has optimizations disabled, the CoreCLR configuration is expected to be overall faster than mono.

[dotnet-policy-service]

Tagging subscribers to this area: @BrzVlad, @janvorli, @kg
See info in area-owners.md if you want to be subscribed.