[mono] Re-enable HAS_CUSTOM_BLOCKS for non-amd64 Mono

[matouskozak]

We've seen a bunch of regressions on WASM, AOT-arm64 and Interp-arm64 microbenchmark after #106764. Rather than reverting the PR and loosing the gains on x64 MonoJIT, we can try limiting the change to x64.

Regressions:

• [mono][Perf] Disabled HAS_CUSTOM_BLOCKS causing regressions on MonoAOT, MonoInterpreter arm64 on 8/23/2024 12:30:35 PM #107308

[dotnet-policy-service]

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

[EgorBo]

oops, class must be public

@EgorBot -mono -arm64

using BenchmarkDotNet.Attributes;

public class Bencha
{
    byte[] data = new byte[512];

    [Benchmark]
    public void Clear() => data.AsSpan().Clear();
}

[EgorBo]

Seems like it regresses Jit-arm64 EgorBot/runtime-utils#65 (comment)

[EgorBo]

Let me try the interp:

@EgorBot -mono -arm64 -x64 --envvars MONO_ENV_OPTIONS:--interpreter

using BenchmarkDotNet.Attributes;

public class Bencha
{
    byte[] data = new byte[512];

    [Benchmark]
    public void Clear() => data.AsSpan().Clear();
}

[matouskozak]

Seems like it regresses Jit-arm64 EgorBot/runtime-utils#65 (comment)

Indeed, Thank you for running the measurements. I'm trying to get some local measurements as well and I'll investigate more. I added NO-MERGE label until we figure if this fix works or not.

[EgorBo]

I suspect that indeed can help interpeter (because 1 block copy is faster than 8 scalars) and maybe LLVM is not able to merge 8 scalars into SIMD (SLP) due to some lack of aliasing/alignment info, at least on x64

[matouskozak]

I've checked locally and this fixes interp arm64 regression on new ReadOnlySpan<Int32>(_array).CopyTo(new Span<Int32>(_destination));

However, I don't like that it regressed MonoJIT-arm64, e.g.:

Seems like it regresses Jit-arm64 EgorBot/runtime-utils#65 (comment)

I will have to investigate more where the block optimizations are made.

[xtqqczze]

Superseded by #107558.

[matouskozak]

Superseded by #107558.

Yes, sorry I forgot to close this PR.

We decided to do full revert based on the investigation in #107308 (comment). We will address the regression in #106822.