Mono-LLVM is not tested?

[EgorBo]

From what I see we currently have two AOT runs for Mono https://pvscmdupload.blob.core.windows.net/reports/allTestHistory/TestHistoryIndexIndex.html

Both don't look like they use LLVM since I know that Mono-LLVM implements S.R.I intrinsics and is expected to benefit when we vectorize something new. Recently, I've vectorized Guid.ToString and it demonstrated visible benefits on CoreCLR:

While for Mono-AOT runs there were no changes. The only explanation for that is that IsSuppored path were not taken:

cc @naricc @fanyang-mono @DrewScoggins

[naricc]

Despite the LLVM false label, the AOT lanes for mono should both be using LLVM. We need to fix the label. Or it broke and stopped running LLVM?

[naricc]

Actually if the perf lanes some how stopped using LLVM it would explain this regression that we can't repro: dotnet/perf-autofiling-issues#11536

[EgorBo]

Maybe it used to use LLVM JIT (that was disabled in Mono some time ago)?
Anyway, the main problem in this issue is that if ((Ssse3.IsSupported || AdvSimd.Arm64.IsSupported) && BitConverter.IsLittleEndian) is never taken in Mono

[fanyang-mono]

Right now the Mono AOT lanes are configured as LLVM AOT with mini JIT fallback. If the method were AOT'ed, then you should see the performance difference. If it falls back go JIT, then there is no LLVM. LLVM JIT was turned off after we integrated LLVM14.

[EgorBo]

https://pvscmdupload.blob.core.windows.net/reports/allTestHistory%2frefs%2fheads%2fmain_arm64_ubuntu%2020.04_LLVM%3dfalse_MonoAOT%3dtrue_MonoInterpreter%3dfalse_RunKind%3dmicro_mono%2fSystem.Buffers.Text.Tests.Base64Tests.ConvertToBase64CharArray(NumberOfBytes%3a%201000).html

Base64 is expected to be vectorized, something happend with mono on Jan5 and that benchmark regressed. I assume it's the same issue. Commit range: dotnet/runtime@9197fd5...92d7381

[fanyang-mono]

@EgorBo AdvSimd.Arm64.IsSupported is true only when it is LLVM full AOT mode. Currentlhy, we are not running microbenchmarks with full aot mode.