[release/6.0] Fix ReadyToRun loading on Apple Silicon

[jgiannuzzi]

Customer Impact

Slow startup of .NET applications on Apple Silicon. ReadyToRun images are not loaded correctly on Apple Silicon. All methods are JITed during startup.

For reference, compiling a medium-size project like ILGPU takes ~17s on a Mac Mini M1 before this change, and ~13s after this change (about 30% faster).

Testing

crossgen2 outerloop tests. Manually verified that the ReadyToRun images are loaded correctly on Apple Silicon.

Risk

Low risk. These fixes have been in dotnet/runtime:main for more than a month, however there is a risk of discovering latent Apple Silicon specific bugs in ReadyToRun with this fix.

---

Fixes #64103.

This PR aims at backporting a subset of the changes introduced in #61938 and #62855. The rationale behind this change is explained in more details in #64103. I believe that this change would be limited to osx-arm64 and introduce a substantial performance gain by enabling ReadyToRun on Apple Silicon, which I imagine was supposed to work in .NET 6 in the first place.

For reference, compiling a medium-size project like ILGPU takes ~17s on a Mac Mini M1 before this change, and ~13s after this change (about 30% faster). Given the focus on performance in .NET 6, I would argue that this has to be fixed in .NET 6.

[dotnet-issue-labeler]

I couldn't figure out the best area label to add to this PR. If you have write-permissions please help me learn by adding exactly one area label.

[dnfadmin]

All CLA requirements met.

[VSadov]

Changes in map and layout look good. Thanks!

[trylek]

LGTM but my experience with Apple Silicon is very limited, I'd feel more comfortable if JanV could take a look as well.

[jkotas]

LGTM! Thank you

[jgiannuzzi]

I am not sure whether the current test failures are due to my changes or not.
@jkotas @trylek @VSadov do you have any pointers to help me solve this?

[janvorli]

The JIT\Directed\rvastatics\RVAOrderingTest is failing on both windows and x64 / arm64 Linux and OSX.
The Loader\binding\tracing\BinderTracingTest.Basic is failing on Windows x64 / arm64 only.
However, the change doesn't touch any Windows and Linux behavior (the changes are under ifdefs so that they occur only on macOS and one of them on Linux too), so it is not clear to me how it could cause these failures.

[jgiannuzzi]

I have tried running the JIT/Directed/rvastatics/RVAOrderingTest test locally with CompositeBuildMode=1 and RunCrossGen2=1.
On Linux x64, Linux arm64, and macOS x64, they fail with or without my changes (e.g. 6de5c5b – which my branch is based on – fails, but also bbfc15c – which is the current tip of release/6.0).
On macOS arm64, they only fail after 6e24b2d is applied (my commit enabling R2R loading), which makes sense as without this commit the composite R2R dll does not get loaded and the test is then equivalent to not having RunCrossGen2=1 set.
I did not try on Windows, but I expect the results would be similar.

[jgiannuzzi]

Update: the same test also fails on main (869ae1a at the time of writing) and on v6.0.0 on Linux arm64 (and potentially more platforms, but I only tested that one).
There seems to be a bug in composite crossgen2 that copies only the referenced static data, but not the rest. It results in having only s_First, s_Last, and s_1 (referenced as s_Another1) being copied, leaving a bunch of data between s_First and s_Last out.

@jkotas @trylek @VSadov @janvorli what is the best course of action for this particular PR? Is there anything more that is needed to get it merged?
I would love to see it part of the next release so that I can finally setup an osx-arm64 CI pipeline for ParquetSharp using a macOS VM!

[jkotas]

what is the best course of action for this particular PR? Is there anything more that is needed to get it merged?

I have filled the request to get it approved for .NET 6 serving. Thank you for your help with this fix!

[jeffschwMSFT]

Approved. We will take for consideration in 6.0.x.

cc @kdubau and @safern

[kdubau]

Thanks @jeffschwMSFT, as of right now we have no immediate plan to enable R2R in VS4Mac. cc @slluis FYI

[jeffschwMSFT]

@kdubau unless the r2r is stripped from NetCoreApp you would benefit from this change as those packages have r2r enabled.

[safern]

@trylek are these crossgen2 failures related? Can we merge this?

[jeffschwMSFT]

@safern I asked the same question. These failures are unrelated, we just ran more outerloops. It is good to merge.

[safern]

Thanks for confirming.